content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
\label{sec:Introduction}
In this paper we investigate the Farrell--Jones Conjecture for Waldhausen's $A$--theory. Our main result is
\begin{theorem}[Main result]
\label{the:main_result}
Let ${\mathcal F}\!{\mathcal J}_A$ be the class of groups for which the Farrell--Jones~Conjecture \ref{con:FJC_for_A-theory_with_c_and_fwp}
for (non-connective) $A$--theory with coefficients and finite wreath products holds.
\begin{enumerate}
\item \label{the:main_result:groups}
The class ${\mathcal F}\!{\mathcal J}_A$ contains the following groups:\
\begin{itemize}
\item Hyperbolic groups
\item $\mathrm{CAT}(0)$--groups
\item Virtually poly-cyclic groups
\item Cocompact lattices in almost connected Lie groups
\item Fundamental groups of (not necessarily compact) $d$--dimensional manifolds (possibly with boundary) for $d \le 3$
\end{itemize}
\item \label{the:main_result:inheritance} The class ${\mathcal F}\!{\mathcal J}_A$ has the following inheritance properties:
\begin{itemize}
\item If $G_1$ and $G_2$ belong to
${\mathcal F}\!{\mathcal J}_A$, then $G_1 \times G_2$ and $G_1 \ast G_2$ belong to ${\mathcal F}\!{\mathcal J}_A$.
\item If $H$ is a subgroup of $G$ and
$G \in {\mathcal F}\!{\mathcal J}_A$, then $H \in {\mathcal F}\!{\mathcal J}_A$.
\item Let $1 \to K \to G \xrightarrow{p} Q \to 1$ be an extension of groups. Suppose
that $K$, $Q$ and $p^{-1}(C)$ for every infinite cyclic subgroup $C \subseteq Q$ belong
to ${\mathcal F}\!{\mathcal J}_A$. Then $G$ belongs to ${\mathcal F}\!{\mathcal J}_A$.
\item If $H \subseteq G$ is a subgroup of $G$ with $[G:H] < \infty$ and $H \in {\mathcal F}\!{\mathcal J}_A$, then
$G \in {\mathcal F}\!{\mathcal J}_A$.
\item Let $\{G_i \mid i\in I\}$ be a
directed system of groups (with not necessarily injective structure maps)
such that $G_i \in {\mathcal F}\!{\mathcal J}_A$ for $i \in I$.
Then $\operatorname{colim}_{i \in I} G_i$ belongs to ${\mathcal F}\!{\mathcal J}_A$.
\end{itemize}
\end{enumerate}
\end{theorem}
The Farrell--Jones Conjecture for $A$--theory aims at the computation of the homotopy groups
of $\bfA(BG)$ for a group $G$, where $\bfA \colon \curs{SPACES} \to \curs{SPECTRA}$ sends a space $X$
to the non-connective $A$--theory spectrum $\bfA(X)$ modeling Waldhausen's $A$--theory space $A(X)$.
More precisely, it predicts the bijectivity of the assembly map
\[
H_n^G(\EGF{G}{{\mathcal V}{\mathcal C}{\mathcal Y}};\bfA^B) \xrightarrow{\cong} H_n^G(G/G;\bfA^B) = \pi_n(\bfA(BG))
\]
induced by the projection of the classifying space $\EGF{G}{{\mathcal V}{\mathcal C}{\mathcal Y}}$ for the family of virtually cyclic subgroups of $G$
to $G/G$. It essentially reduces the computation of $\pi_n(\bfA(BG))$ to the computation of the system $\{\pi_n(\bfA(BV))\}$,
where $V$ ranges over the virtually cyclic subgroups $V$ of $G$.
Following the setup of Davis--L\"uck~\cite{Davis-Lueck(1998)}, we give the precise formulations of the various versions
of the Farrell--Jones Conjecture in \Cref{sec:The_Isomorphism_Conjecture}.
The equivalent original formulation of the Farrell--Jones Conjecture can be found in Farrell--Jones~\cite{Farrell-Jones(1993a)}.
\Cref{sec:Relations_between_the_various_theories} relates the Farrell--Jones Conjecture for $A$--theory
to the corresponding conjectures for other functors.
In particular, we discuss the equivalence of the conjectures for $\bfA$, ${\mathbf W}{\mathbf h}^{\operatorname{CAT}}$ and $\bfP^{\operatorname{CAT}}$,
where the latter denote the non-connective spectra modeling the Whitehead space and the space of stable pseudo-isotopies,
with $\operatorname{CAT}$ being $\operatorname{TOP}$, $\operatorname{PL}$ or $\operatorname{DIFF}$.
As an illustration about the impact of the Farrell--Jones Conjecture for $A$--theory we discuss
applications to the automorphism groups of aspherical closed manifolds in
\Cref{subsec:Topological_automorphism_groups_of_aspherical_closed_manifolds}, where also the proof
of the following \Cref{the:widetilde(TOP)(M)/TOP(M)_and_bfN(A)} is given.
Let ${\mathbf{NA}}(\{\bullet\})$ be the Nil-term occurring in the Bass--Heller--Swan-isomorphisms for non-connective $A$--theory,
see~\cite{Huettemann-Klein-Waldhausen-Williams(2001),Huettemann-Klein-Vogell-Waldhausen-Williams(2002)}.
\begin{equation}
\pi_n(\bfA(S^1)) = \pi_n(\bfA(\{\bullet\})) \oplus \pi_{n-1}(\bfA(\{\bullet\})) \oplus \pi_n({\mathbf{NA}}(\{\bullet\})) \oplus \pi_n({\mathbf{NA}}(\{\bullet\})).
\label{Bass-Heller_Swan_for_A-theory}
\end{equation}
We conclude
$\pi_n({\mathbf{NA}}(\{\bullet\})) = \{0\}$ for $n \le 1$ and $\pi_n({\mathbf{NA}}(\{\bullet\})) \otimes_{{\mathbb Z}} \IQ = \{0\}$ for $n \in {\mathbb Z}$
from \Cref{the:Relating_A-theory_to_algebraic_K_theory} and~\cite[Theorem~0.3]{Lueck-Steimle(2015splitasmb)}.
On the other hand, $\pi_n({\mathbf{NA}}(\{\bullet\}))$
for $n = 2,3$ is an infinite-dimensional ${\mathbb F}_2$--vector space. For more information about $\pi_n({\mathbf{NA}}(\{\bullet\}))$
we refer to~\cite{Grunewald-Klein-Macko(2008),Hesselholt(2009)}.
The next result is already explained in the special case of closed manifolds with negative sectional curvature
in~\cite[Section~6.3]{Weiss-Williams(2001)}, based on the work of
Farrell and Jones~\cite{Farrell-Jones(1990c),Farrell-Jones(1991comp),Farrell-Jones(1990b),Farrell-Jones(1991),Farrell-Jones(1993a)},
and we can extend it to torsionfree hyperbolic groups.
\begin{theorem}\label{the:widetilde(TOP)(M)/TOP(M)_and_bfN(A)}
\
\begin{enumerate}
\item\label{the:widetilde(TOP)(M)/TOP(M)_and_bfN(A)-it1} Let $G$ be a torsionfree hyperbolic group. Then we get an equivalence
\[ {\mathbf W}{\mathbf h}^{\operatorname{TOP}}(BG) \simeq \bigvee_C {\mathbf W}{\mathbf h}^{\operatorname{TOP}}(BC) \simeq \bigvee_C {\mathbf{NA}}(\{\bullet\}) \vee {\mathbf{NA}}(\{\bullet\}), \]
where $C$ ranges over the conjugacy classes of maximal infinite cyclic subgroups of $G$.
In particular, ${\mathbf W}{\mathbf h}^{\operatorname{TOP}}(BG)$ is connective.
\item\label{the:widetilde(TOP)(M)/TOP(M)_and_bfN(A)-it2} Let $M$ be a smoothable aspherical closed manifold of dimension $\ge 10$,
whose fundamental group $\pi$ is hyperbolic.
Then there is a $\mathbb{Z}/2$--action on ${\mathbf W}{\mathbf h}^{\operatorname{TOP}}(B\pi)$ such that we obtain for $1 \le n \leq \min\{( \dim M -7 ) / 2, (\dim M - 4)/3\}$ isomorphisms
\[
\pi_n(\operatorname{TOP}(M)) \cong \pi_{n+2}\Big( E\mathbb{Z}/2_+ \wedge_{\mathbb{Z}/2} \big( \bigvee_C {\mathbf W}{\mathbf h}^{\operatorname{TOP}}(BC) \big) \Big)
\]
and an exact sequence
\[
1 \to \pi_{2}\Big( E\mathbb{Z}/2_+ \wedge_{\mathbb{Z}/2} \big( \bigvee_C {\mathbf W}{\mathbf h}^{\operatorname{TOP}}(BC) \big) \Big) \to\pi_0(\operatorname{TOP}(M)) \to \operatorname{Out}(\pi) \to 1,
\]
where $C$ ranges over the conjugacy classes of maximal infinite cyclic subgroups of $\pi$.
\end{enumerate}
\end{theorem}
\begin{remark}
The $\mathbb{Z}/2$--action we refer to in \Cref{the:widetilde(TOP)(M)/TOP(M)_and_bfN(A)} is induced by the one given in \cite{Weiss-Williams(1988)}.
Vogell described in \cite{Vogell(1984)} another $\mathbb{Z}/2$--action on ${\mathbf W}{\mathbf h}(B\pi)$ which depends on the choice of a spherical fibration over $B\pi$.
It was shown in \cite{Huettemann-Klein-Vogell-Waldhausen-Williams(2002)} that for a certain fibration this action corresponds under the Bass--Heller--Swan decomposition to switching the Nil-terms via a homeomorphism.
If the arguments presented in \cite{Huettemann-Klein-Vogell-Waldhausen-Williams(2002)} carry over to other spherical fibrations and the two actions on ${\mathbf W}{\mathbf h}^{\operatorname{TOP}}(B\pi)$ agree for a suitably chosen fibration, then the homotopy orbits appearing in \Cref{the:widetilde(TOP)(M)/TOP(M)_and_bfN(A)} can be identified as
\[
E\mathbb{Z}/2_+ \wedge_{\mathbb{Z}/2} \big( \bigvee_C {\mathbf W}{\mathbf h}^{\operatorname{TOP}}(BC) \big) \simeq \bigvee_C {\mathbf{NA}}(\{\bullet\}).
\]
These issues will be discussed in \cite{PieperPhD}.
\end{remark}
The rest of the paper is devoted to the proof of \Cref{the:main_result}.
The main technical part of this paper concerns the proof for hyperbolic groups and $\mathrm{CAT}(0)$--groups.
It is given in \Cref{Proof of the Farrell--Jones Conjecture for hyperbolic and CAT(0)--groups} and \Cref{The Transfer: Final Part of the Proof}, and is motivated by the proof of the $K$--theoretic Farrell--Jones Conjecture for $\mathrm{CAT}(0)$--groups given in \cite{Wegner(2012)} based on the method of Bartels--L\"uck \cite{Bartels-Lueck(2012annals)}. Our approach, which is based on~\cite{Ullmann-Winges(2015)}, requires us to define an analog of the transfer on geometric modules which works on Waldhausen categories of controlled retractive spaces. Virtually poly-cyclic groups have already been treated by Ullmann--Winges~\cite{Ullmann-Winges(2015)}.
In conjunction with the inheritance properties in \Cref{the:main_result}~\ref{the:main_result:inheritance},
the case of a cocompact lattice in an almost connected Lie group or a fundamental group of a (not necessarily compact) $d$--dimensional manifold (possibly with boundary) for $d \le 3$
follows via the argument presented in ~\cite{Bartels-Farrell-Lueck(2014)}.
The inheritance properties for the $A$--theoretic conjecture are taken care of in~\Cref{sec:Inheritance_properties_of_the_Isomorphism_Conjectures}.
\begin{remark}[Solvable groups] \label{rem_Solvable_groups} If one can show that
${\mathcal F}\!{\mathcal J}_A$ contains all virtually solvable groups, then it contains any (not necessarily cocompact)
lattice in a second countable locally compact Hausdorff group with finitely many path
components, the groups $GL_n(\IQ)$ and $GL_n(F(t))$ for $F(t)$ the function field over a
finite field $F$, and all $S$--arithmetic groups. The arguments in
\cite{Kammeyer-Lueck-Rueping(2016),Rueping(2016_S-arithmetic)} carry over to show the prerequisites of \Cref{the:main_result} and \Cref{cor:hypcatfollowproof}, respectively.
\end{remark}
\subsection*{Acknowledgments.}
The idea to use the categories of ``cellwise $0$-controlled morphisms'' in
\Cref{The Transfer: Final Part of the Proof} is due to Arthur Bartels
and Paul Bubenzer.
The paper is financially supported by the Leibniz-Preis of
the second author granted by the {DFG}, the ERC Advanced Grant ``KL2MG-interactions''
(no. 662400) of the second author granted by the European Research Council,
the Cluster of Excellence ``Hausdorff Center for Mathematics'' at Bonn, the SFB 647 ``Space -- Time -- Matter'' at Berlin, the SFB 878 ``Groups, Geometry \& Actions'' at M\"unster,
and the Junior Hausdorff Trimester Program ``Topology'' at the Hausdorff Research Institute for Mathematics (HIM).
\tableofcontents
\section{The Isomorphism Conjecture}
\label{sec:The_Isomorphism_Conjecture}
In this section we state various versions of the Isomorphism Conjectures we
want to consider. We assume familiarity with the notion of a $G$-equivariant
homology theory from~\cite{Davis-Lueck(1998)} and the notion of an equivariant
homology theory from~\cite{Lueck(2002b)}. As usual, we use a convenient
category of compactly generated spaces.
\subsection{The Meta-Isomorphism Conjecture for functors from spaces to spectra}
\label{subsec:The_Meta_Isomorphism_Conjecture_for_functors_from_spaces_to_spectra}
Let ${\mathbf S} \colon \curs{SPACES}\to \curs{SPECTRA}$ be a covariant functor. Throughout this section we
will assume that \emph{${\mathbf S}$ respects weak equivalences and disjoint unions}, i.e., a weak
homotopy equivalence of spaces $f \colon X \to Y$ is sent to a weak homotopy equivalence of spectra ${\mathbf S}(f) \colon {\mathbf S}(X) \to {\mathbf S}(Y)$
and for a collection of spaces $\{X_i \mid i \in I\}$
for an arbitrary index set $I$ the canonical map
\begin{equation}
\bigvee_{i \in I} {\mathbf S}(X_i) \to {\mathbf S}\biggl(\coprod_{i \in I} X_i\biggr)
\end{equation}
is a weak homotopy equivalence of spectra. Weak equivalences of spectra are
understood to be the stable equivalences, i.e., the maps which induce
isomorphisms on all stable homotopy groups.
We obtain a covariant functor
\begin{equation}
{\mathbf S}^B \colon \curs{GROUPOIDS} \to \curs{SPECTRA}, \quad {\mathcal G} \mapsto {\mathbf S}(B{\mathcal G}),
\label{bfs_upper_B}
\end{equation}
where $B{\mathcal G}$ is the classifying space of the groupoid ${\mathcal G}$ which is the geometric
realization of the simplicial set given by its nerve and denoted by
$B^{\operatorname{bar}}{\mathcal G}$ in~\cite[page~227]{Davis-Lueck(1998)}. Let
$H^?_n(-;{\mathbf S}^B)$ be the equivariant homology theory in the sense
of~\cite[Section~1]{Lueck(2002b)} which is associated to ${\mathbf S}^B$ by the construction
in~\cite[Proposition~157 on page~796]{Lueck-Reich(2005)}. Equivariant homology theory
essentially means that we get for every group $G$ a $G$--homology theory $H^G_n(-;{\mathbf S}^B)$
satisfying the disjoint union axiom and for every group homomorphism $\alpha \colon H \to G$
and $H$--CW--pair $(X,A)$ we get natural maps compatible with boundary homomorphisms of
pairs $H^H_*(X,A;{\mathbf S}^B) \to H^G_*(\alpha_*(X,A);{\mathbf S}^B)$ which are bijective if the kernel of
$\alpha$ acts freely on $X$. Moreover, for any group $G$, subgroup $H \subseteq G$ and $n \in {\mathbb Z}$
we have canonical identifications
\begin{equation*}
H_n^G(G/H;{\mathbf S}^B) \cong H_n^H(H/H;{\mathbf S}^B) \cong \pi_n({\mathbf S}(BH)).
\end{equation*}
\begin{conjecture}[Meta-Isomorphism Conjecture for functors from spaces to spectra]
\label{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra}
Let ${\mathbf S} \colon \curs{SPACES}\to \curs{SPECTRA}$ be a covariant functor which
respects weak equivalences and disjoint unions. The group $G$ satisfies the
\emph{Meta-Isomorphism Conjecture for ${\mathbf S}$}
with respect to the family ${\mathcal F}$ of subgroups of $G$
if the assembly map induced by the projection $\operatorname{pr} \colon \EGF{G}{{\mathcal F}} \to G/G$
\[
H_n^G(\operatorname{pr};{\mathbf S}^B) \colon H^G_n(\EGF{G}{{\mathcal F}};{\mathbf S}^B) \to H_n^G(G/G;{\mathbf S}^B) \cong \pi_n({\mathbf S}(BG))
\]
is bijective for all $n \in {\mathbb Z}$.
\end{conjecture}
\begin{example}[The $K$-- and $L$--theoretic Farrell--Jones Conjecture]
\label{exa:The_K-and_L-theoreticFJC}
Let $R$ be a ring (with involution). There are covariant
functors~\cite[Theorem 158]{Lueck-Reich(2005)}
\begin{eqnarray*}
\bfK_R \colon \curs{GROUPOIDS} & \to & \curs{SPECTRA},
\\
\bfL_R^{\langle -\infty \rangle} \colon \curs{GROUPOIDS} & \to & \curs{SPECTRA},
\end{eqnarray*}
such that for every group $G$, which we can consider as a groupoid $\underline{G}$ with precisely one
object and $G$ as its group of automorphisms, and $n \in {\mathbb Z}$ we have
\begin{eqnarray*}
K_n(RG)
& = &
\pi_n(\bfK_R(\underline{G})),
\\
L_n^{\langle -\infty \rangle} (RG)
& = &
\pi_n(\bfL^{\langle -\infty \rangle} _R(\underline{G})).
\end{eqnarray*}
Then the $K$--theoretic and $L$--theoretic Farrell--Jones Conjectures, which
were originally
formulated in~\cite[1.6 on page 257]{Farrell-Jones(1993a)}, are equivalent to the
statement that the covariant functors ${\mathbf S} \colon \curs{SPACES} \to \curs{SPECTRA}$, given by the
composition of $\bfK_R$ and$\bfL_R$, with the functor sending a space to its fundamental
groupoid, satisfy the Meta-Isomorphism
Conjecture~\ref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra} for
the family ${\mathcal V}{\mathcal C}{\mathcal Y}$ of virtually cyclic subgroups of $G$.
\end{example}
Our main example is the following case. Let $\bfA \colon \curs{SPACES}\to \curs{SPECTRA}$
be the functor sending a space $X$ to the
spectrum $\bfA(X)$ given by the non-connective version
of Waldhausen's algebraic $K$--theory of spaces in the sense of \cite{Ullmann-Winges(2015)}.
\begin{lemma}\label{lem:bfA_respects_weak_equivalences_and_disjoint_unions}\
\begin{enumerate}
\item \label{lem:bfA_respects_weak_equivalences_and_disjoint_unions:dis_union}
The functor $\bfA \colon \curs{SPACES}\to \curs{SPECTRA}$ respects weak equivalences
and disjoint unions.
\item \label{lem:bfA_respects_weak_equivalences_and_disjoint_unions:hocolim}
For any directed systems of spaces $\{X_i \mid i \in I\}$ indexed over an arbitrary
directed set $I$ the canonical map
\[
\operatorname{hocolim}_{i \in I} \bfA(X_i) \to \bfA\bigl(\operatorname{hocolim}_{i \in I} X_i\bigr)
\]
is a weak homotopy equivalence.
\end{enumerate}
\end{lemma}
\begin{proof}
In the connective case, Waldhausen proved in~\cite[Proposition~2.1.7]{Waldhausen(1985)}) that $A$--theory preserves weak equivalences. The other two properties follow upon inspection of the explicit model as finite retractive CW--complexes.
Note that the algebraic $K$--theory functor which sends $X$ to $K({\mathbb Z}\Pi(X))$ enjoys the properties claimed for $\bfA$. Since Vogell showed that the linearization map $\bfL \colon \bfA \rightarrow \bfK$ induces an isomorphism on all non-positive homotopy groups \cite{Vogell(1991)}, the general case follows.
It will be shown in \cite{PieperPhD} that the non-connective deloopings described by Ullmann--Winges in~\cite{Ullmann-Winges(2015)} and Vogell in~\cite{Vogell(1990)} are equivalent.
\end{proof}
\begin{conjecture}[The Farrell--Jones Conjecture for $A$--theory]
\label{con:FJC_for_A-theory}
A group $G$ satisfies the Farrell--Jones Conjecture for $A$--theory if the
Meta-Isomorphism
Conjecture~\ref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra}
holds for $\bfA \colon \curs{SPACES}\to \curs{SPECTRA}$ and the family ${\mathcal V}{\mathcal C}{\mathcal Y}$, i.e., for
every $n \in {\mathbb Z}$ the projection $\EGF{G}{{\mathcal V}{\mathcal C}{\mathcal Y}} \to G/G$ induces an isomorphism
\[
H_n^G(\operatorname{pr};\bfA^B) \colon H^G_n(\EGF{G}{{\mathcal V}{\mathcal C}{\mathcal Y}};\bfA^B) \to H_n^G(G/G;\bfA^B) = \pi_n(\bfA(BG)).
\]
\end{conjecture}
\subsection{The Meta--Isomorphism Conjecture for functors from spaces to spectra with coefficients}
\label{subsec:The_Meta_Isomorphism_Conjecture_for_functors_from_spaces_to_spectra_with_c}
Let $G$ be a group and $Z$ be a $G$--CW--complex. Define a covariant $\curs{Or}(G)$--spectrum
\begin{equation}
{\mathbf S}^G_Z \colon\curs{Or}(G) \to \curs{SPECTRA},
\quad G/H \mapsto {\mathbf S}(G/H \times_G Z),
\label{bfS:upper_G_Z}
\end{equation}
where $G/H \times_G Z$ is the orbit space of the diagonal $G$--action on $G/H \times Z$.
Notice that there is an obvious homeomorphism $G/H \times_G Z \xrightarrow{\cong} Z/H$.
Denote by $H_n^G(-;{\mathbf S}^G_Z)$ the $G$--homology theory in the sense of~\cite[Section~1]{Lueck(2002b)} which is associated to ${\mathbf S}^G_Z$
by the construction of~\cite[Proposition~156 on page~795]{Lueck-Reich(2005)} and satisfies
$H_n^G(G/H;{\mathbf S}^G_Z) \cong \pi_n({\mathbf S}^G_Z(G/H)) = \pi_n({\mathbf S}(Z/H))$ for any homogeneous $G$--space $G/H$ and $n \in {\mathbb Z}$.
\begin{conjecture}[Meta-Isomorphism Conjecture for functors from spaces to
spectra with coefficients]
\label{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c}
Let ${\mathbf S} \colon \curs{SPACES}\to \curs{SPECTRA}$ be a covariant functor which respects
weak equivalences and disjoint unions. The group $G$ satisfies the
\emph{Meta-Isomorphism Conjecture for ${\mathbf S}$ with coefficients}
\index{Conjecture!Meta-Isomorphism Conjecture for functors from spaces to
spectra with coefficients} with respect to the family ${\mathcal F}$ of subgroups
of $G$ if for any free $G$--CW--complex $Z$ the assembly map
\[
H^G_n(\operatorname{pr};{\mathbf S}^G_Z) \colon H^G_n(\EGF{G}{{\mathcal F}};{\mathbf S}^G_Z) \to
H_n^G(G/G;{\mathbf S}^G_Z) = \pi_n({\mathbf S}(Z/G)),
\]
induced by the projection $\operatorname{pr} \colon \EGF{G}{{\mathcal F}} \to G/G$, is bijective
for all $n \in {\mathbb Z}$.
\end{conjecture}
\begin{example}[$Z = EG$] \label{exa:Z_is_EG} If we take $Z = EG$ in
\Cref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c},
then
\Cref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c}
reduces to
\Cref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra}.
Namely, for a $G$--set $S$ let ${\mathcal T}^G(S)$ be its \emph{transport groupoid} whose set of objects is $S$,
the set of morphisms from $s_1$ to $s_2$ is the set $\{g \in G \mid s_2 = gs_1\}$ and composition comes from the
multiplication in $G$. There is a homotopy equivalence $B{\mathcal T}^G(G/H) \xrightarrow{\simeq} G/H \times_G EG$
which is natural in $G/H$. Hence we get a weak homotopy equivalence of $\curs{Or}(G)$--spectra
${\mathbf S}^B({\mathcal T}^G(G/?)) \xrightarrow{\simeq} {\mathbf S}^G_{EG} $. It induces an isomorphism of $G$--homology theories,
see~\cite[Lemma~4.6]{Davis-Lueck(1998)}
\[
H_*^G(-;{\mathbf S}^B) \xrightarrow{\cong} H_*^G(-;{\mathbf S}^G_{EG}).
\]
\end{example}
\begin{remark}[Relation to the original formulation]
\label{rem:relation_to_original}
In~\cite[Section~1.7 on page~262]{Farrell-Jones(1993a)} Farrell and Jones formulate a
fibered version of their conjectures for a covariant functor ${\mathbf S} \colon \curs{SPACES} \to \curs{SPECTRA}$ for
every (Serre) fibration $\xi \colon Y \to X$ over a
connected CW-complex $X$. In our set-up this corresponds to choosing $Z$ to be the
total space of the fibration obtained from $Y \to X$ by pulling back along the universal
covering $\widetilde{X} \to X$. This space $Z$ is a free $G$--CW--complex for $G= \pi_1 (X)$.
Note that every free $G$--CW--complex $Z$ can always be obtained in this fashion
from the fiber bundle $EG \times_G Z \to BG$ up to $G$--homotopy,
compare~\cite[Corollary~2.2.1 on page~263]{Farrell-Jones(1993a)}.
We sketch the proof of this identification. Let $A$ be a $G$--CW--complex. Let
$f \colon {\mathcal E}(X) \to X$ be the map obtained by taking the quotient of the
$G = \pi_1(X)$--action on the $G$--map $A \times \widetilde{X} \to \widetilde{X}$ given by the
projection. Denote by $\widehat{p} \colon {\mathcal E}(\xi) \to {\mathcal E}(X)$ the pullback of
$\xi$ with $f$. Let $q\colon {\mathcal E}(\xi) \to A/G$ be the composite of
$\widehat{p}$ with the map ${\mathcal E}(X) \to A/G$ induced by the projection $A \times \widetilde{X} \to A$.
This is a stratified fibration and one can consider the spectrum
${\mathbb H}(A/G;{\mathcal S}(q))$ in the sense of Quinn~\cite[Section~8]{Quinn(1979a)}. Put
\[
H_n^G(A;\xi) := \pi_n({\mathbb H}(A/G;{\mathcal S}(q))).
\]
The projection $\operatorname{pr} \colon A \to G/G$ induces a map
\begin{equation}
a(A) \colon {\mathbb H}(A/G;{\mathcal S}(q)) \to {\mathbb H}(G/G;{\mathcal S}(Y\to G/G)) = {\mathbf S}(Y),
\label{assembly_a_la_Quinn}
\end{equation}
which is the assembly map in~\cite[Section~1.7 on page~262]{Farrell-Jones(1993a)} if we take $A = \EGF{G}{{\mathcal V}{\mathcal C}{\mathcal Y}}$. The construction of $H_n^G(A;\xi) := {\mathbb H}(A/G;{\mathcal S}(q))$ is very complicated, but,
fortunately, for us only two facts are relevant. We obtain a $G$--homology
theory $H_n^G(-;\xi)$ and for every $H \subseteq G$ we get a natural
identification $H_n^G(G/H;\xi) = {\mathbf S}^G_Z(G/H)$. Hence the
functor $G\text{-}\curs{CW}\text{-}\curs{COMPLEXES} \to \curs{SPECTRA}$ given by $A \mapsto {\mathbb H}(A/G;{\mathcal S}(q))$ is
weakly excisive and its restriction to $\curs{Or}(G)$ is the functor ${\mathbf S}^G_Z$. We conclude
from~\cite[Theorem~6.3]{Davis-Lueck(1998)} that the map~\eqref{assembly_a_la_Quinn} can
be identified with the map induced by the projection $A \to G/G$
\[
H_n^G(A;{\mathbf S}^G_Z) \to H_n^G(G/G;{\mathbf S}^G_Z) = \pi_n({\mathbf S}(Z/G)) = \pi_n({\mathbf S}(Y)),
\]
which appears in Meta-Isomorphism Conjecture~\ref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c}
for functors from spaces to spectra with coefficients.
\end{remark}
\begin{remark}[The condition free is necessary in
\Cref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c}]
The \label{rem:The_condition_free_is_necessary_in_MC_spaces_spectra_with_coeff}
\Cref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c}
is only true very rarely if we drop the condition that $Z$ is free. Take for instance $Z = G/G$.
Then
\Cref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c}
predicts that the projection $\EGF{G}{{\mathcal F}}/G \to G/G$ induces for all $n \in {\mathbb Z}$ an
isomorphism
\[
H_n(\operatorname{pr};{\mathbf S}(\{\bullet\})) \colon H_n(\EGF{G}{{\mathcal F}}/G;{\mathbf S}(\{\bullet\})) \to H_n(\{\bullet\},{\mathbf S}(\{\bullet\}))
\]
where $H_*(-;{\mathbf S}(\{\bullet\}))$ is the (non-equivariant) homology theory associated to the
spectrum ${\mathbf S}(\{\bullet\})$. This statement is in general wrong, except in extreme cases such
as ${\mathcal F} = {\mathcal A} {\mathcal L}{\mathcal L}$.
\end{remark}
\begin{conjecture}[The Farrell--Jones Conjecture for $A$--theory with coefficients]
\label{con:FJC_for_A-theory_with_c}
A group $G$ satisfies the \emph{Farrell--Jones Conjecture for $A$--theory with coefficients} if the
Meta-Isomorphism Conjecture~\ref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c}
with coefficients
holds for $\bfA \colon \curs{SPACES} \to \curs{SPECTRA} $ and the family ${\mathcal V}{\mathcal C}{\mathcal Y}$,
i.e., for
every $n \in {\mathbb Z}$ and free $G$--CW--complex $Z$ the projection $\EGF{G}{{\mathcal V}{\mathcal C}{\mathcal Y}} \to G/G$ induces an isomorphism
\[
H_n^G(\operatorname{pr};\bfA^G_Z) \colon H^G_n(\EGF{G}{{\mathcal V}{\mathcal C}{\mathcal Y}};\bfA^G_Z)
\to H_n^G(G/G;\bfA^G_Z) = \pi_n(\bfA(Z/G)).
\]
\end{conjecture}
\subsection{The Meta--Isomorphism Conjecture for functors from spaces to spectra with coefficients and finite wreath products}
\label{subsec:The_Meta_Isomorphism_Conjecture_for_functors_from_spaces_to_spectra_with_c_and_fwp}
There are also versions with finite wreath products.
Recall that for groups $G$ and $F$ their \emph{wreath product $G\wr F$} is defined to be the semi-direct product
$\left(\prod_F G \right) \rtimes F$, where $F$ acts on $\prod_F G$ by permuting the factors.
Fix a class of groups ${\mathcal C}$ which is closed under
isomorphisms, taking subgroups and taking quotients. Examples are the classes
${\mathcal F}{\mathcal I}{\mathcal N}$ and ${\mathcal V}{\mathcal C}{\mathcal Y}$ of finite and of virtually cyclic groups.
For a group $G$ define the family of subgroups ${\mathcal C}(G) := \{ K \subseteq G, K \in {\mathcal C}\}$.
\begin{conjecture}[The Meta-Isomorphism Conjecture for functors from spaces to spectra
with coefficients and finite wreath products]
\label{con:The_Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c_and_fwp}
Let ${\mathbf S} \colon \curs{SPACES}\to \curs{SPECTRA}$ be a covariant functor which respects weak
equivalences and disjoint unions. The group $G$ satisfies the \emph{Meta-Isomorphism
Conjecture with coefficients and finite wreath products} for the functor
${\mathbf S} \colon \curs{SPACES} \to \curs{SPECTRA}$ with respect to the class ${\mathcal C}$ of groups,
if, for any finite group $F$, the wreath product $G \wr F$ satisfies the
Meta-Isomorphism Conjecture~\ref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c}
with coefficients for the functor ${\mathbf S} \colon \curs{SPACES} \to \curs{SPECTRA}$ with respect to the family ${\mathcal C}(G \wr F)$ of subgroups of $G$.
\end{conjecture}
\begin{conjecture}[The Farrell--Jones Conjecture for $A$--theory with coefficients and finite wreath products]
\label{con:FJC_for_A-theory_with_c_and_fwp} A group $G$ satisfies
the \emph{Farrell--Jones Conjecture for $A$--theory with coefficients and finite
wreath products} if the Meta-Isomorphism
Conjecture~\ref{con:The_Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c_and_fwp}
with coefficients and finite wreath products holds for $\bfA \colon \curs{SPACES} \to \curs{SPECTRA} $ and the class ${\mathcal V}{\mathcal C}{\mathcal Y}$ of virtually cyclic groups.
\end{conjecture}
The next two lemmas will be needed later.
\begin{lemma}\label{lem:bfS_homology_theory}
Let ${\mathbf E}$ be a spectrum such that ${\mathbf S} \colon \curs{SPACES} \to \curs{SPECTRA}$ is given by $Y \mapsto Y_+ \wedge {\mathbf E}$.
\begin{enumerate}
\item \label{lem:bfS_homology_theory:pr}
Then for any group $G$, any $G$--CW--complex $X$ which is contractible (after forgetting
the $G$--action), and any free $G$--CW--complex $Z$ the projection $X \to G/G$ induces for
all $n \in {\mathbb Z}$ an isomorphism
\[
H_n^G(X;{\mathbf S}^G_Z) \xrightarrow{\cong} H_n^G(G/G;{\mathbf S}^G_Z).
\]
\item \label{lem:bfS_homology_theory:MCwc}
\Cref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra},
\Cref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c}
and \Cref{con:The_Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c_and_fwp}
hold for such an ${\mathbf S}$ for every group $G$ and every family ${\mathcal F}$ of subgroups of $G$.
\end{enumerate}
\end{lemma}
An ${\mathbf S}$ given by $Y \mapsto Y_+ \wedge {\mathbf E}$ is a homology theory, and thus
the lemma states that the conjectures hold for homology theories.
\begin{proof}~\ref{lem:bfS_homology_theory:pr}
There are natural isomorphisms of spectra
\begin{eqnarray*}
\lefteqn{\operatorname{map}_G((G/?),X)_+ \wedge_{\curs{Or}(G)} \bigl((G/? \times_G Z)_+ \wedge {\mathbf E}\bigr)}
\\
& \xrightarrow{\cong} &
\bigl((\operatorname{map}_G((G/?),X) \times_{\curs{Or}(G)} G/?) \times_G Z\bigr)_+ \wedge {\mathbf E}
\\
& \xrightarrow{\cong} &
(X \times_G Z)_+ \wedge {\mathbf E},
\end{eqnarray*}
where the second isomorphism comes from the $G$--homeomorphism
\[
\operatorname{map}_G((G/?),X) \times_{\curs{Or}(G)} G/? \xrightarrow{\cong} X
\]
of~\cite[Theorem~7.4~(1)]{Davis-Lueck(1998)}. Since
$Z$ is a free $G$--CW--complex and $X$ is contractible (after forgetting the group action),
the projection $X \times_G Z \to G/G \times_G Z$ is a homotopy equivalence and hence
induces a weak homotopy equivalence
\[
(X \times_G Z)_+ \wedge {\mathbf E} \xrightarrow{\simeq} (G/G \times_G Z)_+ \wedge {\mathbf E},
\]
Thus we get a weak homotopy equivalence
\[
\operatorname{map}_G((G/?),X)_+ \wedge_{\curs{Or}(G)} \bigl((G/? \times_G Z)_+ \wedge {\mathbf E}\bigr)
\to (G/G \times_G Z)_+ \wedge {\mathbf E}.
\]
Under the identifications coming from the definitions
\begin{eqnarray*}
H_n^G(X;{\mathbf S}^G_Z)
& = &
\pi_n\left(\operatorname{map}_G((G/?),X)_+ \wedge_{\curs{Or}(G)} \bigl((G/? \times_G Z)_+ \wedge {\mathbf E}\bigr)\right),
\\
H_n^G(G/G;{\mathbf S}^G_Z)
& = &
\pi_n\left((G/G \times_G Z)_+ \wedge {\mathbf E}\right),
\end{eqnarray*}
this weak homotopy equivalence induces on homotopy groups the isomorphism
$H_n^G(X;{\mathbf S}^G_Z) \to H_n^G(G/G;{\mathbf S}^G_Z)$.
\\[1mm]~\ref{lem:bfS_homology_theory:MCwc} This follows from assertion~\ref{lem:bfS_homology_theory:pr}.
\end{proof}
\begin{lemma}
\label{lem:cofibration_sequences_of_bfS}
Let ${\mathbf S}, \bfT, {\mathbf U} \colon \curs{SPACES}\to \curs{SPECTRA}$ be covariant functors which respect
weak equivalences and disjoint unions. Let ${\mathbf i} \colon {\mathbf S} \to \bfT$ and $\bfp \colon
\bfT \to {\mathbf U}$ be natural transformations such that for any space $Y$ the sequence of spectra
${\mathbf S}(Y) \xrightarrow{{\mathbf i}(Y)} \bfT(Y) \xrightarrow{\bfp(Y)} {\mathbf U}(Y)$ is up to weak
homotopy equivalence a cofiber sequence of spectra.
\begin{enumerate}
\item \label{lem:cofibration_sequences_of_bfS:long_exact_sequence}
Then we obtain for every group $G$ and all $G$--CW--complexes $X$ and $Z$ a
natural long exact sequence
\begin{multline*}
\cdots \to H_n^G(X;{\mathbf S}_Z^G) \to H_n^G(X;\bfT_Z^G) \to H_n^G(X;{\mathbf U}_Z^G)
\\
\to H_{n-1}^G(X;{\mathbf S}_Z^G) \to H_{n-1}^G(X;\bfT_Z^G) \to H_{n-1}^G(X;{\mathbf U}_Z^G) \to \cdots \;.
\end{multline*}
\item \label{lem:cofibration_sequences_of_bfS:MCwc}
Let $G$ be a group and ${\mathcal F}$ be a family of subgroups of $G$.
Then Meta-Isomorphism Conjecture~\ref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra}
for functors from spaces to spectra holds for all three functors ${\mathbf S}$, $\bfT$ and ${\mathbf U}$ for $(G,{\mathcal F})$
if it holds for two of the functors ${\mathbf S}$, $\bfT$ and ${\mathbf U}$ for $(G,{\mathcal F})$.
The analogous statement is true for the Meta-Isomorphism
Conjecture~\ref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c}
for functors from spaces to spectra with coefficients and for the
Meta-Isomorphism Conjecture~\ref{con:The_Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c_and_fwp}
for functors from spaces to spectra with coefficients and finite wreath products.
\end{enumerate}
\end{lemma}
\begin{proof}~\ref{lem:cofibration_sequences_of_bfS:long_exact_sequence}
The version for spectra of~\cite[Theorem~3.11]{Davis-Lueck(1998)} implies that we obtain, up to weak homotopy equivalence, a cofiber sequence of spectra
\begin{multline*}
\operatorname{map}_G(G/?,X)_+ \wedge_{\curs{Or}(G)} {\mathbf S}(G/? \times_{G} Z) \to \operatorname{map}_G(G/?,X)_+ \wedge_{\curs{Or}(G)} \bfT(G/? \times_{G} Z)
\\
\to \operatorname{map}_G(G/?,X)_+ \wedge_{\curs{Or}(G)} {\mathbf U}(G/? \times_{G} Z).
\end{multline*}%
and passing to its associated long exact sequence of homotopy groups yields the result.
\ref{lem:cofibration_sequences_of_bfS:MCwc}
This follows from assertion~\ref{lem:cofibration_sequences_of_bfS:long_exact_sequence}
and the Five-Lemma.
\end{proof}
\section{Relations between the conjectures for various theories}
\label{sec:Relations_between_the_various_theories}
There are other prominent covariant functors $\curs{SPACES} \to \curs{SPECTRA}$ which
respect weak homotopy equivalences and disjoint unions. Notice in the sequel
that we are always considering the non-connective versions. We are thinking of
the stable pseudo-isotopy spectrum $\bfP^{\operatorname{CAT}}$ and the Whitehead spectrum
${\mathbf W}{\mathbf h}^{\operatorname{CAT}}$, where $\operatorname{CAT}$ can be the topological category $\operatorname{TOP}$, the
$PL$--category $\operatorname{PL}$ or the smooth category $\operatorname{DIFF}$. For the
definition of $\bfP^{\operatorname{CAT}}$
we refer to~\cite{EnkelmannPhD, PieperPhD, Weiss-Williams(1988)}.
Usually, the Whitehead spectrum is defined as a connective spectrum,
see~\cite{Jahren-Rognes-Waldhausen(2013)}, and see \cite[Section~2.2]{Weiss-Williams(1993)} for a definition of the classical assembly
map. We make the obvious generalization.
\begin{dfn}\label{def: non-conn whspectrum}
Let $\operatorname{CAT}$ be $\operatorname{TOP}$ or $\operatorname{PL}$. The \emph{topological or piecewise-linear
non-connective Whitehead spectrum} ${\mathbf W}{\mathbf h}^{\operatorname{CAT}}(X)$ is the homotopy cofiber
of the classical assembly map in non-connective $A$--theory:
\[
X_+ \wedge \bfA(\{\bullet\}) \to \bfA(X) \to {\mathbf W}{\mathbf h}^{\operatorname{CAT}}(X).
\]
Further, we define the \emph{smooth non-connective Whitehead spectrum}
${\mathbf W}{\mathbf h}^{\operatorname{DIFF}}(X)$ as the homotopy cofiber of the sequence
\[
\Sigma^{\infty} X_+ \to \bfA(X) \to {\mathbf W}{\mathbf h}^{\operatorname{DIFF}}(X)
\]
where $\Sigma^{\infty} X_+ \to \bfA(X)$ factors as the unit map $\Sigma^\infty X_+ = X_+\wedge\mathbb{S} \to X_+ \wedge \bfA(\{\bullet\})$ and assembly.
\end{dfn}
\begin{theorem}[Relations between the various functors]
\label{the:relations_between_the_various_functors}
\
\begin{enumerate}
\item \label{the:relations_between_the_various_functors:calp_PL_and_wh_PL}
There is a zigzag of natural equivalences,
\[
\bfP^{\operatorname{CAT}} \stackrel{\simeq}{\longleftrightarrow} \Omega^2{\mathbf W}{\mathbf h}^{\operatorname{CAT}},
\]
where $\operatorname{CAT}$ can be taken to be $\operatorname{TOP}$, $\operatorname{PL}$, or $\operatorname{DIFF}$.
\item \label{the:relations_between_the_various_functors:wh_TOP_and_wh_PL}
The canonical map
\[
\bfP^{\operatorname{PL}} \xrightarrow{\simeq} \bfP^{\operatorname{TOP}}
\]
is a natural equivalence.
\end{enumerate}
\end{theorem}
\begin{proof}
The connective, objectwise case of \ref{the:relations_between_the_various_functors:calp_PL_and_wh_PL} follows from the equivalence $\bfP(M)\simeq \Omega^2{\mathbf W}{\mathbf h}^{PL}(M)$, which was originally stated in~\cite{Waldhausen(1978)} and fully proved in~\cite[Theorem~0.2]{Jahren-Rognes-Waldhausen(2013)}.
There are some issues concerning the full functoriality of pseudo-isotopy, which will be clarified in \cite{EnkelmannPhD} and \cite{PieperPhD}. The full statement will be established in \cite{PieperPhD}.
The objectwise version of \ref{the:relations_between_the_various_functors:wh_TOP_and_wh_PL} has been shown in \cite{Burghelea-Lashof(1974)}. An argument for the full statement will be given in \cite{EnkelmannPhD}.
\end{proof}
\begin{lemma} \label{lem:equivalence_of_FJ_for_the_various_functors}
If the Meta-Isomorphism
Conjecture~\ref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra}
for functors from spaces to spectra holds for the group $G$ and the family ${\mathcal F}$
for one of the functors $ \bfA$, ${\mathbf W}{\mathbf h}^{\operatorname{TOP}}$, ${\mathbf W}{\mathbf h}^{\operatorname{PL}}$, ${\mathbf W}{\mathbf h}^{\operatorname{DIFF}}$,
$\bfP^{\operatorname{TOP}}$, $\bfP^{\operatorname{PL}}$, and $\bfP^{\operatorname{DIFF}}$, then it holds for all of them.
The analogous statement holds for the Meta-Isomorphism
Conjecture~\ref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c}
for functors from spaces to spectra with coefficients and for the Meta-Isomorphism
Conjecture~\ref{con:The_Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c_and_fwp}
for functors from spaces to spectra with coefficients and finite wreath products.
\end{lemma}
\begin{proof}
This follows from \Cref{lem:bfS_homology_theory} and \Cref{lem:cofibration_sequences_of_bfS}.
\end{proof}
\begin{remark}[The non-connective spectrum of stable $h$--cobordisms]
\label{rem:stable_space_of_h-cobordism}
There is also the non-connective stable $h$--cobordism spectrum ${\mathbf H}^{\operatorname{CAT}}(M)$ of a compact manifold (possibly with boundary) $M$. Note that $h$--cobordisms are (usually) only defined as a functor in codimension zero embeddings. As such, they are related to the previous functors. For every compact manifold $M$ (possibly with boundary) there are natural weak homotopy equivalences
\[
{\mathbf H}^{\operatorname{CAT}}(M) \xrightarrow{\simeq} \Omega{\mathbf W}{\mathbf h}^{\operatorname{CAT}}(M)
\]
and
\[
\bfP^{\operatorname{CAT}}(M)\xrightarrow{\simeq} \Omega{\mathbf H}^{\operatorname{CAT}}(M).
\]
For the proof and more information we refer to \cite{Jahren-Rognes-Waldhausen(2013)}.
\end{remark}
Finally, we explain the relationship between $A$--theory and algebraic $K$--theory of integral group rings.
For a space $X$ denote its fundamental groupoid by $\Pi(X)$. There is a so called \emph{linearization map}, natural in $X$,
\begin{eqnarray}
\bfL(X) \colon \bfA(X) & \to & \bfK({\mathbb Z} \Pi_1(X))
\label{bfl_bfA_to_bfK}
\end{eqnarray}
The next result follows combining~\cite[Section~4]{Vogell(1991)}
and~\cite[Proposition~2.2 and Proposition~2.3]{Waldhausen(1978)}.
\begin{theorem}[Connectivity of the linearization map]
\label{the:Connectivity_of_the_linearization_map}
Let $X$ be a CW--complex. Then:
\begin{enumerate}
\item \label{the:Connectivity_of_the_linearization_map:2_connected}
The linearization map $\bfL(X)$ of~\eqref{bfl_bfA_to_bfK} is $2$--connected, i.e., the map
\[
L_n(X) := \pi_n(\bfL(X)) \colon A_n(X) \to K_n({\mathbb Z} \Pi(X))
\]
is bijective for $n \le 1$ and surjective for $n = 2$.
\item \label{the:Connectivity_of_the_linearization_map:rational}
The map $L_n$ is rationally bijective for all $n \in {\mathbb Z}$, provided that each component of $X$ is aspherical.
\end{enumerate}
\end{theorem}
This implies that the $K$--theoretic Farrell--Jones Conjecture for ${\mathbb Z} G$ and the $A$--theoretic Farrell--Jones
Conjecture for $\bfA(BG)$ are equivalent in degree $\le 1$ and rationally equivalent
in all degrees. More precisely, we have
\begin{theorem}[Relating A-theory to algebraic $K$--theory]
\label{the:Relating_A-theory_to_algebraic_K_theory}
Consider a group $G$ and a family ${\mathcal F}$ of subgroups of $G$.
The linearization map~\eqref{bfl_bfA_to_bfK} and the projection $\EGF{G}{{\mathcal F}} \to G/G$ yield
a commutative diagram
\[
\xymatrix{H_n^G(\EGF{G}{{\mathcal F}};\bfA^B) \ar[r] \ar[d]
&
H_n^G(G/G;\bfA^B) = A_n(BG) \ar[d]
\\
H_n^G(\EGF{G}{{\mathcal F}};\bfK_{{\mathbb Z}}) \ar[r]
&
H_n^G(G/G;\bfK_{{\mathbb Z}}) = K_n({\mathbb Z} G)
}
\]
where $\bfK_{{\mathbb Z}} \colon \curs{GROUPOIDS} \to \curs{SPECTRA}$ has been recalled in
\Cref{exa:The_K-and_L-theoreticFJC}. The vertical arrows are
bijective for $n \le 1$ and surjective for $n =2$. They are rationally
bijective for all $n \in {\mathbb Z}$.
\end{theorem}
\section{Some applications to automorphism groups of aspherical closed manifolds}
\label{sec:Some_applications_to_automorphism_groups_of_and_bundles_over_aspherical_closed_manifolds}
Before we begin with the proof of \Cref{the:main_result}, we want to illustrate
the impact of the Farrell--Jones Conjecture by discussing automorphism
groups of aspherical closed manifolds. For rational computations the Farrell--Jones
Conjecture for $K$--theory and $L$--theory suffices. For potential integral computations one
needs the Farrell--Jones Conjecture for $A$--theory and for $L$--theory. More details
about automorphism groups of closed manifolds can be found in~\cite{Weiss-Williams(2001)}.
\subsection{Topological automorphism groups of aspherical closed manifolds}
\label{subsec:Topological_automorphism_groups_of_aspherical_closed_manifolds}
Let $\operatorname{TOP}(M)$ be the topological group of self-homeomorphisms of the closed manifold $M$.
Denote by $\operatorname{G}(M)$ the monoid of self homotopy equivalences $M \to M$. Let $\widetilde{\operatorname{TOP}}(M)$ and
$\widetilde{\operatorname{G}}(M)$ be the block versions, see \cite[p.~168]{Weiss-Williams(2001)} for a survey and further references. There are natural maps making the diagram
\[
\xymatrix{\operatorname{TOP}(M) \ar[d] \ar[r]
&
\widetilde{\operatorname{TOP}}(M) \ar[d]
\\
\operatorname{G}(M) \ar[r]
&
\widetilde{G}(M)
}
\]
commute.
Define $\widetilde{\operatorname{TOP}}(M)/\operatorname{TOP}(M)$, $\widetilde{G}(M)/\operatorname{TOP}(M)$ and $\operatorname{G}(M)/\operatorname{TOP}(M)$ to be the homotopy fibers of the maps
$ B\!\operatorname{TOP}(M) \to B\widetilde{\operatorname{TOP}}(M)$, $B\!\operatorname{TOP}(M) \to B\widetilde{\operatorname{G}}(M)$ and $B\!\operatorname{TOP}(M) \to B\!\operatorname{G}(M)$.
We obtain a commutative diagram with horizontal fiber sequences
\[
\xymatrix{
\widetilde{\operatorname{TOP}}(M)/\operatorname{TOP}(M) \ar[r] \ar[d]
&
B\!\operatorname{TOP}(M) \ar[r] \ar[d]^{\operatorname{id}} &
\ar[d] B\widetilde{\operatorname{TOP}}(M)
\\
\widetilde{\operatorname{G}}(M)/\operatorname{TOP}(M) \ar[r]
&
B\!\operatorname{TOP}(M) \ar[r]
&
B\widetilde{\operatorname{G}}(M)
\\
\operatorname{G}(M)/\operatorname{TOP}(M) \ar[r] \ar[u]
&
B\!\operatorname{TOP}(M) \ar[r] \ar[u]_{\operatorname{id}}
&
B\!\operatorname{G}(M). \ar[u]
}
\]
According to \cite[Theorem~5.8]{Rourke-Sanderson(1968)}, there is no real difference between self homotopy equivalences and their block version.
\begin{lemma}\label{lem:.widetildeG_andG}
The map $G(M) \to \widetilde{G}(M)$ and hence the map $BG(M) \to B\widetilde{G}(M)$ are weak homotopy equivalences.
\end{lemma}
The relative homotopy groups of the map $\widetilde{\operatorname{TOP}}(M) \to \widetilde{\operatorname{G}}(M)$
can be identified with the groups ${\mathcal S}^s(M \times D^n,\partial)$ as
explained in~\cite[page~285]{Farrell(2002)}. The next lemma follows in
combination with \cite[Proposition~0.3]{Bartels-Lueck(2012annals)}. Recall
that a space $X$ is \emph{aspherical} if $\pi_i(X) = 0$ for $i \neq 1$.
\begin{lemma}\label{lem:widetildeTOP_andG}
Suppose that $M$ is an aspherical closed manifold of dimension $\ge 5$ and both the $K$-- and $L$--theoretic Farrell--Jones Conjecture hold for ${\mathbb Z} \pi_1(M)$.
Then ${\mathcal S}^s(M \times D^n,\partial)$ is trivial for $n \ge 0$ and the map
\[
\widetilde{\operatorname{TOP}}(M) \to \widetilde{\operatorname{G}}(M)
\]
is a weak homotopy equivalence.
\end{lemma}
For aspherical spaces $X$, the homotopy groups of $\operatorname{G}(X)$ can be computed from
the long exact sequence of homotopy groups associated to the evaluation map $\operatorname{G}(X) \xrightarrow{ev_{x_0}} X$ for some basepoint $x_0 \in X$:
\begin{lemma} \label{lem:pi_n(G(X)_X_aspherical}
Let $X$ be an aspherical CW--complex. Then
\[\pi_n(\operatorname{G}(X))
\cong
\begin{cases}
\operatorname{Out}(\pi_1(X)) & n = 0,
\\
\operatorname{center}(\pi_1(X)) & n = 1,
\\
0 & n \ge 2.
\end{cases}
\]
\end{lemma}
We conclude from \Cref{lem:.widetildeG_andG}, \Cref{lem:widetildeTOP_andG}
and \Cref{lem:pi_n(G(X)_X_aspherical}:
\begin{corollary}\label{lem:pi_n(TOP(X)_X_aspherical}
If $M$ is an aspherical closed manifold of dimension $\ge 5$ with fundamental group $\pi$,
and both the $K$--theoretic and $L$--theoretic Farrell--Jones Conjecture hold for ${\mathbb Z} \pi$,
then there are natural zig-zags of homotopy equivalences
\[
\widetilde{\operatorname{TOP}}(M) \simeq \operatorname{G}(M)
\]
and
\[
B\!\widetilde{\operatorname{TOP}}(M) \simeq B\!\operatorname{G}(M)
\]
and we get
\[
\pi_n(\widetilde{\operatorname{TOP}}(M))
\cong
\begin{cases}
\operatorname{Out}(\pi) & n = 0,
\\
\operatorname{center}(\pi) & n = 1,
\\
0 & n \ge 2.
\end{cases}
\]
\end{corollary}
\begin{theorem}\label{the:widetilde(TOP)(M)/TOP(M)_and_bfH}
There is a map
\[
\widetilde{\operatorname{TOP}}(M)/\operatorname{TOP}(M) \to \Omega^{\infty} \bigl(E{\mathbb Z}/2_+ \wedge_{{\mathbb Z}/2} \Omega {\mathbf W}{\mathbf h}{\mathbf s}^{\operatorname{TOP}}(M)\bigr)
\]
which is $(k+1)$--connected if $k$ is in the topological concordance stable range for $M$. Here $ {\mathbf W}{\mathbf h}{\mathbf s}^{\operatorname{TOP}}(M)$ denotes the connective cover of the Whitehead spectrum ${\mathbf W}{\mathbf h}^{\operatorname{TOP}}(M)$.
\end{theorem}
\begin{proof}
It suffices to show that the spectrum denoted by ${\mathbf W}{\mathbf h}^{\operatorname{TOP}}(M)$ in \cite{Weiss-Williams(1988)} is a model for the homotopy cofiber of the assembly map, see \Cref{def: non-conn whspectrum}. This follows from combining~\cite[Theorem~A]{Weiss-Williams(1988)}, the equivalence $\bfP^{TOP}(M)\simeq \Omega {\mathbf H}^{TOP}(M)$ and \cite[Theorem~0.2]{Jahren-Rognes-Waldhausen(2013)}.
\end{proof}
We conclude from \Cref{the:relations_between_the_various_functors}~\ref{the:relations_between_the_various_functors:wh_TOP_and_wh_PL}, \Cref{lem:pi_n(TOP(X)_X_aspherical}, \Cref{the:widetilde(TOP)(M)/TOP(M)_and_bfH} and the lower bound on the topological concordance stable range given in ~\cite[Corollary~1.4.2]{Jahren-Rognes-Waldhausen(2013)}:
\begin{theorem}\label{the:widetilde(TOP)(M)/TOP(M)_and_bfWh}
Let $M$ be a smoothable aspherical closed manifold of dimension $\ge 10$ with fundamental group $\pi$. Suppose that
the Farrell--Jones Conjecture for $A$--theory for $B\pi$ and the Farrell--Jones Conjecture for $L$--theory for ${\mathbb Z}\pi$ hold.
Then we obtain for $2 \le n \le \min\{( \dim M -7 ) / 2, (\dim M - 4)/3\}$ isomorphisms
\[
\pi_{n}(\operatorname{TOP}(M)) \xrightarrow{\cong} \pi_{n+2}\bigl(E{\mathbb Z}/2_+ \wedge_{{\mathbb Z}/2} {\mathbf W}{\mathbf h}^{\operatorname{TOP}}(B\pi)\bigr)),
\]
and an exact sequence
\begin{multline*}
1 \to \pi_{3}\bigl(E{\mathbb Z}/2_+ \wedge_{{\mathbb Z}/2} {\mathbf W}{\mathbf h}^{\operatorname{TOP}}(B\pi)\bigr)) \to\pi_1(\operatorname{TOP}(M)) \to \operatorname{center}(\pi)
\\
\to \pi_{2}\bigl(E{\mathbb Z}/2_+ \wedge_{{\mathbb Z}/2} {\mathbf W}{\mathbf h}^{\operatorname{TOP}}(B\pi)\bigr)) \to\pi_0(\operatorname{TOP}(M)) \to \operatorname{Out}(\pi) \to 1.
\end{multline*}
\end{theorem}
Next we give the proof of \Cref{the:widetilde(TOP)(M)/TOP(M)_and_bfN(A)}.
\begin{proof}[Proof of \Cref{the:widetilde(TOP)(M)/TOP(M)_and_bfN(A)}]
Let $G$ be a torsionfree hyperbolic group.
Then the Farrell--Jones Conjecture for $A$--theory for $BG$ and the Farrell--Jones Conjectures
for algebraic $K$--theory and for $L$--theory for ${\mathbb Z} G$ hold by
\Cref{the:main_result} and \cite{Bartels-Lueck(2012annals),Bartels-Lueck-Reich(2008hyper)}.
Since $G$ is torsionfree, we have $K_{-i}({\mathbb Z} G) = 0$ for all $i \ge 1$ and
$\widetilde{K}_0({\mathbb Z} G) = 0$, and thus the spectra under consideration are connective by
\Cref{the:Connectivity_of_the_linearization_map}~\ref{the:Connectivity_of_the_linearization_map:2_connected}. It follows from \Cref{lem:equivalence_of_FJ_for_the_various_functors} that there is a weak homotopy equivalence
\[\mathcal{H}^G(\EGF{G}{{\mathcal V}{\mathcal C}{\mathcal Y}};({\mathbf W}{\mathbf h}^{\operatorname{TOP}})^B) \xrightarrow{\simeq} {\mathbf W}{\mathbf h}^{\operatorname{TOP}}(BG).
\]
The arguments in~\cite[Section~10] {Lueck-Steimle(2015splitasmb)} based
on~\cite[Corollary~2.8 and Example~3.6]{Lueck-Weiermann(2012)} for algebraic $K$--theory carry over to ${\mathbf W}{\mathbf h}^{\operatorname{TOP}}$ and
imply that there is a weak homotopy equivalence induced by the various inclusions $C \to G$ of representatives of the conjugacy classes of maximal cyclic subgroups of $G$
\[
\bigvee_C {\mathbf W}{\mathbf h}^{\operatorname{TOP}}(BC) \xrightarrow{\simeq} {\mathbf W}{\mathbf h}^{\operatorname{TOP}}(BG).
\]
From the Bass--Heller--Swan decomposition~\eqref{Bass-Heller_Swan_for_A-theory} we obtain a weak homotopy equivalence
\[
{\mathbf{NA}}(\{\bullet\}) \vee {\mathbf{NA}}(\{\bullet\}) \xrightarrow{\simeq} {\mathbf W}{\mathbf h}^{\operatorname{TOP}}(BC).
\]
This proves part \ref{the:widetilde(TOP)(M)/TOP(M)_and_bfN(A)-it1} of \Cref{the:widetilde(TOP)(M)/TOP(M)_and_bfN(A)}.
Part \ref{the:widetilde(TOP)(M)/TOP(M)_and_bfN(A)-it2} follows from part \ref{the:widetilde(TOP)(M)/TOP(M)_and_bfN(A)-it1} and \Cref{the:widetilde(TOP)(M)/TOP(M)_and_bfWh} together with the fact that the center of a hyperbolic group which is torsionfree and not cyclic is trivial.
\end{proof}
\Cref{the:Relating_A-theory_to_algebraic_K_theory} and \Cref{the:widetilde(TOP)(M)/TOP(M)_and_bfWh} imply
\begin{theorem}[Rational homotopy groups of $\operatorname{TOP}(M)$ for an aspherical closed manifold]
\label{the:pi_i(TOP(M)_M_aspherical}
Let $M$ be a smoothable aspherical closed manifold of dimension $\ge 10$ with fundamental group
$\pi$. Suppose that the Farrell--Jones Conjecture for $K$--theory and for $L$--theory
for ${\mathbb Z}\pi$ hold.
Then for $1 \leq n\leq \min\{( \dim M -7 ) / 2, (\dim M - 4)/3\}$ we have
\[
\pi_n (\operatorname{TOP}(M)) \otimes_{{\mathbb Z}} \IQ
= \begin{cases}
\operatorname{center}(\pi) \otimes_{{\mathbb Z}} \IQ & \text{if}\; n = 1,
\\
\{0\} & \text{if} \; n \ge 2.
\end{cases}
\]
\end{theorem}
\subsection{Smooth automorphism groups of aspherical closed smooth manifolds}
\label{subsec:Smooth_automorphism_groups_of_aspherical_closed_smooth_manifolds}
Taking the computation of $K_i({\mathbb Z}) \otimes_{{\mathbb Z}} \IQ$ of Borel~\cite{Borel(1972)} into account, we get
from \Cref{the:relations_between_the_various_functors},
\Cref{the:Relating_A-theory_to_algebraic_K_theory}
and~\cite[Theorem~0.3]{Lueck-Steimle(2015splitasmb)}
\begin{theorem} \label{the:computation_of_diff_pseudos} Let $M$ be an aspherical closed
smooth manifold of dimension $\ge 10$ with fundamental group $\pi$. Suppose that the
Farrell--Jones Conjecture for $K$--theory and for $L$--theory for ${\mathbb Z}\pi$ hold.
Then we get for all $n \in {\mathbb Z}$
\begin{eqnarray*}
\pi_n ({\mathbf W}{\mathbf h}^{\operatorname{DIFF}} ( M ) )\otimes_{{\mathbb Z}} \IQ
& \cong &
\bigoplus_{k = 1}^{\infty} H_{n-4k-1} (M ; \IQ ),
\\
\pi_n ( \bfP^{\operatorname{DIFF}} ( M ) )\otimes_{{\mathbb Z}} \IQ
& \cong &
\bigoplus_{k = 1}^{\infty} H_{n-4k+1} (M ; \IQ ).
\end{eqnarray*}
\end{theorem}
For the proof of the next result, which does involve the involutions on higher algebraic $K$--theory,
we refer to~\cite[Lecture~5]{Farrell(2002)},~\cite{Farrell-Hsiang(1978)},
or~\cite[Section~2]{Farrell-Jones(1990b)}.
\begin{theorem}[Rational homotopy groups of $\operatorname{DIFF}(M)$ for an aspherical closed smooth manifold]
\label{the:pi_i(DIFF(M)_M_aspherical}
Let $M$ be an aspherical closed smooth manifold of dimension $\ge 10$ with fundamental group $\pi$.
Suppose that the Farrell--Jones Conjecture for $K$--theory and for $L$--theory
for ${\mathbb Z}\pi$ hold.
Then for $1 \leq n \leq \min\{( \dim M -7 ) / 2, (\dim M - 4)/3\}$ we have
\[
\pi_n (\operatorname{DIFF}(M)) \otimes_{{\mathbb Z}} \IQ =
\begin{cases}
\operatorname{center}(\pi) \otimes_{{\mathbb Z}} \IQ & \mbox{if} \; n=1,
\\
\bigoplus_{j=1}^{\infty} H_{(n +1) - 4j} ( M;\IQ ) & \mbox{if} \; n \ge 2 , \; \dim M \; \mbox{odd},
\\
\{0\} & \mbox{if} \; n \ge 2, \; \dim M \mbox{even}.
\end{cases}
\]
\end{theorem}
\begin{remark}[Surfaces and simply connected manifolds] \label{rem:other_manifolds} There
are very interesting computations of the cohomology of $B\!\operatorname{DIFF}(M)$ in a range and under
stabilization with taking the connected sum with $S^n \times S^n$ for $2$--dimensional
manifolds or simply connected high-dimensional manifolds by Berglund, Galatius, Madsen,
Randal-Williams, Weiss and others, see for
instance~\cite{Berglund-Madsen(2013),Berglund-Madsen(2014),Galatius-Randal-Williams(2014abelian),
Galatius-Randal-Williams(2014Cont), Galatius-Randal-Williams(2014Acta), Galatius-Randal-Williams(2014stability_I),
Galatius-Randal-Williams(2016stability_II), Madsen-Weiss(2005),Madsen-Weiss(2007)}.
The methods used in these papers are quite different. Notice that taking the connected
sum with $S^n \times S^n$ will destroy asphericity except for $n = 1$, so that it is not
clear what stabilization could mean in the context of aspherical manifolds in high
dimensions.
\end{remark}
\section{Inheritance properties of the Isomorphism Conjectures}
\label{sec:Inheritance_properties_of_the_Isomorphism_Conjectures}
The main result of this section is
\begin{theorem}[Inheritance properties of the Meta-Conjecture with coefficients]
\label{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c}
Let ${\mathbf S} \colon \curs{SPACES}\to \curs{SPECTRA}$ be a covariant functor which respects weak
equivalences and disjoint unions. Let ${\mathcal C}$ be a class of groups which is closed under
isomorphisms, taking subgroups and taking quotients.
\begin{enumerate}
\item \label{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c:subgroups}
Suppose that the Meta-Isomorphism
Conjecture~\ref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c}
with coefficients holds for $(G,{\mathcal C}(G))$, i.e., it holds for $G$ with respect to the family of subgroups
${\mathcal C}(G) = \{H \subseteq G \mid H \in {\mathcal C}\}$ of $G$. Let $H \subseteq G$ be a subgroup.
Then \Cref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c}
holds for $(H,{\mathcal C}(H))$.
\item \label{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c:extensions}
Let $1 \to K \to G \xrightarrow{p} Q \to 1$ be an extension of
groups. Suppose that $(Q,{\mathcal C}(Q))$ and $(p^{-1}(H), {\mathcal C}(p^{-1}(H))$ for every $H \in {\mathcal C}(Q)$ satisfy
\Cref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c}.
Then $(G,{\mathcal C}(G))$ satisfies
\Cref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c}.
\item \label{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c:direct_products}
Suppose that \Cref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c}
is true for $(H_1 \times H_2,{\mathcal C}(H_1 \times H_2))$ for every $H_1, H_2 \in {\mathcal C}$.
Then for two groups $G_1$ and $G_2$
\Cref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c}
is true for the direct product $G_1 \times G_2$ and the family ${\mathcal C}(G_1 \times G_2)$,
if and only if is true for $(G_k,{\mathcal C}(G_k))$ for $k = 1,2$.
\item \label{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c:directed_colimits}
Suppose that for any directed systems of spaces $\{X_i \mid i \in I\}$ indexed over an arbitrary
directed set $I$ the canonical map
\[
\operatorname{hocolim}_{i \in I} {\mathbf S}(X_i) \to {\mathbf S}\bigl(\operatorname{hocolim}_{i \in I} X_i\bigr)
\]
is a weak homotopy equivalence.
Let $\{G_i \mid i \in I\}$ be a directed system of groups over a directed set $I$
(with arbitrary structure maps). Put $G = \operatorname{colim}_{i \in I} G_i$.
Suppose that \Cref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c}
holds for $(G_i,{\mathcal C}(G_i))$ for every $i \in I$. Then \Cref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c}
holds for $(G,{\mathcal C}(G))$.
\item \label{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c:overgroups_of_finite_index}
The analogs of assertions~\ref{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c:subgroups},~%
\ref{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c:extensions},~%
\ref{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c:direct_products},
and~\ref{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c:directed_colimits}
hold for the Meta-Isomorphism
Conjecture~\ref{con:The_Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c_and_fwp}
with coefficients and finite wreath products.
Moreover, if $G$ is a group and $H \subseteq G$ is a subgroup of finite index, then
\Cref{con:The_Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c_and_fwp}
holds for $(G,{\mathcal C}(G))$, if and only if
\Cref{con:The_Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c_and_fwp}
holds for $(H,{\mathcal C}(H))$.
\end{enumerate}
\end{theorem}
Let us remark that the case of free products is missing in
\Cref{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c}.
It will be treated in
\Cref{subsec:Proof_of_assertion_ref(the:main_result:inheritance)_of_Theorem_ref(the:main_result)}
below.
\subsection{The Fibered Meta-Isomorphism Conjecture for equivariant homology theories}
\label{subsec:The_Fibered_Meta_Isomorphism_Conjecture_for_equivariant_homology_theories}
Next we introduce the Meta-Conjecture and its fibered version in terms
of $G$--homology theories. In this setting the analog of
\Cref{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c}
has already been proved and we want to reduce the case coming from a
functor from spaces to spectra to this situation.
\begin{conjecture}[Meta-Isomorphism Conjecture] \label{con:Meta_Isomorphisms_Conjecture}
The group $G$ satisfies the \emph{Meta-Isomorphism Conjecture}
\index{Conjecture!Meta-Isomorphism Conjecture}
with respect to the $G$--homology theory ${\mathcal H}^G_*$ and the family ${\mathcal F}$ of subgroups of $G$
if the assembly map
\[
{\mathcal H}_n^G(\operatorname{pr}) \colon {\mathcal H}^G_n(\EGF{G}{{\mathcal F}}) \to {\mathcal H}_n^G(G/G)
\]
induced by the projection $\operatorname{pr} \colon \EGF{G}{{\mathcal F}} \to G/G$ is bijective for all $n \in {\mathbb Z}$.
\end{conjecture}
Let $X$ be a $G$--CW--complex. Let $\alpha \colon H \to G$ be a group homomorphism.
Denote by $\alpha^*X$ the $H$--CW--complex obtained from $X$ by
\emph{restriction with $\alpha$}. Given an $H$--CW--complex $Y$, we denote the
$G$--CW--complex given by \emph{induction} by $\alpha_*Y$.
Fix a group $\Gamma$. An equivariant homology theory $H_*^{?}$ over $\Gamma$ in the sense of
\cite[Definition~2.3]{Bartels-Echterhoff-Lueck(2008colim)} assigns to a group $(G,\psi)$
over $\Gamma$, i.e., a group $G$ together with a homomorphism $\psi \colon G \to \Gamma $,
a $G$--homology theory $H_n^{G,\psi}$, sometimes denoted just by $H_*^G$. For two groups
$(G,\psi)$ and $(G',\psi')$ over $\Gamma$ and a morphism $\phi$ between them,
i.e., a
group homomorphism $\phi\colon G \to G'$ with $\psi' \circ \alpha = \psi$, one obtains
homomorphisms $\operatorname{ind}_{\alpha} \colon H^G_*(X,A) \to H^{G'}_*(\alpha_*(X,A))$ for every
$G$--CW--pair $(X,A)$, which are bijective, if the kernel of $\alpha$ acts freely on
$(X,A)$, and compatible with the boundary homomorphisms associated to pairs. If $\Gamma$ is
trivial, this is just an equivariant homology theory.
\begin{conjecture}[Fibered Meta-Isomorphism Conjecture]
\label{con:Fibered_Meta-Isomorphism_Conjectures_for_calh?_ast}
A group $(G,\psi)$ over $\Gamma$
satisfies the \emph{Fibered Meta-Isomorphism Conjecture with respect to ${\mathcal H}^?_*$ and the
family ${\mathcal F}$ of subgroups of $G$} if for each group homomorphism
$\phi \colon K \to G$ the group $K$ satisfies the Meta-Isomorphism
Conjecture~\ref{con:Meta_Isomorphisms_Conjecture} with respect to the $K$--homology
theory ${\mathcal H}^{K,\psi \circ \phi}_*$ and the family $\phi^*{\mathcal F} = \{H \subseteq G \mid \phi(H) \in {\mathcal F}\}$ of subgroups of $K$.
\end{conjecture}
\begin{lemma} \label{lem:basic_inheritance_property_of_fibered_Isomorphism_conjecture}
Let $(G,\psi)$ be a group over $\Gamma$ and $\phi \colon K \to G$ be a group
homomorphism. If $(G,\psi)$ satisfies
the Fibered Meta-Isomorphism Conjecture~\ref{con:Fibered_Meta-Isomorphism_Conjectures_for_calh?_ast}
with respect to the family ${\mathcal F}$ of subgroups of $G$, then the group $(K,\psi \circ \phi)$ over $\Gamma$ satisfies
the Fibered Meta-Isomorphism Conjecture~\ref{con:Fibered_Meta-Isomorphism_Conjectures_for_calh?_ast}
with respect to the family $\phi^*{\mathcal F}$.
\end{lemma}
\begin{proof}
If $\theta \colon L \to K$ is a group homomorphism, then
$\theta^*(\phi^*{\mathcal F}) = (\phi \circ \theta)^*{\mathcal F}$.
\end{proof}
\subsection{Some adjunctions}
\label{subsec:Some_adjunctions}
Let ${\mathbf S} \colon \curs{SPACES}\to \curs{SPECTRA}$ be a covariant functor. Throughout this section we
will assume that it respects weak equivalences and disjoint unions.
\begin{lemma} \label{lem:adjunctions_for_homology_associated_toS_upper_K}
Let $\psi \colon K_1 \to K_2$ be a group homomorphism.
\begin{enumerate}
\item \label{lem:adjunctions_for_homology_associated_toS_upper_K:restriction_of_X}
If $Z$ is a $K_1$--CW--complex and $X$ is a $K_2$--CW--complex, then there is a natural isomorphism
\[
H_n^{K_1}(\psi^*X;{\mathbf S}^{K_1}_Z) \xrightarrow{\cong} H_n^{K_2}(X;{\mathbf S}^{K_2}_{\psi_*Z}).
\]
\item \label{lem:adjunctions_for_homology_associated_toS_upper_K:induction_of_X}
If $Z$ is a $K_2$--CW--complex and $X$ is a $K_1$--CW--complex, then there is a natural isomorphism
\[
H_n^{K_1}(X;{\mathbf S}^{K_1}_{\psi^*Z}) \xrightarrow{\cong} H_n^{K_2}(\psi_*X;{\mathbf S}^{K_2}_{Z}).
\]
\end{enumerate}
\end{lemma}
\begin{proof}~\ref{lem:adjunctions_for_homology_associated_toS_upper_K:restriction_of_X}
The fourth isomorphism appearing in~\cite[Lemma~1.9]{Davis-Lueck(1998)}
together
with~\cite[Lemma~4.6]{Davis-Lueck(1998)} applied levelwise implies that it suffices to construct a
natural weak homotopy equivalence of $\curs{Or}(K_2)$--spectra
\[
u(\psi,Z) \colon \psi_* {\mathbf S}^{K_1}_Z \xrightarrow{\simeq} {\mathbf S}^{K_2}_{\psi_* Z},
\]
where $\psi_* {\mathbf S}^{K_1}_Z$ is the $\curs{Or}(K_2)$--spectrum obtained by induction
in the sense of~\cite[Definition~1.8]{Davis-Lueck(1998)} with the functor
$\curs{Or}(\psi) \colon \curs{Or}(K_1) \to \curs{Or}(K_2),\; K_1/H_1 \mapsto \psi_* (K_1/H_1)$
applied to the $\curs{Or}(K_1)$--spectrum ${\mathbf S}^{K_1}_Z$. For a homogeneous space
$K_2/H$ we define $u(\psi,Z)(K_2/H)$ to be the composite
\begin{eqnarray*}
\psi_* {\mathbf S}^{K_1}_Z(K_2/H)
& = &
\operatorname{map}_{K_2}(\psi_*(K_1/?),K_2/H)_+ \wedge_{\curs{Or}(K_1)} {\mathbf S}\left(K_1/? \times_{K_1} Z\right)
\\
& \xrightarrow{\cong} &
\operatorname{map}_{K_1}((K_1/?),\psi^*(K_2/H))_+ \wedge_{\curs{Or}(K_1)} {\mathbf S}(K_1/? \times_{K_1} Z)
\\
& \xrightarrow{\simeq} &
{\mathbf S}(\psi^*(K_2/H) \times_{K_1} Z)
\\
& \xrightarrow{\cong} &
{\mathbf S}(K_2/H\times_{K_2} \psi_*Z)
\\
& =: &
{\mathbf S}^{K_2}_{\psi_* Z}(K_2/H).
\end{eqnarray*}
Here the first map comes from the adjunction isomorphism
\[
\operatorname{map}_{K_2}(\psi_*(K_1/?),K_2/H) \xrightarrow{\cong} \operatorname{map}_{K_1}(K_1/?),\psi^*(K_2/H)),
\]
and the third map comes from the canonical homeomorphism
\[
\psi^*(K_2/H) \times_{K_1} Z \xrightarrow{\cong} K_2/H\times_{K_2} \psi_*Z.
\]
The second map is the special case $T = \psi^*K_2/{?}$ of the natural weak homotopy equivalence defined for
any $K_1$--set $T$
\[
\kappa(T) \colon \operatorname{map}_{K_1}((K_1/?),T)_+ \wedge_{\curs{Or}(K_1)} {\mathbf S}\left(K_1/? \times_{K_1} Z\right)
\xrightarrow{\simeq}
{\mathbf S}(T\times_{K_1} Z),
\]
which is given by $(u \colon K_1/? \to T) \times s \mapsto {\mathbf S}(u \times_{K_1}
\operatorname{id}_Z)(s)$. If $T$ is a transitive $K_1$--set, then $\kappa(T)$ is even an
isomorphism by the Yoneda Lemma. The
left-hand side is compatible with disjoint unions in $T$, the right-hand side
is compatible with disjoint unions in $T$ up to homotopy, where we use that
${\mathbf S}$ respects disjoint unions.
As every $K_1$--set is the disjoint union of
homogeneous $K_1$--sets, $\kappa(T)$ is a weak homotopy equivalence for every $K_1$--set $T$.
\\[1mm]~\ref{lem:adjunctions_for_homology_associated_toS_upper_K:induction_of_X}
The third isomorphism appearing in~\cite[Lemma~1.9]{Davis-Lueck(1998)} together
with~\cite[Lemma~4.6]{Davis-Lueck(1998)} implies that it suffices to construct a
natural weak homotopy equivalence of $\curs{Or}(K_1)$--spectra
\[
v(\psi,Z) \colon \psi^* {\mathbf S}^{K_2}_Z \xrightarrow{\simeq} {\mathbf S}^{K_1}_{\psi^* Z},
\]
where $ \psi^* {\mathbf S}^{K_2}_Z$ is the $\curs{Or}(K_1)$--spectrum obtained by restriction in the
sense of~\cite[Definition~1.8]{Davis-Lueck(1998)} with the functor $\curs{Or}(\psi) \colon
\curs{Or}(K_1) \to \curs{Or}(K_2),\; K_1/H \mapsto \psi_* (K_1/H)$ applied to the $\curs{Or}(K_2)$--spectrum
${\mathbf S}^{K_2}_Z$. Actually, we obtain even an isomorphism $v(\psi,Z)$ using the adjunction
\[
\psi_*(K_1/H) \times_{K_2} Z \cong K_1/H \times_{K_1} \psi^* Z
\]
for any subgroup $H \subseteq K_1$.
\end{proof}
\subsection{The Fibered Meta--Isomorphism Conjecture with coefficients for functors from spaces to spectra}
\label{subsec:The_Fibered_Meta_Isomorphism_Conjecture_with_coefficients_for_functors_from_spaces_to_spectra}
Notice that for a homomorphism $\phi \colon H \to G$ the restriction $\phi^* Z$ of a free
$G$--CW--complex $Z$ is free again if and only if $\phi$ is injective. We have already
explained in
Remark~\ref{rem:The_condition_free_is_necessary_in_MC_spaces_spectra_with_coeff} that the
assumption that $Z$ is free is needed in
\Cref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c}.
In the Fibered Meta-Isomorphism Conjecture~\ref{con:Fibered_Meta-Isomorphism_Conjectures_for_calh?_ast}
it is crucial not to require that $\phi \colon H \to G$ is injective since we want to have good inheritance properties.
Therefore we have to blow up $Z$ everywhere by passing to $EG \times Z$ as explained below.
Let $G$ be a group and $Z$ be a $G$--CW--complex. Recall that $\underline{G}$
denotes the groupoid with precisely one object which has $G$ as its
automorphism group. Let $\curs{GROUPOIDS} \downarrow G$ be the category of groupoids
over $\underline{G}$. Objects are groupoids ${\mathcal G}$ together with a functor $P
\colon {\mathcal G} \to \underline{G}$. A morphism from $P \colon {\mathcal G} \to
\underline{G}$ to $P'\colon {\mathcal G}' \to \underline{G}$ is a covariant functor
$F \colon {\mathcal G} \to {\mathcal G}'$ satisfying $P' \circ F = P$. Given a groupoid
${\mathcal G}$, we obtain a contravariant functor $E(? \downarrow {\mathcal G}) \colon {\mathcal G}
\to \curs{SPACES}$ by sending an object $x$ to the classifying space of the category
$x \downarrow {\mathcal G}$ of objects in ${\mathcal G}$ under $x$. We get from $Z$, by
restriction along $P$, a covariant functor $P^*Z \colon {\mathcal G} \to \curs{SPACES}$
where we think of the left $G$--space $Z$ as a covariant functor
$\underline{G} \to \curs{SPACES}$. The tensor product over ${\mathcal G}$,
see~\cite[Section~1]{Davis-Lueck(1998)}, yields a space $E(? \downarrow {\mathcal G})
\times_{{\mathcal G}}P^*Z(?)$. Thus we obtain a covariant functor
\begin{equation}
{\mathbf S}^{\downarrow G}_Z \colon \curs{GROUPOIDS} \downarrow G \to \curs{SPECTRA},
\quad P \colon ({\mathcal G} \to \underline{G}) \mapsto {\mathbf S}(E(? \downarrow {\mathcal G})
\times_{{\mathcal G}} P^*Z(?)).
\label{bfS(G,B)_Z}
\end{equation}
It yields an equivariant homology theory $H_n^?(-;{\mathbf S}_Z^{\downarrow G})$ over
$G$, see~\cite[Lemma~7.1]{Bartels-Echterhoff-Lueck(2008colim)}.
Given a homomorphism $\psi \colon K \to G$ we get an identification of $K$--homology theories
\begin{equation}
H^{K,\psi}_*(-;{\mathbf S}^{\downarrow G}_Z) \cong H_*^K(-;{\mathbf S}^K_{EK \times \psi^* Z}),
\label{identification_of_K-homology_theories}
\end{equation}
which is induced by a homotopy equivalence, natural in
$K/H$,
\begin{equation*}
E(? \downarrow {\mathcal T}^K(K/H)) \times_{{\mathcal T}^K(K/H)}
\psi^*Z(?)\xrightarrow{\simeq} K/H
\times_K(EK \times \psi^*Z)
\end{equation*}
and~\cite[Lemma~4.6]{Davis-Lueck(1998)}, where ${\mathcal T}$ denotes the transport
groupoid from \Cref{exa:Z_is_EG} and $\psi$ also denotes its
induced map ${\mathcal T}^K(K/H) \to \underline{G}$. For any group $\psi \colon K \to G$
over $G$, inclusion $i \colon H \to K$ of a subgroup $H$ of $K$, and $n \in
{\mathbb Z}$ we have canonical identifications
\begin{equation*}
H_n^{K,\psi}(K/H;{\mathbf S}_Z^{\downarrow G})) \xrightarrow{\cong} H_n^{H, \psi
\circ i}(H/H;{\mathbf S}_Z^{\downarrow G}))
\cong \pi_n({\mathbf S}(EH \times_H (\psi \circ i)^*Z)).
\end{equation*}
\begin{lemma} \label{lem:HG_ast(phiastY;bfSdownarrow_G_Z)}
Let $\phi \colon H \to K$ and $\psi \colon K \to G$ be group homomorphisms.
\begin{enumerate}
\item \label{lem:HG_ast(phiastY;bfSdownarrow_G_Z):(1)}
Let $X$ be a $G$--CW--complex and let $Z$ be a $K$--CW--complex. Then we obtain a natural isomorphism
\[
H_n^{H,\phi}(\phi^*\psi^*X;{\mathbf S}^{\downarrow K}_{Z})
\xrightarrow{\cong}
H_n^{G}(X;{\mathbf S}^G_{(\psi \circ \phi)_*(EH \times \phi^*Z)}).
\]
\item \label{lem:HG_ast(phiastY;bfSdownarrow_G_Z):(2)}
Let $X$ be an $H$--CW--complex and let $Z$ be a $G$--CW--complex.
Then we obtain a natural isomorphism
\[
H_n^{H,\phi}(X;{\mathbf S}^{\downarrow K}_{\psi^* Z})
\xrightarrow{\cong}
H_n^{H,\psi \circ \phi}\bigl(X;{\mathbf S}^{\downarrow G}_Z\bigr).
\]
\end{enumerate}
\end{lemma}
\begin{proof}~\ref{lem:HG_ast(phiastY;bfSdownarrow_G_Z):(1)}
We get from~\eqref{identification_of_K-homology_theories}
\[
H_n^{H,\phi}(\phi^*\psi^*X;{\mathbf S}^{\downarrow K}_{Z}) := H_n^H(\phi^*\psi^*X;{\mathbf S}^H_{EH \times \phi^*Z}).
\]
Now apply \Cref{lem:adjunctions_for_homology_associated_toS_upper_K}~%
\ref{lem:adjunctions_for_homology_associated_toS_upper_K:restriction_of_X}.
\\[1mm]~\ref{lem:HG_ast(phiastY;bfSdownarrow_G_Z):(2)}
We get from~\eqref{identification_of_K-homology_theories}
\begin{multline*}
H_n^{H,\phi}(X;{\mathbf S}^{\downarrow K}_{\psi^* Z})
:= H_n^H(X;{\mathbf S}^H_{EH \times \phi^* \psi^* Z})
\\
= H_n^H(X;{\mathbf S}^H_{EH \times (\psi \circ \phi)^* Z})
=: H_n^{H,\psi \circ \phi}\bigl(X;{\mathbf S}^{\downarrow G}_Z\bigr).
\end{multline*}
\end{proof}
\begin{conjecture}[Fibered Meta-Isomorphism Conjecture for a functor from spaces to spectra with
coefficients]
\label{con:Fibered_Meta_Conjecture_for_a_functor_from_spaces_to_spectra_with_c}
\index{Conjecture!Fibered Meta Isomorphisms Conjecture for a functor from spaces to spectra with
coefficients} Let ${\mathbf S} \colon \curs{SPACES}\to \curs{SPECTRA}$, as before, respect
weak equivalences and disjoint unions. We say that ${\mathbf S}$ satisfies the Fibered Meta-Isomorphism Conjecture for a functor
from spaces to spectra with coefficients for the group $G$ and the family of subgroups
${\mathcal F}$ of $G$ if the following holds: For any $G$--CW--complex $Z$ the equivariant homology theory
$H_*^?(-;{\mathbf S}_Z^{\downarrow G})$ over $G$ satisfies the Fibered Meta-Isomorphism
Conjecture~\ref{con:Fibered_Meta-Isomorphism_Conjectures_for_calh?_ast} for the group
$(G,\operatorname{id}_G)$ over $G$ and the family ${\mathcal F}$.
\end{conjecture}
Note that
\Cref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c}
deals with the $G$--homology theory $H^G_*(-;{\mathbf S}^G)$, whereas
\Cref{con:Fibered_Meta_Conjecture_for_a_functor_from_spaces_to_spectra_with_c}
deals with the the equivariant homology theory $H^?_*(-;{\mathbf S}^{\downarrow G})$ over
$G$. Moreover,
\Cref{con:Fibered_Meta_Conjecture_for_a_functor_from_spaces_to_spectra_with_c}
is unchanged if we would additionally require that the $G$--CW--complex $Z$ is
free. Namely, for any $G$--CW--complex $Z$ the $G$--CW--complex $EG \times Z$ is free and
the projection $EG \times Z \to Z$ induces an isomorphism $H_*^?(-;{\mathbf S}_{EG \times
Z}^{\downarrow G}) \xrightarrow{\cong} H_*^?(-;{\mathbf S}_Z^{\downarrow G})$ of equivariant
homology theories over $G$ because of~\eqref{identification_of_K-homology_theories}
and~\cite[Lemma~4.6]{Davis-Lueck(1998)}.
\newcommand{\ConjMICforS}
{C\ref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c}\xspace}
\newcommand{MIC\ref{con:Meta_Isomorphisms_Conjecture}\xspace}{MIC\ref{con:Meta_Isomorphisms_Conjecture}\xspace}
\newcommand{FMIC\ref{con:Fibered_Meta-Isomorphism_Conjectures_for_calh?_ast}\xspace}{FMIC\ref{con:Fibered_Meta-Isomorphism_Conjectures_for_calh?_ast}\xspace}
\newcommand{S\ref{con:Fibered_Meta_Conjecture_for_a_functor_from_spaces_to_spectra_with_c}\xspace}{S\ref{con:Fibered_Meta_Conjecture_for_a_functor_from_spaces_to_spectra_with_c}\xspace}
For the rest of this section, we abbreviate the different conjectures as
follows:
\begin{itemize}
\item
\ConjMICforS is the
Meta-Isomorphism~\Cref{con:Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c}
for functors from spaces to spectra with coefficients. This is the
conjecture we want to know about in the end.
\item MIC\ref{con:Meta_Isomorphisms_Conjecture}\xspace and
FMIC\ref{con:Fibered_Meta-Isomorphism_Conjectures_for_calh?_ast}\xspace denote the
Meta-Isomorphism \Cref{con:Meta_Isomorphisms_Conjecture}, and
the Fibered Meta-Isomophism
\Cref{con:Fibered_Meta-Isomorphism_Conjectures_for_calh?_ast}. These are
statements about a ($G$-)equivariant homology theory.
\item S\ref{con:Fibered_Meta_Conjecture_for_a_functor_from_spaces_to_spectra_with_c}\xspace
denotes the Fibered Meta-Isomorphism
\Cref{con:Fibered_Meta_Conjecture_for_a_functor_from_spaces_to_spectra_with_c}
for a functor from spaces to spectra with coefficients. This takes as
input a functor ${\mathbf S}$ and is the most general version of a conjecture we
are interested it.
\end{itemize}
\begin{lemma} \label{lem:MC_with_coeff_and_FMC_with_coeff}
Let $\psi \colon K \to G$ be a group homomorphism.
\begin{enumerate}
\item \label{lem:MC_with_coeff_and_FMC_with_coeff:MC_implies_FC}
Suppose that \ConjMICforS
holds for the group $G$ and the family ${\mathcal F}$. Then S\ref{con:Fibered_Meta_Conjecture_for_a_functor_from_spaces_to_spectra_with_c}\xspace
holds for the group $K$ and the family $\psi^* {\mathcal F}$.
\item \label{lem:MC_with_coeff_and_FMC_with_coeff:FC_implies_MC} If
S\ref{con:Fibered_Meta_Conjecture_for_a_functor_from_spaces_to_spectra_with_c}\xspace
holds for the group $G$ and the family ${\mathcal F}$, then
\ConjMICforS
holds for the group $G$ and the family ${\mathcal F}$.
\item \label{lem:MC_with_coeff_and_FMC_with_coeff_up_from_G_to_K}
Suppose that
S\ref{con:Fibered_Meta_Conjecture_for_a_functor_from_spaces_to_spectra_with_c}\xspace
holds for the group $K$ and the family ${\mathcal F}$. Then for every $G$--CW--complex
$Z$ FMIC\ref{con:Fibered_Meta-Isomorphism_Conjectures_for_calh?_ast}\xspace
holds for the equivariant homology theory $H_n(-;{\mathbf S}^{\downarrow G}_Z)$ over $G$ for the group
$(K,\psi)$ over $G$ and the family ${\mathcal F}$ of subgroups of $K$.
\end{enumerate}
\end{lemma}
\begin{proof}~\ref{lem:MC_with_coeff_and_FMC_with_coeff:MC_implies_FC} This follows from
\Cref{lem:HG_ast(phiastY;bfSdownarrow_G_Z)}~\ref{lem:HG_ast(phiastY;bfSdownarrow_G_Z):(1)},
since in the notation used there we have $\phi^* \psi^*\EGF{G}{{\mathcal F}} = \phi^*\EGF{K}{\psi^*{\mathcal F}}$ and
$\phi^* \psi^* G/G = H/H$, and $(\psi \circ \phi)_*(EH \times \phi^*Z)$ is a free $G$--CW--complex.
\\[1mm]~\ref{lem:MC_with_coeff_and_FMC_with_coeff:FC_implies_MC} This follows from
applying \Cref{con:Fibered_Meta_Conjecture_for_a_functor_from_spaces_to_spectra_with_c} to the special case
$\psi= \operatorname{id}_G$ and the fact that for a free $G$--CW--complex $Z$ the projection $EG \times Z \to Z$ is a
$G$--homotopy equivalence and hence we get from~\eqref{identification_of_K-homology_theories}
and~\cite[Lemma~4.6]{Davis-Lueck(1998)} natural isomorphisms
\[
H^{G,\operatorname{id}_G}_n(X;{\mathbf S}^{\downarrow G}_Z) \cong H^G_n(X;{\mathbf S}^G_{EG \times Z}) \cong H^G_n(X;{\mathbf S}^G_Z)
\]
for every $G$--CW--complex $X$ and $n \in {\mathbb Z}$.
\\[1mm]~\ref{lem:MC_with_coeff_and_FMC_with_coeff_up_from_G_to_K} This follows from
\Cref{lem:HG_ast(phiastY;bfSdownarrow_G_Z)}~\ref{lem:HG_ast(phiastY;bfSdownarrow_G_Z):(2)}.
\end{proof}
\subsection{Strongly continuous equivariant homology theories over a group}
\label{subsec:Strongly_continuous_equivariant_homology_theories_over_a_group}
Fix a group $\Gamma$ and an equivariant homology theory ${\mathcal H}^?_*$ over
$\Gamma$.
Let $X$ be a $G$--CW--complex and let $\alpha \colon H \to G$ be a group homomorphism.
The functors ${\alpha}_*\colon H{-}CW \rightleftarrows G{-}CW \colon\alpha^*$ are adjoint to one another. In particular,
the adjoint of the identity on $\alpha^* X$ is a natural $G$--map
\begin{equation}
f(X,\alpha) \colon {\alpha}_*\alpha^*X \to X, \quad (g,x) \mapsto gx.
\end{equation}
Consider a map $\alpha \colon (H,\xi) \to (G,\mu)$ of groups over $\Gamma$. Define the
$\Lambda$--map
\begin{eqnarray*}
& a_n = a_n(X,\alpha)\colon {\mathcal H}_n^{H}(\alpha^*X)
\xrightarrow{\operatorname{ind}_{\alpha}} {\mathcal H}_n^G({\alpha}_*\alpha^*X)
\xrightarrow{{\mathcal H}_n^G(f(X,\alpha))}
{\mathcal H}_n^G(X).&
\end{eqnarray*}
If $\beta \colon (G,\mu) \to (K,\nu)$ is another morphism of groups over $\Gamma$, then by
the axioms of an induction structure, see~\cite{Lueck(2002b)}, the composite ${\mathcal H}_n^{H}(\alpha^*\beta^*X)
\xrightarrow{a_n(\beta^*X,\alpha)} {\mathcal H}_n^{G}(\beta^*X) \xrightarrow{a_n(X,\beta)}
{\mathcal H}_n^{K}(X)$ agrees with $a_n(X,\beta \circ \alpha) \colon
{\mathcal H}_n^{H}(\alpha^*\beta^*X) = {\mathcal H}_n^{H}((\beta\circ \alpha)^*X) \to {\mathcal H}_n^{K}(X)$
for a $K$--CW--complex $X$.
Consider a directed system of groups $\{G_i \mid i \in I\}$ with $G = \operatorname{colim}_{i \in I}
G_i$ and structure maps $\psi_i \colon G_i \to G$ for $i \in I$ and $\phi_{i,j} \colon G_i
\to G_j$ for $i,j \in I, i \le j$. We obtain for every $G$--CW--complex $X$ a
system
$a_n(\psi_j^*X,\phi_{i,j}) \colon {\mathcal H}^{G_i}(\psi_i^*X) \to {\mathcal H}^{G_j}(\psi_j^*X)$. We
get a map
\begin{equation}
t_n^G(X)~:=~\operatorname{colim}_{i \in I} a_n(X,\psi_i) \colon \operatorname{colim}_{i \in I}
{\mathcal H}_n^{G_i}(\psi_i^*(X))~\to~{\mathcal H}_n^G(X).
\label{t_nG(X)}
\end{equation}
The next definition is taken from~\cite[Definition~3.3]{Bartels-Echterhoff-Lueck(2008colim)}.
\begin{definition}[Strongly continuous equivariant homology theory over a group]
\label{def:strongly_continuous_equivariant_homology_theory_over_a-group}
An equivariant homology theory ${\mathcal H}^?_*$ over the group $\Gamma$ is called \emph{strongly
continuous} if for every group $(G,\xi)$ over $\Gamma$ and every directed system of
groups $\{G_i \mid i \in I\}$ with $G = \operatorname{colim}_{i \in I} G_i$ and structure maps $\psi_i
\colon G_i \to G$ for $i \in I$ the
map
\[
t^G_n(\{\bullet\}) \colon \operatorname{colim}_{i \in I} {\mathcal H}^{G_i}_n(\{\bullet\}) \to
{\mathcal H}^G_n(\{\bullet\})
\]
is an isomorphism for every $n \in {\mathbb Z}$.
\end{definition}
\begin{lemma} \label{lem:bfS_and_homotopy_colimits} Suppose
that for any directed system of spaces $\{X_i \mid i \in I\}$ indexed over an arbitrary
directed set $I$ the canonical map
\[
\operatorname{hocolim}_{i \in I} {\mathbf S}(X_i) \to {\mathbf S}\bigl(\operatorname{hocolim}_{i \in I} X_i\bigr)
\]
is a weak homotopy equivalence.
Then for every group $\Gamma$ and $\Gamma$--CW--complex $Z$ the equivariant homology theory over
$\Gamma$ given by $H^?_*(-;{\mathbf S}^{\downarrow \Gamma}_Z)$ is strongly continuous.
\end{lemma}
\begin{proof} We only treat the case $\Gamma = G$ and $\psi = \operatorname{id}_G$, the general case of a group
$\psi \colon G \to \Gamma$ over $\Gamma$ is completely analogous.
Consider a directed system of
groups $\{G_i \mid i \in I\}$ with $G = \operatorname{colim}_{i \in I} G_i$. Let $\psi_i \colon G_i \to G$ be the structure map for $i \in I$.
As $I$ is directed, the canonical map
\begin{equation}
\operatorname{hocolim}_{i \in I} {\mathbf S}(EG_i \times_{G_i} \psi_i^* Z) \to {\mathbf S}\bigl(\operatorname{hocolim}_{i \in I} (EG_i \times_{G_i} \psi_i^* Z)\bigr)
\label{lem:bfS_and_homotopy_colimits:map_1}
\end{equation}
is by assumption a weak homotopy equivalence.
We have the homeomorphisms
\begin{eqnarray*}
EG_i \times_{G_i} \psi_i^* Z
& \xrightarrow{\cong} &
(\psi_i)_* EG_i \times_G Z,
\\
\bigl(\operatorname{hocolim}_{i \in I} (\psi_i)_* EG_i\bigr) \times_GZ
&\xrightarrow{\cong} &
\operatorname{hocolim}_{i \in I} \bigl((\psi_i)_* EG_i \times_G Z\bigr).
\end{eqnarray*}
They induce a homeomorphism
\begin{equation}
{\mathbf S}(\operatorname{hocolim}_{i \in I} (EG_i \times_{G_i} \psi_i^* Z)\bigr) \xrightarrow{\cong}
{\mathbf S}\bigl((\operatorname{hocolim}_{i \in I} (\psi_i)_*EG_i) \times_G Z\bigr).
\label{lem:bfS_and_homotopy_colimits:map_2}
\end{equation}
The canonical map
\[
\operatorname{hocolim}_{i \in I} (\psi_i)_*EG_i \to EG
\]
is a $G$--homotopy equivalence. The proof of this fact is a special case of the argument
appearing in the proof of~\cite[Theorem~4.3 on page~516]{Lueck-Weiermann(2012)}. It induces a weak homotopy equivalence
\begin{equation}
{\mathbf S}\bigl((\operatorname{hocolim}_{i \in I} (\psi_i)_*EG_i) \times_G Z) \to {\mathbf S}(EG \times_G Z).
\label{lem:bfS_and_homotopy_colimits:map_3}
\end{equation}
Hence we get by taking the composite of the
maps~\eqref{lem:bfS_and_homotopy_colimits:map_1},~\eqref{lem:bfS_and_homotopy_colimits:map_2}
and~\eqref{lem:bfS_and_homotopy_colimits:map_3} a weak homotopy equivalence
\[
\operatorname{hocolim}_{i \in I} {\mathbf S}(EG_i \times_{G_i} \psi_i^* Z)\to {\mathbf S}(EG \times_G Z).
\]
As $I$ is directed, it induces after taking homotopy groups for every $n \in {\mathbb Z}$ an isomorphism
\[
\operatorname{colim}_{i \in I} \pi_n\bigl({\mathbf S}(EG_i \times_{G_i} \psi_i^* Z)\bigr) \to \pi_n\bigl({\mathbf S}(EG \times_G Z)\bigr),
\]
which can be identified using~\eqref{identification_of_K-homology_theories} with the canonical map
\[
t_n^G(\{\bullet\}) \colon \operatorname{colim}_{i \in I} H_n^{G_i}(\{\bullet\};{\mathbf S}^{\downarrow G}_Z) \to H_n^G(\{\bullet\};{\mathbf S}^{\downarrow G}_Z).
\]
This finishes the proof of \Cref{lem:bfS_and_homotopy_colimits}.
\end{proof}
\subsection{Proof of Theorem~\ref{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c}}
\label{subsec:proof_of_Theorem_ref(the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c)}
In this section we give the proof of
\Cref{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c}.
We use the notation from there.
\begin{proof}
\ref{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c:subgroups}
Consider a free $H$--CW--complex $Z$. Let $i \colon H \to G$ be the inclusion.
Then $i_*Z$ is a free $G$--CW--complex, $i^*\EGF{G}{{\mathcal C}(G)}$ is a model for
$\EGF{H}{{\mathcal C}(H)}$ and $i^*G/G = H/H$. From \Cref{lem:adjunctions_for_homology_associated_toS_upper_K}~%
\ref{lem:adjunctions_for_homology_associated_toS_upper_K:restriction_of_X},
we obtain a commutative diagram with isomorphisms as vertical maps
\[\xymatrix{
H_n^H(\EGF{H}{{\mathcal C}(H)};{\mathbf S}^H_Z) \ar[r] \ar[d]_{\cong}
&
H_n^H(H/H;{\mathbf S}^G_Z) \ar[d]^{\cong}
\\
H_n^G(\EGF{G}{{\mathcal C}(G)};{\mathbf S}^G_{i_*Z}) \ar[r]
&
H_n^G(G/G;{\mathbf S}^G_{i_*Z})
}
\]
where the horizontal maps are induced by the projections. The lower map is bijective by assumption.
Hence the upper map is bijective as well.
\\[1mm]~\ref{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c:extensions}
As \ConjMICforS holds for $(Q, \mathcal{Q})$, by
\Cref{lem:MC_with_coeff_and_FMC_with_coeff}\ref{lem:MC_with_coeff_and_FMC_with_coeff:MC_implies_FC},
S\ref{con:Fibered_Meta_Conjecture_for_a_functor_from_spaces_to_spectra_with_c}\xspace holds for $(G, p^{*}\mathcal{C}(Q))$. By the same
\Cref{lem:MC_with_coeff_and_FMC_with_coeff}\ref{lem:MC_with_coeff_and_FMC_with_coeff:MC_implies_FC},
for every $H\in\mathcal{C}(Q)$, \ConjMICforS holds for $(p^{-1}(H),
\mathcal{C}(p^{-1}(H)))$. Naturally, $p^{-1}(H) \subseteq G$ is a group over
$G$ for which by
\Cref{lem:MC_with_coeff_and_FMC_with_coeff}\ref{lem:MC_with_coeff_and_FMC_with_coeff_up_from_G_to_K}
FMIC\ref{con:Fibered_Meta-Isomorphism_Conjectures_for_calh?_ast}\xspace holds for $H^{?}_n(-; \mathbf{S}^{\downarrow G}_{Z})$ for any $G$-CW
complex $Z$ and the family $\mathcal{C}(p^{-1}(H)) =
\mathcal{C}(G)|_{p^{-1}(H)}$. Let $L \in p^{*}\mathcal{C}(Q)$. Then, using
\Cref{lem:basic_inheritance_property_of_fibered_Isomorphism_conjecture} for
the map $L \to p^{-1}(p(L))$, FMIC\ref{con:Fibered_Meta-Isomorphism_Conjectures_for_calh?_ast}\xspace holds for $(L, \mathcal{C}|_L)$ and
$H^{?}_n(-; \mathbf{S}^{\downarrow G}_{Z})$. As FMIC\ref{con:Fibered_Meta-Isomorphism_Conjectures_for_calh?_ast}\xspace holds for $(G,
p^{*}{C}(Q))$ and for every $L \in p^{*}{C}(Q)$ for $(L, \mathcal{C}|_L)$, the
Transitivity Principle,
see~\cite[Theorem~4.3]{Bartels-Echterhoff-Lueck(2008colim)}, implies that FMIC\ref{con:Fibered_Meta-Isomorphism_Conjectures_for_calh?_ast}\xspace
holds for $(G, \mathcal{C})$. By Lemma
\Cref{lem:MC_with_coeff_and_FMC_with_coeff}\ref{lem:MC_with_coeff_and_FMC_with_coeff:FC_implies_MC},
then also \ConjMICforS holds for $(G, \mathcal{C})$.
\\[1mm]~\ref{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c:direct_products}
If \ConjMICforS
holds for $(G_1 \times G_2,{\mathcal C}(G_1 \times G_2))$, it holds for
$G_k$ and the family ${\mathcal C}(G_k) = {\mathcal C}(G_1 \times G_2)|_{G_k}$ for $k = 1,2
$ by
assertion~\ref{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c:subgroups}.
Suppose that \ConjMICforS
holds for $(G_k,{\mathcal C}(G_k))$ for $k = 1,2$.
By assertion~\ref{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c:extensions}
applied to the split exact sequence $1 \to H_2 \to G_1 \times H_2 \to G_1 \to 1$,
\ConjMICforS
holds
for $(G_1 \times H_2,{\mathcal C}(G_1 \times H_2))$ for every $H_2 \in {\mathcal C}(G_2)$.
By assertion~\ref{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c:extensions}
applied to the split exact sequence $1 \to G_1 \to G_1 \times G_2 \to G_2 \to 1$
\ConjMICforS
holds for $(G_1 \times G_2,{\mathcal C}(G_1 \times G_2))$.
\\[1mm]~\ref{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c:directed_colimits}
Since the \ConjMICforS
holds for $G_i$ and ${\mathcal C}(G_i)$ for every $i \in I$ by assumption, we conclude from \Cref{lem:MC_with_coeff_and_FMC_with_coeff}~%
\ref{lem:MC_with_coeff_and_FMC_with_coeff:MC_implies_FC} that
S\ref{con:Fibered_Meta_Conjecture_for_a_functor_from_spaces_to_spectra_with_c}\xspace
holds for the group $G_i$ and the family ${\mathcal C}(G_i)$ for every $i \in I$.
\Cref{lem:MC_with_coeff_and_FMC_with_coeff}~\ref{lem:MC_with_coeff_and_FMC_with_coeff_up_from_G_to_K}
implies that for every $i \in I$ and $G$--CW--complex $Z$
FMIC\ref{con:Fibered_Meta-Isomorphism_Conjectures_for_calh?_ast}\xspace
holds for the equivariant homology theory $H_n(-;{\mathbf S}^{\downarrow G}_Z)$ over $G$ for the group
$\psi_i \colon G_i \to G$ over $G$ and the family ${\mathcal C}(G_i)$. We conclude
from~\cite[Theorem~5.2]{Bartels-Echterhoff-Lueck(2008colim)}
and \Cref{lem:bfS_and_homotopy_colimits} that for every $G$--CW--complex $Z$
FMIC\ref{con:Fibered_Meta-Isomorphism_Conjectures_for_calh?_ast}\xspace
holds for the equivariant homology theory $H^?_*(-;{\mathbf S}^{\downarrow G}_{Z})$ over $G$ for the group
$(G,\operatorname{id}_G) $ over $G$ and the family ${\mathcal C}(G)$. In other words,
S\ref{con:Fibered_Meta_Conjecture_for_a_functor_from_spaces_to_spectra_with_c}\xspace
holds for the group $G$ and the family ${\mathcal C}(G)$.
\Cref{lem:MC_with_coeff_and_FMC_with_coeff}~\ref{lem:MC_with_coeff_and_FMC_with_coeff:FC_implies_MC} implies
that \ConjMICforS
holds for the group $G$ and the family ${\mathcal C}(G)$.
\\[1mm]~\ref{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c:overgroups_of_finite_index}
The analogs of \ref{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c:subgroups},~%
\ref{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c:extensions},~%
\ref{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c:direct_products},
and~\ref{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c:directed_colimits}
hold for the Meta-Isomorphism Conjecture~\ref{con:The_Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c_and_fwp}
with coefficients and finite wreath products by~\cite[Lemma~3.2, 3.15, 3.16, Satz~3.5]{Kuehl(2009)}.
For a group $G$ and two finite groups $F_1$ and $F_2$ we have $(H \wr F_1 )\wr F_2 \subset
H \wr (F_1 \wr F_2)$ and $F_1 \wr F_2$ is finite. In particular, if $G$
satisfies
\Cref{con:The_Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c_and_fwp} with wreath products, then the same is true for any wreath
product $G \wr F$ with $F$
finite. If $H \subseteq G$ is a subgroup of finite index, then $G$ can be embedded in $H \wr F$ for some
finite group $F$, see \cite[Proof of Proposition 2.17]{Wegner-solvable}. Hence Meta-Isomorphism~\Cref{con:The_Meta_Isomorphisms_Conjecture_for_functors_from_spaces_to_spectra_with_c_and_fwp}
with coefficients passes to supergroups of finite index. This finishes the proof of \Cref{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c}.
\end{proof}
\subsection{Proof of assertion~\ref{the:main_result:inheritance} of Theorem~\ref{the:main_result}}
\label{subsec:Proof_of_assertion_ref(the:main_result:inheritance)_of_Theorem_ref(the:main_result)}
By
\Cref{lem:bfA_respects_weak_equivalences_and_disjoint_unions}, the functor
$\bfA$ satisfies all assumptions of
\Cref{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c}.
The claim of the inheritance properties appearing in
\cref{the:main_result:inheritance} of \Cref{the:main_result} follows
immediately from
\Cref{the:Inheritance_properties_of_the_Meta-Isomorphism_Conjecture_for_bfS_with_c}
except for the statements about extensions, direct products and free
products. For extensions, it follows from the inheritance under finite index
supergroups. For direct products, note that the product of two virtually
cyclic groups is virtually abelian, hence by \cite{Ullmann-Winges(2015)} it
satisfies the conjecture.
For free products, note that
due to the inheritance under filtered colimits, we can assume our groups are
finitely generated, in particular countable.
For $G_1$,
$G_2 \in {\mathcal F}\!{\mathcal J}_A$ consider the canonical map $p \colon G_1 * G_2 \to G_1 \times G_2$. We
already know that $G_1 \times G_2 \in {\mathcal F}\!{\mathcal J}_A$ and hence that it suffices to prove
$p^{-1}(C) \in {\mathcal F}\!{\mathcal J}_A$, where $C$ is the trivial or any infinite cyclic subgroup of $G_1 \times G_2$.
By~\cite[Lemma~5.2]{Roushon(2008FJJ3)} all such $p^{-1}(C)$ are
free and hence hyperbolic, as $G_1 \times G_2$ is countable.
\section{Proof of the Farrell--Jones Conjecture for hyperbolic and CAT(0)--groups}\label{Proof of the Farrell--Jones Conjecture for hyperbolic and CAT(0)--groups}
Thanks to the framework established in \cite{Ullmann-Winges(2015)}, we can
proceed similarly to the linear case as in \cite{Wegner(2012)} and reduce the proof to the construction of a transfer map. This reduction is carried out in this section, while the construction of the transfer occupies \Cref{The Transfer: Final Part of the Proof}.
\subsection{Homotopy coherent actions and homotopy transfer reducibility}
\label{subsec:proof-fj:homotopy-coherent}
The geometric criterion we use to prove the conjecture relies on the notion
of a homotopy coherent diagram, which goes back to Vogt \cite{Vogt(1973)}.
For applications to the Farrell--Jones conjecture, it is enough to consider
the case of a homotopy coherent diagram of shape $G$, regarding $G$ as a
one-object groupoid. In this special case, Vogt's definition was
rediscovered by Wegner \cite[Definition~2.1]{Wegner(2012)}, who called it
``strong homotopy action''.
\begin{dfn}\label{def:hptycohG-action}
A \emph{homotopy coherent $G$--action} of a group $G$ on a topological space $X$ is a continuous map
\[
\Gamma: \coprod_{j=0}^\infty((G \times [0,1])^j\times G\times X)\to X
\]
with the following properties:
\begin{equation*}
\Gamma(\gamma_k,t_k,\dots,\gamma_1,t_1,\gamma_0,x) = \begin{cases}
\Gamma(\dots, \gamma_j, \Gamma(\gamma_{j-1},\dots,x)) & t_j = 0 \\
\Gamma(\dots,\gamma_j\gamma_{j-1},\dots,x) & t_j = 1 \\
\Gamma(\gamma_k,\dots,\gamma_2,t_2,\gamma_1,x) & \gamma_0 = e, 0 < k \\
\Gamma(\gamma_k,\dots,t_{j+1}t_j,\dots,\gamma_0,x) & \gamma_j = e, 1 \leq j < k \\
\Gamma(\gamma_{k-1},t_{k-1},\dots,t_1,\gamma_0,x) & \gamma_k = e, 0 < k
\\
x & \gamma_0 = e, k=0
\end{cases}
\end{equation*}
\end{dfn}
The following definition is adapted from the conditions given in
\cite[Theorem~B]{Bartels(2012)}, which does not use coherence conditions.
We explain some notation below.
\begin{dfn}\label{def:transfer-reducible}
Let $G$ be a discrete group. Let $\mathcal{F}$ be a family of subgroups of $G$.
Then $G$ is \emph{homotopy transfer reducible over $\mathcal{F}$} if there exists a
finite, symmetric generating set $S \subset G$ of $G$ which contains the
trivial element, as well as $N \in \mathbb{N}$ such that there are for every $n
\in \mathbb{N}$
\begin{enumerate}
\item a compact, contractible metric space $(X,d_X)$ such that for every $\epsilon > 0$ there is an $\epsilon$--controlled domination of $X$ by an at most $N$--dimensional, finite simplicial complex.
\item a homotopy coherent $G$--action $\Gamma$ on $X$.
\item a $G$--simplicial complex $\Sigma$ of dimension at most $N$ whose isotropy is contained in $\mathcal{F}$.
\item a continuous map $f \colon X \to \Sigma$ which is \emph{$(S,n)$--equivariant} in the sense that
\begin{itemize}
\item for all $x \in X$ and $s \in S^n$,
\begin{equation*}
d^{\ell^1}(f(\Gamma(s,x)),s \cdot f(x)) \leq \frac{1}{n}.
\end{equation*}
\item for all $x \in X$ and $s_0,\dots,s_n \in S^n$,
\begin{equation*}
\diam \{ f(\Gamma(s_n,t_n,\dots,s_0,x)) \mid (t_1,\dots,t_n) \in [0,1]^n \} \leq \frac{2}{n}.
\end{equation*}
\end{itemize}
\end{enumerate}
\end{dfn}
\begin{notation}
Let us briefly recall some notation used in \Cref{def:transfer-reducible}.
\begin{enumerate}
\item Recall from~\cite[Definition 1.5]{Bartels-Lueck(2012annals)} that an
$\epsilon$--controlled domination of a metric space $(X,d)$ by a finite
simplicial complex $K$ consists of maps $i\colon X \to K$, $p\colon K\to
X$ together with a homotopy $H$ from $p \circ i$ to $\operatorname{id}_X$ such that for
every $x\in X$ the diameter of $\{H(x,t) \mid t\in [0,1]\}$ is at most
$\epsilon$.
\item The $\ell^1$--metric $d^{\ell^1}$ on a simplicial complex is defined
in~\cite[4.2]{Bartels-Lueck-Reich(2008hyper)}.
\item If $S$ is a finite generating set of $G$, we denote by $S^n
\subseteq G$ the set $\{ s_1 s_2 \dots s_n \in G \mid s_i \in S\}$. We
always equip $G$ with the word metric $d_G$ with respect to $S$.
Equivalently, $S^n$ is the $n$--ball around the trivial element with
respect to $d_G$.
\end{enumerate}
\end{notation}
We will show in \Cref{subsec:proof-fj:main-theorem} that a group
satisfying \Cref{def:transfer-reducible} satisfies the Farrell-Jones
Conjecture with coefficients in $A$--theory with respect to the family $\mathcal{F}$,
and we show in
\Cref{subsec:proof-fj:homotopy-transfer-follows-strongly-transfer} that
hyperbolic and $\mathrm{CAT}(0)$--groups satisfy
\Cref{def:transfer-reducible}, thus proving
\Cref{the:main_result}\ref{the:main_result:groups}.
\subsection{Controlled CW--complexes}
\label{subsec:proof-fj:controlled-cw}
Let $G$ be a discrete group and let $\mathcal{F}$ be a family of subgroups of $G$.
In \cite{Ullmann-Winges(2015)} it was shown that \Cref{thm:afjc-transfer-reducible} holds for $G$ if and only if a certain spectrum $\mathbb{F}(G, W, \EGF{G}{{\mathcal F}})$ is weakly contractible for every free $G$--CW--complex $W$. The spectrum $\mathbb{F}(G, W, \EGF{G}{{\mathcal F}})$ is the algebraic $K$--theory of a Waldhausen category of controlled retractive $G$--CW--complexes, similar in spirit to the obstruction category for the Isomorphism Conjecture in algebraic $K$--theory, cf.~\cite{Bartels-Farrell-Jones-Reich(2004)}, \cite[Section~3]{Bartels-Lueck-Reich(2008hyper)}.
Let us recall the relevant definitions from \cite{Ullmann-Winges(2015)} in
this and the next section.
A \emph{coarse structure} is a triple $\mathfrak{Z}=(Z, \mathfrak{C}, \mathfrak{S})$
such that $Z$ is a Hausdorff $G$--space, $\mathfrak{C}$ is a collection of reflexive,
symmetric and $G$--invariant relations on $Z$ which is closed under taking
finite unions and compositions,
see \cite[Definition~2.1]{Ullmann-Winges(2015)}, and $\mathfrak{S}$ is a collection of
$G$--invariant subsets of $Z$ which is closed under taking finite unions.
See \cite[Definition~3.23]{Ullmann-Winges(2015)} for the notion of a
\emph{morphism of coarse structures}.
Fix a coarse structure $\mathfrak{Z}$.
A \emph{labeled $G$--CW--complex relative W}, see \cite[Definition~2.3]{Ullmann-Winges(2015)}, is a pair $(Y, \kappa)$, where $Y$ is a free $G$--CW--complex relative $W$ together with a $G$--equivariant function $\kappa \colon \diamond\, Y \rightarrow Z$. Here, $\diamond Y$ denotes the (discrete) set of relative cells of $Y$.
A \emph{$\mathfrak{Z}$--controlled map} $f \colon (Y_1,\kappa_1) \rightarrow (Y_2, \kappa_2)$ is a $G$--equivariant, cellular map $f \colon Y_1 \rightarrow Y_2$ relative $W$ such that for all $k \in \mathbb{N}$ there is some $C \in \mathfrak{C}$ for which
\[
(\kappa_2,\kappa_1)(\{(e_2,e_1) \mid e_1 \in \diamond\,_k Y_1, e_2\in \diamond\, Y_2, \langle f(e_1) \rangle \cap e_2 \neq \varnothing\}) \subseteq C
\]
holds, where $\langle f(e_1)\rangle$ denotes the smallest non-equivariant
subcomplex of $Y_2$ which contains $f(e_1)$.
A \emph{$\mathfrak{Z}$--controlled $G$--CW--complex relative W} is a labeled $G$--CW--complex $(Y,\kappa)$ relative $W$, such that the identity is a $\mathfrak{Z}$--controlled map and for all $k \in \mathbb{N}$ there is some $S \in \mathfrak{S}$ such that
\[
\kappa(\diamond_k Y)\subseteq S.
\]
A \emph{$\mathfrak{Z}$--controlled retractive space relative $W$} is a
$\mathfrak{Z}$--controlled $G$--CW--complex $(Y,\kappa)$ relative $W$ together with
a $G$--equivariant retraction $r \colon Y \to W$, i.e., a left inverse to
the structural inclusion $W \hookrightarrow Y$. The $\mathfrak{Z}$--controlled
retractive spaces relative $W$ form a category $\mathcal{R}^G(W,\mathfrak{Z})$ in which
\emph{morphisms} are $\mathfrak{Z}$--controlled maps which additionally respect the chosen retractions.
The category of controlled $G$--CW--complexes (relative $W$) and controlled
maps admits a notion of \textit{controlled homotopies}, see
\cite[Definition~2.5]{Ullmann-Winges(2015)} via the objects $(Y
\leftthreetimes [0,1], \kappa \circ pr_Y)$, where $Y \leftthreetimes [0,1]$
denotes the reduced product which identifies $W \times [0,1] \subseteq Y
\times [0,1]$ to a single copy of $W$ and $pr_Y: \diamond\, (Y \leftthreetimes
[0,1]) \rightarrow \diamond Y$ is the canonical projection.
In particular, we obtain a notion of \emph{controlled homotopy equivalence} (or \emph{$h$--equivalence}).
A $\mathfrak{Z}$--controlled retractive space $(Y, \kappa)$ is called \textit{finite} if it is finite-dimensional, the image of $Y \backslash W$ under the retraction meets the orbits of only finitely many path components of $W$ and for each $z \in Z$ there is some open neighborhood $U$ of $z$ such that $\kappa^{-1}(U)$ is finite, see \cite[Definition~3.3]{Ullmann-Winges(2015)}.
A $\mathfrak{Z}$--controlled retractive space $(Y, \kappa)$ is called
\textit{finitely dominated}, if there are a finite $\mathfrak{Z}$--controlled,
retractive space $D$, a morphism $p \colon D \rightarrow Y$ and a
$\mathfrak{Z}$--controlled map $i \colon Y \rightarrow D$ such that $p \circ i$ is
controlled homotopic to $\operatorname{id}_Y$.
The finite, respectively finitely dominated, $\mathfrak{Z}$--controlled retractive spaces form full subcategories $\mathcal{R}^G_f(W,\mathfrak{Z}) \subset \mathcal{R}^G_{fd}(W,\mathfrak{Z}) \subset \mathcal{R}^G(W,\mathfrak{Z})$.
All three of these categories support a Waldhausen category structure in
which inclusions of $G$--invariant subcomplexes up to isomorphism are the
cofibrations and controlled homotopy equivalences are the weak equivalences,
see \cite[Corollary~3.22]{Ullmann-Winges(2015)}. We denote this class of
weak equivalences by $h$.
Note that a controlled homotopy equivalence is a \emph{morphism}, but only
admits a controlled homotopy inverse \emph{map}, which does not need to be
compatible with the retractions to $W$. This is similar to the classical
situation \cite[Section~2.1]{Waldhausen(1985)}.
\subsection{The obstruction category}
\label{subsec:proof-fj:the-obstruction-category}
Let $M$ be a metric space with free, isometric $G$--action. Define the
\emph{bounded morphism control condition on $M$},
$\mathfrak{C}_{bdd}(M)$, to be the collection of all subsets $C \subset M \times M$
which are of the form
\begin{equation*}
C = \{ (m,m') \in M \times M \mid d(m,m') \leq \alpha \}
\end{equation*}
for some $\alpha \geq 0$.
Let $X$ be a $G$--CW--complex. Define further the \emph{$G$--continuous
control condition} $\mathfrak{C}_{Gcc}(X)$ to be the collection of all $C \subset (X \times [1,\infty[) \times (X \times [1,\infty[)$ which satisfy the following:
\begin{enumerate}
\item For every $x \in X$ and every $G_x$--invariant open neighborhood
$U$ of $(x,\infty)$ in $X \times [1,\infty]$, there exists a
$G_x$--invariant open neighborhood $V \subset U$ of $(x,\infty)$ such
that $(((X \times [1,\infty[) \smallsetminus U) \times V) \cap C = \varnothing$.
\item Let $p_{[1,\infty[} \colon X \times [1,\infty[ \to [1,\infty[$ be the projection map. Equip $[1,\infty[$ with the Euclidean metric. Then there exists some $B \in \mathfrak{C}_{bdd}([1,\infty[)$ such that $C \subset p^{-1}_{[1,\infty[}(B)$.
\item $C$ is symmetric, $G$--invariant and contains the diagonal.
\end{enumerate}
We can combine the two morphims control conditions into one set of
conditions on $M \times X \times [1, \infty[$: Let $p_M \colon M \times X
\times [1,\infty[ \to M$ and $p_{X \times [1,\infty[} \colon M \times X
\times [1,\infty[ \to X \times [1,\infty[$ denote the projection maps.
Then $\mathfrak{C}(M,X)$ is the collection of all subsets $C \subset (M \times
X \times [1,\infty[)^2$ which are of the form
\begin{equation*}
C = p_M^{-1}(B) \cap p_{X \times [1,\infty[}^{-1}(C')
\end{equation*}
for some $B \in \mathfrak{C}_{bdd}(M)$ and $C' \in \mathfrak{C}_{Gcc}(X)$.
Finally, define $\mathfrak{S}(M,X)$ to be the collection of all subsets $S \subset M
\times X \times [1,\infty[$ which are of the form $S = K \times [1,\infty[$
for some $G$--compact subset $K \subset M \times X$. Recall that
$\EGF{G}{{\mathcal F}}$ denotes the classifying space of $G$ with respect to
${\mathcal F}$. We also consider $G$ as a metric space with the word metric
induced by a generating set $S$.
\begin{definition}
With the above definitions we obtain a coarse structure
\begin{equation*}
\mathbb{J}(M,X) := (M \times X \times [1,\infty[, \mathfrak{C}(M,X), \mathfrak{S}(M,X)).
\end{equation*}
Define the ``obstruction category'' as the category of finite controlled
CW-complexes relative $W$, i.e., as
\begin{equation*}
\mathcal{R}^G_f(W,\mathbb{J}(G,\EGF{G}{{\mathcal F}})),h),
\end{equation*}
cf.~\cite[Example~2.2 and Definition~6.1]{Ullmann-Winges(2015)}. The
spectrum $\mathbb{F}(G,W,\EGF{G}{{\mathcal F}})$ alluded to before is the non-connective
$K$--theory
spectrum of $\mathcal{R}^G_f(W,\mathbb{J}(G,\EGF{G}{{\mathcal F}}))$ with respect to the
$h$--equivalences, cf.~\cite[Section~5]{Ullmann-Winges(2015)} and
\Cref{def:delooping-coarse-structure} below. If $M = G$, we often
abbreviate $\mathbb{J}(G,X)$ as $\mathbb{J}(X)$.
\end{definition}
By \cite[Corollary~6.11]{Ullmann-Winges(2015)}, a group $G$ satisfies the
Farrell-Jones Conjecture \ref{con:FJC_for_A-theory_with_c} with coefficients in $A$--theory with respect to
$\mathcal{F}$ if and only if $\mathbb{F}(G,W,\EGF{G}{{\mathcal F}})$ is weakly contractible for
every free $G$--CW--complex $W$.
\subsection{The target of the transfer}
\label{subsec:proof-fj:target-of-the-transfer}
Suppose that $G$ is homotopy transfer reducible in the sense of \Cref{def:transfer-reducible}.
The key step in proving the weak contractibility of
$\mathbb{F}(G,W,\EGF{G}{{\mathcal F}})$ will be the construction of a ``transfer map". We
need a generalization of the coarse structure $\mathbb{J}(M,X)$ to define the
target of the transfer.
Suppose that $(M_n)_n$ is a sequence of metric spaces with a free, isometric $G$--action. Let $X$ be a $G$--CW--complex. Following \cite[Section~7]{Ullmann-Winges(2015)}, define the coarse structure
\begin{equation*}
\mathbb{J}((M_n)_n,X) := \big( \coprod_n M_n \times X \times [1,\infty[, \mathfrak{C}((M_n)_n,X), \mathfrak{S}((M_n)_n,X) \big)
\end{equation*}
as follows: Members of $\mathfrak{C}((M_n)_n,X)$ are of the form $C = \coprod_n C_n$
with $C_n \in \mathfrak{C}(M_n,X)$, and we additionally require that $C$ satisfies
the \emph{uniform metric control conditon}: There is some $\alpha > 0$,
independent of $n$, such that for all $((m,x,t)$, $(m',x',t')) \in C$ we
have $d(m,m') < \alpha$. Members of $\mathfrak{S}((M_n)_n,X)$ are sets of the form $S
= \coprod_n S_n$ with $S_n \in \mathfrak{S}(M_n,X)$. The resulting category
$\mathcal{R}^G(W,\mathbb{J}((M_n)_n,X))$ has a canonical faithful functor into the product category $\prod_n \mathcal{R}^G(W,\mathbb{J}(M_n,X))$.
Fix a symmetric, finite generating set $S$ of $G$. Let $d_G$ denote the word metric on $G$ with respect to $S$. Since $G$ is homotopy transfer reducible by assumption, there exists a natural number $N\in \mathbb{N}$ such that we can choose for each $n\in \mathbb{N}$
\begin{enumerate}
\item a compact, contractible metric space $(X_n,d_{X_n})$ such that for every $\epsilon > 0$ there is an $\epsilon$--controlled domination of $X_n$ by an at most $N$--dimensional, finite simplicial complex;
\item a homotopy coherent $G$--action $\Gamma_n$ on $X_n$;
\item a $G$--simplicial complex $\Sigma_n$ of dimension at most $N$ whose isotropy is contained in $\mathcal{F}$;
\item a map $f_n \colon X \to \Sigma_n$ which is $(S,n)$--equivariant,
i.e.,
\begin{enumerate}
\item for all $x \in X_n$ and $s \in S^n$,
\begin{equation}\label{eq:homotopy-trans-reduc:almost-equivariant}
d^{\ell^1}(f(\Gamma_n(s,x)),s \cdot f_n(x)) \leq \frac{1}{n};
\end{equation}
\item for all $x \in X_n$ and $s_0,\dots,s_n \in S^n$,
\begin{equation}\label{eq:homotopy-trans-reduc:diameter-bounded}
\diam \{ f_n(\Gamma_n(s_n,t_n,\dots,s_0,x)) \mid (t_1,\dots,t_n) \in [0,1]^n \} \leq \frac{2}{n}.
\end{equation}
\end{enumerate}
\end{enumerate}
\begin{definition}\label{def:metric-on-Sigma-G}
We equip $\Sigma_n \times G$ with the metric $ n \cdot d^{\ell^1}(x,y) +
d_G(g,h) $.
\end{definition}
Recall that an extended metric satisfies the usual axioms of a metric, but
it is allowed to take the value $\infty$. The following definition will be
used to produce a metric on $X_n \times G$ for each $n \in \mathbb{N}$.
\begin{dfn}\label{def-metric}
Let $(X,d_X)$ be a metric space, $\Gamma$ a homotopy coherent $G$--action
on $X$, and $S \subset G$ a finite subset containing the trivial element.
Let $k \in \mathbb{N}$ and $\Lambda > 0$. Define on $X \times G$ the extended metric
\begin{equation*}
d_{S,k,\Lambda}( (x,g), (y,h) ) \in [0,\infty]
\end{equation*}
to be the infimum over the numbers
\begin{equation*}
l + \sum_{i=0}^l \Lambda \cdot d_X(x_i,z_i),
\end{equation*}
where the infimum is taken over all $l \in \mathbb{N}$, $x_0,\dots,x_l$, $z_0,\dots,z_l \in X$ and $a_1,\dots,a_l$, $b_1,\dots,b_l \in S$ such that
\begin{enumerate}
\item $x_0 = x$ and $z_l = y$;
\item $ga_1^{-1}b_1\dots a_l^{-1}b_l = h$;
\item for each $1 \leq i \leq l$ there are elements $r_0,\dots,r_k,s_0,\dots,s_k \in S$ such that $a_i = r_k\dots r_0$, $b_i = s_k \dots s_0$ and $\Gamma(r_k,t_k,\dots,r_0,z_{i-1}) = \Gamma(s_k,u_k,\dots,s_0,x_i)$ for some $t_1, \dots,t_k,u_1,\dots,u_k \in [0,1]$.
\end{enumerate}
If no such data exist, take the infimum to be $\infty$.
\end{dfn}
This definition is analogous to
\cite[Definition~3.4]{Bartels-Lueck(2012annals)} and
\cite[Definition~2.3]{Wegner(2012)}. Since we only consider the coherent
$G$--action $\Gamma_n$ on $X_n$, we drop $\Gamma_n$ from the notation of
\cite{Wegner(2012)}. The proof of the next lemma is analogous to the one
given in \cite[Lemma~3.5]{Bartels-Lueck(2012annals)}.
\begin{lem}\label{lem:metricisnice}
Let $k \in \mathbb{N}$.
\begin{enumerate}
\item For all $\Lambda > 0$, the function $d_{S,k,\Lambda}$ is an extended metric on $X \times G$ which is $G$--invariant if we let $G$ act on $X \times G$ by $\gamma \cdot (x,g) = (x,\gamma g)$. It is a metric if and only if $S$ generates $G$.
\item We have $d_{S,k,\Lambda}((x,g),(y,h)) < 1$ if and only if $g = h$ and $\Lambda \cdot d_X(x,y) < 1$ holds, in which case we have $d_{S,k,\Lambda}((x,g),(y,h)) = \Lambda \cdot d_X(x,y)$. In particular, the topology induced by $d_{S,k,\Lambda}$ is the product topology. \qed
\end{enumerate}
\end{lem}
\subsection{The actual target of the transfer}
\label{subsec:proof-fj:actual-target-of-the-transfer}
We now specialize the construction of
\Cref{subsec:proof-fj:target-of-the-transfer} to our needs.
Assume that $G$ is homotopy transfer reducible, i.e., it satisfies
\Cref{def:transfer-reducible}. That definition provides us for every $n$
with a metric space $X_n$, as well as $\Gamma_n$, $f_n$ and $\Sigma_n$.
From \Cref{def-metric} and \Cref{lem:metricisnice} we obtain for any
sequence $(\Lambda_n)_n$ a sequence of metric spaces $(X_n \times G, d_{S^n,
n, \Lambda_n})_n$.
Although we do not need to restrict to a specific choice of
$(\Lambda_n)_n$ until a little later, we wish to avoid spreading our choices
throughout the whole proof. Therefore, we will now fix a specific sequence
$(\Lambda_n)_n$.
Since each $X_n$ is compact, $f_n$ is uniformly continuous. Hence, there
exists for each $n \in \mathbb{N}$ some $\delta_n > 0$ such that for all $x,y \in
X_n$ with $d_X(x,y) < \delta_n$ we have $d^{\ell^1}(f_n(x),f_n(y)) <
\frac{1}{n}$.
\begin{definition}\label{def:lambda-delta}
Choose such $\delta_n$ for all $n$ and set
\begin{equation*}
\Lambda_n := \frac{n+1}{\delta_n}.
\end{equation*}
Define a metric $d_n$ on $X_n \times G$ by
\begin{equation*}
d_n((x,g),(y,h)) := d_{S^n,n,\Lambda_n}((x,g),(y,h)) + d_G(g,h).
\end{equation*}
\end{definition}
Then $X_n \times G$ carries a free and isometric $G$--action if we let $G$
act on the right factor. If we make no explicit mention of a metric, we will
view $X_n \times G$ as a metric space with respect to $d_n$ in what follows.
Similarly, $\Sigma_n \times G$ carries a diagonal $G$--action and will
always be understood as a metric space with respect to the metric $n\cdot
d^{\ell^1} + d_G$ from \Cref{def:metric-on-Sigma-G}. Abbreviate $E :=
E_\mathcal{F}(G)$.
The category
$\mathcal{R}^G(W,\mathbb{J}((X_n \times G)_n,E))$ will be the target of the ``transfer''. However,
we need to equip it with another class of weak equivalences.
These $h^{fin}$--equivalences were introduced in the proof of
\cite[Theorem~10.1]{Ullmann-Winges(2015)}. Basically, they ignore the
behavior of an object on finitely many factors and behave like
$h$--equivalences otherwise.
\begin{dfn}\label{def:truncated-object}
Let $(M_n)_n$ be a sequence of metric spaces with free, isometric $G$--action (e.g.~$M_n = X_n \times G$).
Let $(Y_n)_n$ be an object of $\mathcal{R}^G(W, \mathbb{J}((M_n)_n,E))$. For $\nu \in \mathbb{N}$, we denote by $(-)_{n > \nu}$ the endofunctor which sends $(Y_n)_n$ to the sequence $(\widetilde{Y}_n)_n$ with $\widetilde{Y}_n = \ast$ for $n \leq \nu$ and $\widetilde{Y}_n = Y_n$ for $n > \nu$.
A morphism $(f_n)_n \colon (Y_n)_n \rightarrow (Y_n')_n$ is an \emph{$h^{fin}$--equivalence} if there is some $\nu \in \mathbb{N}$, such that $(f_n)_{n>\nu}\colon (Y_n)_{n>\nu} \rightarrow (Y_n')_{n>\nu}$ is an $h$--equivalence.
\end{dfn}
\begin{lem}\label{lem:functoriality-of-obstruction-category}
Let $(M_n)_n$, $(N_n)_n$ be sequences of metric spaces with free, isometric
$G$-action. Let $(g_n)_n\colon (M_n)_n \to (N_n)_n$ be a uniformly expanding
sequence of $G$--equivariant maps, i.e., for every $\alpha > 0$ there is
some $\beta > 0$ such that for all $n \in \mathbb{N}$ and $x,y \in M_n$ we have
$d(g_n(x),g_n(y)) < \beta$ whenever $d(x,y) < \alpha$.
Then $(g_n)_n$ induces a map $\mathcal{R}^G(W, \mathbb{J}((M_n)_n,E)) \to \mathcal{R}^G(W,
\mathbb{J}((N_n)_n,E))$ which also respects $h$-- and $h^{fin}$--equivalences, as
well as finiteness conditions.
\end{lem}
\begin{proof}
As $(g_n)_n$ induces a map on $\mathbb{J}((M_n)_n, E)$ which respects the control
conditions, it also respects the $h$--equivalences. As it maps $M_n$ to
$N_n$, it also respects the $h^{fin}$--equivalences.
\end{proof}
We will discuss the difference between the $h$-- and the
$h^{fin}$--equivalences in \Cref{subsec:proof-fj:squeezing}.
\subsection{Non-connective algebraic $K$--theory of controlled CW--complexes}
\label{subsec:proof-fj:negative-A-groups}
Before we turn to the main theorem, we need to briefly recall the definition
of algebraic $K$--theory in our setting. Let $\mathfrak{Z} = (Z, \mathfrak{C}, \mathfrak{S})$ be a
coarse structure. Then $\mathcal{R}^G_f(W,\mathfrak{Z})$ and its variants are Waldhausen
categories, hence their algebraic $K$--theory is defined
by~\cite{Waldhausen(1985)}. However, we need the non-connective delooping from
\cite[Section~5]{Ullmann-Winges(2015)}, which we briefly recall for
completeness.
\begin{dfn}\label{def:delooping-coarse-structure}
Let $\mathfrak{Z} = (Z, \mathfrak{C}, \mathfrak{S})$ be a coarse structure. For $n \in \mathbb{N}$ define the coarse structure $\mathfrak{Z}(n) = (\mathbb{R}^n \times Z, \mathfrak{C}(n), \mathfrak{S}(n))$ as follows:
A set $C \subset (\mathbb{R}^n \times Z)^2$ is in $\mathfrak{C}(n)$ if and only if:
\begin{enumerate}
\item $C$ is symmetric, $G$--invariant and contains the diagonal.
\item $C \subseteq p_n^{-1}(C')$ for some $C' \in \mathfrak{C}_{bdd}(\mathbb{R}^n)$, where $p_n \colon \mathbb{R}^n \times Z \to \mathbb{R}^n$ is the projection map.
\item For all $K \subset \mathbb{R}^n$ compact, there is a $C' \in \mathfrak{C}$ such that
\begin{equation*}
C \cap ((K \times Z) \times (K \times Z)) \subset p_Z^{-1}(C'),
\end{equation*}
where $p_Z \colon \mathbb{R}^n \times Z \to Z$ is the projection map.
\end{enumerate}
Let $\mathfrak{S}(n)$ be the collection of all $S \subset \mathbb{R}^n \times Z$ such that $S = p_Z^{-1}(S')$ for some $S' \in \mathfrak{S}(Z)$.
\end{dfn}
Consider for all $n$ also the restricted coarse structures
\begin{equation*}
\begin{split}
\mathfrak{Z}(n+1)^+ &:= \mathfrak{Z}(n+1) \cap (\mathbb{R}^n \times \mathbb{R}_{\geq 0} \times Z), \\
\mathfrak{Z}(n+1)^- &:= \mathfrak{Z}(n+1) \cap (\mathbb{R}^n \times \mathbb{R}_{\leq 0} \times Z). \\
\end{split}
\end{equation*}
Note that $\mathfrak{Z}(n+1) \cap (\mathbb{R}^n \times \{ 0 \} \times Z) = \mathfrak{Z}(n)$. The inclusion maps give rise to a commutative square
\begin{equation*}\label{diag:deloopingsquare}
\begin{tikzpicture}
\matrix (m) [matrix of math nodes, column sep=2em, row sep=2em, text depth=.5em, text height=1em]
{h S_\bullet \mathcal{R}^G_f(W,\mathfrak{Z}(n)) & h S_\bullet \mathcal{R}^G_f(W,\mathfrak{Z}(n+1)^+) \\ h S_\bullet \mathcal{R}^G_f(W,\mathfrak{Z}(n+1)^-) & h S_\bullet \mathcal{R}^G_f(W,\mathfrak{Z}(n+1)). \\};
\path[->]
(m-1-1) edge (m-1-2)
(m-1-1) edge (m-2-1)
(m-1-2) edge (m-2-2)
(m-2-1) edge (m-2-2);
\end{tikzpicture}
\end{equation*}
By an Eilenberg swindle, the top right and bottom left corners of this square are contractible. This provides us with structure maps for a spectrum
\begin{equation*}
\mathbb{K}^{-\infty}(\mathcal{R}^G_f(W,\mathfrak{Z}),h)_n := K(\mathcal{R}^G_f(W,\mathfrak{Z}(n)), h)
\end{equation*}
which we call the \emph{non-connective algebraic $K$--theory spectrum} of $\mathcal{R}^G_f(W,\mathfrak{Z})$. The construction is functorial in $\mathfrak{Z}$.
Non-connective algebraic $K$--theory is a functor on coarse structures and
morphisms of coarse structures by \cite[Section~5]{Ullmann-Winges(2015)}.
All the arguments which will follow do not interact with a possible
$\mathbb{R}^n$--coordinate, hence can also be carried out for $n > 0$, similar
to \cite[Section~9]{Ullmann-Winges(2015)}.
From the next section onwards, our proofs will only treat the case $n=0$.
\subsection{The main theorem}
\label{subsec:proof-fj:main-theorem}
In this part we show the following result.
\begin{thm}\label{thm:afjc-transfer-reducible}
Let $G$ be a discrete group and let $\mathcal{F}$ be a family of subgroups of $G$.
If $G$ is homotopy transfer reducible over $\mathcal{F}$, then $G$ satisfies the
Farrell-Jones \Cref{con:FJC_for_A-theory_with_c} with coefficients in $A$--theory with respect to $\mathcal{F}$.
\end{thm}
\Cref{the:main_result}~\ref{the:main_result:groups} follows from
\Cref{thm:afjc-transfer-reducible} in conjunction with the inheritance properties established in \Cref{sec:Inheritance_properties_of_the_Isomorphism_Conjectures}, cf.~the introduction. We derive the validity of \Cref{the:main_result}~\ref{the:main_result:groups} for hyperbolic and $\mathrm{CAT}(0)$--groups in \Cref{cor:hypcatfollowproof} below.
We follow the strategy of \cite[Section~5]{Wegner(2012)}. We construct a
commutative diagram of Waldhausen categories and exact functors
\[
\begin{tikzpicture}
\matrix (m) [matrix of math nodes, column sep=2em, row sep=2em, text depth=.5em, text height=1em, ampersand replacement=\&]
{(\mathcal{R}_{fd}^G(W, \mathbb{J}((X_n \times G)_n, E)), h^{fin}) \&
(\mathcal{R}_{fd}^G(W, \mathbb{J}((\Sigma_n \times G)_n, E)), h^{fin}) \\
\&
(\mathcal{R}_{fd}^G(W, \mathbb{J}((G)_n, E)), h^{fin})\\
(\mathcal{R}^G_f(W, \mathbb{J}(E)),h) \&
(\mathcal{R}^G_{fd}(W, \mathbb{J}(E)),h).\\
};
\path[->]
(m-1-1) edge node[above]{$F$} (m-1-2)
(m-1-1) edge node[below]{$p_{X_n \times G\rightarrow G}\;\;\;\;\;$} (m-2-2)
(m-3-2) edge node[right]{$\Delta$} (m-2-2)
(m-1-2) edge node[right]{$p_{\Sigma_n \times G\rightarrow G}$} (m-2-2)
(m-3-1) edge node[above]{$\incl$} (m-3-2);
\path[dashed][->]
(m-3-1) edge node[left]{$\trans$} (m-1-1);
\end{tikzpicture}
\]
We define the maps $\trans$, $\Delta$ and $F$ below
and show the following.
\begin{prop}\label{prop:main-diag-claims}
\mbox{}
\begin{enumerate}
\item\label{item:main-diag-claims-1} The arrow $\trans$ exists after
applying non-connective algebraic $K$--theory. It will be induced by a
map of spectra whose domain is weakly equivalent to
$\mathbb{K}^{-\infty}(\mathcal{R}^G_f(W, \mathbb{J}(E)),h)$. The square formed by $p_{X_n
\times G\rightarrow G} \circ \trans$ and $\Delta \circ \incl$ commutes up
to levelwise weak equivalence of spectra.
\item\label{item:main-diag-claims-2} The functor $F$, defined below,
is well-defined.
\item\label{item:main-diag-claims-3} The algebraic $K$--theory of $\mathcal{R}_{fd}^G(W, \mathbb{J}((\Sigma_n \times G)_n, E), h^{fin})$ vanishes.
\item\label{item:main-diag-claims-4} The map $\Delta\circ \incl$ is
injective on non-connective algebraic $K$--theory groups.
\end{enumerate}
\end{prop}
Given all of this, the proof can be finished as in
\cite[Section~5]{Wegner(2012)}. It is a diagram chase on the level of
homotopy groups.
Before proving \Cref{prop:main-diag-claims}, let us define the maps of the
diagram. The maps $p$ are induced by the indicated projections on control
spaces, ``$\incl$'' is the inclusion of the finite into the finitely dominated
objects. The functor $\Delta$ is induced by the diagonal map into $\prod_n
\mathcal{R}^G_{fd}(W,\mathbb{J}(G_n,X))$, which factors over $(\mathcal{R}_{fd}^G(W, \mathbb{J}((G)_n, E)),
h^{fin})$ because its image consists of uniformly controlled objects and
maps. Also note that every $h$--equivalence is an $h^{fin}$--equivalence.
The functor $F$ is defined using the maps $f_n$ from
\Cref{subsec:proof-fj:target-of-the-transfer}. It is induced by the maps $F_n
\colon X_n \times G \to \Sigma_n \times G$, $(x,g) \mapsto (g f_n(x), g)$.
We show in \Cref{subsec:proof-fj:squeezing} that with our choices these are
uniformly bounded.
The map ``$\trans$'' is constructed in \Cref{The Transfer: Final Part of
the Proof}.
\subsection{Squeezing}
\label{subsec:proof-fj:squeezing}
All claims made in \Cref{prop:main-diag-claims} except part
\ref{item:main-diag-claims-1} admit fairly short proofs which we give in
this section. Part \ref{item:main-diag-claims-1} will be shown in
\Cref{subsec:transfer-map}
\Cref{prop:main-diag-claims}~\ref{item:main-diag-claims-3} follows from the
``Squeezing Theorem'' \cite[Theorem~10.1]{Ullmann-Winges(2015)} and the fact
that non-connective $K$--theory does not distinguish between finite and
finitely dominated objects \cite[Remark~5.5]{Ullmann-Winges(2015)}.
Indeed,
in the proof of Theorem 10.1 of \cite{Ullmann-Winges(2015)}, in equation (21), a
homotopy fiber sequence
\begin{multline}\label{eq:homotopy-fiber-sequence}
\mathbb{K}^{-\infty}\left(\operatorname{colim}_n \prod_{k=1}^n \mathcal{R}_f^G(W, \mathbb{J}(M_k,E)), h \right)
\longrightarrow \\
\mathbb{K}^{-\infty}(\mathcal{R}_f^G(W, \mathbb{J}((M_n)_n,E)), h) \longrightarrow
\mathbb{K}^{-\infty}(\mathcal{R}_f^G(W, \mathbb{J}((M_n)_n,E)), h^{fin})
\end{multline}
is established for any sequence of metric spaces $(M_n)_n$. Then it is
shown there that under the assumptions from
\Cref{subsec:proof-fj:target-of-the-transfer} on $(\Sigma_n \times G)_n$,
the first map is a weak equivalence by proving that the last object is
weakly contractible.
Let us discuss \Cref{prop:main-diag-claims}~\ref{item:main-diag-claims-2}
next. It suffices to show the following.
\begin{lem}
The map $(F_n)_n \colon \mathbb{J}((X_n \times G)_n, E)\rightarrow \mathbb{J}((\Sigma_n \times G)_n, E)$ is a morphism of coarse structures.
\end{lem}
\begin{proof}
By \Cref{lem:functoriality-of-obstruction-category}, it suffices to check
that $(F_n)_n$ is a uniformly expanding sequence. Since the the proof is
fairly lengthy (though still straightforward), we give the details.
Let us recall the definitions. The metric $d_n$ on $X_n \times G$ was
defined in \Cref{def:lambda-delta}, the metric $n\cdot d^{\ell^1} + d_G$ on
$\Sigma_n \times G$ was chosen in \Cref{def:metric-on-Sigma-G}. Let
$\alpha > 0$.
Let $n \in \mathbb{N}$ and $(x,g)$, $(y,h) \in X_n \times G$. Suppose that
$d_n((x,g),(y,h)) < \alpha$. To prove that $F_n$ is uniformly expanding we
have to show that $n \cdot d^{\ell^1}(gf_n(x),hf_n(y)) + d_G(g,h) \leq
\beta$ for some $\beta > 0$ which is independent of $n$.
In fact, it suffices to show this for $n \geq \alpha$. Then we have by
\Cref{def:lambda-delta}
\[
\frac{\alpha}{\Lambda_n} = \frac{\alpha \delta_n}{n+1} \leq
\frac{n\delta_n}{n+1} < \delta_n.
\]
By definition of $d_n$, we have
$d_G(g,h) < \alpha$ and $d_{S^n,n,\Lambda_n}((x,g),(y,h)) < \alpha$.
Hence, there exist
\begin{itemize}
\item $l \in \mathbb{N}$
\item $x_0,\dots,x_l$, $z_0,\dots,z_l \in X_n$
\item $a_1,\dots,a_l$, $b_1,\dots,b_l \in S^n$
\end{itemize}
such that
\begin{enumerate}
\item $x_0 = x$, $z_l = y$
\item $g a_1^{-1} b_1 \dots a_l^{-1} b_l = h$
\item for each $1 \leq i \leq l$ there are elements
$r_0,\dots,r_n,s_0,\dots,s_n \in S^n$ such that $a_i = r_n\dots r_0$,
$b_i = s_n \dots s_0$ and, for some $t_1, \dots,t_n,u_1,\dots,u_n \in
[0,1]$, $\Gamma_n(r_n,t_n,\dots,t_0,z_{i-1}) =
\Gamma_n(s_n,u_n,\dots,s_0,x_i)$ holds
\item $l + \sum_{i=0}^l \Lambda_n \cdot d_{X_n}(x_i,z_i) < \alpha$
\end{enumerate}
This implies $l < \alpha$, $d_{X_n}(x_i, z_i) < \frac{\alpha}{\Lambda_n} <
\delta_n$. By \Cref{def:lambda-delta} of $\delta_n$, this implies
$d^{\ell^1}(f_n(x_i),f_n(z_i)) < \frac{1}{n}$.
We proceed by induction on $l$. For $l = 0$, we have $g=h$ and
$d^{\ell^1}(f_n(x),f_n(y)) < \frac{1}{n}$.
For the induction step, with $a_1 = r_n\dots r_0$ and $b_1 = s_n \dots s_0$
we have
\begin{equation*}
\begin{split}
d^{\ell^1}(g f_n(x), h f_n(y))
&= d^{\ell^1}(g f_n(x), g a_1^{-1} b_1 \dots a_l^{-1} b_l f_n(y) ) \\
&= d^{\ell^1}(f_n(x), a_1^{-1} b_1 \dots a_l^{-1} b_l f_n(y) ) \\
%
&\leq d^{\ell^1}(f_n(x_0),f_n(z_0)) \\
&\quad + d^{\ell^1}(f_n(z_0), a_1^{-1} f_n(\Gamma_n(r_n,1,\dots,1,r_0,z_0)) ) \\
&\quad + d^{\ell^1}( f_n(\Gamma_n(r_n,1,\dots,1,r_0,z_0)) , f_n(\Gamma_n(r_n,t_n,\dots,t_1,r_0,z_0)) ) \\
&\quad + d^{\ell^1}( f_n(\Gamma_n(s_n,u_n,\dots,u_1,s_0,x_1)), f_n(\Gamma_n(s_n,1,\dots,1,s_0,x_1)) ) \\
&\quad + d^{\ell^1}( f_n(\Gamma_n(s_n,1,\dots,1,s_0,x_1)), b_1 f_n(x_1) ) \\
&\quad + d^{\ell^1}( f_n(x_1), f_n(z_1) ) \\
&\quad + d^{\ell^1}( f_n(z_1), a_2^{-1} b_2 \dots a_l^{-1} b_l f_n(y) )
\end{split}
\end{equation*}
We give an estimate for each summand. We already know
\begin{equation*}
d^{\ell^1}(f_n(x_0),f_n(z_0)) < \frac{1}{n}, \qquad
d^{\ell^1}( f_n(x_1), f_n(z_1) )< \frac{1}{n}.
\end{equation*}
For the second summand, we have by
\Cref{eq:homotopy-trans-reduc:almost-equivariant}
\begin{equation*}
\begin{split}
d^{\ell^1}(f_n(z_0), a_1^{-1} f_n(\Gamma_n(r_n,1,\dots,1,r_0,z_0)) )
&= d^{\ell^1}(a_1 \cdot f_n(z_0), f_n(\Gamma_n(a_1,z_0)))
\leq \frac{1}{n},
\end{split}
\end{equation*}
similarly for $d^{\ell^1}( f(\Gamma_n(s_n,1,\dots,1,s_0,x_1)), b_1 f(x_1) )$.
Furthermore, we have
\begin{equation*}
\begin{split}
d^{\ell^1}( f_n(\Gamma_n(r_n,1,\dots,1,r_0,z_0)) , f_n(\Gamma_n(r_n,t_n,\dots,t_1,r_0,z_0)) ) &\leq \frac{2}{n} \\
d^{\ell^1}( f_n(\Gamma_n(s_n,u_n,\dots,u_1,s_0,x_1)), f_n(\Gamma_n(s_n,1,\dots,1,s_0,x_1)) ) &\leq \frac{2}{n}
\end{split}
\end{equation*}
by \Cref{eq:homotopy-trans-reduc:diameter-bounded}. Finally, we choose the
induction hypothesis to be
\[
d^{\ell^1}( f_n(z_1), a_2^{-1} b_2 \dots a_l^{-1} b_l f_n(y) ) <
\frac{8(l-1) + 1}{n}.
\]
Thus we obtain
\[
d^{\ell^1}(g f_n(x), h f_n(y))<\frac{8l + 1}{n}.
\]
Since we also have $d_G(g,h) < \alpha$, we conclude that
\begin{equation*}
n \cdot d^{\ell^1}(gf_n(x),hf_n(y)) + d_G(g,h) < 9\alpha + 1.\qedhere
\end{equation*}
\end{proof}
\subsection{Injectivity of the $\Delta$-map}
\label{subsec:proof-fj:injectivity-delta}
Now we show \Cref{prop:main-diag-claims}~\ref{item:main-diag-claims-4}. Namely, we have to show that $\Delta$ induces an injective map on algebraic $K$--theory. Our argument is a straightforward adaptation of the argument used in \cite[Section~5]{Wegner(2012)}.
As usual, we abbreviate $\pi_m(\mathbb{K}^{-\infty}(\dots))$ by $K_m(\dots)$.
\begin{lem}
The map
\[
K_m(\Delta)\circ K_m(incl)\colon K_m(\mathcal{R}^G_f(W, \mathbb{J}(E)),h)\rightarrow K_m(\mathcal{R}_{fd}^G(W, \mathbb{J}((G)_n, E)), h^{fin})
\]
is injective for each $m \geq 0$.
\end{lem}
\begin{proof}
The map $K_m(incl)\colon K_m(\mathcal{R}^G_f(W, \mathbb{J}(E)),h)\rightarrow
K_m(\mathcal{R}^G_{fd}(W, \mathbb{J}(E)),h)$ is an isomorphism by
\cite[Remark~5.5]{Ullmann-Winges(2015)}. Hence, we only have to show that
$K_m(\Delta)$ is injective. To increase readability, we shorten
$\mathcal{R}_{fd}^G(W,...)$ to $\mathcal{R}(...)$ in the following commutative diagram:
\[
\begin{tikzpicture}
\matrix (m) [matrix of math nodes, column sep=2em, row sep=2em, text depth=.5em, text height=1em, ampersand replacement=\&]
{\& K_m(\mathcal{R}(\mathbb{J}(E)),h)\\K_m(\prod^{fin}\mathcal{R}( \mathbb{J}(E)),h) \& K_m(\mathcal{R}(\mathbb{J}((G)_n, E)), h) \& K_m(\mathcal{R}(\mathbb{J}((G)_n, E)), h^{fin}) \\ \bigoplus_{n\in\mathbb{N}}K_m(\mathcal{R}(\mathbb{J}(E)),h) \& \prod_{n\in\mathbb{N}}K_m(\mathcal{R}(\mathbb{J}(E)),h) \\};
\path[->]
(m-1-2) edge node[right]{$\Delta_\ast$} (m-2-2)
(m-1-2) edge node[above]{$\Delta_\ast$} (m-2-3)
(m-2-1) edge node[right]{$\cong$} (m-3-1)
(m-2-2) edge node[right]{$\prod_{n\in\mathbb{N}}p_n$} (m-3-2)
(m-2-2) edge node[below]{$\operatorname{id}_\ast$} (m-2-3)
(m-3-1) edge node[above]{$\incl$} (m-3-2)
(m-2-1) edge node[above]{$\incl$} (m-2-2);
\end{tikzpicture}
\]
The middle row is exact due to the homotopy fiber sequence
\eqref{eq:homotopy-fiber-sequence}, where we abbreviated $ \operatorname{colim}_n
\prod_{k=1}^n $ as $\prod^{fin}$. The left vertical map is an isomorphism,
because algebraic $K$--theory commutes with directed colimits and is
compatible with finite products. The map is defined using the projections
onto the factors of the product. Note that after projection on any $n$, the
middle column is the identity. A diagram chase finishes the proof.
\end{proof}
\subsection{Homotopy transfer reducible follows from strongly transfer reducible}
\label{subsec:proof-fj:homotopy-transfer-follows-strongly-transfer}
We can now prove \Cref{the:main_result}~\ref{the:main_result:groups} for hyperbolic and $\mathrm{CAT}(0)$--groups.
In \cite[Definition~3.1]{Wegner(2012)}, Wegner defined when a group $G$ is
\emph{strongly transfer reducible over a family $\mathcal{F}$}. As we will not need
the precise definition here, we refer to \emph{loc.~cit.} for the
definition. We will use the definitions from
\Cref{subsec:proof-fj:homotopy-coherent}.
\begin{thm}\label{thm:strongly-trans-reducible}
Let $G$ be strongly transfer reducible over $\mathcal{F}$.
Then $G$ is homotopy transfer reducible over $\mathcal{F}$ and
the Farrell-Jones \Cref{con:FJC_for_A-theory_with_c} for $A$--theory
with coefficients holds for $G$ relative to $\mathcal{F}$.
\end{thm}
\begin{proof}
Assume that $G$ is strongly transfer reducible. We show it is homotopy
transfer reducible and apply \Cref{thm:afjc-transfer-reducible}.
According to \cite[Proposition~3.6]{Wegner(2012)}, there exists $N \in \mathbb{N}$ such that there are for every $n \in \mathbb{N}$
\begin{enumerate}
\item a compact, contractible metric space $(X,d_X)$ such that for every $\epsilon > 0$ there is an $\epsilon$--controlled domination of $X$ by an at most $N$--dimensional, finite simplicial complex;
\item a homotopy coherent $G$--action $\Gamma$ on $X$;
\item a $G$--simplicial complex $\Sigma$ of dimension at most $N$ whose isotropy is contained in $\mathcal{F}$;
\item a positive real number $\Lambda$;
\item a $G$--equivariant map $\phi \colon G \times X \to \Sigma$ such that
\[
n \cdot d^{\ell^1}(\phi(g,x), \phi(h,y)) \leq d_{S^n,n,\Lambda}((g,x),(h,y))
\]
holds for all $(g,x)$, $(h,y) \in G \times X$, where $G$ acts on the
$G$--factor.
\end{enumerate}
Fix $n \in \mathbb{N}$ and choose $X$, $\Gamma$, $\Sigma$, $\Lambda$ and $\phi$ as
above. Define $f := \phi|_{\{e\} \times X} \colon X \to \Sigma$. Then we
have for all $x \in X$ and $s \in S^n$ \begin{equation*}
\begin{split}
n \cdot d^{\ell^1}(f(\Gamma(s,x)), s \cdot f(x))
&= n \cdot d^{\ell^1}(\phi(e,\Gamma(s,x)),\phi(s,x)) \\
&\leq d_{S^n,n,\Lambda}((e,\Gamma(s,x)),(s,x)) \\
&\leq 1.
\end{split}
\end{equation*}
Similarly, we find for all $x \in X$, $s_0,\dots,s_n \in S^n$ and $t_1,\dots,t_n,u_1,\dots,u_n \in I^n$
\begin{equation*}
\begin{split}
n \cdot d^{\ell^1}&(f(\Gamma(s_n,t_n,\dots,s_0,x)), f(\Gamma(s_n,u_n,\dots,s_0,x)) ) \\
&= n \cdot d^{\ell^1}(\phi(e,\Gamma(s_n,t_n,\dots,s_0,x)),\phi(e,\Gamma(s_n,u_n,\dots,s_0,x)) ) \\
&\leq d_{S^n,n,\Lambda}((e,\Gamma(s_n,t_n,\dots,s_0,x)),(e,\Gamma(s_n,u_n,\dots,s_0,x)) ) \\
&\leq 2.
\end{split}
\end{equation*}
Hence, $G$ is homotopy transfer reducible over $\mathcal{F}$ and we can apply \Cref{thm:afjc-transfer-reducible}.
\end{proof}
\begin{cor}\label{cor:hypcatfollowproof}
The Farrell-Jones \Cref{con:FJC_for_A-theory_with_c_and_fwp} for
$A$--theory with coefficients and finite wreath products is true for
hyperbolic and $\mathrm{CAT}(0)$--groups.
\end{cor}
\begin{proof}
By \cite[Example~3.2 and Theorem~3.4]{Wegner(2012)}, finitely generated
hyperbolic groups as well as $\mathrm{CAT}(0)$--groups are strongly
transfer reducible with respect to the family of virtually cyclic
subgroups. Thus these groups are homotopy
transfer reducible over the same family by
\Cref{thm:strongly-trans-reducible}. If a group is homotopy transfer
reducible with respect to $\mathcal{F}$, then the wreath product $G \wr F$ with a
finite group $F$ is homotopy transfer reducible over $\mathcal{F} \wr F$. This
follows as in \cite[Section 5]{Bartels-Lueck-Reich-Rueping(2014)};
basically, one takes the $F$--fold product of $X$ and $\Sigma$ and uses e.g.
\cite[Lemma 11.14]{Ullmann-Winges(2015)} for the estimate. As the
$A$--theoretic Farrell--Jones Conjecture with coefficients holds for
virtually finitely generated abelian groups
\cite[Proposition~11.9]{Ullmann-Winges(2015)}, the $A$--theoretic
Farrell--Jones Conjecture with coefficients and finite wreath products
holds for hyperbolic and $\mathrm{CAT}(0)$--groups by the
Transitivity Principle, \cite[Proposition~11.2]{Ullmann-Winges(2015)}.
\end{proof}
\section{The transfer: Final part of the proof}\label{The Transfer: Final Part of the Proof}
We turn now to the construction of the transfer map whose existence was claimed in the first part of \Cref{prop:main-diag-claims}~\ref{item:main-diag-claims-1}. This map will be induced by a ``transfer'' construction on controlled retractive spaces. Here we employ an analog of the classical construction of the transfer, see \cite{Bartels-Lueck(2012annals)}, \cite{Bartels-Lueck-Reich(2008hyper)} and \cite{Wegner(2012)}, on each cell
and glue these together according to the CW--structure. Unfortunately, this construction is not fully functorial on $\mathcal{R}^G_f(W,\mathbb{J}(E))$. To avoid this problem, we restrict our attention to subcategories of ``cellwise $0$--controlled morphisms'' (defined below) as a domain for the transfer. The idea to use these categories is due to Arthur Bartels and Paul Bubenzer.
The transfer is defined using the ideas of the linear counterpart.
However, they only work well for one cell at a time. To extend, we need to
refine the idea of crossing a controlled CW-complex with the singular
complex of a metric space, by allowing different, but compatible singular
complexes for each cell. While this provides us with a transfer on objects,
only ``cellwise $0$-controlled morphisms'' behave well with respect to
this construction. We could also transfer any map, but then the target
gets a more lax control condition, and this makes the construction
non-functorial.
We show that transferring only ``cellwise $0$-controlled morphisms'' is
enough to construct the transfer, but transferring the other morphisms is
needed to show that it preserves weak equivalences.
\subsection{The domain of the transfer}
We now define the appropriate subcategories of cellwise $0$--controlled morphisms, which will serve as the source of the transfer.
Let $M$ be a metric space with a free, isometric $G$--action, and consider the category $\mathcal{R}^G(W,\mathbb{J}(M,E))$. For $(Y,\kappa) \in \mathcal{R}^G(W,\mathbb{J}(M,E))$, let $\kappa_M$ denote the composition of the control map $\kappa$ with the projection map $M \times E \times [1,\infty[ \to M$.
\begin{dfn}\label{def:0-controlled-morphism}
Let $f \colon (Y_1,\kappa_1) \to (Y_2,\kappa_2)$ be a morphism in
$\mathcal{R}^G_f(W,\mathbb{J}(M,E))$. We say that $f$ is \emph{regular}, if the image
of each open cell in $Y_1$ is either equal to an open cell in $Y_2$ or completely
contained in $W$. That is, either $f(\mathrm{int}\ e)= \mathrm{int}\ e'$ or
$f(\mathrm{int}\ e) \subseteq W$.
We say that $f$ is \emph{cellwise $0$--controlled over $M$} if $f$ is regular and satisfies the property that $\kappa_{1,M}(e) = \kappa_{2,M}(f(e))$ for all cells $e \in \diamond\, Y_1$.
\end{dfn}
The composition of two morphisms which are cellwise $0$--controlled over $M$ is again cellwise $0$--controlled over $M$, so we can consider the subcategory
\begin{equation*}
\mathcal{R}^G_f(W,\mathbb{J}(M,E))_0 \subset \mathcal{R}^G_f(W,\mathbb{J}(M,E))
\end{equation*}
which has the same objects as $\mathcal{R}^G_f(W,\mathbb{J}(M,E))$, but contains only those
morphisms which are cellwise $0$--controlled over $M$. The category
$\mathcal{R}^G_f(W,\mathbb{J}(M,E))_0$ inherits cofibrations and weak equivalences from $\mathcal{R}^G_f(W,\mathbb{J}(M,E))$. It is a Waldhausen subcategory of $\mathcal{R}^G_f(W,\mathbb{J}(M,E))$.
For $\alpha > 0$, we may further restrict to the full subcategory
\begin{equation*}
\mathcal{R}^G_f(W,\mathbb{J}(M,E))_\alpha \subset \mathcal{R}^G_f(W,\mathbb{J}(M,E))_0
\end{equation*}
consisting only of those objects which are \emph{$\alpha$--controlled over
$M$}, i.e., those $(Y,\kappa)$ such that $\kappa_M(\diamond\, \gen{e}) \subset
B_\alpha(\kappa_M(e))$ for every cell $e \in \diamond\, Y$. (Recall that
$\gen{e}$ denotes the smallest subcomplex of $Y$ containing $e$.) The
category $\mathcal{R}^G_f(W,\mathbb{J}(M,E))_\alpha$ also inherits the structure of a
Waldhausen category, as the pushout of $\alpha$--controlled
complexes along cellwise $0$--controlled morphisms is again
$\alpha$--controlled.
Finally, we can filter $\mathcal{R}^G_f(W,\mathbb{J}(M,E))_\alpha$ by
\begin{equation*}
\mathcal{R}^G_f(W,\mathbb{J}(M,E))_{\alpha,0} \subset \mathcal{R}^G_f(W,\mathbb{J}(M,E))_{\alpha,1} \subset \dots \subset \mathcal{R}^G_f(W,\mathbb{J}(M,E))_\alpha,
\end{equation*}
where $\mathcal{R}^G_f(W,\mathbb{J}(M,E))_{\alpha,d}$ denotes the full subcategory of $\mathcal{R}^G_f(W,\mathbb{J}(M,E))_\alpha$ containing those objects whose dimension is at most $d$. Note that
\begin{equation*}
\mathcal{R}^G_f(W,\mathbb{J}(M,E))_\alpha = \operatorname{colim}_d \mathcal{R}^G_f(W,\mathbb{J}(M,E))_{\alpha,d},
\end{equation*}
as each object in $\mathcal{R}^G_f(W,\mathbb{J}(M,E))$ is finite-dimensional.
\begin{prop}\label{prop:domain-of-transfer}
There is a natural weak equivalence
\begin{equation*}
\operatorname{hocolim}_{\alpha,d} K(\mathcal{R}^G_f(W,\mathbb{J}(M,E))_{\alpha,d}) \xrightarrow{\sim} K(\mathcal{R}^G_f(W,\mathbb{J}(M,E))).
\end{equation*}
\end{prop}
\begin{proof}
We have $\mathcal{R}^G_f(W,\mathbb{J}(M,E))_0 = \operatorname{colim}_{\alpha,d} \mathcal{R}^G_f(W,\mathbb{J}(M,E))_{\alpha,d}$. Since $K$--theory commutes with directed colimits, we obtain a natural weak equivalence
\begin{equation*}
\operatorname{hocolim}_{\alpha,d} K(\mathcal{R}^G_f(W,\mathbb{J}(M,E))_{\alpha,d}) \xrightarrow{\sim} K(\mathcal{R}^G_f(W,\mathbb{J}(M,E))_0).
\end{equation*}
Now consider the inclusion functor $\mathcal{R}^G_f(W,\mathbb{J}(M,E))_0 \hookrightarrow
\mathcal{R}^G_f(W,\mathbb{J}(M,E))$. We show that Waldhausen's Approximation Theorem
\cite[Theorem~1.6.7]{Waldhausen(1985)} applies.
The cylinder functor on $\mathcal{R}^G_f(W,\mathbb{J}(M,E))$ constructed in
\cite[Lemma~3.14]{Ullmann-Winges(2015)} restricts to a cylinder
functor on $\mathcal{R}^G_f(W,\mathbb{J}(M,E))_0$, in particular the inclusion of the
source is always cellwise $0$-controlled. By definition, the inclusion functor
satisfies the first part of the approximation property. To verify the second
part of the approximation property, let $f \colon Y_1 \to Y_2$ be an
arbitrary morphism in $\mathcal{R}^G_f(W,\mathbb{J}(M,E))$. Then the factorization of $f$
via the cylinder functor $Y_1 \rightarrowtail Mf \xrightarrow{\sim} Y_2$
decomposes $f$ into
a cellwise $0$--controlled morphism and a weak equivalence. So the
Approximation Theorem implies that the inclusion functor induces an
equivalence on algebraic $K$--theory.
\end{proof}
\begin{rem}
The upshot of \Cref{prop:domain-of-transfer} is that we do not have to define a ``global'' transfer functor on $\mathcal{R}^G_f(W,\mathbb{J}(E))$. Instead, it suffices to define a transfer functor $\trans^{\alpha, d}\colon \mathcal{R}^G_f(W,\mathbb{J}(E))_{\alpha,d}\rightarrow (\mathcal{R}^G_{fd}(W, \mathbb{J}((X_n\times G)_n, E)), h^{fin})$ on each subcategory, such that the induced diagrams on $K$--theory
\begin{equation*}\label{eq:hocomm}
\begin{tikzpicture}
\matrix (m) [matrix of math nodes, column sep=4.5em, row sep=2em, text depth=.5em, text height=1em]
{K(\mathcal{R}^G_f(W,\mathbb{J}(E))_{\alpha,d}, h)& K(\mathcal{R}^G_{fd}(W, \mathbb{J}((X_n\times G)_n, E)), h^{fin}) \\ K(\mathcal{R}^G_f(W,\mathbb{J}(E))_{\alpha+1,d+1}, h)\\};
\path[->]
(m-1-1) edge node[above]{$K(\trans^{\alpha,d})$} (m-1-2)
(m-1-1) edge (m-2-1)
(m-2-1) edge node[below right]{$K(\trans^{\alpha+1,d+1})$} (m-1-2);
\end{tikzpicture}
\end{equation*}
are homotopy commutative.
\end{rem}
\subsection{Balanced products of CW--complexes}
\label{sec:balanced-products}
We introduce a slight generalization of the balanced products discussed in
\cite{Davis-Lueck(1998)} as a means to define the transfer in
\Cref{subsec:transfer-on-objects}.
Let $W$ be a topological space and $\mathcal{C}$ a small category.
A \emph{$\mathcal{C}$--CW--complex relative $W$} is a functor $Y$ from $\mathcal{C}$ to
topological spaces such that $Y(c)$ is a CW--complex relative $W$ and the
morphisms in $\mathcal{C}$ are mapped to cellular maps relative $W$. A (relative)
free $\mathcal{C}$--$n$--cell based at $c$, $c \in \mathcal{C}$, is a pair $(\eta, \partial
\eta)$ of $\mathcal{C}$--CW--complexes relative $W$, where $\eta = \mathcal{C}(c,-) \times
D^{n} \amalg W$, $\partial \eta =\mathcal{C}(c,-) \times S^{n-1} \amalg W$. Attaching
a free $\mathcal{C}$--cell $\eta$ to $Y$ means taking the pushout along a map
$\partial\eta \to Y$. Note that $W$ itself defines a (constant) covariant
$\mathcal{C}$--CW--complex relative $W$.
We say that $Y$ is a \emph{free} $\mathcal{C}$--CW--complex relative $W$ if it comes
equipped with a filtration $W = \skel{-1}{Y} \subset \skel{0}{Y} \subset
\skel{1}{Y} \subset \dots$ such that $Y = \operatorname{colim}_n \skel{n}{Y}$ and for every
$n \geq 0$
there exists a pushout in the category of $\mathcal{C}$--CW--complexes relative $W$
\begin{equation*}
\begin{tikzpicture}
\matrix (m) [matrix of math nodes, column sep=3em, row sep=2em, text depth=.5em, text height=1em]
{\left(\coprod_{i \in I_n} \mathcal{C}(c_i,-) \times S^{n-1}\right) \amalg W &
\skel{n-1}{Y}(-) \\
\left( \coprod_{i \in I_n} \mathcal{C}(c_i,-) \times D^n \right)
\amalg W& \skel{n}{Y}(-). \\};
\path[->]
(m-1-1) edge (m-1-2)
(m-1-1) edge (m-2-1)
(m-1-2) edge (m-2-2)
(m-2-1) edge (m-2-2);
\end{tikzpicture}
\end{equation*}
Hence, a free $\mathcal{C}$--CW--complex arises by attaching free $\mathcal{C}$--cells. The set of free
$\mathcal{C}$--$n$--cells is in bijection with $I_n$. Note that the attaching map of a
$\mathcal{C}$--$n$--cell based at $c$ is the same as a map $S^{n-1} \to
\skel{n-1}{Y}(c)$, hence we can consider $\eta$ as a map $D^n \to Y(c)$.
Let $Y$ be a covariant $\mathcal{C}$--CW--complex relative $W$ and $X \colon \mathcal{C}^{op}
\to \curs{CW}\text{-}\curs{COMPLEXES}$ be a contravariant $\mathcal{C}$--CW--complex. Define the
\emph{reduced balanced product} $X \rightthreetimes_\mathcal{C} Y$ as
the pushout
\begin{equation*}
\begin{tikzpicture}
\matrix (m) [matrix of math nodes, column sep=2em, row sep=2em, text depth=.5em, text height=1em]
{X \times_\mathcal{C} W & X \times_\mathcal{C} Y \\ * \times_\mathcal{C} W \cong W & X
\rightthreetimes_\mathcal{C} Y. \\};
\path[->]
(m-1-1) edge (m-1-2)
(m-1-1) edge (m-2-1)
(m-1-2) edge (m-2-2)
(m-2-1) edge (m-2-2);
\end{tikzpicture}
\end{equation*}
\begin{prop}
Let $X$ be a contravariant $\mathcal{C}$--space, $Y$ a covariant $\mathcal{C}$--CW--complex
relative $W$ and $Z$ a space relative $W$. There is a natural homeomorphism
\begin{equation*}
\hom^W(X \rightthreetimes_{\mathcal{C}} Y, Z) \cong
\hom_{\mathcal{C}}^W(Y, \hom(X,Z)).
\end{equation*}
Here, $\hom(X,Z)$ is a covariant $\mathcal{C}$--space relative $W$ via the inclusion that sends a
point $w \in W(c)$ to the constant map $X(c) \to \{w\} \subseteq Z$,
$\hom^W_{\mathcal{C}}$ denotes the natural transformations which are
relative $W$,
and $\hom^W$ denotes just the set
of maps relative $W$.
\end{prop}
\begin{proof}
By definition, a map $X \rightthreetimes_{\mathcal{C}} Y \to Z$ is the same as
three compatible maps from $W \leftarrow X \times_{\mathcal{C}} W \to X \times_{\mathcal{C}}
Y$ to $Z$. Using that $\hom(X\times_{\mathcal{C}}Y, Z)$ is isomorphic to
$\hom_{\mathcal{C}}(Y, \hom(X,Z))$, the result is easy to deduce.
\end{proof}
It follows, that $X \rightthreetimes_{\mathcal{C}} Y$ commutes with
colimits in the ``$Y$''-variable. We can therefore determine the cell
structure of $X \rightthreetimes_{\mathcal{C}} Y$. The attachment of a free
$\mathcal{C}$--$n$--cell $\eta$ to $Y$ gives a pushout $X \rightthreetimes_{\mathcal{C}}
\eta(-) \cup_{X \rightthreetimes_{\mathcal{C}} \partial \eta} Y$. Now
\begin{equation*}
X \rightthreetimes_{\mathcal{C}} \eta \cong
((X \times_{\mathcal{C}} \mathcal{C}(c,-) )\times D^n) \amalg W \cong
(X(c) \times D^n) \amalg W
\end{equation*}
and similarly for $\partial \eta$. First, this gives a filtration on $X
\rightthreetimes_{\mathcal{C}} Y$, namely
\begin{equation}
\ldots \subseteq X \rightthreetimes_{\mathcal{C}} \skel{n-1}{Y} \subseteq X
\rightthreetimes_{\mathcal{C}} \skel{n}{Y} \subseteq \ldots
\label{eq:balance_product_skeletal_filtration}
\end{equation}
Second, as $X(c)$ is a CW--complex, we can now read off the cell structure of $X
\rightthreetimes_{\mathcal{C}} Y$:
\begin{prop}[{cf.~\cite[Lemma~3.19(2)]{Davis-Lueck(1998)}}]\label{prop:balanced-product-cells}
Let $Y$ be a free covariant $\mathcal{C}$--CW--complex relative $W$ and $X$ a contravariant $\mathcal{C}$--CW--complex.
Then $X \rightthreetimes_\mathcal{C} Y$ is a CW--complex relative $W$, and there is a canonical identification
\begin{equation*}
\diamond\, (X \rightthreetimes_\mathcal{C} Y) \cong \{ (\xi,\eta) \mid \eta \text{ is a free $\mathcal{C}$--cell based at $c$}, \xi \in \diamond\, X(c) \}.
\end{equation*}
Let $(\xi,\eta) \in \diamond\, (X \rightthreetimes_\mathcal{C} Y)$. If $\Phi \colon D^p \to X(c)$ and $\Psi \colon \mathcal{C}(c,-) \times D^q \to Y(-)$ are characteristic maps for $\xi$ and $\eta$, respectively, then
\begin{equation*}
D^p \times D^q \to X \rightthreetimes_\mathcal{C} Y, \quad (a,b) \mapsto [\Phi(a),\Psi(\operatorname{id}_c,b)]
\end{equation*}
is a characteristic map for $(\xi,\eta)$.
Let $(\xi,\eta)$, $(\xi',\eta') \in \diamond\, (X \rightthreetimes_\mathcal{C} Y)$ be
two cells, with $\eta$ based at $c$ and $\eta'$ based at $c'$. Then
$(\xi,\eta) \subset \gen{(\xi',\eta')}$ if and only if there exists a
morphism $\gamma \colon c \to c'$ such that $\gamma_* \eta \subset
\gen{\eta'} \subset Y(c')$ and $\xi \subset \gen{\gamma^*\xi'}
\subset X(c)$.
\end{prop}
In greater generality, \eqref{eq:balance_product_skeletal_filtration} gives a
filtration for an inclusion $Y_1 \hookrightarrow Y_2$ of $\mathcal{C}$--spaces in which $Y_2$ is obtained from $Y_1$ by the attachment of free $\mathcal{C}$--cells. This observation allows us to translate the constructions for geometric modules to CW--complexes.
\medskip
Let us conclude this section with a short remark about functoriality of the balanced product construction. In addition to the obvious functoriality properties, we have the following: Let $X$ be a contravariant and $Y$ be a covariant $\mathcal{C}$--space. Let $F \colon \mathcal{D} \to \mathcal{C}$ be a functor. Then there is an induced map
\begin{equation*}
\iota_F \colon F^*X \times_\mathcal{D} F^*Y \to X \times_\mathcal{C} Y, \quad [d,x,y] \mapsto [F(d),x,y].
\end{equation*}
This map is functorial in the sense that $\iota_{F_2}\iota_{F_1} =
\iota_{F_2F_1}$ for any two composable functors $F_1$, $F_2$. In particular,
if $F \colon \mathcal{C} \xrightarrow{\cong} \mathcal{C}$ is an automorphism of the indexing
category, then $\iota_F$ is an isomorphism.
\subsection{Conventions}\label{subsec:transfers}
For the following sections, fix the following data:
\begin{enumerate}
\item natural numbers $\alpha,d \in \mathbb{N}$ and a natural number $n > \max \{ d+1, \alpha \}$
\item a natural number $N \in \mathbb{N}$
\item a compact and contractible metric space $(X,d_X)$ such that for every $\epsilon > 0$ there is an $\epsilon$--controlled domination of $X$ by an at most $N$--dimensional, finite simplicial complex
\item a homotopy coherent $G$--action $\Gamma$ on $X$
\item a positive real number $\Lambda$
\end{enumerate}
As before, we consider $X \times G$ equipped with the metric $d_{S^n,n,\Lambda} + d_G$.
\subsection{$Y$ as a $\diamond_+ Y$--CW--complex and a $(\diamond_+
Y)^{op}$--CW--complex from $X$}
\label{subsec:y-spaces}
Let $(Y, \kappa) \in \mathcal{R}^G_f(W,\mathbb{J}(E))_{\alpha,d}$. If $c \in \diamond\, Y$ is a cell of $Y$, we will frequently need to refer to the $G$--component of $\kappa(c)$; we denote this by $\kappa_G(c)$.
Define a relation $\leq$ on the set of cells $\diamond\, Y$ by saying that $c \leq
c'$ if and only if $c \subset \gen{c'}$. Then $\diamond\, Y$ forms a poset under
the relation $\leq$. We define $\diamond_+\, Y$ as the category given by this
poset where we add an additional initial object (which corresponds to $W$).
The complex $Y$ itself gives rise to a covariant $\diamond_+\, Y$--CW--complex (relative $W$) $\mathcal{C}_Y$ by setting
\begin{equation*}
\mathcal{C}_Y(c) := \gen{c}
\end{equation*}
and sending a morphism $c \leq c'$ to the obvious inclusion $\gen{c}
\hookrightarrow \gen{c'}$. Observe that $\mathcal{C}_Y$ is a free $\diamond_+\,
Y$--CW--complex; the set of free $\diamond_+\, Y$--cells of $\mathcal{C}_Y$ is in canonical
bijection with the cells of $Y$. Note that a cellwise
$0$-controlled map $Y \to Y'$ gives rise to a functor $\diamond_+\, Y \to \diamond_+\,
Y'$. Last, each cell in $\diamond_+\, Y$ has a dimension $|c|$,
where we assign the initial object the dimension $-1$.
The metric space $X$ gives rise to a contravariant $\diamond_+\, Y$--CW--complex, but the construction is more involved. We mimic the construction used in \cite{Wegner(2012)}, but do not pass to the cellular chain complex. Instead, we simply stick with the space of controlled simplices.
In the first step, we pass from the homotopy coherent $G$--action $\Gamma$ on $X$ to an honest $G$--action on a closely related space. This is accomplished by strictifying the homotopy coherent diagram $\Gamma$, see \cite[proof of Proposition~5.4]{Vogt(1973)}. Define $M\Gamma$ to be the space
\begin{equation*}
M\Gamma := \left( \coprod_{k \geq 0}
G^{k+1} \times [0,1]^k \times X \right)\big/ \sim,
\end{equation*}
where $\sim$ is the equivalence relation generated by
\begin{equation*}
(\gamma_{k+1}, t_k, \gamma_k,\dots,\gamma_1,x)
\sim
\begin{cases}
(\gamma_{k+1},t_k,\dots,\gamma_2,x) & \gamma_1 = e \\
(\gamma_{k+1},\dots,t_it_{i-1},\dots,\gamma_1,x) & \gamma_i = e, 2 \leq i \leq k \\
(\gamma_{k+1},\dots,\gamma_{i+1}\gamma_i,\dots,\gamma_1,x)
& t_i = 1, 1 \leq i \leq k \\
(\gamma_{k+1},\dots,\gamma_{i+1},
\Gamma(\gamma_i,\dots,\gamma_1,x)) & t_i = 0, 1
\leq i \leq k.
\end{cases}
\end{equation*}
Then $G$ acts on $M\Gamma$ by
\begin{equation*}
g \cdot [\gamma,t_k,\gamma_k,\dots,\gamma_1,x] := [g \gamma, t_k, \gamma_k, \dots, \gamma_1,x].
\end{equation*}
We have a map $X \to M \Gamma$ via $x \mapsto [e,x]$. Let
$R \colon M\Gamma \to X$ be the retraction induced by $\Gamma$; explicitly,
$R([\gamma,t_k,\gamma_k,\dots,\gamma_1,x]) =
\Gamma(\gamma,t_k,\gamma_k,\dots,\gamma_1,x)$. Using the axioms of a homotopy
action from~\Cref{def:hptycohG-action}, one checks this is a well-defined map.
The homotopy
\begin{equation}\label{eq:strong-deformation-rectraction-MGamma}
H \colon M\Gamma \times [0,1] \to M\Gamma, \quad ([\gamma,t_k,\gamma_k,\dots,\gamma_1,x],u) \mapsto [e,u,\gamma,t_k,\gamma_k,\dots,\gamma_1,x]
\end{equation}
then shows that $X$ is a strong deformation retract of $M\Gamma$.
The space $M\Gamma$ comes with a filtration by subspaces $M\Gamma^{l,r}$,
where we set
\begin{equation*}
M\Gamma^{l,r} := \{ [e, t_k, \gamma_k,\dots,\gamma_1,x]
\in M\Gamma \mid k \leq l, \gamma_i \in B_{r}(e) \}.
\end{equation*}
For $\delta > 0$, define $S^\delta_\bullet(M\Gamma^{l,r})$ to be the
subsimplicial set of the singular simplicial set
$S_\bullet(M\Gamma^{l,r})$ containing those singular simplices $\sigma
\colon \Delta^{\abs{\sigma}} \to M\Gamma^{l,r}$ which fulfill
\begin{equation*}
\diam_{X \times G}( (R \circ \sigma)(\Delta^{\abs{\sigma}}) \times \{
e
\} ) \leq \delta,
\end{equation*}
where diameters in $X \times G$ are taken with respect to the metric
$d_{S^n,n,\Lambda}$. Note that we could replace $e$ by any other
group element without changing the diameter, as the metric is $G$-invariant.
Finally, we can define the contravariant $\diamond_+\, Y$--CW--complex
$\sing{\alpha,d}{X,Y}$: Let $\delta_k^{d}:=4(d+1-k)$
and $l_k^{d}:=d+1-k$. Typically, we will omit
$d$ from the notation. On objects, we set
\begin{equation*}
\sing{\alpha,d}{X,Y}(c) := \norm{ S^{\delta_{\abs{c}}}_\bullet(M\Gamma^{l_{\abs{c}},\alpha}) },
\end{equation*}
where $\abs{c}$ denotes the dimension of the cell $c$ and $\norm{-}$ is fat
geometric realization, i.e., the realization after forgetting the
degeneracies. Note that we have the canonical inclusion $\iota_{c'}$ of
$\sing{\alpha,d}{X,Y}(c')$ into $\norm{ S^{}_\bullet(M\Gamma^l) }$. The
latter has an honest $G$--action. For a morphism $c' \to c = c \leq
c'$ in $(\diamond_+\, Y)^{op}$
define $\sing{\alpha,d}{X,Y}(c' \to c)$
as the factorization of
\begin{equation*}
\sing{\alpha,d}{X,Y}(c' \to c) := \kappa_G(c)^{-1}\kappa_G(c') \cdot
\iota_{c'}(-)
\end{equation*}
over $\iota_{c}$.
We have to check that it is well-defined, i.e., that it actually
factors. We require the following observation.
\begin{lem}\label{lem:distance-to-translate}
Let $[e,t_{b},\dots,\gamma_1,x] \in M\Gamma^{l,\alpha}$. Suppose that $l < n$ and $\alpha\leq n$. Let $h \in B_\alpha(e)$. Then
\begin{equation*}
d_{S^n,n,\Lambda}(R([h,t_{b},\dots,\gamma_1,x],g),
(R([e,t_{b},\dots,\gamma_1,x]),gh) ) \leq 2
\end{equation*}
for all $g \in G$.
\end{lem}
\begin{proof}
Note that $b < n$.
We use the definition of the metric. Let
$x_0=z_0=\Gamma(h,t_{b},\gamma_{b},\dots,\gamma_1,x)$,
$x_1=z_1=x$ and $x_2=z_2=
\Gamma(e,t_{b},\gamma_{b},\dots,\gamma_1,x)$.
Furthermore, we set $a_1=e$, $b_1=h\gamma_{b}\dots\gamma_1$,
$a_2=\gamma_{b}\dots\gamma_1$ and $b_2=e$. Now we can estimate
\begin{equation*}
\begin{split}
d&_{S^n,n,\Lambda}\big(
(\Gamma(h,t_{b},\gamma_{b},\dots,\gamma_1,x),g),
(\Gamma(e,t_{b},\gamma_{b},\dots,\gamma_1,x),gh) \big) \\
&\leq 2 + \Lambda \cdot
d_X(\Gamma(h,t_{b},\gamma_{b},\dots,\gamma_1,x),\Gamma(h,t_{b},\gamma_{b},\dots,\gamma_1,x)) \\
&\quad+ \Lambda \cdot d_X(x,x) + \Lambda \cdot
d_X(\Gamma(e,t_{b},\gamma_{b},\dots,\gamma_1,x),\Gamma(e,t_{b},\gamma_{b},\dots,\gamma_1,x)) \\
&=2.
\end{split}
\end{equation*}
\end{proof}
\begin{cor}
Assume $ \alpha \leq n$ and $d + 1 < n$. Then the functor
$\sing{\alpha,d}{X,Y} \colon (\diamond_+\, Y)^{op} \to \mathrm{CW}$ is
well-defined.
\end{cor}
\begin{proof}
Since $M\Gamma$ carries an honest $G$--action, functoriality is clear as soon as we have convinced ourselves that $\sing{\alpha,d}{X,Y}$ is well-defined on morphisms.
Let $c' \to c$ be a morphism in $(\diamond\, Y)^{op}$, and let $\sigma \in S^{\delta_{\abs{c'}}}_\bullet(M\Gamma^{l_{\abs{c'}},\alpha})$.
We only need to check non-identity morphisms. Hence we assume $|c| \geq
|c'| + 1$.
Let $[e,t_{b},\gamma_{b},\dots,\gamma_1,x]$ be a point in the image of $\sigma$. By definition, we have $b \leq l_{\abs{c'}}$ and $\gamma_i \in B_\alpha(e)$ for all $i$.
Set $\gamma_{c,c'} := \kappa_G(c)^{-1}\kappa_G(c')$. Note that $\gamma_{c,c'} \in B_\alpha(e)$ since $Y$ is $\alpha$--controlled over $G$.
Then we obtain
\begin{equation*}
\begin{split}
\gamma_{c,c'} \cdot &[e,t_{b},\gamma_{b},\dots,\gamma_1,x] \\
&= [\gamma_{c,c'},t_{b},\gamma_{b},\dots,\gamma_1,x] \\
&= [e,1,\gamma_{c,c'},t_{b},\gamma_{b},\dots,\gamma_1,x] \in M\Gamma^{l_{\abs{c'}+1},\alpha} \subset M\Gamma^{l_{\abs{c}},\alpha}.
\end{split}
\end{equation*}
Hence, $\gamma_{c,c'} \cdot \sigma$ is a singular simplex in $M\Gamma^{l_{\abs{c}},\alpha}$.
Let $[e,t'_{b'},\gamma'_{b'},\dots,\gamma'_1,x']$ be another point in the image of $\sigma$. Then
\begin{equation*}
R(\gamma_{c,c'} \cdot [e,t_{b},\gamma_{b},\dots,\gamma_1,x]) = \Gamma(\gamma_{c,c'},t_{b},\gamma_{b},\dots,\gamma_1,x),
\end{equation*}
and similarly for $[e,t'_{b'},\gamma'_{b'},\dots,\gamma'_1,x']$. We calculate
\begin{equation*}
\begin{split}
d&_{S^n,n,\Lambda}\big( (\Gamma(\gamma_{c,c'},t_{b},\gamma_{b},\dots,\gamma_1,x),\kappa_G(c)), (\Gamma(\gamma_{c,c'},t'_{b'},\gamma'_{b'},\dots,\gamma'_1,x'),\kappa_G(c)) \big) \\
&\leq d_{S^n,n,\Lambda}\big( (\Gamma(\gamma_{c,c'},t_{b},\gamma_{b},\dots,\gamma_1,x),\kappa_G(c)), (\Gamma(e,t_{b},\gamma_{b},\dots,\gamma_1,x),\kappa_G(c')) \big) \\
&\quad + d_{S^n,n,\Lambda}\big( (\Gamma(e,t_{b},\gamma_{b},\dots,\gamma_1,x),\kappa_G(c')), (\Gamma(e,t'_{b'},\gamma'_{b'},\dots,\gamma'_1,x'),\kappa_G(c')) \big) \\
&\quad + d_{S^n,n,\Lambda}\big( (\Gamma(e,t'_{b'},\gamma'_{b'},\dots,\gamma'_1,x'),\kappa_G(c')), (\Gamma(\gamma_{c,c'},t'_{b'},\gamma'_{b'},\dots,\gamma'_1,x'),\kappa_G(c)) \big) \\
&\leq 2 + \delta_{\abs{c'}} + 2 \\
&= 4(d + 1 - (\abs{c'}-1)) \\
&\leq 4(d + 1 - \abs{c}) = \delta_{\abs{c}},
\end{split}
\end{equation*}
where we used \Cref{lem:distance-to-translate} for the second inequality. This shows that multiplication by $\gamma_{c,c'}$ indeed defines a map
\begin{equation*}
\gamma_{c,c'} \cdot - \colon S^{\delta_{\abs{c'}}}_\bullet(M\Gamma^{l_{\abs{c'}},\alpha}) \to S^{\delta_{\abs{c}}}_\bullet(M\Gamma^{l_{\abs{c}},\alpha}).
\end{equation*}
So the functor $\sing{\alpha,d}{X,Y}$ is well-defined.
\end{proof}
\subsection{The transfer on objects}
\label{subsec:transfer-on-objects}
Recall that by our assumptions in \Cref{subsec:transfers} we have $\alpha < n,
d+1 < n$.
\begin{dfn}\label{dfn:transfer}
The \emph{transfer $\trans^{\alpha,d}_X(Y)$ of $Y$ with respect to $X$} is defined to be
\begin{equation*}
\trans^{\alpha,d}_X(Y) := \sing{\alpha,d}{X,Y}
\rightthreetimes_{\diamond_+\, Y} \mathcal{C}_Y.
\end{equation*}
If $\alpha$, $d$ or both of them are understood, we abbreviate $\trans^{\alpha,d}_X(Y)$ to $\trans^d_X(Y)$, $\trans^\alpha_X(Y)$ or $\trans_X(Y)$, respectively.
\end{dfn}
Since $\mathcal{C}_Y$ is a free $\diamond_+\, Y$--CW--complex, the space $\trans_X(Y)$
is a CW--complex relative $W$ by \Cref{prop:balanced-product-cells}. The
natural transformation $\sing{\alpha,d}{X,Y} \to *$ to the constant functor
with value the one-point space induces a map $\sing{\alpha,d}{X,Y}
\rightthreetimes_{\diamond_+\, Y} \mathcal{C}_Y \to * \rightthreetimes_{\diamond_+\,
Y} \mathcal{C}_Y \cong Y$ of CW--complexes relative $W$. We regard
$\sing{\alpha,d}{X,Y} \rightthreetimes_{\diamond_+\, Y} \mathcal{C}_Y$ as a retractive space via this map.
We equip $\trans_X(Y)$ with a $G$--action as follows. Observe that $G$ acts on
the indexing category $\diamond\, Y$; let $\mu_g \colon \diamond\, Y \to \diamond\, Y$
denote the functor induced by the action of $g \in G$. The action of $g$ on $Y$ induces a natural isomorphism $\mathcal{C}_Y \xrightarrow{\cong} \mathcal{C}_Y \circ \mu_g$, and hence a cellular homeomorphism
\begin{equation*}
\tau_g \colon \sing{\alpha,d}{X,Y} \rightthreetimes_{\diamond_+\, Y} \mathcal{C}_Y
\xrightarrow{\cong} \sing{\alpha,d}{X,Y} \rightthreetimes_{\diamond_+\, Y} (\mathcal{C}_Y \circ \mu_g).
\end{equation*}
Observing that $\sing{\alpha,d}{X,Y} \circ \mu_g = \sing{\alpha,d}{X,Y}$, we
obtain from the functoriality of $\rightthreetimes_{\diamond_+\, Y}$ in \Cref{sec:balanced-products} a cellular homeomorphism
\begin{equation*}
\iota_{\mu_g} \colon \sing{\alpha,d}{X,Y} \rightthreetimes_{\diamond_+\, Y}
(\mathcal{C}_Y \circ \mu_g) = (\sing{\alpha,d}{X,Y} \circ \mu_g)
\rightthreetimes_{\diamond_+\, Y} (\mathcal{C}_Y \circ \mu_g) \xrightarrow{\cong}
\sing{\alpha,d}{X,Y} \rightthreetimes_{\diamond_+\, Y} \mathcal{C}_Y.
\end{equation*}
Define the action map of $g \in G$ as the composition
\begin{equation*}
g \cdot - := \iota_{\mu_g} \circ \tau_g \colon \trans_X(Y) \xrightarrow{\cong} \trans_X(Y).
\end{equation*}
Explicitly, this map is given by $g \cdot [c,x,y] \mapsto [gc,x,gy]$, and defines a group action by cellular homeomorphisms.
Again by \Cref{prop:balanced-product-cells}, we have a canonical identification
\begin{equation*}
\diamond\, \trans_X(Y) \cong \{ (\sigma, c) \mid c \in \diamond\, Y, \sigma \in S^{\delta_{\abs{c}}}_\bullet(M\Gamma^{l_{\abs{c}},\alpha}) \},
\end{equation*}
which translates the $G$--action on the set of cells of $\trans_X(Y)$ to $g \cdot (\sigma, c) = (\sigma, gc)$. Hence, $\trans_X(Y)$ is a free $G$--CW--complex.
Continuing to use the above identification of $\diamond\, \trans_X(Y)$, we define a control map for $\trans_X(Y)$: Let $\beta_p$ denote the barycenter of the standard $p$--simplex. Then set
\begin{equation*}
\trans_X(\kappa) \colon \diamond\, \trans_X(Y) \to X \times G \times E \times [1,\infty[, \quad (\sigma,c) \mapsto ((R \circ \sigma)(\beta_{\abs{\sigma}}),\kappa(c)).
\end{equation*}
\begin{lem}\label{lem:transfer-control-objects}
The pair $(\trans_X(Y),\trans_X(\kappa))$ is an object in $\mathcal{R}^G(W,\mathbb{J}(X \times G,E))$ which is $(\alpha + \delta_0 + 2)$--controlled over $X \times G$.
\end{lem}
\begin{proof}
By construction, the labeled $G$--CW--complex $(\trans_X(Y),
\trans_X(\kappa))$ satisfies the $G$--continuous control condition. It also
has the correct support, since $X$ is compact.
So it is only necessary to check that it satisfies bounded control over $X \times G$. Let $(\sigma,c)$ and $(\sigma',c')$ be cells such that $(\sigma,c) \subset \gen{(\sigma',c')}$. By \Cref{prop:balanced-product-cells}, this is equivalent to the conditions $c \subset \gen{c'}$ and $\sigma \subset \gen{\kappa_G(c)^{-1}\kappa_G(c')\sigma'} \subset \norm{ S^{\delta_{\abs{c}}}_\bullet(M\Gamma^{l_{\abs{c}},\alpha}) }$. Set $\gamma_{c,c'} := \kappa_G(c)^{-1}\kappa_G(c')$. Then we have
\begin{equation*}
\begin{split}
d_{S^n,n,\Lambda}&((\trans_X(\kappa)(\sigma,c), \trans_X(\kappa)(\sigma',c')) \\
&= d_{S^n,n,\Lambda}\big( ((R \circ \sigma)(\beta_{\abs{\sigma}}),\kappa_G(c)), ((R \circ \sigma')(\beta_{\abs{\sigma'}}),\kappa_G(c')) \big) \\
&\leq d_{S^n,n,\Lambda}\big( ((R \circ \sigma)(\beta_{\abs{\sigma}}),\kappa_G(c)), (R \circ \gamma_{c,c'}\sigma')(\beta_{\abs{\sigma'}}), \kappa_G(c)) \big) \\
&\quad + d_{S^n,n,\Lambda}\big( (R \circ \gamma_{c,c'}\sigma')(\beta_{\abs{\sigma'}}), \kappa_G(c)), ((R \circ \sigma')(\beta_{\abs{\sigma'}}),\kappa_G(c')) \big) \\
&\leq \delta_{\abs{c}} + 2 \leq \delta_0 + 2,
\end{split}
\end{equation*}
where the last inequality follows from our assumption and
\Cref{lem:distance-to-translate}. Using $d_G(\kappa_G(c), \kappa_G(c')) \leq
\alpha$, we conclude that $(\trans_X(Y),\trans_X(\kappa))$ is $(\alpha+\delta_0+2)$--controlled over $X \times G$.
\end{proof}
\subsection{The transfer on cellwise $0$-controlled morphisms}
\label{subsec:transfer-cellwise-0}
In the next step, we extend the assignment $(Y,\kappa) \mapsto (\trans_X(Y), \trans_X(\kappa))$ to a functor $\mathcal{R}^G_f(W,\mathbb{J}(E))_{\alpha,d} \to \mathcal{R}^G(W,\mathbb{J}(X \times G, E))$.
Let $f \colon (Y_1,\kappa_1) \to (Y_2,\kappa_2)$ be a morphism in
$\mathcal{R}^G_f(W,\mathbb{J}(E))_{\alpha,d}$. Since $f$ is a regular map, we have an induced
functor $\diamond_+\, f \colon \diamond_+\, Y_1 \to \diamond_+\, Y_2$, which is compatible
with the $G$--actions. ($G$ acts trivially on the initial
object.)
Define a natural transformation $\mathcal{C}_f \colon \mathcal{C}_{Y_1} \to \mathcal{C}_{Y_2} \circ
\diamond_+\, f$ by
\begin{equation*}
\mathcal{C}_{f,c} \colon \mathcal{C}_{Y_1}(c) = \gen{c} \xrightarrow{f} f(\gen{c}) =
\gen{f(c)} = (\mathcal{C}_{Y_2} \circ \diamond_+\, f)(c).
\end{equation*}
Define a natural transformation $\sing{}{f} \colon \sing{\alpha,d}{X,Y_1} \to
\sing{\alpha,d}{X,Y_2} \circ \diamond_+\, f$ by
\begin{equation*}
\sing{}{f,c} \colon \norm{ S^{\delta_{\abs{c}}}(M\Gamma^{l_{\abs{c}},\alpha}) } \subset \norm{ S^{\delta_{\abs{f(c)}}}_\bullet(M\Gamma^{l_{\abs{f(c)}},\alpha}) }.
\end{equation*}
Then $\trans_X(f)$ is defined as the composition
\begin{equation*}
\sing{\alpha,d}{X,Y_1} \rightthreetimes_{\diamond_+\, Y_1} \mathcal{C}_{Y_1}
\xrightarrow{\sing{}{f} \rightthreetimes_{\diamond_+\, Y_1} \mathcal{C}_f}
(\sing{\alpha,d}{X,Y_2} \circ (\diamond_+\, f)) \rightthreetimes_{\diamond_+\,
Y_1} \mathcal{C}_{Y_2} \circ (\diamond_+\, f)) \xrightarrow{\iota_{(\diamond_+\, f)}}
\sing{\alpha,d}{X,Y_2} \rightthreetimes_{\diamond_+\, Y_2} \mathcal{C}_{Y_2}.
\end{equation*}
\begin{lem}\label{lem:transfer-functor}
This defines a functor
\begin{equation*}
\trans_X \colon \mathcal{R}^G_f(W,\mathbb{J}(E))_{\alpha,d} \to \mathcal{R}^G(W,\mathbb{J}(X \times G,E))_0.
\end{equation*}
\end{lem}
\begin{proof}
To see that the construction of $\trans_X(f)$ is functorial, it is best to translate the above formalism again into an explicit mapping rule. Concretely, $\trans_X(f)$ is given by $[c,x,y] \mapsto [f(c),x,f(y)]$, and functoriality becomes obvious.
We also need to check that $\trans_X(f)$ is a controlled map. It suffices to
consider bounded control over $X \times G$. Note that $\trans_X(f)$ is
regular as $f$ and $\sing{}{f,c}$ are regular. Hence it is enough to compute for $(\sigma, c) \in \diamond\, \trans_X(Y_1)$
\begin{equation*}
\begin{split}
d&_{S^n,n,\Lambda}(\trans_X(\kappa_1)(\sigma,c), \trans_X(\kappa_2)(\trans_X(f)(\sigma,c))) + d_G(\kappa_{1,G}(c),\kappa_{2,G}(f(c))) \\
&= d_{S^n,n,\Lambda}\big( ((R \circ
\sigma)(\beta_{\abs{\sigma}}),\kappa_{1,G}(c)), ((R \circ
\sigma)(\beta_{\abs{\sigma}}),\kappa_{2,G}(f(c))) \big) + 0\\
&= 0.
\end{split}
\end{equation*}
So $\trans_X(f)$ is in fact cellwise $0$--controlled over $X \times G$. Hence, we have defined a functor $\trans_X \colon \mathcal{R}^G_f(W,\mathbb{J}(E))_{\alpha,d} \to \mathcal{R}^G(W,\mathbb{J}(X \times G,E))_0$.
\end{proof}
\begin{rem}
One can adapt the constructions presented in this paper to chain complexes over geometric modules to obtain a linear transfer. The linearization map, which assigns to a CW--complex its cellular chain complex, translates our transfer functor into its linear counterpart.
Moreover, the natural inclusion of geometric modules into chain complexes makes these constructions compatible with the transfers defined in \cite{Bartels-Lueck(2012annals)}, \cite{Bartels-Lueck-Reich(2008hyper)} and \cite{Wegner(2012)}. Thus, the transfer for geometric $\mathbb{Z}[G]$--modules corresponds to our construction restricted to $0$--dimensional CW--complexes.
\end{rem}
\subsection{Transferring cofibrations}
Our next aim is to show that the transfer is a functor of categories with cofibrations. Since the cofibrations under consideration are essentially inclusions of CW--subcomplexes it comes as no surprise that this result relies on an analysis of the CW--structure of $\trans_X(Y)$.
Let $(Y, \kappa) \in \mathcal{R}^G_f(W,\mathbb{J}(E))$ as before. It defines a $\diamond_+\,
Y$--CW--complex $\mathcal{C}_Y$ relative $W$. Let $B\subseteq Y$ be a subcomplex. We
get a $\diamond_+\, Y$--CW--complex $\mathcal{C}_Y^B$ relative $W$ by setting
$\mathcal{C}_Y^B(c) := \gen{c} \cap B$. As before, $\operatorname{colim}_{\diamond_+\, Y} \mathcal{C}_Y^B
\cong B$.
Let $A \subseteq B \subseteq Y$ be subcomplexes. Assume that $B$
arises from $A$ by attaching cells $\eta_i, i \in I$. From
\Cref{sec:balanced-products} we get a pushout diagram
\begin{equation*}
\begin{tikzpicture}
\matrix (m) [matrix of math nodes, column sep=3em, row sep=2em, text depth=.5em, text height=1em]
{ {\coprod^W_{i\in I}\partial \eta_i} & \mathcal{C}^A_Y \\
{\coprod^W_{i\in I}\eta_i} & \mathcal{C}^B_Y \\};
\path[->]
(m-1-1) edge (m-1-2)
(m-1-1) edge (m-2-1)
(m-1-2) edge (m-2-2)
(m-2-1) edge (m-2-2);
\end{tikzpicture}
\end{equation*}
in $\diamond_+\, Y$--CW--complexes relative $W$. This becomes a pushout diagram
in retractive spaces, if we equip everything with the retractions into $W$
arising from $Y$. Now, $\sing{\alpha,d}{X,Y} \rightthreetimes_{\diamond_+\,
Y} (-) $ commutes with pushouts, so we get the following result.
\begin{lem}\label{lem:attaching-cells-transfer}
There is a pushout diagram
\begin{equation*
\begin{tikzpicture}
\matrix (m) [matrix of math nodes, column sep=3em, row sep=2em, text depth=.5em, text height=1em]
{\sing{\alpha,d}{X,Y} \rightthreetimes_{\diamond_+\, Y} \big(\coprod_{i \in
I}^W
\partial \eta_i \big) &
\sing{\alpha,d}{X,Y} \rightthreetimes_{\diamond_+\, Y} \mathcal{C}^A_Y \\
\sing{\alpha,d}{X,Y} \rightthreetimes_{\diamond_+\, Y} \big(\coprod_{i \in
I}^W
\eta_i \big) &
\sing{\alpha,d}{X,Y} \rightthreetimes_{\diamond_+\, Y} \mathcal{C}^B_Y \\};
\path[->]
(m-1-1) edge (m-1-2)
(m-1-1) edge (m-2-1)
(m-1-2) edge (m-2-2)
(m-2-1) edge (m-2-2);
\end{tikzpicture}
\end{equation*}
in $\mathcal{R}^G(W,\mathbb{J}(X \times G,E))$. Here the coproducts on the left are
disjoint unions over $W$. As these are cells, the space on the lower
left is isomorphic
to
\begin{equation*}
\coprod_{i \in I} \big( \norm{
S^{\delta^d_{\abs{c_i}}}_\bullet(M\Gamma^{l^d_{\abs{c_i}},\alpha}) }
\times D^{\abs{c_i}} \big) \amalg W ,
\end{equation*}
when $\eta_i$ is a cell
based at $c_i$, and similarly for the upper left. \qed
\end{lem}
This enables us to do inductive arguments over the cells in $Y$. Note that
if $A \subseteq Y$ is a subcomplex, we can interpret $A$ as a $\diamond_+\,
A$--CW--complex, or as $\diamond_+\, Y$--CW--complex. We can define the
transfer also for $A$ as a $\diamond_+\, Y$--CW--complex, and it is canonically
isomorphic to \Cref{dfn:transfer}.
\begin{lem}\label{lem:transfer-cofibrations}
The functor $\trans_X$ preserves the zero object, cofibrations and admissible pushout diagrams, i.e., it is a functor of categories with cofibrations.
\end{lem}
\begin{proof}
Let $f \colon (Y_1,\kappa_1) \rightarrowtail (Y_2,\kappa_2)$ be a cofibration
in $\mathcal{R}^G_f(W,\mathbb{J}(E))_{\alpha,d}$. Without loss of generality, we can assume
that $Y_2$ is obtained from $Y_1$ by attaching free $G$--cells, and that $f$
is the inclusion of the subcomplex $Y_1$ into $Y_2$. Then it follows from
\Cref{lem:attaching-cells-transfer} and interpreting $Y_1$ as a $\diamond_+\,
Y_2$--CW--complex that $\trans_X(f)$ is also a cofibration. For the same
reason, $\trans_X$ preserves all relevant pushout squares. Last, it maps the
zero object $W$ to $W$.
\end{proof}
\subsection{The transfer on general morphisms and weak equivalences}
Next, we construct natural transformations between our transfer functors for various indices. Once we have shown that they are weak equivalences, it follows that the diagram in \Cref{eq:hocomm} is homotopy commutative. In addition, these enter the proof that $\trans^{\alpha,d}$ preserves weak equivalences.
\begin{dfn}\label{def:transfer-vary-constraints}
Let $\alpha' > \alpha$ and $d' > d$ satisfying $n > \max \{d'+1, \alpha'
\}$. Then both $\sing{\alpha,d}{X,Y}$ and $\sing{\alpha',d'}{X,Y}$ are
defined and give rise to transfer functors $\trans^{\alpha,d}_X$ and
$\trans^{\alpha',d'}_X$. For every $(Y,\kappa) \in
\mathcal{R}^G_f(W,\mathbb{J}(E))_{\alpha,d}$, there is a natural transformation
$\sing{\alpha,d}{X,Y} \to \sing{\alpha',d'}{X,Y}$ which is given at $c \in
\diamond_+\, Y$ by the obvious inclusion
\begin{equation*}
\norm{ S^{\delta^{d}_{\abs{c}}}_\bullet(M\Gamma^{l^{d}_{\abs{c}},\alpha}) } \subset \norm{ S^{\delta^{d'}_{\abs{c}}}_\bullet(M\Gamma^{l^{d'}_{\abs{c}},\alpha'}) }.
\end{equation*}
Hence, we obtain an induced natural morphism
\begin{equation*}
\rho^{\alpha,\alpha',d,d'}_Y \colon \trans^{\alpha,d}_X(Y) \to \trans^{\alpha',d'}_X(Y).
\end{equation*}
\end{dfn}
\begin{lem}\label{cor:MXdom}
Let $\delta > 0$. Consider $X$ as a subspace of $M\Gamma^{l,s}$ via the embedding $x \mapsto [e,x]$. There exists a $\delta$--controlled strong deformation retraction
\begin{equation*}
H \colon \norm{S_\bullet^\delta(M\Gamma^{l,s})} \times [0,1] \to \norm{S_\bullet^\delta(M\Gamma^{l,s})}
\end{equation*}
onto $\norm{S^\delta_\bullet(X)}$.
\end{lem}
\begin{proof}
There is a (topological) inclusion
\begin{equation*}
i \colon \norm{S^\delta_\bullet(M\Gamma^{l,s})} \times [0,1] \to \norm{S^\delta_\bullet(M\Gamma^{l,s} \times [0,1])}
\end{equation*}
which maps each prism $\Delta^p \times [0,1]$ to its canonical triangulation.
The strong deformation retraction from
\Cref{eq:strong-deformation-rectraction-MGamma} restricts to a
strong deformation retraction
\begin{equation*}
H' \colon M\Gamma^{l,s} \times [0,1] \to M\Gamma^{l,s}
\end{equation*}
of $M\Gamma^{l,s}$ onto $X$, it is given by
\begin{equation*}
H'([e,t_k,\gamma_k,\dots,\gamma_1,x],u) := [e,u \cdot t_k,\gamma_k, \dots,
\gamma_1, x].
\end{equation*}
It has the property that $R \circ H'(m,u) = R(m)$, so
\begin{equation*}
\diam_{X \times G} \{ (R \circ H'(m,u),g) \mid u \in [0,1] \} = 0
\end{equation*}
for all $m \in M\Gamma^{l,s}$ and $g \in G$. Hence, $H'$ induces a map
\begin{equation*}
H'_* \colon \norm{S^\delta_\bullet(M\Gamma^{l,s} \times [0,1])} \to \norm{S^\delta_\bullet(M\Gamma^{l,s})}.
\end{equation*}
Then $H := H'_* \circ i$ is a strong deformation retraction onto $\norm{S^\delta_\bullet(X)}$. As the target is
$\delta$-controlled, $H$ is $\delta$-controlled.
\end{proof}
\begin{cor}\label{cor:smallvaryconstraints}
Let $\delta > 0$, $l \leq l'$ and $s \leq s'$.
Then the canonical inclusion map $\norm{S_\bullet^\delta(M\Gamma^{l,s})} \hookrightarrow \norm{S_\bullet^{\delta}(M\Gamma^{l',s'})}$ is a $2\delta$--controlled homotopy equivalence (with respect to $d_{S^n,n,\Lambda}$).
\end{cor}
\begin{proof}
The corollary above yields homotopy equivalences $\norm{S_\bullet^\delta(X)}\hookrightarrow \norm{S_\bullet^\delta(M\Gamma^{l,s})}$ and $\norm{S_\bullet^\delta(X)}\hookrightarrow \norm{S_\bullet^{\delta}(M\Gamma^{l',s'})}$ which are $\delta$--controlled. Since the triangle
\[
\begin{tikzpicture}
\matrix (m) [matrix of math nodes, column sep=1.5em, row sep=1.5em, text depth=.5em, text height=1em, ampersand replacement=\&]
{\norm{\mathcal{S}_\bullet^\delta(X)} \&\norm{\mathcal{S}_\bullet^\delta(M\Gamma^{l,s})} \\ \& \norm{\mathcal{S}_\bullet^{\delta}(M\Gamma^{l',s'})} \\};
\path[->]
(m-1-1) edge node[above]{} (m-1-2)
(m-1-1) edge node[below left]{} (m-2-2)
(m-1-2) edge node[below right]{} (m-2-2);
\end{tikzpicture}
\]
commutes, the result follows.
\end{proof}
\begin{prop}\label{prop:transfer-vary-constraints}
The morphisms $\rho^{\alpha,\alpha',d,d'}_Y$ are weak equivalences.
\end{prop}
\begin{proof}
We prove that $\rho^{\alpha,\alpha',d,d'}_Y$ is a weak equivalence for all $Y
\in \mathcal{R}^G_f(W,\mathbb{J}(E))_{\alpha,d}$ by induction over the dimension of $Y$. For
$(-1)$--dimensional objects, which is the start of the induction, the claim
is trivial. For the induction step, we apply
\Cref{lem:attaching-cells-transfer} to see that the inclusion of the
$p$--skeleton into the $(p+1)$--skeleton induces a pushout
\begin{equation*
\begin{tikzpicture}
\matrix (m) [matrix of math nodes, column sep=3em, row sep=2em, text
depth=.5em, text height=1em]
{\sing{\alpha,d}{X,Y} \rightthreetimes_{\diamond_+\, Y}
\big(\coprod_{c \in \diamond\,_{p+1}Y}^W \partial c \big) &
\sing{\alpha,d}{X,Y} \rightthreetimes_{\diamond_+\, Y} \skel{p}{\mathcal{C}_Y} \\
\sing{\alpha,d}{X,Y} \rightthreetimes_{\diamond_+\, Y}
\big(\coprod_{c \in \diamond\,_{p+1}Y}^W c \big) &
\sing{\alpha,d}{X,Y} \rightthreetimes_{\diamond_+\, Y} \skel{p+1}{\mathcal{C}_Y}
\\};
\path[->]
(m-1-1) edge (m-1-2)
(m-1-1) edge (m-2-1)
(m-1-2) edge (m-2-2)
(m-2-1) edge (m-2-2);
\end{tikzpicture}
\end{equation*}
in $\mathcal{R}^G_f(W,\mathbb{J}(X \times G, E))_0$. The lower left corner is, again by
\Cref{lem:attaching-cells-transfer}, identified as
\begin{equation*}
\coprod_{c \in \diamond\,_{p+1}Y} \big( \norm{
S^{\delta^d_{\abs{c}}}_\bullet(M\Gamma^{l^d_{\abs{c}},\alpha}) }
\times D^{\abs{c}} \big) \amalg W,
\end{equation*}
and similarly for the the upper corner, with $D^{\abs{c}}$ replaced with
$\partial D^{\abs{c}}$.
There is an analogous pushout square
with $\alpha$ and $d$ replaced by $\alpha'$ and $d'$, respectively. Moreover,
the former square maps to the latter via the transformation
$\rho^{\alpha,\alpha',d,d'}_Y$. On the left-hand sides this is identified
with the canonical inclusion maps. This transformation is a weak equivalence
on the top right corner of the diagram by induction hypothesis, and it is a
$2\delta_{p+1}$--controlled homotopy equivalence on the top left and bottom
left corners combining \Cref{cor:smallvaryconstraints} and
\Cref{lem:controlled-excision} below. Hence, the gluing
lemma implies that it is also a weak equivalence on the bottom right corner.
This finishes the induction step and finite dimensionality of $Y$ proves the
claim.
\end{proof}
In order to show exactness, we will need that the transfer maps
$h$--equivalences to $h^{fin}$--equivalences later. The following lemma
implies this.
\begin{prop}\label{prop:transfer-exact}
Let $f$ be a weak equivalence in $\mathcal{R}^G_f(W,\mathbb{J}(E))_{\alpha,d}$ which is an $\alpha'$--controlled homotopy equivalence over $G$. Suppose that $n > \max \{d+2, \alpha, \alpha' \}$.
Then $\trans_X(f)$ is a controlled homotopy equivalence.
\end{prop}
\begin{proof}
To show the proposition, we exploit the fact that maps which are not cellwise
$0$--controlled over $G$ can also be transferred, but in a less functorial
fashion. Let $(Y_1,\kappa_1)$ and $(Y_2,\kappa_2)$ be objects of
$\mathcal{R}^G_f(W,\mathbb{J}(E))_{\alpha,d}$, and $f \colon Y_1 \to Y_2$ be an arbitrary
map in $\mathcal{R}^G_f(W,\mathbb{J}(E))$. Choose $\alpha' > 0$ such that $f$ is
$\alpha'$--controlled over $G$. We construct an induced map
\begin{equation*}
\trans_{\alpha,\alpha'}(f) \colon \trans^{\alpha,d}_X(Y_1) \to \trans^{\max \{ \alpha,\alpha' \}, d+1}_X(Y_2).
\end{equation*}
To define $\trans_{\alpha,\alpha'}(f)$, consider for the beginning a single
cell $c \in \diamond_+\, Y_1$, denote by $\eta_c$ the corresponding
$\diamond_+\, Y_1$--$n$--cell of $\mathcal{C}_{Y_2}$. We define the
function
\begin{equation*}
\begin{split}
t_c \colon
\norm{ S^{\delta_{\abs{c}}}_\bullet(M\Gamma^{l_{\abs{c}},\alpha}) }
\rightthreetimes_{\diamond_+\, Y_1} \eta_{c} & \to
\sing{\max \{ \alpha,\alpha' \},d+1}{X,Y_2}
\rightthreetimes_{\diamond_+\, Y_2} \mathcal{C}_{Y_2} \\
(x,y) & \mapsto [\supp(f(y)), \gamma^c_y \cdot x, f(y)],
\end{split}
\end{equation*}
where $\supp(f(y))$ denotes the support of $f(y)$, i.e., the unique open cell
$\supp(f(y))$ of $Y_2$ such that $f(y) \in \supp(f(y))$, and $\gamma^c_y :=
\kappa_{2,G}(\supp(f(y)))^{-1} \kappa_{1,G}(c)$. We will glue the different
$t_c$ together to get the transfer for $f$.
Let us check that the target space is large enough such that $t_c(x,y)$ is
contained in it: Recall that
$\gamma^c_y \cdot x$ is defined via the $G$--action which
$\norm{S_\bullet(M\Gamma)}$ inherits from $M\Gamma$. Since $f$ is
$\alpha'$--controlled, we have $\gamma^c_y \in B_{\alpha'}(e)$. Therefore, we
can regard multiplication with $\gamma^c_y$ as a map
$M\Gamma^{l_{\abs{c}},\alpha} \to M\Gamma^{l_{\abs{c}}+1,\max \{
\alpha,\alpha' \}}$. In addition, $M\Gamma^{l_{\abs{c}}+1,\max \{ \alpha,
\alpha' \}}$ is contained in $M\Gamma^{l_{\abs{\supp(f(y))}} +1, \max \{
\alpha,\alpha' \}}$, so $[\supp(f(y)),\gamma^c_y \cdot x, f(y)]$ defines a
point in the target space.
We need to check that $t_c$ is continuous. It suffices to show continuity on
finite subcomplexes. These are metrizable, so it is enough to show that $t_c$
is sequentially continuous. Let $(x_l,y_l)_l$ be a convergent sequence in
$\norm{ S^{\delta_{\abs{c}}}_\bullet(M\Gamma^{l_{\abs{c}},\alpha}) } \times
\eta_{c}$ with limit point $(x,y)$. As $f$ is continuous, $f(y_l)$ converges
against $f(y)$. Hence $S := \{ \supp(f(y_l)) \mid l \in \mathbb{N} \}$ is a finite
set, and we can assume that for each $s \in S$ there are infinitely many $l$
such that $\supp(f(y_l)) = s$. We treat the $s$ individually and restrict to
the corresponding subsequence. If $s$ happens to be equal to $\supp(f(y))$,
$\gamma^c_{y}= \gamma^c_{y_l}$ and continuity follows. Otherwise, $f(y)$ must
still lie in the closure of the cell $s$, i.e.,
$\supp(f(y)) \subset \gen{s}$. Hence,
\begin{equation*}
\begin{split}
[\supp(f(y)), &\gamma^c_y \cdot x, f(y)] \\
&= [\supp(f(y)),\kappa_{2,G}(\supp(f(y)))^{-1} \kappa_{2,G}(s) \kappa_{2,G}(s)^{-1} \kappa_{1,G}(c) x, f(y)] \\
&= [s, \kappa_{2,G}(s)^{-1} \kappa_{1,G}(c) x, f(y)],
\end{split}
\end{equation*}
and continuity becomes obvious.
Suppose now that $c \leq c'$ in $\diamond\, Y_1$. For $y \in \gen{c}$ and $x \in \norm{ S^{\delta_{\abs{c'}}}_\bullet(M\Gamma^{l_{\abs{c'}},\alpha}) }$ we obtain
\begin{equation*}
\begin{split}
t_{c'}(x,y)
&= [ \supp(f(y)), \gamma^{c'}_y x, f(y)] \\
&= [ \supp(f(y)), \kappa_{2,G}(\supp(f(y)))^{-1} \kappa_{1,G}(c) \kappa_{1,G}(c)^{-1} \kappa_{1,G}(c') x, f(y)] \\
&= t_c(\kappa_{1,G}(c)^{-1} \kappa_{1,G}(c') x,y).
\end{split}
\end{equation*}
Therefore, the collection $\{ t_c \}_{c \in \diamond\, Y_1}$ induces a continuous, cellular map relative $W$
\begin{equation*}
\trans_{\alpha,\alpha'}(f) \colon \trans^{\alpha,d}_X(Y_1) \to \trans^{\max\{ \alpha,\alpha' \},d+1}_X(Y_2).
\end{equation*}
Using \Cref{lem:distance-to-translate} and
\Cref{lem:transfer-control-objects}, it is not hard to show that
$\trans_{\alpha,\alpha'}(f)$ is $(\max \{\alpha,\alpha' \} + \alpha' +
\delta_0 +
4)$--controlled over $X \times G$, as $n > \alpha'$.
Note that
for cellwise $0$--controlled
maps, $\trans_{\alpha,\alpha'}(f)$ agrees with the previous defined transfer
from \Cref{subsec:transfer-cellwise-0}. The only reason we increased $d$ is
the argument which follows, it was not needed in the construction so far.
Suppose now that $f \colon (Y_1,\kappa_1) \to (Y_2,\kappa_2)$ is a weak
equivalence in $\mathcal{R}^G_f(W,\mathbb{J}(E))_{\alpha,d}$ which is $\alpha'$--controlled
over $G$ as a homotopy equivalence, i.e., its inverses and the homotopies
are $\alpha'$--controlled over $G$. Then there exists some
$\alpha'$--controlled
map $\overline{f} \colon (Y_2,\kappa_2) \to (Y_1,\kappa_1)$ such that
$\overline{f}f$ and $f\overline{f}$ are $\alpha'$--controlled homotopic to
the identity. Consider the diagram
\begin{equation*}
\begin{tikzpicture}
\matrix (m) [matrix of math nodes, column sep=10em, row sep=4em, text depth=.5em, text height=1em]
{\trans^{\alpha,d}_X(Y_1) & \trans^{\alpha,d}_X(Y_2) \\
\trans^{\max \{\alpha,\alpha'\},d+1}_X(Y_1) & \trans^{\max \{\alpha,\alpha'\},d+1}_X(Y_2) \\};
\path[->]
(m-1-1) edge node[above]{$\trans^{\alpha,d}_X(f)$} (m-1-2)
(m-2-1) edge node[below]{$\trans^{\max \{ \alpha,\alpha' \}, d+1}_X(f)$} (m-2-2);
\path[>->]
(m-1-1) edge node[left]{$\rho^{\alpha,\max \{ \alpha,\alpha' \},d,d+1}_{Y_1}$} (m-2-1)
(m-1-2) edge node[right]{$\rho^{\alpha,\max \{ \alpha,\alpha' \},d,d+1}_{Y_2}$} (m-2-2);
\path[dashed,->]
(m-1-2) edge node[above left]{$\trans_{\alpha,\alpha'}(\overline{f})$} (m-2-1);
\end{tikzpicture}
\end{equation*}
in which the outer square commutes. The vertical maps $\rho^{\alpha,\max \{\alpha,\alpha'\},d,d+1}_{Y_i}$, $i=1,2$, are weak equivalences by \Cref{prop:transfer-vary-constraints}. We claim that the two triangles involving the dashed diagonal map $\trans_{\alpha,\alpha'}(\overline{f})$ commute up to controlled homotopy. If this is true, it follows that $\trans^{\alpha,d}_X(f)$ is a weak equivalence.
Let $h \colon Y_1 \leftthreetimes [0,1] \to Y_1$ be an $\alpha'$--controlled homotopy from $\overline{f}f$ to $\operatorname{id}_{Y_1}$. Note that $\trans_{\alpha,\alpha'}(\overline{f}) \circ \trans^{\alpha,d}_X(f) = \trans_{\alpha,\alpha'}(\overline{f}f)$. Since $n > d+2$, we can apply $\trans_X^{\alpha,d}$ also to $Y_1 \leftthreetimes [0,1]$ and consider the controlled map
\begin{equation*}
\trans_{\alpha,\alpha'}(h) \colon \trans^{\alpha,d}_X(Y_1 \leftthreetimes [0,1]) \to \trans^{\max \{\alpha,\alpha' \}, d+1}_X(Y_1).
\end{equation*}
The domain of this map is not equal to $\trans^{\alpha,d}_X(Y_1)
\leftthreetimes [0,1]$, but it is contained in $\trans^{\alpha,d}_X(Y_1)
\leftthreetimes [0,1]$ as a controlled strong deformation retract. This
follows by an induction argument similar to
\Cref{prop:transfer-vary-constraints}. Essentially, we can construct both
objects as the balanced products over $\diamond_+\, (Y_1 \leftthreetimes
[0,1])$ and use that the inclusion
$\norm{ S^{\delta_{\abs{c}}}_\bullet(M\Gamma^{l_{\abs{c}},\alpha}) } \to
\norm{ S^{\delta_{\abs{c}}+1}_\bullet(M\Gamma^{l_{\abs{c}+1},\alpha}) } $ is
a controlled deformation retraction by \Cref{cor:smallvaryconstraints} and
\Cref{lem:controlled-excision}. The retraction induces the required
controlled homotopy.
The argument for the second triangle is analogous.
\end{proof}
\subsection{Restricting the target category}
\label{subsect:transfer-restricting-target}
Now we show that the transfer functor factors over the full subcategory of
finitely dominated objects. The following result was already used in the
previous chapter.
\begin{lem}\label{lem:controlled-excision}
Let $(M,d)$ be a metric space.
\begin{enumerate}
\item Let $\delta > 0$. The natural inclusion map $\norm{
S^\delta_\bullet(M) } \to \norm{ S_\bullet(M) }$ is a homotopy
equivalence.
\item Let $0 < \delta \leq \delta'$. Then the inclusion map $\norm{
S^\delta_\bullet(M) } \to \norm{ S^{\delta'}_\bullet(M) }$ is a
$\delta'$--controlled homotopy equivalence (with respect to the metric on
$M$, labelling simplices by the image of their barycenter).
\item Suppose $\abs{K}$ is the realization of an ordered (abstract)
simplicial complex $K$ and suppose that $p \colon \abs{K} \to M$ is
a continuous map. Let $\kappa \colon \diamond\, K \to M$ be the labelling
sending a cell (=simplex) to the image of its barycenter under $p$. Let
$\delta > 0$. Let $S^\delta_\bullet(\abs{K},p)$ denote the
(semi)simplicial set of all singular simplices $\sigma$ in $\abs{K}$ such
that the diameter of $p
\circ \sigma$ is at most $\delta$.
If the characteristic maps of all simplices of $K$ lie in $S^\delta_\bullet(\abs{K},p)$, then the canonical map $\abs{K} \to \norm{ S^\delta_\bullet(\abs{K},p) }$ is a $\delta$--controlled homotopy equivalence (measuring control in $M$ via $p$).
\end{enumerate}
\end{lem}
\begin{proof}
The proof proceeds in analogy to \cite[Lemma~6.7]{Bartels-Lueck-Reich(2008hyper)}. The first part follows directly from an appropriate formulation of excision, e.g.~\cite[Theorem~4.6.9]{Fritsch-Piccinini(1990)}.
For the second part, let $\mathcal{A}$ be the poset of closed subsets of $X$, considered as a category. Then the $\mathcal{A}$--CW--complex $\sing{\delta}{\mathcal{A}}$ given by
\begin{equation*}
\sing{\delta}{\mathcal{A}}(A) := \norm{ S^\delta_\bullet(A) }
\end{equation*}
is a free $\mathcal{A}$--CW--complex, whose free cells are of the form
$\hom_\mathcal{A}(\sigma(\Delta^{\abs{\sigma}}),-)\times D^{\abs{\sigma}}$ since
$\diamond\, \norm{ S^\delta_\bullet(A) } \cong \coprod_{\sigma \in
S^\delta_\bullet(X) } \hom_\mathcal{A}(\sigma(\Delta^{\abs{\sigma}}),A)$. Since both
$\norm{ S^\delta_\bullet(A) } \hookrightarrow \norm{ S_\bullet(A) }$ and
$\norm{ S^{\delta'}_\bullet(A) } \hookrightarrow \norm{ S_\bullet(A) }$ are
homotopy equivalences for every $A \in \mathcal{A}$, so is the inclusion $\norm{
S^\delta_\bullet(A) } \hookrightarrow \norm{ S^{\delta'}_\bullet(A) }$.
Hence, the natural transformation $\sing{\delta}{\mathcal{A}} \to
\sing{\delta'}{\mathcal{A}}$ is a homotopy equivalence of (free) $\mathcal{A}$--CW--complexes
by \cite[Corollary~3.5]{Davis-Lueck(1998)}. This in particular means that
there is an inverse map, compatible with the structure map, as well as
compatible
homotopies. It is easy to check that such a map has the right control.
The third claim follows by similar reasoning, substituting the poset of subcomplexes for the poset of closed subsets.
\end{proof}
\begin{lem}\label{lem:finite-domination-singular-complex}
Suppose that $(X,d_X)$ admits a finite $\epsilon$--domination. Then
$\norm{ S^\delta_\bullet(M\Gamma^{l,\alpha}) }$ is $4 \delta + 6
\Lambda \epsilon$-dominated over $X \times G$ (with respect to the metric
$d_{S^n, n , \Lambda}$).
\end{lem}
\begin{proof}
By \Cref{cor:MXdom}, the complex $\norm{ S^{\delta}_\bullet(X) }$ is a $\delta$--controlled strong deformation retract of $\norm{ S^{\delta}_\bullet(M\Gamma^{l,\alpha})}$. Choose an appropriate controlled retraction $r$.
Pick an $\epsilon$--domination of $X$ by a finite simplicial complex
$\abs{K}$, i.e., a sequence of maps $X \xrightarrow{\iota} \abs{K}
\xrightarrow{\pi} X$ together with a homotopy $h \colon \pi \circ \iota
\simeq \operatorname{id}_X$, and such that the diameter, measured with respect to the
original metric $d_X$ on $X$, of $h(x,[0,1])$ is at most $\epsilon$ for every
$x \in X$. Then the given domination induces maps
\begin{equation*}
\norm{ S^\delta_\bullet(X) } \xrightarrow{\iota_*} \norm{
S^{\delta+2\Lambda\epsilon}_\bullet(\abs{K},\pi) } \xrightarrow{\pi_*}
\norm{ S^{\delta+2\Lambda\epsilon}_\bullet(X) },
\end{equation*}
where similarly to \Cref{subsec:y-spaces} we measure distances of points in
$\abs{K}$ via $\pi$.
These are maps of labeled complexes over $X \times G$: Pick an arbitrary group element $g \in G$. Then we label simplices $\sigma$ in $S^\delta_\bullet(X)$ or $S^{\delta+2\Lambda\epsilon}_\bullet(X)$ by $(\sigma(\beta_{\abs{\sigma}}),g)$ and simplices $\sigma$ in $S^{\delta+2\Lambda\epsilon}_\bullet(\abs{K},\pi)$ by $(\pi(\sigma(\beta_{\abs{\sigma}})),g)$ (cf.~\Cref{cor:MXdom} and \Cref{cor:smallvaryconstraints}).
Choose an iterated barycentric subdivision $K'$ of $K$ such that the characteristic map of each simplex of $K'$ is a simplex in $S^{\delta+2\Lambda\epsilon}_\bullet(\abs{K},\pi)$; note that $K'$ is ordered if we subdivide at least once. Since $\abs{K'}$ is then naturally a subcomplex of $S^{\delta+2\Lambda\epsilon}_\bullet(\abs{K},\pi)$, we endow it with the induced control map.
The canonical inclusion $i \colon\abs{K'} \to \norm{ S^{\delta+2\Lambda\epsilon}_\bullet(\abs{K},\pi) }$ is a $(\delta+2\Lambda\epsilon)$--controlled homotopy equivalence over $X \times G$ by \Cref{lem:controlled-excision}. Choose an appropriate controlled homotopy inverse $p$.
Finally, the inclusion $\norm{ S^\delta_\bullet(X) } \hookrightarrow \norm{ S^{\delta+2\Lambda\epsilon}_\bullet(X) }$ is a $(\delta+2\Lambda\epsilon)$--controlled homotopy equivalence by \Cref{lem:controlled-excision}; let $f$ be an appropriate controlled homotopy inverse.
Then
\begin{align*}
\norm{ S^\delta_\bullet(M\Gamma^{l,\alpha}) } \xrightarrow{r} \norm{ S^\delta_\bullet(X) } \xrightarrow{\iota_*} \norm{ S^{\delta+2\Lambda\epsilon}_\bullet(\abs{K},\pi) } \xrightarrow{p} \abs{K'}
\end{align*}
and
\begin{align*}
\abs{K'} \xrightarrow{i} \norm{ S^{\delta+2\Lambda\epsilon}_\bullet(\abs{K},\pi) } \xrightarrow{\pi_*} \norm{ S^{\delta+2\Lambda\epsilon}_\bullet(X) } \xrightarrow{f} \norm{ S^\delta_\bullet(X) } \hookrightarrow \norm{ S^\delta_\bullet(M\Gamma^{l,\alpha}) }
\end{align*}
yield the desired domination of $\norm{ S^\delta_\bullet(X) }$; from the previous control estimates we see that there is a $(4\delta+6\Lambda\epsilon)$--controlled homotopy between the composition of these two maps and the identity on $\norm{ S^\delta_\bullet(M\Gamma^{l,\alpha}) }$.
\end{proof}
\begin{prop}\label{prop:transfer-finite-domination}
Suppose that $(X,d_X)$ admits a finite $\epsilon$--domination for every
$\epsilon$. Let $(Y,\kappa) \in \mathcal{R}^G_f(W,\mathbb{J}(E))_{\alpha,d}$.
Then $\trans_X(Y,\kappa)$ is controlled finitely dominated, i.e., it defines
an object in $\mathcal{R}^G_{fd}(W,\mathbb{J}(X \times G, E))$. We can choose the control
estimate to be independent of the constants $\Lambda$ and $n$ from the metric.
\end{prop}
\begin{proof}
We prove the claim by induction on the dimension of $Y$. The case of a $(-1)$--dimensional object is trivial and provides the start of the induction.
For the induction step, we use \Cref{lem:attaching-cells-transfer} to obtain
a pushout square
\begin{equation*
\begin{tikzpicture}
\matrix (m) [matrix of math nodes, column sep=3em, row sep=2em, text
depth=.5em, text height=1em]
{
\coprod_{c \in \diamond\,_{p+1}Y} \big( \norm{
S^{\delta^d_{p+1}}_\bullet(M\Gamma^{l^d_{p+1},\alpha}) }
\times \partial D^{\abs{c}} \big) \amalg W &
\sing{\alpha,d}{X,Y} \rightthreetimes_{\diamond_+\, Y} \skel{p}{\mathcal{C}_Y} \\
\coprod_{c \in \diamond\,_{p+1}Y} \big( \norm{
S^{\delta^d_{p+1}}_\bullet(M\Gamma^{l^d_{p+1},\alpha}) }
\times D^{\abs{c}} \big) \amalg W &
\sing{\alpha,d}{X,Y} \rightthreetimes_{\diamond_+\, Y} \skel{p+1}{\mathcal{C}_Y}
\\};
\path[->]
(m-1-1) edge (m-1-2)
(m-1-1) edge (m-2-1)
(m-1-2) edge (m-2-2)
(m-2-1) edge (m-2-2);
\end{tikzpicture}
\end{equation*}
in $\mathcal{R}^G(W,\mathbb{J}(X \times G,E))$.
By induction hypothesis, the object at the top right corner of this square is
finitely dominated. So we only need to find a controlled finite domination
for $\norm{ S^{\delta_{p+1}}_\bullet(M\Gamma^{l_{p+1},\alpha}) }$, as $Y$
itself is (locally) finite. By
\Cref{lem:finite-domination-singular-complex}, such a domination indeed
exists.
Note that the same bound works if we increase $n$, and we can choose
$\epsilon$ to be $\frac{1}{\Lambda}$. Then the estimate of the metric does
not depend on $n$ and $\Lambda$, which finishes the proof.
\end{proof}
Finally, we show that, after forgetting the labelling in $X$, the transfer does not alter the homotopy type of a given object.
\begin{prop}\label{prop:transfer-section}
Let $P \colon \mathcal{R}^G_{fd}(W,\mathbb{J}(X \times G, E)) \to \mathcal{R}^G_{fd}(W,\mathbb{J}(E))$ denote the functor induced by the projection map $X \times G \to G$. Let $(Y,\kappa) \in \mathcal{R}^G_f(W,\mathbb{J}(E))_{\alpha,d}$.
Then there is an $\alpha$--controlled natural weak equivalence
\begin{equation*}
P(\trans_X(Y)) \xrightarrow{\sim} Y.
\end{equation*}
\end{prop}
\begin{proof}
The relevant map is induced by the projection map $M\Gamma \to *$. As in the
proofs of \Cref{prop:transfer-exact} and
\Cref{prop:transfer-finite-domination}, the claim follows by another
induction along the skeleta of $Y$, using \Cref{lem:attaching-cells-transfer} and \Cref{lem:controlled-excision} together with the fact that the projection map $\norm{ S_\bullet(X) } \to *$ is a homotopy equivalence. Since the bounded control is only over $(G,d_G)$, it is not hard to check that the weak equivalence is $\alpha$--controlled.
\end{proof}
\subsection{The transfer map}
\label{subsec:transfer-map}
We combine all of the results established so far to show
\Cref{prop:main-diag-claims}~\ref{item:main-diag-claims-1}.
Let $N \in \mathbb{N}$. Suppose that we have chosen for every $n \in \mathbb{N}$
\begin{enumerate}
\item a compact, contractible metric space $(X_n,d_{X_n})$ such that for every $\epsilon > 0$ there is an $\epsilon$--controlled domination of $X_n$ by an at most $N$--dimensional, finite simplicial complex;
\item a homotopy coherent $G$--action $\Gamma_n$ on $X_n$;
\item a positive real number $\Lambda_n$.
\end{enumerate}
We equip $X_n \times G$ with the metric $d_{S^n, n, \Lambda_n} + d_G$. As in
\Cref{subsec:transfers}, we set
\begin{equation*}
\delta_k := 4(d+1-k), \quad l_k := d+1-k.
\end{equation*}
\begin{prop}\label{prop:full-transfer}
Let $\alpha, d \in \mathbb{N}$. The assignment
\begin{equation*}
\begin{split}
(Y,\kappa) &\mapsto \trans^{\alpha,d}(Y,\kappa) := \big( \trans^{\alpha,d}_{X_n}(Y,\kappa) \big)_{n > \max \{ d+1, \alpha \} } \\
f &\mapsto \big( \trans^{\alpha,d}_{X_n}(f) \big)_{n > \max \{d+1, \alpha \} }
\end{split}
\end{equation*}
defines an exact functor
\begin{equation*}
\trans^{\alpha,d} \colon (\mathcal{R}^G_f(W,\mathbb{J}(E)),h)_{\alpha, d} \to (\mathcal{R}^G_{fd}(W,\mathbb{J}((X_n
\times G)_n,E)), h^{fin}).
\end{equation*}
\end{prop}
\begin{proof}
According to \Cref{lem:transfer-control-objects}, \Cref{lem:transfer-functor} and \Cref{lem:transfer-cofibrations} the assignment yields a functor of categories with cofibrations
\[
\trans^{\alpha,d} \colon \mathcal{R}^G_f(W,\mathbb{J}(E))_{\alpha,d} \to \prod_{n\in \mathbb{N}}\mathcal{R}^G_{fd}(W,\mathbb{J}(X_n \times G,E)).
\]
We have to show that it factors over the subcategory $\mathcal{R}^G_{fd}(W,\mathbb{J}((X_n \times G)_n,E))$. This is the case if all objects and morphisms in the image of $\trans^{\alpha,d}$ are uniformly boundedly controlled over $X_n \times G$. Essentially, we have to see that all of the necessary control estimates are independent of $n\in \mathbb{N}$.
For this, recall that the map
\begin{equation*}
\mathcal{R}^G_{fd}(W,\mathbb{J}((X_n \times G)_n,E)) \to \prod_{n\in
\mathbb{N}}\mathcal{R}^G_{fd}(W,\mathbb{J}(X_n \times G,E))
\end{equation*}
works as follows. Essentially, an object in the source is a CW--complex
relative $W$, where we have a partition of its cells into $\mathbb{N}$--many sets and
no boundary and no map is allowed to hit a cell which is in a different
set. Hence, we can write the object as the coproduct (over $W$), indexed
by $\mathbb{N}$, of
CW--complexes relative $W$. The collection of summands defines an element in
the target. If the transfer satisfies a uniform metric control
condition, it factors over this map. Hence we need to check that the
previous results of this section give uniform bounds for all $n$.
Since $\trans^{\alpha,d}_{X_n}(Y,\kappa)$ is
$(\alpha+\delta_0+2)$--controlled over $X_n \times G$ for every $n$ by
\Cref{lem:transfer-control-objects} and $\trans_{X_n}^{\alpha, d}(f)$ is
cellwise $0$-controlled by \Cref{lem:transfer-functor}, all
objects and morphisms are uniformly bounded, as desired.
\Cref{prop:transfer-finite-domination} shows that each component
$\trans^{\alpha,d}(Y,\kappa)$ is finitely dominated, but we need it
uniformly. For this, note that the proof of
\Cref{prop:transfer-finite-domination} can actually be done with $(X_n)_n$
replacing $X$. Roughly, we would get an extra coproduct over $\mathbb{N}$
everywhere, and everything else would need to get an extra index, which is
why \Cref{prop:transfer-finite-domination} is not stated that way. However,
the control estimations come from applications of the gluing lemma and an
induction over the cells of $Y$. But the gluing lemma preserves
the property of everything being uniformly controlled, and we start the
induction with uniform control arising from $f$ and the $\delta_k$, so we can
do the same induction.
\Cref{prop:transfer-exact} tells us that $\trans^{\alpha,d}_{X_n}$ sends
$h$--equivalences to $h^{fin}$--equivalences since it applies for
sufficiently large $n$. Again, the proof can be done for $(X_n)_n$ instead
of $X$, and the control estimates come from an induction over the cells of
$Y$ and the gluing lemma, so they will be uniform.
\end{proof}
\begin{prop}\label{prop:full-transfer-compatibility}
Let $\alpha, d \in \mathbb{N}$ and $i_{\alpha,d} \colon \mathcal{R}^G_f(W,\mathbb{J}(E))_{\alpha,d} \hookrightarrow \mathcal{R}^G_f(W,\mathbb{J}(E))_{\alpha+1,d+1}$ be the obvious inclusion functor. Then there is a natural $h^{fin}$--equivalence
\begin{equation*}
\trans^{\alpha,d} \xrightarrow{\sim} \trans^{\alpha+1,d+1} \circ i_{\alpha,d}.
\end{equation*}
\end{prop}
\begin{proof}
There is a natural transformation $\trans^{\alpha,d} \to \trans^{\alpha+1,d+1} \circ i_{\alpha,d}$ given by the sequence $(\rho^{\alpha,\alpha+1,d,d+1})_{n > \max \{ d+2,\alpha+1 \} }$ from \Cref{def:transfer-vary-constraints}. These are homotopy equivalences by \Cref{prop:transfer-vary-constraints}, and the control estimates in the proof of \Cref{prop:transfer-vary-constraints} show that they are also uniformly boundedly controlled homotopy equivalences.
\end{proof}
To obtain the transfer map whose existence was claimed in \Cref{prop:main-diag-claims}, we proceed as follows. Let $k \in \mathbb{N}$. Consider the inclusion $j_k \colon \mathcal{R}^G_f(W,\mathbb{J}(E))_{k,k} \hookrightarrow \mathcal{R}^G_f(W,\mathbb{J}(E))_{k+1,k+1}$. By \Cref{prop:full-transfer-compatibility}, there is a natural weak equivalence
\begin{equation*}
\rho_k \colon \trans^{k,k} \xrightarrow{\sim} \trans^{k+1,k+1} \circ j_k.
\end{equation*}
Hence, we obtain an induced homotopy
\begin{equation*}
K(\trans^{k,k}) \simeq K(\trans^{k+1,k+1}) \circ K(j_k).
\end{equation*}
Thinking of $\operatorname{hocolim}_k K(\mathcal{R}^G_f(W,\mathbb{J}(E))_{k,k},h)$ as the mapping telescope
of
\begin{equation*}
K(\mathcal{R}^G_f(W,\mathbb{J}(E))_{1,1},h) \xrightarrow{K(j_1)} K(\mathcal{R}^G_f(W,\mathbb{J}(E))_{2,2},h) \xrightarrow{K(j_2)} \dots,
\end{equation*}
these homotopies serve to define a map
\begin{equation*}
\trans \colon \operatorname{hocolim}_k K(\mathcal{R}^G_f(W,\mathbb{J}(E))_{k,k},h) \to K(\mathcal{R}^G_{fd}(W,\mathbb{J}((X_n \times G)_n,E)),h^{fin}).
\end{equation*}
\begin{prop}
The map $\trans$ satisfies \Cref{prop:main-diag-claims}~\ref{item:main-diag-claims-1}.
\end{prop}
\begin{proof}
That $\trans$ is the required map and that the diagram commutes up to
homotopy is immediate from \Cref{prop:domain-of-transfer} and
\Cref{prop:transfer-section}, noting again that the latter proof can be
done uniformly.
\end{proof}
\typeout{-------------------------------- References -----------------------------------------------}
|
\section{Introduction}
General relativity and quantum mechanics have become two pillars of modern physics. Meanwhile, due to constant developments of technologies, increasingly accurate observations mostly indicate that the Einstein's theory of general relativity has passed all experimental tests with flying colors, especially in the weak-field or slow-motion regime~\cite{EB}. More attractively, black holes, which can probably be regarded as a tie of connecting general relativity and quantum mechanics~\cite{XC}, have been getting more and more attentions. In particular, the thermodynamics of black holes in anti-de Sitter (AdS) spacetime~\cite{JMB,RW,SC} has acquired great progress since the AdS/CFT duality plays a pivotal role in recent developments of theoretical physics~\cite{SH,RGC}.
Conversely, some of the funniest strong-field predictions of the Einstein's theory of general relativity still remain difficult to be understood and verified. In this sense, black holes are ideal candidates to be used as probes of Einstein theory. Most theoretical and observational issues, such as the high-curvature corrections, the origin of curvature singularities, the cosmological constant problem, the dark energy/matter, and so on, strongly recommend that the Einstein's theory of general relativity should be modified, i.e., the so-called modified gravity~\cite{TC,EP}. What is worth mentioning is that Horndeski~\cite{GWH} proposed the most general scalar-tensor modified gravity action which generates equations of motion with second-order derivatives. The Horndeski scalar-tensor modified gravity theory has widely been investigated in astrophysics and cosmology~\cite{ESSVS,CCEJA,AMMHO}. In the investigation of a locally stable solution of Horndeski black holes, the action containing a non-minimal kinetic coupling of one scalar and Einstein tensor has received considerable attentions, and some important progress has been made. A spherically symmetric and static solution has been obtained in ref.~\cite{MR} for the case of a vanishing cosmological constant, and in refs.~\cite{AAJ,MM} for the case of a negative cosmological constant. The no-hair theorem for scalar-tensor gravity theory has been shown in refs.~\cite{LHAN,TPS}. Moreover, the black hole solution in the presence of an electric field~\cite{ACCE,TK,XHHC}, the BTZ black hole solution with a Horndeski source~\cite{MBMH}, and the slowly rotating black hole solutions~\cite{AHME} have been studied. Furthermore, the black hole solution with a time-dependent scalar~\cite{EBCC,MBMH1} and the exact wormhole solutions with a non-minimal kinetic coupling~\cite{RSVS} have also been found. For the topics in other relevant aspects, see, for instance, refs.~\cite{EBKY,HTT,XHHC1,TKNT,RMLLJ,ECCAL,MR2}.
In this paper we revisit thermodynamic properties of Horndeski black holes with a non-minimal kinetic coupling in the presence of an electric field along the line of refs.~\cite{CEJM,DSJT,BPD2,BPD,KM}, namely by considering the cosmological constant as thermodynamic pressure and the mass of black holes as thermodynamic enthalpy. A few thermodynamic quantities have been calculated~\cite{MR,AAJ,MM} for Horndeski black holes, showing that the first law of thermodynamics is satisfied, but the Smarr relation is violated. Hence, we wish to fill up this deficiency from the point of view of thermodynamics. If we take the non-minimal kinetic coupling strength of scalar and tensor fields into account, the relevant term should appear in the Smarr relation and its variation should be included in the first law of thermodynamics. As a result, we obtain the generalized Smarr relation and the first law of thermodynamics in the extended phase space that includes the coupling strength and its conjugate. In other words, we shall deal with thermodynamic behaviors of Horndeski black holes in a new extended phase space. It has been known that the Born-Infeld parameter~\cite{GKM,NB}, the Gauss-Bonnet coupling constant~\cite{RCLY}, and the noncommutative parameter~\cite{YGMX} can be dealt with as a kind of thermodynamic pressure.
Our result further indicates that the coupling strength acts as the thermodynamic pressure in the behavior of thermodynamics. Meanwhile, we show that the class of Horndeski black holes with a non-minimal kinetic coupling presents rich critical phenomena.
The paper is organized as follows. In section \ref{sec2}, the thermodynamics of Horndeski black holes with a non-minimal derivative coupling is analyzed. This section contains two subsections which correspond to the scenarios without and with charge, respectively. Finally, we devote to drawing our conclusion in section \ref{sec4}. In addition, the geometric units, $\hbar=c=k_{B}=G=1$, are adopted throughout this paper.
\section{Thermodynamics of Horndeski black holes}\label{sec2}
At the beginning, we proceed to investigate the class of Horndeski black holes whose action contains a non-minimal kinetic coupling of the massless real scalar $\phi$ and the Einstein tensor $G_{\mu\nu}$. This action, in the presence of an electric field, has the following form~\cite{ACCE,XHHC},
\begin{equation}
I=\int \sqrt{-g} \,\text{d}^4 x \left[(R-2\Lambda)-\frac12(\alpha g_{\mu\nu}-\eta G_{\mu\nu})\nabla^{\mu}\phi \nabla^{\nu}\phi-\frac14 F_{\mu\nu}F^{\mu\nu}\right], \label{act}
\end{equation}
where $\eta$ stands for the coupling strength with the dimension of length square, $\alpha$ a coupling constant, $\Lambda$ the negative cosmological constant, $R$ the scalar curvature, $g_{\mu\nu}$ the metric with mostly plus signatures, and $F_{\mu\nu}$ the electromagnetic field strength defined as $F_{\mu\nu} \equiv \partial_{\mu}A_{\nu}-\partial_{\nu}A_{\mu}$ with $A_{\mu}$ the vector potential.
Making a variation of the action eq.~(\ref{act}) with respect to the metric $g_{\mu\nu}$, the scalar field $\phi$, and the Maxwell field $A_{\mu}$, respectively, one can obtain
\begin{eqnarray}
G_{\mu\nu}+\Lambda g_{\mu\nu}&=&\frac12 \left(\alpha T_{\mu\nu}+\eta \Xi_{\mu\nu}+E_{\mu\nu}\right), \label{gmu}\\
\nabla_{\mu}\left[(\alpha g_{\mu\nu}-\eta G_{\mu\nu})\nabla_{\nu}\phi\right]&=&0, \label{scalarmu}\\
\nabla_{\mu}F^{\mu\nu}&=&0, \label{emu}
\end{eqnarray}
where $T_{\mu\nu}$, $\Xi_{\mu\nu}$, and $E_{\mu\nu}$ are defined as
\begin{eqnarray}
T_{\mu\nu} &\equiv& \nabla_{\mu}\phi \nabla_{\nu}\phi-\frac12 g_{\mu\nu}\nabla_{\rho}\phi \nabla^{\rho}\phi,\\
\Xi_{\mu\nu} &\equiv & \frac12 \nabla_{\mu}\phi \nabla_{\nu}\phi R-2\nabla_{\rho}\phi \nabla_{(\mu}\phi R_{\nu)}^{\rho}-\nabla^{\rho}\phi \nabla^{\lambda}\phi R_{\mu\rho\nu\lambda}\nonumber \\
& &-(\nabla_{\mu}\nabla^{\rho}\phi)(\nabla_{\nu}\nabla_{\rho}\phi)
+(\nabla_{\mu}\nabla_{\nu}\phi)\square \phi+\frac12 G_{\mu\nu}(\nabla \phi)^2\nonumber\\
& &-g_{\mu\nu}\left[-\frac12 (\nabla^{\rho}\nabla^{\lambda}\phi)(\nabla_{\rho}\nabla_{\lambda}\phi)
+\frac12 (\square \phi)^2-\nabla_{\rho}\phi \nabla_{\lambda}\phi R^{\rho\lambda}\right],\\
E_{\mu\nu}&\equiv& F_{\mu}^{\rho}F_{\nu\rho}-\frac12 g_{\mu\nu}F^2.
\end{eqnarray}
In the following we focus on the static solutions with the spherical symmetry in eqs.~(\ref{act})-(\ref{emu}), so the metric is simplified to be
\begin{eqnarray}
\text{d} s^2&=&-f(r)\text{d} t^2+g(r)\text{d} r^2+r^2(\text{d} \theta^2+\sin^2 \theta \text{d}\varphi^2), \label{metric}\\
F&=&\text{d}A,\\
A&=&\Psi \text{d}t,
\end{eqnarray}
where $f(r)$ and $g(r)$ are functions to be determined and $\Psi$ is the electrostatic potential. Under the assumption of spherical symmetry, we only consider a static and isotropic scalar field, i.e., the scalar field is a function of the radial coordinate, $\phi=\phi(r)$. In the two subsections below we shall discuss thermodynamic properties of such a class of Horndeski black holes without and with the Maxwell field, respectively, i.e., under the considerations of the specific forms of $f(r)$, $g(r)$, and $\phi(r)$ in the former subsection, and the specific forms of $f(r)$, $g(r)$, $\phi(r)$, and $\Psi(r)$ in the latter one.
\subsection{Scenario without charge}
For this situation, the analytic solution takes the form~\cite{AAJ,MM},
\begin{eqnarray}
f(r)&=&\frac{\alpha r^2}{3\eta}-\frac{2M}{r}+\frac{3\alpha+\Lambda \eta}{\alpha-\Lambda \eta}+\left(\frac{\alpha+\Lambda \eta}{\alpha-\Lambda \eta}\right)^2\,\frac{\tan^{-1}\left(\sqrt{\frac{\alpha}{\eta}}r\right)}{\sqrt{\frac{\alpha}{\eta}}r},\label{ff}\\
g(r)&=&\frac{\alpha^2[(\alpha-\Lambda \eta) r^2+2\eta]^2}{(\alpha-\Lambda \eta)^2(\alpha r^2+\eta)^2 f(r)},\\
\vspace{0.4cm}
\psi^2(r)&=&-\frac{2\alpha^2 r^2(\alpha+\Lambda \eta)[(\alpha-\Lambda \eta) r^2+2\eta]^2}{\eta(\alpha-\Lambda \eta)^2 (\alpha r^2+\eta)^3 f(r)}, \label{scalar1}
\end{eqnarray}
where $M$ is considered as the mass of black holes, and $\psi \equiv \phi^{\prime}$, where a prime stands for the first order derivative with respect to $r$. This solution requires $\alpha$ and $\eta$ to have the same sign and $\alpha \neq -\Lambda \eta$. Once $\alpha=-\Lambda \eta$, the Schwarzschild-AdS solution is recovered and the scalar field becomes trivial~\cite{MM}. For simplicity but without the loss of generality, we set $\alpha >0$ and $\eta >0$ in the following context.
Based on refs.~\cite{AAJ,MM}, we write the constraint\footnote{This constraint condition does not work on the analysis of thermodynamic properties because we adopt the method of horizon thermodynamics. Incidentally, its corresponding form for the scenario with charge is given by eq.~(\ref{yueshu2}).} that ensures the reality of the scalar field outside the horizon,
\begin{equation}
\alpha+\Lambda \eta <0. \label{yueshu1}
\end{equation}
It is obvious to deduce such a condition from eq.~(\ref{ff}) and eq.~(\ref{scalar1}).
At first, let us see the asymptotic behavior of eq.~(\ref{ff}): $f(r)$ goes to minus infinity under the limit $r \rightarrow 0$; on the other hand, it goes to plus infinity under the limit $r \rightarrow +\infty$. Therefore, the equation $f(r)=0$ has at least one real root and the largest real root can be regarded as the horizon radius $r_h$. Next, it is evident that $f(r)>0$ once $r>r_h$. Hence, the positivity of eq.~(\ref{scalar1}), i.e., the reality of $\phi(r)$ in the regime $r>r_h$ leads of course to the above inequality.
In addition, it is necessary to take a close look at the behavior of the scalar field in the near horizon region because the scalar field seems to be divergent from eq.~(\ref{scalar1}). Due to $f(r_h)=0$ and $f^{\prime}(r_h)\neq 0$, one can get the Taylor expansion of $f(r)$: $f(r)=f_0+f_1(r-r_h)+f_2(r-r_h)^2 +\cdots$. As to $\psi(r)$, see eq.~(\ref{scalar1}), it approximates to $\frac{1}{\sqrt{f(r)}}$ in this region, which gives rise to the form of the scalar field, $\phi(r)=\phi_0+\phi_1(r-r_h)^{1/2}+\phi_2(r-r_h)^{3/2}+\cdots$. As a result, the scalar field remains finite in the near horizon region.
Now let us revisit the thermodynamic properties of this class of Horndeski black holes.
Along the line of refs.~\cite{DSJT,BPD2,BPD,KM}, one can regard the mass of black holes as the thermodynamic enthalpy,
\begin{equation}
M=\frac{\alpha r_h^3}{6\eta}+\frac{(3\alpha+\Lambda \eta)r_h}{2(\alpha-\Lambda \eta)}+\frac12\sqrt{\frac{\eta}{\alpha}}\left(\frac{\alpha+\Lambda \eta}{\alpha-\Lambda \eta}\right)^2 \tan^{-1}\left(\sqrt{\frac{\alpha}{\eta}}r_h\right). \label{enth1}
\end{equation}
The Bekenstein-Hawking entropy is one-fourth of the event horizon area,
\begin{equation}
S =\pi r_h^2, \label{entro}
\end{equation}
and the thermodynamic temperature can be calculated to be
\begin{equation}
T_h=\left(\frac{\partial M}{\partial S}\right)_{\eta,\Lambda}=\frac{1}{4\pi r_h}\left[\frac{\alpha r_h^2}{\eta}+\frac{\eta(\alpha+\Lambda \eta)^2}{(\alpha r_h^2+\eta)(\alpha-\Lambda \eta)^2}+\frac{3\alpha+\Lambda \eta}{\alpha-\Lambda \eta}\right]. \label{temp1}
\end{equation}
When the thermodynamic pressure $P$ is regarded as
\begin{equation}
P=-\frac{\Lambda}{8\pi}, \label{press}
\end{equation}
the extensive variable conjugate to it, i.e., the thermodynamic volume $V$ has the form,
\begin{eqnarray}
V=\left(\frac{\partial M}{\partial P}\right)_{\eta,S}=-\frac{16\pi \alpha \eta}{(\alpha-\Lambda \eta)^3}\left[(\alpha-\Lambda \eta)r_h+\sqrt{\frac{\eta}{\alpha}}(\alpha+\Lambda \eta)\tan^{-1}\left(\sqrt{\frac{\alpha}{\eta}}r_h\right)\right].\label{case1V}
\end{eqnarray}
Next, we introduce a new intensive thermodynamic variable $\Pi$ in terms of the coupling strength $\eta$,
\begin{equation}
\Pi \equiv\frac{\alpha}{8\pi \eta}, \label{new}
\end{equation}
and derive, with the help of eq.~(\ref{enth1}), the extensive variable conjugate to it,
\begin{eqnarray}
\Theta &=&\left(\frac{\partial M}{\partial \Pi}\right)_{P,S} \nonumber\\
&=&\frac{4\pi r_h^3}{3}+\frac{2\pi \eta^2 r_h(\alpha+\Lambda \eta)^2}{\alpha(\alpha r_h^2+\eta)(\alpha-\Lambda \eta)^2} -2\pi \left(\frac{\eta}{\alpha}\right)^{\frac32}\left(\frac{\alpha+\Lambda \eta}{\alpha-\Lambda \eta}\right)^2\tan^{-1}\left(\sqrt{\frac{\alpha}{\eta}}r_h\right) \nonumber\\
& &-\frac{16\pi\Lambda\eta^2 r_h}{(\alpha-\Lambda\eta)^2}\left[1+\frac{\alpha+\Lambda \eta}{\alpha-\Lambda \eta}\,\frac{\tan^{-1}\left(\sqrt{\frac{\alpha}{\eta}}r_h\right)}{\sqrt{\frac{\alpha}{\eta}}r_h}\right].\label{case1Theta}
\end{eqnarray}
As a result, we can write the first law of thermodynamics,
\begin{equation}
\text{d}M=T_h \text{d}S+V\text{d}P+\Theta \text{d}\Pi, \label{law1}
\end{equation}
and the generalized Smarr relation,
\begin{equation}
M=2T_h S-2PV-2 \Pi \Theta. \label{law2}
\end{equation}
We thus provide a possibility of making up the gap in refs.~\cite{MR,AAJ,MM}, i.e., the Smarr relation can be maintained in the extended phase space that contains $\Pi$ and $\Theta$.
The heat capacity at constant pressure is defined by
\begin{equation}
C_p \equiv \left(\frac{\partial M}{\partial T_h}\right)_P=\frac{\partial M}{\partial r_h}\left(\frac{\partial T_h}{\partial r_h}\right)^{-1}, \label{capac1}
\end{equation}
where the two factors can be calculated to be
\begin{eqnarray*}
\frac{\partial M}{\partial r_h} &=& \frac{\alpha r_h^2}{2\eta}+\frac{\eta(\alpha+\Lambda\eta)^2}{(\alpha r_h^2+\eta)(\alpha-\Lambda\eta)^2}+\frac{3\alpha+\Lambda\eta}{\alpha-\Lambda\eta},\\
\vspace{0.4cm}
\frac{\partial T_h}{\partial r_h}&=&-\frac{1}{4\pi r_h^2}\left[-\frac{\alpha r_h^2}{\eta}+\frac{\eta(\alpha+\Lambda\eta)^2}{(\alpha r_h^2+\eta)(\alpha-\Lambda\eta)^2}+\frac{3\alpha+\Lambda\eta}{\alpha-\Lambda\eta}+\frac{2\alpha\eta r_h^2(\alpha+\Lambda\eta)^2}{(\alpha r_h^2+\eta)^2(\alpha-\Lambda\eta)^2}\right].
\end{eqnarray*}
The Gibbs free energy is the Legendre transform of the enthalpy eq.~(\ref{enth1}), i.e. $G\equiv M-T_h S$. Thanks to eqs.~(\ref{enth1})-(\ref{temp1}), we obtain its exact expression,
\begin{equation}
G=-\frac{\alpha r_h^3}{12\eta}+\frac{(3\alpha+\Lambda \eta)r_h}{4(\alpha-\Lambda \eta)}-\frac{\eta r_h(\alpha+\Lambda \eta)^2}{4(\alpha r_h^2+\eta)(\alpha-\Lambda \eta)^2}+\frac12\sqrt{\frac{\eta}{\alpha}}\left(\frac{\alpha+\Lambda \eta}{\alpha-\Lambda \eta}\right)^2 \tan^{-1}\left(\sqrt{\frac{\alpha}{\eta}}r_h\right). \label{gfree1}
\end{equation}
In order to visualize the thermodynamic quantities, we plot the temperature eq.~(\ref{temp1}), the heat capacity at constant pressure eq.~(\ref{capac1}), and the Gibbs free energy eq.~(\ref{gfree1}) in Figures \ref{tu1}, \ref{tu2} and \ref{tu3}. From the three figures, we can see that the thermodynamic behaviors are similar for $\Lambda=0$ and $\Lambda\neq 0$. For case $\Lambda=0$, according to eq.~(\ref{scalar1}), the scalar field outside the horizon is not real and it can be explained as an extra degree of freedom, rather than a matter field~\cite{MR}. Fortunately, for $\Lambda\neq 0$, i.e., a non-vanishing and negative cosmological constant, it is possible to obtain the real scalar field outside the horizon and the scalar field does not become ghostlike with resorting to eq.~(\ref{yueshu1}). Furthermore, we notice that the thermodynamic behavior of this class of Horndeski black holes with or without the negative cosmological constant is similar to that of the Schwarzschild-AdS black hole. In fact, the negative cosmological constant $\Lambda$ not only plays the role of the thermodynamic pressure, but also gives the constraint eq.~(\ref{yueshu1}) to ensure the reality of the scalar field outside the horizon. Moreover, in the light of similar behaviors of the three thermodynamic quantities between the cases of different $\eta$ but fixed $\Lambda$ and the cases of different $\Lambda$ but fixed $\eta$, as shown in Figures \ref{tu1}, \ref{tu2} and \ref{tu3}, we can conclude that the newly introduced intensive thermodynamic variable eq.~(\ref{new}) plays the similar role to that of thermodynamic pressure. At the end of this subsection, it is necessary to mention that the temperature has one local minimum and the heat capacity at constant pressure undergoes only one divergence, see Figures \ref{tu1} and \ref{tu2}. These behaviors imply that there exists only one phase transition for the Horndeski black holes without charge. In the next subsection, we shall point out that the Horndeski black holes with the charge hair present rich thermodynamic behaviors and critical phenomena.
\begin{figure}
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=70mm]{tem-11} &
\includegraphics[width=70mm]{tem-12} \\
\includegraphics[width=70mm]{tem-13} &
\includegraphics[width=70mm]{tem-14}
\end{tabular}
\end{center}
\caption{When $\alpha=1$, plots of the relation of $T_h$ with respect to $r_h$ for $\eta=1$ (\text{Black}), $2$ (\text{Red}), $3$ (\text{Blue}), and $4$ (\text{Purple}) at $\Lambda=0$ (\text{top left}), $\Lambda=-0.5$ (\text{top right}), $\Lambda=-1$ (\text{bottom left}), and $\Lambda=-1.5$ (\text{bottom right}), respectively.}
\label{tu1}
\end{figure}
\begin{figure}
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=70mm]{cap-11} &
\includegraphics[width=70mm]{cap-12} \\
\includegraphics[width=70mm]{cap-13} &
\includegraphics[width=70mm]{cap-14}
\end{tabular}
\end{center}
\caption{When $\alpha=1$, plots of the relation of $C_p$ with respect to $r_h$ for $\eta=1$ (\text{Black}), $2$ (\text{Red}), $3$ (\text{Blue}), and $4$ (\text{Purple}) at $\Lambda=0$ (\text{top left}), $\Lambda=-0.5$ (\text{top right}), $\Lambda=-1$ (\text{bottom left}), and $\Lambda=-1.5$ (\text{bottom right}), respectively.}
\label{tu2}
\end{figure}
\begin{figure}
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=70mm]{free-11} &
\includegraphics[width=70mm]{free-12}
\end{tabular}
\end{center}
\caption{When $\alpha=1$, plots of the relation of $G$ with respect to $T_h$ for $\eta=1$ (\text{Black}), $2$ (\text{Red}), $3$ (\text{Blue}), and $4$ (\text{Purple}) at $\Lambda=0$ (\text{left}) and $\Lambda=-0.5$ (\text{right}), respectively.}
\label{tu3}
\end{figure}
\subsection{Scenario with charge}
For the class of Horndeski black holes with an electric field, the analytic solution reads~\cite{ACCE,XHHC}
\begin{eqnarray}
f(r)&=&\frac{\alpha r^2}{3\eta}-\frac{2M}{r}+\frac{3\alpha+\Lambda \eta}{\alpha-\Lambda \eta}+\left(\frac{\alpha+\Lambda \eta+\frac{\alpha^2 q^2}{4\eta}}{\alpha-\Lambda \eta}\right)^2\,\frac{\tan^{-1}\left(\sqrt{\frac{\alpha}{\eta}}r\right)}{\sqrt{\frac{\alpha}{\eta}}r}\nonumber \\
& &+\frac{\alpha^2 q^2}{(\alpha-\Lambda \eta)^2 r^2}-\frac{\alpha^2 q^4}{48(\alpha-\Lambda \eta)^2 r^4}+\frac{\alpha^3 q^4}{16\eta(\alpha-\Lambda \eta)^2 r^2},\\
\vspace{0.8cm}
g(r)&=&\frac{\alpha^2[4(\alpha-\Lambda \eta) r^4+8\eta r^2-\eta q^2]^2}{16r^4(\alpha-\Lambda \eta)^2(\alpha r^2+\eta)^2 f(r)},\\
\vspace{0.8cm}
\psi^2(r)&=&-\frac{\alpha^2[4(\alpha+\Lambda \eta)r^4+\eta q^2][4(\alpha-\Lambda \eta) r^4+8\eta r^2-\eta q^2]^2}{32\eta r^6(\alpha-\Lambda \eta)^2 (\alpha r^2+\eta)^3 f(r)},\\
\vspace{0.8cm}
\Psi(r)&=&\Psi_0+\frac14 \frac{q\sqrt{\alpha}}{\eta^{\frac32}}\left[\frac{4\eta(\alpha+\Lambda\eta)+\alpha^2 q^2}{\alpha-\Lambda\eta}\right]\tan^{-1}\left(\sqrt{\frac{\alpha}{\eta}}r\right)\nonumber \\
& &+\frac{\alpha q(\alpha q^2+8\eta)}{4\eta r(\alpha-\Lambda\eta)}-\frac{\alpha q^3}{12r^3(\alpha-\Lambda\eta)}, \label{elepo}
\end{eqnarray}
where $\Psi_0$ is an integration constant. In order to obtain a real scalar field outside the horizon, one needs to impose the following constraint of parameters~\cite{ACCE},
\begin{equation}
4(\alpha+\Lambda \eta)r^4+\eta q^2 <0, \label{yueshu2}
\end{equation}
whose derivation is similar to that of the constraint eq.(\ref{yueshu1}), see the analysis in the above subsection.
From the point of view of thermodynamics, the thermodynamic enthalpy can be written as a function of the horizon radius $r_h$,
\begin{eqnarray}
M&=&\frac{\alpha r_h^3}{6\eta}+\frac{(3\alpha+\Lambda \eta)r_h}{2(\alpha-\Lambda \eta)}+\frac12\sqrt{\frac{\eta}{\alpha}}\left(\frac{\alpha+\Lambda \eta+\frac{\alpha^2 q^2}{4\eta}}{\alpha-\Lambda \eta}\right)^2\tan^{-1}\left(\sqrt{\frac{\alpha}{\eta}}r_h\right)\nonumber\\
& &+\frac{\alpha^2 q^2}{2(\alpha-\Lambda \eta)^2 r_h}-\frac{\alpha^2 q^4}{96(\alpha-\Lambda \eta)^2 r_h^3}+\frac{\alpha^3 q^4}{32\eta(\alpha-\Lambda \eta)^2 r_h}. \label{enth2}
\end{eqnarray}
There are four pairs of thermodynamic variables and in each pair the two variables are conjugate to each other. By following the calculations of eqs.~(\ref{temp1}), (\ref{case1V}), and (\ref{case1Theta}), we derive the first three pairs using eq.~(\ref{enth2}). The first pair consists of the Bekenstein-Hawking entropy eq.~(\ref{entro}) and the thermodynamic temperature,
\begin{eqnarray}
T_h&=&\frac{1}{4\pi r_h}\left\{\frac{\alpha r_h^2}{\eta}+\frac{3\alpha+\Lambda \eta}{\alpha-\Lambda \eta}+\frac{\alpha^2 q^4}{16r_h^4(\alpha-\Lambda\eta)^2}-\frac{\alpha^2 q^2}{r_h^2(\alpha-\Lambda\eta)^2}\right.\nonumber\\
& &\left.-\frac{\alpha^3 q^4}{16\eta r_h^2(\alpha-\Lambda\eta)^2}+\frac{[\alpha^2 q^2+4\eta(\alpha+\Lambda \eta)]^2}{16\eta(\alpha r_h^2+\eta)(\alpha-\Lambda \eta)^2}\right\}. \label{temp2}
\end{eqnarray}
The second pair contains the thermodynamic pressure eq.~(\ref{press}) and the thermodynamic volume,
\begin{eqnarray}
V&=&-\frac{16\pi \alpha \eta}{(\alpha-\Lambda \eta)^3}\left\{(\alpha-\Lambda \eta)r_h+\frac{\alpha q^4(3\alpha r_h^2-\eta)}{96\eta r_h^3}+\frac{\alpha q^2}{2r_h}\right.\nonumber\\
& &\left.+\sqrt{\frac{\eta}{\alpha}}\frac{(\alpha q^2+8\eta)[\alpha^2 q^2+4\eta(\alpha+\Lambda \eta)]}{32\eta^2}\tan^{-1}\left(\sqrt{\frac{\alpha}{\eta}}r_h\right)\right\}.
\end{eqnarray}
The last pair is composed of the coupling strength eq.~(\ref{new}) as a new intensive thermodynamic variable and its conjugate extensive variable,
\begin{eqnarray}
\Theta&=&\frac{4\pi r_h^3}{3}+\frac{\pi\alpha\Lambda\eta^2 q^4}{6r_h^3(\alpha-\Lambda\eta)^3}-\frac{8\pi\alpha\Lambda\eta^2 q^2}{r_h(\alpha-\Lambda\eta)^3}-\frac{16\pi\Lambda\eta^2 r_h}{(\alpha-\Lambda\eta)^2}+\frac{\pi\alpha^2 q^4(1-2\Lambda q)}{4r_h(\alpha-\Lambda\eta)^3}\nonumber\\
& &+\frac{2\pi \eta^2 r_h}{\alpha(\alpha r_h^2+\eta)}\left(\frac{\alpha+\Lambda\eta+\frac{\alpha^2 q^2}{4\eta}}{\alpha-\Lambda \eta}\right)^2-2\pi \left(\frac{\eta}{\alpha}\right)^{\frac32}\left(\frac{\alpha+\Lambda \eta+\frac{\alpha^2 q^2}{4\eta}}{\alpha-\Lambda \eta}\right)^2\tan^{-1}\left(\sqrt{\frac{\alpha}{\eta}}r_h\right)\nonumber \\
& &-2\pi\sqrt{\frac{\eta}{\alpha}}\,\frac{\alpha+\Lambda \eta+\frac{\alpha^2 q^2}{4\eta}}{(\alpha-\Lambda \eta)^3}\left(8\Lambda\eta^2+2\alpha\Lambda\eta q^2-\alpha^2 q^3\right)\tan^{-1}\left(\sqrt{\frac{\alpha}{\eta}}r_h\right).
\end{eqnarray}
In addition, due to the presence of an electric filed, the charge $Q$ of black holes reads
\begin{equation}
Q=\frac{\alpha q}{\alpha-\Lambda\eta},
\end{equation}
and its conjugate intensive variable, i.e., the electric potential $\Phi$ can be obtained,
\begin{eqnarray}
\Phi&=&\left(\frac{\partial M}{\partial Q}\right)_{S,P,\Pi}\nonumber\\
&=&\frac14 \frac{q\sqrt{\alpha}}{\eta^{\frac32}}\left(\frac{4\eta(\alpha+\Lambda\eta)+\alpha^2 q^2}{\alpha-\Lambda\eta}\right)\tan^{-1}\left(\sqrt{\frac{\alpha}{\eta}}r_h\right)\nonumber\\
& &+\frac{\alpha q(\alpha q^2+8\eta)}{4\eta r_h(\alpha-\Lambda\eta)}-\frac{\alpha q^3}{12r_h^3(\alpha-\Lambda\eta)}.\label{elepo11}
\end{eqnarray}
We notice that eq.~(\ref{elepo11}) has a good agreement with eq.~(\ref{elepo}) under the condition of the vanishing integration constant $\Psi_0=0$. Eq.~(\ref{elepo}) was obtained by solving the Maxwell equation of motion eq.~(\ref{emu}), i.e., $\nabla_{\mu}F^{\mu\nu}=0$, while eq.~(\ref{elepo11}) is derived by us through thermodynamic relations. The consistency of the two equations, eqs.~(\ref{elepo}) and (\ref{elepo11}), shows that the thermodynamic method we have adopted is reasonable.
Hence, we give the first law of thermodynamics and the generalized Smarr relation as follows,
\begin{eqnarray}
\text{d}M&=& T_h \text{d}S+V\text{d}P+\Phi \text{d}Q+\Theta \text{d}\Pi,\nonumber\\
\vspace{0.8cm}
M&=& 2T_h S-2PV+\Phi Q-2\Pi\Theta.
\label{law3}
\end{eqnarray}
With the help of eqs.~(\ref{entro}), (\ref{enth2}), and (\ref{temp2}), we can derive the heat capacity at constant pressure,
\begin{equation}
C_p \equiv \left(\frac{\partial M}{\partial T_h}\right)_P=\frac{\partial M}{\partial r_h}\left(\frac{\partial T_h}{\partial r_h}\right)^{-1}, \label{capac2}
\end{equation}
where the factors of the numerator and denominator of eq.~(\ref{capac2}) can be calculated to be, respectively,
\begin{eqnarray*}
\frac{\partial M}{\partial r_h} &=& \frac{\alpha r_h^2}{2\eta}+\frac{\eta(\alpha+\Lambda\eta)^2}{(\alpha r_h^2+\eta)(\alpha-\Lambda\eta)^2}+\frac{3\alpha+\Lambda\eta}{\alpha-\Lambda\eta}+\frac{\alpha^2 q^4}{32r_h^4(\alpha-\Lambda\eta)^2}\\
& &-\frac{\alpha^2 q^2}{2r_h^2(\alpha-\Lambda\eta)^2}-\frac{\alpha^3 q^4}{32\eta r_h^2(\alpha-\Lambda\eta)^2},\\
\vspace{0.4cm}
\frac{\partial T_h}{\partial r_h}&=&-\frac{1}{4\pi r_h^2}\left[-\frac{\alpha r_h^2}{\eta}+\frac{3\alpha+\Lambda\eta}{\alpha-\Lambda\eta}+\frac{\eta\left(\alpha+\Lambda\eta+\frac{\alpha^2 q^2}{4\eta}\right)^2}{(\alpha r_h^2+\eta)(\alpha-\Lambda\eta)^2}-\frac{3\alpha^2 q^2}{r_h^2(\alpha-\Lambda\eta)^2}\right.\\
& &\left.+\frac{5\alpha^2 q^4}{16r_h^4(\alpha-\Lambda\eta)^2}-\frac{3\alpha^3 q^4}{16\eta r_h^2(\alpha-\Lambda\eta)^2}+\frac{2\alpha\eta r_h^2\left(\alpha+\Lambda\eta+\frac{\alpha^2 q^2}{4\eta}\right)^2}{(\alpha r_h^2+\eta)^2(\alpha-\Lambda\eta)^2}\right],
\end{eqnarray*}
and the Gibbs free energy
\begin{eqnarray}
G &\equiv& M-T_h S \nonumber\\
&=&-\frac{\alpha r_h^3}{12\eta}-\frac{\eta r_h\left(\alpha+\Lambda \eta+\frac{\alpha^2 q^2}{4\eta}\right)^2}{4(\alpha r_h^2+\eta)(\alpha-\Lambda \eta)^2}+\frac12\sqrt{\frac{\eta}{\alpha}}\left(\frac{\alpha+\Lambda \eta+\frac{\alpha^2 q^2}{4\eta}}{\alpha-\Lambda \eta}\right)^2 \tan^{-1}\left(\sqrt{\frac{\alpha}{\eta}}r_h\right)\nonumber\\
& &+\frac{(3\alpha+\Lambda \eta)r_h}{4(\alpha-\Lambda \eta)}-\frac{5\alpha^2 q^4}{192r_h^3(\alpha-\Lambda\eta)^2}+\frac{3\alpha^2 q^2}{4r_h(\alpha-\Lambda\eta)^2}+\frac{3\alpha^3 q^4}{64\eta r_h(\alpha-\Lambda\eta)^2}. \label{gfree2}
\end{eqnarray}
We plot the thermodynamic temperature eq.~(\ref{temp2}), the heat capacity at constant pressure eq.~(\ref{capac2}), and the Gibbs free energy eq.~(\ref{gfree2}) in Figures \ref{tu4}, \ref{tu5}, \ref{tu6}, and \ref{tu7}. These figures show rich thermodynamic behaviors and critical phenomena if we compare to the case of no charge. In Figure \ref{tu4}, for fixed values of $\alpha$, $\eta$, and $\Lambda$, the thermodynamic temperature presents one local maximum and one local minimum, and the heat capacity at constant pressure undergoes two times of divergence for different values of $q$. As is known, black holes are locally stable for $C_p > 0$, while unstable for $C_p <0$. The behavior of the heat capacity at constant pressure depicted by Figure \ref{tu4} indicates that the Horndeski black holes undergo two phase transitions: the first phase transition happens from a locally stable state to a locally unstable one at the local maximum temperature, and the other phase transition occurs from a locally unstable state to a locally stable one at the local minimum temperature. When $q$ is larger than its critical value $q_c=0.45049$, the temperature has no extrema and the heat capacity at constant pressure has no divergences. In addition, the characteristic swallowtail behavior of the Gibbs free energy disappears once $q>q_c=0.45049$, as shown in Figure \ref{tu6}. All of these phenomena imply that no phase transitions happen when the charge parameter $q$ exceeds its critical value $q_c$. On the other hand, the similar critical phenomenon appears for different values of $\Lambda$ but fixed values of $\alpha$, $\eta$, and $q$, as shown in Figures \ref{tu5} and \ref{tu7}. When the cosmological parameter $\Lambda$ is less than its critical value $\Lambda_c=-3.1631$, no phase transitions occur. Furthermore, we observe that the thermodynamic temperature goes to zero at a very small horizon radius. It is the electric charge that provides negative contributions in eq.~(\ref{temp2}), which makes it possible that the thermodynamic temperature vanishes.
\begin{figure}
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=70mm]{tem-21} &
\includegraphics[width=70mm]{cap-21}
\end{tabular}
\end{center}
\caption{When $\alpha=1$, $\eta=2$, and $\Lambda=-1.5$, plots of the relations of $T_h$ and $C_p$ with respect to $r_h$ for $q=0.3$ (\text{Black}), $q=0.4$ (\text{Red}), $q_c=0.45049$ (\text{Green}), and $q=0.5$ (\text{Purple}), respectively.}
\label{tu4}
\end{figure}
\begin{figure}
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=70mm]{tem-22} &
\includegraphics[width=70mm]{cap-22} \\
\includegraphics[width=70mm]{tem-23} &
\includegraphics[width=70mm]{cap-23}
\end{tabular}
\end{center}
\caption{When $\alpha=1$, $\eta=2$, and $q=0.3$, plots of the relations of $T_h$ and $C_p$ with respect to $r_h$ for $\Lambda=0$ (\text{Black}), $\Lambda=-2.5$ (\text{Red}), $\Lambda_c=-3.1631$ (\text{Green}), and $\Lambda=-3.5$ (\text{Purple}), respectively.}
\label{tu5}
\end{figure}
\begin{figure}
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=70mm]{free-21} &
\includegraphics[width=70mm]{free-22} \\
\includegraphics[width=70mm]{free-23} &
\includegraphics[width=70mm]{free-24}
\end{tabular}
\end{center}
\caption{When $\alpha=1$, $\eta=2$, and $\Lambda=-1.5$, plots of the relation of $G$ with respect to $T_h$ for different values of $q$.}
\label{tu6}
\end{figure}
\begin{figure}
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=70mm]{free-31} &
\includegraphics[width=70mm]{free-32} \\
\includegraphics[width=70mm]{free-33} &
\includegraphics[width=70mm]{free-34}
\end{tabular}
\end{center}
\caption{When $\alpha=1$, $\eta=2$, and $q=0.3$, plots of the relation of $G$ with respect to $T_h$ for different values of $\Lambda$.}
\label{tu7}
\end{figure}
\section{Conclusion}\label{sec4}
Based on the well-accepted consideration~\cite{CEJM,DSJT,BPD2,BPD,KM}, namely the cosmological constant as the thermodynamic pressure and the mass of black holes as thermodynamic enthalpy, we revisit thermodynamic properties of a new class of Horndeski black holes whose action contains a non-minimal kinetic coupling of one massless real scalar and the Einstein tensor. We resort to a new intensive thermodynamic variable, see eq.~(\ref{new}), which originates from the coupling strength $\eta$ with the dimension of length square, and thus deduce both the generalized first law of thermodynamics and the generalized Smarr relation, see eqs.~(\ref{law1}), (\ref{law2}), and (\ref{law3}). By calculation of some thermodynamic quantities, such as the thermodynamic temperature, the heat capacity at constant pressure, and the Gibbs free energy, our result indicates that this class of Horndeski black holes presents rich thermodynamic behaviors and critical phenomena. Especially in the case of the presence of an electric field, the black holes undergo two times of phase transitions: the first phase transition happens from a locally stable state to a locally unstable one, and the other phase transition occurs from a locally unstable state to a locally stable one. Once the charge parameter $q$ exceeds its critical value $q_c$, or the cosmological parameter $\Lambda$ does not exceeds its critical value $\Lambda_c$, no phase transitions happen and the black holes are stable. As a by-product, we indicate that the behavior of the coupling strength acts as that of the thermodynamic pressure, as shown in Figures \ref{tu1}, \ref{tu2}, and \ref{tu3}.
\section*{Acknowledgments}
Y-GM would like to thank H. Nicolai of Max-Planck-Institut f\"ur Gravitationsphysik (Albert-Einstein-Institut) for kind hospitality. This work was supported in part by the National Natural
Science Foundation of China under grant No.11675081. At last, the authors would like to thank the anonymous referee for the helpful comment that indeed greatly improves this work.
|
\section{Introdcution}
Let $M$ be a finite dimensional smooth connected complete and stochastically complete Riemannian manifold $M$ whose Riemannian distance is denoted by $r$.
By stochastic completeness we mean that its minimal heat kernel satisfies that $\int p_t(x,y)dy=1$. Denote by $C([0,1]; M)$ the space of continuous curves: $ \sigma: [0,1]\to M$, a Banach manifold modelled on the Wiener space. A chart containing a path $\sigma$ is given by a tubular neighbourhood of $\sigma$ and the coordinate map is induced from the exponential map given by the Levi-Civita connection on the underlying finite dimensional manifold.
For $x_0, y_0 \in M$ we denote by $C_{x_0}M$ and
$C_{x_0, y_0}M$, respectively, the based and the pinned space of continuous paths over $M$: \begin{equs}C_{x_0}M\;\;&=\{ \sigma \in C([0,1]; M): \quad \sigma(0)=x_0\}, \\
C_{x_0, y_0}M&=\{\sigma \in C([0,1]; M): \quad \sigma(0)=x_0, \quad \sigma(1)=y_0\}.
\end{equs}
The pullback tangent bundle of $C_{x_0}M$ consisting of continuous $v: [0,1]\to TM$ with $v(0)=0$ and $v(t)\in T_{\sigma(t)}M$ where
$\sigma \in C([0,1]; M)$ which for each $\sigma$ can be identified by parallel translation with continuous paths on $T_{x_0}M$, the latter is identified with ${\mathbb R}^n$ with a frame $u_0$.
To define gradient operators we make a choice of a family of $L^2$ sub-spaces together with an Hilbert space structure,
and so we have a family of continuously embedded $L^2$ subspaces ${\mathcal H}_\sigma$ and the $L^2$ sub-bundle ${\mathcal H}:=\cup_\sigma {\mathcal H}_\sigma$.
Firstly denote by $H$ the Cameron-Martin space over ${\mathbb R}^n$, \begin{equation*}
H:= \bigg\lbrace h\in C( [0,1]; {\mathbb R}^n): h(0)=0, |h|_{H^1}: =\left(\int_0^1 |\dot h_s|^2 ds\right)^{\frac 12} < \infty\bigg\rbrace,
\end{equation*}
with $H^0$ its subset consisting of $h$ with $h(1)=0$.
If $\parals_\cdot(\sigma)$ denotes stochastic parallel translation along a path $\sigma$
we denote
by ${\mathcal H}_\sigma$ and ${\mathcal H}^0_\sigma$ the Bismut tangent spaces:
\begin{equs}{\mathcal H}_\sigma &=\{ \parals_\cdot(\sigma) h: h\in H\}, \qquad
{\mathcal H}^0_\sigma =\{ \parals_\cdot(\sigma) h: h\in H, h(1)=0\},
\end{equs} specifying
respectively the `admissible' tangent vectors at $\sigma\in C_{x_0}M$ and vectors at $\sigma\in C_{x_0, y_0}M$.
These vector spaces are given the inner product inherited from the Cameron-Martin space $H$.
For an $L^2$ analysis on $C_{x_0, y_0}M$ we need a probability measure on it which is usually taken to be the probability distribution of the conditioned Brownian motion. The heat kernel measure, the distribution of a Brownian sheet, offers also an alternative measure, see \cite{Malliavin-loop, Driver-Lohrenz, Norris-sheets}.
See also \cite{Li-hypoelliptic-bridge} for a study of the measure induced by a conditioned hypoelliptic stochastic process.
If we suppose that $M$ has a pole $y_0$, by which we mean that the exponential map $\exp_{y_0}: T_{y_0} M\to M$ is a diffeomorphism, another probability measure, the probability distribution of the semi-classical Riemannian bridge, becomes available to us. For a simply connected Riemannian manifold with non-negative sectional curvature, every point is a pole.
We denote this measure by $\nu=\nu_{x_0,{y_0}}$ and denote by $L^2(C_{x_0,y_0}M;{\mathbb R})$ the corresponding $L^2$ space.
A semi-classical Riemannian Brownian bridge $(\tilde{x}_s, s\le 1)$
is a time dependent strong Markov process with generator $\frac{1}{2}\triangle + \nabla \log k_{1-s}(\cdot, y_0)$ where, $$k_t(x_0, y_0):=(2\pi t)^{-\frac n 2}e^{-\frac {r^2(x_0, y_0)}{2t}}
J^{-\frac 12}(x_0),$$
and $J(y)= | \det D_{\exp_{y_0}^{-1}(y)} \exp_{y_0}|$ is the Jacobian determinant of the exponential map at $y_0$.
Semi-classical Riemannian Brownian bridges (semi-classical bridge for short) were introduced by K. D. Elworthy and A. Truman \cite{Elworthy-Truman-81}.
For further explorations in this direction see \cite{Elworthy-Truman-Waltling} and \cite{Ndumn-09}. If $p_t$ is the heat kernel, the Brownian bridge is a Markov process with generator
$\frac 12 \Delta + \nabla \log p_{1-t}(\cdot, y_0)$. Let us consider the two time dependent potential functions that drives the Brownian motion to the terminal value.
They are close to each other as $t\to 1$, by Varadhan's asymptotic relations \cite{Varadhan}: $ (1-t) \log p_{1-t}(x,y_0)\sim-\frac 1 2r^2(x,y_0)$. There is also the relation $\lim_{t\to 1} (1-t) \nabla \log p_{1-t}(x,y)=-\dot \gamma(0)$ where $\gamma$ is normal geodesic from $y_0$ to $x$.
The two drift vector fields
$\nabla \log p_{1-t}(\cdot, y_0)$ and $\nabla \log k_{1-t}(\cdot, y_0)$ differ by $-\frac 12 \nabla \log J$ near the terminal time.
Let us consider the unbounded linear differential operator $d$ on $L^2(C_{x_0,y_0}M;{\mathbb R})$ taking values in $L^2(\cup_\sigma{\mathcal H}_\sigma^*)$ where for $v\in \cup_\sigma{\mathcal H}_\sigma^*$,
$$\|v_\cdot(\cdot)\|:=\left(\int_{C_{x_0,y_0}M }\left(|\parals_\cdot^{-1}v_\cdot(\sigma)|_H\right)^2 d\nu(\sigma)\right)^{\frac 12}.$$
Another norm can be given, taking into accounts of the damping effects of the Riccic curvature, which will be discussed later.
As the distance function from the semi-classical bridge to the pole is precisely the $n$-dimensional Bessel bridge where $n=\dim(M)$, the advantage of the semi-classical Brownian bridge measure is that it is easier to handle, which we demonstrate by studying the elementary property of the divergence operator.
Our main result is an integration by parts formula for $d$. Such a formula is believed to be equivalent to an integration by parts formula. A proof for the equivalence was given in \cite{Elworthy-Li-ibp} for compact manifold and for the Brownian motion measure by induction. The same method should work here. However since it is a bridge measure the current method has its advantages. First order Feyman-Kac type formulas together with estimates for the gradient of the Feyman-Kac kernel using semi-classical bridge process and the damped stochastic parallel translation was obtained in \cite{Li-Thompson}.
Denote by $OM$ the space of orthonormal frames over $M$ and $\lbrace H_i \rbrace$ the canonical horizontal vector fields on $OM$ associated to an orthonormal basis $\{e_i\}$ of ${\mathbb R}^n$ so that $H_i$ is the horizontal lift of $ue_i$. For a tangent vector $v$ on $M$, we will denote by $\tilde v$ the horizontal lift of $v$ to $TOM$. Let $\lbrace \Omega,\mathcal{F},\mathcal{F}_t,\mathbb{P}\rbrace$ be a filtered probability space on which is given a family of independent one-dimensional Brownian motions $\lbrace B^i \rbrace$. We define $B_t=(B_t^1, \dots, B_t^n)$. Let $u_0\in \pi^{-1}(x)$ be a frame at $x$, $u_t$ and $\tilde u_t$ be the solution to the stochastic differential equations,
\begin{equation}\label{horizontal2}
d {u}_s = \sum_{i=1}^n H_i({u}_s)\circ dB^i_s, \quad
d \tilde{u}_s = \sum_{i=1}^n H_i(\tilde{u}_s)\circ dB^i_s + \tilde{A_s}(\tilde{u}_s)ds, \quad \tilde u_0=u_0,
\end{equation}
where $\circ$ denote Stratonovich integration and $A_s=\nabla \log k_{1-s}(\cdot, y_0)$. Then $\tilde{x}_s := \pi(\tilde{u}_s)$ is a semi-classical Brownian bridge from $x_0$ to $y_0$ in time~$1$.
Let ${\mathop{\rm {Ric}}}_x$ denote the Ricci curvature at $x \in M$, by ${{\mathop{\rm {Ric}}}}^{\sharp}_x:T_xM \rightarrow T_xM$ we mean the linear map given by the relation $\< {{\mathop{\rm {Ric}}}}^\sharp_x u,v\> = {{\mathop{\rm {Ric}}}_x}(u,v)$.
Denote $r=r(\cdot, y_0)$ for simplicity. We will need the following geometric conditions.
Set \begin{equation}
\Phi=\frac{1}{2}J^{\frac{1}{2}}\triangle J^{-\frac{1}{2}}=\frac 14|\nabla {\log} J|^2-\frac 1 4 \Delta ({\log }J).
\end{equation}
\noindent
{\bf C1:} The Ricci curvature is bounded.\\
{\bf C2:} $|\nabla \Phi|+|\nabla( \log J)|\le c(e^{ a r^2}+1)$ for some $c>0$ and $a$ is an explicit constant to be given later.\\
{\bf C3:} $\Phi$ is bounded from below.\\
{\bf C4:} For each $t$, $k_t$ and $|\nabla k_t|$ are bounded, $|\nabla \Phi|$ is bounded.\\
The condition that the Ricci curvature is bounded ensures that the solution to the canonical SDE is differentiable in the sense of Malliavin calculus. It also implies that $|\tilde W_t|$ is bounded and that the integration by parts formula holds for
the Brownian motion measure. Observe that $k_t$ and $|\nabla k_t|$ are bounded if $rJ^{-\frac 12} $ and $J^{-\frac 12} \nabla \log J^{-\f12}$ grow at most exponentially. Here we do not strive for the best possible conditions, as the optimal conditions will manifest themselves when Clark-Ocone formula and Poincar\'e inequalities are studied.
Our main results is the following integration by parts theorem.
\begin{theorem}\label{thm1}
Assume {\bf C1- C4} hold.
Then for any $F, G\in{\mathop {\rm Cyl}}$ and $h\in H^0$ the following integration by parts formula hold.
\small
\begin{equs}
&\int_{C_{x_0, y_0}M} G(\tilde x_\cdot)dF\left(\tilde u_\cdot(\sigma) h_\cdot \right) \nu(d\sigma)+\int_{C_{x_0, y_0}M} F(\tilde x_\cdot)dG\left(\tilde u_\cdot(\sigma) h_\cdot \right) \nu(d\sigma)\\
=&{\mathbf E} \left[(FG) (\tilde x_\cdot ) \int_0^1 \<\dot h_s+\frac 1 2 \tilde u_s^{-1}{{\mathop{\rm {Ric}}}}^{\sharp}
(\tilde u_s h_s), d\tilde B_s\>\right]
+{\mathbf E}\left[(FG)(\tilde x_\cdot) \int_0^1 d\Phi(\tilde u_s h_s)ds \right].
\end{equs}
\normalsize
Here $d\tilde{B}_s = dB_s -\tilde u_s^{-1} \nabla \log k_{1-s}(\tilde{x}_s)\,ds$. In particular $d: {\mathop {\rm Cyl}} \subset L^2(C_{x_0, y_0}M)\to L^2(\cup_\sigma {\mathcal H}_\sigma)$ is closable, the domain of $d^*$ contains ${\mathop {\rm Cyl}}$ and
$$d^*G=-dG+G \int_0^1 \<\dot h_s+\frac 1 2 \tilde u_s^{-1}{{\mathop{\rm {Ric}}}}^{\sharp}
(\tilde u_s h_s), d\tilde B_s\>+G\int_0^1 d\Phi(\tilde u_s h_s)ds.$$
\end{theorem}
For based path space over a compact manifold, with Brownian motion measure (the Wiener measure), this was proved in \cite{Driver92}, for
non-compact manifolds see
\cite{Elworthy-Li-ibp, Elworthy-Li-icm},
\cite{Fang-Wang}, \cite{Thalmaier-Wang}, and
\cite{Arnaudon-Driver-Thalmaier}. For pinned manifolds with measure coming from the classical Brownian bridge measure, this was proved in \cite{Driver94} and \cite{Hsu97}.
Let us now clarify the definition of $d$. A common definition for $d$, which we use, is to take its initial domain to be ${\mathop {\rm Cyl}}$, the set of cylindrical functions of the form
$F(\sigma)=f(\sigma_{t_1}, \dots, \sigma_{t_m})$ where $m\in {\mathcal N}$,
$0<t_1<t_2<\dots <t_m< 1$, and $f$ is a $BC^1$ function on the $m$-fold product space of $M$, or ${\mathop {\rm Cyl}}_0$ the subset containing $f(\sigma_{t_1}, \dots, \sigma_{t_m})$ where $f$ is compactly supported.
The $H$-derivative (Malliavin derivative) of $F$ in the direction of
$u_\cdot(\sigma) h_\cdot\in T_\sigma C_{x_0}M$~is:
$$(dF)(\parals_\cdot(\sigma) h_\cdot)=\sum_{k=1}^m \partial_{k}f \big(\parals_{t_k}(\sigma) h_{t_k}\big),$$
where $\partial_{k}f $ denotes the derivative of $f$ in its $k$th component and $/\kern-0.55ex/$ denotes parallel translation and identified with $u$ in the sequel. Denote by
$G(s,t)$ and $G^0(s,t)$, respectively, the Green's functions of $\frac d {ds}$ on $(0,1)$ with suitable Dirichlet conditions: $G(s,t)=s\wedge t$ and $G^0(s,t)=s\wedge t -st$.
Then \begin{equs}(\nabla F)(\sigma)(t)&=\sum_{k=1}^m G(t_k,t) \parals_{t_k,t}(\sigma)\nabla_{k}f (\sigma_{t_1},\sigma_{t_2},\dots, \sigma_{t_m}),\\
(\nabla^0 F)(\sigma)(t)&=\sum_{k=1}^m G^0(t_k,t) \parals_{t_k,t}(\sigma)\nabla_{k}f (\sigma_{t_1},\sigma_{t_2},\dots, \sigma_{t_m}),
\end{equs}
where $\nabla_k f$ denotes the gradient of $f$ in the $k$th variable.
We have
\begin{equs}\|\nabla F\|^2&=\sum_{i,j=1}^m G(t_k, t_j) \</\kern-0.55ex/_{t_k, t_j}\nabla_k f, \nabla_jf\>,\\ \quad \|\nabla^0 F\|^2&=\sum_{i,j=1}^m G^0(t_k, t_j) \</\kern-0.55ex/_{t_k, t_j}\nabla_k f, \nabla_jf\>.
\end{equs}
It is an open problem whether the closure of $d$ with initial domain $BC^1$ agrees wth the closure of $d$ with initial value the cylindrical functions. This is the Markov uniqueness problem, this was studied In \cite{Elworthy-Li-CR} where it was only proved that the latter including $BC^2$.
\section{Proof of Theorem \ref{thm1}}
\label{proof}
To clarify the singularities at the terminal time we first prove a lemma concerning the divergence of a suitable vector field on the path space.
Let $\tilde u_t$ be as defined by (\ref{horizontal2}), $\tilde x_t=\pi(\tilde u_t)$. Recall that $k_t(x_0, y_0)=(2\pi t)^{-\frac n 2}e^{-\frac {\rho^2(x_0, y_0)}{2t}}J^{-\frac 12}(x_0)$
and $\tilde{B}_s = B_s - \tilde u_s^{-1}\nabla \log k_{1-s}(\tilde{x}_s, y_0)\,ds$.
The reference to $y_0$ will be dropped from time to time for simplicity. Define ${\mathop{\rm ric}}_u=u^{-1}{\mathop{\rm {Ric}}}^\sharp u$.
\begin{lemma}\label{lemma1}
Assume stochastic completeness, {\bf C2}, and $h\in H^{0}$. Then the following integral exists,
$$\int_0^1\left \<\dot h_s+\frac 1 2 {\mathop{\rm ric}}_{\tilde u_s} (h_s), d\tilde B_s\right\>.$$
Furthermore,
$$\lim_{t\to 1}{\mathbf E}\left( \left\<\nabla \ {\log} {k_{1-t}(\cdot)}, \tilde u_t h_t\right\>\right)^2=0,$$
$\int_0^t \<\dot h_s+\frac 1 2 {\mathop{\rm ric}}_{\tilde u_s} (h_s), d\tilde B_s\>$ converges, as $t\to 1$, in
$L^2(\Omega, \mathbb{P})$; and
$$\begin{aligned} \int_0^1\left \<\dot h_s+\frac 1 2 {\mathop{\rm ric}}_{\tilde u_s} (h_s), d\tilde B_s\right\>
=&\int_0^1\left \<\dot h_s+\frac 1 2 {\mathop{\rm ric}}_{\tilde u_s} (h_s), d B_s\right\>+\int_0^1 d\Phi (\tilde u_s h_s) ds\\
&+\int_0^1 \nabla d\left( { \log} k_{1-s} (\tilde x_s, y_0)\right) (\tilde u_sdB_s, \tilde u_sh_s).
\end{aligned}$$
\end{lemma}
\begin{proof}
The singularities in the integral $\int_0^1\left \<\dot h_s+\frac 1 2 {\mathop{\rm ric}}_{\tilde u_s} (h_s), d\tilde B_s\right\>$ come from the involvement of $ \nabla \log k_{1-s}(\tilde{x}_s,y_0)$ and we only need to worry about
\begin{equation}
\label{singular}
\alpha_t:=\int_0^t \left\<\dot h_s+\frac 1 2 {\mathop{\rm ric}}_{\tilde u_s} (h_s),\tilde u_s^{-1}\nabla \log k_{1-s}(\tilde{x}_s,y_0)\right\>ds.
\end{equation}
We integrate by parts to deal with $\int_0^t \left\<\dot h_s,\tilde u_s^{-1}\nabla \log k_{1-s}(\tilde{x}_s,y_0)\right\>ds$, which involves the derivative of $h_s$.
Since $\frac D{ds} (u_sh_s)=u_s\dot h_s$, by stochastic calculus applied to
$d \left( {\log} k_{1-s}\right)(u_s h_s)$, where $d$ denotes spatial differentiation with respect to the $M$-valued variable,
we see that
\begin{equs}
& \left\< \nabla {\log} {k_{1-t}(\tilde x_t)}, \tilde u_t h_t\right\>\\
=&\int_0^t \left\<\nabla { \log} k_{1-s} (\tilde x_s), \tilde u_s\dot h_s \right\>\, ds+\sum_{i=1}^n\int_0^t \nabla d\left( { \log} k_{1-s}\right) (\tilde u_se_i, \tilde u_sh_s) dB_s^i\\
&+\int_0^t \nabla d\left( { \log} k_{1-s}) \left(\nabla {\log} k_{1-s} (\tilde x_s), \tilde u_sh_s \right)\right)\, ds\\
&+\int_0^t\left( \frac 1 2\mathop{\rm trace} \nabla^2 + {\partial_r} \right)\left( D \left( {\log} {k_{1-s}(\tilde x_s)}\right)\right)(\tilde u_s h_s)\, ds,
\end{equs}
the first term on the right hand side being $\alpha_t$.
Since $\nabla {\log} {k_{1-s}(x)}=-\frac {r (x) \nabla r(x)}{1-s} +\nabla {\log} (J^{-\frac 12})$, $\Delta r={n-1 \over r} +\<\nabla r, \nabla \log J\>$, the following set of formulas are easy to verify.
\begin{equation}
\label{formulas}
{\begin{split}
\Delta {\log} k_{1-s}&=-\frac n {1-s} -\frac {r\<\nabla r,\nabla \log J\>}{1-s}-\frac 12 \Delta(\log J),\\
{\partial_ r} \log k_{1-s} &=\frac n {2(1-s)}-\frac {r^2} {2(1-s)^2},\\
|\nabla \log k_{1-s}|^2&=\frac {r^2}{(1-s)^2}+\frac 12|\nabla \log J|^2+\frac {r\<\nabla r,\nabla \log J\>}{1-s}.
\end{split}}
\end{equation}
It follows that
$$\left(\frac 1 2 \Delta + {\partial_r}\right)\left( {\log} {k_{1-s}}\right) +\frac 1 2|\nabla \log k_{1-s} |^2=\frac 14|\nabla {\log} J|^2-\frac 1 4 \Delta ({\log }J)=\Phi.$$
Let $\Delta^1:=(dd^*+d^*d)$ denote the Laplace-Beltrami Kodaira operator on differential 1-forms.
By the Weitzenb\"ock formula,
$\left( \frac 1 2\mathop{\rm trace} \nabla^2 + {\partial_r} \right)d =\left( \frac 1 2\Delta^1 d +\frac 1 2 {\mathop{\rm {Ric}}}^{\sharp}d+ {\partial_r}d \right),$
and consequently,
\begin{equs}
&\left( \frac 1 2\mathop{\rm trace} \nabla^2 + {\partial_r} \right)d \left( {\log} {k_{1-s}(\tilde x_s)}\right)\\
=&d\left( \frac 1 2 \Delta + {\partial_r} \right) \left( {\log} {k_{1-s}(\tilde x_s)}\right)
+\frac 1 2 {\mathop{\rm {Ric}}}^{\sharp}\left(d {\log} {k_{1-s}(\tilde x_s)}\right)\\
=&-\frac 1 2 d(|\nabla {\log k}_{1-s}(\cdot) |^2)+d\Phi
+\frac 1 2 {\mathop{\rm {Ric}}}^{\sharp}\left(d {\log} {k_{1-s}(\tilde x_s)}\right).
\end{equs}
Thus
\begin{equs}{}& \nabla d \left( { \log} k_{1-s}) \left(\nabla {\log} k_{1-s} (\tilde x_s), \tilde u_sh_s \right)\right)
+\left( \frac 1 2\mathop{\rm trace} \nabla^2 +\frac {\partial} {\partial r} \right)\left( d \left( {\log} {k_{1-s}}\right)\right)(\tilde u_s h_s)\\
&=d\Phi( \tilde u_sh_s)
+\frac 1 2 {\mathop{\rm {Ric}}}\left(\nabla{\log} {k_{1-s}(\tilde x_s)}, \tilde u_sh_s\right).
\end{equs}
Let us return to $ \left\< \nabla {\log} {k_{1-t}(\tilde x_t)}, \tilde u_t h_t\right\>$:
\begin{equs}
& \left\< \nabla {\log} {k_{1-t}(\tilde x_t)}, \tilde u_t h_t\right\>\\
=&\int_0^t \left\<\nabla { \log} k_{1-s} (\tilde x_s), \tilde u_s\dot h_s \right\>\, ds
+\sum_{i=1}^n\int_0^t \nabla d\left( { \log} k_{1-s} \right) (\tilde u_se_i, \tilde u_sh_s) dB_s^i\\
&+\int_0^t d\Phi( \tilde u_sh_s)\;ds
+\frac 1 2 \int_0^t{\mathop{\rm {Ric}}}\left(\nabla{\log} {k_{1-s}(\tilde x_s)}, \tilde u_sh_s\right)\;ds.
\end{equs}
We thus obtain
the following relation:
\begin{equs}
\alpha_t=&\left\< \nabla {\log} {k_{1-t}(\tilde x_t)}, \tilde u_t h_t\right\>+
\frac 1 2\int_0^t D \log k_{1-s}({\mathop{\rm {Ric}}}^{\sharp}(\tilde u_s h_s))ds\\=&
\left\<\nabla \ {\log} {k_{1-t}(\cdot)}, \tilde u_t h_t\right\>-\int_0^t \left\< \nabla \Phi, \tilde u_sh_s\right\> ds\\
&-\sum_{i=1}^n\int_0^t \nabla d\left( { \log} k_{1-s} \right) (\tilde u_se_i, \tilde u_sh_s) \,dB_s^i.
\end{equs}
We will prove that each of the terms on the right hand side converges as $t$ approaches $1$. Furthermore
$\left\<\nabla \ {\log} {k_{1-t}(\cdot)}, \tilde u_t h_t\right\>$ converges to zero.
We first observe that there exists a constant $C$ such that ${\mathbf E}[ r(\tilde x_t)^p]
\le Ct^{\frac p2}$.
Indeed $r_t:=\rho(\tilde x_t, y_0)$ satisfies
\begin{equs}
r_t-r_0=&
\beta_t +\int_0^t{1\over 2} \Delta r(\tilde{x}_s) ds-
\int_0^t {r(\tilde{x}_s) \over 1-s}ds
-{1\over 2} \int_0^t \left\< \nabla r, \nabla \log J \right\>_{\tilde{x}_s}ds\\
=&\beta_t +\int_0^t \frac {n-1} {2r_s} ds-
\int_0^t \frac{r_s} {1-s}ds,
\end{equs}
where $\beta_t$ is a one dimensional Brownian motion and we have used the fact that
$\Delta r={n-1 \over r} +\<\nabla r, \nabla \log J\>$. Thus $r_s$ is a Bessel bridge starting at $\rho(x_0, y_0)$ and ending at $0$ at time ${1}$. In particular $\lim_{t\uparrow 1} \tilde x_t=y_0$ and $(r_t, t\le 1)$ is a continuous semi-martingale. Furthermore for any $p>1$, ${\mathbf E} [r(\tilde x_t)^p] \le C t^{\frac p 2}$. If $K_t$ denotes the standard Gaussian kernel on ${\mathbb R}^n$ then for $z_1, z_2\in {\mathbb R}^n$ with $|z_1-z_2|=\rho(x_0, y_0)$,
$$
{\mathbf E}[ r(\tilde x_t, y_0)^p]=\frac 1 {K_T(z_1, z_2)}\int_{{\mathbb R}^n} |z-z_2|^p K_t(z_1, z) K_{1-s}(z, z_2) dz\le C |z-z_1|^{\frac p 2}.$$
We also know that
${\mathbf E} [e^{2a r_t^2}]<\infty$ for some $a$ and $t\le 1$, involking condition C2.
We show below that (\ref{singular}) has a limit as $t\to 1$.
Firstly, since $|d\Phi|\le c e^{ar^2}$,
$$\lim_{t\to 1}{\mathbf E} \left[\int_t^1 \left\< \nabla \Phi, \tilde u_sh_s\right\> ds\right]^2= 0.$$
We work with the first term on the right hand side: $$\left\<\nabla \ {\log} {k_{1-t}(\cdot, y_0)}, \tilde u_t h_t\right\>=\frac { r(\tilde x_t) \<\nabla r(\tilde x_t), \tilde u_t h_t\>}{1-t} +
\<\nabla {\log} J_{\tilde x_t}^{-\frac 12},\tilde u_t h_t\>.$$
Since $|d( {\log} J_x^{-\frac 12})|\le c e^{ar(x)^2}$,
$\lim_{t\to 1}\<\nabla {\log} J_{\tilde x_t}^{-\frac 12},\tilde u_t h_t\>$ converges in $L^2(\Omega)$. Thus \begin{equation}\label{2.1}
\lim_{t\uparrow 1} {\mathbf E} \left|\<\nabla {\log} J^{-\frac 12}(\tilde x_t),\tilde u_t h_t\>\right|^2=0,
\end{equation}
using the fact that $h_t\to 1$.
Also, by the symmetry of the Euclidean bridge, ${\mathbf E}[ r^2(\tilde x_t, y_0)]\le C\left( t\wedge (1-t)\right)$ and hence
$${\mathbf E} \left |\frac {r (\tilde x_t) \left\<\nabla r(\tilde x_t),\tilde u_t h_t\right\>}{1-t} \right|^2\le \frac {|h_t|^2} {1-t}.$$
Since $h_1=0$, and $h\in H$,
$$\frac{ |h_t|^2} {1-t}=\frac 1 {1-t}\left| \int_t^1 \dot h_s dr\right|^2\le \int_t^1 |\dot h_s|^2 ds\to 0,$$
as $t\to 1$, using the fact that $h\in H$.
We conclude that
$$\lim_{t\to 1}{\mathbf E} \left[\left\<\nabla \ {\log} {k_{1-t}(\cdot)}, \tilde u_t h_t\right\>\right]^2=0.$$
For the final term we observe that
$$\nabla d\left( { \log} k_{1-s} \right) (\tilde u_se_i, \tilde u_sh_s)
=-\frac {\nabla r(\tilde u_se_i)\nabla r(\tilde u_s h_s)}{1-s}-\frac {r \nabla dr (\tilde u_se_i, \tilde u_sh_s)}{1-s}.$$
We further observe that the Frobenius norm of
the Hessian of the distance function satisfies:
$$\|\nabla d r\|_F:=\left(\sum_{i,j=1}^n \<\nabla_{E_i} {\partial r}, E_j\>\right)^{\frac 12} \le \frac 1{\sqrt{n-1}}\Delta r
\le \frac 1{\sqrt{n-1}} \left({n-1 \over r} +\<\nabla r, \nabla \log J\>\right).$$
Since $|\nabla \log J|\le ce^{a r^2}$, for some constant $C$, which may depend on $n$,
$${\begin{split}
&{\mathbf E} \left[\sum_{i=1}^n\int_0^t \nabla d\left( { \log} k_{1-s} \right) (\tilde u_se_i, \tilde u_sh_s) \,dB_s^i\right]^2\\
&\le C\int _0^t \frac {|h_s|^2} {(1-s)^2}ds\le C\frac { |h_t|^2} {1-t}+4C\int_0^t |\dot h_s|^2 ds.
\end{split}}$$
This follows from the following standard computation, $$\int_0^t \frac {|h_s|^2} {(1-s)^2}ds
= \frac {|h_t|^2} {1-t}-\int_0^t \frac {\<h_s, 2\dot h_s\>} {(1-s)^2}ds
\le \frac { |h_t|^2} {1-t}+\frac 1 2 \int_0^t |h_s|^2 ds+2\int_0^t |\dot h_s|^2 ds.$$
This concludes the proof of the convergence of the integral. The required identity follows from the formula,
given earlier, for $\alpha_t$.
\end{proof}
Let $u_t$ be the solution to the equation
$du_t = \sum_{i=1}^n H_i(u_t) \circ dB^i_t$ with initial value $u_0 \in \pi^{-1}(x_0)$. Then $x_t := \pi (u_t)$ is a Brownian motion on $M$ starting at $x_0$ and
the integration by parts formula holds on $L^2(C_{x_0}M; \mu)$.
For any $F,G\in {\mathop {\rm Cyl}}$, and $h\in H(T_{x_0}M)$ with $h(0)=0$, $d$ is the differential on $L^2(C_{x_0}M)$ with respect to the Brownian motion measure:
\begin{equation}
\label{ibp}
{\mathbf E} [dF(u_\cdot h_\cdot) G]=-{\mathbf E} [FdG(u_\cdot h_\cdot)]
+{\mathbf E} \left[FG\int_0^1 \<\dot h_s+\frac 1 2{u_s^{-1} {{\mathop{\rm {Ric}}}}^{\sharp}(u_sh_s)}, dB_s\>\right].
\end{equation}
If $M$ is compact, see e.g. B. Driver \cite{Driver92}. This is also known to hold if the Ricci curvature is bounded from below.
The divergence of $u_\cdot h_\cdot$ is
$$\mathop{\rm div}(u_\cdot h_\cdot)=\int_0^1 \<\dot h_s+\frac 1 2{u_s^{-1} {{\mathop{\rm {Ric}}}}^{\sharp}_{u_s}(u_sh_s)}, dB_s\>.$$
The following lemma completes the proof of Theorem \ref{thm1}.
\begin{lemma}
Suppose stochastic completeness, {\bf C2-C4}, and suppose that the integration by parts formula (\ref{ibp}) holds for the Brownian motion measure. Then the conclusion of Theorem \ref{thm1} holds.
\end{lemma}
Let $h\in H^{0}$.
Our plan is to pass the integration on the path space to the pinned path space
by a Girsanov transform.
We first observe that if $F\in \mathop{\rm Dom}(d)$, adapted to ${\mathcal G}_t$ where $t<1$, then
$${\mathbf E}[ dF(\tilde u h_\cdot)]={\mathbf E} \left[dF ( u h_\cdot)\frac {k_{1-t}(x_t)}{k_1(x_0)}e^{-\int_0^t \Phi(x_s)ds}\right].$$
In fact, the formula for the probability density between the original probability measure, on ${\mathcal G}_t$, and the one for which
$B_t-\int_0^t \<u_sdB_s, \nabla {\log} k_{1-s}(x_s)\>$ is a Brownian motion, is:
\begin{equs}
M_t=\exp \left[ \sum_{i=1}^m\int_0^{t} \<\nabla {\log} k_{{1}-{s}} ( x_{s}, y_0), u_{s} e_i\> dB_{s}^i -{1\over 2} \int_0^{t} | \nabla {\log} k_{{1}-{s}}(x_{s}, y_0) |^2 d{s}\right].
\end{equs}
By an application of It\^o's formula, and identities (\ref{formulas}) in the proof of Lemma \ref{lemma1},
$$M_t=\frac {k_{1-t}(x_t, y_0)} {k_1(x_0, y_0)} \exp \left(-\int_0^t \Phi(x_s)ds\right). $$
Since the Brownian motion and the semi-classical bridge are conservative, then $(M_s, s\le t)$ is a martingale for any $t<1$.
Since $\Phi$ is bounded from below and has bounded derivative,
$e^{-\int_0^t \Phi(\tilde x_s) ds}$ can be approximated by smooth cylindrical functions in the domain of $d$.
Next we observe that
$$\nabla {k_{1-s}(\cdot, y_0)}= {2\pi (1-s)^{-\frac n 2}}e^{-\frac {r^2}{2(1-s)}}J^{-\frac 12}
\left(-\frac {r \nabla r }{1-s} +\nabla {\log} J^{-\frac 12}\right),$$
is bounded and smooth, so $\frac {k_{1-t}(x_t, y_0)}{k_1(x_0, y_0)}e^{-\int_0^t \Phi(x_s)ds}$ belongs to the domain of $d$.
Consequently, for $F, G$ measurable with respect to the canonical filtration up to time $t<1$, we apply (\ref{ibp}) to see \begin{equs}
&{\mathbf E} [GdF(\tilde u_\cdot h_\cdot] ={\mathbf E} \left [dF ( u _\cdot h_\cdot) G(\tilde x_\cdot)M_t\right]\\
&={\mathbf E} \left[(FG)( x_\cdot ) M_t\mathop{\rm div}(u_\cdot h_\cdot)\right]
-{\mathbf E} \left[(FG) ( x_\cdot ) dM_t(u_\cdot h_\cdot)\right]-{\mathbf E} [F(x_\cdot) dG( uh) M_t]\\
=&{\mathbf E} \left[F( x_\cdot )M_t\int_0^t \<\dot h_s+\frac 1 2{ u_s^{-1} {{\mathop{\rm {Ric}}}}^{\sharp}_{u_s}(h_s)}, dB_s\>\right]-{\mathbf E} [F(\tilde x_\cdot) dG(\tilde uh) ]\\
&-{\mathbf E} \left[F ( x_\cdot ) M_td\left({\log} {k_{1-t}(x_t, y_0)}-\int_0^t \Phi(x_s)ds\right)(u_\cdot h_\cdot)\right]\\
=&{\mathbf E} \left[F (\tilde x_\cdot ) \int_0^t \<\dot h_s+\frac 1 2{\tilde u_s^{-1} {{\mathop{\rm {Ric}}}}^{\sharp}_{\tilde u_s}(h_s)}, d\tilde B_s\>\right]-{\mathbf E} [F(\tilde x_\cdot) dG(\tilde uh) ]\\
&-{\mathbf E} \left[F(\tilde x_\cdot)\< \nabla {\log} k_{1-t} (\tilde x_t, y_0),\tilde u_t h_t\> -F(\tilde x_\cdot) \int_0^t d \Phi(u_s h_s)ds\right].
\end{equs}
We take $t\uparrow 1$, by (\ref{2.1}) and Lemma \ref{lemma1}, $\lim_{t\uparrow 1} \<\nabla {\log} k_{1-t} (\tilde x_t, y_0),\tilde u_t h_t\>=0$ in $L^2$,
\begin{equs}
&{\mathbf E} [GdF(\tilde u_\cdot h_\cdot] +{\mathbf E} [F(\tilde x_\cdot) dG(\tilde uh) ]\\
=&{\mathbf E} \left[(FG) (\tilde x) \int_0^1 \<\dot h_s+\frac 1 2{\tilde u_s^{-1} {{\mathop{\rm {Ric}}}}^{\sharp}(\tilde u_s h_s)}, d\tilde B_s\>\right]
+ {\mathbf E}\left[(FG)(\tilde x) \left( \int_0^1 \;d \Phi(\tilde u_s h_s)ds\right) \right].
\end{equs}
In particular, $\mathop{\rm Dom}(d^*)\supset {\mathop {\rm Cyl}}$,
$$d^*1=\int_0^1 \<\dot h_s+\frac 1 2{\tilde u_s^{-1} {{\mathop{\rm {Ric}}}}^{\sharp}(\tilde u_s h_s)}, d\tilde B_s\>+ \left( \int_0^1 \;d \Phi(\tilde u_s h_s)ds\right),$$
and $d^*$ is a closable operator. This completes the proof of the Lemma.
\subsection{Comment}
Let us consider briefly for which manifolds our assumptions on $\Phi$ hold.
Denote by $\partial r$ the radial curvature which, evaluated at $x\in M$, is the unit vector field tangent to the normal geodesic between $x$ and the pole pointing away from the pole. The Hessian of $r$ describes the change of the Riemannian tensor in the radial directions,
while the change of the volume form in the radial direction is associated to the Laplacian of $r$. More precisely we have:
$$L_{\partial r}g=2 \mathop{\rm Hess} (r), \qquad L_{\partial r} d \text{vol}=\Delta r d \text{vol}, \qquad \Delta r=\frac {n-1} r+d r (\nabla \log J),$$
indicating how the Jacobian determinant adjusts the speed of the convergence so that the semi-classical bridge behaves exactly like
the Euclidean Brownian bridge.
For the Hyperbolic space, $\Phi$ is bounded from the formula below,
$\Phi=-\frac 1 8 (n-1)^2c^2+\frac 1 8(n-1)(n-3) \left(\frac 1 {r^2}- c^2 \sinh^{-2} (rc)\right)$.
If $(N, o)$ is a model space,
its Riemannian metric in the geodesic polar coordinates takes the form $g=dr^2+f^2(r) d\theta^2$, then
on $N\setminus\{o\}$, $\mathop{\rm Hess} (r)=\frac {f'(r)}{f(r)}(g -dr\otimes dr)$.
For the hyperbolic space of constant sectional
curvature $-c^2$, the Riemannian metric is $g=dr^2+(\frac 1 c \sinh (cr))^2 d\theta^2$. Also $\mathop{\rm Hess} (r^2)=2 dr\otimes dr +2cr \coth (cr) (g -dr\otimes dr)$. Furthermore its Jacobian determinant
is $J= (\frac {\sinh (cr)}{c r})^{(n-1)}$.
For manifolds of non-constant curvature we may use the Hessian comparison theorem.
The radial curvature at a point $x\in M$ is the sectional curvature in a plane at $T_xM$ containing the radial vector field $\partial_r$. Let us recall a comparison theorem from \cite[R. E. Greene and H. Wu]{Greene-Wu}:
let $(N, o)$ be another Riemannian manifolds with a pole which we denote by $o$. Suppose that $(\gamma(t), t\in [0, b])$ is a normal geodesic in $M$
with the initial point ${y_0}$ and $(\gamma_2(t): t\in [0, b])$ a normal geodesic in $N$ from $o$.
We suppose that the radial curvature at $\gamma_2(t)$ is greater than or equal to the radial curvatures at $\gamma(t)$.
By this we mean the curvature operator ${\mathcal R}$ on $M$ and ${\mathcal R}_2$ on $N$ satisfy the relation
$\left\<{\mathcal R}(w, \dot \gamma)w, \dot \gamma\right\>\le \left\<{\mathcal R}_2(w_2, \dot \gamma_2)w_2, \dot \gamma_2\right\>$
for any unit vectors $w\in ST_{\gamma(t)}M$ and $w_2\in ST_{\gamma_2(t)}N$,
satisfying the relation $\<w, \partial_r\>=\<w_2, \partial_r\>$ where $\partial_r$ denotes the radial vector fields for both manifolds.
Then for any nondecreasing function $\alpha: {\mathbb R}_+\to {\mathbb R}$, $\mathop{\rm Hess} (\alpha\circ r_2)(\gamma_2(t))\le \mathop{\rm Hess} (\alpha\circ r)(\gamma(t))$,
where $r_2$ is the Riemannian distance function on $N$ from $o$.
\section{Conclusion}
We have proved an integration by parts formula on $L^2(C_{x_0, y_0}, \nu)$ where
$\nu$ is the probability measure induced by the semi-classical bridge.
A probability measure $\mu$ on the path space is said to satisfy the Poincar\'e inequality if there exists a constant $c$ such that
$$\int \left(F- \int F d\mu \right)^2 d\mu \le c\int \left(|\nabla F|_{{\mathcal H}}\right)^2d\mu $$
for all $F\in {\mathcal D}om(d)$ and the inner product on ${\mathcal H}$ can be defined either by stochastic parallel translation or by damped stochastic parallel translation.
{\bf Conjecture.} A Poincar\'e inequality holds for the semi-classical bridge measure on
a class of Cartan-Hadamard manifolds. Of course it is reasonable to assume growth conditions on $J$, $J^{-1}$ and suitable conditions on the range of the sectional curvature.
We remark that, for the Brownian bridge measure the question whether the Poincar\'e inequality holds is not solved satisfactorily. The spectral gap inequality is known to hold for Gaussian measure on ${\mathbb R}^n$ by L. Gross \cite{Gross}, who also
made a conjecture on its validity. The spectral gal inequality has been proven to hold on the hyperbolic space \cite{CLW},
see also \cite{Aida,Gong-Ma,Airault-Malliavin,Fang99,Elworthy-LeJan-Li-book}.
A counter example exists \cite{Eberle}, see also the more recent articles \cite{Hino, Gong-Rockner-Wu}.
\def$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'${$'$}
|
\section{Présentation de la thématique}
L'entropie, en théorie de l'information, est un concept fondamental introduit par C.~Shannon en 1948 \cite{Shannon-1948-entropy}.
A.~Kolmogorov \cite{Kolmogorov-1958-entropy,Kolmogorov-1959-entropy-erratum} l'a utilisée pour définir un invariant non moins fondamental en théorie ergodique~: l'entropie d'une transformation préservant la mesure, dont il a posé les bases entre 1958 et 1962 avec quelques proches mathématiciens, notamment Y.~Sina{\u\i} et V.~Rokhlin.
L'entropie de Shannon d'une partition dénombrable $\alpha=(A_i)_{i\in K}$ de l'espace de probabilité $(X,\mu)$ est définie par
\begin{equation}
H(\alpha)\overset{\mathrm{def}}{=}-\sum_{i\in K} \mu(A_i) \log(\mu(A_i)). \tag{\textbf{Entropie de Shannon}}
\end{equation}
\subsection{Pour le groupe \texorpdfstring{$G=\mathbf{Z}$}{Z}}
Pour un isomorphisme préservant la mesure $T\colon X\to X$, ou autrement dit une action \textbf{p.m.p.} $G\curvearrowright^{\!T}{\!}(X,\mu)$ de $G = \mathbf{Z}$, sur l'espace de probabilité standard, la définition de Kolmogorov nécessite l'existence d'une partition \textbf{génératrice}\footnote{Avec l'aide de l'action de $G$, elle permet de séparer presque tous les points de $X$ ; voir
section~\ref{sect: definitions}.} $\alpha$ d'entropie de Shannon finie et considère les entropies de Shannon des joints des itérés (les raffinements de $\alpha$ obtenus par itérations) et normalisées~:
\begin{equation}
h({G} \curvearrowright^{\!T}{\!} X,\mu)\overset{\mathrm{def}}{=}\lim\limits_{n\to\infty} \frac{1}{n} H(\vee_{i=0}^{n} T^{-i}\alpha).\tag{\textbf{Entropie de Kolmogorov}}
\end{equation}
Le point clef est bien entendu l'indépendance vis-à-vis de la partition génératrice (l'existence de la limite n'est pas difficile).
Y.~Sina{\u\i} \cite{Sinai-1959-entropy} a apporté une amélioration significative en observant que, parmi les partitions d'entropie de Shannon finie, les partitions génératrices, lorsqu'elles existent, maximisent la quantité considérée (c'est ce qu'on appelle le {\em théorème de Kolmogorov-Sina{\u\i}}).
Cela permet de se débarrasser de l'hypothèse génératrice~:
\begin{equation}
h_{\mathrm{KS}}({G} \curvearrowright^{\!T}{\!} X,\mu)\overset{\mathrm{def}}{=}\sup_{\substack{\beta \textrm{ partition } \\ H(\beta)<\infty}} \lim\limits_{n\to\infty} \frac{1}{n} H(\vee_{i=0}^{n} T^{-i}\beta).
\tag{\textbf{Entropie de Kolmogorov-Sina{\u\i}}}
\end{equation}
L'un des tout premiers succès de cette théorie aura été de résoudre le problème, qu'on attribue à J.~von~Neumann, de la conjugaison mesurée des actions par décalage de Bernoulli sur des espaces de base à $2$ points $G\curvearrowright \{1,2\}^G$, respectivement $3$ points $G\curvearrowright \{1,2,3\}^G$ (avec les mesures canoniques). Leur entropie valant $\log(2)$, resp. $\log(3)$, ils ne peuvent pas être isomorphes.
Et on touche de suite à deux propriétés cruciales de l'entropie de Kolmogorov~:
\begin{itemize}
\item [\textbf{(a)}] l'entropie est un \textbf{invariant de conjugaison mesurée},
\item [\textbf{(b)}] l'entropie des \textbf{décalages de Bernoulli} est égale à l'entropie de Shannon de leur base~:
$h_{\mathrm{KS}}({G}\curvearrowright (L,\lambda)^{{G}})= H(\lambda) \overset{\mathrm{def}}{=} \sum_{l\in L} -\lambda(l) \log\lambda(l).$
En particulier, une conjugaison mesurée entraîne que les bases ont même entropie~:
\begin{equation}
G\curvearrowright(K,\kappa)^{G}\simeq G\curvearrowright(L,\lambda)^{G} \Rightarrow H(\kappa)=H(\lambda). \tag{\textbf{Théorème de Kolmogorov}}\end{equation}
\end{itemize}
Ce résultat reste vrai, que la base (et son entropie de Shannon) soit finie ou non\footnote{L'entropie de Shannon d'un espace diffus tel que $[0,1]$ muni la mesure Lebesgue vaut $+\infty$.}. Voir aussi section~\ref{sect: ent Rokhlin} pour l'existence de partitions génératrices finies.
Après Kolmogorov, le problème est devenu celui d'une réciproque. D.~Ornstein a développé de puissantes méthodes qui lui ont permis d'identifier une forme d'ubiquité des actions Bernoulli (on peut consulter l'article de survol \cite{Ornstein-2013-survey-Bernoulli}). Et la réciproque en a découlé \cite{Ornstein-1970-entrop-shift-isom}~:
\begin{equation}
H(\kappa)=H(\lambda) \Rightarrow G\curvearrowright(K,\kappa)^{G}\simeq G\curvearrowright(L,\lambda)^{G}. \tag{\textbf{Théorème d'Ornstein}}
\end{equation}
Ainsi, l'entropie est un invariant complet de conjugaison mesurée parmi les décalages de Bernoulli.
Une propriété saute aux yeux avec le théorème de Kolmogorov-Sina{\u\i} (elle lui est d'ailleurs essentiellement équivalente), c'est le comportement de l'entropie sous \textbf{facteur}\footnote{Les facteurs jouent un grand rôle en théorie ergodique ; ils sont liés aux partitions non génératrices (voir section~\ref{sect: definitions}).}, \emph{i.e.} une application mesurable $\pi\colon (X,\mu)\to (Y,\nu)$ qui pousse $\mu$ sur $\nu$ (essentiellement surjective, donc -- l'image est de mesure pleine) et qui est équivariante pour des actions ${G}\curvearrowright(X,\mu)$ et ${G}\curvearrowright(Y,\nu)$. C'est une troisième propriété fondamentale de l'entropie :
\begin{itemize}
\item[\textbf{(c)}] l'entropie décroît sous \textbf{facteur}~: $h_{\mathrm{KS}}({G}\curvearrowright X,\mu)\geqslant h_{\mathrm{KS}}({G}\curvearrowright Y,\nu)$.
\end{itemize}
En effet, les partitions de $Y$ et les calculs d'entropie de joints se remontent dans $X$.
Signalons une forme de réciproque due à Y.~Sina{\u\i}.
\\
\emph{Si $\mathbf{Z} \curvearrowright^{\! T}\! (X,\mu)$ est une action p.m.p. ergodique, alors elle factorise sur tout décalage de Bernoulli $\mathbf{Z}\curvearrowright (K^{\mathbf{Z}}, \kappa^{\mathbf{Z}})$ à base dénombrable \og entropie compatible \fg, \emph{i.e.} qui vérifie $h_{\mathrm{KS}}(\mathbf{Z} \curvearrowright^{\! T}\! X,\mu)\geqslant h_{\mathrm{KS}}(\mathbf{Z} \curvearrowright K^{\mathbf{Z}}, \kappa^{\mathbf{Z}})$} \cite{Sinai-1962-weak-isom}.
\label{th: Sinai fact sur Bernoulli}
Les applications de l'entropie d'une transformation se sont diffusées dans une grande partie des systèmes dynamiques, et les résultats sont quasiment indénom\-brables.
Nous reviendrons plus loin sur certains d'entre eux. Pour un survol historique détaillé de l'entropie de Kolmogorov, on pourra consulter l'article \cite{Katok-2007-50-yrs-entropy}.
\subsection{Pour les groupes $G$ moyennables}
La théorie s'est également développée pour des actions de groupes plus généraux que le groupe $\mathbf{Z}$. Et on s'est vite aperçu que les choses se passaient bien pour les groupes commutatifs, puis pour le dire rapidement, la théorie s'est étendue à tous les groupes dénombrables moyennables (voir section~\ref{sect:groupes sofiques}), avec le considérable travail fondateur de D.~Ornstein et B.~Weiss \cite{OW87}.
Une particularité de la définition de l'entropie de Kolmogorov pour $G=\mathbf{Z}$ est l'utilisation des intervalles $\{0, 1, \cdots, n\}$ qui ont un petit bord (formé de deux points $0$ et $n$).
La définition de l'entropie pour les actions p.m.p. des groupes moyennables doit maintenant faire appel aux \textbf{suites de F{\o}lner}\footnote{Des parties finies à constante isopérimétrique tendant vers $0$, caractéristiques de la moyennabilité de $G$.}
dans le groupe $G$, mais l'essentiel de la théorie est conservé, et principalement les propriétés (a) (conjugaison mesurée), (b) (décalages de Bernoulli) et (c) (décroissance sous facteur) mises en évidence ci-dessus. Le théorème d'Ornstein reste également valide et l'entropie mesurée est un invariant complet de conjugaison mesurée parmi les décalages de Bernoulli.
\subsection{Pour les groupes $G$ non moyennables}
L'extension de la théorie entropique à des groupes non moyennables est restée une quête inaccessible jusqu'à l'intervention en 2008 de L.~Bowen et de son entropie sofique.
Une quête d'autant plus frustrante que A.~Stepin \cite{Stepin-1975} a montré que le théorème d'Ornstein, la partie réciproque donc, s'étendait facilement des sous-groupes au groupe ambiant. Ainsi,
pour tous les groupes contenant un groupe moyennable infini\footnote{On peut signaler que L.~Bowen \cite{Bowen-2012-almost-Ornstein} a étendu cette propriété
\og si $H(\nu_1)=H(\nu_2)$, alors $G\curvearrowright (K_1^G,\nu_1^G)\simeq G\curvearrowright (K_2^G,\nu_2^G)$\fg\ à tous les groupes infinis dénombrables, à condition qu'aucun des deux espaces de base ne soit constitué de seulement deux atomes.}~:
{\em si les entropies de Shannon des espaces de base sont égales, alors les décalages de Bernoulli associés sont conjugués}, sans qu'on ait été capable de dire si
réciproquement, comme dans le théorème de Kolmogorov, la conjugaison entraînait l'égalité des entropies de Shannon des bases.
\subsection{Le blocage psychologique de l'exemple d'Ornstein-Weiss}
\label{sect: appl Ornstein-Weiss}
Dans leur article \cite{OW87}, D.~Ornstein et B.~Weiss ont exhibé un exemple qui semblait rendre vaine toute tentative au-delà du moyennable\footnote{Rappelons que les prototypes de groupes non moyennables sont les groupes contenant le groupe libre $\mathbf{L}_2$.}.
Soit $\mathbf{L}_2=\langle a, b\rangle$ le groupe libre à deux générateurs et soit $\mathbb{K}$ un groupe fini abélien, par exemple $\mathbb{K}=\mathbf{Z}/2\mathbf{Z}$.
L'application
\begin{equation}
\label{eq: appl OW}
\Theta \colon \left(\begin{array}{cclcc}
\mathbb{K}^{\mathbf{L}_2}&\to &\mathbb{K}^{\mathbf{L}_2}\times \mathbb{K}^{\mathbf{L}_2}\simeq (\mathbb{K}\times \mathbb{K})^{\mathbf{L}_2}\\
\omega &\mapsto &
(\omega(g a)-\omega(g), \omega(g b)-\omega(g) )_{g\in \mathbf{L}_2}
\end{array}\right)
\end{equation}
est surjective et définit une factorisation du décalage de Bernoulli de base $\mathbb{K}$ sur celui de base $\mathbb{K}\times \mathbb{K}$, où les bases sont munies des mesures d'équiprobabilité.
Toute bonne théorie devant leur donner entropie $\log \vert \mathbb{K}\vert $, resp. $\log \vert \mathbb{K}\times \mathbb{K} \vert=2 \log \vert \mathbb{K}\vert $, on aurait affaire à {\em un facteur qui augmente l'entropie} !
Et cette application est loin d'être une pathologie, c'est un homomorphisme continu $\mathbf{L}_2$-équivariant de groupes compacts, qu'on équipe de leurs mesures de Haar. Son noyau est fini~: ce sont les fonctions constantes $\mathbf{L}_2\to \mathbb{K}$.
On peut aussi l'interpréter comme l'application cobord qui va des cochaînes de dimension $0$ à coefficients dans $\mathbb{K}$ dans celles de dimension $1$, pour l'arbre de Cayley $\mathcal{T}$ du groupe libre.
Quelques détails supplémentaires sont fournis en section~\ref{sect:plus sur l'ex de OW}.
Cet exemple a conduit D.~Ornstein et B.~Weiss \cite{OW87} à demander si tous les décalages de Bernoulli sur un groupe non moyennable étaient isomorphes\footnote{Auquel cas, l'entropie de Shannon de la base n'aurait pas été un invariant de conjugaison.}.
Et pour enfoncer le clou\footnote{Le théorème \og entropie compatible \fg\ de Sina{\u\i} (section \ref{th: Sinai fact sur Bernoulli}) qui montrait l'isomorphisme {\em faible} des {$\mathbf{Z}$-décalages} de même entropie, fut considéré comme un jalon important vers le théorème d'Ornstein d'isomorphisme {\em fort} de ces décalages.},
L. Bowen montre que~:
\begin{theo}[Bowen {\cite[Th. 1.3]{Bowen-2011=weak-isom-Bernoulli}}]\label{th:Bowen isom faibles des Bernoulli}
Pour tout groupe $G$ contenant $\mathbf{L}_2$, tous les décalages de Bernoulli factorisent les uns sur les autres (on dit qu'ils sont faiblement isomorphes).
\end{theo}
C'est une catastrophe ?!
\`A moins de renoncer à la propriété de décroissance... et de changer de point de vue.
\subsection{Point de vue externe, modèles finis}
L.~Bowen dans une série de résultats retentissants annoncés à partir de 2008 va faire rebondir complètement le sujet.
Le premier de la série \cite{Bowen-2010-f-invariant} (annoncé en 2008, publié en 2010) traite spécifiquement le cas où $G$ est un groupe libre. Il introduit une quantité qu'il appelle le $f$-invariant\footnote{Le symbole $f$ dans $f$-invariant est utilisé pour évoquer le qualificatif {\em free} de {\em free group}.} dont il montre qu'il est invariant de conjugaison mesurée et qu'il prend la valeur $H(\kappa)$ sur un décalage de Bernoulli $G\curvearrowright (K,\kappa)^G$ dès que $K$ est fini\footnote{Il parvient aussi à traiter des $K$ infinis d'entropie de Shannon finie.}.
Dans cet article et dans le suivant \cite{Bowen-2010-mes-conj-inv-sofic}, il adopte un \og point de vue externe\fg, comme dit D.~Kerr. Il cherche à modéliser sur des ensembles finis $D$ la dynamique de l'action $G\curvearrowright^{\!T}{\!} (X,\mu)$.
Le point de vue adopté est plus proche de celui de l'interprétation statistique de l'entropie de Boltzmann. L'entropie de Shannon $H(\alpha)$ d'une partition $\alpha=(A_k)_{k\in K}$ de $X$ (sans considérer de dynamique)
peut s'obtenir de la façon suivante.
On considère un ensemble fini $D$ (qu'on pense très grand) et on regarde toutes les partitions sur $D$ qui imitent bien $\alpha$ (au sens où les pièces ont les bonnes mesures à $\epsilon$ près) ; on estime leur nombre, puis on prend le taux de croissance exponentielle de ce nombre en la taille de $D$ qu'on fait tendre vers l'infini, puis on serre les $\epsilon$. Plus précisément, c'est une forme simple d'un principe de grande déviation~:
\begin{prop}[Boltzmann, Sanov {\cite{Sanov-1957-large-dev}}]
\label{prop:Boltzmann-Sanov}
\begin{equation*}
H(\alpha)=\lim\limits_{\epsilon\to 0} \lim\limits_{\vert D\vert \to \infty} \frac{1}{\vert D\vert }\log \left\vert
\left\{\textrm{partitions } (V_k)_{k\in K} \textrm{ de } D \ \Big\vert \ \forall k\in K,\ \left\vert \frac{\vert V_k \vert}{\vert D\vert} -\mu(A_k)\right\vert <\epsilon\right\}
\right\vert.
\end{equation*}
\end{prop}
C'est ce point de vue, étendu à la situation où un groupe agit, qui conduit à la notion d'entropie sofique.
On revient en détail sur cet énoncé (en section~\ref{sect: preuve du lem Boltzmann-Sanov})
et on en donne une preuve \og probabiliste\fg\ qui reviendra en écho lors des estimations pour les décalages de Bernoulli (section~\ref{sect: ent sofique Bernoulli}).
\subsection{Introduction de la dynamique, soficité}
Si on cherche maintenant à modéliser la dynamique de $G\curvearrowright^{\!T}{\!} (X,\mu)$ sur des ensembles finis $D$ qui auront vocation à devenir grands, il faudra commencer par concéder que le groupe $G$ possède des actions sur de tels ensembles, c'est-à-dire des homomorphismes \og pas trop triviaux\fg\ dans des groupes symétriques $\mathrm{Sym}(D)$. Un bon cadre est celui des groupes résiduellement finis (voir section~\ref{sect:groupes sofiques}). Mais puisque les erreurs sous-exponentielles seront tolérables\footnote{Voire, en vérité, des erreurs exponentielles mais de taux de croissance arbitrairement petit.}, on sera en mesure d'admettre de petites erreurs dans les homomorphismes, c'est-à-dire considérer des groupes sofiques.
Les \textbf{groupes sofiques} (introduits par M.~Gromov) sont des groupes qui possèdent de bonnes propriétés d'approximation en termes de permutations sur des ensembles finis ; et un peu plus précisément, des suites de {\em presque-homomorphismes-sans-point-fixe} $\Sigma=(\sigma_n \colon G\to \mathrm{Sym}(D_n))_{n}$ dans des groupes de permutations (voir section~\ref{sect:groupes sofiques}).
\begin{defi}[Groupe sofique, approximation sofique] \label{defn: gp approx sofic}
Un groupe dénombrable $G$ est \textbf{sofique} s'il admet une suite
d'ensembles finis $D_n$ et d'applications\footnote{Insistons : pas nécessairement des homomorphismes !}
${\sigma_n \colon G \rightarrow \mathrm{Sym}(D_n)}$ telles que $\sigma_n(1_G) = 1_{\mathrm{Sym}(D_n)}$ et
\begin{enumerate}
\item[\rm (i)] (\textbf{presque action}) \hfil
$\frac{1}{\vert D_n\vert} \vert \{ z \in D_n \ \vert \ \sigma_n(g) \circ \sigma_n(h)(z) = \sigma_n(g h)(z)\} \vert \underset{n\to \infty}{\longrightarrow} 1$, \hfil $\forall g, h \in G$
\item[\rm (ii)] (\textbf{presque libre}) \hfil
$\frac{1}{\vert D_n\vert} \vert\{ z \in D_n \ \vert \ \sigma_n(g)(z) \neq z\}\vert \underset{n\to \infty}{\longrightarrow} 1$, \hfil $\forall g\in G\setminus\{1_G\}$.
\end{enumerate}
Une suite $\Sigma=(\sigma_n \colon G \rightarrow \mathrm{Sym}(D_n))_{n}$ comme ci-dessus, où le cardinal $d_n\overset{\mathrm{def}}{=}\vert D_n\vert$ tend vers l'infini\footnote{Cette condition, automatiquement satisfaite si le groupe $G$ est infini, permet d'éviter certaines pathologies pour les groupes finis.}
est appelée une \textbf{approximation sofique} du groupe $G$.
\end{defi}
Les groupes résiduellement finis rentrent dans cette classe grâce à leurs actions par multiplication sur leurs quotients finis. On peut déjà goûter tout le sel de la notion d'entropie sofique en se restreignant aux actions de ces derniers et on peut lire ce texte en se concentrant sur eux.
L'extension aux groupes sofiques n'est pas que pur désir de généralisation maximale. On aurait deux types de théories de l'entropie avec une intersection non triviale, l'une pour les groupes moyennables, l'autre pour les groupes résiduellement finis. Or, les groupes moyennables sont également sofiques et, pour eux, les invariants obtenus dans le cadre de cette théorie de l'entropie sofique, quoique définis de manière bien différente, se trouvent coïncider avec ceux de la théorie classique de l'entropie \cite{Bowen-2012-sofic-ent-amenab, Kerr-Li-2013-sofic-amenabl-dyn-entrop}.
On disposera ainsi d'un emboîtement de généralisations successives de l'entropie mesurée.
On développe un peu cette notion de soficité et on la relie à la moyennabilité et aux groupes résiduellement finis dans la section~\ref{sect:groupes sofiques}. On signale (voir Théorème~\ref{th: Kerr-Li entrop Gottschalk conject}) la preuve entropique due à D.~Kerr et H.~Li de la conjecture de surjonctivité de Gottschalk pour les groupes sofiques.
\subsection{Entropie sofique mesurée}
Considérons donc un groupe sofique $G$ et une de ses approximations sofiques $\Sigma=(\sigma_n \colon G\to \mathrm{Sym}(D_n))_{n}$ (définition~\ref{defn: gp approx sofic}).
Les ensembles finis $D_n$ de $\Sigma$, équipés de leur {\em presque action}\footnote{Comme dit précédemment, on peut faire semblant de croire que les $\sigma_n$ sont de vrais homomorphismes.} de $G$, sont munis chacun de leur mesure de probabilité uniforme $\mathbf{u}_{n}$. Ils sont envisagés comme des espaces modèles pour les diverses dynamiques produites par les actions de $G$.
Considérons une action p.m.p. $G\curvearrowright^{\!T}{\!} (X,\mu)$.
Soit $\alpha \colon X\to K$ une $K$-partition mesurée finie de $X$ (les pièces sont les $A_k=\alpha^{-1}(k)$, indexées, ou colorées si on veut, par l'ensemble fini $K$). L'ensemble $K^{D_n}=\{a \colon D_n\to K\}$ est alors simplement l'ensemble de toutes les $K$-partitions de $D_n$.
Si $F\subset G$ est une partie finie contenant $\mathrm{id}_G$, alors par itérations, elle définit la partition $F$-raffinée $\alpha^{\vee F}$ et de manière analogue, chaque partition $a \colon D_n\to K$ fournit une partition $F$-raffinée\footnote{Observons que la définition de $a^{\vee F}$ ne nécessite pas que $\sigma_n$ soit un homomorphisme. Une approximation sofique fera aussi bien l'affaire.} $a^{\vee F}$,
\begin{equation}
\alpha^{\vee F} \colon \left(\begin{array}{ccl}
X &\to &K^F
\\
x &\mapsto & (\alpha(T(f)\cdot x))_{f\in F}
\end{array}\right)
\ \textrm{ et } \
a^{\vee F} \colon \left(\begin{array}{ccl}
D_n&\to &K^F
\\
{v} &\mapsto & (a(\sigma_n(f)\cdot {v}))_{f\in F}
\end{array}\right).
\end{equation}
Il s'agit alors de comparer les mesures des pièces de ces partitions dans $X$ et dans $D_n$.
Ou autrement dit de comparer les mesures poussées en avant $\alpha^{\vee F}_{*} \mu$ et $a^{\vee F}_{*} \mathbf{u}_n$ sur $K^F$.
On note
\begin{equation}\label{eq:def M mu-intro}
\mathcal{M}_\mu(\alpha, F, \epsilon, \sigma_n)\overset{\mathrm{def}}{=}\left\{a \in K^{D_n}\ \big\vert \ \Vert \alpha^{\vee F}_{*} \mu-a^{\vee F}_{*} \mathbf{u}_n\Vert_{1} < \epsilon\right\}.
\end{equation}
C'est l'ensemble de toutes les $K$-partitions de $D_n$ qui, une fois itérées par $F\subset G$, fournissent des pièces de mesure proche de celle des pièces obtenues de la même façon pour $\alpha$ sur $X$, et la proximité est mesurée dans la norme $\ell^1$. Ceux qui préfèrent la version ensembliste pourront se reporter à la section~\ref{sect: modeles finis de la dynamique}.
L.~Bowen considère alors le cardinal de cet ensemble et son taux de croissance exponentielle en le cardinal de $D_n$, puis il minimise selon les paramètres $F$ et $\epsilon$.
Le résultat frappant est qu'on obtient une quantité qui est indépendante du choix d'une partition génératrice finie (si elle existe !).
\begin{theodef}[Entropie sofique mesurée, Bowen {\cite{Bowen-2010-mes-conj-inv-sofic}}]
\label{th-def: ent sof mes Bowen}
Soit $\Sigma$ une approximation sofique du groupe sofique $G$. Soit $G\curvearrowright^{\!T}{\!} (X,\mu)$ une action p.m.p. qui admet une partition génératrice finie $\alpha$. Alors, la quantité ci-dessous est indépendante du choix de la partition génératrice finie.
Cette valeur commune est appelée l'\textbf{entropie sofique mesurée} de l'action relativement à $\Sigma$, et on la note~:
\begin{equation}
h_{\mathrm{mes}}^{\Sigma}(G\curvearrowright^{\!T}{\!} X,\mu)\overset{\mathrm{def}}{=}\inf_{\epsilon > 0} \inf_{\substack{F \subset G\\F \text{ fini}}} \limsup_{n \rightarrow \infty} \frac{\log \left| \mathcal{M}_\mu(\alpha, F, \epsilon, \sigma_n) \right |}{\vert D_n \vert}.
\tag{\textbf{Entropie sofique mesurée}}\label{ent sofic Bowen}
\end{equation}
\end{theodef}
L'entropie sofique prend ses valeurs dans $[0, +\infty]\cup \{-\infty\}$. La valeur $-\infty$ correspond aux situations où l'ensemble $\mathcal{M}_\mu(\alpha, F, \epsilon, \sigma_n)$ est vide\footnote{Via la convention $\log \vert \emptyset\vert=-\infty$.} à partir d'un certain rang $n$ pour $F$ grand et $\epsilon$ petit. La valeur $+\infty$ apparaîtra lorsqu'on se sera débarrassé de l'hypothèse génératrice.
On peut en effet noter que si $\alpha \colon X\to K$ est une partition génératrice finie, alors $\mathcal{M}_\mu(\alpha, F, \epsilon, \sigma_n)\subset K^{D_n}$ nous donne $h_{\mathrm{mes}}^{\Sigma}(G\curvearrowright X,\mu) \leqslant \frac {\log\vert K^{D_n} \vert }{\vert D_n \vert}= \log\vert K \vert$ ; c'est un premier pas vers l'entropie de Rokhlin (section~\ref{sect: ent Rokhlin}).
La dépendance en l'approximation sofique $\Sigma$ est l'objet de grandes spéculations.
On dispose d'exemples où elle en dépend. Voir la section~\ref{sect: depend-approx-sofic} pour des informations plus détaillées.
Mais en l'état actuel des connaissances, tous ces exemples reposent de façon plus ou moins directe sur l'utilisation de la propriété $(\tau)$, une variante de la propriété (T) de Kazhdan, et sur le fait que pour certaines actions de certains groupes, on peut trouver des approximations sofiques $\Sigma_1$ qui donnent lieu à des $\mathcal{M}_\mu(\alpha, F, \epsilon, \sigma_n)$ qui auront tendance à être vides (et donneront $h_{\mathrm{mes}}^{\Sigma_1}(G\curvearrowright X,\mu)=-\infty$), tandis que pour d'autres approximations $\Sigma_2$, ils ne seront pas vides.
On ne dispose pas d'exemple d'action pour laquelle deux approximations sofiques donneraient des valeurs réelles distinctes.
Il découle de cette dépendance que la $ \limsup_{n \rightarrow \infty}$ qui intervient dans la définition de l'entropie sofique ne peut pas être remplacée par une limite\footnote{
Penser à une approximation sofique qui \emph{piocherait} alternativement dans $\Sigma_1$ ou $\Sigma_2$.} et ce détail complique très sérieusement la vie.
L.~Bowen \cite[Rem. 1] {Bowen-2010-mes-conj-inv-sofic} signale que sa théorie peut aussi se développer de manière satisfaisante en remplaçant cette $\limsup$ par une $\liminf$ voire par une ultra-limite\footnote{Een introduisant un ultra-filtre comme paramètre supplémentaire. Et ce point de vue serait assez cohérent avec l'approche des groupes sofiques par les ultra-produits développée par G.~Elek et E.~Szab{\'o} dans \cite{2005=Elek-Szabo=hyperlinearity}.}. Une variante sans conséquence fondamentale consisterait à remplacer la norme $\ell^1$ par d'autres normes pour les mesures sur $K^{F}$.
L'entropie sofique vérifie, elle aussi, les propriétés (a) et (b).
\\ (a) C'est un invariant de conjugaison mesurée.
\\ (b) L'entropie sofique des décalages est bien celle de leur base~:
\begin{theo}[Bowen {\cite{Bowen-2010-mes-conj-inv-sofic}}]
Si $G\curvearrowright (K^{G}, \nu^{\otimes G})=(K,\nu)^{G}$ est une action par décalage de Bernoulli d'un groupe sofique, alors pour toute approximation sofique $\Sigma$ de $G$, on a~:
\[h_{\mathrm{mes}}^{\Sigma}(G\curvearrowright K^{G}, \nu^{\otimes G}) =H(\nu).\tag{\textbf{Entropie des Bernoulli}}\]
\end{theo}
\`A la différence de l'entropie de Kolmogorov pour les groupes moyennables, la dé\-mons\-tration de cet énoncé n'est pas immédiate. Il est dû à L.~Bowen \cite{Bowen-2010-mes-conj-inv-sofic} lorsque l'entropie de la base est finie et à D.~Kerr et H.~Li~\cite{Kerr-Li-2011-Bernoulli-infinite-entropy} en général (en anticipant un peu sur la suite de ce texte). On trouvera des indications de preuve en section~\ref{sect: ent sofique Bernoulli}.
En combinant cela avec les résultats d'Ornstein, Stepin et Ornstein-Weiss on obtient~:
\begin{theo}
Pour tout groupe sofique $G$ contenant un sous-groupe moyennable infini\footnote{Ou bien
$G$ sofique et aucun des espaces de base n'est constitué de seulement deux atomes \cite{Bowen-2012-almost-Ornstein}.}, l'entropie de Shannon de la base est un invariant complet de conjugaison mesurée~:
\begin{equation}G\curvearrowright(K,\kappa)^{G}\simeq G\curvearrowright(L,\lambda)^{G} \Longleftrightarrow H(\kappa)=H(\lambda). \tag{\textbf{Invariant complet}}\end{equation}
\end{theo}
On ignore à ce jour si c'est vrai pour tout groupe dénombrable.
En revanche, la propriété (c) est maintenant mise en défaut notamment par l'application d'Ornstein-Weiss (section~\ref{sect: appl Ornstein-Weiss}).
{\em L'entropie sofique peut croître sous facteurs.}
\subsection{Entropie sofique sans partition génératrice}
\label{subsect: Ent sof sans part gen}
Et l'histoire semble se répéter.
Nous voici avec une notion d'entropie bien définie lorsqu'on dispose d'une partition génératrice finie.
On aimerait bien se débarrasser de cette hypothèse, notamment dans la perspective d'un principe variationnel (voir section~\ref{sect: ent topologique Kerr-Li}).
Ce souhait sera réalisé par les travaux de D.~Kerr et H.~Li \cite{Kerr-Li-2011-Variationnal-principle, Kerr-Li-2013-sofic-amenabl-dyn-entrop}. Signalons que L.~Bowen, dans son article fondateur \cite{Bowen-2010-mes-conj-inv-sofic}, étend par un procédé limite une partie de ses résultats aux actions qui possèdent une partition génératrice dénombrable d'entropie de Shannon finie.
D.~Kerr et H.~Li développent une théorie de l'entropie sofique mesurée dans un cadre d'analyse fonctionnelle \cite{Kerr-Li-2011-Variationnal-principle} (où l'hypothèse de génération finie perdra de sa pertinence) et démontrent qu'elle est équivalente à celle de L.~Bowen en présence d'une partition génératrice finie.
Ils introduisent une nouvelle notion, celle d'\textbf{entropie sofique topologique} et démontrent un \textbf{principe variationnel} (voir section~\ref{sect: ent topologique Kerr-Li}).
Ils parviennent ensuite à réintégrer ces notions dans un cadre analogue à celui de L.~Bowen de partitions finies et de dénombrements dans \cite{Kerr-2013-Sofic-meas-ent-via-finite-partitions} pour l'entropie mesurée et dans \cite{Kerr-Li-2013-ent-top-version-combinat} pour l'entropie topologique. C'est plutôt sur ces dernières versions qu'on va se concentrer.
Partitions non génératrices et facteurs étant intimement liés (voir section~\ref{subsubsect:partitions generatrices}), l'exemple d'Ornstein-Weiss, ainsi que le théorème~\ref{th:Bowen isom faibles des Bernoulli} de factorisation les uns sur les autres des décalages de Bernoulli des groupes contenant $\mathbf{L}_2$ \cite[Th. 1.3]{Bowen-2011=weak-isom-Bernoulli}, indiquent qu'une définition à la Sina{\u\i}, en prenant un \emph{supremum} sur toutes les partitions finies de l'entropie sofique introduite par Bowen, conduirait certainement à un invariant peu intéressant (qui donnerait par exemple la valeur $+\infty$ pour tous les décalages des groupes contenant $\mathbf{L}_2$).
L'idée est alors de définir l'entropie sofique d'une partition finie $\beta \colon X \rightarrow L$, à valeurs dans l'ensemble fini $L$, en la confrontant à toutes les partitions finies mesurables $\alpha \colon X \rightarrow K$ qui sont \textbf{plus fines}\footnote{Les pièces de $\beta$ sont obtenues en regroupant entre elles des pièces de $\alpha$.} qu'elle (ce qu'on note $\alpha \geqslant \beta$), c'est-à-dire telles que $\beta = \Theta_{\beta,\alpha} \circ \alpha$ pour une certaine application $\Theta_{\beta,\alpha} \colon K \rightarrow L$ (de fusion des pièces).
\begin{equation}
\left.\begin{array}{rcl} X & \overset{\alpha}{\longrightarrow} & K \\ & {\searrow}{\beta} & \downarrow \Theta_{\beta,\alpha} \\ & & L\end{array}\right.
\left.\begin{array}{rcl} D_n & \overset{a}{\longrightarrow} & K \\ & {\searrow}{b} & \downarrow \Theta_{\beta,\alpha} \\ & & L\end{array}\right.
\tag{\textbf{Fusion des pièces}}
\end{equation}
Les partitions de $D_n$ qui sont de bons $K$-modèles finis de $\alpha$ pour $(F,\epsilon)$, c'est-à-dire les éléments de
$\mathcal{M}_\mu(\alpha, F, \epsilon, \sigma_n)=\left\{a \in K^{D_n}\ \big\vert \ \Vert \alpha^{\vee F}_{*} \mu-a^{\vee F}_{*} \mathbf{u}_n\Vert_{1} < \epsilon\right\}$,
fournissent également de bons $L$-modèles pour la partition plus grossière $\beta$
via
\begin{equation}
\left(\begin{array}{ccc}\mathcal{M}_\mu(\alpha, F, \epsilon, \sigma_n) & \longrightarrow & \mathcal{M}_\mu(\beta, F, \epsilon, \sigma_n) \\ a & \mapsto & b=\Theta_{\beta,\alpha}\circ a\end{array}\right).
\end{equation}
Et ce sont ces modèles images qu'on dénombre. En d'autres termes, on considère le nombre de $L$-partitions de $D_n$ dans $\mathcal{M}_\mu(\beta, F, \epsilon, \sigma_n)$ qui sont susceptibles de se raffiner en des $K$-partitions de $D_n$ dans $\mathcal{M}_\mu(\alpha, F, \epsilon, \sigma_n)$, et on en étudie le taux de croissance exponentielle en la taille de $D_n$.
Cela conduit à une définition générale de l'entropie mesurée, sans hypothèse d'existence d'une partition génératrice finie.
Cette définition est due à D.~Kerr \cite{Kerr-2013-Sofic-meas-ent-via-finite-partitions}. Elle est équivalente aux définitions de Kerr--Li \cite{Kerr-Li-2011-Variationnal-principle, Kerr-Li-2013-sofic-amenabl-dyn-entrop}, et toutes généralisent la définition due à L.~Bowen \cite{Bowen-2010-mes-conj-inv-sofic}.
\begin{defi}[Entropie sofique]
L'\textbf{entropie sofique mesurée} de $G \curvearrowright (X, \mu)$ relativement à $\Sigma$ est définie comme
\begin{equation}
h_{\mathrm{mes}}^{\Sigma}(G\curvearrowright X,\mu) \overset{\mathrm{def}}{=} \sup_\beta \ \ \inf_{\alpha \geqslant \beta} \ \ \inf_{\epsilon > 0} \ \ \inf_{\substack{F \subset G\\F \text{ fini}}}\ \ \limsup_{n \rightarrow \infty} \frac{ \log \left| \Theta_{\beta,\alpha} \circ \mathcal{M}_\mu(\alpha, F, \epsilon, \sigma_n) \right|}{\vert D_n \vert},\tag{\textbf{Entropie sofique mesurée}}
\end{equation}
où $\alpha$ et $\beta$ parcourent les partitions mesurables finies de $X$.
\end{defi}
Force est d'admettre que cette formule est assez épouvantable. Mais en présence de partitions génératrices, on a de sérieuses simplifications. On donne quelques éléments supplémentaires en section~\ref{sect: ent sof. Kerr-Li th. generateur}.
Lorsque le groupe $G$ est moyennable infini, bien que ces définitions soient externes, utilisant des modèles finis, il se trouve qu'elles vont néanmoins coïncider avec les versions classiques de l'entropie \cite{Bowen-2012-sofic-ent-amenab, Kerr-Li-2013-sofic-amenabl-dyn-entrop}~:
\begin{theo}[Entropie pour les groupes moyennables]
\label{th: entropie sof. gp moyennable=KS}
Soit $G$ un groupe \textbf{moyennable} infini.
Pour toute action p.m.p. $G\curvearrowright^{\!T}{\!}(X, \mu)$, l'entropie mesurée classique de Kolmogorov-Sina{\u\i} et Ornstein-Weiss coïncide avec l'entropie mesurée sofique relativement à n'importe quelle approximation sofique $\Sigma$~:
\begin{equation}
h_{\mathrm{mes}}^{\Sigma}(G\curvearrowright X,\mu)=h_{\mathrm{KS}} (G\curvearrowright X,\mu).\tag{\textbf{Entropie des moyennables}}
\end{equation}
\end{theo}
\subsection{Entropie de Rokhlin}
\label{sect: ent Rokhlin}
L'origine de la notion d'\textbf{entropie de Rokhlin} est à chercher dans la majoration\footnote{Immédiate vu les propriétés de sous-additivité de la fonction $H$.} $h_{\mathrm{KS}}(\mathbf{Z}\curvearrowright X,\mu)\leqslant H(\alpha)$, pour toute partition génératrice $\alpha$, et dans l'optimisation qu'en constitue le théorème des {\em générateurs de Rokhlin}\footnote{Concernant la manière d'orthographier son nom dans l'alphabet latin, observons que la plupart de ses articles en anglais sont publiés sous le nom de Rohlin.
Il est cependant très plausible qu'il ait fini par préférer le nom de Rokhlin, sous lequel il a signé quelques-uns de ses derniers travaux, sous lequel il est référencé dans MathSciNet, dans le {\em Mathematics Genealogy Project} ou sur Wikipedia. C'est également cette orthographe qui est utilisée pour la traduction de sa notice nécrologique, dans les articles historiques rédigés par A.~Vershik, et par son propre fils Vladimir Rokhlin Jr, professeur en informatique à Yale.}
(\cite{Rohlin-1963-generators} ou les notes~\cite{Rohlin-1967-lectures-entropy})
~:
\begin{theo}[des générateurs de Rokhlin]\label{th:gen Rokhlin}
Si $G\curvearrowright(X,\mu)$ est une action p.m.p. libre ergodique de $G=\mathbf{Z}$, alors son entropie de Kolmogorov-Sina{\u\i} est l'infimum des entropies de Shannon de ses partitions génératrices~:
\begin{equation}
h_{\mathrm{KS}}(G\curvearrowright X,\mu)=\inf\{H(\alpha): \textrm{ $\alpha$ partition génératrice dénombrable}\}.
\tag{\textbf{Rokhlin}}
\end{equation}
\end{theo}
En particulier, si l'entropie est finie, alors il doit exister une partition génératrice dénom\-brable d'entropie finie.
En fait, le théorème du {\em générateur fini} de Krieger \cite{Krieger-1970-finite-generator} affirme l'existence, dans ce contexte, d'une partition finie à $k$ pièces sitôt que $h_{\mathrm{KS}}(\mathbf{Z}\curvearrowright X,\mu)\leqslant \log k$.
Le théorème de Rokhlin a connu plusieurs généralisations à des actions d'autres groupes,
au premier rang desquels les \textbf{groupes abéliens} par J.-P.~Conze \cite{Conze-1972}.
Ce n'est que récemment qu'il a été formellement étendu à tous les groupes $G$ infinis moyennables par B.~Seward et R.~Tucker-Drob \cite{Seward-Tucker-Drob-2014-arxiv}.
Cela conduit à la définition suivante d'entropie qui pourrait potentiellement rendre les mêmes services que les notions vues précédemment, sans restriction aucune sur la nature du groupe dénombrable. Elle est introduite et étudiée dans une série d'articles extrêmement prometteurs de B.~Seward \cite{Seward-2014-Krieger-finite-th-Rokhlin-1, Seward-2015-Krieger-finite-th-Rokhlin-2}.
\begin{defi}[Entropie de Rokhlin {\cite{Seward-2014-Krieger-finite-th-Rokhlin-1}}]\label{def:Entropie de Rokhlin(Seward)}
L'\textbf{entropie de Rokhlin} d'une action ergodique p.m.p. $G\curvearrowright^{\! T}{\!} (X,\mu)$ d'un groupe dénombrable infini quelconque est
définie comme~:
\begin{equation}
h_{\mathrm{Rok}}(G\curvearrowright^{\! T}{\!} X,\mu)\overset{\mathrm{def}}{=}\inf\{H(\alpha): \textrm{ $\alpha$ partition génératrice dénombrable}\}.
\tag{\textbf{Entropie de Rokhlin}}
\end{equation}
\end{defi}
Une version non ergodique à l'étude \cite{Seward-2016-Krieger-finite-th-Rokhlin-3} fait appel à l'entropie de Shannon conditionnelle relativement à la sous-$\sigma$-algèbre des parties $G$-invariantes.
Observons que si $\alpha \colon X\to K$ est une partition finie à $\vert K \vert$ pièces, alors le cardinal de $\mathcal{M}_\mu(\alpha, F, \epsilon, \sigma_n)$ est trivialement majoré par le nombre $\vert K \vert^{\vert D_n\vert}$ de toutes les $K$-partitions sur $D_n$, ce qui conduit à la majoration $h_{\mathrm{mes}}^{\Sigma}(G\curvearrowright X,\mu)\leqslant \log \vert K \vert$.
Une estimée plus précise découle de l'article de L.~Bowen \cite[Prop. 5.3]{Bowen-2010-mes-conj-inv-sofic}~:~pour toute action ergodique p.m.p. $G\curvearrowright^{\! T}{\!} (X,\mu)$ et toute approximation sofique $\Sigma$ de $G$,
on a
\begin{equation}
h_{\mathrm{mes}}^{\Sigma}(G\curvearrowright^{\! T}{\!} X,\mu)\leqslant h_{\mathrm{Rok}}(G\curvearrowright^{\! T}{\!} X,\mu).\tag{\textbf{Entropie sofique vs Rokhlin}}
\end{equation}
L'entropie de Rokhlin prend manifestement ses valeurs dans $[0, +\infty]$.
On ignore si, en dehors des cas où $h_{\mathrm{mes}}^{\Sigma}(G\curvearrowright X,\mu)=-\infty$, l'inégalité ci-dessus peut être remplacée par une égalité.
C'est tout de même ce qui se produit pour les décalages de Bernoulli à base finie ou dénombrable des groupes sofiques, puisque pour la partition canonique $\alpha \colon x\mapsto x(\mathrm{id}_{G})$, on a
$H(\alpha)= h_{\mathrm{mes}}^{\Sigma}(G\curvearrowright K^G,\nu^{\otimes G})\leqslant h_{\mathrm{Rok}}(G\curvearrowright K^G,\nu^{\otimes G})\leqslant H(\alpha)$.
Si une partition d'entropie finie réalise l'infimum dans la définition de l'entropie de Rokhlin, on a une réciproque.
\begin{theo}[Seward {\cite[Cor. 1.4]{Seward-2015-Krieger-finite-th-Rokhlin-2}}]
Soit $G\curvearrowright^{\! T}{\!} (X,\mu)$ une action p.m.p. libre ergodique d'un groupe infini dénombrable et $\alpha$ une partition génératrice dénombrable. Si $h_{\mathrm{Rok}}(G\curvearrowright^{\! T}{\!} X,\mu)=H(\alpha)<\infty$, alors l'action est conjuguée à un\footnote{Elle est en fait conjuguée au décalage de Bernoulli évident.} décalage de Bernoulli.
\end{theo}
\begin{theo}[Seward {\cite{Seward-2015-Krieger-finite-th-Rokhlin-2}}]
Soit $G$ un groupe infini dénombrable.
Si $G$ admet des actions libres ergodiques d'entropie de Rokhlin finie, arbitrairement grande, alors~:
\begin{itemize}
\item [(i)] l'entropie de Rokhlin des décalages de Bernoulli de $G$ est égale à l'entropie de Shannon de leur base ;
\item[(ii)] les facteurs des Bernoulli de $G$ sont d'entropie de Rokhlin non nulle ;
\item[(iii)] $G$ satisfait la conjecture de surjonctivité de Gottschalk\footnote{Voir section~\ref{sect:groupes sofiques}.}.
\end{itemize}
\end{theo}
De plus, B. Seward montre \cite[Cor. 1.14]{Seward-2015-Krieger-finite-th-Rokhlin-2} que si tout groupe dénombrable admet une action libre ergodique avec $h_{\mathrm{Rok}}(G\curvearrowright X,\mu)>0$, alors les propriétés (i, ii, iii) sont vérifiées pour tout groupe infini dénombrable.
Après des améliorations quantitatives pour $\mathbf{Z}$, dues notamment à Denker \cite{Denker-1974} et Grillenberger et Krengel \cite{Grillenberger-Krengel-1976-Krieger-gen-th}, le théorème du générateur fini de Krieger a été poussé à degré de généralité optimale par B.~Seward (pour l'entropie de Rokhlin et un groupe non nécessairement sofique), qui montre une certaine flexibilité dans les partitions génératrices (et incidemment qu'on peut concocter une version de l'entropie de Rokhlin avec des partitions finies).
\begin{theo}[Seward {\cite{Seward-2014-Krieger-finite-th-Rokhlin-1}}]
Soit $G$ un groupe infini dénombrable et $G\curvearrowright^{\! T}{\!} (X,\mu)$ une action p.m.p. ergodique mais pas nécessairement libre sur $(X,\mu)$ sans atome.
Pour tout vecteur de probabilité (fini ou infini) $\bar{p}=(p_i)_{i\in K}$ tel que $\protect{h_{\mathrm{Rok}}(G\curvearrowright^{\! T}{\!} X,\mu)< H(\bar{p})}$, il existe une partition génératrice dont les pièces sont exactement de mesure $\mu(A_i)=p_i$ pour tout~$i\in K$.
\end{theo}
\subsection{Entropie topologique et principe variationnel}
\label{sect: ent topologique Kerr-Li}
L'entropie s'est invitée en dynamique topologique avec l'article \cite{Adler-Konheim-McAndrew-1965-top-entropy} de R.~Adler, A.~Konheim et M.~McAndrew qui introduisent l'entropie topologique pour les homéomorphismes $S \colon X\to X$ d'un espace compact métrisable\footnote{Ils considèrent plus généralement une application continue d'un espace topologique, mais pour ce qui nous intéresse, ce degré de généralité n'est pas requis.}, guidés par la méthode de Kolmogorov-Sina{\u\i}. Il s'agit d'un invariant de conjugaison topologique.
Le rôle des partitions est joué par les recouvrements ouverts $\mathfrak{A}$ de $X$, auxquels on attache un nombre $N(\mathfrak{A})\overset{\mathrm{def}}{=}$ le cardinal minimal d'un sous-recouvrement. Ils définissent~:
\begin{align*}
h_{\mathrm{top}}(\mathbf{Z}\curvearrowright^{\! S}{\!} X; \mathfrak{A})& \overset{\mathrm{def}}{=} \lim\limits_{n\to \infty} \frac {1}{n}\log N(\mathfrak A\vee S^{-1}\mathfrak A\vee\cdots\vee S^{-n}\mathfrak A),
\\
h_{\mathrm{top}}(\mathbf{Z}\curvearrowright^{\! S}{\!} X)& \overset{\mathrm{def}}{=} \sup\limits_{\mathfrak{A}} \{h_{\mathrm{top}}(S\curvearrowright X; \mathfrak{A}) : \mathfrak{A}\textrm{ recouvrement ouvert} \}\tag{\textbf{Entropie topologique}}.
\end{align*}
Une variante est proposée par R.~Bowen\footnote{Rufus Bowen, apparemment sans lien de famille avec Lewis Bowen, l'inventeur de l'entropie sofique mesurée.} \cite{Bowen-1971-ent-top} et E.~Dinaburg \cite{Dinaburg-1970-announct-of-Dinaburg-1971, Dinaburg-1971-conn-var-ent-charac} ; et c'est plutôt celle-ci qui inspirera D.~Kerr et H.~Li. Il s'agit de compter le nombre de segments d'orbites qui sont \textbf{$\kappa$-séparés} via une distance auxiliaire $\rho$.
Plus précisément, soit $N_{\kappa}(n, \rho_{\infty})$ le nombre maximum (de points $x\in X$ et) de fonctions\footnote{\em Segments d'orbites.} \[\phi_{x} \colon \{0, 1, \cdots, n\}\to X, \ i\mapsto S^{i}(x)\] qui soient deux à deux à $\rho_{\infty}$-distance $\geqslant \kappa>0$
où
\[\rho_{\infty}(\phi_x, \phi_y) \overset{\mathrm{def}}{=} \max\limits_{i\in \{0,1, \cdots, n\}} \rho(S^{i}(x), S^{i}(y)).\]
R.~Bowen \cite{Bowen-1971-ent-top} et E.~Dinaburg \cite{Dinaburg-1971-conn-var-ent-charac} montrent que le taux de croissance exponentielle en $n$ permet de retrouver l'entropie topologique d'Adler-Konheim-McAndrew et ce, indépendamment du choix d'une distance $\rho$ compatible avec la topologie~:
\begin{equation}
h_{\mathrm{top}}(\mathbf{Z}\curvearrowright^{\! S}{\!} X)=\lim\limits_{\kappa\to 0} \limsup\limits_{n\to \infty} \frac{\log N_{\kappa}(n, \rho_{\infty})}{n}.
\tag{\textbf{Bowen-Dinaburg}}
\end{equation}
D.~Kerr et H.~Li étendent la notion aux actions continues des groupes sofiques, sur les compacts métrisables. Leur première version \cite{Kerr-Li-2011-Variationnal-principle}, exprimée en termes d'algèbres d'opérateurs, est exprimée dans \cite{Kerr-Li-2013-sofic-amenabl-dyn-entrop} en termes dynamiques sur l'espace.
Soit $G$ un groupe sofique et $\Sigma = (\sigma_n \colon G \rightarrow \mathrm{Sym}(D_n))_{n}$ une approximation sofique de $G$.
Soit $G \curvearrowright X$ une action continue de $G$ sur un espace compact métrisable $X$.
Soit $\rho \colon X \times X \rightarrow [0, \infty[$ une \textbf{pseudo-distance}\footnote{Cette généralisation est particulièrement pertinente lorsqu'on regarde un décalage de Bernoulli $K^G$ et une pseudo-distance induite par sa partition canonique $\alpha \colon K^G\to K$.}
continue sur $X$ qui soit \textbf{génératrice} ; \emph{i.e.} $\rho$ est symétrique, satisfait l'inégalité triangulaire, et pour tout $x \neq y \in X$ il existe un $g \in G$ avec $\rho(g \cdot x, g \cdot y) > 0$. Pour des applications $\phi, \phi' \colon D_n \rightarrow X$ on définit
\begin{equation*}
\rho_2(\phi, \phi') \overset{\mathrm{def}}{=} \left( \frac{1}{\vert D_n \vert} \sum_{v\in D_n} \rho(\phi(v), \phi'(v))^2 \right)^{1 / 2}
\hskip10pt \textrm{ et } \hskip10pt
\rho_\infty(\phi, \phi') \overset{\mathrm{def}}{=} \max_{v \in D_n} \rho(\phi(v), \phi'(v)).
\end{equation*}
Pour une partie finie $F \subset G$ et $\delta > 0$, soit
\begin{equation*}
\mathrm{Map}(\rho, F, \delta, \sigma_n) \overset{\mathrm{def}}{=} \left\{\phi \colon D_n \rightarrow X:
\forall f\in F,\ \ \rho_2 \left( \phi \circ \sigma_n(f), \ f \cdot \phi \right) < \delta\right\}.
\end{equation*}
C'est la collection des applications $\phi \in X^{D_n}$ qui sont {\em presque équivariantes}
(à $\delta$ près, sous la \og presque action\fg\ $\sigma_n$ restreinte à la partie finie $F\subset G$).
Finalement, on pose~:
\begin{equation}
N_\kappa(\mathrm{Map}(\rho, F, \delta, \sigma_n), \rho_\infty)
\end{equation}
le cardinal maximal d'un ensemble \textbf{$(\rho_\infty, \kappa)$-séparé} ; \emph{i.e.}
un ensemble tel que pour toute paire d'éléments $\phi$ et $\phi'$ on ait $\rho_\infty(\phi, \phi') \geqslant \kappa$.
\begin{defi}[Kerr-Li {\cite[Def. 2.3]{Kerr-Li-2013-sofic-amenabl-dyn-entrop}}]
L'\textbf{entropie sofique topologique} de l'action continue $G \curvearrowright X$ sur le compact métrisable $X$, relativement à l'approximation sofique $\Sigma$, est définie comme~:
\begin{equation}
h_{\mathrm{top}}^{\Sigma}(G\curvearrowright X) = \sup_{\kappa > 0} \inf_{\delta > 0} \inf_{\substack{F \subset G \\ F \text{ fini}}} \limsup_{n \rightarrow \infty} \ \frac{\log N_\kappa \left( \mathrm{Map}(\rho, F, \delta, \sigma_n), \rho_\infty \right)}{\vert D_n \vert} .
\tag{\textbf{Entropie sofique topologique}}
\end{equation}
\end{defi}
D.~Kerr et H.~Li \cite{Kerr-Li-2013-sofic-amenabl-dyn-entrop} démontrent que la valeur de $h_{\mathrm{top}}^{\Sigma}(G\curvearrowright X)$ ne dépend pas du choix de la pseudo-distance continue génératrice $\rho$ (bien qu'elle puisse dépendre de $\Sigma$). Observons que de nouveau $h_{\mathrm{top}}^{\Sigma}(G\curvearrowright X)\geqslant 0$ ou bien $h_{\mathrm{top}}^{\Sigma}(G\curvearrowright X) = -\infty$.
Lorsque le groupe $G$ est moyennable, ils montrent qu'on retrouve la notion classique d'entropie topologique, et ce pour tout choix d'approximation sofique
\cite{Kerr-Li-2013-sofic-amenabl-dyn-entrop}. C'est la version topologique du théorème~\ref{th: entropie sof. gp moyennable=KS}.
\medskip
Ce qu'on appelle le \textbf{principe variationnel} est un énoncé qui affirme que l'entropie topologique d'une action continue $G \curvearrowright^{\! S}\! X$ est le \emph{supremum} des entropies mesurées pour tous les éléments de $M(G \curvearrowright^{\! S}\! X)$~: l'ensemble des mesures boréliennes de probabilité $G$-invariantes.
Sa version classique pour les actions continues de $\mathbf{Z}$ est due à
T.~Goodman \cite{Goodman-1971-princ-variat} et repose sur des résultats de E.~Dinaburg \cite{Dinaburg-1970-announct-of-Dinaburg-1971, Dinaburg-1971-conn-var-ent-charac} et W. Goodwyn\footnote{Qui démontre l'inégalité $\geqslant$.} \cite{Goodwyn-1969-maj-ent-top-by-mes}.
\begin{theo}[Principe variationnel, Kerr-Li {\cite[Th. 6.1]{Kerr-Li-2011-Variationnal-principle}}]
Soit $G \curvearrowright^{\! S}\! X$ une action continue, sur le compact métrisable $X$, du groupe sofique $G$ et soit $\Sigma$ une approximation sofique de $G$ ; alors
\begin{equation}
h_{\mathrm{top}}^{\Sigma}(G \curvearrowright^{\! S}\! X)=\sup\left\{h_{\mathrm{mes}}^{\Sigma}(G \curvearrowright^{\! S}\! X, \mu) : \mu \in M(G \curvearrowright^{\! S}\! X)\right\}.\tag{\textbf{Principe variationnel}}
\end{equation}
\end{theo}
Observons qu'il se peut que de telles actions n'aient pas de mesure invariante. Il s'agit alors d'un cas où $h_{\mathrm{top}}^{\Sigma}(G \curvearrowright^{\! S}\! X)=-\infty$. Cela n'arrive jamais si $G$ est moyennable et d'ailleurs l'existence d'une action continue telle que $M(G \curvearrowright^{\! S}\! X)=\emptyset$ est un critère de non-moyennabilité.
La recherche et l'identification de mesure qui réalise le \emph{supremum} dans le principe variationnel, et les liens avec le nombre de points périodiques constituent des thèmes récurrents en entropie topologique.
On fournit quelques éléments dans ce sens dans le cadre sofique.
Par exemple, pour un décalage de Bernoulli $K^G$ de base finie d'un groupe sofique, l'entropie sofique mesurée maximale $=\log\vert K\vert$ est réalisée par la mesure $\nu_{\mathbf{u}}^{\otimes G}$ provenant de la probabilité uniforme $\nu_{\mathbf{u}}$ sur $K$.
On peut observer qu'il s'agit de la mesure de Haar sur $\mathbf{K}^G$ lorsque $K=\mathbf{K}$ est un groupe fini.
\begin{theo}[Gaboriau-Seward {\cite[Th. 8.2]{Gab-Seward-2015-arxiv}}]
\label{th:Gab-Sew haar-ent max}
Si $H$ est un groupe profini sur lequel le groupe sofique $G$ agit par automorphismes continus, de sorte que le sous-groupe homocline soit dense, alors la mesure de Haar de $H$ est d'entropie maximale~:
\[h_{\mathrm{top}}^{\Sigma}(G\curvearrowright H)=h_{\mathrm{mes}}^{\Sigma}(G\curvearrowright X, \mathrm{Haar}),\]
pour toute approximation sofique $\Sigma$ de $G$.
\end{theo}
On rappelle que le sous-groupe \textbf{homocline} est le sous-groupe des points $h\in H$ tels que $g_n.h\to \mathrm{id}_{H}$ pour toute suite injective $(g_n)_{n}$ dans $G$, c'est par exemple les éléments de support fini dans $H<\mathbf{K}^{G}$.
\begin{theo}[Gaboriau-Seward {\cite[Th. 4.6]{Gab-Seward-2015-arxiv}}]
\label{th: Gab-Sew ent et pt fixes}
Soit $G$ un groupe résiduellement fini et $\Sigma$ une approximation sofique associée à une chaîne $(G_n)_{n}$ de sous-groupes d'indice fini. Soit $\mathbf{K}$ un groupe fini et $X\subset \mathbf{K}^G$ un sous-groupe compact $G$-invariant
qui soit un sous-décalage de type fini.
Alors
\[h_{\mathrm{top}}^{\Sigma}(G\curvearrowright X)=\limsup \frac{1}{[G:G_n]} \ \log \vert \mathrm{Fix}_{G_n} (X) \vert. \]
\end{theo}
Rappelons qu'une partie $X\subset \mathbf{K}^G$ est un \textbf{sous-décalage de type fini} (défini par une partie finie $W$ du groupe $G$ et une partie $P\subset \mathbf{K}^W$)
si c'est un fermé $G$-invariant qui est maximal sous la condition
que pour tout $x\in X$, l'application $W\to \mathbf{K}$, $w\mapsto \alpha(w.x)$ appartienne à $P$.
Si $\mathbf{K}$ et $\mathbf{L}$ sont des groupes finis et $\Phi \colon \mathbf{K}^G\to \mathbf{L}^G$ est un homomorphisme $G$-équi\-variant continu, alors on peut appliquer
le Théorème~\ref{th:Gab-Sew haar-ent max} à l'image de $\Phi$ et le Théorème~\ref{th: Gab-Sew ent et pt fixes} au noyau de $\Phi$.
\section{Définitions \& notations}
\label{sect: definitions}
Si $Y$ est un ensemble fini, on note $\vert Y\vert $ son cardinal.
Pour un ensemble fini non vide $D$, on note $\mathbf{u}$ la \textbf{mesure uniforme} $\mathbf{u}(A)=\frac{\vert A\vert}{\vert D\vert}$ sur $D$.
On désigne par $\mathrm{Sym}(Y)$ le groupe symétrique (de toutes les permutations) de $Y$.
\subsection{Partitions}
Une \textbf{partition} $\alpha=(A_k)_{k\in K}$ d'un ensemble $X$ est une famille
de parties de $X$ mutuellement disjointes et qui forme un recouvrement de $X$.
En particulier, une permutation des indices conduit en général à des partitions distinctes.
Pour insister sur ce point, on utilise parfois la terminologie \textbf{$K$-partition}.
Un autre point de vue intéressant consiste à considérer la partition $\alpha$ de $X$ comme la fonction $\alpha \colon X\to K$ qui à $x\in X$ associe l'indice de la pièce qui le contient, et donc $A_k=\alpha^{-1}(k)$.
Dans ce contexte, $\alpha$ est parfois appelée une \textbf{observable}, notamment chez L.~Bowen \cite{Bowen-2010-mes-conj-inv-sofic}.
Dans le cadre d'un espace de probabilité $(X,\mathcal{B}, \mu)$, c'est-à-dire un ensemble $X$ muni d'une tribu $\mathcal{B}$ (qu'on omet de mentionner lorsqu'une confusion nous paraît improbable)
et d'une mesure de probabilité $\mu$ sur $\mathcal{B}$, on se restreint à des partitions au plus dénombrables, dont les pièces sont mesurables.
L'\textbf{entropie de Shannon} d'une partition finie ou dénombrable $\alpha=(A_k)_{k\in K}$ est définie par la formule~:
\begin{equation}
H(\alpha)\overset{\mathrm{def}}{=}-\sum_{k\in K} \mu(A_k)\log \mu(A_k).
\end{equation}
Elle représente, en théorie de l'information, la quantité d'information ($-\log \mu(A_k)$ pour chaque pièce $A_k$) contenue en moyenne dans les pièces de la partition.
Le sort des pièces de mesure nulle est réglé en convenant que $0 \log(0)=0$.
Si $\nu$ est une mesure de probabilité sur un ensemble fini ou dénombrable $K$, on note encore $H(\nu)$
l'entropie de la partition en singletons.
On note $\alpha\vee \beta \colon
\left(\begin{array}{ccc}
X & \to & K\times L
\\
x&\mapsto &(\alpha(x), \beta(x))
\end{array}\right)$ le \textbf{joint} des partitions $\alpha \colon X\to K$ et $\beta \colon X\to L$, c'est-à-dire la partition formée des intersections
$A_{k}\cap B_{l}=\alpha^{-1}(k)\cap \beta^{-1}(l)$.
La fonction $t\mapsto -t\log t$ pour $t\in ]0,1]$ est concave. Lorsqu'on raffine une partition, son entropie augmente.
On a toujours,
$H((A_k)_{k\in K})\leqslant \log \vert K\vert $ avec égalité si et seulement si les parties $A_k$
ont même mesure.
Par ailleurs, $H(\alpha\vee \beta)\leqslant H(\alpha)+H(\beta)$ avec égalité si et seulement si les partitions sont indépendantes~: $\mu(A_i\cap B_j)=\mu(A_i)\mu(B_j)$ pour tout $(i,j)\in K\times L$.
\subsection{Partitions et actions de groupes}
De manière générale, une \textbf{conjugaison} entre deux actions $G\curvearrowright^{\! T} \! X$ et $G\curvearrowright^{\! S} \! Y$ est un isomorphisme $\Phi \colon X\to Y$
tel que
\begin{equation}
\tag{\'Equivariance}
g\in G, \ \ \Phi\circ T(g)(x)=S(g)\circ \Phi(x).
\end{equation}
Les actions sont alors dites \textbf{conjuguées}.
Si on parle d'actions continues, on demande que $\Phi$ soit un homéomorphisme.
Si on parle d'action p.m.p. $G\curvearrowright^{\! T} \! (X, \mu)$ et $G\curvearrowright^{\! S} \! (Y, \nu)$, on demande une bijection bimesurable préservant la mesure entre deux parties $X'\subset X$ et $Y'\subset Y$ de mesure pleine telle que la condition d'équivariance soit vérifiée pour tout $x\in X'$.
Dans le cadre mesuré, un \textbf{facteur} est une application mesurable équivariante $\Phi \colon (X,\mu)\to (Y,\nu)$ essentiellement surjective~: $Y\setminus \Phi(X)$ est négligeable.
\subsection{Décalages de Bernoulli}
De façon générale, si $K$ est un ensemble et $V$ un ensemble dénombrable muni d'une action $G\curvearrowright V$ d'un groupe dénombrable, alors l'espace $K^{V}=\prod_{v\in V} K$ des fonctions $V\to K$ est muni de l'action par décalage~:
\[\forall x\in K^G, \forall g\in G, \ g\cdot x(v)= x(g^{-1} v), \ \forall v\in V.\]
Lorsque $V=G$ sur lequel $G$ agit par multiplication à gauche, alors $G\curvearrowright K^{G}$ est appelé \textbf{décalage de Bernoulli} de base $K$.
Si $K$ est un espace topologique métrisable séparable, $K^{V}$ est équipé de la topologie produit et l'action par décalage est continue.
Une mesure borélienne $\nu$ sur $K$ délivre la mesure borélienne produit $\nu^{\otimes G}$ sur $K^G$. Elle est invariante sous l'action de $G$. On parle encore de \textbf{décalage de Bernoulli} (mesuré, cette fois-ci)~:
\[ G\curvearrowright (K^{G}, \nu^{\otimes G}).\]
Lorsque $K$ est fini ou dénombrable muni d'une mesure $\nu$, la partition \textbf{canonique} $\alpha$ est définie par l'évaluation
$\alpha \colon \left(\begin{array}{ccc}
K^{G}&\to &K\\
x&\mapsto &x(\mathrm{id}_{G})
\end{array}\right)$ en l'élément neutre du groupe.
Elle est génératrice au sens ci-dessous.
\subsection{Partitions génératrices}
\label{subsubsect:partitions generatrices}
Une action p.m.p. $G\curvearrowright (X,\mathcal{B}_X,\mu)$ étant donnée, une $K$-partition finie ou dénombrable $\alpha \colon X\to K$ nous parle en réalité d'un \textbf{facteur} de cette action.
Elle nous fournit la sous-$\sigma$-algèbre $\mathcal{S}_{\alpha}$ (engendrée par les $\cup_{F\subset G, F \textrm{ finie }} \alpha^{\vee F}$)
laissée globalement invariante par $G$ et à laquelle correspond un certain facteur.
On peut introduire explicitement le facteur~: c'est l'application $G$-équivariante naturelle $\pi \colon X\to K^{G}, x\mapsto (\alpha(g^{-1}(x)))_{g\in G}$, où $G$ agit sur $K^G$ par décalage de Bernoulli et laisse invariante la mesure $\nu\overset{\mathrm{def}}{=}\pi_{*}(\mu)$, image directe de la mesure $\mu$.
Une partition $\alpha$ est \textbf{génératrice} \index{génératrice! partition}\index{partition! génératrice} s'il existe une partie $X'\subset X$ de mesure $1$ telle que pour tout $ x\not= y\in X'$, il y a un $g\in G$ pour lequel $\alpha$ sépare $g.x$ de $g.y$ \emph{i.e.} $\alpha(g.x)\not=\alpha(g.y)$.
De manière équivalente, la partition $\alpha$ est génératrice si et seulement si $\pi$ est essentiellement injectif (\emph{i.e.} injectif sur une partie de mesure pleine).
Inversement, pour un facteur $\pi \colon (X,\mathcal{B}_X,\mu)\to (Y,\mathcal{B}_Y,\nu)$ non essentiellement injectif, toute partition finie mesurée $\xi$ de $Y$ se relève en une partition $\alpha=\pi^{-1}(\xi)$ non génératrice~: la sous-$\sigma$-algèbre engendrée n'est pas capable de séparer les points d'une même fibre.
\section{Groupes sofiques}\label{sect:groupes sofiques}
Un groupe dénombrable est résiduellement fini s'il admet une \textbf{chaîne normale}, c'est-à-dire une suite décroissante de sous-groupes normaux d'indices finis $(G_n)_{n}$ telle que $\cap_n G_n=\{1_G\}$.
Ce qu'on va en retenir, c'est qu'on dispose d'ensembles finis ${D_n\overset{\mathrm{def}}{=}G_n\backslash G}$ (disons les classes à droite) et d'applications dans leurs groupes symétriques $\sigma_n \colon G \rightarrow \mathrm{Sym}(D_n)$ (obtenues par multiplication à droite par l'inverse\footnote{Voir ci-dessous l'interprétation en termes de graphes, pour une explication sur ce choix.}), telles que
\begin{enumerate}
\item[\rm (i)] (action) les $\sigma_n$ sont des homomorphismes de groupes,
\item[\rm (ii)] (liberté) les $\sigma_n$ séparent les éléments de $G$ en un sens fort~: pour tout $g\in G\setminus\{1_G\}$ et $n$ assez grand, l'image $\sigma_n(g)$ agit sans point fixe sur l'ensemble fini $D_n$.
\end{enumerate}
Ce sont ces conditions qu'on imite, en les relaxant, dans la définition~\ref{defn: gp approx sofic} pour obtenir la notion de groupe sofique, en leur demandant d'être satisfaites asymptotiquement.
On demande que la proportion de points de $D_n$ sur lesquels elles sont vérifiées tende vers~$1$.
Si on préfère une version non asymptotique, on dira que $G$ est \textbf{sofique} si
pour toute partie finie $F\subset G$ et tout réel $\delta>0$, il existe une \textbf{$(F,\delta)$-approximation}, c'est-à-dire un ensemble fini $D$ et une application $\sigma \colon G \rightarrow \mathrm{Sym}(D)$ tels que
\begin{enumerate}
\item[\rm (i)] (presque action) \hfil
$\frac{1}{\vert D\vert}\, \vert \{ {v} \in D \ \vert \ \sigma_n(g) \circ \sigma_n(h)({v}) = \sigma_n(g h)({v})\} \vert \geqslant 1-\delta$, \hfil $\forall g, h \in F$
\item[\rm (ii)] (presque libre) \hfil
$\frac{1}{\vert D\vert}\, \vert\{ {v} \in D \ \vert \ \sigma_n(g)({v}) \neq {v}\}\vert \geqslant 1-\delta$, \hfil $\forall g\in F\setminus\{1_G\}$.
\end{enumerate}
On constate que la condition de normalité de la chaîne de sous-groupes d'indices finis n'est pas indispensable pour que les $G_n\backslash G$ fournissent une approximation sofique. La condition optimale (parfois appelée \textbf{condition de Farber}) est que $\forall g \in G \setminus \{1_G\}$,
$ \lim_{n \rightarrow \infty} \mathbf{u}_{G_n\backslash G} \left(\{G_n v \ \vert \ v \in G, \textrm{ t. q. } \ g \in v^{-1} G_n v\}\right)= 0$.
Dans ce cas, on dira que $(G_n)_{n}$ est une \textbf{chaîne sofique}.
Les groupes résiduellement finis, par exemple les groupes linéaires de type fini (Malcev 1940) et notamment les groupes libres, sont sofiques.
On dispose d'une autre grande classe de groupes sofiques~: les groupes moyennables.
Rappelons qu'un groupe dénombrable est \textbf{moyennable} s'il admet une \textbf{suite de F{\o}lner}, c'est-à-dire une suite de parties finies $F_n\subset G$ qui vérifient
pour tout $g\in G$
\begin{equation*}
\lim\limits_{n\to \infty}\frac{\vert F_n\cdot g^{-1}\Delta F_n\vert}{\vert F_n\vert}=0.
\end{equation*}
Autrement dit, l'action de $G$ sur lui-même, par multiplication à droite par l'inverse, laisse les parties $F_n$ asymptotiquement invariantes.
On obtient alors une approximation sofique de $G$ en observant que, par multiplication à droite par l'inverse, tout élément $g$ de $G$ définit une bijection de l'ensemble fini $D_n\overset{\mathrm{def}}{=}F_n\subset G$, à une partie asymptotiquement négligeable près. On pose alors $\sigma_n(g)(f)=f g^{-1}$ lorsque $f$ et $f g^{-1}$ appartiennent à $F_n$ et on l'étend à souhait en une bijection de $F_n$ tout entier.
Voici une liste de quelques propriétés de stabilité pour la soficité.
Un groupe est sofique si et seulement si tous ses sous-groupes de type fini sont sofiques.
Si un groupe $G$ possède un sous-groupe normal sofique tel que le quotient soit moyennable, alors $G$ est sofique.
Les produits directs de groupes sofiques sont sofiques.
Un produit amalgamé ou une HNN-extension de groupes sofiques au-dessus d'un groupe moyennable est sofique (G.~Elek et E.~Szab{\'o} \cite{Elek-Szabo-2011-sofic-amalg-over-amenable} et indépendamment L.~ P{\u{a}}unescu \cite{Paunescu-2011-sofic-act-equiv-rel}).
Une interprétation agréable, notamment lorsque le groupe $G$ est engendré par une partie finie $S$, consiste à considérer $D_n$ comme l'ensemble des sommets d'un graphe $\mathcal{G}_n$ dont les arêtes, orientées et étiquetées par $s\in S$, joignent chaque $v$ à son image par $\sigma_n(s)$~:
\begin{equation}
\mathcal{G}_{n}=(D_n, ([v, \sigma_n(s).v])_{v\in D_n, s\in S}).
\end{equation}
Les conditions (i) et (ii) de la définition de soficité reviennent à dire que les $v\in D_n$ depuis lesquels la boule de centre $v$, de rayon $R$, dans $\mathcal{G}_n$ est isomorphe\footnote{Comme graphe orienté, étiqueté.} à la boule de même rayon\footnote{Disons centrée en $\mathrm{id}_{G}$, mais les boules de même rayon sont toutes isomorphes.} dans le graphe de Cayley\footnote{Où $\rho$ est l'action à gauche de $G$ sur lui-même par multiplication à droite par l'inverse. Un \textbf{graphe de Cayley} est un graphe orienté, équipé d'une action de $G$ simplement transitive sur les sommets. Le choix d'un point base permet d'identifier les sommets avec $G$ et fournit une action $\rho$ qui commute avec la première, et dont découle l'étiquetage.}
\begin{equation}
\mathrm{Cayley}(G, S, \rho)=\mathcal{G}=(G, ([v, \rho(s).v])_{g\in G, s\in S})
\tag{\textbf{Graphe de Cayley}}
\end{equation}
forment une partie de $D_n$ dont, à $R$ fixé, la proportion tend vers $1$ lorsque $n$ tend vers l'infini~:
\begin{equation}
\forall R>0, \ \ \lim\limits_{n \to \infty}\mathbf{u}_{n}\left(\left\{ v\in D_n\ \vert\ B_{\mathcal{G}_{n}}(v,R)\simeq B_{\mathcal{G}}(v,R)\right\}\right)=1.
\tag{\textbf{Soficité \& graphes}}
\end{equation}
Les exemples décrits ci-dessus s'interprètent alors de la manière suivante.
\\
-- Si $G$ est résiduellement fini et $(G_n)_{n}$ est une chaîne sofique de sous-groupes, alors $\mathcal{G}_{n}$ n'est autre que ce qu'on appelle le graphe de Schreier, le graphe quotient~:
\begin{equation}
\mathcal{G}_{n}=G_n\backslash \mathrm{Cayley}(G, S, \rho).
\tag{\textbf{Graphe de Schreier}}
\end{equation}
-- Si $G$ est moyennable, alors $\mathcal{G}_{n}$ est formé à partir de la restriction du graphe $\mathrm{Cayley}(G, S, \rho)$ à la partie de F\o lner $F_n\subset G$, en bricolant les arêtes du bord, c'est-à-dire en reliant bijectivement pour chaque $s\in S$ les sommets $\{g\in F_n \ \vert\ gs^{-1}\not\in F_n\}$ aux sommets $\{gs^{-1} \in F_n \ \vert\ g\not\in F_n\}$.
\\
-- Les graphes associés aux groupes libres, pour $S$ partie génératrice libre, sont ceux dont le tour de taille tend vers l'infini (voir section~\ref{sect:plus sur l'ex de OW}).
\medskip
Les groupes sofiques ont été introduits (sous un autre nom\footnote{Groupes à graphes de Cayley {\em initially subamenable}.}) par M.~Gromov \cite{Gromov-1999-symbolic-algebraic-varieties},
pour lesquels il a montré la validité de la \textbf{conjecture de surjonctivité de Gottschalk}~: toute application continue $G$-équivariante injective $\Phi \colon K^G\to K^G$ est automatiquement surjective, où $K$ est fini et $G$ agit par décalage de Bernoulli.
D.~Kerr et H.~Li en donnent une preuve entropique.
\begin{theo}[Kerr-Li {\cite[Th. 4.2]{Kerr-Li-2011-Variationnal-principle}}]
\label{th: Kerr-Li entrop Gottschalk conject}
Soit $G \curvearrowright K^{G}$ l'action (continue) d'un groupe sofique $G$ par décalage de Bernoulli de base finie.
Toute restriction de cette action à une partie fermée propre $G$-invariante, est d'entropie sofique topologique strictement inférieure à celle de $G\curvearrowright K^G$.
\end{theo}
En particulier, si $\Phi \colon X=K^{G}\to K^{G}$ est une application continue $G$-équivariante injective, son image est un fermé $G$-invariant de même entropie que $G\curvearrowright K^{G}$.
Cette image ne peut pas être propre ; $\Phi$~doit être surjective.
\medskip
La terminologie \textbf{sofique}, dérivée d'un mot hébreu signifiant \og fini\fg, a été introduite par B.~Weiss \cite{Weiss-2000-sofic-gp}.
Il faut noter qu'à ce jour, on ne connaît aucun exemple de groupe qui ne soit pas sofique.
Pour une jolie introduction aux groupes sofiques, on pourra consulter \cite{Pestov-2008-sofic-gps-survey}. Voir aussi l'ouvrage \cite{Capraro-Lupini-LNM-sofic-hyperlin-gps}.
\section{Entropie sofique, le point de vue externe}
\subsection{Un exemple de \og point de vue externe\fg\ sur l'entropie, sans dynamique}
\label{sect: preuve du lem Boltzmann-Sanov}
On revient sur la proposition~\ref{prop:Boltzmann-Sanov} \og Boltzmann-Sanov\fg\ et l'entropie sans action de groupe.
Soit $\alpha=(A_k)_{k\in K}$ une $K$-partition finie d'un espace de probabilité $(X,\mu)$.
On considère, sur un ensemble fini\footnote{Qu'on pense grand ; et ce qu'on regarde ne dépend que de son cardinal.} $D$ de cardinal $d$, la famille de toutes les partitions $(V_k)_{k\in K}$ qui imitent bien $\alpha$ en termes de mesure des pièces pour la mesure uniforme sur $D$
\begin{equation}
{\mathcal{M}(\alpha,\epsilon,d)}\overset{\mathrm{def}}{=}\left\{K\textrm{-partitions } (V_k)_{k\in K} \textrm{ de } D \ \Big\vert \ \forall k\in K,\ \left\vert \frac{\vert V_k \vert}{\vert D\vert} -\mu(A_k)\right\vert <\epsilon\right\}, \label{eq: estimee exp nb bon. part.}
\end{equation}
pour un certain $\epsilon>0$ petit. On compare le taux de croissance exponentielle en $d$ de son cardinal à l'entropie $H(\alpha)$.
Plus précisément~:
\begin{lemm}[Entropie de Shannon et modèles finis, Boltzmann-Sanov]
\label{lem:entropie sans groupe}
Dans ce contexte,
$\forall\delta>0$,
$\exists\epsilon_0>0$ tel que $\forall \epsilon \in ]0,\epsilon_0[$,
$\exists d_0\in \mathbf{N}$ tel que pour tout ensemble fini $D$ de cardinal $d>d_0$~:
\begin{equation}
\mathrm{e}^{ (H(\alpha)-\delta) d}
<
\left\vert
\mathcal{M}(\alpha,\epsilon,d)
\right\vert
< \mathrm{e}^{ (H(\alpha)+\delta) d}.
\end{equation}
\end{lemm}
Et cela conduit à une définition \og externe\fg\ de l'entropie de Shannon de $\alpha$, en extrayant le taux de croissance exponentielle du nombre de bons modèles (Proposition~\ref{prop:Boltzmann-Sanov}).
On a des preuves purement combinatoires de ce lemme~\ref{lem:entropie sans groupe}, en utilisant la formule de Stirling et les estimées standard.
Mais, en supposant que les $\mu(A_k)$ sont tous non nuls, on a aussi une preuve, qui dans un esprit de principe de grande déviation adopte une mesure adéquate, et qui s'avérera utile en section~\ref{sect: ent sofique Bernoulli}.
Au fond, en interprétant les $K$-partitions de $D$ comme des fonctions $a \colon D\to K$,
\begin{equation*}
\mathcal{M}(\alpha,\epsilon,d)=\left\{
a\in K^{D} \ \Big\vert \ \forall k\in K, \ \left\vert \frac{\vert a^{-1}(k)\vert}{\vert D\vert} - \mu(A_k) \right\vert < \epsilon
\right\}
\end{equation*}
et
cette mesure $\frac{\vert a^{-1}(k)\vert}{\vert D\vert} =\frac{\vert \{v\in D \ \vert \ a(v)=k\} \vert}{\vert D\vert}$ n'est autre que la fréquence d'apparition de la lettre $k$ dans la \og suite\fg\ $(a(v))_{v\in D}$. Et ce qu'on requiert dans (\ref{eq: estimee exp nb bon. part.}), c'est que cette fréquence soit proche de $\mu(A_k)$. Or on dispose d'une mesure naturelle sur $K^D$ qui est bien adaptée à cette question, via la loi des grands nombres~:
quand on effectue un grand nombre de tirages aléatoires indépendants dans $(K,\nu_{\alpha})$, où la mesure $\nu_{\alpha}$ sur $K$ est définie par $\nu_{\alpha}(\{k\})=\mu(A_k)\overset{\mathrm{def}}{=}p_k$, une grande proportion de ces tirages verra la fréquence d'apparition de la lettre $k$ proche de $\mu(A_k)$. En d'autres termes, avec la mesure produit ${\nu_{\alpha}}^{\otimes D} $ sur $K^D$, on a~:
$\forall \delta>0, \forall\epsilon>0$,
il existe une taille $d_1$ de $D$ à partir de laquelle~:
\begin{align*}
{\nu_{\alpha}}^{\otimes D} \left(\left\{a\in K^D\ \Bigg\vert \
{ \forall k
\textrm{ la fréquence de $k$ dans $(a(v))_{v\in D}$ }
\atop
\textrm{ est dans $]p_k-\epsilon, p_k+\epsilon[$}
}\right\}
\right)
>
1 -\delta.
\end{align*}
Mais maintenant qu'on en connaît à peu près les fréquences, chaque atome de notre ensemble $\mathcal{M}(\alpha,\epsilon,d)$ a une mesure qui vérifie~:
\begin{equation}
\prod_{k\in K} p_k^{(p_k+\epsilon) d} < {\nu_{\alpha}}^{\otimes D} (\textrm{atome de } \mathcal{M}(\alpha,\epsilon,d)) < \prod_{k\in K } p_k^{(p_k-\epsilon) d} ;
\end{equation}
soit, $\forall \delta>0$, dès que $0<\epsilon< \frac{\delta}{-2 \sum_{k\in K} p_k}$, pour tout $d\geqslant d_1$ l'encadrement suivant~:
\begin{equation}\label{eq: mesure d'un atome de Bernoulli Dn}
\mathrm{e}^{(-H(\alpha) -\delta/2)d}
< {\nu_{\alpha}}^{\otimes D} (\textrm{atome de } \mathcal{M}(\alpha,\epsilon,d)) <
\mathrm{e}^{(-H(\alpha)+\delta/2)d}.
\end{equation}
Ce qui nous délivre les estimations sur le cardinal~:
\begin{equation}
\mathrm{e}^{(-H(\alpha) -\delta/2)d} \vert \mathcal{M}(\alpha,\epsilon,d)\vert
< {\nu_{\alpha}}^{\otimes D} ( \mathcal{M}(\alpha,\epsilon,d))
< \mathrm{e}^{(-H(\alpha) +\delta/2)d}
\vert \mathcal{M}(\alpha,\epsilon,d)\vert.
\end{equation}
Et pour $d$ suffisamment grand~:
\begin{equation}
\mathrm{e}^{(H(\alpha) -\delta)d}< (1-\delta)
\mathrm{e}^{(H(\alpha) -\delta/2)d}
< \vert \mathcal{M}(\alpha,\epsilon,d)\vert
<
\mathrm{e}^{(H(\alpha) +\delta/2)d}
<
\mathrm{e}^{(H(\alpha) +\delta)d}.
\end{equation}
\subsection{Les modèles finis de la dynamique}
\label{sect: modeles finis de la dynamique}
On ajoute l'ingrédient d'une approximation sofique $\Sigma$ d'un groupe sofique $G$ et
une action p.m.p. $G\curvearrowright (X,\mu)$.
Au risque d'être redondant, une partition finie $\alpha \colon X \rightarrow K$ et la partie finie $F \subset G$ délivrent pour chaque $x\in X$ un certain élément $p\in K^{F}$ ; le \textbf{$F$-parcours} $f\mapsto \alpha(f\cdot x)$ de $x$.
De façon analogue, une partition $a \colon D_n \rightarrow K$ délivre pour tout ${v}\in D_n$ un certain élément $p\in K^{F}$ ;
le $F$-(pseudo)-parcours $f\mapsto a(\sigma_n(f)({v}))$ de $v$.
Ces données définissent ainsi des partitions de $X$ (resp. $D_n$), selon le parcours associé, en les pièces suivantes indexées par les $p\in K^{F}$~:
\begin{align}
U_{p}(\alpha,F)&\overset{\mathrm{def}}{=}\{x\in X\ \vert\ \forall f\in F, \ \alpha(f\cdot x)=p(f)\},\label{def:U p alpha F}
\\
U_{p}(a,F,n)&\overset{\mathrm{def}}{=}\{{v}\in D_n\ \vert\ \forall f\in F, \ a(\sigma_n(f) \cdot {v})=p(f)\} \label{def:U p a F n}.
\end{align}
Et on compare les mesures $\mu(U_{p}(\alpha,F))$ et $\mathbf{u}_n(U_{p}(a,F,n))$ de ces pièces. Plus précisément, pour tout $\epsilon > 0$, posons~:
\begin{equation}\label{eq:def M mu}
\mathcal{M}_\mu(\alpha, F, \epsilon, \sigma_n)\overset{\mathrm{def}}{=}\left\{a \in K^{D_n}\ \Big\vert \ \sum_{p \in K^F}
\left| \mu(U_{p}(\alpha,F)) - \mathbf{u}_n(U_{p}(a,F,n))
\right| < \epsilon\right\},
\end{equation}
l'ensemble de toutes les partitions $a\in K^D$ qui sont $(F,\epsilon)$-plausibles dans l'approximation $\sigma_{n} \colon G\to \mathrm{Sym}(D_{n})$ de $G$. Et on s'intéresse à leur nombre $\vert \mathcal{M}_\mu(\alpha, F, \epsilon, \sigma_n)\vert $.
On a des propriétés de monotonie immédiates~:
\begin{equation}
\textrm{si $F'\subset F$ et $\epsilon\leqslant \epsilon'$, alors }
\mathcal{M}_\mu(\alpha, F, \epsilon, \sigma_n)\subset \mathcal{M}_\mu(\alpha, F', \epsilon', \sigma_n).
\end{equation}
\subsection{Entropie sofique sans partition génératrice}
\label{sect: ent sof. Kerr-Li th. generateur}
On rappelle que si $\alpha \colon X \rightarrow K$ est une partition plus fine que
$\beta \colon X \rightarrow L$, l'application de fusion des pièces est notée $\Theta_{\beta,\alpha} \colon K \rightarrow L$
(voir section~\ref{subsect: Ent sof sans part gen}).
\begin{defi}[Entropie sofique mesurée]\label{def:ent sof mes Kerr}
Soit $\Sigma$ une approximation sofique du groupe $G$ et soit $G\curvearrowright(X, \mathcal{B}_X,\mu)$ une action action p.m.p.
L'\textbf{entropie sofique mesurée} de l'action relativement à $\Sigma$ est définie comme~:
\begin{equation*}
h_{\mathrm{mes}}^{\Sigma}(G\curvearrowright X, \mu)\overset{\mathrm{def}}{=}\sup_{\beta}\
\underbrace{\inf_{\alpha\geqslant \beta}\
\underbrace{\inf_{F} \inf_{\epsilon>0}
\underbrace{\limsup_{n\to \infty} \frac{\log \vert \Theta_{\beta,\alpha} \circ \mathcal{M}_\mu(\alpha, F, \epsilon, \sigma_n) \vert}{\vert D_n\vert}}_{h_{\mathrm{mes}}^{\Sigma}(\beta, \alpha, F, \epsilon)}}_{h_{\mathrm{mes}}^{\Sigma}(\beta, \alpha)}}_{h_{\mathrm{mes}}^{\Sigma}(\beta)},
\end{equation*}
où $\alpha$ et $\beta$ parcourent les partitions mesurables finies et $\alpha$ est plus fine que $\beta$.
\end{defi}
\begin{theo}[Générateur]
\label{th: generateur -Kerr}
Soit $\Sigma$ une approximation sofique du groupe $G$ et soit $G\curvearrowright(X, \mathcal{B}_X,\mu)$ une action action p.m.p.
Si $\beta$ est une partition génératrice finie, alors elle réalise le \emph{supremum}~:
\begin{equation}
h_{\mathrm{mes}}^{\Sigma}(G\curvearrowright X, \mu)=h_{\mathrm{mes}}^{\Sigma}(\beta).
\end{equation}
De plus, pour toute partition finie plus fine $\alpha\geqslant \beta$, on a~:
\begin{equation}
h_{\mathrm{mes}}^{\Sigma}(\beta)=h_{\mathrm{mes}}^{\Sigma}(\beta,\alpha)=h_{\mathrm{mes}}^{\Sigma}(\beta, \beta).
\end{equation}
\end{theo}
\begin{coro}
Si $\alpha$ est une partition génératrice, alors on retrouve bien la définition de L.~Bowen~:
\begin{equation*}
h_{\mathrm{mes}}^{\Sigma}(G\curvearrowright X,\mu) = \inf_{\epsilon > 0} \inf_{\substack{F \subset G\\F \text{ fini}}} \limsup_{n \rightarrow \infty} \frac{\log \left| \mathcal{M}_\mu(\alpha, F, \epsilon, \sigma_n) \right|}{\vert D_n \vert} .
\end{equation*}
\end{coro}
Le théorème~\ref{th: generateur -Kerr} fait usage des quatre ingrédients suivants, dont les preuves reposent sur des arguments de dénombrement sans énorme surprise, mais assez délicats.
Les entités mises en jeu sont des partitions finies $\alpha, \beta, \xi$ telles que $\alpha\geqslant \beta$ et $W$ est une partie finie du groupe, qui contient $\mathrm{id}_{G}$.
\begin{enumerate}
\item
\label{ingredient iterer beta}
$h_{\mathrm{mes}}^{\Sigma}(\beta)=h_{\mathrm{mes}}^{\Sigma}(\beta^{\vee W})$.
\item
\label{ingredient : ineg de Rokhlin}
$h_{\mathrm{mes}}^{\Sigma}(\xi)\leqslant h_{\mathrm{mes}}^{\Sigma}(\beta) +H(\xi \vert \beta)$ pour toute $\xi$ (\og Inégalité de Rokhlin\fg).
\item $h_{\mathrm{mes}}^{\Sigma}(\beta,\alpha^{\vee W})=h_{\mathrm{mes}}^{\Sigma}(\beta, \alpha)$.
\label{ingredient iterer alpha}
\item
\label{ingredient partitions dans sous-algebre}
Soit $\mathcal{S}$ une sous-algèbre dense dans $\mathcal{B}_X$ qui contient $\beta$. Pour toute $\alpha\geqslant \beta$, il existe une partition $\xi$ dans $\mathcal{S}$ telle que $\xi\geqslant \beta$
telle que $h_{\mathrm{mes}}^{\Sigma}(\beta, \xi)\leqslant h_{\mathrm{mes}}^{\Sigma}(\beta, \alpha)$.
\end{enumerate}
Les ingrédients \ref{ingredient iterer beta} et \ref{ingredient iterer alpha} étudient le comportement des quantités lorsque les partitions sont itérées sous $W$.
Les deux autres permettront des arguments d'approximations de partitions quelconques par des partitions de la forme $\beta^{\vee W}$.
Le point~\ref{ingredient partitions dans sous-algebre} notamment est rassurant puisqu'il autorisera à effectuer les calculs dans une sous-algèbre de parties.
\\
Les ingrédients \ref{ingredient iterer beta} et \ref{ingredient : ineg de Rokhlin} suffisent\footnote{Et c'est désormais le même genre d'argument que pour le théorème de Kolmogorov-Sina{\u\i}.}
à montrer que les partitions finies génératrices réalisent l'entropie sofique (le \emph{supremum} des $h_{\mathrm{mes}}^{\Sigma}(\xi)$).
En effet si $\beta$ est génératrice, alors pour toute partition $\xi$, l'entropie relative $H(\xi \vert \beta^{\vee W})$ peut être rendue aussi petite que l'on souhaite en choisissant $W$ partie finie assez grande de $G$. Alors,
\begin{equation*}
h_{\mathrm{mes}}^{\Sigma}(\xi)\leqslant h_{\mathrm{mes}}^{\Sigma}(\beta^{\vee W}) +H(\xi \vert \beta^{\vee W})=h_{\mathrm{mes}}^{\Sigma}(\beta) +H(\xi \vert \beta^{\vee W})\underset{''W\nearrow G''}{\longrightarrow}h_{\mathrm{mes}}^{\Sigma}(\beta).
\end{equation*}
Les ingrédients \ref{ingredient iterer alpha} et \ref{ingredient partitions dans sous-algebre} fournissent les outils qui permettront les calculs.
Si $\beta$ est génératrice, l'algèbre $\cup_{\substack{W\subset G \\ W \textrm{ finie}}} \beta^{\vee W}$ peut jouer le rôle de $\mathcal{S}$ dans \ref{ingredient partitions dans sous-algebre}.
Ainsi, pour toute $\alpha\geqslant \beta$, on peut trouver $W$ et $\beta^{\vee W} \geqslant \xi\geqslant \beta$ tels que $h_{\mathrm{mes}}^{\Sigma}(\beta, \xi) \leqslant h_{\mathrm{mes}}^{\Sigma}(\beta, \alpha)$ ; et alors
\begin{equation*}
\begin{array}{ccccccc}
h_{\mathrm{mes}}^{\Sigma}(\beta,\beta) \overset{(\ref{ingredient iterer alpha})}{=}h_{\mathrm{mes}}^{\Sigma}(\beta,\beta^{\vee W}) & \overset{(*)}{\leqslant} & h_{\mathrm{mes}}^{\Sigma}(\beta, \xi) &\overset{(\ref{ingredient partitions dans sous-algebre})}{\leqslant}& h_{\mathrm{mes}}^{\Sigma}(\beta, \alpha) & \overset{(*)}{\leqslant} & h_{\mathrm{mes}}^{\Sigma}(\beta,\beta), \end{array}
\end{equation*}
où les deux inégalités $(*)$ reposent sur la monotonie évidente~: si $\alpha'\geqslant \alpha$, alors $h_{\mathrm{mes}}^{\Sigma}(\beta,\alpha')\leqslant h_{\mathrm{mes}}^{\Sigma}(\beta, \alpha)$.
\subsection{Actions Bernoulli}\label{sect: ent sofique Bernoulli}
On donne des éléments de preuve permettant le calcul de l'entropie sofique mesurée des décalages de Bernoulli.
\begin{theo}[Bowen {\cite{Bowen-2010-mes-conj-inv-sofic}}]
Soit $G$ un groupe sofique infini et $\Sigma$ une approximation sofique de $G$.
Soit $K$ un ensemble fini, muni d'une mesure de probabilité $\nu$ qui charge chacun de ses points
et $G\curvearrowright (K^{G}, \nu^{\otimes G})$ l'action par décalage de Bernoulli associée. Alors,
\begin{equation}
h_{\mathrm{mes}}^{\Sigma}\left(G\curvearrowright K^{G}, \nu^{\otimes G}\right)=H(\nu).
\end{equation}
\end{theo}
Considérons la partition canonique $\alpha \colon K^{G}\to K$, $x\mapsto x(\mathrm{id}_{G})$.
Elle est génératrice et son entropie de Shannon vaut
$H(\alpha)=H(\nu)$.
On se restreint au cas où le groupe $G$ est résiduellement fini et où l'approximation sofique vérifie~:
\\
-a- les $\sigma_n \colon G\to \mathrm{Sym}(D_n)$ sont des homomorphismes, et
\\
-b- pour toute partie finie $F\subset G$, si $n$ est assez grand, alors \textbf{pour tout} $v\in D_n$ et tout $s,t\in F$, si $s\not=t$ alors
$\sigma_n(s).v\not= \sigma_n(t).v$.
D'après le Théorème \&\ Définition~\ref{th-def: ent sof mes Bowen}, on cherche à estimer le cardinal de l'ensemble (voir (\ref{def:U p alpha F}), (\ref{def:U p a F n}) et (\ref{eq:def M mu}))~:
\begin{equation}
\mathcal{M}_\mu(\alpha, F, \epsilon, \sigma_n)\overset{\mathrm{def}}{=}\left\{a \in K^{D_n}\ \Bigg\vert \ \sum_{p \in K^F}
\left| \mu\left(U_{p}(\alpha,F)\right) - \mathbf{u}_n\left(U_{p}(a,F,n)\right)
\right| < \epsilon\right\}.
\end{equation}
On le regarde comme partie de l'espace probabilisé $(K^{D_n}, \nu^{\otimes D_{n}})$.
On sait déjà que, pour cette mesure déjà considérée $\nu^{\otimes D_{n}}$, quand $\mathrm{id}_{G}\in F$ chaque atome est de mesure approximativement $\mathrm{e}^{-H(\nu) \vert D_n\vert}$ (voir section~\ref{sect: preuve du lem Boltzmann-Sanov}).
On va montrer, par une variante standard de la loi des grands nombres (avec dépendance limitée),
que
pour toute partie finie $F\subset G$ et pour tout $\epsilon>0$, avec une très grande probabilité lorsque $D_n$ est grand, tous les $a \in K^{D_n}$ vont convenir.
Plus précisément :
pour tout $p\in K^{F}$, pour tout $\epsilon>0$,
\begin{equation}
\nu^{\otimes D_{n}}\left(\left\{a \in K^{D_n}\ \vert\
\left| \mu(U_{p}(\alpha,F)) - \mathbf{u}_n(U_{p}(a,F,n))
\right|< \epsilon\right\}\right)\underset{n\to \infty}{\longrightarrow} 1.
\end{equation}
Fixons $p\in K^{F}$.
On cherche à estimer, pour chaque $a\in K^{D_n}$, le cardinal
\begin{equation}
\left\vert U_{p}(a,F,n)\right\vert =\left\vert \left\{{v}\in D_n\ \vert\ \ a^{\vee F}(v)=p\right\}\right\vert =\sum_{v\in D_n}
\mathbf{1}_{a^{\vee F}(v)=p},
\end{equation}
qu'on regarde comme variable aléatoire sur $(K^{D_n}, \nu^{\otimes D_{n}})$
\begin{equation}
Z \colon a\mapsto \sum_{v\in D_n} Z_{v}(a)
\ \ \textrm{ en posant } \ \ Z_{v}(a)\overset{\mathrm{def}}{=} \mathbf{1}_{a^{\vee F}(v)=p}.
\end{equation}
On va utiliser l'inégalité de Bienaymé-Tchebychev\footnote{Si $V$ une variable aléatoire d'espérance $m$ et de variance finie $\sigma^2$,
alors pour tout réel strictement positif $\delta$, on a~: $P\left(\left|V-m\right| \geqslant \delta \right) \leqslant \frac{\sigma^2}{\delta^2}$.}
qui majore la probabilité de l'écart à la moyenne à l'aide de la variance.
\noindent\textbf{Espérance de $Z$.}
\begin{equation}
\mathbb{E}_{a}(Z)= \mathbb{E}_{a}\left(\sum_{v\in D_n} Z_{v}(a)\right)=\sum_{v\in D_n} \mathbb{E}_{a}\left(Z_{v}(a)\right)=\vert D_n\vert \ \mu\left(U_p(\alpha, F)\right).
\end{equation}
En effet, pour $n$ suffisamment grand, pour chaque $v\in D_n$, les images $\sigma_{n}(f).v$ pour $f\in F$ sont deux à deux distinctes. Du coup $\mathbb{E}_{a}(Z_{v}(a))$, qui est la mesure pour $\nu^{\otimes D_n}$ du cylindre de $K^{D_n}$ dont les valeurs aux coordonnées $\sigma_n(f_i).v$ sont fixées à $p(f_i)$, vaut exactement $\mu(U_p(\alpha, F))$, \emph{i.e.} la mesure d'un cylindre analogue pour $\nu^{\otimes G}$.
\medskip
\noindent\textbf{Variance de $Z$.}
Si $v,w\in D_n$ sont deux points dont les $F$-parcours sont disjoints $\{\sigma_n(f) \cdot v: f\in F\}\cap \{\sigma_n(f) \cdot w: f\in F\}=\emptyset$, alors les cylindres qu'ils peuvent définir portant sur des coordonnées distinctes,
les variables aléatoires $Z_v$ et $Z_{w}$ sont indépendantes~: $\mathbb{E}_{a}(Z_{v}Z_{w})=\mathbb{E}_{a}(Z_{v})\ \mathbb{E}_{a}(Z_{w})$.
Le nombre de paires $(v,w)\in D_n^2$ pour lesquelles $Z_v$ et $Z_{w}$ ne sont pas indépendantes est donc inférieur à $\vert D_n\vert \vert F\vert^2$.
Ainsi puisque chacun des $ \mathbb{E}_{a}(Z_{v}\ Z_{w})\in [0,1]$, on a~:
\[\mathbb{E}_{a}(Z^2)=\sum_{(v,w)\in D_n} \mathbb{E}_{a}(Z_{v}\ Z_{w})\leqslant
\sum_{(v,w)\in D_n} \mathbb{E}_{a}(Z_{v})\ \mathbb{E}_{a}(Z_{w})+ \vert D_n\vert \vert F\vert^2.\]
Cela donne une majoration de la variance de~$Z$~:
\[\mathrm{Var}(Z)= \mathbb{E}_{a}(Z^2)- \mathbb{E}_{a}(Z)^2\leqslant \vert D_n\vert \vert F\vert^2.\]
L'inégalité de Bienaymé-Tchebychev donne alors~:
\begin{align*}
\mathbb{P}_{a} \left( \left\vert \frac{Z}{\vert D_n\vert} -\frac{\mathbb{E}_{a}(Z)}{\vert D_n\vert} \right\vert \geqslant \epsilon_1 \right) &\leqslant \frac{\vert F\vert ^2}{\epsilon_1 ^2 \vert D_n\vert}, \textrm{ \emph{i.e.} }
\\
\nu^{\otimes D_{n}}\left(\left\{a \in K^{D_n}\ \vert\
\left| \mu\left(U_{p}(\alpha,F)\right) - \mathbf{u}_n\left(U_{p}(a,F,n)\right)
\right|< \epsilon_1
\right\}\right)
&> 1-\frac{\vert F\vert ^2}{\epsilon_1 ^2 \vert D_n\vert}
\end{align*}
D'où avec $\epsilon_1=\frac{\epsilon}{\vert K\vert^{\vert F\vert}}$
\[
\nu^{\otimes D_{n}}\left(\mathcal{M}_\mu(\alpha, F, \epsilon, \sigma_n)\right) > 1- \frac{\vert F\vert ^2}{\epsilon ^2 \vert D_n\vert}.
\]
Alors exactement comme en section~\ref{sect: preuve du lem Boltzmann-Sanov}, l'estimée sur la mesure de $\mathcal{M}_\mu(\alpha, F, \epsilon, \sigma_n)$ et celle des atomes permet de conclure.
Pour tout $\delta>0$, il existe $\epsilon_0$ tel que pour tout $\epsilon\in]0, \epsilon_0[$, pour toute partie finie $F\subset G$ avec $\mathrm{id}_{G}\in F$, il existe $n_0$ tel que pour tout $n\geqslant n_0$ on a~:
\begin{equation}
\mathrm{e}^{(H(\alpha) -\delta)d}
< \vert \mathcal{M}_\mu(\alpha, F, \epsilon, \sigma_n)\vert
< \mathrm{e}^{(H(\alpha) +\delta)d}.
\end{equation}
Et donc
\begin{equation}
H(\nu)=H(\alpha)=\inf\limits_{\epsilon>0}\ \inf\limits_{\substack{F\subset G\\ \textrm{ fini}}}\ \limsup\limits_{n\to \infty} \frac{ \log\vert \mathcal{M}_\mu(\alpha, F, \epsilon, \sigma_n)\vert }{\vert D_n\vert}.
\end{equation}
\section{Dépendance en l'approximation sofique}\label{sect: depend-approx-sofic}
On montre des situations où la valeur de l'entropie sofique dépend de l'approximation sofique.
Soit $G$ un groupe $G$ avec la \textbf{propriété (T) de Kazhdan}. C'est-à-dire qu'il vérifie~:
$\forall \delta>0$, il existe une partie finie $F\subset G$ et $\epsilon>0$ tels que pour toute représentation unitaire $\pi$, si $\xi$ est un vecteur unitaire $(F,\epsilon)$-invariant (\emph{i.e.} $\Vert \pi(f) \xi-\xi\Vert ^2<\epsilon$), alors il existe un vecteur unitaire $G$-invariant $\xi_0$ tel que $\Vert \xi_0-\xi\Vert^2<\delta$.
Si $G$ a la propriété (T) de Kazhdan et est résiduellement fini, alors une approximation sofique $\Sigma$ associée à une chaîne $(G_n)_{n}$ de sous-groupes normaux d'indices finis n'est pas du tout encline à modéliser des actions non ergodiques de $G$.
Plus précisément, si $G\curvearrowright(X,\mu)$ se décompose en deux parties $G$-invariantes $X=X_1\sqcup X_2$ de mesures non nulles, alors pour toute partition $\alpha$ qui raffine cette décomposition, pour $\epsilon$ assez petit et $F$ assez grand on a $ \mathcal{M}_\mu(\alpha, F, \epsilon, \sigma_n)=\emptyset$, pour tout $n$, et donc~:
\[h_{\mathrm{mes}}^{\Sigma}(G\curvearrowright X,\mu)=-\infty.\]
Cela repose sur le fait suivant~: considérons une action transitive $\sigma \colon G\curvearrowright D$ sur un ensemble fini (par exemple sur $G_n\backslash G$). Pour $\delta>0$, si $A\subset D$ est $(F,\epsilon)$-invariante (pour $(F,\epsilon)$ donnés par la propriété de Kazhdan) au sens où pour tout $f^{-1}\in F$
\begin{equation*}
\vert f^{-1}A\Delta A\vert <\epsilon \vert A\vert,
\end{equation*}
alors $A$ occupe une grande proportion de $D$~:
\begin{equation*}
\vert A\vert > (1-\delta/2)^2 \ \vert D\vert.
\end{equation*}
En effet, la représentation unitaire associée sur $\ell^2(D)$ définie pour $g\in G, \xi \in \ell^2(D)$ et $v\in D$ par $(\pi(g).\xi )(v)=\xi(\sigma(g). v)$ possède deux vecteurs unitaires $G$-invariants~:~$\pm \frac{\mathbf{1}_{D}}{\sqrt{\vert D\vert}}$.
Puisque $\Vert \mathbf{1}_{A}-\mathbf{1}_{B}\Vert^2=\vert A\Delta B\vert$ pour toutes parties $A,B\subset D$, on~a~:
$
\vert f^{-1}A\Delta A\vert =\Vert \mathbf{1}_{A}-\mathbf{1}_{f^{-1}(A)}\Vert^2=\Vert \mathbf{1}_{A}-\pi(f^{-1}).\mathbf{1}_{A}\Vert^2<\epsilon \Vert \mathbf{1}_{A}\Vert^2 =\epsilon \vert A\vert.
$
La propriété (T) montre alors que $\Vert \frac{\mathbf{1}_{D}}{\sqrt{\vert D\vert}}-\frac{\mathbf{1}_{A}}{\sqrt{\vert A\vert}} \Vert^2<\delta$ soit $\frac{\sqrt{\vert A\vert}}{\sqrt{\vert D\vert}}=
\langle \frac{\mathbf{1}_{D}}{\sqrt{\vert D\vert}},\frac{\mathbf{1}_{A}}{\sqrt{\vert A\vert}} \rangle
> 1-\delta/2$.
Admettons maintenant que les restrictions de notre action $G\curvearrowright (X,\mu)$ à $X_1$ et $X_2$ sont ergodiques (par exemple des décalages de Bernoulli de base finie) et disons que $(\mu(X_1), \mu(X_2))=(2/5,3/5)$.
L'approximation sofique $\Sigma'$ consistant en cinq copies de $\Sigma$, \emph{i.e.} $D'_{n}=D_{n}\times\{1,2,3, 4, 5\}$ et $\sigma_{n}' \colon G\to \mathrm{Sym}(D_{n}\times\{1,2,3, 4, 5\})$ est induite par $\sigma_{n}$ sur la première coordonnée, sera elle bien adaptée.
Pour $\epsilon$ assez petit et $F$ assez grand, les bons modèles pour $G\curvearrowright (X_1\sqcup X_2,\mu_{1}+\mu_{2})$ vont se décomposer en bons modèles pour $G\curvearrowright (X_1,\mu_1)$ (sur deux des $D_{n}\times \{i\}$) et pour
$G\curvearrowright (X_2,\mu_2)$ (sur les trois autres $D_{n}\times \{j\}$). Leur nombre sera approximativement~:
\[ C_5^2\,\left(\mathrm{e}^{\vert D_n\vert \,h_{\mathrm{mes}}^{\Sigma} (G\curvearrowright X_1, \mu_1)}\right)^{2} \left(\mathrm{e}^{\vert D_n\vert \,h_{\mathrm{mes}}^{\Sigma} (G\curvearrowright X_2, \mu_2)}\right)^{3}.\]
Et finalement pour cette deuxième approximation sofique $\Sigma'$, on a~:
\begin{eqnarray*}
h_{\mathrm{mes}}^{\Sigma'}\left(G\curvearrowright X, \mu\right)=\frac{2}{5} h_{\mathrm{mes}}^{\Sigma} \left(G\curvearrowright X_1, \mu_1\right) + \frac{3}{5} h_{\mathrm{mes}}^{\Sigma} \left(G\curvearrowright X_2, \mu_2\right),
\end{eqnarray*}
dont on peut facilement prescrire des valeurs finies $h_{\mathrm{mes}}^{\Sigma} \left(G\curvearrowright X_i, \mu_i\right)=t_i$ pour des décalages de Bernoulli.
Observons qu'on n'a pas utilisé toute la force de la propriété (T), mais seulement la propriété $(\tau)$
(voir le livre de A.~Lubotzky \cite{Lubotzky=book-exp-graph-1994}) du groupe $G$ vis-à-vis de sa suite de sous-groupes d'indice fini $(G_n)_{n}$ ; plus précisément que la représentation unitaire
$G\curvearrowright \oplus_n \ell^{2}_{0}(G_n\backslash G)$ ne contient pas faiblement la représentation triviale, où $\ell^{2}_{0}(G_n\backslash G)$ est l'orthogonal des fonctions constantes.
En d'autres termes, on a utilisé le caractère \textbf{expanseur} des graphes de Schreier $\mathcal{G}_n$ de la section~\ref{sect:groupes sofiques}.
Le théorème suivant de A.~Carderi est frappant dans ce sens qu'il met en évidence une forme de rigidité dans l'adéquation chaîne sofique/action.
\begin{theo}[Carderi {\cite[Th. D]{Carderi-2015-arxiv}}]
Soit $G$ un groupe libre ou bien $\mathrm{PSL}_{r}(\mathbf{Z})$ pour $r\geqslant 2$.
Il existe un continuum de chaînes normales $(H_n^{t})_{n\in \mathbf{N}}$ de sous-groupes de $G$ (indexées par $t\in \mathbf{R}$) telles que les entropies sofiques mesurées et topologiques, relativement à l'approximation sofique $\Sigma_t$ associée à la chaîne $(H_n^{t})_{n\in \mathbf{N}}$, de l'action profinie $G\curvearrowright \left(\varprojlim(H_n^{s})_{n}, \mu_s\right)$ associée à la chaîne $(H_n^{s})_{n\in \mathbf{N}}$ (avec son unique mesure invariante $\mu_s$) vérifient
\[
h_{\mathrm{top}}^{\Sigma_t}(G\curvearrowright \varprojlim(H_n^{s})_{n})=h_{\mathrm{mes}}^{\Sigma_t}(G\curvearrowright \varprojlim(H_n^{s})_{n},\mu_s)=\begin{cases} 0 & \textrm{ si } t=s,\\ -\infty & \textrm{ si } t\not=s. \end{cases}\]
\end{theo}
Le résultat repose via \cite{Abert-Elek=profinite-act-2012} sur la propriété $(T)$ de Kazhdan ou sur le fait que les sous-groupes de congruence $N_i$ de $\mathrm{PSL}_{2}(\mathbf{Z})$ ont la propriété $(\tau)$.
Les familles de sous-groupes $H_n=N_{i_1}\cap N_{i_2}\cap\cdots\cap N_{i_n}$ associées à des suites infinies incomparables (au sens de l'inclusion) d'indices $I=\{i_1, i_2, i_3, \cdots\}\subset \mathbf{N}$ conviendront.
Notons que pour les actions profinies générales $G \curvearrowright \varprojlim(G_n)_{n}$, l'entropie sofique topologique (ou mesurée pour l'unique mesure invariante) prend ses valeurs dans $\{-\infty, 0\}$.
\section{Miscellanées}
\subsection{Un peu plus sur l'exemple d'Ornstein-Weiss}
\label{sect:plus sur l'ex de OW}
Soit $\mathbf{L}_r$ le groupe libre à $r$ générateurs engendré par $S=\{a_1, a_2, \cdots, a_r\}$ et soit $\mathbb{K}$ un corps fini.
Soit $\mathcal{G}$ le graphe (l'arbre) de Cayley associé. L'application
\begin{equation*}
\Theta_{r} \colon \left(\begin{array}{ccl}
\mathbb{K}^{\mathbf{L}_r}&\to & (\mathbb{K}^{r})^{\mathbf{L}_r}\\
\omega &\mapsto &
(\omega(g a_1)-\omega(g), \omega(g a_2)-\omega(g), \cdots, \omega(g a_r)-\omega(g))_{g\in \mathbf{L}_r}
\end{array}\right)
\end{equation*}
de la section~\ref{sect: appl Ornstein-Weiss} (\ref{eq: appl OW}) s'interprète comme l'application linéaire cobord \[\delta^{1} \colon C^{0}(\mathcal{G}, \mathbb{K})\to C^{1}(\mathcal{G}, \mathbb{K})\] entre les cochaînes en dimension $0$ et les cochaînes en dimension $1$ à coefficients dans $\mathbb{K}$.
On constate facilement qu'elle est surjective.
L'entropie sofique topologique (ou mesurée pour les mesures de Haar) vaut $\log \vert \mathbb{K}\vert$ pour $\mathbb{K}^{\mathbf{L}_r}$, resp. $r\,\log \vert \mathbb{K}\vert$ pour $\mathrm{im}\, \delta^{1}$.
Les approximations sofiques $\Sigma=(\sigma_n \colon \mathbf{L}_r\to \mathrm{Sym}(D_n))_{n}$ du groupe libre $\mathbf{L}_r$ correspondent à des graphes $\mathcal{G}_{n}=(D_n, ([v, \sigma_n(s).v])_{v\in D_n, s\in S})$ dont le tour de taille (longueur minimale de cycles) tend vers l'infini avec $n$. Chaque sommet est de valence $2r$ et est le sommet initial et le sommet terminal d'une arête étiquetée $a_i$ pour chaque $i=1, 2, \cdots, r$.
On peut considérer l'application cobord
\begin{equation}
\label{eq: appl cobord}
\begin{array}{rccc}
\delta^{1}_{n} \colon &C^{0}(\mathcal{G}_n, \mathbb{K})&\to& C^{1}(\mathcal{G}_n, \mathbb{K})
\\
&\simeq \mathbb{K}^{\vert D_n \vert} & & \simeq \mathbb{K}^{r\,\vert D_n \vert}
\end{array}
\end{equation}
sur ces graphes.
Des considérations de dimensions, avec $\dim_{\mathbb{K}}C^{0}(\mathcal{G}_n, \mathbb{K})=\vert D_n\vert$, $\dim_{\mathbb{K}}C^{1}(\mathcal{G}_n, \mathbb{K})=r\,\vert D_n\vert$ et $\dim_{\mathbb{K}}\ker \delta^{1}_{n}=1$,
nous indiquent par le théorème du rang ($\dim_{\mathbb{K}}\mathrm{im}\, \delta^{1}_{n}= \vert D_n\vert -1$) que l'image est loin d'être surjective !
On a vu que dans $C^{0}(\mathcal{G}_n, \mathbb{K})\simeq \mathbb{K}^{\vert D_n \vert}$, la plupart des points (au sens de la mesure de Haar) fournissent de bons modèles pour le calcul de l'entropie sofique de $G\curvearrowright \mathbb{K}^G$.
En revanche, leurs images par $\delta^{1}_{n}$ sont en nombre insuffisant pour représenter tous les bons modèles de $G\curvearrowright (\mathbb{K}^r)^{G}$.
Si l'approximation sofique est donnée par une chaîne de sous-groupes normaux $(G_n)_{n}$ d'indice fini, on peut tester le Théorème~\ref{th: Gab-Sew ent et pt fixes}~:
les points fixes de $G_n\curvearrowright C^{0}(\mathcal{G}, \mathbb{K})$ et de $G_n\curvearrowright C^{1}(\mathcal{G}, \mathbb{K})$ sont précisément les relevés de
$C^{0}(G_n\backslash\mathcal{G}, \mathbb{K})$ et de $C^{1}(G_n\backslash\mathcal{G}_n, \mathbb{K})$ tandis que $\delta^{1}\left(\mathrm{Fix}_{G_n}C^{0}(\mathcal{G}, \mathbb{K})\right)$ est le relevé de
$\mathrm{im}\,\delta^{1}_{n}$~: bien trop \emph{petit}.
Observons que ces applications $\delta^{1}, \delta^{1}_{n}$ de (\ref{eq: appl cobord}) se généralisent à tout groupe avec $r$ générateurs\footnote{Mais $\delta^{1}$ n'est pas surjective en général.} pour le graphe de Cayley $\mathcal{G}$ associé, et qu'en vérité, l'action image s'identifie à
\[G\curvearrowright C^{0}(\mathcal{G}, \mathbb{K})/\ker \delta^{1}\simeq \mathbb{K}^{G}/\mathbb{K}\] et ne dépend donc pas du système générateur.
Le \og défaut de commutation dans le diagramme~: pousser par $\delta$, puis prendre les modèles finis ou prendre d'abord les modèles finis, puis pousser par $\delta$\fg\
a été exploité dans \cite[Th. 9.4]{Gab-Seward-2015-arxiv} pour interpréter la croissance de l'entropie sofique par le facteur $\delta^{1} \colon \mathbb{K}^{G}\to \mathbb{K}^{G}/\mathbb{K}$ en termes de \emph{coût\footnote{On renvoie à \cite{Gab00a} pour cette notion.} du groupe $G$} ou en termes de $\beta_{(2)}^{1}(G)$, son \emph{premier nombre de Betti} $\ell^2$, via le théorème d'approximation de L\"uck et sa généralisation aux approximations sofiques \cite{Luc94b,Thom=Diophantine-approx=2008}. En particulier, lorsque $G$ est de type fini, $\delta_1$ fait croître l'entropie sitôt que le premier nombre de Betti $\ell^2$ de $G$ est non nul~:
\[(1+\beta_{(2)}^{1}(G))\, \log\vert \mathbb{K}\vert \leqslant h_{\mathrm{mes}}^{\Sigma}(G\curvearrowright \mathbb{K}^{G}/\mathbb{K})=h_{\mathrm{top}}^{\Sigma}(G\curvearrowright \mathbb{K}^{G}/\mathbb{K}).\]
La question du caractère Bernoulli\footnote{Est-elle conjuguée à un décalage de Bernoulli ?} de l'image est très largement ouverte.
Lorsque $G$ est moyennable, les facteurs des Bernoulli sont des Bernoulli \cite{OW87}. En revanche, des travaux de S.~Popa et R.~Sasyk \cite{PS07,Popa-1-cohomology-2006} montrent que pour un groupe infini avec la propriété (T), l'action $G\curvearrowright \mathbb{K}^{G}/\mathbb{K}$ n'est pas Bernoulli. Cela passe par le calcul explicite du premier groupe de cohomologie de ces actions (égal au groupe fini $\mathrm{Char}(G)$ des caractères de $G$ pour Bernoulli~; isomorphe à $\mathrm{Char}(G)\times \mathbb{K}$ pour le quotient).
\subsection{Produits}
L'entropie de Kolmogorov-Sina{\u\i} est additive sous produits cartésiens.
Considérons deux actions p.m.p. $G \curvearrowright^{\! T}\! (X,\mu)$ et $G \curvearrowright^{\! S}\! (Y,\nu)$.
Si $G=\mathbf{Z}$, alors
$$h_{\mathrm{KS}}(G \curvearrowright^{\! T\times S}\! X\times Y,\mu\times \nu)=h_{\mathrm{KS}}(G \curvearrowright^{\! T}\! X,\mu) + h_{\mathrm{KS}}(G \curvearrowright^{\! S}\! Y,\nu).$$
{\c C}a n'est plus vrai pour l'entropie sofique. Tim Austin \cite{Austin-2015-Add-prod-sofic} a montré l'inégalité
$$h_{\mathrm{mes}}^{\Sigma}(G \curvearrowright^{\! T\times S}\! X\times Y,\mu\times \nu)\leqslant h_{\mathrm{mes}}^{\Sigma}(G \curvearrowright^{\!T}\! X,\mu) + h_{\mathrm{mes}}^{\Sigma}(G \curvearrowright^{\! S}\! Y,\nu),$$
et donné des contre-exemples à l'égalité.
Cependant, si l'un des deux facteurs est un décalage de Bernoulli, d'entropie de Shannon de base finie, alors on a égalité
\cite[Th. 8.1]{Bowen-2010-mes-conj-inv-sofic}.
\subsection{Actions non libres}
Soit $G\curvearrowright (X,\mu)$ une action p.m.p. ergodique d'un groupe sofique et $\Sigma$ une approximation sofique.
Si $h_{\mathrm{mes}}^{\Sigma}(G\curvearrowright X, \mu)>0$, alors le stabilisateur de $\mu$-presque tout point est fini \cite{Meyerovitch-2015-ent-stabilizer}.
L'entropie de Rokhlin n'a aucune prise sur ce genre de question, puisqu'elle ne voit pas les stabilisateurs.
\subsection{Actions algébriques et déterminant de Fuglede-Kadison}
Un autre thème récurrent en théorie classique de l'entropie consiste à relier l'entropie de systèmes dynamiques d'origine algébrique aux valeurs propres d'une l'application linéaire sous-jacente.
Par exemple, pour un automorphisme $\phi$ linéaire hyperbolique\footnote{Les valeurs propres sont de module $\not=1$.} du tore $\mathbf{R}^{n}/\mathbf{Z}^{n}$
$$h_{\mathrm{top}}(G\curvearrowright X_{\phi})=h_{\mathrm{KS}}(G\curvearrowright X_{\phi}, \mathrm{Haar})=\log{ \det}^{+} (\Phi),$$
où ${\det}^{+} (\Phi)$ est le produit des valeurs propres de module $>1$.
De plus, la mesure de Haar est l'unique mesure borélienne d'entropie maximale.
Considérons un élément $\phi=\sum_{h\in G} \phi_{h} h$ dans l'anneau entier $\mathbf{Z}[G]$ d'un groupe dénombrable $G$ et le quotient $\mathbf{Z}[G]/\mathbf{Z}[G] \phi$ par l'idéal à gauche engendré par $\phi$. Le dual de Pontryagin de ce groupe abélien discret est un groupe abélien compact muni d'une action continue par automorphismes de groupe, induite par la multiplication à gauche
\[G\curvearrowright X_{\phi}\overset{\mathrm{def}}{=}\widehat{(\mathbf{Z}[G]/\mathbf{Z}[G] \phi)}.\]
C'est le fermé $G$-invariant du décalage de Bernoulli $\widehat{\mathbf{Z}[G]}=\left( \mathbf{R}/\mathbf{Z}\right)^{G}$ formé des suites
\[X_{\phi}=\{(x_{g})_{g}\in \left( \mathbf{R}/\mathbf{Z}\right)^{G} \vert \sum_{h\in G} \phi_{h} x_{gh}=0, \forall g\in G\}.\]
On a toute une série de travaux qui permettent d'exprimer l'entropie topologique d'une telle action.
S.~Juzvinski{\u\i} \cite{Jusvinskii-1967-ent-gp-endom} dans le cas de $G=\mathbf{Z}$, l'exprime comme logarithme du produit des racines de module $\geqslant 1$ de $\Phi$.
Le lemme de Mahler \cite{Mahler-1960-app-Jensen-Form,Mahler-1962-ineq-sev-var} permet d'interpréter un tel produit en analyse complexe comme une intégrale (\og mesure de Mahler logarithmique \fg). D.~Lind, K.~ Schmidt et T.~Ward \cite{Lind-Schmidt-Ward-1990-Mahler-mes} exprimeront l'entropie dans le cas de $G=\mathbf{Z}^p$ en ces termes, considérant $\Phi$ comme polynôme de Laurent
\begin{equation}
h_{\mathrm{top}}(\mathbf{Z}^p\curvearrowright X_{\Phi})=\int_{(\mathbf{R}/\mathbf{Z})^p} \log\vert \Phi(\mathrm{e}^{2\pi \mathrm{i}\theta})\vert \, d\theta.
\end{equation}
C.~Deninger \cite{Deninger-2006-Fuglede-Kadison-det} observe que ces quantités admettent des généralisations dans le cadre non commutatif à l'aide du déterminant de Fuglede-Kadison. Il s'agit d'un objet d'analyse fonctionnelle, concocté à l'aide du calcul fonctionnel et de la trace de von Neumann défini sur des opérateurs $u$ de
$G$-modules de Hilbert et qui joue le rôle du déterminant positif classique ${\det}^{+}$~:
\[ {\det}^{+}_{\mathrm{vN}(G)} (u) \overset{\mathrm{def}}{=} \exp\left(\int_{]0,\infty[} \log(t) \, d\lambda_{\vert u\vert} (t)\right).\]
où $\lambda_{\vert u\vert}$ représente la fonction de densité spectrale de l'opérateur ${\vert u\vert}$.
Le domaine de validité de l'égalité entre l'entropie topologique et ce déterminant positif est peu à peu étendu, sous des hypothèses plus ou moins fortes sur $\phi$ (positivité, diverses formes d'inversibilité,...), à des classes de plus en plus grandes de groupes (croissance polynomiale, moyennable résiduellement fini, ...)
\cite{Deninger-2006-Fuglede-Kadison-det, Deninger-Schmidt-2007-exp-alg-act-entropy,Li-2012-Fuglede-Kadison-det, Li-Thom-2014-entropy-det-L2-torsion}
jusqu'à atteindre une forme optimale\footnote{En effet, puisque $\det^{+}_{\mathrm{vN}(G)} (\phi)$ est fini, une égalité ne sera envisageable que lorsque $h_{\mathrm{top}}^{\Sigma}(G\curvearrowright X_{\phi})<\infty$.}, due à B.~Hayes.
Ces travaux montrent également que la mesure de Haar maximise l'entropie mesurée.
\begin{theo}
\cite[Th. 1.1]{Hayes-2014-Fuglede-Kadison-sofic-ent}
Soit $G$ un groupe dénombrable sofique et $\Sigma$ une approximation sofique de $G$.
Soit $\phi\in \mathrm{Mat}_{p,q} (\mathbf{Z}[G])$ et $G\curvearrowright X_{\phi}$ l'action algébrique associée.
On a les propriétés suivantes.
\begin{itemize}
\item [(i)] L'entropie sofique topologique $h_{\mathrm{top}}^{\Sigma}(G\curvearrowright X_{\phi})$ est finie si et seulement si $\phi$ est injective comme opérateur $\ell^{2}(G)^{\oplus q}\to \ell^{2}(G)^{\oplus p}$.
\end{itemize}
Supposons que $\phi$ est injective comme opérateur $\ell^{2}(G)^{\oplus q}\to \ell^{2}(G)^{\oplus p}$.
\begin{itemize}
\item[(ii)] Si $p=q$,
alors
$h_{\mathrm{top}}^{\Sigma}(G\curvearrowright X_{\phi})=h_{\mathrm{mes}}^{\Sigma}(G\curvearrowright X_{\phi}, \mathrm{Haar})=\log \det^{+}_{\mathrm{vN}(G)} (\phi)$.
\item[(iii)] Si $p\not =q$, alors
$h_{\mathrm{top}}^{\Sigma}(G\curvearrowright X_{\phi})\leqslant h_{\mathrm{mes}}^{\Sigma}(G\curvearrowright X_{\phi}, \mathrm{Haar}) \leqslant \log \det^{+}_{\mathrm{vN}(G)} (\phi)$.
\end{itemize}
\end{theo}
\subsection{Entropie d'Abért-Weiss}
Le calcul de l'entropie sofique mesurée des décalages de Bernoulli (voir section~\ref{sect: ent sofique Bernoulli}) fait intervenir une mesure auxiliaire judicieusement choisie sur l'ensemble des partitions $K^{D_n}$ avant comptage, dans un esprit de \og grande déviation\fg.
Cette idée a conduit M.~Abért et B.~Weiss (communication personnelle de M.~Abért et annonce \cite{Weiss2015-survey-sofic-ent})
à une approche un peu différente de l'entropie sofique mesurée, dont on donne ici quelques éléments.
Soit $G\curvearrowright (X,\mu)$ une action p.m.p. du groupe sofique $G$ et $\alpha \colon X\to K$ une partition mesurée finie génératrice.
Pour chaque partie finie $F\subset G$, on cherche à imiter la mesure poussée en avant $\mu_{F}=\alpha^{\vee F}_{*} \mu$ sur $K^{F}$.
Ce qu'on a fait jusqu'ici consistait, partant d'une application
\begin{equation}
\Upsilon \colon \left(\begin{array}{ccl}
D\times K^{D} &\to &K^F
\\
({v}, a) &\mapsto & (a(\sigma(f)\cdot {v}))_{f\in F}
\end{array}\right)
\end{equation}
à considérer, pour chaque partition $a\in K^{D}$ de $D$, la mesure poussée en avant
$\Upsilon(\cdot, a)_{*}\mathbf{u} $
sur $K^{F}$ de la mesure uniforme sur $D$, puis à compter les bons $a$ (ceux pour lesquels cette mesure est proche de $\mu_{F}$).
La démarche de M.~Abért et B.~Weiss consiste à considérer des mesures de probabilité $\nu\in M(K^{D})$ sur l'ensemble fini $K^{D}$. Elles ont chacune une certaine entropie de Shannon $H(\nu)$ qui aura tendance à croître sous-linéairement en $\vert D\vert$, d'où la pertinence d'une normalisation $\frac{H(\nu)}{\vert D\vert}$. Pour chaque point $v\in D$, ils considèrent la mesure poussée en avant $\nu_{v,F}=\Upsilon(v, \cdot)_{*}\nu$. C'est une mesure sur $K^{D}$ qu'ils comparent avec $\mu_{F}$ pour la norme $\ell^1$, en moyenne sur $D$. Et ils définissent une notion d'entropie
qui satisfait elle aussi les conditions (a) et (b).
\begin{equation}
h_{\mathrm{AW}}(G\curvearrowright X,\mu)\overset{\mathrm{def}}{=}\inf_{\epsilon > 0} \ \ \inf_{\substack{F \subset G\\F \text{ fini}}}
\sup \frac{H(\nu)}{\vert D\vert},
\tag{\textbf{Entropie selon Abért-Weiss}}
\end{equation}
où le \emph{supremum} est pris sur toutes les $(F,\epsilon)$-approximations sofiques $\Sigma=(\sigma \colon G\to \mathrm{Sym}(D))$ et pour toutes les mesures de probabilité $\nu\in M(K^{D})$ telles que $\frac{1}{\vert D\vert} \sum_{v\in D} \Vert \nu_{v,F}-\mu_{F}\Vert_1 <\epsilon$.
Il peuvent montrer que $h_{\mathrm{AW}}(G\curvearrowright X,\mu)\leqslant h_{\mathrm{mes}}^{\Sigma}(G\curvearrowright X,\mu)$ et L.~Bowen a exhibé des exemples\footnote{Qui encore une fois reposent sur la propriété $(\tau)$, voir section~\ref{sect: depend-approx-sofic}.} où on a une inégalité stricte.
\section{Remerciements}
Je veux exprimer ma gratitude à Alessandro Carderi et Mika\"el de la Salle dont l'aide m'a été très précieuse pour comprendre certaines références et pour la prépa\-ra\-tion et des relectures de ce texte.
Je tiens à remercier chaleureusement Mikl\'os Abért, Lewis Bowen, David Kerr et Brandon Seward qui m'ont initié à l'entropie sofique, m'ont fait part de leur vision, et ont répondu à mes nombreuses questions.
David Kerr et Hanfeng Li m'ont permis d'accéder à une version préliminaire du chapitre 9 \og {\em Entropy for actions of sofic groups} \fg\ de leur livre en préparation \cite{Kerr-Li-book}, je leur en suis reconnaissant.
Merci également à toutes les personnes m’ayant signalé diverses coquilles, omissions ou imprécisions, notamment Pierre-Emmanuel Caprace, Véronique Chen Ai Ti, \'Etienne Ghys, Julien Melleray,
Romain Tessera ainsi qu'à toutes les autres personnes avec qui j'ai pu discuter de ces sujets, parmi lesquelles Tim Austin, Tullio Ceccherini-Silberstein, Grégory Miermont, Jean-François Quint,...
\thanks{Travail soutenu par le C.N.R.S. et par le projet
ANR-14-CE25-0004 GAMME.}
\bibliographystyle{alpha}
|
\section{Introduction}
As is well known, the Fibonacci numbers are given by
the numbers in the following integer sequnce:
\[
1,1,2,3,5,8,13,21,34,55,89,144,\dots.
\]
The sequence $F_{n}$ of Fibonacci numbers is defined by the recurrence
relation as follows:
\begin{equation}
F_{0}=1,\quad F_{1}=1,\quad F_{n}=F_{n-1}+F_{n-2},\quad\left(n\ge2\right),\quad\left(\text{see \cite{key-1,key-2,key-3,key-4,key-5,key-6,key-7,key-8}}\right).\label{eq:1}
\end{equation}
The sequence can be extended to negative index $n$ arising from the
re-arranged recurrence relation
\begin{equation}
F_{n-2}=F_{n}-F_{n-1},\quad\left(\text{see \cite{key-1,key-2,key-3,key-4,key-5,key-6,key-7,key-8,key-9,key-10,key-11,key-12,key-13}}\right)\label{eq:2}
\end{equation}
which yields the sequence of ``negafibonacci'' numbers satisfying
\begin{equation}
F_{-n}=\left(-1\right)^{n+1}F_{n},\quad\left(\text{see \cite{key-11,key-12}}\right).\label{eq:3}
\end{equation}
It is well known that the generating function of Fibonacci numbers
is given by
\begin{equation}
\frac{1}{1-t-t^{2}}=\sum_{n=0}^{\infty}F_{n}t^{n},\quad\left(\text{see \cite{key-3,key-4,key-5,key-6}}\right).\label{eq:4}
\end{equation}
The convolved Fibonacci numbers $p_{n}\left(x\right)$, $\left(n\ge0\right)$,
are defined by the generating function
\begin{equation}
\left(\frac{1}{1-t-t^{2}}\right)^{x}=\sum_{n=0}^{\infty}p_{n}\left(x\right)\frac{t^{n}}{n!},\quad\left(x\in\mathbb{R}\right).\label{eq:5}
\end{equation}
From (\ref{eq:4}) and (\ref{eq:5}), we note that
\begin{equation}
\frac{p_{n}\left(1\right)}{n!}=F_{n},\quad\left(n\ge0\right).\label{eq:6}
\end{equation}
In this paper, we present a new approach to the convolved
Fibonacci numbers arising from the generating function of them and give
some new and explicit identities for the convolved Fibonacci numbers.
\section{Convolved Fibonacci numbers and their applications}
From (\ref{eq:51}), we note that
\begin{align}
\sum_{n=0}^{\infty}p_{n}\left(x\right)\frac{t^{n}}{n!} & =\left(\frac{1}{1-t-t^{2}}\right)^{x}=\left(\frac{1}{1-t-t^{2}}\right)\left(\frac{1}{1-t-t^{2}}\right)^{x-1}\label{eq:7}\\
& =\left(\sum_{l=0}^{\infty}p_{l}\left(1\right)\frac{t^{l}}{l!}\right)\left(\sum_{m=0}^{\infty}p_{m}\left(x-1\right)\frac{t^{m}}{m!}\right)\nonumber \\
& =\sum_{n=0}^{\infty}\left(\sum_{l=0}^{n}\binom{n}{l}p_{l}\left(1\right)p_{n-l}\left(x-1\right)\right)\frac{t^{n}}{n!}.\nonumber
\end{align}
By comparing the coefficients on both sides of (\ref{eq:7}),
we obtain the following proposition.
\begin{prop}
\label{prop:1}For $n\ge0$, $x\in\mathbb{R}$, we have
\[
p_{n}\left(x\right)=\sum_{l=0}^{n}\binom{n}{l}p_{l}\left(1\right)p_{n-l}\left(x-1\right).
\]
\end{prop}
Let us take $x=r\in\mathbb{N}$. Then, by Proposition \ref{prop:1},
we get
\begin{align}
p_{n}\left(r\right) & =\sum_{l_{1}=0}^{n}\binom{n}{l_{1}}p_{l_{1}}\left(1\right)p_{n-l_{1}}\left(r-1\right)\label{eq:8}\\
& =\sum_{l_{1}=0}^{n}\sum_{l_{2}=0}^{n-l_{1}}\binom{n}{l_{1}}\binom{n-l_{1}}{l_{2}}p_{l_{1}}\left(1\right)p_{l_{2}}\left(1\right)p_{n-l_{1}-l_{2}}\left(r-2\right)\nonumber \\
& =\sum_{l_{1}=0}^{n}\sum_{l_{2}=0}^{n-l_{1}}\sum_{l_{3}=0}^{n-l_{1}-l_{2}}\binom{n}{l_{1}}\binom{n-l_{1}}{l_{2}}\binom{n-l_{1}-l_{2}}{l_{3}}p_{l_{1}}\left(1\right)p_{l_{2}}\left(1\right)\nonumber\\
&\relphantom{=}\times p_{l_{3}}\left(1\right)p_{n-l_{1}-l_{2}-l_{3}}\left(r-3\right)\nonumber \\
& \vdots\nonumber \\
& =\sum_{l_{1}=0}^{n}\sum_{l_{2}=0}^{n-l_{1}}\cdots\sum_{l_{r-1}=0}^{n-l_{1}-\cdots-l_{r-2}}\binom{n}{l_{1}}\binom{n-l_{1}}{l_{2}}\cdots\nonumber\\
&\relphantom{=}\times \binom{n-l_{1}-l_{2}-\cdots-l_{r-2}}{l_{r-1}}\nonumber\\
&\times \left(\prod_{k=1}^{r-1}p_{l_{k}}\left(1\right)\right)p_{n-l_{1}-l_{2}-\cdots-l_{r-1}}\left(1\right).\nonumber
\end{align}
Therefore, by (\ref{eq:8}), we obtain the following corollary.
\begin{cor}
\label{cor:2} For $r\in\mathbb{N}$ and $n\ge0$, we have
\begin{align*}
p_{n}\left(r\right)&=\sum_{l_{1}=0}^{n}\sum_{l_{2}=0}^{n-l_{1}}\cdots\sum_{l_{r-1}=0}^{n-l_{1}-\cdots-l_{r-2}}\binom{n}{l_{1}}\binom{n-l_{1}}{l_{2}}\cdots\\
&\relphantom{=}\times\binom{n-l_{1}-l_{2}-\cdots-l_{r-2}}{l_{r-1}}\left(\prod_{k=1}^{r-1}p_{l_{k}}\left(1\right)\right)p_{n-l_{1}-l_{2}-\cdots-l_{r-1}}\left(1\right).
\end{align*}
\end{cor}
We observe that
\begin{align}
\left(\frac{1}{1-t-t^{2}}\right)^{x} & =\left(\frac{1}{1-t-t^{2}}\right)^{r}\left(\frac{1}{1-t-t^{2}}\right)^{x-r}\label{eq:9}\\
& =\left(\sum_{l=0}^{\infty}p_{l}\left(r\right)\frac{t^{l}}{l!}\right)\left(\sum_{m=0}^{\infty}p_{m}\left(x-r\right)\frac{t^{m}}{m!}\right)\nonumber \\
& =\sum_{n=0}^{\infty}\left(\sum_{l=0}^{n}\binom{n}{l}p_{l}\left(r\right)p_{n-l}\left(x-r\right)\right)\frac{t^{n}}{n!}.\nonumber
\end{align}
Therefore, by (\ref{eq:51}) and (\ref{eq:9}), we obtain the following
theorem.
\begin{thm}
\label{thm:3} For $n\ge0$, $r\in\mathbb{N}$, we have
\[
p_{n}\left(x\right)=\sum_{l=0}^{n}\binom{n}{l}p_{l}\left(r\right)p_{n-l}\left(x-r\right)=\sum_{l=0}^{n}\binom{n}{l}p_{n-l}\left(r\right)p_{l}\left(x-r\right).
\]
\end{thm}
Let us take $x=r+1$ in Theorem \ref{thm:3}. Then, we have
\begin{align}
p_{n}\left(r+1\right) & =\sum_{l=0}^{n}\binom{n}{l}p_{n-l}\left(r\right)p_{l}\left(1\right)\label{eq:10}\\
& =\sum_{l=0}^{n}\binom{n}{l}p_{n-l}\left(r\right)l!\frac{p_{l}\left(1\right)}{l!}\nonumber \\
& =\sum_{l=0}^{n}\left(n\right)_{l}p_{n-l}\left(r\right)F_{l},\nonumber
\end{align}
where $\left(x\right)_{n}=x\left(x-1\right)\cdots\left(x-n+1\right)$,
$\left(n\ge1\right)$, $\left(x\right)_{0}=1$.
\begin{cor}
\label{cor:4} For $r\in\mathbb{N}$, $n\ge0$, we have
\[
p_{n}\left(r+1\right)=\sum_{l=0}^{n}\left(n\right)_{l}p_{n-l}\left(r\right)F_{l}.
\]
\end{cor}
Taking $r=1$ in Corollary \ref{cor:4}, we have
\begin{align}
p_{n}\left(2\right) & =\sum_{l=0}^{n}\left(n\right)_{l}p_{n-l}\left(1\right)F_{l}\label{eq:11}\\
& =\sum_{l=0}^{n}\left(n\right)_{l}\left(n-l\right)!\frac{p_{n-l}\left(1\right)}{\left(n-l\right)!}F_{l}\nonumber \\
& =n!\sum_{n=0}^{n}\binom{n}{l}\binom{n}{l}^{-1}F_{n-l}F_{l}\nonumber \\
& =n!\sum_{l=0}^{n}F_{l}F_{n-l}.\nonumber
\end{align}
Thus, by (\ref{eq:11}), we get
\begin{equation}
\frac{p_{n}\left(2\right)}{n!}=\sum_{l=0}^{n}F_{l}F_{n-l},\quad\left(n\ge0\right).\label{eq:12}
\end{equation}
From (\ref{eq:12}), we can also derive the following equation.
\begin{align}
p_{n}\left(3\right) & =\sum_{l_{1}=0}^{n}\left(n\right)_{l_{1}}p_{n-l_{1}}\left(2\right)F_{l_{1}}\label{eq:13}\\
& =\sum_{l_{1}=0}^{n}\left(n\right)_{l_{1}}\left(n-l_{1}\right)!\frac{p_{n-l_{1}}\left(2\right)}{\left(n-l_{1}\right)!}F_{l_{1}}\nonumber \\
& =n!\sum_{l_{1}=0}^{n}\sum_{l_{2}=0}^{n-l_{1}}F_{l_{1}}F_{l_{2}}F_{n-l_{1}-l_{2}}.\nonumber
\end{align}
Thus, by (\ref{eq:13}), we get
\begin{equation}
\frac{p_{n}\left(3\right)}{n!}=\sum_{l_{1}=0}^{n}\sum_{l_{2}=0}^{n-l_{1}}F_{l_{1}}F_{l_{2}}F_{n-l_{1}-l_{2}}.\label{eq:14}
\end{equation}
For $r=3$ in Corollary \ref{cor:4}, we have
\begin{align}
p_{n}\left(4\right) & =\sum_{l_{1}=0}^{n}\left(n\right)_{l_{1}}p_{n-l_{1}}\left(3\right)F_{l_{1}}\label{eq:15}\\
& =n!\sum_{l_{1}=0}^{n}\frac{p_{n-l_{1}}\left(3\right)}{\left(n-l_{1}\right)!}F_{l_{1}}\nonumber \\
& =n!\sum_{l_{1}=0}^{n}\sum_{l_{2}=0}^{n-l_{1}}\sum_{l_{3}=0}^{n-l_{1}-l_{2}}F_{l_{1}}F_{l_{2}}F_{l_{3}}F_{n-l_{1}-l_{2}-l_{3}}.\nonumber
\end{align}
From (\ref{eq:15}), we note that
\begin{equation}
\frac{p_{n}\left(4\right)}{n!}=\sum_{l_{1}=0}^{n}\sum_{l_{2}=0}^{n-l_{1}}\sum_{l_{3}=0}^{n-l_{1}-l_{2}}F_{l_{1}}F_{l_{2}}F_{l_{3}}F_{n-l_{1}-l_{2}-l_{3}}.\label{eq:16}
\end{equation}
Continuing this process, we have
\begin{equation}
\frac{p_{n}\left(r+1\right)}{n!}=\sum_{l_{1}=0}^{n}\sum_{l_{2}=0}^{n-l_{1}}\cdots\sum_{l_{r}=0}^{n-l_{1}-\cdots-l_{r-1}}F_{l_{1}}F_{l_{2}}\cdots F_{l_{r}}F_{n-l_{1}-l_{2}-\cdots-l_{r}},\label{eq:17}
\end{equation}
where $r\in\mathbb{N}$.
\begin{thm}
\label{thm:5} For $r\in\mathbb{N}$ and $n\ge0$, we have
\[
\frac{p_{n}\left(r+1\right)}{n!}=\sum_{l_{1}=0}^{n}\sum_{l_{2}=0}^{n-l_{1}}\cdots\sum_{l_{r}=0}^{n-l_{1}-\cdots-l_{r-1}}F_{l_{1}}F_{l_{2}}\cdots F_{l_{r}}F_{n-l_{1}-l_{2}-\cdots-l_{r}}.
\]
\end{thm}
Let
\begin{align}
F\left(t,x\right) & =\left(1-t-t^{2}\right)^{-x}\label{eq:18}\\
& =e^{-x\log\left(1-t-t^{2}\right)}.\nonumber
\end{align}
Then, by (\ref{eq:18}), we get
\begin{align}
F^{\left(1\right)}\left(t,x\right) & =\frac{dF}{dt}\left(t,x\right)\label{eq:19}\\
& =x\left(1+2t\right)\left(1-t-t^{2}\right)^{-x-1}\nonumber \\
& =x\left(1+2t\right)F\left(t,x+1\right),\nonumber
\end{align}
\begin{align}
F^{\left(2\right)}\left(t,x\right) & =\frac{dF^{\left(1\right)}}{dt}\left(t,x\right)\label{eq:20}\\
& =2xF\left(t,x+1\right)+\left\langle x\right\rangle _{2}\left(1+2t\right)^{2}F\left(t,x+2\right),\nonumber
\end{align}
where $\left\langle x\right\rangle _{n}=x\left(x+1\right)\cdots\left(x+n-1\right)$,
$\left(n\ge1\right)$, $\left\langle x\right\rangle _{0}=1$.
From (\ref{eq:20}), we note that
\begin{align}
F^{\left(3\right)}\left(t,x\right) & =\frac{dF^{\left(2\right)}}{dt}\left(t,x\right)\label{eq:21}\\
& =6\left\langle x\right\rangle _{2}\left(1+2t\right)F\left(t,x+2\right)+\left\langle x\right\rangle _{3}\left(1+2t\right)^{3}F\left(t,x+3\right).\nonumber
\end{align}
\begin{align}
F^{\left(4\right)}\left(t,x\right) & =\frac{dF^{\left(3\right)}}{dt}\left(t,x\right)\label{eq:22}\\
& =12\left\langle x\right\rangle _{2}F\left(t,x+2\right)+12\left\langle x\right\rangle _{3}\left(1+2t\right)^{2}F\left(t,x+3\right)\nonumber \\
& \relphantom =+\left\langle x\right\rangle _{4}\left(1+2t\right)^{4}F\left(t,x+4\right),\nonumber
\end{align}
\begin{align}
F^{\left(5\right)}\left(t,x\right) & =\frac{dF^{\left(4\right)}}{dt}\left(t,x\right)\label{eq:23}\\
& =60\left\langle x\right\rangle _{3}\left(1+2t\right)F\left(t,x+3\right)+20\left\langle x\right\rangle _{4}\left(1+2t\right)^{3}F\left(t,x+4\right)\nonumber\\
&\relphantom{=}+\left\langle x\right\rangle _{5}\left(1+2t\right)^{5}F\left(t,x+5\right)\nonumber
\end{align}
and
\begin{align}
F^{\left(6\right)}\left(t,x\right) & =\frac{dF^{\left(5\right)}}{dt}\left(t,x\right)\label{eq:24}\\
& =120\left\langle x\right\rangle _{3}F\left(t,x+3\right)+180\left\langle x\right\rangle _{4}\left(1+2t\right)^{2}F\left(t,x+4\right)\nonumber \\
& \relphantom =+30\left\langle x\right\rangle _{5}\left(1+2t\right)^{4}F\left(t,x+5\right)+\left\langle x\right\rangle _{6}\left(1+2t\right)^{6}F\left(t,x+6\right).\nonumber
\end{align}
Thus, we are led to put
\begin{align}
&\relphantom{=}F^{\left(N\right)}\left(t,x\right)=\left(\frac{d}{dt}\right)^{N}F\left(t,x\right)\label{eq:25}\\
&=\sum_{i=0}^{\left[\frac{N}{2}\right]}a_{i}\left(N\right)\left\langle x\right\rangle _{N-i}\left(1+2t\right)^{N-2i}F\left(t,x+N-i\right)\nonumber
\end{align}
where $N\in\mathbb{N}$.
Taking the derivatives of (\ref{eq:25}) with respect to $t$, we have
\begin{align}
F^{\left(N+1\right)}\left(t,x\right) & =\sum_{i=0}^{\left[\frac{N}{2}\right]}a_{i}\left(N\right)\left\langle x\right\rangle _{N-i}\left(1+2t\right)^{N-2i}F^{\left(1\right)}\left(t,x+N-i\right)\label{eq:26}\\
& \relphantom =+\sum_{i=0}^{\left[\frac{N}{2}\right]}a_{i}\left(N\right)\left\langle x\right\rangle _{N-i}2\left(N-2i\right)\left(1+2t\right)^{N-2i-1}F\left(t,x+N-i\right)\nonumber \\
& =\sum_{i=0}^{\left[\frac{N}{2}\right]}2\left(N-2i\right)a_{i}\left(N\right)\left\langle x\right\rangle _{N-i}\left(1+2t\right)^{N-2i-1}F\left(t,x+N-i\right)\nonumber \\
& \relphantom =+\sum_{i=0}^{\left[\frac{N}{2}\right]}a_{i}\left(N\right)\left\langle x\right\rangle _{N-i+1}\left(1+2t\right)^{N-2i+1}F\left(t,x+N-i+1\right)\nonumber \\
& =\sum_{i=1}^{\left[\frac{N}{2}\right]+1}2\left(N-2i+2\right)a_{i-1}\left(N\right)\left\langle x\right\rangle _{N-i+1}\nonumber\\
&\relphantom{=}\times\left(1+2t\right)^{N-2i+1}F\left(t,x+N-i+1\right)\nonumber \\
& \relphantom =+\sum_{i=0}^{\left[\frac{N}{2}\right]}a_{i}\left(N\right)\left\langle x\right\rangle _{N-i+1}\left(1+2t\right)^{N-2i+1}F\left(t,x+N-i+1\right).\nonumber
\end{align}
On the other hand, by replacing $N$ by $N+1$ in (\ref{eq:25}),
we get
\begin{equation}
F^{\left(N+1\right)}\left(t,x\right)=\sum_{i=0}^{\left[\frac{N+1}{2}\right]}a_{i}\left(N+1\right)\left\langle x\right\rangle _{N-i+1}\left(1+2t\right)^{N-2i+1}F\left(t,x+N-i+1\right).\label{eq:27}
\end{equation}
\emph{Case 1. }Let $N$ be an even number. Then we have
\begin{align}
& \sum_{i=1}^{\frac{N}{2}+1}2\left(N-2i+2\right)a_{i-1}\left(N\right)\left\langle x\right\rangle _{N-i+1}\left(1+2t\right)^{N-2i+1}F\left(t,x+N-i+1\right)\label{eq:28}\\
& +\sum_{i=0}^{\frac{N}{2}}a_{i}\left(N\right)\left\langle x\right\rangle _{N-i+1}\left(1+2t\right)^{N-2i+1}F\left(t,x+N-i+1\right)\nonumber \\
& =\sum_{i=0}^{\frac{N}{2}}a_{i}\left(N+1\right)\left\langle x\right\rangle _{N-i+1}\left(1+2t\right)^{N-2i+1}F\left(t,x+N-i+1\right).\nonumber
\end{align}
Comparing the coefficients on both sides of (\ref{eq:28}), we
get
\begin{align}
a_{0}\left(N+1\right) & =a_{0}\left(N\right),\label{eq:29}\\
a_{i}\left(N+1\right) & =2\left(N-2i+2\right)a_{i-1}\left(N\right)+a_{i}\left(N\right),\quad\left(1\le i\le\frac{N}{2}\right).\label{eq:30}
\end{align}
\emph{Case 2}. Let $N$ be an odd number. Then, we have
\begin{align}
& \sum_{i=1}^{\frac{N+1}{2}}2\left(N-2i+2\right)a_{i-1}\left(N\right)\left\langle x\right\rangle _{N-i+1}\left(1+2t\right)^{N-2i+1}F\left(t,x+N-i+1\right)\label{eq:31}\\
& +\sum_{i=0}^{\frac{N-1}{2}}a_{i}\left(N\right)\left\langle x\right\rangle _{N-i+1}\left(1+2t\right)^{N-2i+1}F\left(t,x+N-i+1\right)\nonumber \\
& =\sum_{i=0}^{\frac{N+1}{2}}a_{i}\left(N+1\right)\left\langle x\right\rangle _{N-i+1}\left(1+2t\right)^{N-2i+1}F\left(t,x+N-i+1\right).\nonumber
\end{align}
Comparing the coefficients on both sides of (\ref{eq:31}), we
have
\begin{equation}
a_{0}\left(N+1\right)=a_{0}\left(N\right),\quad a_{\frac{N+1}{2}}\left(N+1\right)=2a_{\frac{N-1}{2}}\left(N\right),\label{eq:32}
\end{equation}
and
\begin{equation}
a_{i}\left(N+1\right)=2\left(N-2i+2\right)a_{i-1}\left(N\right)+a_{i}\left(N\right),\quad\left(1\le i\le\frac{N-1}{2}\right).\label{eq:33}
\end{equation}
In addition, we have the following ``initial conditions'':
\begin{equation}
F^{\left(0\right)}\left(t,x\right)=F\left(t,x\right)=a_{0}\left(0\right)F\left(t,x\right).\label{eq:34}
\end{equation}
Thus, by (\ref{eq:34}), we get $a_{0}\left(0\right)=1$.
From (\ref{eq:19}) and (\ref{eq:25}), we note that
\begin{equation}
F^{\left(1\right)}\left(t,x\right)=a_{0}\left(1\right)x\left(1+2t\right)F\left(t,x+1\right)=x\left(1+2t\right)F\left(t,x+1\right).\label{eq:35}
\end{equation}
Thus, by (\ref{eq:35}), we see that $a_{0}\left(1\right)=1$.
By (\ref{eq:20}) and (\ref{eq:25}), we easily get
\begin{align}
F^{\left(2\right)}\left(t,x\right) & =\sum_{i=0}^{1}a_{i}\left(2\right)\left\langle x\right\rangle _{2-i}\left(1+2t\right)^{2-2i}F\left(t,x+2-i\right)\label{eq:36}\\
& =a_{0}\left(2\right)\left\langle x\right\rangle _{2}\left(1+2t\right)^{2}F\left(t,x+2\right)+a_{1}\left(2\right)xF\left(t,x+1\right)\nonumber \\
& =2xF\left(t,x+1\right)+\left\langle x\right\rangle _{2}\left(1+2t\right)^{2}F\left(t,x+2\right).\nonumber
\end{align}
Thus, by comparing the coefficients on both sides of (\ref{eq:36}),
we get
\begin{equation}
a_{0}\left(2\right)=1,\quad\text{and}\quad a_{1}\left(2\right)=2.\label{eq:37}
\end{equation}
In (\ref{eq:25}), it is not difficult to show that
\begin{equation}
a_{\frac{N+1}{2}}\left(N\right)=0,\quad\text{for all positive integers }N.\label{eq:38}
\end{equation}
From (\ref{eq:38}), we note that
\begin{equation}
a_{1}\left(1\right)=a_{2}\left(3\right)=a_{3}\left(5\right)=a_{4}\left(7\right)=\cdots=0.\label{eq:39}
\end{equation}
By (\ref{eq:38}), we get
\begin{equation}
F^{\left(N\right)}\left(t,x\right)=\sum_{i=0}^{\left[\frac{N+1}{2}\right]}a_{i}\left(N\right)\left\langle x\right\rangle _{N-i}\left(1+2t\right)^{N-2i}F\left(t,x+N-i\right),\label{eq:40}
\end{equation}
where
\begin{equation}
a_{0}\left(N+1\right)=a_{0}\left(N\right),\quad a_{\frac{N+1}{2}}\left(N\right)=0,\quad\text{for all positive integers }N,\label{eq:41}
\end{equation}
and
\begin{equation}
a_{i}\left(N+1\right)=2\left(N-2i+2\right)a_{i-1}\left(N\right)+a_{i}\left(N\right),\quad\left(1\le i\le\left[\frac{N+1}{2}\right]\right).\label{eq:42}
\end{equation}
From (\ref{eq:41}), we note that
\begin{equation}
a_{0}\left(N+1\right)=a_{0}\left(N\right)=a_{0}\left(N-1\right)=\cdots=a_{0}\left(1\right)=1.\label{eq:43}
\end{equation}
For $i=1,2,3$ in (\ref{eq:42}), we have
\begin{align}
a_{1}\left(N+1\right) & =2\sum_{k=0}^{N-1}\left(N-k\right)a_{0}\left(N-k\right),\label{eq:44}\\
a_{2}\left(N+1\right) & =2\sum_{k=0}^{N-3}\left(N-2-k\right)a_{1}\left(N-k\right),\label{eq:45}
\end{align}
and
\begin{equation}
a_{3}\left(N+1\right)=2\sum_{k=0}^{N-5}\left(N-4-k\right)a_{2}\left(N-k\right).\label{eq:46}
\end{equation}
Thus, we can deduce that, for $1\le i\le\left[\frac{N+1}{2}\right],$
\begin{align}
a_{i}\left(N+1\right) & =2\sum_{k=0}^{N-2i+1}\left(N-k-2i+2\right)a_{i-1}\left(N-k\right)\label{eq:47}\\
& =2\sum_{k=0}^{N+2-2i}ka_{i-1}\left(k+2i-2\right).\nonumber
\end{align}
Now, we give explicit expressions for $a_{i}\left(N+1\right)$.
From (\ref{eq:43}), (\ref{eq:44}), (\ref{eq:45}) and (\ref{eq:46}),
we have
\begin{align}
a_{1}\left(N+1\right) & =2\sum_{k_{1}=1}^{N}k_{1}a_{0}\left(k_{1}\right)=2\sum_{k_{1}=1}^{N}k_{1}\label{eq:48}\\
a_{2}\left(N+1\right) & =2\sum_{k_{2}=1}^{N-2}k_{2}a_{1}\left(k_{2}+2\right)=2^{2}\sum_{k_{2}=1}^{N-2}\sum_{k_{1}=1}^{k_{2}+1}k_{2}k_{1},\label{eq:49}\\
a_{3}\left(N+1\right) & =2\sum_{k_{3}=1}^{N-4}k_{3}a_{2}\left(k_{3}+4\right)=2^{3}\sum_{k_{3}=1}^{N-4}\sum_{k_{2}=1}^{k_{3}+1}\sum_{k_{1}=1}^{k_{2}+1}k_{3}k_{2}k_{1}\label{eq:50}
\end{align}
and
\begin{equation}
a_{4}\left(N+1\right)=2^{4}\sum_{k_{4}=1}^{N-6}\sum_{k_{3}=1}^{k_{4}+1}\sum_{k_{2}=1}^{k_{3}+1}\sum_{k_{1}=1}^{k_{2}+1}k_{4}k_{3}k_{2}k_{1}.\label{eq:51}
\end{equation}
Continuing this process, we have
\begin{equation}
a_{i}\left(N+1\right)=2^{i}\sum_{k_{i}=1}^{N-2i+2}\sum_{k_{i-1}=1}^{k_{i}+1}\cdots\sum_{k_{1}=1}^{k_{2}+1}\left(\prod_{l=1}^{i}k_{l}\right),\quad\left(1\le i\le\left[\frac{N+1}{2}\right]\right).\label{eq:52}
\end{equation}
Therefore, by (\ref{eq:52}), we obtain the following theorem.
\begin{thm}
\label{thm:6} For $N=0,1,2,\dots$, the family of differential equations
\begin{align*}
F^{\left(N\right)}\left(t,x\right)&=\left(\frac{d}{dt}\right)^{N}F\left(t,x\right)\\
&=\left(\sum_{i=0}^{\left[\frac{N+1}{2}\right]}a_{i}\left(N\right)\left\langle x\right\rangle _{N-i}\left(1+2t\right)^{N-2i}\left(1-t-t^{2}\right)^{-N+i}\right)F\left(t,x\right)
\end{align*}
have a solution
\[
F\left(t,x\right)=\left(1-t-t^{2}\right)^{-x}
\]
where
\[
a_{0}\left(N\right)=1,\quad a_{\frac{N+1}{2}}\left(N\right)=0,\quad\text{for all positive integers }N,
\]
and
\[
a_{i}\left(N\right)=2^{i}\sum_{k_{i}=1}^{N-2i+1}\sum_{k_{i-1}=1}^{k_{i}+1}\cdots\sum_{k_{1}=1}^{k_{2}+1}\left(\prod_{l=1}^{i}k_{l}\right),\quad\left(1\le i\le\left[\frac{N}{2}\right]\right).
\]
\end{thm}
From (\ref{eq:4}), we note that
\begin{equation}
1=\sum_{k=0}^{\infty}F_{k}t^{k}\left(1-t-t^{2}\right)=\sum_{k=0}^{\infty}F_{k}t^{k}-\sum_{k=1}^{\infty}F_{k-1}t^{k}-\sum_{k=2}^{\infty}F_{k-2}t^{k}.\label{eq:53}
\end{equation}
Comparing the coefficients on the both sides of (\ref{eq:53}), we
have
\begin{equation}
F_{0}=1,\quad F_{1}-F_{0}=0\iff F_{1}=F_{0}=1,\label{eq:54}
\end{equation}
and
\begin{equation}
F_{k}-F_{k-1}-F_{k-2}=0\quad\text{if }k\ge2.\label{eq:55}
\end{equation}
By (\ref{eq:4}), we easily get
\begin{equation}
F\left(t,x\right)=\left(1-t-t^{2}\right)^{-x}=\sum_{k=0}^{\infty}p_{k}\left(x\right)\frac{t^{k}}{k!},\label{eq:56}
\end{equation}
and
\begin{equation}
F^{\left(N\right)}\left(t,x\right)=\left(\frac{d}{dt}\right)^{N}F\left(t,x\right)=\sum_{k=0}^{\infty}p_{k+N}\left(x\right)\frac{t^{k}}{k!}.\label{eq:57}
\end{equation}
On the other hand, by Theorem \ref{thm:6}, we get
\begin{align}
F^{\left(N\right)}\left(t,x\right) & =\sum_{i=0}^{\left[\frac{N+1}{2}\right]}a_{i}\left(N\right)\left\langle x\right\rangle _{N-i}\left(1+2t\right)^{N-2i}F\left(t,x+N-i\right)\label{eq:58}\\
& =\sum_{i=0}^{\left[\frac{N+1}{2}\right]}a_{i}\left(N\right)\left\langle x\right\rangle _{N-i}\left(1+2t\right)^{N-2i}\sum_{m=0}^{\infty}p_{m}\left(x+N-i\right)\frac{t^{m}}{m!}\nonumber \\
& =\sum_{i=0}^{\left[\frac{N+1}{2}\right]}a_{i}\left(N\right)\left\langle x\right\rangle _{N-i}\sum_{l=0}^{\infty}\left(N-2i\right)_{l}2^{l}\frac{t^{l}}{l!}\sum_{m=0}^{\infty}p_{m}\left(x+N-i\right)\frac{t^{m}}{m!}\nonumber \\
& =\sum_{i=0}^{\left[\frac{N+1}{2}\right]}a_{i}\left(N\right)\left\langle x\right\rangle _{N-i}\sum_{k=0}^{\infty}\left(\sum_{l=0}^{k}\binom{k}{l}\left(N-2i\right)_{l}2^{l}p_{k-l}\left(x+N-i\right)\right)\frac{t^{k}}{k!}\nonumber \\
& =\sum_{k=0}^{\infty}\left(\sum_{i=0}^{\left[\frac{N+1}{2}\right]}\sum_{l=0}^{k}\binom{k}{l}\left(N-2i\right)_{l}2^{l}a_{i}\left(N\right)\left\langle x\right\rangle _{N-i}p_{k-l}\left(x+N-i\right)\right)\frac{t^{k}}{k!}.\nonumber
\end{align}
Therefore, by comparing the coefficients on both sides of (\ref{eq:57})
and (\ref{eq:58}), we obtain the following theorem.
\begin{thm}
\label{thm:7} For $k,N=0,1,2,\dots$, we have
\[
p_{k+N}\left(x\right)=\sum_{i=0}^{\left[\frac{N+1}{2}\right]}\sum_{l=0}^{k}\binom{k}{l}\left(N-2i\right)_{l}2^{l}a_{i}\left(N\right)\left\langle x\right\rangle _{N-i}p_{k-l}\left(x+N-i\right),
\]
where
\begin{align*}
a_{0}\left(N\right) & =1,\quad a_{\frac{N+1}{2}}\left(N\right)=0,\quad\text{for all positive integers }N,\\
a_{i}\left(N\right) & =2^{i}\sum_{k_{i}=1}^{N-2i+1}\sum_{k_{i-1}=1}^{k_{i}+1}\cdots\sum_{k_{1}=1}^{k_{2}+1}\left(\prod_{l=1}^{i}k_{l}\right),\quad\left(1\le i\le\left[\frac{N}{2}\right]\right).
\end{align*}
\end{thm}
When $k=0$ in Theorem \ref{thm:7}, we have the following corollary.
\begin{cor}
\label{cor:8} For $N=0,1,2,\dots$, we have
\[
p_{N}\left(x\right)=\sum_{i=0}^{\left[\frac{N+1}{2}\right]}a_{i}\left(N\right)\left\langle x\right\rangle _{N-i}.
\]
\end{cor}
Let us take $x=1$ in Corollary \ref{cor:8}. Then, we easily get
\begin{equation}
p_{N}\left(1\right)=\sum_{i=0}^{\left[\frac{N+1}{2}\right]}a_{i}\left(N\right)\left(N-i\right)!=N!+\sum_{i=1}^{\left[\frac{N+1}{2}\right]}a_{i}\left(N\right)\left(N-i\right)!\label{eq:59}
\end{equation}
Thus, by (\ref{eq:59}), we get
\begin{align}
\frac{p_{N}\left(1\right)}{N!} & =1+\frac{1}{N!}\sum_{i=1}^{\left[\frac{N+1}{2}\right]}a_{i}\left(N\right)\left(N-i\right)!\label{eq:60}\\
& =1+\frac{1}{N!}\sum_{i=1}^{\left[\frac{N+1}{2}\right]}\sum_{k_{i}=1}^{N+1-2i}\sum_{k_{i-1}=1}^{k_{i}+1}\cdots\sum_{k_{1}=1}^{k_{2}+1}2^{i}\left(\prod_{l=1}^{i}k_{l}\right)\left(N-i\right)!\nonumber
\end{align}
Therefore, by (\ref{eq:6}) and (\ref{eq:60}), we obtain the following
corollary.
\begin{cor}
\label{cor:9}For $N=0,1,2,\dots$, we have
\[
F_{N}-1=\frac{1}{N!}\left(\sum_{i=1}^{\left[\frac{N+1}{2}\right]}\sum_{k_{i}=1}^{N+1-2i}\sum_{k_{i-1}=1}^{k_{i}+1}\cdots\sum_{k_{1}=1}^{k_{2}+1}2^{i}\left(\prod_{l=1}^{i}k_{l}\right)\left(N-i\right)!\right).
\]
\end{cor}
\begin{rem*}
Recently, several authors have studied special polynomials and sequences
arising from the generating functions (see \cite{key-1,key-2,key-3,key-4,key-5,key-6,key-7,key-8,key-9,key-10,key-11,key-12,key-13,key-14,key-15,key-16}).
\end{rem*}
\subsection*{Acknowledgements}
This paper is supported by grant NO 14-11-00022 of Russian Scientific Fund.
\bibliographystyle{amsplain}
\providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
\providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR }
\providecommand{\MRhref}[2]
\href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
}
\providecommand{\href}[2]{#2}
|
\section{Intro and motivation}
It should be stated right away that {\it a priori} there is no right or wrong regularization scheme. No physical observable should depend on the choice of the procedure, as long as it qualifies as a regularization, meaning that there exists a smooth limit of zero regulator resulting in a local theory with a proper particle/field content and symmetries.
One regularization can be more appropriate than the other in the sense that it allows to reproduce the necessary (specific) Ward identities with lesser effort. In particular symmetry preserving regularizations are usually favored among others, however, one is not forced by any means to make this choice, and if done carefully, computations using different regularizations, lead to the same predictions for physical observables.
One of the regularizations suitable for computing determinants of differential operators (one-loop partition functions) or Casimir energies is zeta function regularization. Even though the subtleties of its uses have been discussed on numerous occasions (for systematic approach see for example~\cite{Elizalde:1994gf, Vassilevich:2003xt, Dunne:2007rt, Elizalde:2012zza}) there are still instances when it is handled in a somewhat cavalier way leading to inconsistencies. We do not claim that results obtained in these cases are wrong, it is rather the matter of interpreting those results.
If a regularization breaking symmetries of a system is used, one may be forced to introduce non-symmetric counterterms to compensate for it in the Ward identities. It means that the space of possible operators in the Lagrangian is not restricted by those respecting the symmetries.
The feature of zeta function regularization is that it is diffeomorphism invariant. Thus, the analysis boils down to listing all possible diffeomorphism invariant operators in the Lagrangian. However, if the regularization is not used in a proper way there is no guarantee that the diffeomorphisms play any role in restricting the terms in the Lagrangian. Thus, the analysis with only diffeomorphism invariant terms may be not complete, leading to a non-complete result.
In this paper we consider a scalar field coupled to a round sphere. We show several examples of how to compute its free energy using different regularizations, like zeta function, Pauli-Villars, the heat kernel, and show that all of them produce the same result. Surprisingly enough sometimes a misuse of zeta function regularization leads to a correct prediction for a finite part of the free energy. We show examples when that happens and explain the reason. Some of our results are necessarily new and can be found in cited textbooks, but we hope
that our presentation and new results provide certain clarification.
The paper is organized as follows. In Section~\ref{zeta_function_intro} we define different zeta function to be used later on. In Sections~\ref{section_PV} and~\ref{section_zeta-function} we show how to use Pauli-Villars and zeta function regularizations correspondingly to compute the free energy of a scalar on a round sphere. In Section~\ref{section_t-expansion} we present yet another regularization using a smoothed cut-off function. Section~\ref{section_examples} is dedicated to examples and in Section~\ref{section_conclusion} we present our conclusions. Some relevant formulas and more detailed computations can be found in Appendices.
\section{Different zeta functions \label{zeta_function_intro}}
In this section we introduce the definitions of different zeta functions along with some of their properties which will be useful in sections to come.
\subsection{Riemann zeta function}
Riemann zeta function $\zeta (s)$ is defined for $\mathrm {Re} \, s >1$ as an infinite convergent sum
\begin{equation}
\zeta (s)=\sum _ {k=1} ^ \infty \frac {1} {k^s}.
\label{Riemann_zeta}
\end{equation}
For $s<1$ one defines zeta function using analytic continuation. From the integral representation of the $\Gamma$-function
\begin{equation}
\Gamma (s) = \int _ 0 ^ \infty dt e ^ {-t} t ^ {s-1}, ~~ \mathrm {Re} \, s >1,
\end{equation}
it is easy to show that for $\mathrm {Re} \, s >1$ the following relation holds
\begin{equation}
\zeta (s) \Gamma (s) = \int _ 0 ^ \infty dt \, \theta (t) \, t ^ {s-1},
\label{zeta_gamma_continuation}
\end{equation}
where we introduced the function $\theta(t)$ (later to be called the heat kernel) as the sum
\begin{equation}
\theta (t) = \sum _ {k=1} ^ \infty e ^ {-k t} = \frac {1} {e ^ t -1}
\end{equation}
convergent for all positive values of $t$. Breaking up the integral in (\ref{zeta_gamma_continuation}) into two parts and doing a trivial integration
\begin{equation}
\int _ 0 ^ 1 + \int _ 1 ^ \infty \frac {t ^ {s-1} \, d t} {e ^ t -1} =
\int _ 0 ^ 1 t ^ {s-1} \, d t \left ( \frac {1} {e ^ t -1} - \frac {1} {t} \right ) + \frac {1} {s - 1} + \int _ 1 ^ \infty \frac {t ^ {s-1} \, d t} {e ^ t -1},
\end{equation}
one gets the expression for the zeta function which can be used for $\mathrm{Re} s > 0$
\begin{equation}
\zeta (s) \Gamma (s) - \frac {1} {s - 1} = \int _ 0 ^ 1 t ^ {s-1} \, d t \left ( \frac {1} {e ^ t -1} - \frac {1} {t} \right ) + \int _ 1 ^ \infty \frac {t ^ {s-1} \, d t} {e ^ t -1}, ~~
\mathrm{Re} (s) > 0.
\end{equation}
The function can be analytically continued to the region $\mathrm{Re} \, s > - N$, where $N$ is an arbitrary positive integer, in a similar manner.
\subsection{Hurwitz zeta function}
A natural generalization of (\ref{Riemann_zeta}) is the Hurwitz zeta function defined as
\begin{equation}
\zeta (s,q)=\sum _ {k=0} ^ \infty \frac {1} {(k+q)^s}, ~~ \mathrm {Re} \, s >1.
\label{Hurwitz_zeta}
\end{equation}
Analytic continuation of the Hurwitz zeta function is obtained analogously to the one described in the previous section. Using the Mellin transform
\begin{equation}
\Phi (z) = \int _ 0 ^ \infty x^ {z-1} (x+q) ^{-s} dx = q ^ {z-s} \frac {\Gamma (s-z) \Gamma (z)} {\Gamma (s)},
\end{equation}
which is holomorphic in the strip $ 0 < \mathrm {Re} \, z < \mathrm {Re} \, s $ one can rewrite for $c > 1$ and $\mathrm {Re} \, s > 1$ the zeta function in the following form
\begin{equation}
\zeta (s,q) = q^{-s} + \sum _ {k=1} ^ \infty \frac {1} {(k+q)^s} = \frac {1} {2 \pi i} \int \limits _ {c - i \infty} ^ {c + i \infty} \sum _ {k=1} ^ \infty k ^ {-z}
\Phi(z) dz = \frac {1} {2 \pi i} \int \limits _ {c - i \infty} ^ {c + i \infty} \zeta (z) \Phi(z) dz.
\end{equation}
Closing the contour of integration in the left half plane the integral can be written as a sum over residues
\begin{eqnarray}
\label{asymptotic_Hurwitz_zeta}
\zeta (s,q) & = & \int _ 0 ^ \infty (x+q) ^{-s} dx +
\frac {q ^ {-s}} {\Gamma (s)} \sum _ {k=0} ^ \infty \mathrm {res} \left [ q^ z \Gamma (s-z) \Gamma (z) \zeta(z), z= - k \right ] \\
& = & - \frac {q ^ {1-s}} {1-s} + \frac {q ^ {-s}} {2} + \sum _ {m=1} ^ \infty \frac {B _ {2m}} {(2m)!} q ^ {1-s - 2m} s (s+1) \dots (s + 2 m - 2), \nonumber
\end{eqnarray}
where the first integral in (\ref{asymptotic_Hurwitz_zeta}) corresponds to the pole of $\zeta(s)$ at $s=1$. This expansion for the zeta function should be understood as an asymptotic one. The result can be used to find a large $L$ expansion of the following finite sum
\begin{equation}
\sum _ {n =1} ^ L \frac {1} {(n+q)^s} = \zeta (s,q + 1) - \zeta (s,L+q+1), ~~ s \neq 1,
\label{finite_two_zeta}
\end{equation}
which will be used in the following sections.
\subsection{Spectral zeta functions \label{heat_kernel_section}}
In general for a differential operator $D$ with positive eigenvalues $\Lambda _ \alpha > 0$ one defines spectral zeta function as the sum
\begin{equation}
\zeta _ D (s) = \sum _ \alpha \Lambda _ \alpha ^ {-s} \equiv \mathrm {Tr} D ^ {-s},
\label{zeta_spectral}
\end{equation}
for sufficiently large $\mathrm {Re} \, s > s _ 0$. The analytic continuation is performed using the corresponding heat kernel, which is given by
\begin{equation}
\theta _ D (t) = \sum _ \alpha e ^ {- t \Lambda _ \alpha} \equiv \mathrm{Tr} e ^ {-D t}.
\label{heat_kernel}
\end{equation}
If the asymptotic expansion of $\theta (t)$ for $t \to 0$ is known
\begin{equation}
\theta _ D (t) \underset{t \to 0} {=} \sum _ {k \geq 0} a _ k t ^ {-s_ k} + O (t),
\label{heat_exp_gen}
\end{equation}
with $s_0>s_1>s_2>\dots>0$ and $a_k$ being constants, then the analytic continuation is done in exactly the same way as for the Riemann zeta function. For instance, the following expression
\begin{equation}
\zeta _ D (s) \Gamma (s) = \frac {a _ 0} {s-s_0} + \int _ 0 ^ 1 dt \, \left ( \theta _ D (t) - \frac{a_0}{t ^ {s_ 0}} \right ) t ^ {s-1} + \int _ 1 ^ \infty dt \, \theta _ D (t) t ^ {s-1},
\label{zeta_analytic_cont}
\end{equation}
can be used to define the spectral zeta function of the differential operator $D$ in the domain~$\mathrm {Re} \, s > s _ 1$. It is clear that the powers $s _ k$ in (\ref{heat_exp_gen}) are in one to one correspondence with the poles of the zeta function $\zeta_D(s)$. In general after analytic continuation the zeta function can be written as
\begin{equation}
\zeta _ D (s) \Gamma (s) = G _ D (s),
\end{equation}
where $G _ D (s)$ is the analytic continuation of the r.h.s. of (\ref{zeta_analytic_cont}). It is straightforward to show that around $s=0$ the following expansion -- provided $\zeta _ D (0)$ is regular -- holds
\begin{equation}
G _ D (s) \underset{s\to 0}{=} \frac {\zeta _ D (0)} {s} + \zeta ' _ D (0) - \gamma \zeta _ D (0) + \dots,
\label{G_s_expansion}
\end{equation}
hence $\zeta _ D (0)$ can be found as a $t$-independent (compare with (\ref{gen_theta_asympt})) coefficient in (\ref{heat_exp_gen}).
\section{Pauli-Villars regularization \label{section_PV}}
Let us consider a free scalar field on a $d$-dimensional Euclidean sphere (not necessarily minimally coupled to a round metric) and compute its partition function. Computations of the free energy of a scalar on spheres can be found in several papers, for instance the case of conformally coupled scalar was considered in~\cite{Klebanov:2011gs}. Since the Ricci curvature is constant for a round sphere $R = d(d-1)/a^2$, the action for both massive and massless scalar on a sphere can be written in a general form as
\begin{equation}
S[\phi,g] = \frac {1} {2} \int d ^ d x \sqrt{g} \left [ \left ( \nabla \phi \right ) ^ 2 + \alpha _ m ^2 (\xi) \phi ^ 2 \right ],
\label{bare_action}
\end{equation}
where $\alpha _ m ^2 (\xi)= m^2 + \xi R$ with $\xi$ being a constant and $a$ the radius of the sphere. For a specific value ($d > 1$)
\begin{equation}
\xi = \frac {d-2} {4(d-1)}
\label{xi_conformal}
\end{equation}
the action becomes Weyl invariant (conformally coupled scalar). The free energy $F$ is by definition a logarithm of the partition function
\begin{equation}
Z = e ^ {-F} = \int \mathcal D \phi e ^ {-S[\phi,g]}, ~~ \text{or} ~~ F = \frac {1} {2} \log \det \Delta,
\label{formal_det}
\end{equation}
where the operator $\Delta$ is given by
\begin{equation}
\Delta = - \nabla ^ 2 + \alpha _ m ^ 2(\xi).
\label{general_delta}
\end{equation}
For an $n$-dimensional sphere this operator has eigenvalues
\begin{equation}
\lambda _ \ell = \frac {\ell (\ell+d-1)} {a^2} + \frac {\alpha _ m ^2(\xi)} {a^2},
\end{equation}
with multiplicity
\begin{equation}
\mu _ \ell = \frac {(2\ell + d-1)(\ell+d-2)!} {(d-1)!d!}, ~~ \ell = 0,1,\dots
\label{degeneracy}
\end{equation}
It is clear that the formal expression (\ref{formal_det}) is divergent and needs regularization. Before showing how zeta functions can be used to compute the free energy we make use of an alternative regularization. In this way we will have a reference point for the computations involving zeta functions. One of the most straightforward regularizations is the Pauli-Villars (PV) one, in which the action (\ref{bare_action}) is regulated by adding sufficient number of regulator fields with mass $M_i$ and statistics $-c_i$ (in general fractional). It can also be viewed as adding higher derivative operators.
\subsection{$S^1$ \label{PV_S1_ex}}
For a one dimensional sphere to regularize the action
\begin{equation}
S = \frac {1} {2} \int d t \left ( \dot \phi ^ 2 + m^ 2 \phi ^ 2 \right ),
\end{equation}
it is enough to introduce one PV regulator with mass $M$ and opposite to $\phi$ statistics.
The corresponding eigenvalues are
\begin{equation}
\lambda _ \ell = \frac {\ell^2} {a^2}+m^2, ~~ \text{and} ~~ \lambda ^ M _ \ell = \frac {\ell^2} {a^2} + M ^ 2, ~~ \ell \in \mathbb Z.
\label{lambda_S1}
\end{equation}
As a result the expression for the regularized free energy reads
\begin{equation}
F ^ R _ {S^2} = \frac {1} {2} \log \prod _ {\ell \in \mathbb Z} \frac {\lambda _ \ell } {\lambda _ \ell ^ M}
= \log \left ( \frac {m} {M} \prod _ {\ell =1} ^ \infty \frac { \ell ^ 2 + m^2 a ^ 2} {\ell ^ 2 + M ^ 2 a^ 2} \right ) = \log \left ( \frac{ \sinh {\pi a m} } {\sinh {\pi a M}} \right ),
\end{equation}
which upon the limit $M \to \infty$ becomes
\begin{equation}
F ^ R _ {S^2} = \log \left ( 2 \sinh {\pi a m} \right ) - \pi a M.
\label{PV_S2}
\end{equation}
The divergent part can be eliminated by the counterterm $\mathrm {Vol} _ {S ^ 2} = 2 \pi a$ with an appropriate coefficient.
\subsection{$S^3$}
In the case of a three dimensional sphere the free energy can be regularized employing only two PV fields
\begin{equation}
2 F ^ R _ {S^3} = \sum _ {\ell = 1} ^ \infty \ell ^ 2 \left \{ \log \left [ \ell ^ 2 - 1 + \alpha _ m ^ 2(\xi) a^2 \right ] -
\sum _ {i=1} ^ 2 c _ i \log \left [ \ell ^ 2 - 1 + \alpha _ {m} ^ 2(\xi) a^2 + M^2_i a^2 \right ] \right \},
\label{FS3}
\end{equation}
with mass parameters satisfying the following relations
\begin{eqnarray}
c_1 + c _ 2 & = & 1 \nonumber \\
c _ 1 M _ 1 ^ 2 + c _ 2 M _ 2 ^ 2 & = & 0.
\label{ci_constr}
\end{eqnarray}
Introducing the cutoff $L \gg M _ i a \gg 1 $ the sum with PV regulators can be computed using Euler-Maclaurin summation formula, while
the first term in (\ref{FS3}) can be found using the zeta function expansion (\ref{asymptotic_Hurwitz_zeta}) and the formula (\ref{finite_two_zeta}). As a result for $M _ 1 = M$ and $M _ 2 = r M$ with $r$ being a number we obtain the following expression for a regularized partition function in the limit $M \to \infty$
\begin{equation}
F ^ R _ {S^3} = - a^3 M ^ 3 \frac {\pi r ^ 2} { 6 (1+r) } - a M \frac {\pi r} {4( 1 + r )} \left [1-\alpha_m^2(\xi) a^2 \right ] + F _ {S ^ 3} \left (\sqrt{\alpha _ m ^2(\xi) a ^ 2-1}\right ).
\label{PV_S3}
\end{equation}
Divergent terms can be removed by the counterterms of the form
\begin{equation}
b_1 M ^ 3 \mathrm {Vol} _ {S ^ 3} + b_2 M \mathrm {Ric} _ {S ^ 3},
\label{counterterms _S3}
\end{equation}
with appropriate coefficients $b_{1,2}$, where $ \mathrm {Ric} _ {S ^ 3}$ is the Ricci curvature. The function $F_ {S^3}(x)$ is given in the Appendix~\ref{Appendix_PV}. For a limiting case of conformal coupling and zero mass it becomes
\begin{equation}
f _ {S ^ 3} = \frac {1} {16} \left ( 2 \log 2 - \frac {3 \zeta (3)} {\pi ^ 2} \right ) = 0.0638,
\label{f_conf_scalar}
\end{equation}
which coincides with the result for the partition function from~\cite{Klebanov:2011gs}.
\section{Zeta function regularization \label{section_zeta-function}}
As was discussed in the introduction different regularizations should lead to one and the same prediction for an observable. However, different procedures (schemes) allow different number of counterterms. For instance, in the previous section we considered the PV regularization which obviously preserves the diffeomorphisms but even in the case of zero mass $m=0$ and conformal coupling (\ref{xi_conformal}) it breaks Weyl invariance due to the presence of a scale parameter (regulator mass). This is precisely what we see from the expressions for the regularized partition function (\ref{PV_S2}) and (\ref{PV_S3}), where the divergent terms are clearly of a diffeomorphism invariant nature but are not Weyl invariant.
In general for the case of zero mass $m=0$ using a regularization preserving the diffeomorphisms but having a scale $M$ the only local counterterms one can write are of the form
\begin{equation}
M ^ d \mathrm {Vol}, ~ M ^ {d-2} \mathrm {Ric}, ~ M ^ {d-4} \mathrm {Ric^2}, \dots
\end{equation}
Thus the power counting tells us that in an odd\footnote{In an even number of dimensions $\log$ divergent terms lead to the Weyl anomaly.} number of dimensions $d$ a counterterm scaling as $M^0$ does not exist, hence, the $M$-independent (finite) part of the regularized partition function is a prediction and can be associated with the (renormalized) partition function of a scalar on a $d$-sphere $f _ {S ^d}$. There is simply no counterterm that may change it. It is not the case for a massive scalar, for even $M$-independent part can be changed using counterterms, for example $m ^ d \mathrm {Vol}$.
If instead for a conformally coupled scalar (still odd number of dimensions) a regularization preserving the Weyl invariance was used, the result for the partition function would be produced automatically without any need to add counterterms for there are none satisfying the constraints of Weyl invariance. To show that we consider a bit more general setup. In this way we will also obtain the expression for the partition function in terms of the zeta function of the corresponding operator. To find a partition function of a scalar field living on a $d$-dimensional manifold $\mathbb M^d$, instead of introducing PV regulators we employ dimensional regularization and put a scalar on
$\mathbb M ^ d \times \mathbb R ^ {2 \varepsilon}$, in the end the limit $\varepsilon \to 0$ should be taken. It follows from (\ref{formal_det}) that the regularized partition function has the form
\begin{equation}
F ^ R = -\frac {1} {2} \frac {\partial} {\partial s} \Big | _ {s=0} \sum _ {\alpha} \int \frac {d^ {2\varepsilon}x \, d^ {2\varepsilon} k} {(2 \pi) ^ {2 \varepsilon}}
\left [ \frac {k ^ 2} {\mu ^ 2} + \frac {\Lambda _ \alpha} {\mu ^ 2} \right ] ^ {-s},
\end{equation}
where $\Lambda _ \alpha$ are eigenvalues of the operator of second variation and $\mu$ is a normalization scale. Computing the integral we obtain
\begin{equation}
F ^ R = -\frac {1} {2} \frac {\partial} {\partial s} \Big | _ {s=0} \frac {\Gamma(s-\varepsilon)} {\Gamma(s)} \sum _ {\alpha} \left ( \frac {\mu ^ 2 L ^ 2} {4 \pi} \right ) ^ \varepsilon
\left ( \frac {\Lambda _ \alpha} {\mu ^ 2} \right ) ^ {\varepsilon-s},
\end{equation}
where $L$ is the size of an auxiliary space (infrared regulator). The above expression should be understood in the sense of an analytic continuation in $\varepsilon$. Equivalently one can view it as an analytic continuation in $s$ and set $\varepsilon=0$ (there are no $1 / \varepsilon$ poles in odd number of dimensions). As a result we get
\begin{equation}
F ^ R = -\frac {1} {2} \frac {\partial} {\partial s} \Big | _ {s=0} \sum _ \alpha \left ( \frac{\Lambda _ \alpha} {\mu ^ 2} \right ) ^ {-s} = - \frac {1} {2} \zeta ' _ D (0)
- \frac {1} {2} \zeta _ D (0) \log \mu ^ 2,
\label{F_zeta_function}
\end{equation}
where $\zeta _ D (s)$ is nothing else but the corresponding zeta function (\ref{zeta_spectral}).
In order to compute a partition function of a conformally coupled scalar on a round $d$-dimensional sphere
we first have to find a small $t$ asymptotic expansion of the corresponding heat kernel
\begin{equation}
\theta_ {S ^ d} (s) = \sum _ {\ell=0} ^ \infty \mu _ \ell \exp \left [ - t \left(\ell + \frac {d} {2} \right) \left(\ell + \frac {d} {2} -1 \right) \right ].
\end{equation}
Using the Euler-Maclaurin formula (see Appendix \ref{Appendix_PV}) it is easy to show that
\begin{equation}
\theta_ {S ^ d} (s) \underset{t\to 0}{=} \sum _ {k=0} ^ {(d+1)/2}a _ {d/2-k} t ^ {-d/2+k}+ O (t).
\end{equation}
Hence we get
\begin{eqnarray}
\zeta _ {S ^ d} (s) \Gamma (s) = G _ {S ^ d}(s) \equiv & \displaystyle \int _ 0 ^ \infty & dt \, \left [ \theta _ {S ^ d} (t)
- \sum _ {k=0} ^ {(d+1)/2}a _ {d/2-k} t ^ {-d/2+k} \right ] t ^ {s-1} \\
& + &\sum _ {k=0} ^ {(d+1)/2} \frac { a _ {d/2-k} } {s - d/2 + k} + \int _ 1 ^ \infty dt \, \theta _ {S ^ d} (t) t ^ {s-1}. \nonumber
\end{eqnarray}
Since there is no pole at $s=0$ it means that for an arbitrary odd $d$ the zeta function vanishes $\zeta _ {S ^ d} (0)=0$. In this case the partition function of a conformally coupled scalar is given by
\begin{equation}
f _ {S ^ d} = - \frac {1} {2} \frac {\partial} {\partial s} \Big | _ {s=0} \sum _ {\ell=0} ^ \infty \mu _ \ell \left [ \left(\ell + \frac {d} {2} \right) \left(\ell + \frac {d} {2} - 1 \right) \right ] ^ {-s} (\mu^2 a^2) ^ s
= - \frac {1} {2} \zeta ' _ {S ^ d} (s),
\end{equation}
and it does not depend on the normalization scale $\mu$. Just as an example for $3$-dimensional sphere we get
\begin{equation}
G _ {S ^ 3}(0) = \int _ 0 ^ \infty \frac{dt} {t} \, \left [ \theta _ {S ^ 3} (t) - \sqrt{\pi } \left ( \frac{1}{4 \, t^{3/2}}+\frac{1}{16 \, t ^ {1/2}} + \frac{t ^ {1/2}}{128} \right ) \right ] - \frac {53\sqrt{\pi }} {192} + \int _ 1 ^ \infty \frac{dt} {t} \, \theta _ {S ^ d} (t).
\label{S3_zeta_function_reg}
\end{equation}
Computing the integrals numerically we obtain the same value as in (\ref{f_conf_scalar}).
\subsection{How not to use zeta function regularization}
We should stress that although it is tempting to continue the formal manipulations with (\ref{formal_det}), it is the expression (\ref{F_zeta_function}) that we obtained for the partition function using the dimensional (in the case at hand it is reduced to the zeta function) regularization and not something else. It is easy to face an inconsistency if the regularization is not used with proper care. We will illustrate it on the example of a $3$-sphere with a conformally coupled massless scalar. The formal expression for the partition function is given by\footnote{We deliberately use capital letter $F _ {S ^ 3}$ to distinguish from the correct expression given by $f _ {S ^ 3}$.}
\begin{eqnarray}
F_ {S^3} & = &
\frac {1} {2} \sum _ {\ell=0} ^ \infty (\ell + 1)^2 \left \{ \log \left [ \left (\ell + \frac {3} {2} \right)\left( \mu a \right )^p \right ] + \log \left [ \left(\ell + \frac {1} {2} \right)\left( \mu a \right ) ^ {2-p} \right ] \right \} \nonumber \\
& = & - \frac {1} {2} \frac {\partial} {\partial s} \Big | _ {s=0} \left [ (\mu a) ^ {-ps} \sum _ {\ell=0} ^ \infty (\ell + 1 ) ^ 2 \left (\ell + \frac {3} {2} \right)^{-s} + (\mu a) ^ {-(2-p)s} \sum _ {\ell=0} ^ \infty (\ell + 1)^2 \left (\ell + \frac {1} {2} \right)^{-s} \right ] \nonumber \\
& = & - \frac {1-p} {24}\log \left( \mu a \right ) + \frac {1} {16} \left ( 2 \log 2 - \frac {3 \zeta (3)} {\pi ^ 2} \right ),
\label{f_formal}
\end{eqnarray}
where $p$ is an arbitrary constant. The term $\log (\mu a)$ signals the presence of a $\log$-divergent term, meaning that -- since there are no $\log$-divergent diffeomorphism invariant counterterms -- the regularization breaks diffeomorphisms. Even for $p=1$ when (\ref{f_formal}) reproduces (\ref{f_conf_scalar}) these manipulations are illegitimate because
\begin{equation}
\sum _ {\ell=0} ^ \infty \frac{(\ell + 1)^2} {\left (\ell + \frac {3} {2} \right)^{s}} + \sum _ {\ell=0} ^ \infty \frac{(\ell + 1)^2} {\left (\ell + \frac {1} {2} \right)^{s}} \neq
\sum _ {\ell=0} ^ \infty \frac{(\ell + 1 ) ^ 2} {\left [ \left(\ell + \frac {3} {2} \right) \left(\ell + \frac {1} {2} \right) \right ] ^ {s}}, ~~ s > 3.
\end{equation}
So it is a wrong function whose analytic continuation to $s=0$ we use in (\ref{f_formal}).
From a purely mathematical perspective it can be understood in the following way. There are different notions of summation like Abel, Ces\'aro, Borel etc. (see~\cite{Hardy}) that are used to give meaning to otherwise divergent series. However, it happens so that some divergent series, for instance
\begin{equation}
1+2+3+\dots,
\end{equation}
cannot be summed using linear and stable methods\footnote{Linear means that the sum of two series is equal to the sum of the series obtained as a combination of the two, and stable means that adding a number to the series increases the sum by the same amount.}. The series in question $\sum _ \ell \mu _ \ell \log \lambda _ \ell$ is one of them, thus the formal step $\log A B = \log A + \log B$ in the sum is inconsistent.
Another example -- considered in~\cite{Brevik:1989fw,Elizalde:1996zk} -- when formal manipulations lead to a wrong result is the computation of the Casimir energy of a piecewise uniform, closed string. The formal expression for the energy is given by
\begin{equation}
E ^ \text{formal} _ \text{Cas} = \sum _ {k=0} ^ \infty \varepsilon _ k = \sum _ {k=0} ^ \infty \left ( k + \beta \right ).
\label{Casimir_string}
\end{equation}
If regularized with the zeta function the energy can be expressed in terms of the Hurwitz zeta function
\begin{equation}
E _ \text{Cas} = \lim _ {s \to -1} \sum _ {k=0} ^ \infty \varepsilon _ k ^ {-s} = \zeta (-1,\beta),
\end{equation}
which is clearly different from the formal substitution
\begin{equation}
E ^ \text{formal} _ \text{Cas} = \sum _ {k=0} ^ \infty \left ( k + \beta \right ) = \zeta (-1) + \beta \left [ 1+ \zeta (0) \right ].
\label{Casimir_string_wrong}
\end{equation}
At the same time imagine that the following sum has to be computed
\begin{equation}
\sum _ {\ell = 1} ^ \infty \lambda _ \ell a ^ 2,
\label{zeta_1_S1}
\end{equation}
with $\lambda _ \ell$ from (\ref{lambda_S1}). Then both methods produce the same result. Indeed, on the one hand formally we have
\begin{equation}
\sum _ {\ell = 1} ^ \infty \left ( \ell ^ 2 + m ^ 2 a ^ 2 \right ) = \zeta (-2)+ \zeta(0) m^ 2 a ^ 2 = -\frac {1} {2} m^ 2 a ^ 2,
\label{Epstein_S1}
\end{equation}
on the other hand using the expression from~\cite{Elizalde:1996zk} for the Epstein zeta function $\zeta _ E (s,q)$
\begin{equation}
\lim _ {s \to - 1} \sum _ {\ell = 1} ^ \infty \left ( \ell ^ 2 + m ^ 2 a ^ 2 \right ) ^ {-s} = \zeta _ E (-1,m^2a^2) = -\frac {1} {2} m^ 2 a ^ 2,
\label{naive_S1}
\end{equation}
leads to the same result.
\section{Heat kernel and zeta function \label{section_t-expansion}}
The zeta function regularized expression (\ref{F_zeta_function}) for the free energy (the determinant of a corresponding operator) can be simply postulated. In the previous section we showed how to justify this definition using the dimensional regularization.
However, the dimensional regularization also may involve rather formal manipulations with integrals that do not converge for any real number of dimensions\footnote{Usually the problem is solved by splitting the integral into IR and UV pieces, computing them separately and analytically continuing the result.}. Therefore, in this section we will present yet another derivation using a more physically motivated cut-off regularization. Keeping in mind that we want to preserve the diffeomorphisms it is not a good idea to use the hard cut-off. Instead we employ a smoothed cut-off for regularization~(see \cite{Dietz:1982uc})
\begin{equation}
\mathrm {Reg} \, \sum _ \alpha h(\Lambda _ \alpha) \equiv \sum _ \alpha h(\Lambda _ \alpha) \eta (t \Lambda _ \alpha),
\label{eta_reg}
\end{equation}
where $h(x)$ is an arbitrary function and $\eta(x)$ is a cutoff function, such that $\eta (0) = 1$ and it decays sufficiently fast at infinity
\begin{equation}
\eta(x) \underset{x\to \infty} \to 0,
\end{equation}
so that the sum (\ref{eta_reg}) is convergent. We will be interested in a very specific class of functions $h(x)$, namely $h(x) = x ^ {-s}$ or $h(x) = (x + q) ^ {-s}$ with $q$ being a constant. The claim is that for the asymptotic $t \to 0$ expansion of the sum the $t$-independent piece does not depend (see Appendix~\ref{Riemann_Hurwitz_heat} for examples) on the function $\eta(x)$ and is given by the corresponding zeta function $\zeta _ D(s)$. It is obvious for sufficiently large $s$ when the sum $\sum _ \alpha \Lambda ^ {-s} _ \alpha$ is convergent. We will show that for a specific choice
\begin{equation}
\eta (x) = e ^ {-x},
\end{equation}
which is convenient for practical reasons. In this case defining
\begin{equation}
\theta _ D (s,t) \equiv \sum _ \alpha \Lambda _ \alpha ^ {-s} e ^ {- t \Lambda _ \alpha},
\end{equation}
which for $h(x)=1$ coincides with the heat kernel defined previously (\ref{heat_kernel}), one can show similarly to (\ref{heat_kernel_section}) that for sufficiently large $s'$ the following relation holds
\begin{equation}
\zeta _ D(s+s') \Gamma (s') = \int _ 0 ^ \infty dt \, \theta _ D (s,t) t ^ {s ' - 1}.
\label{zeta_heat_rel_gen}
\end{equation}
For small $t$ the function $\theta _ D (s,t)$ can be expanded in the asymptotic series\begin{equation}
\theta _ D (s,t) \underset{t \to 0} {=} \sum _ {k \geq 0} a _ k (s) t ^ {-s_ k(s)} + O (t).
\label{generalized_heat}
\end{equation}
Then the analytic continuation of (\ref{zeta_heat_rel_gen}) to a region including $s'=0$ is given by
\begin{eqnarray}
\zeta _ D(s+s') \Gamma (s') = & \displaystyle \int _ 0 ^ 1 & dt \, \left [ \theta _ D (s,t) - \sum _ {s _ k \geq 0} a _ k (s) t ^ {-s_ k(s)} \right ] t ^ {s ' - 1} \\
& + &
\frac {a _ {k_ 0}} {s'} + \sum _ {s _ k > 0} \frac {a _ k (s)} {s' + s _ {k} (s)} + \int _ 1 ^ \infty dt \, \theta _ D (s,t) t ^ {s ' - 1}.
\end{eqnarray}
On the l.h.s the residue of the pole at $s'=0$ equals $\zeta _ D (s)$, while on the r.h.s. it corresponds to the coefficient $a _ {k_0}$ for which
$s_{k_0}(s)=0$. Hence, we proved that
\begin{equation}
\theta _ D (s,t) \underset{t \to 0}{=} \sum _ {s_k(s)>0} a _ k (s) t ^ {-s _ k (s)} + \zeta _ D (s) + \sum _ {\tilde s_k(s)>0} \tilde a _ k (s) t ^ {s _ k (s)}.
\label{gen_theta_asympt}
\end{equation}
This expression gives us yet another way of computing the zeta function of a differential operator. At the same time it gives a more physical perspective on the formula (\ref{F_zeta_function}). All terms except for the $t$-independent one depend on the choice of the cut-off function
$\eta (x)$ (see the Appendix~\ref{Riemann_Hurwitz_heat}). However, the terms in (\ref{gen_theta_asympt}) with positive powers of $t$ are irrelevant for the limit $t \to 0$ and those with negative powers of $t$ can be removed with a proper choice of local counterterms, leaving the $t$-independent term (if there are no counterterms scaling as $t^0 \sim a^0$, i.e. there is no $\log$ in (\ref{F_zeta_function})) as a prediction of an observable (for instance free energy or Casimir energy).
\section{Examples \label{section_examples}}
\subsection{$S ^ 1$}
The regularization discussed in the previous section makes it clear why sometimes obviously not legitimate operations -- like the ones in (\ref{naive_S1}) -- lead to the same result as the analytic continuation (\ref{Epstein_S1}). Let us consider again the sum (\ref{zeta_1_S1}).
Its analogue after the regularization with a smoothed cut-off has the form
\begin{equation}
\sum _ {\ell = 1} ^ \infty \left ( \ell ^ 2 + m ^ 2 a ^ 2 \right ) e ^ {- t \left ( \ell ^ 2 + m ^ 2 a ^ 2 \right )} = e ^ {- t \, m ^ 2 a ^ 2} \left ( \sum _ {\ell = 1} ^ \infty \ell ^ 2 e ^ {- t \ell ^ 2} + m ^ 2 a ^ 2 \sum _ {\ell = 1} ^ \infty e ^ {- t \ell ^ 2} \right ),
\label{ex_1}
\end{equation}
where now one is free to regroup the sum in all possible ways, for it converges. Using the formula (\ref{Riemann_Hurwitz_heat_formula}) it is straightforward to show that the expression in braces on the r.h.s. of (\ref{ex_1}) has the following asymptotic expansion for $t\to 0$
\begin{equation}
\frac{\Gamma(3/2)} {t ^ {3/2}} + \zeta(-2) + m^2a^2 \left ( \frac{\Gamma(3/2)} {t ^ {1/2}} + \zeta(0) \right ).
\label{ex_1_rewritten}
\end{equation}
It is clear that the factor $e ^ {- t \, m ^ 2 a ^ 2}$ cannot change the $t$-independent part in (\ref{ex_1_rewritten}) since its expansion runs only over integer powers of $t$, thus so regularized sum reproduces the result from (\ref{Epstein_S1}).
On the other hand the sum (\ref{Casimir_string}) after regularization becomes
\begin{equation}
e ^ {-t \beta} \left ( \sum _ {k=0} ^ \infty k e ^ {-k t} + \beta \sum _ {k=0} ^ \infty e ^ {-k t} \right )
\underset {t \to 0} {=} e ^ {-t \beta} \left [ \frac {1} {t^2} + \zeta (-1) + \beta \left ( \frac {1} {t} + \zeta (0) + 1 \right ) \right ].
\label{ex_2_rewritten}
\end{equation}
The result (\ref{Casimir_string_wrong}) corresponds to the $t$-independent part of the expression in brackets in (\ref{ex_2_rewritten}). The term $e ^ {-t \beta}$ is crucial, for it leads to the following $t$-independent part of the whole sum
\begin{equation}
\zeta (-1) + \beta \left [ 1+ \zeta (0) \right ] - \frac {\beta^2} {2} = \zeta (-1,\beta),
\end{equation}
coinciding with (\ref{Casimir_string}).
Another aspect of the computation that should be stressed is the following. The regularization (\ref{ex_2_rewritten}) leads to only one divergent piece, $1/t^2$, while if one drops the term $e^{-\beta t}$ -- leading to (\ref{Casimir_string_wrong}) -- there are two divergent terms, namely
\begin{equation}
\frac {1} {t ^ 2} + \frac {\beta} {t}.
\end{equation}
It means that in order to get the finite part one has to introduce additional counterterms, which is signaling that the regularization is not covariant.
\subsection{$S ^ 3$}
Now we move to a $3$-dimensional example of a scalar field conformally coupled to a round metric. The quantity to be computed is the following
\begin{equation}
f _ {S ^3} = - \frac {1} {2} \zeta ' _ {S ^ 3} (0) = F _ {S ^ 3} (t) \Big | _ {t\text{-indep}}.
\end{equation}
with
\begin{equation}
F _ {S ^ 3} (t) = \frac {1} {2} \sum _ {\ell=1} ^ \infty \ell ^ 2
\log \left [ \mu ^ {-2} a ^ {-2} \left(\ell + \frac {1} {2} \right) \left(\ell - \frac {1} {2} \right) \right ] e ^ {-t \left(\ell^2 - {1} / {4} \right)}.
\end{equation}
As we saw before (see (\ref{ex_1}) and (\ref{ex_1_rewritten})) the term with $\log \mu^2 a ^2$ does not contribute to the finite $t$-independent part of the sum. Taking into account that for large $\ell$
\begin{equation}
\ell^2 \log (\ell ^2 -1/4) = \ell^2 \log \ell ^ 2 - \frac {1} {4} + O (\ell ^ {-2}),
\end{equation}
we obtain
\begin{equation}
f _ {S ^3} = \frac {1} {2} \sum _ {\ell=1} ^ \infty \left ( \ell^2 \log \ell ^ 2 - \frac {1} {4} \right ) e ^ {-t \left(\ell^2 - {1} / {4} \right)} \Big | _ {t\text{-indep}} +
\frac {1} {2} \sum _ {\ell=1} ^ \infty \left [\ell ^ 2 \log \left(\ell^2 - \frac {1} {4} \right) - \ell^2 \log \ell ^ 2 + \frac {1} {4} \right ],
\label{t_expansion_zeta}
\end{equation}
where in the second sum we dropped the factor $e ^ {- t (\ell^2-1/4)}$, for it is convergent. The last sum can be computed numerically while using the formulas~(\ref{Riemann_Hurwitz_heat_formula}) we get
\begin{equation}
\sum _ {\ell=1} ^ \infty \left ( \ell^2 \log \ell ^ 2 - \frac {1} {4} \right ) e ^ {-t \ell^2} = - \frac {\sqrt{\pi}} {4 \, t ^ {3/2}} \left ( \log 4 t + \gamma -2 \right ) - \frac {\sqrt{\pi}} {8 \, t ^ {1/2}} - \frac {\zeta (0)} {4} - 2 \zeta' (-2).
\label{naive_S3_log}
\end{equation}
Hence, since the additional -- compared to (\ref{naive_S3_log}) -- factor $e ^ {t/4}$ in the first term in (\ref{t_expansion_zeta}) does not change the $t$-independent part of the sum\footnote{This is the reason why the naive computation leads to the correct result. See also the Section~\ref{section_conclusion}.}, we get the same numerical result as in~(\ref{f_conf_scalar}) and~(\ref{S3_zeta_function_reg})
\begin{equation}
f _ {S ^3} = - \zeta ' (-2) - \frac {1} {8} \zeta (0) + 0.02914 = 0.0638.
\end{equation}
We find it also instructive to compute the free energy, as a $t$-independent term in the sum, in a closed form. To this end we use the following identities
\begin{equation}
\sum _ {\ell=1} ^ \infty \frac {e ^ {-t (\ell ^ 2 - 1/4)}} {(\ell-1/2) ^ s} = 2 ^ s \sum _ {k=1} ^ \infty \frac {1} {k ^ s} e ^ {-t k (k + 2)/4}
- \sum _ {k=1} ^ \infty \frac {1} {k ^ s} e ^ {-t k (k + 1)}
\end{equation}
and
\begin{equation}
\sum _ {\ell=1} ^ \infty \frac {e ^ {-t (\ell ^ 2 - 1/4)}} {(\ell+1/2) ^ s} = 2 ^ s \sum _ {k=2} ^ \infty \frac {1} {k ^ s} e ^ {-t k (k - 2)/4}
- \sum _ {k=1} ^ \infty \frac {1} {k ^ s} e ^ {-t k (k - 1)}.
\end{equation}
Rewriting the sum
\begin{equation}
\sum _ {\ell=1} ^ \infty \frac {\ell ^ 2 \, e ^ {-t (\ell ^ 2 - 1/4)}} {(\ell+q) ^ s} = \sum _ {\ell=1} ^ \infty e ^ {-t (\ell ^ 2 - 1/4)}
\left [ \frac {1} {(\ell+q) ^ {s-2}} + \frac {2(1-q)} {(\ell+q) ^ {s-1}} + \frac {(1-q)^2} {(\ell+q) ^ {s}}\right ]
\end{equation}
and using the formula (\ref{heat_formula_1}) we obtain
\begin{eqnarray}
\sum _ {\ell=1} ^ \infty {\ell ^ 2 \log \left ( \ell^ 2 - 1/2 \right ) e ^ {-t (\ell ^ 2 - 1/4)}} & = & - \frac {\sqrt{\pi}} {16 \, t ^ {3/2}} \left ( \log 4 t + \gamma -2\right ) - \frac {1} {8 \, t} - \frac {\sqrt{\pi}} {64 \, t ^ {1/2}} \left ( \log 4 t + \gamma \right ) \nonumber \\
&& +\frac {1} {96} \left (\gamma + \log t \right ) + \frac {1} {12} \log 2 + \frac {1} {4}\zeta ' (-1) - \frac {9 \zeta (3)} {96 \pi ^ 2},
\label{F12}
\end{eqnarray}
and similarly
\begin{eqnarray}
\sum _ {\ell=1} ^ \infty {\ell ^ 2 \log \left ( \ell^ 2 + 1/2 \right ) e ^ {-t (\ell ^ 2 - 1/4)}} & = & - \frac {\sqrt{\pi}} {16 \, t ^ {3/2}} \left ( \log 4 t + \gamma -2\right ) + \frac {1} {8 \, t} - \frac {\sqrt{\pi}} {64 \, t ^ {1/2}} \left ( \log 4 t + \gamma \right ) \nonumber \\
&& +\frac {1} {96} \left ( \gamma - \log t \right ) + \frac {1} {24} \log 2 - \frac {1} {4}\zeta ' (-1) - \frac {9 \zeta (3)} {96 \pi ^ 2}.
\label{F32}
\end{eqnarray}
Putting all terms together produces the following result for the regularized partition function
\begin{eqnarray}
F _ {S ^ 3} (t) =
- \frac {\sqrt{\pi}} {8 \, t ^ {3/2}} \left ( \log 4 t + \gamma -2 \right ) - \frac {\sqrt{\pi}} {32 \, t ^ {1/2}} \left ( \log 4 t + \gamma \right )
+ \frac {1} {16} \left ( 2 \log 2 - \frac {3 \zeta (3)} {\pi ^ 2} \right ).
\label{FS3_t}
\end{eqnarray}
It is clear that there are no covariant counterterms that can be used to remove divergent pieces in (\ref{F12}) and (\ref{F32}) separately, while for the sum (\ref{FS3_t}) the counterterms of the form (\ref{counterterms _S3}) will be sufficient, thus leading to the free energy (\ref{f_conf_scalar}).
\section{Conclusion \label{section_conclusion}}
It is not uncommon that naively computed expressions for certain quantities are divergent. However, predictions for observables should not depend on a regularization chosen to make sense of infinities. For instance, Ward identities (a manifestation of symmetries) should be satisfied. Some of regularizations preserve symmetries by construction others demand symmetry breaking counterterms to restore the symmetry in the final result. The effective field theory perspective tells us in particular that whatever is not forbidden is possible, therefore, all admissible (consistent with symmetries and field/particle content) operators should be included in the Lagrangian. The advantage of regularizations respecting the symmetries in a manifest way is that in this case the space of possible operators in the Lagrangian is restricted, leading to the result in a more economic way.
In computing the Casimir energy or the partition function regularizations preserving diffeomorphism invariance are, for instance, zeta function or the dimensional regularizations. Another regularization (which is very much reminiscent of the zeta function one) was considered in the present paper. According to this regularization for an operator with a (discrete and positive) spectrum $\Lambda _ \alpha$ the finite value of the following sum
\begin{equation}
\sum _ \alpha h(\Lambda _ \alpha),
\label{general_sum}
\end{equation}
is given by the $t$-independent part of the sum regulated using a smoothed cut-off function~$\eta(x)$
\begin{equation}
\mathrm {Finite} \, \sum _ \alpha h(\Lambda _ \alpha) \equiv \sum _ \alpha h(\Lambda _ \alpha) \eta (t \Lambda _ \alpha) \Big | _ {t-\text{indep}},
\end{equation}
such that $\eta (0) = 1$ and it tends to zero sufficiently fast at infinity. This is well physically motivated, since divergent $1/t^{b}$ pieces can be associated with counterterms. The result does not depend on the choice of $\eta (x)$, but in the current paper we considered the case of $\eta (x) = e ^ {-x}$. This regularization in many cases allows to find the sum numerically in a straightforward and easy way. Let's consider eigenvalues labelled by only one quantum number $\lambda _ \ell$ with multiplicity $\mu _ \ell$. Then the regulated sum becomes
\begin{equation}
\sum _ \ell \mu _ \ell \, h (\lambda _ \ell) e ^ {- t \lambda _ \ell}.
\end{equation}
For large $\ell$ the summand can be expanded in a power\footnote{The terms with $\log \ell$ can be expressed as derivatives $\frac {d} {ds} \ell ^{s}$.} series
\begin{equation}
\mu _ \ell f (\lambda _ \ell) \underset{\ell \to \infty}{=} \sum _ k g_ k \ell ^ {b_k} + O (\ell ^ {-2}).
\end{equation}
Hence, the sum can be rewritten as follows
\begin{equation}
\sum _ \ell \sum _ k g_ k \ell ^ {b_k} e ^ {- t \lambda _ \ell} + \sum _ \ell \left [ \mu _ \ell \, h (\lambda _ \ell) - \sum _ k g_ k \ell ^ {b_k} \right ],
\end{equation}
where in the second term we dropped the factor $e ^ {-\lambda _ \ell t}$, since the sum is convergent and can be computed numerically. The asymptotic behavior of the first term can be found using the formulas (\ref{Riemann_Hurwitz_heat_formula}) or similar, thus, leading to the final result.
From this computation it is clear when an apparent misuse of the zeta function regularization, namely, a formal substitution
\begin{equation}
\sum _ \ell \sum _ k \ell ^ {b_k} = \sum _ k \zeta( -b_k ),
\label{zeta_formal}
\end{equation}
may lead to a correct result (see examples in Section~\ref{section_zeta-function} and Appendix~\ref{Riemann_Hurwitz_heat}). If for large $\ell$ the eigenvalues have the following expansion
\begin{equation}
\lambda _ \ell \underset{\ell \to \infty}{=} \ell ^ {a _ 0} + d _ 1 \ell ^ {a _ 1} + d _ 2 \ell ^ {a _ 2}+ \dots,
\end{equation}
it is clear that the formal result (\ref{zeta_formal}) can be obtained if the sum (\ref{general_sum}) is regulated with $e^{-t \ell ^ {a_0}}$, rather than with (leading to the diffeomorphism invariance) $e ^ {-\lambda _ \ell t}$, since
\begin{equation}
\sum _ \ell \ell ^ {b_k} \exp \left ( - t \ell ^ {a _ 0} \right ) = A t ^ {-(b _ k + 1)/ a _ 0} + \zeta( -b_k ).
\label{zeta_naive_non-diff}
\end{equation}
The mismatch between the sums can be found as an asymptotic ($t \to 0$) expansion from the following integral
\begin{equation}
t ^ {(b_k+1)/ a _ 0} \int dx \, x ^ {b_ k} \, \exp \left ( - x ^ {a_0} \right) \left [ 1 + \left (d_ 1 t ^ {1- a_1/a_0} x ^ {b_1} \right ) +
\frac {1} {2} \left (d_ 1 t ^ {1- a_1/a_0} x ^ {b_1} \right ) ^ 2 + \dots \right ]
\label{corrections_naive}
\end{equation}
As a result the naive expression may be correct if none of the terms in (\ref{corrections_naive}) have $t ^ 0$ behavior thus not changing the
$t$-independent part of the sum (\ref{zeta_naive_non-diff}).
\appendices
\section{Euler-Maclaurin formula \label{Riemann_Hurwitz_heat}}
The Euler-Maclaurin formula is an extremely powerful tool approximating sums by integrals. Its derivation can be found in many textbooks, for instance~\cite{Hardy}. We use the following convention for the Euler-Maclaurin formula
\begin{equation}
\sum _ {k=a+1} ^ b f(k) = \int_ a ^ b dx f (x) - B _ 1 \left [ f(b) - f (a) \right ] +
\sum _ {k=1} ^ N \frac {B_ {2k}} {(2k)!} \left [ f^ {(2k-1)} (b) - f^ {(2k-1)} (a) \right ] + R _ N,
\end{equation}
where the remainder is given by
\begin{equation}
R _ N = \int _ 0 ^ 1 \sum _ {k=a} ^ {b-1} f ^{(2N)} (t+k) B _ {2k}(t) dt,
\end{equation}
with $f ^{k}(x)$ denoting the $k$-th derivative of the function $f(x)$ and $B_{k} (t)$ being Bernoulli polynomials defined as coefficients of the following Taylor series
\begin{equation}
\frac {x e ^ {xt}} {e ^ x-1} = \sum _ {k = 0} ^ \infty \frac {B _ n (t)} {n!} x^n.
\end{equation}
For $t=0$ the polynomials become Bernoulli numbers $B _ {k}(0) = B _ k$.
The formula (\ref{gen_theta_asympt}) allows to find the zeta function of the corresponding operator. For example for eigenvalues of the form
$\lambda _ \ell ^ 0 = \ell$ or $\lambda _ \ell ^ q = \ell+q$ choosing the cut-off function to be one of the two
\begin{equation}
e ^ {- \lambda ^ {(0,q)} _ \ell t} ~~ \text{or} ~~ e ^ {- \sqrt {\lambda ^ {(0,q)} _ \ell} t},
\end{equation}
it is straightforward to show using the Euler-Maclaurin formula that for $s<1$ the following asymptotic expansions hold (see for instance~\cite{TerryTao})
\begin{eqnarray}
\sum _ {\ell=1}^\infty \ell ^ {-s} e ^ {-\ell t} & \underset{t \to 0} {=} & \frac {\Gamma (1-s)} {t ^ {1-s}} + \zeta (s) - t \, \zeta(s-1) + O (t^2), \\
\sum _ {\ell=1}^\infty \ell ^ {-s} e ^ {-\ell^2 t} & \underset{t \to 0} {=} & \frac {\Gamma \left (\frac {1-s}{2} \right)} {2 \, t ^ {(1-s)/2}} + \zeta (s) - t \, \zeta (2-s) + O (t^2), \nonumber \\
\sum _ {\ell=0}^\infty \left ( \ell + q \right )^ {-s} e ^ {- \left ( \ell + q \right )t} & \underset{t \to 0} {=} & \frac {\Gamma (1-s)} {t ^ {1-s}} + \zeta (s,q) - t \, \zeta (s-1,q)+ O (t^2), \nonumber \\
\sum _ {\ell=0}^\infty \left ( \ell + q \right )^ {-s} e ^ {- \left ( \ell + q \right )^ 2t} & \underset{t \to 0} {=} & \frac {\Gamma (1-s)} {2 \, t ^ {1-s}} + \zeta (s,q)- t \, \zeta (2-s,q)
+ O (t^2). \nonumber
\label{Riemann_Hurwitz_heat_formula}
\end{eqnarray}
In fact any term in the expansions above can be related to a derivative of the corresponding sum. For instance, for the coefficient in front of
$t ^ {k+1}$, with $1 \leq k \in \mathrm Z$ one gets correspondingly
\begin{equation}
\left [ \sum _ {\ell=1}^\infty \ell ^ {-s} e ^ {-\ell t} \right ] _ {k+1} = \frac{(-1)^k}{k!}\zeta (k-s), ~~ \text{and} ~~
\left [ \sum _ {\ell=1}^\infty \ell ^ {-s} e ^ {-\ell^2 t} \right ] _ {k+1} = \frac{(-1)^k}{k!} \zeta (2k-s).
\end{equation}
Similarly one can prove that for $s<1$ the following asymptotic expansion is valid
\begin{eqnarray}
\sum _ {\ell=1}^\infty \ell ^ {-s} e ^ {-\ell (\ell+q) t} \underset{t \to 0} {=} \displaystyle \frac {t^{-\frac {1-s} {2}}} {2} &\Bigg [& \Gamma \left ( \frac {1}{2}-\frac{s} {2}\right )
- q \Gamma \left ( 1- \frac {s} {2} \right ) t ^ {1/2} + \frac {q^2} {2} \Gamma \left ( \frac {3}{2}-\frac{s} {2}\right ) t \\
&& - \frac {q^3} {3!} \Gamma \left ( 2 -\frac{s} {2} \right ) t ^ {3/2} + \frac {q^4} {4!} \Gamma \left ( \frac {5}{2}-\frac{s} {2}\right ) t^2 +
O (t ^ 2) \, \, \, \Bigg ] + \zeta (s). \nonumber
\label{heat_formula_1}
\end{eqnarray}
Another example we want to show here is related to a misuse of the zeta function regularization. Imagine that the following sum should be computed for a system on a $2$-sphere~\cite{Hellerman:2015nra}
\begin{equation}
\sum _ {\ell=1}^\infty (2 \ell + 1) \sqrt{\ell (\ell+1)}.
\end{equation}
Formally it can be rewritten as
\begin{equation}
\sum _ {\ell=1}^\infty \left ( 2 \ell^2 + 2 \ell + \frac {1} {4} \right ) + \sum _ {\ell=1}^\infty \left [ (2 \ell + 1) \sqrt{\ell (\ell+1)} - 2 \ell^2 - 2 \ell - \frac {1} {4} \right ]
= - 7/4 - 0.015.
\label{conf_sphere_Q_naive}
\end{equation}
At the same time regulating the sum according to the Section~\ref{section_conclusion} leads to
\begin{eqnarray}
\sum _ {\ell=1}^\infty \left ( 2 \ell^2 + 2 \ell + \frac {1} {4} \right ) e ^ {- t \ell(\ell+1)} +
\sum _ {\ell=1}^\infty \left [ (2 \ell + 1) \sqrt{\ell (\ell+1)} - \left ( 2 \ell^2 + 2 \ell + \frac {1} {4} \right ) \right ] \nonumber \\
= \frac {\sqrt{\pi}}{2 t ^ {3/2}} - 1/4 - 0.015.
\label{conf_sphere_Q}
\end{eqnarray}
The mismatch is due to a nontrivial contribution to the $t$-independent part of the first term in (\ref{conf_sphere_Q}).
\section{Pauli-Villars regularization \label{Appendix_PV}}
In the main text we showed that the PV regularized partition function of a scalar coupled to a round $3$-sphere is given by (\ref{FS3}).
One can show that using the Euler-Maclaurin formula the sum associated to the PV regulators in the limit $L \to \infty$ equals to
\begin{eqnarray}
\sum _ {\ell = 1} ^ L \ell ^ 2 \log \left ([ \ell ^ 2 - 1 + \alpha _ {m} ^ 2(\xi) a^2 + M^2_i a^2 \right ]
& = & \frac{2L^3} {9} \left ( 3 \log L - 1 \right ) + L^2 \log L \\
&& + \left [ L a^ 2M _ i ^ 2 + L \left ( \frac {L} {3} + \alpha _ m ^ 2(\xi) a^2 + \frac {5}{6} \right ) \right ] \nonumber \\
&& + \left \{ -\frac {\pi a^3 M _ i ^3} {3} + \frac {a ^ 2 M_ i^2} {2} -\frac {1} {2} \left [\pi a M_i - 1 \right ] \left [\alpha _ m ^ 2(\xi) a^2 - 1 \right] \right \}. \nonumber
\end{eqnarray}
At the same time rewriting
\begin{equation}
\log \left [ \ell ^ 2 - 1 + \alpha _ m ^ 2(\xi) a^2 \right ] = \log \left ( \ell + i \sqrt{\alpha _ m ^ 2(\xi) a^2 - 1} \right ) \left ( \ell - i \sqrt{\alpha _ m ^ 2(\xi) a^2 - 1} \right ),
\end{equation}
and using the formulas (\ref{asymptotic_Hurwitz_zeta}), (\ref{finite_two_zeta}) and the following relation
\begin{equation}
\zeta ' (0,z) = \log \Gamma (z) - \frac {1} {2} \log 2 \pi, ~~ \mathrm {Re} z >0,
\end{equation}
it is straightforward to show that in the limit $L \to \infty$
\begin{eqnarray}
\sum _ {\ell = 1} ^ L \ell ^ 2 \log \left [ \ell ^ 2 - 1 + \alpha _ m ^ 2(\xi) a^2 \right ] & = & \frac {2 L^3} {9} \left (3 \log L - 1 \right ) + L^2 \log L
+ L \left ( \frac {L} {3} + \alpha _ m ^ 2(\xi) a^2 + \frac {5}{6} \right ) \nonumber \\
&& + F _ {S^3}(\sqrt{\alpha _ m ^ 2(\xi) a^2 - 1}) - \frac{1}{6} \left [ \alpha _ m ^ 2(\xi) a^2 - 1\right ].
\end{eqnarray}
Combining the two results one gets the formula (\ref{PV_S3}) for the regularized free energy with
\begin{eqnarray}
\label{finite_FS3_PV}
F _ {S^3} (x) = & - & \frac{1}{2} \left [ \zeta ^{'}(-2,1+i x) + \zeta ^{'}(-2,1-i x) \right ] \nonumber \\
&+& i x \left [ \zeta ^{'}(-1,1+i x) - \zeta ^{'}(-1,1-i x) \right ] \\
&+&\frac{1}{2} x^2 \left [ \zeta ^{'}(0,1+i x) + \zeta ^{'}(0,1-i x) \right ]. \nonumber
\end{eqnarray}
Although the regulator dependent pieces can be removed with the appropriate counterterms the function (\ref{finite_FS3_PV}) should not be considered as a prediction for the finite value of the partition function, for there is always an ambiguity in adding the counterterms of the form
\begin{equation}
b_1 m ^ 3 \mathrm {Vol} _ {S ^ 3} + b_2 m \mathrm {Ric} _ {S ^ 3}.
\end{equation}
\subsection{General manifold $\mathbb M ^ 2 \times S ^ 1 _ \beta$}
PV regularization can be used to find the free energy of a scalar field on an arbitrary manifold of the form $\mathbb M ^ 2 \times S ^ 1 _ \beta$,
where $\beta$ is the size of $S ^ 1$. Indeed, assuming that the eigenvalues of the corresponding operator on $\mathbb M ^ 2$ are
$\Lambda _ \alpha > 0$ we get
\begin{equation}
F = \frac {1} {2} \sum _ {i=0} ^ N \sum _ {\alpha} \sum _ {k = - \infty} ^ \infty c _ i \log \left [ \left ( \frac {2 \pi k} {\beta} \right ) ^ 2
+ \Lambda _ \alpha + M _ i ^ 2 \right ],
\end{equation}
where $c_ 0 = 1$ is the non-regulator contribution ($M_0=0$) and $N$ is the number of regulators needed. Computing the sum in the way similar to the Section~\ref{PV_S1_ex} results in
\begin{equation}
F = \frac {\beta} {2} \sum _ {i=1} ^ N \sum _ {\alpha} c _ i \sqrt {\Lambda _ \alpha + M _ \mu ^ 2} + \sum _ \alpha \log \left ( 1 - e ^ {- \beta \sqrt {\Lambda _ \alpha} } \right ).
\end{equation}
For instance in the case of a round $2$-sphere $S ^ 2 _ a \times S ^ 1 _ \beta$ the procedure leads to
\begin{equation}
F _ {S ^ 2 \times S ^ 1} = \sum _ {\ell = 0} ^ \infty (2 \ell + 1) \log \left \{ 1 - \exp \left [ - \frac {\beta} {a} \sqrt { (\ell + 1/2) ^ 2 + m ^ 2 a^ 2} \right ] \right \}.
\end{equation}
while for a squashed $2$-sphere $S^2 _ {a,c} \times S^1 _ \beta$ we obtain
\begin{equation}
F _ {\text{sq}} = \sum _ {\ell = 0} ^ {\infty} \sum _ {k = -\ell} ^ {\ell} \log \left \{ 1 - \exp \left [ - \frac {\beta} {a} \sqrt { \left ( \ell + \frac {1} {2} \right ) ^ 2
+ k ^ 2 \left ( \frac {a^2} {c^2} - 1 \right ) + m ^ 2 a ^ 2 } \right ] \right \}.
\end{equation}
\newpage
\bibliographystyle{utphys}
|
\section{Introduction}
\label{sec:intro}
The total cross section for proton--proton ($pp$) interactions characterizes a fundamental process of the strong interaction.
Its energy evolution has been studied at each new range of centre-of-mass energies available.
ATLAS has previously reported a measurement of the total cross section in $pp$ collisions at
$\rts=7 \tev$ \cite{alfa_pub_7}.
This paper details a measurement of the total cross section at $\rts=8 \tev$ using data collected in 2012.
The measurement methodology and analysis
technique are very similar between the two measurements and the technical details are discussed
thoroughly in Ref. \cite{alfa_pub_7}.
Both measurements rely on the optical theorem:
\begin{linenomath*}
\begin{equation}\label{eq:OpticalTheorem}
\sigmatot = 4\,\pi \mbox{Im} \, [ f_{\mathrm{el}}\left(t \rightarrow 0\right)] \,
\end{equation}
which relates the total $pp$ cross section $\sigmatot$ to the elastic-scattering amplitude extrapolated
to the forward direction $f_{\mathrm{el}}(t \rightarrow 0)$,
with $t$ being the four-momentum transfer squared. The total cross section can be extracted in different ways using
the optical theorem.
ATLAS uses the \textit{luminosity-dependent} method which requires a measurement of
the luminosity in order to normalize the elastic cross section. Here the measurement benefits from the high-precision luminosity
measurement that ATLAS provides. With this method, $\sigmatot$ is given by the formula:
\begin{equation}\label{eq:totxs}
\sigmatot^{2} = \frac{16\pi(\hbar c)^2}{1+\rho^2} \left. \frac{\mathrm{d}\sigma_{\mathrm{el}}}{\mathrm{d}t}\right|_{t \rightarrow 0} \; ,
\end{equation}
\end{linenomath*}
where $\rho$ represents a small correction arising from the ratio of the real to
the imaginary part of the elastic-scattering amplitude in the forward direction and is taken from global
model extrapolations \cite{PDG_2014}.
The first measurement of $\sigmatot$ at the LHC at 8 $\tev$ was performed by the TOTEM Collaboration \cite{TOTEM_8TeV}
using a \textit{luminosity-independent }method and using data from the same LHC fill as ATLAS. At 7 $\tev$
measurements of $\sigmatot$ were provided by TOTEM \cite{TOTEM_first,TOTEM_second,TOTEM_lumindep} and ATLAS \cite{alfa_pub_7}.
In a recent publication a measurement in the Coulomb--nuclear interference region at very small $t$ was also
reported by TOTEM \cite{TOTEM_8TeV_1km}.
The inelastic cross section $\sigmainel$ can either be derived from the total and elastic cross section measurements as
in Refs. \cite{TOTEM_8TeV,TOTEM_first,TOTEM_second,TOTEM_lumindep,alfa_pub_7} at 7 and 8 $\tev$, or be
determined directly from the measurement of the inelastic rate without exploiting the optical theorem.
These measurements of $\sigmainel$ were performed
at 7 $\tev$ by all LHC collaborations \cite{ATLAS_inel_7,CMS_inel_7,LHCb_inel_7,ALICE_inel_7,TOTEM_inel} and recently
also at 13 $\tev$ by ATLAS \cite{ATLAS_inel_13}.
\section{Experimental setup}
\label{sec:exp}
The ATLAS detector is described in detail elsewhere \cite{atlas1}. The elastic-scattering data were recorded
with the ALFA sub-detector (Absolute Luminosity For ATLAS) \cite{alfa_pub_7}. It consists of Roman Pot (RP)
tracking-detector stations placed at distances of 237 m (inner station) and 241 m (outer station) on either side of the ATLAS interaction point (IP).
Each station houses two vertically moveable scintillating fibre detectors which are inserted in RPs
and positioned close to the beam for data taking.
Each detector consists of 10 modules of scintillating fibres with 64 fibres on both the front and
back sides of a titanium support plate. The fibres are arranged orthogonally in a $u$--$v$-geometry at $\pm45^\circ$
with respect to the $y$-axis. \footnote{ATLAS uses a right-handed coordinate
system with its origin at the nominal IP in the centre of the detector and the $z$-axis along the
beam pipe. The $x$-axis points from the IP to the centre of the LHC ring and the $y$-axis points upwards.}
The spatial resolution of the detectors is
about 35 $\mu$m. Elastic scattering events are recorded in two independent arms of the spectrometer. Arm 1 consists
of two upper detectors at the left side and two lower detectors at the right side, and arm 2 consists inversely of
two lower detectors at the left and two upper detectors at the right side. Events with reconstructed tracks in all four
detectors of an arm are referred to as ``golden'' events \cite{alfa_pub_7}.
The detectors are supplemented with trigger counters consisting of plain scintillator tiles. The detector geometry
is illustrated in Figure~1 of Ref. \cite{alfa_pub_7}. All scintillation signals are detected by photomultipliers coupled
to a compact assembly of front-end electronics including the MAROC chip \cite{MAROC1,MAROC2}
for signal amplification and discrimination. The entire experimental setup is depicted in Figure~2 of Ref. \cite{alfa_pub_7}.
\section{Experimental method}
\label{sec:method}
\subsection{Measurement principle}
The data were recorded in a single run of the LHC with special beam optics \cite{Note90m,HBLumiday11} of
$\betastar=90$ m.
\footnote{The $\beta$-function determines the variation of
the beam envelope around the ring and depends on the focusing properties of the magnetic lattice; its value at the
IP is denoted by $\betastar$.} The same optics were
used at $7 \tev$ \cite{alfa_pub_7}
and result in a small beam divergence with parallel-to-point focusing in the vertical plane.
The four-momentum transfer $t$ is calculated from the scattering angle $\theta^\star$ and the beam momentum
$p$ by:
\begin{linenomath*}
\begin{equation}
\label{eq:t-basic}
-t = \left(\theta^\star \times p \right)^2 \,\,,
\end{equation}
where for the nominal beam momentum $p=3988 \pm 26 \gev$ is assumed \cite{Wenninger} and the scattering angle is
calculated from the proton trajectories and beam optics parameters. The relevant beam optics parameters
are incorporated in transport matrix elements which describe the particle trajectory from the interaction
point through the magnetic lattice of the LHC to the RPs. Several methods were developed for the reconstruction
of the scattering angle, as detailed in Ref. \cite{alfa_pub_7}. The {\em subtraction method} has the best resolution and
is selected as the nominal method.
It uses only the track positions ($w=\{x,y\}$)
and the matrix element $M_{12}=\sqrt{\beta \times \betastar}\sin\psi$,
where $\psi$
refers to the phase advance of the betatron function at the RP:
\begin{equation}
\label{eq:subtraction}
\theta^\star_w = \frac{w_\mathrm{A} - w_\mathrm{C}}{M_{12,\mathrm{A}} + M_{12,\mathrm{C}}} \; \; .
\end{equation}
\end{linenomath*}
Here A refers to the left side of the IP at positive $z$ and C refers to the right side at negative $z$.
Three alternative methods are defined in detail in Ref. \cite{alfa_pub_7}. The {\em local angle method} uses only the $M_{22}$
matrix element and the track angle between the inner and outer detectors.
The {\em local subtraction method} uses a combination of $M_{11}$ and $M_{12}$ matrix elements and both the local angle
and track position.
The {\em lattice method} also uses both track parameters and reconstructs the scattering angle by
an inversion of the transport matrix.
The alternative methods are used
to impose constraints on the beam optics and to cross-check the subtraction method.
\subsection{Data taking}
The low-luminosity, high $\betastar$ run had
108 colliding bunches with about 7$\times$10$^{10}$ protons per bunch,
but only 3 well-separated bunches of low emittance were selected for triggering.
Precise positioning of the RPs is achieved with a
beam-based alignment procedure which determines the position of the RPs with
respect to the proton beams by monitoring the rate of the LHC beam-loss monitors during the RP insertion.
The data were collected with the RPs at a distance of approximately $7.5$ mm from the beam centre, corresponding
to 9.5 times the vertical beam width.
The beam centre and width monitored by LHC beam position monitors and the ATLAS beam-spot measurement \cite{BeamSpot} were
found to be stable to within $10~\mu$m during the run. The beam emittance was derived from the width
of the luminous region in conjunction with the beam optics. It was supplemented by direct measurement from ALFA in
the vertical plane. The luminosity-weighted average of the emittance in the vertical plane was determined
to be 1.6 $\mu$m for both beams and between 1.8 $\mu$m and 2.5 $\mu$m for beam 1 and beam 2 respectively
in the horizontal plane. The emittance uncertainty is about $10\%$.
To trigger on elastic-scattering events a coincidence was required between the A- and C-sides,
where on each side at least one trigger signal in a detector of the corresponding arm was required.
The trigger efficiency
was determined from a data stream recorded with looser conditions to be $99.9\%$ with negligible uncertainty.
The dead-time fraction of the data acquisition system (DAQ) for the selected period was $0.4\%$.
\subsection{Track reconstruction and alignment}
A well-reconstructed elastic-scattering event consists of local tracks from the
proton trajectory in all four ALFA stations.
The reconstruction method assumes that the protons pass through the fibre detector perpendicularly.
The average multiplicity per
detector is about 23 hits, where typically 18--19 are attributed to the proton trajectory
while the remaining 4--5 hits are due to beam-related background, cross-talk and electronic noise.
Tracks are reconstructed in several steps
from the overlap area of the hit fibres and several selections are
applied \cite{alfa_pub_7} in order to reject events with hadronic shower developments.
The precise detector positions
with respect to the circulating beams are crucial inputs
for the reconstruction of the proton kinematics.
First, the distance between the upper and the lower detectors is determined by the use of
dedicated ALFA overlap detectors which allow simultaneous measurements of the same
particle in the upper and lower half of a station.
Then, the detector positions are directly determined from the elastic-scattering data,
using the fact that the high-$\betastar$ optics and the azimuthal symmetry of the scattering angle result
in elastic hit patterns that have an ellipsoidal shape elongated in the vertical direction.
Three alignment parameters are determined for each detector: the horizontal and vertical offsets and the
rotation angle around the beam axis.
For the horizontal offset the centre of the $x$-distribution is taken and the rotation is obtained
from a linear fit to a profile histogram of the $x$--$y$ correlation.
The vertical offset
is obtained from a comparison of the yields in the upper and lower detectors using the sliding window
technique \cite{alfa_pub_7}.
The above procedures provide an independent alignment of each ALFA station.
The vertical alignment parameters are in addition fine-tuned, exploiting the strong correlations between positions of tracks
measured by different detectors in elastic events. First, the positions measured in one detector are extrapolated to the other
detectors in the same arm using the ratio of the appropriate $M_{12}$ matrix elements.
Then, the extrapolated positions are compared to the corresponding measurements -- the average distance gives information
about residual misalignments. The residuals obtained for all pairs of detectors are combined with the vertical offset and
distance measurements in a global $\chi^2$ fit, resulting in the final alignment parameters.
\section{Model for elastic scattering simulation}
\label{sec:thmc}
Several parameterizations are available \cite{Bethe,WestAndYennie,Bourrely_Spin,Cahn,menon_silva,
Block_and_Cahn_curvature,KFK,Selyugin,PhillipsAndBarger,Fagundes,BourrelyAndSoffer} for the differential elastic $pp$ cross section.
A conventional approach is adopted here by taking the following simplified formulae:
\begin{linenomath*}
\begin{eqnarray}
\label{eq:elamplitudes}
\frac{\mathrm{d}\sigma}{\mathrm{d}t} & = & \frac{1}{16\pi}\left|f_{\mathrm{N}}(t) + f_{\mathrm{C}}(t)\mathrm{e}^{\mathrm{i}\alpha\phi(t)}\right|^2 \; \; , \\
f_{\mathrm{C}}(t) & = & -8\pi\alpha\hbar c\frac{G^2(t)}{|t|} \;\; , \\
f_{\mathrm{N}}(t) & = & \left(\rho + \mathrm{i}\right)\frac{\sigmatot}{\hbar c}\mathrm{e}^{-B|t|/2} \;\; ,
\end{eqnarray}
\end{linenomath*}
where $G$ is the electric form factor of the proton, $B$ the nuclear slope, $f_{\mathrm{C}}$ the Coulomb amplitude
and $f_{\mathrm{N}}$ the nuclear amplitude with $\phi$ their relative phase shift.
The value of $\rho=\mathrm{Re}(f_{\mathrm{el}}) / \mathrm{Im}(f_{\mathrm{el}}) = 0.1362 \pm 0.0034$ is taken
from a global fit to lower-energy data \cite{PDG_2014} and parameterizations for $G$ and $\phi$ are given
in Ref. \cite{alfa_pub_7}. This expression is used to fit the data and extract
$\sigmatot$ and $B$.
Monte Carlo simulation of elastic-scattering events is performed with PYTHIA8~\cite{PYTHIA,PYTHIA6} version 8.186
with a $t$-spectrum generated according to Eq.~(\ref{eq:elamplitudes}).
The simulation is used to calculate acceptance and unfolding corrections.
In the simulation the angular divergence of beams at the IP and the spread of the production vertex
are set to the measured values.
Elastically scattered protons are transported from the interaction
point to the RPs nominally by means of the transport matrix. For studies of systematic uncertainties this was also done by the tracking
module of the MadX~\cite{madx} beam optics calculation program. A fast parameterization of the detector response is
used in the simulation and tuned to reproduce the measured difference in position between the outer detectors and
their position as extrapolated from the inner detectors.
\section{Data analysis}
\label{sec:analysis}
\subsection{Event selection}
Events are required to pass the trigger conditions for elastic-scattering events
and have a reconstructed track in all four detectors of an arm in the golden topology.
The fiducial volume
is defined by cuts on the vertical coordinate of the reconstructed track, which
is required to be at least 90 $\mu$m from the detector edge near the beam and at least 1 mm away from
the shadow of the beam screen, in each of the four detectors.
\footnote{The beam screen is a protection element of the quadrupoles, which limits the acceptance of the detector at large $|y|$.}
The values of cuts are chosen to obtain good agreement between data and simulation in the position distributions.
The back-to-back topology of elastic events is further exploited to clean the sample by imposing
cuts on the left-right acollinearity. The difference between the absolute value of the vertical coordinate at the A- and
C-side is requested to be below 3 mm. For the horizontal coordinate the correlation of the A- and C-sides is used.
Events are selected inside an ellipse with half-axis values of $3.5 \sigma$ of the resolution determined by simulation,
as illustrated in Figure~\ref{fig:evsel}(a).
\begin{figure}[h!]
\centering
\includegraphics[width=0.48\textwidth]{evsel_xlr}
\includegraphics[width=0.48\textwidth]{bg1}
\label{fig:background_antigolden1}
\caption{(a) The correlation between the horizontal coordinates on the A- and C-sides.
Elastic-scattering candidates after data quality, trigger and bunch selection but before
acceptance and background rejection cuts are shown. Identified elastic events are required to lie inside the ellipse.
\noindent (b) The distribution $\mathrm{d}N/\mathrm{d}t$, before corrections, as a function of $t$ in arm 1 compared
to the background spectrum determined using anti-golden events.
The results of a simulation of the DPE background is also shown for comparison.}
\label{fig:evsel}
\end{figure}
Elastic events are concentrated inside a narrow ellipse with negative slope, whereas the
beam-halo background appears in broad uncorrelated bands.
The most efficient selection against background is obtained from
the correlation between the position in the horizontal plane and the local angle between two stations,
where events on either side are again required to be inside an ellipse of $3.5\sigma$ width.
From an initial sample of 4.2 million elastic candidates, 3.8 million golden elastic events were selected after all cuts.
The $t$-spectrum, before corrections, for selected elastic events in one arm is shown in Figure~\ref{fig:evsel}(b).
\subsection{Background estimate}
A small fraction of the events inside the selected elliptical area shown in
Figure~\ref{fig:evsel}(a) are expected to be background, predominantly originating from double-Pomeron exchange (DPE)
according to simulations based on the MBR model~\cite{MBR}. The background is estimated
with a data-driven method \cite{alfa_pub_7} using events in the ``anti-golden'' topology
with two tracks in both upper or both lower detectors at the A- and C-sides. This
sample is free of signal and yields an estimate of background in the elastic sample with the golden topology.
The shape of the $t$-spectrum for background events is obtained by flipping the sign of the vertical coordinate
on either side. The resulting background distribution is shown in Figure~\ref{fig:evsel}(b). In total 4400 background
events are estimated to be in the selected sample, corresponding
to a fraction of $0.12\%$ of the selected events. The systematic uncertainty is about 50\%, as derived in
Ref.~\cite{alfa_pub_7} from a comparison of different methods.
\subsection{Reconstruction efficiency}
The rate of elastic-scattering events is corrected for reconstruction inefficiencies.
These events may not be reconstructed
when protons or halo particles interact with the stations or detectors, causing a shower to develop and
resulting in high fibre hit multiplicities.
This correction is called the event reconstruction efficiency and is given by
\begin{linenomath*}
\begin{equation}\label{eq:eff2}
\varepsilon_{\text{rec}} = \frac{N_{\text{reco}}}{N_{\text{reco}}+N_{\text{fail}}} \; ,
\end{equation}
\end{linenomath*}
for each arm where $N_{\text{reco}}$ is the number of reconstructed events and $N_{\text{fail}}$ the number of events for
which the reconstruction failed because a shower developed. The sample of failed events is split into
different categories depending
on the number of detectors with reconstruction failures, because the event background is different for each category.
The fraction of elastic events in the subsample where one out of four detectors failed to reconstruct a track
is above 99\%,
whereas this fraction is 95\% for the subsample where two detectors failed to reconstruct a track on one side.
The event yields in the
different categories are
calculated with a data-driven method, for which the details are given in Ref. \cite{alfa_pub_7}.
The background fraction in the case with only two detectors with reconstructed tracks is estimated with background
templates of the $x$ distribution, obtained from data by selecting single diffractive events.
In the case of a successful track reconstruction in three detectors,
where a good $t$-measurement
is still possible, the partial reconstruction efficiency was verified to be independent of $t$, which is then also
assumed for the other categories.
Events falling outside the acceptance, but faking a signal through shower development, were eliminated from the
reconstruction efficiency calculation by applying another template analysis using the $y$ distribution obtained
from golden elastic events.
The event reconstruction efficiencies in arm 1 and arm 2 are determined to be
$\varepsilon_{\text{rec,1}} = 0.9050 ~ \pm 0.0003 \stat ~ \pm ~ 0.0034 \syst$ and
$\varepsilon_{\text{rec,2}} = 0.8883 ~ \pm ~ 0.0003 \stat ~ \pm ~ 0.0045 \syst$, respectively.
The lower reconstruction efficiency in arm 2 originates from a different amount of material which
induces a higher probability of shower development.
The systematic uncertainty is estimated
by a variation of the selection criteria and templates,
as described in Ref. \cite{alfa_pub_7}.
\subsection{Beam optics}
The precision of the $t$-reconstruction depends on knowledge of the transport matrix elements.
A data-driven method was developed \cite{alfa_pub_7} to tune the relevant matrix elements using
constraints on the beam optics derived from measured correlations in the ALFA data. These constraints
are incorporated in a fit of the strength of the inner triplet quadrupole magnets Q1 and Q3, which
yields an effective beam optics used in the simulation. The values of the constraints are compatible with those
published in Ref. \cite{alfa_pub_7} within $15\%$ and the resulting magnet strength offsets are in good
agreement with the values found at $7~ \tev$.
\subsection{Acceptance and unfolding}
The acceptance is defined as the ratio of events passing all geometrical
and fiducial acceptance cuts to all generated events, and is
calculated as a function of $t$. The form of the acceptance curve as shown in Figure~\ref{fig:acceptance}
\begin{figure}[h!]
\centering
\includegraphics[width=0.5\textwidth]{Acceptance}
\caption{The acceptance as a function of the true value of $t$ for each arm with total uncertainties shown
as error bars. The lower panels show relative total and statistical uncertainties.}
\label{fig:acceptance}
\end{figure}
results from the different contributions
of the vertical and horizontal scattering angles to the value of $t$ and the impact of the fiducial volume cuts
on these contributions. In particular, the position of the peak depends on the cut at large $|y|$ at
the beam screen, which is slightly different for the two arms. The rise of the acceptance at small
$t$ is different in the two arms because of different detector distances, between 8 and 8.4 mm, to the beam.
The measured $t$-spectrum is affected by detector resolution and beam divergence
effects, which are corrected with an unfolding procedure. The $t$-resolution of the subtraction method
is about $10\%$ at small $t$ and $3\%$ at large $t$. The alternative methods have
a $t$-resolution which is a factor of 2--3 worse \cite{alfa_pub_7}.
The background-subtracted distributions
in each arm are corrected for migration effects using an iterative, dynamically stabilized, unfolding
method~\cite{IDS}, which is based on a simulated transition matrix describing the resolution-induced
migration between bins of the $t$-spectrum. The corrections induced by the unfolding are small ($<2\%$) for
the subtraction method except at small $t$ where they rise to $30\%$.
For the other methods the corrections are generally $t$-dependent and increase to $50\%$
at large $t$.
\subsection{Luminosity}
The ATLAS luminosity measurement at high luminosity ($L>10^{33}$cm$^{-2}$s$^{-1}$) is described in detail in
Ref.~\cite{LumiPaper2012}.
Unlike that measurement, the run in this analysis had an instantaneous luminosity
$L \sim0.05\cdot10^{30}$cm$^{-2}$s$^{-1}$, about five orders of magnitude lower.
Only three bunches were present in this run, whereas more than a thousand bunches are common at high luminosity.
The average number of interactions per bunch-crossing (pile-up) in this sample is $\mu\sim0.1$,
which is also low compared to the values of $\mu$=10--40 reached routinely in normal conditions.
At such low values of the luminosity,
some of the standard algorithms are unusable due to lack of sensitivity.
On the other hand, an additional method based on vertex counting in the inner
detector (ID) can be exploited, which is most effective at low pile-up.
Another consequence of the low luminosity is the relative importance of the background sources:
the beam--gas contribution, normally negligible, can become
comparable with the collision rate, while the “afterglow” background (see Ref.~\cite{LumiPaper2012})
becomes conversely less important, due to the small number of colliding bunches.
In 2012, the beam conditions monitor (BCM) was used as the baseline detector for luminosity measurements.
It consists of diamond-sensor detectors placed
on both sides of the IP. It measures the luminosity using an event-counting
method based on the requirement of having activity in either side (BCM\_EventOR).
LUCID (LUminosity measurement with a Cherenkov Integrating Detector) is also located
on both sides of the IP and uses the same algorithm to measure the luminosity (LUCID\_EventOR).
A third method for measuring the per-bunch luminosity is provided by the ID.
Tracks are reconstructed requiring at least nine hits and no missing hits along the track trajectory, and a
transverse momentum $p_{\mathrm{T}}>$ 900 $\MeV$. Then, at least five selected tracks are required to form a primary vertex (VTX5).
The number of primary vertices per event is proportional to the luminosity and provides an
independent method with respect to LUCID and BCM.
The absolute luminosity scale of each algorithm was calibrated by the van der Meer (vdM) method ~\cite{svdm}
at an intermediate luminosity regime ($L\sim10^{30}$cm$^{-2}$s$^{-1}$).
The treatment of both afterglow and beam--gas background is described in detail in
Ref.~\cite{LumiPaper2012}. The first is evaluated by measuring the detector activity
in unfilled bunches preceding the colliding bunches, while the second is estimated from the so-called unpaired bunches,
in which bunches in only one of the two beams are filled and no beam--beam collisions occur.
In the high-$\betastar$ run and for BCM\_EventOR, the afterglow background is evaluated to be 0.05\% and
the beam--gas contribution is 0.4\%.
BCM\_EventOR was chosen as the baseline algorithm for the luminosity determination,
whereas the LUCID\_EventOR and VTX5 methods are only used for the evaluation of systematic uncertainties.
It proved to be the most stable, both by comparing the various vdM calibration
sessions performed during the year and by studying its long-term behaviour at high luminosity. This choice
also ensures maximum compatibility with the high-luminosity case.
By comparing the LUCID\_EventOR and VTX5 results with BCM\_EventOR, a maximum
difference of 0.3\% is found. No change
of this difference with time, or equivalently $\mu$, is observed.
The following contributions to the systematic uncertainty of the luminosity determination are considered:
\begin{itemize}
\item The absolute luminosity scale, common to all algorithms, is determined by the vdM method. Its uncertainty of 1.2\%
is dominated by the beam conditions. This uncertainty is fully correlated between low- and high-luminosity
data sets~\cite{LumiPaper2012}.
\item The BCM calibration stability between the high-$\betastar$ run and the vdM session is estimated to be 0.8\%
by comparing with the VTX5 method among the various vdM scans.
\item The afterglow background uncertainty is conservatively taken as 100\% of the afterglow level itself, which leads to an
uncertainty of 0.05\% in the luminosity.
\item The beam--gas background uncertainty is obtained using LUCID by comparing the difference in the off-time activity
(i.e. produced by beam--gas interactions and not by collisions at the IP) between the colliding and the unpaired bunches.
It is estimated to be 0.3\%.
\end{itemize}
The total systematic uncertainty is therefore 1.5\%.
The final integrated luminosity is measured to be ${L_\mathrm{{int}}} = 496.3 ~ \pm ~ 0.3 ~ (\mathrm{stat.}) ~ \pm ~ 7.3 ~ (\mathrm{syst.})~\mu$b$^{-1}$.
\section{Results}
\label{sec:result}
\subsection{Elastic cross section}
The differential elastic cross section in a given bin $t_i$ is calculated from the following formula:
\begin{linenomath*}
\begin{equation}\label{eq:cross-section}
\frac{\mathrm{d}\sigmael}{\mathrm{d}t_i} = \frac{1}{\Delta t_i}\times \frac{{\cal M}^{-1}[N_i - B_i]}{A_i \times \epsilon^{\mathrm{reco}} \times \epsilon^{\mathrm{trig}} \times \epsilon^{\mathrm{DAQ}} \times L_{\mathrm{int}} }\; \; ,
\end{equation}
\end{linenomath*}
where $\Delta t_i$ is the width of the bins in $t$, ${\cal M}^{-1}$ symbolizes the unfolding procedure applied to the
background-subtracted number of events $N_i - B_i$, $A_i$ is the acceptance,
$\epsilon^{\mathrm{reco}}$ is the event reconstruction efficiency, $\epsilon^{\mathrm{trig}}$ is the trigger efficiency,
$\epsilon^{\mathrm{DAQ}}$ is the dead-time correction and $L_{\mathrm{int}}$ is the integrated luminosity. The binning
in $t$ is chosen to yield a purity above $50\%$, which corresponds to 1.5 times the resolution at small $t$.
It is enlarged at large $t$ in order to account for the lower number of events. The numerical values for the resulting
differential elastic cross section are given in Table~\ref{tab:differential_elastic_crosssection}.
\begin{table}[!htb]
\scriptsize
\begin{center}
\begin{tabular}{S[table-format=1.4] S[table-format=1.4] S[table-format=1.4] S[table-format=3.2] S[table-format=2.2] S[table-format=2.2] S[table-format=2.2]}
\hline \hline
{Low $|t|$ edge} & {High $|t|$ edge} & {Central $|t|$} & {$\mathrm{d}\sigma_{\mathrm{el}}/\mathrm{d}t$} & {Stat. uncert.} & {Syst. uncert.} & {Total uncert.}\\
{[\GeV$^2$]} & {[\GeV$^{2}$]} & {[\GeV$^{2}$]} & {[mb/\GeV$^2$]} & {[mb/\GeV$^2$]} & {[mb/\GeV$^2$]} & {[mb/\GeV$^2$]} \\ \hline
0.0090 & 0.0120 & 0.0105 & 387 & 29 & 14 & 32 \\
0.0120 & 0.0140 & 0.0130 & 370 & 5.6 & 12 & 13 \\
0.0140 & 0.0175 & 0.0157 & 352.3 & 1.4 & 8.7 & 8.9 \\
0.0175 & 0.0210 & 0.0192 & 329.8 & 0.8 & 6.5 & 6.5 \\
0.0210 & 0.0245 & 0.0227 & 306.9 & 0.6 & 5.7 & 5.8 \\
0.0245 & 0.0285 & 0.0265 & 284.6 & 0.5 & 5.0 & 5.1 \\
0.0285 & 0.0330 & 0.0307 & 261.7 & 0.4 & 4.6 & 4.6 \\
0.0330 & 0.0375 & 0.0352 & 239.3 & 0.4 & 4.1 & 4.1 \\
0.0375 & 0.0425 & 0.0400 & 218.0 & 0.4 & 3.6 & 3.6 \\
0.0425 & 0.0475 & 0.0450 & 197.3 & 0.3 & 3.3 & 3.3 \\
0.0475 & 0.0530 & 0.0502 & 178.0 & 0.3 & 3.0 & 3.0 \\
0.0530 & 0.0590 & 0.0559 & 158.8 & 0.2 & 2.7 & 2.7 \\
0.0590 & 0.0650 & 0.0619 & 141.1 & 0.2 & 2.4 & 2.4 \\
0.0650 & 0.0710 & 0.0679 & 126.0 & 0.2 & 2.2 & 2.2 \\
0.0710 & 0.0780 & 0.0744 & 111.1 & 0.2 & 2.0 & 2.0 \\
0.0780 & 0.0850 & 0.0814 & 96.8 & 0.2 & 2.0 & 2.0 \\
0.0850 & 0.0920 & 0.0884 & 84.7 & 0.2 & 1.7 & 1.7 \\
0.0920 & 0.1000 & 0.0959 & 72.9 & 0.2 & 1.6 & 1.6 \\
0.1000 & 0.1075 & 0.1037 & 62.7 & 0.2 & 1.5 & 1.5 \\
0.1075 & 0.1150 & 0.1112 & 54.1 & 0.2 & 1.3 & 1.4 \\
0.1150 & 0.1240 & 0.1194 & 46.11 & 0.14 & 1.13 & 1.13 \\
0.1240 & 0.1330 & 0.1284 & 38.76 & 0.14 & 1.0 & 1.01 \\
0.1330 & 0.1420 & 0.1374 & 32.60 & 0.12 & 0.92 & 0.93 \\
0.1420 & 0.1520 & 0.1468 & 27.10 & 0.11 & 0.82 & 0.83 \\
0.1520 & 0.1620 & 0.1568 & 22.48 & 0.11 & 0.74 & 0.74 \\
0.1620 & 0.1720 & 0.1668 & 18.48 & 0.10 & 0.68 & 0.68 \\
0.1720 & 0.1820 & 0.1768 & 15.25 & 0.09 & 0.67 & 0.68 \\
0.1820 & 0.1930 & 0.1873 & 12.36 & 0.08 & 0.57 & 0.58 \\
0.1930 & 0.2030 & 0.1978 & 10.08 & 0.08 & 0.48 & 0.48 \\
0.2030 & 0.2140 & 0.2083 & 8.20 & 0.07 & 0.43 & 0.43 \\
0.2140 & 0.2250 & 0.2193 & 6.58 & 0.06 & 0.33 & 0.33 \\
0.2250 & 0.2360 & 0.2303 & 5.34 & 0.06 & 0.27 & 0.28 \\
0.2360 & 0.2490 & 0.2422 & 4.28 & 0.05 & 0.24 & 0.24 \\
0.2490 & 0.2620 & 0.2552 & 3.30 & 0.05 & 0.22 & 0.23 \\
0.2620 & 0.2770 & 0.2691 & 2.47 & 0.04 & 0.18 & 0.18 \\
0.2770 & 0.3000 & 0.2877 & 1.69 & 0.03 & 0.14 & 0.14 \\
0.3000 & 0.3200 & 0.3094 & 1.06 & 0.03 & 0.10 & 0.1 \\
0.3200 & 0.3500 & 0.3335 & 0.62 & 0.02 & 0.08 & 0.08 \\
0.3500 & 0.3800 & 0.3635 & 0.36 & 0.04 & 0.04 & 0.05 \\ \hline
\end{tabular}
\caption{The measured values of the differential elastic cross section with statistical and systematic uncertainties.
The central $t$-values in each bin are calculated from simulation, in which a slope parameter of $B = 19.7 \GeV^{-2}$ is used.}
\label{tab:differential_elastic_crosssection}
\end{center}
\end{table}
The experimental systematic uncertainties are derived according to the methods detailed in Ref. \cite{alfa_pub_7}
as follows:
\begin{itemize}
\item The value of the beam momentum used in the $t$-reconstruction (Eq.~(\ref{eq:t-basic})) and in the simulation is
varied by $0.65\%$, as recommended in Ref.~\cite{Wenninger}.
\item The uncertainty in the luminosity of $1.5\%$ is applied to the cross-section normalization.
\item The event reconstruction efficiency is varied by its uncertainty of about $0.5\%$ and the uncertainty in the tracking efficiency
is estimated by varying the reconstruction criteria.
\item The uncertainties originating from the effective beam optics are calculated from variations of the
optics constraints, of the strength of the quadrupoles not adjusted in the fit, and of the quadrupole alignment
constants. Additional uncertainties are related to the error of the optics fit,
to the beam transport scheme used in the simulation, and to the impact from a residual
beam crossing angle assumed to vary within its uncertainty of
$\pm$10 $\mu$rad.
\item The uncertainties from the alignment of the ALFA detectors are evaluated by varying the correction
constants for horizontal and vertical offsets as well as the rotation
within their uncertainties as determined from variations of the
alignment procedures, and by taking the difference between different optimization configurations for the vertical alignment parameters.
\item The background normalization uncertainty of $50\%$ is applied in the background subtraction and
the background shape
is varied by inverting the sign of different detector combinations.
\item The detector resolution values in the fast simulation are replaced by estimates from GEANT4~\cite{GEANT41,GEANT42} and test-beam
measurements, and a $y$-dependent resolution is used instead of a constant value.
\item The value of the nuclear slope in the simulation is varied around the nominal value of
$19.7 \gev^{-2}$ by $\pm 1 \gev^{-2}$, corresponding to about five times the uncertainty of the measured $B$ value.
\item The beam emittance value in the simulation is varied by its uncertainty of about $7\%$. Additionally, the ratio of
the emittance in beam 1 to the emittance in beam 2, which are measured by wire scans after injection only, is set to unity.
\item The intrinsic unfolding uncertainty is estimated from a data-driven closure test.
\end{itemize}
The main sources of systematic uncertainty are the beam momentum uncertainty and the luminosity uncertainty. For each
systematic uncertainty source the shift of the cross-section value in each $t$-bin is recorded.
The most important shifts are shown in Figure~\ref{fig:tfit}(a).
\subsection{Total cross section}
A profile fit~\cite{profile} is used to determine $\sigmatot$.
It includes statistical and systematic uncertainties
and their correlations across the $t$-spectrum. For each shift due to a systematic uncertainty a nuisance parameter is fitted in
a procedure described in Ref.~\cite{alfa_pub_7}.
The theoretical prediction of Eq.~(\ref{eq:elamplitudes}) including the Coulomb and interference terms is fitted to the data
to extract $\sigmatot$ and $B$ alongside the nuisance parameters, as shown in Figure~\ref{fig:tfit}(b).
\begin{figure}[h!]
\centering
\includegraphics[width=0.48\textwidth]{shift_summary}
\includegraphics[width=0.48\textwidth]{tfit}
\caption{(a) Relative shifts in the differential elastic cross section as a function of $t$ for
selected systematic uncertainty sources. Shown are the uncertainties related to the beam energy, to the crossing angle, to the
modelling of the detector resolution in the simulation (MC resolution), to the beam optics (kQ5Q6, magnet strength), to the value
of $B$ in the simulation (Physics model) and to the emittance.
(b) The fit of the theoretical prediction to the
differential elastic cross section with $\sigmatot$ and $B$ as free parameters.
In the lower plot the points represent the relative difference between fit and data,
the yellow area represents the total experimental uncertainty and the hatched area the statistical component.
The red line indicates the fit range; the fit result is extrapolated in the lower plot outside the fit range.
The upper right insert shows a zoom of the data and fit at small $t$.
}
\label{fig:tfit}
\end{figure}
The fit range is chosen to be from $-t = 0.014 \GeV^2$ to $-t = 0.1 \GeV^2$,
where the lower bound is set by requiring the acceptance to exceed $10\%$ and the upper
bound is chosen to exclude the large-$t$ region where theoretical models predict
deviations from a single exponential function \cite{per_theory}.
The fit yields
$\sigmatot = 96.07\pm0.86~\mbox{mb}$ and $ B = 19.74\pm0.17 \GeV^{-2}$ with $\chi^2/N_{\mbox{dof}}=17.8/14$
and the uncertainties include all statistical and experimental systematic contributions. The most important
uncertainty component is the luminosity error for $\sigmatot$ and the beam energy error for $B$.
Additional uncertainties arising from the extrapolation $t\rightarrow 0$ are
estimated from a variation of the upper end of the fit range respectively up to $-t=0.152 \GeV^2$
and up to $-t=0.065 \GeV^2$, and from a variation of the lower end,
i.e. from $-t=0.009 \GeV^2$ to $-t=0.0245 \GeV^2$. Further theoretical uncertainties considered include:
a variation of the $\rho$-parameter in Eq. (\ref{eq:OpticalTheorem}) by $\pm 0.0034$; the replacement of the dipole parameterization by a double-dipole
parameterization~\cite{A1} for the proton
electric form factor; the replacement of the Coulomb phase from West and Yennie \cite{WestAndYennie}
by parameterizations from Refs.~\cite{Cahn,KFK}; the inclusion of a term related to the magnetic moment
of the proton in the Coulomb amplitude~\cite{Bourrely_Spin}. The dominant extrapolation uncertainty is induced
by the fit range variation. The final results for $\sigmatot$ and $B$ are:
\begin{eqnarray}\label{eq:profilefit}
\sigmatot & = & \mbox{96.07} \; \pm \mbox{0.18} \; (\mbox{stat.}) \pm \mbox{0.85} \; (\mbox{exp.}) \pm \mbox{0.31} \; (\mbox{extr.}) \; \mbox{mb} \; , \\
B & = & 19.74 \; \pm \mbox{0.05} \; (\mbox{stat.}) \pm \mbox{0.16} \; (\mbox{exp.}) \pm \mbox{0.15} \; (\mbox{extr.}) \; \GeV^{-2} \; .
\end{eqnarray}
A summary of the results for $\sigmatot$ from four different $t$-reconstruction methods
is given in Table~\ref{tab:sigma_tot_syst_summary}.
\begin{table}
\begin{center}
\begin{tabular}{l|S[table-format=2.2] | S[table-format=2.2] S[table-format=2.2] S[table-format=2.2]}
\hline \hline
& \multicolumn{4}{c}{$\sigmatot$ [mb]} \\
& {Subtraction} & {Local angle} & {Lattice} & {Local subtraction} \\ \hline
Total cross section & 96.07 & 96.52 & 96.56 & 96.58 \\
Statistical error & 0.18 & 0.15 & 0.16 & 0.15 \\
Experimental error & 0.85 & 0.94 & 0.88 & 0.89 \\
Extrapolation error & 0.31 & 0.42 & 0.23 & 0.23 \\ \hline
Total error & 0.92 & 0.98 & 0.93 & 0.93 \\ \hline \hline
\end{tabular}
\caption{The total cross section and uncertainties for four different
$t$-reconstruction methods. The nominal results are based on the subtraction method, quoted in the second column.}
\label{tab:sigma_tot_syst_summary}
\end{center}
\end{table}
The results from the nominal subtraction method are in good agreement with the other methods, considering the
uncorrelated uncertainty of 0.3--0.4 mb. The alternative methods are correlated through the common use
of the local angle variable.
Further stability checks are carried out in order to cross-check the fitting method. A fit using only the covariance
matrix of statistical uncertainties yields $\sigmatot = 96.34 \; \pm \mbox{0.07} \; (\mbox{stat.})$
in good agreement with the results from the profile fit Eq.~(\ref{eq:profilefit}).
The same fit with only statistical uncertainties was also performed for the two arms of ALFA independently
and gave consistent results within one standard deviation of the statistical uncertainty. The data sample was split into
ten sub-periods with roughly equal numbers of selected events and no dependence of the measured
value of $\sigmatot$ on time was observed.
Also, the data from the three different bunches were investigated independently
and found to give consistent results.
Finally the stability of the analysis was tested by a wide variation of
the event selection cuts.
The largest change of $\sigmatot$ with
these cut variations was observed for the cut on the correlation between $x$ and $\theta_x$. That produced a change of
$\pm 0.3$ mb, well within the $t$-dependent experimental systematic uncertainty of about $0.5$ mb.
Several alternative parameterizations
~\cite{WestAndYennie, Block_and_Cahn_curvature, Selyugin, KFK, PhillipsAndBarger,Fagundes, BourrelyAndSoffer} of the
differential elastic cross section, including non-exponential forms at large $t$, were used to fit the spectrum
up to $-t=0.3 \GeV^2$ in order to assess the impact on the value of the total cross section. The RMS of the values
obtained is $0.28$ mb, in good agreement with the quoted extrapolation uncertainty of $0.31$ mb assigned to the simple
exponential form.
The TOTEM Collaboration exploited data from the same LHC fill for a measurement of $\sigmatot$ using the
luminosity-independent method. Their result is $\sigmatot = 101.7 \pm 2.9$ mb \cite{TOTEM_8TeV}, higher than
the measurement presented here. The difference corresponds to $1.9 \sigma$ assuming uncorrelated uncertainties.
Better agreement is observed in the nuclear slope measurement, where TOTEM reports $B=19.9 \pm 0.3 \GeV^{-2}$,
a value very close to the present result $B=19.74 \pm 0.19 \GeV^{-2}$, which indicates that the difference is confined to
the normalization.
The measurements of ATLAS and TOTEM are compared to measurements at lower energy and to a global
fit ~\cite{PDG_2014} in Figure~\ref{fig:CrossSectionS}(a) for $\sigmatot$ and in Figure~\ref{fig:CrossSectionS}(b) for $B$.
\begin{figure}[h!]
\centering
\includegraphics[width=0.48\textwidth]{running_plot}
\includegraphics[width=0.48\textwidth]{BslopeS}
\caption{(a) Comparison of total and elastic cross-section measurements presented here with other
published measurements~\cite{PDG_2014,TOTEM_second,Auger,ARGO-YBJ,AKENO,FlysEye} and model predictions as a function
of the centre-of-mass energy. (b) Comparison of the measurement of the nuclear slope $B$ presented here with other
published measurements at the ISR, at the S$p\bar{p}$S, at RHIC, at the Tevatron
and with the measurement from TOTEM at the LHC.
The red line shows a model calculation \cite{Scheg}, which contains a linear term and quadratic term in $\ln s$.}
\label{fig:CrossSectionS}
\end{figure}
TOTEM also reported evidence of non-exponential behaviour of the differential elastic cross section
\cite{TOTEM_nonexp} in the $-t$-range below $0.2\GeV^2$, where deviations from the single exponential form
of the order of one percent are observed. Such effects cannot be substantiated with this data set
because their size is below the systematic uncertainties of the present measurement.
As well as the total cross section, the total integrated elastic cross section can
be calculated,
provided that the
Coulomb amplitude is neglected. In this case, $\sigmael$ can be obtained from the formula
\begin{linenomath*}
\begin{equation}
\label{eq:elastic_extrapolation}
\sigmael \; = \;
\frac{\sigmatot^2}{B} \; \frac{1+\rho^2}{16\pi(\hbar c)^2} \; ,
\end{equation}
\end{linenomath*}
and the result is $\sigma_{\mathrm{el}} = 24.33 \pm 0.04 \; (\mbox{stat.}) \pm 0.39 \; (\mbox{syst.}) \; \mbox{mb}$.
The measured integrated elastic cross section in the fiducial range from $-t=0.009$ \GeV$^2$ to $-t=0.38$ \GeV$^2$
corresponds to $80\%$ of this total elastic cross section
$\sigmael^{\mathrm{observed}} = 19.67 \pm 0.02 \;(\mbox{stat.}) \pm 0.33 \; (\mbox{syst.}) \; \mbox{mb}$.
The total inelastic cross section is determined by subtraction of the total elastic cross section
from the total cross section. The resulting value is
$\sigmainel = 71.73 \pm 0.15 \; (\mbox{stat.}) \pm 0.69 \; (\mbox{syst.}) \; \mbox{mb}$.
\FloatBarrier
\section{Conclusion}
\label{sec:conclusion}
ATLAS has performed a measurement of the total cross section from elastic $pp$ scattering at $\rts=8 \TeV$.
The measurement is based on $500$ $\mu$b$^{-1}$ of collision data collected in a high-$\betastar$ run at the LHC
in 2012 with the ALFA Roman Pot sub-detector. The optical theorem
is used to extract the total cross section from the differential elastic cross section by extrapolating
$t \rightarrow 0$. The differential cross section is also used to determine the nuclear slope.
The analysis uses data-driven methods to determine relevant
beam optics parameters and event reconstruction efficiency, and to tune the simulation.
The detailed evaluation of the associated systematic uncertainties is supplemented by a comparison of
$t$-reconstruction methods with different sensitivities to beam optics.
The absolute luminosity for this run is determined in a dedicated analysis,
taking into account the special conditions with a very low number of interactions per bunch crossing.
The total cross section at $\rts=8 \TeV$ is determined to be
\begin{eqnarray*}
\sigmatot(pp\rightarrow X) & = & \mbox{96.07} \; \pm \mbox{0.18} \; (\mbox{stat.}) \pm \mbox{0.85} \; (\mbox{exp.}) \pm \mbox{0.31} \; (\mbox{extr.}) \; \mbox{mb} \; ,
\end{eqnarray*}
where the first error is statistical, the second accounts for all experimental systematic uncertainties and the last
is related to uncertainties in the extrapolation $t\rightarrow 0$.
In addition, the slope of the elastic differential cross section at small $t$ is determined to be
$B = \mbox{19.74} \; \pm \mbox{0.05} \; (\mbox{stat.}) \pm \mbox{0.23} \; (\mbox{syst.}) \; \GeV^{-2}$.
The total elastic cross section is extracted from the fitted parameterization as
$\sigma_{\mathrm{el}}(pp\rightarrow pp) = \mbox{24.33} \; \pm \mbox{0.04} \; (\mbox{stat.}) \pm \mbox{0.39} \; (\mbox{syst.}) \; \mbox{mb}$
and the inelastic cross section is obtained by subtraction from the total cross section as
$\sigma_{\mathrm{inel}} = \mbox{71.73} \; \pm \mbox{0.15} \; (\mbox{stat.}) \pm \mbox{0.69} \; (\mbox{syst.})\; \mbox{mb}$.
The measurements at $8\tev$ are significantly more precise than the previous measurements at $7\tev$ because
of the smaller luminosity uncertainty and a larger data sample.
\section*{Acknowledgements}
\input{Acknowledgements}
\printbibliography
\newpage
\input{atlas_authlist}
\end{document}
|
\section{Introduction and Summary of Results}
Slow-roll inflation relies on flat scalar potentials, making axion-like fields ideal inflaton candidates. This is especially true in the context of large-field inflation. The latter is of particular interest since, on the one hand, it is arguably the most natural form of inflation and, on the other hand, it will be discovered or experimentally ruled out in the foreseeable future.
The flatness of axion potentials (we denote the axion henceforth by $\theta$) is protected by a shift symmetry which is only broken non-perturbatively, i.e.~by instantons. However, possible problems with consistently embedding axionic models of inflation in quantum gravity are an issue of continuing concern \cite{9502069, 0303252, 0601001,12035476,14022287,14095793, 14123457,150300795,150303886,150304783,150307853,150307912,150400659, 150403566,150603447,150800009,150906374,150907049,151002005,151105119, 151200025,151203768,160206517,160505311,160608437,160608438,160706105}. In particular, the focus has recently been on the Weak Gravity Conjecture \cite{0601001}. In the context of axions, it states that with growing axion decay constant $f_{\text{ax}}$ the action $S$ of the `lightest' instanton decreases, such that the flatness of the potential is spoiled by corrections $\sim \exp(-S)$.
However, the Weak Gravity Conjecture has not been firmly established. In particular, its validity remains unclear outside the domain of UV completions of quantum gravity provided by the presently understood string compactifications. This is even more true for the extension to axions. Moreover, the prefactors of the $\exp(-S)$ corrections mentioned above may be parametrically small, especially if SUSY or the opening up of extra dimensions come to rescue just above the inflationary Hubble scale.
Thus, it is useful to pursue the related but complementary approach of constraining axionic potentials on the basis of gravitational instantons. Indeed, the very fundamental statement that quantum gravity forbids global symmetries is, in the context of shift symmetries, explicitly realised by instantonic saddle points of the path integral of Euclidean quantum gravity. These are also known as Giddings-Strominger wormholes \cite{Giddings:1987cg}. If, as proposed in \cite{150303886}, gravitational instantons yield significant contributions to the axion potential, some models of natural inflation would be under pressure (at least those with one or only few axions like alignment scenarios), while axion-monodromy inflation models seem to be unaffected.\footnote{Natural inflation \cite{Freese:1990rb} with one axion requires a transplanckian axion field space. Ideas for realising natural inflation in a subplanckian field space of multiple axions were proposed in \cite{0409138, 0507205, 14047496, 150301015}. For models implementing these ideas see e.g.~\cite{07103883, 9804177, 14012579, 14037507, 14044268, 14045235, 14046209, 14046988, 14047127, 14047773, 14047852, 14060341, 14072562, 14091221,14104660, 14114768, 150301777, 150302965, 150307183, 150307912, 160506299}. Axion monodromy inflation was introduced in \cite{08033085, 08080706} (for a field theory implementation see \cite{08111989,11010026}). A realisation of this idea with enhanced theoretical control is $F$-term axion monodromy \cite{14043040,14043542,14043711}. For further work in this context see \cite{09121341, 14053652, 14057044, 14097075,14112032,14115380, 150301607, 150402103, 150507871, 151001522, 151008768, 151108820, 151204463, 160701680}.}
It is our goal to study the effect of Euclidean wormholes and that of related instantonic solutions in detail. In particular, in the spirit of what was said above, we want to be as model-independent and general as possible, ideally relying only on Einstein gravity and the additional axion. The goal is to constrain large classes of string models or even any model with a consistent UV completion. As we go along, we will however be forced to consider certain model-dependent features and take inspiration from the known part of the string theory landscape.
The aim of this paper is thus to determine the strongest constraints on axion inflation due to gravitational instantons. One important aspect of our analysis is that -- to be as model-independent as possible -- calculations are performed in an effective 4-dimensional Einstein-axion(-dilaton) theory. However, this theory is only valid up to an energy-scale $\Lambda$ and, for consistency, we have to make sure that our analysis only includes gravitational instanton solutions within the range of validity of our effective theory.
This leads to the following challenge pointed out in \cite{150307912} (see also \cite{9502069}) and which we will repeat here. Given an energy cutoff $\Lambda$, gravitational instantons within the range of theoretical control contribute at most as $\delta V \sim e^{-S} \sim e^{-M_p^2 / \Lambda^2}$ to the axion potential. Then, gravitational instantons are dangerous for inflation if their contribution to the potential is comparable to the energy density in the inflationary sector, i.e.~$\delta V \sim H^2$. If the cutoff $\Lambda$ is not much above $H$ gravitational instantons are clearly harmless. However, if $\Lambda$ is close to $M_p$ gravitational instantons can easily disrupt inflation. As a result, the importance of gravitational instantons for inflation hinges on a good understanding of the scale $\Lambda$ where the 4-dimensional Einstein-axion(-dilaton) theory breaks down.
To arrive at a quantitative expression for $\Lambda$ requires some knowledge about the UV completion of our theory. Here, we take string theory as our model of a theory of quantum gravity, i.e.~we assume that the effective Einstein-axion(-dilaton) theory is derived from string theory upon compactification. String compactifications give rise to a hierarchy of scales as shown in \autoref{Fig:scales}. Inflation is assumed to take place below the moduli scale $m_{\textrm{mod}}$ where only gravity and one or more axions are dynamical. Above $m_{\textrm{mod}}$ further scalars in the form of moduli become dynamical. As a result, if we want to work with a Einstein-axion theory the cutoff $\Lambda$ is the moduli scale.
\begin{figure}
\centering
\begin{overpic}[width=0.9\textwidth]{scales.pdf}
\put (8,2) {$H$} \put (28,2) {$m_{\textrm{mod}}$} \put (49,2) {$m_{KK}$} \put (71,2) {$m_s$} \put (90,2) {$M_{p}$}
\end{overpic}
\caption{Hierarchy of scales in a string model of inflation.}
\label{Fig:scales}
\end{figure}
Here, we want to do better. An analysis using 4-dimensional gravitational instantons can in principle be valid up to the Kaluza-Klein (KK) scale $m_{KK}$, at which a description in terms of a 4-dimensional theory breaks down. However, to be able to go beyond $m_{\textrm{mod}}$ we have to allow for dynamical moduli. Hence, for this purpose Einstein-axion theories are insufficient and we have to study gravitational instantons in Einstein-axion-moduli theories instead.
These considerations give rise to the following structure of our paper. We start by recalling the Giddings-Strominger or Euclidean wormhole solution \cite{Giddings:1987cg} in \autoref{Section:Gravitational Instanton Solutions }. This is a classical solution of the axion-gravity system which gives space-time a `handle' with cross-section $S^3$. In fact, this solution can be interpreted as a real saddle point of the path integral only in the dual 2-form theory. We take some care to describe the relevant subtleties of the dualisation procedure in \autoref{Subsection: Dualisation}. Subsequently, we generalise to the case with an additional dilatonic scalar in \autoref{GravInstScalar}. Now extremal as well cored instanton solutions \cite{0406038,150906374} also exist. The situation with a dilaton is important for us as a model of the realistic string-phenomenology case with light moduli. \autoref{Subsection: Integration constant} focusses on the way in which cored and extremal gravitational instantons may arise from a Euclidean black 0-brane in an underlying 5d theory. In this way we obtain a UV-completion of cored and extremal gravitational instantons, which can then be understood by parameters of the 5d theory.
\autoref{Section:Instanton Potentials from Wormholes} is devoted to the crucial issue whether a scalar potential is induced by Euclidean wormholes. We will provide an explicit computation of the contributions to the axion potential from Euclidean wormholes. Thereby, we describe how to circumvent a recent counter-argument given in \cite{150906374}, suggesting that Euclidean wormholes could not break the axionic shift-symmetry. Thus, we stress that Euclidean wormholes are by no means less important than cored or extremal gravitational instantons.
In \autoref{Section:Cored Grav Inst} we calculate the instanton actions for Euclidean wormholes as well as for cored and extremal gravitational instantons. We also give a quantitative answer to the question which gravitational instantons can be trusted within our effective theory with cutoff $\Lambda$. The result is as follows. As in the case of gauge instantons one can associate gravitational instantons with an instanton number $n$. Given an energy cutoff $\Lambda$ one can then only trust gravitational instantons with a sufficiently high instanton number $n \gg f_{\textrm{ax}} M_p / \Lambda^2$, where $f_{\textrm{ax}}$ is the axion decay constant.\footnote{Note that this implies that we neglect potentially more severe, but incalculable contributions due to instantons with low instanton numbers.}
In \autoref{Section:Moduli stabilisation} we take first steps towards studying gravitational instantons in the presence of dynamical moduli. We argue that the case with one light modulus coupled to the Einstein-axion theory can be modelled by an Einstein-axion-dilaton theory with massless dilaton. For one, in \autoref{sec:addmass} we show that for our purposes the modulus potential can be neglected if there is a sufficient hierarchy between the modulus mass and the cutoff $\Lambda$. The reason is that deep inside the `throat' of a gravitational instanton the modulus mass only gives a subleading contribution to the stress-energy tensor, while curvature and gradient terms dominate. As this region is also the source of the dominant part of the instanton action, we conclude that the action obtained for a massless modulus will remain a good approximation even in the massive case. We then motivate our restriction to moduli with dilatonic couplings. This implies that the modulus $\varphi$ is coupled to the axion $\theta$ through the kinetic term for the axion as $e^{\alpha \varphi} (\partial \theta)^2$. In \autoref{Subsection: Dilaton coupling from string theory} we review that dilatonic couplings arise frequently in string compactification.
In \autoref{Section:Inflation} we analyse possible constraints for inflation due to gravitational instantons. To this end we identify the instantons with the largest contributions to the axion potential in \autoref{sec:mostdanger}. We arrive at the strongest constraint if the cutoff $\Lambda$ is as high as possible. In \autoref{sec:selfdual} we identify the highest possible cutoff $\Lambda_{\textrm{max}}$ for an effective 4-dimensional theory arising from a string compactification. This is given by the KK scale of a compactification with smallest possible compactification volume, which we take as the self-dual volume under T-duality. Unfortunately, there is an ambiguity in this definition of $\Lambda_{\textrm{max}}$ up to factors of $\pi$, which can be crucial. We then determine the maximal contribution $\delta V$ to the axion potential due to gravitational instantons and compare this to the scale of inflation in models of large-field axion inflation. Our main result is as follows. We find that gravitational instantons do not give rise to strong \emph{model-independent} constraints on axion inflation. Extremal gravitational instantons may be important for inflation, but this is model-dependent, as the size of their contribution depends on the value of the dilaton coupling $\alpha$.
Last, in \autoref{Section: WGC} we record some observations regarding the Weak Gravity Conjecture (WGC) \cite{0601001} in the context of gravitational instantons. We pick up the idea from \cite{150906374} that extremal instantons play the role extremal charged black holes for the WGC. We then find that the WGC appears to be satisfied due to the existence of Euclidean wormholes. This either hints at a realisation of the WGC in the context of gravitational instantons, or implies a different definition of the WGC in the presence of wormholes.
We summarise our findings in \autoref{Section:Conclusions} and point out directions for future work. Various appendices contain detailed computations on which some of our results are based, or clarify subtleties which are not absolutely essential for the understanding of the main body of the paper.
Overall, our analysis leaves us with the following: a semi-classical approach to quantum gravity via gravitational instantons does not give rise to strong constraints for large-field inflation. Thus, if quantum gravity has anything to say about large field inflation, the quantum part will have to speak.
\section{Gravitational Instanton Solutions} \label{Section:Gravitational Instanton Solutions }
A model of axion inflation will necessarily involve an axionic field coupled to gravity. One feature of such a system is that it may allow for \textit{gravitational instantons}, i.e.~finite-action solutions to the equations of motion of the Euclidean axion-gravity theory.
Our starting point is the Euclidean action for an axionic field $\theta$ coupled to gravity, which takes the form ($M_p=1$)
\begin{equation} \label{Action with axion}
S = \int d^4x \sqrt{g} \left[-\frac{1}{2}R + \frac{1}{2}K g^{\mu\nu} \partial_{\mu} \theta \partial_{\nu} \theta \right] \ .
\end{equation}
The prefactor $K$ can in principle depend on further fields. In this section we ignore the Gibbons-Hawking-York boundary terms, because we will be focussing on the dynamics of the system. Instead of working with the axionic field $\theta$, one can write the action in terms of the dual 2-form $B$ and its field strength $H=dB$:
\begin{equation} \label{Action with H}
S = \int d^4x \sqrt{g} \left[-\frac{1}{2}R + \frac{1}{2}\mathcal{F} H_{\mu \nu \rho} H^{\mu \nu \rho} \right] \ ,
\end{equation}
where $\mathcal{F}= 1/(3! K)$. The field strength $H$ is related to $d \theta$ via
\begin{equation}
H = K \star d \theta \ .
\end{equation}
The dualisation from \eqref{Action with axion} to \eqref{Action with H} must be done under the path integral using Lagrange multipliers. We will explain this in the following subsection.
In Euclidean space the theory of the 3-form $H$ coupled to gravity \eqref{Action with H} then has non-trivial solutions. In particular, \textit{gravitational instantons} are rotationally symmetric solutions with metric
\begin{equation} \label{Metric}
ds^2 = \left( 1+ \frac{C}{r^4} \right)^{-1} dr^2 + r^2 d \Omega_3^2,
\end{equation}
where the parameter $C$ arises as a boundary condition or integration constant (see \autoref{Appendix: metric structure}).
For $C<0$ this is known as a \textit{Giddings-Strominger or Euclidean wormhole} \cite{Giddings:1987cg}: for large $r$ it approaches flat space, while for decreasing $r$ the geometry exhibits a throat with cross-section $S^3$. At $r=|C|^{1/4}$ one encounters a coordinate singularity, where another solution of this type can be attached (see e.g.~\autoref{Figure: Wormhole and 1-Cycle} and \ref{Figure: Types of Grav. Inst.}(a) for two possibilities). Gravitational instanton solutions for $C=0$ and $C>0$ can also be found if a dilaton-type field is included \cite{0406038,150906374}.
\begin{figure}
\centering
\includegraphics[width=0.70\linewidth]{wormhole_2}
\caption{This picture illustrates a Euclidean wormhole, whose two ends are connected to the \textit{same} asymptotically flat space. Then there is a non-trivial 1-cycle (dotted line) passing through the wormhole. The cycle orthogonal to this 1-cycle is a $S^3$ (symbolised by the dashed line around the right-hand throat).}
\label{Figure: Wormhole and 1-Cycle}
\end{figure}
Before we extend our system to dilaton-type couplings, we review and discuss several subtleties involved in the aforementioned dualisation between $\theta$ and $B$ in Euclidean space.
\subsection{Dualisation} \label{Subsection: Dualisation}
For the sake of clarity, in this subsection we index the field variables by their rank, i.e.~we write $\theta_0$ and $B_2$. Those fields are sourced by an instanton and a microscopic string, respectively. We start from the two Euclidean actions in 4d:\footnote{The appearance of the $i$-factor in front of the coupling terms can be understood by writing these terms as $\int_M f_p \wedge j_{4-p}$ with $p$-form field $f_p$ and source current $j_{4-p}$. One of the relevant tensor components of either $f_p$ or $j_{4-p}$ then always carries a zero-index and hence acquires an $i$-factor by Wick rotation.}
\begin{align} \label{Action Theta}
S[\theta_0] &= \int_M \frac{1}{2g_{\theta}^2} F_1 \wedge \star F_1 + iQ_{\theta} \int_I \theta_0, ~~~~~~~~F_1 = d \theta_0, \\ \label{Action B}
S[B_2] &= \int_M \frac{1}{2g_{B}^2} H_3 \wedge \star H_3 + iQ_B \int_{\sigma} B_2, ~~~~H_3 = dB_2,
\end{align}
where $M$ denotes our 4-manifold, $I$ the set of points where the instantons are located, and $\sigma$ is the surface swept out by the string. One can identify the kinetic terms of \eqref{Action Theta} and \eqref{Action B} by imposing
\begin{equation} \label{Dual relation for B and theta}
H_3 = g_B^2 \star F_1
\end{equation}
and $g_B^2 = 1/g_{\theta}^2$. This now becomes a single theory with both strings and instantons allowed and either $\theta_0$ or $B_2$ to be used locally as the appropriate field variable.
Note that the $H_3$-flux is quantised by
\begin{equation} \label{Charge quantisation condition}
\int_{S^3} H = n \in \mathbbm{Z},
\end{equation}
as we review in \autoref{Appendix: Charge Quantisation} in the context of the existence of fundamental strings and instantons.
We now couple the 1-form/3-form theory to gravity. It is well-known that choosing either $\theta_0$ or $B_2$ as the fundamental field leads to Einstein equations differing by an overall sign \cite{Giddings:1987cg}. Indeed, the action of \eqref{Action Theta} gives the energy-momentum tensor
\begin{equation} \label{Energy-Momentum tensor theta}
T_{\mu\nu}^{(\theta)} = \frac{1}{g_{\theta}^2} \left(- \frac{1}{2} g_{\mu\nu}(\partial \theta_0)^2 + \partial_{\mu} \theta_0 \partial_{\nu}\theta_0 \right)\,,
\end{equation}
while \eqref{Action B} leads to
\begin{align} \label{Energy-Momentum tensor B}
T_{\mu\nu}^{(B)} &= \frac{1}{g_{B}^2} \left(- \frac{1}{2 \cdot 3!} g_{\mu\nu} H_3^2 + \frac{1}{2} H_{\mu \rho \sigma}H_{\nu}^{~\rho \sigma}\right) = \\ \nonumber
&= -\frac{1}{g_{\theta}^2} \left(- \frac{1}{2} g_{\mu\nu}(\partial \theta_0)^2 + \partial_{\mu} \theta_0 \partial_{\nu}\theta_0 \right) = - T_{\mu\nu}^{(\theta)}\,.
\end{align}
In the second line we used \eqref{Dual relation for B and theta} together with $g_B^2 = 1/g_{\theta}^2$.
The above sign difference implies that Euclidean wormholes exist in the $B_2$ but not in the $\theta_0$ formulation. Technically, this is due to the Hodge star being introduced before or after the variation w.r.t. the metric. Also at the intuitive level the difference is clear: The $H_3$-flux on the transverse $S^3$, which is fixed due to the Bianchi identity, supports the finite-radius throat. By contrast, the dual quantity $\theta^{\prime} \equiv \partial_r \theta$, i.e. the variation of $\theta$ along the throat, is \textit{not} fixed by the dual Bianchi identity and the solution is lost.
We note that the \textit{Minkowski-space} Einstein equations remain the same on both sides of the duality. However, we are interested in the path integral in the \textit{Euclidean} theory with gravity, so this observation does not help.
Thus, one may wonder whether Giddings-Strominger wormholes do contribute to the action or whether the dual descriptions are really fully equivalent. This problem has been intensively investigated in the past, see e.g.~\cite{Giddings:1987cg, Grinstein:1988ja,Lee:1988ge,Abbott:1989jw,Brown:1989df,Burgess:1989da, Coleman:1989zu,9502069,9511080,9604038,9608065,9701093,0510048,Collinucci:2005opa, 07052768,10116301} and our present understanding mainly derives from \cite{0510048,Collinucci:2005opa,07052768}.
Indeed, it should be possible to resolve the problem by dualising under the Euclidean path integral and following the fate of the instanton solution. We review the dualisation following \cite{Burgess:1989da,0510048,Collinucci:2005opa}. To be specific, let $M$ be a cylinder, $M=S^3\times I$, with an interval $I\subset \mathbb{R}$. This is the simplest relevant topology since the $S^3$ can carry $H_3$-flux, supporting a narrow throat somewhere within $I$.
Starting on the $B_2$-side, the partition function reads
\begin{equation}
Z \sim \int_{\text{b.c.}} d[B_2] \exp \left(- \int_M \frac{1}{2g_{B}^2} dB_2 \wedge \star dB_2\right)\,,
\end{equation}
where ``b.c.'' denotes the boundary conditions $B_2(S^3_I) \equiv B_2^{(I)}$ and $B_2(S^3_F) \equiv B_2^{(F)}$ at the initial and final boundaries $S^3_I$ and $S^3_F$. The possibility of a non-trivial flux, $\int_{S^3} H_3\neq 0$, can as usual be implemented by defining $B_2$ in patches over the transverse $S^3$ and choosing appropriate transition functions.
One can also express $Z$ as a path integral over $H_3$, imposing $dH_3=0$ with the help of a Lagrange-multiplier $\theta_0$:
\begin{equation} \label{Partition function with Lagrange multiplier}
Z \sim \int_{\text{b.c.}} d[H_3]d[\theta_0] \exp \left\{- \int_M \frac{1}{2g_{B}^2} \left(H_3 \wedge \star H_3 + 2 ig_B^2 \theta_0 dH_3 \right) \right\}\,.
\end{equation}
The previous $B_2$-boundary conditions now translate into boundary conditions on the pullback\footnote{
This is \textit{not} the same as $H_3$ at the position of the boundaries, which contains time-derivatives of $B_2$ and should not be constrained.
}
of $H_3$ to the initial and final boundary, i.e.~$H_3(S^3_I) \equiv H_3^{(I)}$ and $H_3(S^3_F) \equiv H_3^{(F)}$. In this language, the information about a possible $H_3$-flux is simply part of the $H_3$ boundary conditions. The $\theta_0$-integral is unconstrained. The $i$ in front of the Lagrange-multiplier is needed to get a delta-functional $\delta(dH_3)$ in the path integral after integrating out $\theta_0$. Hence, we have $dH_3=0$ and Stokes theorem yields $H_3^{(I)} = H_3^{(F)}$. In other words $Z \sim \delta(H_3^{(I)} - H_3^{(F)})$.
Equation \eqref{Partition function with Lagrange multiplier} can be rewritten
by integrating the second term by parts and completing the square:
\begin{align} \label{Gaussian Path Integral H_3}
Z \sim \int_{\text{b.c.}} & d[H_3]d[\theta_0] \exp \left\{ - i \int_{\partial M} \theta_0H_3 \right\} \\ \nonumber
&\exp \left\{- \int_M \frac{1}{2g_{B}^2} \left[ \left(H_3 - ig_B^2 \star d \theta_0\right) \wedge \star \left(H_3 - ig_B^2 \star d \theta_0\right) +g_B^4 d\theta_0 \wedge \star d \theta_0 \right]\right\}.
\end{align}
According to \cite{Burgess:1989da,0510048,Collinucci:2005opa} one can now shift the variable $H_3 \to \tilde{H}_3\equiv H_3 - ig_B^2 \star d\theta_0$ and trivially perform the Gaussian integral. One may however also be concerned about this step since, for any fixed $\theta_0$, the boundary conditions, e.g. $\tilde{H}_3(S^3_I)=H_3(S^3_I)- ig_B^2 \star d\theta_0$, clash with the saddle point value $\tilde{H}_3=0$ of the Gaussian integral in the interior of $M$.
To make this issue more explicit, let us write $H_3=\langle H_3\rangle + \delta H_3$, where $\langle H_3\rangle$ is constant along the $S^3$ but time dependent. Its boundary values are determined by the $H_3$-flux. Furthermore, decompose $\delta H_3$ into spherical harmonics on $S^3$. If the cylinder $M$ were flat and gravity non-dynamical, we would now simply have a quantum mechanical system of infinitely many, independent oscillators. The dualisation process sketched above would correspond, as is well known from $T$-duality for a scalar field on the cylinder $S^1\times\mathbbm{R}$, to a canonical transformation ($p\leftrightarrow q$) for each oscillator. In our case, the dual variables are coefficients of the spherical harmonic decomposition of $\theta_0$.
Let us focus on the most interesting subsystem (see also the discussion in \cite{07052768}) with the variable $\langle H_3\rangle\sim p$ and the dual variable $\langle \theta_0\rangle\sim q$. Thus, we first restrict our attention to the question whether it is correct to naively integrate out $q$ in
\begin{equation}
Z\sim \int_{\mathrm{b.c.}}d[p]\,\int d[q]\,\exp\left\{-\frac{1}{2}\int_{t_i}^{t_f} dt\,\left[(p-i\dot{q})^2+\dot{q}^2\right]\right\}\,.
\end{equation}
Based on an explicit, discretised calculation in \autoref{Appendix: Dualisation}, we claim this is indeed the case. One can now argue that, also for the full system \eqref{Gaussian Path Integral H_3} including all oscillators and gravity, this formal manipulation with path integrals is correct. It will then also remain correct if, as argued in \autoref{Appendix: Charge Quantisation}, $\langle H_3\rangle$ is initially quantised, i.e. $\int_{S^3}\langle H_3 \rangle = n \in \mathbb{Z}$. Indeed, this quantisation is `neutralised' once the Lagrange multiplier is introduced and the now continuous variable $\langle H_3\rangle $ is integrated out as above.
As a result of all this the partition function can eventually be given as
\begin{equation} \label{Partition function theta}
Z \sim \int d[\theta_0] \exp \left(-\int_M \frac{1}{2g_{\theta}^2} d\theta_0 \wedge \star d \theta_0 - i \int_{\partial M} \theta_0H_3\right),
\end{equation}
where $g_{\theta}^2=1/g_B^2$. We emphasise that the sign of the kinetic term is the one required for a well-defined Euclidean path integral. This sign will become important below. We also note that this procedure can be straightforwardly generalised to any $p$-form in arbitrary dimensions $d >p$. Moreover, we observe that despite the shift $H_3 \to H_3 - ig_B^2 \star d\theta_0$, the field $\theta_0$ can be kept real (see also \cite{Collinucci:2005opa}).\footnote{In other references, e.g.~\cite{9701093,10116301}, the axion field was taken to be imaginary. Then, however, we do not see how to ensure $dH_3=0$ using \eqref{Partition function with Lagrange multiplier}.}
Varying the action in \eqref{Partition function theta},
\begin{equation}
\delta S = \int_M\frac{1}{g_{\theta}^2} \delta \theta_0 d \star d \theta_0 - \int_{\partial M}\frac{1}{g_{\theta}^2} \delta \theta_0 \star d \theta_0 - i \int_{\partial M} \delta \theta_0H_3 \stackrel{!}{=}0\,,
\end{equation}
we find the equation of motion $d \star d \theta_0=0$ in the bulk and
\begin{equation} \label{Complex saddle}
H_3(\partial M) = \frac{i}{g^2_{\theta}} \star d \theta_0 (\partial M)
\end{equation}
at the boundary. Thus, the $\theta_0$ path integral has only complex saddle points \cite{0510048,Collinucci:2005opa}.\footnote{
For
a treatment of path integrals with complex phase space or complex saddles, see e.g.~\cite{10096032} and \cite{151000978}, respectively.
}
Indeed, the possibility of taking $\theta_0$ imaginary at stationary points was discussed before, see e.g.~\cite{Burgess:1989da,Coleman:1989zu}.
To summarise, dualisation leads to a Euclidean path integral in which $\theta_0$ is a priori real and the kinetic term has the standard sign. However, a semi-classical evaluation is only possible on the basis of complex saddles. Crucially, the relevant field-theory solutions then also solve Einstein equations because imaginary $\theta_0$ flips the sign of $T_{\mu\nu}^{(\theta)}$ (cf. \eqref{Energy-Momentum tensor B}). Thus, one can expect gravitational instantons to contribute consistently both in the $B_2$ and the $\theta_0$ formulation. Nevertheless, it is natural to use the $B_2$ path integral to keep the saddle points real \cite{0510048,Collinucci:2005opa}, and we will do so in what follows.
\subsection{Gravitational Instantons in the Presence of a Massless Scalar Field}
\label{GravInstScalar}
One goal of this paper is to study the effect gravitational instantons can have on geometric moduli of string compactifications. In the 4-dimensional theory these moduli appear as scalar fields. Consequently, we will study systems of an axion $\theta$ and a scalar $\varphi$ coupled to gravity.\footnote{A string compactification will typically give rise to many axionic fields and many geometric moduli. We focus here on one, potentially super-Planckian, light axion which may be identified with the inflaton. Similarly, the scalar can be identified with the lightest modulus. Note that the analysis in this subsection neglects any mass term for the modulus $\varphi$, which will be included only later in \autoref{Section:Moduli stabilisation}.} The relevant Euclidean action then takes the form
\begin{equation} \label{Action with complex scalar}
S = \int d^4x \sqrt{g} \left[-\frac{1}{2}R + \frac{1}{2} K(\varphi) g^{\mu\nu} \partial_{\mu} \theta \partial_{\nu} \theta + \frac{1}{2} g^{\mu\nu} \partial_{\mu} \varphi \partial_{\nu} \varphi \right].
\end{equation}
Here we already canonically normalised the field $\varphi$. At 2-derivative level, the axion $\theta$ can only enter the action through a term $\partial_{\mu} \theta \partial^{\mu} \theta$ due to its shift symmetry. There is no such symmetry for $\varphi$ and hence the kinetic term for $\theta$ can in general depend on $\varphi$. This situation is typically encountered in string compactifications, see \autoref{Subsection: Dilaton coupling from string theory} for examples. In this subsection we consider a massless scalar field $\varphi$ and apply the subsequent results to the case of a massive scalar in \autoref{Section:Moduli stabilisation}.
As we are interested in gravitational instantons, we should consider the dual formulation of the above theory. The relevant Euclidean action is then
\begin{equation} \label{Action without potential
S = \int d^4x \sqrt{g} \left[-\frac{1}{2}R + \frac{1}{2} \mathcal{F}(\varphi)H^2 + \frac{1}{2} g^{\mu\nu} \partial_{\mu}\varphi \partial_{\nu} \varphi \right],
\end{equation}
where $\mathcal{F}=1/(3!K)=1/(3!f_{\text{ax}}^2)$. Here $f_{\text{ax}}$ is the $\varphi$-dependent analogue of the familiar axion decay constant.
In the following we will review explicit solutions of this system corresponding to gravitational instantons. Following \cite{150906374} we will construct solutions to the equations of motion for the metric, the 3-form $H$ and the scalar $\varphi$.
\subsubsection*{General solution}
For completeness, let us recall the metric given in \eqref{Metric}:
\begin{equation}
\nonumber ds^2 = \left( 1+ \frac{C}{r^4} \right)^{-1} dr^2 + r^2 d \Omega_3^2 .
\end{equation}
The derivation of the functional form of $g_{rr}$ can be found in \autoref{Appendix: metric structure}.
There we show that the equation of motion for $g_{rr}$ decouples from the equations of motion of the massless fields $\varphi$ and $B$. In particular, the form of the metric \eqref{Metric} is independent of the functional form of the kinetic terms of these fields.
The constant $C$ can a priori be negative, positive or zero. Depending on the sign of this parameter $C$, this solution has the following interpretations. Using the terminology of \cite{0406038,150906374} we can distinguish between three types of gravitational instantons (see \autoref{Figure: Types of Grav. Inst.} for an illustration).
\begin{itemize}
\item \textit{Euclidean wormholes} ($C<0$):\newline
The case $C<0$ leads to a geometry with a throat and we call this solution a Euclidean wormhole. The divergence of $g_{rr}$ at $r =r_0 \equiv |C|^{1/4}$ is only a coordinate singularity. The Ricci scalar $R$ is
\begin{equation}
R = 6 \frac{C}{r^6}
\end{equation}
and thus it is finite for all $r \geq r_0$. The locus $r=r_0$ can then be interpreted as the end of one wormhole throat. We can then attach another solution of this type at $r=r_0$ which can either be attached to our universe (see \autoref{Figure: Wormhole and 1-Cycle}) or a different universe (see \autoref{Figure: Types of Grav. Inst.}). In this paper we will only consider wormholes which close again in our universe, i.e.~we are dealing with pairs of holes each connected by a ``handle''.
\item \textit{Extremal instantons} ($C=0$):\newline
The solution for $C=0$ is called an extremal gravitational instanton \cite{0406038,150906374}. Even though space is flat in that case, the fields $\varphi$ and $\theta$ still exhibit a nontrivial profile. This is possible due to a complete cancellation of terms in the energy-momentum tensor \cite{Rey:1989xj}.
\item \textit{Cored gravitational instantons} ($C>0$):\newline
The case $C>0$ gives rise to a geometry with a curvature singularity at $r=0$. Such solutions are called \textit{cored gravitational instantons} \cite{150906374}.
\end{itemize}
\begin{figure}
\subfigure[Wormhole connected to another universe]{\includegraphics[width=0.32\textwidth]{wormhole_1}}
\subfigure[Extremal gravitational instanton]{\includegraphics[width=0.32\textwidth]{extremal_instanton}}
\subfigure[Cored gravitational instanton]{\includegraphics[width=0.32\textwidth]{cored_instanton}}
\caption{The three types of gravitational instantons are depicted. (a) Euclidean wormhole connecting two asymptotically flat spaces. It is also possible to connect both ends to the same space as shown in \autoref{Figure: Wormhole and 1-Cycle}. (b) Extremal gravitational instanton: in this case space is flat everywhere. The cross in the middle indicates the locus $r=0$. (c) Cored gravitational instanton: there is a curvature singularity at $r=0$.}
\label{Figure: Types of Grav. Inst.}
\end{figure}
Having reviewed the solution for the metric, we will now solve the equation of motion for $H$ without specifying $\mathcal{F}(\varphi)$. From \eqref{Action without potential} we obtain the equation of motion:
\begin{equation} \label{EoM for B}
d \star H = - \frac{\mathcal{F}^{\prime}(\varphi)}{\mathcal{F}(\varphi)} d \varphi \wedge \star H.
\end{equation}
We expect that solutions for $\varphi$ and $H$ should exist that respect the spherical symmetry of the background. We thus propose that $\varphi = \varphi(r)$. Similarly, following \cite{Giddings:1987cg}, we make the ansatz
\begin{equation} \label{Ansatz for H}
H = h(r) \epsilon
\end{equation}
with $\epsilon$ the volume form on $S^3$ such that
\begin{equation}
\int_{S^3} \epsilon = 2 \pi^2 r^3 \ .
\end{equation}
From \eqref{Ansatz for H} it follows that $\star H \sim h(r) dr$ and the LHS of \eqref{EoM for B} vanishes. As we have chosen $\varphi= \varphi(r)$ the RHS of \eqref{EoM for B} equally vanishes and the equation of motion for $H$ is satisfied.
In addition, $H$ also has to satisfy the Bianchi identity $dH=0$. This enforces
\begin{equation} \label{Solution for h(r)}
h(r) = \frac{n}{Ar^3},
\end{equation}
with $A \equiv A(S^3)=2\pi^2$ the area of the unit sphere. Charge quantisation \eqref{Charge quantisation condition} implies that $n \in \mathbbm{Z}$.
In order to find the solution for $\varphi$ it is sufficient to consider the $rr$-component of the
Einstein equations, $G_{rr} = T_{rr}$, which can be shown to be equivalent to the Klein-Gordon equation for $\varphi$. It reads
\begin{equation} \label{Einstein equation without potential}
\frac{1}{2} \left(1+ \frac{C}{r^4}\right) (\varphi^{\prime})^2 - \frac{3 \mathcal{F}(\varphi)n^2/A^2 + 3C}{r^6} = 0 \ ,
\end{equation}
where we already used the solution for $H$. We also defined $\varphi' \equiv \partial \varphi / \partial r$. The solution for $\varphi$ can then be found by integrating this differential equation.
\subsubsection*{Model-dependent solutions}
From \eqref{Einstein equation without potential} it is clear that explicit solutions for $\varphi$ will depend on the functional form of the term $\mathcal{F}(\varphi)$. In this subsection we will restrict our attention to functions of the form $\mathcal{F}(\varphi) \sim \exp(- \alpha \varphi)$, where we choose without loss of generality $\alpha >0$, as this functional form arises frequently in string compactifications. For example, this behaviour is observed when $\varphi$ is identified with the dilaton. Similarly, the same functional form appears if $\varphi$ corresponds to the volume modulus in setups with large compactification volume (e.g.~\cite{0502058}) or if $\varphi$ is a complex structure modulus at large complex structure. We will study such examples in \autoref{Subsection: Dilaton coupling from string theory}.
To be specific, we take
\begin{equation}
\mathcal{F}(\varphi) =\frac{1}{3!f_{\text{ax}}^2 } \exp(-\alpha \varphi) \ ,
\end{equation}
where $f_{\text{ax}}$ is from now on a constant. The value of the parameter $\alpha$ will depend on the type of geometric modulus. We can assume $\lim_{r \to \infty} \varphi(r) =0$ without loss of generality. Then $f_{\text{ax}}$ will correspond to the asymptotic value of the axion decay constant.
In the following, we will summarise the explicit solutions for $\varphi$ for the Euclidean wormhole, the extremal instanton and for the cored instanton. Further details can be found in \autoref{Appendix: analytical solution to Einstein}.
\begin{itemize}
\item \textit{Euclidean Wormhole} ($C<0$):\newline
The analytical solution to \eqref{Einstein equation without potential} in this case is \cite{Giddings:1987cg,0406038}
\begin{equation} \label{Solution for C<0, with constant}
e^{\alpha \varphi (r)} = \frac{1}{\cos^2(K_-)} \cos^2 \left(K_- + \frac{\alpha}{2} \sqrt{\frac{3}{2}} \arcsin \left(\frac{\sqrt{|C|}}{r^2}\right)\right) \ .
\end{equation}
Here, we already implemented the boundary condition $\lim_{r \to \infty}\varphi(r) =0$, which also implies that
\begin{equation} \label{C for wormholes}
C = - \frac{n^2}{3!f_{\text{ax}}^2A^2} \cos^2(K_-).
\end{equation}
The integration constant $K_-$ is not a free parameter. This can be seen as follows. When the field reaches the wormhole throat at $r=r_0 \equiv |C|^{1/4}$, the factor $(1+C/r^4)$ in \eqref{Einstein equation without potential} vanishes, hence
\begin{equation} \label{Condition finiteness of velocity}
3 \mathcal{F}(\varphi(r_0))n^2/A^2 + 3C = 0 \ .
\end{equation}
Using \eqref{C for wormholes}, this translates to
\begin{equation}
\cos^2 \left(K_-+ \frac{\alpha \pi}{4}\sqrt{\frac{3}{2}}\right) =1,
\end{equation}
and thus
\begin{equation}
K_- = - \frac{\alpha \pi}{4}\sqrt{\frac{3}{2}}.
\end{equation}
Inserting this back into the solution yields
\begin{equation} \label{Solution for C<0 w/o constant K_-}
e^{\alpha \varphi (r)} = \frac{1}{\cos^2(\sqrt{3/2} \alpha \pi/4 )} \cos^2 \left(\frac{\alpha}{2} \sqrt{\frac{3}{2}} \arccos \left(\frac{\sqrt{|C|}}{r^2}\right)\right).
\end{equation}
To see that one can take two wormhole solutions and glue them together, let us now change coordinates by writing $r= a(t)$ such that the metric becomes
\begin{equation} \label{Metric with coordinate t}
ds^2 = dt^2 + a^2(t) d \Omega_3^2 \ .
\end{equation}
One can show that $a(t)$ and $\varphi(t)$ are symmetric under $t \to - t$. This implies the existence of a ``handle'' as shown in \autoref{Figure: Wormhole and 1-Cycle}, assuming also that the two throats are very distant in $\mathbb{R}^4$.
Interestingly, not all values for $\alpha$ will lead to physically acceptable solutions.
Note that $\varphi(r)$ is regular everywhere on $r \in [|C|^{1/4}, +\infty)$ only for dilaton couplings in the range $0 \leq \alpha < 2 \sqrt{2/3}$. For $\alpha > 2 \sqrt{2/3}$ there is always a value of $r> |C|^{1/4}$, where $e^{\alpha \varphi(r)}=0$, i.e.~$\varphi(r) \to - \infty$. This is consistent with \cite{Giddings:1987cg,9502069,0406038}. In our case the field $\varphi$ corresponds to the string coupling or a geometric modulus of the string compactification. A runaway behaviour $\varphi(r) \to - \infty$ is then pathological as it would correspond to a limit of decompactification or vanishing string coupling. In all these cases new light states will appear resulting in a loss of control over the effective theory. This pathology is avoided for $\alpha = 2 \sqrt{2/3}$. However, in this case we obtain $C=0$ which will be discussed next. Overall, we find that only the range $0 \leq \alpha < 2 \sqrt{2/3}$ is physically allowed for Euclidean wormholes.
Last, note that the limit $\alpha \to 0$ can be identified with the Giddings-Strominger wormhole \cite{Giddings:1987cg} which exhibits a constant dilaton profile.
\item \textit{Extremal Instanton} ($C=0$):\newline
For the case of an extremal instanton we find
\begin{equation} \label{Solution for C=0}
e^{\alpha \varphi(r)} = \left(1+ \frac{\alpha n}{4A f_{\text{ax}}} \frac{1}{r^2}\right)^2,
\end{equation}
which is valid for all $\alpha > 0$. (For $\alpha = 2 \sqrt{2/3}$ this solution agrees with \eqref{Solution for C<0 w/o constant K_-}. A plot of the dilaton profile in this case can be found in \autoref{Fig: Dilaton profile for C<0}.). The result can be obtained most easily by solving \eqref{Einstein equation without potential} for $C=0$. Notice that \eqref{Solution for C=0} with a minus sign in the bracket would in principle also be a solution (see \autoref{Appendix: analytical solution to Einstein}), but then there would again be a value of $r>0$ so that $e^{\alpha \varphi}=0$, leading to the same problems as described above. We hence exclude this possibility.
\begin{figure}
\subfigure[]{\includegraphics[width=0.50\textwidth]{PlotWormholes2_2}}
\subfigure[]{\includegraphics[width=0.50\textwidth]{PlotWormholes2_3}}
\caption{Illustration of dilaton profiles. The values of $r$ and $\varphi$ are in Planck units. \newline (a) Euclidean wormhole ($C<0$): \newline Here we choose $n / f_{\text{ax}}$ such that $C= - \cos^2 \left( \alpha \pi \sqrt{3/2}/4\right)$ and plot for $\alpha=1$. \newline (b) Extremal instanton ($C=0$) with $\alpha= 2 \sqrt{2/3}$.}
\label{Fig: Dilaton profile for C<0}
\end{figure}
\item \textit{Cored gravitational instantons} ($C>0$):\newline
Finally, for the case of cored gravitational instantons $C>0$ one finds \cite{0406038,150906374}
\begin{equation} \label{Solution for C>0}
e^{\alpha \varphi(r)} = \frac{1}{\sinh^2(K_+)} \sinh^2 \left(K_+ + \frac{\alpha}{2} \sqrt{\frac{3}{2}} \text{arcsinh} \left(\frac{\sqrt{C}}{r^2}\right)\right),
\end{equation}
where we again ensured $\lim_{r \to \infty} \varphi(r) =0$ by demanding
\begin{equation} \label{C for C>0}
C = \frac{n^2}{3!f_{\text{ax}}^2A^2} \sinh^2(K_+) \ .
\end{equation}
In \autoref{Fig: Dilaton profile C>0} two plots of the dilaton profile are presented. The integration constant $K_+$ should be positive in order to again avoid a divergence of $\varphi$ for some $r>0$, but is otherwise unconstrained. This is different compared to wormholes or extremal instantons, which do not exhibit a free parameter. From this 4d effective theory one is lead to believe that there exists a whole family of cored instanton solutions parametrised by $K_+$. However, by considering the microscopic origin of gravitational instanton solutions, one finds evidence that only certain values of $K_+$ are allowed, as we will now discuss.
\end{itemize}
\begin{figure}
\centering
\includegraphics[width=0.7\linewidth]{PlotCoredInstantons}
\caption{This plot shows dilaton profiles for the cored gravitational instanton with $\alpha = 15$ (solid line) and $\alpha=0.1$ (dashed line). Again, $r$ and $\varphi$ are given in Planck units. For the purpose of illustration we have chosen $ K_+ =0.5$ and $n/f_{\text{ax}}$ such that $C/\sinh^2 K_+=1$.}
\label{Fig: Dilaton profile C>0}
\end{figure}
\subsection{Interpretation of the Integration Constant $K_+$} \label{Subsection: Integration constant}
The integration constant $K_{+}$, or equivalently $C$, seems to be a free and continuous parameter giving rise to a family of solutions. We want to argue that this is not the case. Note that the cored gravitational instanton solutions are UV-sensitive and therefore a naive 4d field theory treatment is not sufficient. Instead, it is crucial to understand those solutions in a UV-complete theory, such as string theory. In this context the role of the integration constant $K_+$ becomes clear. Specifically, it was pointed out in \cite{0406038} that the parameter $C$ is determined by the mass $M$ and charge $Q$ of a dilatonic black brane wrapping internal cycles in a higher-dimensional theory, whose dimensionally reduced action coincides with \eqref{Action without potential}. This holds true at least for some values of $\alpha$. Consequently, we conjecture that $C$ and $K_+$ generically take discrete and well-defined values determined by the underlying microscopic theory. Further, if the Weak Gravity Conjecture holds in 5d, cored gravitational instantons may not be stable.
We support this conjecture by providing a specific toy-example borrowed from \cite{0406038}.
Following their results, we can consider a five-dimensional model with Euclidean action in 5d Planck units
\begin{equation} \label{Action in 5d}
S = \int d^5x \sqrt{\hat{g}} \left[-\frac{1}{2}\hat{R} + \frac{1}{2} (\partial \hat{\phi})^2 + \frac{1}{4}e^{a \hat{\phi}} \hat{F}^2\right],
\end{equation}
where $\hat{F}=d \hat{A}$ is a 2-form field strength tensor. For the dimensional reduction to a 4d theory we choose\footnote{For the purpose of compactification we switch to Euclidean time $\tau$ by Wick-rotation. For simplicity we choose the periodicity $\tau \sim \tau +1$. Later in this subsection we allow the circumference of the $S^1$ to have length $\ell >0$. This will then have to be taken into account in order to determine the axion-decay constant.}
\begin{equation} \label{Metric ansatz in 5D}
d\hat{s}_{(5)}^2 = e^{2 \beta_1 \psi} d\tau^2 + e^{2\beta_2 \psi} ds_{(4)}^2
\end{equation}
together with $\hat{A}= \theta d\tau$ and $\hat{\phi}=\phi$, i.e.~the fields $\theta$ and $\hat{\phi}$ do not depend on the extra-dimensional coordinate $\tau$. In Einstein-frame with canonically normalised kinetic terms dimensional reduction fixes the constants $\beta_1$ and $\beta_2$ to
\begin{equation}
\beta_1 = - 2 \beta_2, ~~~~~~\beta_2 = \frac{1}{\sqrt{6}}.
\end{equation}
After redefining the fields $\phi$ and $\psi$ via a rotation in the $(\phi, \psi)$-plane we get
\begin{equation} \label{Action compactified to 4d}
S = \int d^4x \sqrt{g} \left[-\frac{1}{2}R + \frac{1}{2} (\partial \tilde{\phi})^2 + \frac{1}{2} (\partial \tilde{\psi})^2 + \frac{1}{2}e^{\alpha \tilde{\phi}} (\partial \theta)^2\right],
\end{equation}
where $g$ denotes the metric corresponding to the 4d-line-element. Setting $\tilde{\psi}=0$ in this action one obtains the model considered in \eqref{Action with complex scalar} with dilatonic dependence in the kinetic term of $\theta$. The 4d dilaton coupling $\alpha$ is related to the 5d dilaton coupling $a$ via\footnote{Notice that our normalisation of $\phi$ is such that the prefactors of the Ricci scalar $R$ and the kinetic term $(\partial \phi)^2$ are equal, while in \cite{0406038} the prefactor of the dilaton has a factor $1/2$ relative to $R$. This is why our dilaton-coupling $\alpha$ differs by a factor of $\sqrt{2}$.}
\begin{equation}
\alpha ^2 = a^2 + \frac{8}{3}.
\end{equation}
Therefore, the interpretation of the 4d theory in terms of a 5d theory is only possible if $\alpha \geq 2\sqrt{2/3}$.
Let us now explicitly relate the integration constant $C$ to microscopic properties of a higher-dimensional theory.
In \cite{0406038} it was shown that for $\alpha = 2\sqrt{2/3}$ the solutions of the 4d model \eqref{Action compactified to 4d} can be uplifted to a five-dimensional Reissner-Nordstr\"om (RN) black hole solution
\begin{equation} \label{RN metric}
ds^2_{(5)} = g_+(\rho)g_-(\rho)d\tau^2 + \frac{d\rho^2}{g_+(\rho)g_-(\rho)} + \rho^2 d\Omega_3^2, ~~~~ \hat{F}_{\tau \rho} = \sqrt{6} \frac{Q}{\rho^3}
\end{equation}
where
\begin{equation}
g_{\pm}(\rho)= 1 - \frac{\rho_{\pm}^2}{\rho^2}
\end{equation}
with
\begin{equation}
\rho_{\pm}^2 = M \pm \sqrt{M^2-Q^2}.
\end{equation}
We take this as a simple toy-model to argue that $C$ is generically fixed by properties of a black brane wrapping internal cycles. The RN black hole can be interpreted as $N$ particles or 0-branes (or just one 0-brane wrapping the cycle $N$ times) of total mass $M$ and total charge $Q$. Note however that the ADM-mass $M_{\text{ADM}}$ is related to the mass parameter $M$ by $M_{\text{ADM}}=6 \pi^2 M$. Nevertheless, we henceforth call $M$ the mass of the RN-black hole. The charge $Q$ is defined such that $M=Q$ sets the extremality bound. That is, $Q=N\hat{q} \sqrt{6}/(6 \pi^2)$, where the charge $\hat{q}$ is defined by $N\hat{q}= 1/2 \int_{S^3} \star_5 \hat{F}$.\footnote{For the normalisation we found it useful to translate the conventions in \cite{Myers:1986un,Ortin:2015hya} to our situation.}
Upon toroidal dimensional reduction along the coordinate $\tau$ with the identification $\tau \sim \tau + \ell$ and the circumference $\ell >0$ of the compactified dimension, the 5d solution \eqref{RN metric} turns into an instanton solution \eqref{Metric}. Note that the coordinate singularity at $\rho=\rho_+$ of the 5d solution becomes a curvature singularity (at $r=0$) in the 4d solution \eqref{Metric}.
In the subsequent computation we show that our integration constant $C$ is simply given by $C= \ell^2 (M^2-Q^2)$ in 4d Planck units.
Denote by $g_{MN}^{(5)}$ the RN-metric \eqref{RN metric}, where $M,N$ run over the coordinates of the 4d space and the extra-dimensional coordinate $\tau$. Now, rescale the metric as follows: $\tilde{g}_{MN}^{(5)}=g_{MN}^{(5)}/(g_+g_-)$. From the canonical Einstein-Hilbert term we then get:
\begin{align} \nonumber
\int d^5x \sqrt{g^{(5)}}R[g_{MN}^{(5)}] &= \int d^5x (g_+g_-)^{3/2} \sqrt{\tilde{g}^{(5)}}R[\tilde{g}_{MN}^{(5)}] + ... = \\
&= \int d^4x\ell (g_+g_-)^{3/2} \sqrt{\tilde{g}^{(4)}}R[\tilde{g}_{\mu\nu}^{(4)}] + ...
\end{align}
The last term occurs in the compactified 4d theory using the identification $\tau \sim \tau+\ell$.
We want to point out that for generic $\ell >0$ there is a conical singularity at the outer horizon $\rho=\rho_+$. In principle, one could avoid such a conical singularity by choosing the periodicity of $\tau$ appropriately (it would be the inverse of the Hawking-temperature \cite{Hawking:1982dh}), but this would mean to fix the compactification radius. Instead, we accept the conical singularity as a necessary feature of Euclidean branes wrapped on cycles of the compact space.\footnote{Note that in the so-called dual frame metric discussed in \cite{0406038} non-extremal instantons with $\alpha=2\sqrt{2/3}$ can be interpreted as sections of constant time of the RN black hole metric. In this frame one recovers a wormhole geometry connecting two asymptotically flat regions smoothly. One pays the price of rescaling by a divergent factor. The above is technically different from our approach of obtaining gravitational instantons by compactification of a 5d black hole solution on an $S^1$. In our case the RN black hole solution \eqref{RN metric} in general yields conical singularities.} Since it is known that Euclidean branes wrapped on non-trivial cycles give rise to instantonic terms (see e.g. \cite{Wen:1985jz,Dine:1986zy}), we assume that the corresponding conical spacetimes are saddle-points of the Euclidean path integral.
We go to the Einstein frame (with 4d Planck mass $M_p=1$) by rewriting the Einstein-Hilbert term using the rescaled metric $g_{\mu\nu}^{(4)} = \ell (g_+g_-)^{3/2}\tilde{g}_{\mu\nu}^{(4)}$. The compactified 4d line-element then reads
\begin{equation}
ds_{(4)}^2 = \ell \frac{d\rho^2}{\sqrt{g_+g_-}} + \ell \sqrt{g_+g_-} \rho^2 d\Omega_3^2.
\end{equation}
For the comparison with the metric \eqref{Metric}, the obvious coordinate transformation to be made is simply $r^2 = \rho^2 \ell \sqrt{g_+g_-}$. Using the definitions of $g_{\pm}$ it follows
\begin{equation}
rdr = \ell^2 \frac{\rho(\rho^2-M)}{r^2}d \rho.
\end{equation}
Together with $(\rho^2-M)^2 = r^4/\ell ^2 + (M^2-Q^2)$, this implies:
\begin{equation}
\ell \frac{d\rho^2}{\sqrt{g_+g_-}} = \frac{dr^2}{1+ \ell^2 (M^2-Q^2)/r^4}.
\end{equation}
Hence, we find the simple relationship
\begin{equation}
C=\ell^2(M^2-Q^2)
\end{equation}
in 4d Planck units. Upon dimensional reduction of \eqref{Action in 5d} and using the periodicity of the Wilson line $\hat{A}_{\tau} \cong \hat{A}_{\tau} + \pi/(\hat{q}\ell)$ one can easily check that the axion decay constant reads $f_{\text{ax}} =1/(2\hat{q}\ell)$ for an axion $\theta$ with $2\pi$-periodicity. It follows that
\begin{equation}
C = \frac{N^2}{24 \pi^4 f_{\text{ax}}^2}\left[\left(\frac{M}{Q}\right)^2-1\right].
\end{equation}
We can compare this result to our previous expression \eqref{C for C>0}. First, we can identify the wrapping number/number of 0-branes $N$ with the flux number $n$. We then find that the integration constant $K_+$ in \eqref{C for C>0} is completely determined by the parameters $M$ and $Q$ describing black holes/branes in the 5d theory. An immediate result is that $K_+$ and hence $C$ are not free parameters. The possible range of values is determined by the spectrum of black branes in the higher-dimensional theory. Furthermore, as $M$ and $Q$ are discrete quantities it follows that $C$ can also only take discrete values (for a given value of $f_{\text{ax}}$). This property is only important as long as $M$ and $Q$ are small. In the macroscopic regime of large $M$ and $Q$ the value of $C$ can be dialed to any positive value and it becomes effectively continuous. We come back to this in \autoref{Section:Cored Grav Inst}.
Notice that the case of $M=Q$, which gives $C=0$, corresponds to an extremal Reissner-Nordstr\"om black hole. In this sense, the name \textit{extremal instanton} for flat $4$d solutions \eqref{Metric} is justified. In \autoref{Section:Cored Grav Inst} we comment on how to express the extremal instanton action in terms of $\ell$ and $M_{\text{ADM}}$, consistent with, for instance, \cite{09052844,10116301,150906374}.
This example illustrates nicely how $4$d cored or extremal instanton solutions can be obtained from black holes/branes with mass $M$ and charge $Q$. Of course, one could also go beyond such simple toy-models we just discussed, allowing also for dilaton couplings $\alpha \neq 2\sqrt{2/3}$. We expect the relation $C=\ell^2(M^2-Q^2)$ to be modified by the corresponding parameter $a \neq 0$ in this more general case. Furthermore, one would expect that after SUSY-breaking extremal objects in string theory would appear as non-extremal instantons in the 4d effective theory.
Last, let us remark on possible implications for cored gravitational instantons arising from the Weak Gravity Conjecture. In particular, if the Weak Gravity Conjecture holds in the 5d model we expect that objects with $M > Q$ can in principle decay. As cored gravitational instantons arise from such unstable objects upon dimensional reduction, one may wonder whether this instability is then inherited by the instantons. Here `unstable instanton' means that two instantons exist which cause the same flux change but have smaller total action. In this sense, the contribution of cored instantons to the Euclidean path integral is subdominant if cored instantons are `unstable'. This point will me made more precise in \autoref{Section: WGC}.
\section{Instanton Potentials from Euclidean Wormholes} \label{Section:Instanton Potentials from Wormholes}
The goal of this section is to show that the one-instanton action, describing a Giddings-Strominger wormhole, gives rise to an instanton potential of the structure $\cos \theta e^{-S}$.
We begin with a brief review of Coleman's derivation \cite{ColemanBook,Vainshtein:1981wh} of the energy eigenvalues for a simple one-dimensional quantum mechanical system with periodic potential $V$, e.g.~$V(x) \sim \sin^2(2\pi x)$. These considerations can be applied to quantum field theory and in particular to our system as well.
The Hamiltonian is $H=p^2/2+V(x)$. An instanton or an anti-instanton correspond to tunnelling events from $x$ to $x+1$ or $x-1$, respectively. Using the dilute-gas approximation we can distribute instantons and anti-instantons freely in time. Let us introduce a basis of states $\ket{j}$ in which the particle is localised at $x \simeq j$. Then for some time interval $T>0$, transition amplitudes are \cite{ColemanBook}
\begin{equation} \label{Coleman formula}
\braket{j_+|e^{-HT}|j_-} = \left(\frac{\omega}{\pi}\right)^{1/2} e^{-\omega T/2} \sum_{N=0}^{\infty} \sum_{\bar{N}=0}^{\infty} \frac{1}{N! \bar{N}!} (Ke^{-S_0}T)^{N+\bar{N}} \delta_{(N-\bar{N})-(j_+-j_-)},
\end{equation}
where $j_-$ and $j_+$ are the positions of the initial and final state, respectively. $N$ and $\bar{N}$ count the number of instantons and anti-instantons. Moreover, $\omega$ is defined by $\omega = V^{\prime \prime}(0)$. $K$ is the familiar determinant factor, which depends on details of the potential $V$. $S_0$ denotes the instanton action. The Kronecker delta can be rewritten as
\begin{equation}
\delta_{ab} = \int_0^{2\pi} \frac{d\theta}{2\pi} e^{i(a-b)\theta},
\end{equation}
and thus, after performing the summation,
\begin{equation}
\braket{j_+|e^{-HT}|j_-} = \left(\frac{\omega}{\pi}\right)^{1/2} e^{-\omega T/2} \int_0^{2\pi} \frac{d\theta}{2\pi} e^{i(j_--j_+)\theta} \exp \left(2KT \cos \theta e^{-S_0}\right).
\end{equation}
From this we can read off that the system has an energy eigenbasis
\begin{equation}
\ket{\theta} = \sum_j e^{ij \theta} \ket{j}
\end{equation}
with eigenvalues
\begin{equation}
E(\theta) = \frac{1}{2} \omega - 2K \cos \theta e^{-S_0}.
\end{equation}
This derivation reveals the logic behind the famous contribution $\sim \cos \theta e^{-S}$ to the axion potential in quantum field theory, where the centres of the instantons are not distributed on a time interval but instead in a region of spacetime with volume $\mathcal{V}$. One then simply has to replace the variable $T$ by the volume $\mathcal{V}$.
\bigskip
In the following we explain how this computation can be used to derive an instanton potential induced by Euclidean wormholes. In the previous \autoref{Section:Gravitational Instanton Solutions } we reviewed that Euclidean wormholes exist in the presence of a non-vanishing $3$-form flux $H$ with quantised charge $n \in \mathbbm{Z}$. An instanton would then correspond to a transition from $n$ to $n+1$. By the logic of Coleman's computation above, this should induce a shift-symmetry breaking potential.
In \cite{150906374} this was questioned, because Euclidean wormholes appear as conduits and charges would not disappear. In other words, one always has an instanton and an anti-instanton, thus preserving $n$.
We argue that this issue is more subtle: the two ends of a Euclidean wormhole do not necessarily have to end at the same hypersurface of constant Euclidean time, but can also close on distant hypersurfaces. Similarly, the two ends can have very large spatial separation such that, from a local perspective, a potential \`a la Coleman should be induced. Then, a Minkowskian observer would only see either the instanton or anti-instanton part of the wormhole and thus find a change in the charge $n$, see \autoref{Figure: Wormholes}. This invalidates the reasoning in \cite{150906374}, and hence we do not see any argument against the breaking of the shift-symmetry due to Euclidean wormholes.
\begin{figure}[t]
\centering
\includegraphics[width=1.00\textwidth]{wormhole_3}
\caption{This picture presents a wormhole which opens at some initial time $t_i$ and closes at $t_f>t_i$. The dotted line indicates the separation of the two events. }
\label{Figure: Wormholes}
\end{figure}
We want to make this mathematically more precise. This requires to compute the path integral contribution of all possible wormhole configurations.
This allows us to infer the effective potential $V_{\text{eff}}(\theta)$ for the axion field $\theta$. The logic behind the computation of $V_{\text{eff}}(\theta)$ is the following. The expectation value of any observable $\mathcal{O}(\theta)$ is given by
\begin{equation}
\braket{\mathcal{O}(\theta)} \sim \int d[\theta] \mathcal{O}(\theta) \exp \left(- \frac{1}{2}\int d^4x f_{\text{ax}}^2 (\partial \theta)^2\right) Z_{\text{wh}}[\theta] \ ,
\end{equation}
where the path integral contribution of wormholes is schematically (i.e.~no combinatioral factors included yet) given by
\begin{equation} \label{Wormholesum}
Z_{\text{wh}}[\theta] \sim \sum_{w} \prod_{n=1}^{w} \prod_{m=1}^{w} \int d^4x_n \int d^4x_m e^{-S}e^{i \theta(x_n)}e^{-S}e^{-i \theta(x_m)} \ .
\end{equation}
Here $w$ denotes the number of wormholes of a configuration. Note that the phase difference occurs because one factor is for the instantons, the other for anti-instantons. Those factors arise from the second term of \eqref{Action Theta}, where $Q_{\theta} = \pm 1$ ($+$ for instantons, $-$ for anti-instantons). We write out these factors explicitly, because they finally give rise to the $\cos$-potential for $\theta$. The contribution $Z_{\text{wh}}[\theta]$ induces a change $\delta S_{\text{ind}}(\theta)$ of the action for the axion and we expect
\begin{equation}
\braket{\mathcal{O}(\theta)} \sim \int d[\theta] \mathcal{O}(\theta) \exp \left(- \frac{1}{2}\int d^4x f_{\text{ax}}^2 (\partial \theta)^2 - \delta S_{\text{ind}}(\theta)\right) \ ,
\end{equation}
where $\delta S_{\text{ind}}(\theta)$ contains by definition the effective potential of $\theta$ plus higher derivative corrections:
\begin{equation}
\delta S_{\text{ind}}(\theta) = \int d^4x \left(V_{\text{eff}}(\theta) + \text{higher derivative terms}\right) \ .
\end{equation}
Hence, the effective potential $V_{\text{eff}}(\theta)$ can be determined by computing $Z_{\text{wh}}$ using any field configuration $\theta$ for which $V_{\text{eff}}(\theta)$ dominates all derivative terms. We choose a smooth version of the profile
\begin{equation} \label{Testfunction}
\theta(x) = \begin{cases}
\theta_0 & \text{for}~ x \in I \times \mathbbm{R}^3 \\
0 & \text{else} \ ,
\end{cases}
\end{equation}
i.e.~a profile which is only non-zero in a small Euclidean time interval $I$ (see \autoref{fig:wormholesummation}) and goes to zero smoothly at the boundary of $I$. This is illustrated in \autoref{fig:testfunction}.
\begin{figure}[t]
\centering
\begin{overpic}[width=0.85\linewidth]{wormholesummation}
\put (97,20) {$\mathbbm{R}^3$}
\end{overpic}
\caption{This illustration shows pairs of connected black and white dots, each representing an end of a wormhole (black if the end corresponds to an instanton and white for an anti-instanton). Only few wormholes lie completely inside the shaded region $I \times \mathbbm{R}^3$.}
\label{fig:wormholesummation}
\end{figure}
The volumes of $I \times \mathbbm{R}^3$ and of the remaining part of Euclidean space are denoted by $\mathcal{W}$ and $\mathcal{V}$, respectively. We assume that $\mathcal{W} \ll \mathcal{V}$ with $\mathcal{W}$ being large enough to typically contain many wormhole ends.
We first check that the derivative terms can indeed be made subdominant with respect to the effective potential. For simplicity, we work near the minimum and use the approximation $V_{\text{eff}} \sim m^2\theta^2$. It is crucial that our axion profile at the boundary of $I$ features a smooth transition of characteristic length $\ell$ from 0 to $\theta_0$ with $\ell \ll L$, where $L$ is the length of the Euclidean interval $I$.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{testfunction}
\caption{We illustrate the smooth axion profile corresponding to the approximation in \eqref{Testfunction}. In the Euclidean time interval of length $L$ the axion field is constant (value $\theta_0$) and outside this interval the field decays smoothly. The characteristic length of this transition from $\theta_0$ to $0$ is denoted by $\ell$.}
\label{fig:testfunction}
\end{figure}
We then have $\partial \theta \sim \theta_0/\ell$ close to the boundary (and zero elsewhere) and the comparison of $\int d^4x f_{\text{ax}}^2 (\partial \theta)^2$ with $\int d^4x V_{\text{eff}}$ should yield
\begin{equation}
\ell V_3 f_{\text{ax}}^2 \frac{\theta_0^2}{\ell^2} \ll L V_3 m^2 \theta_0^2 \ ,
\end{equation}
where $V_3$ denotes the corresponding 3-volume of the Euclidean spacetime regions we consider. It follows that
\begin{equation}
L \gg \frac{f_{\text{ax}}^2}{m^2 \ell}
\end{equation}
has to be imposed. Note that $\ell$ cannot be arbitrarily small as we have to ensure that also higher-derivative terms must be subdominant. Comparing $f_{\text{ax}}^2 (\partial \theta)^2$ with $(\partial \theta)^4$ yields
\begin{equation}
\ell \gg \frac{\theta_0}{f_{\text{ax}}} \sim \frac{1}{f_{\text{ax}}} \ .
\end{equation}
It is not hard to see that these two conditions together with $L \gg \ell$ can be satisfied simultaneously. Thus, our field configuration \eqref{Testfunction} is suitable for the calculation of the effective potential given below.
For this computation we find it useful to group the sum in the above expression \eqref{Wormholesum} according to the position of the wormhole ends, see again \autoref{fig:wormholesummation}.
Let us assume that a wormhole corresponds, from the perspective of the $\theta$-field theory, simply to an instanton-anti-instanton pair. Then, denoting by $\theta$ the field configuration of \autoref{fig:testfunction}, we can write:
\begin{equation} \label{Partition function wormhole naiv}
Z_{\text{wh}}[\theta] \sim \sum_{n} \frac{1}{n!n!}\prod_{j=1}^{n} e^{-2S}K^2 \left(\mathcal{V} + \mathcal{W}e^{i\theta_0}\right) \left(\mathcal{V} + \mathcal{W}e^{-i\theta_0}\right) \ .
\end{equation}
Note that this differs from the toy model \eqref{Coleman formula} in the sense that the sector with $N \neq \bar{N}$ is not contained in \eqref{Partition function wormhole naiv}. But this is precisely our point: we want to find out whether a potential can still be generated if we impose $N=\bar{N}$, i.e.~an equal number of instantons and anti-instantons.
The combinatorial factor $1/(n!)^2$ is due to the indistinguishability of instantons and anti-instantons, respectively. The cross-terms $\mathcal{V}\mathcal{W}e^{\pm i \theta_0}$ correspond to wormholes where only one end is within $I \times \mathbbm{R}^3$.
The sum can be expressed as a Bessel function $I_0$
\begin{equation}
Z_{\text{wh}} \sim I_0(x)
\end{equation}
with
\begin{align} \label{Bessel sum}
I_0(x) &= \sum_{m=0}^{\infty} \frac{1}{m!m!} \left(\frac{x}{2}\right)^{2m} \\
x & \simeq 2Ke^{-S} \mathcal{V}\left(1+(\mathcal{W}/\mathcal{V})\cos \theta_0 + \mathcal{O}\left((\mathcal{W}/\mathcal{V})^2\right) \right) \ .
\end{align}
Furthermore, there is an integral expression for $I_0$:
\begin{equation} \label{Bessel integral}
I_0(x) = \frac{1}{2\pi} \int_{0}^{2\pi} d \phi e^{-x \cos \phi} \ .
\end{equation}
Hence, we arrive at
\begin{equation}
Z_{\text{wh}} \sim \frac{1}{2\pi} e^{x} \int_{0}^{2\pi} d\phi e^{-x(1+\cos \phi)} \simeq \frac{1}{2\pi} e^x \int_{0}^{2\pi} d\phi e^{-x(\phi-\pi)^2/2} \simeq \frac{1}{\sqrt{2\pi}} \frac{e^{x}}{\sqrt{x}} \ ,
\end{equation}
where we relied on the fact that in our case $x$ is large (because $\mathcal{V}$ and $\mathcal{W}$ are large). Since we are interested in the effective potential $-\mathcal{W}V_{\text{eff}}(\theta_0) \simeq \ln Z_{\text{wh}} $, we can focus on the exponential factor:
\begin{align} \label{Paritition function wormhole contribution}
Z_{\text{wh}} \sim \exp \left[2Ke^{-S}\left(\mathcal{V} + \mathcal{W} \cos \theta_0\right)\right] \ .
\end{align}
From here it is clear that, to explain the $\mathcal{W}\cos \theta_0$ term, the effective action for $\theta$ must contain a potential (which in our case contributes only in the region $I \times \mathbbm{R}^3$):
\begin{equation}
V_{\text{eff}}(\theta_0) \sim 2Ke^{-S} \cos \theta_0 \ .
\end{equation}
The $\theta$-dependency is as in the case of an instanton-anti-instanton gas without any constraints imposed. Actually, the term generated by wormhole instantons, which looks like a potential term for the field configuration \eqref{Testfunction}, is in fact a non-local interaction term. This can be seen by modifying \eqref{Testfunction} such that $\theta(x) = \theta_{\mathcal{V}} \neq 0$ for $x \in \mathcal{V}$. Then, $\cos(\theta_0)$ in \eqref{Paritition function wormhole contribution} is replaced by $\cos(\theta_0-\theta_{\mathcal{V}})$. Indeed the exponential now contains non-local terms. Globally, this non-local term preserves shift-symmetry. Nevertheless, we observe a crucial effect induced by Giddings-Strominger wormholes: The change of the action due to a local fluctuation, $S[\theta+\delta\theta] - S[\theta]$, corresponds to that induced by a potential $V (\delta\theta) \sim 2 K \exp(-S)\cos(\delta\theta)$.
This can also be seen by applying Coleman's computation \cite{ColemanBook} to our problem. We are then interested in computing the partition function $Z=\sum_{n}\braket{n|e^{-HT}|n}$. It is therefore sufficient to focus on transition functions $\braket{n|e^{-HT}|n}$, although transitions from $\ket{m}$ to $\ket{n}$ will occur as well (the Hamilitonian $H$ is in general non-diagonal is this basis). Due to the trace we do not need to consider the latter in our calculation. The result of this computation has to be compared with the partition function $Z=\int_{0}^{2\pi} d\theta \braket{\theta|e^{-HT}|\theta}/(2\pi)$ in the $\theta$-language.
In the free theory $n$ is the dual variable to our axion field $\theta$, which can be seen as follows: The free theory action for the axion is given by $S= \int d^4x f^2 \dot{\theta}^2/2$ or, after integrating out the spatial directions, $S= \int dt A \dot{\theta}^2/2$ with $A \equiv f^2 V_3$, where $V_3$ is the 3-volume ($\theta$ is then the zero mode). The canonical momentum $p$ is then given by $p=A \dot{\theta}$. As it is well known from quantum mechanics on $S^1$, $p$ is quantised as $p=n \in \mathbbm{Z}$. One can therefore relate $\ket{\theta}$-states to $\ket{n}$-states via
\begin{equation}
\ket{\theta} = \sum_n e^{in\theta} \ket{n}
\end{equation}
in the free theory (see also \cite{07052768}), where we chose the normalisation $\braket{\theta|\theta^{\prime}}= 2\pi \delta(\theta-\theta^{\prime})$.
The Hamiltonian of the free theory is then given by $H=n^2/(2A)$ and in the free theory we have the transition amplitude
\begin{equation} \label{Transition amplitude free theory}
\braket{n|e^{-HT}|n}_{\text{free}} = e^{-Tn^2/(2A)} \ .
\end{equation}
Let us now return to interacting theory and take into account the effects of the wormhole gas induced by the coupling of $\theta$ to gravity. We assume that instantons and anti-instantons are randomly distributed and that wormhole ends can have arbitrarily long separation with no physical effect. By applying Coleman's formula \eqref{Coleman formula} to our situation and taking into account \eqref{Transition amplitude free theory}, we find:
\begin{equation} \label{Transition amplitude with instantons}
\braket{n|e^{-HT}|n} = e^{-Tn^2/(2A)} \sum_{N=0}^{\infty} \frac{1}{N! N!} (Ke^{-S_0}T)^{2N}.
\end{equation}
We emphasise once more that off-diagonal elements $\braket{m|e^{-HT}|n}$ are in general non-zero.
For instance, if $T$ corresponds to half of the time interval of \autoref{Figure: Wormholes}, the instanton number clearly changes by unity. However, such off-diagonal elements never appear explicitly in our calculation, which relies solely on the partition function.
We can once again express the sum in \eqref{Transition amplitude with instantons} by $I_0$ via \eqref{Bessel sum} and then use the integral expression \eqref{Bessel integral} with integration variable $\theta$. We find:
\begin{equation}
\braket{n|e^{-HT}|n} = e^{-Tn^2/(2A)} \int_{0}^{2\pi} \frac{d\theta}{2\pi} \exp \left(-2KT\cos \theta e^{-S_0}\right) \ .
\end{equation}
For the partition function $Z=\sum_{n}\braket{n|e^{-HT}|n}$ one then obtains
\begin{equation} \label{Partition function for free theory + instantons}
Z(T) \simeq \sqrt{\frac{2\pi A}{T}} \int_{0}^{2\pi} \frac{d\theta}{2\pi} \exp \left(-2KT\cos \theta e^{-S_0}\right) \ .
\end{equation}
As we already mentioned, this should be compared with
\begin{equation} \label{Partition function general}
Z=\int_{0}^{2\pi} \frac{d\theta}{2\pi} \braket{\theta|e^{-HT}|\theta} \ .
\end{equation}
For a first naive comparison of \eqref{Partition function for free theory + instantons} and \eqref{Partition function general} we ignore the non-exponential $T$-dependence in the prefactor of \eqref{Partition function for free theory + instantons}.\footnote{Our parametrisation is then equivalent to $Z(T)=\int dE \rho(E)e^{-ET}$. Thus, we found the partition function. The latter characterises the system unambiguously.} Then $\ket{\theta}$ is an eigenbasis of the Hamiltonian with eigenvalues $V(\theta)$. We see by comparison with \eqref{Partition function for free theory + instantons} that $ V(\theta) =2 \cos \theta Ke^{-S_0}$.
We can, however, be more precise and understand also the prefactor. To do so we observe that \eqref{Partition function for free theory + instantons} was derived on the basis of \eqref{Transition amplitude with instantons}, and in this equation a non-trivial approximation was made: Indeed, we placed the factor $\exp(-Tn^2/(2A))$ outside the instanton sum. In general, that is not justified for the following reason. If we start at $t=0$ with flux number $n$, and the first instanton occurs e.g. at $t=T_1$, we get a factor $\exp(-T_1n^2/(2A))$ from the kinetic term. If then the next instanton occurs at $T_2$, we get a further factor $\exp(-T_2(n+1)^2/(2A))$ and so on. The times $T_i$ have to be integrated over and these prefactors can not be extracted from the instanton sum. However, we can find conditions under which it is safe to approximate the $(n+N_i)^2$ in the exponents (where $N_i$ is the number of instantons present at some given time) simply by $n^2$. To do so we note that, on the one hand, the instanton sum is dominated by instanton numbers of the order of
\begin{equation}\
\braket{N} \sim Ke^{-S_0}T \ .
\end{equation}
On the other hand typical values of $n$ dominating the sum over $\exp(-Tn^2/(2A))$ are of the order of $n\sim \sqrt{A/T}$. Thus, disregarding $N_i$ relative to $n$ in the $(n+N_i)^2$-terms will be justified if
\begin{equation}
\braket{N} \ll \sqrt{A/T}\qquad\mbox{or}\qquad A\gg \langle N\rangle^2 T \ .
\end{equation}
Given that we anyway choose $T$ large enough to ensure $\langle N\rangle\gg 1$, this implies in particular $A\gg T$.
With this in mind, we return to the corresponding quantum mechanical model. We conjecture that the instanton dynamics is captured by an effective potential $V(\theta)=2\cos(\theta)Ke^{-S_0}$. To confirm this, we calculate the partition function
\begin{equation} \label{Partition function quantum particle}
Z(T) \simeq \frac{1}{Z(0)}\int d\theta \int_{\tilde{\theta}(0)= \theta}^{\tilde{\theta}(T) = \theta} d[\tilde{\theta}] \exp \left[ -\int_0^T dt \left(A \frac{\dot{\tilde{\theta}}^2}{2} + V(\tilde{\theta})\right)\right] \ .
\end{equation}
It has to be compared to \eqref{Partition function for free theory + instantons} to establish the correctness of the chosen effective description and, in particular, the potential. But working out \eqref{Partition function quantum particle} in the regime $A\gg T$ is easy. Indeed, if we first disregard the potential, we are simply dealing with a 1-dimensional system on the interval $(0,2\pi)$ and a kinetic-term prefactor $A$. This prefactor sets the minimum time by which any wave packet unavoidably spreads to an $\mathcal{O}(1)$ width due to quantum dynamics. In addition, the potential has a maximal steepness $|V'|\sim Ke^{-S_0}$, leading to a displacement of $Ke^{-S_0}T^2/A\sim \braket{N} T/A$ during a time interval $T$. Our previously derived conditions on $A$, which underly our derivation of \eqref{Partition function for free theory + instantons}, are sufficient to ensure that the particle moves only by a distance $\Delta\theta\ll 1$ during the time $T$. Hence, in evaluating \eqref{Partition function quantum particle} we can approximate $V(\tilde{\theta}(t))\simeq V(\theta)$. The path integral then becomes that of free particle, to be evaluated on times too short for the periodicity of the configuration space to be relevant. One obtains the well-known time-dependence $\sim \sqrt{A/T}$ of the amplitude, to be multiplied by the integral over $\exp(-TV(\theta))$. This is now in perfect agreement with \eqref{Partition function for free theory + instantons}.\footnote{While our analysis establishes the quantum mechanical model with effective potential $V(\theta)=2\cos(\theta)Ke^{-S_0}$ only for a certain range of $T$, we expect it to be valid also for $T \to \infty$.}
Thus, we find that Giddings-Strominger wormholes give rise to an effective potential $V(\theta)\sim 2Ke^{-S} \cos \theta$ in two independent approaches. We wish to remark that in both approaches we can be agnostic about details of the interpretation of wormholes connecting to baby-universes. Crucially, the axionic shift-symmetry is broken locally even if the condition of having equally many instantons and anti-instantons is imposed on the global space-time.
\bigskip
Finally, we wish to remark that the correct choice of the combinatorial factors is a subtle issue. We interpreted a configuration of $N$ wormholes as an instanton-anti-instanton-gas with (anti-)instantons randomly distributed. It is then plausible to include the combinatorial factor $1/(N!)^2$. However, one might argue that each instanton has a corresponding anti-instanton and therefore we should multiply by $N!$ to account for the number of possible pairings. If we assume that the right combinatorial factor is just $1/N!$, we can still do the computation starting with \eqref{Wormholesum}. We then still get $\cos \theta_0$, but this time the energy density in $I \times \mathbbm{R}^3$ scales with
\begin{equation} \label{Rho with 1/N!}
V(\theta_0) \sim K^2\mathcal{V}e^{-2S} \cos \theta_0 \ ,
\end{equation}
which diverges as $\mathcal{V} \to \infty$. Possibilities to avoid this divergence were discussed in \cite{Giddings:1988cx,Coleman:1988cy,Preskill:1988na,Klebanov:1988eh}, mostly in the baby-universe interpretation of Giddings-Strominger wormholes. It is possible to express the partition function as an integral over a parameter $\alpha$, which is an eigenvalue of a baby-universe operator \cite{Coleman:1988cy}.
We rather follow Preskill \cite{Preskill:1988na} to sketch the idea of how to evade the divergence:
For a combinatorial factor $1/N!$ the partition function reads
\begin{equation}
Z \sim \sum_{N=0}^{\infty} \frac{C^N}{N!} = e^C \ ,
\end{equation}
where
\begin{align}
C \sim \bar{z}z ~, ~~~~~ z \equiv K\mathcal{V}e^{-S}e^{i \theta} \ .
\end{align}
Clearly, $C \sim \mathcal{V}^2$. But formally, we can write
\begin{equation}
Z \sim \int d \alpha d \bar{\alpha} e^{-\bar{\alpha} \alpha + \alpha \bar{z} + \bar{\alpha} z } \ .
\end{equation}
If $\alpha$ is integrated out we obtain the divergent result \eqref{Rho with 1/N!}. (To see $\cos \theta_0$ coming in one has to group terms carefully as in our first computation.) If, as suggested by Preskill \cite{Preskill:1988na}, one has to fix $\alpha$ to a certain value, the energy density is simply given by\footnote{Nevertheless, the divergence remains disconcerting. For instance, the expectation value of the number of wormholes in a certain space-time region scales as $\braket{N} \sim \mathcal{V}^2$ and it is questionable whether the wormhole gas can be dilute in the limit $\mathcal{V} \to \infty$. See also discussions in e.g.~\cite{Preskill:1988na,Fischler:1988ia,Coleman:1989ky,Polchinski:1989ae}.}
\begin{equation}
\rho \sim \alpha e^{-S} \cos \theta \ .
\end{equation}
In any case, no matter which combinatorial factor is correct and no matter how to interpret $\alpha$, we always find that a term $\cos \theta$ arises in the effective action.
To summarise, we conclude that Euclidean wormholes are expected to induce an instanton potential $\sim \cos \theta e^{-S}$. Shift-Symmetry appears to be broken locally.
It would be interesting to study whether this can be seen more directly by building an analogy between gravitational and gauge instantons, where the role of the term $\theta \text{Tr}(F \wedge F)$ is played by $\theta \text{Tr}( R \wedge R)$.
\bigskip
In the following we apply the presented derivation of the instanton potential to cases of $S=nS_0$, giving rise to potentials of the form $\sum_n \cos(n\theta)e^{-nS_0}$.
\section{The Limit of Validity of Gravitational Instanton Actions}
\label{Section:Cored Grav Inst}
In this section we summarise the instanton actions for all cases $C<0$, $C=0$ and $C>0$ and find limits for the validity of the computation. Qualitatively, we have
\begin{equation}
S \sim \frac{n}{f_{\text{ax}}}
\end{equation}
in all three cases. This is of course already known for Euclidean wormholes, see e.g.~\cite{Giddings:1987cg,Grinstein:1988ja,Lee:1988ge,Abbott:1989jw,Brown:1989df,Burgess:1989da,Coleman:1989zu,0406038,150303886} and also for $C \geq 0$, see e.g.~\cite{0406038,150906374}.
Furthermore, we address one concern raised in \cite{0406038}: the cored gravitational instanton solutions have a singularity at $r=0$ and hence it is unclear whether these solutions can be trusted all the way to the limit $r \to 0$. In fact, we expect a breakdown of the solutions at some radius $r = r_c >0$, which will be estimated in \autoref{Section:Inflation}. We expect such a cutoff radius to be present in any extra-dimensional theory independently of whether a curvature singularity exists or not. Therefore, even the extremal instantons, which do not have singularities, should only be trusted down to $r = r_c$. The situation is different for the Euclidean wormhole solutions, where we can have full control over the solution as long as $r_0 \gtrsim r_c$, with $r_0 \equiv |C|^{1/4}$ being the radius of the wormhole throat at the centre.
The limit of validity affects the computation of the instanton action. In the case of $C \geq 0$ one would usually integrate from $r=0$ to infinity, but instead we can only rely on the contribution from the interval $(r_c,+ \infty)$. Whenever a significant fraction of the action comes from $(0,r_c)$, we cannot trust the instanton actions computed in \cite{0406038,150906374} and we will discard these cases.
Thus, the initial task of this section is the evaluation of the on-shell contribution of the integral in \eqref{Action without potential}.
We proceed by using the equations of motion successively. Details of the computations are presented in \autoref{Appendix:Instanton Action}.
At first, by tracing Einstein's equations, we can express the Ricci scalar by the trace of the energy-momentum tensor:
\begin{equation}
R = - T.
\end{equation}
One can then rewrite \eqref{Action without potential} as
\begin{equation}
S = \int_M d^4x \sqrt{g} \mathcal{F}(\varphi)H^2.
\end{equation}
However, this is not yet the full contribution to the instanton action, because the Gibbons-Hawking-York boundary term has to be taken into account. It is
\begin{equation} \label{Gibbons-Hawking-York}
S_{\text{GHY}} = - \oint_{\partial M} d^3x \sqrt{h}(K-K_0),
\end{equation}
where $h$ is the determinant of the induced metric on $\partial M$. $K$ and $K_0$ are the traces of the extrinsic curvatures of $\partial M$ in $M$ and flat space, respectively.
Then, the instanton action is computed as
\begin{equation}
S_{\text{inst}} = S + S_{\text{GHY}} = \int_M d^4x \sqrt{g} \mathcal{F}(\varphi)H^2 - \oint_{\partial M} d^3x \sqrt{h}(K-K_0).
\end{equation}
Henceforth, we restrict to the case $\mathcal{F}(\varphi)= \exp(-\alpha \varphi)/(3!f_{\text{ax}}^2)$. Using this together with the equation of motion \eqref{EoM for B} and \eqref{Solution for C<0, with constant}, \eqref{Solution for C=0} or \eqref{Solution for C>0} depending on the choice of $C$, one can rewrite the first term, $S$, in the instanton action as an integral over $\varphi$. The contribution from $S_{\text{GHY}}$ is computed by considering a surface of constant $r$, see \cite{150906374} or \autoref{Appendix:Instanton Action}.
In the following we analyse the instanton action case by case:
\subsubsection*{Case $C=0$:}
Extremal instanton solutions go along with a flat metric ($C=0$). Thus, we have
\begin{equation}
S_{\text{GHY}}=0.
\end{equation}
However, the fields $\varphi$ and $B$ have a non-trivial profile giving rise to non-vanishing contributions to the instanton action.
The full contribution from $r=0$ to $r=\infty$ is given by
\begin{equation} \label{Instanton action C=0}
S_{\text{inst}} = - \frac{n}{f_{\text{ax}}} \int_{\varphi(0)}^{\varphi(\infty)} d \varphi \exp(-\alpha \varphi/2) = \frac{2}{\alpha} \frac{n}{f_{\text{ax}}}.
\end{equation}
As we explained previously, we cannot trust this computation for radii $r<r_c$. Nevertheless, as long as the main contributions to the action come from the regime $r>r_c$ the result \eqref{Instanton action C=0} can still be used to estimate contributions to the instanton potential in \autoref{Section:Inflation}. One should therefore compare the contribution $\Delta S$ from the regime $r<r_c$ with \eqref{Instanton action C=0}. Unfortunately, the contribution $\Delta S$ is UV-sensitive. We assume, however, that the actual UV-contribution $\Delta S$ can be parametrically estimated by the naive formula
\begin{equation}
\label{Delta S for C=0}
\Delta S = - \frac{n}{f_{\text{ax}}} \int_{\varphi(0)}^{\varphi(r_c)} d \varphi \exp(-\alpha \varphi/2) = \frac{2}{\alpha} \frac{n}{f_{\text{ax}}} \left(1 + \frac{\alpha n}{4A f_{\text{ax}}} \frac{1}{r_c^2}\right)^{-1} \ ,
\end{equation}
where we used \eqref{Solution for C=0} in the second step.
Demanding that $\Delta S \ll S_{\text{inst}}$ implies
\begin{equation} \label{Condition dS << Sinst}
\frac{\alpha n}{4A f_{\text{ax}}} \gg r_c^2 \ ,
\end{equation}
which in turn can be rewritten as a lower bound on $S_{\text{inst}}$:
\begin{equation}
S_{\text{inst}} \gg \frac{8 A}{\alpha^2} r_c^2 \ .
\end{equation}
This bound depends on the cutoff $r_c$ and the dilaton coupling $\alpha$. Interestingly, the bound gets weaker for larger $\alpha$ such that contributions from gravitational instantons become increasingly important with increasing $\alpha$. However, as we will describe in \autoref{Subsection: Dilaton coupling from string theory}, a regime of large dilaton coupling $\alpha$ may not be attainable in string theory. We find that only rather small values of $\alpha \sim \mathcal{O}(1)$ arise from the simplest string compactifications.
Before addressing the next case, we want to point out that \eqref{Instanton action C=0} can be rewritten as
\begin{equation} \label{Instanton action C=0 and ADM Mass}
S_{\text{inst}} = \ell M_{\text{ADM}}
\end{equation}
in the case of $\alpha = 2 \sqrt{2/3}$, where $M_{\text{ADM}}$ is the ADM-mass of our \textit{extremal} 5d RN-black hole of \autoref{Subsection: Integration constant}, which is consistent with e.g.~\cite{09052844,10116301,150906374}.\footnote{For the derivation of \eqref{Instanton action C=0 and ADM Mass} we used that the black hole charge $Q$ is related to $n$ by $n=2\pi^2 \sqrt{6} Q$. This can be obtained by dimensional reduction of the term $1/(2 \cdot 3!) \int (\star_5 \hat{F})^2$ together with $\hat{F}_{\tau \rho}=\sqrt{6}Q/\rho^3$ and \eqref{Charge quantisation condition}.}
\subsubsection*{Case $C>0$:}
The Gibbons-Hawking-York boundary term yields
\begin{equation}
S_{\text{GHY}} = \left.-3Ar^2 \left(\sqrt{1+ \frac{C}{r^4}}-1\right) \right|_{r_c}^{\infty} = 3Ar_c^2\left(\sqrt{1+ \frac{C}{r_c^4}}-1\right),
\end{equation}
where $A=2 \pi^2$.
The contribution from \eqref{Action without potential} is given by the integral
\begin{align} \nonumber
S &= - \frac{n^2}{Af_{\text{ax}}^2} \int_{\varphi(r_c)}^{\varphi(\infty)} \frac{\exp(-\alpha \varphi)}{\sqrt{n^2 \exp(-\alpha \varphi)/(A^2f_{\text{ax}}^2)+6C}} = \\
&= \left.\frac{2n}{\alpha f_{\text{ax}}} \sqrt{\exp(-\alpha \varphi)+\sinh^2 K_+} \right|_{\varphi(r_c)}^{\varphi(\infty)},
\end{align}
where we used \eqref{C for C>0}. Combining those two results and taking $r_c \to 0$, we obtain the instanton action
\begin{equation} \label{Instanton action for C>0}
S_{\text{inst}} = \frac{2}{\alpha} \frac{n}{f_{\text{ax}}} \left(e^{-K_+} + \frac{\alpha}{2} \sqrt{\frac{3}{2}} \sinh K_+\right).
\end{equation}
As before, we need to ensure that the integral from $r=0$ to $r=r_c$ only gives a minor contribution to the full instanton action \eqref{Instanton action for C>0}. This contribution is
\begin{align}
\Delta S &\equiv \left.(S+ S_{\text{GHY}})\right|_{r=0}^{r=r_c} = \\
&= \frac{2}{\alpha}\frac{n}{f_{\text{ax}}} \left[\sqrt{\exp(-\alpha \varphi(r_c))+\sinh^2 K_+}- \left(1- \frac{\alpha}{2} \sqrt{\frac{3}{2}}\right) \sinh K_+ \right] \nonumber \\
&~~~~~- 3Ar_c^2 \left(\sqrt{1+ \frac{C}{r_c^4}}-1\right). \nonumber
\end{align}
In the limit $r_c^2 / \sqrt{C} \ll 1$ this can be simplified to
\begin{align}
\nonumber \Delta S & = \frac{4 n}{\alpha f_{\text{ax}}} \sinh K_+ \left(\frac{r_c^2}{2 \sqrt{C}}\right)^{\frac{\alpha}{\sqrt{2/3}}} + 3Ar_c^2 + \ldots \\
& = \frac{2 n}{\alpha f_{\text{ax}}} \sinh K_+ \left[ 2{\left( \frac{r_c^2}{\sqrt{C}} \right)}^{\frac{\alpha}{2 \sqrt{2/3}}} + \frac{\alpha}{2} \sqrt{\frac{3}{2}} \left( \frac{r_c^2}{\sqrt{C}} \right) \right] + \ldots \ ,
\end{align}
where omitted terms decrease as $r_c^4/C$. The condition $\Delta S \ll S_{\text{inst}}$ turns out to be self-consistent with the imposed regime $r_c^2 / \sqrt{C} \ll 1$. More precisely, by choosing $\sqrt{C}$ sufficiently large one can always ensure that $\Delta S \ll S_{\text{inst}}$. According to \eqref{C for C>0} this is equivalent to choosing $(n \sinh K_+)/f_{\textrm{ax}}$ sufficiently large. This is very similar to the parametric situation encountered above for $C=0$.
To determine the strongest constraints on inflation we are interested in identifying the instantons with the smallest action. For a given value of $n/f_{\textrm{ax}}$ and at a fixed dilaton coupling $\alpha$ cored gravitational instantons correspond to a family of solutions parameterised by $K_+$ (see \autoref{GravInstScalar}). We wish to identify the instanton with the smallest action in this family. As pointed out in \autoref{Subsection: Integration constant}, while $K_+$ is expected to take discrete values, it can be effectively treated as a continuous parameter in the limit of macroscopic objects. Hence we can determine the solutions with the smallest action by formally extremising \eqref{Instanton action for C>0} with respect to $K_+$ as it was done in \cite{150906374}. For $\alpha \geq 2 \sqrt{2/3}$ the action of cored instantons is always bigger than that of extremal instantons. If $0<\alpha < 2 \sqrt{2/3}$, the smallest cored instanton action is as big as the extremal instanton action for $\alpha = 2 \sqrt{2/3}$. To summarise, we obtain
\begin{equation}
S_{\text{cored}}(\alpha) \geq \begin{cases}
S_{\text{extremal}}(\alpha) & \text{for}~ \alpha \geq 2 \sqrt{2/3} \\
S_{\text{extremal}}(\alpha = 2 \sqrt{2/3}) & \text{for}~ \alpha < 2 \sqrt{2/3}
\end{cases} \ ,
\end{equation}
where the extremal instanton action was computed above in \eqref{Instanton action C=0}.
The upshot is that the contributions to the axion potential due to cored gravitational instantons will always be subleading compared to the effects due to a suitable extremal instanton. As we are interested in determining the strongest constraints on axion inflation, we will hence neglect cored instantons in the following analyses and focus on extremal instantons and Euclidean wormholes instead.
\subsubsection*{Case $C<0$:}
For Euclidean wormholes the coordinate $r$ is defined on $r \in [r_0,+ \infty)$, where $r_0 \equiv |C|^{1/4}$ is the size of the wormhole at the centre. As long as $r_0 \gtrsim r_c$ one can safely integrate from $r=r_0$ to $r= \infty$. As $r_0 \equiv |C|^{1/4} \propto n/ f_{ax}$ (see \eqref{C for wormholes}) the condition $r_0 \gtrsim r_c$ can be fulfilled by choosing $n/f_{ax}$ sufficiently large.
As pointed out in \autoref{GravInstScalar} we will only consider wormholes with dilaton couplings $\alpha < 2 \sqrt{2/3}$ in order to have regular solutions for $\varphi$. We then proceed with calculating the action. The Giddings-Hawking-York boundary term vanishes \cite{Giddings:1987cg},
\begin{equation}
S_{\text{GHY}} = 0,
\end{equation}
since two asymptotically flat regions are connected by a handle and thus the integral gives zero.
The on-shell contribution from \eqref{Action without potential} for only half of the wormhole\footnote{To get the instanton action, we have to divide the full wormhole action by two, as the wormhole represents a pair of instanton and anti-instanton. For more details, see \autoref{Appendix:Instanton Action}.} is given by
\begin{align} \nonumber
S_{\text{inst}} &= - \frac{n^2}{Af_{\text{ax}}^2} \int_{\varphi(r_0)}^{\varphi(\infty)} \frac{\exp(-\alpha \varphi)}{\sqrt{n^2 \exp(-\alpha \varphi)/(A^2f_{\text{ax}}^2)-6|C|}} = \\ \label{Instanton action C<0}
&= \frac{2}{\alpha} \frac{n}{f_{\text{ax}}} \sin \left(\frac{\alpha \pi}{4}\sqrt{\frac{3}{2}}\right),
\end{align}
where we used the solutions for $C<0$ from \autoref{GravInstScalar}.
Notice that the limit $\alpha \to 0$ corresponds to the Giddings-Strominger wormhole \cite{Giddings:1987cg}, and we have
\begin{equation} \label{Instanton action Giddings-Strominger}
S_{\text{inst}} = \frac{\pi \sqrt{6}}{4} \frac{n}{f_{\text{ax}}}.
\end{equation}
Furthermore, in the limit $\alpha \to 2 \sqrt{2/3}$ we find the instanton action of an extremal instanton with $\alpha =2 \sqrt{2/3}$.
\subsubsection*{Summary}
We summarise our results for the instanton action.
For one, the instanton action $S_{\text{inst}}$ scales as $S_{\text{inst}} \sim n/f_{\text{ax}}$ for all three types of gravitational instanton. Results were obtained in an effective theory with a cutoff at a length scale $r_c$. The existence of this cutoff implies that not all gravitational instanton solutions can be trusted in the framework of the effective theory. One can derive a criterion for deciding which gravitational instantons to include. While numerical factors may vary, this condition exhibits the same parametric behaviour for all three types of gravitational instantons: given a cutoff at a length scale $r_c$ one has to choose $n /f_{\textrm{ax}} \gg r_c^2$ for being able to trust the instanton action computed in the effective theory.
In order to determine the importance of such gravitational instantons it is crucial to estimate the size of the cutoff scale $r_c$. The first step is to see whether moduli stabilisation places a lower bound on $r_c$.
\section{Gravitational Instantons and Moduli Stabilisation} \label{Section:Moduli stabilisation}
We now want to make progress towards realistic string
compactifications. The pure Einstein-axion system is relevant only
below the moduli scale. Above that scale, moduli can play the role of an additional scalar $\varphi$ with dilatonic coupling to the axion or 2-form kinetic term. We will make use of our detailed discussion of this extended system in \autoref{Section:Gravitational Instanton Solutions } and \autoref{Section:Cored Grav Inst}.
\subsection{Gravitational Instantons in the presence of a potential}
\label{sec:addmass}
We only consider the lightest modulus, which we will call $\varphi$. For instance, it could be the saxion associated with the axion $\theta$.
We will assume stabilisation at $\varphi = 0$. In the throat region of the instanton, the modulus will be typically driven away from this value. We will assume that this displacement is small enough so that the potential of the modulus can be approximated by a mass term, i.e.~$V=m^2 \varphi^2/2$.
The obvious extension of \eqref{Action with H} is then
\begin{equation} \label{Action with potential}
S = \int d^4x \sqrt{g} \left[-\frac{1}{2}R + \frac{1}{2}\mathcal{F}(\varphi)H^2 + \frac{1}{2} g^{\mu\nu} \partial_{\mu}\varphi \partial_{\nu} {\varphi} + V(\varphi) \right].
\end{equation}
We take ${\cal F}$ to be exponential, which is the case discussed in detail earlier and which is typical for string-derived models (see \autoref{Subsection: Dilaton coupling from string theory}). Nevertheless, due to the presence of the potential, solutions are more complicated than before. We make the most general ansatz respecting spherical symmetry
\begin{equation} \label{Metric Ansatz General}
ds^2 = \lambda(r)dr^2 + r^2 d\Omega_3^2,
\end{equation}
as in \autoref{Appendix: metric structure}. From the derivation therein it becomes clear that $\lambda(r)$ is no longer given by $(1+C/r^4)^{-1}$.
However, we will see that for $r\ll r_*\equiv 1/m$ the mass term is negligible and we can use the approximation $\lambda(r) \simeq (1+C/r^4)^{-1}$ (cf. the related discussion in \cite{Abbott:1989jw}). Thus, the three types of gravitational instantons analysed above remain relevant.
The fact that the mass term is negligible close to the centre of the instanton is intuitively clear: The field strength contribution to the energy-momentum tensor increases as one approaches the centre and hence, for sufficiently small $r$, the contribution from the mass term becomes subdominant. This will become more explicit below.
Employing \eqref{Metric Ansatz General}, the Einstein equation $G_{rr}=T_{rr}$ and the Klein-Gordon equation read
\begin{align}
\frac{1}{2}(\varphi^{\prime})^2 - \lambda(r) V(\varphi)+ \frac{3}{r^2}(\lambda(r)-1) - 3 \lambda (r) \mathcal{F}(\varphi) \frac{n^2}{A^2 r^6}&= 0 \label{Einstein eq} \\
\varphi^{\prime \prime} + \left( \frac{3}{r} - \frac{\lambda^{\prime}(r)}{2 \lambda(r)} \right) \varphi^{\prime} - \lambda(r) V^{\prime}(\varphi) - 3 \lambda(r) \mathcal{F}^{\prime}(\varphi) \frac{n^2}{A^2 r^6} &= 0. \label{KG eq}
\end{align}
Here we also used \eqref{Ansatz for H} and \eqref{Solution for h(r)}, which specify the profile of $H$.
\subsubsection*{Approximation}
As already sketched above, the strategy is as follows: Let $\varphi_0(r)$ and $\lambda_0 (r) \equiv (1+C/r^4)^{-1}$ be the field and metric profiles for $V\equiv 0$. Then we work out the conditions under which
\begin{equation}
T_{rr}(\varphi_0, \lambda_0) \gg V(\varphi_0)\,.
\end{equation}
This specifies the regime where we can expect the $\varphi_0(r)$ and $\lambda_0(r)$ to provide good approximations to the true solutions $\varphi(r)$ and $\lambda(r)$.
We now go into more detail: The full energy-momentum tensor of \eqref{Action with potential} reads
\begin{equation}
T_{\mu\nu} = - g_{\mu\nu} \left[\frac{1}{2}\mathcal{F}(\varphi)H^2 + \frac{1}{2} \partial_{\rho}\varphi \partial^{\rho}\varphi + V(\varphi)\right] + 3 \mathcal{F}(\varphi)H_{\mu\rho\sigma}H_{\nu}^{~\rho\sigma} + \partial_{\mu}\varphi \partial_{\nu}\varphi.
\end{equation}
Taking $\varphi = \varphi_0$ and $\lambda = \lambda_0$ we find
\begin{equation}
T_{rr}(\varphi_0, \lambda_0) = \frac{3C}{r^6(1+C/r^4)} - \frac{V(\varphi_0)}{1+C/r^4},
\end{equation}
where \eqref{Einstein equation without potential} was used. We see that the potential is negligible compared to the curvature contributions if (for $C \neq 0$)
\begin{equation} \label{Criterion for neglecting the potential}
\left|\frac{3C}{r^6}\right| \gg \frac{1}{2}m^2 \varphi_0^2\,.
\end{equation}
Appealing again to \eqref{Einstein equation without potential}, we first consider the regime $r\gg |C|^{1/4}$. Then $\varphi_0^{\prime}(r) \sim 1/r^3$ and hence
\begin{equation} \label{Phi asymptotically}
\varphi_0 (r) \sim \frac{1}{r^2}\,.
\end{equation}
Here we treat $n/A$ and $C$ as `${\cal O}(1)$ factors' and disregard them.
We explain this below.
With this, \eqref{Criterion for neglecting the potential} translates to
\begin{equation}
\label{rcriterion}
r \ll r_*\equiv \frac{1}{m} \ .
\end{equation}
Now, our interest is in the case $m \ll1$, i.e.~in moduli much lighter than the Planck scale. This implies $r_* \gg 1$ so that $r_* \gg |C|^{1/4}$, giving us a large validity range for our approximation $\varphi_0 \sim 1/r^2$. Crucially, while $|C|$ also figured as a large parameter in other parts of this paper, here the much stronger hierarchy $1/m\gg 1$ dominates and our crude approximation concerning `${\cal O}(1)$ factors' is justified.
Next, we need to consider the region $r\lesssim |C|^{1/4}$. While here the profile $\varphi_0(r)$ is more complicated, we are now deeply inside the regime of large field strength. It is easy to convince oneself that the potential $\sim m^2\varphi^2$ remains subdominant. What is less obvious is whether the $m^2 \varphi^2$ approximation remains justified, given that the field now moves significantly away from zero. This will be discussed later.
Finally, the extremal instanton with $C=0$ requires an extra comment. In this case the energy-momentum tensor vanishes and the criterion \eqref{Criterion for neglecting the potential} is no longer applicable. Instead, we require that the mass term in \eqref{Einstein eq} should be subdominant compared to every other term in this equation, i.e.
\begin{equation}
m^2 \varphi_0^2 \ll \frac{3 \mathcal{F}(\varphi_0)n^2}{A^2r^6},
\end{equation}
which yields again the condition $r \ll 1/m$ (here we used that $\mathcal{F}$ is approximately constant for large $r$). Note that in this case the behaviour of $\varphi_0$ at large $r$ is specified by \eqref{Solution for C=0} and the role of the `largish' parameter $|C|^{1/4}$ is taken over by $n/f_{\mathrm{ax}}$.
To summarise, we have now argued rather generally that the gravitational solutions found in the absence of a potential are good approximations for $r \ll 1/m$. We will not need the behaviour of $\varphi$ outside that region, at $r \to \infty$. Indeed, by redefining $\varphi$ we can, as argued before, always ensure that the $\varphi_0$ asymptotically approaches the minimum of the potential at $\varphi=0$. Thus, even while the actual profile of $\varphi(r)$ can significantly deviate from $\varphi_0(r)$ at $r\gg 1/m$, there is no doubt that the fundamental property of $\varphi$ approaching zero at large $r$ will be maintained. Crucially, since $1/m\gg |C|^{1/4}$ and the action integral is dominated by the region $r\lesssim |C|^{1/4}$, we can also trust the zero-potential solutions for evaluating the action, independently of the large-$r$ region.
\subsection{Dilaton Couplings from String Compactifications} \label{Subsection: Dilaton coupling from string theory}
The gravitational solutions in \autoref{GravInstScalar} were obtained for scalars with dilatonic couplings, i.e.~where the prefactor of the kinetic term for the axion is given by $\mathcal{F}(\varphi)=e^{-\alpha \varphi}/(3! f_{\textrm{ax}}^2)$. This form frequently occurs for effective theories obtained from string theory compactifications. The value of $\alpha$ will depend on the precise identification of the axion and scalar with the corresponding fields in the string compactifications. In the following, we will provide relevant examples.
\subsubsection*{The Axio-Dilaton}
Let us first consider the case where both the axion and the scalar descend from the axio-dilaton field $S= C_0 + i/g_s$ with string coupling $g_s$ and universal axion $C_0$. It appears in the K\"ahler potential as
\begin{equation}
\mathcal{K} = - \ln \left(-i(S-\bar{S})\right).
\end{equation}
The kinetic term of the Lagrangian $\mathscr{L} \supset \mathcal{K}_{S\bar{S}} \partial_{\mu} S \partial^{\mu} \bar{S}$ then becomes
\begin{equation}
\mathscr{L} \supset \frac{g_s^2}{4} (\partial C_0)^2 + \frac{1}{4g_s^2} (\partial g_s)^2 \ .
\end{equation}
Canonical normalisation of our saxion gives $g_s = g_s^0 \exp(\sqrt{2}\varphi)$. Thus, the field strength coupling reads
\begin{equation}
\mathcal{F}(\varphi) = \frac{1}{2 \cdot 3!}\frac{1}{\mathcal{K}_{S\bar{S}}} = \frac{1}{3(g_s^0)^2} \exp(-2\sqrt{2}\varphi),
\end{equation}
so in our notation the dilaton coupling $\alpha$ is $\alpha=2 \sqrt{2}$. Notice that $\varphi \to \infty$ corresponds to the strong coupling limit, while the weak coupling limit is given by $\varphi \to -\infty$.
\subsubsection*{K\"ahler Moduli at Large Volume}
Let us now consider the K\"ahler moduli sector at large volume. In particular, consider the case where the volume is dominated by one K\"ahler modulus $T$. For example, this arises in the scheme of moduli stabilisation known as the Large Volume Scenario (LVS) \cite{0502058}. The relevant part of the K\"ahler potential is
\begin{equation}
\mathcal{K} = -2 \ln \mathcal{V} = - 3 (T+ \bar{T}) + \ldots \ .
\end{equation}
Here we wish to identify the saxion with $\textrm{Re}(T)$ and the axion with $\textrm{Im}(T)$. The leading contribution to the kinetic term for the saxion and axion is then given by
\begin{equation}
\mathcal{K}_{T \bar{T}} = \frac{3}{(T+\bar{T})^2} \ .
\end{equation}
Canonical normalisation gives
\begin{equation}
\mathrm{Re}(T) = \exp \left(-\sqrt{\frac{2}{3}} \varphi\right),
\end{equation}
and hence
\begin{equation}
\mathcal{F}(\varphi) \sim \exp \left(-2\sqrt{\frac{2}{3}} \varphi\right) \ .
\end{equation}
The dilaton coupling is thus $\alpha = 2 \sqrt{2/3}$.
\subsubsection*{Complex Structure Moduli in the Large Complex Structure Limit (LCS)}
It is well-known that complex structure moduli in the LCS limit give rise to a shift-symmetric structure in the K\"ahler potential. Let $u$ be a complex structure modulus in the LCS regime and $z$ denote the remaining complex structure moduli. Then we have
\begin{align} \nonumber
\mathcal{K} &= -\ln \left(\kappa_{uuu}(u+\bar{u})^3 + \kappa_{uui}(u+\bar{u})^2(z_i+ \bar{z}_i) + \frac{\kappa_{uij}}{2!}(u+\bar{u})(z_i+ \bar{z}_i) (z_j+ \bar{z}_j) \right.+ \\
&\left.+ \frac{\kappa_{ijk}}{3!}(z_i+ \bar{z}_i) (z_j+ \bar{z}_j)(z_k+ \bar{z}_k) + f(z_i) \right),
\end{align}
where the $\kappa_{ijk}$ denote the intersection numbers of the mirror-dual Calabi-Yau three-fold and $f$ is a function of the remaining complex structure moduli $z_i$ and accounts for instantonic corrections to the K\"ahler potential. For the moment only $u$ shall be stabilised in the LCS limit, i.e.
\begin{equation}
\mathrm{Re}(u) > 1 \ .
\end{equation}
Thus, one obtains
\begin{equation}
\mathcal{K}_{u\bar{u}} = \frac{3}{(u+\bar{u})^2}
\end{equation}
at leading order as long as $\kappa_{uuu} \neq 0$. Omitted terms scale as $(u+\bar{u})^{-3}$. Therefore, canonical normalisation yields
\begin{equation}
\label{eq:cannormsaxu} \mathrm{Re}(u) = \exp \left(-\sqrt{\frac{2}{3}} \varphi\right),
\end{equation}
and hence $\alpha = 2\sqrt{2/3}$.
In the situations studied so far the saxion and axion arose from the same complex scalar field. However, one may also consider the case where the saxion and axion originate from different moduli. To give just one example, let us again consider the complex structure sector of a CY threefold, but now we will assume that two complex structure moduli $u,v$ are in the LCS regime. Further, we assume the following hierarchy
\begin{equation}
\mathrm{Re}(u) \gg \mathrm{Re}(v) \gg 1 \ .
\end{equation}
We will now consider the axionic field $\textrm{Im}(v)$ and study the coupling to the saxion $\textrm{Re}(u)$.
As before, the leading contribution to the kinetic term of the saxion is
\begin{equation}
\mathcal{K}_{u\bar{u}} = \frac{3}{(u+\bar{u})^2} \ ,
\end{equation}
and the canonically normalised saxion is given by \eqref{eq:cannormsaxu}. The leading contribution to the kinetic term for the axion is
\begin{equation}
\mathcal{K}_{v\bar{v}} \sim \frac{1}{(u+\bar{u})^2} \sim \exp \left(-2\sqrt{\frac{2}{3}} \varphi\right) \ ,
\end{equation}
where omitted terms decrease as $(u + \bar{u})^{-3}$. While both $\mathcal{K}_{u\bar{u}}$ and $\mathcal{K}_{v\bar{v}}$ scale as $(u + \bar{u})^{-2}$, this behaviour has different origins in the two cases. The leading contribution to $\mathcal{K}_{u\bar{u}}$ comes from $\kappa_{uuu} (u + \bar{u})^3$, whereas it is the terms $\kappa_{uvv} (u + \bar{u})(v + \bar{v})^2$ and $\kappa_{uuv} (u + \bar{u})^2(v + \bar{v})$ which contribute to $\mathcal{K}_{v\bar{v}}$ at leading order.
Despite these differences we again find
\begin{equation}
\mathcal{F}(\varphi) \sim \exp \left(-2\sqrt{\frac{2}{3}} \varphi\right) \ ,
\end{equation}
and $\alpha = 2\sqrt{2/3}$.\footnote{This would be different if $\kappa_{uuu}$ and $\kappa_{vvv}$ were the only non-vanishing intersection numbers. Then we would still have $\mathcal{K}_{u \bar{u}} \simeq 3/(u + \bar{u})^2$ but now $\mathcal{K}_{v\bar{v}} \sim (v+ \bar{v})/(u+\bar{u})^3$ for $\text{Re}(u) \gg \text{Re}(v)$. Assuming that $\text{Re}(v)$ is stabilised such that we can take it as constant, we would now find a dilaton coupling $\alpha = \sqrt{6}$.}
Note that in all the cases examined above the dilaton coupling is just outside the range allowing for Euclidean wormhole solutions $0 \leq \alpha < 2 \sqrt{2/3}$. This observation and a possible way out have been pointed out in \cite{0412183, 07052768}. The idea is as follows. Even if wormholes charged under individual axions do not exist, one can nevertheless find solutions which are charged under more than one axion (see also \cite{Tamvakis:1989aq}).
We will conclude this section with an example that allows for the existence of Euclidean wormhole solutions and may be useful to illustrate and develop the above idea. Let us consider both the axio-dilaton sector and the complex structure moduli sector of a CY 3-fold at LCS:
\begin{equation}
\mathcal{K} = - \ln \left(-i(S-\bar{S})\right) - \ln \left(\kappa_{uuu}(u+\bar{u})^3\right) \ .
\end{equation}
In the spirit of \cite{Tamvakis:1989aq, 0412183, 07052768} we could now investigate Euclidean wormhole solutions charged under both the universal axion as well as the complex structure axion. Alternatively, we may assume that we can stabilise moduli such that $S=iu$. Then we effectively have the theory of one 4-fold complex structure modulus and we obtain
\begin{equation}
\mathcal{K}_{u\bar{u}} = \frac{4}{(u+\bar{u})^2} \ .
\end{equation}
Taking the saxion as $\textrm{Re}(u)$ and the axion as $\textrm{Im}(u)$ we now find $\alpha = \sqrt{2}$ which lies within the range allowing for wormholes. We leave it for future work to investigate whether this pattern of moduli stabilisation can be realised in a realistic compactification.
\subsection{Maximal Field Displacements of Dilatonic Fields} \label{Subsection: Maximal field displacement}
In the previous sections we have made progress towards studying gravitational instantons in the presence of moduli. The results of \autoref{Subsection: Dilaton coupling from string theory} imply that a restriction to moduli with dilatonic couplings is well-motivated from string compactifications. We also made progress towards understanding the role of the potential stabilising the modulus in \autoref{sec:addmass}. In particular, in the regime $r \ll 1/m$ the potential can be ignored and gravitational instanton solutions for a massless dilaton will be good approximations.
There is another effect which we need to take into consideration. When approaching the core of a gravitational instanton, the value of the dilaton increases. In the case of a Euclidean wormhole it reaches a maximum at the wormhole throat, while for extremal and cored instantons the dilaton diverges for $r \rightarrow 0$. However, we cannot afford arbitrarily large field displacements, as this will take us outside the range of validity of our effective theory.
To be specific, consider the effective theory of the axio-dilaton at weak string coupling. When approaching the centre of a gravitational instanton solution the string coupling increases compared to its asymptotic value. If it becomes too large the supergravity regime breaks down and we cannot trust our solutions. A similar argument can be made for any effective dilaton-axion theory from string compactifications.
This gives us an additional criterion to decide which gravitational instantons to trust and which ones to disregard. We will analyse this condition focussing on Euclidean wormholes and extremal gravitational instantons. In the following, we will denote by $\varphi_{\textrm{max}}$ the threshold value at which the effective theory breaks down.
\subsubsection*{Euclidean Wormholes}
For Euclidean wormholes the displacement becomes maximal at the throat of the wormhole. Using our solution for $\varphi(r)$ \eqref{Solution for C<0 w/o constant K_-} one finds:
\begin{equation}
\varphi(r_0) = - \frac{1}{\alpha} \ln \cos^2 \left(\frac{\alpha \pi}{4} \sqrt{\frac{3}{2}}\right).
\end{equation}
To trust the wormhole solution we require $\varphi(r_0) < \varphi_{\textrm{max}}$. The maximal displacement only depends on the dilaton coupling $\alpha$, which is not a free parameter, but a property of the physical system studied. As a result the maximal displacement is model-dependent.
Recall that Euclidean wormhole solutions only exist for dilaton couplings in the range $0 \leq \alpha < 2 \sqrt{\tfrac{2}{3}}$. The maximal displacement at the wormhole throat is smallest for $\alpha =0$, grows when $\alpha$ is increased and eventually diverges for $\alpha \rightarrow 2 \sqrt{\tfrac{2}{3}}$. To give just one example, the value $\alpha = \sqrt{2}$ yields a displacement $\varphi(r_0) - \varphi(\infty) = \varphi(r_0) \simeq 2.2$ in Planck units, which may already be critical.
Another important result from this section is that the maximal displacement $\varphi (r_0)$ is independent of the ratio $n/f_{\text{ax}}$, or, equivalently, the wormhole radius at the throat $r_0$. Hence we do not get any additional constraints on these quantities due to the displacement of the saxion.
\subsubsection*{Extremal Instantons}
For extremal gravitational instantons the $\varphi$-profile exhibits a divergent behaviour for $r \to 0$. As laid out in \autoref{Section:Cored Grav Inst}, we will nevertheless trust such solutions as long as the dominant part of the instanton action arises from the region $r > r_c$, where $r_c$ is the length cutoff of our effective theory. This cutoff will be discussed in more detail in \autoref{sec:selfdual}. Here we will show that the displacement of the saxion gives an independent condition for the reliability of our action.
Let us be more precise. Given a threshold value $\varphi_{\textrm{max}}$ beyond which our theory breaks down, we can identify a radius $r_{\textrm{min}}$ at which the dilaton crosses this value: $\varphi(r_{\textrm{min}}) = \varphi_{\textrm{max}}$. This can be made explicit using \eqref{Solution for C=0}:
\begin{equation}
e^{\alpha \varphi_{\mathrm{max}}} = \left(1+ \frac{\alpha n}{4A f_{\mathrm{ax}}} \frac{1}{r_{\mathrm{min}}^2}\right)^2 \ ,.
\end{equation}
To trust our solution we need to ensure that $\Delta S / S \ll 1$, where $\Delta S$ is the contribution to the instanton action from the region $r < r_{\mathrm{min}}$. Using \eqref{Instanton action C=0} and \eqref{Delta S for C=0} we find
\begin{equation}
\frac{\Delta S}{S} = \left(1+ \frac{\alpha n}{4A f_{\text{ax}}} \frac{1}{r_{\mathrm{min}}^2}\right)^{-1} = \exp \left(- \tfrac{\alpha }{2} \varphi_{\textrm{max}} \right) \,.
\end{equation}
Hence the relevant condition is
\begin{equation}
\exp \left(- \tfrac{\alpha }{2} \varphi_{\textrm{max}} \right) \ll 1 \,,
\end{equation}
which gives an additional (model-dependent) constraint.
Last, let us return to one aspect encountered for the case of Euclidean wormholes. There we observed that for $\alpha \rightarrow 2 \sqrt{\tfrac{2}{3}}$ the saxion displacement at the wormhole throat grows without bound and would exceed any finite value $\varphi_{\text{max}}$. Note that this does not necessary constitute a pathology. Rather, the behaviour observed for a wormhole becomes similar to that of an extremal instanton. In fact, in the limit $\alpha \rightarrow 2 \sqrt{\tfrac{2}{3}}$ the Euclidean wormhole becomes a pair of extremal instantons. We can then deal with the divergence of $\varphi$ as in the case of extremal instantons and cut our solution off at some $r= r_{\textrm{min}}$.
\section{Consequences for Large Field Inflation}
\label{Section:Inflation}
In this section we will analyse to what extent gravitational instantons constrain axion inflation. The idea is as follows: we will check whether the contribution to the axion potential $\delta V$ due to gravitational instantons can be large enough to disrupt inflation.
To be specific, gravitational instantons contribute as
\begin{equation} \label{Delta V of gravitational instantons}
\delta V = \mathcal{A} \cos (n\theta) e^{-S} \ ,
\end{equation}
where $S \sim n/f_{\text{ax}}$ (see \autoref{Section:Cored Grav Inst}). Whether such instanton corrections can have significant influence on the slow-roll dynamics clearly depends both on the size of the instanton action $S$ and the prefactor $\mathcal{A}$. The latter is quoted to be of order $M_p^4$ \cite{150300795,150303886}. However, in \autoref{Appendix: Prefactor} we give arguments why the prefactor $\mathcal{A}$ can be significantly below the Planck scale in more realistic string compactification models. Specifically, we expect $\mathcal{A}$ to scale as $\mathcal{A} \sim \mathcal{V}^{-5/3}$ with compactification volume $\mathcal{V}$.
We then compare $\delta V$ with the size of the axion potential during inflation. For large field inflation the scale of inflation is of the order \cite{150202114}
\begin{equation}
V_{\textrm{inflation}} \sim 10^{-8}.
\end{equation}
Hence, whenever we find $\delta V \sim 10^{-8}$ we will conclude that the effects of gravitational instantons on the axion potential are in principle large enough to spoil inflation.
In what follows we compute $\delta V$ only for the case of a single axion, but our results can be straightforwardly extended to models of $N$-flation, kinetic alignment and the Kim-Nilles-Peloso mechanism, see \cite{150303886} for more details.
\subsection{Action of the most `dangerous' Gravitational Instantons}
\label{sec:mostdanger}
To check whether gravitational instantons are dangerous for inflation, we want to focus on the instantons with the smallest action. At the same time, we need to ensure that these are solutions which we can trust within the framework of our effective theory. In brief, we are interested in the most relevant instanton within the regime of validity of our theory.
The breakdown of our effective gravity theory is crucial in this context, because it will put a lower bound on the instanton action $S$. As explained in \autoref{Section:Cored Grav Inst}, in a theory with length cutoff $r_c$ we can only trust gravitational instanton solutions with $n/ f_{\textrm{ax}} \gg r_c^2$. This translates into a lower bound on the instanton action as $S \sim n/ f_{\textrm{ax}}$.
To calculate the contributions of gravitational instantons to the axion potential we hence need to determine the cutoff $r_c$. In \autoref{sec:selfdual} we will estimate the smallest possible value of $r_c$ at which the description in terms of a 4-dimensional theory may hold. Before doing this it will be instructive to check how large $r_c$ can be so that gravitational instantons still induce a sizeable contribution to the inflaton potential.
Note that gravitational instanton solutions for the case of a massless dilaton will be sufficient for our analysis, despite the fact that we are interested in the case of massive dilaton fields. As described in \autoref{sec:addmass} the non-zero potential does not affect the action significantly.
\subsubsection*{Euclidean Wormholes}
For any $n$ and $f_{\text{ax}}$ the Euclidean wormhole action is computed in \eqref{Instanton action C<0}. At the same time the wormhole radius $r_0$ scales as $r_0 \sim (n/f_{\text{ax}})^{1/2}$ according to \eqref{C for wormholes}. As we require $r_0 \gtrsim r_c$ we get
\begin{equation}
S_{\text{inst}} \gtrsim (2\pi^2) \sqrt{6} \frac{2}{\alpha} \tan \left(\frac{\alpha \pi}{4} \sqrt{\frac{3}{2}} \right) r_c^2 \ .
\end{equation}
On the allowed interval $0 \leq \alpha < 2 \sqrt{2/3}$ the instanton action as a function of $\alpha$ increases monotonically. Therefore, the most dangerous wormhole corresponds to the Giddings-Strominger instanton with $\alpha=0$. Hence
\begin{equation}
\label{DangerWH}
S_{\text{inst}} \geq S_{\text{inst}} (\alpha=0) \gtrsim 3 \pi^3 r_c^2.
\end{equation}
Demanding that $e^{-S} \gtrsim 10^{-8}$ implies $r_c \lesssim 0.4$ (in Planck units).
In \autoref{Subsection: Dilaton coupling from string theory} we found that $\alpha=\sqrt{2}$ can be obtained from string compactifications and still lies in the allowed range of dilaton-couplings appropriate to allow for Euclidean wormholes. This example requires $r_c \lesssim 0.2$ in order to get a contribution of at least $\delta V \sim 10^{-8}$.
Note that the prefactor $\mathcal{A}$ (see \autoref{Appendix: Prefactor}) may potentially lower the size of the contribution to the inflaton potential.
\subsubsection*{Extremal Gravitational Instantons}
The action for extremal gravitational instantons is obtained from \eqref{Instanton action C=0}. However, we have to take into account the computability condition \eqref{Condition dS << Sinst} for the action. It follows that
\begin{equation}
\label{DangerExt}
S_{\text{inst}} > \frac{8 \cdot (2\pi^2)}{\alpha^2}r_c^2 \ .
\end{equation}
In string theory $\alpha$ cannot be chosen arbitrarily large. The largest $\alpha$ we could obtain from string compactifications was $\alpha = 2\sqrt{2}$. Extremal gravitational instantons then become relevant if $r_c \lesssim 1$. Hence, extremal gravitational instantons may turn out to be somewhat more dangerous for axion-inflation than Euclidean wormholes.
We do not consider cored gravitational instantons, for which a similar analysis could be made. The reason is that their action is always larger than that of a suitable extremal instanton (see \autoref{Section:Cored Grav Inst}).
\bigskip
The question we want to address now is how small $r_c$ can be in any controlled model of quantum gravity. Knowing that moduli displacements are not an issue, one would naively expect that $r_c \simeq 1$ can be problematic as we reach already Planck regime. Notice however that it is important to determine $r_c$ as precisely as possible, because due to $\delta V \sim e^{-S} \sim e^{-r_c^2}$ the instanton contributions are very sensitive to every $\mathcal{O}(1)$-factorchange in the cutoff radius.
\subsection{Estimating the Critical Radius $r_c$}
\label{sec:selfdual}
Let us take string theory as our model of quantum gravity. String compactifications then yield a hierarchy of scales in the effective theory as depicted in \autoref{Fig:scales}. We expect that going beyond the Kaluza-Klein scale will render our effective description insufficient. The reason is that the gravitational instanton solutions we consider are obtained in a 4-dimensional effective theory which arises from a more fundamental description upon compactification. For the 4-dimensional picture to remain valid, we require the length scale $r_c$ associated with our 4-dimensional solution not to be smaller than the length scale associated to the compactified extra-dimensions.
But how small can this length scale be? In string theory it cannot be arbitrarily small. String compactifications exhibit a property termed T-duality which states that a compactification with a small volume describes the same physics as another compactification with large volume. This gives rise to the notion of a smallest length scale at the self-dual value of the compactification volume.
Putting everything together, we arrive at the smallest possible value $r_c$ where we can trust our effective 4-dimensional analysis. We find that $r_c$ should be related to the length scale of the compact dimensions at self-dual volume $\mathcal{V}_{\text{sd}}$ of the compactification space. In this way, we push the KK-scale as close to the Planck scale as possible, allowing us to consider the lightest gravitational instantons we can obtain within the regime of validity of our description.
What we mean by ``related'' is at this naive level ambiguous. There are at least two ``canonical'' possibilities to make the definition of $r_c$ more precise. They differ by factors of $\pi$, which are unfortunately crucial when comparing $e^{-S}$ with $V_{\text{inflation}}$. Given the volume $\mathcal{V}_{\text{sd}}$ of the six-dimensional compact space at the self-dual point we can define a length scale as $\mathcal{V}_{\text{sd}}^{1/6}$ and a 3-volume by $\mathcal{V}_{\text{sd}}^{1/2}$. Two possible definitions of $r_c$ are then:
\begin{enumerate}
\item The volume of the $S^3$ of our wormhole solution should not be smaller than $\mathcal{V}_{\text{sd}}^{1/2}$, i.e.
\begin{equation} \label{Self dual radius - Def1}
2\pi^2r_c^3 = \mathcal{V}_{\text{sd}}^{1/2}.
\end{equation}
\item More generously, the great circle of $S^3$ should not be smaller than $\mathcal{V}_{\text{sd}}^{1/6}$, i.e.
\begin{equation} \label{Self dual radius - Def2}
2\pi r_c = \mathcal{V}_{\text{sd}}^{1/6}.
\end{equation}
\end{enumerate}
As a toy-model to compute $\mathcal{V}_{\text{sd}}$ we take $T^6$ and apply T-duality six times for each $S^1$ to get $\mathcal{V}_{\text{sd}}(T^6) = \ell_s^6= (2 \pi)^6 (\alpha')^3$. To convert this into Planck units, recall that (see e.g.~\cite{Conlon:2006gv})
\begin{equation}
M_p^2 = \frac{4\pi \mathcal{V}}{g_s^2 \ell_s^8} \ .
\end{equation}
In the following we also go to the S-self-dual point $g_s=1$.
The first criterion \eqref{Self dual radius - Def1} then gives $r_c M_p = \sqrt{4\pi} \cdot (2\pi^2)^{-1/3} \simeq 1.3$. Using \eqref{DangerWH} and \eqref{DangerExt}, which are both in 4d Planck units, the contributions to the axion potential due to gravitational instantons are then:
\begin{align}
\nonumber \textrm{Giddings-Strominger wormhole:}& \qquad e^{-S} \simeq 10^{-68} \ , \\
\nonumber \textrm{Extremal instantons:}& \qquad e^{-S} \lesssim 10^{-15} \quad \textrm{for } \alpha=2\sqrt{2} \ .
\end{align}
Hence, in both cases the gravitational instantons appear to be irrelevant for inflation.
If we apply the second criterion \eqref{Self dual radius - Def2} we have $r_c M_p = 1/\sqrt{\pi} \simeq 0.56$. This yields
\begin{align}
\nonumber \textrm{Giddings-Strominger wormhole:}& \qquad e^{-S} \simeq 10^{-13} \ , \\
\nonumber \textrm{Extremal instantons:}& \qquad e^{-S} \lesssim 10^{-3} \quad \textrm{for } \alpha=2\sqrt{2} \ .
\end{align}
Again, Euclidean wormholes contribute to the axion potential too weakly to interfere significantly with inflation. However, extremal instantons can in principle be important, but this will depend on the value of the dilaton coupling $\alpha$. Note that for $\alpha = 2\sqrt{2/3}$ we still get $e^{-S} \lesssim 6 \cdot 10^{-9}$ for extremal instantons, which is marginal as far as the significance for inflation is concerned. However, we want to emphasise that our numerical results should be taken with a grain of salt. In particular, given a value of a length cutoff $r_c$ we only have a lower bound \eqref{DangerExt} for the action of the most important trustworthy extremal instanton. However, $\delta V$ is exponentially sensitive to the instanton action. Thus, unless the instanton action is close to saturating the inequality \eqref{DangerExt} the contributions from extremal instantons can quickly become irrelevant for inflation.
Of course, the instanton contribution $\delta V = \mathcal{A}e^{-S} \cos(n\theta)$ also involves the prefactor $\mathcal{A}$, which we estimate in \autoref{Appendix: Prefactor}. We expect $\mathcal{A} \sim \mathcal{V}^{-5/3}$, which is $\mathcal{O}(1)$ at the self-dual point. Note that in more realistic scenarios away from the self-dual point (i.e.~compactifications with a hierarchy of scales) it would suppress the gravitational instanton contributions even further.
\bigskip
Our results can be summarised as follows: overall, we find that gravitational instantons do not give rise to strong \emph{model-independent} constraints on axion inflation, even if we push the KK-scale as close to the Planck-scale as possible. Extremal gravitational instantons may be important for inflation, but this is model-dependent, as the size of their contribution depends on the value of the dilaton coupling $\alpha$.
\section{Gravitational Instantons and the Weak Gravity Conjecture}
\label{Section: WGC}
Finally, we want to make further remarks on the relation between gravitational instantons and the Weak Gravity Conjecture (WGC) \cite{0601001, 150303886,150906374}. The original form of the WGC requires that the particle spectrum of a consistent, UV-complete gravitational theory with a $U(1)$ gauge field contains at least one particle whose charge-to-mass ratio is larger or equal to that of an extremal black hole \cite{0601001}. There exists a straightforward generalisation to gravitational theories with an axion coupling to instantons. In the following we will argue in analogy with the WGC for particles with U(1) charges, i.e.~we will treat instantons like particles with axion charge.
We start with the theory of an axion with an instanton-induced potential:
\begin{equation}
\mathscr{L} = \frac{1}{2} (\partial \theta)^2 - \sum_i \Lambda_i^4 e^{-S_i} \cos (\tfrac{n_i}{f_{\text{ax}}} \theta) \ .
\end{equation}
The WGC then requires the existence of an instanton with
\begin{equation}
z \equiv \frac{n_i/f_{\text{ax}}}{S_i} > z_{0} \ ,
\end{equation}
for some $z_0$ to be specified shortly.\footnote{The WGC can be made more precise by adding further attributes to the condition $z>z_0$ \cite{0601001}. A more careful definition becomes important when several U(1) group factors (or axion species) are present. See \cite{14022287, 150303886, 150307853, 150400659, 150906374, 160608437} for more details.} The quantity $z$ is the equivalent of the charge-to-mass ratio for the instanton, where the charge is given by $n / f_{\text{ax}}$ and the mass corresponds to $S$. When working with black holes an object satisfying $z > z_0$ is referred to as superextremal, while a black hole with $z < z_0$ is termed subextremal. It will be useful to extend this nomenclature to the case of instantons. The WGC then requires the existence of superextremal instantons.
To define the WGC for instantons it is hence important to determine $z_0$. In the black hole case $z_0$ is the charge-to-mass ratio of an extremal RN black hole. By analogy, we will define $z_0$ as the charge-to-mass ratio of an extremal gravitational instanton as suggested in \cite{150906374}. There is further support for this assertion. In \autoref{Subsection: Integration constant} we saw that gravitational instantons in 4d are related to RN black holes in 5d. More specifically, the relation $C= \ell^2(M^2-Q^2)$ implies that extremal black holes ($M^2=Q^2$) are in one-to-one correspondence with extremal instantons ($C=0$). It is thus plausible that extremal instantons play the role of extremal black holes in the WGC. Using our expression \eqref{Instanton action C=0} for the action of an extremal gravitational instanton we find
\begin{equation}
z_{0} = \frac{n/f_{\text{ax}}}{S_{\text{extremal}}} = \frac{\alpha}{2} .
\end{equation}
\begin{figure}
\subfigure[]{\begin{overpic}[width=0.45\textwidth]{PlotCored1}
\put (84,0) {$n/ f_{\textrm{ax}}$} \put (0,67) {$S$} \put (0,12) {$1$} \put (57,22) {$\alpha \geq 2 \sqrt{2/3}$} \end{overpic}}
\subfigure[]{\begin{overpic}[width=0.50\textwidth]{PlotCored2}
\put (80,0) {$n/ f_{\textrm{ax}}$} \put (3,60) {$S$} \put (3,11) {$1$} \put (57,32) {$\alpha < 2 \sqrt{2/3}$} \end{overpic}}
\caption{Instanton action $S$ vs.~$n/ f_{\textrm{ax}}$ for \textbf{(a)} $\alpha \geq 2 \sqrt{2/3}$ and \textbf{(b)} $\alpha < 2 \sqrt{2/3}$.}
\label{fig:ChargeToMass}
\end{figure}
Let us now compute the charge-to-mass ratio $z$ for cored gravitational instantons and Euclidean wormholes to see how they fit into this picture. We begin with cored gravitational instantons. For fixed $n/f_{\text{ax}}$ we have
\begin{equation}
S_{\text{cored}}(\alpha) \geq \begin{cases}
S_{\text{extremal}}(\alpha) & \text{for}~ \alpha \geq 2 \sqrt{2/3} \\
S_{\text{extremal}}(2 \sqrt{2/3}) & \text{for}~ \alpha < 2 \sqrt{2/3},
\end{cases}
\end{equation}
and thus
\begin{equation}
z_{\text{cored}} \leq \begin{cases} z_0 & \text{for}~ \alpha \geq 2 \sqrt{2/3} \\ \frac{2 \sqrt{2/3}}{\alpha} \, z_0 & \text{for}~ \alpha < 2 \sqrt{2/3} \end{cases}
\end{equation}
We can make the following observation. For $\alpha \geq 2 \sqrt{2/3}$ cored gravitational instantons are strictly subextremal and do not satisfy the WGC condition $z > z_0$. They hence play a role akin to subextremal black holes in the WGC for particles. This is consistent with the finding that for $\alpha \geq 2 \sqrt{2/3}$ cored gravitational instantons are related to subextremal black holes in higher dimensions (see \cite{0406038} and \autoref{Subsection: Integration constant}). The situation is different for $\alpha < 2 \sqrt{2/3}$. The lightest cored instantons are now superextremal. We illustrate our findings in \autoref{fig:ChargeToMass}.
Next, let us turn to Euclidean wormholes. From \eqref{Instanton action C<0} we find
\begin{equation}
z_{\text{wh}} = \frac{n/f_{\text{ax}}}{S_{\text{wh}}} = \frac{\alpha}{2\sin \left(\frac{\alpha \pi}{4}\sqrt{\frac{3}{2}}\right)} > \frac{\alpha}{2} = z_{0}
\end{equation}
for $0 \leq \alpha <2 \sqrt{2/3}$, which is the allowed range for wormhole solutions. We find that Euclidean wormholes are strictly superextremal. In addition, one can also show that $z_{\text{wh}} > z_{\textrm{cored}}$. This is displayed in \autoref{fig:ChargeToMass} (b).
\begin{figure}
\centering
\begin{overpic}[width=0.65\textwidth,tics=10]{WGC.pdf}
\put (34,-1) {$n / f_{\textrm{ax}}$} \put (13,35) {extremal} \put (13,31) {instantons} \put (62,35) {extremal} \put (62,31) {instantons} \put (4,35) {$S$} \put (85,-1) {$n / f_{\textrm{ax}}$} \put (53,35) {$S$}
\end{overpic}
\caption{Possible realisations of the WGC for gravitational instantons \cite{ReeceTalk}. Red dots denote extremal gravitational instantons while blue dots correspond to additional superextremal instantons required by the WGC. }\label{fig:WGC}
\end{figure}
What can one learn from these results about the WGC? We will discuss this question for the two cases $\alpha \geq 2 \sqrt{2/3}$ and $\alpha < 2 \sqrt{2/3}$ separately.
For $\alpha \geq 2 \sqrt{2/3}$ the spectrum of gravitational instantons does not contain any superextremal objects that could satisfy the WGC. This is not surprising. Our analysis is restricted to macroscopic gravitational instantons, while it is expected that microscopic physics is responsible for satisfying the WGC. If the WGC is true, it could be realised in two different ways which are shown in \autoref{fig:WGC}. For one, extremal gravitational instantons (red dots in \autoref{fig:WGC}) could satisfy the WGC on their own. This occurs if quantum corrections decrease the instanton action for small $n$ such that they naively become superextremal (LHS of \autoref{fig:WGC}). If this is not the case (see RHS of \autoref{fig:WGC}) the WGC requires the existence of additional superextremal instantons (blue dot). At the moment it is not clear which implementation of the WGC, if any, is realised. Unfortunately, our analysis is not suitable for resolving this issue.
Let us move on to $\alpha < 2 \sqrt{2/3}$. Interestingly, the set of gravitational instanton solutions now contains superextremal objects in the form of wormholes and cored instantons. It thus seems that the WGC is satisfied in Einstein-axion-dilaton systems in virtue of cored instantons and Euclidean wormholes. Note that this is different to the situations shown in \autoref{fig:WGC}. Here the WGC would be satisfied by an infinite tower of superextremal macroscopic objects.
Another interpretation of our findings is that the statement of the WGC has to be modified for $\alpha < 2 \sqrt{2/3}$. In this regime the `lightest' macroscopic object with given charge $n/f_{\textrm{ax}}$ is not the extremal instanton but the wormhole. Also the correspondence between extremal instantons and extremal black holes in higher dimensions is lost for $\alpha < 2 \sqrt{2/3}$. This may imply that the WGC condition is now set by the charge-to-mass ratio of the wormhole rather than that of the extremal instanton. To satisfy the WGC one would then require the existence of states with $z > z_{\textrm{wh}}$. We leave further investigations on this topic for future work.
Last, there may be further implications for gravitational instantons if the WGC for axions is true: gravitational instantons may not be `stable' in the following sense. To be specific, consider a cored instanton with action $S$ and axion charge $n$ in a theory with $\alpha > 2 \sqrt{2/3}$. This corresponds to a tunnelling process between two configurations differing by $n$ units of axion charge. Let us then assume that the WGC is true and implies the existence of instantons with charge-to-mass ratio $z > z_0$, where $z_0$ is the charge-to-mass of an extremal gravitational instanton as before. An immediate consequence is that a tunnelling process will then always be dominated by the instantons predicted by the WGC. For our example this works as follows. The instantons needed to satisfy the WGC have $z > z_0 \geq z_{\textrm{cored}}$. Let two such instantons have $n_1, S_1$ and $n_2, S_2$, such that $n_1+n_2=n$. Since $z > z_{\textrm{cored}}$ it follows that $S_1+S_2 < S$ and tunnelling via two such instantons will dominate over tunnelling via the cored instanton.\footnote{Note that this is equivalent to the statement that (sub-)extremal black holes can in principle decay if the WGC for particles holds.}
Hence, we do not expect `unstable' gravitational instantons to be relevant in the path integral computation of the instanton potential as the major contributions should arise from the instantons satisfying the WGC. We leave a more rigorous analysis of this issue for future work.
\section{Conclusions and Outlook} \label{Section:Conclusions}
It is of great interest to understand whether quantum gravity forbids periodic scalars with large field range and flat potential. The obvious way in which this can happen is via instanton-induced corrections. In detail, there are two specific options: On the one hand, quantum gravity may demand the presence of instantons with certain actions and charges, via a generalized weak gravity conjecture. This is rather indirect: One tries to show that certain things `go wrong' unless the relevant particles (or instantons) exist.
There is, however, also a more direct approach: gravity itself supplies, in a rather direct or `constructive' way the instantons which may lift the flat potential. In the present paper, we have tried to push this direct approach as far as possible, striving also for maximal model-independence.
Our results are as follows. We observe that in a pure axion-gravity system a potential for the axion is generated by Giddings-Strominger wormholes and that this potential is parametrically unsuppressed if the cutoff is at the Planck scale. Trying to be more precise about this, we encountered a surprise: If, as a model of high-cutoff quantum gravity, we take string theory at self-dual coupling and self-dual compactification radius, we are still left with a purely numerical suppression factor of $\exp(-3\pi^2)\simeq 10^{-13}$. Such a result makes it hard to hope for a strong constraint on inflation, even after further refining the analysis.
Furthermore, we continued to ask for generic 4d constraints, but assuming more concretely that the 4d theory arises from string theory with a potentially low moduli scale. First, we found that in this setting nothing too dramatic happens to gravitational instantons: One linear combination of the moduli acts as a 4d dilaton governing the axion coupling; the instantons become more diverse in that extremal and cored gravitational instantons exist in addition to wormholes; the calculation still breaks down only at the Kaluza-Klein scale, which can of course still be high.
Unfortunately, the predictions now become model dependent as the coupling strength of the 4d dilaton to the axion (an ${\cal O}(1)$ numerical factor) enters. Taking the highest value for this factor that we could obtain in the simplest models results in a less severe instanton suppression factor of $\exp(-2\pi)\simeq 10^{-3}$. This is of course highly relevant for inflation, but easily avoided by considering models with different dilaton coupling.
In both of the above approaches, the suppression factors start out small and further fall as $\exp(-r^2)$, with $r$ an appropriately normalised compactification radius in 4d Planck units. As a result, while we do believe that gravitational instantons are the most fundamental and model-independent way to constrain field ranges, the numbers appear to allow for enough room for realistic large-field inflation.
Finally, we have attempted to connect our analysis of the various types of gravitational instantons, including their dependence on the axion-dilaton coupling, to the ongoing discussion of the weak gravity conjecture. In particular, we found a intriguing regime where wormholes are the objects with highest charge-to-mass ratio and may thus be sufficient to satisfy the instanton-axion weak gravity conjecture.
\bigskip
There are many directions for further investigations. By limiting our analysis to gravitational instantons in 4-dimensional Einstein-axion-dilaton theories we were unable to arrive at strong constraints on inflation. While this approach allows us to remain ignorant about the detailed UV completion, we are forced to neglect potentially more important contributions. These would arise from gravitational instantons with low instanton numbers, which are incalculable in the 4-dimensional Einstein-axion-dilaton theory. However, a quantitative analysis may be possible if one assumes that UV physics is described by string theory. It is expected that gravitational instantons will correspond to non-perturbative effects such as D-brane instantons in string theory. To arrive at stronger constraints a better understanding of non-perturbative effects in string theory is desirable. In particular, it is expected that poorly understood non-BPS instantons may become important during inflation \cite{150303886}.
There is a related question that is worthy of further examination. While more important instanton contributions to the axion potential may exist, it is possible that the overall effect on the axion potential vanishes once all such contributions are included. To calculate contributions from `more important' instantons is equivalent to studying instantons in a theory at a higher energy scale. However, taking string theory as our UV completion, we would expect the theory to become supersymmetric and/or higher-dimensional at some scale. It is then possible that, once we work above the supersymmetry scale, there are cancellations between the various instanton contributions to the axion potential. This is somewhat analogous to the cancellation between boson and fermion loops in supersymmetric field theory. We regard it as important to determine whether such cancellations can occur.
While we were unable to arrive at strong model-independent constraints on inflation, gravitational instantons may be important for inflation in models where the dilaton coupling takes sufficiently large values. In the effective 4-dimensional theory the dilaton coupling is a free parameter. However, one would expect that its value is constrained by possible UV completions. Indeed, by considering simple axion-dilaton systems in string compactifications, we find that the dilaton coupling typically takes $\mathcal{O}(1)$ values, i.e.~it can neither be very small nor very large. It would be interesting to examine to what extent these results are generic.
The upshot of these points is clear: It is imperative to understand the ultraviolet end of the instanton spectrum.
\subsubsection*{Acknowledgements}
We thank T.~Bachlechner, A.~Collinucci, A.~Font, B.~Heidenreich, K.-M.~Lee, E.~Kiritsis, E.~Palti, M.~Reece, F.~Rompineve, T.~Rudelius, M.~Salmhofer, J.~Stout, I.~Valenzuela and T.~Weigand for helpful discussions. We are also grateful to the referee for valuable comments on the manuscript. This work was partly supported by the DFG Transregional Collaborative Research Centre TRR 33 ``The Dark Universe''. P.M.~acknowledges financial support by the Studienstiftung des deutschen Volkes. S.T.~would like to thank the Institute for Theoretical Physics in Heidelberg for hospitality during several visits and the Klaus-Tschira foundation for generous support. He also thanks A.~Kashani-Poor and R.~Minasian for hospitality at IHP in Paris.
\begin{appendices}
\section{Derivation of the Metric Structure of Gravitational Instantons} \label{Appendix: metric structure}
We present a derivation of the metric \eqref{Metric} following \cite{07052768}, which also shows that $C$ arises as an integration constant. The most general $4$d line element with rotational symmetry is
\begin{equation} \label{Metric ansatz general}
ds^2 = \lambda(r)dr^2 + r^2 d\Omega_3^2,
\end{equation}
where $d\Omega_3^2$ represents the metric on $S^3$. Let us be more generic than in \eqref{Action with axion} and consider a set of moduli $\phi^I$ on moduli space with metric $G_{IJ}$:
\begin{equation}
S = \int d^4x \sqrt{g} \left[- \frac{1}{2} R + \frac{1}{2}G_{IJ}(\phi) g^{\mu\nu} \partial_{\mu} \phi^I \partial_{\nu}\phi^J\right].
\end{equation}
Due to the rotational symmetry of our system we take $\phi^I=\phi^I(r)$. Variation of $S$ with respect to $\phi^K$ yields the equation of motion
\begin{equation}
\left(\sqrt{g}g^{rr}G_{KJ}\phi^{\prime J}\right)^{\prime} - \frac{1}{2}\sqrt{g}g^{rr} \partial_K G_{JL} \phi^{\prime J} \phi^{\prime L} = 0.
\end{equation}
Here, the prime denotes the derivative with respect to the coordinate $r$ and $\partial_K$ the derivative with respect to $\phi^K$. Let us introduce a new variable $\tau$ such that $dr/d \tau = \sqrt{g}g^{rr}$. The equation of motion above can then be rewritten as the geodesic equation on moduli space, i.e.
\begin{equation}
\partial_{\tau}^2 \phi^I + \Gamma^{I}_{JL} \partial_{\tau}\phi^J\partial_{\tau}\phi^L = 0,
\end{equation}
with Christoffel-symbols $\Gamma^{I}_{JL}$ for the metric $G_{IJ}$. Along the geodesics we then have
\begin{equation}
\partial_{\tau} \left(G_{IJ} \partial_{\tau}\phi^I \partial_{\tau}\phi^J\right)=0
\end{equation}
or, expressed in the coordinate $r$,
\begin{equation} \label{Constant along geodesics}
G_{IJ} \phi^{\prime I}\phi^{\prime J} = \frac{k}{(\sqrt{g}g^{rr})^2} = \frac{k \lambda(r)}{r^6},
\end{equation}
where we introduced a constant $k$ and used \eqref{Metric ansatz general}.
Furthermore, the $rr$-component of the energy-momentum tensor is
\begin{equation}
T_{rr} = \frac{1}{2} G_{IJ} \phi^{\prime I} \phi^{\prime J} \stackrel{\eqref{Constant along geodesics}}{=} \frac{k \lambda(r)}{2r^6}.
\end{equation}
The algebraic form of $\lambda(r)$ can now be read off from the $rr$-component of Einstein's equations. The $rr$-component of the Einstein tensor is
\begin{equation}
G_{rr} = \frac{3}{r^2} (1-\lambda(r))
\end{equation}
and hence $G_{rr}=T_{rr}$ yields
\begin{equation} \label{Lambda}
\lambda(r) = \left(1+ \frac{C}{r^4}\right)^{-1},
\end{equation}
where $C= k/6$ is indeed an integration constant. It is interesting to note that the metric component $g_{rr}$ is determined independently of the functional form of $G_{IJ}(\phi)$.
Also note that the metric is asymptotically flat, because $\lambda(r) \to 1$ as $r \to \infty$.
Finally, we want to remark that for the creation of Euclidean wormholes ($C<0$) it is necessary to have $G_{IJ} \phi^{\prime I}\phi^{\prime J} <0$ (see \eqref{Constant along geodesics}). While one cannot simply put a wrong sign into the kinetic term of the scalar fields, one can instead consider a Lagrangian with a 2-form gauge field, whose dual field is an axion. According to our discussions of quantum mechanical dualisation in \autoref{Subsection: Dualisation}, this axion is imaginary at the saddle-point of the path integral and effectively obtains an opposite sign in the kinetic term. Moreover, for solutions with $C \geq 0$ one necessarily needs to include dynamical scalar fields so that $G_{IJ} \phi^{\prime I}\phi^{\prime J} >0$.
\section{Charge Quantisation} \label{Appendix: Charge Quantisation}
Let us first recall how flux and charge quantisation usually work in a $B_2$-/$\theta_0$-theory with strings and fundamental instantons. For any 3-cycle $S^3$ we have
\begin{equation} \label{Dirac quantisation H}
Q_B \int_{S^3} H_3 = 2 \pi n
\end{equation}
with integer $n$.\footnote{This follows from assuming gauge invariance of the coupling term in \eqref{Action B}, i.e.~one can define $B_2$ with either the south- or north pole of $S^3$ removed, getting the same result in both cases. This is another argument to see the necessity of the $i$-factor in front of the coupling terms.}
Analogously, for any 1-cycle $S^1$, we have
\begin{equation}
Q_{\theta} \int_{S^1} F_1 = 2 \pi m, \label{Dirac quantisation F}
\end{equation}
$m \in \mathbbm{Z}$. Obviously, $n$ and $m$ can only be non-zero if the relevant cycle is either non-trivial in $M$ or if it encloses the appropriate charged object.
The above are just the familiar flux quantisation conditions. In order to derive charge quantisation, we temporarily go back to Minkowskian space and use the equations of motion of
\begin{equation}
S= -\int_M \frac{1}{2g_B^2} H_3 \wedge \star H_3 + Q_B \int_M B_2 \wedge j_2,
\end{equation}
where $j_2$ is the current modelling the distribution of strings. It can be defined explicitly by $\int_{\Sigma}j_2 = N$, where $N$ is the number of strings intersecting some surface $\Sigma$. Without loss of generality we choose $N=1$. From the equation of motion for $B_2$,
\begin{equation}
d (1/g_B^2 \star H_3) = -Q_Bj_2,
\end{equation}
we find, using Stokes theorem:
\begin{equation}
Q_B= -\int_{\partial \Sigma} 1/g_B^2\star H_3 = - \int_{\partial \Sigma} F_1 = \frac{2\pi m}{Q_{\theta}}.
\end{equation}
In the last step we used $F_1$-flux quantisation. Thus, we see that
\begin{equation} \label{Dirac quantisation - Full}
Q_B Q_{\theta} = 2\pi m,
\end{equation}
which is the well-known Dirac quantisation condition. For the following, we take the freedom to choose $Q_{\theta}=1$, i.e.~the periodicity of the axion field is in this case $\theta_0 \to \theta_0 + 2\pi$.
Then, combining \eqref{Dirac quantisation - Full} with $m=1$ (here we assume that a string with smallest charge exists) and \eqref{Dirac quantisation H}, we find that the quantisation condition on $H_3$ can simply be expressed as
\begin{equation} \label{Charge quantisation condition - Appendix}
\int_{S^3}H_3 = n.
\end{equation}
This flux quantisation condition \eqref{Charge quantisation condition - Appendix} is at the heart of gravitational instanton solutions.
Now, we are actually interested in potentials introduced by gravitational instantons, i.e., in shift symmetry breaking by quantum gravity. Hence, assuming the existence of fundamental instantons defeats the purpose. So let us see how far we get with the logic above if we abandon the source term in (\ref{Action Theta}).
First, if we allow for geometries with non-trivial 3-cycles, the $H_3$ flux quantisation condition (\ref{Dirac quantisation H}) can still be derived. All we need is the existence of strings coupled to $B_2$. This then also implies that $Q_B$ is quantised. By contrast, (\ref{Dirac quantisation F}) cannot be derived without assuming the existence of fundamental instantons.
However, if we allow for geometries which also have non-trivial 1-cycles (see \autoref{Figure: Wormhole and 1-Cycle}), and if we postulate that the dual potential $\theta_0$ is
a globally defined function taking values on $S^1$ (i.e.~$\theta_0
\equiv \theta_0+2\pi$)), then both (\ref{Dirac quantisation F}) and charge quantisation, (\ref{Dirac quantisation - Full}) and (\ref{Charge quantisation condition - Appendix}), follow.
\section{Dualisation under the Path Integral} \label{Appendix: Dualisation}
In \autoref{Subsection: Dualisation} we are interested in computing
\begin{equation}
\bra{H_3^{(F)}}e^{-HT}\ket{H_3^{(I)}} \sim \int_{\text{b.c.}} d[H_3]d[\theta_0] \exp \left\{- \int_M \frac{1}{2g_{B}^2} \left(H_3 \wedge \star H_3 + 2 ig_B^2 \theta_0 dH_3 \right) \right\},
\end{equation}
which is \eqref{Partition function with Lagrange multiplier}. Here, $T \equiv t_F-t_I$. At the end we want to obtain a path integral over the variable $\theta_0$, i.e. \eqref{Partition function theta}. This is nothing but dualising from a set of canonical momentum variables to their generalised coordinates.
Thus, we illustrate the subtleties of the computation leading to \eqref{Partition function theta} by considering the quantum-mechanical harmonic oscillator, i.e.~$H=q^2/2+p^2/2$. The momentum $p$ then corresponds to the background flux $\braket{H_3}$ or, more precisely, to the quantised charge $n$, while the position variable $q$ corresponds to $\theta_0$. The transition amplitude from state $\ket{q_I}$ to $\ket{q_F}$ reads
\begin{equation} \label{Transition amplitude q}
\bra{q_F} e^{-HT} \ket{q_I} = \int d[p] \int_{\text{b.c.}} d[q] \exp\left\{\int_{t_I}^{t_F} dt \left(ip \dot{q} - H(q,p)\right)\right\},
\end{equation}
with boundary conditions $q(t_I) = q_I$ and $q(t_F)=q_F$ imposed.
In fact we rather want to compute $\bra{p_F}e^{-HT}\ket{p_I}$, which is expressed similarly:
\begin{align} \label{Transition amplitued p}
\bra{p_F}e^{-HT}\ket{p_I} &= \int d[q] \int_{\text{b.c.}} d[p] \exp\left\{\int_{t_I}^{t_F} dt \left(-iq \dot{p} - H(q,p)\right)\right\} \\ \nonumber
&= \int d[q] \int_{\text{b.c.}} d[p] \exp\left\{\int_{t_I}^{t_F} dt \left(ip \dot{q} - H(q,p)\right)\right\} \exp \left\{-i(q_Fp_F - q_Ip_I)\right\},
\end{align}
where we impose again $p(t_I) = p_I$ and $p(t_F)=p_F$. In the second step we integrated the first term of the exponential by parts. In our case we have $H=q^2/2+p^2/2$ which allows us to complete the square. Integrating out $p$ without worrying about the boundary conditions to be imposed yields the desired result
\begin{equation} \label{Path integral over q}
\bra{p_F}e^{-HT}\ket{p_I} \sim \int d[q] \exp \left(-i(q_Fp_F - q_Ip_I)\right)e^{-S[q]},
\end{equation}
see also \cite{Feynman:1951gn} for comments on the integration over the momentum.
We wish to have a closer look at this decisive step.
To do so, we write the amplitude $\bra{p_F}e^{-HT}\ket{p_I}$ as:
\begin{equation} \label{Transition amplitute Fourier transform}
\bra{p_F}e^{-HT}\ket{p_I} = \int dq_I dq_F \braket{p_F|q_F}\bra{q_F} e^{-HT} \ket{q_I} \braket{q_I|p_I}.
\end{equation}
Now let us assume that the two dual relations \eqref{Transition amplitude q} and \eqref{Transition amplitued p} hold. Then, in particular, \eqref{Transition amplitude q} implies
\begin{equation}
\bra{q_F} e^{-HT} \ket{q_I} = \int_{\text{b.c.}} d[q]e^{-S[q]},
\end{equation}
and the result \eqref{Path integral over q} follows immediately (use $\braket{p|q}=e^{-ipq}$). The operation of integrating out $p$ while disregarding its boundary conditions is thereby indirectly justified.
Finally, we can demonstrate this directly and explicitly by writing\footnote{We are grateful to K.-M.~Lee for pointing out this possibility and for further discussions on this issue. See also \cite{Coleman:1989zu}.}
\begin{align} \nonumber
& \bra{p_F}e^{-HT}\ket{p_I} \\
\nonumber =& \int \prod_{m=0}^{N}dq_m \prod_{n=0}^{N-1} dp_n \braket{p_F|q_N}\braket{q_N|e^{-H \epsilon}|p_{N-1}}\braket{p_{N-1}|q_{N-1}}\braket{q_{N-1}|e^{-H \epsilon}|p_{N-2}} \ldots \\
&\hphantom{\int \prod_{m=0}^{N}dq_m \prod_{n=0}^{N-1} dp_n} \ldots \braket{p_1|q_1}\braket{q_1|e^{-H \epsilon}|p_0}\braket{p_0|q_0}\braket{q_0|p_I},
\end{align}
where $\epsilon \equiv T/(N+1)$ and $q_0=q_I$, $q_N=q_F$.
This becomes the discretised version of \eqref{Transition amplitued p}:
\begin{align} \nonumber
\bra{p_F}e^{-HT}\ket{p_I} = \int \prod_{m=0}^{N}dq_m \prod_{n=0}^{N-1} dp_n \, & e^{-iq_N(p_F-p_{N-1})-H(q_N,p_{N-1})\epsilon} \ldots \\
& \ldots e^{-iq_1(p_1-p_0)-H(q_1,p_0)\epsilon}e^{-iq_0(p_0-p_I)}.
\end{align}
For the harmonic oscillator (and in fact for more general potentials $V(q)$) we can integrate out $p_0,...,p_{N-1}$ (after completing the square for each $p_m$). As a result we find
\begin{align} \nonumber
\bra{p_F}e^{-HT}\ket{p_I} \sim \int \prod_{m=0}^{N}dq_m \, &\exp \left\{-iq_Np_F\right\} \exp \left\{-\frac{q_N^2}{2}\epsilon - \frac{(q_N-q_{N-1})^2}{2 \epsilon}\right\} \ldots \\
&\ldots \exp \left\{-\frac{q_1^2}{2}\epsilon - \frac{(q_1-q_{0})^2}{2 \epsilon}\right\} \exp \left\{iq_0p_I\right\}.
\end{align}
This is precisely the discretised version of \eqref{Path integral over q}. Hence, integrating out the momenta from \eqref{Transition amplitued p} to \eqref{Path integral over q} without considering the boundaries is indeed justified.
\section{Analytical Solutions to Einstein's Equation} \label{Appendix: analytical solution to Einstein}
Einstein's equation \eqref{Einstein equation without potential} which follows from the action \eqref{Action without potential} can be solved analytically. We explain how to arrive at solutions \eqref{Solution for C<0, with constant}, \eqref{Solution for C=0} and \eqref{Solution for C>0} for $C<0$, $C=0$ and $C>0$, respectively.
\bigskip
First of all, \eqref{Einstein equation without potential} can be rewritten as:
\begin{equation}
\pm \int d \varphi \frac{1}{\sqrt{ \mathcal{F}(\varphi)n^2/A^2 + C }} = \sqrt{6}\int dr \frac{1}{r^3 \sqrt{1+C/r^4}}.
\end{equation}
Thus, integral representations can in principle be obtained for any $\mathcal{F}$. For $\mathcal{F}(\varphi)=1/(3!f_{\text{ax}}^2) \exp(-\alpha \varphi)$, explicit solutions exist.
For the RHS, one finds
\begin{equation}
\sqrt{6}\int dr \frac{1}{r^3 \sqrt{1+C/r^4}} = \begin{cases}
- \sqrt{\frac{3}{2|C|}} \arcsin \left(\frac{\sqrt{|C|}}{r^2}\right) + \text{const} & \text{for} ~ C<0 \\
- \sqrt{\frac{3}{2C}} \text{arcsinh} \left(\frac{\sqrt{C}}{r^2}\right) + \text{const} & \text{for} ~ C>0 \\
- \sqrt{\frac{3}{2}} \frac{1}{r^2} + \text{const} & \text{for}~ C=0,
\end{cases}
\end{equation}
where we use the substitution $y= \sqrt{|C|}/r^2$ for $C \neq 0$.
The integral on the LHS can be rewritten as
\begin{equation}
\pm \frac{1}{\sqrt{|C|}} \int d \varphi \frac{1}{\sqrt{k\exp(-\alpha \varphi) \pm 1 }},
\end{equation}
with $k \equiv n^2/(3!|C| A^2 f_{\text{ax}}^2)$. If $C>0$ ($C<0$), the positive (negative) sign under the square root applies. In the case of $C>0$ we substitute
\begin{equation}
\sinh y = \frac{1}{\sqrt{k}} \exp(\alpha \varphi /2),
\end{equation}
and for $C<0$ we take
\begin{equation}
\sin y = \frac{1}{\sqrt{k}} \exp \left(\alpha \varphi/2\right).
\end{equation}
Using appropriate identities for the hyperbolic or trigonometric functions, one arrives at
\begin{align}
\pm \int d \varphi & \frac{1}{\sqrt{ \exp(-\alpha \varphi)n^2/(3!f_{\text{ax}}^2A^2) + C }} = \\ \nonumber
&=
\begin{cases}
\pm \frac{2}{\sqrt{|C|} \alpha} \left[\arcsin \left(\frac{1}{\sqrt{k}}\exp\left(\alpha \varphi/2\right)\right) - \text{const} \right] & \text{for} ~ C<0 \\
\pm \frac{2}{\sqrt{C} \alpha} \left[\text{arcsinh} \left(\frac{1}{\sqrt{k}}\exp(\alpha \varphi/2)\right) - \text{const}\right] & \text{for} ~ C>0\\
\pm \frac{2\sqrt{6}Af_{\text{ax}}}{n \alpha} \exp \left( \alpha \varphi /2\right) + \text{const} & \text{for} ~ C=0.
\end{cases}
\end{align}
From here one can read off the solutions, which can be rewritten as \eqref{Solution for C<0, with constant}, \eqref{Solution for C=0} or \eqref{Solution for C>0}.
\section{Computation of the Instanton Action} \label{Appendix:Instanton Action}
We present further details of the computation of the instanton action in \autoref{Section:Cored Grav Inst}. The computation consists of determining the on-shell contribution from the action and the contribution coming from the Gibbons-Hawking-York boundary term. We begin by looking at the latter, where we follow \cite{150906374}.
\subsubsection*{Gibbons-Hawking-York boundary term:}
The Gibbons-Hawking-York boundary term is
\begin{equation}
S_{\text{GHY}} = - \oint_{\partial M} d^3x \sqrt{h}(K-K_0),
\end{equation}
as described around \eqref{Gibbons-Hawking-York}. Starting from our metric ansatz \eqref{Metric} we choose hypersurfaces of constant $r$. The normal unit vector $n$ is then
\begin{equation}
n = \sqrt{1+ \frac{C}{r^4}} \frac{\partial}{\partial r}.
\end{equation}
The trace of the extrinsic curvature is
\begin{equation}
K = \nabla_{\mu}n^{\mu} = \partial_{\mu}n^{\mu} + \Gamma^{\mu}_{\mu\nu}n^{\nu}
\end{equation}
where $\nabla$ is the Levi-Civita connection on $M$. One finds
\begin{equation}
\Gamma^{\mu}_{\mu r} = \frac{2C}{r^5} \left(1+ \frac{C}{r^4}\right)^{-1} + \frac{3}{r},
\end{equation}
and therefore
\begin{equation}
K=\nabla_{\mu}n^{\mu} = \frac{3}{r}\left(1+ \frac{C}{r^4}\right)^{1/2}.
\end{equation}
By taking $C=0$ we can also read off the trace of the extrinsic curvature of $\partial M$ embedded in $\mathbb{R}^4$:
\begin{equation}
K_0 = \frac{3}{r}.
\end{equation}
It then follows
\begin{equation}
S_{\text{GHY}} = - \oint_{\partial M} \epsilon_{S^3} (K-K_0) = - 3Ar^2 \left.\left[\left(1+ \frac{C}{r^4}\right)^{1/2}-1\right]\right|_{\text{boundary}},
\end{equation}
with surface area $A=2\pi^2$ of $S^3$. Recall that according to our conventions the volume form on $S^3$ contains a factor $r^3$. Clearly, for $C=0$ we have $S_{\text{GHY}}=0$. For $C>0$ the boundary is at $r=0$ and at $r=\infty$,
\begin{equation}
S_{\text{GHY}} = 3AC^{1/2}, ~~~~~~~~ C>0.
\end{equation}
In the case of $C<0$ the integral vanishes, because we always consider instanton-anti-instanton pairs, so $S_{\text{GHY}} =0$.
These are the results used in \autoref{Section:Cored Grav Inst}.
\subsubsection*{On-shell contribution:}
We now evaluate the bulk action \eqref{Action without potential} on-shell, i.e.~we plug in the equations of motion successively. As described in \autoref{Section:Cored Grav Inst}, the first step is to express the Ricci scalar $R$ by the trace of the energy-momentum tensor using Einstein's equations:
\begin{equation}
R=-T.
\end{equation}
The energy-momentum tensor $T_{\mu\nu}$ from the action \eqref{Action without potential} is
\begin{equation}
T_{\mu\nu} = - g_{\mu\nu} \left[\frac{1}{2}\mathcal{F}(\varphi)H^2 + \frac{1}{2} \partial_{\rho}\varphi \partial^{\rho}\varphi \right] + 3 \mathcal{F}(\varphi)H_{\mu\rho\sigma}H_{\nu}^{~\rho\sigma} + \partial_{\mu}\varphi \partial_{\nu}\varphi.
\end{equation}
Consequently,
\begin{equation}
T = g^{\mu\nu}T_{\mu\nu} = \mathcal{F}(\varphi)H^2 - (\partial \varphi)^2,
\end{equation}
and then \eqref{Action without potential} becomes simply
\begin{equation}
S = \int d^4x \sqrt{g} \mathcal{F}(\varphi)H^2 = A \int dr \frac{r^3}{\sqrt{1+C/r^4}} \mathcal{F}(\varphi)H^2,
\end{equation}
where we used the rotational symmetry of our system. Next, we plug in the solution to \eqref{EoM for B},
\begin{equation}
H=\frac{n}{Ar^3} \epsilon,
\end{equation}
and restrict ourselves to $\mathcal{F}(\varphi)=\exp(-\alpha \varphi)/(3!f_{\text{ax}}^2)$, for which we know the analytical solutions:
\begin{equation}
S = \frac{n^2}{Af_{\text{ax}}^2} \int dr \frac{1}{r^3\sqrt{1+C/r^4}} \exp(-\alpha \varphi).
\end{equation}
It is then convenient to rewrite the action as an integral over $d\varphi$ using Einstein's equation \eqref{Einstein equation without potential}. We consider only regular solutions. They are monotonically decreasing and therefore we have $\varphi^{\prime}(r)<0$ everywhere. Hence,
\begin{equation}
S = - \frac{n^2}{Af_{\text{ax}}^2} \int d \varphi \frac{\exp(-\alpha \varphi)}{\sqrt{n^2 \exp(-\alpha \varphi)/(A^2f_{\text{ax}}^2)+6C}}.
\end{equation}
The integral has to be evaluated case by case.
For extremal gravitational instantons with $C=0$ we have
\begin{equation}
S = - \frac{n}{f_{\text{ax}}} \int_{\varphi(0)}^{\varphi(\infty)} d \varphi \exp(-\alpha \varphi/2) = \frac{2 n}{\alpha f_{\text{ax}}}.
\end{equation}
In the case of $C>0$ we obtain
\begin{align} \nonumber
S &= - \frac{n^2}{Af_{\text{ax}}^2} \int_{\varphi(0)}^{\varphi(\infty)} d \varphi \frac{\exp(-\alpha \varphi)}{\sqrt{n^2 \exp(-\alpha \varphi)/(A^2f_{\text{ax}}^2)+6C}} \\
&= \frac{2n}{\alpha f_{\text{ax}}} \left.\sqrt{\exp(-\alpha \varphi)+ \sinh^2 K_+} \right|_{\varphi(0)}^{\varphi(\infty)} = \frac{2n}{\alpha f_{\text{ax}}} e^{-K_+},
\end{align}
where we used \eqref{C for C>0} and took $K_+>0$. Combining this with the GHY boundary term yields the desired instanton action \eqref{Instanton action for C>0}.
Finally, for Euclidean wormholes, i.e.~for $C<0$, we have
\begin{equation}
S = \int d^4x \sqrt{g} \mathcal{F}(\varphi)H^2 = 2 \times A \int_{r_0}^{\infty} dr \frac{r^3}{\sqrt{1-|C|/r^4}} \mathcal{F}(\varphi)H^2,
\end{equation}
where the factor of two occurs because the left integral is over the whole Euclidean space, and hence accounts for the whole wormhole and thus for the instanton and anti-instanton, while the integral on the RHS integrates from the centre of the wormhole to one end. The appearance of this factor may be seen more easily by evaluating the integral on the LHS using the $t$-coordinate \eqref{Metric with coordinate t} and then changing coordinates from $t$ to $r$. As was noted in \cite{150307853}, this contribution has to be divided by two, because the instanton action $S_{\text{inst}}$ should only take into account half of the full wormhole action. Consequently, using the equations of motion as in the previous cases,
\begin{align} \nonumber
S_{\text{inst}} &= - \frac{n^2}{Af_{\text{ax}}^2} \int_{\varphi(r_0)}^{\varphi(\infty)} d \varphi \frac{\exp(-\alpha \varphi)}{\sqrt{n^2 \exp(-\alpha \varphi)/(A^2f_{\text{ax}}^2)-6|C|}} \\
&= \frac{2n}{\alpha f_{\text{ax}}} \left| \sin \left(\frac{\alpha \pi}{4}\sqrt{\frac{3}{2}}\right) \right|.
\end{align}
Hence, \eqref{Instanton action C<0} follows, where we can drop the modulus due to the restriction to $0 \leq \alpha < 2 \sqrt{2/3}$.
\section{Estimating the Size of the Prefactor $\mathcal{A}$ in the Instanton Potential} \label{Appendix: Prefactor}
The contribution of gravitational instantons to the axion potential is given by $\delta V = \mathcal{A}e^{-S}\cos (n\theta)$. While it has been proposed e.g.~in \cite{150303886,150300795} that $\mathcal{A} \sim 1$ (in Planck units), we attempt a somewhat more precise estimate. This is inspired by the analogies between gravitational instantons and instantons arising from Euclidean branes wrapping an internal cycle of the compactification manifold (see e.g.~\cite{150300795,150303886,150906374}). Let us start by recalling how the latter contributes to the supergravity $F$-term potential in a simple setup.
We consider a Euclidean brane instanton modifying the perturbative superpotential $W_0$ as
\begin{equation}
W = W_0 + A(z) e^{-a T},\label{ww0}
\end{equation}
where $z$ denotes the complex structure moduli and $T$ is a K\"ahler modulus.
Then the supergravity $F$-term potential
\begin{equation}
V_F = e^K \left(K^{i \bar{\jmath}} D_iW D_{\bar{\jmath}} \overline{W} - 3 |W|^2\right)
\end{equation}
is corrected at leading order by
\begin{equation}
\delta V \sim e^K W_0 A(z) e^{-a \tau} \ ,
\end{equation}
where $\tau$ is the real part of $T$. Recall that $K = - 2 \ln \mathcal{V} + ... $, which gives a suppression by $1/\mathcal{V}^2$. Furthermore, we rewrite the above expression in terms of the gravitino mass $m_{3/2} \sim W_0/ \mathcal{V}$ and the KK-scale $m_{\text{KK}} \sim 1/\mathcal{V}^{2/3}$:
\begin{equation}
\delta V \sim \frac{1}{\mathcal{V}^{5/3}} \frac{m_{3/2}}{m_{\text{KK}}} A(z) e^{-a \tau}.
\end{equation}
If we were allowed to compare this with \eqref{Delta V of gravitational instantons} then, using $m_{3/2} \lesssim m_{\text{KK}}$, we would conclude that
\begin{equation}
\mathcal{A} \lesssim \frac{A(z)}{\mathcal{V}^{5/3}}\label{calae}
\end{equation}
in Planck units. Here we identified $\exp(-a\tau)$ with $\exp(-S)$ motivated by the obvious analogy: Indeed, the Euclidean brane action is proportional to the brane tension and the volume of the cycle. Similarly, the action of a cored gravitational instanton is proportional to the ADM tension of a black brane wrapping a cycle in a higher-dimensional version of the gravitational instanton system, see e.g.~\cite{150906374} for an example.
Nevertheless, our proposal to estimate ${\cal A}$ by \eqref{calae} remains nontrivial. Indeed, we first need to consider a large wrapping number $n$ to relate to the calculable regime on the gravitational side. This is unproblematic in the present case since these higher instantons will contribute to $W$ analogously to \eqref{ww0}. Next, we are \textit{not} interested in Euclidean brane instantons (their effect is well-known) but in some possibly very different type of instanton arising in a string model and not having a simple microscopic description. The claim or proposal implicit in \eqref{calae} is then that this instanton may, conservatively, also be suppressed by a factor ${\cal A}$ which becomes small as the KK-scale and SUSY breaking scales go down. This appears to be reasonable since, beyond the simple Euclidean brane case discussed here, higher-dimensional and SUSY-based cancellations are expected to occur above those scales.
Accepting the above proposal, compactification volumes in the range $\mathcal{V} \sim 10^2$ to $10^3$ imply $\mathcal{A} \sim 10^{-4}$ and $10^{-5}$, respectively, assuming that $A(z) = \mathcal{O}(1)$. Note that in order to avoid destabilisation of the K\"ahler moduli the compactification volume is at most of order $\mathcal{O}(10^3)$, see e.g.~\cite{14043542,14043711,14112032}. Nevertheless, the suppression by $e^{-S}$ remains dominant in all regimes we considered.
\end{appendices}
\newpage
\begingroup
\raggedright
\sloppy
\bibliographystyle{apsrev4-1}
\nocite{apsrev41control}
|
\section{}
\section{Introduction}
Synthetic genetic networks provide a potential tool to design useful biological functions targeted to perform specific tasks.\cite{hasty2001,elowitz2010,benenson2012} The early stages of research in this direction were to envision and understand simple networks which provided the basic components for building more complex functional devices. Emphasis was first given to the design of a genetic toggle switch\cite{gardner2000} and an oscillator known as the Repressilator consisting of a 3-gene inhibitory ring that has been expressed in \textit{E. coli}.\cite{elowitz2000} Later, electronic circuits were suggested and used to study the dynamics of synthetic genetic networks.\cite{mason2004,wagemakers2006,buldu2007,tokuda2010} Electronic circuits, in general, allow precise control of system parameters and provide a minimal set-up for experimenting with a dynamical behavior in the presence of intrinsic and extrinsic noises. This option is useful, to predict various desired functional behaviors in electronic analogs of synthetic genetic networks which are difficult to control in real biological experiments.
We have designed electronic circuits, in the past, to model genetic networks configured to investigate dynamical behaviors of the Repressilator\cite{hellen2011, hellen2013} and to perform noise-aided logic operations.\cite{hellen2013b} In the Repressilator studies, we first considered an isolated Repressilator and verified the functional form of the predicted oscillations.\cite{hellen2011} Then we incorporated a bacterial-inspired method of quorum sensing (QS) coupling\cite{Garcia-Ojalvo2004} into our Repressilator circuit by adding a feedback chain to the 3-gene inhibitory ring. This additional pathway led to a rich variety of dynamical behavior, including multistability, for the QS-modified isolated Repressilator.\cite{hellen2013} Simulations of this single Repressilator system have even demonstrated period doubling chaotization.\cite{potapov2012} The next step of allowing the QS mechanism to couple Repressilators together as has been done in simulation\cite{Garcia-Ojalvo2004,ullner2007,ullner2008} proved difficult using our previous circuit models. This difficulty leads us to make improvements of the circuit including a complete redesign of the QS circuitry, which we present here in detail. The improved design allowed us to investigate the more complex dynamics that exist for coupled Repressilators [in prep] and to access the full QS-parameter range of the mathematical model. Apart from their potential use in synthetic biological devices, coupled Repressilators are of interest because they belong to the field of coupled nonlinear oscillators which is essential for the understanding of a wide variety of biological phenomena.\cite{strogatz1993}
It is crucial to have a precise control of the initial conditions when studying a multistable system like the QS-coupled Repressilators so that all of the coexisting attractors for a given set of parameters can be captured. We describe the use of an analog switch to set the initial conditions by initializing capacitor voltages to the desired values. Multistability also opens the possibility of noise-induced transitions from one attractor to another. Therefore we use our previous noise circuit\cite{hellen2013b} and the genetic network circuit as a test-bed to demonstrate noise-induced transitions between attractors within the QS-coupled Repressilator system.
We begin with the mathematical model and the analog circuit for the genetic network of Repressilators coupled via QS. Then we present our circuit analysis to relate the circuit with the mathematical model, use the QS circuit to verify the numerical predictions, and show results for coupled Repressilator circuits. Finally, we describe how to set the initial conditions and incorporate additive noise in the electronic circuit.
\begin{figure}
\includegraphics[width=0.5\textwidth]{Rep_full2.eps}
\caption{\label{fig:rep-qs}Repressilator with quorum sensing feedback. mRNA (a,b,c) and their expressed proteins (A,B,C) form the 3-gene inhibitory loop referred to as the Repressilator. Quorum sensing is provided by the additional feedback loop of the small auto-inducer molecule which can diffuse through the cell membrane thereby exchanging with the external medium. }
\end{figure}
\section{Model: Repressilator with quorum sensing}
We present here the mathematical model and the circuit model for the genetic network of our interest. The following sections show our analysis which connects the circuits to the equations.
Figure \ref{fig:rep-qs} shows a Repressilator with a QS feedback loop. The mRNA (a,b,c) and their expressed proteins (A,B,C) form the 3-gene inhibitory loop referred to as the Repressilator.\cite{elowitz2000} It is named Repressilator because each gene's output``represses" the next gene's expression, resulting in stable oscillations of protein concentrations over a very broad interval of parameter values. Thus the 3-gene ring network works as a genetic oscillator. The QS feedback loop uses a small auto-inducer (AI) molecule to provide an indirect activation path from $B$ to $C$ to compete with the direct inhibition.\cite{ullner2008} This network structure generally leads to an anti-phase synchronization of two coupled Repressilators, meaning there is a $180^o$ phase difference between the protein oscillations of the two Repressilators. A different network structure placing the feedback loop from $A$ to $C$ has also been employed,\cite{Garcia-Ojalvo2004} which generally leads to in-phase synchrony. Interestingly, the network structure does not fully determine the type of synchronization observed between coupled Repressilators as both of these structures are birhythmic--capable of both types of synchrony--depending on the model's parameter values.\cite{potapov2011} This birhythmic property may be of use in the design of task-oriented devices.
We use our reduced mathematical model for QS-coupled Repressilators\cite{hellen2013} which is based on previous models\cite{elowitz2000,ullner2008} and applies to the case of fast mRNA kinetics compared to protein kinetics. The model uses standard chemical kinetics $(\beta,\alpha,\kappa,n,k_i)$ including Hill function inhibition, $1/(1+x^n)$, and is
\begin{subequations}
\begin{align}
\frac{dA}{dt}&=\beta_1\left(-A+\frac{\alpha}{1+C^n}\right)\\
\frac{dB}{dt}&=\beta_2\left(-B+\frac{\alpha}{1+A^n}\right)\\
\frac{dC}{dt}&=\beta_3\left(-C+\frac{\alpha}{1+B^n}+\frac{\kappa S}{1+S}\right)
\label{ode-c}\\
\frac{dS}{dt}&=-k_{s0}S+k_{s1}B-\eta\left(S-S_{ext}\right).
\label{ode-s}
\end{align}
\label{rep-eqns}
\end{subequations}
$(A,B,C)$ are the protein concentrations for the Repressilator, and $S$ is the concentration of the AI molecule. The AI can diffuse (diffusion constant $\eta$) through the cell membrane into the external medium, unlike the proteins which are confined inside the cell. $S_{ext}$ is the AI concentration in the external medium and is a diluted average of the contributions from all the Repressilators, $S_{ext}=QS_{ave}$, where $Q$ is the dilution factor. For results presented here we use $k_{s0}=1$, $k_{s1}=0.01$, and $\eta=2$ as taken previously.\cite{ullner2008}
The circuit for a single inhibitory gene shown in Fig.~\ref{fig:gene-circuit} is a modification of the previous one.\cite{hellen2011} The transistor current represents the rate of gene expression and the voltage $V_i$ represents the concentration of expressed protein. $V_{i-1}$ represents the concentration of the repressor, and the $V_{cth}$ adjusts the affinity of the repressor binding to the gene's DNA. The Hill function inhibition in Eq.\ \eqref{rep-eqns} is accounted for by the dependence of the transistor current on repressor concentration voltage $V_{i-1}$. This dependence is derived in the next section.
\begin{figure}
\includegraphics[width=0.5\textwidth]{rgn_gene2.eps}
\caption{\label{fig:gene-circuit}Single-gene circuit. Inhibitory input at
$V_{i-1}$. Expressed protein concentration is represented by
$V_{i}$. Dual op-amp is LF412 supplied by $\pm5$ V. The
\textit{pnp} transistor is 2N3906. Resistor $R_{hill}$ is adjusted to achieve desired Hill-function $n$ value. Capacitor value $C_i=0.1\mu f$ is for $\beta=1$. }
\end{figure}
The circuit for a Repressilator with quorum sensing feedback shown in Fig.~\ref{fig:rep-circuit} is a complete redesign of that presented previously.\cite{hellen2013} The Repressilator consists of the closed 3-gene loop with op-amp buffers between the genes. The QS circuitry takes input from current source I(B) controlled by the repressilator's $B$-protein voltage, and feeds back to the Repressilator's $C$-protein via source I(S). The feedback activation in the mathematical model is through the binding-site occupation term $S/(1+S)$. We show below that the circuit accounts for the activation via QS by using a piece-wise continuous linear behavior, modeled by $\min(0.8S,1)$ and hence we replace Eq.\ \eqref{ode-c} by
\begin{equation}
\frac{dC}{dt}=\beta_3\left(-C+\frac{\alpha}{1+B^n}+\kappa \: \min(0.8S,1)
\right)
\label{new-C-eqn}
\end{equation}
In Fig.~\ref{fig:rep-circuit} $S_1$ is the AI concentration belonging to the shown Repressilator. Coupling this Repressilator to a second Repressilator (not shown) is accomplished by adding their respective AI concentrations, $S_1$ and $S_2$, thus creating $S_{ext}$, the concentration of AI in the external medium. Figure \ref{fig:rep-circuit} shows the connection of $S_2$ to the op-amp at the bottom of the figure and the combination with $S_1$ to produce $S_{ext}$.
\begin{figure}
\includegraphics[width=0.475\textwidth]{erep_qs_circuit.eps}
\caption{\label{fig:rep-circuit}Circuit for Repressilator with QS feedback. The Repressilator consists of the closed ring of genes A, B, and C. The quorum sensing loop is from $B$ through $S_1$ to $C$. Protein $B$ creates auto-inducer $S_1$ via the voltage-controlled current source $I(B)$, and $S_1$ activates production of $C$ via $I(S)$. Each ``gene" triangle corresponds to the single gene circuit in Fig.~\ref{fig:gene-circuit}. $S_2$ is the contribution from a second Repressilator (not shown) and $S_{ext}$ is the auto-inducer concentration in the external medium. The op-amps have low offset voltage (below 0.5 mV).}
\end{figure}
\subsection{Single Gene Circuit with Hill-function}
We now analyse the circuit for a single gene and show how the inhibitory Hill function behavior is reproduced. In the process we find useful results: how to connect model parameters $n$ and $\alpha$ to circuit parameters, and the minimum accessible value of $n$.
Applying current conservation to the capacitor voltage in Fig.~\ref{fig:gene-circuit}, and normalizing by a scaling parameter $V_{th}$ gives,
\begin{equation}
R_CC_0\frac{dx_i}{dt}=\frac{C_0}{C_i}\left(-x_i +\frac{I_tR_C}{V_{th}}\right)
\label{gene-circ-eqn}
\end{equation}
where $x_i=V_i/V_{th}$ is the dimensionless protein concentration and $I_t$ is the transistor's current collector. $R_CC_0$ is the time-scale and it normalizes the time variable, thereby making $t$ dimensionless. A comparison with Eq.\ \eqref{rep-eqns} gives a useful relation between the model parameters and the circuit values,
\begin{equation}
\beta_i=\frac{C_0}{C_i},\: \alpha=\frac{I_{max}R_C}{V_{th}}
\label{beta-alpha}
\end{equation}
where $I_{max}$ is the maximum transistor current and its relation to $I_t$ is defined below clearly to derive the Hill function behavior in the circuit.
The gene inhibition in Eq.\ \eqref{rep-eqns} is controlled by the Hill function
\begin{equation}
H(x)=\frac{1}{1+x^n}
\label{Hill}
\end{equation}
where $x$ is the dimensionless inhibitory protein concentration. The scaling parameter $V_{th}$ accounts for the inhibitor's equilibrium binding constant. Comparing Eqs.\ \eqref{rep-eqns} and \eqref{gene-circ-eqn} shows that the Hill function behavior must be accounted for in the circuit by the transistor current's dependence on input voltage $V_{i-1}$. In this section we derive this current-voltage dependence. The key elements are to get the correct slope at $x=1$ where $H(x=1)=0.5$ and to approximate the Hill function's positive curvature decay to zero.
The op-amp U2 in Fig.~\ref{fig:gene-circuit} has different gains, $G_{-2}$ when $V_{i-1} < V_{cth}$, and $G_{+2}$ when $V_{i-1} > V_{cth}$. For the selected component values in the circuit, the subtraction op-amp U1 has a gain $G_1 = -6.8$, and inverting op-amp U2 has $G_{-2} = -22$ and $G_{+2}$ is an amplitude-dependent diminishing gain due to the three diodes in the feedback for U2. The diodes create the positive curvature decay of the Hill function.
The gene inhibition in the circuit corresponds to $V_{i-1}$ surpassing $V_{cth}$, which causes the output of U2 to go positive and thereby turns off the \textit{pnp} transistor resulting in no current from the collector. The maximum output voltage of U2 is about 2.0 V when the three diodes are fully conducting in their forward biased state. The resistors $R_{b1}$ and $R_{b2}$ are chosen such that an output voltage at U2 of 2.0 V causes a drop of $(0.42/2.62)(5 - 2) = 0.48$ V across $R_{b1}$ which is small enough so that the transistor current is essentially zero. Maximal protein expression in the circuit corresponds to $V_{i-1} = 0$ which results in U2 output going negative with a limit at the lower saturation level $V_{-sat} = -3.5$ V for the dual op-amp LF412 supplied with $\pm 5$ V. We assume that the gain $G_1G_{-2}$ is large enough so that the output of U2 reaches $V_{-sat}$ when $V_{i-1} = 0$. Later we determine a practical restriction on Hill coefficient $n$ imposed by this assumption.
We predict the transistor's collector current in Fig.~\ref{fig:gene-circuit} when the output of U2 varies between -3.5 and 2.0 V. The collector current is essentially the current in $R_E$ since the transistor is in the active region. The voltage across $R_{b1}$ is $f(5-G\Delta V)$ where the fraction $f = 0.42/2.62 = 0.160$ is the voltage divider gain, $\Delta V = (V_{i-1}-V_{cth})$, and $G$ is the overall gain of the 2 op-amps. The current in $R_E$, and therefore the transistor current, is
\begin{equation}
I_t=\frac{f(5-G\Delta V)-V_{eb}}{R_E}
\end{equation}
where $V_{eb}$ is the emitter-base voltage. $V_{eb}$ varies from about 0.5 V when there is essentially zero transistor current ($G\Delta V\approx 2$ V) to a maximum of about $V_{ebmx}=0.70$ V at maximum current ($G\Delta V=V_{-sat}$). Maximal protein expression occurs for $V_{i-1} = 0$ (no inhibition) and thus $G\Delta V = V_{-sat}$ giving the maximum transistor current
\begin{equation}
I_{max}=\frac{f(5-V_{-sat})-V_{ebmx}}{R_E}.
\label{imax}
\end{equation}
For our chosen circuit components we measure $I_{max}=2.95$ mA and $V_{ebmx}=0.70$ V. This agrees well with the prediction using the large-signal transistor model with saturation current $I_S=7$ fA (which we measured for the 2N3906 transistors), $V_{ebmx}=V_T\ln(I/I_S)=0.026\ln(2.95\textrm{ mA}/7\textrm{ fA})=0.696$ V. The resulting voltage drop across $R_C$ is easily measured by setting $V_{i-1}=0$, and agrees with that predicted by Eq.\ \eqref{imax} flowing into $R_C=1$ k$\Omega$, $I_{max}R_C= 2.97$ V.
In the circuit, the Hill function Eq.\ \eqref{Hill} corresponds to the normalized transistor current
\begin{equation}
\frac{I_t}{I_{max}}=\frac{f(5-G\Delta V)-V_{eb}}{f(5-V_{-sat})-V_{ebmx}}.
\label{it-norm}
\end{equation}
As presented previously,\cite{hellen2011} the circuit approximation of the Hill function is accomplished by setting the slope of the normalized current equal to the slope of the Hill function $dH/dx$ at $x = 1$. Setting the slopes of Eqs.\ \eqref{Hill} and \eqref{it-norm} equal, using $\Delta V = V_{th}(x_{i-1} - V_{cth}/V_{th})$ with $x_{i-1} = 1$, gain $G = G_1G_{-2}$, Eq.\ \eqref{imax},
and $V_{th} = I_{max}R_C/\alpha$ provides a useful result connecting important model parameters $n$ and $\alpha$ to circuit parameters.
\begin{equation}
n\alpha=\frac{4fR_CG_1G_{-2}}{R_E}
\label{n-alpha}
\end{equation}
Using our circuit values $f=0.160$, $R_C=1$ $k\Omega$, and $R_E = 222\Omega$, we determine
$n\alpha=2.88G_1G_{-2}$. Equation \eqref{n-alpha} allows desired model parameters $n$ and $\alpha$ to be achieved in the circuit by adjusting gains $G_1$ and $G_{-2}$.
\begin{figure}
\includegraphics[width=0.475\textwidth]{eHill_n3pt2a218bpt5Rh4cp.eps}
\caption{\label{fig:Hill}Hill inhibition approximation for the single gene circuit in Fig.~\ref{fig:gene-circuit}. Numerical Hill inhibition (solid red) and experimentally measured (blue) normalized transistor current. $n=3.2, \alpha=218, R_{hill}=4 k\Omega$. Data was collected with capacitor $C_i$ removed.}
\end{figure}
Next we find the relationship between the binding constant scaling voltage $V_{th}$ and the circuit value $V_{cth}$. At $x = 1$ the Hill function has a value of 0.5. The corresponding condition for the circuit is that the normalized transistor current be 0.5 when $V_{i-1} = V_{th}$. By setting Eq.\ \eqref{it-norm} equal to 0.5, letting $\Delta V = (V_{th} - V_{cth})$ and solving gives
\begin{equation}
V_{cth}=V_{th}+\frac{\left(2V_{eb}-V_{ebmx}-f\left(5+V_{-sat}\right)\right)}{2fG_1G_{-2}}.
\label{Vcth}
\end{equation}
$V_{eb}$ at half the maximal current is predicted by using 1.5 mA for the transistor current resulting in $V_{eb}=V_T\ln(I/I_S)=0.026\ln(1.5\textrm{ mA}/7\textrm{ fA})=0.678$ V. For the circuit in Fig.~\ref{fig:gene-circuit}, $G_1G_{-2} = (-6.8)×(-22)$, $f=0.160$, $V_{-sat}=-3.5$ V, and using $V_{eb}=0.68$ V and $V_{ebmx}=0.70$ V gives $V_{cth} = V_{th} + 8.8$ mV.
Figure \ref{fig:Hill} shows the measured approximation of the Hill inhibition for the single gene circuit of Fig.~\ref{fig:gene-circuit} for $n=3.2$, $\alpha=218$, and $R_{hill}=4k\Omega$. The blue dots are the normalized output voltage $V_i/V_{th}$ as a function of normalized input voltage $V_{i-1}/V_{th}$. It is apparent that as the input voltage surpasses $V_{th}$ (at $x=1$) the transistor current shuts off, closely following the numerically plotted Hill function (solid red line). The location (at $x=1$) and slope of the drop are set by Eqs.\ \eqref{n-alpha} and \eqref{Vcth}, but the positive curvature decay to zero is controlled by $R_{hill}$ in Fig.~\ref{fig:gene-circuit}. The value of $R_{hill}$ is varied to match the transistor current's decay to that of the Hill function. Our previous circuit model for a single gene\cite{hellen2011} used a piecewise-linear approximation to the Hill function and therefore did not include a positive curvature decay to zero.
The assumption that the output of op-amp U2 is saturated at $V_{-sat}$ when $V_{i-1} = 0$ (no inhibition) means that $G_1G_{-2}V_{cth} > -V_{-sat}$. Using the relations between $V_{cth}$ and $V_{th}$ (Eq.\ \eqref{Vcth}), between $V_{th}$ and $\alpha$ (Eq.\ \eqref{beta-alpha}), and between $G_1G_{-2}$ and $n\alpha$ (Eq.\ \eqref{n-alpha}), we find the restriction on the Hill coefficient
\begin{equation}
n>\frac{2\left(f(5-V_{-sat})-2V_{be}+V_{bemx}\right)}{f(5-V_{-sat})-V_{bemx}}.
\end{equation}
For our circuit values this gives a minimum Hill coefficient of $n=2.12$. This restriction is generally not a problem since the Repressilator in Eq.\ \eqref{rep-eqns} for $\kappa=0$ has a stable fixed point and therefore is not an oscillator for $n<2$ over a wide range of $\alpha$ and identical $\beta$.
\begin{figure}
\includegraphics[width=0.5\textwidth]{rgn_n3pt2a218bpt5_R4k_420.eps}
\caption{\label{fig:rep-time}Repressilator time series. Numerical (dashed) and circuit measurements (colored) for Repressillator with no quorum sensing ($\kappa = 0$ in Eq.\ \eqref{rep-eqns}). $n=3.2, \alpha=218, R_{hill}=4 k\Omega$. }
\end{figure}
The Repressilator consisting of the 3-gene ring in Fig.~\ref{fig:rep-qs} is modeled by connecting three single-gene circuits in a closed loop depicted by the 3 gene-triangles (A,B,C) in Fig.~\ref{fig:rep-circuit}. Figure \ref{fig:rep-time} shows the measured time series and simulations (dashed lines) for a Repressilator demonstrating the stable protein oscillations ($A,B,C$) with different amplitudes that occur for different protein time-scales $\beta_1=0.5$, $\beta_2=0.1$, and $\beta_3=0.1$.
\begin{figure}
\includegraphics[width=0.3\textwidth]{B-current-circuit.eps}
\caption{\label{fig:B-current}Voltage controlled current source I(B). Protein $B$ voltage controls the current source to the $S$-voltage. $I(B)=V_B/2.2k\Omega$. Op-amp has low offset voltage (below 0.5 mV).}
\end{figure}
\subsection{Circuit for Repressilator with Quorum Sensing}
Figure \ref{fig:rep-circuit} shows the circuit for a Repressilator with QS feedback. The circuit is a modification of the earlier version.\cite{hellen2013} The feedback from $B$ through the current source $I(B)$ to $S_1$, then through $I(S)$ to $C$ corresponds to the AI feedback loop between $B$ and $c$ in Fig.~\ref{fig:rep-qs}. We analyse the circuit to derive relations between the mathematical model and circuit values. Figures \ref{fig:B-current} and \ref{fig:S-current} show the circuits for the voltage dependent current sources $I(B)$ and $I(S)$ used in Fig.~\ref{fig:rep-circuit}.
The circuit equation corresponding to Eq.\ \eqref{ode-s} comes from circuit analysis for the voltage $V_S$ across the capacitor $C_S$ in Fig.~\ref{fig:rep-circuit}
\begin{equation}
R_{S1}C_S\frac{dV_{S1}}{dt}=-V_{S1}+R_{S1}I(B)-\frac{R_{S1}}{R_d}\left(V_{S1}-V_{ext}\right).
\end{equation}
$V_{S1}$ and $V_{ext}$ correspond to the scaled voltages $S_1$ and $S_{ext}$ in Fig.~\ref{fig:rep-circuit}. Multiplying both sides by $k_{S0}$, setting $k_{S0}R_{S1}C_S$ to be the same as the time-scale $R_CC_0$ defined for the single-gene circuit, using $I(B)=V_B/2.2k\Omega$ from Fig.~\ref{fig:B-current}, $V_B=V_{th}B$, and dividing by a scaling factor $V_{sth}$ gives
\begin{equation}
\frac{dS_1}{dt}=-k_{S0}S_1+k_{S0}\frac{R_{S1}}{2.2k}\frac{V_{th}}{V_{sth}}B-k_{S0}\frac{R_{S1}}{R_d}\left(S_1-S_{ext}\right)
\end{equation}
where $V_S=V_{sth}S$ and $V_{ext}=V_{sth}S_{ext}$. Comparison with Eq.\ \eqref{ode-s} gives relations for the activation rate $k_{S1}$ of auto-inducer and the membrane diffusion parameter $\eta$.
\begin{equation}
k_{S1}=k_{S0}\frac{R_{S1}V_{th}}{2.2kV_{sth}},\:\eta=k_{S0}\frac{R_{S1}}{R_d}
\label{Vsth}
\end{equation}
Equation \eqref{Vsth} sets the scaling factor $V_{sth}$.
\begin{figure}
\includegraphics[width=0.475\textwidth]{S-current-circuit.eps}
\caption{\label{fig:S-current}Voltage controlled current source I(S). Auto-inducer $S$-voltage controls the current source feeding back to protein $C$-voltage. For small $S$ the diodes are not conducting and $I(S)=G_SV_S/R_k$. For large $S$ the diodes are forward biased and the voltage $V'$ remains close to 2 V causing the output $I(S)$ to level off. Op-amps have low offset voltage (below 0.5 mV).}
\end{figure}
The equation for the protein $C$ voltage is found in the same way as Eq.\ \eqref{gene-circ-eqn} with the addition of the current $I(S)$ from the feedback loop in Fig.~\ref{fig:rep-circuit}.
\begin{equation}
R_CC_0\frac{dC}{dt}=\frac{C_0}{C_3}\left(-C+\frac{I_tR_C}{V_{th}}+\frac{I(S)R_C}{V_{th}}\right)
\end{equation}
Comparison with Eq.\ \eqref{new-C-eqn} shows that
\begin{equation}
\frac{I(S)R_C}{V_{th}}=\kappa \:\min(0.8S,1)
\label{kappa-act}
\end{equation}
Equation \eqref{kappa-act} imposes two constraints. First, the maximum value of $I(S)$ must correspond to the right-hand-side maximum $\kappa$ occurring for $S\geq 1.25$, giving
\begin{equation}
I(S\geq 1.25)\equiv I_{Smax}=\frac{\kappa V_{th}}{R_C}.
\label{ISmax-kap}
\end{equation}
The maximum current is implemented by adjusting the gain $G_S$ in Fig.~\ref{fig:S-current} so that the $S=1.25$ input voltage $V_S=1.25V_{sth}$ creates a current of 1 mA in the series diodes causing $V'=3\times 0.63=1.9$ V. The required op-amp output is $G_S(1.25)V_{sth}\approx 1.9+(1mA)(300\Omega)=2.2$ V which provides the appropriate value for $G_S$. Secondly, for currents below the maximum value, Eq.\ \eqref{kappa-act}'s slopes must be the same. From Fig.~\ref{fig:S-current} the current source is $I(S)=G_SV_S/R_\kappa$. For currents below $I_{Smax}$ we use Eq.\ \eqref{kappa-act} and the relation for $V_{sth}$ in Eq.\ \eqref{Vsth} to find the relation between model parameter $\kappa$ and circuit value $R_\kappa$,
\begin{equation}
R\kappa =k_{s0}\frac{G_SR_CR_S}{0.8(2.2k\Omega )k_{s1}\kappa}.
\label{kapa}
\end{equation}
All the values on the right-hand-side except $\kappa$ have been previously determined, therefore Eq.\ \eqref{kapa} provides a direct link between parameter $\kappa$ and circuit value $R_\kappa$. For the values used here the result is $R_\kappa=(56.8G_S)/\kappa$ in $k\Omega$.
\begin{figure}
\includegraphics[width=0.475\textwidth]{S-function_ers.eps}
\caption{\label{fig:S-function}Measured normalized AI activated current(green) and models; piece-wise-linear $\text{min}(0.8S,1)$ (red) and hyperbola $S/(1+S)$ (blue). For $\kappa=21, \alpha=135$.}
\end{figure}
Figure \ref{fig:S-function} shows the measured normalized current $I(S)/I_{Smax}$ from the circuit in Fig.~\ref{fig:S-current}, the piece-wise-linear model $\min(0.8S,1)$ (used in Eq.\ \ref{new-C-eqn}), and the hyperbola $S/(1+S)$ (Eq.\ \ref{ode-c}) for $\kappa=21.3$, $V_{th}=0.0219$ V, and gain $G_s=1.76$. The piece-wise-linear function intersects the hyperbola at $S=0$ and $0.25$.
We now consider the circuit which creates the AI concentration in the external medium $S_{ext}$. Each Repressilator circuit contributes its intracellular AI concentration $S_i$ to the external concentration $S_{ext}$. Figure \ref{fig:rep-circuit} shows how two Repressilators are coupled by concentrations $S_1$ and $S_2$ combining to produce
\begin{equation}
S_{ext}=\frac{2R_Q}{4.7k}S_{ave}=QS_{ave}
\end{equation}
where $S_{ave}= (S_1+S_2)/2$. $Q=2R_Q/4.7k\Omega$ is a dilution factor which in a biological setting ranges from 0 to 1. For purposes of exploring dynamics in the full parameter range of the mathematical system, we use a $5k\Omega$ potentiometer for $R_Q$ so that we can vary $Q$ from 0 to 2. Our previous circuit design\cite{hellen2013} limited $Q$ variation from 0 to 1.
\subsection{Selection of Circuit Values}
Here we summarize the practical results for choosing circuit values in Figs.\ \ref{fig:gene-circuit} and \ref{fig:rep-circuit}. The model parameters are $n$, $\alpha$, $\beta 's$, $\kappa$, $k_{s1}$, and $\eta$. Some circuit values are chosen independent of the model parameters. We choose $R_C=R_S=1k\Omega$ and $C_0=C_S=0.1 \mu f$ for characteristic time $0.10$ ms, $R_E=222\Omega$, $V_{-sat}=-3.5$ V (for the LF412 op-amp powered by $\pm 5$ V), and the voltage divider fraction ($R_{b1}$ and $R_{b2}$) in Fig.~\ref{fig:gene-circuit} as $f=420/2620=0.160$. Resulting measured quantities for the transistor are $I_{max}=2.95$ mA at $V_{ebmx}=0.70$ V, and 1.5 mA at $V_{eb}=0.68$ V. These currents were shown to be consistent with predictions using the standard transistor model $I(V_{eb})=I_S\exp (V_{eb}/V_T)$.
For Fig.~\ref{fig:gene-circuit}, Eq.\ \eqref{beta-alpha} gives $V_{th}$ and $C_i$, Eq.\ \eqref{n-alpha} gives overall gain $G_1G_{-2}$, and Eq.\ \eqref{Vcth} gives $V_{cth}$. For Fig.~\ref{fig:rep-circuit}, Eq.\ \eqref{kapa} gives $R_\kappa$, op-amp gain $G_S=0.8\times2.2V/V_{sth}$, where $V_{sth}$ is given by Eq.\ \eqref{Vsth}. The only circuit value not determined by the model parameters is $R_{hill}$ in Fig.~\ref{fig:gene-circuit}. It is convenient to incorporate trim-pots into $R_{hill}$ to adjust the Hill function's positive curvature decay to zero.
For many choices of parameters the AI concentration $S$ stays below 1, in which case the $S$ activation term $\min(0.8S,1) \rightarrow 0.8S$ meaning there is no need to amplify $V_S$ to impose saturation of $I(S)$. Thus, the current source $I(S)$ in Fig.~\ref{fig:S-current} can be simplified by leaving out the non-inverting op-amp at the input and the 3 diodes, so that $V_S$ connects directly to the $300+3k=3.3k\Omega$. In this case $G_S=1$ in Eq.\ \eqref{kapa}.
\subsection{Setting Initial Conditions}
The ability to set initial conditions is crucial when studying systems with multistability so that all attractors can be captured. We use the 4066 quad analog switch to impose initial conditions by momentarily connecting ``protein" capacitor voltages to desired initial values set by trim-pot voltage dividers with op-amp followers. The 4066 is gated by the output of a 555 timer controlled by a push-button momentary switch (circuit not shown). Improved performance of the 4066 switch is achieved by powering it with 0 and +15 V, compared to the synthetic genetic network circuits powered by $\pm 5$ V.
\subsection{Other Design Considerations}
The inexpensive 2N3906 \textit{pnp} transistors used in the gene circuits were selected from a large batch to have nearly the same saturation current, $I_S=7\pm1$ fA, by performing in-house measurements.
For the case of coupled Repressilator circuits, care was taken to distribute the $\pm 5$ V power rails and ground paths symmetrically to both Repressilators. The measured voltage difference during operation between respective rails and respective grounds of the two Repressilators was less than 1 mV.
\section{Measurments: Quorum Sensing Circuit}
We now present experimental results incorporating the new QS circuitry. We begin with a single Repressilator with QS feedback, followed by two coupled Repressilators.
The case of a single Repressilator with QS feedback corresponds to setting $S_{ext}=QS$ in Eq.\ \eqref{ode-s}. Measured results from the QS circuit are compared to predictions from numerical simulations using the XPPAUT software.\cite{ermentrout} The desired goal is that the circuit and the simulations have the same structure of dynamical behaviors. A convenient way to do this dynamical comparison is to compare their $Q$-continuation bifurcation diagrams shown in Fig.~\ref{fig:Q-bif-sim}. These diagrams show the possible amplitudes of $B$ for different $Q$-values. Steady-state (SS) is either stable (red) or unstable (black), and the limit cycle (LC) oscillations are stable (green). The $B$-values for the circuit were obtained by normalizing the measured voltage amplitudes by $V_{th}=15.5$ mV which corresponds to the parameter values used in the simulation; $n=3.0$, $\alpha=190$, $\beta_i=0.5,0.1,0.1$, and $\kappa=10$. $R_{hill}=2.7k\Omega$.
\begin{figure}
\includegraphics[width=0.4\textwidth]{sgl_n3a190bpt5k10_no_S2.eps}
\includegraphics[width=0.4\textwidth]{sgl_data_n3a190bpt5k10Rh2pt7.eps}
\caption{\label{fig:Q-bif-sim}Numerical (top) and measured (bottom) $Q$-continuation bifurcation diagrams showing amplitude of protein $B$ for a single Repressilator with quorum sensing. Stable (red) and unstable (black) steady-state. Limit cycle (green). }
\end{figure}
In both simulation and circuit measurements Fig.~\ref{fig:Q-bif-sim} shows that increasing $Q$ causes the LC to decrease in amplitude until reaching the low-$B$-SS, and there is coexistence of high-$B$-SS and LC over a broad range of $Q$-values from approximately 0.6 to 1.3. Both bifurcation diagrams predict that decreasing $Q$ will cause a transition to LC for a system starting from the high-$B$-SS. Figure \ref{fig:SS_to_LC} shows an oscilloscope screenshot of this $Q$-induced high-$B$-SS to LC transition when $Q$ was slowly decreased by adjusting the trim-pot in Fig.~\ref{fig:rep-circuit}. The transition occurred at a value of $1.5k\Omega$ corresponding to $Q=2\times 1.5/4.7=0.64$ agreeing well with the left-side endpoint of the high-$B$-SS in the bifurcation diagrams.
\begin{figure}
\includegraphics[width=0.475\textwidth]{SS_to_LC_n3a190bpt5k10q1pt3Rh2pt7cp.eps}
\caption{\label{fig:SS_to_LC}Screenshot (oscilloscope) of high-$B$-SS to LC transition caused by decreasing $Q$ at 0.64 for a single Repressilator circuit with QS feedback. Protein $B$ voltage shown. Parameters: $n=3$, $\alpha=190$, $\kappa=10$, $R_{hill}=2.7$ k$\Omega$.}
\end{figure}
The agreement between the circuit and simulation results is not exact in Fig.~\ref{fig:Q-bif-sim}, however, the qualitative structure and relative location of dynamical behaviors are the same. For the circuit the low-$B$-SS was stable over a $Q$-range narrower than the resolution of $Q$-values and therefore appears as a single data point at the end of the LC-branch. We note that the simulations are able to find the unstable SS (black lines in Fig.~\ref{fig:Q-bif-sim}), whereas the circuit, of course, can only find stable dynamics. We conclude that the quorum sensing circuit achieves the goal of having the same dynamical behavior as the mathematical model.
\begin{figure}
\includegraphics[width=0.4\textwidth]{AP_n4a143bpt5k4pt8q0pt6Rh9_ccp.eps}
\includegraphics[width=0.4\textwidth]{IP_n4a143bpt5k4pt8q0pt65Rh9_ccp.eps}
\caption{\label{fig:AP}Screenshots of anti-phase (left) and in-phase (right) oscillations for two QS-coupled Repressilator circuits. The protein $B$ voltages from each Repressilator circuit are shown. Both screenshots use the same circuit values, thus demonstrating the coexistence of AP and IP states.}
\end{figure}
The motivation for the circuit improvements presented here is to extend our previous investigations to coupled Repressilators. Figure \ref{fig:AP} shows examples of the oscillations for two coupled Repressilators; more extensive investigation results are in preparation. The screen-shots show the $B$-protein voltages of the two Repressilator circuits. The coupling scheme in Eqs.\ \eqref{rep-eqns} produces a multistable system whose stable oscillations are predominantly anti-phase (AP)\cite{ullner2008} like those in the top screen-shot of Fig.~\ref{fig:AP}. Interestingly, under appropriate parameter values it is possible to find stable in-phase oscillations (IP) like those in the bottom screen-shot of Fig.~\ref{fig:AP}, which coexist with AP. Both screen-shots use the same parameters ($n=4$, $\alpha=143$, $\beta_i=0.5,0.1,0.1$, $\kappa=4.8$, $R_{hill}=9$ k$\Omega$) and both the AP and IP can be accessed simply by smoothly varying the coupling strength $Q$. AP is the sole stable state at small $Q$ and as $Q$ is increased the amplitude of the AP decreases until the AP becomes unstable and transitions to a stable steady-state characterized by both $B$-proteins being at the high value. When $Q$ is then decreased there is a transition to IP at the endpoint of the stable steady-state (similar to the decreasing Q induced transition for the single Repressilator in Fig.~\ref{fig:SS_to_LC}).
\section{Incorporation of Additive Noise}
Additive noise may be included using a simple noise circuit shown in Fig.~\ref{fig:noise_circuit} based on the breakdown of a reverse biased base-emitter junction as described previously.\cite{hellen2013b} Noise is added to a protein by disconnecting its $R_C$ from ground in Fig.~\ref{fig:gene-circuit} and connecting it to the noise circuit output as shown in Fig.~\ref{fig:noise_circuit}. The potentiometer at the second op-amp adjusts the noise amplitude. Using the same procedure used to find Eq.\ \eqref{gene-circ-eqn}, the equation for the gene's protein voltage $V_i$ is easily found to be
\begin{equation}
(1k\Omega)C_i\frac{dV_i}{dt}=V_i-V_{noise}+I_t(1k\Omega).
\end{equation}
The noise is symmetric about zero and therefore the minus sign is irrelevant, thus accomplishing the task of adding noise to the protein voltage $V_i$.
\begin{figure}
\includegraphics[width=0.5\textwidth]{noise_circuitcp.eps}
\caption{\label{fig:noise_circuit}Noise circuit. 2N3904 \textit{npn} on left has no connection at its collector. 2N3906 \textit{pnp} on right is from the gene circuit. Op-amps are OPA228 powered by $\pm 12$ V. }
\end{figure}
Comparison of the noise-influenced dynamical results from circuit measurements and numerical predictions requires careful connection of the electronically generated noise characteristics to the simulated noise. Here we summarize those connections, which were derived previously.\cite{hellen2013b} In simulations additive noise is typically represented by $D\eta(t)$ where $\eta(t)$ is a zero-mean Gaussian noise with unit variance and the amplitude $D$ is the noise strength. The electronically generated noise is characterized by its $rms$ amplitude $V_{Nrms}$ and its frequency bandwidth $f_c$. The relation between the simulated noise strength $D$ and the measured strength $V_{Nrms}$ is\cite{hellen2013b}
\begin{equation}
D=\frac{V_{Nrms}}{V_{th}\sqrt{\gamma (RC)f_c}}
\label{noise_strength}
\end{equation}
where $RC$ is the characteristic time of the Repressilator, and $\pi/4 \leq \gamma \leq \pi/2$ depending on the gain of the second amplifier in Fig.~\ref{fig:noise_circuit}. The noise bandwidth is determined by the op-amp's gain-bandwidth product (33 MHz for the OPA228) and the gain of the non-inverting amplifier in Fig.~\ref{fig:noise_circuit} (about $20\times$) resulting in a noise bandwidth of $f_c=33/20\approx 1.5$ MHz.
As a demonstration, we add independent noises to each $B$-protein for the case of coexistence of AP and IP states used for Fig.~\ref{fig:AP} ($n=4$, $\alpha=143$, $\kappa=4.8$). Multiple transitions between the states were observed. Figure \ref{fig:nz_AP_IP} shows a noise-induced transition from AP to IP. The top two traces are the added noises with $rms$-amplitudes of $V_{Nrms}=0.156$ V. Equation \eqref{noise_strength} gives the corresponding noise strength for simulation $D\approx0.6$, found using $\alpha=143$ and $I_{max}R_C=2.95$ V to give $V_{th}=20.6$ mV, characteristic time $RC=0.1$ ms, and taking $\gamma=1$.
\begin{figure}
\includegraphics[width=0.475\textwidth]{nz_AP_IP_n4a143k4pt8Rh9q0pt6nzpt16cp.eps}
\caption{\label{fig:nz_AP_IP}Screen-shot of noise-induced transition from anti-phase (AP) to in-phase (IP) oscillation for two QS-coupled Repressilator circuits. Independent noise (top two traces) was added to each $B$-protein voltage. }
\end{figure}
\section{Conclusion}
We presented a revised design for our electronic circuit model of a synthetic genetic network comprised of Repressilators coupled together by quorum sensing. Connections between mathematical parameters and circuit values were improved, in part, by including the large-signal transistor model in the derivation. The all-new quorum sensing circuitry allowed expansion of the quorum sensing circuit's accessible parameter range to match that of the mathematical model. Important features include the incorporation of Hill function binding kinetics and the ability to set initial conditions. Circuit behavior was verified by comparing bifurcation diagrams obtained from measurements and numerical simulation. The circuit revisions were important because they allow us to extend previous investigations to the case of coupled Repressilators. An example of this extension demonstrated the coexistence of IP and AP oscillatory states, and noise-induced transitions between these states. A more extensive investigation of the coupled Represilators is undertaken and to be presented in the future.
\begin{acknowledgments}
S.K.D. acknowledges support by the CSIR Emeritus Scientist scheme. The authors thank Evgeny Volkov for valuable contributions.
\end{acknowledgments}
|
\section{Introduction}\label{sec:intro}
Active galactic nuclei (AGN) emit light across the whole
electromagnetic spectrum, often dominating the emission of their host
galaxy. Due to accretion onto the central supermassive black hole
(SMBH), they can produce so-called ``jets'', highly relativistic plasma
outflows. They belong to the most fascinating objects in the Universe,
but the underlying physics is still not
fully understood. The knowledge is crucial in context of AGN
feedback and multimessenger astronomy. Multiwavelength observations
are useful tools to address open questions concerning the formation,
acceleration and the mechanism(s) behind the broadband emission up to the
highest energies.
Blazars are a subset of radio-loud AGN, where the jet is observed at a
small angle to the line of sight, such that the jet emission is strongly
Doppler boosted. They belong to the most luminous and highly variable
sources \citep{Urry1996variability}, typically showing superluminal motion in the pc-scale
radio jet \citep[e.g.,][]{Lister2013}.
With the detection of $\gamma$-ray emission of AGN jets by
\textsl{EGRET} \citep{Hartman1999}
various models were considered in order to explain the broadband
emission \citep[e.g.,][and many more]{Marscher1985,Mannheim1993,Sikora1994,Dermer1997,Dermer2012,Boettcher2013}.
A typical radio to $\gamma$-ray SED of a blazar shows a double-humped spectral shape from the
radio up to the $\gamma$-ray regime \citep{Fossati1998}. While the low-energy peak can be well
explained by synchrotron emission, it is still discussed which
emission processes are responsible for the high-energy peak. It is
contentious whether it is due to synchrotron self-Compton
up-scattering and/or inverse Compton
scattering with external photons. Furthermore, the composition of the
ejected plasma, leptons or hadrons or the combination of both, plays
an important role in modeling the broadband emission.
Single-zone leptonic models have been very successful in describing
the broadband spectrum, however they fail to explain observations revealing rapid flaring and
multiple emission zones. In that case models need to take the jet
geometry into account \citep[like spine-sheath configuration,
e.g.,][]{Tavecchio2008}. Hadronic models, on the other hand, attempt
to explain the high-energy hump due to accelerated hadrons inducing
pion-photo production resulting in a electromagnetic cascade.
The combination of simultaneous broadband data allows us to study the
spectral energy distribution (SED) and the variability across the
bands. This provides information on the different radiating
components, e.g., the disk, broad line region or the jet, all together
making up the overall spectrum. Since these sources show strong
variability across all wavelengths, simultaneity of the data is
essential, i.e., contemporaneous monitoring at different wavelengths
is required.
In addition to the broadband spectral data, we use high-resolution
radio data from Very Long Baseline Interferometry (VLBI). It is a
unique tool to address the innermost regions of extragalactic jets at
milliarcsecond (mas) scales. It provides the highest angular
resolution and insights into regions close to the jet base where the
high-energy emission is thought to be produced. VLBI images reveal the morphology
of the jets at (sub-)pc scales. Typical blazar jet morphologies
are compact or one-sided, while for
larger jet inclination angles, where relativistic beaming effects are small, the jet and the counterjet can be
detected \citep[see, e.g.,][]{Kadler2004}.
Most objects detected with the Large Area Telescope (LAT) onboard of the \textsl{Fermi} Gamma-ray Space Telescope are classified as blazars
\citep{2fgl,3fgl}. Only few of the so-called ``misaligned'' objects (radio galaxies) with
jets seen edge-on
\citep{Abdo2010_misaligned} are bright in the $\gamma$-rays. However,
these objects are of particular interest and challenge theoretical jet emission
models, which typically explain the high-energy spectral component with
high beaming factors. Their study can help to determine the
$\gamma$-ray emission region(s) and to constrain emission models,
because the broadband emission is less dominated by the beamed jet
emission \citep{Abdo2010_cenacore}.
In the radio regime, these misaligned objects can be
divided into evolved (e.g., as Centaurus~A or M\,87) and young radio
galaxies. The jets of the former have sizes up to several hundred
kiloparsecs, while the latter are typically more compact and smaller than
1\,kpc. Therefore, these sources are also called Compact
Symmetric Objects \citep[CSO,][]{ODea1998,Readhead1996a,Readhead1996b}. Because of
their intrinsic power, theoretical models predicted $\gamma$-ray emission from CSOs
\citep{Stawarz2008,Kino2007,Kino2009}, but no detection is
confirmed yet.
Here, the multiwavelength and VLBI study of extragalactic jets on
the Southern Hemisphere is presented. This work was performed in the
framework of the multiwavelength monitoring program TANAMI
(Sect.~\ref{sec:tanami}). After a short introduction to the project
and the sample results (Sect.~\ref{sec:tanamisources}), the properties
of two particular
sources are discussed, PMN\,J1603$-$4904\xspace (Sect.~\ref{sec:pmn}) and Centaurus~A (Sect.~\ref{sec:cena}).
\section{TANAMI - The multiwavelength monitoring program of
extragalactic jets in the Southern Hemisphere}\label{sec:tanami}
In 2007, the TANAMI\footnote{\textit{Tracking Active Galactic Nuclei with
Austral Milliarcsecond Interferometry}
\url{http://pulsar.sternwarte.uni-erlangen.de/tanami/}} program
started monitoring the brighest extragalactic jets in the Southern
sky (below $-30^\circ$ declination) using the combination of
high-resolution VLBI observations at 8.4\,and 22.3\,GHz \citep{Ojha2010a}
and corresponding radio
monitoring with the Australian Compact Array
\citep[ATCA,][]{Stevens2012} and the Ceduna telescope \citep{Blanchard2012a}, in the optical/UV with
\textsl{Swift}/UVOT and \textsl{Rapid
Eye Mount} \citep{Nesci2013}, in the X-rays with \textsl{Swift}/XRT and all-sky observations at $\gamma$-rays by
\textsl{Fermi}/LAT.
In addition, pointed observations for particular
sources are conducted with \textsl{XMM}-Newton, \textsl{Suzaku}
\citep[see Sect.~\ref{sec:pmn} and e.g.,][]{Mueller2015a,Kreikenbohm2016} and
\textsl{INTEGRAL} in the X-rays.
The TANAMI VLBI observations are performed with the Australian Long
Baseline Array (LBA), including additional telescopes at NASA's Deep
Space Network (DSN) located at Tidbinbilla, the South-African
Hartebeeshoeck antenna, the German Antarctic Receiving Station (GARS)
in O'Higgins (Antarctica), and the Transportable Integrated Geodetic
Observatory (TIGO) in Chile \citep{Ojha2010a}. Since 2011, the
$(u,v)$-coverage (see as an example Fig.~\ref{fig:uvplot}) at
intermediate baselines is significantly improved by the Warkworth (New
Zealand), Katherine and Yarragadee (Australia) antennas \citep{Kadler2015}. TANAMI
observations typically have an angular resolution of about a few mas,
down to less than 1 mas, with the largest baselines to the
transoceanic antennas Hartebeeshoeck, TIGO and O'Higgins.
The VLBI data are recorded on the LBADRs (Long Baseline Array Disk
Recorders) and correlated on the DiFX software correlator at Curtin
University in Perth, Western Australia \citep{Deller2007}. The
correlated data are inspected, edited and fringe fitted
in AIPS \citep[National Radio Astronomy Observatory’s
Astronomical Image Processing System software;][]{Greisen2003_AIPS}. The
amplitude calibration is performed using known flux values of prior
observed sources. Hybrid imaging and subsequent model fitting is performed
in the program DIFMAP \citep{Shepherd1997}.
For more details on the data reduction see \citet{Ojha2010a}, where we presented
the first-epoch images of the intial sample.
\begin{figure}
\includegraphics[width=\columnwidth]{uvplot.pdf}
\caption{The typical TANAMI $(u,v)$-coverage at 8\,GHz has improved due to the
addition of new
telescopes since 2011 compared to the initial array configuration
presented in \citet{Ojha2010a}. Here, the
intermediate $(u,v)$-range is covered due to baselines to Warkworth. The
long baselines are provided by TIGO and Hartebeesthoek. The
displayed $(u,v)$-coverage corresponds to the observation of
Centaurus~B (PKS\,1343--601), shown in Fig.~\ref{fig:new}.}
\label{fig:uvplot}
\end{figure}
The TANAMI sample of extragalactic jets was defined as a hybrid radio
and $\gamma$-ray selected sample. Initially, it consisted of 43
objects, while to date $\sim$90 sources are regularly monitored.
We aim to investigate the emission and formation mechanism of AGN jets.
Our observational setup provides structural and spectral information
at mas-scales, and, in addition, broadband spectral
properties with time \citep{Krauss2016}.
The long-term VLBI monitoring yields information on jet properties,
such as apparent speed, inclination, opening angles, and structural changes
with time. Furthermore, the dual-frequency approach provides spatial spectral
index distributions of individual jet features.
The simultaneous broadband observations address the jet
activity and spectral changes across all wavelengths.
Since the start of TANAMI, new flat-spectrum radio sources have been
added to the sample when associated with a $\gamma$-ray
detection by \textsl{Fermi}/LAT. For most of these sources, TANAMI
provides the first high-resolution VLBI images (M\"uller et al., in
prep.). Figure~\ref{fig:new} shows a selection of these mas-scale
CLEAN images from
8\,GHz VLBI observations.
We are
particularly interested in the so called
``radio-to-gamma-connection'', i.e., the correlation of high-energy
emission with changes in pc-scale properties seen in jets. This is
linked to the open question of the production sites and mechanisms of
high-energy photons. Furthermore, the multiwavelength analysis enables
us to study the broadband emission, to test different emission models,
related to the composition of jets \citep[e.g.,][]{Dutka2013}.
Beyond that, TANAMI has a strong multimessenger component, trying to
explain the high-energy neutrinos observed with IceCube and ANTARES
with the broadband emission from $\gamma$-ray loud TANAMI blazars. The details
can be found in \citet{Krauss2014a},
\citet{AntaresTANAMI2015}, and \citet{Kadler2016}, as well as in \citet{FritschPhDT}, \citet{FehnPhDT} and
\citet{Mueller2014PhDT}, but it is
beyond the scope of this paper.
\begin{figure*}
\includegraphics[width=\columnwidth]{0235-618_X_X_color.pdf}\hfill
\includegraphics[width=\columnwidth]{0447-439_X_L_color.pdf}\hfill
\includegraphics[width=\columnwidth]{2142-758_X_AA_color.pdf}\hfill
\includegraphics[width=\columnwidth]{1343-601_X_Z_color.pdf}
\caption{First-epoch VLBI images from newly added TANAMI sources
(M\"uller et al., in prep.). \textit{From top left to bottom right}:
Parsec-scale morphology of
\mbox{PKS\,0235--618}, one of the blazar sources consistent with one
of the PeV neutrino events detected by IceCube \citep[see][for more
details]{Krauss2014a}, the TeV-blazar PKS\,0447--439, the flat-spectrum $\gamma$-ray quasar PKS\,2142-758
\citep[see][for details on the multiwavelength
properties]{Dutka2013}, and the parsec-scale jet of
the radio galaxy
Centaurus~B (PKS\,1343--601) also detected at $\gamma$-rays
\citep{Katsuta2013}. The color scale displays the flux density
distribution in mJy/beam. Contours are scaled logarithmically,
increased by a factor of 2, with the lowest contour set to the
$3\sigma$-noise-level (negative contours are dashed). The restoring
beam is shown as gray ellipse in the lower left corner. }
\label{fig:new}
\end{figure*}
\section{Gamma-ray loudness and milliarcsecond-scale properties of
TANAMI sources}\label{sec:tanamisources}
Using one year of contemporaneous TANAMI VLBI and \textsl{Fermi}/LAT
data, we addressed the high-energy properties and
radio-$\gamma$-correlation in our sample \citep{Boeck2016}. We
analyzed the radio and $0.1-100$\,GeV $\gamma$-ray data obtained during
the first 11\,months of \textsl{Fermi}/LAT monitoring. More
than 70\% of the TANAMI sources are associated with $\gamma$-ray
emitters detected by \textsl{Fermi}/LAT. Upper limits on the
$\gamma$-ray flux were determined for the remaining sources, yielding
three new significant detections, which were later confirmed by the
\textsl{Fermi}/LAT team \citep{2fgl}.
We find increasing brightness temperatures $T_B$ of the radio cores, obtained
from VLBI measurements, with 1\,year average $\gamma$-ray
luminosity. The $\gamma$-ray undetected sources have lower
$T_B$ values and probably
$\gamma$-ray luminosities close to the determined upper limits.
Comparing the morphologies of the $\gamma$-ray detected versus
undetected TANAMI sources, we find that the $\gamma$-ray loud objects
are generally more core dominated, i.e., are more compact and have a
higher core-to-jet flux ratios (M\"uller et al., in prep.).
These results are consistent with the general picture of strong
Doppler boosting in $\gamma$-ray bright sources.
Most $\gamma$-ray sources in our
sample show compact or one-sided jet morphologies. Two exceptional objects,
PMN\,J1603$-$4904\xspace and Centaurus~A, are discussed in more
detail in the following sections. Their double-sided\footnote{Note
that, as in \citet{Ojha2010a}, we use the morphology classification scheme by
\citet{Kellermann1998}, classifying sources with the most compact
component in the middle of the emission as double-sided, without taking
spectral information into account.} morphology and
$\gamma$-ray loudness are in particular interesting and provide
important insights into the jet physics.
\section{The unusual jet source PMN\,J1603$-$4904\xspace}\label{sec:pmn}
The radio source PMN\,J1603$-$4904\xspace has been known as a calibrator source (PKS\,1600$-$48)
for Southern Hemisphere radio observations, but it attracted attention due to
its association with one of the brightest, flat-spectrum $\gamma$-ray
sources detected during the first few months of \textsl{Fermi}/LAT
monitoring \citep{0fgl,Kovalev2009_LBAS}. Since then, it is
is a bright, significantly detected $\gamma$-ray source \citep[3FGL\,1603.9--4903;][]{1fgl,2fgl,3fgl,1fhl,2fhl}.
Based on sparse optical and broadband data, the source was classified
as a low-synchrotron peaked BL\,Lac object without known redshift
\citep{Shaw2013a}.
Due to the high-confidence association of the known radio source with the
$\gamma$-ray object 1FGL\,J1603.8$-$4903 \citep{1fgl}, it was included in the
TANAMI sample. The first VLBI observations were performed in 2009 and
revealed unusual pc-scale properties \citep{Mueller2014a}. This result
triggered
multiwavelength follow-up observations further questioning its classification as a blazar
\citep{Mueller2015a,Goldoni2016}.
In the following these observations and their conclusions will
be presented.
\begin{figure}
\includegraphics[width=\columnwidth]{1600-489_X_KL_color.pdf}
\caption{Naturally weighted 8\,GHz VLBI image of PMN\,J1603$-$4904\xspace. The brightest, most compact
component is the central feature. The overall brightness
distribution is almost symmetric and constant over a monitoring
period of $\sim$4\,years. The color scale displays the flux density
distribution in mJy/beam. Contour lines scale logarithmically and
increase by a factor of 2, with the lowest contour set to the
$3\sigma$-noise-level. The restoring
beam is shown as a gray ellipse in the lower left corner. }
\label{fig:pmn}
\end{figure}
\subsection{VLBI observations}
The TANAMI 8\,GHz VLBI observations of PMN\,J1603$-$4904\xspace yielded the first
image at mas-scale resolution of its radio morphology \citep{Mueller2014a}.
Since 2009 the source has been observed about twice a year at 8\,GHz,
with one (quasi-)simultaneous observation at 22\,GHz in 2010
May. Figure~\ref{fig:pmn} shows the mas-scale brightness distribution of the
source at 8\,GHz. It is almost symmetrical with a total correlated
flux density of $\sim 600$\,mJy. It is resolved and shows an East-West
orientation (position angle of $-80^\circ$) with three distinct emission
features. The brightest, most compact component is located at the
center, with a brightness temperature of $T_B\geq 1\times 10^{10}$\,K.
Using spectral index information from combined 8\,GHz and 22\,GHz
VLBI datasets, we concluded that this central component is the most
plausible `core' of PMN\,J1603$-$4904\xspace.
TANAMI monitoring reveals no significant proper motion of the Eastern
and Western components. The source appearance as in
Fig.~\ref{fig:pmn} remains effectively constant. By modeling the source with three Gaussian
emission regions, we find relative motions of $v_\mathrm{app}< 0.2\,\mathrm{mas\,yr^{-1}}$ \citep{Mueller2014a,Mueller2014PhDT}, later
confirmed by \citet{Hekalo2015} using observations
spanning four years.
\subsection{Broadband spectrum and variability}
The ATCA monitoring between 1\,and 40\,GHz reveals only very minor flux
variability at higher frequencies. The ATCA spectral index of $\alpha=-0.4$
is consistent with the VLBI results for
the Eastern and Western features. However, ATCA measures about
$\sim$200\,mJy more flux as detected with VLBI, possibly indicating a
diffuse extended emission component
which is resolved out by the TANAMI VLBI array.
Continuous monitoring by \textsl{Fermi}/LAT shows no major flaring
activity as typically observed in blazar sources \citep{2fgl,3fgl}. PMN\,J1603$-$4904\xspace has been
reported in both high-energy catalogs \citep[1FHL, 2FHL][]{1fhl,2fhl}, being a potential
candidate for detection in the TeV range by ground based
Cherenkov telescopes.
Using all available (including archival) multiwavelength data, we
constructed a non-simultaneous broadband SED (see Fig.~\ref{fig:sed}). It includes data from
TANAMI VLBI and ATCA in the radio, \textsl{2MASS} and \textsl{WISE}
in the infrared, \textsl{Swift}/UVOT, GMOS and NTT in the optical,
\textsl{Swift}/XRT, \textsl{XMM-Newton}, \textsl{Suzaku} in the X-rays and
\textsl{Fermi/LAT} in the $\gamma$-rays. We parametrize the broadband
spectrum with two logarithmic parabolas. In the
infrared, we see a strong excess, which can be modeled with a black
body spectrum.
\subsection{\textsl{XMM} and \textsl{Suzaku} results and follow-up
VLT/X-shooter observations}
In order to better constrain the X-ray spectrum than with
\textsl{Swift}/XRT-only data, we performed \textsl{XMM}-Newton and \textsl{Suzaku}
observations in 2013 \citep{Mueller2015a}. The \mbox{2--10\,keV}
spectrum (Fig.~\ref{fig:pmnspec})
was simultaneously modeled with an absorbed power-law component
($N_\mathrm{H}=2.05_{-0.12}^{0.14}\times10^{22}\mathrm{cm^{-2}}$, $\Gamma=2.07_{-0.12}^{0.04}$),
cross-calibration constants for the different detectors, and
a Gaussian emission line at $5.44\pm0.05$\,keV. We interpret this X-ray line as the most
prominent emission line in AGN X-ray spectra, the
Fe~K$\alpha$ fluorescence line with a restframe energy of
6.4\,keV. Therefore, the redshift of PMN\,J1603$-$4904\xspace can be determined to be
$z=0.18\pm0.01$. Adopting this redshift, the brightness distribution
in the VLBI image (Fig.~\ref{fig:pmn}) has an extent of $\sim$46\,pc.
Triggered by these intriguing results, \citet{Goldoni2016} observed
PMN\,J1603$-$4904\xspace with the UV-NIR VLT/X-shooter spectrograph in 2014. The optical continuum
spectrum is mostly featureless, not compatible with a stellar origin,
but likely non-thermal. Three emission lines are detected, allowing
us to determine a redshift of \mbox{$z=0.2321\pm0.0004$}. The spectral
features do not follow the definition of a BL\,Lac object. The new
redshift measurement implies that the X-ray emission line detected by
\textsl{XMM}-Newton and \textsl{Suzaku} should be interpreted as a
6.7\,keV line. This result is very peculiar, due to a missing
6.4\,keV line and is possibly indicating that
the 6.7\,keV emission is due to collisionally ionised plasma, and
merits further follow-up observations of PMN\,J1603$-$4904\xspace.
\begin{figure}
\includegraphics[width=\columnwidth]{pmn_spec_AGproc.pdf}
\caption{Simultaneous fit to the (quasi-)simultaneous X-ray data from \textsl{XMM}-Newton and
\textsl{Suzaku} of PMN\,J1603$-$4904\xspace \citep{Mueller2015a}. The data are
fitted by an absorbed power-law component and an emission line at
$\sim$5.44\,keV (highlighted by the shaded region). a) Counts spectrum for all individual detectors. b)
Ratio of data to model for the best fit the Gaussian
emission line. The shaded region marks the position of the emission
line. }
\label{fig:pmnspec}
\end{figure}
\subsection{An unusual blazar or a $\gamma$-ray loud young radio
galaxy?}
PMN\,J1603$-$4904\xspace has been classified as a BL\,Lac object, i.e., a jet source
pointing towards the observer. The emission of these objects are
dominated by relativistic beaming effects and spectral features
like emission lines are usually outshined by the non-thermal jet
emission. Furthermore, blazars are typically highly variable across
the electromagnetic spectrum and VLBI measurements reveal compact,
sometimes one-sided jet structures, showing high polarization and
relativistic motion.
Compared to this, the VLBI and multiwavelength data of PMN\,J1603$-$4904\xspace indicate very
unsual properties for a blazar source.
Our VLBI data show a symmetric brightness distribution on mas-scales, which is
constant over 4\,years of monitoring. The brightest component in the
center has the flattest spectrum and highest brightness temperature
values. We find no significant proper motion.
Similarly, multiwavelength monitoring
reveals no rapid flaring activity. Only long-term variability is
reported at GeV energies \citep{1fhl,3fgl} and ATCA
monitoring over more than ten years shows some minor long timescale
flux density variations at higher radio frequencies. ATCA data also
give upper limits of $<1.2$\% on the polarization \citep{Murphy2010_AT20G},
though higher polarization fractions at VLBI scales cannot be excluded
so far. The broadband SED shows an unexpected strong excess in the
infrared and a significant X-ray spectral line and high intrinsic
absorption. These are all features which are not observed in typical
blazar spectra. Finally, the recent VLT/X-shooter data suggest a non-BL\,Lac
nature.
This peculiar appearance of PMN\,J1603$-$4904\xspace is questioning its blazar
classification, hence, we consider alternative interpretations
\citep{Mueller2014a}, favouring a young radio galaxy seen at a
larger angle to the line of sight.
Its VLBI structure resembles a double-sided morphology as seen in
FR\,I radio galaxies, but on smaller scales, i.e. a young
version. These compact symmetric objects typically show low
variability and polarization \citep{Peck2000} and VLBI observations
suggest that they are seen edge-on.
Our multiwavelength data are in agreement with the young radio galaxy
scenario. Adopting the measured redshift
\citep{Mueller2015a,Goldoni2016} and assuming edge-on geometry, the
source size probed by VLBI is well below the canonical limit for young radio sources of
1\,kpc, while the ATCA data limits the total size to be less than 3\,kpc.
Further multiwavelength data will help to substantiate this
alternative classification. For example, VLBI monitoring can reveal
opposed apparent motion, as expected for a jet-counterjet
system. Low-frequency radio observations below 1\,GHz can give important information about
a potential spectral turnover, as expected for young radio sources.
If
confirmed, PMN\,J1603$-$4904\xspace will add to the class of misaligned $\gamma$-ray
bright sources. Moreover, the confirmed detection of a $\gamma$-ray
loud CSO would help to determine the location of the high-energy
emission using broadband SED models.
\begin{figure}
\includegraphics[width=\columnwidth]{1600-489_black_cyan.pdf}
\caption{Broadband $\nu F_\nu$ spectral energy distribution of PMN\,J1603$-$4904\xspace,
including VLBI total fluxes from TANAMI observations \citep{Mueller2014a}, data by
MGPS \citep{Murphy2007_MGPS}, ATCA, WISE and 2MASS, by GMOS and \textsl{Swift}/UVOT
\citep{Mueller2014a} and NTT \citep{Shaw2013a}, and 4\,years data from
\textsl{Fermi}/LAT from 3FGL \citep{3fgl}. The combined X-ray data
by \textsl{Swift} XRT \citep{Mueller2014a}, \textsl{XMM}
and \textsl{Suzaku} \citep{Mueller2015a} are shown in purple. We parametrize the data
with two logarithmic parabolas (gray lines), absorbed by photoelectric absorption
in the X-rays, and a blackbody component. The black symbols
mark all data included in the fit, data marked by cyan symbols are excluded. The gray symbols represent
the optical/IR and X-ray data corrected for extinction and
absorption, respectively. The dashed lines show the unabsorbed log
parabolas and the extinction corrected blackbody. In the lower panel
we show the residuals of the fit in units of standard deviation of
individual data points. A physical SED model is required to investigate
the broadband emission in more detail.
}
\label{fig:sed}
\end{figure}
\section{Zooming into the closest radio-loud AGN: Centaurus
A}\label{sec:cena}
At a distance of only 3.8\,Mpc \citep{Harris2010}, the radio galaxy
Centaurus~A (Cen~A) is an ideal target to study the innermost region
of an AGN and jet physics at highest linear resolutions \citep[for a
detailed review on this source see,
e.g.,][]{Israel1998}. At this distance, an angular resolution of
1\,milliarcsecond corrseponds to only $\sim$0.018\,pc. Cen~A is
detected from the radio up to TeV energies
\citep{Aharonian2009} and considered as a source candidate of
ultra-high energy cosmic rays \citep{Clay2010} and extragalactic
neutrinos \citep{Icecube2009}.
In the framework of the TANAMI program Cen~A has been monitored about
twice a year with VLBI at 8\,GHz since 2007, including one
simultaneous epoch at 22\,GHz \citep{Mueller2014b,Mueller2011a}. Our
observations result in the highest resolved images of Cen~A (down to
$\sim$0.4\,mas corresponding to $\sim$0.007\,pc), allowing us to study
the jet-counterjet system in unprecedented
detail. Figure~\ref{fig:cena2} shows the imaging results for the
dual-frequency observation
\citep{Mueller2011a}. Figure~\ref{fig:epochs} shows the time evolution
of the first seven 8\,GHz images at natural weighting.
Table~\ref{tablecena} provides an overview of the observation logs for
these observations and corresponding image parameters.
Here, a summary of the results from this VLBI monitoring is presented,
based on the publications by \citet{Mueller2011a} and
\citet{Mueller2014b}.
\begin{table*}
\caption{Details on TANAMI observations of Centaurus~A and image parameters}
\footnotesize
\label{tablecena}
\begin{center}
\begin{tabular}{l c c c c c c r }
\hline\hline
Obs. Date \& Frequency & Array Configuration$^{a}$ & $S_\mathrm{peak}$$^{b}$ & RMS$^{c}$ & $S_\mathrm{total}$$^{d}$ & $b_\mathrm{maj}$$^{e}$ & $b_\mathrm{min}$$^{f}$ & P.A.$^{g}$ \\
(yyyy-mm-dd) & & (Jy beam$^{-1}$) & (mJy beam$^{-1}$) & (Jy) & (mas) & (mas) & ($^\circ$) \\
\hline
\textbf{natural weighting} & & & & & & & \\
2007-11-10 (8.4\,GHz)& PKS-HART-CD-HO-MP-AT & 0.60 & $0.40\pm0.06 $ & 2.61 & 1.64 & 0.41 & 8 \\
2008-06-09 (8.4\,GHz)& PKS-HART-CD-HO-MP-AT & 1.06 & $0.63\pm0.09$ & 3.11 & 2.86 & 1.18 & $-123$\\
2008-11-27 (8.4\,GHz)& PKS-CD-HO-MP-AT-DSS43-TC-OH & 0.74 & $0.37\pm0.06 $ & 3.91 & 0.98 & 0.59 & 31 \\
2009-09-05 (8.4\,GHz)& PKS-CD-HO-MP-AT-DSS43-TC-OH & 0.76 & $0.45\pm0.07$ & 3.97 & 2.29 & 0.58 & 16 \\
2009-12-13 (8.4\,GHz)& PKS-CD-HO-MP-AT-TC & 1.03 & $0.18\pm0.03 $ & 3.82 & 3.33 & 0.78 & 26 \\
2010-07-24 (8.4\,GHz)& PKS-CD-HO-MP-AT-TC & 1.21 & $0.38\pm0.06 $ & 4.20 & 2.60 & 0.87 & 21\\
2011-04-01 (8.4\,GHz) & HART-CD-HO-MP-AT-DSS43-WW & 0.63 & $0.31\pm0.05 $ & 5.10 & 2.31 & 0.51 &$-1$\\
\hline\hline
\textbf{uniform weighting} & & & & & & & \\
2008-11-27 (8.4\,GHz)&PKS-CD-HO-MP-AT-DSS43-TC-OH& 0.48 & $0.45\pm0.02$ & 3.2 &0.68 &0.43 & 33 \\
2008-11-29 (22.3\,GHz)& PKS-CD-HO-MP-AT-DSS43& 1.46 & $1.20\pm0.10$ & 3.3&1.55 &1.21 & $-75$ \\
\hline
\end{tabular}\\
\end{center}
{\footnotesize
$^{a}$ AT: Australia Telescope Compact Array, CD: Ceduna, HART:
Hartebeesthoek, HO: Hobart, MP: Mopra, OH: GARS/O'Higgins, PKS:
Parkes, TC: TIGO, TI: DSS43 - NASA's Deep Space Network Tidbinbilla
(70\,m), WW: Warkworth; $^{b-g}$ image parameters: peak flux density, RMS noise, total
flux density, major and minor axis and position angle of the
restoring beam. We estimate a flux density uncertainty of $\sim$15\%.
}
\end{table*}
\subsection{High-resolution imaging}
The high-resolution images show a well collimated, straight jet with a small
opening angle of $\lesssim12^\circ$ \citep[compared to the rapid broadening
observed in M\,87,][]{Junor1999}. The counterjet is significantly
detected in all images.
The jet is resolved in unprecedented detail and into distinct
features. The emission is detected up to a maximum extent of
$\sim$70\,mas from the image phase center, i.e., TANAMI observations
probe the inner parsec of Cen~A with an angular resolution of less than
$\sim$0.01\,pc.
The VLBI core of the jet is identified as the brightest feature. Next
to the core, the second brightest, isolated jet component (at a
distance of
$\sim$3.5\,mas) is found to be not moving and stable in flux. It can be identified with
the stationary component discussed in
\citet{Tingay2001b}. Extragalactic jets often show stationary
components which can be explained as locally beamed emission or
standing shocks \citep{Lister2009a}. A possible interpretation of this
feature is a local pressure enhancement (like a jet nozzle) or a
cross-shock in the jet flow as observed in theoretical simulations of
over-pressured jets \citep{Mimica2009a}. Further (dual-frequency) VLBI
monitoring is required to investigate this in detail.
\subsection{The ``tuning fork''}
At a projected distance of $\sim$0.4\,pc from the jet core, the VLBI
images show a ``tuning-fork'' like structure, where the jet flow is
interrupted and widens up, but remains collimated downstream. The
brightness profile along the jet \citep[see Fig.~3 in][]{Mueller2014b}
shows a sharp gap. The surface brightness locally decreases and no
positional change is observed, i.e., the feature remains stationary
over the monitoring period. In \citet{Mueller2014b}, we discuss this
feature and possible interpretations in detail. We conclude that while
it resembles a recollimation shock, decelerating and separating the
jet flow as seen in simulations by \citet{Perucho2007}, the overall
structure is difficult to reconcile with this scenario. As a
different explanation we consider a standing disturbance, causing an
interaction with the jet. In particular, we discuss the penetration of
a gas cloud or massive star. Simulations by \citet{BoschRamon2012},
describing jet-star interactions in AGN, show similar resulting
bow-shock structures. Note that such interactions in Cen~A's jet have
already been considered by \citet{Hardcastle2003} to explain the
observed X-ray knots in the kpc-scale jet. We find that a red giant
with a stellar wind of $v\approx100\,\mathrm{km\,s^{-1}}$ and a mass
loss rate of $10^{-8}\mathrm{M_\odot yr^{-1}}$ could create an
equilibrium condition, such that it can penetrate the jet flow without
disruption. Detailed estimations and calculations can be found in
\citet{Mueller2014b}.
In \citet{Mueller2015b}, we discuss the consequences of such an
interaction event. Theoretical simulations expect $\gamma$-rays from
the collision of a red giant or a massive object with the jet
plasma
\citep{Araudo2013,Khangulyan2013,BoschRamon2012,Barkov2010}, producing
variability on short time scales of hours to days. Cen~A is a bright
$\gamma$-ray source, but it shows no significant variability
\citep{Abdo2010_cenacore,3fgl}. We estimate that this persistent emission could
be partly produced by multiple jet-star-interaction events.
\begin{figure}
\includegraphics[width=\columnwidth]{CenA_V252I_uniform.pdf}\\
\includegraphics[width=\columnwidth]{CenA_V252J_uniform.pdf}
\caption{Highest-resolution images of sub-pc scale morphology of
Centaurus~A (top: 8.4\,GHz, bottom: 22.3\,GHz). Clean contour images
from the dual-frequency observation in 2008 November
\citep{Mueller2011a} using uniform
weighting. The ellipse in the lower
left corner of each panel indicates the restoring beam.
}
\label{fig:cena2}
\end{figure}
\subsection{Sub-parsec scale jet kinematics}
The time evolution of the pc-scale jet of Cen~A had been well
studied over a period of about twelve years by
\citet{Tingay1998b,Tingay2001b}. Tracking two moving components, they
determined an apparent speed of $\sim 2\,\mathrm{mas\,yr^{-1}}$
(corresponding to $\beta_\mathrm{app}=0.12$). The 3.5\,years of
TANAMI monitoring allow us to study Cen~A's jet in more
detail. With about ten times better angular resolution we can
resolve the innermost parsec of the jet into multiple individual
components. The time evolution of the jet is presented in
Fig.~\ref{fig:epochs}.
To parametrize and track the individual jet features, we
use the \texttt{modelfit} task in DIFMAP \citep{Shepherd1997} to fit
Gaussian emission model components to the self-calibrated visibility
data.
Figure~6 in \citet{Mueller2014b} shows the kinematics, i.e., the time dependent seperation of each
component from the core. A linear regression fit gives the
proper motion for each component in mas/yr.
As an illustration of the plasma
motion in the jet, we produced a movie using the individual VLBI
images interpolated in time (Fig.~\ref{fig:epochs}), as reported in \citet{Kadler2015}. It
can be accessed at \url{http://www.aip.de/AN/movies}.
Besides the core and the stationary component next to it, we can
identify eight moving components showing a range of speeds from $\sim
1.8\,\mathrm{mas\,yr^{-1}}$ to $\sim 5\,\mathrm{mas\,yr^{-1}}$, with a
mean apparent speed of $\sim 2.98\,\mathrm{mas\,yr^{-1}}$.
Using the tapered visibility data we can show \citep{Mueller2014b} that this result is
consistent with the previous measurements by \citet{Tingay1998b,Tingay2001b}.
Furthermore the comparison of the naturally weighted and tapered
images show that we can interpret this as an underlying jet flow with
faster substructure.
We further find that the outer components have higher apparent speeds
than the ones closer to the core, i.e., showing apparent
acceleration downstream. This result connects to the speed of $\sim 0.5\,c$ measured at
$\sim$100\,pc from the core by \citet{Hardcastle2003}, already suggesting
intrinsic acceleration from pc to kpc scales.
Furthermore, this region coincides with the optically thin part of the
jet \citep{Mueller2011a}. This correlation can naturally be explained
in the context of a spine-sheath structure of the jet: a faster, inner
spine, surrounded by a slower sheath. The larger speeds measured
in optically thin regions would correspond to the faster spine.
With the derived apparent speeds and the flux density
jet-to-counterjet ratios from the individual and stacked images, we can set
constraints on the intrinsic jet speed and jet inclination angle \citep[see also
Fig.~9 in][]{Mueller2014b}, following:
\begin{equation}\label{eq:superluminal}
v_\mathrm{app}=\frac{v \sin\theta}{1-\frac{v}{c}\cos\theta}
\end{equation}
and
\begin{equation}\label{eq:R}
R=\frac{S_\mathrm{jet}}{S_\mathrm{counterjet}}=\left(\frac{1+\beta\cos\theta}{1-\beta\cos\theta}\right)^{2-\alpha}\quad,
\end{equation}
with the apparent speed $v_\mathrm{app}$, the brightness ratio $R$ of
the approaching jet to the counterjet,
the inclination angle $\theta$, the intrinsic speed $\beta =
\frac{v}{c}$, and the spectral index $\alpha$.
TANAMI observations limit the intrinsic jet speed to
$\beta\sim0.24-0.37$ and the angle to the line of sight to
$\theta\sim12^\circ-45^\circ$. The upper limits for the inclination
angle are consistent within the uncertainties with the results for the
pc-scale jet from \citet{Jones1996} and \citet{Tingay1998b}, but
our limits better match the results by \citet{Hardcastle2003} at larger
scales. We find that a change of the angle to the line of sight
cannot explain the difference in speed. Therefore, the larger
speed at kpc scales can consistently be explained by intrinsic
acceleration.
\begin{figure}
\includegraphics[width=\columnwidth]{1322-428kinematics_illustration_color_cut.png}
\caption{Time evolution of the innermost parsec of Cen~A's jet. Shown
are the individual VLBI images from TANAMI monitoring at 8\,GHz. The
image parameters are given in Table~\ref{tablecena}. Note the
different angular resolutions and beam shapes for each epoch due to
the different configurations of the TANAMI VLBI array. See also the
corresponding proper motion movie presented
in \citet{Kadler2015}.
}
\label{fig:epochs}
\end{figure}
\subsection{On the high-energy emission origin}
Cen~A is bright across the electromagnetic spectrum with a blazar-like
broadband SED of its core emission. The high-energy emission has been extensively
studied, but the origin and mechanism(s) are still debated. The radio
to $\gamma$-ray emission can be well
described by a single-zone synchrotron self-compton model
\citep{Abdo2010_cenacore}, but it fails to model the (non-simultaneous) TeV
emission. In order to better constrain theoretical models, we need to
disentangle the individual emission components and study them in detail.
In \citet{Mueller2011a}, we presented the first dual-frequency observation resulting
in the first sub-pc scale spectral index map for this source. The spectral index changes along the jet from flat/inverted
in the core, to steep further downstream. The central part of the
jet-counterjet region indicates free-free absorption as previously
discussed by \citet{Tingay2001a}. The spectral index distribution along
the jet shows multiple bright, compact, and optically thick regions,
which can be interpreted as possible production sites of highly
energetic photons. As the origin of the $\gamma$-ray emission in
extragalactic jets is still debated, this result suggests that
multi-zone emission models need to be considered.
Furthermore the origin of the hard X-ray emission is still not clear. Several results
indicate the jet as the possible origin
\citep{Tingay1998b,Fukazawa2011,Beckmann2011a}. In \citet{Tingay1998b}
the coincidence of two ejection events with X-ray high-flux states are
discussed, suggesting a relation of jet and X-ray activity.
Following this study, we use archival X-ray monitoring data by
\textsl{Swift}/BAT, \textsl{RXTE}/ASM, \textsl{RXTE}/PCA,
\textsl{CGRO}/BATSE, and the 90-350\,GHz lightcurve from SEST
\citep{Israel2008} to compare to the VLBI flux density variability at
8\,GHz from \citet{Tingay2001b} and TANAMI monitoring \citep[see
Fig.~10 in][]{Mueller2014b}.
The analysis of the jet
kinematics using TANAMI data revealed two new components ejected into
the jet during the 3.5\,year monitoring period. Their ejection time
can be estimated to be between 2007 and 2009. We find a partial
overlap of higher X-ray activity and these jet ejections, although
we cannot claim a common origin based on this result; more correlated
events need to be found.
Recently, \citet{Fuerst2016} performed a detailed study of the
3--78\,keV emission using simultaneous observations by \textsl{NuSTAR}
and \textsl{XMM-Newton}.
The sub-arcmin imaging with \textsl{NuSTAR} results in no evidence for
a jet detection above 10\,keV .
The combined spectrum can be fitted either with an absorbed power-law
component ($\Gamma = 1.815\pm0.005$) or a single-temperature
Comptonization spectrum, and an additional flourescent Fe\,K$\alpha$
emission line. The physical origin of the observed hard X-ray spectrum
is discussed in detail, concluding that the emission can be explained
by synchrotron self-Compton emission from the inner jet or by an
advection-dominated accretion flow or a combination of both.
Further multi-epoch, multiwavelength observations will help us to
disentangle these scenarios.
\section{Conclusion and Outlook}
It has been discussed how combined multiwavelength and VLBI studies of
extragalactic jets can
shed light on the physics of these powerful objects. These observations
provide both, monitoring of source activity and changes in the
spectrum as well as
highly resolved images of the innermost regions, where the power is
thought to be released.
The monitoring of the TANAMI program is set up to address open
questions in jet physics. Two TANAMI sources have been studied in
great detail, namely PMN\,J1603$-$4904\xspace and Cen~A.
Both sources present ideal objects to study the high-energy emission
and formation of jets.
PMN\,J1603$-$4904\xspace is one of the brightest sources in the $\gamma$-ray sky, but
shows no major flaring activity. Its
unusual broadband properties question its classification as a blazar
and open room for an alternative interpretation.
Future observations can confirm the CSO classification. Only recently
\citet{Migliori2016} presented the first $\gamma$-ray detection of a
confirmed CSO (PKS\,1718-649). Since PMN\,1603-4904 has a hard $\gamma$-ray spectrum, it is a
likely candidate source for TeV instruments like H.E.S.S. or in future
CTA, and
therefore it could play an important role in investigating the
high-energy properties in misaligned sources.
The sub-pc scale imaging of Cen~A provides unprecedented insights
into the properties of the inner region of an AGN jet. We observe complex jet dynamics,
which, together with long-term light curves can help to constrain SED model
parameters. The overall jet structure can be well explained by a
spine-sheath configuration. Connecting our results for the
pc-scale jet and the observations at hundreds of parsecs requires
intrinsic acceleration between these scales.
Individual jet features can be studied in detail. The jet widening at
a distance of $\sim$0.4\,pc from the core could arise from a jet-star
interaction.
Thanks to the recent developments in VLBI at millimeter
wavelengths (mm-VLBI), we will be able to further study southern
extragalactic jets at even higher angular resolution. Future mm-VLBI
observations will include the Atacama Large Millimeter Array (ALMA) in
Chile, providing for the first time enough sensitivity and suitable
$(u,v)$-coverage to image sources below $-30^\circ$ declination at
millimeter wavelengths.
In particular Cen~A presents an ideal target due to its proximity,
such that we can obtain insights into regions that are self-absorbed
at longer wavelengths and are located even closer to the jet base.
\acknowledgements
I thank the committee of the German Astronomical Society for
awarding me with the Doctoral Thesis Prize 2015.
I thank Matthias Kadler, J\"orn Wilms, and Roopesh Ojha for their support, and all
collaborators of the TANAMI, \textsl{Fermi}/LAT and ANTARES teams for the
fruitful discussions and cooperation. Special thanks
to Robert Schulz and Felicia Krau\ss for proofreading the manusscript,
and to all
colleagues of the Remeis Observatory in Bamberg, at ECAP in Erlangen, and at the Chair of
Astronomy in W\"urzburg for the inspiring working environment.
I acknowledge the funding through a PhD fellowship from the
Studienstiftung des Deutschen Volkes and the support of the
Bundesministerium für Wirtschaft und Technologie (BMWi) through the
Deutsches Zentrum für Luft- und Raumfahrt (DLR) grant 50OR1404.
This research has made use of a collection of ISIS scripts provided by
the Dr.~Karl Remeis observatory, Bamberg, Germany, at
\url{http://www.sternwarte.uni-erlangen.de/isis/}. The Long Baseline
Array and Australia Telescope Compact Array are part of the Australia
Telescope National Facility, which is funded by the Commonwealth of
Australia for operation as a National Facility managed by CSIRO.
|
\section{Introduction}
Let $X=(X_t ,t\geq 0)$ be an one-dimensional diffusion which satisfies
It\^{o}'s stochastic differential equation (SDE) of the
form
\begin{equation}\begin{array}{l}
X_t =x_0 +\int_0^t\mu (X_s ,\vartheta )\, ds+
\int_0^t \sqrt{\sigma}\, b(X_s )\, dW_s ,\quad t>0.
\end{array}\label{e1}
\end{equation}
Here, $W=(W_t ,t\geq 0)$ is an one-dimensional standard Brownian
motion,
$\mu$ and $b$ are real functions such that they ensure the
uniqueness in law of a solution to (\ref{e1})
and $x_0$ is a given deterministic initial value of $X$ (see e.g.\
\cite{Yor} as a reference for SDE).
The problem is to estimate unknown vector parameter $\theta =(\vartheta , \sigma )$
of $X$, given a discrete observation $(X_{t_i},0\leq i\leq n)$ of a
trajectory $(X_t ,t\in [0,T])$ over a time interval subdivision $0=:t_0<t_1 <\cdots < t_n :=T$, ($n$
is a positive integer) with diameter $\delta_{n,T}:=\max_{0\leq i<n}(t_{i+1}-t_i )$, $T>0$ being fixed.
Component $\vartheta$ of $\theta$ is a (vector) drift parameter, and $\sigma$ is a
diffusion coefficient parameter. We assume that $\vartheta$ belongs to
drift parameter space $\Theta$, which is an open and convex
set in Euclidean space $\mathbb {R}^d$, and that $\sigma$ is a positive
real number. Hence, $\theta =(\vartheta ,\sigma )$ is an element of
open and convex parameter space $\Psi$ $:=\Theta\times\langle 0,+\infty\rangle$.
Diffusion parameter estimation problems based on discrete
observations have been discussed by many authors (see
\cite{AitSahalia,Bibby,DCFZ,Dohnal,FZ,Kessler,Kloeden3,Breton,Yoshida}).
Although the maximum likelihood estimator (MLE) has the usual good
properties (see \cite{DCFZ}), it may not be possible to calculate it
explicitly because the transition density of process $X$ is
generally unknown and so the likelihood function (LF) of the
discrete process is unknown as well. Hence, other methods of
estimations have to be considered.
The method of parameter estimation which is discussed in this paper
and described in Section 3 below,
is based on a Gaussian approximation of the transition density and can be interpreted as based
on maximization of a discretized continuous-time log-likelihood function (LLF) as well.
Such methods
are usually called quasi-likelihood or approximate maximum likelihood (AML) methods, and in these ways
obtained estimators we will briefly call approximate maximum likelihood estimators (AMLEs).
Motivation for analyzing the method described in Section 3 is in the fact that it can provide us with
useful estimators of the parameters. It is well known that in a such way obtained AMLE of diffusion coefficient parameter is consistent and asymptotically normally distributed over fixed observational time interval $[0,T]$ when $\delta_{n,T}\rightarrow 0$ (see \cite{Dohnal} in case where all drift parameters are known, and see \cite{Jacod} in general cases). The same holds in ergodic
diffusion cases when $T\rightarrow +\infty$ in a way that $\delta_{n,T}=T/n\rightarrow 0$
for appropriate equidistant sampling (see e.g.\ \cite{FZ} or \cite{Kessler}). Local asymptotic properties of the AMLE of drift parameters over fixed interval $[0,T]$ and when $\delta_{n,T}\rightarrow 0$ are less known especially in more general cases,
particularly when drift is nonlinear in its parameters (see \cite{Bishwal}).
Although a knowledge of local asymptotic properties of drift parameter AMLEs does not imply their consistency or asymptotic normality necessarily it may help in further analysis of the AMLEs which might include, for example, measuring effects of discretization on the estimator's standard errors with applications in simulation studies. In ergodic diffusion cases it is well known that the AMLE of drift (vector) parameter is consistent and asymptotically normal and efficient when $T \rightarrow +\infty$ in a way that $T\delta_{n,T}^2\rightarrow 0$ for equidistant sampling (see e.g.\ \cite{FZ} for one-dimensional case and \cite{Kessler} for vector and more general cases) but the local properties of drift parameters AMLEs (for example the rate of convergence
of $\hat{\vartheta}_{n,T}-\hat{\vartheta}_T$ to zero) are still less investigated as well. Let us stress that the
problems of statistical inferences about diffusion drift parameters are very important especially in biomedical modeling
(see \cite{HuzakI}).
In this paper we state general conditions for proving and prove: (1.) existence and measurability of the AMLEs, (2.)
that $\hat{\vartheta}_{n,T}-\hat{\vartheta}_T$ converges to zero with rate
$\sqrt{\delta_{n,T}}$ in probability
when $\delta_{n,T}\rightarrow 0$ over fixed bounded observational time interval $[0,T]$,
and (3.) that $\hat{\vartheta}_{n,T}-\hat{\vartheta}_T$ converges to zero with rate
$\sqrt{\delta_{n,T}}$ in probability
when $T\rightarrow +\infty$ in a way that $\delta_{n,T}=T/n \rightarrow 0$ in an ergodic diffusion case and equidistant sampling. We apply these findings in
proving (4.) measurability, consistency and asymptotic normality of diffusion coefficient parameter AMLEs when $\delta_{n,T}\rightarrow 0$ in both cases: when $T$ is fixed, and in an ergodic diffusion case when $T\rightarrow +\infty$ and $T\delta_{n,T}=T^2/n\rightarrow 0$ with equidistant sampling, and (5.)
consistency and asymptotic normality and efficiency of drift parameter AMLEs in an ergodic case when $T\rightarrow +\infty$ in a way that $T\delta_{n,T}\rightarrow 0$ with equidistant sampling.
Properties (1.-2.) for drift parameter AMLEs were proved
in \cite{Breton} in cases when drift depended linearly on its parameters. For detailed review of liner case see \cite{Bishwal}. The first nonlinear case was covered by the author in his Ph.D. thesis \cite{Huzak}. The main assumption was that the drift was an analytic function in its parameters with properly bounded derivatives of all orders. In this paper we only assume that the drift has at least $d+3$ continuous derivatives with respect to the drift parameters ($d$ is a dimension of the drift parameter vector). The main difficulty was in proving core technical Theorem \ref{tm:new} of Section 5. Although facts (4.-5.) have been already known we included these alternative proofs for completeness and the illustrative purposes of the applicability of the findings (1.-3.) and in this paper developed methods. We belive that other discretization schemes (for example, of higher order) can be analyzed similarly by using the techniques of this paper.
The paper is organized in the following way. In the next section we introduce
notation used through the paper. The discussed method of
estimation is described in Section 3. The main results are presented in Section 4.
An example is provided in Section 5.
The proofs of the main results are in the last section. Lemmas are proved in Appendix.
\section{Notations}
Let
$|\cdot |$ denote Euclidean norm in $\mathbb {R}^d$ and its induced operator norm, and let
$|\cdot |_\infty$ be max-norm. If $f$ is a bounded
real function, $\|f\|_\infty:=\sup_\mathbf{z} |f(\mathbf{z})|$ is a sup-norm of $f$. Let $L^p (\mathbb {P} )$ be
the Banach space of all random variables with finite $p$-th moment
and let $\|\cdot\|_{L^p(\mathbb {P} )}$ denote its
norm.
If $(x,\vartheta )\mapsto f(x,\vartheta )$
is a real function defined on an open subset of $\mathbb {R}\times\mathbb {R}^d$,
then we denote by $D_{\vartheta }^m f(x,\vartheta )$ the $m$-th partial derivative
with respect to $\vartheta$. Let
$|D_{\vartheta }^m f(x,\vartheta ) |_\infty :=
\max_{j_1 +\cdots +j_d =m}|\frac{\partial^m f}{\partial\vartheta_1^{j_1} \ldots\partial
\vartheta_d^{j_d}}|$. In this case we say that $D_{\vartheta }^m f(x,\vartheta )$ is bounded
if all partial derivatives $\frac{\partial^m f}{\partial\vartheta_1^{j_1} \ldots\partial
\vartheta_d^{j_d}}(x,\vartheta )$
are bounded, and $\|D_{\vartheta }^m f\|_\infty
:=\max_{j_1 +\cdots +j_d =m}\|\frac{\partial^m f}{\partial\vartheta_1^{j_1} \ldots\partial
\vartheta_d^{j_d}}\|_\infty$.
The notation $D^2_{\vartheta } f(x,\vartheta )<\O$ means that the
Hessian $D^2_{\vartheta } f(x,\vartheta )$ is a negatively definite matrix. Similarly for a positively definite matrix.
$D_{\mathbf{z}}^0 f\equiv f$ by convention.
The $m$-th derivative of $f$ at a point $\mathbf{z}$ we simply denote by $D^m f(\mathbf{z})$.
Let ${\cal K}$ and $\Theta$ be open sets in $\mathbb {R}^d$.
The closure and the boundary of ${\cal K}$ will be denoted by
$\overline{\cal K}$ and $\partial {\cal K}$ respectively, and the $\sigma$-algebra of Borel subsets of $\Theta$ by
${\cal B}(\Theta )$. If ${\cal K}\subset\Theta$ is an open set such that $\overline{\cal K}$ is compact in $\Theta$
then we will say that ${\cal K}$ is a relatively compact set in $\Theta$.
Let $(\gamma_n ,n\geq 1 )$ be a sequence of positive numbers and let
$(Y_n ,n\geq 1 )$ be a sequence of random variables defined on some probability space.
We will say
that $(Y_n ,n\geq 1 )$ is $O_\mathbb {P} (\gamma_n)$, and write
$Y_n =O_\mathbb {P} (\gamma_n )$, if the sequence $(Y_n /\gamma_n ,n\geq 1)$
is bounded in probability, i.e.\ if
\[\begin{array}{l}
\lim_{A\rightarrow +\infty}\overline{\lim_n}\,\mathbb {P}\{ \gamma_n^{-1}|Y_n|> A\} =0.
\end{array}\]
\section{Estimation method}
Let $0=t_0<t_1<\cdots <t_n=T$ be discrete times at which diffusion $X$
is observed, and let us denote by $\Delta$ the difference operator
defined in the following way: if $F$ is a function defined on
$[0,T]$ then $\Delta_i F:=F(t_{i+1})-F(t_i)$, $0\leq i<n$.
Let us discretize SDE (\ref{e1}) over interval $[t_i,t_{i+1}]$ by
using the Euler approximation of the both types of integrals:
\[\begin{array}{l}
X_{t_{i+1}}-X_{t_i}\approx\mu (X_{t_i},\vartheta )
(t_{i+1}-t_{i})+\sqrt{\sigma}\, b(X_{t_i}) (W_{t_{i+1}}-W_{t_i}).
\end{array}\]
In this way the following stochastic difference equation is
obtained:
\begin{equation}\begin{array}{l}
\Delta_i Z=\mu (Z_i ,\vartheta )\, \Delta_it+\sqrt{\sigma}\, b(Z_i
)\,\Delta_i W
\end{array}\label{e2}
\end{equation}
for $0\leq i<n$, and $Z_0 =x_0$.
Solution to (\ref{e2}) is a time-discrete process
$Z=(Z_0,Z_1,\ldots,Z_n)$ that is an approximation of $X$ over
$[0,T]$.
Up to the constant not depending on the parameters
a LLF of the process $Z$ is
\begin{equation}\begin{array}{l}
-\frac{1}{2}\sum_{i=0}^{n-1}\left(\frac{
(\Delta_i Z-\mu (Z_i ,\theta )\Delta_i t)^2}{\sigma b^2 (Z_i )\Delta_i t}+
\log\sigma\right).
\end{array}\label{LLFd}
\end{equation}
Criterion function
\begin{equation}\begin{array}{l}
{\cal L}_{n,T} (\theta )={\cal L}_{n,T} (\vartheta ,\sigma ):=
-\frac{1}{2}\sum_{i=0}^{n-1}\left(\frac{(\Delta_i X-\mu (X_{t_i} ,
\vartheta )\Delta_it )^2}{\sigma b^2 (X_{t_i} )\Delta_it}+
\log\sigma\right)
\end{array}\label{e3}
\end{equation}
is obtained from (\ref{LLFd}) by substituting $(Z_i ,0\leq i\leq n)$
with discrete observations $(X_{t_i} ,0\leq i\leq n)$ of
diffusion $X$. Notice that
\[\begin{array}{l}
{\cal L}_{n,T} (\vartheta ,\sigma )=-\frac{1}{2\sigma}\sum_{i=0}^{n-1}
\frac{(\Delta_i X)^2}{b^2 (X_{t_i}) \Delta_it}-\frac{n}{2}\log\sigma +
\frac{1}{\sigma}{\ell}_{n,T} (\vartheta)
\end{array}\]
where
\begin{equation}\begin{array}{l}
\ell_{n,T} (\vartheta )=\sum_{i=0}^{n-1} \frac{\mu (X_{t_i},\vartheta )}{b^2
(X_{t_i})}\Delta_i X-\frac{1}{2}\sum_{i=0}^{n-1}
\frac{\mu^2 (X_{t_i},\vartheta )}{b^2 (X_{t_i})}\Delta_it \label{lnizraz}
\end{array}\end{equation}
depends only on drift parameter $\vartheta$.
A point of maximum $\hat{\theta}_{n,T} =(\hat{\vartheta}_{n,T} ,
\hat{\sigma}_{n,T} )$ of function (\ref{e3}) in $\Psi$
is an AMLE of vector parametar $\theta$ if it exists. Notice that if AMLE
exists then necessary
\begin{equation}
D{\cal L}_{n,T} (\hat{\vartheta}_{n,T},\hat{\sigma}_{n,T} )=0
\,\Leftrightarrow\,\left\{
\begin{array}{l} D\ell_{n,T} (\hat{\vartheta}_{n,T})=0\\
\hat{\sigma}_{n,T} =\frac{1}{n}\sum_{i=0}^{n-1}\frac{(\Delta_i X-\mu (X_{t_i},\hat{\vartheta}_{n,T})\Delta_it )^2
}{b^2 (X_{t_i})\Delta_it}.
\end{array}\right. \label{defv}
\end{equation}
Hence every stationary point $\hat{\vartheta}_{n,T}$ of function $\ell_{n,T}$
uniquely determines second component $\hat{\sigma}_{n,T}$ of stationary point $\hat{\theta}_{n,T}=(\hat{\vartheta}_{n,T},\hat{\sigma}_{n,T})$ of
function ${\cal L}_{n,T}$
by the following expression:
\begin{equation}\begin{array}{l}
\hat{\sigma}_{n,T} =\frac{1}{n}\sum_{i=0}^{n-1}\frac{(\Delta_i X-\mu (X_{t_i},
\hat{\vartheta}_{n,T} )\Delta_it )^2}{b^2 (X_{t_i})\Delta_it}.
\end{array}\label{hatvn}
\end{equation}
Moreover,
if $\hat{\vartheta}_{n,T}$ is a unique point of the global maximum of
function $\ell_{n,T}$ then $\hat{\theta}_{n,T}$ is a unique point of the global
maximum of function ${\cal L}_{n,T}$. Hence to prove existence of a measurable AMLE $\hat{\theta}_{n,T}$
it is sufficient to prove that there exists a measurable point of maximum of function $\ell_{n,T}$.
\section{Main results}
\subsection{{\em Fixed maximal observational time case\/}}
Let the following assumptions be satisfied.
\vspace{5pt}
{\sc (H1a):} For all $\theta =(\vartheta ,\sigma )\in\Psi$,
there exists a strong solution $(X,W)$ of the SDE (\ref{e1}) on time interval
$[0,+\infty\rangle$
with values in open interval
$E\subseteq\mathbb {R}$. \vspace{5pt}
{\sc (H2a):} For all $\vartheta\in\Theta$, $\mu (\cdot,\vartheta )\in C^2(E)$
and
$b\in C^3 (E)$.
Moreover for all $x\in E$, $b(x)\neq 0$ and $\hbox{\rm sign}\, b =$ const.
\vspace{5pt}
For example, by Theorem 5.2.2 in \cite{Friedman},
{\sc (H1a)} will be satisfied if in addition to
{\sc (H2a)} we assume that for all $\vartheta\in\Theta$ SDE (\ref{e1})
satisfies so called the bounded linear growth assumption, i.e.\ that
there exists a positive constant $C$ such that for all $x\in E$,
$
|\mu (x,\vartheta )|+|b(x)|\leq C(1+|x|)
$
More precisely, {\sc (H2a)} states that the functions $x\mapsto b(x)$ and
$x\mapsto\mu (x,\vartheta )$, $\vartheta\in\Theta$, are continuously
differentiable in $E$ and hence locally Lipschitz. In this case there
exists a strong, continuous and pathwise unique solution to SDE (\ref{e1})
on time interval $[0,+\infty\rangle$.
However, there are some SDEs which satisfy {\sc (H1a)} and {\sc (H2a)} but do not
satisfy the linear growth assumption
(see e.g. the example of Section 5).
\vspace{5pt}
{\sc (H3a):} For all $(x,\vartheta )\in E\times\Theta$ and all $1\leq m\leq d+3$, there
exists partial derivatives $D_\vartheta^m\mu (x,\vartheta )$,
$\frac{\partial}{\partial x}D_\vartheta^m\mu (x,\vartheta )$,and
$\frac{\partial^2}{\partial x^2}D_\vartheta^m\mu (x,\vartheta )$
of drift function $\mu$.
Moreover, for all $0\leq m\leq d+3$,
$D_\vartheta^m\mu$, $\frac{\partial}{\partial x}D_\vartheta^m\mu$, $\frac{\partial^2}{\partial x^2}D_\vartheta^m\mu\in C(E\times\Theta)$.
\vspace{5pt}
Let $\mathbb {P}_\theta$ denote the law of $X$ for $\theta\in\Psi$.
We assume that probabilities $\mathbb {P}_\theta$, $\theta
\in\Psi$, are defined on filtered space
$(\Omega ,({\cal F}_T^0, T\geq 0) )$ where $\Omega$ is a set of continuous functions $\omega :[0,+\infty\rangle\rightarrow
E$ such that $\omega (0)=x_0$, ${\cal F}_T^0$ is a
$\sigma$-algebra generated by the coordinate functions up to the time $T$, and the filtration is augmented in so called the usual way
(see e.g.\ I.4 in \cite{Yor}).
On this space, coordinate process $(\omega\mapsto\omega (t),
t\geq 0)$ is a canonical version of $X$ (see \cite{Yor}, I.\S3).
Hence, for each $T>0$ we assume that $X$ is defined on the measurable space $(\Omega ,
{\cal F}_T^0 )$ as a canonical process with law $\mathbb {P}_\theta$.
For the moment, let us assume that we are able to observe the process
$(X_t ,0\leq t\leq T)$ continuously. Because diffusion coefficient
parameter $\sigma$ can be uniquely determined through equation
\begin{equation}\begin{array}{l}
\sigma =\frac{\lim_n \sum_{i=1}^{2^n}(X_{jT2^{-n}}-X_{(j-1)T2^{-n}})^2}{
\int_{0}^{T}b^2 (X_t )\, dt}\quad (\mbox{\rm a.s.}\;\mathbb {P}_\theta )
\end{array}\label{sigma}
\end{equation}
(see \cite{BH}) since $b^2>0$ by {\sc (H2a)}, the estimation problem from continuously
observed process can be reduced to an estimation problem for drift
parameter $\vartheta\in\Theta$. In this case
for every fixed diffusion parameter $\sigma$ assumed to be known,
and every two different $\vartheta_1 ,\vartheta_2\in\Theta$, probability
measures $\mathbb {P}_{(\vartheta_1 ,\sigma )}$ and $\mathbb {P}_{(\vartheta_2 ,\sigma )}$
are equivalent on ${\cal F}_T^0$, and
\[\begin{array}{l}
\log\frac{d\mathbb {P}_{(\vartheta_2 ,\sigma )}}{d\mathbb {P}_{(\vartheta_1 ,\sigma )}}\! =\!
\frac{1}{\sigma}(\!\int_0^T\!\!\!\frac{\mu (X_t ,\vartheta_2 )\! -\!
\mu (X_t ,\vartheta_1 )}{b^2 (X_t )} dX_t\! -\!\frac{1}{2}
\int_0^T\!\!\!\frac{\mu^2 (X_t ,\vartheta_2 )\! -\!\mu^2 (X_t ,\vartheta_1 )}{
b^2 (X_t )} dt)
\end{array}\]
where
$\frac{d\mathbb {P}_{(\vartheta_2 ,\sigma )}}{d\mathbb {P}_{(\vartheta_1 ,\sigma )}}$
denotes Radon-Nikodym derivative of $\mathbb {P}_{(\vartheta_2 ,\sigma)}$
with respect to $\mathbb {P}_{(\vartheta_1 ,\sigma)}$ on ${\cal F}_T^0$ (see \cite{Feigin}).
If we fix some $\vartheta_*\in\Theta$, a continuous-time LLF is
$\vartheta\mapsto\log\frac{d\mathbb {P}_{(\vartheta ,\sigma)}}{
d\mathbb {P}_{(\vartheta_* ,\sigma )}}$. Up to the constant and factor not depending on $\vartheta$,
function
\begin{equation}\begin{array}{l}
\ell_T (\vartheta ):=
\int_0^T\frac{\mu (X_t ,\vartheta )}{b^2 (X_t )}\, dX_t -
\frac{1}{2}\int_0^T\frac{\mu^2 (X_t ,\vartheta )}{b^2 (X_t )}\, dt.
\end{array}\label{ell}
\end{equation}
is equal to the LLF.
Hence, $\ell_T$ will be called a
continuous-time LLF (see \cite{Lanska}). Assumption
{\sc (H3a)} implies that $\ell_T$ is at least three-times continuously
differentiable function on $\Theta$, and for $1\leq m\leq d+3$, its derivatives are equal
to (see \cite{Lanska} for $m\leq 2$)
\begin{equation}
\begin{array}{l}
D^m\ell_T (\vartheta )=
\int_0^T\frac{1}{b^2 (X_t )}D_\vartheta^m\mu (X_t ,\vartheta )\, dX_t -
\frac{1}{2}\int_0^T\frac{1}{b^2 (X_t )}D_\vartheta^m\mu^2 (X_t ,\vartheta )\,
dt
\end{array}\label{Dell}
\end{equation}
{\sc (H4a):} For all $\omega\in\Omega$, function
$\vartheta\mapsto\ell_T (\vartheta )=\ell_T (\vartheta, \omega )$
has a unique point of global maximum
$\hat{\vartheta}_T=\hat{\vartheta}_T(\omega )$ in $\Theta$.
Moreover, $D_{\vartheta}^2 \ell_T (\hat{\vartheta}_T)<\O$.
\vspace{5pt}
Assumption {\sc (H4a)}
enables property ($ii$) in
Theorem \ref{tm:30} below, to be proved. If {\sc (H3a)} and {\sc (H4a)} hold then Lemma 4.1.\
from \cite{huzakMLE} implies that $(\omega,\vartheta)\mapsto\ell_T(\vartheta)(\omega)$ is
an ${\cal F}_T^0\otimes {\cal B}(\Theta )$-measurable function, and continuous-time MLE
$\hat{\vartheta}_T$ is an ${\cal F}_T^0$-measurable random variable.
Let ${\cal F}_{n,T}$ be a $\sigma$-subalgebra of ${\cal F}_T^0$ generated by discrete observation
$(X_{t_i}, 0\leq i\leq n)$ of process
$(X_t ,0\leq t\leq T)$. Notice that if {\sc (H3a)} holds then $(\omega,\vartheta )\mapsto{\ell}_{n,T} (\vartheta, \omega )$ (given by \ref{lnizraz})
is an ${\cal F}_{n,T}\otimes {\cal B}(\Theta)$ measurable function by Lemma 4.1.\ in \cite{huzakMLE}.
If $\ell_{n,T}$ is a concave function on $\Theta$ then a stationary point $\hat{\vartheta}_{n,T}$ is an unique point of maximum of $\ell_{n,T}$ on $\Theta$ and hence it is ${\cal F}_{n,T}$-measurable by e.g. Lemma 4.1.\ in \cite{huzakMLE}.
If $\ell_{n,T}$ is not a concave function on $\Theta$,
for proving ${\cal F}_{n,T}$-measurability of estimators
$\hat{\vartheta}_{n,T}$ (and so $\hat{\theta}_{n,T}$) introduced in Section 3 we
need additional assumptions:
\vspace{5pt}
{\sc (H5a):} $\Theta$ is a relative compact set in $\mathbb {R}^d$, and for each $0\leq m\leq d+3$,
$D_\vartheta^m\mu$, $\frac{\partial}{\partial x}D_\vartheta^m\mu$, $\frac{\partial^2}{\partial x^2}D_\vartheta^m\mu \in C(E\times\overline{\Theta})$.\vspace{5pt}
{\sc (H6a):}
For all $\omega\in\Omega$ and some $r>0$,
\[\begin{array}{l}
\ell_T (\hat{\vartheta}(\omega ), \omega)>\sup_{|x|\geq r}\ell_T (\hat{\vartheta}(\omega )+x, \omega).
\end{array}\]
\noindent
Assumption {\sc (H6a)} holds if {\sc (H5a)} holds and $\hat{\vartheta}_T$ is the unique point of maximum of
$\ell_T$ on compact $\overline{\Theta}$.
\begin{theorem}\label{tm:30}
Let us assume that {\sc (H1a-4a)} hold and $T>0$ be fixed.
Then there exists a sequence $(\hat{\vartheta}_{n,T}, n\geq 1 )$ of ${\cal F}_T^0$-measurable random vectors
such that for all $\theta =(\vartheta ,\sigma )\in\Psi$
and when $\delta_{n,T}\downarrow 0$,
\begin{romannum}
\item $\lim_n\mathbb {P}_{\theta}(D{\ell}_{n,T}
(\hat{\vartheta}_{n,T})=\O )=1$
\item $(\mathbb {P}_{\theta})\lim_n\hat{\vartheta}_{n,T} =
\hat{\vartheta}_T$
\item $\hat{\vartheta}_{n,T} -\hat{\vartheta}_T=O_{\mathbb {P}_\theta}(\sqrt{\delta_{n,T}})$, $n\rightarrow +\infty$
\item If $(\tilde{\vartheta}_{n,T},n\geq 1)$ is an ${\cal F}_T^0$-measurable sequence in $\Theta$ that satisfies $(i-ii)$ then
$\lim_n\mathbb {P}_{\theta} (\tilde{\vartheta}_{n,T} =\hat{\vartheta}_{n,T})=1$.
\end{romannum}
If either for $n\geq 1$ and
almost all $\omega\in\Omega$ function $\vartheta\mapsto
{\ell}_{n,T}(\vartheta, \omega )$ has a unique point of local
maximum which is a point of the global maximum as well,
or the hypotheses {\sc (H5a-6a)} are satisfied,
then $\hat{\theta}_{n,T}$ can be chosen to be
${\cal F}_{n,T}$-measurable.
\end{theorem
\begin{corollary}\label{cor:30a}
Let {\sc (H1a-4a)} hold, $T>0$ be fixed, and $(\hat{\sigma}_{n,T},n\geq 1)$
be given by $(\ref{hatvn})$. Then
\begin{romannum}
\item $(\mathbb {P}_{\theta})\lim_n\hat{\sigma}_{n,T}
=\sigma$;
\item $(\sqrt{n}\frac{1}{\sigma\sqrt{2}
}(\hat{\sigma}_{n,T} -\sigma),n\geq 1)$ converges in law w.r.t.\ $\mathbb {P}_\theta$ to the standard normal
distribution $N(0,1)$ when $n\rightarrow +\infty$.
\end{romannum}
Moreover, if $\hat {\vartheta}_{n,T}$ is ${\cal F}_{n,T}$-measurable then $\hat{\sigma}_{n,T}$ is ${\cal F}_{n,T}$-measurable too.
\end{corollary}
\begin{remark}\label{rem1}{\rm
Theorem \ref{tm:30} still holds if we replace {\sc (H1a)} with the assumption that $T<\xi$ a.s. where $\xi$ is a maximal random time such that SDE (\ref{e1}) has a solution on $[[0,\xi[[=\{(\omega,t)\in\Omega\times [0,+\infty\rangle\,:\;0\leq t<\xi (\omega)\}$.
$\xi$ exists by assumption {\sc (H2a)} and the existence and uniqueness theorem for SDEs (see e.g.\ \cite{Friedman} or \cite{Yor}).
}
\end{remark}
\begin{remark}\label{rem2}{\rm
Theorem \ref{tm:30} still holds if the drift and diffusion coefficient functions depend on time variable too (non autonomous case: $(t,x)\mapsto\mu (t,x,\vartheta )$, $\sigma b(t,x)$) in a way that assumptions {\sc (H2a)} and {\sc (H3a)} hold for $\mu$ and $b$
with $x$ and $E$ replaced with $(t,x)$ and $\tilde{E}=[0,+\infty\rangle\times E$ respectively.
}
\end{remark}
\subsection{{\em Ergodic diffusions case}}
Let the coefficient diffusion function parameter $\sigma >0$ be fixed. We need
the following assumptions.\vspace{5pt}
{\sc (H1b):} {\sc (H1a)} holds, and $X$ is an ergodic diffusion with stationary distribution $\pi_\vartheta (dx)$,
$\vartheta\in\Theta$. \vspace{5pt}
{\sc (H2b):} {\sc (H2a)} holds, and for all $\vartheta\in\Theta$ functions $\mu (\cdot ,\vartheta )b'/b, (b')^2, b''b\in L^{16} (\pi_\vartheta )$, $b^2b'''\in L^{8} (\pi_\vartheta )$, and there exist a function $c\in L^{1} (\pi_\vartheta )$ and a number $h_0>0$ such that
\[\begin{array}{l}
\sup_{0<h\leq h_0}\mathbb {E}_{(\vartheta,\sigma)}\exp\left( 8\int_0^{h}\left(2\frac{\mu_0 b'}{b} +\sigma (b''b+15b'^2)\right)(X_s)\,ds\right)\leq c(x_0).
\end{array}\]
{\sc (H3b):} {\sc (H3a)} and {\sc (H5a)} hold, and there exist nonnegative functions $g_{0},g_1,g_2:E\rightarrow\mathbb {R}$ such that
for all $\vartheta_0\in\Theta$, $g_0\in L^{32} (\pi_{\vartheta_0} )\cap C^1(E)$ such that $g_0' b\in L^{16} (\pi_{\vartheta_0} )$,
$g_1\in L^{16} (\pi_{\vartheta_0} )\cap C (E)$, $g_2\in L^{8} (\pi_{\vartheta_0} )\cap C (E)$, and for all $x\in E$ and $0\leq m\leq d+3$,
\[\begin{array}{lcl}
\sup_{\vartheta\in\overline{\Theta}}|D^m_\vartheta\mu (x,\cdot )/b(x)|_\infty &\leq &g_0(x)\\
\sup_{\vartheta\in\overline{\Theta}}|\frac{\partial}{\partial x}D^m_\vartheta\mu (x,\cdot )|_\infty
&\leq & g_1(x)\\
\sup_{\vartheta\in\overline{\Theta}}|\frac{\partial^2}{\partial x^2}D^m_\vartheta\mu (x,\cdot )b(x)|_\infty &\leq & g_2(x).
\end{array}\]
{\sc (H4b):} For all $\vartheta\in\Theta$,
\begin{equation}\begin{array}{l}
(\forall \vartheta'\in\overline{\Theta})\; \vartheta'\neq\vartheta\Rightarrow
\int_E \frac{(\mu (x,\vartheta)-\mu(x,\vartheta'))^2}{b^2 (x)}\, \pi_\vartheta (dx) >0.
\end{array}\label{identiassump}
\end{equation}
{\sc (H5b):} For all $\vartheta\in\Theta$, functions $\frac{\partial\mu }{\partial\vartheta_i}(\cdot,\vartheta )/b$, $1\leq i\leq d$, are linearly independent in $L^2 (\pi_\vartheta )$.
\vspace{5pt}
\noindent
$\Theta$ is a relatively compact set in $\mathbb {R}^d$ by assumption {\sc (H5a)} since {\sc (H3b)} holds.
Assumptions {\sc (Hb1-b3)} imply that for all $\vartheta_0\in{\Theta}$ and $\vartheta\in\overline{\Theta}$, $\mathbb {P}_{(\vartheta_0,\sigma)}$-a.s.
\begin{equation}\begin{array}{l}
\lim_{T\rightarrow +\infty}\frac{1}{T}\ell_T (\vartheta )=\frac{1}{2}\int_E\frac{\mu (x,\vartheta_0)^2 -
(\mu (x,\vartheta_0 )-\mu (x,\vartheta ))^2}{b^2(x)}\,\pi_{\vartheta_0} (dx)
=:\ell_{\vartheta_0} (\vartheta )
\end{array}\label{liml0}
\end{equation}
by ergodic property of the diffusion and the law of large numbers for continuous martingales (see e.g. \cite{Yor}, Chapters V and X).
Function $\ell_{\vartheta_0} :\overline{\Theta}\rightarrow\mathbb {R}$ defined for every $\vartheta_0\in
\Theta$ by formula (\ref{liml0})
is at least three times continuously differentiable on compact $\overline{\Theta}$ by {\sc (H3b)}, and
\[
\begin{array}{l}
D\ell_{\vartheta_0} (\vartheta )=\int_E \frac{
(\mu (x,\vartheta_0 )-\mu (x,\vartheta ))}{b^2(x)}D_\vartheta \mu (x,\vartheta )\,\pi_{\vartheta_0} (dx)\\
D^2\ell_{\vartheta_0} (\vartheta )=\int_E\left(\frac{
(\mu (x,\vartheta_0 )-\mu (x,\vartheta ))}{b^2(x)}D_\vartheta^2 \mu (x,\vartheta )
-\frac{1}{b^2 (x)}(D_\vartheta^\tau \mu D_\vartheta\mu )(x,\vartheta)\right)\,\pi_{\vartheta_0} (dx).
\end{array
\]
Hence, by the same argument as for (\ref{liml0}),
for any fixed $\vartheta\in\overline{\Theta}$, $\mathbb {P}_{(\vartheta_0,\sigma)}$-a.s.
\begin{equation}\begin{array}{lcl}
\lim_{T\rightarrow +\infty}\frac{1}{T}D\ell_T (\vartheta ) &=& D\ell_{\vartheta_0} (\vartheta ),\\
\lim_{T\rightarrow +\infty}\frac{1}{T}D^2\ell_T (\vartheta ) &=& D^2\ell_{\vartheta_0} (\vartheta ).
\end{array}\label{liml0dd2}
\end{equation}
If $\vartheta\neq\vartheta_0$ then $\ell_{\vartheta_0}(\vartheta )<\ell_{\vartheta_0} (\vartheta_0 )$
by (\ref{liml0}),
and
{\sc (H4b)}. Hence $\vartheta_0$ is the unique point of maximum of
$\ell_{\vartheta_0}$ on $\overline{\Theta}$. This implies identifiability property of the model: let $\vartheta_1,\vartheta_2\in\Theta$ be such that $\mathbb {P}_{(\vartheta_1,\sigma )}=\mathbb {P}_{(\vartheta_2,\sigma )}$. Then
$\pi_{\vartheta_1}=\pi_{\vartheta_2}$ and so $\ell_{\vartheta_1}\equiv\ell_{\vartheta_2}$ by (\ref{liml0}). Hence
$\vartheta_1 =\vartheta_2$. Moreover, {\sc (H5b)} implies that the Fisher information matrix is positive definite, i.e.\
\[\begin{array}{l}
I(\vartheta_0 )=-D^2\ell_{\vartheta_0} (\vartheta_0 )=\int_E\frac{1}{b^2 (x)}(D_\vartheta^\tau \mu D_\vartheta\mu )(x,\vartheta_0 )\,\pi_{\vartheta_0} (dx)>\O.
\end{array}\]
The next theorem states that the continuous-time MLE of drift parameters exists, is consistent and asymptotically efficient, and satisfies assumptions {\sc (H4a)} and {\sc (H6a)} a.s.\ for almost all observational times.
\begin{theorem}\label{tm:cont}
Let us assume that {\sc (H1b-5b)} hold. Then
there exists an $({\cal F}_T^0, T>0)$-adapted process $(\hat{\vartheta}_{T}, T>0 )$ of random vectors
such that for every $\theta =(\vartheta ,\sigma )\in\Psi$ the following holds:
\begin{romannum}
\item $\mathbb {P}_{\theta}$-a.s.\ there exists $T_0 >0$ such that
for all $T\geq T_0$, $\hat{\vartheta}_{T}\in\Theta$ is the unique point of maximum of $\ell_T$ on $\overline{\Theta}$, and $D^2 \ell_T (\hat{\vartheta}_T)<\O$ in a way that $\min_{|y|=1}y^\tau (-\frac{1}{T}D^2\ell_T(\hat{\vartheta}_T))y\geq\frac{1}{2}\min_{|y|=1}y^\tau I(\vartheta )y$.
\item $\lim_{T\rightarrow +\infty}\hat{\vartheta}_T=\vartheta$ $\mathbb {P}_\theta$-a.s.
\item $(\sqrt{T}(\hat{\vartheta}_T-\vartheta ), T>0)$ converges in law w.r.t.\ $\mathbb {P}_{\theta}$ to normal law $N(\O,\sigma I(\vartheta)^{-1})$ with expectation $\O$ and covariance matrix
$\sigma I(\vartheta)^{-1}$.
\end{romannum}
\end{theorem}\vspace{5pt}
\noindent
The following theorem is a version of Theorem \ref{tm:30} for ergodic diffusions. In addition it states that AMLEs are consistent and asymptotically efficient when both maximal observational time and number of discrete observational time points tend to infinity for appropriate sampling schemes. Hence in its statement '$\lim_{n,T}$' denotes the limit when
both $T\rightarrow +\infty$ and $n\rightarrow +\infty$.
\begin{theorem}\label{tm:erg}
Let us assume that {\sc (H1b-5b)} hold.
Then there exists a process $(\hat{\vartheta}_{n,T}; n\geq 1, T>0 )$ of ${\cal F}_{n,T}$-measurable random vectors
$\hat{\vartheta}_{n,T}$ such that for all $\theta =(\vartheta ,\sigma )\in\Psi$ and $\pi_\vartheta$-a.s.\ nonrandom initial conditions,
and all equidistant samplings such that $\delta_{n,T}=T/n\rightarrow 0$, the following holds.
\begin{romannum}
\item $\lim_{n,T}\mathbb {P}_{\theta}(D{\ell}_{n,T}
(\hat{\vartheta}_{n,T})=\O )=1$.
\item $(\mathbb {P}_{\theta})\lim_{n,T}(\hat{\vartheta}_{n,T}-
\hat{\vartheta}_T)=\O$,
\item $\hat{\vartheta}_{n,T} -\hat{\vartheta}_T=O_{\mathbb {P}_\theta}(\sqrt{\delta_{n,T}})$, $n\rightarrow +\infty$, $T\rightarrow +\infty$
\item If $(\tilde{\vartheta}_{n,T};n\geq 1,T>0)$ is a process of random vectors in $\Theta$ that satisfies $(i-ii)$ then
$\lim_{n,T}\mathbb {P}_{\theta} (\tilde{\vartheta}_{n,T} =\hat{\vartheta}_{n,T})=1$.
\item $(\mathbb {P}_{\theta})\lim_{n,T}\hat{\vartheta}_{n,T}=
\vartheta$, and if
in addition $\lim_{n,T}T\delta_{n,T}=0$ then
\[\begin{array}{l}
\sqrt{T}(\hat{\vartheta}_{T,n}-\vartheta )\stackrel{{\cal L}-\mathbb {P}_{\theta}}{\longrightarrow} N(\O,\sigma I(\vartheta)^{-1}),\; T\rightarrow +\infty, n\rightarrow +\infty.
\end{array}\]
\item $(\mathbb {P}_{\theta})\lim_{n,T}\sigma_{n,T}=\sigma$, and if in addition $\lim_{n,T}T\delta_{n,T}=0$ then
\[\begin{array}{l}\sqrt{n}\frac{1}{\sigma\sqrt{2}
}(\hat{\sigma}_{n,T} -\sigma)\stackrel{{\cal L}-\mathbb {P}_{\theta}}{\longrightarrow} N(0,1),\;
T\rightarrow +\infty,n\rightarrow +\infty.\end{array}\]
\end{romannum}
\end{theorem
\section{Example}
Let the stochastic generalized logistic model be given with the following SDE:
\begin{equation}\begin{array}{l}
dX_t =(\alpha-\beta X_t^\gamma)X_t\, dt + \sqrt{\sigma} X_t\, dW_t,\; X_0 =x_0>0\label{logSDE}
\end{array}\end{equation}
where $\vartheta =(\alpha,\beta,\gamma)$ ($\gamma>0$) is a drift vector parameter. By using the methods of
stochastic calculus it is possible to explicitly solve (\ref{logSDE}) that proves
that there exists pathwise unique, continuous and strong solution to this SDE
with $X$ defined on $\Omega\times [0,+\infty\rangle$ and values in $E=\langle 0,+\infty\rangle$.
Moreover, it turns out that for drift parameters such that
$\alpha>\sigma/2$, $\beta>0$ and $\gamma >0$, generalized logistic process $X$ is recurrent and ergodic
with a such stationary distribution $\pi_\vartheta$ that for stationary $X$, $X_t^\gamma$ follows
$\Gamma$-distribution with parameters $A:=2(\alpha-\sigma/2)/(\gamma\sigma)$ and $B:=\gamma\sigma/(2\beta)$ (i.e.
$\mathbb {E} X_t^\gamma =AB$, $\mathbb {E} (X_t^{\gamma})^2 =AB(B+1)$) by e.g. Theorem 7.1, pp. 219-220 in \cite{Friedman}.
Hence, assumption {\sc (H1b)} holds.
In generalized logistic model, drift function is equal $\mu (x,\vartheta )=(\alpha-\beta x^\gamma )x$, and up to the
diffusion parameter $\sigma >0$, diffusion coefficient function is $b(x)=x >0$ on $E$. Hence $b'\equiv 1$,
$bb''=b^2 b'''\equiv 0$ that are trivially integrable with respect to any probability law. Let
$f (x,\vartheta )=\mu (x,\vartheta)/b(x)=\alpha-\beta x^\gamma $. Notice that any partial derivatives of $f$ with respect
to $\vartheta$ are of the form $-\beta^n x^\gamma \log^m x$ where $n\in\{ 0,1\}$, $m\in\mathbb {N}_0$. Of the same forms
are components of $b^{k} \frac{\partial^k}{\partial x^k}D_\vartheta^m f$ for $k=1, 2$. Finally, any $p$-th power of
their absolute values ($p$ is a positive integer) are of the form $x^c |\log x|^m$ up to a constant, where $c >0$ is a real
number and $m$ is a nonnegative integer. These functions are
integrable with respect to $\pi_\vartheta$. If we choose a relative compact $\Theta$ of drift parametric set
$\langle\frac{\sigma}{2},+\infty\rangle\times\langle 0,+\infty\rangle^2$ then there exist $\alpha_0>\sigma/2$,
$\beta_0 >0$ and $\gamma_0>0$ such that for all $\vartheta\in\overline{\Theta}$, $x>0$, and all $0\leq m\leq 6$, $k\in\{0,1,2\}$
and integers $j_\alpha$, $j_\beta$, $j_\gamma$ such that $j_\alpha+j_\beta +j_\gamma=m$,
\[\begin{array}{l}
|b^k(x) \frac{\partial^{m+k}}{\partial^k x\partial\alpha^{j_\alpha}\partial\beta^{j_\beta}\partial\gamma^{j_\gamma}} f(x,\vartheta)|\leq g(x):=\alpha_0+\beta_0 x^{\gamma_0}(1+\log^2 x+\log^4 x +\log^6 x).
\end{array}\]
Then $g\in L^p (\pi_\vartheta)\cap C^1 (E)$ for all $p\geq 1$ and $\vartheta\in\overline{\Theta}$ which implies
partially {\sc (H2b)} and {\sc (H3b)} by simple calculation (see the proof of Corollary \ref{tm:cor2} below). To finish the proof of {\sc (H2b)} notice that for all $h_0>0$, and all $0<h\leq h_0$,
\[\begin{array}{l}
\exp(16\int_0^h((\alpha -\beta X_t^\gamma+\frac{15\sigma}{2})\, dt)\leq \exp((16\alpha_0+120\sigma) h_0)=c(x_0)=\mbox{\rm constant}
\end{array}\]
since $X_t >0$ for all $t\geq 0$ and $\beta >0$. This implies the same inequalities for expectations with respect to any
initial conditions $X_0 =x_0$. Hence {\sc (H2b)} is proved.
To show that {\sc (H4b)} holds, let us assume that
\[\begin{array}{l}\int_E (\mu (x,\vartheta_1)-\mu (x,\vartheta_2))^2/b^2(x)\pi_{\vartheta_1}(dx)=0
\end{array}\]
for some $\vartheta_1\in\Theta$ and $\vartheta_2\in\overline{\Theta}$. Since $\pi_{\vartheta_1}$ is absolutely continuous w.r.t.
Lesbegues measure $\lambda$ on $E$, this implies that $\mu(x,\vartheta_1)=\mu (x,\vartheta_2)$ for a.s. $x>0$ w.r.t. $\lambda$.
Hence, smooth function $u(x):=\beta_1 x^{\gamma_1}-\beta_2 x^{\gamma_2}$ must be a constant function for $\lambda$-a.s. $x>0$. This
implies that $\gamma_1=\gamma_2$ and hence $\vartheta_1 =\vartheta_2$. This proves {\sc (H4b)}.
Finally, {\sc (H5b)} holds since $\frac{\partial}{\partial\alpha} \mu (x, \vartheta)=1$, $\frac{\partial}{\partial\beta} \mu (x, \vartheta)=-x^\gamma$, and
$\frac{\partial}{\partial\gamma} \mu (x, \vartheta)=-\beta x^\gamma\log x$ are obviously linearly independent functions in $L^2(\pi_\vartheta )$.\vspace{5pt}
\section{Proofs}
Basically the
proof of Theorem \ref{tm:30} is based on the so-called general
theorem on approximate maximum likelihood estimation and its
corollary that are stated and proved in \cite{huzakMLE} as Theorem 3.1
and Corollary 3.2. The proof of Theorem \ref{tm:erg} is a modification of
the proof of the same theorem based on Theorem \ref{tm:cont}.
But first we need to state and prove Theorem \ref{tm:new}, and
its Corollaries \ref{tm:23} and \ref{tm:cor2} that are needed in applying the general theorem in
this context. Proofs of some technical lemmas are in Appendix.
Let us suppose that $X=(X_t, t\geq 0)$ is a diffusion satisfying {\sc (H1a-2a)} with true parameter
$\theta_0 =(\vartheta_0,\sigma )\in\Theta$, and such that $\mathbb {P} (X_0=x_0)=1$ for $x_0\in E$. Here
$\mathbb {P}\equiv\mathbb {P}_{\theta_0}$, $\mathbb {E}\equiv\mathbb {E}_{\theta_0}$ and $L^2\equiv L^2(\mathbb {P} )$. Denote $\mu_0 =\mu (\cdot,\vartheta_0)$, $\nu =\sqrt{\sigma} b$,
${\cal A}a:=a'\mu_0+a''\nu^2/2$, and $\bar{a}:=|{\cal A}a|+|a'\nu|$ for $a\in C^2(E)$.
\begin{theorem}\label{tm:new}
Let $\Theta\subset\mathbb {R}^d$ be an open convex set, and let $f:E\times\Theta\rightarrow\mathbb {R}$,
$a:E\rightarrow\mathbb {R}$ be functions. Let $0=t_0<t_1<\cdots <t_n=T$ be subdivisions of intervals $[0,T]$, $T>0$, such that
$\delta_{n,T}\downarrow 0$. Assume the following:
\begin{nitemizetm}
\item[{\sc (B1):}] $a\in C^2(E)$ and there exist constants $C_a>0$, $T_a\geq 0$, and $n_a\in\mathbb {N}$ such that
\[\begin{array}{l}
(\forall\, T >T_a)(\forall n\geq n_a)\;\;\frac{1}{T}\mathbb {E}\left(\!\!\int_{0}^{T}\!\!(a^4\!+\!\bar{a}^4)(X_t))dt
+\sum_{i=0}^{n-1}\!a^4 (X_{t_i})\Delta_i t\right)\leq C_a.
\end{array}\]
\item[{\sc (B2):}] For all $\vartheta\in\Theta$, $f(\cdot,\vartheta )\in C^2 (E)$, and for all
$(x,\vartheta)\in E\times\Theta$ and $1\leq m\leq d+1$ there exists partial derivatives
$D^m_\vartheta f(x,\vartheta )$, $\frac{\partial}{\partial x}D^m_\vartheta f(x,\vartheta )$, and
$\frac{\partial^2}{\partial x^2}D^m_\vartheta f(x,\vartheta )$. Moreover,
\[\begin{array}{l}
(\forall\, 0\leq m\leq d+1)\;\;
D^m_\vartheta f,\, \frac{\partial}{\partial x}D^m_\vartheta f,\, \frac{\partial^2}{\partial x^2}D^m_\vartheta f\in C(E\times\Theta).
\end{array}\]
\item[{\sc (B3):}] For any relatively compact set ${\cal K}$ in $\Theta$ there exist: a positive measurable function
$g:E\rightarrow\mathbb {R}$ such that for all $0\leq m\leq d+1$,
\[\begin{array}{l}
\sup_{\vartheta\in\overline{\cal K}}\left(\!|D^m_\vartheta f (\cdot ,\vartheta )|_\infty\!+\!|\frac{\partial}{\partial x}D^m_\vartheta f (\cdot ,\vartheta )|_\infty(|\mu_0 |\! +\! |\nu |)\!+\!
|\frac{\partial^2}{\partial x^2}D^m_\vartheta f (\cdot ,\vartheta )\nu^2 |_\infty\right)\! \leq\! g,
\end{array}\]
and constants $C_g>0$, $T_g\geq 0$, and $n_g\in\mathbb {N}$, such that
\[\begin{array}{ll}
(\forall\, T>T_g)(\forall n\geq n_g)\;\;\frac{1}{T}\mathbb {E}\left(\!
\int_{0}^{T}g^4\! (X_t)\,dt +
\sum_{i=0}^{n-1}g^4\! (X_{t_i})\Delta_i t\!\right) & \leq C_g\;\&\\
\quad\quad\frac{1}{T}\mathbb {E}
\sum_{i=0}^{n-1}\left(\! (ga)^4\! (X_{t_i})\Delta_i t +
g^4\!(X_{t_i})\int_{t_i}^{t_{i+1}}(a^4+\bar{a}^4)\!(X_t)\,dt\!\right) &\leq C_{g}.
\end{array}\]
\item[{\sc (B4):}] There exist: a measurable function $c: E\rightarrow\mathbb {R}$ and constants $h_0>0$, $C_c>0$, $T_c\geq 0$, and $n_c\in\mathbb {N}$
such that for $r:=|\frac{\mu_0 b'}{b}| +|b''b|+|b'|$,
\[\begin{array}{ll}
\sup_{0<h\leq h_0}\mathbb {E}\exp\left( 8\int_0^{h}\left(2\frac{\mu_0 b'}{b} +\sigma (b''b+15b'^2)\right)(X_s)\,ds\right) &\leq c(x_0),\\
(\forall\, T>T_c)(\forall n\geq n_c)\;\;\frac{1}{T}\mathbb {E}\left(\!
\sum_{i=0}^{n-1}c(X_{t_i})\Delta_i t +
\int_0^T r^8(X_t)dt\right) &\leq C_c.
\end{array}\]
\end{nitemizetm}
Then there exist constants $C_1>0$, $C_2>0$, $T_0 \geq 0$, and $n_0\in\mathbb {N}$, possible dependent on ${\cal K}$, $d$, and $a$, such that
for all $T> T_0$, and $n\geq n_0$,
\begin{eqnarray}
\mathbb {E}\sup_{\vartheta\in\overline{\cal K}}\left(\frac{1}{T\sqrt{\delta_{n,T}}}\sum_{i=0}^{n-1}\int_{t_i}^{t_{i+1}}\!\!\!(f(X_t,\vartheta ) -f(X_{t_i},\vartheta))a(X_t)\,dt\right)^2\!
&\leq & \! C_1\label{ineq1}\\
\mathbb {E}\sup_{\vartheta\in\overline{\cal K}}\left(\frac{1}{\sqrt{T\delta_{n,T}}}\sum_{i=0}^{n-1}\int_{t_i}^{t_{i+1}}\!\!\!(f(X_t,\vartheta ) -f(X_{t_i},\vartheta))a(X_t)\,dW_t \right)^2\!
&\leq & \! C_1\label{ineq2}\\
\mathbb {E}\sup_{\vartheta\in\overline{\cal K}}\left(\frac{1}{T\sqrt{\delta_{n,T}}}\sum_{i=0}^{n-1}\int_{t_i}^{t_{i+1}}\!\!\!f(X_{t_i},\vartheta )\left(\frac{b(X_t )}{b(X_{t_i})}-1\right) a(X_t)\,dt \right)^2\! &\leq &\! C_2 \label{b2}\\
\mathbb {E}\sup_{\vartheta\in\overline{\cal K}}\left(\frac{1}{\sqrt{T\delta_{n,T}}}\sum_{i=0}^{n-1}\int_{t_i}^{t_{i+1}}\!\!\!f(X_{t_i},\vartheta )\left(\frac{b(X_t )}{b(X_{t_i})}-1\right) a(X_t)\,dW_t \right)^2\! &\leq &\! C_2 \label{b3}
\end{eqnarray}
\end{theorem}
\begin{remark}\label{rem3}{\rm If function $f$ that satisfies (B2) and all its partial derivatives from (B2) are bounded on $E\times\overline{\cal K}$ then (B3) holds if $a$ is bounded too.
Similarly if $a\in C^2(E)$
is bounded then satisfies (B1).
If in addition $\mu_0$, $b$, $b'$, and $b''$ are bounded then (B4) holds for constant function $c\equiv\exp(\gamma h_0)$ where $\gamma >0$ and $h_0>0$ are constants. In this case the statements of Theorem \ref{tm:new} hold for $T_0=0$, and hence for
all $T>0$ obviously from the proof of Theorem \ref{tm:new}.
}
\end{remark}
For a moment let us assume that ${\cal K}=\prod_{i=1}^d\langle a_i,b_i\rangle$ is an open and bounded
$d$-dimensional rectangular in $\Theta$. Then there
exists $\varepsilon >0$ such that ${\cal K}_\varepsilon:=\prod_{i=1}^d\langle a_i-\varepsilon,b_i+\varepsilon\rangle$ is an open and bounded $d$-rectangular in $\Theta$ too. Let $\phi:\mathbb {R}^d\rightarrow\mathbb {R}$ be a $C^\infty$-function such that
$\phi\equiv 1$ on ${\cal K}$ and $\phi\equiv 0$ on ${\cal K}_\varepsilon^c$. Such a function exists (see e.g.\ \cite{Borisovich}, Lemma IV.4.4, p.\ 176). Then function $(x,\vartheta )\mapsto \tilde{f}(x,\vartheta):=f(x,\vartheta )\cdot\phi (\vartheta )$
satisfies (B2-3) if $f$ satisfies the same assumption (with rescaled function $g$). Namely, $\tilde{f}\equiv f$ on $E\times{\cal K}$
and $\tilde{f}\equiv 0$ on $\partial{\cal K}_\varepsilon$. The same holds for all partial derivatives of $\tilde{f}$ that exist, and
$\tilde{f}$ satisfies (B1) obviously. Since
$\phi$ and all of its derivatives are bounded, $\tilde{f}$ satisfies (B3) too with $Cg$ instead of $g$ with a constant $C$ depending on $\phi$.
Obviously, statements (\ref{ineq1}-\ref{b3}) hold for a function $f$ that satisfies (B2-3), and a rectangular
${\cal K}$
if (\ref{ineq1}-\ref{b3}) hold for $\tilde{f}$ and the rectangular ${\cal K}_\varepsilon$. Moreover, notice that if (\ref{ineq1}-\ref{b3}) hold for an arbitrary open and bounded $d$-dimensional rectangular ${\cal K}$, then the same statements
hold for every relatively compact set in $\Theta$. Hence it is sufficient to prove (\ref{ineq1}-\ref{b3}) for an open and bounded $d$-dimensional rectangular ${\cal K}\subset\Theta$, and a function $f$ satisfying (B2-3) and the following additional assumption.\vspace{5pt}
{\sc (B ${\cal K}$)}: For all $x\in E$ and all $0\leq m\leq d+1$,
$D^m_\vartheta f(x,\cdot)\equiv\O$, $\frac{\partial}{\partial x}D^m_\vartheta f(x,\cdot )\equiv\O$
and $\frac{\partial^2}{\partial x^2}D^m_\vartheta f(x,\cdot )\equiv\O$
on $\partial{\cal K}$.\vspace{5pt}
Moreover, let $A$ be an invertible affine mapping of $\mathbb {R}^d$, and let $f$ be a function
on $E\times\Theta$ that satisfies (B2-3) and (B ${\cal K}$). Then the function $\bar{f}$ defined on
$E\times A(\Theta )$ by the rule $\bar{f}(x,\eta ):=
f(x,A^{-1}\eta )$, satisfies (B2-3) and (B $A{\cal K}$) too. Since the left hand side of (\ref{ineq1}-\ref{b3})
do not change by the change of variable $\vartheta\mapsto \eta=A\vartheta$, it is sufficient
to prove (\ref{ineq1}-\ref{b3}) for ${\cal K}_0:=\langle -\pi,\pi\rangle^d$ and a function $f$ that satisfies
(B2-3) and (B ${\cal K}_0$).
Now, let $f$ be a function satisfying (B2-3) and (B ${\cal K}_0$). For $x\in E$,
$\hbox{\bf k}=(k_1,\ldots,k_d)\in\mathbb {Z}^d$, $\vartheta =(\vartheta_1,\ldots,\vartheta_d)\in\Theta$, and $\mathbf{j}=(j_1,\ldots,j_d)$
where $j_1$,..., $j_d$ are nonnegative integers such that $m:=j_1+\cdots+j_d\leq d+1$, let us define Fourier coefficients of $f$ by
\[\begin{array}{l}
C_{\hbox{\bf k}}(x)\;\; :=\frac{1}{(2\pi)^d}\int_{\overline{\cal K}_0}f(x,\vartheta )e^{-i\langle\hbox{\bf k}|\vartheta\rangle}d\vartheta,\\
C^{(\mathbf{j} )}_{\hbox{\bf k}}(x) :=\frac{1}{(2\pi)^d}\int_{\overline{\cal K}_0}\frac{\partial^mf}{\partial\vartheta_1^{j_1}\cdots
\partial\vartheta_d^{j_d}}(x,\vartheta )e^{-i\langle\hbox{\bf k} |\vartheta\rangle}d\vartheta .
\end{array}\]
Let $\hbox{\bf k}^{\mathbf{j}}:=k_1^{j_1}\cdots k_d^{j_d}$. Since (B ${\cal K}_0$) holds, it is well known
that
$C^{(\mathbf{j} )}_{\hbox{\bf k}}(x)=i^m \hbox{\bf k}^{\mathbf{j}} C_{\hbox{\bf k}}(x)$ for each fixed $x\in E$ (see e.g.\ \cite{Taylor}, pp.\ 177-178). This relation is used in the proof of the next few lemmas (see Appendix).
\begin{lemma}\label{lema4} Let $x,y\in E$. Then for all $\hbox{\bf k}\in\mathbb {Z}^d$,
\[\begin{array}{l}
|C_{\hbox{\bf k}} (x)|\leq g (x
\left(\frac{d+1}{1+|k_1|+\cdots+|k_d|}\right)^{d+1}.
\end{array}\]
\end{lemma}
\begin{lemma}\label{lema:4mom} Let $f\in C(E)$.
Then for all $0\leq t_0<t$,
\[\begin{array}{ll}
\mathbb {E}{\left(\!\!\int_{t_0}^t\!\! f(X_s)\, dW_s\!\right)\!\!}^4 &\leq 3e^{3(t-t_0)}\mathbb {E}\!\int_{t_0}^t\! f^4 (X_s )\, ds\leq\\
&\leq 24(e^{3(t-t_0)}\mathbb {E}\!\!\int_{t_0}^t\!\! (f(X_s )\! -\! f(X_{t_0}))^4\, ds+\mathbb {E}[f^4(X_{t_0})](t-t_0)^2).
\end{array}\]
\end{lemma}
\begin{lemma}\label{lema1} Let
{\sc (B4)} hold. If $c_0 :=(1+c)/2$ then
\begin{equation}\begin{array}{l}
(\forall t\geq 0)\;\; \sup_{0< h\leq h_0}\mathbb {E}\left(\frac{b(X_{t+h})}{b(X_t)}\right)^8 \leq \mathbb {E}\, c_0 (X_t).
\end{array}\label{b1}
\end{equation}
\end{lemma}
\begin{lemma}\label{lema6} There exist constants $K_1>0$, $K_2>0$, $T_0\geq 0$, and $n_0\in\mathbb {N}$, depending on ${\cal K}_0$, $g$ and $a$,
and such that for all $\hbox{\bf k}\in\mathbb {Z}^d$, $T>T_0$, $n\geq n_0$ and subdivisions $0=t_0<t_1<\cdots <t_n=T$ (with $\delta_{n,T}\downarrow 0$)
the following hold:
\begin{eqnarray}
\|\!\frac{1}{T\!\sqrt{\delta_{n,T}}}\!\!\sum_{i=0}^{n-1}\!\!\!\int_{t_i}^{t_{i+1}}\!\!\!\!\!\!\!\!\!\!
(C_{\hbox{\bf k}}(X_t)\! -\! C_{\hbox{\bf k}}(X_{t_i})) a(X_t)\,dt\|_{L^2}\!
&\leq &\! K_1\cdot K_\hbox{\bf k}\label{ineq3}\\
\|\!\frac{1}{\sqrt{T\delta_{n,T}}}\!\!\sum_{i=0}^{n-1}\!\!\!\int_{t_i}^{t_{i+1}}\!\!\!\!\!\!\!\!\!\!
(C_{\hbox{\bf k}}(X_t)\! -\!C_{\hbox{\bf k}}(X_{t_i})) a(X_t)\,dW_t \|_{L^2}\!
&\leq &\!K_1\cdot K_\hbox{\bf k}\label{ineq4}\\
\|\!\frac{1}{T\!\sqrt{\delta_{n,T}}}\!\!\sum_{i=0}^{n-1}\!\!\!\int_{t_i}^{t_{i+1}}\!\!\!\!\!\!\!\!\!\!
C_{\hbox{\bf k}}(X_{t_i})\!\left(\!\frac{b(X_t )}{b(X_{t_i})}\!-\!1\!\right)\! a(X_t)\,dt\|_{L^2}\!
&\leq &\! K_2\cdot K_\hbox{\bf k}\label{ineq5}\\
\|\!\frac{1}{\sqrt{T\delta_{n,T}}}\!\!\sum_{i=0}^{n-1}\!\!\!\int_{t_i}^{t_{i+1}}\!\!\!\!\!\!\!\!\!\!
C_{\hbox{\bf k}}(X_{t_i})\!\left(\!\frac{b(X_t )}{b(X_{t_i})}\!-\!1\!\right)\! a(X_t)\,dW_t \|_{L^2}\!
&\leq &\!K_2\cdot K_\hbox{\bf k},\label{ineq6}
\end{eqnarray}
where $K_\hbox{\bf k} :=((d+1)/(1+|k_1|+\cdots+|k_d|))^{d+1}$.
\end{lemma}
Let $S_N(x,\vartheta ):=\sum_{|\hbox{\bf k}|\leq N} C_{\hbox{\bf k}}(x)e^{i\langle\hbox{\bf k} |\vartheta\rangle}$ for $x\in E$,
$\vartheta\in\overline{\cal K}_0$ and $N$ be a positive integer. Then it can be proved that
$\lim_N|S_N(x,\vartheta )-f(x,\vartheta )|=0$ uniformly in $\vartheta\in\overline{\cal K}_0$
by the methods of Fourier analysis (see e.g.\ \cite{Taylor}, pp.\ 180-183).
\begin{lemma}\label{lema5}
$\sum_{\hbox{\bf k}\in\mathbb {Z}^d}|C_{\hbox{\bf k}}(x)|\leq K g (x)$, and
$\sup_{N, \vartheta\in\overline{\cal K}_0}|S_N(x,\vartheta )-f(x,\vartheta )|\leq K g (x)$ for a positive and finite constant
\begin{equation}\begin{array}{l}
K=\sum_{\hbox{\bf k}\in\mathbb {Z}^d}{\left(\!\!\frac{d+1}{1+|k_1|+\cdots+|k_d|}\!\right)\!.}^{\!\! d+1}
\end{array} \label{K}
\end{equation}
\end{lemma}
\begin{lemma}\label{lema3} Let $a\in C^1 (E)$ and let $f$ be a function that satisfies {\sc (B2)}. Then
for a.s. $\omega\in\Omega$, function $\vartheta\mapsto\int_0^T f(X_t,\vartheta ) a(X_t)\, dW_t(\omega )$ is continuous on $\Theta$.
\end{lemma}
{\em Proof of Theorem \ref{tm:new}}.
Let us prove (\ref{ineq2}) and (\ref{b3}). The proofs of
(\ref{ineq1}) and (\ref{b2}) go in the same way but we have to obtain expressions
of form (\ref{L2}) below with respect to
Lesbegues' instead of Winner's integral, and to apply Lemma \ref{lema6} (\ref{ineq3}) and (\ref{ineq5}).
Without loosing generality let us assume that
${\cal K}={\cal K}_0=\prod_{i=1}^d\langle -\pi,\pi\rangle$ and let $f$ satisfy (B2-3) and (B ${\cal K}_0$).
For fixed $\vartheta\in {\cal K}_0$, $T>0$ and a subdivision
$0=t_0<t_1<\cdots <t_n=T$ we define the following processes:
\[\begin{array}{l}
U_t\quad:=\sum_{i=0}^{n-1}(f(X_t,\vartheta)-f(X_{t_i},\vartheta )) a(X_t)\1_{\langle t_i,t_{i+1}]}(t),\; t\in [0,T],\\
U_t^{(N)}:=\sum_{i=0}^{n-1}(S_N(X_t,\vartheta)-S_N(X_{t_i},\vartheta )) a(X_t)\1_{\langle t_i,t_{i+1}]}(t),\; t\in [0,T],\;\; N\in\mathbb {N},
\end{array}\]
and
\[\begin{array}{l}
V_t\quad:=\sum_{i=0}^{n-1}f(X_{t_i},\vartheta )\left(\frac{b(X_t)}{b(X_{t_i})}-1\right) a(X_t)\1_{\langle t_i,t_{i+1}]}(t),\; t\in [0,T],\\
V_t^{(N)}:=\sum_{i=0}^{n-1}S_N(X_{t_i},\vartheta )\left(\frac{b(X_t)}{b(X_{t_i})}-1\right) a(X_t)\1_{\langle t_i,t_{i+1}]}(t),\; t\in [0,T],\;\; N\in\mathbb {N},
\end{array}\]
Then $\lim_N |U_t^{(N)} -U_t|=0$, $\lim_N |V_t^{(N)} -V_t|=0$, and $\sup_N |U_t^{(N)}-U_t|\leq K^2 (g^2 (X_t) + g^2 (X_{t_i})) + a^2 (X_t)/2$, $\sup_N |V_t^{(N)}-V_t|\leq (K/2)g^2 (X_{t_i})\,a^2 (X_t)+2(b(X_t)/b(X_{t_i}))^2+2$,
for $ t\in \langle t_i,t_{i+1}]$ by Lemma \ref{lema5}. Since (B1-4) hold and hence Lemma \ref{lema1} holds there exist $T_1\geq 0$ and $n_1\in\mathbb {N}$ such that for all $T>T_1$, $n\geq n_1$ integrals
$\int_0^T g^2 (X_t)\, dW_t$, $\sum_{i=0}^{n-1} g^2 (X_{t_i})\Delta_i W$,
$\sum_{i=0}^{n-1} g^2 (X_{t_i})\int_{t_i}^{t_{i+1}}a^2 (X_t)\,dW_t$, $\sum_{i=0}^{n-1}\int_{t_i}^{t_{i+1}}(b(X_t)/b(X_{t_i}))^2\, dW_t$, \\and $\int_0^T a^2 (X_t)\, dW_t$ are well defined, and so
\[\begin{array}{l}
I_{N}(\vartheta ):=\int_0^TU_t^{(N)}\, dW_t\stackrel{\mathbb {P}}{\rightarrow}\int_0^T U_t\,dW_t=:I(\vartheta ),\; N\rightarrow +\infty,\\
J_{N}(\vartheta ):=\int_0^TV_t^{(N)}\, dW_t\stackrel{\mathbb {P}}{\rightarrow}\int_0^T V_t\,dW_t=:J(\vartheta ),\; N\rightarrow +\infty,
\end{array}\]
by the dominated convergence theorem for stochastic integrals (see e.g.\ \cite{Yor}, Theorem (2.12), pp.\ 134-135).
First, let us consider sequence $(I_N(\vartheta))$. For every
$\vartheta\in {\cal K}_0\cap\mathbb {Q}^d$
there exists a subsequence $(N_p)\equiv(N_p(\vartheta))$ and an event $A(\vartheta)$ of the probability 1 such that for all $\omega\in A(\vartheta)$, $\lim_p I_{N_p}(\vartheta)(\omega)=I(\vartheta )(\omega)$. Let us recall that
\[\begin{array}{lcl}
I_{N}(\vartheta)&=&\sum_{i=0}^{n-1}\int_{t_i}^{t_{i+1}}\!(S_N(X_t,\vartheta )-S_N(X_{t_i},\vartheta)) a(X_t)\, dW_t,
\;\; N\in\mathbb {N},\\
I(\vartheta)&=&
\sum_{i=0}^{n-1}\int_{t_i}^{t_{i+1}}\!(f(X_t,\vartheta )-f(X_{t_i},\vartheta)) a(X_t)\, dW_t.
\end{array}\]
Let $\Omega_0:=\cap_{\vartheta\in {\cal K}_0\cap\mathbb {Q}^d}A(\vartheta )$. Then on this event of probability 1,
for all $\vartheta\in {\cal K}_0\cap\mathbb {Q}^d$, the following holds:
\[\begin{array}{lcl}
|I(\vartheta )| &\leq& |I(\vartheta)-I_{N_p(\vartheta )}(\vartheta )|+|I_{N_p(\vartheta )}(\vartheta )|\leq
|I(\vartheta)-I_{N_p(\vartheta )}(\vartheta )|+\\
&& +\sum_{\hbox{\bf k}\in\mathbb {Z}^d}|\sum_{i=0}^{n-1}\int_{t_i}^{t_{i+1}}
(C_{\hbox{\bf k}}(X_t)-C_{\hbox{\bf k}}(X_{t_i})) a(X_t)\, dW_t|.
\end{array}\]
By taking limit when $p\rightarrow +\infty$, we get the following inequality:
\[\begin{array}{l}
|I(\vartheta )|\leq\sum_{\hbox{\bf k}\in\mathbb {Z}^d}|\sum_{i=0}^{n-1}\int_{t_i}^{t_{i+1}}
(C_{\hbox{\bf k}}(X_t)-C_{\hbox{\bf k}}(X_{t_i}))a(X_t)\, dW_t|.
\end{array}\]
Since $\vartheta\mapsto I(\vartheta )$ is a continuous function by Lemma \ref{lema3}, it turns out that
$\sup_{\vartheta\in\overline{\cal K}_0}|I(\vartheta )|=\sup_{\vartheta\in {\cal K}_0\cap\mathbb {Q}^d}|I(\vartheta )|$, and
so $\sup_{\vartheta\in\overline{\cal K}_0}|I(\vartheta )|$ is a random variable. Hence
\begin{equation}\begin{array}{l}
\sup_{\vartheta\in\overline{\cal K}_0}|I(\vartheta )|\leq\sum_{\hbox{\bf k}\in\mathbb {Z}^d}\left|\sum_{i=0}^{n-1}\int_{t_i}^{t_{i+1}}
(C_{\hbox{\bf k}}(X_t)-C_{\hbox{\bf k}}(X_{t_i})) a(X_t)\, dW_t\right|\;\mbox{\rm a.s.}
\end{array}\label{L2}
\end{equation}
Since there exist $T_0\geq T_1$ and $n_0\geq n_1$ such that for all $T>T_0$, $n\geq n_0$ and subdivisions of $[0,T]$
with $\delta_{n,T}\downarrow 0$,
\[\begin{array}{l}
\sum_{\hbox{\bf k}\in\mathbb {Z}^d}\|\sum_{i=0}^{n-1}\int_{t_i}^{t_{i+1}}(C_{\hbox{\bf k}}(X_t)-C_{\hbox{\bf k}}(X_{t_i}))a(X_t)\,dW_t\|_{L^2}\leq
K_1 K
\sqrt{T\delta_{n,T}},
\end{array}\
by Lemma \ref{lema6} and (\ref{K}), the series on the righthand side of (\ref{L2}) converges a.s.\ and
in $L^2$-norm to a.s.\ equal limits (see Proposition 2.10.1.\ in
\cite{Brockwel}, p.\ 68). Hence $\|\sup_{\vartheta\in {\cal
K}_0}|I(\vartheta )|\|_{L^2}$ $\leq C_1\sqrt{T\delta_{n,T}}$ for $C_1:=K_1 K$. That proves
(\ref{ineq2}). The proof of (\ref{b3}) goes in a similar way considering sequence $(J_N(\vartheta))$. \endproof\vspace{5pt}
We need following lemma for proving consistency and asymptotic normality of
diffusion coefficient parameter estimator.\vspace{5pt}
\begin{lemma}\label{difkoef} Let {\sc (B4)} hold, and let $b\in C^3(E)$. Moreover, let there exist constants $C_b>0$ and $T_b\geq 0$ such that
\[\begin{array}{l}
(\forall T>T_b)\;
\frac{1}{T}\mathbb {E}(\int_0^T ((b^2b''')^2+r^{16})(X_t)\, dt+\sum_{i=0}^{n-1}r^4(X_{t_i})\Delta_it)\leq C_b.
\end{array}\]
Then there exist constants
$C>0$,
$T_0 \geq 0$, and $n_0\in\mathbb {N}$, such that for all $T> T_0$, and $n\geq n_0$,
\[\begin{array}{l}
\frac{1}{T}\mathbb {E} \left|\sum_{i=0}^{n-1}\frac{1}{\Delta_i t}\left(\left(\int_{t_i}^{t_{i+1}}\frac{b(X_t)}{b(X_{t_i})}\, dW_t\right)^2 -(\Delta_i W)^2\right)\right|\leq C.
\end{array}\]
\end{lemma}
\begin{remark}\label{rem4}{\rm If $b$ and its derivatives up to the third order are bounded then the statement of
Lemma \ref{difkoef} hold for all $T>T_0 =0$ by the same arguments as in Remark \ref{rem3}.
}\end {remark}
\subsection{\em Fixed maximal observational time case}
Let $T>0$ be fixed, and let $0=t_0<\cdots t_n=T$, $n\in\mathbb {N}$, be subdivisions of $[0,T]$ such that
$\delta_{n,T}=\max_{0\leq i\leq n-1}\Delta_it\downarrow 0$ when $n\rightarrow +\infty$.
We need the next corollary to Theorem \ref{tm:new}.
\begin{corollary}\label{tm:23}
Let $X$ be a diffusion such that {\sc(H1a-4a)} hold and let
${\cal K}\subset\Theta$ be a relatively compact set. Then for all $\theta_0=(\vartheta_0,\sigma)\in
\Psi$, $T>0$, and $r=0,1,2$,
\begin{equation}\begin{array}{l}
\sup_{\vartheta\in\overline{\cal K}}|D^{r}\ell_{n,T} (\vartheta )-D^{r}\ell_T (\vartheta)|=
O_{\mathbb {P}_{\theta_0}}(\sqrt{\delta_{n,T}}),\;\; n\rightarrow +\infty.
\end{array}\label{drugac}
\end{equation}
\end{corollary}
{\em Proof of Corollary \ref{tm:23}}. We prove (\ref{drugac}) for $r=0$.
Statement (\ref{drugac}) for cases $r=1$ and $r=2$ can be proved similarly.
Let $\theta_0 =(\vartheta_0 ,\sigma)\in\Psi$ be arbitrary, and let
$\mu_0 :=\mu(\cdot,\vartheta_0)$. Moreover, let $f(\cdot,\vartheta):=\mu(\cdot,\vartheta)/b$,
$\vartheta\in\overline{\cal K}$, and $f_0:=\mu_0/b$. Then for any $n$,
\begin{equation}\begin{array}{lcl}
\!\!&\! \!&\! \ell_{n,T} (\vartheta )-\ell_T (\vartheta ) =\\%\nonumber\\
\!\!&\!=\!&\!\!\sum_{i=0}^{n-1}\!\!\int_{t_i}^{t_{i+1}}\!
(\frac{\mu (X_{t_i},\vartheta )}{b^2 (X_{t_i})}\! -\!
\frac{\mu (X_{t},\vartheta )}{b^2 (X_{t})})\, dX_t
-\frac{1}{2}\sum_{i=0}^{n-1}\!\!\int_{t_i}^{t_{i+1}}\!
(\frac{\mu^2\! (X_{t_i},\vartheta )}{b^2 (X_{t_i})}\! -\!
\frac{\mu^2\! (X_{t},\vartheta )}{b^2 (X_{t})})\, dt=\\%\nonumber\\
\!\!&\!=\!&\!\!
\sum_{i=0}^{n-1}\!\!\int_{t_i}^{t_{i+1}}\!\!( (f(X_t,\vartheta)\!-\!f(X_{t_i},\vartheta))f_0(X_t)
+f(X_{t_i},\vartheta)\left(\!\frac{b(X_t )}{b(X_{t_i})}\!-\!1\!\right)f_0(X_t))\,dt +\\%\nonumber\\
\!\!&\! \!&\!\! +\sqrt{\sigma}\sum_{i=0}^{n-1}\!\!\int_{t_i}^{t_{i+1}}\!\!( (f(X_t,\vartheta)\!-\!f(X_{t_i},\vartheta))
+f(X_{t_i},\vartheta)\!\left(\!\frac{b(X_t )}{b(X_{t_i})}\!-\!1\!\right)\!)\,dW_t -\\%\nonumber\\
\!\!&\! \!&\!\! -\frac{1}{2}\sum_{i=0}^{n-1}\!\!\int_{t_i}^{t_{i+1}}\!\!
(f^2(X_t,\vartheta)\!-\!f^2(X_{t_i},\vartheta))\, dt
\end{array}\label{lnld0}
\end{equation}
by the definitions of $\ell_T$ and $\ell_{n,T}$, and (\ref{e1}).
Let us assume for a moment that functions $f_0$, $b$, $b'$, $b''$, are bounded on $E$, and $f$ and
its partial derivatives $D^m_\vartheta f$, $\frac{\partial}{\partial x}D^m_\vartheta f$, and
$\frac{\partial^2}{\partial x^2}D^m_\vartheta f$ are bounded on $E\times\overline{\cal K}$ for $0\leq m\leq d+1$.
Then $f$ and $f^2$
satisfy condition (B2) from Theorem \ref{tm:new}, and $f_0$ and a constant function $1$ satisfy (B1), since {\sc (H2a-3a)} hold. Hence,
by Remark \ref{rem3} the statements of Theorem \ref{tm:new} holds for these functions, and any $T>0$.
By applying this conclusion to (\ref{lnld0}), the following holds:
\begin{equation}\begin{array}{l}
\|\sup_{\vartheta\in {\cal K}}|\ell_{n,T} (\vartheta )-\ell_T (\vartheta )|\|_{L^2(\mathbb {P}_{\theta_0})}
\leq C\sqrt{\delta_{n,T}},
\end{array}\label{L2bound}
\end{equation}
for any $T>0$ and subdivisions of $[0,T]$ with $\delta_{n,T}\leq h_0$, and a constant $C>0$ which depends on $T$, $X$ and ${\cal K}$.
Now, let $X$, $\mu$ and $b$ satisfy assumptions {\sc (H1a-3a)}, and let $x_0$ be the initial state of $X$.
Moreover, let $(E_m ,m\geq 1)$ be a sequence
of open and bounded subintervals of $E$ such that for all $m$, $\overline{E}_m\subset E_{m+1}$,
$x_0\in E_1$, and $\bigcup_{m=1}^{+\infty}E_m =E$, and let
$(\phi_m ,m\geq 1)$ be a sequence of
$C^{\infty}$-functions on $E$ such that for all $m$,
$0\leq\phi_m\leq 1$,
$\phi_m (x)=1$ for $x\in\overline{E}_m$ and $\phi_m\equiv0$ on
${E}_{m+1}^c$. Let us define the following bounded functions for each $m$:
$\mu_m (x,\vartheta ):=
\phi_m (x)\mu (x,\vartheta )$, $(x,\vartheta )\in E\times\Theta$, $b_m (x):=\phi_m (x)b(x)+c_m (1-\phi_m(x))$,
$x\in E$ where $c_m:=\hbox{\rm sign}\, b\cdot\max_{x\in\overline{E}_ {m+1}}|b(x)|$.
Since $\mu$ and $b$ satisfy {\sc (Ha2-a3)}, $b_m \in C^2(E)$, and $b_m$, $b_m'$, $b_m''$ are bounded on $E$,
and
$(x,\vartheta)\mapsto\mu_m (x,\vartheta )/b (x)$, $\mu_m^2 (x,\vartheta )/b^2 (x)$ satisfy (B2) and are bounded on $E\times\overline{\cal K}$, and hence satisfy (B3) too, for each $m$.
Moreover, let
$\tau_m :=\inf\{ t\geq 0 : X_t \in E_{m}^c \}$, $m\geq 1$.
Since $X$ is a continuous process, $(\tau_m ,m\geq 1)$
is an increasing sequence of stopping times (see \cite{Yor}) such that
$\tau_m\uparrow +\infty $ a.s., when $m\rightarrow +\infty$.
Let $m$ be fixed and let diffusion $X^m=(X^m_t; t\geq 0)$ be defined as solution to SDE:
\[\begin{array}{l}
X_t^m =x_0 +\int_0^t\mu_m (X_s^m,\vartheta_0 )\, ds+\sqrt{\sigma}\int_0^t b_m (X_s^m )\,
dW_s,\; t>0.
\end{array}\]
By Theorem V.11.2 in \cite{RW} (Vol.\ 2, p.\ 128) such a diffusion exists and is
a.s.\ unique.
Moreover, for almost all $\omega\in\Omega$ and $t\in [0, \tau_m (\omega )]$,
$X_t (\omega )=
X^{m}_t(\omega )$ by Corollary V.11.10 in \cite{RW} (Vol.\ 2, p.\ 131).
This implies (see \cite{Yor2}) that for an arbitrary number $A>0$,
\begin{equation}\begin{array}{lcl}
&&\!\mathbb {P}_{\theta_0}\{\sup_{\vartheta\in {\cal K}}|\ell_{n,T} (\vartheta )-
\ell_T (\vartheta )|>\! A\sqrt{\delta_{n,T}}\}
\leq \\
&\leq&\!\! \mathbb {P}_{\theta_0}\{ \tau_m\!\leq T\}+
\frac{1}{A\sqrt{\delta_{n,T}}}\|\sup_{\vartheta\in {\cal K}}|
\ell_{n,T}^m (\vartheta )-\ell_T^m (\vartheta )|\|_{L^2(\mathbb {P}_{\theta_0})}
\end{array}\label{ineqlln}
\end{equation}
where $\ell_T^m$ and $\ell_{n,T}^m$ are LLF (\ref{ell}) and its Euler approximation (\ref{lnizraz}) respectively, both
based on diffusion $X^m$ with drift $\mu_m (\cdot,\vartheta_0 )$, and diffusion coefficient function
$\sqrt{\sigma}b_m$. Now, (\ref{L2bound}) holds for functions $\ell_T^m$ and $\ell_{n,T}^m$ with constant $C=C_m$. Hence the righthand side
of (\ref{ineqlln}) is dominated by expression
$ \mathbb {P}_{\theta_0}\{ \tau_m\!\leq T\}+\frac{1}{A} C_m$.
First, let us take a limit when $n\rightarrow +\infty$, and then when $A\rightarrow +\infty$. Next, we take a limit when $m\rightarrow +\infty$, and hence we prove (\ref{drugac}).
\endproof\vspace{5pt}
{\em Proof of Theorem \ref{tm:30}.\/}
We need to show that the model and random
functions $\ell_T$ and $\ell_{n,T}$, $n\geq 1$, for fixed $T>0$, satisfy conditions (A1-5)
of Theorem 3.1 of \cite{huzakMLE}.
Let ${\cal F}_{n,T}$ be
$\sigma$-subalgebras of ${\cal F}_T^0$ that are introduced in
Section 4. We recall from the same section that $\ell_T$ is a ${\cal
F}_T^0\otimes {\cal B}(\Theta )$-measurable function. In the same
way, $\ell_{n,T}$ is ${\cal F}_{n,T}\otimes {\cal B}(\Theta )$-measurable,
for each $n$. Hence (A1) is satisfied. Corollary \ref{tm:23} implies
that functions $\ell_T$ and $\ell_{n,T}$, $n\geq 1$, satisfy (A3). The same corollary and {\sc (H5a)} imply (A4) and (A5). Condition (A2) is the same as assumption {\sc (H4a)}. Hence by Theorem 3.1 of \cite{huzakMLE} there exists a sequence of ${\cal F}_T^0$-measurable random vectors
$(\hat{\vartheta}_{n,T} ,n\geq 1)$ such that the statements of Theorem \ref{tm:30} hold.\endproof\vspace{5pt}
For proving Corollary \ref{cor:30a} we need the following lemma.
\begin{lemma}\label{lm:25}
Let {\sc (H1a-2a)} hold, and $T>0$ be fixed. Then for $\theta=(\vartheta,\sigma)\in\Psi$,
\begin{equation}\begin{array}{l}
\sum_{i=0}^{n-1}\frac{(\Delta_i X -\mu (X_{t_i},\vartheta )\Delta_it )^2}{
b^2 (X_{t_i})\Delta_it}-\sigma \sum_{i=0}^{n-1}\frac{(\Delta_i W)^2}{\Delta_it}=O_{\mathbb {P}_{\theta}}(1),\;\;n\rightarrow +\infty.
\end{array}\label{n1}
\end{equation}
\end{lemma
{\em Proof of Corollary \ref{cor:30a}.\/}
Notice that $(ii)$ implies the consistency (i.e.\ $(i)$) of
$\hat{\sigma}_n$. Let us prove $(ii)$. Since
\begin{equation}\begin{array}{l}
\sqrt{n}(\hat{\sigma}_{n,T} -\sigma)=\sqrt{n}(\hat{\sigma}_{n,T} -
\frac{\sigma}{n}\sum_{i=0}^{n-1}\frac{(\Delta_i W)^2}{\Delta_it})+\sigma\sqrt{2}
\cdot\frac{1}{\sqrt{2n}}\sum_{i=0}^{n-1}\frac{(\Delta_i W)^2 -\Delta_it}{\Delta_it
\end{array}\label{dekomp}
\end{equation}
and $({\cal L})\lim_n
\frac{1}{\sqrt{2n}}\sum_{i=0}^{n-1}\frac{(\Delta_i W)^2 -\Delta_it}{\Delta_it} =N(0,1)$,
for $(ii)$ to hold it is sufficient to prove
that for all $\epsilon >0$,
\begin{equation}\begin{array}{l}
\lim_n \mathbb {P}_{\theta}\{\sqrt{n}(\hat{\sigma}_{n,T} -
\frac{\sigma}{n}\sum_{i=0}^{n-1}\frac{(\Delta_i W)^2}{\Delta_it})\geq\epsilon\} =0.
\end{array}\label{doh}
\end{equation}
Let $\epsilon >0$ and $\eta >0$ be any numbers and let ${\cal K}$
be a relatively compact set in $\Theta$. If ${\cal K}+\eta:=\{\vartheta\in\Theta : (\exists \vartheta'\in {\cal K})\,
|\vartheta -\vartheta'|<\eta\}$ then
on event
\[\begin{array}{c}
A=
\{|\frac{1}{\sqrt{n}}(\sum_{i=0}^{n-1}\frac{(\Delta_i X -\mu (X_{t_i},
\vartheta)\Delta_it )^2}{b^2 (X_{t_i}) \Delta_it}-\sigma\sum_{i=0}^{n-1}
\frac{(\Delta_i W)^2}{\Delta_it})|<\frac{\epsilon}{5},\;
\hat{\vartheta}_T\in {\cal K}\}\cap \\ \cap
\{|\hat{\vartheta}_{n,T} -\hat{\vartheta}_T| <\eta,\;
|\ell_T
(\vartheta)-\ell_T (\hat{\vartheta}_T)|<\sqrt{n}\frac{\epsilon}{10},\;
|\ell_{n,T} (\vartheta)-\ell_T (\vartheta)|<\sqrt{n}\frac{\epsilon}{10}\}
\cap\\
\cap\{\sup_{\vartheta'\in {\cal K}+\eta }| D\ell_T (\vartheta' )| <
\frac{\sqrt{n}}{\eta}\frac{\epsilon}{10},\;
\sup_{\vartheta'\in
{\cal K}+\eta }|\ell_{n,T} (\vartheta' )-\ell_T (\vartheta' )|<
\sqrt{n}\frac{\epsilon}{10}\} ,
\end{array}\]
the following holds: $|\sqrt{n}(\hat{\sigma}_{n,T} -\frac{\sigma}{n}\sum_{i=0}^{n-1}\frac{(
\Delta_i W)^2}{\Delta_it})|<\epsilon$.
This implies that
$A\subseteq\{\sqrt{n}(\hat{\sigma}_{n,T}-\frac{\sigma}{n}\sum_{i=0}^{n-1}\frac{(
\Delta_i W)^2}{\Delta_it})<\epsilon\}$.
Hence
\[\begin{array}{rl}
&\mathbb {P}_{\theta}\{\sqrt{n}(\hat{\sigma}_{n,T} -\frac{\sigma }{n}
\sum_{i=0}^{n-1}\frac{(\Delta_i W)^2}{\Delta_it})\geq\epsilon\}\leq\\
\leq &\mathbb {P}_{\theta}\{|\frac{1}{\sqrt{n}}(\sum_{i=0}^{n-1}\frac{(\Delta_i X -
\mu (X_{t_i},\vartheta )\Delta_it )^2}{b^2 (X_{t_i}) \Delta_it}-\sigma\sum_{i=0}^{n-1}
\frac{(\Delta_i W)^2}{\Delta_it})|\geq\frac{1}{5}\epsilon\} +\\
&+\mathbb {P}_{\theta}\{\hat{\vartheta}_T\in {\cal K}^c\}+\mathbb {P}_{\theta}
\{ |\ell_T (\vartheta )-
\ell_T (\hat{\vartheta}_T)|\geq\sqrt{n}\frac{\epsilon}{10}\}
+\mathbb {P}_{\theta}\{ |\hat{\vartheta}_{n,T} -\hat{\vartheta}_T|\geq\eta\}+\\
&+\mathbb {P}_{\theta}\{\sup_{\vartheta'\in {\cal K}+\eta}| D\ell_T (\vartheta' )|
\geq\frac{\sqrt{n}}{\eta}\frac{\epsilon}{10}\} +\mathbb {P}_{\theta}\{
|\ell_{n,T}(\vartheta)-\ell_T(\vartheta )| \geq\sqrt{n}\frac{\epsilon}{10}\}
+\\ &+\mathbb {P}_{\theta}\{\sup_{\vartheta'\in {\cal K}+\eta}
|\ell_{n,T}(\vartheta' )-\ell_T (\vartheta' )|\geq
\sqrt{n}\frac{\epsilon}{10}\} .
\end{array}\]
By Lemma \ref{lm:25}, Corollary \ref{tm:23}, property $(ii)$ of
$\hat{\vartheta}_{n,T}$ from Theorem \ref{tm:30}, and arbitrariness of ${\cal K}$, (\ref{doh}) follows.
\endproof
\subsection{\em Ergodic case}
For all $T>0$ let $0=t_0<\cdots <t_n=T$, $n\in\mathbb {N}$, be equidistant subdivisions of $[0,T]$ such that
$\delta_{n,T}=T/n\rightarrow 0$ when $T\rightarrow +\infty$ and $n\rightarrow +\infty$.
We need the following corollary to Theorem \ref{tm:new}.
\begin{corollary
\label{tm:cor2}
Let $X$ be a diffusion such that {\sc(H1b-3b)} hold. Then for all $\theta_0=(\vartheta_0,\sigma)\in
\Psi$, $\pi_{\vartheta_0}$-a.s.\ nonrandom initial conditions, and $r=0,1,2$,
\begin{equation}\begin{array}{l}
\sup_{\vartheta\in\overline{\Theta}}|\frac{1}{T}D^{r}\ell_{n,T} (\vartheta )-\frac{1}{T}D^{r}\ell_T (\vartheta)|=
O_{\mathbb {P}_{\theta_0}}(\sqrt{\delta_{n,T}}),\;\;T\rightarrow +\infty,\; n\rightarrow +\infty.
\end{array}\label{drugac2}
\end{equation}
\end{corollary}
{\em Proof of Corollary \ref{tm:cor2}}. Similarly to the proof of Corolarlly \ref{tm:23} it is sufficient to prove (\ref{drugac2}) for $r=0$ since the statement of the corollary for cases $r=1$ and $r=2$ can be proved in the same way.
Let $\theta_0 =(\vartheta_0 ,\sigma)\in\Psi$ be arbitrary, and let
$\mu_0 :=\mu(\cdot,\vartheta_0)$, $\nu:=\sqrt{\sigma} b$, and $\mathbb {P}\equiv\mathbb {P}_{\theta_0}$, $\mathbb {E}\equiv \mathbb {E}_{\theta_0}$. Let us recall expression (\ref{lnld0}) from the proof of Corolarlly \ref{tm:23} where
$f(\cdot,\vartheta)=\mu(\cdot,\vartheta)/b$,
$\vartheta\in\overline{\Theta}$, and $f_0 =\mu_0/b$. Notice that $f$ and $f^2$ satisfy (B2) since {\sc (H2a-3a)} hold by {\sc (H2b-3b)}. Let us show
that $f_0$ satisfies (B1) and $f$ satisfies (B3) with respect to $a\equiv f_0$ and compact $\overline{\Theta}$, and that $f^2$ satisfies (B3) with respect to constant function $a\equiv 1$ and the same compact (notice that constant function trivially satisfies (B1)). If we fix $\vartheta\in\overline{\Theta}$, $m$ such that $0\leq m\leq d+1$, and
nonnegative integers $j_1$,..., $j_d$ such that $j_1+\cdots +j_d=m$ then let $\tilde{f}:=\frac{\partial^m}{\partial\vartheta_1^{j_1}\cdots\partial\vartheta_d^{j_d}}f(\cdot,\vartheta )$, and
$\tilde{\mu}:=\frac{\partial^m}{\partial\vartheta_1^{j_1}\cdots\partial\vartheta_d^{j_d}}\mu (\cdot,\vartheta )$. By {\sc (H3a)}, $\tilde{f},\tilde{\mu}\in C^2 (E)$. Since {\sc (H3b)} holds it follows that $|\tilde{f}|\leq g_0\in L^{32}(\pi_{\vartheta_0})\subset
L^{8}(\pi_{\vartheta_0})$, and
\[\begin{array}{lcl}
|\tilde{f}'b|&=&|\tilde{\mu}'-\tilde{f}b'|\leq g_1+g_0 |b'|=:g_{01}\in L^{16}(\pi_{\vartheta_0})\subset L^{8}(\pi_{\vartheta_0})\\
|\tilde{f}'\mu_0|&=&|(\tilde{f}'b)f_0|\leq g_{01}g_1 =:g_{02} \in L^{8}(\pi_{\vartheta_0})\\
|\tilde{f}''b^2|&=&|\tilde{\mu}'' b-2(\tilde{f}'b)b'-f(b''b)|\leq g_2+2g_{01}|b'|+g_0 |b''b|=:g_{03}\in L^{8}(\pi_{\vartheta_0})
\end{array}\]
by {\sc (H2b-3b)}. Then function $g_{00}:=g_0+\sqrt{\sigma}g_{01}+g_{02}+\sigma g_{03}$ is such that $g_{00}\in L^{8}(\pi_{\vartheta_0})\subset L^{4}(\pi_{\vartheta_0})$
and
\[\begin{array}{l}
\sup_{\vartheta\in\overline{\Theta}}\left(\!|D^m_\vartheta f (\cdot ,\vartheta )|_\infty\!+\!|\frac{\partial}{\partial x}D^m_\vartheta f (\cdot ,\vartheta )|_\infty(|\mu_0 |\! +\!|\nu |)\!+\!
|\frac{\partial^2}{\partial x^2}D^m_\vartheta f (\cdot ,\vartheta )\nu^2 |_\infty\right)\!\! \leq\! g_{00}
\end{array}\]
for all $0\leq m\leq d+1$. This implies that $f$ satisfies the first part of (B3) with $g\equiv g_{00}$. This also implies that $|f_0|+|\bar{f}_0|\leq g_{00}$ and hence $f_0$, $\bar{f}_0\in L^{8}(\pi_{\vartheta_0})$.
By Chacon-Ornstein theorem, ergodic theorem for additive functionals and its corollary (e.g. Theorem (A.5.2) on p. 504, Theorem (X.3.12)
on p. 397, and Exercise (X.3.18) on p. 399 in \cite{Yor}), for $\pi_{\vartheta_0}$-a.s. initial values $x_0\in E$,
\begin{equation}\begin{array}{lcl}
&&\lim_{T\rightarrow +\infty}\mathbb {E} (\frac{1}{T}\int_0^T f_0^8(X_t)\, dt)=
\lim_{n,T}\mathbb {E} (\frac{1}{T}\sum_{i=0}^{n-1}f_0^8(X_{t_i})\Delta_i t)=\\
&=&\lim_n\mathbb {E} (\frac{1}{n}\sum_{i=0}^{n-1}f_0^8(X_{t_i}))=
\int_E f_0^8 (x)\pi_{\vartheta_0}(dx)\leq \int_E g_{00}^8 (x)\pi_{\vartheta_0}(dx)<+\infty
\end{array}\label{f0}
\end{equation}
since {\sc (H1b)} holds, and subdivisions are equidistant ($\Delta_i t=T/n$ for each $i$). Moreover, since $f_0\in L^{4}(\pi_{\vartheta_0})$ too, the same holds for 4th powers of $f_0$, i.e. if we substitute $f_0^4$ instead of $f_0^8$
in (\ref{f0}). Finally, the both conclusions hold for $\bar{f}_0$ too. Hence $f_0$ satisfies (B1). It remains to show
that $g_{00}$ satisfies the limiting properties from (B3).
Using the same arguments as in proving (\ref{f0}) it follows that (\ref{f0}) holds for 8th and hence for 4th power of
$g_{00}$.
Moreover, since $f_0, g_{00}\in L^{8}(\pi_{\vartheta_0})$ implies $f_0g_{00}\in L^{4}(\pi_{\vartheta_0})$,
and (\ref{f0}) (with respect to $\bar{f}_0$ and $g_{00}$ too)
holds, it follows that
\[\begin{array}{lcl}
&
\lim_{n,T}\mathbb {E} (\frac{1}{T}\sum_{i=0}^{n-1}(f_0 g_{00})^4(X_{t_i})\Delta_i t) =
\int_E (f_0g_{00})^4 (x)\pi_{\vartheta_0}(dx)<+\infty,\\
&&\overline{\lim}_{n,T}\mathbb {E} (\frac{1}{T}\sum_{i=0}^{n-1}g_{00}^4(X_{t_i})\int_{t_i}^{t_{i+1}}(f_0^4 +\bar{f}_0^4)(X_t)\, dt \leq\\
&\leq& \frac{1}{2}\lim_{n,T}\!\mathbb {E} (\frac{1}{T}\!\sum_{i=0}^{n-1}\!\! g_{00}^8(X_{t_i})\Delta_i t)+
\lim_{T\rightarrow +\infty}\!\mathbb {E} (\frac{1}{T}\!\int_0^T\!\!(f_0^8 + \bar{f}_0^8) (X_t)dt)\!<\!+\infty.
\end{array}
\]
Hence $f$ satisfies (B3) for $\pi_{\vartheta_0}$-a.s. nonrandom initial conditions. It remains to show that $f^2$ satisfies
(B3) with respect to function $a\equiv 1$. Let $g:=7\cdot 2^{d+1}g_{00}^2\in L^{4}(\pi_{\vartheta_0})$.
Notice that uniformly with respect to $\vartheta\in\overline{\Theta}$,
\[\begin{array}{l}
|f^2|\!+\!|\frac{\partial}{\partial x}(f^2)|\!+\!|\frac{\partial^2}{\partial x^2}(f^2)|\leq
|f^2|\!+\!2|f\frac{\partial}{\partial x}f|\!+\!2|\left(\frac{\partial}{\partial x}f\right)^2\!+\!f\frac{\partial^2}{\partial x^2}f|\leq
7g_{00}^2\leq g.
\end{array}\]
Let us put $\hat{f}:=\frac{\partial^m}{\partial\vartheta_1^{j_1}\cdots\partial\vartheta_d^{j_d}}(f^2)(\cdot,\vartheta )$ for
fixed $\vartheta\in\overline{\Theta}$, $m$ such that $0\leq m\leq d+1$, and
nonnegative integers $j_1$,..., $j_d$ such that $j_1+\cdots +j_d=m$. Then by induction
\[\begin{array}{l}
|\hat{f}|+|\frac{\partial}{\partial x}\hat{f}|+|\frac{\partial^2}{\partial x^2}\hat{f}|\leq
7\cdot 2^m g_{00}^2\leq g.
\end{array}\]
Then
(\ref{f0}) (for 4th powers of $g_{00}$) implies that $f^2$ satisfies (B3) with respect to $a\equiv 1$, for $\pi_{\vartheta_0}$-a.s. nonrandom initial conditions. Finally, (B4) holds for $\pi_{\vartheta_0}$-a.s. nonrandom initial conditions since {\sc (H1b-H2b)} hold. Hence we
can apply Theorem \ref{tm:new} to (\ref{lnld0}) to conclude that there exists constants $C>0$, $T_0\geq 0$, and $n_0\in\mathbb {N}$,
such that for all $T>T_0$ and $n\geq n_0$, and arbitrary $A>0$,
\[\begin{array}{cl}
&\mathbb {P}_{\theta_0}\{\frac{1}{\sqrt{\delta_{n,T}}}
\sup_{\vartheta\in\overline{\Theta}}|\frac{1}{T}\ell_{n,T}(\vartheta )-\frac{1}{T}\ell_T (\vartheta )|\geq A \}\leq\\
\leq & \frac{1}{A^2}\mathbb {E}{\left(\frac{1}{\sqrt{\delta_{n,T}}}\sup_{\vartheta\in\overline{\Theta}}|\frac{1}{T}\ell_{n,T}(\vartheta )-\frac{1}{T}\ell_T (\vartheta )|\!\right)\!}^2\leq\frac{C}{A^2}.
\end{array}\]
Hence
\[\begin{array}{l}
\lim_{A\rightarrow +\infty}\overline{\lim}_{n,T}\mathbb {P}_{\theta_0}\{\frac{1}{\sqrt{\delta_{n,T}}}
\sup_{\vartheta\in\overline{\Theta}}|\frac{1}{T}\ell_{n,T}(\vartheta )-\frac{1}{T}\ell_T (\vartheta )|\geq A \} = 0
\end{array}\]
which proves the corollary. \endproof\vspace{5pt}
In order to prove Theorems \ref{tm:cont}-\ref{tm:erg} we need the following lemmas.
\begin{lemma}\label{lema1b}
Let {\sc (H1b-3b)} hold. Then for all $\theta_0=(\vartheta_0,\sigma)\in\Psi$ there exist constants $C_r>0$ ($r=0,1,2$)
such that $\mathbb {P}_{\theta_0}$-a.s. there exists $T_0>0$ such that for all $\vartheta_1,\vartheta_2\in\overline{\Theta}$, and all $T\geq T_0$,
\[\begin{array}{cl}
|\frac{1}{T}D^r\ell_T (\vartheta_1)-\frac{1}{T}D^r\ell_T (\vartheta_2)| &\leq C_r |\vartheta_1-\vartheta_2|,\; r=0,1,2,\\
|\frac{1}{T}D\ell_T (\vartheta_1)-\frac{1}{T}D\ell_T (\vartheta_2)-
\frac{1}{T}D^2\ell_T (\vartheta_2)(\vartheta_1-\vartheta_2)|&\leq\frac{1}{2} C_2 |\vartheta_1-\vartheta_2|^2,\; \mbox{ and}\\
\sup_{\vartheta\in\overline{\Theta}}\frac{1}{T}|D^3\ell_T (\vartheta)| & \leq C_2.
\end{array}\]
\end{lemma}
\begin{lemma}\label{lema2b}
Let {\sc (H1b-3b)} hold. Then for all $\theta_0=(\vartheta_0,\sigma)\in\Psi$, $\mathbb {P}_{\theta_0}$-a.s.
\[\begin{array}{l}
\lim_{T\rightarrow +\infty}\sup_{\vartheta\in\overline{\Theta}}|\frac{1}{T}\ell_T (\vartheta )-\ell_{\vartheta_0}(\vartheta )|=0.
\end{array}\]
\end{lemma}
{\em Proof of Theorem \ref{tm:cont}}. Let $\theta_0=(\vartheta_0,\sigma)\in\Psi$ be arbitrary.
Since $\Theta$ is an open set there exists $\varepsilon_0>0$ such that $K(\vartheta_0,\varepsilon_0)\subset\Theta$.
Let $\ell_{\vartheta_0}$ be function (\ref{liml0})
and let $
\lambda_0:=\min_{|y|=1}y^\tau I(\vartheta_0 )y=-\max_{|y|=1}y^\tau D^2 \ell_{\vartheta_0}(\vartheta_0 )y >0$
be the minimal eigenvalue of the Fisher information matrix $I(\vartheta_0)$ since it is positive definite by {\sc (H5b)}. Moreover, let $C_r>0$ ($r=0,1,2$) be constants from Lemma \ref{lema1b}, and
let $\Omega_0$ be an intersection of the events from Lemmas \ref{lema1b}-\ref{lema2b}, and the events such that (\ref{liml0}) and
(\ref{liml0dd2}) hold for $\vartheta_0$. Hence $\mathbb {P}_{\theta_0} (\Omega_0)=1$, and for $\omega\in\Omega_0$, let $T_0\equiv T_0 (\omega)>0$ be a such that the statements of Lemma \ref{lema1b} hold for $T\geq T_0$.
Let $\varepsilon >0$ be such that $\varepsilon\leq \varepsilon_0\wedge\lambda_0/(4 C_2)$. Then $K(\vartheta_0,\varepsilon)\subset\Theta$. Let $\omega\in\Omega_0$ be fixed.
Since (\ref{liml0dd2}) holds, there exists $T_1\geq T_0$ such that for all
$T\geq T_1$, $|\frac{1}{T}D^2\ell_T (\vartheta_0)-D^2\ell_{\vartheta_0} (\vartheta_0)| <\frac{\lambda_0}{4}$ and
$|\frac{1}{T}D\ell_T (\vartheta_0)-D\ell_{\vartheta_0} (\vartheta_0)| <\frac{\lambda_0}{4}\varepsilon$.
Then for all $y\in\mathbb {R}^d$, $|y|=1$, $T\geq T_1$, and $\vartheta\in K(\vartheta_0,\varepsilon)$,
\[\begin{array}{lcl}
y^\tau (\frac{1}{T}D^2\ell_T (\vartheta )) y &\leq &|\frac{1}{T}D^2\ell_T (\vartheta )-\frac{1}{T}D^2\ell_T (\vartheta_0)|+
|\frac{1}{T}D^2\ell_T (\vartheta_0)-D^2\ell_{\vartheta_0} (\vartheta_0)|+\\
&&+ y^\tau D^2\ell_{\vartheta_0} (\vartheta_0 ) y <C_2 |\vartheta-\vartheta_0|+\frac{\lambda_0}{4} -\lambda_0\leq\\
&\leq & C_2 \frac{\lambda_0}{4C_2}+\frac{\lambda_0}{4} -\lambda_0 =-\frac{\lambda_0}{2}.
\end{array}\]
Hence $\vartheta\mapsto\frac{1}{T}\ell_T(\vartheta)$ is a strictly concave function on $K(\vartheta_0,\varepsilon)$. Moreover,
if $z\in\mathbb {R}^d$ is such that $|z|=\varepsilon$, then for $y:=z/|z|$ and $T\geq T_1$,
\[\begin{array}{lcl}
\frac{1}{T}D\ell_T (\vartheta_0+z ) z &= &\frac{1}{T}D\ell_T (\vartheta_0 )z+z^\tau(\frac{1}{T}\!\!\int_0^1\!\!
D^2\ell_T (\vartheta_0+tz)\,dt)z\leq\\
&\leq&|\frac{1}{T}D\ell_T (\vartheta_0 )-D\ell_{\vartheta_0}(\vartheta_0)|\varepsilon +y^\tau(\frac{1}{T}\!\!\int_0^1\!\!
D^2\ell_T (\vartheta_0+tz)dt)y\varepsilon^2\leq\\
&\leq&\frac{\lambda_0}{4}\varepsilon^2 -\frac{\lambda_0}{2}\varepsilon^2 =-\frac{\lambda_0}{4}\varepsilon^2<0.
\end{array}\]
Then there exists $\hat{\vartheta}_T\in K(\vartheta_0,\varepsilon)$ such that $D\ell_T(\hat{\vartheta}_T)=\O$ (see e.g.
Lemma 4.3. in \cite{huzakMLE}), and $D^2\ell_T(\hat{\vartheta}_T)<\O$ since
$\min_{|y|=1}y^\tau (-\frac{1}{T}D^2\ell_T (\vartheta )) y\geq\frac{\lambda_0}{2}=\frac{1}{2}\min_{|y|=1}y^\tau I(\vartheta_0 )y$
for all $\vartheta\in K(\vartheta_0,\varepsilon)$ obviously. Since $\varepsilon >0$ is an arbitrary small number, these imply statement $(ii)$ of the theorem. Notice that $\hat{\vartheta}_T$ is the unique point of maximum of function $\ell_T$ on
$K(\vartheta_0,\varepsilon)$ since $\ell_T$ is strictly concave on this set. To finish the proof of statement $(i)$ we have to prove that there exists $T_2\geq T_1$ such that $\hat{\vartheta}_T$ is the unique point of global maximum of $\ell_T$ on $\overline{\Theta}$. Since for all $\vartheta\in\overline{\Theta}\setminus\{\vartheta_0\}$, $\ell_{\vartheta_0}(\vartheta_0 )>\ell_{\vartheta_0}(\vartheta )$, $\ell_{\vartheta_0} \in C(\overline{\Theta})$, and $\overline{\Theta}\setminus K(\vartheta_0,\varepsilon)$ is a compact set,
it follows that $\ell_{\vartheta_0}(\vartheta_0 )>\sup_{|y|\geq\varepsilon}\ell_{\vartheta_0}(\vartheta_0+y )$. By Lemma 4.4. in \cite{huzakMLE} there exists a number $0<s(\varepsilon)<\varepsilon$ such that
\[\begin{array}{l}
\Delta (\vartheta_0,\varepsilon):=\inf_{|x|\leq s(\varepsilon)}\ell_{\vartheta_0}(\vartheta_0+x )-\sup_{|y|\geq\varepsilon}\ell_{\vartheta_0}(\vartheta_0+y )>0.
\end{array}\]
Since Lemma \ref{lema2b} holds there exists $T_2\geq T_1$ such that for
$T\geq T_2$,
\[\begin{array}{l}
\sup_{\vartheta\in\overline{\Theta}}|\frac{1}{T}\ell_T(\vartheta )-\ell_{\vartheta_0}(\vartheta )|<\frac{\Delta (\vartheta_0,\varepsilon)}{4}.
\end{array}\]
If $x,y\in\mathbb {R}^d$ such that $|x|\leq s(\varepsilon )$ and $|y|\geq\varepsilon $ then
\[\begin{array}{lcl}
\frac{1}{T}\ell_T(\vartheta_0 +x) &=&\frac{1}{T}\ell_T(\vartheta_0 +x)-\ell_{\vartheta_0}(\vartheta_0 +x) +\ell_{\vartheta_0}(\vartheta_0 +x) -\ell_{\vartheta_0}(\vartheta_0 +y)+\\
&&+\ell_{\vartheta_0}(\vartheta_0 +y) -\frac{1}{T}\ell_T(\vartheta_0 +y)+
\frac{1}{T}\ell_T(\vartheta_0 +y)\geq\\
&\geq &-\frac{\Delta (\vartheta_0,\varepsilon)}{4}+\inf_{|x|\leq s(\varepsilon)}\ell_{\vartheta_0}(\vartheta_0+x )-\sup_{|y|\geq\varepsilon}\ell_{\vartheta_0}(\vartheta_0+y )+\\
&& -\frac{\Delta (\vartheta_0,\varepsilon)}{4}+\frac{1}{T}\ell_T(\vartheta_0 +y)\geq\\
&\geq &\frac{\Delta (\vartheta_0,\varepsilon)}{2}+\frac{1}{T}\ell_T(\vartheta_0 +y)
\end{array}\]
implying that
\begin{equation}\begin{array}{l}
\inf_{|x|\leq s(\varepsilon)}\frac{1}{T}\ell_{T}(\vartheta_0+x )-\sup_{|y|\geq\varepsilon}\frac{1}{T}\ell_{T}(\vartheta_0+y )\geq \frac{\Delta (\vartheta_0,\varepsilon)}{2} >0
\end{array}\label{deltapotez}\end{equation}
and hence $\ell_T(\vartheta_0)> \sup_{|y|\geq\varepsilon}\ell_T(\vartheta_0+y )$. Finally, $(i)$ follows. To prove statement $(iii)$,
first notice that
\begin{equation}\begin{array}{l}
\frac{1}{\sqrt{T}}D\ell_{T}(\vartheta_0)=\frac{\sqrt{\sigma}}{\sqrt{T}}\int_0^T\frac{1}{b(X_t)}D\mu_0 (X_t)\, dW_t\stackrel{{\cal L}-\mathbb {P}_{\theta_0}}{\longrightarrow}N(\O,\sigma I(\vartheta_0)),\; T\rightarrow +\infty
\end{array}\label{elnorm}\end{equation}
by Theorem 1 in \cite{BH} since {\sc (H1b-5b)} hold, and second notice that for $\bar{\vartheta} (s):=s\vartheta_T +(1-s)\vartheta_0$,
\begin{equation}\begin{array}{l}
D\ell_{T}(\hat{\vartheta}_T)\! =\! D\ell_{T}(\vartheta_0)\!+\!D^2\ell_{T}(\vartheta_0) (\hat{\vartheta}_T\!-\!\vartheta_0)\!+\! \int_0^1\!\!\int_0^1\!
D^3\ell_{T}(\bar{\vartheta}(st))\, ds\, tdt (\hat{\vartheta}_T\!-\!\vartheta_0)^2.
\end{array}\label{eltaylor}\end{equation}
Let $H_T(\vartheta_0):=\frac{1}{T}D^2\ell_{T}(\vartheta_0)+\frac{1}{T}\int_0^1\!\!\int_0^1\!
D^3\ell_{T}(\bar{\vartheta}(st))\, ds\,tdt (\hat{\vartheta}_T\!-\!\vartheta_0)$, and let us recall $\omega\in\Omega_0$ and $T_1=T_1(\omega )$ from the first part od the proof. Notice that $H_T(\vartheta_0)$ is a symmetric matrix. Then from Lemma \ref{lema1b}, for $T\geq T_1$,
\[\begin{array}{l}
|H_T(\vartheta_0)-\frac{1}{T}D^2\ell_T (\vartheta_0)|\leq\sup_{\vartheta\in\overline{\Theta}}|\frac{1}{2T}D^3\ell_{T}(\vartheta )|
|\hat{\vartheta}_T\!-\!\vartheta_0|\leq\frac{C_2}{2}|\hat{\vartheta}_T\!-\!\vartheta_0|
\end{array}\]
and hence, for $y\in\mathbb {R}^d$ such that $|y|=1$,
\[\begin{array}{l}
y^\tau H_T (\vartheta_0 ) y \leq |H_T(\vartheta_0)-\frac{1}{T}D^2\ell_T (\vartheta_0)|+y^\tau (\frac{1}{T}D^2\ell_T (\vartheta )) y\leq
-\frac{3\lambda_0}{8}
\end{array}\]
implying that $H_T (\vartheta_0 )$ is a negative definite matrix, and $|H_T (\vartheta_0 )^{-1}|\leq\frac{8}{3\lambda_0}$. Since
$|I(\vartheta_0)^{-1}|=1/\lambda_0$,
\[\begin{array}{cl}
&|H_T(\vartheta_0)^{-1}+I(\vartheta_0)^{-1}|\leq |H_T(\vartheta_0)^{-1}|\cdot |H_T(\vartheta_0)+I(\vartheta_0)|
\cdot |I(\vartheta_0)^{-1}| \leq\\
\leq & \frac{8}{3\lambda_0^2}(\frac{C_2}{2}|\hat{\vartheta}_T\!-\!\vartheta_0|+|\frac{1}{T}D^2\ell_T (\vartheta_0)-D^2\ell_{\vartheta_0}(\vartheta_0)|),
\end{array}\]
and $(ii)$ and (\ref{liml0dd2}) hold, it follows that $\mathbb {P}_{\theta_0}$-a.s.
\begin{equation}\begin{array}{l}
\lim_{T\rightarrow +\infty}H_T(\vartheta_0)^{-1}=-I(\vartheta_0)^{-1}.
\end{array}\label{limH}\end{equation}
Finally, since $D\ell_{T}(\hat{\vartheta}_T)=\O$ and $I(\vartheta_0)$ is nonrandom, (\ref{elnorm}-\ref{limH}) imply that
\[\begin{array}{l}
\sqrt{T}(\hat{\vartheta}_T-\vartheta_0)=-H_T(\vartheta_0)^{-1}\frac{1}{\sqrt{T}}D\ell_T(\vartheta_0)\stackrel{{\cal L}-\mathbb {P}_{\vartheta_0}}{\longrightarrow}
N(\O,\sigma I(\vartheta_0 )^{-1}),\; T\rightarrow +\infty.\;\mbox{\endproof}
\end{array}\]
{\em Proof of Theorem \ref{tm:erg}}. Let $\theta_0=(\vartheta_0,\sigma)\in\Psi$ be arbitrary, and let
$C_r>0$ ($r=0,1,2$) be constants from Lemma \ref{lema1b}. Moreover, let $\Omega_0$ be a $\mathbb {P}_{\theta_0}$-probability one
event from Lemmas \ref{lema1b}-\ref{lema2b} and Theorem \ref{tm:cont} ($i$-$ii$). Let $\omega\in\Omega_0$ be
fixed. Let $\varepsilon_0>0$ be a such number that $K(\vartheta_0,\varepsilon_0)\subset\Theta$, and let $\lambda_0>0$
be the minimal eigenvalue of Fisher matrix $I(\vartheta_0)$.
Then there exists $T_0=T_0(\omega)\geq 0$ such that for all $T>T_0$, $\hat{\vartheta}_T\in K(\vartheta_0,\varepsilon_0/2)$ and $\lambda_T:=\min_{|y|=1}y^\tau (-\frac{1}{T}D^2\ell_T(\hat{\vartheta}_T))y\geq\lambda_0/2>0$, and the statements of Lemma \ref {lema1b} hold.
Let $\varepsilon >0$ be an arbitrary small number such that $\varepsilon<
\frac{\varepsilon_0}{2}\wedge \frac{\lambda_0}{8C_2}$. Then $K(\hat{\vartheta}_T,\varepsilon)\subset K(\hat{\vartheta}_T,\varepsilon_0/2)\subset
K(\vartheta_0,\varepsilon_0)\subset\Theta$. Moreover, on event
\[\begin{array}{l}
\Omega_{n,T}:=\{ \sup_{\vartheta\in\overline{\Theta}}|\frac{1}{T}D^r\ell_{n,T}(\vartheta)-\frac{1}{T}D^r\ell_{T}(\vartheta)|\leq
\frac{\lambda_0}{8}(1\wedge\frac{\lambda_0}{8C_2}),\; r=1,2\},
\end{array}\]
for $\vartheta\in K(\hat{\vartheta}_T,\varepsilon)$ and $z\in\mathbb {R}^d$ such that $|z|=\varepsilon$, and $y:=z/|z|$,
the following holds:
\[\begin{array}{lcl}
y^\tau D^2\ell_{n,T} (\vartheta ) y &\leq &|D^2\ell_{n,T} (\vartheta )-D^2\ell_T (\vartheta)|\!+\!
|D^2\ell_T (\vartheta)-D^2\ell_{T} (\hat{\vartheta}_T)|\!+\\
&&+ y^\tau D^2\ell_{T} (\hat{\vartheta}_T) y <(\frac{\lambda_0}{4}+C_2 \frac{\lambda_0}{8C_2} -\frac{\lambda_0}{2})T =-\frac{\lambda_0}{4}T<0,\\
D\ell_{n,T} (\hat{\vartheta}_T+z ) z &= &D\ell_{n,T} (\hat{\vartheta}_T )z+z^\tau(\!\!\int_0^1\!\!
D^2\ell_{n,T}(\hat{\vartheta}_T+tz)\,dt) z\leq\\
&\leq&|D\ell_{n,T} (\hat{\vartheta}_T )-D\ell_T (\hat{\vartheta}_T )|\varepsilon +y^\tau(\int_0^1\!\!
D^2\ell_T (\hat{\vartheta}_T+tz)dt)y\varepsilon^2\leq\\
&\leq&\varepsilon\frac{\lambda_0}{8C_2}(\frac{\lambda_0}{8} -\frac{\lambda_0}{4})T =-\varepsilon\frac{\lambda_0}{8C_2}\frac{\lambda_0}{8}T<0.
\end{array}\]
Hence $\vartheta\mapsto\ell_{n,T}(\vartheta)$ is a strictly concave function on $K(\hat{\vartheta}_T,\varepsilon)$,
and there exists $\hat{\vartheta}_{n,T}\in K(\hat{\vartheta}_T,\varepsilon)$ such that $D\ell_{n,T}(\hat{\vartheta}_{n,T})=\O$, and $\hat{\vartheta}_{n,T}$ is the unique stationary point and a point of
maximum of $\ell_{n,T}$ at $K(\hat{\vartheta}_T,\varepsilon)$. These imply that $\hat{\vartheta}_{n,T}$ is a random vector. Since $\lim_{n,T}\mathbb {P}_{\theta_0}(\Omega_{n,T}^c)=0$ by Corollary \ref{tm:cor2},
and $\Omega_{n,T}\subset\{ D\ell_{n,T}(\hat{\vartheta}_{n,T})=\O\}\cap\{|\hat{\vartheta}_{n,T}-\hat{\vartheta}_T|<\varepsilon\}$, statements $(i)$ and $(ii)$ of the theorem follow. Moreover if process $(\tilde{\vartheta}_{n,T})$ satisfies $(i)$ and $(ii)$ then
statement $(iv)$ follows since
\[\begin{array}{l}
\Omega_{n,T}\cap\{D\ell_{n,T}(\tilde{\vartheta}_{n,T})=\O\}\cap\{|\tilde{\vartheta}_{n,T}-\hat{\vartheta}_T|<\varepsilon\}
\subseteq\{\hat{\vartheta}_{n,T}=\tilde{\vartheta}_{n,T}\}
\end{array}\]
by uniqness of a stationary point of $\ell_{n,T}$ on $K(\hat{\vartheta}_{T},\varepsilon)$. To prove $(iii)$, let $A>0$
be an arbitrary number, and let
$\Omega_{n,T}(A):=\{
\sup_{\vartheta\in\overline{\Theta}}|\frac{1}{T}D\ell_{n,T}(\vartheta)-\frac{1}{T}D\ell_{T}(\vartheta)|\leq
\frac{\lambda_0}{4}A\sqrt{\delta_{n,T}}\}$. Then on event $\Omega_{n,T}(A)\cap\Omega_{n,T}$,
\[\begin{array}{ccl}
|\hat{\vartheta}_{n,T}\!-\!\hat{\vartheta}_T|&\!\leq\! &|(D^2\ell_{T}(\hat{\vartheta}_T))^{-1}|\!\cdot \! |D^2\ell_{T}(\hat{\vartheta}_T)
(\hat{\vartheta}_{n,T}-\hat{\vartheta}_T)|\leq\\
&\!\leq\! &|(D^2\ell_{T}(\hat{\vartheta}_T))^{-1}|\!\cdot \!
|D\ell_{T}(\hat{\vartheta}_{n,T})\!-\!D\ell_{T}(\hat{\vartheta}_T)\!-\!D^2\ell_{T}(\hat{\vartheta}_T)
(\hat{\vartheta}_{n,T}\!-\!\hat{\vartheta}_T)|+\\
&&+|(D^2\ell_{T}(\hat{\vartheta}_T))^{-1}|\!\cdot \!|D\ell_{n,T}(\hat{\vartheta}_{n,T})-D\ell_{T}(\hat{\vartheta}_{n,T})|\leq\\
&\leq&\frac{2}{\lambda_0 T}\frac{C_2}{2}\frac{\lambda_0}{2C_2}T|\hat{\vartheta}_{n,T}-\hat{\vartheta}_T|+\frac{2}{\lambda_0 T}\frac{\lambda_0 T}{4}A\sqrt{\delta_{n,T}}\leq\\
&\leq &\frac{1}{2}|\hat{\vartheta}_{n,T}-\hat{\vartheta}_T|+\frac{1}{2}A\sqrt{\delta_{n,T}}\\
\Rightarrow & & |\hat{\vartheta}_{n,T}\!-\!\hat{\vartheta}_T|\leq A\sqrt{\delta_{n,T}}
\end{array}\]
by Lemma \ref{lema1b} and since $\varepsilon\leq\frac{\lambda_0}{2C_2}$. Hence $\Omega_{n,T}(A)\cap\Omega_{n,T}\subseteq\{
|\hat{\vartheta}_{n,T}\!-\!\hat{\vartheta}_T|\leq A\sqrt{\delta_{n,T}}\}$, and
\[\begin{array}{cl}
0&\leq\overline{\lim}_{A\rightarrow +\infty}\overline{\lim}_{n,T}\mathbb {P}_{\theta_0}\{|\hat{\vartheta}_{n,T}\!-\!\hat{\vartheta}_T|\leq A\sqrt{\delta_{n,T}}\}\leq\\
&\leq \lim_{A\rightarrow +\infty}\overline{\lim}_{n,T}\mathbb {P}_{\theta_0}(\Omega_{n,T}(A)^c)+\overline{\lim}_{n,T}\mathbb {P}_{\theta_0}(\Omega_{n,T}^c)=0
\end{array}\]
by Corollary \ref{tm:cor2}, and $(iii)$ follows. Consistency of $\hat{\vartheta}_{n,T}$ (the first part of statement $(v)$) follows directly from $(ii)$ and Theorem \ref{tm:cont} $(ii)$. To prove its asymptotic normality (the second part of $(v)$) notice that
\[\begin{array}{l}
|\sqrt{T}(\hat{\vartheta}_{n,T}\!-\!\vartheta_0)-\sqrt{T}(\hat{\vartheta}_{T}\!-\!\vartheta_0)|=
\sqrt{T\delta_{n,T}}\frac{1}{\sqrt{\delta_{n,T}}}|\hat{\vartheta}_{n,T}\!-\!\hat{\vartheta}_T|\stackrel{\mathbb {P}_{\theta_0}}{
\longrightarrow} 0
\end{array}\]
when $\lim_{n,T}T\delta_{n,T}=0$ since $(iii)$ holds. Then the second part of $(v)$ follows by Slutsky theorem since Theorem \ref{tm:cont} $(iii)$ holds. To prove statement $(vi)$, first we need to prove that
\begin{equation}\begin{array}{l}
\frac{1}{T}(\sum_{i=0}^{n-1}\!\frac{(\Delta_i X -\mu (X_{t_i},\vartheta )\Delta_it )^2}{
b^2 (X_{t_i})\Delta_it}-\sigma\! \sum_{i=0}^{n-1}\!\frac{(\Delta_i W)^2}{\Delta_it})=O_{\mathbb {P}_{\theta}}(1),\;T\rightarrow +\infty, n\rightarrow +\infty
\end{array}\label{opb}\end{equation}
for $\pi_{\vartheta_0}$-a.s. initial conditions.
This follows from Lemma \ref{difkoef}, the proof of Lemma \ref{lm:25}, and the fact that the functions $f:=\mu(\cdot,\vartheta)/b$ and $b$ satisfies (B1-4) which is proved in Corollary \ref{tm:cor2}. The proof of asymptotic normality of $\hat{\sigma}_{n,T}$ is the same as in the proof of Corollary \ref{cor:30a} since
\[\begin{array}{cl}
&\frac{1}{\sqrt{n}}(\sum_{i=0}^{n-1}\!\frac{(\Delta_i X -\mu (X_{t_i},\vartheta )\Delta_it )^2}{
b^2 (X_{t_i})\Delta_it}-\sigma\! \sum_{i=0}^{n-1}\!\frac{(\Delta_i W)^2}{\Delta_it})=\\
=&
\frac{\sqrt{T\delta_{n,T}}}{T}(\sum_{i=0}^{n-1}\!\frac{(\Delta_i X -\mu (X_{t_i},\vartheta )\Delta_it )^2}{
b^2 (X_{t_i})\Delta_it}-\sigma\! \sum_{i=0}^{n-1}\!\frac{(\Delta_i W)^2}{\Delta_it})\rightarrow 0
\end{array}\]
when $T\rightarrow +\infty$ such that $T\delta_{n,T}\rightarrow 0$, and since $(i-v)$, Corollary \ref{tm:cor2},
and Lemma \ref{lema2b} hold.
Similarly consistency of $\hat{\sigma}_{n,T}$
follows from decomposition (\ref{dekomp}) in the proof of Corollary \ref{cor:30a} (but without factor "$\sqrt{n}$") by
using (\ref{opb}) which appears with factor "$\delta_{n,T}$" (notice that $\delta_{n,T}/T =1/n$), and by the strong low of large numbers instead of CLT. In this case it is sufficient to assume that $\delta_{n,T}\rightarrow 0$ when $T\rightarrow +\infty$.
Finally, for proving ${\cal F}_{n,T}^0$-measurability of $\hat{\vartheta}_{n,T}$ (and hence $\hat{\sigma}_{n,T}$ too) it is sufficient to
prove that $\hat{\vartheta}_{n,T}$ is a unique point of maximum of $\ell_{n,T}$ on $\overline{\Theta}$. This proof
follows in the similar way as proof of uniqness of $\hat{\vartheta}_T$ as global point of maximum of $\ell_T$ on $\overline{\Theta}$ by replacing $\ell_T$ with $\ell_{n,T}$ and $\ell_{\vartheta_0}$ with $\ell_T$.\endproo
{\footnotesize
|
\section{#1}\inputtex{#2}}
\providecommand{\isubsection}[2] {\subsection{#1}\inputtex{#2}}
\providecommand{\etal} {\emph{et al\@.}\xspace}
\providecommand{\ie} {\emph{i.e\@.}\xspace}
\providecommand{\eg} {\emph{e.g\@.}\xspace}
\providecommand{\cf} {\emph{cf\@.}\xspace}
\providecommand{\defacto} {\emph{de facto}\xspace}
\providecommand{\adhoc} {\emph{ad hoc}\xspace}
\providecommand{\apriori} {\emph{a priori}\xspace}
\providecommand{\myurl}[1][] {\texttt{web.eecs.umich.edu/$\sim$fessler#1}\xspace}
\providecommand{\myweb} {\url{web.eecs.umich.edu/~fessler}\xspace}
\providecommand{\onweb}[1] {Available from \myurl.}
\let\comments=\comment
\let\endcomments=\endcomment
\long\def\comment#1{}
\providecommand{\uncomment}[1] {#1}
\providecommand{\bcent} {\begin{center}}
\providecommand{\ecent} {\end{center}}
\providecommand{\benum} {\begin{enumerate}}
\providecommand{\eenum} {\end{enumerate}}
\providecommand{\bitem} {\begin{itemize}}
\providecommand{\eitem} {\end{itemize}}
\providecommand{\bvers} {\begin{verse}}
\providecommand{\evers} {\end{verse}}
\providecommand{\btab} {\begin{tabbing}}
\providecommand{\etab} {\end{tabbing}}
\providecommand{\lfill} {\mbox{}\hfill}
\providecommand{\rfill} {\hfill\mbox{}}
\newcommand{\cent}[1] {\lfill{#1}\rfill}
\newcommand{\stacktext}[2][c] {\begin{tabular}{#1}#2\end{tabular}}
\newcommand{\fsbox}[2][c] {\fbox{\stacktext[#1]{#2}}}
\providecommand{\ul}[1] {\underline{#1}}
\providecommand{\ebox}[1] {\mbox{\fbox{$\displaystyle#1$}}}
\providecommand{\cbox}[1] {\[\ebox{#1}\]}
\newcounter{blist}
\providecommand{\blistmark} {\makebox[0pt]{$\bullet$}}
\providecommand{\blistitemsep} {0pt}
\providecommand{\blist}[1][] {%
\begin{list}{\blistmark}{%
\usecounter{blist}%
\setlength{\itemsep}{\blistitemsep}%
\setlength{\parsep}{0pt}%
\setlength{\parskip}{0pt}%
\setlength{\partopsep}{0pt}%
\setlength{\topsep}{0pt}%
\setlength{\leftmargin}{1.2em}%
\setlength{\labelsep}{0.5\leftmargin
\setlength{\labelwidth}{0em}%
#1
}
\providecommand{\elist} {\end{list}}
\providecommand{\elistup} {\elist\vspace{-\parskip}}
\providecommand{\blistitemsep} {0pt}
\providecommand{\bjfenum}[1][] {%
\begin{list}{\bcolor{\arabic{blist}.} }{%
\usecounter{blist}%
\setlength{\itemsep}{\blistitemsep}%
\setlength{\parsep}{0pt}%
\setlength{\parskip}{0pt}%
\setlength{\partopsep}{0pt}%
\setlength{\topsep}{0pt}%
\setlength{\leftmargin}{0.0em}%
\setlength{\labelsep}{1.0\leftmargin
\setlength{\labelwidth}{0pt}%
#1
}
\newcounter{blistAlph}
\providecommand{\blistAlph}[1][]
{\begin{list}{\makebox[0pt][l]{\Alph{blistAlph}.}}{%
\usecounter{blistAlph}%
\setlength{\itemsep}{0pt}\setlength{\parsep}{0pt}%
\setlength{\parskip}{0pt}\setlength{\partopsep}{0pt}%
\setlength{\topsep}{0pt}%
\setlength{\leftmargin}{1.2em}%
\setlength{\labelsep}{1.0\leftmargin
\setlength{\labelwidth}{0.0\leftmargin}#1}%
}
\newcounter{blistRoman}
\providecommand{\blistRoman}[1][]
{\begin{list}{\Roman{blistRoman}.}{%
\usecounter{blistRoman}%
\setlength{\itemsep}{0.5em}\setlength{\parsep}{0pt}%
\setlength{\parskip}{0pt}\setlength{\partopsep}{0pt}%
\setlength{\topsep}{0pt}%
\setlength{\leftmargin}{4em}%
\setlength{\labelsep}{0.4\leftmargin}
\setlength{\labelwidth}{0.6\leftmargin}#1}%
}
\providecommand{\myheadheight} {0in}
\providecommand{\myheadsep} {0in}
\providecommand{\Halfheight} {5.5in}
\providecommand{\Halfwidth} {4.25in}
\providecommand{\pagesize}[3][\parindent]{
\setlength{\parindent}{#1}
\setlength{\textwidth}{#2}
\setlength{\textheight}{#3}
\setlength{\headsep}{\myheadsep}
\setlength{\headheight}{\myheadheight}
\setlength{\oddsidemargin}{\Halfwidth}
\addtolength{\oddsidemargin}{-0.5\textwidth}
\addtolength{\oddsidemargin}{-1in}
\setlength{\evensidemargin}{\oddsidemargin}
\setlength{\topmargin}{\Halfheight}
\addtolength{\topmargin}{-0.5\textheight}
\addtolength{\topmargin}{-0.5\headheight}
\addtolength{\topmargin}{-0.5\headsep}
\addtolength{\topmargin}{-1in}
}
\providecommand{\pagesizeland}[3][\parindent]{
\setlength{\parindent}{#1}
\setlength{\textwidth}{#2}
\setlength{\textheight}{#3}
\setlength{\headsep}{0.0in}
\setlength{\headheight}{0.0in}
\setlength{\oddsidemargin}{5.5in}
\addtolength{\oddsidemargin}{-0.5\textwidth}
\addtolength{\oddsidemargin}{-1in}
\setlength{\evensidemargin}{\oddsidemargin}
\setlength{\topmargin}{4.25in}
\addtolength{\topmargin}{-0.5\textheight}
\addtolength{\topmargin}{-1in}
}
\providecommand{\pagehead}{%
\setlength{\headsep}{\topmargin}%
\addtolength{\headsep}{1in}%
\setlength{\headsep}{0.5\headsep}%
\setlength{\headheight}{\headsep}%
\setlength{\topmargin}{-1in}%
}
\section{Introduction}
\label{intro}
First-order methods are favorable
for solving large-scale problems
because their computational complexity per iteration
depends mildly on the problem dimension.
In particular, Nesterov's fast gradient method (FGM)
\cite{nesterov:83:amf,nesterov:05:smo}
achieves the optimal worst-case rate $O(1/N^2)$
for decreasing smooth convex functions after $N$ iterations
\cite{nesterov:04},
and thus has been widely used in (large-scale) applications.
Recently,
the optimized gradient method (OGM)~\cite{drori:14:pof,kim:16:ofo,kim:17:otc}
was found to achieve the optimal worst-case cost function bound
of first-order methods
(with either fixed-step or adaptive-step approaches)
for smooth convex minimization in~\cite{drori:17:tei},
whereas FGM achieves that bound only up to constant.\footnote{
There is a backtracking line-search version of FGM~\cite{nesterov:04}
that also achieves the optimal worst-case function bound up to constant,
which is sometimes more useful than the fixed-step FGM in practice.
However, such backtracking line-search version of OGM
with a fast worst-case bound
is yet unknown (unlike the fixed-step OGM~\cite{drori:14:pof,kim:16:ofo,kim:17:otc}),
while recently an exact line-search version of OGM
is developed in~\cite{drori:18:efo}.
}
Building upon~\cite{drori:14:pof,kim:16:ofo,kim:17:otc},
this paper presents two different ways of generalizing OGM and its development.
First, this paper specifies a parameterized family of
algorithms that generalizes OGM,
and provides worst-case bounds
on the function and gradient norm values
for this family.
Like the generalized forms of FGM~\cite{chambolle:15:otc,nesterov:05:smo}
being widely used and studied
(\eg,~\cite{attouch:18:fco,chambolle:15:otc,su:16:ade}),
we believe introducing the generalized OGM here
can be potentially useful.
Second, this paper optimizes
the step coefficients of
fixed-step first-order methods
with respect to the rate of decrease of the cost function's gradient norm,
leading to a new algorithm called OGM-OG (OG for optimized over a gradient).
This development expands the choice of worst-case rate metrics for
optimizing first-order methods in~\cite{drori:14:pof,kim:16:ofo,kim:17:otc}
that focused on the cost function decrease leading to OGM.
We next briefly review the Performance Estimation Problem (PEP)~\cite{drori:14:pof}
that was used in~\cite{drori:14:pof,kim:16:ofo,kim:17:otc} to develop OGM
and that we extensively use throughout the paper.
Drori and Teboulle~\cite{drori:14:pof}
cast a worst-case analysis
into an optimization problem
called PEP\footnote{
The original PEP was intractable to solve,
so a series of relaxation on the PEP was introduced in~\cite{drori:14:pof}
to make it possibly solvable,
which we review in Sec.~\ref{sec:pep,cost,relax}.}
\cite{drori:14:pof}
that examines the maximal absolute cost function inaccuracy
over all possible inputs (cost functions) to the optimization algorithm.
(See \eg,~\cite{deklerk:17:otw,drori:18:efo,drori:16:aov,kim:16:ofo,kim:17:otc,kim:18:ala,kim:18:ote-arxiv,lessard:16:aad,taylor:17:ewc-composite,taylor:17:ssc} for its extensions.)
Moreover, Drori and Teboulle~\cite{drori:14:pof}
optimized \emph{numerically} the step coefficients of first-order methods using PEP
for smooth convex minimization,
and found an algorithm
whose worst-case bound is lower than that of FGM,
but it required too much computation and memory to be appealing
for large-scale problems.
Building on their work,
the authors~\cite{kim:16:ofo,kim:17:otc}
found computationally and memory-wise efficient version,
called OGM,
and showed analytically that OGM satisfies
an analytical worst-case bound
that is twice smaller than that of FGM.
Drori~\cite{drori:17:tei} showed that the OGM
is optimal
for large-dimensional smooth convex minimization
over a general class of first-order methods
with either fixed or adaptive step sizes~\cite{drori:17:tei}.
This OGM has been numerically extended for nonsmooth composite convex problems
in~\cite{taylor:17:ewc-composite}.
In addition, this OGM-type algorithm was already studied
in the context of a proximal point method
\cite[Appendix]{guler:92:npp}.
Using the PEP approach~\cite{drori:14:pof},
this paper proposes
a generalized version of OGM (GOGM)
and analyzes its worst-case rates
in terms of \emph{both}
the decrease of the cost function
and the decrease of the norm
of the gradient of the cost function.
The results complement the worst-case analysis
of the OGM~\cite{kim:16:ofo,kim:17:otc},
and expands our understanding of OGM-type first-order methods.
This paper analyzes
the worst-case rate of the gradient norm
(in addition to that of the cost function)
because it
is important
when dealing with dual problems,
considering that
the dual gradient norm
corresponds to the primal distance to feasibility
(see \eg,~\cite{devolder:12:dst,necoara:16:ica,nesterov:12:htm}).
While FGM has not been shown previously
to satisfy a rate $O(1/N^{1.5})$
for decreasing the gradient norm,
modified versions of FGM with such rate were studied
in~\cite{ghadimi:16:agm,monteiro:13:aah,nesterov:12:htm}.
This paper proves that FGM in fact does have that rate,
building upon~\cite{taylor:17:ssc}
that numerically conjectured such rate for FGM
using the gradient norm version of PEP.
For further acceleration of the worst-case gradient norm rate,
we optimize the step coefficients of first-order methods
with respect to the gradient norm using PEP
and propose an algorithm named OGM-OG
that belongs to the GOGM family
and has the best known analytical worst-case bound
on rate of decrease of the gradient norm
among fixed-step first-order methods.
One can extend some aspects
of the approaches for generalizing OGM described in this paper
to other optimization algorithms and problems.
One direction we have already taken in~\cite{kim:18:ala}
aims to improve the fast iterative shrinkage/thresholding algorithm (FISTA)
\cite{beck:09:afi}
(that reduces to FGM for smooth convex problems)
for nonsmooth composite convex problems.
Naturally, this paper and~\cite{kim:18:ala}
use some similar approaches,
but they are different;
the methods in~\cite{kim:18:ala}
when simplified to the smooth case correspond to a generalization of FGM
that differs from the GOGM.
Another direction we have recently taken in
\cite{kim:18:ote-arxiv}
focuses on optimizing the step coefficients of first-order methods
with respect to the gradient norm
under the initial bounded function condition
that is different from the initial bounded distance condition
used in this paper.
Sec.~\ref{sec:prob}
defines the smooth convex problem
and the first-order methods.
Sec.~\ref{sec:conv} reviews and discusses worst-case analyses
of a gradient method (GM), FGM, and OGM
for both the function value and the gradient norm.
Sec.~\ref{sec:conv}
also reviews first-order methods that guarantee an $O(1/N^{1.5})$ rate
for the gradient decrease.
Sec.~\ref{sec:pep,cost} reviews
the cost function form of PEP~\cite{drori:14:pof}
and reviews how the OGM~\cite{kim:16:ofo}
is derived using such PEP.
Sec.~\ref{sec:pep,cost} then proposes a generalized version of OGM (GOGM)
using the cost function form of PEP,
and Sec.~\ref{sec:pep,sgnorm}
provides a worst-case gradient norm bound for the GOGM
using the gradient form of PEP.
Then, Sec.~\ref{sec:pep,sgnorm}
optimizes the step coefficients using the gradient form of PEP
and proposes the OGM-OG that belongs to the GOGM family.
Sec.~\ref{sec:pep,sgnorm}
also proves that FGM decreases the gradient norm
with a rate $O(1/N^{1.5})$.
Sec.~\ref{sec:disc}
and Sec.~\ref{sec:conc} provide discussion and conclusion.
\section{Smooth convex problem and first-order methods}
\label{sec:prob}
\subsection{Smooth convex problem}
We focus on the following smooth convex minimization problem
\begin{align}
\min_{\x\in\Reals^d} \;&\; f(\x)
\label{eq:prob} \tag{M}
,\end{align}
where the following additional conditions are assumed:
\begin{itemize}[leftmargin=40pt]
\item
$f\;:\;\Reals^d\rightarrow\Reals$
is a convex function of the type \cC, i.e., continuously differentiable with
Lipschitz continuous gradient:
\begin{align}
||\nabla f(\x) - \nabla f(\y)|| \le L||\x - \y||, \quad \forall \x, \y \in \Reals^d,
\end{align}
where $L > 0$ is the Lipschitz constant.
\item
The optimal set $X_*(f)=\argmin{\x\in\Reals^d} f(\x)$ is nonempty,
\ie, the problem~\eqref{eq:prob} is solvable.
\end{itemize}
We use \cF to denote the class of functions that satisfy the above conditions.
We also assume that
the distance between an initial point $\x_0$
and an optimal solution $\x_* \in X_∗(f)$ is bounded by
some $R > 0$, i.e.,
\begin{align}
||\x_0 - \x_*|| \le R
.\end{align}
\subsection{First-order methods}
To solve~\eqref{eq:prob},
we consider the following class of
\emph{fixed-step} (or non-apdative-step)
first-order methods (\FO),
where the update step at $(i+1)$th iteration
is a weighted sum of the previous and current gradients
$\{\nabla f(\x_k)\}_{k=0}^i$
scaled by $\frac{1}{L}$
with fixed constant step coefficients $\{h_{i+1,k}\}_{k=0}^i$
that are not adaptive to the given $f$ and $\x_0$ (and thus $L$ and $R$).
This class~\FO includes GM, FGM, OGM,
and the methods proposed in this paper,
but excludes line-search-type methods.
\fbox{
\begin{minipage}[t]{0.85\linewidth}
\vspace{-10pt}
\begin{flalign}
&\quad \text{\bf Algorithm Class~\FO} & \nonumber \\
&\qquad \text{Input: } f\in \cF,\; \x_0\in\Reals^d. & \nonumber \\
&\qquad \text{For } i = 0,\ldots,N-1 & \nonumber \\
&\qquad \qquad \x_{i+1} = \x_i - \frac{1}{L}\sum_{k=0}^i \hkip \nabla f(\x_k). & \nonumber
\end{flalign}
\end{minipage}
} \vspace{5pt}
\section{Review of the worst-case analysis of~\FO}
\label{sec:conv}
This section reviews the worst-case analysis of existing~{\FO}s
(and simple variants thereof)
in terms of bounds on the cost function and gradient norm.
Sec.~\ref{sec:conv_fv} reviews
the worst-case cost function decrease of GM, FGM, and OGM.
Sec.~\ref{sec:grad}
presents both~{\FO}s (including GM, FGM, OGM and some variants)
that have either $O(1/N)$ or $O(1/N^{1.5})$ rate
for the worst-case gradient decrease;
it also reviews an $O(1/N^2)$ lower bound of the worst-case rates of first-order methods
for decreasing the gradient norm.
\subsection{Function value worst-case analysis of~\FO}
\label{sec:conv_fv}
The simplest example of a~\FO is the following GM
that uses only the current gradient and the Lipschitz constant $L$
for the update.
\fbox{
\begin{minipage}[t]{0.85\linewidth}
\vspace{-10pt}
\begin{flalign}
&\quad \text{\bf Algorithm GM} & \nonumber \\
&\qquad \text{Input: } f\in \cF,\; \x_0\in\Reals^d. & \nonumber \\
&\qquad \text{For } i = 0,1,\ldots & \nonumber \\
&\qquad \qquad \x_{i+1} = \x_i - \frac{1}{L}\nabla f(\x_i) & \nonumber
\end{flalign}
\end{minipage}
} \vspace{5pt}
\noindent
This GM monotonically decreases the cost function~\cite{nesterov:04}
and satisfies the following tight\footnote{
A \emph{tight} worst-case bound denotes
an inequality
where the equality holds for some function $f$.
For example,~\cite[Thm.~2]{drori:14:pof}
shows the bound~\eqref{eq:fv_gm} is tight.
}
worst-case bound~\cite[Thm.~1]{drori:14:pof}, for any $i\ge 0$,
\begin{align}
f(\x_i) - f(\x_*) \le \frac{LR^2}{4i + 2}
\label{eq:fv_gm}
.\end{align}
Among the class~\FO, the following two equivalent forms
of FGM~\cite{nesterov:83:amf,nesterov:05:smo}
have been used widely
because they decrease the cost function
with the optimal rate $O(1/N^2)$.
\begin{center}
\fbox{
\small
\begin{minipage}[t]{0.45\textwidth}
\vspace{-10pt}
\begin{flalign}
&\text{\bf Algorithm FGM1} & \nonumber \\
&\text{Input: } f\in \cF,\; \x_0=\y_0\in\Reals^d, & \nonumber \\
& \hspace{30pt}
t_0 = 1. & \nonumber \\
&\text{For } i = 0,1,\ldots & \nonumber \\
&\quad \y_{i+1} = \x_i - \frac{1}{L}\nabla f(\x_i) & \nonumber \\
&\quad t_{i+1} = \frac{1+\sqrt{1+4t_i^2}}{2}, & \nonumber \\
&\quad \x_{i+1} = \y_{i+1}
+ \frac{t_i - 1}{t_{i+1}}(\y_{i+1} - \y_i) \nonumber
\end{flalign}
\end{minipage}\vline\hspace{8pt
\begin{minipage}[t]{0.45\textwidth}
\vspace{-10pt}
\begin{flalign}
&\text{\bf Algorithm FGM2} & \nonumber \\
&\text{Input: } f\in \cF,\; \x_0=\y_0\in\Reals^d, & \nonumber \\
& \hspace{30pt}
t_0 = 1. & \nonumber \\
&\text{For } i = 0,1,\ldots & \nonumber \\
&\quad \y_{i+1} = \x_i - \frac{1}{L}\nabla f(\x_i) & \nonumber \\
&\quad \z_{i+1} = \x_0 - \frac{1}{L}\sum_{k=0}^it_k \nabla f(\x_k) & \nonumber \\
&\quad t_{i+1} = \frac{1+\sqrt{1+4t_i^2}}{2}, & \nonumber \\
&\quad \x_{i+1} = \paren{1 - \frac{1}{t_{i+1}}}\y_{i+1}
+ \frac{1}{t_{i+1}}\z_{i+1} & \nonumber
\end{flalign}
\end{minipage}
} \vspace{5pt}
\end{center}
\noindent
Specifically, the FGM1 and FGM2 iterates satisfy
the following worst-case cost function bounds
\cite{kim:16:ofo,nesterov:83:amf,nesterov:05:smo} for any $i\ge 1$:
\begin{align}
f(\y_i) - f(\x_*) \le \frac{LR^2}{2t_{i-1}^2}
\le \frac{2LR^2}{(i+1)^2},
\quad\text{and}\quad
f(\x_i) - f(\x_*) \le \frac{LR^2}{2t_i^2}
\le \frac{2LR^2}{(i+2)^2}
\label{eq:fv_fgm}
,\end{align}
where the parameter $t_i$ satisfies
\begin{align}
t_i^2 = \sum_{l=0}^i t_l
\quad\text{and}\quad
t_i \ge \frac{i+2}{2}
\quad\text{ for all } i.
\label{eq:t_rule}
\end{align}
A generalized form of FGM in~\cite{chambolle:15:otc}
uses parameters $t_i$
satisfying $t_0=1$ and $t_i^2 \le t_{i-1}^2 + t_i$,
including the choice $t_i = \frac{i+a}{a}$ for any $a\ge2$.
There is another generalized form of FGM in~\cite{nesterov:05:smo},
and these generalized forms of FGM
have been widely used and studied
(\eg,~\cite{attouch:18:fco,chambolle:15:otc,su:16:ade}).
Similarly this paper studies generalizations of
the OGM.
Building upon~\cite{drori:14:pof}
that optimized numerically the step coefficients over the cost function form of PEP,
the authors~\cite{kim:16:ofo} developed
the following two equivalent forms of OGM,
as reviewed in Sec.~\ref{sec:pep,cost}.
\begin{center}
\fbox{
\small
\begin{minipage}[t]{0.45\textwidth}
\vspace{-10pt}
\begin{flalign*}
&\text{\bf Algorithm OGM1} & \\
&\text{Input: } f\in \cF,\; \x_0=\y_0\in\Reals^d, & \\
& \hspace{30pt} \theta_0 = 1. & \\
&\text{For } i = 0,\ldots,N-1 & \\
&\quad \y_{i+1} = \x_i - \frac{1}{L}\nabla f(\x_i) & \\
&\quad \theta_{i+1}
= \begin{cases}
\frac{1+\sqrt{1+4\theta_i^2}}{2}, & i \le N-2 \\
\frac{1+\sqrt{1+8\theta_i^2}}{2}, & i = N-1
\end{cases} & \\
&\quad \x_{i+1} = \y_{i+1}
+ \frac{\theta_i - 1}{\theta_{i+1}}(\y_{i+1} - \y_i) & \\
&\hspace{70pt}
+ \frac{\theta_i}{\theta_{i+1}}(\y_{i+1} - \x_i) &
\end{flalign*}
\end{minipage}\vline\hspace{8pt
\begin{minipage}[t]{0.45\textwidth}
\vspace{-10pt}
\begin{flalign*}
&\text{\bf Algorithm OGM2} & \\
&\text{Input: } f\in \cF,\; \x_0=\y_0\in\Reals^d, & \\
& \hspace{30pt} \theta_0 = 1. & \\
&\text{For } i = 0,\ldots,N-1 & \\
&\quad \y_{i+1} = \x_i - \frac{1}{L}\nabla f(\x_i) & \\
&\quad \z_{i+1} = \x_0 - \frac{1}{L}\sum_{k=0}^i2\theta_k \nabla f(\x_k) & \\
&\quad \theta_{i+1}
= \begin{cases}
\frac{1+\sqrt{1+4\theta_i^2}}{2}, & i \le N-2 \\
\frac{1+\sqrt{1+8\theta_i^2}}{2}, & i = N-1
\end{cases} & \\
&\quad \x_{i+1} = \paren{1 - \frac{1}{\theta_{i+1}}}\y_{i+1}
+ \frac{1}{\theta_{i+1}}\z_{i+1} &
\end{flalign*}
\end{minipage}
} \vspace{5pt}
\end{center}
\noindent
The OGM iterates
satisfy the following worst-case
cost function bounds~\cite{kim:16:ofo,kim:17:otc}:
\begin{align}
f(\y_i) - f(\x_*) &\le \frac{LR^2}{4\theta_{i-1}^2}
\le \frac{LR^2}{(i+1)^2},
\label{eq:fv_ogm_}
\end{align}
for any $1\le i\le N$, and
\begin{align}
f(\x_N) - f(\x_*) &\le \frac{LR^2}{2\theta_N^2}
\le \frac{LR^2}{(N+1)(N+1+\sqrt{2})}
\label{eq:fv_ogm}
.\end{align}
The parameter sequence $\theta_i$ satisfies
\begin{align}
\theta_i^2 = \begin{cases}
\sum_{l=0}^{i+1}\theta_l, & i \le N-1, \\
2\sum_{l=0}^{N-1}\theta_l + \theta_N, & i = N,
\end{cases}
\quad\text{and}\quad
\theta_i \ge \begin{cases}
\frac{i+2}{2}, & i \le N-1, \\
\frac{i+1}{\sqrt{2}}, & i = N,
\end{cases}
\label{eq:theta_rule}
\end{align}
which is equivalent to $t_i$~\eqref{eq:t_rule}
except at the final iteration.
The bounds~\eqref{eq:fv_ogm_} and~\eqref{eq:fv_ogm} of OGM
are about twice smaller than the bounds~\eqref{eq:fv_fgm} of FGM,
so OGM decreases the cost function faster than FGM
in the worst case
(and often in practice~\cite{kim:15:aof}).
In addition, the bound~\eqref{eq:fv_ogm} on the final iterate $\x_N$
is tight and
satisfies the optimal worst-case bound of
general first-order methods
including both~\FO and adaptive-step first-order methods,
when the condition $d\ge N+1$ holds~\cite{drori:17:tei}.
The additional term $\frac{\theta_i}{\theta_{i+1}}(\y_{i+1} - \x_i)$ of OGM1
and the additional constant $2$ for the update of $\z_i$ of OGM2,
compared to FGM1 and FGM2 respectively,
along with the parameter $\theta_N$,
are what make
OGM optimal (for $d \ge N+1$).
One of the main goals of this paper is to
generalize the form of OGM
and analyze the worst-case rate of such generalized OGM
in terms of both the function value and the gradient norm,
complementing the bounds~\eqref{eq:fv_ogm_} and~\eqref{eq:fv_ogm}
on the function value of OGM.
The next section studies the worst-case rate of the gradient of~\FO.
\subsection{Gradient norm worst-case analysis of~\FO}
\label{sec:grad}
When tackling dual problems,
it is known that the gradient norm worst-case rate is important
in addition to the function value worst-case rate
because the dual gradient norm
is related to the primal distance to feasibility
(see \eg, \cite{devolder:12:dst,necoara:16:ica,nesterov:12:htm}).
One simple way to find a (loose) worst-case bound for the gradient norm
is to use the well-known convex inequality
for convex functions with $L$-Lipschitz continuous gradients~\cite{nesterov:04}:
\begin{align}
\frac{1}{2L}||\nabla f(\x)||^2
\le f(\x) - f\paren{\x - \frac{1}{L}\nabla f(\x)}
\le f(\x) - f(\x_*), \quad \forall \x\in\Reals^d
\label{eq:g_upper}
,\end{align}
as discussed in~\cite{nesterov:04,taylor:17:ssc}.
Combining the bounds~\eqref{eq:fv_gm},~\eqref{eq:fv_fgm}
and the inequality~\eqref{eq:g_upper},
for any $i\ge1$,
the GM iterates satisfy
\begin{align}
||\nabla f(\x_i)|| \le \sqrt{2L(f(\x_i) - f(\x_*))} \le \frac{LR}{\sqrt{2i+1}}
\label{eq:gg_gm}
,\end{align}
and the iterates of FGM satisfy
\begin{align}
||\nabla f(\y_i)|| \le \frac{LR}{t_{i-1}} \le \frac{2LR}{i+1},
\quad\text{and}\quad
||\nabla f(\x_i)|| \le \frac{LR}{t_i} \le \frac{2LR}{i+2}
\label{eq:g_fgm}
.\end{align}
Similarly for any $1\le i\le N$,
the OGM iterates with the bounds~\eqref{eq:fv_ogm_} and~\eqref{eq:fv_ogm}
satisfy
\begin{align}
||\nabla f(\y_i)|| \le \frac{LR}{\sqrt{2}\theta_{i-1}} \le \frac{\sqrt{2}LR}{i+1},
\quad\text{and}\quad
||\nabla f(\x_N)|| \le \frac{LR}{\theta_N} \le \frac{\sqrt{2}LR}{N+1}
\label{eq:g_ogm}
.\end{align}
Unfortunately, using the inequality~\eqref{eq:g_upper}
provides at best an $O(1/N)$ bound
due to the optimal rate $O(1/N^2)$ of the function decrease.
Furthermore, in general using~\eqref{eq:g_upper}
need not lead to tight worst-case bounds
on the gradient norm.
Using a different approach,
a smaller $O(1/N)$ worst-case bound for the gradient norm of GM
was derived in~\cite{nesterov:12:htm},
as reviewed in next section.
While the bounds~\eqref{eq:gg_gm},~\eqref{eq:g_fgm}, and~\eqref{eq:g_ogm}
are not guaranteed to be tight,
the next section shows that
the worst-case gradient bound~\eqref{eq:g_ogm}
on the final iterate $\x_N$ of OGM
is in fact tight
and thus has the same disappointingly slow $O(1/N)$ worst-case bound
on the gradient norm as GM.
\subsubsection{\FO with rate $O(1/N)$ for decreasing the gradient norm}
This section uses the following lemma
stating that GM monotonically decreases the gradient.
\begin{lemma}
\label{lem:gmono}
\cite[Lemma~2.4]{necoara:16:ica}
The GM monotonically decreases the gradient norm, i.e.,
\begin{align}
\bigg|\bigg|\nabla f\paren{\x - \frac{1}{L}\nabla f(\x)}\bigg|\bigg| \le ||\nabla f(\x)||
\label{eq:g_desc}
.\end{align}
\end{lemma}
The following theorem reviews
a simple proof in~\cite{nesterov:12:htm}
that provides a worst-case gradient norm bound for GM
with rate $O(1/N)$
that is smaller than~\eqref{eq:gg_gm},
where~\cite[Thm.~6.1]{necoara:16:ica}
additionally considers Lemma~\ref{lem:gmono}.
\begin{theorem}
\cite[Thm.~6.1]{necoara:16:ica},~\cite{nesterov:12:htm}
Let $f\;:\;\Reals^d\rightarrow\Reals$ be $\cF$
and let $\x_0,\cdots,\x_N\allowbreak\in\Reals^d$ be
generated by GM. Then for any $N\ge1$,
\begin{align}
\min_{i\in\{0,\ldots,N\}} ||\nabla f(\x_i)|| = ||\nabla f(\x_N)||
\le \frac{\sqrt{2}LR}{\sqrt{N(N+2)}}
\label{eq:g_gm}
.\end{align}
\begin{proof}
Lemma~\ref{lem:gmono}
implies the first equality in~\eqref{eq:g_gm}.
Using~\eqref{eq:fv_gm},~\eqref{eq:g_upper},~\eqref{eq:g_desc} yields:
\begin{align*}
\frac{LR^2}{4m+2}
&\stackrel{\eqref{eq:fv_gm}}{\ge} f(\x_m) - f(\x_*)
\stackrel{\eqref{eq:g_upper}}{\ge}
f(\x_{N+1}) - f(\x_*) + \frac{1}{2L}\sum_{i=m}^N||\nabla f(\x_i)||^2 \\
&\stackrel{\eqref{eq:g_desc}}{\ge} \frac{N-m+1}{2L} ||\nabla f(\x_N)||^2
,\end{align*}
which is equivalent to~\eqref{eq:g_gm}
using $m = \fNh$ for which $m \ge \frac{N-1}{2}$ and $N-m \ge \frac{N}{2}$.
\end{proof}
\end{theorem}
Inspired by the conjecture in~\cite[Sec.~4.1.3]{taylor:17:ssc},
the following theorem shows that
the $O(1/N)$ rate of the worst-case gradient norm bound~\eqref{eq:g_gm} of GM
is tight up to a constant.
\begin{theorem}
Let $\x_0,\cdots,\x_N \in \Reals^d$ be generated by
GM. Then for any $N\ge1$,
\begin{align}
\frac{LR}{N+1}
\le \max_{\substack{f\in\cF, \\ \x_* \in X_∗(f), \\ ||\x_0 - \x_*||\le R}}\min_{i\in\{0,\ldots,N\}} ||\nabla f(\x_i)||
= \max_{\substack{f\in\cF, \\ \x_* \in X_∗(f), \\ ||\x_0 - \x_*||\le R}}||\nabla f(\x_N)||
\label{eq:g_gm_low}
,\end{align}
where the
inequality in~\eqref{eq:g_gm_low} is achieved
by the following function in \cF:
\begin{align}
\psi(\x) = \begin{cases}
\frac{LR}{N+1}||\x|| - \frac{LR^2}{2(N+1)^2}, & ||\x|| \ge \frac{R}{N+1}, \\
\frac{L}{2}||\x||^2, & ||\x|| < \frac{R}{N+1}.
\end{cases}
\end{align}
\begin{proof}
Lemma~\ref{lem:gmono}
implies the equality in~\eqref{eq:g_gm_low}.
Starting from $\x_0 = R\bmnu$, where $\bmnu$ is a unit vector,
the GM iterates are
\begin{align*}
\x_i = \paren{1 - \frac{i}{N+1}}R\bmnu, \quad
\nabla \psi(\x_i) = \frac{LR}{N+1}\bmnu, \quad i=0,\ldots,N,
\end{align*}
which implies the inequality~\eqref{eq:g_gm_low}.
\end{proof}
\end{theorem}
We next show that the bound~\eqref{eq:g_ogm}
for the gradient norm at the final iterate $\x_N$ of OGM is tight
and and that its worst-case function is a simple quadratic function.
Note that OGM was derived by optimizing a worst-case bound
on the cost function decrease
and its behavior in terms of gradient norms
was not investigated previously.
\begin{theorem}
\label{thm:g_ogm_tight}
Let $\x_0,\cdots,\x_N\in\Reals^d$ be
generated by OGM. Then for any $N\ge1$,
\begin{align}
\max_{\substack{f\in\cF, \\ \x_* \in X_∗(f), \\ ||\x_0 - \x_*||\le R}} \min_{i\in\{0,\ldots,N\}}
||\nabla f(\x_i)||
= \max_{\substack{f\in\cF, \\ \x_* \in X_∗(f), \\ ||\x_0 - \x_*||\le R}} ||\nabla f(\x_N)||
= \frac{LR}{\theta_N}
\;\bigg(\! \le \frac{\sqrt{2}LR}{N+1}\bigg)
\label{eq:g_ogm_tight}
,\end{align}
where the worst-case function in \cF for OGM in terms of the gradient norm is
the quadratic function
$\phi(\x) = \frac{L}{2}||\x||^2$.
\begin{proof}
See Appendix~\ref{appen1}.
\end{proof}
\end{theorem}
Comparing~\eqref{eq:g_gm} and~\eqref{eq:g_ogm_tight},
we see that GM and OGM
have essentially similar
worst-case gradient norm bounds.
This is a dilemma
because OGM is the fastest~\FO in terms of the worst-case cost function bound,
but is as slow as GM in terms of the worst-case gradient norm bound.
Therefore, one of the main goals of this paper
is to study optimizing the step coefficients of~\FO
using PEP with respect to the gradient norm
in Sec.~\ref{sec:opt,grad,pep}.
We next discuss
the specific~\FO in~\cite{nesterov:12:htm}
that decreases the gradient norm
with a faster $O(1/N^{1.5})$ rate.
\subsubsection{\FO with rate $O(1/N^{1.5})$
for decreasing the gradient norm}
Searching for a~\FO that decreases the gradient norm faster
than the $O(1/N)$ rate of GM (and OGM),
Nesterov~\cite{nesterov:12:htm}
(among other variants of FGM~\cite{ghadimi:16:agm,monteiro:13:aah})
considered performing
FGM for the first \m iterations,
and GM for the remaining iterations.
He showed that this method,
which we denote FGM-\m,
satisfies a fast rate $O(1/N^{1.5})$
for decreasing the gradient norm.
In~\cite{devolder:12:dst,necoara:16:ica,nesterov:12:htm},
FGM-\m for $m=\fNh$ was used to solve dual problems.
To pursue a faster worst-case rate (in terms of the constant factor),
we consider here another variant
that performs OGM for the first \m iterations
and GM for the remaining iterations,
which we denote OGM-\m.
\noindent
\hspace{5pt}
\fbox{
\small
\begin{minipage}[t]{0.85\linewidth}
\vspace{-10pt}
\begin{flalign*}
&\quad \text{\bf Algorithm OGM-\m} & \\
&\qquad \text{Input: } f\in \cF,\; \x_0=\y_0\in\Reals^d,\;\vartheta_0 = 1,\;
m\in\{1,\ldots,N-1\}. & \\
&\qquad \text{For } i = 0,\ldots,\m - 1 & \\
&\qquad \qquad \y_{i+1} = \x_i - \frac{1}{L}\nabla f(\x_i) & \\
&\qquad \qquad \vartheta_{i+1} = \begin{cases}
\frac{1+\sqrt{1+4\vartheta_i^2}}{2}, & i\le\m - 2 \\
\frac{1+\sqrt{1+8\vartheta_i^2}}{2}, & i=\m - 1 \\
\end{cases}
& \\
&\qquad \qquad \x_{i+1}
=
\y_{i+1} + \frac{\vartheta_i-1}{\vartheta_{i+1}}(\y_{i+1} - \y_i)
+ \frac{\vartheta_i}{\vartheta_{i+1}}(\y_{i+1} - \x_i)
& \\
&\qquad \text{For } i = \m,\ldots,N-1 & \\
&\qquad \qquad \x_{i+1} = \x_i - \frac{1}{L}\nabla f(\x_i) &
\end{flalign*}
\end{minipage}
}
\noindent
The following theorem bounds the gradient norm of the OGM-\m iterates,
inspired by the proof in~\cite{necoara:16:ica,nesterov:12:htm}
for the worst-case gradient norm bound of the FGM-\m iterates.
The worst-case bound of FGM-\m
in~\cite{necoara:16:ica,nesterov:12:htm}
is asymptotically $\sqrt{2}$-times larger than
the following new bound~\eqref{eq:g_ogmh}
for OGM-\m.
\begin{theorem}
\label{thm:ogmh}
Let $f\;:\;\Reals^d\rightarrow\Reals$ be $\cF$
and let $\x_0,\cdots\x_N\in\Reals^d$ be
generated by OGM-\m for $1 \le m \le N-1$. Then for any $N\ge1$,
\begin{align}
\min_{i\in\{0,\ldots,N\}} ||\nabla f(\x_i)||
\le ||\nabla f(\x_N)||
\le \frac{\sqrt{2}LR}{(m+1)\sqrt{N-m+1}}
\label{eq:g_ogmh}
.\end{align}
\begin{proof}
Using~\eqref{eq:fv_ogm},~\eqref{eq:g_upper},~\eqref{eq:g_desc} yields:
\begin{align*}
\frac{LR^2}{2\vartheta_m^2}
&\stackrel{\eqref{eq:fv_ogm}}{\ge} f(\x_m) - f(\x_*)
\stackrel{\eqref{eq:g_upper}}{\ge}
f(\x_{N+1}) - f(\x_*) + \frac{1}{2L}\sum_{i=m}^N||\nabla f(\x_i)||^2 \\
&\stackrel{\eqref{eq:g_desc}}{\ge} \frac{N-m+1}{2L}||\nabla f(\x_N)||^2
,\end{align*}
which is equivalent to~\eqref{eq:g_ogmh}
using
$\vartheta_m \ge \frac{m+1}{\sqrt{2}}$
that is implied by~\eqref{eq:theta_rule}.
\end{proof}
\end{theorem}
The bound~\eqref{eq:g_ogmh}
is minimized
at a point close to $m=\fNt$,
leading to its (approximately) smallest constant $\frac{3\sqrt{6}}{2}$
with the rate $O(1/N^{1.5})$.
Other variants of FGM
having $O(1/N^{1.5})$ worst-case gradient bounds
were derived in~\cite{ghadimi:16:agm,monteiro:13:aah}.
Such variations of FGM (including FGM-$m$) were derived
since, prior to this paper, it was unknown
whether or not FGM
decreases the gradient norm with the rate $O(1/N^{1.5})$;
this rate for the gradient norm of FGM
was conjectured numerically in~\cite{taylor:17:ssc}.
Sec.~\ref{sec:pep,sgnorm,fgm} below
uses the PEP to show for the first time
the rate $O(1/N^{1.5})$ for the gradient decrease of the FGM.
The bound~\eqref{eq:g_ogmh} of OGM-\m for decreasing the gradient
is smaller than the bounds for the FGM variants
in~\cite{ghadimi:16:agm,monteiro:13:aah},
and
Sec.~\ref{sec:opt,grad,pep} below shows that our proposed methods
have worst-case bounds even lower than~\eqref{eq:g_ogmh}.
The preceding sections have focused on tight or upper worst-case bounds
of the gradient norm decrease of first-order methods,
whereas the next section
reviews a lower bound for the worst-case gradient norm decrease in
\cite{nemirovsky:92:ibc},
illustrating the best achievable worst-case rate
of the gradient norm decrease
for any first-order method
(with either fixed-step or adaptive-step approaches).
\subsubsection{A lower bound of the worst-case rates of
first-order methods
for decreasing the gradient norm}
\label{sec:low,grad}
For completeness, this section reviews
a lower bound on the worst-case rate of any first-order method
in terms of the gradient norm values
for smooth convex \emph{quadratic} functions~\cite{nemirovsky:92:ibc}.
Lower bounds on the function value
were studied
for convex \emph{quadratic} functions in~\cite{nemirovsky:92:ibc},
and for smooth convex functions
in~\cite{drori:17:tei,nesterov:04}.
When the condition $d\ge 2N+3$ holds,
a worst-case gradient norm bound
of any first-order method generating $\x_N$ after $N$ iterations
has rate $O(1/N^2)$ at best, for convex quadratic $f$,
\ie,
has the following lower bound~\cite[Sec.~2.3.B]{nemirovsky:92:ibc}:
\begin{align}
\frac{LR}{4e^2(N+1)^2}
\le \max_{\substack{f\in\cQ, \\ \x_* \in X_∗(f), \\ ||\x_0 - \x_*||\le R}}||\nabla f(\x_N)||
\label{eq:low}
,\end{align}
where
$
\cQ := \left\{f\;:\;
f(\x) \xmath{\equivsave} \frac{1}{2}\x^\top\Q\x + \p^\top\x + \rr
\text{ for } \x\in\reals^d,\;
\Q\succeq\Zero,\; ||\Q||\le L
\right\}
.$
Since $\cQ\subset\cF$, the lower bound~\eqref{eq:low}
for convex quadratic functions
also applies to smooth convex functions.
A regularization technique in~\cite{nesterov:12:htm}
achieves the rate $O(1/N^2)$ up to a logarithmic factor.
However, its adaptive step coefficients
require knowing $R$ in advance which is undesirable in practice.
To our knowledge,
whether there exists any \FO satisfying such rate
is an open question.
Instead, this paper discusses a way to develop \FO
that achieves an $O(1/N^{1.5})$ gradient norm bound
with the smallest constant
among known \FO.
\section{Relaxation and optimization of the cost function form of PEP}
\label{sec:pep,cost}
This section reviews
a relaxation of the cost function form of PEP~\cite{drori:14:pof} and
reviews how~\cite{drori:14:pof,kim:16:ofo,kim:17:otc}
optimized the step coefficients of the~\FO class
over the cost function form of PEP,
leading to OGM.
Then, we
propose a parameterized family of algorithms
that generalizes OGM,
and analyze the worst-case cost function decrease of the generalized OGM family.
\subsection{Review: Relaxation for the cost function form of PEP}
\label{sec:pep,cost,relax}
The worst-case bound on the cost function
for a~\FO having given step coefficients $\bmh:=\{h_{i+1,k}\}$
corresponds to a solution of the following PEP problem~\cite[Prob.~(P)]{drori:14:pof}:
\begin{align}\mathcal{B}_{\mathrm{P}}(\bmh,N,d,L,R) :=\;
& \max_{\substack{f\in\cF, \\
\x_0,\cdots,\x_N\in\Reals^d, \\
\x_*\in X_*(f), \\
||\x_0 - \x_*|| \le R}}
f(\x_N) - f(\x_*)
\label{eq:PEP} \tag{P} \\
&\st \; \x_{i+1} = \x_i - \frac{1}{L}\sum_{k=0}^i \hkip \nabla f(\x_k),
\quad i=0,\ldots,N-1. \nonumber
\end{align}
Since problem~\eqref{eq:PEP} is impractical to solve
due to its functional constraint
$f\in\cF$,
\cite{drori:14:pof} relaxed it by
the following finite set of inequalities
satisfied by $f$~\cite[Thm.~2.1.5]{nesterov:04}:
\begin{align}
\frac{1}{2L}||\nabla f(\x_i) - \nabla f(\x_j)||^2
\le f(\x_i) - f(\x_j) - \inprod{\nabla f(\x_j)}{\x_i - \x_j}
\label{eq:ineq}
\end{align}
for $i,j=0,1,\ldots,N,*$.
Then, a matrix
$\G = [\g_0,\cdots,\g_N]^\top \in \Reals^{(N+1)\times d}$
and a vector $\del = [\delta_0,\cdots,\delta_N]^\top \in \Reals^{N+1}$
with
\begin{align*}
\g_i := \frac{1}{L||\x_0 - \x_*||}\nabla f(\x_i),
\text{ and }
\delta_i := \frac{1}{L||\x_0 - \x_*||^2}(f(\x_i) - f(\x_*))
\end{align*}
are introduced
to represent gradient vectors and function values respectively
in the set of~\eqref{eq:ineq}.
This leads to
a finite-dimensional relaxation of problem~\eqref{eq:PEP}
\cite[Prob.~(Q)]{drori:14:pof}:
\begin{align}\mathcal{B}_{\mathrm{P1}}(\bmh,N,d,L,R) :=\;
& \max_{\substack{\G\in\Reals^{(N+1)\times d}, \\
\del\in\Reals^{N+1}}}
LR^2\delta_N
\label{eq:relPEP} \tag{P1} \\
&\st \; \Tr{\G^\top\A_{i,j}(\bmh)\G} \le \delta_i - \delta_j,
\quad i<j=0,\ldots,N, \nonumber \\
&\quad\;\;\; \Tr{\G^\top\B_{i,j}(\bmh)\G} \le \delta_i - \delta_j,
\quad j<i=0,\ldots,N, \nonumber \\
&\quad\;\;\; \Tr{\G^\top\C_i\G} \le \delta_i,
\quad i=0,\ldots,N, \nonumber \\
&\quad\;\;\; \Tr{\G^\top\D_i(\bmh)\G + \bmnu\u_i^\top\G} \le - \delta_i,
\quad i=0,\ldots,N, \nonumber
\end{align}
for any given unit vector $\bmnu\in\Reals^d$,
where $\u_i = \e_{i+1} \in\Reals^{N+1}$
is the $(i+1)$th standard basis vector.
Note that
$\Tr{\G^\top\u_i\u_j^\top\G} = \inprod{\g_i}{\g_j}$ by definition.
The matrices $\A_{i,j}(\bmh), \B_{i,j}(\bmh), \C_i, \D_i(\bmh)$ are defined as
\begin{align}
\begin{cases}
\A_{i,j}(\bmh) :=
\frac{1}{2}(\u_i - \u_j)(\u_i - \u_j)^\top
+ \frac{1}{2}\sum_{l=i+1}^j\sum_{k=0}^{l-1}
\hkl (\u_j\u_k^\top + \u_k\u_j^\top), & \\
\B_{i,j}(\bmh) :=
\frac{1}{2}(\u_i - \u_j)(\u_i - \u_j)^\top
- \frac{1}{2}\sum_{l=j+1}^i\sum_{k=0}^{l-1}
\hkl (\u_j\u_k^\top + \u_k\u_j^\top), & \\
\C_i := \frac{1}{2}\u_i\u_i^\top, & \\
\D_i(\bmh) :=
\frac{1}{2}\u_i\u_i^\top + \frac{1}{2}\sum_{j=1}^i\sum_{k=0}^{j-1}
\hkj(\u_i\u_k^\top + \u_k\u_i^\top). &
\end{cases}
\label{eq:ABCDF}
\end{align}
In~\cite[Prob.~(Q$'$)]{drori:14:pof},
problem~\eqref{eq:relPEP} is further relaxed
by discarding some constraints to yield
\begin{align}\mathcal{B}_{\mathrm{P2}}(\bmh,N,d,L,R) :=\;
& \max_{\substack{\G\in\Reals^{(N+1)\times d}, \\
\del\in\Reals^{N+1}}}
LR^2\delta_N
\label{eq:pPEP} \tag{P2} \\
&\st \; \Tr{\G^\top\A_{i-1,i}(\bmh)\G} \le \delta_{i-1} - \delta_i,
\quad i=1,\ldots,N, \nonumber \\
&\quad\;\;\; \Tr{\G^\top\D_i(\bmh)\G + \bmnu\u_i^\top\G} \le - \delta_i,
\quad i=0,\ldots,N, \nonumber
\end{align}
for any given unit vector $\bmnu\in\Reals^d$.
We explicitly illustrate the relaxation
from~\eqref{eq:relPEP} to~\eqref{eq:pPEP}
because Sec.~\ref{sec:pep,sgnorm}
uses a similar but different relaxation.
Taylor~\etal~\cite{taylor:17:ssc} avoided this step
to analyze a tight worst-case bound of~\eqref{eq:PEP}
(under a large-scale condition $d\ge N+2$
\cite[Thm.~5]{taylor:17:ssc});
however, this relaxation~\eqref{eq:pPEP}
facilitates the analysis in~\cite{drori:14:pof,kim:16:ofo,kim:17:otc} and in this paper.
Replacing $\max_{\G,\del} LR^2\delta_N$
by $\min_{\G,\del} \{-\delta_N\}$ for convenience in~\eqref{eq:pPEP},
the Lagrangian of the corresponding constrained minimization problem
with dual variables $\bmlam = (\lambda_1,\cdots,\lambda_N)^\top \in\Reals^N$
and $\bmtau = (\tau_0,\cdots,\tau_N)^\top\in\Reals^{N+1}$
for the first and second constraint inequalities of~\eqref{eq:pPEP}
respectively becomes
\begin{align}
\cL(\G,\del,\bmlam,\bmtau;\bmh)
=& - \delta_N + \sum_{i=1}^N\lambda_i(\delta_i - \delta_{i-1})
+ \sum_{i=0}^N\tau_i\delta_i \\
&\qquad\qquad\qquad\qquad + \Tr{\G^\top\S(\bmh,\bmlam,\bmtau)\G + \bmnu\bmtau^\top\G},
\nonumber
\end{align}
where
\begin{align}
\S(\bmh,\bmlam,\bmtau) := \sum_{i=1}^N \lambda_i\A_{i-1,i}(\bmh)
+ \sum_{i=0}^N\tau_i\D_i(\bmh).
\label{eq:S}
\end{align}
Then, we have the following dual problem of~\eqref{eq:pPEP}
that one could use to compute a valid upper bound of~\eqref{eq:PEP}
using a semidefinite program (SDP)
for given $\bmh$
\cite[Prob.~(DQ$'$)]{drori:14:pof}:
\begin{align}
\mathcal{B}_{\mathrm{D}}(\bmh,N,L,R) :=\;
&\min_{\substack{(\bmlam,\bmtau)\in\Lambda, \\ \gamma\in\Reals}}
\left\{
\frac{1}{2}LR^2\gamma\;:\;
\left(\begin{array}{cc}
\S(\bmh,\bmlam,\bmtau) & \frac{1}{2}\bmtau \\
\frac{1}{2}\bmtau^\top & \frac{1}{2}\gamma
\end{array}\right)
\succeq\Zero
\right\}
\label{eq:D} \tag{D}
,\end{align}
where
\begin{align}
\Lambda = \left\{(\bmlam,\bmtau)\in\Reals_+^{2N+1}
\;:\;
\begin{array}{l}
\tau_0 = \lambda_1,\;\; \lambda_N + \tau_N = 1, \\
\lambda_i - \lambda_{i+1} + \tau_i = 0, \; i=1,\ldots,N-1,
\end{array}
\right\}
\label{eq:Lam}
.\end{align}
The next section reviews the analytical solution to this upper bound~\eqref{eq:D}
for OGM,
instead of using a numerical SDP solver.
\subsection{Review: Optimizing step coefficients
for the cost function form of PEP}
\label{sec:pep,cost,opt}
Drori and Teboulle~\cite{drori:14:pof} optimized numerically
the step coefficients $\bmh$ over
the simple SDP problem~\eqref{eq:D} as follows
\cite[Prob.~(BIL)]{drori:14:pof}:
\begin{align}
\hat{\bmh}_{\mathrm{D}}
:=
\argmin{\bmh\in\Reals^{N(N+1)/2}} \mathcal{B}_{\mathrm{D}}(\bmh,N,L,R)
\tag{HD}
\label{eq:HD}
.\end{align}
The problem~\eqref{eq:HD} is bilinear,
and \cite[Thm.~3]{drori:14:pof}\footnote{
\cite[Thm.~3]{drori:14:pof} has typos
that are fixed in~\cite[Eq. (6.3)]{kim:16:ofo}.}
used a convex relaxation technique
to make it solvable by numerical methods.
In~\cite[Lemma~4]{kim:16:ofo},
we solved~\eqref{eq:HD} analytically
yielding the optimized step coefficients
\begin{align}
\hkip &= \begin{cases}
\frac{1}{\theta_{i+1}}
\paren{2\theta_k - \sum_{j=k+1}^i\hkj}, & k=0,\ldots,i-1, \\
1 + \frac{2\theta_i - 1}{\theta_{i+1}}, & k=i,
\end{cases}
\label{eq:hh_ogm}
\end{align}
for $\theta_i$ in~\eqref{eq:theta_rule}.
Fortuitously, the optimized coefficients~\eqref{eq:hh_ogm}
lead to equivalent computationally
efficient OGM1 and OGM2 forms~\cite[Prop.~3, 4 and 5]{kim:16:ofo},
and the bound~\eqref{eq:fv_ogm}
for the final secondary iterate $\x_N$ of
OGM is implied by~\cite[Lemma~4]{kim:16:ofo}.
Recently,
Drori~\cite{drori:17:tei}
showed that the OGM is optimal for $d \ge N+1$,
implying that
optimizing over the relaxed bound~\eqref{eq:D} in~\eqref{eq:HD}
for simplicity
is equivalent to optimizing
over the exact worst-case cost function bound~\eqref{eq:PEP}
when $d\ge N+1$.
One could use a SDP solver to compute a numerical bound
from~\eqref{eq:D}
for any~\FO; however,
deriving an analytical bound using~\eqref{eq:D} is difficult
for the primary sequence $\{\y_i\}$ of OGM.
Therefore, we devised a new relaxed bound in~\cite{kim:17:otc}
similar to~\eqref{eq:D},
which we review next.
\subsection{Review: Another cost function form of relaxed PEP for the primary sequence of OGM}
An upper bound of the worst-case bound on $f(\y_{N+1}) - f(\x_*)$
for~\FO with step coefficients \bmh
and $\y_{N+1} = \x_N - \frac{1}{L}f(\x_N)$
could be computed using~\eqref{eq:D} by a SDP solver.
However, we found it difficult to find its analytical worst-case bound
for the primary sequence $\{\y_i\}$ of OGM,
so~\cite[Prob.~(D$'$)]{kim:17:otc} provided
the following alternate upper bound
on $f(\y_{N+1}) - f(\x_*)$:
\begin{align}
\mathcal{B}_{\mathrm{D'}}(\bmh,N,L,R) :=\;
\min_{\substack{(\bmlam,\bmtau)\in\Lambda, \\ \gamma\in\Reals}}
\left\{
\frac{1}{2}LR^2\gamma\;:\;
\left(\begin{array}{cc}
\S(\bmh,\bmlam,\bmtau) + \frac{1}{2}\u_N\u_N^\top & \frac{1}{2}\bmtau \\
\frac{1}{2}\bmtau^\top & \frac{1}{2}\gamma
\end{array}\right)
\succeq\Zero
\right\}
\label{eq:D_} \tag{D$'$}
,\end{align}
which led to the bound~\eqref{eq:fv_ogm_}
for the primary sequence $\{\y_i\}$ of OGM
in~\cite{kim:17:otc}.
Similar to~\cite[Lemma~4]{kim:16:ofo},
we found a feasible point of~\eqref{eq:D_}
in~\cite[Lemma~3.1]{kim:17:otc},
along with feasible step coefficients \bmh of a~\FO:
\begin{align}
\hkip &= \begin{cases}
\frac{1}{t_{i+1}}
\paren{2t_k - \sum_{j=k+1}^i\hkj}, & k=0,\ldots,i-1, \\
1 + \frac{2t_i - 1}{t_{i+1}}, & k=i,
\end{cases}
\label{eq:hh_ogm_}
\end{align}
for $t_i$ in~\eqref{eq:t_rule}.
Then,~\cite[Thm.~3.1]{kim:17:otc}
showed the bound~\eqref{eq:fv_ogm_}
using~\cite[Lemma~3.1]{kim:17:otc}.
The step coefficients~\eqref{eq:hh_ogm} and~\eqref{eq:hh_ogm_}
are identical except the final iteration,
since $t_i$~\eqref{eq:t_rule}
and $\theta_i$~\eqref{eq:theta_rule} are equivalent for $i<N$.
We are now done reviewing the portions of papers~\cite{kim:16:ofo,kim:17:otc}
that are the ingredients for
specifying a parameterized family of algorithms that generalizes OGM
in the next two sections.
\subsection{Feasible points of~\eqref{eq:D} and~\eqref{eq:D_}
for the generalized OGM}
This section
specifies feasible points of~\eqref{eq:D} and~\eqref{eq:D_}
that lead to a generalized version of OGM.
Specifically, the following lemma
presents additional feasible points
of~\eqref{eq:D};
this lemma reduces to~\cite[Lemma~4]{kim:16:ofo}
(and the step coefficients~\eqref{eq:hh_ogm} of OGM)
when $\theta_i^2 = \Theta_i$ for all $i$.
\begin{lemma}
\label{lem:gogm}
For the following step coefficients:
\begin{align}
\hkip &= \begin{cases}
\frac{\theta_{i+1}}{\Theta_{i+1}}
\paren{2\theta_k - \sum_{j=k+1}^i\hkj},
& i=0,\ldots,N-1,\;k=0,\ldots,i-1, \\
1 + \frac{(2\theta_i - 1)\theta_{i+1}}{\Theta_{i+1}},
& i=0,\ldots,N-1,\;k=i,
\end{cases}
\label{eq:hh_gen_ogm}
\end{align}
the choice of variables:
\begin{align}
\gamma = \frac{1}{2}\tau_0,
\quad
\lambda_i &= \Theta_{i-1} \tau_0,\;\; i=1,\ldots,N,
\quad
\tau_i = \begin{cases}
\frac{2}{\Theta_N}, & i=0, \\
\theta_i\tau_0 & i=1,\ldots,N-1, \\
\frac{\theta_N}{2}\tau_0, & i=N,
\end{cases}
\label{eq:par_gen_ogm}
\end{align}
is a feasible point of~\eqref{eq:D}
for any choice of $\theta_i$ such that
\begin{align}
\theta_0 = 1,
\quad
\theta_i > 0,
\quad\text{and}\quad
\theta_i^2 \le \Theta_i := \begin{cases}
\sum_{l=0}^i \theta_l, & i=0,\ldots,N-1, \\
2\sum_{l=0}^{N-1}\theta_l + \theta_N, & i=N.
\end{cases}
\label{eq:gen_ogm_rule}
\end{align}
\begin{proof}
See Appendix~\ref{appen2}.
\end{proof}
\end{lemma}
The following lemma also specifies some feasible points of~\eqref{eq:D_};
this lemma reduces to~\cite[Lemma~3.1]{kim:17:otc}
(and~\eqref{eq:hh_ogm_})
when $t_i^2 = T_i$ for all $i$.
\begin{lemma}
\label{lem:gogm_}
For the following step coefficients:
\begin{align}
\hkip &= \begin{cases}
\frac{t_{i+1}}{T_{i+1}}
\paren{2t_k - \sum_{j=k+1}^i\hkj},
& i=0,\ldots,N-1,\;k=0,\ldots,i-1, \\
1 + \frac{(2t_i - 1)t_{i+1}}{T_{i+1}},
& i=0,\ldots,N-1,\;k=i,
\end{cases}
\label{eq:hh_gen_ogm_}
\end{align}
the choice of variables:
\begin{align}
\gamma = \frac{1}{2}\tau_0,
\quad
\lambda_i &= T_{i-1} \tau_0,\;\; i=1,\ldots,N,
\quad
\tau_i = \begin{cases}
\frac{1}{T_N}, & i=0, \\
t_i\tau_0 & i=1,\ldots,N,
\end{cases}
\label{eq:par_gen_ogm_}
\end{align}
is a feasible point of~\eqref{eq:D_}
for any choice of $t_i$ such that
\begin{align}
t_0 = 1,
\quad
t_i > 0,
\quad\text{and}\quad
t_i^2 \le T_i := \sum_{l=0}^i t_l.
\label{eq:gen_ogm_rule_}
\end{align}
\begin{proof}
See Appendix~\ref{appen3}.
\end{proof}
\end{lemma}
Similar to the relationship between
the step coefficients~\eqref{eq:hh_ogm} and~\eqref{eq:hh_ogm_},
the step coefficients~\eqref{eq:hh_gen_ogm} and~\eqref{eq:hh_gen_ogm_}
are identical (when $\theta_i = t_i$ for $i<N$)
except for the final iteration,
implying that the iterates $\{\x_i\}_{i=0}^{N-1}$
of the two~{\FO}s
with~\eqref{eq:hh_gen_ogm} and~\eqref{eq:hh_gen_ogm_} are equivalent;
only the final iterate $\x_N$ is different.
The step coefficients~\eqref{eq:hh_ogm} and~\eqref{eq:hh_ogm_}
lead to computationally efficient equivalent OGM forms;
similarly the next section provides
computationally efficient generalized forms of OGM
that each correspond to a~\FO
with either~\eqref{eq:hh_gen_ogm} or~\eqref{eq:hh_gen_ogm_},
and we analyze their cost function worst-case bounds.
\subsection{Generalized OGM}
This section proposes a generalized OGM
using lemmas~\ref{lem:gogm} and~\ref{lem:gogm_}.
The \FO with the step coefficients~\eqref{eq:hh_gen_ogm}
has the following two equivalent efficient generalized forms of OGM,
named GOGM1 and GOGM2,
that reduce to the standard OGM when $\theta_i^2 = \Theta_i$ for all $i$.
\begin{center}
\fbox{
\footnotesize
\begin{minipage}[t]{0.45\textwidth}
\vspace{-10pt}
\begin{flalign*}
&\text{\bf Algorithm GOGM1} & \\
&\text{Input: } f\in \cF,\; \x_0=\y_0\in\Reals^d, & \\
& \hspace{30pt} \theta_0=\Theta_0 = 1. & \\
&\text{For } i = 0,\ldots,N-1 & \\
&\quad \y_{i+1} = \x_i - \frac{1}{L}\nabla f(\x_i) & \\
&\quad \text{Choose } \theta_{i+1}>0 & \\
&\quad \text{s.t. }
\theta_{i+1}^2
\le \Theta_{i+1}, \text{ where} & \\
&\quad \Theta_{i+1} =
\begin{cases}
\sum_{l=0}^{i+1} \theta_l, & i \le N-2 \\
2\sum_{l=0}^{N-1}\theta_l + \theta_N, & i = N-1
\end{cases} & \\
&\quad \x_{i+1} = \y_{i+1}
+ \frac{(\Theta_i - \theta_i)\theta_{i+1}}{\theta_i\Theta_{i+1}}
(\y_{i+1} - \y_i) & \\
&\hspace{55pt}
+ \frac{(2\theta_i^2 - \Theta_i)\theta_{i+1}}{\theta_i\Theta_{i+1}}(\y_{i+1} - \x_i) &
\end{flalign*}
\end{minipage}\vline\hspace{8pt
\begin{minipage}[t]{0.45\textwidth}
\vspace{-10pt}
\begin{flalign*}
&\text{\bf Algorithm GOGM2} & \\
&\text{Input: } f\in \cF,\; \x_0=\y_0\in\Reals^d, & \\
& \hspace{30pt} \theta_0=\Theta_0 = 1. & \\
&\text{For } i = 0,\ldots,N-1 & \\
&\quad \y_{i+1} = \x_i - \frac{1}{L}\nabla f(\x_i) & \\
&\quad \z_{i+1} = \x_0 - \frac{1}{L}\sum_{k=0}^i2\theta_k \nabla f(\x_k) & \\
&\quad \text{Choose } \theta_{i+1}>0 & \\
&\quad \text{s.t. }
\theta_{i+1}^2
\le \Theta_{i+1}, \text{ where} & \\
&\quad \Theta_{i+1} =
\begin{cases}
\sum_{l=0}^{i+1} \theta_l, & i \le N-2 \\
2\sum_{l=0}^{N-1}\theta_l + \theta_N, & i = N-1
\end{cases} & \\
&\quad \x_{i+1} = \paren{1 - \frac{\theta_{i+1}}{\Theta_{i+1}}}\y_{i+1}
+ \frac{\theta_{i+1}}{\Theta_{i+1}}\z_{i+1} &
\end{flalign*}
\end{minipage}
} \vspace{5pt}
\end{center}
\begin{proposition}
\label{prop:gogm}
The sequence $\{\x_0,\cdots,\x_N\}$ generated
by the~\FO with
\eqref{eq:hh_gen_ogm}
is identical to the corresponding sequence
generated by GOGM1 and GOGM2.
\begin{proof}
See Appendix~\ref{appen4}.
Note that this proof is independent of the choice of
$\theta_i$ and $\Theta_i$.
\end{proof}
\end{proposition}
Because the proof of Prop.~\ref{prop:gogm}
for the~\FO with step coefficients~\eqref{eq:hh_gen_ogm}
is independent of the choice of $\theta_i$ and $\Theta_i$,
it is straightforward to show that
the~\FO with step coefficients~\eqref{eq:hh_gen_ogm_}
has the following two efficient equivalent forms,
named GOGM1$'$ and GOGM2$'$,
that reduce to~\cite[Alg. OGM1$'$ and OGM2$'$]{kim:17:otc}
when $t_i^2 = T_i$ for all $i$.
\begin{center}
\fbox{
\footnotesize
\begin{minipage}[t]{0.45\textwidth}
\vspace{-10pt}
\begin{flalign*}
&\text{\bf Algorithm GOGM1$'$} & \\
&\text{Input: } f\in \cF,\; \x_0=\y_0\in\Reals^d, & \\
& \hspace{30pt} t_0=T_0 = 1. & \\
&\text{For } i = 0,\ldots,N-1 & \\
&\quad \y_{i+1} = \x_i - \frac{1}{L}\nabla f(\x_i) & \\
&\quad \text{Choose } t_{i+1}>0 & \\
&\quad \text{s.t. }
t_{i+1}^2
\le T_{i+1} = \sum_{l=0}^{i+1} t_l & \\
&\quad \x_{i+1} = \y_{i+1}
+ \frac{(T_i - t_i)t_{i+1}}{t_iT_{i+1}}
(\y_{i+1} - \y_i) & \\
&\hspace{55pt}
+ \frac{(2t_i^2 - T_i)t_{i+1}}{t_iT_{i+1}}(\y_{i+1} - \x_i) &
\end{flalign*}
\end{minipage}\vline\hspace{8pt
\begin{minipage}[t]{0.45\textwidth}
\vspace{-10pt}
\begin{flalign*}
&\text{\bf Algorithm GOGM2$'$} & \\
&\text{Input: } f\in \cF,\; \x_0=\y_0\in\Reals^d, & \\
& \hspace{30pt} t_0=T_0 = 1. & \\
&\text{For } i = 0,\ldots,N-1 & \\
&\quad \y_{i+1} = \x_i - \frac{1}{L}\nabla f(\x_i) & \\
&\quad \z_{i+1} = \x_0 - \frac{1}{L}\sum_{k=0}^i2t_k \nabla f(\x_k) & \\
&\quad \text{Choose } t_{i+1}>0 \\
&\quad \text{s.t. }
t_{i+1}^2
\le T_{i+1} = \sum_{l=0}^{i+1} t_l & \\
&\quad \x_{i+1} = \paren{1 - \frac{t_{i+1}}{T_{i+1}}}\y_{i+1}
+ \frac{t_{i+1}}{T_{i+1}}\z_{i+1} &
\end{flalign*}
\end{minipage}
} \vspace{5pt}
\end{center}
\noindent
Clearly when $\theta_i = t_i$ for $i<N$,
the primary iterates $\{\y_i\}_{i=0}^N$
and the intermediate secondary iterates $\{\x_i\}_{i=0}^{N-1}$
of GOGM and GOGM$'$ are equivalent.
Although illustrating two similar algorithms GOGM and GOGM$'$
might seem redundant,
presenting both formulations with lemmas~\ref{lem:gogm} and~\ref{lem:gogm_}
completes the story of generalized OGM
here and in Sec.~\ref{sec:pep,sgnorm}.
Using lemmas~\ref{lem:gogm} and~\ref{lem:gogm_},
the following theorem
bounds the cost function decrease of the GOGM and GOGM$'$ iterates.
\begin{theorem}
\label{thm:gogm}
Let $f\;:\;\Reals^d\rightarrow\Reals$ be $\cF$
and let $\y_0,\cdots,\y_N,\x_N\in\Reals^d$ be
generated by GOGM1 and GOGM2. Then for any $1\le i \le N$,
\begin{align}
f(\y_i) - f(\x_*) &\le \frac{LR^2}{4\Theta_{i-1}},
\label{eq:gogm_conv_} \\
f(\x_N) - f(\x_*) &\le \frac{LR^2}{2\Theta_N}
\label{eq:gogm_conv}
.\end{align}
The iterates $\y_0,\cdots,\y_N\in\Reals^d$
generated by GOGM1$'$ and GOGM2$'$
also satisfy the bound~\eqref{eq:gogm_conv_}
when $\theta_i = t_i$ for $i<N$.
\begin{proof}
Using Lemma~\ref{lem:gogm_},
the~\FO with \bmh in~\eqref{eq:hh_gen_ogm_}
of GOGM$'$ satisfies
\begin{align}
f(\y_{N+1}) - f(\x_*) \le \mathcal{B}_{\mathrm{D'}}(\bmh,N,L,R)
= \frac{1}{2}LR^2\gamma
= \frac{LR^2}{4T_N}
\label{eq:fv_gogm_}
,\end{align}
where $\y_{N+1} = \x_N - \frac{1}{L}\nabla f(\x_N)$.
Since the coefficients \bmh in~\eqref{eq:hh_gen_ogm_}
are recursive and do not depend on a given $N$,
we can extend~\eqref{eq:fv_gogm_} for all iterations.
By letting $\theta_i = t_i$ for $i<N$,
the bound~\eqref{eq:fv_gogm_} also satisfies
for the iterates $\{\y_i\}$ of GOGM,
as in~\eqref{eq:gogm_conv_}.
Using Lemma~\ref{lem:gogm},
the~\FO with the step \bmh in~\eqref{eq:hh_gen_ogm}
of GOGM satisfies
\begin{align}
f(\x_N) - f(\x_*) \le \mathcal{B}_{\mathrm{D}}(\bmh,N,L,R)
= \frac{1}{2}LR^2\gamma
= \frac{LR^2}{2\Theta_N}
,\end{align}
which is equivalent to~\eqref{eq:gogm_conv}.
\end{proof}
\end{theorem}
GOGM and Thm.~\ref{thm:gogm}
reduce to OGM and its bounds~\eqref{eq:fv_ogm_} and~\eqref{eq:fv_ogm},
when $\theta_i^2 = \Theta_i$ for all $i$.
Similar to general forms of FGM in~\cite{chambolle:15:otc,nesterov:05:smo},
the GOGM family includes the choice
$
\theta_i = \begin{cases}
\frac{i+a}{a}, & i < N, \\
\frac{\sqrt{2}(N+a-1)}{a}, & i= N
\end{cases}
$
for any $a\ge2$, because such parameter $\theta_i$ satisfies
the following conditions for GOGM:
\begin{align}
\Theta_i - \theta_i^2 = \frac{(i+1)(i+2a)}{2a} - \frac{(i+a)^2}{a^2}
= \frac{(a-2)i^2 + a(2a-3)}{2a^2}
\ge 0,
\label{eq:Thth}
\end{align}
for $i<N$, and
$
\Theta_N - \theta_N^2 = 2\Theta_{N-1} + \theta_N - \theta_N^2
\ge 2\theta_{N-1}^2 + \theta_N - \theta_N^2
= \theta_N \ge 0.
$
Similarly, the GOGM$'$ family includes the choice $t_i=\frac{i+a}{a}$
for any $a\ge2$, which we denote as OGM-$a$.
\begin{corollary}
\label{cor:ogma}
Let $f\;:\;\Reals^d\rightarrow\Reals$ be $\cF$
and let $\y_0,\cdots,\y_N\in\Reals^d$ be
generated by GOGM$'$ with $t_i = \frac{i+a}{a}$ (OGM-$a$)
for any $a\ge2$. Then for any $1\le i \le N$,
\begin{align}
f(\y_i) - f(\x_*) \le \frac{aLR^2}{2i(i+2a-1)}
\label{eq:gogm_cor}
.\end{align}
\begin{proof}
Thm.~\ref{thm:gogm} implies~\eqref{eq:gogm_cor},
since $T_i = \frac{(i+1)(i+2a)}{2a}$
and the condition $T_i - t_i^2\ge0$ satisfies in~\eqref{eq:Thth}
for any $a\ge2$.
\end{proof}
\end{corollary}
\section{Relaxation and optimization of the gradient form of PEP}
\label{sec:pep,sgnorm}
This section analyzes
a worst-case bound for the gradient of any GOGM (and GOGM$'$)
using the gradient form of PEP.
We use relaxations on the gradient form of PEP
that are similar but slightly different
from those of PEP for the cost function in the previous section.
Using this relaxed PEP,
we prove that FGM has an $O(1/N^{1.5})$ rate for the worst-case gradient decrease,
and analyze the worst-case gradient bound for the GOGM.
Then, we optimize the step coefficients with
respect to the gradient form of PEP
and propose an algorithm named OGM-OG
that lies in the GOGM family
and that has the best known analytical worst-case bound
for decreasing the gradient norm among the class~\FO.
\subsection{Relaxation for the gradient form of PEP}
\label{sec:pep,grad}
To analyze a worst-case bound on the gradient
for a~\FO with a given \bmh,
we consider the following gradient-form version of PEP
that is similar to~\eqref{eq:PEP}:
\begin{align}
\mathcal{B}_{\mathrm{P''}}(\bmh,N,d,L,R) :=\;
& \max_{\substack{f\in\cF, \\
\x_0,\cdots,\x_N\in\Reals^d, \\
\x_*\in X_*(f), \\
||\x_0 - \x_*|| \le R}}
\min_{i\in\{0,\ldots,N\}}
||\nabla f(\x_i)||^2
\label{eq:PEP__} \tag{P$''$} \\
\st \; &\x_{i+1} = \x_i - \frac{1}{L}\sum_{k=0}^i \hkip \nabla f(\x_k),
\quad i=0,\ldots,N-1. \nonumber
\end{align}
Here, we use the smallest gradient norm squared among all iterates
($\min_{i\in\{0,\ldots,N\}}\allowbreak||\nabla f(\x_i)||^2$)
as a criteria, as considered in~\cite[Sec.~4.3]{taylor:17:ssc}.
We could instead consider the final gradient norm squared
($||\nabla f(\x_N)||^2$) as a criteria,
but our proposed relaxation on~\eqref{eq:PEP__} in this section
for such criteria
provided only an $O(1/N)$ worst-case bound at best
even for the corresponding optimized step coefficients
(results not shown);
we leave studying the gradient form of the tight PEP as future work.
As in~\cite{taylor:17:ssc},
we replace $\min_{i\in\{0,\ldots,N\}} ||\nabla f(\x_i)||^2$ in~\eqref{eq:PEP__}
by $L^2||\x_0-\x_*||^2\alpha$ with the condition
$\alpha \le \frac{1}{L^2||\x_0-\x_*||^2}||\nabla f(\x_i)||^2
= \Tr{\G^\top(\u_i\u_i^\top)\G}$
for all $i$.
Then, we relax this reformulated~\eqref{eq:PEP__}
similar to the relaxation from~\eqref{eq:PEP} to~\eqref{eq:pPEP}
with the additional constraint $\Tr{\G^\top\C_N\G} \le \delta_N$ in~\eqref{eq:relPEP}
as follows:
\begin{align}
\mathcal{B}_{\mathrm{P2''}}(\bmh,N,d,L,R) :=\;
&\max_{\substack{\G\in\Reals^{(N+1)d}, \\ \del\in\Reals^{N+1}, \\ \alpha\in\Reals}}
L^2R^2\alpha
\label{eq:pPEP__} \tag{P2$''$} \\
&\st \; \Tr{\G^\top\A_{i-1,i}(\bmh)\G} \le \delta_{i-1} - \delta_i,
\quad i=1,\ldots,N, \nonumber \\
&\quad\;\;\;
\Tr{\G^\top\C_N\G} \le \delta_N, \nonumber \\
&\quad\;\;\; \Tr{\G^\top\D_i(\bmh)\G + \bmnu \u_i^\top\G} \le -\delta_i,
\quad i=0,\ldots,N, \nonumber \\
&\quad\;\;\; \Tr{\G^\top(-\u_i\u_i^\top)\G} \le - \alpha,
\quad i=0,\ldots,N. \nonumber
\end{align}
Replacing $\max_{\G,\del} L^2R^2\alpha$
by $\min_{\G,\del} \{-\alpha\}$ for convenience,
the Lagrangian of the corresponding constrained minimization problem
with dual variables $\bmlam\in\Reals_+^N$, $\eta\in\Reals_+$,
$\bmtau\in\Reals_+^{N+1}$, and $\bmbeta = (\beta_0,\cdots,\beta_N)^\top\in\Reals_+^{N+1}$
for the first, second, third, and fourth set of constraint
inequalities of~\eqref{eq:pPEP__} respectively becomes
\begin{align}
\cL''(\G,\del,\bmlam,\eta,\bmtau,\bmbeta;\bmh)
=& - \alpha + \sum_{i=1}^N\lambda_i(\delta_i - \delta_{i-1})
- \eta\delta_N
+ \sum_{i=0}^N\tau_i\delta_i
+ \sum_{i=0}^N\beta_i\alpha \\
&\qquad\qquad\qquad
+ \Tr{\G^\top\S''(\bmh,\bmlam,\eta,\bmtau,\bmbeta)\G + \bmnu\bmtau^\top\G}
\nonumber
,\end{align}
where
\begin{align}
\S''(\bmh,\bmlam,\eta,\bmtau,\bmbeta)
:= \S(\bmh,\bmlam,\bmtau)
+ \frac{1}{2}\eta\u_N\u_N^\top
- \sum_{i=0}^N\beta_i\u_i\u_i^\top
\label{eq:S__}
.\end{align}
Then similar to~\eqref{eq:D}, we have
the following dual problem of~\eqref{eq:pPEP__}
that one could use to compute an upper bound of the PEP~\eqref{eq:PEP__}
of the smallest gradient norm squared among all iterates
by a numerical SDP solver:
\begin{align}
\mathcal{B}_{\mathrm{D''}}(\bmh,N,L,R) :=\;
&\min_{\substack{(\bmlam,\eta,\bmtau,\bmbeta)\in\Lambda'', \\ \gamma\in\Reals}}
\left\{
\frac{1}{2}L^2R^2\gamma\;:\;
\left(\begin{array}{cc}
\S''(\bmh,\bmlam,\eta,\bmtau,\bmbeta) & \frac{1}{2}\bmtau \\
\frac{1}{2}\bmtau^\top & \frac{1}{2}\gamma
\end{array}\right)
\succeq\Zero
\right\}
\label{eq:D__} \tag{D$''$}
,\end{align}
where
\begin{align}
\Lambda'' := \left\{(\bmlam,\eta,\bmtau,\bmbeta)\in\Reals_+^{3N+3}
\;:\;
\begin{array}{l}
\tau_0 = \lambda_1, \;\; \lambda_N + \tau_N = \eta, \;\;
\sum_{i=0}^N \beta_i = 1, \\
\lambda_i - \lambda_{i+1} + \tau_i = 0, \; i=1,\ldots,N-1
\end{array}\right\}
\label{eq:Lam__}
.\end{align}
The next two sections use a valid upper bound~\eqref{eq:D__} of~\eqref{eq:PEP__}
for given step coefficients \bmh,
providing an analytical solution to~\eqref{eq:D__}
for the step coefficients \bmh of FGM and GOGM$'$,
superseding the use of a SDP solver.
\subsection{A worst-case bound for the gradient norm of FGM}
\label{sec:pep,sgnorm,fgm}
FGM is equivalent to a~\FO with the step coefficients
\cite[Prop.~1]{kim:16:ofo}:
\begin{align}
h_{i+1,k}
&= \begin{cases}
\frac{1}{t_{i+1}}
\left(t_k - \sum_{j=k+1}^i h_{j,k}\right), & k=0,\ldots,i-1, \\
1 + \frac{t_i - 1}{t_{i+1}}, & k=i,
\end{cases}
\label{eq:hh_fgm}
\end{align}
for $t_i$ in~\eqref{eq:t_rule}.
The following lemma provides a feasible point of~\eqref{eq:D__}
associated with the step coefficients~\eqref{eq:hh_fgm} of FGM
to provide a worst-case bound for the gradient of FGM.
\begin{lemma}
\label{lem:fgm}
For the step coefficients~\eqref{eq:hh_fgm},
the following choice of variables:
\begin{align}
\gamma &= \tau_0, \quad
\lambda_i = t_{i-1}^2\tau_0, \quad i=1,\ldots,N, \quad
\tau_i =
\begin{cases}
\left(\frac{1}{2}\sum_{k=0}^N t_k^2\right)^{-1}, & i = 0, \\
t_i\tau_0, & i = 1,\ldots,N,
\end{cases}
\label{eq:par_fgm} \\
\eta &= t_N^2 \tau_0, \quad
\beta_i = \frac{1}{2}t_i^2\tau_0, \quad i=0,\ldots,N,
\label{eq:gam_fgm}
\end{align}
is a feasible point of~\eqref{eq:D__} for $t_i$ in~\eqref{eq:t_rule}.
\begin{proof}
See Appendix~\ref{appen5}.
\end{proof}
\end{lemma}
Using Lemma~\ref{lem:fgm},
the following theorem bounds the gradient norm of the FGM iterates,
proving for the first time an $O(1/N^{1.5})$ rate of decrease.
\begin{theorem}
\label{thm:fgm}
Let $f\;:\;\Reals^d\rightarrow\Reals$ be $\cF$
and let $\y_0,\cdots,\y_N,\x_0,\cdots,\x_N\in\Reals^d$ be
generated by FGM. Then for any $N\ge1$,
\begin{align}
\min_{i\in\{0,\ldots,N+1\}}||\nabla f(\y_i)||
&\le \min_{i\in\{0,\ldots,N\}}||\nabla f(\x_i)||
\label{eq:fgm_g} \\
&\le \frac{LR}{\sqrt{\sum_{k=0}^Nt_k^2}}
\le \frac{2\sqrt{3}LR}{\sqrt{(N+1)(N^2 + 6N + 12)}}
\nonumber
,\end{align}
where $\y_{N+1} = \x_N - \frac{1}{L}\nabla f(\x_N)$.
\begin{proof}
Lemma~\ref{lem:gmono}
implies the first inequality in~\eqref{eq:fgm_g}.
Using Lemma~\ref{lem:fgm},
the~\FO with the step coefficients \bmh~\eqref{eq:hh_fgm} of FGM satisfies
\begin{align}
\min_{i\in\{0,\ldots,N\}}||\nabla f(\x_i)||^2 \le \mathcal{B}_{\mathrm{D''}}(\bmh,N,L,R)
= \frac{1}{2}L^2R^2\gamma
= \frac{L^2R^2}{\sum_{k=0}^Nt_k^2}
,\end{align}
which is equivalent to~\eqref{eq:fgm_g} using
$
\sum_{k=0}^N t_k^2 \ge \sum_{k=0}^N \frac{(k+2)^2}{4}
= \frac{(N+1)(2N^2 + 13N + 24)}{24}
.$
\end{proof}
\end{theorem}
\subsection{A worst-case bound for the gradient norm of GOGM}
Having established
the gradient bound~\eqref{eq:fgm_g}
for FGM,
this section and the next
seek to improve on it
by studying GOGM.
To bound the gradient decrease of GOGM (and GOGM$'$),
the following lemma illustrates one possible set
of feasible points of~\eqref{eq:D__}.
\begin{lemma}
\label{lem:gogm_feas_}
For the step coefficients~\eqref{eq:hh_gen_ogm_},
the following choice of variables:
\begin{align}
\gamma = \frac{1}{2}\tau_0, \quad\!
\lambda_i &= T_{i-1}\tau_0, \quad\! i=1,\ldots,N, \quad\!
\tau_i =
\begin{cases}
\left(\sum_{k=0}^N\left(T_k - t_k^2\right)\right)^{-1}, & i = 0, \\
t_i\tau_0, & i = 1,\ldots,N,
\end{cases}
\label{eq:par_ogmsg} \\
\eta &= T_N \tau_0, \quad
\beta_i = \left(T_i - t_i^2\right)\tau_0,\quad i=0,\ldots,N
\label{eq:gam_ogmsg}
\end{align}
is a feasible point of~\eqref{eq:D__} for any choice of $t_i$ and $T_i$
that satisfies~\eqref{eq:gen_ogm_rule_}
and for which there exists some $i$ such that $t_i^2 < T_i$.
\begin{proof}
See Appendix~\ref{appen6}.
\end{proof}
\end{lemma}
Using Lemma~\ref{lem:gogm_feas_},
the following theorem bounds the worst-case gradient norm
for the iterates of GOGM and GOGM$'$.
\begin{theorem}
\label{thm:gogm_g}
Let $f\;:\;\Reals^d\rightarrow\Reals$ be $\cF$
and let $\y_0,\cdots,\y_N,\x_0,\cdots,\x_N\in\Reals^d$ be
generated by GOGM$'$. Then for any $N\ge1$,
\begin{align}
\min_{i\in\{0,\ldots,N+1\}} ||\nabla f(\y_i)||
\le \min_{i\in\{0,\ldots,N\}} ||\nabla f(\x_i)||
\le \frac{LR}{2\sqrt{\sum_{k=0}^N\left(T_k - t_k^2\right)}}
\label{eq:gogm_g}
,\end{align}
where $\y_{N+1} = \x_N - \frac{1}{L}\nabla f(\x_N)$.
The bound~\eqref{eq:gogm_g} can be generalized
to the intermediate iterates $\{\x_i\}_{i=0}^{N-1}$ and $\{\y_i\}_{i=0}^N$
of both GOGM and GOGM$'$ when $\theta_i = t_i$ (for $i<N$).
\begin{proof}
Lemma~\ref{lem:gmono}
implies the first inequality in~\eqref{eq:gogm_g}.
Using Lemma~\ref{lem:gogm_feas_},
\FO with the step coefficients \bmh~\eqref{eq:hh_gen_ogm_} of GOGM$'$ satisfies
\begin{align*}
\min_{i\in\{0,\ldots,N\}} ||\nabla f(\x_i)||^2
\le \mathcal{B}_{\mathrm{D''}}(\bmh,N,L,R)
= \frac{1}{2}L^2R^2\gamma
= \frac{L^2R^2}{4\sum_{k=0}^N\left(T_k - t_k^2\right)}
,\end{align*}
which implies~\eqref{eq:gogm_g}.
Since the iterates of GOGM$'$ are recursive and do not depend on a given $N$,
the bound~\eqref{eq:gogm_g}
easily generalizes to the intermediate iterates
of GOGM$'$ (and GOGM when $\theta_i = t_i$).
\end{proof}
\end{theorem}
\subsection{Optimizing step coefficients over the gradient form of PEP}
\label{sec:opt,grad,pep}
In search of a~\FO that decreases the gradient norm the fastest,
we optimize the step coefficients
in terms of the gradient form of the relaxed~\eqref{eq:D__}
by solving the following problem:
\begin{align}
\hat{\bmh}_{\mathrm{D''}}
:= \argmin{\bmh\in\Reals^{N(N+1)/2}} \mathcal{B}_{\mathrm{D''}}(\bmh,N,L,R)
\tag{HD$''$}
\label{eq:HD__}
.\end{align}
The problem~\eqref{eq:HD__} is bilinear, similar to~\eqref{eq:HD},
and a convex relaxation technique~\cite[Thm.~3]{drori:14:pof}
makes this problem solvable using numerical methods.
We solved~\eqref{eq:HD__} for many choices of $N$
using a numerical SDP solver~\cite{cvxi,gb08},
and observed that the following choice of $t_i$:
\begin{align}
t_i &= \begin{cases}
1, & i = 0, \\
\frac{1+\sqrt{1 + 4t_{i-1}^2}}{2}, & i = 1, \ldots, \fNh - 1, \\
\frac{N-i+1}{2}, & i = \fNh,\ldots,N,
\end{cases}
\label{eq:opt_t}
\end{align}
makes the feasible point in Lemma~\ref{lem:gogm_feas_} optimal
for the problem~\eqref{eq:HD__}.
Based on that numerical evidence,
we conjecture that $\hat{\bmh}_{\mathrm{D''}}$ in~\eqref{eq:HD__}
corresponds to the step coefficients~\eqref{eq:hh_gen_ogm_} with
the parameter $t_i$~\eqref{eq:opt_t}.
The $t_i$ factors in~\eqref{eq:opt_t}
start decreasing after $i=\fNh-1$,
whereas the usual $t_i$ in~\eqref{eq:t_rule}
and $t_i = \frac{i+a}{a}$ for any $a\ge2$
increase with $i$ indefinitely.
In addition, we found numerically that minimizing
the gradient bound~\eqref{eq:gogm_g} of GOGM$'$, \ie,
solving the following constrained quadratic problem:
\begin{align}
\max_{\{t_i\}} \sum_{k=0}^N\paren{\sum_{l=0}^kt_l - t_k^2}
\quad\st\quad t_i \text{ satisfies~\eqref{eq:gen_ogm_rule_} for all } i
\label{eq:quad}
,\end{align}
is equivalent to solving the problem~\eqref{eq:HD__}.
In other words, the solution of~\eqref{eq:quad} numerically
appears equivalent to~\eqref{eq:opt_t},
the (conjectured) solution of~\eqref{eq:HD__}.
The unconstrained maximizer of the cost function of~\eqref{eq:quad}
is $t_i = \frac{N-i+1}{2}$,
and this term
partially appears in the constrained maximizer~\eqref{eq:opt_t}
for $\fNh \le i\le N$.
We denote the resulting GOGM$'$ with~\eqref{eq:opt_t}
as OGM-OG (OG for optimized over gradient).
The following theorem bounds
the cost function and gradient norm of the OGM-OG iterates.
\begin{theorem}
Let $f\;:\;\Reals^d\rightarrow\Reals$ be $\cF$
and let $\y_0,\cdots,\y_N,\x_0,\cdots,\x_N\in\Reals^d$ be
generated by OGM-OG. Then,
\begin{align}
f(\y_{N+1}) - f(\x_*)
&\le \frac{2LR^2}{(N+2)^2},
\label{eq:ogmog_fv} \\
\min_{i\in\{0,\ldots,N+1\}} ||\nabla f(\y_i)||
&\le \min_{i\in\{0,\ldots,N\}} ||\nabla f(\x_i)||
\le \frac{\sqrt{6}LR}{N\sqrt{N+1}}
\label{eq:ogmog_g}
,\end{align}
where $y_{N+1} = \x_N - \frac{1}{L}\nabla f(\x_N)$.
\begin{proof}
OGM-OG is an instance of GOGM$'$
and thus Thm.~\ref{thm:gogm} implies that OGM-OG satisfies
\begin{align*}
f(\y_{N+1}) - f(\x_*) \le \frac{LR^2}{4T_N},
\end{align*}
which is equivalent to~\eqref{eq:ogmog_fv},
since
\begin{align*}
T_N
&= T_m + \sum_{l=m+1}^Nt_l
= t_m^2 + \frac{(N-m)(N-m+1)}{4} \\
&\ge \frac{(N+3)^2 + N(N+2)}{16}
= \frac{2N^2 + 8N + 9}{16}
\end{align*}
for $m=\fNh$,
using $m\ge\frac{N-1}{2}$, $N-m \ge\frac{N}{2}$,
and $t_m\ge\frac{m+2}{2}\ge\frac{N+3}{4}$ in~\eqref{eq:t_rule}.
Thm.~\ref{thm:gogm_g} implies~\eqref{eq:ogmog_g},
using
the above inequalities for $m$,
the equality $t_i^2 = T_i$ for $i\le m$,
and
\begingroup
\allowdisplaybreaks
\begin{align*}
&\sum_{k=m+1}^N \left(T_k - t_k^2\right)
= \sum_{k=m+1}^N \left(t_m^2 + \sum_{l=m+1}^k t_l - t_k^2\right) \\
=& \left(N - m\right)t_m^2
+ \sum_{k=m+1}^N\left(\sum_{l=m+1}^k\frac{N-l+1}{2}
- \left(\frac{N-k+1}{2}\right)^2\right) \\
=& \left(N - m\right)t_m^2
+ \sum_{k'=1}^{N-m}\left(\sum_{l'=1}^{k'}\frac{N-l'-m+1}{2}
- \left(\frac{N-k'-m+1}{2}\right)^2\right) \\
=& \left(N - m\right)t_m^2 \\
&\quad + \sum_{k=1}^{N-m}\left(\frac{2(N-m+1)k - k(k+1)}{4}
- \frac{(N-m+1)^2 - 2(N-m+1)k + k^2}{4}\right) \\
=& \left(N - m\right)t_m^2
+ \sum_{k=1}^{N-m}\left(
- \frac{k^2}{2}
+ (N-m+3/4)k
- \frac{(N-m+1)^2}{4}
\right) \\
=& \left(N - m\right)t_m^2
- \frac{(N-m)(N-m+1/2)(N-m+1)}{6} \\
&\quad + \frac{(N-m)(N-m+3/4)(N-m+1)}{2}
- \frac{(N-m)(N-m+1)^2}{4} \\
\ge& \frac{(N-m)(m + 2)^2}{4}
+ \frac{(N-m)^2(N-m+1)}{3}
- \frac{(N-m)(N-m+1)^2}{4} \\
\ge& \frac{(N-m)^2(N-m+1)}{3}
\ge \frac{1}{24}N^2(N+1)
.\end{align*}
\endgroup
\end{proof}
\end{theorem}
\noindent
The gradient bound~\eqref{eq:ogmog_g} of OGM-OG
is asymptotically $\sqrt{2}$-times smaller than that of FGM
in Thm.~\ref{thm:fgm}
and $1.5$-times smaller than that of OGM-$m\!=\!\fNt$
in Thm.~\ref{thm:ogmh}.
Regarding the cost function decrease,
the bound~\eqref{eq:ogmog_fv} of OGM-OG
is asymptotically the same as the bound~\eqref{eq:fv_fgm} of FGM,
and both are twice larger than
the bounds~\eqref{eq:fv_ogm_} and~\eqref{eq:fv_ogm} of OGM.
\subsection{Decreasing the gradient norm with rate $O(1/N^{1.5})$
using GOGM without selecting $N$ in advance}
Although OGM-OG satisfies a small worst-case gradient bound
with a rate $O(1/N^{1.5})$,
OGM-OG (and FGM-\m and OGM-\m) must select $N$ in advance,
unlike FGM.
Using Thm.~\ref{thm:gogm_g},
the following corollary shows that OGM-$a$ with $a>2$
can decrease the gradient with a rate $O(1/N^{1.5})$
without selecting $N$ in advance.
(Cor.~\ref{cor:ogma} showed that OGM-$a$ algorithm with $a\ge2$
can decrease the cost function with an optimal rate $O(1/N^2)$.)
\begin{corollary}
Let $f\;:\;\Reals^d\rightarrow\Reals$ be $\cF$
and let $\y_0,\cdots,\y_N,\x_0,\cdots,\x_N\in\Reals^d$ be
generated by GOGM$'$ with $t_i = \frac{i+a}{a}$ (OGM-$a$) for any $a\ge2$.
Then for $N\ge1$,
\begin{align}
\min_{i\in\{0,\ldots,N+1\}} ||\nabla f(\y_i)||
&\le \min_{i\in\{0,\ldots,N\}} ||\nabla f(\x_i)|| \label{eq:ogma_g} \\
&\le \frac{a\sqrt{6}LR}{2\sqrt{N(N+1)\paren{(a-2)N + (3a^2 - 4a - 2)}}}
\nonumber
,\end{align}
where $\y_{N+1} = \x_N - \frac{1}{L}\nabla f(\x_N)$.
\begin{proof}
Using $T_i = \frac{(i+1)(i+2a)}{2a}$ and~\eqref{eq:Thth},
Thm.~\ref{thm:gogm_g} implies~\eqref{eq:ogma_g} since
\begin{align*}
\sum_{k=0}^N(T_k - t_k^2)
&= \sum_{k=0}^N\frac{(a-2)k^2 + a(2a-3)k}{2a^2} \\
&= \frac{N(N+1)\paren{(a-2)N + (3a^2 - 4a - 2)}}{6a^2}
.\end{align*}
\end{proof}
\end{corollary}
\noindent
OGM-$a$ for any $a>2$ has a gradient bound~\eqref{eq:ogma_g}
that is about $\frac{a}{2\sqrt{a-2}}$-times larger
than the bound~\eqref{eq:ogmog_g} of OGM-OG.
This constant factor minimizes to $\sqrt{2}$ when $a=4$,
and this OGM-$a\!=\!4$ has a worst-case gradient bound
that is asymptotically equivalent to the bound~\eqref{eq:fgm_g} of FGM.
Therefore, when one does not want to select $N$ in advance,
both FGM and OGM-$a\!=\!4$
(and OGM-$a$ for any $a>2$)
will be useful
for decreasing the gradient with a rate $O(1/N^{1.5})$.
\section{Discussion}
\label{sec:disc}
This section summarizes analytical worst-case bounds of~\FO
discussed in the previous sections.
This section also reports tight numerical worst-case bounds
for exact comparison of algorithms
because many of the analytical bounds are not guaranteed to be tight.
\subsection{Summary of analytical worst-case bounds
on the cost function and gradient norm}
Table~\ref{tab:rate} summarizes
the asymptotic rate
of analytical worst-case bounds of all algorithms
described in this paper.
As discussed, OGM and OGM-OG have the best known worst-case bounds
for the cost function and gradient decrease respectively in Table~\ref{tab:rate}.
However, since OGM has a slow worst-case rate for the gradient decrease,
other algorithms such as FGM, OGM-\m, OGM-OG, and OGM-$a$
that satisfy both the optimal rate $O(1/N^2)$ for the function decrease
and a fast rate $O(1/N^{1.5})$ for the gradient decrease
could be preferable over OGM when one is interested
in both the gradient decrease
as well as the function decrease,
particularly when solving dual problems.
In addition, when one does not want to choose $N$ in advance,
FGM and OGM-$a$ could be preferable.
\begin{table}[!h]
\centering
\begin{tabular}{|l|c|c|c|}
\hline
\multirow{2}{*}{Algorithm}
& \multicolumn{2}{c|}{Asymptotic worst-case bound}
& Require selecting \\
\cline{2-3}
& Cost function & Gradient norm & $N$ in advance \\
\hline
GM
& $\frac{1}{4} N^{-1}$ & $\sqrt{2} N^{-1}$ & No \\ \hline
FGM
& $2 N^{-2}$ & $2\sqrt{3} N^{-1.5}$ & No \\ \hline
{\bf OGM}
& $N^{-2}$ & $\sqrt{2} N^{-1}$ & No \\ \hline
OGM-$m\!=\!\fNt$
& $\frac{9}{4} N^{-2}$
& $\frac{3\sqrt{6}}{2} N^{-1.5}$ & Yes \\ \hline
{\bf OGM-OG}
& $2 N^{-2}$ & $\sqrt{6} N^{-1.5}$ & Yes \\ \hline
OGM-$a$ ($a > 2$)
& $\frac{a}{2} N^{-2}$ & $\frac{a\sqrt{6}}{2\sqrt{a-2}} N^{-1.5}$
& \multirow{2}{*}{No} \\
OGM-$a\!=\!4$
& $2 N^{-2}$ & $2\sqrt{3} N^{-1.5}$ & \\
\hline
\end{tabular}
\caption{
Asymptotic worst-case bounds
on the cost function $\frac{1}{LR^2}(f(\x_N) - f(\x_*))$
and the gradient norm $\min_{i\in\{0,\ldots,N\}}\frac{1}{LR}||\nabla f(\x_i)||$
of GM, FGM, OGM, OGM-\m, OGM-OG, and OGM-$a$.
(The worst-case cost function bound for OGM-\m in the table corresponds
to the bound for OGM after $\m$ iterations,
because we do not have
an analytical bound for the final iterate.}
\label{tab:rate}
\end{table}
\vspace{-10pt}
\subsection{Tight worst-case bounds
on the cost function and the gradient norm}
Since many worst-case bounds presented in Table~\ref{tab:rate}
are not guaranteed to be tight,
we used the code
in Taylor \etal~\cite{taylor:17:ssc}
(with SDP solvers~\cite{Lofberg2004,sturm:99:us1})
to compare tight (numerical) worst-case bounds
for $N=1,2,4,10,20,30,40,47,50$.
These numerical worst-case bounds
are guaranteed to be tight,
\ie, equivalent to the bounds of either~\eqref{eq:PEP} or~\eqref{eq:PEP__},
when the large-scale condition $d\ge N+2$
is satisfied~\cite[Thm.~5]{taylor:17:ssc},
and we assume this condition hereafter.
Tables~\ref{tab:costbound} and~\ref{tab:sgbound}
provide tight worst-case bounds
for the decrease of the cost function $f(\x_N) - f(\x_*)$
and the gradient norm decrease
$\min_{i\in\{0,\ldots,N\}}||\nabla f(\x_i)||$ respectively.
Although most of the bounds in Table~\ref{tab:rate}
are not guaranteed to be tight,
the worst-case rate formulas in Table~\ref{tab:rate}
are similar to the tight numerical results
in Tables~\ref{tab:costbound} and~\ref{tab:sgbound},
except that the gradient bounds of OGM-\m in Table~\ref{tab:rate}
are relatively looser
than those of OGM-\m in Table~\ref{tab:sgbound}.
In particular,
the tight numerical gradient bound of OGM-\m
is smaller than that of FGM in Table~\ref{tab:sgbound},
which was not expected
from their known (possibly loose) analytical bounds in Table~\ref{tab:rate}.
\begin{table}[!h]
\centering
\begin{tabular}{|c|l|l|l|l|l|l|}
\hline
$N$ & GM & FGM & OGM & OGM-\m
& OGM-OG & OGM-$a$
\\
\hline
1 & $6.0$ & $6.0$ & $8.0$ & $6.0$ & $7.3$ & $6.5$ \\
2 & $10.0$ & $11.1$ & $16.2$ & $12.0$ & $13.2$ & $15.1$ \\
4 & $18.0$ & $24.7$ & $39.1$ & $24.2$ & $28.6$ & $32.3$ \\
10 & $42.0$ & $90.7$ & $159.1$ & $86.6$ & $99.9$ & $106.4$ \\
20 & $82.0$ & $283.6$ & $525.1$ & $275.3$ & $310.4$ & $308.9$ \\
30 & $122.0$ & $578.6$ & $1095.6$ & $565.1$ & $604.9$ & $610.9$ \\
40 & $162.0$ & $975.1$ & $1869.2$ & $899.0$ & $1009.9$ & $1012.8$ \\
47 & $190.0$ & $1312.9$ & $2531.1$ & $1227.9$ & $1352.8$ & $1353.6$ \\
50 & $202.0$ & $1472.8$ & $\bm{2845.1}$ & $1374.4$ & $1516.0$ & $1514.6$ \\
\hline
Empi. $O(\cdot)$
& $N^{-1.0}$
& $N^{-1.9}$
& $N^{-1.9}$
& $N^{-1.8}$
& $N^{-1.8}$
& $N^{-1.8}$ \\
\hline
Known $O(\cdot)$
& $N^{-1}$
& $N^{-2}$
& $N^{-2}$
& $N^{-2}$
& $N^{-2}$
& $N^{-2}$ \\
\hline
\end{tabular}
\caption{Tight worst-case bounds on
$\frac{LR^2}{f(\x_N) - f(\x_*)}$,
the reciprocal of the cost function,
of GM, FGM, OGM, OGM-$m\!=\!\fNt$, OGM-OG, and OGM-$a\!=\!4$.
We computed empirical rates
by assuming that the bounds follow the form $bN^{-c}$ with constants $b$ and $c$,
and then by estimating $c$
from points $N=47,50$.
Note that the corresponding empirical rates are underestimated
due to its simple modeling of bounds.
}
\label{tab:costbound}
\end{table}
\begin{table}[!h]
\centering
\begin{tabular}{|c|l|l|l|l|l|l|}
\hline
$N$ & GM & FGM & OGM & OGM-\m
& OGM-OG
& OGM-$a$
\\
\hline
1 & $2.0$ & $2.0$ & $2.0$ & $2.0$ & $2.3$ & $1.8$ \\
2 & $3.0$ & $3.3$ & $2.8$ & $3.5$ & $3.7$ & $3.3$ \\
4 & $5.0$ & $5.9$ & $4.4$ & $6.4$ & $6.8$ & $5.7$ \\
10 & $11.0$ & $13.8$ & $8.9$ & $18.0$ & $18.9$ & $15.3$ \\
20 & $21.0$ & $32.8$ & $16.2$ & $43.1$ & $45.4$ & $35.2$ \\
30 & $31.0$ & $56.4$ & $23.4$ & $74.4$ & $78.6$ & $59.4$ \\
40 & $41.0$ & $83.6$ & $30.6$ & $110.7$ & $116.9$ & $87.1$ \\
47 & $48.0$ & $104.7$ & $35.6$ & $138.9$ & $146.6$ & $108.4$ \\
50 & $51.0$ & $114.2$ & $37.7$ & $151.4$ & $\bm{160.0}$ & $118.0$ \\
\hline
Empi. $O(\cdot)$
& $N^{-1.0}$ & $N^{-1.4}$ & $N^{-0.9}$
& $N^{-1.4}$ & $N^{-1.4}$ & $N^{-1.4}$ \\
\hline
Known $O(\cdot)$
& $N^{-1}$ & $N^{-1.5}$ & $N^{-1}$
& $N^{-1.5}$ & $N^{-1.5}$ & $N^{-1.5}$ \\
\hline
\end{tabular}
\caption{Tight worst-case bounds on
$\frac{LR}{\min_{i\in\{0,\ldots,N\}}||\nabla f(\x_i)||}$,
the reciprocal of the gradient norm,
of GM, FGM, OGM, OGM-$m\!=\!\fNt$, OGM-OG, and OGM-$a\!=\!4$.
Empirical rates were computed
as described in Table~\ref{tab:costbound}.
}
\label{tab:sgbound}
\end{table}
\subsection{Tight worst-case bounds
on the gradient norm at the final iterate}
To be clear,
Sec.~\ref{sec:pep,sgnorm} and Tables~\ref{tab:rate},~\ref{tab:sgbound}
have focused on analyzing the \emph{smallest} gradient norm among all iterates
using the gradient form of PEP,
whereas the gradient analysis in Sec.~\ref{sec:grad}
considers the \emph{final} gradient
in addition to the \emph{smallest} gradient among all iterates.
As mentioned before,
we have not yet found
a relaxation on the \emph{final} gradient form of the PEP
that provides as comparable results
as for the relaxation
on the \emph{smallest} gradient form of the PEP~\eqref{eq:PEP__}
in Sec.~\ref{sec:pep,sgnorm}.
To complete comparisons on the worst-case gradient bounds,
Table~\ref{tab:lgbound} uses the code provided by Taylor~\etal~\cite{taylor:17:ssc}
(with SDP solvers~\cite{Lofberg2004,sturm:99:us1})
to compare tight (numerical) worst-case bounds on the \emph{final} gradient
of the~{\FO}s presented in this paper.\footnote{
Table~\ref{tab:lgbound} reports tight worst-case gradient bounds
for both the \emph{final} primary iterate $\y_N$
and the \emph{final} secondary iterate $\x_N$ (if necessary),
unlike Tables~\ref{tab:costbound} and~\ref{tab:sgbound}.
We observed that numerical tight worst-case cost function bounds
on both final iterates $\y_N$ and $\x_N$
have similar values
for the algorithms in Table~\ref{tab:costbound} (unlike Table~\ref{tab:lgbound}),
so we did not report the bounds on $\y_N$ for simplicity.
We also did not report numerical tight \emph{smallest} worst-case gradient norm bounds
$\min_{i\in\{0,\ldots,N\}}||\nabla f(\y_i)||$
of the primary iterates $\{\y_i\}$
because
the code provided by Taylor~\etal~\cite{taylor:17:ssc}
does not support computing their values,
unlike that of the secondary iterates $\{\x_i\}$ in Table~\ref{tab:sgbound}.
}
The worst-case \emph{smallest} gradient norm bounds
\eqref{eq:g_gm} and~\eqref{eq:g_ogmh}
of GM and OGM-\m respectively (among algorithms considered)
extend to the \emph{final} gradient bounds.
In Table~\ref{tab:lgbound}, FGM and OGM-$a\!=\!4$
have slow $O(1/N)$ tight worst-case bounds on the final gradient,
unlike OGM-$m\!=\!\fNt$ and OGM-OG roughly having $O(1/N^{1.5})$ bounds
for both the smallest and final gradients.
Thm.~\ref{thm:ogmh} has shown that the final gradient of OGM-\m
satisfies a worst-case rate $O(1/N^{1.5})$,
but this is unknown yet for OGM-OG, which we leave as future work.
We also leave as future work the challenge of developing a~\FO
that has $O(1/N^{1.5})$ or even faster worst-case rates
for the final gradient decrease
that are lower than those of OGM-\m and OGM-OG,
possibly without requiring to choose $N$ in advance.
\begin{table}[!h]
\centering
\begin{tabular}{|c|l|l|l|l|l|l|l|l|l|}
\hline
\multirow{2}{*}{$N$}
& \multirow{2}{*}{GM}
& \multicolumn{2}{c|}{FGM}
& \multicolumn{2}{c|}{OGM}
& OGM
& OGM
& \multicolumn{2}{c|}{OGM-$a$} \\
\cline{3-6}\cline{9-10}
& & $\y_N$
& $\x_N$ & $\y_N$
& $\x_N$ & -\m
& -OG & $\y_N$
& $\x_N$ \\
\hline
1 & $2.0$ & $2.0$ & $2.0$ & $2.0$ & $2.0$ & $2.0$ & $2.3$ & $2.0$ & $1.8$ \\
2 & $3.0$ & $3.0$ & $3.3$ & $3.2$ & $2.8$ & $3.5$ & $3.7$ & $3.6$ & $3.3$ \\
4 & $5.0$ & $5.8$ & $5.9$ & $5.5$ & $4.4$ & $6.4$ & $6.8$ & $6.8$ & $5.1$ \\
10 & $11.0$ & $15.1$ & $8.2$ & $11.9$ & $8.9$ & $18.0$ & $18.9$ & $15.9$ & $8.7$ \\
20 & $21.0$ & $25.1$ & $13.1$ & $22.2$ & $16.2$ & $43.1$ & $44.4$ & $26.3$ & $13.8$ \\
30 & $31.0$ & $35.1$ & $18.2$ & $32.4$ & $23.4$ & $74.4$ & $74.1$ & $36.3$ & $18.8$ \\
40 & $41.0$ & $45.2$ & $23.2$ & $42.5$ & $30.6$ & $110.7$ & $107.0$ & $46.3$ & $23.8$ \\
47 & $48.0$ & $52.2$ & $26.7$ & $49.6$ & $35.6$ & $138.9$ & $131.6$ & $53.3$ & $27.3$ \\
50 & $51.0$ & $55.3$ & $28.2$ & $52.6$ & $37.7$ & $\bm{151.4}$ & $142.6$ & $56.3$ & $28.8$ \\
\hline
Empi. $O(\cdot)$
& $N^{-1.0}$ & \multicolumn{2}{c|}{$N^{-0.9}$} & \multicolumn{2}{c|}{$N^{-0.9}$}
& $N^{-1.4}$ & $N^{-1.3}$ & \multicolumn{2}{c|}{$N^{-0.9}$} \\
\hline
Known $O(\cdot)$
& $N^{-1}$ & \multicolumn{2}{c|}{$N^{-1}$} & \multicolumn{2}{c|}{$N^{-1}$}
& $N^{-1.5}$ & $N^{-1}$ & \multicolumn{2}{c|}{$N^{-1}$} \\
\hline
\end{tabular}
\caption{
Tight worst-case bounds on
$\frac{LR}{||\nabla f(\x_N)||}$
$\paren{\text{and } \frac{LR}{||\nabla f(\y_N)||}}$,
the reciprocal of the final gradient norm,
of GM, FGM, OGM, OGM-$m\!=\!\fNt$, OGM-OG, and OGM-$a\!=\!4$.
Empirical rates were computed
as described in Table~\ref{tab:costbound}.
The known bounds of OGM-OG and OGM-$a\!=\!4$
are derived based on Sec.~\ref{sec:grad},
where the empirical bounds of OGM-OG
are comparable to the bounds of OGM-$m\!=\!\fNt$
with known rate $O(1/N^{1.5})$.
}
\label{tab:lgbound}
\end{table}
\subsection{Non-optimality
of OGM-OG in terms of the worst-case gradient bound}
Because OGM is optimal in terms of the function decrease
when $d \ge N+1$~\cite{drori:17:tei},
one might hope that the OGM-OG would achieve
the optimal worst-case bound
in terms of the gradient decrease,
since OGM-OG is also derived by optimizing the step coefficients
over the gradient form of relaxed PEP.
However, the OGM-OG is apparently not optimal as explained next.
Taylor~\etal~\cite{taylor:17:ssc}
numerically studied an optimal fixed-step GM
using their tight PEP
in terms of both the cost function and gradient decrease.
In other words,
they searched for an optimal step $h$ of GM:
\begin{align*}
\x_{i+1} = \x_i - \frac{h}{L}\nabla f(\x_i)
\end{align*}
for $i=0,\ldots,N-1$ and a given $N$
with respect to either $f(\x_N) - f(x_*)$ or $||f(\x_N)||$.
In the special case of $N=1$,
Taylor~\etal~\cite{taylor:17:ssc} numerically conjectured that
the step size $h=1.5$ is optimal in terms of the cost function decrease.
The corresponding GM is equivalent to OGM for $N=1$,
and this (numerically) confirms the optimality of OGM
\cite{drori:17:tei}
for $N=1$.
They also numerically conjectured that
the optimal step size of GM for $N=1$
in terms of the gradient decrease is $h=\sqrt{2}$
with a worst-case bound
\begin{align}
||\nabla f(\x_1)|| \le \frac{LR}{\sqrt{2}+1}
\approx \frac{LR}{2.4}
.\end{align}
However, OGM-OG for $N=1$ reduces to GM with $h=\frac{4}{3}\approx1.3$
with a bound $\frac{LR}{2.3}$ in Tables~\ref{tab:sgbound} and~\ref{tab:lgbound},
implying that OGM-OG is not optimal even for $N=1$
based on the numerical evidence in~\cite{taylor:17:ssc}.
This analysis for $N=1$ illustrates that there is still room for improvement
in accelerating the worst-case rate of first-order methods
in terms of gradients,
which we leave as future work possibly with a tighter relaxation
on the gradient form of PEP.
In addition,
we leave as future work studying
the optimal worst-case bound for the gradient decrease of first-order methods
building upon~\cite{drori:17:tei,nemirovsky:92:ibc},
and developing a~\FO that achieves such optimal bound.
Nevertheless,
the OGM-OG is the best known~\FO for decreasing the gradient norm
among the class~\FO,
and will be useful when decreasing the gradient is key.
\comment{
\subsection{
Generalizing OGM for nonsmooth composite convex problems}
One can extend some aspects
of the approaches for generalizing OGM described in this paper
to other optimization algorithms and problems.
One direction we have already taken in~\cite{kim:18:ala}
aims to improve the fast iterative shrinkage/thresholding algorithm (FISTA)
\cite{beck:09:afi}
(that reduces to FGM for smooth convex problems)
for nonsmooth composite convex problems.
Naturally, this paper and~\cite{kim:18:ala}
use some similar approaches,
but they are different in the following two aspects.
First, the methods in~\cite{kim:18:ala}
when simplified to the smooth case correspond to a generalization of FGM
that differs from the methods proposed here.
Second,~\cite{kim:18:ala} uses a relaxation for PEP
that is looser than the relaxation in this paper
for the smooth convex case,
and thus the worst-case analysis in~\cite{kim:18:ala}
does not lead to an $O(1/N^{1.5})$ worst-case gradient rate for FGM here.
Nevertheless, the FISTA generalizations in~\cite{kim:18:ala}
provide new insights on first-order methods
for nonsmooth composite convex problems,
as this paper does for the smooth convex problem.
Finding a straightforward extension of this paper
to nonsmooth composite convex problems
remains as future work.
}
\section{Conclusion}
\label{sec:conc}
We generalized the formulation of OGM
and analyzed its worst-case bounds on the function value and gradient,
using the cost function form and the gradient form of relaxed PEP.
We then proposed OGM-OG
by optimizing the step coefficients of~\FO using a relaxed PEP
with respect to the gradient,
similar to the development of the (optimal) OGM.
To the best of our knowledge,
the worst-case bound on the gradient of the OGM-OG
is the best known analytical worst-case bound
for decreasing the gradient norm among the class~\FO.
However, this OGM-OG is not optimal
for decreasing the gradient norm,
and further accelerating the worst-case rate of~\FO in terms of the gradient
possibly with a tight relaxation on the gradient form of PEP
is a possible research direction.
On the other hand,
deriving an optimal worst-case bound
for the gradient norm of first-order methods,
similar to that for the function decrease~\cite{drori:17:tei}
will be useful.
Nonetheless, the proposed OGM-OG (and OGM-$a$)
may be useful when one finds minimizing gradients important,
particularly in dual problems.
In addition, we used the proposed gradient form of PEP
to show that FGM decreases the (smallest) gradient
with a rate $O(1/N^{1.5})$,
implying that FGM is comparable in a big-O sense to OGM-\m, OGM-OG and OGM-$a$
for the gradient decrease.
Our analysis considers unconstrained smooth convex minimization;
extending such gradient norm worst-case analysis
to constrained problems or nonsmooth composite convex problems
is a natural direction to pursue,
which is studied for FGM (or FISTA~\cite{beck:09:afi})
by the authors~\cite{kim:18:ala}.
In addition,
extending the analyses on the general form of FGM
in~\cite{attouch:18:fco,chambolle:15:otc,su:16:ade}
to GOGM
is a possible research direction.
Lastly,
investigating a new relaxation of the PEP approach
that allows adaptive step size
such as backtracking line-search
or exact line-search~\cite{deklerk:17:otw,drori:18:efo}
is of interest.
\section*{Software}
\url{https://gitlab.eecs.umich.edu/michigan-fast-optimization}
has
Matlab codes for the algorithms considered and
the SDP approaches
in Sec.~\ref{sec:opt,grad,pep} and Sec.~\ref{sec:disc}.
|
\section{Introduction}
Since the launch in June 2008 of {\it Fermi} LAT \cite{Atwood2009}, a high-energy (HE) satellite measuring $\gamma$-rays in the range 20 MeV$-$300 GeV, two pulsar catalogues (1PC, \cite{Abdo2010}; 2PC, \cite{Abdo2013}) discussing the light curve and spectral properties of 117 pulsars have been released. The vast majority of the {\it Fermi}-detected pulsars display exponentially cutoff spectra with cutoffs around a few GeV. These spectra are believed to be due to curvature radiation (CR), which is assumed to be the dominating emission process in the GeV band (see Section~\ref{subsection:radiationmechanisms}).
\subsection{Standard pulsar emission models}\label{subsection:emissionmodels}
There exist several physical radiation models that can be used to study HE emission from pulsars. These include the polar cap (PC; \cite{Daugherty1982}), slot gap (SG; \cite{Arons1983}), outer gap (OG; \cite{Cheng1986b}), and the pair-starved polar cap (PSPC; \cite{Harding2005}) models, which can be distinguished from each other based on the different assumptions of the geometry and location of the `gap regions'. The `gap region' is where particle acceleration takes place due to an unscreened, rotation-induced $E$-field parallel to the local $B$-field, as well as subsequent emission by these particles.
In PC models emission from HE particles is assumed to originate close to the neutron star (NS) surface. These particles are accelerated by large $E$-fields near the magnetic poles (known as the magnetic PCs) only up to a few stellar radii. In SG models, the radiation comes from narrow gaps close to the last open field lines (the field lines that are tangent to the light cylinder where the corotation speed equals the speed of light $c$), with the gaps extending from the NS surface up to high altitudes. In the OG model, the gap region extends from the null-charge surface, where the Goldreich-Julian charge density is zero \cite{Goldreich1969} up to high altitudes, also close to the last open field lines. The PSPC model involves a gap region that extends from the NS surface to the light cylinder over the full open volume \cite{Harding2005}, since the potential is unscreened in this case, so that there are not enough pairs to fully screen the $E$-field.
\subsection{Radiation and pair creation processes}\label{subsection:radiationmechanisms}
To explain HE emission in the standard models, one has to take detailed particle transport and radiation mechanisms into account. These mechanisms include CR, synchrotron radiation (SR), and inverse Compton scattering (ICS). CR occurs whenever charged particles are constrained to move along curved paths, e.g., along curved $B$-field lines (e.g., \cite{Harding1981}), therefore involving a change in their longitudinal kinetic energy. When the emitted CR photon energy and the local $B$-field are high enough, magnetic pair production may occur (where an HE photon converts into an electron-positron pair, $e^\pm$), leading to a cascade of $e^\pm$ pairs which may screen the parallel $E$-field outside the gaps. The pair cascade is characterized by the so-called multiplicity, i.e., the number of pairs spawned by a single primary. The pairs may radiate SR if they have velocity components perpendicular to the local $B$-field so that this process involves a change in the particles' transverse kinetic energy. Also, ICS occurs due to the relativistic particles which upscatter soft photons (e.g., originating at a heated PC), which results in the ``boosting'' of the photon energies up to very high energies. ICS photons may also be converted into $e^\pm$ pairs. Two-photon pair creation ($\gamma\gamma$-absorption) may also occur, in particular in OG models. Cyclotron emission combined with subsequent ICS has also been considered by \cite{Lyutikov2013} to explain the broadband spectrum of the Crab pulsar.
\subsection{Historic perspective of VHE spectral modelling}
\begin{figure}[t]\centering
\includegraphics[width=22.3pc,height=13pc]{figure1.eps}\hspace{0.5cm}
\caption{\label{Romani96} Early prediction of the phase-averaged spectrum for the Vela pulsar. The solid lines represent the spectral components for CR, SR, and the thermal surface flux (kT). The dashed curve represents the TeV pulsed spectral component associated with the ICS of SR of the primary $e^\pm$ (SSC). Adapted from \cite{Romani1996}.}
\end{figure}
Early modelling, assuming the standard OG model, predicted spectral components in the VHE regime when estimating the ICS of primary electrons on SR or soft photons. This resulted in a natural bump around a few TeV (involving $\sim 10$ TeV particles) in the extreme Klein-Nishina limit as seen in figure~\ref{Romani96} and \ref{Hirotani01}. However, these components may not survive up to the light cylinder and beyond \cite{Cheng1986b,Romani1996,Hirotani2001a}, since $\gamma\gamma$ pair creation leads to absorption of the TeV $\gamma$-ray flux.
\begin{figure}[t]\centering
\includegraphics[width=20pc,height=15pc]{figure2.eps}\hspace{0.5cm}
\caption{\label{Hirotani01} Expected TeV spectra for three bright pulsars including the Crab (solid lines), PSR B0656$+$14 (dashed lines), and PSR B1509$-$58 (dotted lines). The thick and thin curves represent inclination angles (between the spin and magnetic axis) of $\alpha=30^\circ$ and $45^\circ$, respectively \cite{Hirotani2001a}.}
\end{figure}
Other studies assumed CR to be the dominant radiation mechanism producing $\gamma$-ray emission when performing spectral modelling and found spectral cutoffs of up to 50 GeV. For example, \cite{Bulik2000} modelled the cutoffs of millisecond pulsars (MSPs). These are pulsars which possess relatively low $B$-fields and short periods. Their model assumed a static dipole $B$-field and a PC geometry, and predicted CR from the primary electrons that are released from the PC. Their predicted CR spectral component cut off at $\sim100$~GeV. The CR photons may undergo pair production in the intense low-altitude $B$-fields, and the newly formed electron-positron secondaries will emit SR in the optical and X-ray band (see \cite{Harding2002}). Therefore, they concluded that the HE CR from MSPs occurred in an energy band that was above the detection range of satellite detectors like \textit{Energetic Gamma-Ray Experiment Telescope (EGRET)} and below that of ground-based Cherenkov detectors such as the \textit{High Energy Stereoscopic System (H.E.S.S. I)}.
Later studies investigated the X-ray and $\gamma$-ray spectrum of rotation-powered MSPs using a PSPC model \cite{Harding2005}, and found CR cutoffs of $\sim10-50$~GeV (see \cite{Frackowiak2005,Venter2005}). Optical to $\gamma$-ray spectra were also modelled by \cite{Harding2008} assuming an SG accelerator and a retarded vacuum dipole (RVD) $B$-field, for the Crab pulsar. They found spectral cutoffs of up to a few GeV. Another study modelled the phase-resolved spectra of the Crab pulsar using the OG and SG models, and found HE cutoffs of up to $\sim25$~GeV \cite{Hirotani2008}. Cutoffs around $\sim10$~GeV were found for the OG model using the RVD $B$-field \cite{Tang2008}.
\section{Observational revolution}
In view of the above theoretical paradigm it was not expected that a pulsar should be visible in the VHE regime. It was therefore surprising when the \textit{Very Energetic Radiation Imaging Telescope Array System (VERITAS)} announced the detection of pulsed emission from the Crab pulsar above $\sim100$~GeV \cite{Aliu2011}, followed by the detection by the \textit{Major Atmospheric Gamma-Ray Imaging Cherenkov (MAGIC)} of emission up to $\sim400$~GeV (soon after their initial detection of emission at $\sim25$~GeV) \cite{Aleksic2012,Aliu2008}. The \textit{MAGIC} Collaboration has since reported the detection of pulsed photons with energies up to 1.5~TeV \cite{Ansoldi2016}. Ground-based Cherenkov telescopes are now searching for more examples of VHE pulsars.
\begin{figure}[t]\centering
\includegraphics[width=18pc,height=18pc]{figure3.eps}\hspace{0.5cm}
\caption{\label{MAGIC} The observed and modelled spectra for phase-resolved $\gamma$-ray emission, including both emission peaks P1 and P2, from the Crab pulsar as measured by \textit{MAGIC} (dark red squares). The plot also contains measurements from \textit{Fermi} LAT (orange diamonds) and \textit{VERITAS} (light blue squares and solid line). An OG model including emission from pairs are assumed. The systematic error of the \textit{MAGIC}-Stereo measurement is shown and corresponds to a shift of $\pm$17\% in energy and $\pm$19\% in flux. From \cite{Aleksic2012}.}
\end{figure}
In figure~\ref{MAGIC} the phase-resolved $\gamma$-ray spectrum for the Crab pulsar for both light curve peaks P1 and P2, as measured by \textit{MAGIC}, is shown. The spectrum also contains measurements by \textit{Fermi} LAT and \textit{VERITAS}. The observations by \textit{MAGIC} are in good agreement with those by \textit{VERITAS}. The \textit{VERITAS} data fit a broken power law, and raised important questions whether the CR component is extended or if a second component is required to explain the observed spectra.
The detection of the Crab pulsar above several GeV prompted {\it Fermi} to search for pulsed emission at HEs. They detected significant pulsations above 10~GeV from 20 pulsars and above 25~GeV from 12 pulsars \cite{Ackermann2013}. The Crab pulsar is the first source which have been detected over almost all energies ranging from radio to VHE $\gamma$-rays. More recently, pulsed emission was detected from the Vela pulsar above 30 GeV with the \textit{H.E.S.S.} \cite{Stegmann2014} and up to $\sim80$~GeV (at the $4\sigma$-level) with the \textit{Fermi} LAT \cite{Leung2014}. \textit{MAGIC} furthermore detected no emission from Geminga above 50 GeV~\cite{Ahnen2016}; neither did \textit{VERITAS} above 100~GeV~\cite{Aliu2015}. A stacking analysis involving 115 {\it Fermi}-detected pulsars (excluding the Crab pulsar) was performed by \cite{McCann2015} using {\it Fermi} data. However no emission above 50 GeV was detected, implying that VHE pulsar detections may be rare, given current telescope sensitivities. From all these observations there are three effects visible in the energy-dependent pulse profiles: the peaks remain at the same phase, the P1/P2 ratio decreases as energy increases, and the pulse width decreases with increasing energy. More VHE pulsars may be found by the \textit{Cherenkov Telescope Array (CTA)} which will have a ten-fold increase in sensitivity compared to present-day Cherenkov telescopes.
\section{Theoretical ideas}
All of the standard pulsar emission models (see Section~\ref{subsection:emissionmodels}) predicted HE spectral cutoffs between a few GeV and up to $\sim100$~GeV, assuming $B$-fields such as the static dipole and RVD solutions. Clearly, refinements to these radiation models and $B$-fields are needed to explain the observed VHE emission from the Crab. There are a few ideas for such refinements. One is a revised OG model by \cite{Hirotani2008} which can produce IC radiation of up to $\sim400$~GeV due to secondary and tertiary pairs upscattering infrared to ultraviolet photons \cite{Aleksic2012}. In this OG model the IC flux depends sensitively on the $B$-field structure near the light cylinder.
Another idea was proposed by \cite{Lyutikov2012}, invoking the SSC radiation process. This is indeed a promising radiation mechanism, where relativistic particles upscatter the SR photons emitted by the same population. The SSC radiation mechanism was applied by \cite{Harding2015} to predict optical to X-ray, and $\gamma$-ray spectra (see figure~\ref{SSC}) assuming an SG model and a force-free $B$-field. This process relies critically on the assumed electrodynamics and the magnetospheric structure. They performed simulations for the Crab and Vela pulsars, as well as two MSPs, i.e., B1821$-$24 and B1937$+$21. However, the only significant predicted SSC component was for the Crab pulsar. They also found that the pair SR matched the observed X-ray spectrum of the MSPs. They furthermore tested the addition of an HE power law extension to the pair spectrum (dashed lines in figure 4) whose SR spectrum would account for the observed emission in the $1-100$~MeV range. However, the resulting SSC component exceeded the observed \textit{MAGIC} and \textit{VERITAS} points, implying that the observed $1-100$~MeV emission is not produced by the same particles that produce the SSC emission.
\begin{figure}[t]\centering
\includegraphics[width=22.3pc]{figure4.eps}\hspace{0.5cm}
\caption{\label{SSC} Modelled spectrum of phase-averaged pulsed emission from the Crab pulsar. The spectral components from primary electrons and pairs (as labelled) are for a magnetic inclination angle $\alpha=45^\circ$, observer angle $\zeta= 60^\circ$ and pair multiplicity $M_{+} = 3\times10^5$. From \cite{Harding2015}.}
\end{figure}
\section{Conclusions and future prospects}
The abovementioned detections of the VHE pulsed emission from pulsars and the explanation thereof implies that this emission may yield strong constraints on $\gamma$-ray radiation mechanisms, the location of acceleration regions, and the $B$-field structure. There is thus an urgent need for refinements and extensions of standard pulsar emission models and radiation mechanisms, including more realistic $B$-fields. Some examples are discussed above, but there are many more.
SG model refinements include photon-photon pair production attenuation within the model and also more realistic $B$-fields such as dissipative magnetospheric solutions. One could also model the emission pulse profiles as a function of energy.
Ground-based Cherenkov telescopes are now searching for more examples of VHE pulsars. New pulsar models will assist us in predicting the level of VHE emission expected from them, which would be very important for the upcoming \textit{CTA}. The low threshold energy of \textit{CTA} will provide an overlap with the {\it Fermi} energy range and will help to discriminate between CR and a potentially new spectral component.
\ack
This work is based on research supported by the National Research Foundation (NRF) of South Africa (Grant Numbers 90822, 93278, and 99072). Any opinions, findings, and conclusions or recommendations expressed are that of the authors, and the NRF accepts no liability whatsoever in this regard. A.K.H.\ acknowledges the support from the NASA Astrophysics Theory Program.
\section*{References}
|
\section{Introduction}\label{Intro}
Over the last few decades, and in particular in the last 10 to 15 years, our understanding of active galactic nuclei (AGN), their underlying physical mechanisms, their environments, and their observational properties, has greatly increased. Although the unification model proposed by \citet{Antonucci1993} still holds true in many aspects, subsequent revisions \citep[see e.g.][]{Netzer2015} illustrate what we have learned about the structure of the obscuring torus, the mechanisms that provide feedback, the variability timescales involved, and where the radio-loud sources fit (or do not fit) in the grand AGN unification scheme. We are living in what could be considered a golden era of surveys, which allow us, for the first time, to construct large, consistent, multiwavelength samples of AGN with the potential to push our understanding of these objects even further.
Although only $\sim$10--20 per cent of the AGN we observe are classified as radio-loud, recent evidence shows that jets and lobes could be far more ubiquitous than we previously thought. There is an increasingly large number of Seyfert galaxies, and even QSOs, where jets and lobes, or excess radio emission, have been detected \citep[e.g.][]{Hota2006,Gallimore2006,DelMoro2013,Singh2015,Harrison2015}, throwing into question the radio-loud/quiet classification, which, being based on optical (B band) to radio (5 GHz) flux ratios \citep[e.g.][]{Kellerman1989}, classifies most of these objects as radio-quiet. This `jet mode' or `radio mode' is fundamental to our understanding of the AGN/host relationship, not only for very powerful sources in clusters, where the jet-driven shocks can offset radiative cooling of the gas \citep[e.g.][]{McNamara2012,HLarrondo2015}, but especially for low power sources ($L_{1.4 GHz} \le 10^{23}$ W Hz$^{-1}$ sr$^{-1}$), because it is in these systems that the effect of the AGN on the surrounding interstellar gas (on 10--100 kpc scales) can have the largest potential impact on the evolution and star formation history of the host galaxy \citep[e.g.][]{Cattaneo2009,Croston2011,Mingo2011,Mingo2012}.
Radio-loud sources are also useful in that they allow us to unequivocally identify sources in the radiatively inefficient accretion regime \citep{Narayan1995}. This population, originally identified as low excitation radio galaxies (LERGs) by \citet{Hine1979}, lacks the `traditional' AGN disc and torus, shows very low Eddington rates \citep{Hardcastle2007b,Hardcastle2009,DeGasperin2011,Best2012,Mingo2014,Paggi2016}, and seems to be channeling most of the gravitational energy into jets rather than radiative output, in a similar manner to the low/hard state of low mass X-ray binaries \citep[see e.g. the review by][]{Fender2014}. In the optical, radiatively inefficient sources are typically classified as LINERs (low ionisation nuclear emission line regions) as initially proposed by \citet{Heckman1980}, or even appear as fully `quiescent' (i.e. not containing an AGN) galaxies \citep[see e.g.][]{Kimball2008}. This classification, however, is misleading, in the sense that other processes such as shocks or emission from an old stellar population can also produce low ionisation spectra \citep[see e.g.][and references therein]{Balmaverde2015}. Therefore, finding low ionisation optical emission lines does not guarantee the presence of a radiatively inefficient AGN, while finding active radio jets in an otherwise `quiescent' looking galaxy does. As radiatively inefficient AGN only produce soft X-rays related to the jet \citep[e.g.][]{Hardcastle1999}, their typical X-ray luminosity is $10^{39}-10^{41}$ erg/s, which precludes them from being included in most X-ray selected AGN surveys.
Recent results show that the interplay between AGN activity, outflows, and star formation may be more complex than we previously thought, and fundamental to understanding galaxy evolution and black hole growth \citep[e.g.][]{Alexander2012,Magliocchetti2014,Davies2014}. Although we are beginning to better understand the transition between the regimes in which AGN and star formation activity dominate, and how radio AGN activity, in particular, affects star formation \citep[e.g.][]{Smolcic2009,Dicken2012,DelMoro2013,Hardcastle2013b,Kalfountzou2014,Villarroel2014,Gurkan2015,Rawlings2015,Hardcastle2016,Drouart2016,Tadhunter2016}, there is still a distinct lack of agreement on how and when AGN activity influences star formation \citep{Harrison2012,Ishibashi2012,Symeonidis2013,Symeonidis2014,AlonsoHerrero2013,Heckman2014,Balmaverde2016,Brusa2015,Rosario2013,Rosario2015,Stanley2015,Bernhard2016,Alberts2016}. Although the large timescales involved probably cause part of this confusion \citep{Georgakakis2008,Wild2010,RamosAlmeida2013,Best2014}, and it is clear that we still do not fully understand long AGN variability timescales \citep[see e.g.][]{Hickox2014}, it is also true that dedicated samples that encompass sources in both regimes, as well as the transition, still tend to be limited either in wavelength, scope, redshift, or size.
Obtaining large multiwavelength samples of radio-loud AGN is challenging for several reasons, the main two being the extended nature of radio emission and the low sky density of radio-loud AGN, and the number of sources decreases rapidly if selections in more than two bands are required. These surveys also tend to focus on particular populations of radio-loud AGN (or star-forming galaxies). There is a wealth of on-going and upcoming instruments and surveys that will open a wide field of potential exploration in both fields: LOFAR, SKA, e-MERLIN, JVLA, in the radio; e-Rosita, and Athena in the X-rays, CTA in the gamma-ray band, LSST, and JWST and ALMA at infrared and sub-mm wavelengths, respectively. Now is the perfect time to assess which questions our current data can and cannot answer, to set a framework and potential diagnostic tools for the next generation of results.
The ARCHES FP7 collaboration\footnote{\url{http://www.arches-fp7.eu/}} is a project dedicated to fully exploiting the capabilities of the 3XMM catalogue of X-ray sources, by creating multiwavelength products (cross-correlated catalogues and tools, spectral energy distributions, and a cluster catalogue and finder tool). As part of this collaboration, we have built and describe in this paper the MIXR sample: a systematic, large sample of sources detected in the Mid-IR (WISE all-sky survey), X-rays (3XMM DR5) and Radio (FIRST/NVSS). By requiring a detection in all three bands, we find a wide range of populations: from radiatively inefficient (LERG/LINER) systems in otherwise quiescent galaxies, to low luminosity Seyfert-like sources where the host emission dominates in some bands, to nearby starburst objects, to high luminosity radio-loud and radio-quiet Seyferts and QSOs. The MIXR sample allows us to derive efficient diagnostics for star formation and AGN activity (both radiatively efficient, as seen in `traditional' AGN, and radiatively inefficient, as seen in LERG/LINER), even in host-dominated sources that are normally considered quiescent and discarded from most mid-IR and X-ray AGN samples. We also test the radiative (luminosity) versus kinetic (jet) output in our AGN, to explore the extent and possible causes for the scatter we observed in \citet{Mingo2014}, in contradiction with the well-known correlation of \citet{Rawlings1991}. Our analysis also helps us pinpoint several sources of bias that affect selections performed in one or more of the bands we use, helping us better understand what AGN populations are included and excluded in each selection.
In section \ref{Data} we discuss in detail the MIXR sample construction. In section \ref{Diagnostics} we use WISE colours to pre-classify the sources, and carry out a series of early diagnostics to test these classifications, using hardness ratios, radio versus X-ray `loudness', and flux/magnitude diagrams. In section \ref{z} we add redshift information from SDSS, which we use in section \ref{LDiag} to derive luminosities for the MIXR sources, and extend our diagnostics to verify the underlying type of activity for the MIXR sources. In section \ref{RL_RQ_section} we re-classify the sources based on their activity (radio-quiet and radio-loud AGN, including LERGs/LINERs, and galaxies). For sections \ref{Eddington} and \ref{Power} we focus on the AGN, assessing their Eddington rates and their radiative versus kinetic (jet) output, to highlight the strengths and limitations of current surveys, and address some of the open questions on AGN variability and its impact on the AGN/host relationship.
For this work we have used the latest cosmological values released by the Planck collaboration \citep{Planck2015}: $H_{0} = 67.74$ km/s/Mpc, $\Omega_{m} = 0.3089$, and $\Omega_{\Lambda} = 0.6911$. The catalogue we describe in this paper is available on-line for download at \url{http://www.arches-fp7.eu/index.php/tools-data/downloads/mixr-catalogue} and will be made available on VizieR.
\section{Data and Sample Construction}\label{Data}
Our aim is to select a large, clean (i.e. avoiding mis-classifications, but also contaminants, such as stars) sample of sources, with data that will allow us to characterise the accretion properties of the AGN population, as well as to explore the extent of star formation present; we need large, uniform surveys at wavelengths where AGN and star formation activity can be detected unequivocally: Mid-Infrared (3.4--12 $\mu$m), X-rays (0.2--10 keV) and Radio (1.4 GHz) (MIXR).
X-ray and mid-IR emission are very good probes of accretion in AGN, the former being produced in the accretion disc and hot corona in the inner regions of the AGN, and the latter being the region of the spectrum where the bulk of the thermal (blackbody) emission from the dusty torus peaks \citep[see e.g.][]{Horst2008}. Although it is possible to obtain clean selections of samples using only X-ray and mid-IR data, some caution must be applied to eliminate X-ray binaries and galaxies with X-ray and infrared emission associated with star formation, rather than an AGN. The process typically involves cuts in mid-IR colours and X-ray hardness ratio \citep[e.g.][]{Assef2010,Assef2013,Stern2012,Mateos2012,Rovilos2014}. While these studies are extremely successful in characterising the properties of QSO-like and bright Seyfert-type sources, they cannot include fainter galaxies, where AGN emission cannot be detected unequivocally, as well as radiatively inefficient (LINER or LERG) sources where most of the energy is channelled through a jet, rather than as radiative output \citep[see e.g][]{Best2012,Hardcastle2009,Mingo2014}. It is also important to keep in mind that a strict hardness ratio cut can eliminate sources with a soft excess, most relevantly radio-loud AGN, where jet-related emission produces soft X-rays.
An additional radio selection could prove very advantageous in this context. There are two mechanisms that can produce bulk radio emission: star formation \citep[free-free emission from HII regions, some synchrotron radiation from particle acceleration in winds and supernova explosions, and some thermal emission from cold gas and dust, plus HI at 21cm, see e.g.][]{Harwit1975,Condon1992} and AGN activity (synchrotron radiation from jets, hotspots and lobes). While thermal emission from dust becomes very relevant at higher frequencies ($>$10 GHz), low to intermediate frequency radio production from star formation is remarkably inefficient \citep[e.g.][]{Bell2003} and is generally detected only for very nearby starburst galaxies. It is now known that star formation-related radio emission may be more significant at sub-mJy level \citep{Padovani2011A,Bonzini2013} \citep[see also the recent LOFAR results of][]{Williams2016}, but even at low fluxes AGN processes may still dominate the emission in systems with moderate star formation, rather than powerful starbursts \citep{White2015}. Recent evidence also shows that the fraction of radiatively efficient (`traditional', radiative mode, IR and X-ray bright) AGN that show accretion-related radio emission inversely correlates with radio power \citep{Padovani2015}. AGN radio emission is also unaffected by obscuration, and thus relatively unbiased with respect to orientation \citep[there is a slight bias towards favouring core-dominated, face-on sources, see e.g. the discussion by][]{Mingo2014}.
Aside from the obvious bias introduced by selecting only sources that produce radio emission, the main downside of requiring radio detections is a substantial reduction in the number of sources in the final sample, given the low sky density of the radio sky. This disadvantage, however, is more than made up for by the fact that, without any additional filtering, a combination of radio, mid-IR and X-rays can produce a clean, uniform selection of AGN across all luminosities, host types and accretion modes, as well as identifying nearby starburst galaxies. As such, our study focuses mostly on AGN, but the star-forming galaxies provide the necessary framework to quantify star formation and AGN activity in sources where both contributions are hard to disentangle.
Although using radio data would guarantee a very clean selection of extragalactic sources \citep[the number of individual stars identified at 1.4 GHz is very low, see e.g.][]{McMahon2002,FIRST2015} we decided to minimise the incidence of Galactic sources across all catalogues by imposing a high Galactic latitude cut ($|b_{II}| \ge 20^{\circ}$).
Most multi-survey samples use positional matching techniques to cross-correlate the sources across the different catalogues. For MIXR we have used the statistical \textsc{xmatch} cross-correlation tool developed for the ARCHES collaboration, described in more detail in section \ref{Sample}, which allowed us to quantitatively, efficiently and simultaneously establish the source associations across the three catalogues we used to create MIXR.
\subsection{X-rays: 3XMM}\label{3XMM}
For our X-ray data we have used the DR5 release of the 3XMM catalogue \citep{Rosen2016}. This catalogue comprises results from 7781 individual pointings taken between February 2000 and the end of 2013, resulting in 565962 individual detections and 396910 unique sources covered by XMM-Newton's EPIC cameras (pn, MOS1, MOS2) in the 0.2--12 keV band, making it the largest X-ray catalogue ever produced. The sources in 3XMM are resolved on scales of $\sim6$ arcsec, and the typical positional error is $\sim1.5$ arcsec.
The fluxes in the catalogue are calculated for 5 bands (0.2--0.5, 0.5--1, 1--2, 2--4.5, and 4.5--12 keV) from the count rate of each instrument in each individual observation \citep[see][]{Mateos2009,Watson2009,Rosen2016}, and, when more than one observation per source exists, combined for each source using weighted average based on the flux errors. For the diagnostic plots in section \ref{HR} we have combined the fluxes in the first three bands (0.2--2 keV) to obtain the soft X-ray flux, and the fluxes of the fourth and fifth band (2--12 keV) for the hard X-ray flux.
As 3XMM has the smallest sky area of all our catalogues ($\sim800$ square degrees), it is the limiting factor in this respect. However, X-ray observations are essential to diagnose AGN activity, particularly in complex samples such as ours. Sources that appear extended in the 3XMM catalogue, at high Galactic latitudes, typically fall under two categories: very nearby galaxies or relatively nearby galaxy clusters \citep[the intracluster gas is very hot and typically emits soft X-rays, with a spectral shape that is a combination of bremmstrahlung, recombination and 2-photon radiation, and peaks around 2--5 keV, see e.g.][]{Boehringer2010,Ineson2013,Ineson2015}. By eliminating extended X-ray sources from our sample we can avoid some of the potentially problematic sources in our final sample. Given the method we used to combine FIRST sources (see section \ref{Radio}), this also minimises any cases in which we might have combined radio components from distinct sources in the same cluster.
After eliminating extended sources, and those with very low detection probabilities (by imposing SC\_EXTENT $\le0.0$, and SC\_DET\_ML $\ge10$), as well as applying the high Galactic latitude cut ($|b_{II}| \ge 20^{\circ}$), we are left with $\sim150000$ 3XMM sources.
It is worth mentioning that 3XMM is not corrected for pile-up effects. Pile-up is only relevant for fluxes above $10^{-11}$ erg cm$^{-2}$ s$^{-1}$. Given the flux distribution of our sources, and the fact that we are using catalogue fluxes, rather than performing full spectroscopic fits, we do not expect pile-up issues to affect our results.
\subsection{Mid-IR: WISE all-sky catalogue}\label{WISE}
The WISE catalogue covers the entire sky in four mid-IR bands, 3.4, 4.6, 12, and 22 $\mu$m (W1 to W4, respectively), with a spatial resolution of 6.1--6.5 arcsec for the first three bands, and 12 arcsec for the fourth band \citep{WISE2010,WISE2011}. As such, it is ideally suited to probe AGN activity, and to characterise the host galaxies of systems where AGN activity is not the dominant source of emission in one or more of the bands selected in our catalogue. Given the lower sensitivity and larger pass band of W4, we have focused our analysis on the first three WISE bands, imposing a signal/noise cut of 5 on W1, W2, and of 3 for W3.
For our work we have used the allWISE IPAC release from November 2013 \citep{Cutri2014}, adding up to a total of nearly 750 million sources. As the entire WISE catalogue is very large, and given that our statistical cross-matching tool requires matching sky areas between all catalogues (see section \ref{Sample}), we worked with a subset of WISE sources, obtained by uploading the list of $\sim150000$ pre-selected (see section \ref{3XMM} for the selection criteria) 3XMM source positions to the IPAC allWISE query form\footnote{\url{http://irsa.ipac.caltech.edu/cgi-bin/Gator/nph-scan?submit=Select&projshort=WISE}}, and searching for WISE sources within 60 arcsec of each 3XMM source. The average separation between 3XMM sources (in our sample) is of the order of twice that value; hence such a selection radius guaranteed that the WISE cutout would cover roughly the same sky area as our 3XMM pre-selection. The resulting WISE subset has $\sim$ 1.8 million sources, a small fraction of the original number. We carried out the WISE signal/noise cuts after cross-correlating the catalogues, to keep as many sources for as long as possible. The S/N cuts in W1 and W2 do not reduce the number of sources by a large amount ($\sim8$ per cent), but W3 is far less sensitive, and even a required detection on a 3$\sigma$ level, rather than 5$\sigma$, cuts our sample size by half. While the 12$\mu$m band is essential to characterise the AGN emission related to the torus, there is a large amount of diagnostics and science that can be carried out simply with W1 and W2, particularly for sources not dominated by the AGN.
For the following sections we have considered, separately, those sources that pass the signal to noise cut in W1 and W2, but not W3 (Full Sample, Full Redshift Sample, see section \ref{z}), and those that pass also the W3 signal to noise cut (W3 Sample, W3 Redshift Sample).
The vast majority of the WISE sources in our samples are classified as point-like in the catalogue (ext\_flg=0), and in those cases we have used the standard apertures provided (w1mpro, w2mpro, w3mpro). For sources labelled as extended (ext\_flg=3 and ext\_flg=5, $\sim0.6$ per cent and $\sim17$ per cent of our sources, respectively) we used the provided 2MASS corrected elliptical apertures (w1gmag, w2gmag, w3gmag) instead of the standard apertures, as suggested in the on-line documentation, as they are likely to give more accurate results. We also checked the quality flags for potential problems, and found them to be good after the S/N cuts were implemented.
\subsection{Radio: combining FIRST and NVSS}\label{Radio}
At low to intermediate radio frequencies, specifically at 1.4 GHz, there are two large radio surveys that would be ideally suited for our purposes: FIRST \citep[Faint Images of the Radio Sky at Twenty centimetres][]{FIRST1995,FIRST2015} and NVSS \citep[NRAO VLA Sky Survey][]{NVSS1998}. FIRST has higher spatial resolution ($\sim5$ arcsec, versus $\sim45$ arcsec for NVSS) and thus better positional accuracy, and it is deeper (1 mJy detection level, versus 2 mJy for NVSS) but it covers a smaller area of the sky, as it was designed to coincide with the Sloan Digital Sky survey ($\sim10500$ square degrees, while NVSS covers 82 per cent of the sky, all the area north of $\delta=-40^{\circ}$). Most importantly, given the frequently extended and multi-component nature of radio sources, FIRST can split sub-components from resolved FRI and FRII \citep{FR1974} galaxies into several catalogue entries, which could create false matches in a cross-correlation with a higher sky density catalogue (e.g. independent optical sources matched to the core and lobes of the same radio galaxy) and, in some correct matches, yield only partial integrated fluxes (e.g. the radio core is matched to a counterpart at other wavelengths, but the lobes are not). The lower resolution of NVSS can avoid these problems, but when combined with its lower positional accuracy, especially at low fluxes, it can result in off-centre positions and inaccurate errors, which in turn increase the risk of missed matches. We also want to reach the lowest possible fluxes, to include objects with small jets and lobes that would normally be classified as radio-quiet, as well as objects where the radio emission is produced by star formation, rather than AGN activity.
There have been some notable attempts to combine these two radio surveys before. The Unified Radio Catalogue of \citet{Kimball2008,Kimball2014} also includes sources from the Green Bank 6 cm (GB6) and Westerbork Northern Sky (WENSS) 92 cm survey, which could potentially be useful to estimate the spectral indices of some sources, but it only provides lists of possible counterparts for each entry in those catalogues, leaving to the user how to group and use them. The catalogue of \citet{Best2005}, and its later improved version by \citet{Donoso2009}, uses a very reliable method to group multi-component sources, but relies on prior assumptions by first cross-matching the NVSS sources with Sloan (SDSS) detected optical sources. As a general rule, we prefer to avoid imposing prior cuts on the data, as it is possible to assess the nature of the matches at a later stage and minimise the bias. We therefore decided to combine NVSS and FIRST using criteria that would suit our specific purposes.
We used the latest version of NVSS, which contains 1773484 sources with integrated fluxes and positional errors, and is available through the NVSS public ftp server\footnote{\url{ftp://nvss.cv.nrao.edu/pub/nvss/CATALOG/FullNVSSCat.text}}, and through VizieR. For FIRST we used the March 2014 release, which includes 946432 sources\footnote{\url{http://sundog.stsci.edu/first/catalogs/readme_14mar04.html}}. The catalogue does not provide specific position errors, so we used uncertainties of 0.5 arcsec for fluxes larger than 3mJy, and 1 arcsec at lower fluxes, as suggested by the on-line documentation. While this is likely an overestimation, the positional errors are still much smaller than those in NVSS, and comparable to those in the X-ray survey we will be using (see section \ref{3XMM}).
The first step was to choose a suitable way to combine potential FIRST sub-components, in cases with and without an NVSS potential match. Although the clustering of the radio sky is low \citep[see e.g.][]{Magliocchetti2016b}, and the likelihood of finding more than one (relatively powerful) radio galaxy in a cluster is also low, there is always a risk of grouping distinct sources into one. We chose a conservative grouping radius, 30 arcsec, which is compatible with separations found in earlier studies based on FIRST \citep[e.g.][]{Cress1996,Magliocchetti1998,Gubanov2003}. These studies also explored the clustering of the radio sky at larger radii, finding typical scales of a few arcmin.
For very nearby sources this 30 arcsec radius might prove too small to include all the subcomponents of an extended radio source. Our selection criteria for the X-ray sources, however, minimise this problem (we excluded extended X-ray sources, see section \ref{3XMM}). The X-ray selection excludes not only nearby clusters, but also nearby galaxies, which may appear extended in 3XMM, thus we minimise the presence of very nearby radio sources that might have bright radio components with very large separations on the sky. We decided to use a single collapsing radius, rather than the flux-dependent approach used by \citet{Magliocchetti1998}, for simplicity, as using larger radii for bright FIRST sources would have required us to group NVSS sources as well (the beam used for NVSS had a 45 arcsec FWHM, so for separations around and beyond that scale the possibility of NVSS groups would have to be considered), and would have badly impacted the positional accuracy of the resulting combined sources.
\begin{figure}
\centering
\includegraphics[width=0.46\textwidth]{RadioFlux}
\caption{Ratio between the integrated flux from FIRST (red) or the combined catalogue (blue, overlaid) over the NVSS integrated flux, as a function of the NVSS flux. This plot illustrates the result of applying our 5$\sigma$ criterion when selecting the combined flux, and serves as a direct comparison to Fig. 11 in \citet{FIRST2015}.}\label{RadioFlux}
\end{figure}
\begin{table}\small
\caption{Number of FIRST subcomponents within 30 arcsec, collapsed into a single source. Only $\sim10$ per cent of the sources have multiple components.}\label{nComp}
\centering
\begin{tabular}{ccc}\hline
Subcomponents&Number of cases&Percentage of total groups\\\hline
2&80105&81.2\\
3&14659&14.9\\
4&3145&3.2\\
$\geq 5$&738&0.7\\\hline
\end{tabular}
\end{table}
We therefore grouped FIRST sources within 30 arcsec of an NVSS source, or, where no NVSS counterpart was present, within 30 arcsec of another FIRST source. For each combination we co-added the fluxes, and combined the positions using a flux-weighted average, which is likely to give the best estimate for the core position in all cases \citep[the brightest lobe is closer to the core in most FRI and FRII sources, see][]{Magliocchetti1998}. The positional errors were assumed to be the larger between a flux-weighted sum of the individual positional errors and the standard deviation of the individual positions, for both of which we assessed RA and DEC separately. To minimise issues with the inaccurate positions in NVSS at low fluxes, for the combined entries in the catalogue we used FIRST fluxes and positions unless an NVSS match was present and had a flux larger by at least 5$\sigma$. Fig. \ref{RadioFlux} illustrates the effect of this selection. In this Figure we have plotted $F_{FIRST}/F_{NVSS}$ versus $F_{NVSS}$ in red, and overlaid $F_{Combined}/F_{NVSS}$ versus $F_{NVSS}$ in blue. For the combined (blue) distribution, any sources where we have used the NVSS value follow the 1:1 horizontal line, and any sources where we used a FIRST flux (different from the NVSS flux) will deviate from the 1:1 line. So any areas of the plot where the red FIRST distribution appears represent sources where the FIRST fluxes were smaller by at least 5$\sigma$, and NVSS fluxes were used. The plot shows a very large scatter at low NVSS flux values, a consequence of the large uncertainties in the NVSS fluxes near the detection limit, thus our method selected the (grouped) FIRST values for the majority of these cases. At fluxes around 0.05--0.1 Jy, some FIRST fluxes are noticeably smaller than their NVSS counterparts, perhaps due to our conservative collapsing radius, so the NVSS fluxes were used. For larger fluxes FIRST and NVSS tend to agree, and the distribution becomes narrower around the 1:1 line. While the effect is subtle, our plot shows some differences from Fig. 11 in \citet{FIRST2015}. It is difficult to determine how much of this difference is caused by the much larger number of sources we are using, and how much it is due to the fact that we are grouping FIRST subcomponents (this should narrow the distribution).
Using these criteria, we found 98647 groups ($\sim 4.6$ per cent of the total number of sources, 2129340, $\sim 10$ per cent of the number of FIRST sources). The number of subcomponents per collapsed source is given in Table \ref{nComp}. These numbers give us a rough idea of the number of resolved FRI and FRII galaxies in FIRST, but they are not representative of the entire population, as many FRI and some distant FRII are not split into separate objects in FIRST, even if they show resolved structures ($\sim35$ per cent of the sources have resolved structures on scales of 2--30 arcsec). The number of subcomponents is also not a reliable diagnostic for the radio morphology, as for many FRI sources the core and one of the lobes might be grouped due to orientation effects or Doppler suppression, making them appear as doubles in Table \ref{nComp}, rather than triples.
As for WISE (section \ref{WISE}), only a small fraction of the radio sources fall within 60 arcsec of a 3XMM source, $\sim17500$ (see the details of the area of overlap at the end of the next section). When we compare this number to those in 3XMM and the equivalent fraction of the WISE catalogue, the low radio sky density becomes immediately apparent.
As our combined FIRST+NVSS catalogue might prove useful to other researchers, and only a subset of its sources are used in the final MIXR sample, we have made it available on-line as a stand-alone file at \url{http://www.arches-fp7.eu/index.php/tools-data/downloads/combined-radio-catalogue} and will also upload it to VizieR. As with any catalogue, it may not suit every purpose: please consider carefully the caveats described in this section.
\subsection{MIXR Sample Construction}\label{Sample}
For our sample construction we used a cross-correlation tool developed as part of the ARCHES collaboration products \citep[Pineau et al. 2016, subm.; ][]{Pineau2015}. This tool is based on an earlier version, tested on the 2XMMi and SDSS DR7 catalogues \citep{Pineau2011}. Our version of the tool also uses a chi-square ($\chi^{2}$) statistical hypothesis test to select probable candidates, but applied on n-catalogues instead of only two. The likelihoods we use to compute Bayesian probabilities are $\chi$ distributions of various degrees of freedom and multidimensional Poisson distributions. These distributions are normalised so their integration over the $\chi^{2}$-test region of acceptance equals 1. Priors are derived both from local densities of sources in each catalogue and from the results of the cross-correlation of each possible subset of catalogues. Currently, the tool is able to provide probabilities for up to 8 catalogues. For more information see Pineau et al. (2016, subm.), and the ARCHES website\footnote{\url{http://www.arches-fp7.eu/index.php/tools-data/online-tools/cross-match-service}}. The \textsc{xmatch} tool is available as a web service through the ARCHES website, and will eventually supersede the two-catalogue tool currently available at the CDS cross-match service\footnote{\url{http://cdsxmatch.u-strasbg.fr/xmatch}}.
The association probabilities for any given tuple of sources depend on the normalised distances of the individual sources from the averaged position (up to the equivalent of the 1-dimension 3$\sigma$ level, i.e. 99.7 per cent completeness), as well as the sky density for each given catalogue (which is why it is fundamental that the sky coverage of all the cross-matched catalogues coincide and are accurate).
Although WISE covers the entire sky, FIRST and NVSS only cover latitudes north of $-40^{\circ}$, and 3XMM covers patches throughout the entire sky. The area of overlap between the three individual catalogues is roughly 135 square degrees. We carried out a simultaneous match of the overlapping sections of the combined FIRST/NVSS catalogue ($\sim17500$ sources), WISE ($\sim1.8$ million sources) and the cleaned-up 3XMM ($\sim150000$ sources), using two inner joins (i.e. keeping only the tuples that had a candidate in each catalogue). For the sources resulting from the three-catalogue cross-correlation, we aimed for maximum completeness, requiring an association probability greater than 1$\sigma$ ($\sim70$ per cent), and thus obtaining a sample of 2753 sources (with a reliability of 90.15 per cent), reduced to 2529 in the full sample, and 1575 in the W3 sample.
\section{Activity diagnostics}\label{Diagnostics}
To accurately characterise the multiwavelength behaviour of an extragalactic source, we need to know its distance (redshift), from which we can derive its luminosity in each band. However, redshifts are not always available, consistent, or accurate. In this section we demonstrate how it is possible to pre-emptively diagnose the type of activity present in a sample of sources, using the three bands in MIXR. These diagnostics are very useful to test the accuracy of single-band activity markers commonly employed in other surveys, especially those that do not require a radio selection, as well as to better constrain an a priori range of models for systematic spectral or SED fitting, which can yield more accurate redshift values than those obtained by cross-correlation with an extra catalogue. In sections \ref{LDiag} and \ref{RL_RQ_section} we will verify the accuracy of these preliminary diagnostics.
Please note that, because we want these diagnostic plots to be as straightforward as possible for the end user, we have not converted between flux and magnitude systems: for WISE we plot aperture-corrected magnitudes; for the X-rays we plot fluxes in cgs (erg cm$^{-2}$ s$^{-1}$); for the radio, we plot fluxes in mJy.
\subsection{The WISE colour/colour plot}\label{C_C}
\begin{figure*}
\centering
\includegraphics[width=0.92\textwidth]{C_C_all}
\caption{Colour/colour diagram for the sources in our Full Sample (see Table \ref{nSources} for details on sample and source type statistics). We have plotted with empty symbols the W3 rejected sources, and with full symbols the sources in the W3 Sample. The cross and the arrow by the legend indicate the typical size of the errors and the direction of the W3 upper limits, respectively.}\label{C_C_all}
\end{figure*}
The first diagnostic we tested involves mid-IR colours, and is based on the work by \citet{Lake2012}. For their work, Lake et al. generated a series of synthetic SEDs for a wide range of astronomical populations, and plotted them on the WISE colour/colour diagram (W1-W2 versus W2-W3 magnitudes, see their Fig. 1) to test which regions of the parameter space they occupied. While this diagnostic is extremely useful for a first approach to study what the AGN and the host galaxy are doing, it is important to keep in mind that there is overlap between populations, even more so when obscuration and redshift evolution are taken into account \citep[see e.g. Fig. 1 of][]{Hainline2014}, and that several colour cuts have been proposed to identify the AGN population in particular \citep[e.g.][]{Ashby2009,Assef2010,Assef2013,Stern2012,Mateos2012}.
\begin{table*}\small
\caption{Activity table. For each of our source types, selected on the WISE colour/colour plot, this table shows the types of activity most likely to be found at each wavelength. Please note that, for each colour category, several combinations of the elements in columns 2--4 may be possible, e.g. in the first group, an elliptical galaxy in a cluster, with a radiatively inefficient AGN in X-rays, and a LERG in radio. LINER stands for low-ionisation nuclear emission-line region. ULIRG stands for ultraluminous infrared galaxy. LERG stands for low excitation radio galaxy; high excitation sources (HERG) include NLRG (narrow line radio galaxies) and BLRG (broad line radio galaxies). Please see also Table \ref{nSources} for the statistics of each subset.}\label{Activity}
\centering
\begin{tabular}{lllll}\hline
Label&WISE colour selection&Mid-IR/Optical&X-rays&Radio\\\hline
\multirow{3}{*}{Elliptical}&\multirow{3}{*}{$W1-W2<0.5$; $0<W2-W3<1.6$}&Elliptical galaxy (isolated)&\multirow{2}{*}{Rad. inefficient AGN}&\multirow{3}{*}{LERG}\\
&&Elliptical galaxy (cluster)&\multirow{2}{*}{Hot ICM gas}&\\
&&LINER&\multirow{-1}{*}{}&\\\hline
\multirow{3}{*}{Spiral}&\multirow{3}{*}{$W1-W2<0.5$; $1.6\leq W2-W3<3.4$}&\multirow{2}{*}{Star-forming galaxy}&\multirow{2}{*}{Star formation}&Star formation\\
&&\multirow{2}{*}{Star-forming galaxy + AGN}&\multirow{2}{*}{Seyfert galaxy}&Low-L NLRG\\
&&&&LERG\\\hline
\multirow{2}{*}{Starburst}&\multirow{2}{*}{$W1-W2<0.5$; $W2-W3 \geq 3.4$}&Starburst galaxy&Star formation&Star formation\\
&&ULIRG&Seyfert galaxy&Low-L NLRG\\\hline
\multirow{3}{*}{AGN/QSO}&\multirow{3}{*}{$W1-W2 \geq 0.5$; $W2-W3<4.4$}&\multirow{3}{*}{AGN}&Luminous Seyfert galaxy&NLRG\\
&&&BL-Lac&BLRG\\
&&&QSO&QSO\\\hline
\end{tabular}
\end{table*}
At our flux and magnitude limits, and the high Galactic latitude we are working with, we do not expect to see the stellar objects that appear in the diagram of \citet{Lake2012}, but we pre-emptively excluded 15 sources with W2-W3 values smaller than zero. We also excluded 24 sources in the ULIRG/obscured AGN locus (W1-W2$>0.5$, W2-W3$\geq 4.4$), as this area of the diagram shows severe contamination from resolved star formation regions in extremely nearby galaxies, and it is unclear, for the extragalactic sources in this area, whether they could be treated systematically as AGN or starburst galaxies (including obscured AGN would require us to increase the range of $N_H$ values we use to calculate the X-ray luminosities in section \ref{LDiag}, skewing the results for the entire sample). We give the rest of the sources a rough characterisation based on the labels in the work by Lake et al.; the resulting colour/colour diagram is shown in Fig. \ref{C_C_all}. Table \ref{Activity} describes in detail the boundaries we imposed, and what type of activity we expect to find in each population \citep[see e.g. the source distributions on the equivalent WISE colour-colour plots of][]{Gurkan2014,Yang2015}. Table \ref{nSources} (in section \ref{z}) shows the statistics for each source type. Please note that, until we know more about the underlying properties of our sources, the categories in Table \ref{Activity} are only meant as a rough guide. Galaxies, like almost everything in the Universe, do not fall into neatly cut categories \citep[note how the classifications of][overlap on their colour/colour plot, as well as the size of the errors in Fig. \ref{C_C_all}]{Lake2012}. Thus, some sources may not fall into the general behaviour expected for their assigned categories (e.g. a source with the `elliptical' classification that shows signs of star formation or a radiatively efficient, bright AGN). Also note that, while radio sources are traditionally organised based on the Fanaroff-Riley classification \citep{FR1974}, we have not used this classification for Table \ref{Activity}, as the FRI/FRII divide is based both on morphology and on radio power, both of which depend heavily on the environment through which the jet and lobes propagate \citep[see e.g.][]{Hardcastle2013,Hardcastle2014,English2016}.
We have also pre-emptively labelled the sources that lack a reliable detection in the W3 band, represented by empty symbols in Fig. \ref{C_C_all} to signal that they are upper limits (the arrow next to the legend on the plot indicates the direction of the upper limits). We will show in the next Subsections that, overall, these sources behave very similarly to those in the same region of the colour/colour plot that do have a W3 S/N>3, demonstrating that they belong to the same populations, and that their faintness in the 12$\mu$m band is due to their larger distance or lower luminosity, rather than a mis-classification. We checked the W4 results, for consistency, and found even fewer detections than for W3, as expected.
In Table \ref{Activity} we see that the expected AGN classifications, both for the radio and the X-rays, are not clear-cut for each mid-IR population. Throughout this work we aim to study how accurate our mid-IR labels are with respect to the underlying activity (e.g. what fraction of moderately star-forming galaxies and what fraction of AGN we find among the `spiral' sources in Fig. \ref{C_C_all}). The diagnostic plots in the following Subsections aim to shed some light on this topic, but to get a clearer idea of what type of AGN each host harbours we need X-ray, bolometric and jet kinetic luminosities, which we will study in sections \ref{z} to \ref{Power}.
\subsection{X-ray hardness ratios and radio versus X-ray `loudness'}\label{HR}
\begin{figure*}
\centering
\begin{subfigure}{0.48\textwidth}
\centering
\includegraphics[width=.98\linewidth]{HR_all}
\caption{Hardness ratio - radio/soft X-ray flux}
\label{HR_all}
\end{subfigure
\begin{subfigure}{0.48\textwidth}
\centering
\includegraphics[width=.98\linewidth]{HR_all_2}
\caption{Hardness ratio - radio/hard X-ray flux}
\label{HR_all_2}
\end{subfigure}
\caption{X-ray hardness ratio versus ratio of the 1.4 GHz radio flux (in mJy) to the X-ray flux (left: soft, 0.2--2 keV; right: hard, 2--10 keV, in erg cm$^{-2}$ s$^{-1}$) for all the sources. Colours and symbols are the same as in Fig. \ref{C_C_all}. Only sources in the W3 sample are plotted, for clarity.}
\label{HR_all_comb}
\end{figure*}
Many X-ray selected samples rely on hardness ratio cuts to eliminate non-AGN sources, as well as to determine the obscuration and distance of a given set of AGN. Normally it is preferable to use net (background-subtracted) counts, rather than fluxes, to estimate the hardness ratio. However, due to the nature of our X-ray catalogue, where more than one observation with multiple instruments can be present for any given source, we decided to use the averaged (over all the observations for each source), net fluxes for our analysis. The hardness ratio we use is defined as:
\begin{equation}\label{HR_eq}
HR=(F_{2-12 keV}-F_{0.2-2 keV})/(F_{2-12 keV}+F_{0.2-2 keV})
\end{equation}
The fluxes in our catalogue are biased by the model assumed to derive them from the raw counts \citep[see section \ref{LDiag} and][for details]{Mateos2009}, and thus the following plots should be considered carefully, particularly for sources with X-ray spectral shapes very different from the assumed spectral shape used to represent AGN emission ($\Gamma = 1.7$, $n_H\sim 10^{21}$ cm$^{-2}$). Reassuringly, star-forming sources do not greatly deviate from this approximation on average \citep{Ranalli2012}, but radiatively inefficient and Compton-thick AGN may be represented less accurately in these hardness ratio plots, the first due to their being dominated by the soft, jet-related component \citep[e.g.][]{Hardcastle1999}, the latter, which are not common in our sample (see section \ref{RL_RQ_section}), because of the heavy absorption and Compton reflection. Please note that we have used 2--12 keV fluxes, as we were constrained to the bands defined in the catalogue. At this point of the analysis our catalogue fluxes are also not corrected for foreground absorption, but given that we are working at high Galactic latitudes, the effect of Galactic obscuration should be unimportant in this band.
Fig. \ref{HR_all_comb} shows the distribution of hardness ratios for all the sources on the Y axis, and the ratio of the radio to the relative X-ray (soft and hard, Figs. \ref{HR_all} and \ref{HR_all_2}, respectively) flux on the X axis. This is a very good way to quickly assess the `radio loudness' and `X-ray loudness' of the sources, as well as to establish whether a soft excess or deficit may be related to the same processes that produce the radio emission.
\begin{figure}
\centering
\includegraphics[width=0.46\textwidth]{Fr_Fx_histo}
\caption{Histogram of the radio to soft (red) and hard (blue, dashed) X-ray flux, for all the sources regardless of the W3 S/N cut (full sample).}\label{Fr_Fx_histo}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}{0.238\textwidth}
\centering
\includegraphics[width=.95\linewidth]{Fr_Fx_histo_soft}
\caption{}
\end{subfigure
\begin{subfigure}{0.238\textwidth}
\centering
\includegraphics[width=.95\linewidth]{Fr_Fx_histo_hard}
\caption{}
\end{subfigure}
\caption{Normalised histogram of the radio to soft (left, red) and hard (right, blue) X-ray flux, for the sources that pass (empty bars) and do not pass (full bars) the W3 S/N cut.}
\label{Fr_Fx_W3}
\end{figure}
Despite the large number of points, it is quite clear in these plots that most of the sources have rather high ($>0$) hardness ratios (as a guide: the `typical HR cut' barrier in our diagrams corresponds to an unabsorbed spectrum with $\Gamma=2$; an unabsorbed AGN with $\Gamma=1.7$ would have a HR of 0.17; an AGN with $N_{H}=10^{22}$ cm$^{-2}$, $\Gamma=1.7$, $z=1$, would have a HR of 0.53). The overall trend also varies depending on whether the hard or the soft X-ray flux is used for the ratio on the x axes: for soft X-rays (Fig. \ref{HR_all}) the ratio seems fairly constant, as evidenced by the aggregation of sources around a vertical line at $F_{1.4 GHz}/F_{0.2-2 keV}\sim10^{14}-10^{15}$, with some outliers, especially on the left side of the plot (larger relative X-ray fluxes). For the hard X-rays (Fig. \ref{HR_all_2}), the overall behaviour is slightly different, and there seems to be a slight negative trend between the radio/hard X-ray flux and the hardness ratio, meaning that more `hard X-ray loud' (less `radio-loud') sources have harder spectra. These behaviours become more evident when we plot the flux ratio histograms for both distributions (Fig. \ref{Fr_Fx_histo}). This negative trend probably arises from a combination of factors: the known radio-soft X-ray correlation of \citet{Hardcastle1999}, which will push radio-louder (softer) sources to the right of the plot, and radio-quieter (harder) sources slightly to the left; a higher intrinsic absorption for the hardest sources, which would hide a similar negative trend (pushing the hard sources to the right) in the soft X-rays; uncertainties derived from the underlying 3XMM flux derivation; higher intrinsic hard X-ray fluxes for the harder sources (e.g. from higher Eddington rates). Given that we are working with flux-limited samples, it is not possible to analyse the strength of a possible anticorrelation between the hardness ratio and the radio/hard X-ray flux ratio, as we do not know what sources may be missing from these plots beyond the flux limits.
\begin{figure*}
\begin{subfigure}[b]{0.5\linewidth}
\centering
\includegraphics[width=0.95\linewidth]{HR_ell_2}
\caption{Ellipticals}
\label{HR_ell_2}
\vspace{2ex}
\end{subfigure
\begin{subfigure}[b]{0.5\linewidth}
\centering
\includegraphics[width=0.95\linewidth]{HR_spi_2}
\caption{Spirals}
\label{HR_spi_2}
\vspace{2ex}
\end{subfigure}
\begin{subfigure}[b]{0.5\linewidth}
\centering
\includegraphics[width=0.95\linewidth]{HR_sb_2}
\caption{Starburst}
\label{HR_sb_2}
\end{subfigure
\begin{subfigure}[b]{0.5\linewidth}
\centering
\includegraphics[width=0.95\linewidth]{HR_agn_2}
\caption{AGN}
\label{HR_agn_2}
\end{subfigure}
\caption{Hardness ratio versus ratio of the 1.4 GHz radio flux (in mJy) to the hard X-ray flux (2--12 keV, in erg cm$^{-2}$ s$^{-1}$) for the various source populations. The colours and symbols follow the same scheme as those in Fig. \ref{C_C_all}, with full symbols representing the sources that pass the W3 S/N cut, and empty symbols those that do not. Sources labelled as variable on the 3XMM catalogue are plotted with larger, magenta squares, only in Fig. \ref{HR_agn_2}.}
\label{HR_subsets_hard}
\end{figure*}
Another effect that becomes more apparent when plotting some of the individual populations, and that we have displayed in more detail in Fig. \ref{Fr_Fx_W3}, is that the W3 S/N cut skews the sample slightly towards `radio-quieter' sources. This is expected, as the W3 cut essentially imposes a distance limit, and we know that, overall, AGN were radio-louder in the past \citep[see e.g.][and references therein]{Best2014,Williams2015}, and a majority of our sources are AGN. However, when studying the individual populations (Figs. \ref{HR_ell_2} to \ref{HR_agn_2}) we see that the distributions for the sources above and below the W3 S/N cut are similar enough that it is clear that we are essentially sampling the same types of sources, just at slightly different redshifts. We will study the effect of the W3 S/N cut and the redshift selection in more detail in section \ref{z}.
For brevity, we have only included the hard X-ray (2--12 keV) plots in Fig. \ref{HR_subsets_hard} for the individual populations. Overall, the various populations seem to behave as expected. The elliptical galaxies (Fig. \ref{HR_ell_2}) are `radio-loudest' (largest radio/X-ray flux ratios) and have the largest number of soft sources of all the groups, which is consistent with the idea that many of them host radiatively inefficient AGN. The spirals (Fig. \ref{HR_spi_2}) show quite a lot of scatter, which is consistent with the heterogeneous population of star-forming galaxies, low-luminosity Seyferts and LERG with spiral hosts, as we introduced in Table \ref{Activity}. All these populations are expected to quickly become undetectable in W3 as redshift increases, which is why our spirals are hit the hardest by the W3 cut. The starburst sources (Fig. \ref{HR_sb_2}) seem to have the narrowest distribution in the radio/X-ray flux ratio of all the populations, and the best consistency between the W3 accepted/rejected sources. This is reassuring, as we would expect a fairly clean selection for these sources, and a rather tight correlation between X-rays and radio where only star formation processes are responsible for both types of emission. The large range of hardness ratios covered by the starburst sources may seem surprising, but it is consistent with the picture presented by e.g. \citet{Ranalli2012}. The AGN (Fig. \ref{HR_agn_2}) show the largest scatter of all populations, and also seem to, overall, have the largest HR values, which is consistent with their expected spectral shape, $z$ evolution, and varying degrees of nuclear obscuration. There are probably several factors introducing scatter in the AGN plot, but along the X axis probably the most relevant one is the known scatter between jet output and radiative output, which we discuss in more detail in section \ref{Power}. The AGN sources are by far the most numerous at this point of the analysis, but they are also the ones most reduced by the introduction of redshifts (see Table \ref{nSources}), which is part of the motivation behind performing these diagnostics prior to carrying out the additional cross-correlation with SDSS.
The 3XMM catalogue includes a label for variable sources. These are sources that show X-ray variability within a single observation, thus in timescales of minutes to hours \citep[substantial variability on longer timescales is probably present for a large number of sources, but it is not described in the catalogue - see e.g.][]{Strotjohann2016} for examples of long-term variability from the XMM Slew Survey, and the EXTraS collaboration results\footnote{\url{http://www.extras-fp7.eu/index.php}} for examples across all the available XMM EPIC data). We have indicated with larger, magenta squares the sources with this classification that are retrieved in our sample, in Fig. \ref{HR_agn_2}, as most of them coincide with sources we have classified as AGN. Interestingly, the vast majority of these sources lie on the `X-ray louder' side of both the soft and hard X-ray HR plots. The fact that we do not find rapid X-ray variability for the radio-louder sources might be explained by a combination of factors. The variability timescales of the jet tend to be longer than those of the corona, where variations in the accretion flow are reflected quickly and abruptly. It is also possible that some of the radio flux is self-absorbed, that the relativistic boosting of the jet affects the radio and X-rays differently, or that the jet contribution is diluted in the X-rays.
\subsection{Flux correlations}\label{FCorr}
\begin{figure*}
\centering
\includegraphics[width=0.95\textwidth]{w3_combFlux}
\caption{1.4 GHz radio flux (mJy) versus W3 (12 $\mu$m) magnitude. The blue line represents the best linear correlation for the starburst sources ($r\sim 0.70$, where $r$ is the correlation coefficient). The colours and symbols follow the same scheme as those in Fig. \ref{C_C_all}. Please refer to Appendix \ref{ExtraFigures} for the W1 (3.4 $\mu$m) and W2 (4.6 $\mu$m) versions of this plot, and to Table \ref{nSources} for the source statistics.}\label{w3_combFlux}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.95\textwidth]{w3_hardX}
\caption{Hard X-ray (2--12 keV) flux (erg cm$^{-2}$ s$^{-1}$) versus W3 (12$\mu$m) magnitude. The blue line represents the best linear correlation for the starburst sources ($r\sim 0.70$) and the black line the best attempt at a linear correlation for the AGN ($r\sim 0.55$). The colours and symbols follow the same scheme as those in Fig. \ref{C_C_all}. Upper limits (only for the X-rays) are represented with grey arrows. Please refer to Appendix \ref{ExtraFigures} for the equivalent plot for soft X-rays (0.2--2 keV), and to Table \ref{nSources} for the source statistics.}\label{w3_hardX}
\end{figure*}
To establish possible correlations between different types of activity in the MIXR sources, we plotted the radio and X-ray fluxes, and mid-IR magnitudes. As we introduced at the beginning of this section, in order to allow readers to establish an easier, more direct comparison, we have plotted the flux or magnitude values in the catalogues, with no transformations, other than the aperture corrections for WISE. Only the most relevant plots are displayed in this section, please refer to Appendix \ref{ExtraFigures} for details on the other flux correlations.
To accurately triage sources according to their activity, we need to assess three properties: star formation, radiative AGN output, and kinetic (radio jet and lobes) AGN output. To do so we need to simultaneously consider where the sources fall on the three plots presented in this section, as well as the information from their mid-IR colours and the plots in section \ref{HR}.
Fig. \ref{w3_combFlux} shows the distribution of radio versus W3 flux for the populations we defined from Fig. \ref{C_C_all}. It is clear, at a first glance, that the starburst sources follow a correlation, which is likely to be a direct extension into the mid-IR of the well-known far-IR/radio correlation for star formation \citep[see e.g.][and sections \ref{LDiag} and \ref{RL_RQ_section} for more details]{Gruppioni2003}. A few of the spiral galaxies also follow this correlation, but a large fraction seem to prefer the locus inhabited by most of the AGN and the elliptical galaxies (which we know are likely to harbour radiatively inefficient AGN). Interestingly, many sources with the AGN classification also follow the star formation correlation: these are likely to be radio-quiet AGN. The correlation we derived from the starburst sources, plotted in Fig. 7, is slightly flatter than expected. This is probably caused by the presence of outliers, especially mis-classified AGN, which might be exerting a leverage on the fit, but we have not excluded these points, as doing so might introduce further bias in the subset. We have excluded outliers from the luminosity correlations in section \ref{LCorr}.
It is worth highlighting here that radio-quiet does not mean radio silent \citep[e.g.][]{Wong2016}: because the radio-loud/quiet classification is traditionally based on optical (or other bands) to radio flux ratios \citep[e.g.][]{Kellerman1989}, AGN with large radiative outputs, and small jets and lobes, are often classified as radio-quiet, as an increasingly large number of Seyfert galaxies shows \citep[e.g.][]{Hota2006,Gallimore2006,Croston2008,Mingo2011}. In this work, and in particular in section \ref{RL_RQ_section}, we refer to radio-quiet AGN as sources where the radio emission we detect from them is likely to originate mainly from stellar processes, accelerated particles in wind-driven shocks \citep[see][]{Nims2015,Zakamska2016} or, if arising from a jet and lobes, they are small and faint, and the AGN produces the bulk of its emission as radiative output in the other bands. Conversely, we refer to radio-loud AGN as those that have a substantial kinetic output in the form of jet and lobes, which we measure as radio emission well above the star formation correlation. The radiatively inefficient LERG/LINER sources also follow these criteria, so they are a subset of radio-loud AGN.
For the radio-loud AGN, as well as the potential LERG/LINER sources, there seems to be a wide range of possible radio fluxes for a given 12$\mu$m magnitude. This is partly due to the fact that most AGN have W3 magnitudes close to the detection limit, but it hints at what we observed in \citet{Mingo2014} for the 2Jy and 3CRR samples, when we found a large amount of scatter in the relation between radiative and kinetic output in radio-loud AGN, in apparent contradiction with the correlation proposed by \citet{Rawlings1991}. We will discuss this point in further detail in sections \ref{RL_RQ_section} and \ref{Power}.
Fig. \ref{w3_hardX} shows the hard X-ray versus W3 flux for our sources. There is a clear distinction between the starburst and AGN populations: they follow nearly parallel distributions, but the starburst galaxies have systematically lower X-ray fluxes. This is in agreement with what we know of the mid-IR/X-ray correlation for AGN and star-forming galaxies \citep[e.g.][]{Gandhi2009,Mateos2015}, and it illustrates why it is so difficult to distinguish the break between both populations in luminosity/luminosity plots. The presence of mis-classified sources in the AGN subset introduces scatter, and weakens the correlation we obtain, but the data show that it is clearly there. At this stage we have not wanted to re-classify sources based on their fluxes, we will do so after we obtain their luminosities, in section \ref{RL_RQ_section}.
The elliptical galaxies are systematically fainter in W3 than the starburst galaxies, and also have X-ray fluxes that are systematically lower than those of the AGN subset, reinforcing the conclusion that these sources harbour radiatively inefficient AGN. The spiral galaxies seem to be split between the AGN and the starburst loci, with a few sources falling in the gap between them. We repeated this same plot without the radio selection, and found that sources with spiral colours fill the entire gap between the AGN and starburst galaxies, making it impossible to distinguish between non-active spiral galaxies with different levels of star formation, and Seyferts with a range X-ray luminosities. Only with the radio selection is it possible to easily distinguish between Seyferts and non-active galaxies for sources in the spiral region of the WISE colour/colour plot.
There is a weak correlation between the X-ray and radio fluxes for the starburst sources, which appears mainly for the soft X-rays, but it is not very strong, especially if the sources with radio fluxes greater than 100 mJy are removed. The situation is also less clear for the other populations, even the AGN present a lot of scatter. Although we know that there is X-ray emission arising from the jet \citep{Hardcastle1999}, it appears mainly in the soft X-ray band. What is readily apparent in the X-ray/radio plots (Figs. \ref{combFlux_softX} and \ref{combFlux_hardX}) is the previously mentioned scatter between radiative and kinetic output that we also observed in the 2Jy and 3CRR sources, as well as the LERG/LINER nature of the elliptical sources.
Overall, these plots present a picture consistent with what we outlined in Table \ref{Activity} and section \ref{HR}, allowing us to diagnose the different types of activity present in each source type. These diagnostics need to be confirmed, however, using redshifts to derive the luminosities of the sources in each band. The redshifts will also help us determine the nature of any outliers in the diagnostic plots, as well as to assess the effects of evolution within the populations, which may not be negligible \citep[see e.g. the evolutionary tracks of][]{Assef2010}.
\section{SDSS redshifts}\label{z}
\begin{table*}\small
\caption{Number of sources in each subset. See also Table \ref{Activity} for the detailed WISE colour source classification. For all subsets we have applied the quality cuts for FIRST/NVSS and 3XMM, and the S/N cut for W1 and W2, but we treat the W3 S/N cut separately, as described in section \ref{WISE}. Throughout the text we refer to the full set of sources that passes the W3 S/N cut as W3 Sample (or W3 z sample for the subset that also have redshifts). For the source type subsets we use the prefix `z' for sources with redshifts, and the prefix `rejected' for sources that do not pass the W3 S/N cut, for example the galaxies with elliptical colours that have redshifts but do not pass the W3 S/N cut are referred to as `z Rejected Ellipticals' in the plots.}\label{nSources}
\centering
\begin{tabular}{lcccccc}\hline
Subset name&\multicolumn{6}{c}{Number of sources}\\\hline
&All sources&W3 S/N$\ge3$&W3 S/N$<3$&With $z$&With $z$ + W3 S/N$\ge3$&With $z$ + W3 S/N$<3$\\\hline
Full sample&2529&1575&954&1367&947&420\\
Ellipticals&203&145&58&137&94&43\\
Spirals&507&222&285&323&149&174\\
Starburst&268&174&94&168&114&54\\
AGN&1510&1008&502&721&577&144\\\hline
\end{tabular}
\end{table*}
To find redshifts for our sample, with as uniform a coverage as possible, the ideal choice is the Sloan Digital Sky Survey (SDSS), more so considering that FIRST was initially designed to cover the same sky area, so the overlap between MIXR and SDSS is very large ($\sim 70$ per cent in sky area). We decided to use the latest data release, DR 12 \citep{Gunn2006,Eisenstein2011AJ,Dawson2013}, to maximise the number of possible counterparts to the sources in our catalogue with photometric or spectroscopic redshifts. We did not use the ARCHES xmatch tool in this case, but rather a simple normalised distance histogram (between the \textsc{xmatch} averaged position for the WISE+3XMM+FIRST/NVSS MIXR sources, and the SDSS positions), making use of the astronomy software TOPCAT \citep{Taylor2005} to carry out the cross-match.
We initially selected SDSS sources within 10 arcsec of our merged catalogue positions. The distance distribution histogram (see Fig. \ref{SDSS_dist_histo} in the Appendix) showed a very clean selection at distances under 2--3 arcsec. To further test our selection we manually checked about 150 sources that had 4 or more SDSS matches, using the on-line SDSS finding charts, and found the nearest match to clearly be the best choice in nearly all cases (the other possible matches were stars or had much larger separations). We only found two potentially dubious cases, where the merged position coincided with an optical galaxy cluster, and the nearest and second-nearest match had similar separations, around 2--4 arcsec. We thus decided to only consider redshifts for matches with separations below 3 arcsec. We have included in the catalogue a column with the URL of the SDSS finding chart for each source (column SDSS\_URL), for the users to explore.
Roughly half of the sources with redshifts in SDSS had spectroscopic redshifts, so we used these values whenever possible. We used the sources that had both values to study the reliability of photometric redshifts, and found them to be fairly reliable in most cases. It is possible that the sources that only have photometric information, because they are fainter or more distant, also have less reliable redshifts. To mitigate this effect, and the fact that some photometric redshifts have fairly large error bars, we decided to take these uncertainties into account, when possible, when calculating the luminosities (see section \ref{LDiag}).
Our manual check also confirmed that most optical spectroscopic classifications coincide with those we derived from mid-IR colours. The optical spectra revealed several broad-line AGN in objects classified as spirals in our sample, where there is also contribution from star formation. These objects tend to fall in the AGN locus in the X-ray/W3 diagnostic plot (Fig. \ref{w3_hardX}), but, notably, some of them fall in the star formation correlation for the Radio/IR plot (Fig. \ref{w3_combFlux}). If their luminosities are consistent with this assessment, these sources exhibit behaviour typical of radio-quiet AGN, as shown in the results of \citet{Padovani2011A,Bonzini2013}, where the bulk of the radio emission in (moderately luminous) radio-quiet AGN and star-forming galaxies is produced by star formation.
The fraction of objects with SDSS counterparts is rather large, around $70$ per cent of the sample, although the fraction of objects with good redshift measurements (no upper limits, small separations) falls to $\sim50$ per cent, with the sources classified as AGN suffering the greatest loss. The limiting factor in our overall selection still seems to be the WISE W3 band, but the requirement of a mid-IR counterpart to the radio and X-ray selection clearly plays an important role on finding optical counterparts for our sources as well. The fraction of sources that pass the W3 cut and have redshifts in SDSS is $66$ per cent, but it is quite dependent on the source type. Table \ref{nSources} details the names, definition, and statistics of each subsample of sources.
We saw in section \ref{HR} (see Fig. \ref{Fr_Fx_W3} in particular) that the W3 selection introduces a slight skew in the distribution in terms of radio (or X-ray) loudness. Fig. \ref{SDSS_selection} is useful to also study the possibility of a selection bias introduced by the SDSS selection. The histograms represent the distribution of radio to hard X-ray flux for four different subsets of sources: all sources (initial sample, see Table \ref{nSources}), all sources with redshifts (full z sample), all sources with no redshifts (full sample, no z) all sources that pass the W3 S/N cut and have SDSS counterparts (W3 z sample), and all sources that pass the W3 cut and do not have SDSS redshifts (W3 sample, no z). At a glance the histograms look rather similar, but after carrying out a Kolmogorov--Smirnov test for the $F_{1.4 GHz}/F_{2-12keV}$ distributions for several subsets (Table \ref{KS_table}) we see that the W3 and redshift cuts do indeed change the shape of the original distribution. Interestingly, both cuts seem to skew the distribution in similar ways, as the `W3 z' and `W3 no z' distributions are the most similar in Table \ref{KS_table}. Overall, the W3 cut seems to have a larger effect on the distribution than the $z$ cut, but the combination of both seems to skew the distribution even further.
\begin{figure}
\centering
\includegraphics[width=0.46\textwidth]{SDSS_rej_test_histo}
\caption{Histogram of radio to X-ray fluxes to test possible selection biases introduced by SDSS and WISE. From top to bottom the curves represent all the sources in the initial sample (black, thin line) full z sample (blue line), full sample with no redshifts (red dashed line), W3 z sample (magenta line with longer dashes), and W3 sample with no redshifts (cyan line). See table \ref{nSources} for subsample definitions and statistics.}\label{SDSS_selection}
\end{figure}
\begin{table}\small
\caption{Results of the Kolmogorov--Smirnov test for the various distributions. Please see Table \ref{nSources} for the statistics of each subset and Section \ref{Sample} for the subset definitions. The columns show, respectively, the subsets compared, the KS statistic, and the p-value (to test the null hypothesis probability). The `z' and `no z' denote whether a subset has or does not have SDSS redshifts.}\label{KS_table}
\centering
\begin{tabular}{lcc}\hline
Subsets tested&D&p\\\hline
Initial - W3&0.11&$1.76\times10^{-11}$\\
Initial - full z&0.06&$5.68\times10^{-3}$\\
Initial - full no z&0.05&$1.53\times10^{-2}$\\
Initial - W3 z&0.13&$1.86\times10^{-11}$\\
Initial - W3 no z&0.10&$7.78\times10^{-5}$\\
Full z - Full no z&0.11&$2.18\times10^{-7}$\\
W3 z - W3 no z&0.07&$5.14\times10^{-2}$\\
Full z - W3 z&0.08&$1.89\times10^{-3}$\\
Full no z - W3 no z&0.14&$2.64\times10^{-7}$\\\hline
\end{tabular}
\end{table}
This is to be expected, as we know that both cuts impose, in essence, a distance limit, but the W3 cut is more severe. We also know that AGN were radio-louder in the past \citep{Best2014,Williams2015}, and that most non-AGN galaxies in our sample must be at fairly low $z$, so the skew is consistent with what we expect. Looking at the histogram in Fig. \ref{SDSS_selection} the differences are subtle indeed, despite the numbers in Table \ref{KS_table}. There seems to be a slight bias in terms of X-ray loudness when only the SDSS selection is applied: the z sample histogram deviates more from the overall (full sample) distribution for sources with an intermediate radio/X-ray ratio. Looking at the sources that occupy this range of $F_{1.4 GHz}/F_{2-10 keV}$ in Fig. \ref{HR_all_2}, it seems that with the SDSS selection we must be eliminating some AGN and spiral galaxies, perhaps more distant or overall fainter in the optical than the others (see also section \ref{RL_RQ_section}). When comparing the `full z' and `full no z' histograms, they look very similar, indicating that the SDSS selection is mostly unbiased, except around $F_{1.4 GHz}/F_{2-12 keV}\sim10^{13}$, where more sources are preserved than discarded by the SDSS selection, meaning that there is a slight favouring of more X-ray bright sources with respect to the radio-bright sources on the other wing of the distribution.
For the distributions that apply the W3 S/N cut, `W3 z sample' and `W3 sample, no z', we see that the W3 cut is good at preserving the sources at intermediate values of $F_{1.4 GHz}/F_{2-12 keV}$, but it also appears to be more biased towards X-ray bright sources than the SDSS selection, eliminating a larger fraction of radio-bright sources.
\begin{figure}
\centering
\includegraphics[width=0.46\textwidth]{z_histo_all}
\caption{Histogram of redshift distributions for the sources with SDSS redshifts. The dashed grey area represents the distribution of all the sources with SDSS redshifts (full z sample), while the full magenta bars show the distribution of sources with SDSS redshifts that also pass the W3 S/N cut (W3 z sample), as detailed in Table \ref{nSources}.}\label{z_histo_all}
\end{figure}
\begin{figure*}
\begin{subfigure}[b]{0.5\linewidth}
\centering
\includegraphics[width=0.95\linewidth]{z_histo_ell}
\caption{Ellipticals}
\label{z_histo_ell}
\vspace{2ex}
\end{subfigure
\begin{subfigure}[b]{0.5\linewidth}
\centering
\includegraphics[width=0.95\linewidth]{z_histo_spi}
\caption{Spirals}
\label{z_histo_spi}
\vspace{2ex}
\end{subfigure}
\begin{subfigure}[b]{0.5\linewidth}
\centering
\includegraphics[width=0.95\linewidth]{z_histo_sb}
\caption{Starburst}
\label{z_histo_sb}
\end{subfigure
\begin{subfigure}[b]{0.5\linewidth}
\centering
\includegraphics[width=0.95\linewidth]{z_histo_agn}
\caption{AGN}
\label{z_histo_agn}
\end{subfigure}
\caption{Histogram of redshift distributions for all the sources with SDSS redshifts (see Table \ref{nSources} for details). The dashed magenta outlines represent the distribution of all the sources with SDSS redshifts that also pass the W3 S/N cut (W3 z sample), and is plotted as a reference. For each subset, the full bars show the sources with SDSS redshifts that pass the W3 S/N cut, while the thick empty outlines show the sources that do not pass the W3 cut.}
\label{z_histo_subsets}
\end{figure*}
Fig. \ref{z_histo_all} shows the redshift distribution of sources for the full z sample (see Table \ref{nSources} for source statistics) and those on the W3 z sample. The W3 cut seems to preserve most sources at $z<0.1$, but there is a progressive and sharp increase in the number of sources lost at larger redshifts, confirming our earlier suspicions, with the greatest effect achieved at $z\sim 0.5$. After $z\sim$1--2 both distributions seem to converge again, indicating that the main limiting factor is not W3, but one (or more) of the other bands. The number of sources in both samples decreases very quickly after $z\sim 3$, which means that for the vast majority of our sources the colour cuts we have used for classification should be fairly reliable \citep[AGN, in particular, have bluer colours at larger $z$, see e.g.][]{DiPompeo2015}.
The redshift distributions for the different source subclasses are very different, as illustrated in Fig. \ref{z_histo_subsets} (where we have also plotted the W3 z sample for reference, as it serves to illustrate the relative contributions of each source type). For all the populations the W3 rejected sources can be found at higher $z$ than their W3 detected counterparts, in agreement with what we introduced in Section \ref{HR}. The ellipticals (Fig. \ref{z_histo_ell}) start disappearing from our sample at lower redshifts ($z\sim0.1$) than the other populations, as expected from radiatively inefficient, non-starforming sources, especially since we eliminated the clusters with our initial selection. The redshift distribution for the spirals (Fig. \ref{z_histo_spi}) reflects, again, the heterogeneous nature of this population, and confirms our suspicion that low-luminosity AGN with spiral mid-IR colours \citep[due to either star formation or evolutionary effects, as described by e.g.][]{Assef2010} are not detected by W3 even at moderate $z$. The starburst sources (Fig. \ref{z_histo_sb}) show a markedly bimodal distribution in terms of the W3 filter. It is difficult to speculate how much of this effect arises from genuinely different underlying populations, but it is likely that the W3 rejected sources with $z>0.1$ contain AGN. The AGN sources (Fig. \ref{z_histo_agn}) are still the largest population in our sample, despite the trim suffered by the cross-correlation with SDSS (see Table \ref{nSources}), as for this population, the W3 and redshift cuts seem to mostly overlap, probably because optically bright AGN are also expected to have a substantial contribution from the torus to the W3 band. As expected, the $z$ distribution for the AGN peaks at the highest value of all the populations, and is also the broadest.
\begin{figure}
\centering
\includegraphics[width=0.46\textwidth]{z_histo_pass}
\caption{Summary of the relative contributions (Figs. \ref{z_histo_ell} to \ref{z_histo_agn}) to the W3 z sample (dotted magenta outline). Elliptical galaxies are represented with vertical red lines, spirals with full green bars, starburst with empty blue bars, and AGN with a continuous black distribution, following the colour scheme of Fig. \ref{C_C_all}.}\label{z_histo_pass}
\end{figure}
Fig. \ref{z_histo_pass} summarises the results of Figs. \ref{z_histo_ell} to \ref{z_histo_agn}, by showing the relative contributions of the different source populations to the overall W3 z sample, and explaining some of the bimodality that appears in the latter. Regular galaxies and LERGs are likely to contribute $\sim 50$ per cent to the first peak of the W3 z sample distribution, with radiatively efficient AGN gradually taking over and making up most of the second peak of the distribution. The W3 cut clearly eliminates some of the sources that would fill the gap between both peaks of the distribution, but it is likely that the other selections, especially the radio, also contribute to create the bimodal shape.
\section{Luminosity diagnostics}\label{LDiag}
We have used different techniques to correct the flux densities to rest-frame for each wavelength, in order to obtain the respective luminosities. With these luminosity plots we can determine how true the sources in each population are to the labels we assigned to them based in Fig. \ref{C_C_all}.
For the radio we assumed a spectral index ($\alpha$, where $S_{\nu}\propto\nu^{-\alpha}$) of 0.8, which is consistent with what is found in most star-forming galaxies \citep[e.g.][]{Magnelli2015,Magliocchetti2014}, and a large fraction of AGN. Using a single value of $\alpha$ for AGN is not ideal, as this population can exhibit quite substantial variation in their spectral indices, but without data at other frequencies it is a necessary compromise, and, as we will see in section \ref{RL_RQ_section}, moderate changes in $\alpha$ ($\pm 0.2$) have very little impact on our results. We had initially planned to include low-frequency data from VLSSr \citep[the VLA Low-frequency Sky Survey Redux, ][]{Lane2012} in our sample, to more accurately estimate the spectral index and jet kinetic energy for each source class, but doing so would have drastically reduced the number of sources (we only found $\sim$150 VLSSr-MIXR matches within a 30 arcsec radius, for our full sample). We consider that this assumption does not introduce a larger degree of uncertainty than any of the others we have used throughout this work.
For the WISE values we first calculated the flux densities using the zero point magnitude values given in the on-line documentation and the work of \citet{WISE2010}, adopting the additional colour correction for starburst sources. We then used SED (spectral energy distribution) fitting software to correct the flux densities to rest-frame values. We used the SED code developed by Ruiz et al. (2016, in prep.), which uses additional torus templates and stellar emission, as well as the simple templates for elliptical, spiral, starburst and AGN galaxies from the SWIRE library \citep{Polletta2007}. We included all the SDSS (aperture and reddening corrected) magnitude measurements in the fit, to better constrain the contribution from the host. Once we obtained the redshift-corrected fluxes from the SED curves and filter profiles, we calculated the luminosities and extrapolated the flux and redshift errors to estimate their uncertainties.
The X-ray corrections were somewhat more problematic. The fluxes in the 3XMM DR5 source catalogue are calculated with the same method established for 2XMM \citep{Watson2009}, which assumes a series of corrections based on a power law fit with $\Gamma = 1.7$ and a fixed foreground $N_H$ column \citep[see also][]{Mateos2009}. Working with this assumption would introduce a bias, as a fraction of our sources are likely to deviate quite substantially from this model, especially at low energies \citep[see e.g. the work by ][on the XMM-Newton spectral fit database]{Corral2015}. The alternative would be to use the detections version of the catalogue, which lists the count rates and instruments used for each observation of each source, and use different models for each population. This solution, however, would still have required a model assumption, as it would not be possible to obtain reliable spectra for the faintest sources to carry out proper spectral fitting. It would also have required assumptions on the instrument observation modes and responses to use in each case, a work that was already done to obtain the 3XMM fluxes. As such, we decided to work with the catalogue fluxes, using a model very similar to that assumed by \citet{Watson2009}, but slightly more flexible, and to limit our luminosities to the (rest-frame) 2--10 keV range, where divergences from our assumed model should be minor.
To calculate the 2--10 keV luminosities we used the X-ray spectral analysis tool XSPEC, with an X-ray model consisting of a foreground absorption $N_H$ column (\textit{tbabs}) set to the Galactic value, an intrinsic absorption column (\textit{ztbabs}), and a powerlaw with $\Gamma = 1.7$. We used the method of \citet{Willingale2013} to calculate the Galactic extinction column. This method is innovative in that it takes into account both the atomic (HI) and the molecular (H$_{2}$) Hydrogen absorption columns; the first is calculated from the 21 cm Leiden/Argentine/Bonn maps of \citet{Kalberla2005}, while the second is obtained using the dust maps of \citet{Schlegel1998} and constraints from Gamma-ray burst afterglows detected by Swift. We used the abundance values from \citet{Wilms2000} and cross-sections from \citet{Verner1996}. We fixed the foreground and intrinsic $N_H$, and the powerlaw slope, for each source and set the powerlaw normalisation to 1.0. , calculated the 2--12 keV flux, and used its ratio to the catalogue 2--12 keV flux (SC\_EP\_FLUX\_4 + SC\_EP\_FLUX\_5 = F$_{2-4.5 keV}$ + F$_{4.5-12 keV}$) to rescale the 2--10 keV luminosity.
To fully appreciate the range of uncertainty present in our X-ray luminosity estimations, rather than just propagating the redshift and flux errors, we calculated lower and upper boundaries for each source including a variation in the intrinsic $N_H$ between zero and $10^{22}$ cm$^{-2}$, with the nominal value at $10^{21}$ cm$^{-2}$. Thus, for the upper luminosity boundary we used the highest intrinsic $N_H$ value ($10^{22}$ cm$^{-2}$), the highest possible redshift ($z + z_{err}$), and the largest flux value (F$_{2-12 keV}$ + F$_{2-12 keV, \ err}$), while for the lowest boundary we used the lowest intrinsic $N_H$ (zero), the lowest redshift ($z - z_{err}$), and the lowest flux value (F$_{2-12 keV}$ - F$_{2-12 keV, \ err}$). Our chosen range of $N_H$ does not encompass heavily obscured and Compton-thick sources. Although we know that WISE is very good at detecting obscured AGN \citep[e.g][]{Stern2012,Assef2013,Mateos2013}, we only expect a fraction of them to be detected by the other catalogues we used for our sample, in particular 3XMM. To be safe, however, we excluded most remaining potential cases from our sample with our WISE colour cuts (see section \ref{C_C}), as it would mean imposing a very restrictive criterion, which only affects a fraction of AGN \citep[e.g.][]{Wilkes2013}, over the entire sample. As we will see in section \ref{RL_RQ_section}, a few potentially Compton-thick sources may remain, but the fraction is very low and should not affect our conclusions.
Unlike for the radio and mid-IR, we do have some upper limits for a few X-ray fluxes, which we have represented with down-pointing arrows on our plots. This is due to the fact that in 3XMM only a detection in one of the bands (and a minimum overall detection likelihood) is required, upper limits can be derived for the other bands, while for the radio there is a single band, and our S/N filters in WISE ensure that we have detections in all the bands involved. For the X-ray upper limits we have nonetheless calculated positive errors as well, so as to fully reflect the other sources of uncertainty (redshift and $N_H$). The larger flux errors and the introduction of a further uncertainty (from the $N_H$ values) also contribute to the X-ray data having, in most cases, larger error bars than the mid-IR and radio data.
\subsection{Luminosity correlations}\label{LCorr}
\begin{figure*}
\centering
\includegraphics[width=0.95\textwidth]{LR_LI_acc_errors}
\caption{1.4 GHz versus 12$\mu$m luminosity for the W3 z sample subsets listed in Table \ref{nSources}. Colours and symbols as in Fig. \ref{C_C_all}. The yellow and grey dotted line shows the radio/mid-IR star formation correlation of \citet{Gruppioni2003}; The magenta line shows our best correlation fit for the starburst sources (eq. \ref{SFcorr1}).}\label{LR_LI}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.95\textwidth]{LX_LI_acc_errors}
\caption{2--10 keV versus 12$\mu$m luminosity for the W3 z sample subsets listed in Table \ref{nSources}. Colours and symbols as in Fig. \ref{C_C_all}. The horizontal cyan line indicates the X-ray luminosity above which sources are classified as AGN (which we have set at $5\times10^{41}$ erg s$^{-1}$, see the text). Upper limits (only for the X-rays) are indicated with grey arrows. The dashed pink line shows shows the X-ray/mid-IR AGN correlation of \citet{Hardcastle2009}. The yellow and grey dotted line shows the X-ray/mid-IR AGN correlation of \citet{Gandhi2009}. The magenta line shows our best correlation fit for the AGN sources (eq. \ref{AGNCorr1}).}\label{LX_LI}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.95\textwidth]{LX_LR_acc_errors}
\caption{2--10 keV versus 1.4 GHz luminosity for the W3 z sample subsets listed in Table \ref{nSources}. Colours and symbols as in Fig. \ref{C_C_all}. The horizontal cyan line indicates the X-ray luminosity above which sources are classified as AGN (which we have set at $5\times10^{41}$ erg s$^{-1}$, see the text). Upper limits (only for the X-rays) are indicated with grey arrows. The yellow and grey dotted line shows the X-ray/radio star formation correlation of \citet{Ranalli2003}. The magenta line shows our best correlation fit for the starburst sources (eq. \ref{XSFCorr}).}\label{LX_LR}
\end{figure*}
There is a danger, when assessing luminosity/luminosity plots, to forget that both quantities have a common dependence with redshift. This is evident when comparing Figs. \ref{LR_LI} to \ref{LX_LR} with their flux counterparts in section \ref{FCorr}. For this reason, we have carried out a partial correlation analysis, to test the strength of the luminosity correlations for various subsets. Partial correlation analysis measures the degree of association of two variables (in our case, the two luminosities), when the effect of a third variable (in our case, redshift) is removed. The method we have used for this work is based on Kendall's rank correlation coefficient, which is non-parametric, meaning that it does not rely on any assumptions for the two variables tested. The derivation is described in detail by \citet{Akritas1996}. An advantage of this method is that it works with censored data (upper limits), allowing us to keep our X-ray upper limits. We found the results to agree quite well with those of \citet{Hardcastle2006,Hardcastle2009}, and \citet{Mingo2014}. The most relevant results are listed in Table \ref{parcorr}.
\begin{table}\small
\caption{Results of the partial correlation analysis for the most relevant source subsets. The number of sources in each subset is given in column 3. Column 4 lists the values for Kendall's $\tau$; column 5 shows the square root of the variance; in column 6, $\tau/\sigma$ gives an idea of the strength of the correlation between the luminosities in column 1 in the presence of redshift. We consider the correlation significant if $\tau/\sigma>3$.}\label{parcorr}
\centering
\begin{tabular}{llcccc}\hline
L tested&Subset&n&$\tau$&$\sigma$&$\tau/\sigma$\\\hline
\multirow{2}{*}{L$_{1.4 GHz}$-L$_{12 \mu m}$}&Starburst&114&0.58&$6.36\times10^{-2}$&9.15\\
&AGN&577&0.27&$2.38\times10^{-2}$&11.35\\\hline
\multirow{2}{*}{L$_{2-10 keV}$-L$_{12 \mu m}$}&Starburst&114&0.21&$4.81\times10^{-2}$&4.36\\
&AGN&577&0.30&$2.26\times10^{-2}$&13.50\\\hline
L$_{2-10 keV}$-L$_{1.4 GHz}$&AGN&577&0.31&$2.33\times10^{-2}$&13.50\\\hline
\end{tabular}
\end{table}
Fig. \ref{LR_LI} shows the 1.4 GHz luminosity versus the 12 $\mu$m luminosity for all the sources in the W3 z sample, using the same classifications we derived from Fig. \ref{C_C_all}. As we had already anticipated from the flux distributions in Fig. \ref{w3_combFlux}, and as we can see in Table \ref{parcorr}, the starburst sources, as well as some of the spirals, follow a strong correlation. We find $\tau / \sigma =9.13$, a highly significant correlation. Some of the AGN seem to follow an extension of this correlation (radio-quiet AGN), while others have larger radio luminosities (radio-loud AGN), that also seem to span a wide range of values. The elliptical galaxies and some of the spirals also seem to have a radio excess with respect to the starburst sources, indicating that they host radio-loud AGN given that, as we remarked in section \ref{FCorr}, the correlation derived from the starburst sources represents the maximum degree of star formation we can detect with the flux limits of FIRST and NVSS.
We calculated the radio/mid-IR star formation correlation \citep[an extension of the radio/FIR correlation originally described by][]{VDKruit1973,Condon1982,DeJong1985} \citep[see also e.g. the NVSS/IRAS results of][]{Yun2001} for the starburst sources in Fig. \ref{LR_LI}. For this and all subsequent linear fits we used the Bayesian MCMC (Markov chain Monte Carlo) code developed by \citet{Hardcastle2009}, which can work with upper limits, when present. We excluded the three most obvious outliers (which have high redshifts and AGN-like X-ray luminosities). The resulting correlation will also be used in section \ref{RL_RQ_section}, as a baseline to establish the break between star-forming sources and radio AGN. The correlation we found is:
\begin{equation}\label{SFcorr1}
\log(L_{1.4 GHz})=(0.86\pm0.04) \log(L_{12 \mu m}) + (1.4\pm1.5)
\end{equation}
The MCMC fit also provides a measure of the intrinsic scatter, $\Delta_{LR-LI}=0.54\pm0.05$ (in linear units), such that e.g. a $3\sigma$ distance from the line fit would be the equivalent of multiplying the linear equivalent of eq. \ref{SFcorr1} by $(1+\Delta_{LR-LI})^{3}$ (see also Section \ref{RL_RQ_section} for more details).
We have also plotted in Fig. \ref{LR_LI} the correlation originally obtained by \citet{Gruppioni2003}, which would translate in our units as:
\begin{equation}\label{GruppioniCorr}
\log(L_{1.4 GHz})=(1.09\pm0.05) \log(L_{12 \mu m}) -(8.76\pm0.54)
\end{equation}
Our results are not entirely consistent with those of \citet{Gruppioni2003}. We find a flatter slope, but this could be due to the different selection criteria and redshift ranges covered by both samples, as well as the limited range of luminosities spanned. In terms of redshift-corrected fluxes, the slope in eq. \ref{SFcorr1} corresponds to a value of the IR/radio flux ratio $q_{12}\sim0.78$, which seems compatible with the results obtained (at 24 $\mu$m) from Spitzer data \citep[e.g.][]{Appleton2004,Garrett2015}. The FIR/radio star formation correlation extrapolates linearly quite well into the mid-IR since, even though both the IR and the radio can underestimate star formation at low galaxy luminosities, they do so in a way that the correlation is preserved \citep{Bell2003}. However, recent results show that there may be a dust temperature dependence \citep{Smith2014}, so the results need to be carefully checked for each sample.
It is interesting to note in Fig. \ref{LR_LI} that although not many sources with QSO-like 12 $\mu$m luminosities ($\ge10^{45}$ erg s$^{-1}$) seem to follow the extrapolation of the star-formation correlation, as most luminous sources also seem to be fairly radio-loud, there are indeed a few that do so. This is probably one of the factors driving the correlation we see in Table \ref{parcorr}. Even without a detailed analysis of their star formation rate it is difficult to see how such high radio luminosities could be achieved purely through star formation, and indeed if these sources, which predominantly inhabit the higher end of our redshift distribution, would be detected at all in FIRST/NVSS based solely on their star formation. It is very possible that their emission is also arising from jets and lobes, albeit less powerful ones than those of their more radio-luminous counterparts, or that shocks driven by powerful radiative winds are producing relativistic particles that, in turn, produce synchrotron radio emission, as suggested by \citet{Zakamska2016,Nims2015}, or a combination of both factors. In any case, it is clear that it might not be wise to use the mid-IR/radio star formation correlation to draw conclusions on the star formation rate of very luminous AGN.
Our conclusions are reinforced by what we observe in Fig. \ref{LX_LI}, which shows the 12 $\mu$m luminosity versus the 2--10 keV luminosity for our sources. As we saw in our previous work, all the sources over the AGN barrier seem to follow a fairly tight correlation (Table \ref{parcorr}), which holds even when we consider the common dependence with redshift, with a few outliers that may suffer from beaming (if they have an X-ray excess) or heavy obscuration (if they have an IR excess). This makes sense, as both the mid-IR and the X-rays are expected to be very good proxies for AGN activity. We have plotted a horizontal line at $L_{2-10 keV}=5\times10^{41}$ erg s$^{-1}$ as a reference, to indicate the point above which the X-ray emission we observe is most likely to originate in AGN activity (only very high star formation rates, $>100$ M$_{\odot}$ yr$^{-1}$, can produce X-ray luminosities around this break without any AGN contribution). The exact luminosity at which this happens is a matter of debate, as many (radio-quiet) AGN studies place it at $10^{42}$ erg s$^{-1}$, but we have seen in our previous work that the break between HERGs (high excitation or radiatively efficient radio galaxies) and LERGs seems to occur closer to $10^{41}$ erg s$^{-1}$, so we have plotted the line at an intermediate value. We discuss some of the implications of this choice in section \ref{RL_RQ_section}.
We see that, despite some of them following the tail of the starburst sources on the previous plot, the vast majority of the AGN have X-ray luminosities that leave no doubt as to the nature of the X-ray emission. The few outliers we see are likely luminous infrared galaxies that have W1-W2 values slightly larger than 0.5, as the diagram from \citet{Lake2012} shows that there is some overlap between the populations in the WISE colour/colour diagram. A surprisingly large fraction of spiral galaxies also seem to harbour X-ray luminous AGN, more than was immediately apparent from the fluxes in Fig. \ref{w3_hardX}. Looking again at Fig. \ref{LR_LI}, the fraction of spiral galaxies that also have substantial radio emission seems to be smaller than that of spirals with bright X-ray AGN, thus the number of fairly radio-loud Seyferts we find is not large, but it proves again that these sources do indeed exist. In fact, given that the spiral sources are hit the hardest by the W3 S/N cut, it seems that we really need deeper mid-IR (and X-ray) data to properly study this population.
We carried out a MCMC linear fit to the AGN sources, using the method of \citet{Hardcastle2009}, as described above, to compare our results with those obtained by \citet{Hardcastle2009} and \citet{Gandhi2009} for radio-loud AGN and Seyfert galaxies. We excluded three outliers with X-ray luminosities $4\sigma$--$5\sigma$ above the correlation, which artificially steepened it (it is possible that the redshifts or fluxes for these sources are not entirely correct, or that they are relativistically beamed). Although it is clear at this point that several of the sources in our `spiral' category harbour Seyfert nuclei, we decided to work exclusively with the sources labelled as `AGN', for consistency with our earlier selections. The correlation of \citet{Hardcastle2009} is:
\begin{equation}\label{HcastleCorr}
\log(L_{12 \mu m})=(0.97^{+0.23}_{-0.12})\log(L_{2-10 keV})+(0.91^{+5.35}_{-10.13})
\end{equation}
which has large constraints due to the lower number of sources in the 3CRR sample. Translated to our units, the correlation of \citet{Gandhi2009} is:
\begin{equation}\label{GandhiCorr}
\log(L_{12 \mu m})=(1.11\pm0.07)\log(L_{2-10 keV})-(5.54\pm0.05)
\end{equation}
The best linear fit that we obtained for all the AGN was:
\begin{equation}\label{AGNCorr1}
\log(L_{12 \mu m})=(0.82\pm0.01)\log(L_{2-10 keV})+(7.6\pm0.5)
\end{equation}
which is flatter than those of \citet{Hardcastle2009} and those of \citet{Gandhi2009}, as can clearly be seen in Fig. \ref{LX_LI}. This difference may be due, at least in part, to our X-ray luminosity cut, as well as the different selections arising from the instruments employed in each sample, the fact that we are including quasars, but excluding `spiral' AGN in our correlation, and the presence in our sample of some sources with high mid-IR luminosities, which could be contributing to the flat slope. The scatter for our correlation is (in linear units) $\Delta_{LX-LI}=1.25\pm0.08$, larger than that found by \citet{Hardcastle2009} ($0.32\pm0.05$), which is consistent with the fact that with MIXR we have sampled a broader range of luminosities and populations.
Back in Fig. \ref{LX_LI}, the starburst sources seem to follow a distribution parallel to that of the AGN, as we saw on the flux plot of Fig. \ref{w3_hardX}, although there is some scatter, and several of the starburst sources have only upper limits for their X-ray fluxes; this is reflected by the relatively weak correlation in Table \ref{parcorr}. The elliptical galaxies seem to fall off both the AGN and the star formation correlations, which is consistent with what we have previously observed in LERGs \citep{Hardcastle2009,Mingo2014}.
On Fig. \ref{LX_LR} we can see the 2--10 keV luminosity versus the 1.4 GHz luminosity for our sources. We have again drawn the horizontal `AGN barrier' at $L_{2-10 keV}=5\times10^{41}$ erg s$^{-1}$, as a reference. There seems to be a broad correlation for the AGN and the spiral sources with high X-ray luminosities (Seyferts), which was not evident on the flux plot of Fig. \ref{combFlux_hardX}, and which seems to hold in the presence of redshift (Table \ref{parcorr}). We will see in section \ref{RL_RQ_section} that this correlation is much weaker when we also consider the AGN with W1-W2$<0.5$ on the WISE colour/colour plot. The quiescent spirals, ellipticals and starburst galaxies fall off the correlation for AGN, as expected. Again, the ellipticals display higher radio luminosities than the other two groups, arising from the presence of radiatively inefficient AGN.
Although the X-ray/radio star formation correlation is not very strong for our sample ($r\sim0.5$), we carried out a MCMC linear fit to the starburst sources, again excluding three very X-ray bright outliers, to compare the results with those of \citet{Ranalli2003}, who found (in our units):
\begin{equation}
\log(L_{2-10 keV})=(1.08\pm0.09) \log(L_{1.4 GHz}) -(0.4\pm2.7)
\end{equation}\label{RanalliCorr}
Our best fit shows:
\begin{equation}
\log(L_{2-10 keV})=(1.37\pm0.06) \log(L_{1.4 GHz}) -(12.3\pm2.5)
\end{equation}\label{XSFCorr}
Our result is steeper than that of \citet{Ranalli2003}, but compatible if we consider the large underlying uncertainties and the small range of luminosities covered, as well as the fact that we probed lower X-ray luminosities with our sample \citep[see also the more recent results of][]{Ranalli2012}. The scatter for this correlation is (in linear units) $\Delta_{LX-LR}=1.2\pm0.3$, also rather large.
Figs. \ref{LR_LI} to \ref{LX_LR} reinforce the conclusions we reached in section \ref{FCorr} about the nature of the emission in the different source populations, suggesting that our early diagnostics were correct. The spiral galaxies, being a mixed population, require extra care for classification. Data from the next generation of radio and X-ray instruments will prove invaluable to better study this area of the WISE colour/colour plot, but deeper surveys will also reveal additional complexity: our MIXR sample, complex as it already is, is still dealing with relatively shallow flux limits in all the bands, particularly in the radio, and picking up fairly bright sources. Further data need to be obtained, and new techniques need to be developed to better study host-dominated AGN, particularly those with radio jets and lobes. This is a fairly neglected population in AGN studies, that may provide key clues to further our understanding of the life cycles of AGN feedback and its impact on the star formation activity of AGN hosts, as we have seen in sources such as Circinus \citep{Mingo2012} and NGC 6764 \citep{Croston2008}.
\section{Revisiting the diagnostics: radio-loudness, accretion mode, and star formation}\label{RL_RQ_section}
\begin{figure*}
\centering
\includegraphics[width=0.95\textwidth]{RL_RQ_selection}
\caption{1.4 GHz versus 12 $\mu$m luminosity log/log plot for the MIXR sources, illustrating the radio/IR star formation correlation, and the method we used to define the radio-loud and radio-quiet AGN, LERG/LINER and galaxy samples. Only the sources that pass the WISE W3 cut are plotted. The equation for the $3\sigma$ line is: $\log(L_{1.4 GHz})=0.86 \log(L_{12 \mu m}) + 2.0$. We have also plotted a line at 5$\sigma$ above the star formation correlation, for reference (equation: $\log(L_{1.4 GHz})=0.86 \log(L_{12 \mu m}) + 2.3$).}\label{RL_RQ}
\end{figure*}
\begin{table*}\small
\caption{Number of sources in each subset after the new classifications defined from Fig. \ref{RL_RQ}. See also Table \ref{nSources} for the statistics with the old source classifications.}\label{nSources_RL_RQ}
\centering
\begin{tabular}{llcc}\hline
Subset name&Description&\multicolumn{2}{c}{Number of sources}\\\hline
&&W3 S/N$\geq 3$&W3 S/N$< 3$\\\hline
RL AGN&$\log(L_{1.4 GHz})/\log(L_{12 \mu m})\geq(0.86+2.0)$&505&413\\
RQ AGN&$\log(L_{1.4 GHz})/\log(L_{12 \mu m})<(0.86+2.0)$; $L_{2-10 keV}\geq 5\times 10^{41}$ erg s$^{-1}$&218&0\\
LERG/LINER&subset of RL AGN with $L_{2-10 keV} < 5\times 10^{41}$ erg s$^{-1}$&69&34\\
Galaxies&$\log(L_{1.4 GHz})/\log(L_{12 \mu m})<(0.86+2.0)$; $L_{2-10 keV} < 5\times 10^{41}$ erg s$^{-1}$&211&2\\\hline
\end{tabular}
\end{table*}
In the previous sections we have shown that our early diagnostics are very efficient to pre-classify sources based on their various fluxes, mid-IR colours, and hardness ratios. However, and as introduced in section \ref{FCorr}, within each of the groups we defined from Fig. \ref{C_C_all} we see a range of properties that point to a mix of underlying populations (see also Table \ref{Activity}). Now that we know how to identify the different types of emission in each of them, it might be more efficient to redefine the populations based on their physical properties, rather than their mid-IR colours.
Based on their activity we can distinctly identify non-active star-forming galaxies, radio-quiet AGN, and radio-loud AGN. As a reminder of what we introduced in section \ref{FCorr}, we refer to radio-quiet AGN as sources where the radio emission we detect is likely to be originated mainly from stellar processes, accelerated particles in wind-driven shocks \citep[see][]{Nims2015,Zakamska2016} or, if arising from a jet and lobes, they are small and faint, and the AGN produces the bulk of its emission as radiative output in the other bands. Conversely, we refer to radio-loud AGN as those that have substantial kinetic output in the form of jet and lobes, which we measure as radio emission well above the star formation correlation. The radiatively inefficient LERG/LINER sources also follow these last criteria, so they are a subset of radio-loud AGN.
We used the radio/mid-IR star formation correlation we derived in section \ref{LCorr} (eq. \ref{SFcorr1}) as the basis of our new classification. As illustrated in Fig. \ref{RL_RQ}, any sources with radio emission in excess of $3\sigma$ over the correlation most likely harbour luminous radio jets and/or lobes, and thus we can classify them as `RL AGN' (this category includes both LERGs/LINERs and the radio-loud fraction of the radiatively efficient AGN, or HERGs). $\sigma$ is defined from the MCMC linear fit intrinsic scatter, $\Delta_{LR-LI}$, as detailed in eq. \ref{SFcorr1}, and the equation for the $3\sigma$ line is given in the caption of Fig. \ref{RL_RQ}. In Figs. \ref{LX_LI} and \ref{LX_LR} we used a barrier to define a reference X-ray luminosity above which a source must host an X-ray AGN (L$_{2-10 keV}=5\times10^{41}$ erg s$^{-1}$); applying that barrier to the data in Fig. \ref{RL_RQ}, we can distinguish between `galaxies' and `RQ AGN', and also define the subset of RL AGN that are clearly LERG/LINER sources. Our `radio-loud barrier' is fairly conservative, and likely to classify as non-active galaxies several sources with weak but non-negligible jets and lobes, but such an approach might be necessary for the brighter, QSO-like sources, where some radio emission could arise in radiatively-driven shocks, as mentioned above. Even though the break between LERG and HERG for our sources is determined by the X-ray luminosity, we can also see it clearly in the mid-IR in Fig. \ref{RL_RQ}, and it happens at the same range of mid-IR luminosities as those observed by \citet{Gurkan2014}. The statistics for the new source subsets are listed in Table \ref{nSources_RL_RQ}.
The percentage of non-active galaxies in our sample is $\sim17$ per cent ($\sim27$ per cent if we only include the sources that pass the W3 cut, which eliminates mostly RL AGN, see Table \ref{nSources_RL_RQ}). This fraction is smaller than those found with FIRST and optical data by \citet{Magliocchetti2002,Ivezic2002}, but that is not entirely surprising, considering that we are using flux cuts in more wavelengths, and NVSS data as well as FIRST. Comparing Figs. \ref{RL_RQ} and \ref{LR_LI}, we can also see that the LERG/LINER in our sample inhabit mostly sources with elliptical colours (although there are a few outliers).
\begin{table*}\small
\caption{Results of the partial correlation analysis for the most relevant source subsets. The number of sources in each subset is given in column 3. Column 4 lists the values for Kendall's $\tau$; column 5 shows the square root of the variance; in column 6, $\tau/\sigma$ gives an idea of the strength of the correlation between the luminosities in column 1 in the presence of redshift. We consider the correlation significant if $\tau/\sigma>3$. Please see the source classifications and statistics in Table \ref{nSources_RL_RQ}. HERG are RL AGN that are not LINERs/LERGs.}\label{parcorr2}
\centering
\begin{tabular}{llcccc}\hline
L tested&Subset&n&$\tau$&$\sigma$&$\tau/\sigma$\\\hline
\multirow{4}{*}{L$_{1.4 GHz}$-L$_{12 \mu m}$}&HERG&436&0.32&$2.94\times10^{-2}$&10.93\\
&RQ AGN&218&0.49&$4.95\times10^{-2}$&9.90\\
&HERG+RQ AGN&654&0.24&$2.19\times10^{-2}$&11.04\\
&LERG/LINER&69&0.29&$7.87\times10^{-2}$&3.69\\\hline
\multirow{4}{*}{L$_{2-10 keV}$-L$_{12 \mu m}$}&HERG&436&0.39&$2.78\times10^{-2}$&13.85\\
&RQ AGN&218&0.23&$4.18\times10^{-2}$&5.60\\
&HERG+RQ AGN&654&0.35&$2.18\times10^{-2}$&15.78\\
&Galaxies&211&0.11&$3.63\times10^{-2}$&2.94\\\hline
\multirow{4}{*}{L$_{2-10 keV}$-L$_{1.4 GHz}$}&HERG&436&0.38&$2.72\times10^{-2}$&13.95\\
&RQ AGN&218&0.13&$3.88\times10^{-2}$&3.40\\
&HERG+RQ AGN&654&0.29&$2.12\times10^{-2}$&13.78\\
&Galaxies&211&0.12&$3.42\times10^{-2}$&3.42\\\hline
\end{tabular}
\end{table*}
When we extended our partial correlation analysis to the new classifications we also found some interesting results (Table \ref{parcorr2}). Our results illustrate how important large statistics are when studying populations with the amount of scatter we observe. The RQ AGN have more scatter and lower statistics in the relation between L$_{2-10 keV}$ and L$_{12 \mu m}$ that their RL counterparts, but when added to the radiatively efficient RL AGN (HERGs), they strengthen the overall correlation, which is consistent with our previous results on the 2Jy and 3CRR samples \citep{Hardcastle2006,Hardcastle2009,Mingo2014}, and it highlights the fact that both quantities are very good proxies for radiatively efficient accretion. In the relation between L$_{1.4 GHz}$ and L$_{12 \mu m}$, the RQ AGN follow a slightly different correlation from that of the HERGS (as evident in fig. \ref{RL_RQ}), and thus they don't add much to the overall AGN correlation. The correlation between L$_{2-10 keV}$ and L$_{1.4 GHz}$ shows a similar situation, due to the much larger scatter and a lower reange of L$_{1.4 GHz}$ covered by the RQ AGN with respect to their RL counterparts. These results reinforce our conclusions about the scatter in the relation between radiative and kinetic power in AGN, which we discuss in detail in section \ref{Power}. The L$_{2-10 keV}$--L$_{1.4 GHz}$ correlation for the galaxies is fairly weak, as expected from our earlier results. The low number of sources and the presence of many upper limits in the X-rays are diluting the underlying star formation correlation.
Figs. \ref{z_histo_RL_RQ} and \ref{z_histo_RL_RQ_rej} show the redshift distributions with the new classifications, for sources that pass and do not pass the W3 S/N cut, respectively. What immediately draws attention in Fig. \ref{z_histo_RL_RQ} are the very different redshift distributions for RL and RQ AGN, with the former spanning a broader range and peaking at higher $z$ than the latter. This is a selection effect caused by our radio selection, and, as we will see on section \ref{Eddington}, it has some repercussions for the luminosities and Eddington rates we observe for both populations. It is also interesting that the number of RL AGN drops quite quickly below $z\sim0.1$, while that of RQ AGN does not, and that almost all the RL AGN we detect at these low redshifts are LERG/LINER sources. This is probably caused both by selection and evolutionary effects, as the number and power of RL AGN drop quite quickly at lower $z$, and LERGs come to dominate the RL AGN population at low redshifts \citep{Best2014,Williams2015}. Unsurprisingly, the star-forming, non-active galaxies have a redshift distribution that peaks at lower $z$ values than any of the others, and all but disappear after $z\sim0.1$.
It is also interesting that there are no RQ AGN and two galaxies that do not pass the W3 cut; the bulk of sources lost in this manner are RL AGN at $z\sim$0.1--2, as we will see in the next sections, most likely Seyfert-like sources with fairly luminous radio structures.
\begin{figure}
\centering
\includegraphics[width=0.46\textwidth]{z_histo_RL_RQ}
\caption{Redshift distribution histogram for the samples defined in Fig. \ref{RL_RQ}. Only sources that pass the W3 cut are considered.}\label{z_histo_RL_RQ}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.46\textwidth]{z_histo_RL_RQ_rej}
\caption{Redshift distribution histogram for the sources that do not pass the WISE W3 S/N cut, using the sample selection criteria defined in Fig. \ref{RL_RQ}.}\label{z_histo_RL_RQ_rej}
\end{figure}
Fig. \ref{HR_RL_RQ} is essentially the equivalent of Fig. \ref{HR_all_2} with the new classifications and using luminosities instead ot fluxes. This plot shows even more clearly than those in section \ref{HR} that a mere hardness ratio cut is not enough to eliminate contamination from non-active galaxies. Overall, the RQ AGN have slightly higher hardness ratios than the RL ones, and the LERGs/LINERs are, as a population, softer than the others, which is consistent with what we know about the radio/soft X-ray correlation for jet emission \citep{Hardcastle1999}, and with the radiatively inefficient nature of the LERG. The fraction of galaxies with high hardness ratios (HR$>0$) is still large after our re-classification, $\sim57$ per cent. It is possible that some of these sources are harbouring low-luminosity AGN. Lowering the AGN barrier to L$_{2-10 keV}=10^{41}$ erg s$^{-1}$ would result in $\sim23$ per cent of the galaxies (and also $\sim 48$ per cent of the LERG/LINER) in Fig. \ref{HR_RL_RQ} being re-classified as AGN. Most of these would-be-AGN galaxies ($\sim 85$ per cent, which correspond to $\sim 19$ per cent of all galaxies) have HR$>0$. While this would not completely solve the conundrum of the `hard' galaxies, as the overall fraction of galaxies with HR$>0$ is much larger, it certainly sheds some light on the fraction of possible low-L AGN that might be present in these sources, even considering the uncertainties. By contrast, raising the threshold to L$_{2-10 keV}=10^{42}$ erg s$^{-1}$ would re-classify $\sim 4$ per cent of the RL AGN as LERG/LINER and $\sim 13$ per cent of the RQ AGN as galaxies.
\begin{figure}
\centering
\includegraphics[width=0.46\textwidth]{HR_all_RL_RQ}
\caption{Hardness ratios versus the ratio of $L_{1.4 GHz}/L_{2-10 keV}$ for the new samples defined in Fig. \ref{RL_RQ} (only those that pass the W3 cut).}\label{HR_RL_RQ}
\end{figure}
Now that we have obtained a reasonably clean separation between the AGN and the star-forming galaxies, we can also test our assumption that we have a very small fraction of obscured AGN. As shown by e.g. \citet{Alexander2008,Georgantopoulos2011}, the $L_{2-10 keV}$/$L_{6 \mu m}$ ratio is a good proxy to detect Compton-thick sources. We have not extrapolated the corresponding 6 $\mu$m luminosities, as all our other calculations are based on WISE, rather than Spitzer, but we can use our 12 $\mu$m luminosities to compare our results with those of \citet{Rovilos2014}, as shown in Fig. \ref{CT_RL_RQ_1}. We see that the fraction of sources with $L_{2-10 keV}$/$L_{12 \mu m}\leq 0.01$ (potentially Compton-thick AGN) is fairly small, more so if we consider the uncertainties at lower X-ray luminosities (see the error bars in Fig. \ref{LX_LI}), meaning that our chosen range of $N_H$ was appropriate, as such a small fraction of (potentially) Compton-thick AGN should not bias our results. The distributions for RL and RQ AGN are fairly similar, but the former tend to have marginally larger values of $L_{2-10 keV}$/$L_{12 \mu m}$. This could be caused by the fact that the RL sources have additional (soft) X-ray emission arising from the jet, some of which could be contaminating the 2--10 keV band, particularly in beamed objects. We have not plotted the LERGs/LINERs for clarity, but they would occupy the left-most end of the RL AGN distribution. Although radiatively inefficient sources produce soft X-ray emission related to the jet, but no substantial emission in the mid-IR because they have no tori, our LERG/LINER sources still show mid-IR emission from the old stellar population in the host galaxy (arising from the R-J tail of stars, not heated dust in the ISM), as we saw for the 2Jy and 3CRR sources \citep{Mingo2014}, and as studied in detail by e.g. \citet{Mason2012}.
\begin{figure}
\centering
\includegraphics[width=0.46\textwidth]{CT_RL_RQ_1}
\caption{$L_{2-10 keV}$/$L_{12 \mu m}$ as a proxy for AGN intrinsic obscuration, for the RL and RQ AGN. The distribution for the full z sample (all sources with redshifts, regardless of their W3 S/N, see Table \ref{nSources}) is also plotted as a dashed line, for reference. Please see also Appendix \ref{ExtraFigures} for the equivalent plot using the 4.6 $\mu$m luminosities.}\label{CT_RL_RQ_1}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=0.92\textwidth]{C_C_all_RL_RQ}
\caption{WISE colour/colour diagram, illustrating the mid-IR colours of the new subsets defined from Fig. \ref{RL_RQ}, both for the sources that do and do not pass the W3 S/N cut. The cross and arrow next to the legend indicate the typical size of the errors and the direction of the W3 upper limits (for the sources that do not pass the W3 cuts). We have plotted lines to indicate the cuts that we used to separate the various populations in Fig. \ref{C_C_all}, as a reference.}\label{C_C_all_RL_RQ}
\end{figure*}
Finally, Fig. \ref{C_C_all_RL_RQ} illustrates the distribution of our newly defined subsets (both with and without the W3 cut) on the WISE colour/colour diagram. This Figure illustrates the extent to which mid-IR AGN selections that use WISE are biased against low-luminosity AGN. In Fig. 3 of the work of \citet{Gurkan2014}, a large fraction of NLRGs (narrow-line radio galaxies, roughly speaking, the radio-loud equivalent to Seyferts 1.5--2) fall below the $W1-W2 \ge 0.5$ cut, with roughly half of them falling outside the more conservative wedge of \citet{Mateos2012}. We see a very similar situation for the RL AGN in Fig. \ref{C_C_all_RL_RQ}. Our plot also illustrates how the low sensitivity of W3 makes the situation even worse for the lower luminosity RL AGN, as most of them do not pass the S/N cut. The situation does not look as dramatic for the radio-quiet sources on the plot, but that is probably due to the smaller number and redshift range of RQ AGN we are able to detect in our sample; it is very likely that a similar fraction of low-luminosity, intermediate Seyferts are also excluded from WISE-selected AGN samples. Conversely, the wedge of \citet{Mateos2012} (see their Fig. 2) would be very efficient, with our sample, at eliminating contamination from red, non-active galaxies, whereas a simple $W1-W2 \ge 0.5$ cut would not be sufficient.
It might be interesting to test with an independent sample if the W3 cut has as dramatic an impact on low-luminosity RQ AGN as it does on the (non-LERG/LINER) RL AGN in our sample. These low-luminosity RQ sources are scarce in the MIXR sample because we require a radio detection, and if they are truly radio-quiet, they can only be detected based on their star formation, which limits their host type and redshift. The RQ sources at the same luminosity and redshift ranges as our W3 rejected RL AGN should also have similar W3 fluxes and S/N values, otherwise it would mean that the accretion structures are different for both populations. Extrapolating the fraction of W3 eliminated RL AGN to their RQ counterparts, even without considering the various AGN selection wedges and cuts, it becomes clear that the fraction of AGN excluded in WISE-selected samples is not trivial, by any means.
The bottom line is that, while mid-IR AGN selections, and in particular thorough methods like those of \citet{Mateos2012,Secrest2015}, are very good at selecting clean samples of bright AGN, one must keep in mind that they are biased against lower luminosity sources, particularly at higher redshifts \citep[see also e.g.][]{Rovilos2014}, and thus their conclusions cannot be extrapolated to the entire AGN population. This bias is made much worse by the fact that many of these low luminosity sources are too faint for the W3 and W4 bands. Auxiliary methods, like our radio selection, can be used to partly rectify this bias, but deeper observations are also needed.
\section{Eddington rates}\label{Eddington}
In this section we aim to test the Eddington rates for our redefined populations (excluding the non-active galaxies), to assess whether there are any systematic differences with the results of \citet{Mingo2014,Gurkan2014,Best2012}. To do so, we need to first calculate the bolometric luminosities and black hole masses for our sample.
We calculated the bolometric luminosity for our sources from the X-ray 2--10 keV rest-frame corrected fluxes, using the correlations of \citet[][eq. 21 ]{Marconi2004}:
\begin{equation}\label{Lbol}
\log(L/L_{2-10 keV})=1.54+0.24 \mathcal{L} +0.012 \mathcal{L}^{2} - 0.0015 \mathcal{L}^{3}
\end{equation} where $\mathcal{L}=(\log(L)-12)$, and $L$ is the bolometric luminosity in units of $L_{\odot}$. The bolometric luminosities used for the data points from \citet{Mingo2014} and \citet{Punsly2011} were derived from [OIII] measurements, which, unfortunately, we do not have for the entirety of our current sample. However, we do not expect the results to be systematically different, especially considering the uncertainties involved. We calculated the $L_{bol,X}/L_{bol,[OIII]}$ ratio for the 2Jy, 3CRR and Punsly sources, and found it to be 0.98, with a standard deviation of 0.11.
Given that only about half of our sources have spectroscopic information, and because we want to use a consistent method for the entire sample, we cannot use optical line widths to calculate the black hole masses. We thus derived B band magnitudes using the equations of \citet{Jester2005} and the redshift-corrected fluxes for the g and r bands of SDSS, and calculated the black hole masses using the relations from \citet{Graham2007}. This method for deriving the black hole mass, originally derived from the work of \citet{Kormendy1995,Magorrian1998} is not accurate when the source is a QSO, as the emission from the QSO completely dominates over that of the host. We eliminated the potential QSOs by applying a (conservative) luminosity cut at $L_{2-10 keV}=5\times10^{44}$ erg s$^{-1}$, based on where we found the break between broad-line radio galaxies and QSOs in our previous work \citep{Mingo2014,Hardcastle2009}, and excluding sources above this luminosity. After this cut we were left with 347 RL AGN and 215 RQ AGN (and 372 W3 rejected RL AGN). When we compare these numbers with those in Table \ref{nSources_RL_RQ}, we clearly see that the overwhelming majority of the QSO-like sources in our sample classify as radio-loud (see also section \ref{Power}).
Although some recent results \citep{Kormendy2013b,Kormendy2013} cast doubt on both the $M_{BH}$/L relation and the M-$\sigma$ relation of \citet{Ferrarese2000,Gebhardt2000}, they are still the best indirect methods we have to estimate the black hole masses of AGN.
The distribution of inferred black hole masses is rather similar for the RQ and RL AGN in our sample, as Fig. \ref{BH_masses_RL_RQ_noQSO} illustrates. Although the former seem to peak at slightly lower masses and show a narrower distribution than the latter, the difference is very subtle, and only evident when we consider that the LERG/LINER sources, which are a subset of the RL AGN, take up a large fraction of the lower-mass end of the distribution. This is quite consistent with the fact that RL AGN tend to inhabit hosts with larger $M_{BH}$, but the black hole mass distributions of RL and RQ sources do not look very different in our sample.
\begin{figure}
\centering
\includegraphics[width=0.46\textwidth]{BH_masses_RL_RQ_noQSO}
\caption{Black hole masses for the source subsets defined from Fig. \ref{RL_RQ}, excluding QSOs. Only the sources that pass the W3 S/N cut are considered.}\label{BH_masses_RL_RQ_noQSO}
\end{figure}
Fig. \ref{z_histo_RL_RQ_noQSO}, showing the distribution of sources in $z$, helps us to understand why the difference between the black hole mass distributions for RL and RQ AGN is not more evident. As we saw in Fig \ref{z_histo_RL_RQ}, because of a combination of evolution effects and our radio selection, we are selecting RQ AGN at lower redshifts than the RL AGN. Eliminating the QSOs from our sample mitigates this bias, as luminous QSOs tend to appear at larger $z$, but it does not eliminate it completely. Moreover, if we consider that the LERGs now make up a larger fraction of the RL AGN, the difference between the radiatively efficient RL and RQ source distributions is even more marked.
\begin{figure}
\centering
\includegraphics[width=0.46\textwidth]{z_histo_RL_RQ_noQSO}
\caption{Redshift distributions for the source subsets defined from Fig. \ref{RL_RQ}, excluding QSOs. Only the sources that pass the W3 S/N cut are considered.}\label{z_histo_RL_RQ_noQSO}
\end{figure}
That difference in redshift distributions, and thus black hole masses and luminosities, has a clear impact on the results we obtain for the Eddington rates, as shown in Fig. \ref{Ledd_RL_RQ_noQSO} (See Appendix \ref{ExtraFigures} for equivalents to Figs. \ref{Ledd_RL_RQ_noQSO} and \ref{Ledd_RL_RQ_noQSO_rej} with that include the jet power). While the LERG/LINER sources have Eddington rates comparable to those we obtained for the 2Jy and 3CRR sources \citep{Mingo2014}, and compatible with the results of e.g. \citet{Best2012}, the HERG (RQ AGN + non-LERG RL AGN), in bulk, seem to have higher Eddington rates than the radio-quiet sources.
We highlight that this is not likely to be an underlying physical difference, but a selection effect. With the catalogues we use to build our sample, we are selecting bright radio-loud sources at high redshift (thus with still growing black holes) and faint radio-quiet sources at low redshift (where the black hole masses are larger). This should serve as a warning when comparing samples of radio-loud and radio-quiet sources: a comparison based purely on one criterion, be it luminosity, redshift, or black hole mass, is unlikely to truly compare similar sources. Several factors need to be taken into account to minimise bias.
\begin{figure}
\centering
\includegraphics[width=0.46\textwidth]{Ledd_RL_RQ_noQSO}
\caption{Eddington rates for the source subsets defined from Fig. \ref{RL_RQ}, excluding QSOs. Only the sources that pass the W3 S/N cut are considered (see also Appendix \ref{ExtraFigures} for an equivalent Figure including the jet power in the Eddington luminosity).}\label{Ledd_RL_RQ_noQSO}
\end{figure}
Fig. \ref{Ledd_RL_RQ_noQSO_rej} shows that, by eliminating most of the Seyfert-like RL AGN with the W3 cut, we are also essentially eliminating those RL sources that are most similar to the RQ AGN in terms of their Eddington rate (though not entirely in terms of $z$, see again Fig. \ref{z_histo_RL_RQ_rej}).
\begin{figure}
\centering
\includegraphics[width=0.46\textwidth]{Ledd_RL_RQ_noQSO_rej}
\caption{Eddington rates for the LERG/LINER and radio-quiet AGN that do not pass the W3 S/N cut. QSOs are excluded (see also Appendix \ref{ExtraFigures} for an equivalent Figure including the jet power in the Eddington luminosity).}\label{Ledd_RL_RQ_noQSO_rej}
\end{figure}
\section{Jet vs radiative output}\label{Power}
In our previous work \citep[e.g.][]{Hardcastle2009,Mingo2014} we have observed that the `radio loudness' of a source is not easily determined based just on its radiative power. More than a sharp dichotomy, the radio-loud/quiet transition seems to be gradual, hinting at underlying mechanisms that regulate how accretion power is transformed into radiative output (luminosity) and kinetic output (jets or winds). However, this effect is very difficult to observe with small samples that use monochromatic flux limits or that study just a subset of AGN. MIXR, due to its mixed population, is ideally suited to study the relationship between jet and radiative output from LERGs, through Seyferts, to QSOs, and across a wide range of radio powers.
The question of how the jet and the radiative output are regulated is a very complex one. While a parallel between X-ray binaries and AGN is frequently drawn \citep[e.g.][]{Connolly2016}, there are some important considerations to take into account. LERG can be likened to the low/hard, radiatively inefficient state in low-mass X-ray binaries \citep{Fender2014,Yuan2014}, while, on a first approach, the parallel of the radiatively efficient transition state between the low/hard and high/soft states in low-mass X-ray binaries (LMXRB) would be the HERG (non-LERG, radiatively efficient RL AGN). In LMXRB only sources in the high/soft state can produce winds, and they never have jets. However, we know that there there are many (radiatively efficient) Seyferts and QSOs with jets, sometimes coexisting with winds \citep[e.g.][]{Nesvadba2008,Mullaney2013,Collet2016,Harrison2015}, and that they show examples of both steady, slow jets and fast, relativistic ones. To complicate matters further, we also now know that the radio luminosity we measure depends on the environment around the host \citep{Hardcastle2013,Hardcastle2014,English2016}. And it is also important to keep in mind that the timescales involved in AGN activity are very long; as such, low frequency radio observations, used to study the jet efficiency, are often reflecting the activity level of the AGN on scales of Myr, while X-ray and even mid-IR observations provide measurements of the core activity on much shorter timescales (essentially instantaneous, in the case of the X-rays).
The fact that we can observe the same type of jets in LERG and HERG means that the accretion disc \citep[in the classical][sense, at least]{Shakura1973} cannot be the element responsible for jet generation. In the model of \citet{Blandford1977}, a jet is expected when both the spin of the black hole and the magnetisation of the surrounding material are high. It is possible that in AGN, because of the larger volumes of gas and more inhomogeneous feeding rates compared to XRBs, it is feasible to accrete enough magnetic flux to launch a jet even in the radio-quiet regime. The role of the spin has been recently brought to light \citep[see e.g.][]{McNamara2011,Done2014,Done2016}, especially for sources with very powerful jets \citep{Tchekhovskoy2011}. It is also possible, however, that the magnetic flux accretion alone is the main driving mechanism, and that the episodic accretion of hot or cold gas is what truly drives the difference between inefficient and efficient sources \citep{Hardcastle2007b,Sikora2013,Ineson2013, Ineson2015}, or it could be a combination of both, at least in some sources \citep[see e.g.][]{Nemmen2014,Nemmen2015}.
What we observe is a large scatter in the plots of e.g. \citet{Punsly2011,Mingo2014}, where for a given bolometric luminosity (radiative output) there is a wide range of possible jet powers, and vice versa. This contradicts the conclusions of \citet{Rawlings1991}, who established a tight correlation between both quantities. It is possible that their correlation arises as a selection effect, as most of the work carried out on the topic involves flux-limited samples that only select a particular subset of the radio-loud AGN population, in particular, in the case of \citet{Rawlings1991}, the most luminous radio-loud sources in the Universe. The recent Fermi results of \citet{Chen2015}, for example, seem to agree with \citet{Rawlings1991}, but if we plotted them together with our current or previous results, the scatter in the plot would be too large to support a strong correlation.
The MIXR sample is ideal to test our previous conclusions about the correlation of \citet{Rawlings1991}, as it contains a large number of sources, with a large range of radio powers and bolometric luminosities. In Figs. \ref{Punsly} and \ref{Punsly2} we have plotted the previous data points from \citet{Mingo2014}, which include the 2Jy and 3CRR sources as well as the SDSS quasars from \citet{Punsly2011}, and all the sources from our current sample, using both the WISE source classifications and those we derived in section \ref{RL_RQ_section}. Please see again Tables \ref{nSources} and \ref{nSources_RL_RQ} for the statistics of each population. Although it is difficult to see in Figs. \ref{Punsly} and \ref{Punsly2}, due to the large number of points, there is an overlap between the MIXR, 2Jy+3CRR, and \citet{Punsly2011} sources. The MIXR sources also bridge the gap between the other two samples, which have more restrictive selection criteria: powerful radio sources for the 2Jy+3CRR sample, SDSS-classified QSOs with no X-ray selection in the case of the sample from \citet{Punsly2011}.
The calculation of the bolometric luminosity for our sources is described in section \ref{Eddington} (eq. \ref{Lbol}). For the jet output we used again the method by \citet[][eq. 12]{Willott1999} that we applied in \citet{Mingo2014}:
\begin{equation}\label{Willott}
Q=3 \times 10^{38}f^{3/2}L_{151}^{6/7} W
\end{equation} where $L_{151}$ is the luminosity at 151 MHz, in units of $10^{28}$ W Hz$^{-1}$ sr$^{-1}$, and f=15 \citep[see the discussion by][for the origin and possible values of f]{Hardcastle2007b}. We derived the 151 MHz fluxes by extrapolating the rest-frame corrected 1.4 GHz fluxes, using a spectral index of 0.8 for all the sources (see the discussion in section \ref{LDiag}). We calculated, and present here, $Q$ and $L_{bol}$ in Watts, rather than erg s$^{-1}$, for easier comparison with previous work.
\begin{figure*}
\centering
\includegraphics[width=0.92\textwidth]{Punsly}
\caption{Radiative ($L_{bol}$) versus jet ($Q$) output for our sources (colours and symbols as in Fig. \ref{C_C_all}). The magenta squares represent the 2Jy and 3CRR sources from \citet{Mingo2014}, and the cyan inverted triangles represent the SDSS QSOs from \citet{Punsly2011}. Error bars are omitted for clarity, their sizes are comparable to the X-ray luminosity error bars in Figs. \ref{LX_LI} and \ref{LX_LR}.}\label{Punsly}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.92\textwidth]{Punsly2}
\caption{Radiative versus jet output for the MIXR sources that pass the W3 cut, as in Fig. \ref{Punsly}, but using the new classifications from Fig. \ref{RL_RQ}. Again, we have plotted the 2Jy and 3CRR sources from \citet{Mingo2014}, and the SDSS QSOs from \citet{Punsly2011}, for reference. Error bars are omitted for clarity, their sizes are comparable to the X-ray luminosity error bars in Figs. \ref{LX_LI} and \ref{LX_LR}.}\label{Punsly2}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=0.46\textwidth]{Punsly2_rej}
\caption{Same as Fig. \ref{Punsly2}, for sources that do not pass the W3 cut. No RQ AGN and only 2 Galaxies are rejected by the W3 cut, see Table \ref{nSources_RL_RQ}. Error bars are omitted for clarity, their sizes are comparable to the X-ray luminosity error bars in Figs. \ref{LX_LI} and \ref{LX_LR}.}\label{Punsly2_rej}
\end{figure}
Although star-forming galaxies should not be plotted in Fig. \ref{Punsly}, as their radio emission does not originate from AGN activity, we have included the elliptical, spiral, and starburst sources on the plot, to show where mis-classified sources might lie (as we have seen, using the WISE colour/colour classification some AGN are classified as non-active galaxies, and vice versa). The difference between the loci for the bulk of the starburst and elliptical populations also illustrates how the latter have excess radio emission that cannot be accounted for by even the most powerful star formation, confirming again that our elliptical galaxies host LINER/LERG AGN. Interestingly, a few spiral galaxies also seem to have excess radio emission. Using the new classifications in Fig. \ref{Punsly2}, we see a much clearer separation between the LERG and the non-active galaxies. The overlap between RL and RQ AGN in this plot reflects the degree of uncertainty present in both quantities plotted, but also the fact that the transition between both classes is gradual, rather than abrupt. Fig. \ref{Punsly2} also shows that most of the sources with QSO-like luminosities ($\sim10^{45}$ erg s$^{-1}$, or $\sim10^{38}$ W, and above) are classified as RL AGN.
Figure \ref{Punsly2_rej} is interesting, because it illustrates exactly what types of radio source are eliminated by the W3 cut: RL AGN with Seyfert-like radiative luminosities, but jet outputs that rival those of the SDSS QSOs. These sources are very interesting, as they, presumably, do not have the high accretion rate of luminous QSOs available to produce jets, and yet they can output a similar amount of radio power. It is also now clear that the RQ AGN with similar bolometric luminosities pass the W3 cut because they are found at lower redshifts (see Figs. \ref{z_histo_RL_RQ} and \ref{z_histo_RL_RQ_rej}).
In light of the large scatter in Figs. \ref{Punsly} and \ref{Punsly2}, and knowing that both quantities have a common dependence with redshift, we carried out a partial correlation test on $L_{bol}$ and $Q$, including all the sources from \citet{Punsly2011} and \citet{Mingo2014}. We found a strong correlation in the presence of redshift ($\tau/\sigma = 16.94$), although it is partially driven by the brightest sources (removing the QSOs lowers the ratio to $\tau/\sigma = 12.29$), as we observed previously for the 2Jy and 3CRR sources. This is reassuring, as, intuitively, we do expect these quantities to be tied through their common dependence on accretion, but it is important to note that we are still working with flux-limited samples: our plot is not complete, especially on the left side (where the low radio luminosity sources we are not detecting with FIRST/SDSS would likely fall). However strong the results of the partial correlation, it is undeniable that the amount of scatter we observe (4--5 orders of magnitude in each direction) presents a very different picture from that of the tight correlation of \citet{Rawlings1991}. A critical question in radio-loud AGN research remains: what is mediating between the radiative and jet output in AGN, and how can we measure it?
Some care must be taken when considering measurements of the kinetic output of an AGN, even beyond any scatter we may have introduced by using a single spectral index to extrapolate our fluxes (e.g. increasing/decreasing the spectral index to 1.0/0.6 would increase/decrease $Q$ by a factor of $\sim$1.5--2, the effect increasing with $z$). The recent results of \citet{Godfrey2016} show that the observed correlation between radio power and $Q$ on which the work of \citet{Cavagnolo2010} is based may not be as strong as previously believed, due to effects such as the different behaviour of low (FRI) and high (FRII) power radio galaxies \citep[see e.g.][]{Croston2005,Godfrey2013}, the dependence of radio luminosity with environmental density \citep[see e.g.][]{Hardcastle2013,Hardcastle2014,English2016}, and the common distance dependence of $Q$ and the radio luminosity \citep{Godfrey2016}, on which the correlation is based. However, the $Q$ parameter of \citet{Willott1999} is based on an analytical model, and even a different dependence of $Q$ with $L_{radio}$ for FRI and FRII is unlikely to decrease the scatter in Fig. \ref{Punsly} to the extent needed to be consistent with the correlation of \citet{Rawlings1991}, especially if high and low power sources are considered separately.
It is possible that regulating mechanisms for the jet, such as the spin, mentioned above, are introducing scatter between $Q$ and $L_{bol}$ along the X axis of Figs. \ref{Punsly} to \ref{Punsly2_rej}, although the interplay between black hole spin, disc magnetic fields, accretion rate, and jet properties is likely to be fairly complex in and of itself \citep[e.g.][]{Hawley2015}. Most importantly, we must also consider the aforementioned uncertainty in the relationship between $Q$ and the measured radio luminosity (which depends on factors such as the age and the environment of the source). Along the Y axis, it is possible that long-term variability is an important factor; as mentioned earlier, $L_{bol}$ is essentially an instantaneous measurement of the AGN power, while $Q$ (or any $L_{radio}$ at low frequencies) is a time-averaged measurement on timescales of up to a few Myr.
Fig. \ref{Punsly2} is also useful to understand the role of the individual selections in sampling different regions on the $Q/L_{bol}$ diagram: the QSOs of \citet{Punsly2011} were optically selected; the 3CRR and 2Jy sources of \citet{Mingo2014} were radio selected, and our MIXR sample was X-ray selected (additional constraints driven by the other catalogues we used, as we have seen). The fact that the three samples cover different ranges of $Q$ values (with some overlap) indicates that there are genuine differences in the underlying $Q$ (the SDSS QSOs and the 2Jy+3CRR AGN span a relatively similar range of luminosities), so differences in $Q$ must play a role in the scatter we observe on the $Q/L_{bol}$ diagram. However, the large range of $Q$ values covered by the individual populations, in particular the MIXR RQ AGN, seems to indicate that changes in $L_{bol}$ may also play a role, whose importance we need to assess.
If variability on $\sim$Myr timescales is indeed a factor introducing scatter in Fig. \ref{Punsly2}, it might have very powerful implications for our understanding of AGN feedback and its impact on the star formation history of AGN hosts. Recent works like those of \citet{Hickox2014,Stanley2015} highlight the difficulty of studying the interplay of AGN activity and star formation when the AGN is varying, although they might be correlated on longer timescales \citep[e.g.][]{DelVecchio2014} due to their mutual dependence on reservoirs of cold gas \citep[see also][]{Wild2010}. However, if Fig. \ref{Punsly2} is to be believed, the radiative output of AGN can vary by far more than the two orders of magnitude generally considered in these works, and, more importantly, so can their jet output, which is more likely to influence the host on large scales.
Although there is increasing evidence for `radiative mode' (or `radio-quiet mode') winds and powerful outflows, and a lively ongoing debate on their impact on AGN hosts \citep[see e.g. the review by][and references therein]{King2015}, it is not clear yet whether these winds can significantly affect star formation beyond the bulge, even in the most powerful sources, and how ubiquitous they really are \citep[see e.g.][]{VillarMartin2016}. We know, however, that small radio sources are very efficient at transporting enormous amounts of energy to the interstellar medium (ISM) through jet-ISM driven shocks \citep[see e.g. the energy calculations for NGC 3801 by][]{Croston2007}, with a much larger potential to disrupt star formation on galaxy scales. But that effect on star formation, either as triggering or quenching, takes several Myr to become observable, by which point the radio lobes have long faded out. Even considering star formation on longer timescales ($\sim100$ Myr), such an injection of energy has the potential to alter the overall energy budget, especially if there are periodic recurrences. Recent results show as well that at $z$ greater than $\sim1$, and unlike in the local Universe, the hosts of moderately-powerful radio-loud AGN are very actively star-forming \citep{Magliocchetti2016}, but are these sources truly the ancestors of local radio-loud AGN? If both the bolometric luminosity and the jet output of an AGN can vary by 3--5 orders of magnitude in the space of a few Myr, and we cannot extrapolate their life cycles from those we know about from LMXRB, and if the star formation rates measured are influenced by consecutive radio outflows \citep[see e.g.][on the recurrence of radio outflows]{Saikia2010} that are no longer detectable, and completely unrelated to the current radiative and jet properties of the AGN, how can we analyse the interplay between AGN activity and star formation?
We clearly need to better understand the life cycles of the radio-loud phase of AGN, and to start taking into account the `radio mode' feedback for small, host-scale sources, as well as the larger cluster-sized ones, in simulations of galaxy dynamics and evolution. The first step could be an assessment of how many sources with radiatively driven winds also have radio emission, as recent evidence seems to point out to a frequent coexistence of both \citep[see e.g.][]{Nesvadba2008,Collet2016,Harrison2015}, but we still need to assess whether in the most luminous sources the radio emission is actually produced by jets and lobes or just by particles accelerated in wind-driven shocks \citep[e.g.][]{Nims2015,Zakamska2016}, and what fraction of the observed outflows is actually produced by star formation, rather than the AGN \citep{Sarzi2016}. We also need to better understand the conditions for a coexistence of small radio outflows and star formation in gas-rich hosts \citep[e.g.][]{Frank2016}, as recent studies highlight how challenging it is both to trace black hole growth in low-luminosity AGN with starforming hosts \citep{Jones2016} and to measure star formation in brighter AGN \citep{Symeonidis2016}, even without considering their radio properties. Although we detect a number of radio-loud sources in star-forming hosts (spiral galaxy colours), we are severely hindered by the lack of sensitivity of W3 and do not have the statistics to study this population individually. A dedicated study of Seyfert-type sources with radio emission would be necessary to assess what the true fraction of radio-loud star-forming galaxies is, and how they differ from the more radio-loud gas-depleted systems, to assess the dynamical impact of the jets and lobes on the host and its star formation properties \citep[see also][]{Kaviraj2015,Kaviraj2015b}.
\begin{figure}
\centering
\includegraphics[width=0.46\textwidth]{RL_RQ_dichotomy}
\caption{Histogram of the L$_{1.4 GHz}$/L$_{2-10 keV}$ ratio for the MIXR RL AGN, RQ AGN, and LERG, showing the overlap between the distributions.}\label{RL_RQ_dichotomy}
\end{figure}
It also seems clear that we need to revisit the radio-loud/radio-quiet dichotomy. While the radio-loud and radio-quiet definitions can work well for cases where one regime (kinetic vs radiative) dominates, the definitions based on flux ratios can be very misleading, and there is increasing evidence that transition between both regimes is not as abrupt as we thought in the past (see Fig. \ref{RL_RQ_dichotomy}).
To further study this effect, we carried out a few quick tests on the 80 sources between the $3\sigma$ and $5\sigma$ lines in Fig. \ref{RL_RQ}, to assess where they stood in terms of the RL/RQ division. We checked the properties of these sources in terms of hardness ratio, redshift, black hole mass, Eddington rate, $L_{X}/L_{radio}$, and $L_{bol}/Q$. Overall, the 24 sources with $L_{X}<5\times10^{41}$ erg s$^{-1}$ seemed to show characteristics more similar to those of our LERG/LINER sources than those of our galaxies (with a few outliers), in agreement with our choice of the $3\sigma$ line as a division between both populations. For the AGN ($L_{X}\geq 5\times10^{41}$ erg s$^{-1}$) the division is less clear: overall, the 56 sources exhibited characteristics clearly in the transition between the RL and the RQ AGN populations (with a few outliers as well), for all the parameters we tested. Although the number of sources was, unavoidably, too small to carry out proper statistical tests, this `intermediate' behaviour is consistent with the overlap between both populations that we observe in Fig. \ref{RL_RQ_dichotomy}.
It is quite clear that the RL and RQ regimes are very complicatedly interwoven, and we need to better understand their relationship. With the wealth of radio data that will be made available with ongoing and future surveys, it might finally be possible to revisit the RL/RQ classifications and define better criteria to assess the interplay of the radiative and kinetic output in AGN, and its effect on AGN hosts.
\section{Summary and Conclusions}\label{Conclusions}
We have used the ARCHES xmatch statistical tool to create a large cross-correlated sample of AGN and star-forming galaxies, using the largest, most uniform catalogues available in the Mid-IR (WISE), X-rays (3XMM) and Radio (FIRST+NVSS) bands. The MIXR sample we thus obtain provides efficient and broad-reaching diagnostic tools to classify sources based on their type of activity (radio-loud and radio-quiet AGN, and star formation), even in the absence of redshifts. The techniques we have developed for MIXR can be used to triage sources for any extragalactic sample with measurements that can be translated to these bands, paving the way for classification techniques that will allow us to fully exploit the vast amounts of data that the next generation of instruments will make available.
We pre-classify our sources based on their mid-IR colours, using the WISE colour/colour plot and the results of \citet{Lake2012}, as elliptical, spiral, starburst and AGN sources. While these initial classifications provide a general idea of the type of underlying activity we can expect in our sources, there is a great deal of overlap between populations (see Table \ref{Activity}). We use first flux and magnitude plots, and then luminosity plots, to triage our sources based on their emission in each band, clearly separating star-forming, non-active galaxies \citep[for which we recover the radio/IR correlation of][]{DeJong1985,Appleton2004,Garrett2015} from radio-loud AGN, both of the radiatively efficient and inefficient \citep[LERG/LINER, see][]{Narayan1995} varieties, and from radio-quiet AGN, where the bulk of the radio emission we detect is produced by star formation, or particle acceleration in shocks \citep{Zakamska2016,Nims2015}, but which could also host minor jets and lobes.
Our results show that WISE-colour selected AGN samples are heavily biased against Seyfert-type, moderate- to low-luminosity AGN. This selection bias occurs in two ways: WISE colour cuts such as those of \citet{Assef2010,Mateos2012,Stern2012} are very efficient at selecting clean samples of luminous AGN, but necessarily omit those sources where the host contributes a substantial fraction of the total emission, as only with additional proxies, such as radio emission, it is possible to distinguish between non-active galaxies and AGN at low luminosities; the WISE W3 (and W4) band is also not sensitive enough to detect faint AGN at redshifts beyond $\sim0.1$, which is particularly detrimental to radio-loud Seyfert-like sources, as these tend to appear at higher redshifts than radio-quiet sources of similar bolometric luminosity. In fact, our sample size is cut by $\sim40$ per cent simply by imposing requirement for a signal/noise of 3 in W3, with radio-loud AGN suffering the bulk of the cut (we lose another $40$ per cent of the sample when requiring SDSS redshifts for the second part of our diagnostics).
We find that RL and RQ AGN of similar bolometric luminosities and Eddington rates are found at different redshifts, with the RL sources being found at slightly larger $z$, and our sources become `radio-louder' with increasing redshift, up to our detection limit. Our sample is biased against RQ AGN, as we require a radio detection, limiting the redshift more quickly for RQ sources than for RL ones. As a consequence, when considering both populations as a whole, our RL AGN are more luminous and have larger Eddington rates. This is clearly very likely to be a selection effect, and it illustrates one of the easiest causes of bias that can be incurred when comparing RL and RQ AGN: it is not enough to match both samples exclusively on luminosity, redshift, or Eddington rate; all these variables (plus their environments) must be taken into account to ensure that we are comparing like with like.
Perhaps the most crucial result of this work is the confirmation of the scatter we observed in the 2Jy and 3CRR sources \citep{Mingo2014} between their radiative (bolometric luminosity) and kinetic (jet) output, in contradiction with the tightness of the long-standing correlation of \citet{Rawlings1991}. These two quantities must have a common underlying mechanism, as they are both tied to accretion, but either jet regulating mechanisms \citep[e.g.][]{Done2014,Tchekhovskoy2011,Hawley2015}, dispersion in the jet power/L$_{radio}$ relationship, long-term AGN variability \citep[e.g.][]{Hickox2014,Stanley2015}, or a combination of all three, are introducing the 4--5 order of magnitude scatter we observe in our plots.
Given what we know about the potential impact of small-scale radio sources on the energetics of their hosts \citep[e.g.][]{Croston2007}, and the recently found coexistence of radiative winds and radio outflows \citep{Nesvadba2008,Harrison2015}, which has no parallel in X-ray binaries, we may need to reassess what we know about the interplay between AGN activity and star formation. If both the bolometric luminosity and the jet output of an AGN can vary by 3--5 orders of magnitude in the space of a few Myr, and we cannot extrapolate their life cycles from those we know about from LMXRB, how can we analyse the interplay between AGN activity and star formation? Although star formation occurs on longer timescales, jet-driven shocks can carry enough energy, far enough into the ISM, to potentially change the course of an on-going episode of star formation. However, we may not be able to detect whether the star formation rates we measure are influenced by consecutive radio outflows that have long faded out, and are completely unrelated to the current radiative and jet properties of the AGN because of the short timescale and wide range of AGN variability.
Radio-loud Seyferts may hold the key both to understanding the details of the jet-ISM interaction, and the mechanisms regulating the jet. Some of these sources can produce jet outputs similar to those of luminous QSOs, but at values of $L_{bol}$ and $L_{Edd}$ that are orders of magnitude lower. Unfortunately, these are exactly the sources that W3 is not sensitive enough to reliably detect, as they have similar mid-IR luminosities to those of radio-quiet Seyferts at $z\sim$0.1--0.3, but are far more distant ($z\sim$0.8--1).
We clearly need to better understand the life cycles of the radio-loud phase of AGN, both from a theoretical and from an observational point of view. A sensible first step might be to assess the fraction and properties of sources with radio emission in samples that do not use radio selections, supplemented with dedicated studies of moderate- to low-luminosity AGN, to establish larger samples that we can systematically study from a broad perspective that includes the hosts. We could also focus on samples such as those of e.g. \citet{Lonsdale2015}, for which the radio emission is compact, and the AGN and star formation are acting on similar spatial and time scales. We have also seen that it might be time to revisit and redefine the radio-loud/quiet classifications, as we have shown that the distribution of sources in terms of L$_{1.4 GHz}$/L$_{2-10 keV}$ displays a gradual transition between both regimes, rather than a dichotomy, showing that in many AGN there is a coexistence of two complicatedly interwoven regimes (kinetic and radiative), both with the potential to influence the host galaxy in different ways.
\subsection*{Acknowledgements}
We thank the anonymous referee for insightful and constructive comments that greatly improved the paper. We also thank the TOPCAT developer Mark Taylor for programming and releasing a software patch so that we could improve our Figures as the referee requested, and Dr B. Punsly, for allowing us to plot the data of \citet{Punsly2011} again in Figs. \ref{Punsly} to \ref{Punsly2_rej}. This work has made use of data/facilities and financial support from the ARCHES project (7th Framework of the European Union n$^{\circ}$ 313146). SM, FJC and AR acknowledge financial support by the Spanish Ministry of Economy and Competitiveness through grant AYA2012-31447, which is partly funded by the FEDER programme. FJC also acknowledges financial support through grant AYA2015-64346-C2-1-P (MINECO/FEDER). This work is based on observations obtained with \textit{XMM-Newton}, an ESA science mission with instruments and contributions directly funded by ESA Member States and NASA. It also makes use of data products from the Wide-field Infrared Survey Explorer, which is a joint project of the University of California, Los Angeles, and the Jet Propulsion Laboratory/California Institute of Technology, funded by the National Aeronautics and Space Administration. We acknowledge the use of the FIRST and NVSS catalogues, provided by the NRAO. Optical magnitudes and redshifts were obtained from the Sloan Digital Sky Survey Data Release 12. Funding for the Sloan Digital Sky Survey IV has been provided by the Alfred P. Sloan Foundation and the Participating Institutions. SDSS-IV acknowledges support and resources from the Center for High-Performance Computing at the University of Utah. The SDSS web site is \url{www.sdss.org}.
\bibliographystyle{mnras}
|
\section{Introduction}
\label{headings}
\label{gen_inst}
Support Vector Regression (SVR) has recently received much attention due to its competitive performance [1][2] compared with other regression approaches,
including the method of least squares [3], the XCSF algorithm [4], the logistic regression [5], and Neural Networks (NN) [6].
In this paper, we aim to develop a novel SVR approach by considering the recent progress in the support vector (SV) theory and addressing the limitations of existing SVR approaches.
In general, SVR constructs decision functions in high dimensional space for linear regression while the training data are mapped to a higher dimension in kernel Hilbert space.
$\varepsilon$-SVR [7][8] (the symbol $\varepsilon$ is replaced by e in the rest of this paper, and thus e-SVR is used) is the first popular SVR strategy and the other one is v-SVR [9]. e-SVR aims to find a function whose deviation is not more than e, thus forming the so-called e-tube, to fit all training data. In order to find the best fitting surface, e-SVR tries to maximize the minimum
margin containing data points in the e-tube as much as possible, which is similar to SVM. As for v-SVR, it adds a parameter v to the original e-SVR to control the number of support vectors and adjust
the parameter e automatically [10].
However, both e-SVR and v-SVR are susceptible to the data on the boundary (i.e. the support vectors). In fact, the optimization objective greatly depends on the margin between support vectors,
and this makes the final fitting function heavily rely on the distribution of the support vectors: if the distribution of the whole data within the e-tube is very different from the direction
of the support vectors, the final regression function may not be reliable.
Recent progress in the SV theory suggests that maximizing the minimum margin, that is, the shortest distance from the instances to the separating hyperplane, is not the only optimization goal for
achieving better learning performance. Different from traditional SVMs, Distance-weighted Discrimination (DWD)[11] maximizes the mean of the functional margin of all data [12],
and it uses the distances of all data to define the separating hyperplane,
thus greatly improving the performance. Meanwhile, Large Margin Distribution Machine (LDM) [13][14] is based on the novel theory of minimizing the margin distribution,
and it employed the dual coordinate descent (CD) and the averaged stochastic gradient descent (ASGD) strategies to solve the optimization function.
Considering the above, we propose a novel SVR approach called e-Distance Weighted Support Vector Regression (e-DWSVR).
e-DWSVR optimizes the minimum margin and the mean of functional margin at the same time, and it also uses both CD and ASGD strategies as in LDM to improve its scalability. A comparison of our e-DWSVR
with several other popular regression methods (including e-SVR, NN, linear regression and logistic) indicates that our e-DWSVR fits better the whole data distribution in most cases,
especially for those datasets with strong interference noise.
\section{Background on support vector theory}
Let $S = \left\{ {({x_1},{y_1}),({x_2},{y_2}),...,({x_n},{y_n})} \right\}$ be a training set of $n$ samples, where ${x_i} \in {R^m}$ are the input samples and ${y_i} \in R$ are
the corresponding target values. For regression, the objective function is $f(x) = w \cdot \phi (x) + b$, where $w \in {R^m}$ and $\phi$ is the mapping function induced by a kernel $K$,
i.e., $K({x_i},{x_j}) = \phi ({x_i}) \cdot \phi ({x_j})$, which projects the data to a higher dimensional space.
\subsection{The e-SVR process.}
According to [7][8], the objective function $f(x)$ is represented by the following constrained minimization problem:
\[\begin{array}{l}
\mathop {\min }\limits_{w,{\bf{\xi }},{{\bf{\xi }}^*}} \frac{1}{2}{\left\| w \right\|^2} + C\sum\limits_{i = 1}^n {({\xi _i} + \xi _i^*)} \\
s.t.{\rm{\ }}{y_i} - w \cdot \phi ({x_i}) - b \le \varepsilon + {\xi _i},{\rm{\ }}w \cdot \phi ({x_i}) + b - {y_i} \le \varepsilon + \xi _i^*{\rm{,\ \ }}{\xi _i},\xi _i^* \ge 0,{\rm{\ }}i = 1,2,...,n.
\end{array}\]
In the above, $C$ is a parameter that denotes the trade-off between the flatness of $f(x)$ and sums up to which deviations larger than e are tolerated. ${\bf{\xi }} = {[{\xi _1},{\xi _2},...,{\xi _n}]^T}$
and ${{\bf{\xi }}^{\bf{*}}} = {[\xi _1^*,\xi _2^*,...,\xi _n^*]^T}$ are regarded as the slack variables measuring the distances of the training samples lying outside the e-tube
from the tube itself. After applying the Lagrange multiplier, the minimization problem can be handled as the dual optimization problem:
\[\begin{array}{l}
\mathop {\max }\limits_{\alpha ,{\alpha ^*}} - \frac{1}{2}{\left( {\alpha - {\alpha ^*}} \right)^T}K\left( {{x_i},{x_j}} \right)\left( {\alpha - {\alpha ^*}} \right) - \varepsilon \sum\limits_{i = 1}^n {\left( {{\alpha _i} + \alpha _i^*} \right)} + \sum\limits_{i = 1}^n {{y_i}\left( {{\alpha _i} - \alpha _i^*} \right)} \\
s.t.{\rm{\ }}\sum\limits_{i = 1}^n {\left( {{\alpha _i} - \alpha _i^*} \right)} = 0,{\rm{\ \ }}0 \le {\alpha _i},\alpha _i^* \le C,{\rm{\ \ }}i = 1,2,...,n.
\end{array}\]
And the final objective function becomes $f(x) = \sum\limits_{i = 1}^n {({\alpha _i} - \alpha _i^*)} K({x_i},x) + b. $
\subsection{Recent progress in the SV theory}
SVM aims to maximize the minimum margins, which denotes the smallest distances of all instances to the separating
hyperplane [1][2]. The optimization problem is represented as follows:
\[\mathop {\min }\limits_{w,{\bf{\xi }}} \frac{1}{2}{\left\| w \right\|^2} + C\sum\limits_{i = 1}^n {{\xi _i}} {\rm{\ \ \ }}s.t.{\rm{\ }}{y_i}\left( {w \cdot \phi ({x_i}) + b} \right) \ge 1 - {\xi _i},{\rm{\ }}{\xi _i} \ge 0,{\rm{\ }}i = 1,2,...,n.\]
The DWD method [11] uses a new criterion for the optimization problem [12].
We denote the functional margin as ${u_i} = {y_i}(w \cdot \phi ({x_i}) + b)$ and
let ${r_i} = {y_i}\left( {w \cdot \phi ({x_i}) + b} \right) + {\xi _i}$ be the adjusted distance of
the $i$-th data to the separating hyperplane. So the solution of DWD is given below:
\[\mathop {\min }\limits_{w,b,{\bf{\xi }}} \sum\limits_{i = 1}^n {\left( {\frac{1}{{{r_i}}} + C{\xi _i}} \right)} {\rm{\ \ \ }}s.t.{\rm{\ }}{r_i} = {y_i}\left( {w \cdot \phi ({x_i}) + b} \right) + {\xi _i},{\rm{\ }}{r_i} \ge 0,{\rm{\ }}{\xi _i} \ge 0,{\rm{\ }}{\left\| w \right\|^2} \le 1,{\rm{\ }}i = 1,2,...,n.\]
Furthermore, the LDM method optimizes the margin distribution to solve the optimization problem
[13][14]. Thus the optimization problem becomes
\begin{equation}
\begin{array}{l}
\mathop {\min }\limits_{w,{\bf{\xi }}} \frac{1}{2}{\left\| w \right\|^2} + \frac{{2{\lambda _1}}}{{{n^2}}}(n{w^T}X{X^T}w - {w^T}Xy{y^T}{X^T}w) - \frac{{{\lambda _2}}}{n}w \cdot (Xy) + C\sum\limits_{i = 1}^n {{\xi _i}} \\
s.t.{\rm{\ }}{y_i}w \cdot \phi ({x_i}) \ge 1 - {\xi _i},{\rm{\ }}{\xi _i} \ge 0,{\rm{\ }}i = 1,2,...,n,
\end{array}
\end{equation}
where $X = \left[ {\phi ({x_1}),\phi ({x_2}),...,\phi ({x_n})} \right]$, $y = {\left[ {{y_1},{y_2},...,{y_n}} \right]^T}$,
$y = \{ - 1, + 1\}$ is the label set. LDM offers two strategies to solve Formula (1):
for small and medium datasets, Formula (1) can be solved by the CD method [13], and for large-scale problem, ASGD is used [13][15][16].
ASGD solves the optimization objective by computing a noisy unbiased estimate of the gradient and it randomly
samples a subset of the training instances rather than all the data. Our method uses a similar implementation
of the CD and ASGD methods, which will be introduced in detail in the next section.
\section{The proposed e-DWSVR}
In this section, we describe the novel e-DWSVR method, which applies the idea of the mean of functional margin and
employs a similar solution as in LDM, that is, we use the CD strategy to handle general conditions and adopt the
ASGD strategy to deal with large-scale problems.
\subsection{Formulation for e-DWSVR }
For regression problems, the input is $X = \left[ {\phi ({x_1}),\phi ({x_2}),...,\phi ({x_n})} \right]$ and the
target is $y = {\left[ {{y_1},{y_2},...,{y_n}} \right]^T}$, where $y \in \{ - \infty , + \infty \}$. In order to
simplify the complexity brought by the bias term, we enlarge the dimension of the vectors $w$ and $\phi ({x_i})$
to handle the bias term [17], i.e., $w \leftarrow {[w,b]^T}$, $\phi ({x_i}) \leftarrow [\phi ({x_i}),{\bf{1}}]$.
Thus the form of the regression function becomes $f(x) = w \cdot \phi (x)$. Then the margin in the regression analysis
will be the distance of a point to the fitting hyperplane, i.e. $\left| {w \cdot \phi ({x_i}) - {y_i}} \right|/\left\| w \right\|.$
Based on the concept of margin, we give the definition of functional margin.
\textbf{Definition 1:} functional margin in regression: $\gamma = {\left( {w \cdot \phi ({x_i}) - {y_i}} \right)^2},{\rm{\ \ }}i = 1,...,n.$
The functional margin can describe the difference between the real values and the estimated value. It also has a significant
connection with the geometrical distance. In fact, the functional margin is the adjusted distance of the data to the
separating hyperplane. If the value of $w$ is determined, the ranking of all data points to the fitting surface with
respect to the margin can be decided by functional margin. Then the mean of the functional margin in regression is defined by Definition 2 as follows.
\textbf{Definition 2:} mean of the functional margin in regression: $\bar \gamma = \frac{1}{n}\sum\limits_{i = 1}^n {{{\left( {w \cdot \phi ({x_i}) - {y_i}} \right)}^2}}.$
Based on the above definitions, we will not only maximize the minimum margin, but also minimize the mean of the
functional margin at the same time, in order to obtain a better tradeoff between the distribution of the whole data
and the distribution of support vectors. e-DWSVR considers the influence of all data to the fitting surface, as this
is closer to the actual distribution of the internal data.
To illustrate the difference of the optimal objectives between e-SVR and e-DWSVR, we use an example for comparison among linear regression, e-SVR and e-DWSVR on an artificial dataset. The distribution of the whole data is shown in Figure 1(a), where those sparse quadrate points are the noise and the red solid line with points is drawn by linear regression. Figure 1(b) is the enlarged version of the dotted box in Figure 1(a) for better observing the differences among the three methods. In Figure 1(b) the points show the distribution of most non-noisy data and the black dotted lines represent the data on the boundary. The red solid line with points is part of the line in Figure 1(a). The blue dashed curve is drawn by e-SVR and the red solid curve is drawn by e-DWSVR.
Obviously, the curve drawn by linear regression is far from the dataset, which indicates that the linear regression is more sensitive to noisy points, and this makes the curve in a wrong direction. e-SVR and e-DWSVR are not affected easily by noise points, so the blue dashed curve and the red solid curve are within area of non-noisy data. However, the regression model implemented by e-SVR is controlled by the support vectors (those boundary points). Once the internal data distribution is different from the edge points (which is the case in Figure 1(b)), the regression model will not achieve good performance. e-SVR produces the blue dashed curve which is different from the red solid curve. Because e-DWSVR considers the influence of all data to the fitting surface, it is obvious the red solid curve drawn by e-DWSVR is closer to the actual distribution of the internal data.
\begin{figure}[H]
\setlength{\abovecaptionskip}{0cm}
\setlength{\belowcaptionskip}{-0.4cm}
\centering
\setlength{\fboxrule}{0pt}
\fbox{\rule[0cm]{0cm}{0cm}
\includegraphics[height=5cm,width=12cm]{diagram.png}}\\
\quad\quad\quad(a) The whole data with noise\quad\quad\quad (b) The distribution of data without noise\\
\quad
\caption{The fitting curves drawn by linear regression, e-SVR, and e-DWSVR. In Figure 1(b), 82.6\% of data points are evenly distributed across the line A, 16.5\% are evenly distributed on line B, where they are in parallel and have the different slope with the red solid curve. The other 0.8\% of data are outliers in Figure 1(a). Due to the outliers, the line produced by linear regression will be different from the correct curve (line A). The blue dashed curve produced by SVR will be very close to line B. While the red solid curve produced by eDWSVR is closer to line B, thus a better result can be produced.}
\end{figure}
As in the soft-margin e-SVR [1], we also consider complicated conditions. So, the final optimization function has the following form:
\[\begin{array}{l}
\mathop {\min }\limits_{w,\xi ,{\xi ^*}} \frac{1}{2}{\left\| w \right\|^2} + {\lambda _1}\bar \gamma + C\sum\limits_{i = 1}^n {\left( {{\xi _i} + \xi _i^*} \right)} \\
s.t.{\rm{\ }}{y_i} - w \cdot \phi ({x_i}) \le \varepsilon + {\xi _i}{\rm{,\ \ }}w \cdot \phi ({x_i}) + {y_i} \le \varepsilon + \xi _i^*{\rm{,\ \ }}{\xi _i},\xi _i^* \ge 0,{\rm{\ \ }}i = 1,2,...,n.
\end{array}\]
\subsection{e-DWSVR for medium-scale regression with kernel.}
Considering the mean of the functional margin in the constrained minimization problem, we can obtain the following form:
\begin{equation}
\begin{array}{l}
\mathop {\min }\limits_{w,{\bf{\xi }},{{\bf{\xi }}^*}} \frac{1}{2}{\left\| w \right\|^2} + \frac{{{\lambda _1}}}{n}\left( {{w^T}X{X^T}w - 2{{\left( {Xy} \right)}^T}w} \right) + C\sum\limits_{i = 1}^n {\left( {{\xi _i} + \xi _i^*} \right)} \\
s.t.{\rm{\ }}{y_i} - w \cdot \phi ({x_i}) \le \varepsilon + {\xi _i},{\rm{\ \ }}w \cdot \phi ({x_i}) + {y_i} \le \varepsilon + \xi _i^*,{\rm{\ \ }}{\xi _i},\xi _i^* \ge 0,{\rm{\ \ }}i = 1,2,...,n.
\end{array}
\end{equation}
Here we omit the $y{y^T}$ term in $\bar \gamma$ because it is regarded as a constant in an optimization problem.
Obviously, the high dimensionality of $\phi$ and its complicated form makes Formula (2) more intractable.
To simplify this formula, we take the suggestion from [18] and the optimal solution $w$ in
LDM [13]. We first give the following theorem.
\textbf{Theorem 1.} The optimal solution $w$ for Formula (2) can be represented as the following form:
where $\alpha = {\left[ {{\alpha _1},{\alpha _2},...,{\alpha _n}} \right]^T}$ and
${\alpha ^{\rm{*}}} = {\left[ {{\alpha _1}^{\rm{*}},{\alpha _2}^{\rm{*}},...,{\alpha _n}^{\rm{*}}} \right]^T}$ are the parameters of e-DWSVR.
\textbf{PROOF.} We assume that $w$ can be divided into the span of $\phi ({x_i})$ and an orthogonal vector, i.e.,
\[w = \sum\limits_{i = 1}^n {\left( {{\alpha _i} - \alpha _i^*} \right)} \cdot \phi ({x_i}) + u = X\left( {\alpha - {\alpha ^*}} \right) + u, \]
where $u$ satisfies $\phi ({x_k}) \cdot u = 0$ for all $u$, that is, $X \cdot u = 0.$ Then we obtain the following formula:
Hence, the second term of Formula (2) is independent of $u$ considering Formula (4). Furthermore, note that the
constraints of Formula (2) are also independent of $u$, so the last term of Formula (2) can be considered as
independent of $u$. To simplify the first term, we have
In the above "$\ge $" becomes "${\rm{ = }}$" if and only if $u = 0$. Thus, setting $u = 0$ strictly reduces the first
term of Formula (2) without affecting the rest terms. Based on all above, $w$ in Formula (2) can be represented
in the form of Formula (3). \textbf{Q.E.D.}
According to Theorem 1, Formula (2) can be cast as
where $Q = 2{\lambda _1}{G^T}G/n + G$, $p = - 2{\lambda _1}Gy/n$ and $G = {X^T}X.$ The minimization problem which
has been simplified can be transformed into a dual formulation with the Lagrange multipliers, so the Lagrange
function of Formula (5) leads to
where $\eta ,{\eta ^*},\beta ,{\beta ^*}$ are the Lagrange multipliers. To satisfy the KKT condition,
we set the partial derivatives $(\alpha ,{\alpha ^*},\xi ,{\xi ^*})$ to zero and obtain the following formulas:
By substituting Formulas (7) and (8) into Formula (6) and inspired by reference [19], Formula (5)
can be written as follows to compute the values of $\left[ {\beta ,{\beta ^*}} \right]$ separately.
where $H = G{Q^{ - 1}}G$, and ${Q^{ - 1}}$ stands for the inverse matrix of $Q$ and $e$ means the all-one vector.
So Formula (9) can be solved by the CD method as in [20][21]. This method continuously
selects one variable for minimization and keeps others as constants. In our situation,
we minimize $\beta _k^{'} \in {\beta ^{'}}$ by fixing other $\beta _{l \ne k}^{'}$,
where ${\beta ^{'}} = \left[ {\beta ,{\beta ^*}} \right]$, and $\mathop {\min }\limits_t f({\beta ^{'}} + t{e_k}){\rm{\ \ }}s.t.{\rm{\ }}0 \le {\beta _k}^{'} + t \le C$
needs to be solved, where ${e_k} = {[\underbrace {0,...,0}_{k - 1},1,\underbrace {0,...,0}_{2n - k}]^T}$, and the
form of this sub-problem is
where ${h_{kk}}$ is the diagonal entry of $\left[ \begin{array}{l}
H{\rm{\ \ }} - H\\
- H{\rm{\ \ \ }}H
\end{array} \right]$, and ${\left[ {\nabla f({\beta ^{'}})} \right]_k}$ is the $k$-th element of the gradient $\nabla f({\beta ^{'}})$. Obviously $f({\beta ^{'}})$ is independent of $t$, so we omit this part of Formula (10). Hence $f({\beta ^{'}} + t{e_k})$ is transformed into a simple quadratic function.
After calculating ${\beta ^{'}} = \left[ {\beta ,{\beta ^*}} \right]$, we can obtain $\left( {\alpha - {\alpha ^*}} \right)$ according to Formula (7) as $\left( {{\alpha} - \alpha^*} \right) = {Q^{ - 1}}\left( {G\left( {\beta - {\beta ^*}} \right) - p} \right) = {Q^{ - 1}}G\left( {\frac{{{\lambda _2}}}{n}Y + \left( {\beta - {\beta ^*}} \right)} \right).$
Therefore, we obtain the final fitting function as $f(x) = \sum\limits_{i = 1}^n {({\alpha _i} - \alpha _i^*)} K({x_i},x).$
Algorithm 1 presents the detailed steps of the CD method used for efficiently updating ${\beta ^{'}} = \left[ {\beta ,{\beta ^*}} \right]$.
\begin{algorithm}
\footnotesize
\caption{\footnotesize Kernel e-DWSVR}
\textbf{Input:} Dataset $X$, $\lambda_1$, $C$, $\varepsilon$\\
\textbf{Output:} $\alpha - {\alpha ^*}$\\
\textbf{Initialization:} ${\beta ^{'}} = {\bf{0}},{\rm{\ }}\left( {\alpha - {\alpha ^*}} \right) = \frac{{2{\lambda _1}}}{n}{Q^{ - 1}}Gy,{\rm{\ }}A = {Q^{ - 1}}G,{\rm{\ }}{h_{kk}} = e_k^TG{Q^{ - 1}}G{e_k};$
\begin{algorithmic}[1]
\For { $inter = 1,2,...,maxInter$ }
\For { $k = 1,2,...,2n$ }
\State ${\left[ {\nabla f({\beta ^{'}})} \right]_k} \leftarrow \varepsilon + \left( {G\left( {\alpha - {\alpha ^*}} \right) - {y_k}} \right);$ \quad\quad\quad if\quad $k = 1,2,...,n$
\State ${\left[ {\nabla f({\beta ^{'}})} \right]_k} \leftarrow \varepsilon - \left( {G\left( {\alpha - {\alpha ^*}} \right) - {y_{k - n}}} \right);$ \quad\quad if\quad $k = n+1,n+2,...,2n$
\State $\beta _k^{'old} \leftarrow \beta _k^{'};$
\State $\beta _k^{'} \leftarrow \min (\max (\beta _k^{'} - \frac{{{{\left[ {\nabla f({\beta ^{'}})} \right]}_k}}}{{{h_{kk}}}},0),C);$
\For { $i = 1,2,...,n$ }
\State $\left( {{\alpha _i} - \alpha _i^*} \right) \leftarrow \left( {{\alpha _i} - \alpha _i^*} \right) + \left( {\beta _k^{'} - \beta _k^{'old}} \right)A{e_k};$\quad\quad if\quad $k = 1,2,...,n$
\State $\left( {{\alpha _i} - \alpha _i^*} \right) \leftarrow \left( {{\alpha _i} - \alpha _i^*} \right) - \left( {\beta _k^{'} - \beta _k^{'old}} \right)A{e_k};$ \quad\quad if\quad $k = n+1,n+2,...,2n$
\EndFor
\EndFor
\If {${\beta ^{'}}$ converge} break; \EndIf
\EndFor
\end{algorithmic}
\end{algorithm}
\subsection{e-DWSVR for large-scale regression.}
Although the CD method can solve e-DWSVR efficiently, it is not a best strategy for dealing with large-scale problems.
To further improve the scalability of e-DWSVR, we also apply ASGD method to e-DWSVR.
We reformulate Formula (2) as follows to solve large-scale problems,
\begin{align*}
\begin{array}{l}
\mathop {\min }\limits_w g(w) = \frac{1}{2}{\left\| w \right\|^2} + \frac{{{\lambda _1}}}{n}\left( {{w^T}{X^T}Xw - 2{{\left( {Xy} \right)}^T}w} \right)\\
+ C\sum\limits_{i = 1}^n {\max \left\{ {0,{y_i} - w \cdot {x_i} - \varepsilon ,w \cdot {x_i} - {y_i} - \varepsilon } \right\}}.\tag{11}
\end{array}
\end{align*}
Computing the gradient of $w$ in Formula (11) is time consuming because we need all the training instances for
computation. Considering this issue, we use Stochastic Gradient Descent (SGD) [22] to compute a noisy
unbiased estimation of the gradient, and this is done by randomly sampling part of the training instances.
Therefore, we give an unbiased estimation of the gradient $\nabla g(w)$ in our case. For denoting the last term of (11)
formally, we define a function $s(w)$ that has different values under different constraint conditions, as shown below:
\[s(w) = \left\{ \begin{array}{l}
- {x_i},{\rm{\quad\quad\ }}i \in {I_1}\\
{x_i},{\rm{\quad\quad\quad}}i \in {I_2}\\
0,{\rm{\quad\quad\quad}}otherwise
\end{array} \right.{\rm{\ \ }},i = 1,2,...,n,\]
where ${I_1} \equiv \left\{ {i\left| {{y_i} - w \cdot {x_i} \le \varepsilon } \right.} \right\}{\rm{,\ }}{I_2} \equiv \left\{ {i\left| {w \cdot {x_i} - {y_i} \le \varepsilon } \right.} \right\}$.
\textbf{Theorem 2.} An unbiased estimation of the gradient $\nabla g(w)$ has the following form:
\[\nabla g(w,{x_i}) = 2{\lambda _1}{x_i}x_i^Tw + w - 2{\lambda _1}{y_i}{x_i} + nC \cdot s(w),\]
where one instance $({x_i},{y_i})$ is sampled randomly from the training set.
\textbf{PROOF:} The gradient of $g(w)$ has the form $\nabla g(w) = Qw + p + C\sum\limits_{i = 1}^n {s(w)}$, where $Q = 2{\lambda _1}{X^T}X/n + I$ and ${\rm{ }}p = - 2{\lambda _1}Xy/n$. Note that
Considering the linearity of expectation, and with Formula (12), we have
This leads to a conclusion that $\nabla g(w,{x_i})$ is a noisy unbiased estimation of the gradient. \textbf{Q.E.D.}
Then, the stochastic gradient can be updated iteratively with the following form:
where ${\varphi _t}$ is the learning rate at the $t$-th iteration.
The good choice for ${\varphi _t}$ can be obtained by the form ${\varphi _t} = {\varphi _0}{(1 + a{\varphi _0}t)^{ - c}}$ by ASGD,
where $a$, ${\varphi _0}$, and $c$ are constants [23]. And we compute ${\bar w_t}$ at each iteration in
addition to updating the ordinary stochastic gradient in Formula (13) as
${\bar w_t}{\rm{ = }}\frac{1}{{t - {t_0}}}\sum\limits_{i = {t_0} + 1}^t {{w_t}}$, where ${t_0}$ decides when to perform
the averaging process. This average can also be calculated in a recursive manner as ${\bar w_{t + 1}} = {\delta _t}{w_{t + 1}} + (1 - {\delta _t}){\bar w_t}$,
where ${\delta _t}$ is the rate of averaging and ${\delta _t} = 1/\max \{ 1,t - {t_0}\}.$
Finally, Algorithm 2 provides the detailed steps of ASGD for large-scale problems, where $T*n$ determines the number of iteration. $T$ means the adjustment of the whole times of iteration which is set as same as [13]; $u$ means $\nabla g(w,{x_i})$, and its initial value is 0.
\begin{algorithm}[H]
\footnotesize
\caption{\footnotesize Large-scale Kernel e-DWSVR}
\textbf{Input:} Dataset $X$, $\lambda_1$, $\varepsilon $\\
\textbf{Output:} $\bar w$\\
\textbf{Initialization:} $u = 0,t = 1$
\begin{algorithmic}[1]
\While {$t \le T*n$}
\State Randomly select one instance $({x_i},{y_i})$ from the training set;
\State Compute $\nabla g(w,{x_i})$;
\State ${w_{t + 1}} \leftarrow {\varphi _t}\nabla g(w,{x_i}) + {w_t}$;
\State ${\bar w_{t + 1}} \leftarrow {\delta _t}{w_{t + 1}} + (1 - {\delta _t}){\bar w_t}$;
\State $t + + $;
\EndWhile
\end{algorithmic}
\end{algorithm}
\section{Experiments}
In this section, we report the effectiveness of e-DWSVR in comparison with other regression methods
to assess whether our method has better fitting quality.
\subsection{Experimental setup}
We select sixteen datasets from UCI [24] to evaluate these methods. This includes eight
medium-scale datasets and eight large-scale datasets. The numbers of samples and features for each dataset
are indicated in Table 1 and 2. For instance, Slump (103 / 7) means the sample size is 103 and the number of
features is 7 for the Slump dataset. All features of datasets and target set are normalized into [0,1] for
balancing the influence of each feature. After normalization, we take a preprocess of the PCA with the
contribution to 95\% for feature extraction. During the construction of the regression model, we divide
the datasets into training sets and testing sets by 5-fold cross validation, and the experiments are
repeated for 30 times. For medium-scale datasets, we evaluate both the linear and RBF
kernels [1], and the average values and standard deviation are recorded. For large-scale
datasets, only the linear kernel is evaluated. In addition, we record the time cost.
e-SVR, NN [6], linear regression and logistic are compared with our method. For e-SVR and e-DWSVR,
parameters $C$ and $e$ are both needed when building models. So these parameters are optimized during
the experiments to select the best parameters for regression. For e-DWSVR, parameter $\lambda_1$ is fixed
to 1; parameter $T$, ${t_0}$, and ${\varphi _t}$ have the same values as in [13]. All optimization are processed on testing sets.
The mean square error (MSE) [25] and R-square (R2) [26] are commonly used evaluation metrics in regression, and thus they are chosen in our research for evaluation. All experiments are tested with MATLAB on a PC, which has a 2.50 GHz CPU and 8 GB memory.
\subsection{Result and discussion}
For medium-scale datasets, Figure 1 shows the results of MSE on all methods,
including linear kernel function and RBF kernel function for e-DWSVR and e-SVR. Table 1 summarizes the results of R2 on all methods.
In most datasets, e-DWSVR performs better than other methods. Obviously, the fitting quality of e-DWSVR is much better than e-SVR, which
means e-DWSVR is more competitive than e-SVR. In Table 2, the best R2 on each dataset is indicated in bold.
\begin{figure}[H]
\setlength{\abovecaptionskip}{0cm}
\setlength{\belowcaptionskip}{0cm}
\centering
\setlength{\fboxrule}{0pt}
\fbox{\rule[0cm]{0cm}{0cm}
\includegraphics[height=7.5cm,width=12cm]{mse.png}}\\
\caption{The evaluation of MSE on the medium-scale datasets. It can be seen that e-DWSVR and e-SVR present the advanced results for RBF kernel. Also, e-DWSVR achieves the smallest MSE and largest R2 in most datasets,
which means that e-DWSVR can provide better fitting quality. In addition, e-DWSVR performs significantly better than e-SVR for both linear kernel and RBF kernel. Values of x axis (from left to right) are comparison methods:
1. e-DWSVR-RBF (red), 2. e-SVR-RBF (blue), 3. e-DWSVR-Linear (green), 4. e-SVR-Linear (purple), 5. linear regression (cyan), 6. NN (yellow), 7. logistic (black).}
\end{figure}
\begin{table}[H]
\setlength{\abovecaptionskip}{0cm}
\setlength{\belowcaptionskip}{0cm}
\footnotesize
\caption{The evaluation of R2 on the medium-scale datasets.}
\label{sample-table}
\centering
\begin{tabular}{llllllll}
\cmidrule{1-8}
\tabincell{l}{Dataset\\(Samples / Features)} &\tabincell{l}{eDWSVR\\(RBF)} &\tabincell{l}{eSVR\\(RBF)} &\tabincell{l}{eDWSVR\\(Linear)} &\tabincell{l}{eSVR\\(Linear)} &LINEAR &NN &Logistic\\
\midrule
Slump (103 / 7) &\textbf{0.9539} &0.9224 &0.8739 &0.8641 &0.8067 &0.8220 &0.8336\\
Automobile (205 / 26) &\textbf{0.8865} &0.8320 &0.7847 &0.7689 &0.7748 &0.7526 &0.7592\\
Yacht (308 / 7) &\textbf{0.8684} &0.7526 &0.7397 &0.7083 &0.6574 &0.6990 &0.7007\\
Auto MPG (398 / 8) &\textbf{0.6900} &0.6874 &0.6839 &0.6837 &0.6843 &0.6782 &0.6673\\
Housing (506 / 14) &0.6955 &0.6634 &0.6851 &0.6522 &\textbf{0.7048} &0.6091 &0.6326\\
Stock (536 / 9) &\textbf{0.5687} &0.5566 &0.5344 &0.5241 &0.4055 &0.3719 &0.4146\\
Concrete (1030 / 9) &\textbf{0.5430} &0.4590 &0.4300 &0.4289 &0.3810 &0.3583 &0.3710\\
Music (1059 / 68) &\textbf{0.3865} &0.3628 &0.3084 &0.3074 &0.3331 &0.2485 &0.1472\\
\bottomrule
\end{tabular}
\end{table}
For large-scale datasets, Figure 2 and Table 2 summarize the results for the linear kernel. As one can see,
e-DWSVR performs better than other regression methods on some datasets, and it always performs better than
e-SVR. In addition, linear regression did not return the results on some datasets after 48 hours
(indicated as NA in Table 2).
\begin{figure}[H]
\setlength{\abovecaptionskip}{0cm}
\setlength{\belowcaptionskip}{-0.5cm}
\centering
\setlength{\fboxrule}{0pt}
\fbox{\rule[0cm]{0cm}{0cm} \rule[0cm]{0cm}{0cm}
\includegraphics[height=7.5cm,width=12cm]{large-mse.png}}\
\caption{The evaluation of MSE on the large-scale datasets. Values of x axis (from left to right) are comparison methods: 1. e-DWSVR (red), 2. e-SVR (blue), 3. linear regression (green), 4. NN (black), 5. logistic (purple).}
\end{figure}
\begin{table}[H]
\footnotesize
\setlength{\abovecaptionskip}{0cm}
\setlength{\belowcaptionskip}{0cm}
\caption{The evaluation of R2 on the large-scale datasets.}
\centering
\begin{tabular}{llllll}
\cmidrule{1-6}
\tabincell{l}{Dataset\\(Samples / Features)}&e-DWSVR &e-SVR &LINEAR &NN &Logistic\\
\midrule
Crime (1994 / 128) &\textbf{0.5613} &0.5348 &0.4430 &0.5244 &0.5350\\
SkillCraft (3338 / 18) &\textbf{0.7275} &0.7018 &0.7013 &0.6843 &0.6147\\
CCPP (9568 / 4) &\textbf{0.8634} &0.8327 &0.8246 &0.7334 &0.8286\\
Drift (13910 / 129) &0.5620 &0.5636 &\textbf{0.5888} &0.5490 &0.5574\\
Bike sharing (17389 / 16) &0.5867 &0.5825 &\textbf{0.5938} &0.5395 &0.5465\\
ONP (39797 / 61) &\textbf{0.4598} &0.4590 &0.3856 &0.3544 &0.2113\\
CASP (45730 / 9) &\textbf{0.2755} &0.2608 &NA &0.1836 &0.1657\\
Buzz (140000 / 77) &\textbf{0.2976} &0.2591 &NA &0.1529 &0.1373\\
\bottomrule
\end{tabular}
\end{table}
\subsection{Time cost and parameter effect}
In Figure 3 we present a comparison of CPU time taken between e-SVR and e-DWSVR on each dataset.
e-SVR for large-scale problems was implemented by the LIBLINEAR [17] package and e-DWSVR was implemented by ASGD. According to Figure 3, e-DWSVR cost less time than e-DWSVR on most datasets,
and it is only slightly slower than e-SVR on two datasets.
Furthermore, Figure 4 shows that the MSE on the medium-scale datasets does not have much difference with the change of the parameters.
This indicates that the performance of e-DWSVR is not sensitive to the parameters, which shows the robustness of e-DWSVR.
\begin{figure}[H]
\setlength{\abovecaptionskip}{-0.2cm}
\setlength{\belowcaptionskip}{-0.7cm}
\centering
\begin{minipage}{.48\linewidth}
\setlength{\fboxrule}{0pt}
\fbox{\rule[0cm]{0cm}{0cm} \rule[0cm]{0cm}{0cm}
\includegraphics[height=2.5cm,width=5.5cm]{time-cost.png}}\\
\caption{CPU time on the large-scale datasets.}
\end{minipage}
\begin{minipage}{.48\linewidth}
\setlength{\fboxrule}{0pt}
\fbox{\rule[0cm]{0cm}{0cm} \rule[0cm]{0cm}{0cm}
\includegraphics[height=2.5cm,width=6.5cm]{parameter3.png}}\\
\caption{Parameter effect on eDWSVR.}
\end{minipage}
\end{figure}
\section{Conclusion and future work}
In this paper, we propose e-DWSVR, a novel and promising SVR algorithm. e-DWSVR overcomes the limitations of existing SVR systems and outperforms other regression methods on several benchmark datasets.
We envision great application potential of e-DWSVR in various problems, including feature extraction, anomaly detection, and complex data interpretation. In the near future, we will apply e-DWSVR to solve
several real-world problems, including bioinformatics anaysis and financial data prediction and anomaly detection.
\section{Acknowledgement}
We gratefully thank Dr Teng Zhang and Prof Zhi-Hua Zhou for providing the source code of “LDM” source code and their kind technical assistance. We also thank Prof Chih-Jen Lins team for providing the Libsvm and Liblinear packages and their support. This work is supported by the National Natural Science Foundation of China (Nos. 61472159, 61572227) and Development Project of Jilin Province of China (Nos.20140101180JC, 20160204022GX). This work is also partially supported by the 2015 Scottish Crucible Award funded by the Royal Society of Edinburgh and the 2016 PECE bursary provided by the Scottish Informatics \& Computer Science Alliance (SICSA).
\section*{References}
[1] V. Vapnik C. Cortes. Support vector machine. Machine Learning, 7(2002):1–28, 1995.
[2] J. D. Brown, M. F. Summers, and B. A. Johnson. Prediction of hydrogen and carbon chemical shifts from
rna using database mining and support vector regression. Journal of Biomolecular Nmr, 63(1):1–14, 2015.
[3] C. Y. Deng and H. W. Lin. Progressive and iterative approximation for least squares b-spline curve and
surface fitting. Computer-Aided Design, 47(1):32–44, 2014.
[4] M. V. Butz, G. K. M. Pedersen, and P. O. Stalph. Learning sensorimotor control structures with xcsf:
redundancy exploitation and dynamic control. In Proceedings of the 11th Annual Conference on Genetic
and Evolutionary Computation, pages 1171–1178, 2009.
[5] D. Liu, T.R. Li, and D. C. Liang. Incorporating logistic regression to decision-theoretic rough sets for
classifications. International Journal of Approximate Reasoning, 54(1):197–210, 2014.
[6] K. Hagiwara, T. Hayasaka, N. Toda, S. Usui, and K. Kuno. Upper bound of the expected training error of
neural network regression for a gaussian noise sequence. Neural Networks, 14(10):1419–1429, 2001.
[7] A. J. Smola and B. Scholkopf. A tutorial on support vector regression. Statistics and Computing,
14(3):199–222, 2004.
[8] B. Demir and L. Bruzzone. A multiple criteria active learning method for support vector regression. Pattern
Recognition, 47(7):2558–2567, 2014.
[9] B Scholkopf, A. J. Smola, R. C. Williamson, and P. L. Bartlett. New support vector algorithms. Neural
Computation, 12(5):1207–1245, 2000.
[10] B. Gu, V. S. Sheng, Z. J. Wang, D. Ho, S. Osman, and S. Li. Incremental learning for v-support vector
regression. Neural Networks the Official Journal of the International Neural Network Society, 67(C):140–
150, 2015.
[11] J. S. Marron. Distance-weighted discrimination. Journal of the American Statistical Association,
102(12):1267–1271, 2007.
[12] X. Y. Qiao and L. S. Zhang. Distance-weighted support vector machine. Statistics and Its Interface,
8(3):331–345, 2013.
[13] T. Zhang and Z. H. Zhou. Large margin distribution machine. In Proceedings of the 20th ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining, pages 313–322, 2014.
[14] Z. H. Zhou. Large margin distribution learning. In Proceedings of the 6th Artificial Neural Networks in
Pattern Recognition, pages 1–11, 2014.
[15] B. T. Polyak and A. B. Juditsky. Acceleration of stochastic approximation by averaging. Siam Journal on
Control and Optimization, 30(4):838–855, 1992.
[16] T. Zhang. Solving large scale linear prediction problems using stochastic gradient descent algorithms. In
Proceedings of the 21st International Conference on Machine Learning, pages 919–926, 2004.
[17] R. E. Fan, K. W. Chang, C. J. Hsieh, X. R. Wang, and C. J. Lin. Liblinear: A library for large linear
classification. Journal of Machine Learning Research, 9(12):1871–1874, 2010.
[18] B. Scholkopf and A. Smola. Learning with kernels : support vector machines, regularization, optimization,
and beyond. MIT Press, 2002.
[19] C. C. Chang and C. J. Lin. Libsvm: A library for support vector machines. Acm Transactions on Intelligent
Systems and Technology, 2(3):389–396, 2011.
[20] G. X. Yuan, C. H. Ho, and C. J. Lin. Recent advances of large-scale linear classification. Proceedings of
the IEEE, 100(9):2584–2603, 2012.
[21] Z. Q. Luo and P. Tseng. On the convergence of the coordinate descent method for convex differentiable
minimization. Journal of Optimization Theory and Applications, 72(1):7–35, 1991.
[22] L. Bottou. Large-Scale Machine Learning with Stochastic Gradient Descent. Physica-Verlag HD, 2010.
[23] W. Xu. Towards optimal one pass large scale learning with averaged stochastic gradient descent. Computer
Science, 2011.
[24] M. Lichman. UCI machine learning repository. http://archive.ics.uci.edu/ml, 2013.
[25] D. N. Guo, S. Shamai, and S. Verdu. Mutual information and minimum mean-square error in gaussian
channels. IEEE Transactions on Information Theory, 51(4):1261–1282, 2005.
[26] W. Wen, R. Luo, X. J. Tang, L. Tang, H. X. Huang, X. Y. Wen, S. Hu, and B. Peng. Age-related
progression of arterial stiffness and its elevated positive association with blood pressure in healthy people.
Atherosclerosis, 238(1):147–152, 2015.
\end{document}
|
\section{Introduction}
High-velocity clouds (HVCs) provide a unique window into the coolest component of the circumgalactic medium and the processes of Galactic inflow and outflow. HVCs, and the complexes into which they are arranged, are found by their emission in H{\sc i}~ or absorption in numerous metal lines, and have radial velocities inconsistent with Galactic rotation \citep{Wakker_1997}. The precise origin of most HVCs is unknown, and some mix of Galactic fountain \citep[e.g.][]{Bregman_1980}, multiphase accretion \citep[e.g.][]{Fern_ndez_2012}, and gas stripping from satellites is typically invoked \citep{Putman_2012}. The exception is the Magellanic stream, which was stripped from the large and small Magellenic clouds, and which we will exclude from our discussion in this work. HVCs with negative radial velocities, which are metal enriched in the range of 10\% to 30\% of the solar metallicity, are likely a tracer of the process by which material accretes onto the Galaxy, though the total rate of this accretion is very uncertain. Less explored are the HVCs with positive radial velocities, most of which are in the inner two quadrants of the Galactic sky. These include the Wannier complexes WA, WB, WD, WE, and the Smith Cloud \citep{1991A&A...250..509W}. The Smith cloud has received significant attention of late, for its strongly cometary appearance which provides enough information to infer past trajectories, and make some inference as to its origin \citep{Lockman_2008, Fox_2015}.
Complex WD is the largest-area positive velocity HVC Complex, covering 310 square degrees with a total H{\sc i}~ flux of 1.2 $\times 10^7$ K km s$^{-1}$ arcmin$^2$, and a maximum H{\sc i}~ column density of $\sim 1.2 \times 10^{20}$ cm$^{-2}$. It is by far the largest complex that exists in the inner two Galactic quadrants, where a small fraction of HVC flux is detected. With a range of velocities between +90 and +130 km s$^{-1}$, it is consistent with cylindrical rotation on the far side of the inner Galaxy, 20 kpc from the sun with a mass of 6 $\times 10^7 M_\odot$. This would make it very similar in mass, Galactocentric radius, and height to Complex C, the largest area and brightest HVC complex \citep{Thom_2008}.
One major issue in gaining a better physical understanding of these enigmatic clouds is their unknown distance. Since there are no objects of fixed luminosity in HVCs, there are effectively no intrinsic distance measures. H{\sc i}~ emission or optical and UV absorption lines toward extragalactic background sources only provide distance-independent column densities. HVC distances not only give us a masses for these structures, but also a context; the spatial relationship between the cloud and the nearby spatial and kinematic structure of the disk gives us insight as to its origin.
There are a number of indirect methods for measuring the distance to an HVC complex, including H$\alpha$ emission and kinematic structure \citep{Putman_2003, Peek_2007}, but the only proven direct distance measure is stellar absorption. By observing stars with measured distances at medium or high spectral resolution, one can look for absorption lines in Na {\sc i}, Ca{\sc ii}~ H \& K, Ti {\sc ii}, and numerous ultraviolet absorption lines at the velocity of H{\sc i}~ emission from HVCs \cite{1995A&A...302..364S}. By finding detections and non-detections of these absorption lines along lines of sight toward H{\sc i}~ emitting HVCs, distances can be robustly measured. A number of clouds have well-measured distances using this method, but complex WD is not among them \citep{Wakker_2001, Wakker_2007, Thom_2008, Wakker_2008}.
In this work we report the first distance upper limit on Complex WD using medium resolution absorption line spectroscopy toward a blue horizontal branch star. We extend the methods of \citet{Sirko_2004} to find the spectral type of the star, and thus put a precise distance limit. We use this to make some inferences as to the possible origin of Complex WD, and how it fits into the structure of Galactic HVCs as a whole.
\section{Data}
\subsection{New Observations}
The observations of our target, USNO-A0600-15865535, were obtained at the ESO Very Large Telescope (VLT) at Cerro Paranal, Chile on the nights of May 6th and May 7th, 2016. The target was observed as a part of our program ``Mapping the Cool Circumgalactic Medium with Calcium II'' (097.A-0552, PI: Peek) that uses the FLAMES/GIRAFFE spectrograph, a fiber-fed multiobject spectrograph mounted on the Nasmyth focus of UT2 \citep{Pasquini_2003}. USNO-A0600-15865535 is a bright ($g= 14.2$) very blue ($g-r = -0.21$) source, unresolved in Pan-STARRS imaging \citep{Magnier_2013, Schlafly_2012,Tonry_2012}. Through a combination of overoptimism and clerical error, it was targeted as a quasar candidate behind the circumgalactic medium of M83.
The GIRAFFE High Resolution mode was used in the H395.8 setup (HR02), which gives access to the 385.4 to 404.9nm wavelength range in the near UV at a spectral resolving power $\lambda / \delta \lambda$ of 22,700. A total of 4.5 h of integration were obtained, split in 3 exposures of 90 min. Given that most of the program targets were faint, the non-standard 50 khz,1$\times$1, high gain readout mode was used \citep{2008Msngr.133...17M}. Corresponding calibration frames were obtained within 24h of the science observations. The raw scientific data were bias- and dark-subtracted, flat-fielded and wavelength-calibrated using the instrument's pipeline v12.14.2 under the esorex environment. The source spectra were extracted in the SUM mode. Finally the individual exposures were summed up to improve the signal-to-noise ratio of the extracted data.
\subsection{Archival Data}
In addition to the new data taken with FLAMES, we also used archival H{\sc i}~ data from the GASS survey \citep{Kalberla_2010}. GASS is a survey of Galactic H{\sc i}~ taken with the Parkes antenna, with a beam size of 16$^\prime$, a spectral resolution of 1 km s$^{-1}$, and an rms noise of 57 mK. These data allow us to map out the structure of Complex WD on the sky, and compare the spectrum of the H{\sc i}~ emission to Ca{\sc ii}~ H \& K absorption along the line of sight. We present in Figure \ref{fig:WD} an image of Complex WD from GASS, along with the positions of clouds found in the \citet{1991A&A...250..509W} catalog, and the line-of-sight toward USNO-A0600-15865535.
\begin{figure*}[]
\begin{center}
\includegraphics[scale=3.5]{Complex_WD_chart}
\caption{{\label{fig:WD} Complex WD. The grayscale is H{\sc i}~ integrated column density between 98 km s$^{-1}$ and 131 km s$^{-1}$ on a log scale in Galactic coordinates. The feature below 5$^\circ$ is the Galactic plane, while the rest of the structure is Complex WD. The asterixes are the locations of the WD clouds in the Wakker van Woerden 1991 catalog. The cross is the location of USNO-A0600-15865535. The faint, parallel arc-like structures are systematic errors in the GASS data reduction.%
}}
\end{center}
\end{figure*}
\section{Methods \& Results}
\subsection{Stellar Classification}
The R $= 22700$, high SNR spectra of USNO-A0600-15865535 showed it to have very strong H$\epsilon$ and H$\zeta$ at $\sim 0$ km s$^{-1}$ in the local standard of rest (LSR) frame, and thus was very likely to be a halo star in the Milky Way, rather than the quasar we had originally attempted to target. Such stars are typically either blue horizontal branch (BHB) stars or blue stragglers (BS) stars, though can occasionally be hot main sequence (MS) stars.
To distinguish between these possibilities, we use a modification of the method developed by \citet{Sirko_2004} and later explored by \citet{Xue_2008}. In these works the shape of the H$\gamma$ and H$\delta$ lines are used to distinguish between these populations. These Balmer lines are fit with a standard Sersi\'c profile,
\begin{equation}\label{sersic}
\rm y = 1 - a~ exp \left[-\left(\frac{|\lambda-\lambda_0|}{b}\right)^{c}\right]
\end{equation}
and the parameters measured for these lines are distinct between the populations. Unfortunately, we do not have observations of the H$\gamma$ or H$\delta$ line, only the H$\epsilon$ and H$\zeta$ line. To determine whether we can use the (better resolved) H$\epsilon$ line as a similar discriminator, we first assemble a sub-population of BHB, BS, and MS drawn from the catalog presented in \citet{Xue_2008}, selected to be brighter than 16th magnitude in $g$. These targets have already been fit, and classified, but to test our method we refit the H$\gamma$ line in the SDSS DR10 spectra using Equation \ref{sersic}. We find that we can indeed reproduce the bifurcation between BHB and BS stars. We then apply this same fitting procedure to the H$\epsilon$ line. The results for both fits are shown in Figure \ref{hgammahepsilon}. The H$\epsilon$ line fit Sersi\'c parameters b and c do not as clearly delineate between BHB and BS stars, but the differentiation is still very much in place. We then apply this same fit to our VLT/FLAMES spectrum of USNO-A0600-15865535, and overplot the b and c parameters. USNO-A0600-15865535 is clearly a BHB star.
BHB stars have the useful property of being accurate standard candles. There is a direct relationship between BHB color and absolute magnitude reported in \citet{Xue_2008} (these values are originally reported in \citet{Sirko_2004} with a typographical error). The observed color in Pan-STARRS1 imaging is $g-r = -0.21$, and \citet{Schlegel_1998} report a reddening of $E\left(B-V\right) = 0.068$ toward this sightline, giving us a corrected color of $g-r = -0.28$ using the parameters from \citet{Schlafly_2011}. Using the correspondence reported in \citet{Sirko_2004}, this translates to an absolute magnitude of $M_g = 0.8$. USNO-A0600-15865535's measured $g$ magnitude of 14.2, reddening corrected to 14.0, gives a final distance of 4.4 kpc, with distance errors for BHB stars typically quoted at 10\%.
\begin{figure}[h!]
\begin{center}
\includegraphics[scale=0.8]{BHB_baby_ed}
\caption{{The Sersi\'c parameters of the H$\gamma$ (top) and H$\epsilon$ (bottom) lines as measured in 1637 stars from \citet{Xue_2008} with g $<$ 16. Overplotted in the bottom panel in a red cross is the result for the FLAMES spectrum of USNO-A0600-15865535, clearly a BHB star. \label{hgammahepsilon}%
}}
\end{center}
\end{figure}
\subsection{ISM Line Measuement}
We perform Voigt profile fitting of the calcium II H \& K lines after dividing out the continuum spectrum and, in the case of the calcium II H lines, dividing out the Sersi\'c fit to the strong H$\epsilon$ line, which is centered at 230 km s$^{-1}$ LSR in the calcium II H frame. The column densities were determined by simultaneously fitting Voigt profiles to both lines of the Ca{\sc ii}~ doublet with the \texttt{VPFIT} software \footnote{Available at \href{http://www.ast.cam.ac.uk/~rfc/vpfit.html}{http://www.ast.cam.ac.uk/$\sim$rfc/vpfit.html}}. For our Voigt profile fit analysis, the intrinsic model profiles are convolved with a Gaussian of FWHM 14 km s$^{-1}$ to account for the instrument resolution. The Ca{\sc ii}~ line close to the Milky Way rest frame is best fit with two Voigt profile components centered at $v_{LSR} \,\approx$ -6 and 1 km s$^{-1}$, respectively. The redshifted high-velocity absorption component is best fit with two Voigt profile components centered at $v_{LSR}\, \approx$ 92.3 and 108 km s$^{-1}$, respectively. All four individual Voigt profile compoents to Ca{\sc ii}~ K components agree with the equivalent Ca{\sc ii}~ H parameters to within errors. The individual Voigt profile fits are reported in Table \ref{tab:vp} and are shown in Figure 3.
\begin{table}[ht]
\centering
\label{tab:vp}
\begin{tabular}{lcc}
b [km s$^{-1}$] & $v_{LSR}$ [km s$^{-1}$] & log(N [cm$^{-2}$]) \\
\hline
6.3 $\pm$ 1.8 & -6.6 $\pm$ 0.6 & 12.05 $\pm$ 0.08 \\
20.7 $\pm$ 1.7 & 0.7 $\pm$ 1.7 & 12.15 $\pm$ 0.07 \\
11.8 $\pm$ 22.2 & 92.3 $\pm$ 36.7 & 11.58 $\pm$ 0.80 \\
10.4 $\pm$ 4.1 & 108.1 $\pm$ 7.1 & 12.17 $\pm$ 0.46 \\
\hline
\end{tabular}
\caption{{Voigt profile fit parameters for the Calcium II H \& K lines.}}
\end{table}
Three main H{\sc i}~ components are detected -- one near $v_{LSR} = 0$, one corresponding to the HVC at $v_{LSR} = 100$ and an intermediate velocity component. The H{\sc i}~ column density for each of these components are found by integrating the flux under the line under the optically thin assumption. We find a column density of $4.6 \times 10^{20}$ cm$^{-2}$, $2.2 \times 10^{19}$ cm$^{-2}$, and $8.6 \times 10^{20}$ cm$^{-2}$ for the low, intermediate, and high velocity components respectively.
\begin{figure}[h!]
\begin{center}
\includegraphics[scale=0.3]{BHB_Spectrum_stack1}
\caption{{H{\sc i}~ and Ca{\sc ii}~ toward USNO-A0600-15865535. The top panel is the GASS spectrum toward USNO-A0600-15865535, stretched to show the faint HVC feature at 100 km s$^{-1}$. The two bottom panels are the Ca{\sc ii}~ K and H absorption line features, clearly showing the HVC absorption. The vertical red ticks show the location of the centroids of individual Voigt profile fits.%
}}
\end{center}
\end{figure}
\section{Discussion}
The clear detection of Ca{\sc ii}~ H \& K absorption at +100 km s$^{-1}$ in USNO-A0600-15865535, coincident with H{\sc i}~ emission from Complex WD, indicates that the complex is closer than the star, at a distance 4.4 kpc. The column of Ca{\sc ii}~ measured is comparable between the cloud and the disk, even though they have wildly different H{\sc i}~ columns. This is consistent with the weak \citep{Wakker_2000} or non-existent \citep{Bekhti_2012} correlation between H{\sc i}~ and Ca{\sc ii}~ column density, which also explains the non-detection of the small intermediate velocity cloud along the line of sight at 60 km s$^{-1}$. Unfortunately, this lack of correlation makes it impossible to infer anything about the metallicity of the cloud from these metal absorption lines. Future metallicity measurements, perhaps toward USNO-A0600-15865535, will be critical in determining the origin of Complex WD, as gas of extragalactic origin typically has lower metallicities.
The kinematics and location of Complex WD do give us some clues as to its origin. Originally, \citet{1991A&A...250..509W} suggested that because complex WD is at positive velocity it is likely part of the structure of the Galaxy itself co-rotating with the disk. We call this the ``far" scenario in Figure \ref{fig:contour}, and it is emphatically ruled out by our detection of absorption. The appeal of the scenario is quite clear from the Figure -- a cloud at 20 kpc could be corotating with the disk. We now know that Complex WD is mostly inside the solar circle toward the fourth quadrant. Along the line of sight to USNO-A0600-15865535, Complex WD sits above a portion of the disk moving at $-$30 km s$^{-1}$ LSR if we assume that it is at the maximal distance of 4.4 kpc, decreasing to 0 km s$^{-1}$ LSR as we assume a closer distance. Complex WD is therefore strongly not in corotation with the disk. This is in rather stark contrast with other HVCs; a simplified model of HVCs with known distances found that they rotated with the disk at 77 km s$^{-1}$ -- slower than Galactic rotation, but with the same sense \citep{Putman_2012}.
A number of scenarios could account for an overall difference in velocity between the cloud and the disk. An accreting cloud could easily have a much lower accretion velocity than the rotation speed of the disk, and the positive velocity observed could be an artifact of the solar motion. Similarly, it is possible that Complex WD is material ejected from star-forming regions closer to Galactic center \citep[e.g.][]{Ford_2010}, and thus the high positive velocity is an effect of the lower specific angular momemtum of that material. Both of these scenarios suffer from the fine tuning required to meet the very small LSR velocity gradient found in the complex. A flux-weighted first-order polynomial fit to the velocity gradient in the Wakker \& van Woerden 1991 catalog of WD clouds find $-0.072 \pm 0.146$ km s$^{-1}$ per degree of Galactic longitude. The reflex velocity of the solar motion represents 100 km s$^{-1}$ across 40 degrees of Complex WD -- unless the Cloud is conspiring to thwart our detection of a velocity gradient, we should see some effect of the solar motion. While we cannot fully rule out the ``intermediate" scenario, where the bulk of the cloud is at $\sim 4$ kpc, this velocity structure puts very tight constraints on any future model.
Finally we examine a ``near" scenario, where Complex WD is only 1-2 kpc away. In this scenario the cloud originated from an area near the sun, and thus has inherited the overall solar motion, largely solving the fine-tuning of the LSR velocity. In the ``near" scenario the cloud is ejected from the disk by some kind of impulsive event, perhaps connected to star formation in the Gould Belt or Saggitarus Arm, imparting an overall 100 km s$^{-1}$ bulk velocity. A cloud this far away would only be about 0.5 kpc above the disk, which is quite low for most known HVCs, and would make it distinct from all other known HVCs in its origin. The shearing effect of differential rotation is much weaker close to the sun, which makes this impulsive scenario more credible for a closer cloud. The somewhat symmetric Complex WE, with a similar velocity and location but at negative Galactic latitude, could conceivably have been generated by the same event.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=1.0\columnwidth]{disk_grads_ann}
\caption{{\label{fig:contour} A top-down view of the Galaxy. Lines of constant Galactic longitude are shown in black, meeting at the sun, marked by $\odot$. The colored lines represent the LSR velocity of the gas in the Galactic disk expected from a flat, 220 km s$^{-1}$ rotation curve, with bluer contours representing more negative velocities, and redder representing more positive. The asterisks are the Complex WD clouds found in \citet{1991A&A...250..509W}, colored according to their LSR velocities with the same scheme as the disk. We show three distance scenarios: the original ``far" scenario, which this work has ruled out, an ``intermediate" scenario, where all clouds are put at 4.4 kpc, and a ``near" scenario, where all clouds are 1 kpc from the sun.%
}}
\end{center}
\end{figure}
\section{Conclusions}
In this work we have found that while the H$\epsilon$ line is not as powerful a discriminant between BHB, BS, and MS stars as the H$\gamma$ or H$\delta$ line, it can be used to show that USNO-A0600-15865535 is a BHB star approximately 4.4 kpc from the sun. We demonstrated that this star has clear Calcium H \& K absorption lines at a velocity coincident with Complex WD, and that therefore Complex WD must be closer than 4.4 kpc. We used this fact to rule out the originally assumed model of Complex WD, that is a Complex C-like cloud on the far side of the Galaxy, corotating with disk. Furthermore we investigated an intermediate distance scenario in which the Complex resides at $\sim 4$ kpc and found it difficult to reconcile the fixed LSR velocity of the cloud with the strong gradients implied from both the reflex solar motion and differential Galactic rotation. A "near" scenario, wherein a the complex was ejected from the solar vicinity to a distance of a few kpc seemed the most likely, though it would make the HVC the lowest in altitude known.
Future observations toward USNO-A0600-15865535 would enable precise determinations of metallicities using other elements, which would help determine whether a disk-origin for this cloud is likely. Further observations towards closer stars, and across the face of the cloud could give us much more detailed information about the distance and three-dimensional morphology of the cloud, which would also help us understand how the cloud came to be, and how it relates to the accretion and feedback story of the Milky Way.
\section{Acknowledgements}
The authors thank Jessica Werk and Alis Deason for advice on the BHB spectroscopic determination method. This work is based on observations made with ESO Telescopes at the La Silla Paranal Observatory under program ID 097.A-0552. This work used PanSTARRS1 data for targeting USNO-A0600-15865535 and accurate $g$ and $r$ band photometry. The Pan-STARRS1 Surveys have been made possible through contributions of the Institute for Astronomy, the University of Hawaii, the Pan-STARRS Project Office, the Max-Planck Society and its participating institutes, the Max Planck Institute for Astronomy, Heidelberg and the Max Planck Institute for Extraterrestrial Physics, Garching, The Johns Hopkins University, Durham University, the University of Edinburgh, Queen's University Belfast, the Harvard-Smithsonian Center for Astrophysics, the Las Cumbres Observatory Global Telescope Network Incorporated, the National Central University of Taiwan, the Space Telescope Science Institute, the National Aeronautics and Space Administration under Grant No. NNX08AR22G issued through the Planetary Science Division of the NASA Science Mission Directorate, the National Science Foundation under Grant No. AST-1238877, the University of Maryland, and Eotvos Lorand University (ELTE) and the Los Alamos National Laboratory. Partial support for this work was provided by NASA through Hubble Fellowship grant \#51354 awarded by the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., for NASA, under contract NAS 5-26555.
\bibliographystyle{apj}
|
\section{Conclusion and Future Work}
In this paper, we propose a hybrid deep framework by incorporating a hierarchical structure and joint attention model to the two stream convnet approach for human action recognition. The experimental results suggest that the proposed framework outperforms the two stream convnet approach. Despite using the only optical flow images as input, HAN achieves comparable performance with the state-of-the art method TDD that is much more computationally expensive. These results further support that (1) the hierarchical structure in HAN is important because it can model the frame transitions as well as long video segments and (2) the joint visual attention can help HAN focus on the important video regions and reduce the effect of noisy background. HAN is powerful in sequence modeling thus we would like to explore more applications for HAN in the future such as video event detection since a video event usually contains many sub-events and these sub-events have high dependencies to each.
\section{Experiments}
In this section, we first present the details of datasets and the evaluation protocol. Then, we describe the details of the implementation of our method. Finally, we present the experimental results with discussions.
\subsection{Datasets and evaluation protocol}
The evaluation is conducted on two public benchmark datasets, i.e., UCF-101~\cite{soomro2012ucf101} and HMDB51\cite{kuehne2011hmdb}. These two datasets are among the largest available annotated video action recognition datasets that have been used in~\cite{karpathy2014large,sharma2015action,simonyan2014two,wang2013action,wang2015action}. Specifically, UCF-101 contains $~13,000$ videos annotated into $101$ action classes with each class having at least 100 videos. HMDB51 is composed of $~6,700$ videos from $51$ action categories and each category has at least 100 video clips. For both datasets, the evaluation protocol is the same -- we follow the train/test splits provided by the corresponding organizers. The performance is measured by the mean of accuracies across all the splits in each dataset.
\subsection{Experiments Setting}
\textbf{Training two stream CNNs and HANs:}\quad Compared to image classification and detection, training a good deep convolutional neural network for videos understanding is more challenging. Similar to \cite{wang2015action,simonyan2014two}, we use the training data in UCF 101 split to train two stream CNNs. In our implementation, we use the Caffe toolbox \cite{jia2014caffe} and the layer configuration is the same as \cite{simonyan2014very}. All hidden layers use the rectification activation functions and max pooling is performed over $2\times 2$. Finally, each of the two networks contains 13 convolutional layers and 3 fully connected layers. The training procedure is similar to \cite{wang2015action,simonyan2014two}, where we use mini-batch stochastic gradient descent with momentum (0.9). The learning rate is initially set to $10^{-2}$, then changed to $10^{-3}$ after $10,000$ iterations and stopped after $30,000$ iterations and $10,000$ iterations for spatial and temporal nets, respectively. We use the Theano toolbox for HAN implementation and the model is trained by using Adadelta ~\cite{zeiler2012adadelta}. The dimension of LSTM is $1,024$ and the batch size is fixed to $128$ . Techniques of dropout \cite{dahl2013improving} and BPTT are used.
\textbf{Optical flow:}\quad The optical flow is computed by the off-the-shelf OpenCV toolbox with GPU implementation of \cite{zach2007duality}. Since the computational cost of optical flow is the bottleneck for the two stream CNN training. We pre-computed all the optical flow images and stored the horizontal and vertical components. The optical flow is computed by the adjacent two frames. In the testing stage, we fix the number of frames with the equal temporal window between them.
\subsection{ Results and Analysis}
We compare our models with a set of baselines proposed recently ~\cite{sharma2015action,yue2015beyond,simonyan2014two,wang2015action,karpathy2014large,peng2014action} including shallow video representation methods and deep ConvNets methods.
\begin{table}[!h]
\caption{Average accuracy over three splits on UCF-101 and HMDB51}
\label{tab: HAN}
\begin{tabular}{c||c||c}
\textbf{Model} & \textbf{UCF-101} & \textbf{HMDB51} \\ \hline\hline
Full HAN (spatial CNN cube+temporal CNN cube) &92.7\% & 64.3\% \\ \hline
HAN without attention$^{1}$(spatial CNN cube +temporal CNN cube) & 90.6\% & 62.0\% \\\hline
HAN without attention$^{2}$(spatial CNN 4096+ temporal CNN 4096) & 91.1\% & 62.7\% \\ \hline\hline
Spatial HAN (spatial CNN cube) & 75.1\% & 47.7\%\\\hline
Temporal HAN (temporal CNN cube) & 85.4\% &58.3\% \\\hline\hline
\end{tabular}
\end{table}
We first evaluate our proposed HAN on UCF-101 and HMDB51 datasets by comparing HAN with different settings to show the importance of each key component in HAN in Table ~\ref{tab: HAN}. Then, we further compare HAN with state-of-the art methods and experimental results are reported in Table~\ref{tab: state-of-the art}. From the tables, we can make the following observations:
\begin{itemize}
\item The proposed method with hierarchical LSTM outperforms methods without hierarchical structures \cite{yue2015beyond,karpathy2014large,simonyan2014two}. These results support that (1) the usage of LSTM can capture video sequences by considering the order of the motion transitions; and (2) the proposed hierarchical structure can effectively model the complex and long time range actions in videos.
\item Compared with methods without the attention components, the proposed HAN encourages the model to focus on the important regions in frames during the learning process, which improves the discriminative ability for classification. For example, in Figure ~\ref{fig:comparison} (b) and Figure ~\ref{fig:comparison} (e), we can see that our model can learn the important regions for actions more accurately.
\item The temporal and spatial features are complementary. First, by combining them together, both of them have been improved significantly. Second, compared with \cite{sharma2015action} that only considers attention in spatial, HAN can predict more motion related regions in the videos. Third, compared to TDD, the proposed HAN achieves comparable results without considering the iDT information, which suggests that the learned attention regions can have the similar ability to dense trajectory points and reduce the negative impact of background noises.
\item Compared to state-of-the art methods on UCF and HMDB51, HAN outperforms them remarkably except \cite{wang2015action}. The major reason for the exception is that the dataset HMDB is relatively small and the content is unconstrained, while the method in~\cite{wang2015action} incorporates iDT features that are computationally expensive.
\end{itemize}
\begin{figure}[!ht]
\centering
\subfigure[The sampled frame]{
\label{fig:att_orig}
\includegraphics[width=40mm,height=35mm]{exp_10.png}}
\subfigure[Attention results from HAN]{
\label{fig:att_pro}
\includegraphics[width=40mm,height=35mm]{exp_11.png}}
\subfigure[Attention results from \cite{sharma2015action}]{
\label{fig:att_other}
\includegraphics[width=40mm,height=35mm]{exp_12.png}} \\
\subfigure[The sampled frame]{
\label{fig:att1_orig}
\includegraphics[width=40mm,height=35mm]{exp_20.png}}
\subfigure[Attention results from HAN]{
\label{fig:att1_pro}
\includegraphics[width=40mm,height=35mm]{exp_21.png}}
\subfigure[Attention results from \cite{sharma2015action}]{
\label{fig:att1_other}
\includegraphics[width=40mm,height=35mm]{exp_22.png}}
\caption{Visual attention comparison between HAN and soft attention model in \cite{sharma2015action}, the green and red circles highlight the most important region learned by HAN and \cite{sharma2015action} respectively.}
\label{fig:comparison}
\end{figure}
\begin{table}[t]
\centering
\caption{Comparison with state of the art methods on UCF101 and HMDB51.}
\label{tab: state-of-the art}
\begin{tabular}{c || c || c}
\textbf{Model} & \textbf{UCF-101} & \textbf{HMDB51} \\ \hline\hline
Histogram of Oriented Gradient & 72.4 \% & 40.2\% \\\hline
Improved dense trajectories (iDT) \cite{wang2013action} & 85.9\%& 57.2\% \\ \hline
iDT + Stack Fish Vector \cite{peng2014action} & N/A & 66.8\% \\\hline\hline
spatial-temporal CNN \cite{karpathy2014large} & 65.4\% &N/A \\\hline
two stream CNN \cite{simonyan2014two} & 88.0\% & 59.4\% \\\hline
two stream CNN+LSTM \cite{yue2015beyond} & 88.6\% & N/A \\\hline
two stream CNN + iDT \cite{wang2015action} & 91.5\% & 65.9\% \\\hline
Soft Attention +LSTM \cite{sharma2015action} & 84.96\% & 41.3\% \\\hline \hline
Hierarchical Attention Networks & 92.7\% & 64.3\% \\\hline\hline
\end{tabular}
\end{table}
\section{Introduction}
Understanding human actions in wild videos can advance many real-world applications such as social activity analysis, video surveillance and event detection. Earlier works typically rely on hand craft features to represent videos~\cite{schuldt2004recognizing,wang2013action}. They often consist of two steps: motion detection and feature extraction. First, motion detectors are applied to detect the informative motion regions in the videos and then, hand craft descriptors such HOG ~\cite{dalal2005histograms}, SIFT, or improved Dense Trajectories (iDT)~\cite{wang2013action} extract the feature patterns from those motion regions to represent the video. In contrast to hand-craft shallow video representation, recent efforts try to learn video representation automatically from large scale labeled video data~\cite{karpathy2014large,simonyan2014two,donahue2015long,wang2015action,ji20133d}. For example, In~\cite{karpathy2014large}, authors stack the video frames as the input for convolution neural networks (CNN) and two stream CNNs ~\cite{simonyan2014two} combine optical flow and RGB video frames to train CNN and achieve comparable results with the state-of-the art hand craft based methods. Very recently, dense trajectory pooled CNN that combines iDT and two stream CNNs via the pooling layer achieves the state-of-the-art performance. However, ~\cite{simonyan2014two} and ~\cite{wang2015action} merely use short term motions that cannot capture the order of motion segments and semantic meanings.
The challenges of action recognition in wild videos are three-fold. First, there are large intra-class appearance and motion variances in the same action due to different viewpoints, motion speeds, backgrounds, etc. Second, wild videos are often collected from movies, TV shows and media platforms and usually have very low resolutions and noise background clutters, which exacerbate the difficulty for video understanding. Third, long range temporal dependencies are very difficult to capture. For example, the Optical Flow, iDT and 3D ConvNets ~\cite{ji20133d} are computed within a short-time window. Long Short Term Memory (LSTM) has been recently applied to video analysis \cite{donahue2015long} that provides a memory cell for long temporal information. However, it has been shown that the favorable time range of LSTM is around 40 frames ~\cite{sharma2015action,yue2015beyond}. In this work, we aim to develop a novel framework to tackle these obstacles for action recognition in videos.
\begin{figure}[t]
\centering
\includegraphics[width=6in]{hierarchical_two_stream_LSTM_with_attention.png}
\caption{An Illustration of the Proposed Model. The block LSTM means a LSTM cell, whose structure is given in Figure \ref{fig:lstm_one_cell}. The block ATTN indicates the operation to calculate attention weights by using the encoded features from both LSTMs. The block WA represents the weighted average of the input features with the weights from ATTN.}
\label{fig:hierarchical_two_stream_LSTM_with_attention}
\end{figure}
In this paper, we study the problem of video representation learning for action recognition. In particular, we investigate -- (1) how to utilize the temporal structures in the video to handle intra-appearance variances and background clutters by capturing the informative spatial regions; and (2) how to model the short-term as well as long-term motion dependencies for action recognition. Providing answers to these two research questions, we propose a novel Hierarchical Attention Network (HAN) that employs a hierarchical structure with recurrent neural network unit e.g., LSTM and a soft spatial temporal attention mechanism for video action recognition. Our contributions can be summarized as below:
\begin{itemize}
\item We propose a novel deep learning framework HAN for video action recognition which can explicitly capture both short-term and long-term motion information in an end to end process.
\item A soft attention is adopted on the spatial-temporal input features with LSTM to learn the important regions in a frame and the crucial frames in the videos.
\item We conduct an extensive set of experiments to demonstrate that the proposed framework HAN is superior to both state-of-the art shallow video representation based approaches and deep video representation based approaches on benchmark datasets.
\end{itemize}
The rest of this paper is organized as follows. Section 2 reviews related works. Section 3 describes the proposed hierarchical attention
deep learning framework in detail. Experimental results and comparisons are discussed in Section 4, followed by conclusions in Section 5.
\section{Related Work}
\textbf{Hand crafted features}: Video action recognition is a longstanding topic in computer vision community. Many hand-crafted features are used in still images. For example, ~\cite{schuldt2004recognizing} extends 2D harri corner detector to spatial temporal 3D interest points and achieves good performance with SVM classifier. Later, HOG3D that is based on HOG feature \cite{dalal2005histograms} shows its effectiveness by using integral images. Then improved Dense trajectories ~\cite{wang2013action} has dominated the filed of action recognition. It densely samples interest points and tracks them within a short time period. For each point, local descriptors such as HOG, MBH and HOF are extracted for representation. Then, all the features are encoded by the fish vector as the final video representation.
\textbf{Deep learned features}: Deep learning such as convolutional neural network has been shown its success in object detection and image classification in recent years. Based on image CNN, ~\cite{ji20133d,karpathy2014large} extends the CNN framework to videos by stacking video frames. However, the performance is lower than iDT based approaches. In order to better incorporate temporal information, ~\cite{simonyan2014two} proposes two stream CNNs and achieves comparable results with the state-of-the art performance of hand craft based feature representations \cite{wang2013action}. To consider the time dependency, \cite{yue2015beyond} proposes a LSTM based recurrent neural network to model the video sequences. Different from \cite{yue2015beyond} that uses a stack based LSTM, the proposed HAN proposes a hierarchical structure to model the video sequences in a multi-scale fashion.
\textbf{Recurrent visual attention model}: Visual attention model aims to capture the property of human perception mechanism by identifying the interesting regions in the images. Saliency detection \cite{itti1998model} typically predicts the human eye movement and fixations for a given image. In \cite{xu2015show}, a recurrent is proposed to understand where the model focuses on image caption generations. Moreover, recurrent attention model has been applied to other sequence modeling such as machine translation \cite{luong2015effective} and image generation\cite{gregor2015draw}.
\subsubsection*{References}
\bibliographystyle{plain}
\begingroup
\renewcommand{\section}[2]{}%
\section{The Proposed Method}
The overall architecture of HAN is shown in ~\ref{fig:hierarchical_two_stream_LSTM_with_attention}. We describe the three major components of HAN in this section -- the input appearance and motion CNN feature extraction, temporal sequence modeling and hierarchical attention model.
\subsection{Appearance and Motion Feature Extraction}
In general, HAN can adopt any deep convolution networks~\cite{ji20133d,simonyan2014two,wang2015action} for feature extraction. In this paper, we use two stream ConvNets ~\cite{simonyan2014two} to extract both appearance and motion features. Specifically, we train a VGG net \cite{simonyan2014very} to extract feature map $f_{P}$ and $f_{Q}$ for the $t$-th frame image $P^{t}$ and the corresponding optical flow image $Q^{t}$:
\begin{equation}
\begin{aligned}
f_{P^{t}}=CNN_{vgg}(P^{t}) \\
f_{Q^{t}}=CNN_{vgg}(Q^{t})
\end{aligned}
\end{equation}
Unlike two stream ConvNets \cite{simonyan2014two} that employs the last fully connected layer as the input feature, we use the features for $f_{P}$ and $f_{Q}$ from the last convolutional layer after pooling, which contains the spatial information of the input appearance and motion images. The input appearance and motion images are first rescaled to $224 \times 224$ and the extracted feature maps from the last pooling layer have the dimension of $D\times K\times K$ ($512\times 14\times14$ used in VGG net). $K\times K$ is the number of regions in the input image and $D$ is the number of the feature dimensions. Thus at each time step, we extract $K^2D$ dimension feature vectors for both appearance and motion images. We refer these feature vectors as feature cube shown in Figure ~\ref{fig:attention}. Then, the feature maps $f_{P^t}$ and $f_Q$ can be denoted in matrix forms as $\mathbf{P}^{t} = [\mathbf{p}_1 ^{t}, \mathbf{p}_2^{t}, \dots, \mathbf{p}_{K^2}^t ] \in \mathbb{R}^{D \times K^2}$ and $\mathbf{Q}^{t} = [\mathbf{q}_1 ^{t}, \mathbf{q}_2^{t}, \dots, \mathbf{q}_{K^2}^t ] \in \mathbb{R}^{D \times K^2}$, respectively.
\subsection{Recurrent Neural Network}
Long-short term memory (LSTM), which has the ability to preserve sequence information over time and capture long-term dependencies, has become a very popular model for sequential modeling tasks such as speech recognition~\cite{graves2013hybrid}, machine translation~\cite{bahdanau2014neural} and program execution~\cite{zaremba2014learning}. Recent advances in computer vision also suggest that LSTM has potentials to model videos for action recognition~\cite{sharma2015action}. We follow the LSTM implementation in~\cite{zaremba2014recurrent}, which is given as follows
\begin{figure}[h]
\centering
\includegraphics[width=110mm,height=70mm]{LSTM_one_cell.png}
\caption{Illustration of One LSTM Cell}
\label{fig:lstm_one_cell}
\end{figure}
\begin{align}
\mathbf{i}_t & = \sigma(\mathbf{W}_{ix} \mathbf{x}_t + \mathbf{W}_{ih} \mathbf{h}_{t-1} + \mathbf{b}_i) \nonumber \\
\mathbf{f}_t & = \sigma(\mathbf{W}_{fx} \mathbf{x}_t + \mathbf{W}_{fh} \mathbf{h}_{t-1} + \mathbf{b}_f) \nonumber \\
\mathbf{o}_t & = \sigma(\mathbf{W}_{ox} \mathbf{x}_t + \mathbf{W}_{oh} \mathbf{h}_{t-1} + \mathbf{b}+o) \nonumber \\
\mathbf{g}_t & = \tanh(\mathbf{W}_{gx} \mathbf{x}_t + \mathbf{W}_{gh} \mathbf{h}_{t-1} + \mathbf{b}_g) \nonumber \\
\mathbf{c}_t & = \mathbf{f}_t \odot \mathbf{c}_{t-1} + \mathbf{i}_t \odot \mathbf{g}_t \nonumber \\
\mathbf{h}_t & = \mathbf{o}_t \odot \tanh(\mathbf{c}_t)
\end{align}
where $\mathbf{i}_t$ is the input gate, $\mathbf{f}_t$ is the forget gate, $\mathbf{o}_t$ is the forget fate, $\mathbf{c}_t$ is the memory cell state at $t$ and $\mathbf{x}_t$ is the input features at t. $\sigma(\cdot)$ means the sigmoid function and $\odot$ denotes the Hadmard product. The main idea of the LSTM model is the memory cell $\mathbf{c}_t$, which records the history of the inputs observed up to $t$. $\mathbf{c}_t$ is a summation of -- (1) the previous memory cell $\mathbf{c}_{t-1}$ modulated by a sigmoid gate $\mathbf{f}_t$, and (2) $\mathbf{g}_t$, a function of previous hidden states and the current input modulated by another sigmoid gate $\mathbf{i}_t$. The sigmoid gate $\mathbf{f}_t$ is to selectively forget its previous memory while $\mathbf{i}_{t}$ is to selectively accept the current input. $\mathbf{i}_t$ is the gate controlling the output. The illustration of a cell of LSTM at the time step $t$ is shown in Figure \ref{fig:lstm_one_cell}. Next we will introduce how to model both appearance and motion features using LSTM and how to integrate attention by using encoded features.
\subsection{Hierarchical LSTMs to Capture Temporal Structures}
One natural way of modeling videos is to feed features $\mathbf{P}^t$ and $\mathbf{Q}^t$ into two LSTMs and then put a classifier at the output of LSTMs for classification. However, this straightforward method doesn't fully utilize the structure of actions. In real world, an action is usually composed of a set of sub-actions, which means that video temporal structures are intrinsically layered. For example, a video about long jump consists of three sub-actions -- pushing off the board, flying over the pit and landing. As the three actions take place sequentially, there are strong temporal dependencies among them thus we need to appropriately model the temporal structure among the three actions. In the meantime, the temporal structure within each action is composed of multiple actions. For example, pushing off the board is composed of running and jump. In other words, the actions we want to recognize are layered and we need to model video temporal structure with multiple granularities. However, directly applying LSTM cannot capture this property. To fully capture the video temporal structure, we develop a hierarchical LSTM. An illustration of the hierarchical LSTM is shown in the purple rectangle in Figure \ref{fig:hierarchical_two_stream_LSTM_with_attention}. This hierarchical LSTM is composed of two layers -- the first layer accepts the appearance feature of each frame as the input and the output of the first layer LSTM is used as the input of the second layer LSTM. To capture the dependencies between different sub-actions such as dependencies between pushing off the board, flying over the pit and landing, we skip every $k$ encoded features from LSTM and use that as the input to the second layer. In addition to capturing the video temporal structure, another advantage of layered LSTM is to increase the learning capability of LSTM. By adding another layer in LSTM, we allow LSTM to learn higher level and more complex features, which is a common practice proven to work well in other deep architectures such as CNN, DNN and DBM. Thus, as shown in Figure \ref{fig:hierarchical_two_stream_LSTM_with_attention}, we use two hierarchical LSTMs to model the appearance and motion features, respectively.
\subsection{Attention Model to Capture Spatial Structures}
\begin{figure}[h]
\centering
\includegraphics[width=120mm,height=100mm]{Video_features.png}
\caption{attention}
\label{fig:attention}
\end{figure}
The $K^2$ vectors in the appearance $\mathbf{P}^t$ (or motion features $\mathbf{Q}^t$) correspond to $K^2$ regions in the $t$-th frame, which essentially encode the spatial structures. For action recognition, not every region of the frames are relevant for the task at hand. Obviously, we want to focus on the regions where the action is happening. For the action shown in Figure \ref{fig:attention}, we want to mainly focus on hands and legs that are useful for identifying the action; while the background is noisy as a person can perform the same action at different locations. Therefore we could confuse the classifier if we also target on backgrounds. Thus, it is natural for us to assign different attention weights to different regions of the frame. Since video frames are sequential, neighboring frames have strong dependencies, which suggests that we can use the encoded features at time $t-1$ to predict the attention weights at time $t$ and then use the attention weights to refine the input. Specifically, at each time step $t-1$, we use a softmax function over $K \times K$ locations to predict the importance of the $K^2$ locations in the frame, which is written as:
\begin{equation}
l_i^t = \frac{\exp(\mathbf{w}_i^T \mathbf{h}_{t-1})}{\sum_{j=1}^{K^2} \exp(\mathbf{w}_j^T \mathbf{h}_{t-1})}
\end{equation}
where $l_i^t$ is the importance weight of the $i$-th region of the $t$-th frame, $\mathbf{W} = \{ \mathbf{w}_1, \mathbf{w}_2, \dots, \mathbf{w}_{K^2}\} \in \mathbb{R}^{2D \times K^2}$ are the weights of the softmax function and $\mathbf{h}_{t-1}$ is the concatenation of $\mathbf{h}_{t-1}^{p,1}$ and $\mathbf{h}_{t-1}^{q,1}$, i.e., the encoded appearance and motion features of the $(t-1)$-th frame from the first layer LSTM. Note that we use the encoded appearance feature and motion feature jointly to compute the attention weights instead of computing two attention weights by using the two features, separately. Its advantages are two fold. First, the flow and appearance features capture different aspects of the frame but the attention location on the video should be the same, thus we do not need to calculate two sets of attentions for appearance and optical LSTM separately, which may introduce more computational cost. Second, appearance and motion features provide complimentary information that may help predict more accurate attention. With the attention weights given above, the inputs of the two LSTMs are the weighted average of different locations as:
\begin{equation}
\mathbf{x}_t^{p} = \sum_{i=1}^{K^2} l_i^t \mathbf{p}_{i}^t ~~~ \text{and} ~~~ \mathbf{x}_t^{q} = \sum_{i=1}^{K^2} l_i^t \mathbf{q}_{i}^t
\end{equation}
\subsection{Action Recognition with HAN}
We use $\mathbf{h}_T^{p,i}$ to denote the encoding of the the video by the $i$-the layer LSTM for the appearance features and $\mathbf{h}_T^{q,i}$ for motion features. As mentioned above, the hierarchical LSTM captures multi-granularity of video temporal structures, thus, encoded features in different levels (different $i$) provide distinct descriptions of different granularity about actions, which are all useful for action recognition. In addition, the two LSTMs encode complementary information from appearance and motion features, thus encoded features from appearance and motion are also relevant for action recognition. Therefore, we concatenate these features as $\mathbf{h}_f = [\mathbf{h}_T^{p,1}, \dots, \mathbf{h}_T^{p,L}, \mathbf{h}_T^{q,L}, \dots, \mathbf{h}_T^{q,L}]$, where $L$ is the number of layers in each LSTM. We then use the softmax function to predict the probability that the video $v_i$ is classified into the class $c$ as
\begin{equation}
p(c|v_i, \text{HAN}) = \text{softmax}(\text{HAN}(v_i))
\end{equation}
and the loss function is
\begin{equation}
\max_{\text{HAN}, \mathbf{W}_s} \sum_{i=1}^N \log p( y_i |v_i, \text{HAN})
\end{equation}
where $N$ is number of videos, $y_i$ is the label of $v_i$ and $\mathbf{W}_s$ are the weights of the softmax classifier.
|
\section{Introduction: Separating Homophily from Social Influence}
It is an ancient observation that people are influenced by others (nearby) in
their social network --- that is, the behavior of one node in a social network
adapts or responds to that of neighboring nodes. Such social influence is not
just a curiosity, but of deep theoretical and empirical importance across the
social sciences. It is also of great importance to various kinds of social
engineering, e.g., marketing (especially but not only ``viral'' marketing),
public health (over-coming ``peer pressure'' to engage in risky behaviors, or
using it to spread healthy ones), education (``peer effects'' on learning),
politics (``peer effects'' on voting), etc. Conversely, it is an equally
ancient observation that people are not randomly assigned their social-network
neighbors. Rather, they {\em select} them, and tend to select as neighbors
those who are already similar to themselves\footnote{This is not necessarily
because they {\em prefer} those who are similar; all more-desirable potential
partners might have already been claimed
\citep{JLMartin-on-social-structures}.}. This {\bf homophily} means that
network neighbors are informative about latent qualities a node possess,
providing an alternative route by which a node's behavior can be predicted from
their neighbors. Efforts to separate homophily from influence have a long
history in studies of networks \citep{Leenders-structure-and-influence}.
Motivated by the controversy over \citet{Christakis-Fowler-spread-of-obesity},
\citet{Homophily-contagion-confounded} showed that unless {\em all} of the
nodal attributes which are relevant both to social-tie formation {\em and} the
behavior of interest are observed, then social-influence effects are generally
unidentified. The essence of this result is that a social network
is a machine for {\em creating} selection bias\footnote{A turn of phrase
gratefully borrowed from Ben Hansen.}.
\citet[\S 4.3]{Homophily-contagion-confounded} did conjecture a possible
approach for identification of social influence, even in an homophilous
network. When a network forms by homophily, a node is likely to be similar to
its neighbors. Following this logic, these neighbors are likely to be similar
to {\em their} neighbors and therefore the original node. In the simplest
situations, where there are only a limited number of node types, this means
that a homophilous network should tend to exhibit clusters with a high
within-cluster tie density and a low density of ties across clusters. Breaking
the network into such clusters might, then, provide an observable proxy for the
latent homophilous attributes. The same idea would work, {\em mutatis
mutandis}, when those attributes are continuous.
\citet{Homophily-contagion-confounded} therefore conjectured that, under
certain assumptions on the network-growth process (which they did not specify),
unconfounded causal inferences could be obtained by controlling for {\em
estimated} locations in a latent space. More recently,
\citet{Davin-Gupta-Piskorski-separating-homophily} and
\citet{Worrall-homophily} have shown that, in limited simulations, such
controls can indeed reduce the bias in estimates of social influence, at least
when the network grows according to certain, particularly well-behaved, models.
In this paper, we complement these simulation studies by establishing
sufficient conditions under which controlling for estimated latent locations
leads to {\em asymptotically} unbiased and consistent estimates of social
influence effects. For certain network models, we also show that the remaining
finite-sample bias shrinks exponentially in the size of the network, and
potentially can be upper-bounded by solving a quadratic optimization problem.
To the best of our knowledge, our results provide the first {\em theoretical}
guarantees of consistent estimation of social-influence effects from
non-experimental data, in the face of latent homophily.
Section \ref{sec:setting} lays out the basics of the setting we are working in,
starting with assumptions about the processes of network formation and social
influence (and the links between them), and rehearsing relevant results from the
prior literature on latent community models (\S \ref{sec:communities}) and
continuous latent space models (\S \ref{sec:continuous-space}). Section
\ref{sec:community-control} presents our main results about the asymptotic
estimation of social influence in the presence of latent homophily. (Some of
the longer proofs are deferred to \S \ref{sec:proofs}.)
Section \ref{sec:discussion} discusses the strengths and limits of our
results in the context of the related literature.
\section{Setting and Assumptions}
\label{sec:setting}
We are interested in the patterns of a certain behavior or outcome over time,
across a social network of $n$ nodes. The behavior of node $i \in \{1, \ldots,
n\}$ at time $t \in \{1, \ldots, \infty\}$ is represented by random variable
$Y(i,t) \in \mathbb{R}$. Social network ties or links will be represented
through an $n \times n$ adjacency matrix $A$, with $A_{ij} = 1$ if $i$ receives
a tie from $j$, and $A_{ij} = 0$ otherwise. (In many contexts these ties are
undirected, so $A_{ij} = A_{ji}$, but our results do not require this.) As this
notation suggests, we will assume that the network of social ties does not
change, at least over the time-scale of the observations\footnote{Latent space
modeling of dynamic networks is still in its infancy. For some preliminary
efforts, see, e.g., \citet{DuBois-Butts-Smyth-blockmodeling-of-dynamics,
Ghasemian-et-al-community-in-dynamic-networks} for block models, and
\citet{Sarkar-Moore-dynamical-social-network} for continuous-space models.}.
Additionally, each node has an unchanging set of latent covariates $X_i$.
The crucial assumption of our analysis is that for node $i$, there is a
vector-valued latent variable $C_i$ which controls their location in the
network, i.e., their probabilities of having ties with any other node
$j ~(\neq i)$. For our analysis, we use $C$ to represent the array
$[C_1, C_2, \ldots C_n]$. Furthermore, we assume that
$\Prob{A_{ij}=1|C} = w(C_i, C_j)$ for some measurable function $w$, and that
$A_{ij}$ is conditionally independent given $C$, $\forall i,j$. Models that
satisfy this assumption --- i.e., that all ties are conditionally independent
given the latent variables for each node --- are sometimes called ``graphons''
or ``$w$-random graphs'' and are clearly exchangeable (permutation-invariant)
over nodes. Conversely, the Aldous-Hoover theorem
\citep[ch. 7]{Kallenberg-symmetries} shows that this condition is, in fact, the
generic form of exchangeable random networks. The graphical causal
model\footnote{We do not mean to take sides in the dispute between the
partisans of graphical causal models and those of the potential-outcomes
formalism. The expressive power of the latter is strictly weaker than that
of suitably-augmented graphical models \citep{Richardson-Robins-SWIGs}, but
we could write everything here in terms of potential outcomes, albeit at a
great cost in space and notation.} capturing social influence in our setting
is shown in Figure \ref{fig:graphical-model}.
\begin{figure}
\includegraphics[width=\textwidth]{graphical-model}
\caption{The graphical causal model for our setting. For simplicity, we have
not written out the cross-terms between $j$ and $k$.}
\label{fig:graphical-model}
\end{figure}
The linear\footnote{See \S \ref{sec:nonlinear} for a discussion of non-linear models.} structural-equation model is thus
\begin{equation}
Y_{i,t+1} = \alpha_0 + \alpha_1 Y_{i,t} + \beta\sum_{j}{\left(Y_{j,t}A_{ij}\right)} + \gamma_1^T C_i + \gamma_2^T X_{i} + \epsilon \label{eqn:structural-equation} ~,
\end{equation}
with $X_{i}$ being a vector of un-changing, network-irrelevant attributes for
each node; $\epsilon$ representing noise uncorrelated with any of the other
variables on the right-hand side; and $\gamma_1$ and $\gamma_2$ serving as
appropriately-sized vectors of coefficients. Our goal is to identify, and
estimate, $\beta$, the coefficient for social influence.
When estimating social influence, we do not observe either $X_i$ or $C_i$, so
we cannot estimate a model of the form \eqref{eqn:structural-equation}.
But we {\em can} estimate
\begin{equation}
Y_{i,t+1} = \alpha_0 + \alpha_1 Y_{i,t} + \beta\sum_{j}{\left(Y_{j,t}A_{ij}\right)} + \gamma_0^T \hat{C}_i + \eta \label{eqn:effective-model} ~,
\end{equation}
where $\hat{C}_i$ is an estimated or discovered location for node $i$ and the
noise term $\eta$ is now
\begin{equation}
\eta = \epsilon + \gamma_2^T X_{i} + (\gamma_1^T C_i - \gamma_0^T \hat{C}_i) ~.
\end{equation}
It is well known, by standard arguments for linear models, that
\eqref{eqn:effective-model} results in unbiased, unconfounded inference for
$\beta$ if $\eta$ is uncorrelated with $Y_{j,t}$, conditional on the controls
included in the model. (It is not necessary that $\eta$ have zero mean, since
that would just be incorporated into the intercept.)
We will make the following assumptions:
\begin{eqnarray}
X_i & \rotatebox{90}{\ensuremath{\models}} & C_i \label{eqn:screening-egos-attributes-from-alters-behavior-1}\\
X_{i} & \rotatebox{90}{\ensuremath{\models}} & Y_{j, t} | \hat{C}_i \label{eqn:screening-egos-attributes-from-alters-behavior-2}
\end{eqnarray}
We note that these are essentially behavioral rather than statistical
assumptions, and therefore must be justified on substantive grounds in the
specific context of the study where network influence is being
estimated. Condition
\eqref{eqn:screening-egos-attributes-from-alters-behavior-1} is that the $C_i$
variable captures {\em all} the attributes of a node which are relevant to
their location in the network. Other persistent attributes which might be
relevant to the behavior of interest are independent of network location. The
second and related assumption
\eqref{eqn:screening-egos-attributes-from-alters-behavior-2} is that given our
estimated locations, we learn nothing about a node's unobserved,
network-irrelevant attributes by observing a neighbor's behavior.
\eqref{eqn:screening-egos-attributes-from-alters-behavior-2} implies that the
$\gamma_2^T X_{i}$ contribution to $\eta$ is uncorrelated with $Y_{j, t}$,
therefore it does not add any bias to the estimates of $\beta$; instead it just
increases the variance of the noise term. We have therefore to consider the
other contribution to $\eta$, $(\gamma_1 C_i - \gamma_0 \hat{C}_i)$, and
whether it is uncorrelated with $Y_{j, t}$ given $\hat{C}_i$. \label{error-term-from-c-vs-hat-c}
It is intuitively clear (and formally shown in Lemma
\ref{lemma:no-error-no-problem} below) that if $\hat{C} = C$, estimating
\eqref{eqn:effective-model} is just as good, for finding $\beta$, as estimating
\eqref{eqn:structural-equation}. It should further be plausible (and is
formally shown in \S \ref{sec:community-control} below) that if $\hat{C}$ is a
``good enough'' estimate of $C$, i.e., one which is consistent and converges
sufficiently rapidly, the covariance between $\eta$ and $Y_{j, t}$ shrinks fast
enough that \eqref{eqn:effective-model} will still yield asymptotically
unbiased and consistent estimates of $\beta$.
We do not (yet) know how to get such ``good enough'' estimates of latent node
locations $\hat{C}$ for arbitrary graphons. For this reason, we specialize to
two settings, latent community (stochastic block) models and latent space
models, where the latent node locations $C$ and the link-probability function
$w$ take particularly tractable forms, which have been extensively explored in
the literature. It is by building on results for these models that we can find
regimes where the social-influence coefficients can be estimated consistently.
\paragraph{On $X_i$ and $C_i$} The relatively permanent attributes of node $i$
can be divided in two cross-cutting ways. On the one hand, some attributes are
(in a given study) observable or manifest, and others are latent. On the other
hand, a given attribute could be a cause of the behavior of interest ($Y_i$),
or a cause of network ties ($A_{ij}$), or of both. (Attributes which are
irrelevant to both behavior and network ties are mere distractions here, and we
will ignore them.) One of our key assumptions is that {\em all} of the
network-relevant attributes of node $i$ can be represented by a single $C_i$,
whether or not they are also relevant to behavior, and that this is a latent
variable. There might be attributes that are incorporated into $C_i$ which are relevant {\em only} to
network ties, not behavior, and independent of the other attributes; these are
of no concern to us, and can be regarded as part of the noise in the
tie-formation process. Our subsequent assumptions are, speaking roughly, that
observing the whole network gives us so much information about these $C_i$ that
we learn nothing (in the limit) about $C_i$ from also observing $Y_i$.
\subsection{The Latent Communities Setting}
\label{sec:communities}
In our first setting, we presume that nodes split into a finite number of
discrete types or classes ($k$), which in this context are called {\bf blocks},
{\bf modules} or {\bf communities}. More precisely, there exist a function
$\sigma \colon \{1,\dots, n\} \mapsto \{1,\dots, k\}$ assigning nodes to
communities. We specifically assume that the network is generated by a {\bf
stochastic block model}, which is to say that there are $k$ communities, that
$\sigma(i) \overset{iid}{\sim} \rho$, for some fixed\footnote{Strictly, some of the theory
referenced below allows $k$ to grow with $n$, though with {\em a priori}
known rates.} (but unknown) multinomial distribution $\rho$, and that $w$ is
given by a $k\times k$ {\bf affinity matrix}, so that
\[
\Prob{A_{ij}=1|\sigma(i)
= a, \sigma(j) = b} = w_{ab} ~.
\]
We may translate between $\sigma$ (a sequence of categorical variables) and our
earlier $C$ (a real-valued matrix) by the usual device of introducing indicator
or ``dummy'' variables for $k-1$ of the communities, so that $C_i$ is a $k-1$
binary vector which is a function of $\sigma(i)$ and vice versa.
The objective of community detection or community discovery is to provide an
accurate estimate $\hat{\sigma}$ or $\hat{C}$ from the observed adjacency
matrix $A$, i.e., to say which community each node comes from, subject to a
permutation of the label set. (``Accuracy'' here is typically measured as the
proportion of mis-classification.) Since the problem was posed by
\citet{Girvan-MEJN-community-structure} a vast literature has emerged on the
topic, spanning many fields, including physics, computer science, and
statistics; see \citet{Fortunato-community-detection} for a review. However,
we may summarize the relevant findings of the most recent work as follows.
\begin{enumerate}
\item For networks which are generated from latent community models, under very
mild regularity conditions, it is possible to recover the communities
consistently, i.e., as $n \rightarrow \infty$,
$\Prob{\hat{C} \neq C} \rightarrow 0$ \citep{Bickel-Chen-on-modularity,
Zhao-Levina-Zhu-consistency-of-community-detection}. That is, with
probability tending to one, we can get {\em all} of the community assignments
right\footnote{Naturally, we allow for a global permutation of the labels
between $C$ and $\hat{C}$.}.
\item Such consistent community discovery can be achieved by algorithms whose
running time is polynomial in $n$.
\item The minimax rate of convergence is in fact exponential in $n$ (and can be
achieved by the algorithms mentioned below).
\end{enumerate}
These points, particularly the last, will be important in our argument below,
and so we now elaborate on them.
Recently, \citet{Zhang-sbm_minimax} proved that under very mild regularity
conditions the minimax rate of convergence for networks generated from latent
community models is in fact exponential in $n$. Furthermore,
\citet{Gao-sbm-minimax_algo} exploits techniques provided by
\citet{Zhang-sbm_minimax} to propose an algorithm polynomial in $n$ that
achieves this minimax rate, under slightly modified but equally mild regularity
conditions. More precisely, \citet{Gao-sbm-minimax_algo} considers a general
stochastic block model, parametrized by $n$, the number of nodes; $k$, the
number of communities; $a$ and $\alpha \ge 1$, where
$\frac{a}{n} = \min_{i}{w(i,i)} \le \max_{i}{w(i,i)} \le \frac{\alpha a}{n}$,
ensuring that within-community edges are ``sufficiently'' dense; $b$, where
$\frac{b \alpha}{n} \le \frac{1}{k(k-1)} \sum_{i \ne j}{w(i,j)} \le \max_{i \ne
j}{w(i,j)} \frac{b}{n}$, with $0 < \frac{b}{n} < \frac{a}{n} < 1$, ensuring
that between-community edges are ``sufficiently'' sparse; and $\beta \ge 1$,
where the number of nodes in community $k$,
$n_k \in \left[\frac{n}{\beta k}, \frac{\beta n}{k} \right]$, ensuring that
community sizes are ``sufficiently'' comparable. \citet{Zhang-sbm_minimax} and
\citet{Gao-sbm-minimax_algo} diverge slightly as the former only requires
$\max_{i \ne j}{w(i,j)} \le \frac{b}{n}$ and
$\frac{a}{n} \le \min_{i}{w(i,i)}$. Additionally, the latter slightly
restricts the parameter space by requiring the $k^{th}$ singular value of the
affinity matrix $w$ to be greater than some parameter $\lambda$. The general
context of the theory described in \citet{Zhang-sbm_minimax,
Gao-sbm-minimax_algo} is defined for absolute constant $\beta \ge 1$ and also
in \citet{Gao-sbm-minimax_algo} for absolute constant $\alpha \ge 1$, while
$k$, $a$, $b$, and $\lambda$ are functions of $n$ and therefore vary as $n$
grows. However, in the context of our work, we only consider latent community
where $k$, $\frac{a}{n}$, $\frac{b}{n}$, and $\lambda$ are also absolute
constants. We shall refer to this whole set of restrictions on the latent
community model as ``the GMZZ conditions''.
Given a latent community model satisfying the GMZZ conditions, the minimax rate
of convergence for the expected {\em proportion} of errors is
\begin{align}
\label{eq:converg-rate-2}
\exp{\left(-(1+o(1))\frac{nI}{2}\right)}, & \quad k=2 \\
\label{eq:converg-rate-3}
\exp{\left(-(1+o(1))\frac{nI}{\beta k}\right)}, & \quad k \ge 3,
\end{align}
where $I$ is the \citet{Renyi-introduces-Renyi-entropy}
divergence of order $\frac{1}{2}$ between two
Bernoulli distributions with success probabilities $\left(\frac{a}{n}\right)$
and $\left(\frac{b}{n}\right)$:
$D_{\frac{1}{2}}\left(\text{Ber}\left(\frac{a}{n}\right) \|
\text{Ber}\left(\frac{b}{n}\right) \right)$. Recall, that $\beta$ in
addition to $k$, $\frac{a}{n}$, $\frac{b}{n}$ are, in our context, constant in
$n$; therefore, \eqref{eq:converg-rate-2} and \eqref{eq:converg-rate-3} both
reduce to $\exp{\left(-O(n)\right)}$. The algorithm of
\citet{Gao-sbm-minimax_algo} achieves just this rate at a computational cost
polynomial in $n$. More specifically, the time complexity of their algorithm
is (by our calculations) at most $O(n^3)$, but we do not know whether this is
tight. It would be valuable to know whether this rate is also a lower bound on
the computational cost of obtaining minimax error rates, and if the complexity
could be reduced in practice for very large graphs via parallelization.
\eqref{eq:converg-rate-2}--\eqref{eq:converg-rate-3} are, as mentioned,
for the expected proportion of errors. We require a bound on the probability
of making any errors at all, but this is, in fact, implicit in the results
of \citet{Zhang-sbm_minimax, Gao-sbm-minimax_algo}. To see this, let $M$
stand for the number of errors. Then their results show that
\begin{equation}
\Expect{M/n} \leq e^{-cn}
\end{equation}
for an appropriate constant $c > 0$ (and large enough $n$). Hence
\begin{equation}
\Expect{M} \leq n e^{-cn}
\end{equation}
Now we apply Markov's inequality: the probability of making any errors at all
is the probability that $M \geq 1$, and
\begin{equation}
\Prob{M \geq 1} \leq \Expect{M}/1 \leq n e^{-cn} = e^{-cn+\log{n}} = e^{-O(n)}
\label{eqn:probability-of-any-errors-is-exp-small}
\end{equation}
which goes to zero exponentially fast in $n$. Indeed, since
$\sum_{n}{n e^{-cn}}$ is finite\footnote{To see this, differentiate the
geometric series $\sum_{n}{e^{-cn}}$ with respect to $c$.}, the
Borel-Cantelli lemma \citep[Theorem 7.3.10a, p.\ 288]{Grimmett-Stirzaker} tells
us that with probability 1, $M \geq 1$ only finitely often, i.e., that
$M \rightarrow 0$ almost surely. But this strong consistency is more than we
need.
\subsection{The Continuous Latent Space Setting}
\label{sec:continuous-space}
The second setting we consider is that of continuous latent space models. In
this setting, the latent variable on each node, $C_i$, is a point in a
continuous metric space (often but not always $\mathbb{R}^d$ with the Euclidean
metric), and $w(C_i, C_j)$ is a decreasing function of the distance between
$C_i$ and $C_j$, e.g., a logistic function of the distance. This
link-probability function is often taken to be known {\em a priori}. The
latent locations $C_i \overset{iid}{\sim} F$, where $F$ is a fixed but unknown
distribution, or, more rarely, a point process. Different distributions over
networks thus correspond to different distributions over the continuous latent
space, and vice versa.
Parametric versions of this model have been extensively developed since
\citet{Hoff-Raftery-Handcock}, especially in Bayesian contexts. Less attention
has been paid to the consistent estimation of the latent locations in such
models than to the estimation of community assignments in latent community
models. Recent results by \citet[ch. 3]{Asta-thesis}, however, show that when
$w$ is a smooth function of the metric whose logit transformation is bounded,
the maximum likelihood estimate $\hat{C}$ converges on $C$, with the
probability of an error of size $\epsilon$ or larger is
$O(\exp{\left(-\kappa \epsilon n^2\right)})$, where the constant $\kappa$
depends on the purely geometric properties of the space (see \S
\ref{sec:continuous-latent-space-setting} below). This result holds across
distributions of the $C_i$, but may not be the best possible rate.
\section{Control of Confounding}
\label{sec:community-control}
Recall that we wish to find sufficient conditions under which estimating
$\beta$ from \eqref{eqn:effective-model}, with estimated or discovered node
locations $\hat{C}$, will yield consistent, asymptotically unbiased estimates
of $\beta$ from \eqref{eqn:structural-equation}, which uses the true locations
$C$.
Let us first establish a baseline by considering the case where the estimates
of node locations are perfect, $\Prob{C \neq \hat{C}} = 0$.
\begin{lemma}
\label{lemma:no-error-no-problem}
Under the assumptions
\eqref{eqn:screening-egos-attributes-from-alters-behavior-1}--\eqref{eqn:screening-egos-attributes-from-alters-behavior-2},
if $\Prob{C \neq \hat{C}} = 0$, then the the ordinary least squares estimate
of $\beta$ in \eqref{eqn:effective-model} is also an unbiased and consistent
estimate of the social-influence coefficient from
\eqref{eqn:structural-equation}.
\end{lemma}
\textsc{Proof:} Clearly, $\gamma_0 = \gamma_1$, and $(\gamma_1^T C_i -
\gamma_0^T \hat{C}_i) = \gamma_1^T (\hat{C}_i - \hat{C}_i) = 0$. Indeed, even
if some value other than $\gamma_1 = \gamma_0$ where used, it still follows
that $(\gamma_1 - \gamma_0)^T \hat{C}_i$, which is a function of $\hat{C}_i$,
will therefore be independent of $Y_{j,t}$ given $\hat{C}_i$. So, when locations
are inferred with no error (in the limit), we find that $\eta$ is uncorrelated
with $Y_{j,t}$, and hence \eqref{eqn:effective-model} provides unbiased and
consistent estimates of the social-influence coefficient $\beta$. $\Box$
Let us now consider the properties of \eqref{eqn:effective-model} for finite
$n$. The covariance of interest is that between $Y_{j,t}$ and the contribution
to the error arising from using the estimated rather than the real communities.
We have seen (p.\ \pageref{error-term-from-c-vs-hat-c}), that, under assumption
\eqref{eqn:screening-egos-attributes-from-alters-behavior-2}, this term is just
$\gamma_1^T C_i - \gamma_0^T \hat{C}_i$. We will only need that covariance
conditional on $\hat{C}_i$ and $\hat{C}_j$.
\begin{lemma}
\label{lemma:covariance-between-altrs-behavior-and-estimation-errors}
Suppose that \eqref{eqn:structural-equation},
\eqref{eqn:screening-egos-attributes-from-alters-behavior-1} and
\eqref{eqn:screening-egos-attributes-from-alters-behavior-2} all hold. Then
\begin{equation}
\Cov{Y_{j,t}, (\gamma_1^T C_i - \gamma_0^T \hat{C}_i) | \hat{C}_i, \hat{C}_j}
= \sum_{j}{A_{ij}\gamma_1^T \Cov{C_i, C_j| \hat{C}_i, \hat{C}_j} \gamma_1} \label{eqn:covariance-between-altrs-behavior-and-estimation-errors}
\end{equation}
where by $\Cov{C_i, C_j}$ we mean the $d\times d$ matrix of coordinate-wise
covariances, which is fixed $\forall~C_i,C_j$ pairs.
\end{lemma}
\textsc{Proof:} See \S \ref{sec:proof-of-lemma:covariance-between-altrs-behavior-and-estimation-errors}. $\Box$
There are two individually sufficient (but not necessary) conditions for
\eqref{eqn:covariance-between-altrs-behavior-and-estimation-errors} to be zero:
\begin{enumerate}
\item $C_i \rotatebox{90}{\ensuremath{\models}} C_j | \hat{C_i},\hat{C_j}$, i.e., $C_i$ and $C_j$ are independent given their estimates,
\item $C_i = \hat{C_i}$ and $C_j = \hat{C_j}$, i.e., $C_i$ and $C_j$ are equal to their estimates.
\end{enumerate}
The second condition will generally not be true at any finite $n$. The first
condition is also very strong; it implies that $\hat{C}$ is a sufficient
statistic for $C$, since even learning the true location of node $i$, $C_i$,
would carry no information about the location of any other node not already
contained in $\hat{C}$. We are not aware of any estimates of latent node
locations in network models which have such a sufficiency property, and we
strongly suspect this is because they generally are {\em not}
sufficient\footnote{To get a sense of what would be entailed, suppose that
$A_{ij}=1$, and we knew we were dealing with a homophilous latent community
model. Then $\hat{C}$ would have to be so informative that even if an Oracle
told us $C_i$, our posterior distribution over $C_j$ would be unchanged.}.
We may, however, make further progress in the two specific settings of
latent communities and of continuous latent spaces.
\subsection{Control of Confounding with Latent Communities}
Under the conditions laid out in \S \ref{sec:communities}, we have that, with
probability tending to one as $n\rightarrow\infty$, $\hat{C} = C$. It follows
that {\em asymptotically} in $n$,
$\Cov{C_i, C_j| \hat{C}_i, \hat{C}_j} \rightarrow 0$. Ordinary least squares
estimation of \eqref{eqn:effective-model} will thus {\em in the limit} deliver
unbiased and consistent estimates of the social-influence parameter $\beta$.
\paragraph{Finite-sample bounds on the bias}
We can in fact go somewhat further, to bound the pre-asymptotic bias. Let $G$
be the indicator variable for the event that $\hat{C} = C$.
We know that
$\Expect{G} \geq 1- \delta(n)$, and that $\delta(n) = e^{-O(n)}$ from
\eqref{eqn:probability-of-any-errors-is-exp-small}.
Then we have the following:
\begin{lemma}
\label{lemma:covariance-for-block-models}
In the latent community setting,
\begin{eqnarray}
\label{eqn:final-covariance-for-block-models}
\Cov{C_i, C_j|\hat{C}_i, \hat{C}_j} & = & \delta(n) \left(\Cov{C_i, C_j|\hat{C}_i, \hat{C}_j, G=0} + \right.\\
\nonumber & & \left. (1-\delta(n))\tilde{C}_i \tilde{C}_j^T - \hat{C}_i\hat{C}_j^T - \hat{C}_i \tilde{C}_j^T - \tilde{C}_i \hat{C}_j^T \right)
\end{eqnarray}
where $\tilde{C}_i$ abbreviates $\Expect{C_i|\hat{C}_i, \hat{C}_j, G=0}$.
Thus, $\Cov{C_i, C_j|\hat{C}_i, \hat{C}_j} = O(\delta(n))$.
\end{lemma}
\textsc{Proof:} See \S \ref{sec:proof-of-lemma:covariance-for-block-models}.
$\Box$
The lemmas lead to a number of important conclusions.
\begin{theorem}
Suppose that the network forms according to a latent community model
satisfying the GMZZ conditions, and that
\eqref{eqn:screening-egos-attributes-from-alters-behavior-1}--\eqref{eqn:screening-egos-attributes-from-alters-behavior-2}
hold. Then estimating $\beta$ from \eqref{eqn:effective-model} provides
asymptotically unbiased and consistent estimates of the social-influence
coefficient in \eqref{eqn:structural-equation}, and the pre-asymptotic bias is
exponentially small in $n$.
\label{theorem:asymptotic-consistent-for-block-models}
\end{theorem}
\textsc{Proof:} We have just seen, in
\eqref{eqn:final-covariance-for-block-models}, that
$\Cov{C_i, C_j|\hat{C}_i, \hat{C}_j} = O(\delta(n))$. Under the theorem's
assumptions about the latent community model, we know that $\delta(n)$ can be
made exponentially small, and in only a polynomial cost in computational time
(\S \ref{sec:communities} above). Since, by standard arguments for linear
regression, the bias in $\hat{\beta}$ will be directly proportional to
$\gamma_1^T \Cov{C_i, C_j|\hat{C}_i, \hat{C}_j} \gamma_1$, the bias is itself
exponentially small in $n$. Hence $\hat{\beta}$ will be asymptotically
unbiased and consistent as $n\rightarrow\infty$. $\Box$
\begin{corollary}
Under the conditions of Theorem
\ref{theorem:asymptotic-consistent-for-block-models}, if $\gamma_1$ is known
or its magnitude is bounded, the pre-asymptotic bias can be bounded by the
value of a quadratic programming problem.
\end{corollary}
\textsc{Proof:} If $\gamma_1$ is known, or its magnitude can be bounded,
then even if $\tilde{C}_i$ is not known, we can bound the magnitude of the
bias in $\hat{\beta}$. From the proof of the theorem, the magnitude of the
bias is proportional to
\begin{equation}
\gamma_1^T \Cov{C_i, C_j|\hat{C}_i, \hat{C}_j} \gamma_1
\label{eqn:upper-bound-on-bias}
\end{equation}
which by \eqref{eqn:final-covariance-for-block-models} is a quadratic form in
$\tilde{C}_i$. Since $\tilde{C}_i$ is the expectation of an indicator vector,
it must lie in a $(k-1)$-dimensional simplex. Hence, maximizing
\eqref{eqn:upper-bound-on-bias} is maximizing a quadratic form under linear
constraints. This is a quadratic programming problem (and hence can be solved
in polynomial time, \citealt{Boyd-Vandenberghe-convex-optimization}),
and the value of the program will be the maximum of \eqref{eqn:upper-bound-on-bias}. $\Box$
In the more realistic situation where $\gamma_1$ is not vouchsafed to us by an
Oracle, the estimate of $\gamma_0$ is nonetheless asymptotically unbiased and
consistent, and so a feasible proxy would be to use it in the maximization
problem.
\subsection{Control of Confounding with Continuous Latent Space}
\label{sec:continuous-latent-space-setting}
Our treatment of the latent community setting relies on the fact that,
probability tending to one, the estimated communities match the actual
communities exactly, $\Prob{\hat{C} \neq C} \rightarrow 0$. This is not known
to happen for continuous latent space models, and seems very implausible for
estimates of continuous quantities.
As mentioned in \S \ref{sec:continuous-space}, \citet[ch.\ 3]{Asta-thesis} has
shown that if the link-probability function is known and has certain natural
regularity properties (detailed below), then the probability that the {\em sum}
of the distances between true locations and their maximum likelihood estimates
exceeds $\epsilon$ goes to zero exponentially in $\sqrt{\epsilon} n^2$ (at
least). More specifically, the result requires the link-probability function
to be smooth in the underlying metric and bounded on the logit scale, and
requires the latent space's group of isometries\footnote{An isometry is a
transformation of a metric space which preserves distances between points.
These transformations naturally form groups, and the properties of these
groups control, or encode, the geometry of the metric space
\citep{Brannan-Esplen-Gray}.} to have a bounded number of connected
components. (This is true for Euclidean spaces of any finite dimension, where
the bound is always $2$.) Then
\begin{equation}
\Prob{\sum_{i=1}^{n}{d(\hat{C}_i, C_i)} \geq \epsilon} \leq \mathcal{N}(n, \epsilon) e^{-\kappa n^2 \epsilon}
\label{eqn:asta-bound}
\end{equation}
where the $\mathcal{N}$ is a known function, polynomial in $n$ and
$1/\epsilon$, depending only on the isometry group of the metric, and $\kappa$
is a known constant, calculable from the isometry group and the bound on the
logit. Since the maximum of $n$ distances is at most the sum of those
distances, this further implies that
\begin{equation}
\Prob{\max_{i\in 1:n}{d(\hat{C}_i, C_i)} \leq \epsilon)} \geq 1 - \mathcal{N}(n, \epsilon) e^{-\kappa n^2 \epsilon}
\end{equation}
This is enough for the following asymptotic result.
\begin{theorem}
Assume that the network grows according to a continuous latent space model
satisfying the conditions of the previous paragraph, and that
\eqref{eqn:screening-egos-attributes-from-alters-behavior-1}--\eqref{eqn:screening-egos-attributes-from-alters-behavior-2}
hold. Then if $\hat{C}$ is estimated by maximum likelihood, $\hat{\beta}$ is
asymptotically unbiased and consistent; moreover, the pre-asymptotic bias is
polynomially small in $n$.
\end{theorem}
\textsc{Proof:} Fix a sequence $\epsilon_n > 0$ such that
$\epsilon_n \rightarrow 0$ as $n\rightarrow \infty$ while
$\epsilon_n n^2 \rightarrow \infty$. Let $G_n$ indicate the event when the
estimated locations are within $\epsilon_n$ of the true locations. Conditional
on $G_n=1$, $C_i$ is thus within a ball of radius $\epsilon_n$ around
$\hat{C}_i$, and so is $C_j$, so their conditional covariance is
$O(\epsilon_n^2)$. When $G_n=0$, we do not have a similar control of their
covariance, but such events are of low (exponentially-small) probability.
Indeed, if the $C_i$ are drawn iid from an arbitrary distribution, their
covariance conditional on $G_n=0$ is at most the the variance of that
distribution, and so $O(1)$ in $n$. Hence we obtain an over-all value for the
covariance of $O(\epsilon^2_n)$, which tends to zero by assumption. Since the
covariance between $Y_{j,t}$ and the noise term $\eta$ is tending to zero, OLS
delivers asymptotically unbiased and consistent estimates of $\beta$. $\Box$
We suspect that it is possible in principle not only to prove this exponential
rate of decay for a finite-$n$ bounds on the bias in continuous latent space
models, but also provide a solution for its precise computation, along the
lines done above for latent community models. However, our efforts suggest
that the bound does not lend itself to computation through a simple
optimization problem.
\subsection{Nonlinear Models}
\label{sec:nonlinear}
The arguments above go through almost unchanged if the structural and estimated
equations are not linear but merely additive. Writing the structural equation
as
\begin{equation}
Y_{i,t+1} = \alpha_0 + \alpha_1(Y_{i,t}) + \beta\left(\sum_{j}{\left(Y_{j,t}A_{ij}\right)}\right) + \gamma_1(C_i) + \gamma_2(X_{i}) + \epsilon ~,
\end{equation}
and the estimable model as
\begin{equation}
Y_{i,t+1} = \alpha_0 + \alpha_1(Y_{i,t}) + \beta\left(\sum_{j}{\left(Y_{j,t}A_{ij}\right)}\right) + \gamma_0(\hat{C}_i) + \eta ~,
\end{equation}
where the $\alpha$s, $\beta$s and $\gamma$s are now functions,
arguments parallel to those of \S \ref{sec:community-control} show that we
obtain asymptotically unbiased and consistent estimates of $\beta$ when
$Y_{j,t}$ is uncorrelated with $\eta$, which in turn requires that
$\Cov{\gamma_1(C_i), \gamma_2(C_j)|\hat{C}_i, \hat{C}_j} \rightarrow 0$.
Assuming smooth partial response functions, however, a Taylor series (``delta
method'') argument shows that this in turn amounts to $\Cov{C_i, C_j|\hat{C}_i,
\hat{C}_j} \rightarrow 0$.
We {\em suspect} that this pattern of argument carries over to fully non-linear
models with arbitrary interactions between regressors, provided the conditional
expectation function in the equivalent of \eqref{eqn:structural-equation} is
sufficiently smooth, but we leave this interesting and important topic to
future work.
\section{Discussion}
\label{sec:discussion}
What we have shown is that if a social network is generated either by a member
of a large class of latent community models or by a continuous latent space
model, and the pattern of influence over that network then follows an additive
model, it is possible to get consistent and {\em asymptotically} unbiased
estimates of the social-influence parameter by controlling for estimates of the
latent location of each node.
These are, to our knowledge, the first theoretical results which establish
conditions under which social influence can be estimated from non-experimental
data without confounding, even in the presence of latent homophily. Previous
suggestions for providing such estimates by means of controlling for lagged
observations \citep{Valente-diffusion-of-innovations}, matching
\citep{aral2009distinguishing} or instrumental variables which are also
associated with network location \citep{Tucker-network-externalities} are in
fact all invalid in the presence of {\em latent} homophily
\citep{Homophily-contagion-confounded}. An alternative to full identification
is to provide {\em partial identification}
\citep{Manski-identification-for-prediction}, i.e., bounds on the range of the
social-influence coefficient.
\citet{VanderWeele-sensitivity-analysis-for-contagion} provides such bounds
under extremely strong parametric assumptions\footnote{Among other things,
$C_i$ must be binary and it must not interact with anything.};
\citet{Ver-Steeg-Galstyan-ruling-out-latent-homophily,
Ver-Steeg-Galstyan-tests-for-contagion} provide non-parametric bounds, also
as the solution to an optimization problem, but must assume that each $Y(i,t)$
evolves as a homogeneous Markov process, i.e., that there is no aging in the
behavior of interest. None of these limitations apply to our approach.
Without meaning to diminish the value of our results, we feel it is also
important to be clear about their limitations. The following assumptions were
essential to our arguments:
\begin{enumerate}
\item The social network was generated {\em exactly} according to either a
latent community model or a continuous latent space model.
\item We knew whether it was a latent community model or a continuous latent
space model.
\item We knew either how many blocks there were\footnote{Or that the number of
blocks grows at an appropriate rate.}, or the latent space, its metric, and
its link-probability function.
\item Fixed attributes of the nodes relevant to the behavior were either {\em
fully} incorporated into the latent location, {\em or} stochastically
independent of the location.
\item All of the relevant conditional expectation functions are either linear
or additive.
\end{enumerate}
We suspect --- though we have no proofs --- that similar results will hold for
a somewhat wider class of well-behaved graphon network models\footnote{Graphon
estimation is an active topic of current research
\citep{Choi-Wolfe-consistency-of-co-clustering,
Wolfe-Olhede-nonparametric-graphon-estimation}, but it has focused on
estimating the link-probability function $w$, rather than the latent
locations $C$ (though see \citet{MEJN-Peixoto-generalized-communities} for a
purely-heuristic treatment).}, and for smooth conditional-expectation
functions quite generally. But we feel it important to emphasize that there
are many network processes which are perfectly well-behaved, and even very
natural, which fall outside the scope of our results; if, for instance, both
ties $A_{ij}$ and behaviors $Y_{i,t}$ are influenced by a latent variable $C_i$
which has {\em both} continuous and discrete coordinates, there is no currently
known way to consistently estimate the whole of $C_i$.
We must also re-iterate the point about only proving asymptotic lack of bias.
Even if all the other assumptions hold, the adversary can make the bias at any
finite $n$ as large as they like, by increasing the magnitude of $\gamma_1$.
This might seem implausible, but the scientific community knows little about
how big $\gamma_1$ might be in situations of {\em latent} homophily. Since
$\gamma_1$ must stay constant as $n$ increases, the adversary cannot {\em keep}
the bias large, and at a finite $n$ we have the ability to obtain a
probabilistic bound on the magnitude of the bias. Moreover, we have indicated
how (at least in the community setting) an estimate of $\gamma_1$ can be used
to obtain a direct bound the bias, but there is still a potentially serious
inferential problem here.
Despite these disclaimers, we wish to close by emphasizing the following point.
In general, the strength of social influence cannot be estimated from
observational social network data, because any feasible distribution over the
observables can be achieved in infinitely many ways that trade off influence
against latent homophily. What we have shown above is that {\em if} the
network forms according to either of two standard models, and the rest of our
assumptions hold, this result can be evaded, because the network itself makes
all the relevant parts of the latent homophilous attributes manifest. To the
best of our knowledge, this is the {\em first} situation in which the strength
of social influence can be consistent estimated in the face of latent homophily
--- the first, but we hope not the last.
\section{Selected Proofs}
\label{sec:proofs}
This section collects a number of lengthy, but in principle straightforward,
proofs of subsidiary results.
\subsection{Proof of Lemma \ref{lemma:covariance-between-altrs-behavior-and-estimation-errors}}
\label{sec:proof-of-lemma:covariance-between-altrs-behavior-and-estimation-errors}
First, recognize that
\begin{equation*}
\Cov{\sum_{j}{\left(Y_{j,t}A_{ij}\right)}, (\gamma_1^T C_i - \gamma_0^T \hat{C}_i) | \hat{C}_i, \hat{C}_j}
= \sum_{j}{A_{ij}\Cov{Y_{j,t}, (\gamma_1^T C_i - \gamma_0^T \hat{C}_i) | \hat{C}_i, \hat{C}_j}},
\end{equation*}
by the linearity of covariance and the fact that $A_{ij}$ is a constant. Therefore, we consider
\begin{eqnarray}
\lefteqn{\Cov{Y_{j,t}, (\gamma_1^T C_i - \gamma_0^T \hat{C}_i) | \hat{C}_i, \hat{C}_j}} & &\\
\nonumber & = & \Expect{Y_{j,t} (\gamma_1^T C_i - \gamma_0^T \hat{C}_i)|\hat{C}_i, \hat{C}_j } - \Expect{Y_{j,t}|\hat{C}_i, \hat{C}_j}\Expect{(\gamma_1^T C_i - \gamma_0^T \hat{C}_i)|\hat{C}_i, \hat{C}_j}\\
& = & \gamma_1^T \Expect{Y_{j,t} C_i|\hat{C}_i, \hat{C}_j} - \gamma_0^T \hat{C}_i\Expect{Y_{j,t}|\hat{C}_i, \hat{C}_j}\\
\nonumber & & - \Expect{Y_{j,t}|\hat{C}_i, \hat{C}_j} \gamma_1^T \Expect{C_i|\hat{C}_i, \hat{C}_j} + \Expect{Y_{j,t}|\hat{C}_i, \hat{C}_j} \gamma_0^T \hat{C}_i\\
& = & \gamma_1^T \left(\Expect{Y_{j,t} C_i|\hat{C}_i, \hat{C}_j} - \Expect{Y_{j,t}|\hat{C}_i, \hat{C}_j}\Expect{C_i|\hat{C}_i, \hat{C}_j}\right)\\
& = & \label{eqn:product-of-yj-and-ci} \gamma_1^T \left(\Expect{\Expect{Y_{j,t} C_i|\hat{C}_i, \hat{C}_j, C_i, C_j}|\hat{C}_i, \hat{C}_j} \right. \\
\nonumber & & \left. - \Expect{\Expect{Y_{j,t}|\hat{C}_i, \hat{C}_j, C_i, C_j}|\hat{C}_i, \hat{C}_j}\Expect{C_i|\hat{C}_i, \hat{C}_j} \right)\\
& = & \label{eqn:product-of-gamma-and-cj-and-ci} \gamma_1^T \left(\Expect{\gamma_1^T C_j C_i | \hat{C}_i, \hat{C}_j} - \Expect{\gamma_1^T C_j| \hat{C}_i, \hat{C}_j}\Expect{C_i|\hat{C}_i, \hat{C}_j}\right)\\
& = & \gamma_1^T \Cov{C_i, C_j| \hat{C}_i, \hat{C}_j} \gamma_1
\end{eqnarray}
where we use \eqref{eqn:structural-equation} to go from
\eqref{eqn:product-of-yj-and-ci} to \eqref{eqn:product-of-gamma-and-cj-and-ci}.
Finally,
\begin{equation*}
\Cov{\sum_{j}{\left(Y_{j,t}A_{ij}\right)}, (\gamma_1^T C_i - \gamma_0^T \hat{C}_i) | \hat{C}_i, \hat{C}_j}
= \sum_{j}{A_{ij}\gamma_1^T \Cov{C_i, C_j| \hat{C}_i, \hat{C}_j} \gamma_1},
\end{equation*}
$\Box$
\subsection{Proof of Lemma \ref{lemma:covariance-for-block-models}}
\label{sec:proof-of-lemma:covariance-for-block-models}
Applying the
conditional decomposition of covariance,
\begin{eqnarray}
\lefteqn{\Cov{C_i, C_j| \hat{C}_i, \hat{C}_j}} & & \\
\nonumber & = & \Expect{\Cov{C_i, C_j| \hat{C}_i, \hat{C}_j, G}} + \Cov{\Expect{C_i|\hat{C}_i, \hat{C}_j, G}, \Expect{C_j|\hat{C}_i, \hat{C}_j, G}}\\
& = & 0 + \delta(n) \Cov{C_i, C_j|\hat{C}_i, \hat{C}_j, G=0} \\
\nonumber & & + \Cov{\Expect{C_i|\hat{C}_i, \hat{C}_j, G}, \Expect{C_j|\hat{C}_i, \hat{C}_j, G}}
\end{eqnarray}
As for the second part of the covariance, start with the conditional
expectations:
\begin{equation}
\Expect{C_i|\hat{C}_i, \hat{C}_j, G} = G\hat{C}_i + (1-G)\Expect{C_i|\hat{C}_i, \hat{C}_j, G=0}
\end{equation}
and similarly for $C_j$. Abbreviate $\Expect{C_i|\hat{C}_i, \hat{C}_j, G=0}$ by
$\tilde{C}_i$ (leaving the dependence on $\hat{C}_i, \hat{C}_j$ implicit).
Then
\begin{eqnarray}
\Expect{C_i|\hat{C}_i, \hat{C}_j, G} & = & G\hat{C}_i + (1-G)\tilde{C}_i\\
& = & \tilde{C}_i + G(\hat{C}_i - \tilde{C}_i)
\end{eqnarray}
and, again, similarly for $C_j$.
Thus we can calculate the pieces of the covariance between conditional
expectations:
\begin{eqnarray}
\Expect{\Expect{C_i|\hat{C}_i, \hat{C}_j, G} | \hat{C}_i, \hat{C}_j}
& = & \tilde{C}_i + (\hat{C}_i - \tilde{C}_i) \Expect{G|\hat{C}_i, \hat{C}_j}\\
& = & \tilde{C}_i + (\hat{C}_i - \tilde{C}_i)(1-\delta(n))\\
& = & \hat{C}_i +\delta(n) \tilde{C}_i
\end{eqnarray}
so
\begin{equation}
\Expect{\Expect{C_i|\hat{C}_i, \hat{C}_j, G} | \hat{C}_i, \hat{C}_j} \Expect{\Expect{C_j|\hat{C}_i, \hat{C}_j, G} | \hat{C}_i, \hat{C}_j}^T = \hat{C}_i\hat{C}_j^T + \delta(n) (\hat{C}_i \tilde{C}_j^T + \tilde{C}_i \hat{C}_j^T) + \delta^2(n) \tilde{C}_i \tilde{C}_j^T
\end{equation}
Meanwhile,
\begin{eqnarray}
\lefteqn{\Expect{\Expect{C_i|\hat{C}_i, \hat{C}_j, G} \Expect{C_j|\hat{C}_i, \hat{C}_j, G}^T| \hat{C}_i, \hat{C}_j}} & & \\
\nonumber & = & \Expect{\left(G\hat{C}_i + (1-G)\Expect{C_i|\hat{C}_i, \hat{C}_j, G=0}\right)\left(G\hat{C}_j + (1-G)\Expect{C_j|\hat{C}_i, \hat{C}_j, G=0}\right)^T|\hat{C}_i, \hat{C}_j}\\
& = & \Expect{G^2 \hat{C}_i\hat{C}_j^T + (1-G)^2\tilde{C}_i\tilde{C}_j^T + G(1-G)(\hat{C}_i \tilde{C}_j^T + \tilde{C}_i\hat{C}_j^T) | \hat{C}_i, \hat{C}_j}
\end{eqnarray}
Since $G$ and $1-G$ are complementary indicator functions, $G^2=G$, $(1-G)^2=(1-G)$ and $G(1-G)=0$, so this simplifies to
\begin{eqnarray}
\lefteqn{\Expect{\Expect{C_i|\hat{C}_i, \hat{C}_j, G} \Expect{C_j|\hat{C}_i, \hat{C}_j, G}^T| \hat{C}_i, \hat{C}_j}} & & \\
\nonumber & = & \Expect{G \hat{C}_i\hat{C}_j^T + (1-G)\tilde{C}_i\tilde{C}_j^T | \hat{C}_i, \hat{C}_j}\\
& = & (1-\delta(n))\hat{C}_i\hat{C}_j^T + \delta(n) \tilde{C}_i \tilde{C}_j^T
\end{eqnarray}
Combining,
\begin{eqnarray}
\lefteqn{\Cov{\Expect{C_i|\hat{C}_i, \hat{C}_j, G}, \Expect{C_j|\hat{C}_i, \hat{C}_j, G}| \hat{C}_i, \hat{C}_j}} & & \\
\nonumber & = & (1-\delta)\hat{C}_i\hat{C}_j^T + \delta(n) \tilde{C}_i \tilde{C}_j^T - \hat{C}_i\hat{C}_j^T - \delta(n) (\hat{C}_i \tilde{C}_j^T + \tilde{C}_i \hat{C}_j^T) - \delta^2(n) \tilde{C}_i \tilde{C}_j^T\\
& = & \delta(n)((1-\delta(n))(\tilde{C}_i \tilde{C}_j^T - \hat{C}_i\hat{C}_j^T - \hat{C}_i \tilde{C}_j^T - \tilde{C}_i \hat{C}_j^T)
\end{eqnarray}
Finally, \eqref{eqn:final-covariance-for-block-models} follows. $\Box$
\subsection*{Acknowledgments}
We thank Andrew C. Thomas and David S. Choi for many valuable discussions on
these and related ideas over the years, and Dena Asta and Hannah Worrall for
sharing \citet{Asta-thesis} and \citet{Worrall-homophily}, respectively, and
Max Kaplan for related programming assistance. CRS was supported during this
work by grants from the NSF (DMS1207759 and DMS1418124) and the Institute for
New Economic Thinking (INO1400020).
|
\section{Introduction}
As a strong-weak duality, AdS/CFT correspondence \cite{Mal97}-\cite{Witten:1998qj} is very powerful in applications which use weakly coupled gravity to study strongly coupled field theory. However, this makes the non-trivial checks of this correspondence very hard since one needs to compute some quantities in the strong coupling limit of field theory to compare with results from the gravity side. Supersymmetric localization \cite{Pestun} and integrability \cite{Beisert:2010jr} are two very important tools to perform such field theoretical computations. These two approaches are complemented by each other. Localization can be utilized beyond the planar limit but the quantities which it can compute usually should be invariant under the supercharges on which the localization based. When the integrable structure exists, we can compute some quantities which are even non-supersymmetric. However, such theories are quite rare and integrable structure usually only appears in the large N limit. These two tools also permit us to compute certain quantities at the intermediate values of the coupling constant where neither perturbative gauge theory nor weakly coupled gravity is applicable.
Both four-dimensional ${\cal N}=4$ super Yang-Mills (SYM) theory and three-dimensional Aharony-Bergman-Jafferis-Maldacena (ABJM) theory \cite{ABJM} are integrable in the planar limit \cite{Minahan:2002ve}-\cite{c6}. It is very interesting to see how far one can go by reducing the supersymmetries of the original theory while keeping integrable structure at the same time. For four dimensional case, people have explored a lot through at least three approaches including marginal deformations \cite{Roiban:2003dw}-\cite{Mansson:2008xv}, orbifolding \cite{c4}-\cite{deLeeuw:2012hp} and adding flavors \cite{Chen:2004mu}-\cite{Erler:2005nr}. Excellent reviews on these results include \cite{Zoubos:2010kh, vanTongeren:2013gva}. However in three-dimensional case, similar exploration is limited. In \cite{He:2013hxd}, integrability of planar $\beta$- and $\gamma$-deformed ABJM theories were established at two-loop order in the scalar sector. The anomalous dimension matrices can be expressed as a Hamiltonian acting on an alternative spin chain\footnote{Notice that the $\gamma$-deformation studied in \cite{He:2013hxd} is different from the one in \cite{Imeroni:2008cr}. The integrability of the latter theory will be discussed in detail in \cite{chenliuwu}.}. The obtained Hamiltonians have identical form for these theories in the scalar sector, though the former theory has only one deformation parameter, while the latter has three. Comparing with the two-loop scalar-sector Hamiltonian from planar ABJM theory, now in each summand of the Hamiltonian for $\beta$-deformed ABJM theory, the next-to-nearest permutation term attains a certain phase depending on the charges of the three involved sites under two global $U(1)$'s which are used to perform the $\beta$-deformations. To obtain the needed transfer matrices, we need to deform the four R-matrices by similar phase factors to satisfy Yang-Baxter equations and produce the wanted Hamiltonian at the same time. This deformation is of Drinfeld-Reshetikhin form. A double scaling limit of $\gamma$-deformed ABJM theory was considered in \cite{CGK} which leads to an integrable theory of interacting fermions and scalars following four-dimensional consideration in \cite{Gurdogan:2015csr} (some subtleties of this limit were also studied in \cite{Sieg:2016vap}). This showed that integrable Chern-Simons-matter theories with less supersymmetry can have new interesting feature.
And as in four dimensional case \cite{Zoubos:2010kh}, in $\beta$/$\gamma$-deformed and orbifold ABJM theories, states with single magnon can be physical and detailed study on them may be simpler in many aspects than on the excited states in ABJM theory where at least two magnons are needed.
In this paper, we will focus on integrability of planar orbifold ABJM theories. Orbifolding is a widely used technique to obtain gauge theories from a parent one \cite{Kachru:1998ys, Bershadsky:1998mb, Bershadsky:1998cb}. It is carried out by starting with a discrete subgroup of the global symmetry group of the original theory. One can get various quiver gauge theories with less supersymmetry based on different discrete subgroups when the former one is supersymmetric. One of the advantages of the orbifolding operation
is that the obtained theories inherit some good properties of the parent theory. In this paper, the parent theory is the ABJM theory \cite{ABJM} which is the low energy effective theory on the worldvolume of N coincident M2-branes at the $\mathbb{C}^4/\mathbb{Z}_k$ orbifold singularity. The global (super)symmetry of ABJM theory is $OSp(6|4)\times U(1)_b$, direct product of a simple supergroup and a $U(1)$ factor.
This is distinct from the global symmetry of $\mathcal{N}=4$ SYM which is just a simple supergroup $PSU(2, 2|4)$. In \cite{c1}, two concrete quiver gauge theories with the residual $\mathcal{N}=4$ supersymmetries (non-chiral orbifold) and $\mathcal{N}=2$ supersymmetries (chiral orbifold) have been established through two different $\mathbb{Z}_n$ orbifoldings in ABJM field theory. Other orbifold ABJM theories are discussed in \cite{Terashima:2008ba, Imamura:2008ji, Berenstein}. In this paper, we only consider the case that $\Gamma$ is isomorphic to $\mathbb{Z}_n$.
We start with planar two-loop order and focus on the scalar sector which is closed at this order. We consider the generic case with $\Gamma<SU(4)_R\times U(1)_b$. The composite operators of the
orbifold theory can be expressed compactly using the fields in the parent ABJM theory with twist matrix inserted in the trace and with the projection condition imposed on the fields in the parent theory .
A straightforward computation shows that only two terms of the Hamiltonian were twisted by some phase factors whose precise forms depend on the charges of the involved sites under the action of $\Gamma$. To get transfer matrices which can produce this new Hamiltonian, we only need to insert certain constant matrices which act on the auxiliary spaces inside the traces. One can demonstrate that choosing the inserted matrices to be diagonal will make the RTT relations hold. By suitable choices of such matrices, we can produce the desired Hamiltonian. This completes the proof the integrability of general orbifold ABJM theories at planar two loop order in the scalar sector. Using algebraic Bethe ansatz, we find the Bethe ansatz equations (BAEs) in this sector at two-loop and give the constraints from the trace property and twist condition. The eigenvalues of the anomalous dimension matrix (ADM) are expressed using the Bethe roots.
Then we concentrate on the case with $\Gamma<SU(4)_R$ and generalize the above results to proposals for all-sector and all-loop order.
The leading-order all-sector results can be employed based on the prescription of Beisert-Roiban \cite{c2} after obtaining the charges for each simple root of the superalgebra and the vacuum. The all-order asymptotic results are obtained similarly based on all-loop asymptotic BAEs for planar ABJM theory \cite{c9}. As non-trivial consistency checks, we show that the BAEs we obtained satisfy both the fermionic duality and dynamic duality conditions. Finally we analyse the condition on the charges for the orbifolding to preserve $\mathcal{N}=2$ and $\mathcal{N}=4$ supersymmetries. We also confirm these results by using the fact that the orbifold ABJM theory is the low energy effective theory of $N$ membranes placed at the orbifold singularity $\mathbb{C}^4/(\Gamma\times \mathbb{Z}_{|\Gamma|k})$ \cite{Berenstein}.
The remaining part of this paper is organized as follows, in the next section and section~\ref{section3} we study in detail the integrability of orbifold ABJM theories in the scalar sector at two loop level. In sections~\ref{section4} and \ref{section5}, we will obtain the two-loop all-sector and all-loop results. Finally, we will discuss the supersymmetric orbifold theories. Some technical details will be put in three appendices.
\section{Two-loop Hamiltonian from orbifold ABJM theories}
As mentioned in the introduction, in this section, we will consider orbifold based on group $\Gamma(\simeq \mathbb{Z}_n)< SU(4)_R\times U(1)_b$ and focus on the scalar sector which is closed at two-loop order.
\subsection{Basic Ingredients of Orbifolding in Gauge Theory}
Now we will set up some necessary knowledge of orbifold gauge theory and our notation will follow that of \cite{c2,c3} closely.
We consider to perform orbifolding using discrete subgroup $\Gamma\simeq \mathbb{Z}_n$ of $SU(4)_R\times U(1)_b$ which means to start with ABJM theory with gauge group $U(nN)\times U(nN)$ and impose the following projection condition on gauge fields and the scalar fields
\begin{eqnarray}
\gamma(g) A \gamma^{-1}(g)&=&A,\\
\gamma(g)\hat{A} \gamma^{-1}(g)&=&\hat{A},\\
\gamma(g)\left(R(g)^I_J Y^J\right)\gamma^{-1}(g)&=&Y^I,
\end{eqnarray}
where $R(g)$ is a matrix representation of $\Gamma$ acting on the indices $I,J=1 \cdots 4$ of $Y^I$ and $\gamma(g)$ is acting on the color space with the color indices suppressed.
The projection condition on fermions is
\begin{equation} \gamma(g)\left(R^\prime(g)^I_J\bar{\psi}^J\right)\gamma^{-1}(g)=\bar{\psi}^I. \end{equation}
Notice that when $g=(g_1, g_2)\in SU(4)_R\times U(1)_b$, we have $R(g)=R(g_1)R(g_2)$ and $R^\prime(g)=R^\prime(g_1)R^\prime(g_2)=R(g_1)R(g_2^{-1})$, since $\bar{\psi}^I$ and $Y^I$
have opposite $U(1)_b$ charges.
The resulting theory is a quiver theory with gauge group $U(N)^{2n}$. If the element $g$ is the generator of $\mathbb{Z}_n$, the matrix representation $\gamma(g)$ will have the form
\begin{eqnarray}
\gamma(g)=\mbox{diag}\left(\mathbb{I}_{N\times N}, \omega \mathbb{I}_{N\times N},\cdots \omega^{n-1}\mathbb{I}_{N\times N}\right),\qquad \omega=e^{\frac{2\pi i}{n}}.
\end{eqnarray}
For the sake of simplicity, we also require that the field $Y^I$ has definite $\Gamma(<SU(4)_R\times U(1)_b)$ charge, then $R(g)$ will take the diagonal form $R(g)^I_J=\delta^I_J \omega^{s_I}$ and the constraint on the field $Y^I$ becomes
\begin{eqnarray}
Y^I=\omega^{s_I}\gamma Y^I \gamma^{-1}. \label{orbicon}
\end{eqnarray}
Here and the following by $\gamma$, we always mean $\gamma(g)$.
\par By orbifolding, the field $Y^I$ can be viewed as a $n \times n$ matrix with elements also being $N \times N$ matrices and only some components will survive due to the condition (\ref{orbicon}). Then the orbifold theory can be formulated in terms of those decomposed fields however the action turns out to be quite complicated \cite{c1,c4}. In our paper, we will use the field $Y^I$ in the parent theory and focus on the following single trace operators,
\begin{eqnarray}
\mbox{Tr} \left(\gamma^m Y^{I_1}Y^{\dagger}_{J_1}\cdots Y^{I_L} Y^{\dagger}_{J_L}\right),\quad m=0,1,\cdots n-1,\quad L\geq 2. \label{obasis}
\end{eqnarray}
Operators with the same $m$ constitute the $m$-th twisted sector and $m=0$ corresponds to the untwisted sector. If we move one $\gamma$ to pass all the fields behind and use the cyclic property of the trace to move it back, we find an overall phase factor appear. The composite operators will have the possibility to be non-vanishing only when this phase factor is trivial, hence lead to the twist constraint
\begin{eqnarray}
\frac{1}{n}\left(-\sum_{k}^Ls_{I_k}+\sum_{k}^Ls_{J_k}\right) \in \mathbb{Z}.
\end{eqnarray}
Furthermore, this local operator can also be seen as a closed alternating spin chain state
\begin{eqnarray}
|\mathcal{O}\rangle=|\gamma^m;I_1,\bar{J}_1,\cdots,I_{L},\bar{J}_L\rangle.
\end{eqnarray}
\subsection{Anomalous Dimension Matrix of Composite Operators in Twisted Sector}
We now find the anomalous dimensions for these gauge invariant scalar operators. An important fact is that the operators belonging to different twisted sectors do not mix with each other. Thus in the following discussions we will stay in a fixed $m$-th twisted sector. Before any further computations, let us recall that for parent ABJM theory, in the planar limit and at 2-loop order, the anomalous dimension matrix $\Gamma$ consists of local Hamiltonian of three adjacent sites \cite{c5,c6},
\begin{eqnarray}
\Gamma=\frac{\lambda^2}{2}\sum^{2L}_{i=1}\left(2-2P_{i,i+2}+P_{i,i+2}K_{i,i+1}+K_{i,i+1}P_{i,i+2}\right)=\frac{\lambda^2}{2}\sum^{2L}_{i=1} H_{i,i+1,i+2}.
\end{eqnarray}
where $\lambda={N}/{k}$ and $P$, $K$ are the permutation and trace operators acting on the tensor product of two vector spaces defined as
\begin{eqnarray}
P_{i_1,i_2}^{j_1,j_2}=\delta^{j_2}_{i_1}\delta^{j_1}_{i_2},\quad K_{i_1,j_2}^{j_1,i_2}=\delta^{i_2}_{i_1}\delta^{j_1}_{j_2}.
\end{eqnarray}
For the orbifold ABJM theories, the anomalous dimension matrix is obtained by expressing local interaction terms $\mathcal{H}$ of ABJM theory in the operator basis (\ref{obasis}). If $\gamma^m$ do not appear in the interaction region, we get the same local Hamiltonian as the parent ABJM theory,
\begin{eqnarray}
\mathcal{H}\circ Y^{I_i}Y^{\dagger}_{I_{i+1}}Y^{I_{i+2}}=\left(H_{i,i+1,i+2}\right)^{I_i,J_{i+1},I_{i+2}}_{J_{i},I_{i+1},J_{i+2}}Y^{J_i}Y^{\dagger}_{J_{i+1}}Y^{J_{i+2}}.
\end{eqnarray}
If $\gamma^m$ is present in the interaction region, we should move it away either to the left or to the right for the convenience of solving the problem. In our case, the non-trivial interactions only reside in the first and the last two sites of the spin chain and the modified local Hamiltonian are derived as follows. For the interactions among the $(2L-1)$-th,the $2L$-th and the 1st site, on one hand, we have,
\begin{eqnarray}
{\mathcal{H}}_{2L-1,2L,1}\circ Y^{I_{2L-1}}Y^{\dagger}_{I_{2L}}\gamma^m Y^{I_1}=({H}^{orbi}_{2L-1,2L,1})^{I_{2L-1},J_{2L},I_1}_{J_{2L-1},I_{2L},J_1}Y^{J_{2L-1}}Y^{\dagger}_{J_{2L}}\gamma^m Y^{J_1}.
\end{eqnarray}
where ${H}^{orbi}$ represents the orbifold Hamiltonian. On the other hand, we also have
\begin{eqnarray}
{\mathcal{H}}_{2L-1,2L,1}\circ Y^{I_{2L-1}}Y^{\dagger}_{I_{2L}}\gamma^m Y^{I_1}&=&\omega^{-ms_{I_1}}{\mathcal{H}}_{2L-1,2L,1}\circ Y^{I_{2L-1}}Y^{\dagger}_{I_{2L}} Y^{I_1} \gamma^m\\\nonumber
&=&\omega^{-ms_{I_1}}({H}_{2L-1,2L,1})^{I_{2L-1},J_{2L},I_1}_{J_{2L-1},I_{2L},J_1}Y^{J_{2L-1}}Y^{\dagger}_{J_{2L}} Y^{J_1} \gamma^m\\\nonumber
&=&\omega^{-ms_{I_1}+ms_{J_1}}({H}_{2L-1,2L,1})^{I_{2L-1},J_{2L},I_1}_{J_{2L-1},I_{2L},J_1}Y^{J_{2L-1}}Y^{\dagger}_{J_{2L}}\gamma^m Y^{J_1},
\end{eqnarray}
where we have used the relation $\gamma^m Y^I=\omega^{-ms_I}Y^I\gamma^m$ deduced from (\ref{orbicon}). Finally we get
\begin{eqnarray}
({H}^{orbi}_{2L-1,2L,1})^{I_{2L-1},J_{2L},I_1}_{J_{2L-1},I_{2L},J_1}=\omega^{-ms_{I_1}+ms_{J_1}}({H}_{2L-1,2L,1})^{I_{2L-1},J_{2L},I_1}_{J_{2L-1},I_{2L},J_1}. \label{tbc1}
\end{eqnarray}
Similarly, when acting on the $2L$-th ,the 1st and the 2nd sites, we shift the generator $\gamma^m$ to the left side and obtain
\begin{eqnarray}
({H}^{orbi}_{2L,1,2})^{J_{2L},I_{1},J_2}_{I_{2L},J_{1},I_2}=\omega^{-ms_{I_{2L}}+ms_{J_{2L}}}({H}_{2L,1,2})^{J_{2L},I_{1},J_2}_{I_{2L},J_{1},I_2}. \label{tbc2}
\end{eqnarray}
Therefore, the orbifold ABJM Hamiltonian reads
\begin{eqnarray}
H^{orbi}=\frac{\lambda^2}{2}\sum^{2L-2}_{i=1} H_{i,i+1,i+2}+\frac{\lambda^2}{2}\left({H}^{orbi}_{2L-1,2L,1}+{H}^{orbi}_{2L,1,2}\right). \label{twistbc}
\end{eqnarray}
\section{Algebraic Bethe Ansatz of the Orbifold ABJM Model}\label{section3}
The Hamiltonian derived above can be seen as a spin chain Hamiltonian with twisted boundary conditions. In this section we will give an explicit construction to show the integrability of this model and compute the eigenvalues of the Hamiltonian.
\subsection{Integrability of Orbifold ABJM Hamiltonian}
In order to demonstrate the integrability, the starting object is the R-matrix which satisfy the Yang-Baxter equation (YBE). For the orbifold ABJM theories, we use the same four R-matrices as those defined in the case of period spin chain \cite{c5,c6},
\begin{eqnarray}
R_{ab}(u)=u-P_{ab}:\qquad V_a\otimes V_b \rightarrow V_a\otimes V_b,\\
R_{\bar{a}\bar{b}}(u)=u-P_{\bar{a}\bar{b}}:\qquad V_{\bar{a}}\otimes V_{\bar{b}} \rightarrow V_{\bar{a}}\otimes V_{\bar{b}},\\
R_{a\bar{b}}(u)=u+K_{a\bar{b}}:\qquad V_a\otimes V_{\bar{b}} \rightarrow V_a\otimes V_{\bar{b}},\\
R_{\bar{a}b}(u)=u+K_{\bar{a}b}:\qquad V_{\bar{a}}\otimes V_b \rightarrow V_{\bar{a}}\otimes V_b.
\end{eqnarray}
where $V_i$ and $V_{\bar{i}}$ denote the fundamental and anti-fundamental representation space of $SU(4)$ respectively. The R-matrices satisfy the following six YBEs \cite{c5,c6},
\begin{eqnarray}
R_{ab}(u-v)R_{ac}(u)R_{bc}(v)&=&R_{bc}(v)R_{ac}(u)R_{ab}(u-v),\\
R_{ab}(u-v)R_{a\bar{c}}(u)R_{b\bar{c}}(v)&=&R_{b\bar{c}}(v)R_{a\bar{c}}(u)R_{ab}(u-v),\\
R_{\bar{a}\bar{b}}(u-v)R_{\bar{a}c}(u)R_{\bar{b}c}(v)&=&R_{\bar{b}c}(v)R_{\bar{a}c}(u)R_{\bar{a}\bar{b}}(u-v),\\
R_{\bar{a}\bar{b}}(u-v)R_{\bar{a}\bar{c}}(u)R_{\bar{b}\bar{c}}(v)&=&R_{\bar{b}\bar{c}}(v)R_{\bar{a}\bar{c}}(u)R_{\bar{a}\bar{b}}(u-v),\\
R_{a\bar{b}}(u-v-2)R_{ac}(u)R_{\bar{b}c}(v-2)&=&R_{\bar{b}c}(v-2)R_{ac}(u)R_{a\bar{b}}(u-v-2),\\
R_{a\bar{b}}(u-v-2)R_{a\bar{c}}(u-2)R_{\bar{b}\bar{c}}(v)&=&R_{\bar{b}\bar{c}}(v)R_{a\bar{c}}(u-2)R_{a\bar{b}}(u-v-2).
\end{eqnarray}
By the standard procedure, the next step is to construct the monodromy matrices using these R-matrices, we have
\begin{eqnarray}
T_0(u)=M_0 R_{01}(u)R_{0\bar{1}}(u-2)R_{02}(u)R_{0\bar{2}}(u-2)\cdots R_{0L}(u)R_{0\bar{L}}(u-2), \label{mono1}\\
T_{\bar{0}}(u)=\bar{M}_{\bar{0}} R_{{\bar{0}}1}(u-2)R_{{\bar{0}}\bar{1}}(u)R_{{\bar{0}}2}(u-2)R_{{\bar{0}}\bar{2}}(u)\cdots R_{{\bar{0}}L}(u-2)R_{{\bar{0}}\bar{L}}(u). \label{mono2}
\end{eqnarray}
where $0$ and $\bar{0}$ refer to auxiliary spaces in the SU(4) fundamental and anti-fundamental representations respectively. Comparing with the T-matrices for the periodic spin chain, we modify them by inserting two additional matrices $M$ and $\bar{M}$ in the auxiliary spaces $V_0$ and $V_{\bar{0}}$ so that they can generate the twisted boundary terms in equation (\ref{twistbc}) \cite{Zabrodin:2007rq}. The precise form of these two matrices will be determined later by demanding that the obtained Hamiltonian is the same as the one from the orbifold ABJM theories (up to an overall constant factor and shifting by term proportional to identity operator). Here we first show that when $M$ and $\bar{M}$ are diagonal, the obtained Hamiltonian is integrable. In this case it is easy to show that
\begin{eqnarray}
\left[R_{ab}(u), M_a M_b\right]=0,\\
\left[R_{a\bar{b}}(u), M_a \bar{M}_{\bar{b}}\right]=0,\\
\left[R_{\bar{a}\bar{b}}(u), \bar{M}_{\bar{a}}\bar{M}_{\bar{b}}\right]=0.
\end{eqnarray}
where the indices of $M$ and $\bar{M}$ denote on which site they act. Therefore we have the following important equations known as the RTT relations in the literature,
\begin{eqnarray}
R_{ab}(u-v)T_a(u)T_b(v)&=&T_b(v)T_a(u)R_{ab}(u-v), \label{fcr1}\\
R_{\bar{a}\bar{b}}(u-v)T_{\bar{a}}(u)T_{\bar{b}}(v)&=&T_{\bar{b}}(v)T_{\bar{a}}(u)R_{\bar{a}\bar{b}}(u-v),\\
R_{a\bar{b}}(u-v-2)T_a(u)T_{\bar{b}}(v)&=&T_{\bar{b}}(v)T_a(u)R_{a\bar{b}}(u-v-2).
\end{eqnarray}
By tracing over the auxiliary spaces of monodromy T-matrices, we obtain the transfer matrices
\begin{eqnarray}
\tau(u)=\mbox{Tr}_0 T_0(u),\qquad \bar{\tau}(u)=\mbox{Tr}_{\bar{0}} T_{\bar{0}}(u).
\end{eqnarray}
Then the above RTT relations lead to
\begin{eqnarray}
\left[\tau(u),\tau(v)\right]=0,\\
\left[\bar{\tau}(u),\bar{\tau}(v)\right]=0,\\
\left[\tau(u),\bar{\tau}(v)\right]=0.
\end{eqnarray}
for arbitrary $u$ and $v$. Expanding $\tau(u)$ and $\bar{\tau}(u)$ in terms of $u$, we find that the coefficients are mutually commuting and can be seen as the conserved charges. Of our interests is a certain combination of these conserved quantities given below because they will correspond to the Hamiltonians of our system,
\begin{eqnarray}
&&H_{1}=\tau(u)^{-1}\frac{d}{du}\tau(u)\bigg|_{u=0},\\
&&H_{2}=\bar{\tau}(u)^{-1}\frac{d}{du}\bar{\tau}(u)\bigg|_{u=0}.
\end{eqnarray}
After some computations ,we find
\begin{eqnarray}
H&=&H_1+H_2\\\nonumber
&=&\frac{1}{2}\sum_{i=1}^{2L-2}\left(-2-2P_{i,i+2}+P_{i,i+2}K_{i,i+1}+K_{i,i+1}P_{i,i+2}\right)\\\nonumber
&-&1-M^{-1}_1P_{2L-1,1}M_1+\frac{1}{2}K_{2L-1,2L}M^{-1}_1K_{2L,1}M_1+\frac{1}{2}M^{-1}_1K_{2L,1}M_1K_{2L-1,2L}\\\nonumber
&-&1-P_{2L,2}\bar{M}^{-1}_{2L}\bar{M}_2+\frac{1}{2}P_{2L,2}\bar{M}^{-1}_{2L}\bar{M}_2 K_{2,1} + \frac{1}{2}K_{2,1} P_{2L,2}\bar{M}^{-1}_{2L}\bar{M}_2. \label{totalH}
\end{eqnarray}
The details of the computations are put in the Appendix~\ref{appendixa}. We would like to know the component forms of the boundary terms of the above Hamiltonian and for this purpose we first clarify our convention for the matrix indices as follows\footnote{We only demonstrate this convention for case when
all indices are in the $\bf{4}$ representation. The convention for other cases is similar.},
\begin{eqnarray}
(AB)^{i_1,i_2}_{j_1,j_2}=(A)^{a,\,\,b}_{j_1,j_2}(B)^{i_1,i_2}_{a,\,\,b}.
\end{eqnarray}
Hence we have
\begin{eqnarray}
(M^{-1}_1 P_{1,2L-1} M_1)^{I_{2L-1},I_1}_{J_{2L-1},J_1}&=&(M^{-1}_1)^b_{J_1}(P_{1,2L-1})^{a,I_{2L-1}}_{b,J_{2L-1}}(M_1)^{I_1}_a\\\nonumber
&=&m_{I_1}\delta^{I_1}_a\cdot \delta^a_{J_{2L-1}}\delta^{I_{2L-1}}_b\cdot m^{-1}_{J_1}\delta^b_{J_1}\\\nonumber
&=&m_{I_1} m^{-1}_{J_1}(P_{1,2L-1})^{I_{2L-1},I_1}_{J_{2L-1},J_1}.
\end{eqnarray}
\begin{eqnarray}
(K_{2L-1,2L} M^{-1}_1 K_{2L,1} M_1)^{I_{2L-1},J_{2L},I_1}_{J_{2L-1},I_{2L},J_1}&=&(K_{2L-1,2L})^{J_{2L},I_{2L-1}}_{\,\,a,\,J_{2L-1}}(M^{-1})^c_{J_1}(K_{2L,1})^{b,a}_{c,I_{2L}}(M)_b^{I_1}\\\nonumber
&=&(K_{2L-1,2L})^{J_{2L},I_{2L-1}}_{\,\,a,\,J_{2L-1}}m_{I_1} m^{-1}_{J_1}\delta^a_{J_1}\delta^{I_1}_{I_{2L}}\\\nonumber
&=&m_{I_1} m^{-1}_{J_1}(K_{2L-1,2L} K_{2L,1} )^{I_{2L-1},J_{2L},I_1}_{J_{2L-1},I_{2L},J_1}.
\end{eqnarray}
\begin{eqnarray}
(P_{2L,2}\bar{M}^{-1}_{2L}\bar{M}_2)^{J_{2L},J_2}_{I_{2L},I_2}&=&(P_{2L,2})^{J_{2L},J_2}_{\,\,b,\,\,\,a}(\bar{M}^{-1}_{2L})^{b}_{I_{2L}}(\bar{M}_2)^{a}_{I_2} \\\nonumber
&=&\delta^{J_{2L}}_a\delta^{J_2}_b\bar{m}^{-1}_{I_{2L}}\delta^b_{I_{2L}}\bar{m}_{I_2}\delta^a_{I_2}\\\nonumber
&=&\bar{m}^{-1}_{I_{2L}}\bar{m}_{I_2}(P_{2L,2})^{J_{2L},J_2}_{I_{2L},I_2},
\end{eqnarray}
where $m_i$ and $\bar{m}_i$ $i=1,2,3,4$ are the diagonal elements of $M$ and $\bar{M}$. Comparing these results with the equations (\ref{tbc1}) and (\ref{tbc2}), one can fix the matrices $M$ and $\bar{M}$ as
\begin{eqnarray}
M&=&\mbox{diag}\,(\omega^{-ms_{1}},\omega^{-ms_{2}},\omega^{-ms_{3}},\omega^{-ms_{4}}),\\
\bar{M}&=&\mbox{diag}\,(\omega^{ms_{1}},\omega^{ms_{2}},\omega^{ms_{3}},\omega^{ms_{4}}).
\end{eqnarray}
So our conclusion is that by inserting the above two diagonal matrices into the monodromy matrices, we derived a Hamiltonian nearly the same as the one obtained in the field theory side only up to a shift of $3L \mathbb{I}$ and an overall factor $\lambda^2$ which do not affect the integrability of the model. This completes the proof of the integrability of planar orbifold ABJM theories in the scalar sector at the two-loop order.
\subsection{Eigenvalues of Spin Chain Hamiltonian and Bethe Ansatz Equations}
In this section we consider the diagonalisation of the corresponding transfer matrices. In the seminal paper \cite{c7}, the eigenstates of the Hamiltonian for a very general inhomogeneous spin chain with different spin on each site were constructed by means of the nested algebraic Bethe ansatz method. We find the related results can also apply to our alternating spin chain with twisted boundary conditions. However, here, we will use a much simpler method to obtain the Bethe ansatz equations.\footnote{Such treatment for $SU(N)$ spin chain can be found in the lecture notes by N.~Beisert \cite{c8}.} First we select the ground state as
\begin{eqnarray}
|\Omega\rangle=|\gamma^m; 1\bar{4}\cdots 1\bar{4}\rangle.
\end{eqnarray}
which corresponds to the chiral primary operator $\mbox{Tr} (\gamma^m(Y^1Y^{\dagger}_4)^L)$. Then we write the monodromy matrix as
\begin{eqnarray}
T_0=\left(
\begin{array}{cccc}
T_1&B_1&*&*\\C_1&T_2&B_2&*\\{*}&C_2&T_3&B_3\\{*}&{*}&C_3&T_4
\end{array}
\right)
\end{eqnarray}
For this selected vacuum, we find the three super-diagonal elements $B_1=T^1_2$,$B_2=T^2_3$,$B_3=T^3_4$ serve as the creation operators while the other three sub-diagonal ones $C_1=T^2_1$,$C_2=T^3_2$,$C_3=T^4_3$ as the annihilation operators. They also correspond to the simple roots of $SU(4)$ Lie algebra.
\par The excited states can be constructed by acting three kinds of creation operators on the vacuum state,
\begin{eqnarray}
\prod_{k=1}^{K_r}B_2(u_{2k})\prod_{j=1}^{K_u}B_1(u_{1j})\prod_{n=1}^{K_v}B_3(u_{3n})|\Omega\rangle, \label{exs}
\end{eqnarray}
where $u_{1j}=i u_j+1/2,u_{2k}=ir_k+1,u_{3n}=iv_n+3/2$ with $1\leq j\leq K_u, 1\leq k\leq K_r, 1\leq n\leq K_v$ are three sets of Bethe roots.
Then the eigenvalue of the transfer matrix $\tau(u)$ can be found by using the commutation relations between $T_i,i=1,2,3,4$ and $B_i,i=1,2,3$ originated from the eq.~(\ref{fcr1}) by throwing the unwanted terms,
\begin{eqnarray}
\Lambda(u)&=&\omega^{-ms_1}(u-1)^L(u-2)^L\prod_{j=1}^{K_u}\frac{u-i u_j+\frac{1}{2}}{u-i u_j-\frac{1}{2}}\\\nonumber
&+&\omega^{-ms_2}u^L(u-2)^L\prod_{j=1}^{K_u}\frac{u-i u_j-\frac{3}{2}}{u-i u_j-\frac{1}{2}}\prod_{k=1}^{K_r}\frac{u-i r_k}{u-i r_k-1}\\\nonumber
&+&\omega^{-ms_3}u^L(u-2)^L\prod_{n=1}^{K_v}\frac{u-i v_n-\frac{1}{2}}{u-i v_n-\frac{3}{2}}\prod_{k=1}^{K_r}\frac{u-i r_k-2}{u-i r_k-1}+\omega^{-ms_4}u^L(u-1)^L\prod_{n=1}^{K_v}\frac{u-i v_n-\frac{5}{2}}{u-i v_n-\frac{3}{2}}.
\end{eqnarray}
For the eigenvalue of $\bar{\tau}(u)$, it can be found from the conjugation condition $\bar{\Lambda}(u)=\Lambda(2-u^*)^*$ \cite{c5}
\begin{eqnarray}
\bar{\Lambda}(u)&=&\omega^{ms_1}u^L(u-1)^L\prod_{j=1}^{K_u}\frac{u-i u_j-\frac{5}{2}}{u-i u_j-\frac{3}{2}}\\\nonumber
&+&\omega^{ms_2}u^L(u-2)^L\prod_{j=1}^{K_u}\frac{u-i u_j-\frac{1}{2}}{u-i u_j-\frac{3}{2}}\prod_{k=1}^{K_r}\frac{u-i r_k-2}{u-i r_k-1}\\\nonumber
&+&\omega^{ms_3}u^L(u-2)^L\prod_{n=1}^{K_v}\frac{u-i v_n-\frac{3}{2}}{u-i v_n-\frac{1}{2}}\prod_{k=1}^{K_r}\frac{u-i r_k}{u-i r_k-1}+\omega^{ms_4}(u-2)^L(u-1)^L\prod_{n=1}^{K_v}\frac{u-i v_n+\frac{1}{2}}{u-i v_n-\frac{1}{2}}.
\end{eqnarray}
The Bethe ansatz equations (BAEs) can be readily obtained by demanding that the residue vanishes at each potential pole of $\Lambda(u)$,
\begin{eqnarray}
\omega^{-ms_1+ms_2}\left(\frac{u_j+i/2}{u_j-i/2}\right)^L=\prod_{k\neq j}^{K_u}\frac{u_j-u_k+i}{u_j-u_k-i}\prod_{k=1}^{K_r}\frac{u_j-r_k-i/2}{u_j-r_k+i/2},\label{BAE1}\\
\omega^{-ms_2+ms_3}=\prod_{k\neq j}^{K_r}\frac{r_j-r_k+i}{r_j-r_k-i}\prod_{k=1}^{K_u}\frac{r_j-u_k-i/2}{r_j-u_k+i/2}\prod_{k=1}^{K_v}\frac{r_j-v_k-i/2}{r_j-v_k+i/2},\label{BAE2}\\
\omega^{-ms_3+ms_4}\left(\frac{v_j+i/2}{v_j-i/2}\right)^L=\prod_{k\neq j}^{K_v}\frac{v_j-v_k+i}{v_j-v_k-i}\prod_{k=1}^{K_r}\frac{v_j-r_k-i/2}{v_j-r_k+i/2}.\label{BAE3}
\end{eqnarray}
The consistency of the theory guarantees that we could get the same sets of Bethe ansatz equations from $\bar{\Lambda}(u)$ instead, as one can easily check.
\par Now let us investigate the twist constraint for the excited state which is largely due to an implicit charge conservation condition. Note that the component of monodromy matrix T is
\begin{eqnarray}
&&\left(T_0(u)\right)^{b;i_1,j_2,\cdots,i_{2L-1},j_{2L}}_{a;j_1,i_2,\cdots,j_{2L-1},i_{2L}}\\\nonumber
&=&\left(M_0\right)^{c_1}_{a}\left(R_{01}(u)\right)^{c_2,i_1}_{c_1,j_1}\left(R_{02}(u-2)\right)^{c_3,j_2}_{c_2,i_2}\cdots\left(R_{0,2L-1}(u)\right)^{c_{2L},i_{2L-1}}_{c_{2L-1},j_{2L-1}}\left(R_{0,2L}(u-2)\right)^{b,j_{2L}}_{c_{2L},i_{2L}},
\end{eqnarray}
where $a,b,c_n,n=1,\cdots,2L$ represent the indices of the auxiliary space and $i_n,j_n,n=1,\cdots,2L$ are the indices of quantum spaces. If allocating each index $i$ of $V_i$ a phase $s_i$ and $i'$ of $\bar{V}_{i'}$ a phase $\bar{s}_{i'}$ with obvious relation $\bar{s}_{i'}=-s_{i'}$, we find the total phases are conserved under the action of three braiding operators $I,P$ and $K$,
\begin{eqnarray}
(I)^{k,l}_{i,j}=\delta^k_i\delta^l_j\quad &\rightarrow& \quad s_k+s_l=s_i+s_j,\\
(P)^{k,l}_{i,j}=\delta^k_j\delta^l_i\quad &\rightarrow& \quad s_k+s_l=s_j+s_i,\\
(K)^{k,j}_{i,l}=\delta^k_l\delta^j_i\quad &\rightarrow& \quad s_k+\bar{s}_l=s_k-s_l=0,\\\nonumber
&&\quad s_i+\bar{s}_j=s_i-s_j=0.
\end{eqnarray}
Since the building blocks of the monodromy matrix are R-matrices which entirely consists of these three operators and $M$ is diagonal, the whole process obey the phase conservation law,
\begin{eqnarray}
s_b+\sum_{k=1}^{2L}s_{i_k}=s_a+\sum_{k=1}^{2L}s_{j_{k}}.
\end{eqnarray}
So the net phase of the quantum space is $s_a-s_b$ under the action of $T^b_a$. Note that the phase of vacuum state is $L(s_1-s_4)$, then the phase of the excited state (\ref{exs}) become
\begin{equation}\label{twistconstraint}
K_u(s_2-s_1)+K_r(s_3-s_2)+K_v(s_4-s_3)+L(s_1-s_4).
\end{equation}
Therefore the twist constraint turn out to be
\begin{eqnarray}
\frac{1}{n}\left(K_u(s_2-s_1)+K_r(s_3-s_2)+K_v(s_4-s_3)+L(s_1-s_4)\right) \in \mathbb{Z}.
\end{eqnarray}
\par The shift operator and the corresponding total momentum are defined as
\begin{eqnarray}
\Pi=e^{2iP}=\frac{1}{2^{2L}}\tau(0)\bar{\tau}(0).
\end{eqnarray}
In the Appendix~\ref{appendixb} we will show that the shift operator acts trivially on physical state. Now given the eigenvalues above, we find
\begin{equation}\label{ZMC}
1=\frac{1}{2^{2L}}\Lambda(0)\bar{\Lambda}(0)=\omega^{m(s_4-s_1)}
\prod_{i=1}^{K_u}\frac{u_i+\frac{i}{2}}{u_i-\frac{i}{2}}\prod_{j=1}^{K_v}\frac{v_j+\frac{i}{2}}{v_j-\frac{i}{2}},
\end{equation}
which is the zero momentum condition for the twisted spin chain. As mentioned above, by a shift of $3L$ and then multiplied by $\lambda^2$, we find the energy of the spin chain which is dual to the anomalous dimension $\gamma$ of the orbifold ABJM theories,
\begin{eqnarray}
E=\lambda^2\left(3L+\frac{d}{du}\log (\Lambda(u)\bar{\Lambda}(u))|_{u=0}\right)=\lambda^2\left(\sum_{j=1}^{K_u}\frac{1}{u_j^2+\frac{1}{4}}+\sum_{j=1}^{K_v}\frac{1}{v_j^2+\frac{1}{4}}\right).
\end{eqnarray}
\section{Orbifold Bethe Ansatz}\label{section4}
Having obtained the orbifold Bethe equations for $SU(4)$ sector, now we go toward all-sector results. From now on, we will restrict to the case
with $\Gamma< SU(4)_R$.
The leading order\footnote{For Chern-Simons-matter theories, leading order means two loop level.} Bethe ansatz equations for ABJM theory read \cite{c5, c6},
\begin{eqnarray}\label{LOBAE}
\left(\frac{u_{j,k}-\frac{i}{2} V_j}{u_{j,k}+\frac{i}{2} V_j}\right)^L\mathop{\prod_{j'=1}^{J}\prod_{k'=1}^{K_{j'}}}_{(j',k')\neq(j,k)}\frac{u_{j,k}-u_{j',k'}+\frac{i}{2} M_{j,j'}}{u_{j,k}-u_{j',k'}-\frac{i}{2} M_{j,j'}}=1,\qquad
\prod_{j=1}^{J}\prod_{k=1}^{K_j}\frac{u_{j,k}+\frac{i}{2} V_j}{u_{j,k}-\frac{i}{2} V_j}=1.
\end{eqnarray}
where $J=5$
is the rank of the algebra $osp(6|4)$, $M_{j,j'}$ is the symmetric Cartan matrix and $V_j$ are the Dynkin labels which specify the representation of spin sites. The distinguished simple root system is
\begin{eqnarray}\label{DSTS}
\Delta^{0}=\{\delta_1-\delta_2,\delta_2-\epsilon_1,\epsilon_1-\epsilon_2,\epsilon_2-\epsilon_3,\epsilon_2+\epsilon_3\}.
\end{eqnarray}
we label the simple roots as $\alpha_1,\alpha_2,\alpha_3,\alpha_4,\alpha_{\bar4}$ in the given order above. For more details for the algebra $osp(6|4)$, see Appendix~\ref{appendixc}. As shown in Fig.~\ref{dynkin},
\begin{equation} V_j=(0, 0, 0, 1, 1). \end{equation}
These equations can be written in a compact form,
\begin{eqnarray} \label{BEA}
\mathop{\prod_{j'=0}^{J}\prod_{k'=1}^{K_j'}}_{(j',k')\neq(j,k)}S_{j,j'}(u_{j,k},u_{j',k'})=1,
\end{eqnarray}
with
\begin{eqnarray}
S_{j,j'}=\frac{u_{j,k}-u_{j',k'}+\frac{i}{2} M_{j,j'}}{u_{j,k}-u_{j',k'}-\frac{i}{2} M_{j,j'}},\qquad S_{j,0}=S_{0,j}^{-1}=\frac{u_{j,k}+\frac{i}{2} V_j}{u_{j,k}-\frac{i}{2} V_j},\qquad S_{0,0}=1, \qquad K_0=L.
\end{eqnarray}
\begin{figure}\label{DDynkin}
\centering
\includegraphics[width=8cm]{Dynkin.png}
\caption{The distinguished Dynkin diagram of the algebra $osp(6|4)$.}\label{dynkin}
\end{figure}
\subsection{Orbifolding the Bethe Ansatz}
The leading order orbifold Bethe ansatz equations has the general form \cite{c2} (in the twist $m$ sector for $\mathbb{Z}_n$ orbifold),
\begin{equation}\label{LOOBAE}
e^{2\pi imq_j/n}\left(\frac{u_{j,k}-\frac{i}{2} V_j}{u_{j,k}+\frac{i}{2} V_j}\right)^L
\mathop{\prod_{j'=1}^{J}\prod_{k'=1}^{K_j'}}_{(j',k'\neq(j,k))}
\frac{u_{j,k}-u_{j',k'}+\frac{i}{2} M_{j,j'}}{u_{j,k}-u_{j',k'}-\frac{i}{2} M_{j,j'}}=1,\\
\end{equation}
\begin{equation}\label{zmc}
e^{2\pi imq_0/n}\prod_{j'=1}^{J}\prod_{k'=1}^{K_{j'}}
\frac{u_{j',k'}+\frac{i}{2} V_{j'}}{u_{j',k'}-\frac{i}{2} V_{j'}}=1,\\
\end{equation}
\begin{equation}\label{TwistConstraint}
e^{-2\pi iLq_0/n}\prod_{j'=1}^{J}e^{-2\pi iK_{j'}q_{j'}/n}=1.
\end{equation}
Where the $q_j, j=1,2,3,4,\bar4$ are the $SU(4)$ charges of simple roots under orbifolding, and $q_0$ is the charge of vacuum. In the distinguished simple root system, the charges are,
\begin{equation}
\bm{q}=\(-t_2-t_3|0,-t_1,2t_1-t_2-t_3,-t_1+2t_2,-t_1+2t_3\).
\end{equation}
with $t_1,t_2,t_3$ integers. The $\bm{q}$ is related the charges $s_I$ by
\begin{equation}\label{sq}
s_I=\(t_2,t_1-t_2,-t_1+t_3,-t_3\), q_0=s_4-s_1, q_2=-s_1-s_2, q_3=s_2-s_3, q_4=s_1-s_2, q_{\bar4}=s_3-s_4.
\end{equation}
If we restrict to the scalar sector, one can recover the eqs.~(\ref{BAE1}-\ref{BAE3}) from (\ref{LOOBAE}), the eq.~(\ref{TwistConstraint}) is equivalent to the twist constraint (\ref{twistconstraint}) and eq.~(\ref{zmc}) is the zero momentum condition (\ref{ZMC}).
The energy is given by,
\begin{equation}
E=\lambda^2 \sum_{j=0}^{J}\sum_{k=1}^{K_j}\left(\frac{i}{u_{j,k}+\frac{i}{2} V_j}-\frac{i}{u_{j,k}-\frac{i}{2} V_j}\right).
\end{equation}
It is useful to represent the twist field $\gamma$ by a new type of quasi-excitation $j=-1$, with excited number $K_{-1}=m$. The phase shift
\begin{equation}
S_{j,-1}=1/S_{-1,j}=\exp(2\pi iq_j/n),\qquad j=0,\dots J,\qquad
S_{-1,-1}=1.
\end{equation}
Then the leading order Bethe equations for a $\mathbb{Z}_n$ orbifold theory can also be written in a compact form,
\begin{equation} \label{OBEA}
\mathop{\prod_{j'=-1}^{J}\prod_{k'=1}^{K_j'}}_{(j',k')\neq(j,k)}S_{j,j'}(u_{j,k},u_{j',k'})=1.
\end{equation}
We now consider a simple example in the $SU(2)\times SU(2)$ sector at two loops to verify our orbifold Bethe ansatz. The $SU(2)\times SU(2)$ sector is made of the elementary excitations $(Y^2|Y_3^{\dagger})$ on the even and odd sites above the vacuum $\text{Tr}((Y^1Y_4^{\dagger})^L)$, and it is closed at any order \cite{GGY, GHO}. At leading order of ABJM theory, the Hamiltonian reduces to the sum of two decoupled Heisenberg $XXX_{1/2}$ Hamiltonians, one acting on the even sites and the other acting on the odd sites\footnote{More precisely speaking, these two chains are only coupled by the zero momentum condition.}
\begin{equation}
H=\lambda^2\sum_{l=1}^{2L}(1-P_{l,l+2}).
\end{equation}
In orbifold case, the $l$-th term in the Hamiltonian is the same as above for $1\le l\le 2L-2$, and the $2L-1$-th term and the $2L$-th term are multiplied by the phases indicated in eqs.~(\ref{tbc1}) and (\ref{tbc2}), respectively. We consider two excitations above the ``twist vacuum'' $\text{Tr}(\gamma^m(Y^1Y_4^{\dagger})^L)$, one on the even sites and another on the odd sites. The obtained operators are $\text{Tr}(\gamma^m(Y^2Y_3^{\dagger})(Y^1Y_4^{\dagger})^{L-1})$ and the ones with permutations among even sites and odd sites independently.
For the above operator to be non-vanishing, the twist constraint
\begin{equation}
\frac{m[(L-1)(s_4-s_1)-s_2+s_3]}{n}\in\mathbb{Z},
\end{equation}
must be imposed.
For concreteness, we take $L=3$. In the basis,
\begin{eqnarray}
\mathcal{O}_1=\text{Tr}(\gamma^mY^2Y_3^{\dagger}Y^1Y_4^{\dagger}Y^1Y_4^{\dagger}), \mathcal{O}_2=\text{Tr}(\gamma^mY^2Y_4^{\dagger}Y^1Y_3^{\dagger}Y^1Y_4^{\dagger}),
\mathcal{O}_3=\text{Tr}(\gamma^mY^2Y_4^{\dagger}Y^1Y_4^{\dagger}Y^1Y_3^{\dagger}).
\end{eqnarray}
The Hamiltonian takes the form,
\begin{equation}
H=\lambda^2\matr{ccc}{
4&-(1+\omega^{-mq_0})&-\omega^{mq_{\bar4}}(1+\omega^{mq_0})\\
-(1+\omega^{mq_0})&4&-(1+\omega^{-mq_0})\\
-\omega^{-mq_{\bar4}}(1+\omega^{-mq_0})&-(1+\omega^{mq_0})&4
}.
\end{equation}
To write the Hamiltonian in a compact form, we have used the eq.~(\ref{sq}).
With the aid of {\it Mathematica}, it is easy to find the eigenvalues
\begin{equation}\label{E}
E=4\lambda^2[\text{sin}^2(\frac{m\pi q_{\bar4}}{3n}+\frac{k\pi}{3})+\text{sin}^2(\frac{m\pi (q_{\bar4}+3q_0)}{3n}+\frac{k\pi}{3})], \, k=0,1,2.
\end{equation}
Let's compute it using our orbifold Bethe ansatz equations. In the above simple case $L=3$, we have all excitation numbers to be zero except $K_4=K_{\bar4}=1$. The Bethe equations are simplified to be
\begin{eqnarray}
\(\frac{u+\frac{i}{2}}{u-\frac{i}{2}}\)^3=\text{e}^{2\pi imq_4/n},\label{u}\\
\(\frac{v+\frac{i}{2}}{v-\frac{i}{2}}\)^3=\text{e}^{2\pi imq_{\bar4}/n},\label{v}\\
\frac{u+\frac{i}{2}}{u-\frac{i}{2}}\frac{v+\frac{i}{2}}{v-\frac{i}{2}}=\text{e}^{-2\pi imq_0/n}.\label{uv}
\end{eqnarray}
However, these three equations are not independent if we impose the twist constraint
\begin{equation}
\frac{m(3q_0+q_4+q_{\bar4})}{n}\in\mathbb{Z}.
\end{equation}
The energy is given by
\begin{equation}\label{energy}
E=\lambda^2\(\frac{1}{u^2+\frac14}+\frac{1}{v^2+\frac14}\).
\end{equation}
The solutions of eqs. (\ref{u})-(\ref{uv}) are
\begin{equation} u=-\frac12\cot\left(\frac{m\pi(3q_0+q_{\bar4})}{3n}+\frac{k\pi}{3}\right), v=\frac12\cot\left(\frac{m\pi q_{\bar4}}{3n}+\frac{k \pi}{3}\right).
\end{equation}
Substituting this in eq.~(\ref{energy}) reproduces the result (\ref{E}) obtained by diagonalising the Hamiltonian directly.
\section{Higher Loops}\label{section5}
We want to generalize our orbifold Bethe equations to higher loops. Firstly, we know the all loop $AdS_4/CFT_3$ asymptotic Bethe equations read \cite{c9},
\begin{equation}\label{Allloop}
\begin{aligned}
1&=\prod_{j=1}^{K_2}\frac{u_{1,k}-u_{2,j}+\frac{i}{2}}{u_{1,k}-u_{2,j}-\frac{i}{2}}
\prod_{j=1}^{K_{4}}\frac{1-1/x_{1,k}x^+_{4,j}}{1-1/x_{1,k}x_{4,j}^-}
\prod_{j=1}^{K_{\bar 4}}\frac{1-1/x_{1,k} x^+_{\bar 4,j}}{1-1/x_{1,k}x_{\bar 4,j}^-} \,,\\
1&=\prod_{j=1,j\neq k}^{K_2}\frac{u_{2,k}-u_{2,j}-i }{u_{2,k}-u_{2,j}+i}
\prod_{j=1}^{K_1}\frac{u_{2,k}-u_{1,j}+\frac{i}{2}}{u_{2,k}-u_{1,j}-\frac{i}{2}}
\prod_{j=1}^{K_3}\frac{u_{2,k}-u_{3,j}+\frac{i}{2}}{u_{2,k}-u_{3,j}-\frac{i}{2}}\,,\\
1&=\prod_{j=1}^{K_2}\frac{u_{3,k}-u_{2,j}+\frac{i}{2}}{u_{3,k}-u_{2,j}-\frac{i}{2}}
\prod_{j=1}^{K_4}\frac{x_{3,k}-x^+_{4,j}}{x_{3,k}-x^-_{4,j}}
\prod_{j=1}^{K_{\bar4}}\frac{x_{3,k}-x^+_{\bar4,j}}{x_{3,k}-x^-_{\bar4,j}} \,,\\
\(\frac{x^+_{4,k}}{x^-_{4,k}}\)^{L}
&=\prod_{j=1,j\neq k}^{K_4}\frac{u_{4,k}-u_{4,j}+i}{u_{4,k}-u_{4,j}-i}\,
\prod_{j=1}^{K_1}\frac{1-1/x^-_{4,k}x_{1,j}}{1-1/x^+_{4,k} x_{1,j}}
\prod_{j=1}^{K_3}\frac{x^-_{4,k}-x_{3,j}}{x^+_{4,k}-x_{3,j}}\times\\
&\times\prod_{j=1,j\neq k}^{K_4}\sigma_{BES}(u_{4,k},u_{4,j})
\prod_{j=1}^{K_{\bar4}}\sigma_{BES}(u_{4,k},u_{\bar4,j})\,, \\
\(\frac{x^+_{\bar4,k}}{x^-_{\bar4,k}}\)^{L}
&=\prod_{j=1}^{K_{\bar4}}\frac{u_{\bar4,k}-u_{\bar4,j}+i}{u_{\bar4,k}-u_{\bar4,j}-i}\,
\prod_{j=1}^{K_1}\frac{1-1/x^-_{\bar4,k} x_{1,j}}{1-1/x^+_{\bar4,k} x_{1,j}}
\prod_{j=1}^{K_3}\frac{x^-_{\bar4,k}-x_{3,j} }{x^+_{\bar4,k}-x_{3,j}}\times\\
&\times\prod_{j=1,j\neq k}^{K_{\bar4}}\sigma_{BES}(u_{\bar4,k},u_{\bar4,j})
\prod_{j=1}^{K_{4}}\sigma_{BES}(u_{\bar4,k},u_{4,j})\,.
\end{aligned}
\end{equation}
where the $x^\pm$ are Zhukowski variables,
\begin{equation}\label{X}
x+\frac{1}{x}=\frac{u}{h(\lambda)}\;\;,\;\;x^\pm+\frac{1}{x^\pm}=\frac{1}{h(\lambda)}\(u\pm\frac{i}{2}\) \,.
\end{equation}
Note that the eqs.~(\ref{Allloop}) still have the form of eq.~(\ref{BEA}) except that one uses the rapidities $x_{j,k}$ instead of $u_{j,k}$ and the scattering phases $S_{j,j'}(x_{j,k},x_{j',k'})$ between the various Bethe roots are modified to accommodate the higher-loop interactions.\\
Unlike the leading order Bethe equations, not all simple root systems are possible for writing down higher loops Bethe ansatz equations. One of the possible Dynkin diagrams is shown in Fig.~\ref{Dynkin}.
\begin{figure}
\centering
\includegraphics[width=8cm]{Higher.png}\\
\caption{The ``higher" Dynkin diagram for $AdS_4/CFT_3$}\label{Dynkin}
\end{figure}
Another possible ``Higher'' Dynkin diagram is given by performing Weyl reflections with respect to the 1st and the 3rd simple roots in succession, and the result is the diagram on the right side in Fig.~\ref{twograding}. See Appendix~\ref{appendixc} for details. The two corresponding all loop Bethe equations are mapped to each other by ``fermionic duality'' which is consistent with odd Weyl reflection.
The all loop Bethe equations has another ``dynamic transformation'' symmetry which transform the Bethe roots of type 1 into type 3 and change the spin chain length \cite {c10}. We will prove these two dualities after we give the all-order Bethe ansatz equations. The study of fermionic duality makes sure that these two simple root systems do give the equivalent BAEs for orbifold ABJM theories and helps us to identify simple root systems which can be used at all loop level. The valid of dynamic duality admits the dynamical nature of the higher loop BAEs which takes into the fact that some operators with different length can mix with each other at higher loop level.
\subsection{The All loop Orbifold Bethe Equations and Dualites}
\subsubsection{The all loop Orbifold Bethe equations}
The Cartan matrix of two gradings $\eta=\pm1$ can be summarized as
\begin{equation}
M_{jj'}=\matr{ccccc}{
& +\eta& & & \\
+\eta& -2\eta& +\eta& & \\
& +\eta& & -\eta& -\eta \\
& & -\eta& 1+\eta& -1+\eta \\
& & -\eta& -1+\eta& 1+\eta
}
\end{equation}
\begin{figure}[H]
\centering
\subfigure{
\begin{minipage}{6cm}
\includegraphics[width=6cm]{Grading.png}
\end{minipage}
}
\subfigure{
\begin{minipage}{6cm}
\includegraphics[width=6cm]{Grading1.png}
\end{minipage}
}
\caption{\footnotesize{Two choice of Dynkin daigrams for higher loops Bethe equations.}}
\label{twograding}
\end{figure}
and the charges for $\eta=+1$,
\begin{equation}\label{qplus}
\bm{q}^+=(-t_2-t_3|t_1,0,t_1-t_2-t_3,-t_1+2t_2,-t_1+2t_3),
\end{equation}
for $\eta=-1$,
\begin{equation}\label{qminus}
\bm{q}^-=(-t_2-t_3|-t_1,2t_1-t_2-t_3,-t_1+t_2+t_3,t_2-t_3,-t_2+t_3).
\end{equation}
Now we give a bit details of the derivation of $\bm{q}^+$, the one of $\bm{q}^-$ is similar.
To do this we need the Cartan matrix for the Dynkin diagram in Fig.~\ref{Dynkin},
\begin{equation}
M_{jj'}=\matr{ccccc}{
& +1 & & & \\
+1& -2& +1 & & \\
& +1& & -1& -1 \\
& & -1& +2& \\
& & -1& & +2
}.\label{cartan2}
\end{equation}
Then we begin with the distinguished simple root system (\ref{DSTS}). First we apply $w_{\alpha_2}$ and give
\begin{equation}
w_{\alpha_2}(\Delta^0)=\{\delta_1-\epsilon_1,-\delta_2+\epsilon_1,\delta_2-\epsilon_2,\epsilon_2-\epsilon_3,\epsilon_2+\epsilon_3\},
\end{equation}
then apply another $w_{\alpha_1}$ with $\alpha_1$ being the first simple root $\delta_1-\epsilon_1$ in new basis
\begin{equation}
w_{\alpha_1}(w_{\alpha_2}(\Delta^0))=\{-\delta_1+\epsilon_1,\delta_1-\delta_2,\delta_2-\epsilon_2,\epsilon_2-\epsilon_3,\epsilon_2+\epsilon_3\}.
\end{equation}
Now we get the ``higher'' simple root system with the Dynkin diagram shown in the Fig.~\ref{Dynkin}.
The original three $SO(6)$ simple roots $\epsilon_1-\epsilon_2,\epsilon_2-\epsilon_3,\epsilon_2+\epsilon_3$ can be found in this basis as
\begin{equation}
\begin{aligned}
\epsilon_1-\epsilon_2=&\alpha_1+\alpha_2+\alpha_3,\\
\epsilon_2-\epsilon_3=&\alpha_4,\\
\epsilon_2+\epsilon_3=&\alpha_{\bar{4}}.
\end{aligned}
\end{equation}
Now, adding the first and two rows in eq.~(\ref{cartan2}) to the third one, and multiplying the obtained matrix from the right to $(0, 0, t_1, t_2, t_3)$, we get
\begin{equation} (q_1, q_2, q_3, q_4, q_{\bar{4}})=(t_1,0,t_1-t_2-t_3,-t_1+2t_2,-t_1+2t_3). \end{equation}
and the non-vanishing Dynkin labels are the same with the distinguished simple root system because we have merely dualized the first and the second simple root. Also because of this, $q_0$ does not change and we get eq.(\ref{qplus}).
From eqs.~(\ref{qplus}) and (\ref{qminus}), we observe that
\begin{equation}\label{ChargeRelation1}
q^{+\eta}_3=q^{+\eta}_1+\eta q^{+\eta}_0.
\end{equation}
\begin{equation}\label{ChargeRelation2}
q_4^++q_3^+=q_4^-,\, q_{\bar 4}^++q_3^+=q_{\bar 4}^-.
\end{equation}
\begin{equation}\label{ChargeRelation3}
q_0^{+\eta}-2q_3^+=q_2^+-q_2^-.
\end{equation}
\begin{equation}\label{ChargeRelation4}
q_1^++q_1^-=0.
\end{equation}
\begin{equation}\label{ChargeRelation5}
q_3^++q_3^-=0.
\end{equation}
The all loop $AdS_4/CFT_3$ orbifold Bethe equations read,
\begin{equation}\label{AllloopOrbifold}
\begin{aligned}
e^{-2\pi imq_1^{+\eta}/n}&=\prod_{j=1}^{K_2}\frac{u_{1,k}-u_{2,j}+\frac{i}{2}\eta}{u_{1,k}-u_{2,j}-\frac{i}{2}\eta}
\prod_{j=1}^{K_4}\frac{1-1/x_{1,k} x^{+\eta}_{4,j}}{1-1/x_{1,k}x_{4,j}^{-\eta}}
\prod_{j=1}^{K_{\bar 4}}\frac{1-1/x_{1,k} x^{+\eta}_{\bar 4,j}}{1-1/x_{1,k}x_{\bar 4,j}^{-\eta}}\,,\\
e^{-2\pi imq_2^{+\eta}/n}&=\prod_{j=1,j\neq k}^{K_2}\frac{u_{2,k}-u_{2,j}-i\eta}{u_{2,k}-u_{2,j}+i\eta}
\prod_{j=1}^{K_1}\frac{u_{2,k}-u_{1,j}+\frac{i}{2}\eta}{u_{2,k}-u_{1,j}-\frac{i}{2}\eta}
\prod_{j=1}^{K_3}\frac{u_{2,k}-u_{3,j}+\frac{i}{2}\eta}{u_{2,k}-u_{3,j}-\frac{i}{2}\eta}\,,\\
e^{-2\pi imq_3^{+\eta}/n}&=\prod_{j=1}^{K_2}\frac{u_{3,k}-u_{2,j}+\frac{i}{2}\eta}{u_{3,k}-u_{2,j}-\frac{i}{2}\eta}
\prod_{j=1}^{K_4}\frac{x_{3,k} -x^{+\eta}_{4,j}}{x_{3,k} -x^{-\eta}_{4,j}}
\prod_{j=1}^{K_{\bar 4}}\frac{x_{3,k} -x^{+\eta}_{\bar 4,j}}{x_{3,k} -x^{-\eta}_{\bar 4,j}} \,,\\
e^{-2\pi imq_4^{+\eta}/n}\(\frac{x^+_{4,k}}{x^-_{4,k}}\)^{L}&=
\prod_{j=1,j\neq k}^{K_4}\frac{x^+_{4,k}-x^-_{4,j}}{x^{-\eta}_{4,k}-x^{+\eta}_{4,j}}
\frac{1-1/x^+_{4,k}x^-_{4,j}}{1-1/x^-_{4,k}x^+_{4,j}}\sigma_{BES}(x_{ 4,k},x_{ 4,j})\,
\prod_{j=1}^{K_1}\frac{1-1/x^{-\eta}_{4,k} x_{1,j}}{1-1/x^{+\eta}_{4,k} x_{1,j}}\\
&\times\prod_{j=1}^{K_3}\frac{x^{-\eta}_{4,k}-x_{3,j} }{x^{+\eta}_{4,k}-x_{3,j}}
\prod_{j=1}^{K_{\bar 4}}\frac{x^{+\eta}_{4,k}-x^+_{\bar4,j}}{x^+_{4,k}-x^{+\eta}_{\bar 4,j}}
\sigma_{BES}(x_{4,k},x_{\bar 4,j}) \,,\\
e^{-2\pi imq_{\bar4}^{+\eta}/n}\(\frac{x^+_{\bar 4,k}}{x^-_{\bar 4,k}}\)^{L}
&=\prod_{j=1,j\neq k}^{K_{\bar 4}}
\frac{x^+_{\bar 4,k}-x^-_{\bar 4,j}}{x^{-\eta}_{\bar 4,k}-x^{+\eta}_{\bar 4,j}}
\frac{1-1/x^+_{\bar 4,k}x^-_{\bar 4,j}}{1-1/x^-_{\bar 4,k}x^+_{\bar 4,j}}
\sigma_{BES}(x_{\bar 4,k},x_{\bar 4,j})\,
\prod_{j=1}^{K_1}\frac{1-1/x^{-\eta}_{\bar 4,k} x_{1,j}}{1-1/x^{+\eta}_{\bar 4,k} x_{1,j}}\\
&\times\prod_{j=1}^{K_3}\frac{x^{-\eta}_{\bar 4,k}-x_{3,j} }{x^{+\eta}_{\bar 4,k}-x_{3,j}}
\prod_{j=1}^{K_{4}}\frac{x^{+\eta}_{\bar4,k}-x^+_{4,j}}{x^+_{\bar4,k}-x^{+\eta}_{4,j}}
\sigma_{BES}(x_{\bar4,k},x_{4,j}) \,.\\
\end{aligned}
\end{equation}
subject to zero momentum constraint,
\begin{equation}\label{Momentum}
e^{-2\pi imq_0^{+\eta}/n}=\prod_{j=1}^{K_4}\frac{x^+_{4,j}}{x^-_{4,j}}
\prod_{j=1}^{K_{\bar 4}}\frac{x^+_{\bar 4,j}}{x^-_{\bar 4,j}}\,,
\end{equation}
and the twist condition (\ref{TwistConstraint})
Here $\sigma_{BES}$ is the BES kernel for ABJM theory whose concrete expression can be found in \cite{c9}.
The spectrum of energy is
\begin{equation} E=\sum_{j=1}^{K_4}\frac12\left(\sqrt{1+16h(\lambda)^2\sin^2\frac{p_j}{2}}-1\right)
+\sum_{j=1}^{K_{\bar{4}}}\frac12\left(\sqrt{1+16h(\lambda)^2\sin^2\frac{\bar{p}_j}{2}}-1\right), \end{equation}
where\begin{equation} p_j=\frac{1}{i}\log\frac{x_{4, j}^+}{x_{4, j}^-}, \, \bar{p}_j=\frac{1}{i}\log\frac{x_{\bar{4}, j}^+}{x_{\bar{4}, j}^-}, \end{equation}
and $h(\lambda)$ is an interpolating function \cite{GGY, NT, GHO, Gromov} and it also replaces $\sqrt{\lambda}/(4\pi)$ appearing in the BES kernel for 4d $\mathcal{N}=4$ Super Yang-Mills theory.
We have the following relations,
\begin{equation}\label{Xu}
\begin{aligned}
u_k-u_j&=h(\lambda)(x_k-x_j)(1-1/x_kx_j)=h(\lambda)(x^\pm_k-x^\pm_j)(1-1/x^\pm_kx^\pm_j)\,,\\
u_k-u_j\pm\frac{i}{2}&=h(\lambda)(x^\pm_k-x_j)(1-1/x^\pm_kx_j)=h(\lambda)(x_k-x^\mp_j)(1-1/x_kx^\mp_j)\,,\\
u_k-u_j\pm i&=h(\lambda)(x^\pm_k-x^\mp_j)(1-1/x^\pm_kx^\mp_j)\,.
\end{aligned}
\end{equation}
They are easily confirmed using the definition (\ref{X}).
\subsubsection{Dynamic Duality}
The equation for $x_3$ is,
\begin{equation}
e^{-2\pi imq_3^{+\eta}/n}=\prod_{j=1}^{K_2}\frac{u_{3,k}-u_{2,j}+\frac{i}{2}\eta}{u_{3,k}-u_{2,j}-\frac{i}{2}\eta}
\prod_{j=1}^{K_4}\frac{x_{3,k} -x^{+\eta}_{4,j}}{x_{3,k} -x^{-\eta}_{4,j}}
\prod_{j=1}^{K_{\bar 4}}\frac{x_{3,k} -x^{+\eta}_{\bar 4,j}}{x_{3,k} -x^{-\eta}_{\bar 4,j}} \,.
\end{equation}
We now transform one type 3 root $1/x_{3,k}\rightarrow x_{1,k}$. For this transformation, $u_{3,k}\rightarrow u_{1,k}$, and one of the $x_3$ equations transforms as,
\begin{equation}
\prod_{j=1}^{K_2}\frac{u_{1,k}-u_{2,j}+\frac{i}{2}\eta}{u_{1,k}-u_{2,j}-\frac{i}{2}\eta}\prod_{j=1}^{K_4}\frac{x^{+\eta}_{4,j}}{x^{-\eta}_{4,j}}
\frac{1-1/x_{1,k} x^{+\eta}_{4,j}}{1-1/x_{1,k}x_{4,j}^{-\eta}}\prod_{j=1}^{K_{\bar 4}}
\frac{x^{+\eta}_{\bar 4,j}}{x^{-\eta}_{\bar 4,j}}
\frac{1-1/x_{1,k} x^{+\eta}_{\bar 4,j}}{1-1/x_{1,k}x_{\bar 4,j}^{-\eta}}=e^{-2\pi imq_3^{+\eta}/n}\,.
\end{equation}
Using the momentum condition (\ref{Momentum}), we find
\begin{equation}
\prod_{j=1}^{K_2}\frac{u_{1,k}-u_{2,j}+\frac{i}{2}\eta}{u_{1,k}-u_{2,j}-\frac{i}{2}\eta}
\prod_{j=1}^{K_4}\frac{1-1/x_{1,k} x^{+\eta}_{4,j}}{1-1/x_{1,k}x_{4,j}^{-\eta}}
\prod_{j=1}^{K_{\bar 4}}\frac{1-1/x_{1,k} x^{+\eta}_{\bar 4,j}}{1-1/x_{1,k}x_{\bar 4,j}^{-\eta}}
=e^{-2\pi imq_3^{+\eta}/n+2\pi im\eta q_0^{+\eta}/n}=e^{-2\pi imq_1^{+\eta}/n}\,.
\end{equation}
where the relation(\ref{ChargeRelation1}) has been used.
We recognize that this is the equation for $x_{1}$ with the same grading.
Under this transformation, the scattering phases in $x_4$ and $x_{\bar 4}$ equations also get changed. For example,
\begin{equation}
\frac{x^{-\eta}_{4,k}-x_{3,j} }{x^{+\eta}_{4,k}-x_{3,j}}\rightarrow \frac{x^{-\eta}_{4,k}}{x^{+\eta}_{4,k}}
\frac{1-1/x^{-\eta}_{4,k} x_{1,j}}{1-1/x^{+\eta}_{4,k} x_{1,j}}\,,
\end{equation}
and the equation for $x_4$ transforms as,
\begin{equation}
\begin{aligned}
e^{-2\pi imq_4^{+\eta}/n}\(\frac{x^+_{4,k}}{x^-_{4,k}}\)^{L+\eta}
&=\prod_{j=1,j\neq k}^{K_4}\frac{x^+_{4,k}-x^-_{4,j}}{x^{-\eta}_{4,k}-x^{+\eta}_{4,j}}
\frac{1-1/x^+_{4,k}x^-_{4,j}}{1-1/x^-_{4,k}x^+_{4,j}}
\sigma_{BES}(x_{ 4,k},x_{ 4,j})\,
\prod_{j=1}^{K_1+1}\frac{1-1/x^{-\eta}_{4,k} x_{1,j}}{1-1/x^{+\eta}_{4,k} x_{1,j}}\\
&\times\prod_{j=1}^{K_3-1}\frac{x^{-\eta}_{4,k}-x_{3,j} }{x^{+\eta}_{4,k}-x_{3,j}}
\prod_{j=1}^{K_{\bar 4}}\frac{x^{+\eta}_{4,k}-x_{ \bar 4,j}}{x^+_{4,k}-x^{+\eta}_{\bar 4,j}}
\sigma_{BES}(x_{4,k},x_{\bar 4,j}) \,,\\
\end{aligned}
\end{equation}
thus with this transformation in addition to the following replacements which is called dynamic transformation,
\begin{equation}
K_3\rightarrow K_3-1, K_1\rightarrow K_1+1, L\rightarrow L+\eta\,.
\end{equation}
the all loop Bethe eqs.~(\ref{Allloop}) remain invariant with the same grading. The momentum conservation condition and the expression for the total energy are not changed under the dynamic duality.
This dynamic duality is closely related to properties of the all-loop S-matrix. We only demonstrate this for grading $\eta=1$.
In fact, for the ABJM case, the needed property is
\begin{equation} \label{Phase}
S_{j,3}(x,x_3)=S_{j,1}(x,x_1)S_{j,0}(x)\qquad\text{for}\;j=1,2,3,4,\bar{4}.
\end{equation}
for $x_1x_3=1$.
This can be checked directly. We only give the proof for the case with $j=4,\bar{4}$, other cases are trivial.
\begin{equation}
\begin{aligned}
S_{4,3}=\frac{x_4^--x_3}{x_4^+-x_3},\qquad S_{4,1}=\frac{1-1/x_4^-x_1}{1-1/x_4^+x_1},\qquad S_{4,0}=\frac{x_4^-}{x_4^+}.\\
\end{aligned}
\end{equation}
Using $x_3x_1=1$, we find,
\begin{equation}
\frac{x_4^--1/x_1}{x_4^+-1/x_1}=\frac{x_4^-}{x_4^+}\frac{1-1/x_4^-x_1}{1-1/x_4^+x_1}\,.
\end{equation}
The case for $j=\bar{4}$ is similar.
For the orbifold theories, we need the relation(\ref{Phase}) holds for $j=-1$ as well. This is the case because $q_3^+=t_1-t_2-t_3=q_1^++q_0^+$.
\subsubsection{The Fermionic Duality}
We now prove that two choices of grading in (\ref{AllloopOrbifold}) are equivalent based on some fermionic duality. In order to investigate this duality of the eqs.(\ref{AllloopOrbifold}), we rewrite the equation of $x_3$ for $\eta=+1$ as,
\begin{equation}\label{X3}
\prod_{j=1}^{K_2}\frac{x_{3,k}-x^-_{2,j}}{x_{3,k}-x^+_{2,j}}
\prod_{j=1}^{K_2}\frac{x_{3,k}-1/x^-_{2,j}}{x_{3,k}-1/x^+_{2,j}}
\prod_{j=1}^{K_4}\frac{x_{3,k} -x^{+}_{4,j}}{x_{3,k} -x^{-}_{4,j}}
\prod_{j=1}^{K_{\bar 4}}\frac{x_{3,k} -x^{+}_{\bar 4,j}}{x_{3,k} -x^{-}_{\bar 4,j}}
=e^{-2\pi imq_3^{+\eta}/n}\,.
\end{equation}
We further introduce the following polynomial $P(x)$,
\begin{equation}\label{P}
\begin{aligned}
P(x)&=e^{2\pi imq_3^{+}/n}\prod_{j=1}^{K_2}(x-x^-_{2,j})
\prod_{j=1}^{K_2}(x-1/x^-_{2,j})
\prod_{j=1}^{K_4}(x -x^{+}_{4,j})
\prod_{j=1}^{K_{\bar 4}}(x-x^{+}_{\bar 4,j})\\
&-\prod_{j=1}^{K_2}(x-x^+_{2,j})
\prod_{j=1}^{K_2}(x-1/x^+_{2,j})
\prod_{j=1}^{K_4}(x -x^{-}_{4,j})
\prod_{j=1}^{K_{\bar 4}}(x-x^{-}_{\bar 4,j})\,.
\end{aligned}
\end{equation}
Obviously there already exists $K_3+K_1$ roots of $P(x)$,
\begin{equation}
P(x_{3,k})=0,k=1,\dots, K_3,\qquad P(1/x_{1,k})=0, k=1, \dots, K_1\,.
\end{equation}
The remaining solutions can also be grouped into two classes, type 3 roots and type 1 roots.
\begin{equation}\label{P1}
P(x)\sim \prod_{j=1}^{K_3}(x-x_{3,j})\prod_{j=1}^{K_1}(x-1/x_{1,j})\prod_{j=1}^{\tilde{K}_3}(x-\tilde{x}_{3,j})\prod_{j=1}^{\tilde{K}_1}(x-1/\tilde{x}_{1,j})\,,
\end{equation}
where
\begin{equation}\label{Tilde}
\tilde{K}_3=K_2+K_4+K_{\bar4}-K_3,\qquad \tilde{K}_1=K_2-K_1\,.
\end{equation}
We now calculate $P(x^\pm_{4,k})$ using two equivalent expressions of $P(x)$,
\begin{equation}
\begin{aligned}
\frac{P(x^+_{4,k})}{P(x^-_{4,k})}&=
e^{-2\pi imq_3^{+}/n}
\prod_{j=1,j\neq k}^{K_4}\frac{x^+_{4,k}-x^-_{4,j}}{x^-_{4,k}-x^+_{4,j}}
\prod_{j=1 }^{K_{\bar4}}\frac{x^+_{4,k}-x^-_{\bar4,j}}{x^-_{4,k}-x^+_{\bar4,j}}
\prod_{j=1}^{K_2}\frac{x^+_{4,k}-x^+_{2,j}}{x^-_{4,k}-x^-_{2,j}}
\prod_{j=1}^{K_2}\frac{x^+_{4,k}-1/x^+_{2,j}}{x^-_{4,k}-1/x^-_{2,j}}\\
&=\prod_{j=1}^{K_3}\frac{x^+_{4,k}-x_{3,j}}{x^-_{4,k}-x_{3,j}}
\prod_{j=1}^{K_1}\frac{x^+_{4,k}-1/x_{1,j}}{x^-_{4,k}-1/x_{1,j}}
\prod_{j=1}^{\tilde{K}_3}\frac{x^+_{4,k}-\tilde{x}_{3,j}}{x^-_{4,k}-\tilde{x}_{3,j}}
\prod_{j=1}^{\tilde{K}_1}\frac{x^+_{4,k}-1/\tilde{x}_{1,j}}{x^-_{4,k}-1/\tilde{x}_{1,j}}\,.
\end{aligned}
\end{equation}
Using the relations (\ref{Xu}) and (\ref{ChargeRelation2}), we find,
\begin{equation}
\begin{aligned}
&e^{-2\pi imq_3^{+}/n}\prod_{j=1,j\neq k}^{K_4}\frac{x^+_{4,k}-x^-_{4,j}}{x^-_{4,k}-x^+_{4,j}}
\prod_{j=1}^{K_{\bar4}}\frac{x^+_{4,k}-x^-_{\bar4,j}}{x^-_{4,k}-x^+_{\bar4,j}}
\prod_{j=1}^{K_2}\frac{x^+_{4,j}}{x^-_{4,j}}
=\prod_{j=1}^{K_1+\tilde{K}_1}\frac{x^+_{4,j}}{x^-_{4,j}}
\prod_{j=1}^{K_3}\frac{x^+_{4,k}-x_{3,j}}{x^-_{4,k}-x_{3,j}}\\
&\times \prod_{j=1}^{\tilde{K}_3}\frac{x^+_{4,k}-\tilde{x}_{3,j}}{x^-_{4,k}-\tilde{x}_{3,j}}
\prod_{j=1}^{K_1}\frac{1-1/x_{1,j}x^+_{4,k}}{1-1/x_{1,j}x^-_{4,k}}
\prod_{j=1}^{\tilde{K}_1}\frac{1-1/\tilde{x}_{1,j}x^+_{4,k}}{1-1/\tilde{x}_{1,j}x^-_{4,k}}\,.
\end{aligned}
\end{equation}
Using the relation~(\ref{Tilde}), we arrive at,
\begin{equation}
\begin{aligned}
&e^{-2\pi imq_3^{+}/n}\prod_{j=1,j\neq k}^{K_4}\frac{x^+_{4,k}-x^-_{4,j}}{x^-_{4,k}-x^+_{4,j}}
\prod_{j=1}^{K_3}\frac{x^-_{4,k}-x_{3,j}}{x^+_{4,k}-x_{3,j}}
\prod_{j=1}^{K_1}\frac{1-1/x_{1,j}x^-_{4,k}}{1-1/x_{1,j}x^+_{4,k}}
=\prod_{j=1}^{\tilde{K}_3}\frac{x^+_{4,k}-\tilde{x}_{3,j}}{x^-_{4,k}-\tilde{x}_{3,j}}\\
&\times\prod_{j=1}^{\tilde{K}_1}\frac{1-1/\tilde{x}_{1,j}x^+_{4,k}}{1-1/\tilde{x}_{1,j}x^-_{4,k}}
\prod_{j=1 }^{K_{\bar4}}\frac{x^-_{4,k}-x^+_{\bar4,j}}{x^+_{4,k}-x^-_{\bar4,j}}\,.
\end{aligned}
\end{equation}
Thus the equation for $x_4$ in the grading $\eta=+1$ is equivalent to one in the grading $\eta=-1$ and similar calculations can be done to show the equivalence of two gradings for $x_{\bar 4}$ equation.
It still remains to prove the equivalence for other equations. For this purpose, we calculate the combination $P(x^{\pm}_{2,k})P(1/x^{\pm}_{2,k})$in two ways,
\begin{equation}
\begin{aligned}
&\frac{P(x^+_{2,k})}{P(x^-_{2,k})}\frac{P(1/x^+_{2,k})}{P(1/x^-_{2,k})}
=e^{4\pi imq_3^{+}/n}\prod_{j=1,j\neq k}^{K_2}\frac{x^+_{2,k}-x^-_{2,j}}{x^-_{2,k}-x^+_{2,j}}
\prod_{j=1,j\neq k}^{K_2}\frac{x^+_{2,k}-1/x^-_{2,j}}{x^-_{2,k}-1/x^+_{2,j}}
e^{-2\pi imq_0^{+\eta}/n}\\
&\times\prod_{j=1}^{K_2}\frac{1/x^+_{2,k}-x^-_{2,j}}{1/x^-_{2,k}-x^+_{2,j}}
\prod_{j=1,j\neq k}^{K_2}\frac{1/x^+_{2,k}-1/x^-_{2,j}}{1/x^-_{2,k}-1/x^+_{2,j}}
=e^{-2\pi im(q_0^{+\eta}-2q_3^+)/n}\(\prod_{j=1,j\neq k}^{K_2}\frac{u_{2,k}-u_{2,j}+i}{u_{2,k}-u_{2,j}-i}\)^2\\
&=\prod_{j=1}^{K_3}\frac{x^+_{2,k}-x_{3,j}}{x^-_{2,k}-x_{3,j}}\frac{1/x^+_{2,k}-x_{3,j}}{1/x^-_{2,k}-x_{3,j}}
\prod_{j=1}^{K_1}\frac{x^+_{2,k}-1/x_{1,j}}{x^-_{2,k}-1/x_{1,j}}\frac{1/x^+_{2,k}-1/x_{1,j}}{1/x^-_{2,k}-1/x_{1,j}}\\
&\times\prod_{j=1}^{\tilde{K}_3}\frac{x^+_{2,k}-\tilde{x}_{3,j}}{x^-_{2,k}-\tilde{x}_{3,j}}\frac{1/x^+_{2,k}-\tilde{x}_{3,j}}{1/x^-_{2,k}-\tilde{x}_{3,j}}
\prod_{j=1}^{\tilde{K}_1}\frac{x^+_{2,k}-1/\tilde{x}_{1,j}}{x^-_{2,k}-1/\tilde{x}_{1,j}}\frac{1/x^+_{2,k}-1/\tilde{x}_{1,j}}{1/x^-_{2,k}-1/\tilde{x}_{1,j}}\,.\\
\end{aligned}
\end{equation}
Using the relations (\ref{Xu}) and(\ref{ChargeRelation3}), we can rewrite the above equation as,
\begin{equation}
\begin{aligned}
&e^{2\pi imq_2^{+}/n}\prod_{j=1,j\neq k}^{K_2}\frac{u_{2,k}-u_{2,j}-i}{u_{2,k}-u_{2,j}+i}
\prod_{j=1}^{K_1}\frac{u_{2,k}-u_{1,j}+\frac{i}{2}}{u_{2,k}-u_{1,j}-\frac{i}{2}}
\prod_{j=1}^{K_3}\frac{u_{2,k}-u_{3,j}+\frac{i}{2}}{u_{2,k}-u_{3,j}-\frac{i}{2}}\\
&=e^{2\pi imq_2^{-}/n}\prod_{j=1,j\neq k}^{K_2}\frac{u_{2,k}-u_{2,j}+i}{u_{2,k}-u_{2,j}-i}
\prod_{j=1}^{\tilde{K}_1}\frac{u_{2,k}-\tilde{u}_{1,j}-\frac{i}{2}}{u_{2,k}-\tilde{u}_{1,j}+\frac{i}{2}}
\prod_{j=1}^{\tilde{K}_3}\frac{u_{2,k}-\tilde{u}_{3,j}-\frac{i}{2}}{u_{2,k}-\tilde{u}_{3,j}+\frac{i}{2}}\,.\\
\end{aligned}
\end{equation}
which proves the equivalence of two gradings of the type 2 equation. From eq.(\ref{P1}), we know that $P(\tilde{x}_{3,k})=P(1/\tilde{x}_{1,k})=0$. While substituting back to (\ref{P}), we find $\tilde{x}_3,\tilde{x}_1$ satisfy the same equation as $x_3,x_1$. Now flip the fractions in the $\tilde{x}_3$ and $\tilde{x}_1$ equations, while using the relations~(\ref{ChargeRelation4}-\ref{ChargeRelation5}), we get equations for the alternative grading. In the end, we have proved that (\ref{Allloop}) are equivalent for the two choices of grading. Because $q_0^+=q_0^-$, the momentum conservation condition and the expression for the total energy are not changed under the fermionic duality as well. Notice that the relations among charges (\ref{ChargeRelation1}-\ref{ChargeRelation5}) play important roles in the verification of fermionic duality and dynamics duality. These relations are automatically satisfied by the charges calculated from the Cartan matrices, instead of imposing by hands in the twisted Bethe ansatz equations studied in \cite{Gromov:2007ky}. In this sense, the check of these two duality for orbifold ABJM theories is a non-trivial check of these all loop BAEs, especially the computations of these charges.
\subsection{Two applications}
As an application, we now compare our results with the all-loop BAE equations for the $\beta$-deformed ABJM theory \cite{chenliuwu}. For $\eta=1$ grading, the phase factors appearing in eqs.~(\ref{Momentum}) and (\ref{AllloopOrbifold})
\begin{eqnarray}
-2\pi i \bm{q}^+m/n&=&(2\pi i(t_2+t_3)m/n|-2\pi i t_1m/n, 0, -2\pi i (t_1-t_2-t_3)m/n,\nonumber\\
&&-2\pi i (-t_1+2 t_2)m/n, -2\pi i (-t_1+2 t_3)m/n), \end{eqnarray}
are replaced by
\begin{eqnarray} (-\pi i \beta (K_4-K_{\bar{4}})|0, 0, -\pi i \beta (K_4-K_{\bar{4}}), \pi i \beta(K_3-2 K_{\bar{4}}+L), -\pi i \beta (K_3-2 K_4+L)). \end{eqnarray}
It is easy to see that if \begin{eqnarray} t_1&=&0,\\t_2&=&-\frac{n}{4m}\beta (K_3-2K_{\bar{4}}+L), \\ t_3&=&\frac{n}{4m}\beta (K_3-2 K_4+L), \end{eqnarray}
these two groups of phases are the same. This means if these conditions are satisfied, the all-loop BAEs (for $\eta=1$ grading) for orbifold ABJM theories and $\beta$-deformed ABJM theory coincide for states with these excitation numbers. Notice here $\beta$ should be a rational number and $t_1$ should vanish.
As will be discussed in the next section, $t_1=0$ is the condition for the orbifold theory to have at least $\mathcal{N}=2$ supersymmetry. This condition is not surprising since the $\beta$-deformed ABJM theory is $\mathcal{N}=2$ supersymmetric \cite{Imeroni:2008cr}.
We now turn to relation between cusp anomalous dimension in orbifold ABJM theories and orbifold ${\mathcal{N}}=4$ SYM theories. As in \cite{c9}, we start with grading $\eta=-1$ and focus on the solutions to all loop BAEs with only non-vanishing roots $u_{4, k}=u_{\bar{4}, k}$. Then the consistency of the BAEs leads to $q_4^-=q_{\bar{4}}^-$. However from eq.~(\ref{qminus}), we have already $q_4^-=t_2-t_3=-q_{\bar{4}}^-$. Then we are restricted to the case with $q_4^-=q_{\bar{4}}^-=0$. We also demand the phase in the zero momentum condition is trivial,
\begin{equation} \exp(2\pi i m q_0^-/n)=1.\end{equation}
The above conditions leads to \begin{equation} t_2=t_3, \, \exp(4\pi i mt_2/n)=1. \label{condition1} \end{equation}
As for the orbifold SYM side with $sl(2)$ grading (corresponding to $\eta_1=\eta_2=-1$ in \cite{Beisert:2005fw}), the phase for the momentum-carrying node is automatically zero (see eq.~(3.15) of \cite{Beccaria:2011qd}). The triviality of the phase in the zero momentum condition gives
\begin{equation} \exp(2\pi imt_2^{SYM}/n)=1, \label{condition2}\end{equation}
where $t_2^{SYM}$ is one of the parameters appearing in the orbifold SYM theory.
Under the conditions in eqs.~(\ref{condition1}-\ref{condition2}), we can get the following relations
\begin{equation} f_{\text{orb. ABJM}}(\lambda)=\frac12f_{\text{orb. SYM}}(\lambda)|_{\frac{\sqrt{\lambda}}{4\pi}\to h(\lambda)}, \end{equation}
as the one obtained in \cite{c9}, under the assumption that wrapping contributions for twist operators are still subleading in the large spin limit with twist being finite.
\section{Supersymmetric orbifold theories}
Let us finally discuss the supersymmetric orbifold theories. Based on the results in previous sections, all we need is to determine the $t_i$'s (or equivalently $\bm{q}$'s in the distinguished simple root system) which are compatible with certain number of supersymmetries. Here we follow the argument of \cite{c2}. We also check the result by determining the spinors of $SO(8)$ preserved by the orbifolding.
\subsection{$\mathcal{N}=2$ Orbifolds}
To get an $\mathcal{N}=2$ theory, we need at least one fermionic (odd) generator commuting with the orbifold action. For example, when considering $E^{\alpha_2}$ corresponding to the only odd simple root $\alpha_2=-\epsilon_1+\delta_2$ in the distinguished simple root system, this is equivalent to set $q_2=-t_1=0$.
We now demonstrate that this is enough. $\Gamma$ can be naturally embedded into a $U(1)$ subgroup of $OSp(6|4)$. Denote the generator of this $U(1)$ as $\mathcal{P}$, we have that $\mathcal{P}^{\dagger}=\mathcal{P}$. $q_2=0$ means $[E^{-\epsilon_1+\delta_2},\mathcal{P}]=0$. Then $[E^{\epsilon_1-\delta_2},\mathcal{P}]=0$, as $(E^{-\epsilon_1+\delta_2})^{\dagger}=E^{\epsilon_1-\delta_2}$.
Note that $E^{\epsilon_1-\delta_2}$ locates in the first line of weight diagram in Fig.~\ref{Odd}, and $E^{-\epsilon_1+\delta_2}$ locates in the last line. By using $Sp(4)$ invariance\footnote{Precisely speaking, this $Sp(4)$ is in fact $Sp(2, 2)$ which is the double cover of $SO(2, 3)$, conformal group of three dimensional spacetime.}, we obtain that all generators in these two lines commute with $\mathcal{P}$, then we get an $\mathcal{N}=2$ theory with the charges,
\begin{equation}
\bm{q}=(-t_2-t_3|0, 0; -t_2-t_3, 2t_2, 2t_3).
\end{equation}
We can also get the charges of $\bm{4}$ of $SO(6)$ as $(t_2,-t_2,t_3,-t_3)$, where $t_2, t_3$ are arbitrary integers except ones satisfying $t_2\pm t_3=0$, because such $t_1, t_2$ will give ${\cal N}=4$ supersymmetries, as we will show below.
This includes the chiral orbifold theory in \cite{c1} as a special case.
We further demonstrate this method is indeed correct by counting the spinors of $SO(8)$ preserved by the orbifold.
Notice that the orbifold ABJM theory is the low energy effective theory of $N$ coincident M2-branes at $\mathbb{C}^4/(\Gamma\times \mathbb{Z}_{|\Gamma|k})$ orbifold singularity where $|\Gamma|$ is the order of $\Gamma$
and $\mathbb{Z}_{|\Gamma|k}$ acts as overall phase rotations of the four complex coordinates \cite{Berenstein}.
Under the action of the generator of finite group $\Gamma$ as $(Y^1,Y^2,Y^3,Y^4)\rightarrow(\omega^{t_2}Y^1,\omega^{-t_2}Y^2,\omega^{t_3}Y^3,\omega^{-t_3}Y^4)$,
the $SO(8)$ spinor transforms like $\epsilon\rightarrow\omega^{(\mathfrak{s_1}t_2-\mathfrak{s_2}t_2+\mathfrak{s_3}t_3-\mathfrak{s_4}t_3)}\epsilon$, where $\mathfrak{s}_{1,2,3,4}=\pm 1/2$.
The equation
\begin{equation}
\mathfrak{s_1}t_2-\mathfrak{s_2}t_2+\mathfrak{s_3}t_3-\mathfrak{s_4}t_3\in n\mathbb{Z}
\end{equation}
subject to
\begin{equation}
\mathfrak{s_1}+\mathfrak{s_2}+\mathfrak{s_3}+\mathfrak{s_4} \in k|\Gamma|\mathbb{Z}
\end{equation}
has exact two solutions
\begin{equation} (\mathfrak{s_1}, \mathfrak{s_2}, \mathfrak{s_3}, \mathfrak{s_4})=\pm(1/2, 1/2, -1/2, -1/2), \end{equation}
for generic $t_2, t_3$ and $n$, and this demonstrates our conclusion above\footnote{We assume $k\geq 3$ here.}.
\subsection{$\mathcal{N}=4$ Orbifold}
From the above example, we note supersymmetric orbifold ABJM theories always preserve an $\mathcal{N}=even$ supersymmetry as an consequence of the special structure of $osp(6|4)$ algebra while which is not the case in orbifolds of $\mathcal{N}=4$ SYM theory. This results can be confirmed by spinor counting. If $(\mathfrak{s_1}, \mathfrak{s_2}, \mathfrak{s_3}, \mathfrak{s_4})$ satisfies the projection condition, so does $(-\mathfrak{s_1}, -\mathfrak{s_2}, -\mathfrak{s_3}, -\mathfrak{s_4})$.
We now consider $\mathcal{N}=4$ orbifold. To find the conditions, without loss of generality we can first demand $q_2=-t_1=0$, then further demand $q_3=2t_1-t_2-t_3=0$, while $q_4=-t_1+2t_2\neq0$,\,$q_{\bar4}=-t_1+2t_3\neq0$. We then have $[E^{\alpha_3},\mathcal{P}]=0$, then $[\[E^{-\epsilon_1+\delta_2},E^{\alpha_3}\],\mathcal{P}]=0$, this is $[E^{-\epsilon_2+\delta_2},\mathcal{P}]=0$ and according to the argument in $\mathcal{N}=2$ case, we further have $[E^{\epsilon_1-\delta_2},\mathcal{P}]=0$ and $[E^{\epsilon_2-\delta_2},\mathcal{P}]=0$, together with the $Sp(4)$ symmetry we get an $\mathcal{N}=4$ theory. Solving these constraints including $q_2=-t_1=0$, we find
$\bm{q}=(0|0, 0; 0, 2t_2, -2t_2)$, with the charges of $\bm{4}$ as $(t_2, -t_2, -t_2, t_2)$.
Using the spinor counting method we can also demonstrate our conclusion is correct.
\section{Discussions}
In this paper, we studied the integrability of planar orbifold ABJM theories. We first carried out perturbative computations of ADM
in the scalar sector at two-loop order. We found that in the corresponding spin chain Hamiltonian, only two terms are deformed by certain phases. This deformation can be expressed in terms of twisted boundary condition. By inserting certain diagonal matrices inside the transfer matrices, we proved the integrability of this Hamiltonian. BAEs and eigenvalues of ADM were obtained through algebraic Bethe ansatz method. Restricting $\Gamma$ to be inside $SU(4)_R$, we obtained the all-loop all-sector BAEs which pass some non-trivial consistency checks.
There are several interesting directions worth pursuing. One of them is that to explore all-loop BAEs for general $\Gamma$ in $SU(4)_R\times U(1)_b$. This study is beyond the framework of Beisert-Roiban \cite{c2} since $U(1)_b$ does not correspond to a node in the Dynkin diagram used for BAEs. To obtain some hints for the structure of the result, it may be helpful to first perturbatively compute the ADM of composite operators involving fermions as the computation in ABJM theory \cite{Minahan:2009te}. It is also interesting to find some solutions in the thermodynamical limit and study their holographic dual in term of semi-classical string/membrane solutions in the dual string/M theories.
Supersymmetric condition for the orbifold was studied in this framework of integrability. The obtained condition is consistent with the result that orbifold ABJM theory is the low energy effective theory of $N$ membranes put at $\mathbb{C}^4/(\Gamma\times \mathbb{Z}_{|\Gamma|k})$ \cite{Berenstein}. However the study in the integrability side seems only give condition for $\mathbb{Z}_n$ orbifolds which is $\mathcal{N}=2$ or $\mathcal{N}=4$ simultaneously for all $n$. Let us consider the following examples taken from \cite{Terashima:2008ba}. Take $n$ to be even. The cases with $(t_1, t_2, t_3, t_4)=\pm(n/2, n/2, (-1)^l n, 0), l=0, 1$ is $\mathcal{N}=2$ supersymmetric and the case with $(t_1, t_2, t_3, t_4)=\pm(n/2, -n/2, (-1)^l n, 0), l=0, 1$ is $\mathcal{N}=4$ supersymmetric. The preserved supersymmetries can be easily obtained by counting the $SO(8)$ spinors
preserved by the orbifolds. Also notice that all these cases satisfy $\Gamma<SU(4)_R$. We speculate that these cases do not appear in the analysis of supersymmetric orbifold here because they only appear for even $n$,
not for all integer $n$. It is still interesting to see whether we can probe such cases through some refinements of the studies here. We leave this and directions mentioned previously as suggestions for further studies.
\section*{Acknowledgments}
It is our great pleasure to thank Bin Chen, Jun-Peng Cao, Jian-Xin Lu, Wei Song, Yu-Peng Wang, Gang Yang, Wen-Li Yang, Konstantinos Zoubos for very helpful discussions. We would like to express our special thanks to the anonymous referees for valuable suggestions to improve the paper.
JW would also like to thank the participants of the advanced workshop ``Dark Energy and Fundamental Theory'' supported by the Special Fund for Theoretical Physics from NSFC with Grant No.~11447613 for stimulating discussion. We thank Institute of Modern Physics, Northwest University and School of Physics and Astronomy, Sun Yat-Sen University for hospitality in visits during this project. This work was in part supported by NSFC Grant No. 11575202(NB, HHC, DSL, JW), No. 11475116(XCD) and No. 11222549 (NB, HHC, DSL, JW). JW also gratefully acknowledges the support of K.~C.~Wong Education Foundation.
\begin{appendix}
\section{Hamiltonian of the twisted spin chain}\label{appendixa}
In this appendix, we give the detailed derivation of eq.~(\ref{totalH}). We employ a new set of indices $i=1,2,\cdots ,2L$ to relabel the quantum spaces of the alternating spin chain. Then the monodromy matrices in eqs.~(\ref{mono1}) and (\ref{mono2}) are rewritten as
\begin{eqnarray}
T_0(u)=M_0 R_{01}(u)R_{02}(u-2)R_{03}(u)R_{04}(u-2)\cdots R_{0,2L-1}(u)R_{0,2L}(u-2), \\
T_{\bar{0}}(u)=\bar{M}_{\bar{0}} R_{{\bar{0}}1}(u-2)R_{{\bar{0}}2}(u)R_{{\bar{0}}3}(u-2)R_{{\bar{0}}4}(u)\cdots R_{{\bar{0}},2L-1}(u-2)R_{{\bar{0}},2L}(u).
\end{eqnarray}
At the special point $u=0$, the transfer matrices become
\begin{eqnarray}
{\tau}(0)&=&\mbox{Tr}_0 (-)^L M_0 P_{01}(-2+K_{02})\cdots P_{0,2L-1}(-2+K_{0,2L})\\\nonumber
&=&\mbox{Tr}_0 (-)^L M_0 (-2+K_{12}) P_{13}(-2+K_{14})\cdots P_{1,2L-1}(-2+K_{1,2L})P_{01}\\\nonumber
&=&(-)^L(-2+K_{12})\prod_{j=2}^LP_{1,2j-1}(-2+K_{1,2j})M_1,\\
\bar{\tau}(0)&=&\mbox{Tr}_{\bar{0}} (-)^L \bar{M}_{\bar{0}}(-2+K_{\bar{0}1})P_{\bar{0}2}\cdots(-2+K_{\bar{0},2L-1})P_{\bar{0},2L}\\\nonumber
&=&\mbox{Tr}_{\bar{0}} (-)^L \bar{M}_{\bar{0}}P_{\bar{0},2L}(-2+K_{2L,1})P_{2L,2}\cdots (-2+K_{2L,2L-1})\\\nonumber
&=&(-)^L\bar{M}_{2L}\prod_{j=1}^{L-1}(-2+K_{2L,2j-1})P_{2L,2j}(-2+K_{2L,2L-1}).
\end{eqnarray}
and
\begin{eqnarray} \label{expan1}
&&\frac{d}{du}\tau(u)|_{u=0}\\\nonumber
&=&\mbox{Tr}_0 \sum_{i=1}^{L-1}M_0\left(\prod_{j=1}^{i-1}(-P_{0,2j-1})(-2+K_{0,2j})\right)(-2+K_{0,2i})\left(\prod_{k=i+1}^{L}(-P_{0,2k-1})(-2+K_{0,2k})\right)\\\nonumber
&+&\mbox{Tr}_0\sum_{i=1}^{L-1}M_0\left(\prod_{j=1}^{i-1}(-P_{0,2j-1})(-2+K_{0,2j})\right)(-P_{0,2i-1})\left(\prod_{k=i+1}^{L}(-P_{0,2k-1})(-2+K_{0,2k})\right)\\\nonumber
&+&\mbox{Tr}_0M_0\left(\prod_{j=1}^{L-1}(-P_{0,2j-1})(-2+K_{0,2j})\right)(-2+K_{0,2L})\\\nonumber
&+&\mbox{Tr}_0M_0\left(\prod_{j=1}^{L-1}(-P_{0,2j-1})(-2+K_{0,2j})\right)(-P_{0,2L-1})\\\nonumber
&=&\Sigma_1+\Sigma_2+\Sigma_3+\Sigma_4,
\end{eqnarray}
\begin{eqnarray} \label{expan2}
&&\frac{d}{du}\bar{\tau}(u)|_{u=0}\\\nonumber
&=&\mbox{Tr}_{\bar{0}} \sum_{i=1}^{L-1}\bar{M}_{\bar{0}}\left(\prod_{j=1}^{i-1}(-2+K_{\bar{0},2j-1})(-P_{\bar{0},2j})\right)(-2+K_{\bar{0},2i-1})\left(\prod_{k=i+1}^{L}(-2+K_{\bar{0},2k-1})(-P_{\bar{0},2k})\right)\\\nonumber
&+&\mbox{Tr}_{\bar{0}} \sum_{i=1}^{L-1}\bar{M}_{\bar{0}}\left(\prod_{j=1}^{i-1}(-2+K_{\bar{0},2j-1})(-P_{\bar{0},2j})\right)(-P_{\bar{0},2i})\left(\prod_{k=i+1}^{L}(-2+K_{\bar{0},2k-1})(-P_{\bar{0},2k})\right)\\\nonumber
&+&\mbox{Tr}_{\bar{0}} \bar{M}_{\bar{0}}\left(\prod_{j=1}^{L-1}(-2+K_{\bar{0},2j-1})(-P_{\bar{0},2j})\right)(-2+K_{\bar{0},2L-1})\\\nonumber
&+&\mbox{Tr}_{\bar{0}} \bar{M}_{\bar{0}}\left(\prod_{j=1}^{L-1}(-2+K_{\bar{0},2j-1})(-P_{\bar{0},2j})\right)(-P_{\bar{0},2L})\\\nonumber
&=&\bar{\Sigma}_1+\bar{\Sigma}_2+\bar{\Sigma}_3+\bar{\Sigma}_4,
\end{eqnarray}
where we use $\Sigma_i$,$\bar{\Sigma}_i$,i=1,$\cdots 4$ to label each part in eq.(\ref{expan1}) and (\ref{expan2}) for the convenience of writing. Then let us first deal with $\tau'(0)$ and give some intermediate results of the calculations
\begin{eqnarray}\nonumber
\Sigma_1&=&(-)^{L-1}\sum_{i=1}^{L-1}(-2+K_{12})\prod_{j=2}^{i-1}P_{1,2j-1}(-2+K_{1,2j})(-2+K_{1,2i})\prod_{k=i+1}^{L}P_{1,2k-1}(-2+K_{1,2k})M_1,\\
\Sigma_2&=&(-)^{L}\sum_{i=1}^{L-1}(-2+K_{12})\prod_{j=2}^{i-1}P_{1,2j-1}(-2+K_{1,2j})P_{1,2i-1}\prod_{k=i+1}^{L}P_{1,2k-1}(-2+K_{1,2k})M_1.
\end{eqnarray}
\begin{eqnarray}
\Sigma_3&=&(-)^{L-1}(-2+K_{12})\prod_{j=2}^{L-1}P_{1,2j-1}(-2+K_{1,2j})(-2+K_{1,2L})M_1,\\
\Sigma_4&=&(-)^{L}(-2+K_{12})\prod_{j=2}^{L-1}P_{1,2j-1}(-2+K_{1,2j})P_{1,2L-1}M_1.
\end{eqnarray}
Thus we find
\begin{eqnarray}
\tau(0)^{-1}\Sigma_1&=&-\sum_{i=1}^{L-1}\left(P_{2i-1,2i+1}-\frac{1}{2}K_{2i-1,2i}K_{2i,2i+1}-\frac{1}{2}K_{2i,2i+1}K_{2i-1,2i}+\frac{1}{4}K_{2i,2i+1}\right), \label{ad1}\\
\tau(0)^{-1}\Sigma_2&=&\sum_{i=1}^{L-1}\left(-\frac{1}{2}+\frac{1}{4}K_{2i,2i+1}\right),\\
\tau(0)^{-1}\Sigma_3&=&-M^{-1}_1\left(P_{1,2L-1}-\frac{1}{2}K_{2L-1,2L}K_{1,2L}-\frac{1}{2}K_{1,2L}K_{2L-1,2L}+\frac{1}{4}K_{1,2L}\right)M_1,\\
\tau(0)^{-1}\Sigma_4&=&M_1^{-1}\left(-\frac{1}{2}+\frac{1}{4}K_{1,2L}\right)M_1. \label{ad2}
\end{eqnarray}
Similarly, for $\bar{\tau}'(0)$ we have
\begin{eqnarray}
\bar{\Sigma}_1&=&(-)^{L-1}\sum_{i=1}^{L-1}\bar{M}_{2L}\prod_{j=1}^{i-1}(-2+K_{2L,2j-1})P_{2L,2j}(-2+K_{2L,2i-1})
\\
&\times&\prod_{k=i+1}^{L-1}(-2+K_{2L,2k-1})P_{2L,2k}(-2+K_{2L,2L-1}),\nonumber \\
\bar{\Sigma}_2&=&(-)^{L}\sum_{i=1}^{L-1}\bar{M}_{2L}\prod_{j=1}^{i-1}(-2+K_{2L,2j-1})P_{2L,2j}P_{2L,2i}\\
&\times&\prod_{k=i+1}^{L-1}(-2+K_{2L,2k-1})P_{2L,2k}(-2+K_{2L,2L-1}),\nonumber
\end{eqnarray}
\begin{eqnarray}
\bar{\Sigma}_3&=&(-)^{L-1}(-2+K_{2L-2,2L-1})\bar{M}_{2L-2}\prod_{j=1}^{L-2}(-2+K_{2L-2,2j-1})P_{2L-2,2j}(-2+K_{2L-2,2L-3}),\\
\bar{\Sigma}_4&=&(-)^L\bar{M}_{2L}\prod_{j=1}^{L-1}(-2+K_{2L,2j-1})P_{2L,2j}.
\end{eqnarray}
Therefore,
\begin{eqnarray}
\bar{\tau}(0)^{-1}\bar{\Sigma}_1&=&-\sum_{i=1}^{L-1} \left(P_{2i,2i+2}-\frac{1}{2}K_{2i,2i+1}K_{2i+1,2i+2}-\frac{1}{2}K_{2i+1,2i+2}K_{2i,2i+1}+\frac{1}{4}K_{2i+1,2i+2}\right) \label{ad3}\\
\bar{\tau}(0)^{-1}\bar{\Sigma}_2&=&\sum_{i=1}^{L-1}\left(-\frac{1}{2}+\frac{1}{4}K_{2i-1,2i}\right)\\
\bar{\tau}(0)^{-1}\bar{\Sigma}_3&=&-P_{2L,2}\left(\bar{M}^{-1}_{2L}\bar{M}_{2}-\frac{1}{2}\bar{M}^{-1}_{2L}\bar{M}_{2} K_{12}-\frac{1}{2}K_{2L,1}\bar{M}^{-1}_{2L}\bar{M}_{2}+\frac{1}{4}K_{2L,1}\bar{M}^{-1}_{2L}\bar{M}_{2} K_{12}\right) \label{spterm1}\\
\bar{\tau}(0)^{-1}\bar{\Sigma}_4&=&-\frac{1}{2}+\frac{1}{4}K_{2L-1,2L} \label{ad4}
\end{eqnarray}
We note that the last term in eq.(\ref{spterm1}) can be simplified as
\begin{eqnarray}
&&\left(P_{2L,2}K_{2L,1}\bar{M}_{2L}^{-1}\bar{M}_2K_{12}\right)^{i_1,j_2,j_{2L}}_{j_1,i_2,i_{2L}}\\\nonumber
&=&(P_{2L,2})^{j_{2L},j_2}_{\,e,\,\,\,c}(K_{2L,1})^{a,e}_{j_1,d}(\bar{M}^{-1})^d_{i_{2L}}(\bar{M})^c_b(K_{12})^{i_1,b}_{a,i_2}\\\nonumber
&=&\delta^{j_2}_e\delta^{j_{2L}}_c\delta^a_d\delta^e_{j_1}\bar{m}^{-1}_{i_{2L}}\delta^d_{i_{2L}}\bar{m}_b\delta^c_b\delta^b_a\delta^{i_1}_{i_2}\\
&=&\delta^{i_1}_{i_2}\delta^{j_2}_{j_1}\delta^{j_{2L}}_{i_{2L}}=(K_{12})^{i_1,j_2}_{j_1,i_2}.
\end{eqnarray}
So it turns out that the nearest neighbor interactions still cancels even for the twisted spin chain. Finally, by adding up eqs.(\ref{ad1})-(\ref{ad2}) and (\ref{ad3})-(\ref{ad4}), we get the Hamiltonian in eq.(\ref{totalH}).
\section{Zero momentum condition}\label{appendixb}
We change the transfer matrices into a form much easier for us to compute by means of permutation operators.
\begin{eqnarray}
\tau(0)&=&(-)^L\mbox{Tr}_0 M_0 P_{01}(-2+K_{02})\cdots P_{0,2L-1}(-2+K_{0,2L})\\\nonumber
&=&(-)^L\mbox{Tr}_0 P_{0,2L-1}M_{2L-1} P_{2L-1,1}(-2+K_{2L-1,2})\cdots P_{2L-1,2L-3}(-2+K_{2L-1,2L-2})(-2+K_{0,2L})\\\nonumber
&=&(-)^{L}(-2+K_{2L-1,2L-2})M_{2L-1}\prod_{i=1}^{L-1}P_{2L-2i+1,2L-2i-1}\prod_{j=1}^{L-1}(-2+K_{2j+1,2j}).\\\nonumber
\bar{\tau}(0)&=&(-)^L\mbox{Tr}_{\bar{0}}\bar{M}_{\bar{0}}(-2+K_{\bar{0}1})P_{\bar{0}2}\cdots (-2+K_{\bar{0},2L-1})P_{\bar{0},2L}\\\nonumber
&=&(-)^L(-2+K_{23})P_{24}\cdots (-2+K_{2,2L-1})P_{2,2L}\left[\mbox{Tr}_{\bar{0}}\bar{M}_{\bar{0}}(-2+K_{\bar{0}1})P_{\bar{0}2}\right]\\\nonumber
&=&(-)^L\prod_{i=1}^{L-1}(-2+K_{2i,2i+1})\prod_{j=1}^{L-1}P_{2L-2j,2L-2j+2}\bar{M}_2(-2+K_{21}).
\end{eqnarray}
Therefore after some cancellations, we get
\begin{eqnarray}
&&\tau(0)\bar{\tau}(0)\\\nonumber
&=&2^{2(L-1)}\prod_{i=1}^{L-1}P_{2L-2i+1,2L-2i-1}(-2+K_{1,2L})M_1\bar{M}_{2L}(-2+K_{2L,1})\prod_{j=1}^{L-1}P_{2L-2j,2L-2j+2}.
\end{eqnarray}
We can obtain the component of the above operator by acting on a given basis
\begin{eqnarray}
&&\left[\tau(0)\bar{\tau}(0)\right]^{I_1,J_2,\cdots I_{2L-1},J_{2L}}_{J_1,I_2,\cdots J_{2L-1},I_{2L}}\\\nonumber
&=&2^{2(L-1)}\left(\prod_{i=1}^{L-1}P_{2L-2i+1,2L-2i-1}\right)^{\,b\,,\,\,I_3,\cdots I_{2L-1}}_{J_1,J_3,\cdots J_{2L-1}}\left[(-2+K_{1,2L})M_1\bar{M}_{2L}(-2+K_{2L,1})\right]^{I_1,J_{2L}}_{\,\,b,\,\,\,a}\\\nonumber
&\times& \left(\prod_{j=1}^{L-1}P_{2L-2j,2L-2j+2}\right)^{J_2,J_4,\cdots J_{2L-2},\,a}_{I_2,I_4,\cdots I_{2L-2},I_{2L}}.
\end{eqnarray}
Since
\begin{eqnarray}
\left(\prod_{j=1}^{L-1}P_{2L-2j,2L-2j+2}\right)^{J_2,J_4,\cdots J_{2L-2},\,a}_{I_2,I_4,\cdots I_{2L-2},I_{2L}}&=&\delta^a_{I_2}\delta^{J_2}_{I_4}\delta^{J_4}_{I_6}\cdots \delta^{J_{2L-2}}_{I_{2L}},\\
\left(\prod_{i=1}^{L-1}P_{2L-2i+1,2L-2i-1}\right)^{\,b\,,\,\,I_3,\cdots I_{2L-1}}_{J_1,J_3,\cdots J_{2L-1}}&=&\delta^{I_3}_{J_1}\delta^{I_5}_{J_3}\cdots \delta^{I_{2L-1}}_{J_{2L-3}}\delta^{\,\,b}_{J_{2L-1}}.
\end{eqnarray}
\begin{eqnarray}
\left[(-2+K_{1,2L})M_1\bar{M}_{2L}(-2+K_{2L,1})\right]^{I_1,J_{2L}}_{\,\,b,\,\,\,a}
=2^2m_{I_1}\bar{m}_{J_{2L}}\delta^{I_1}_b\delta^{J_{2L}}_a,
\end{eqnarray}
we find
\begin{eqnarray}
\left[\tau(0)\bar{\tau}(0)\right]^{I_1,J_2,\cdots I_{2L-1},J_{2L}}_{J_1,I_2,\cdots J_{2L-1},I_{2L}}
=2^{2L}\omega^{-ms_{I_1}+ms_{I_2}}\delta^{J_{2L}}_{I_2}\delta^{J_2}_{I_4}\delta^{J_4}_{I_6}\cdots \delta^{J_{2L-2}}_{I_{2L}}\cdot \delta^{I_3}_{J_1}\delta^{I_5}_{J_3}\cdots \delta^{I_{2L-1}}_{J_{2L-3}}\delta^{I_1}_{J_{2L-1}}.
\end{eqnarray}
So
\begin{eqnarray}
&&\left[\tau(0)\bar{\tau}(0)\right]\cdot \mbox{Tr}(\gamma^m Y^{I_1}Y^{\dagger}_{I_2}\cdots Y^{I_{2L-1}}Y^{\dagger}_{I_{2L}})\\\nonumber
&=&\left[\tau(0)\bar{\tau}(0)\right]^{I_1,J_2,\cdots I_{2L-1},J_{2L}}_{J_1,I_2,\cdots J_{2L-1},I_{2L}}\cdot \mbox{Tr}(\gamma^m Y^{J_1}Y^{\dagger}_{J_2}\cdots Y^{J_{2L-1}}Y^{\dagger}_{J_{2L}})\\\nonumber
&=&2^{2L}\omega^{-ms_{I_1}+ms_{I_2}}\mbox{Tr}(\gamma^m Y^{I_3}Y^{\dagger}_{I_4}\cdots Y^{I_{2L-1}}Y^{\dagger}_{I_{2L}}Y^{I_{1}}Y^{\dagger}_{I_{2}})\\\nonumber
&=&2^{2L}\omega^{-ms_{I_1}+ms_{I_2}}\mbox{Tr}(Y^{I_{1}}Y^{\dagger}_{I_{2}}\gamma^m Y^{I_3}Y^{\dagger}_{I_4}\cdots Y^{I_{2L-1}}Y^{\dagger}_{I_{2L}})\\\nonumber
&=&2^{2L}\mbox{Tr}(\gamma^m Y^{I_1}Y^{\dagger}_{I_2}\cdots Y^{I_{2L-1}}Y^{\dagger}_{I_{2L}}).
\end{eqnarray}
which leads to
\begin{eqnarray}
\frac{1}{2^{2L}}\tau(0)\bar{\tau}(0)=\mathbb{I}.
\end{eqnarray}
\section{The $osp(6|4)$ algebra}\label{appendixc}
According to Kac's classification of Lie superalgebra, the $osp(6|4)$ belongs to $D(3,2)$
basic Lie superalgebra,
\begin{eqnarray}
\mathscr{G}=osp(6|4),\quad
\mathscr{G}_{\bar{0}}=so(6)\oplus sp(4) ,\quad
\mathscr{G}_{\bar{1}}=(\bm{6},\bm{4}).
\end{eqnarray}
The $\bar{0},\bar{1}$ refer to the $\mathbb{Z}_2$ grading, and the $\bm{6},\bm{4}$ means that the odd part generators $\mathscr{G}_{\bar{1}}$ are in the $\bm{6}$ and $\bm{4}$ representations of the even part $\mathscr{G}_{\bar{0}}$, $\textit{i.e.}$ in the $\bm{6}$ of $so(6)$ and $\bm{4}$ of $sp(4)$. The total 24 odd generators are presented on the Fig.~\ref{Odd}, where we denote $E^{\alpha}$ as the generators of the algebra, for $\alpha\in\Delta$.\\
\par The rank of the $osp(6|4)$ algebra is 5, and the root system is
\begin{eqnarray}
{\Delta}_{\bar{0}}&=&\{\pm\epsilon_1\pm\epsilon_2,\pm\epsilon_2\pm\epsilon_3,\pm2\delta_1,\pm2\delta_2,\pm\delta_1\pm\delta_2\},\\
{\Delta}_{\bar{1}}&=&\{\pm\epsilon_1\pm\delta_1,\pm\epsilon_2\pm\delta_1,\pm\epsilon_3\pm\delta_1,\pm\epsilon_1\pm\delta_2,\pm\epsilon_2\pm\delta_2,\pm\epsilon_3\pm\delta_2\}.
\end{eqnarray}
where $\delta_{1,2},\epsilon_{1,2,3}$ are two basis satisfy $(\delta_i,\delta_j)=-\delta_{ij}$, $(\epsilon_i,\epsilon_j)=\delta_{ij}$, $(\delta_i,\epsilon_j)=0$.
The distinguished simple root system is
\begin{eqnarray}
\Delta^{0}=\{\delta_1-\delta_2,\delta_2-\epsilon_1,\epsilon_1-\epsilon_2,\epsilon_2-\epsilon_3,\epsilon_2+\epsilon_3\}.
\end{eqnarray}
we label the simple roots as $\alpha_1,\alpha_2,\alpha_3,\alpha_4,\alpha_{\bar4}$ in above giving order. The distinguished simple root system has exactly one odd root, other possible simple root systems can be obtained by odd Weyl reflections.
For our purpose, the \textit{symmetric} Cartan matrix is more useful than the asymmetric definitions and is defined by,
\begin{equation}
M_{jj'}=(\alpha_i,\alpha_{j'}).
\end{equation}
In the distinguished simple root system, it has the form,
\begin{equation}
M_{jj'}=\matr{ccccc}{
-2& +1 & & & \\
+1& & -1 & & \\
& -1& +2 & -1& -1 \\
& & -1& +2& \\
& & -1& & +2
}.
\end{equation}
\def\g#1{\save
[].[dr]!C="g#1"*[F]\frm{}\restore}%
\begin{figure}[H]
\begin{equation}
\xymatrix{
&E^{\epsilon_1+\delta_2}\ar[d]^{\alpha_3}&&E^{\epsilon_1+\delta_1}\ar[d]^{\alpha_3}\ar[l]^{-\delta_2+\delta_1}&&E^{\epsilon_1-\delta_1}\ar[d]^{\alpha_3}\ar[l]^{-2\delta_1}&&E^{\epsilon_1-\delta_2}\ar[d]^{\alpha_3}\ar[l]^{-\delta_2+\delta_1}&\\
&E^{\epsilon_2+\delta_2}\ar[dl]^{\alpha_4}\ar[d]^{\alpha_{\bar4}}&&E^{\epsilon_2+\delta_1}\ar[dl]^{\alpha_4}\ar[d]^{\alpha_{\bar4}}&&E^{\epsilon_2-\delta_1}\ar[dl]^{\alpha_4}\ar[d]^{\alpha_{\bar4}}&&E^{\epsilon_2-\delta_2}\ar[dl]^{\alpha_4}\ar[d]^{\alpha_{\bar4}}&\\
E^{\epsilon_3+\delta_2}\ar[dr]^{\alpha_{\bar4}}&E^{-\epsilon_3+\delta_2}\ar[d]^{\alpha_4}&E^{\epsilon_3+\delta_1}\ar[dr]^{\alpha_{\bar4}}&E^{-\epsilon_3+\delta_1}\ar[d]^{\alpha_4}&E^{\epsilon_3-\delta_1}\ar[dr]^{\alpha_{\bar4}}& E^{-\epsilon_3-\delta_1}\ar[d]^{\alpha_4}&E^{\epsilon_3-\delta_2}\ar[dr]^{\alpha_{\bar4}}&E^{-\epsilon_3-\delta_2}\ar[d]^{\alpha_4}&\\
&E^{-\epsilon_2+\delta_2}\ar[d]^{\alpha_3}&&E^{-\epsilon_2+\delta_1}\ar[d]^{\alpha_3}&&E^{-\epsilon_2-\delta_1}\ar[d]^{\alpha_3}&&E^{-\epsilon_2-\delta_2}\ar[d]^{\alpha_3}&\\
&E^{-\epsilon_1+\delta_2}&&E^{-\epsilon_1+\delta_1}&&E^{-\epsilon_2-\delta_1}&&E^{-\epsilon_2-\delta_2}&
}\\\nonumber
\end{equation}
\caption{\footnotesize{The weight diagram for $\mathscr{G}_{\bar{1}}$, where the four vertical weight sub-diagram are the weight diagram $\bm{6}$ of $so(6)$, while themselves are in the $\bm{4}$ of $sp(4)$ which correspond to the horizontal sub-diagram.}}
\label{Odd}
\end{figure}
\subsection{Odd Weyl reflections}
We know that Dynkin diagram is not unique for simple Lie superalgebra. We extend the ordinary Weyl reflections (reflections with respect to even roots),
\begin{equation}
w_\alpha\beta=\beta-2\frac{(\beta,\alpha)}{(\alpha,\alpha)}\alpha,
\end{equation}
for $\beta\in\Delta, \alpha\in\Delta_0$, to include the case with respect to odd roots as well,
\begin{equation}
\begin{aligned}
&w_\alpha\beta=\beta-2\frac{(\beta,\alpha)}{(\alpha,\alpha)}\alpha,&\text{if}\;(\alpha,\alpha)\neq0,\\
&w_\alpha\beta=\beta+\alpha, \qquad\qquad&\text{if}\;(\alpha,\alpha)=0\;\text{and}\;(\alpha,\beta)\neq0,\\
&w_\alpha\beta=\beta, \qquad\qquad&\text{if}\;(\alpha,\alpha)=0\;,\;(\alpha,\beta)=0\;\text{and}\;\beta\neq\alpha,\\
&w_\alpha\alpha=-\alpha.
\end{aligned}
\end{equation}
Begin with the distinguished simple root system, using the odd root Weyl reflections upon each root, we get a new simple root system and this procedure goes on and on. Here we give some examples. The distinguished simple root system of $osp(6|4)$ is
\begin{eqnarray}
\Delta^{0}=\{\delta_1-\delta_2,\delta_2-\epsilon_1,\epsilon_1-\epsilon_2,\epsilon_2-\epsilon_3,\epsilon_2+\epsilon_3\}.
\end{eqnarray}
\begin{figure}[H]
\centering
\includegraphics[width=5cm]{Dynkin.png}
\end{figure}
Applying the Weyl reflection with respect to the second simple root, we get
\begin{equation}
w_{\delta_2-\epsilon_1}(\Delta^0)=\{\delta_1-\epsilon_1,-\delta_2+\epsilon_1,\delta_2-\epsilon_2,\epsilon_2-\epsilon_3,\epsilon_2+\epsilon_3\}.
\end{equation}
\begin{figure}[H]
\centering
\includegraphics[width=5cm]{Dynkin1.png}
\end{figure}
Other examples are
\begin{equation}
w_{\delta_1-\epsilon_1}(w_{\delta_2-\epsilon_1}(\Delta^0))=\{-\delta_1+\epsilon_1,\delta_1-\delta_2,\delta_2-\epsilon_2,\epsilon_2-\epsilon_3,\epsilon_2+\epsilon_3\},
\end{equation}
\begin{figure}[H]
\centering
\includegraphics[width=5cm]{Dynkin2.png}
\end{figure}
\begin{equation}
w_{\delta_2-\epsilon_2}(w_{\delta_2-\epsilon_1}(\Delta^0))=\{\delta_1-\epsilon_1,\epsilon_1-\epsilon_2,-\delta_2+\epsilon_2,\delta_2-\epsilon_3,\delta_2+\epsilon_3\},
\end{equation}
\begin{figure}[H]
\centering
\includegraphics[width=5cm]{Dynkin3.png}
\end{figure}
\begin{equation}
w_{\delta_2-\epsilon_2}(w_{\delta_1-\epsilon_1}(w_{\delta_2-\epsilon_1}(\Delta^0)))=\{-\delta_1+\epsilon_1,\delta_1-\epsilon_2,-\delta_2+\epsilon_2,\delta_2-\epsilon_3,\delta_2+\epsilon_3\},
\end{equation}
\begin{figure}[H]
\centering
\includegraphics[width=5cm]{Dynkin4.png}
\end{figure}
\begin{equation}
w_{\delta_2+\epsilon_3}(w_{\delta_2-\epsilon_2}(w_{\delta_2-\epsilon_1}(\Delta^0)))=\{\delta_1-\epsilon_1,\epsilon_1-\epsilon_2,\epsilon_3+\epsilon_2,-\delta_2-\epsilon_3,2\delta_2\}.
\end{equation}
\begin{figure}[H]
\centering
\includegraphics[width=6.5cm]{Dynkin5.png}
\end{figure}
\end{appendix}
|
\section{Introduction}
We consider in this paper the incompressible second grade fluid equations:
\begin{equation}\label{maineq}
\partial_t (u-\alpha\Delta u)-\nu\Delta u +u\cdot\nabla(u-\alpha\Delta u)+\sum_j(u-\alpha\Delta u)_j\nabla u_j=-\nabla p, \qquad \dive u=0,
\end{equation}
where $\alpha$ and $\nu$ are some non-negative constants. The fluid is assumed to be enclosed in a bounded smooth region $\Omega$ of $\mathbb{R}^3$ and the homogeneous Dirichlet boundary conditions are imposed
\begin{equation}\label{2}
u(t,\cdot)\bigl|_{\partial\Omega}=0\qquad \text{for all }t\geq0.
\end{equation}
The initial value problem is considered and we denote by $u_0$ the initial velocity:
\begin{equation}\label{3}
u(0,x)=u_0(x).
\end{equation}
The equations \eqref{maineq} were deduced in \cite{dunn_thermodynamics_1974} from physical principles. Let us just mention here that the second grade fluids are characterized by the following fact: the stress tensor is a polynomial of degree two in the first two Rivlin-Ericksen tensors which are the deformation tensor $D$ and the tensor $(\partial_t+u\cdot\nabla)D$. The vanishing viscosity case $\nu=0$ is also known under the name $\alpha$--Euler or Euler--$\alpha$ equations and was later obtained via an averaging procedure performed on the classical incompressible Euler equations.
Two main boundary conditions were used for \eqref{maineq} in the mathematical literature: the no-slip boundary conditions and the frictionless slip Navier boundary conditions where the fluid is allowed to slip on the boundary without friction. The second boundary condition is more complex but allows for better mathematical results; also it has less physical relevance. The classical well-posedness results for \eqref{maineq} are the following (see \cite{cioranescu_weak_1997,cioranescu_existence_1984,galdi_existence_1993,galdi_further_1994} for the Dirichlet boundary conditions and \cite{busuioc_second_2003} for the Navier boundary conditions):
\begin{itemize}
\item In dimension two there exists a unique global $H^3$ solution if $u_0\in H^3$.
\item In dimension three there exists a unique local $H^3$ solution if $u_0\in H^3$. The solution is global if $u_0$ is small in $H^3$.
\end{itemize}
We call $H^3$ solution a divergence free vector field verifying the boundary conditions and the PDE \eqref{maineq} and who is bounded in time (up to time $t=0$) with values in $H^3(\Omega)$. Let us also mention the paper \cite{bresch_existence_1998} where solutions in $W^{2,p}$, $p>3$, are constructed.
Let us observe that when $\alpha=0$ relation \eqref{maineq} becomes the Navier-Stokes equations
\begin{equation}\label{NS}
\partial_t u-\nu\Delta u+u\cdot\nabla u=-\nabla p
\end{equation}
and when $\alpha=\nu=0$ it becomes the Euler equations. It is interesting to know if the solutions of \eqref{maineq} converge to the solutions of the limit equation when $\alpha\to0$ and $\nu>0$ is fixed or when $\alpha,\nu\to0$. This was already studied in several papers as we shall see below.
Let us first mention that in the absence of boundaries one can obtain $H^3$ estimates uniform in $\alpha$ and $\nu$ in both dimensions two and three and pass to the limit. This was performed in \cite{linshiz_convergence_2010}, see also \cite{busuioc_incompressible_2012} for a simpler proof. But such a result cannot hold true on domains with boundaries. Indeed, if the solutions of \eqref{maineq}-\eqref{2} are bounded in $H^3$ uniformly in $\alpha$ then one can easily pass to the limit $\alpha\to0$ and obtain at the limit a solution of the Navier-Stokes equations which must also be bounded in $H^3$. For such a solution to exist, the initial data must verify a compatibility condition. Indeed, one can apply the Leray projector to \eqref{NS} to obtain that $\partial_t u-\nu\mathbb{P}\Delta u+\mathbb{P}(u\cdot\nabla u)=0$. Since $u$ vanishes at the boundary, so does $\partial_t u$. We infer that $-\nu\mathbb{P}\Delta u+\mathbb{P}(u\cdot\nabla u)=0$ at the boundary. Observe that $u$ being in $H^3$ implies that these two terms are in $H^1$ so the trace at the boundary makes sense. By time continuity we infer that the initial data must verify the compatibility condition $\nu\mathbb{P}\Delta u_0=\mathbb{P}(u_0\cdot\nabla u_0)$ at the boundary. This is of course in general not verified if we only assume that $u_0\in H^3$ is divergence free and vanishing on the boundary.
We review now the results available for domains with boundary.
Concerning the limit $\alpha,\nu\to0$, we proved in \cite{busuioc_incompressible_2012} the expected convergence in 2D for weak $H^1$ solutions in the case of the Navier boundary conditions. We also proved the convergence in 3D but under the additional hypothesis that the solutions exist on a time interval independent of $\alpha$ and $\nu$. We proved in \cite{busuioc_uniform_2016} that the hypothesis of existence of a uniform time existence is verified if $\nu=0$ and $\alpha\to0$. In the case of the Dirichlet boundary conditions, there is only the paper \cite{lopes_filho_convergence_2015} which shows convergence in 2D. The case of Dirichlet boundary conditions in 3D is open.
Concerning the limit $\alpha\to0$ and $\nu>0$ fixed, it was proved in \cite{iftimie_remarques_2002-1} the expected convergence for weak $H^1$ solutions. That result is stated for $\Omega=\mathbb{R}^n$, $n=2,3$, but the proof relies only on energy estimates and standard compactness arguments so it goes through to bounded domains without difficulty (Dirichlet and Navier boundary conditions likewise), see also \cite[Remarque 4]{iftimie_remarques_2002-1}. There is however a major drawback to the result of \cite{iftimie_remarques_2002-1}: the author assumes that the sequence of weak $H^1$ solutions exist on a time interval independent of $\alpha$ and proves convergence on any such time interval. This raises the question of proving the existence of such a uniform time interval. Let us also mention the paper \cite{arada_convergence_2016} where the author considers the 2D case with Navier boundary conditions and shows stronger convergence of solutions together with some estimates for the rate of convergence. Observe however that in dimension two the solutions are global in time so the hypothesis assumed in \cite{iftimie_remarques_2002-1} that the solutions exist on a uniform time interval is automatically satisfied.
In this paper we aim to prove that the hypothesis of \cite{iftimie_remarques_2002-1} about the existence of solutions on a uniform time interval is verified in various situations for Dirichlet boundary conditions. Since global existence of solutions holds true in dimension two for Dirichlet boundary conditions, in 2D the hypothesis of \cite{iftimie_remarques_2002-1} is automatically satisfied and the problem is settled. So we restrict ourselves to the 3D case. Recall that the Navier-Stokes equations in dimension three are locally well-posed for large data and globally well-posed for small data. Likewise, we will prove two results on the uniform time of existence: a local result for large data and a global result for small data.
Surprisingly, we find that if the $H^1$ norm of the initial data is sufficiently small and if $\alpha$ is sufficiently small too, then the solutions of \eqref{maineq} are global. Let us emphasize that the smallness of the initial data is measured only in the $H^1$ norm (in fact in a weaker space, see Theorem \ref{globalthm} in the next section) and not in the $H^3$ norm as required by the classical global well-posedness result for \eqref{maineq}. It was expected to find that, if the $H^1$ norm of the initial data is small then the maximal time of existence of \eqref{maineq} goes to infinity as $\alpha\to0$. Indeed the limit equations, \textit{i.e.} the Navier-Stokes equations, are globally well-posed for small $H^1$ initial data and so we expect convergence for the times of existence of solutions. But we did not expect the maximal time of existence to actually be infinite if $\alpha$ is sufficiently small.
For large data, we prove in particular that if $\alpha$ is sufficiently small then the time existence of the solution has a lower bound that depends only on the $H^2$ norm of the initial velocity. This should be compared with previous results on local existence of solutions where the time of existence depends on the $H^3$ norm or the $W^{2,p}$, $p>3$, norm of the initial velocity. When $\alpha\to0$ we obtain convergence of solutions of \eqref{maineq} towards solutions of \eqref{NS} up to the time $C/\nh1{u_0}^4$.
The plan of the paper is the following. In the next section we state and comment our results. In Section \ref{globalsect} we prove the global result for small data. In the last section, we prove our ``uniform local existence'' result.
\section{Statement of the results}
Since we are interested in the limit $\alpha\to0$ with $\nu$ fixed, we will assume throughout this paper that $\alpha\leq1$ and $\nu>0$.
Let us first state the result of \cite{iftimie_remarques_2002-1} which motivates the present work.
\begin{theorem}[see \cite{iftimie_remarques_2002-1}] \label{teorema}
Let $\nu>0$, $T>0$ and $u_\alpha$ some $H^1$ solutions of \eqref{maineq} defined up to the time $T$ such that $\nh1{u_\alpha(0)}\leq C\alpha^{-\frac12}$ and $u_\alpha(0)$ converges weakly in $L^2$ to some $u_0$. Then there exists a weak Leray solution $u$ of the Navier-Stokes equations with initial velocity $u_0$ such that, after extraction of a sub-sequence,
\begin{equation*}
u_\alpha\rightharpoonup u
\text{ in } L^\infty_{loc}([0,T);L^2) \text{ weak* and in }
L^2_{loc}([0,T);H^1) \text{ weak.}
\end{equation*}
\end{theorem}
Our result about global existence of solutions reads as follows.
\begin{theorem}\label{globalthm}
There exists some small constant $\varepsilon=\varepsilon(\Omega)$ depending only on $\Omega$ such that if the initial velocity $u_0$ belongs to $H^3(\Omega)$, is divergence free, vanishes at the boundary and verifies the following smallness conditions
\begin{gather}
\nl2{u_0}\nh1{u_0}\leq \varepsilon^2\nu^2\label{small1}\\
\intertext{and}
\nl2{u_0}\nh2{u_0}\leq \varepsilon^2\nu^2\alpha^{-\frac12},\qquad
\nh1{u_0}\nh2{u_0}\leq\varepsilon^2\nu^2\alpha^{-1}\quad \text{and}\quad \nh3{u_0} \leq\varepsilon\nu\alpha^{-\frac54},\label{small2}
\end{gather}
then the solution $u$ of \eqref{maineq}-\eqref{3} is global: $u\in L^\infty(\mathbb{R}_+;H^3(\Omega))$.
\end{theorem}
The important thing to observe here is that the smallness conditions in \eqref{small2} involve only negative powers of $\alpha$, so they disappear when $\alpha\to0$. Only the smallness condition \eqref{small1} subsists, and this is in accordance with the classical global well-posedness results for small data for the Navier-Stokes equations. Note that condition \eqref{small1} is scaling invariant and that it implies the smallness of the $H^{\frac12}$ norm of $u_0$ by the interpolation inequality $\nh{\frac12}{u_0}\leq\nl2{u_0}^{\frac12}\nh1{u_0}^{\frac12}$. This in turn gives the existence of a unique global solution of \eqref{2}-\eqref{NS} by the celebrated result of Fujita and Kato \cite{fujita_navier-stokes_1964}. We have the following immediate corollary.
\begin{corollary}
Suppose that $u_0\in H^3(\Omega)$ is divergence free and vanishes on the boundary. There exists some small constant $\varepsilon=\varepsilon(\Omega)$ depending only on $\Omega$ and some constant $\alpha_0>0$ such that if
\begin{equation*}
\nl2{u_0}\nh1{u_0}\leq\varepsilon^2\nu^2\qquad\text{and}\qquad\alpha\leq\alpha_0
\end{equation*}
then the $H^3$ solution of \eqref{maineq}-\eqref{3} is globally defined.
\end{corollary}
Combining this corollary with Theorem \ref{teorema} we immediately obtain the global convergence of solutions of \eqref{maineq} towards solutions of the Navier-Stokes equations if we assume that the initial velocity $u_0$ does not depend on $\alpha$, belongs to $H^3$ and is small in $H^1$. This is just an example of convergence result, more general results can be obtained by allowing the initial data to depend on $\alpha$ and combining Theorems \ref{teorema} and \ref{globalthm}.
Finally, let us conclude our results on global solutions with the observation that the smallness conditions \eqref{small1} and \eqref{small2}
are implied by the simpler but less general condition:
\begin{equation*}
\nh1{u_0}\leq\varepsilon\nu,\qquad \nh2{u_0}\leq\varepsilon\nu\alpha^{-\frac12}\qquad \text{and}\qquad \nh3{u_0} \leq\varepsilon\nu\alpha^{-\frac54}.
\end{equation*}
Indeed, if the above relation holds true one can readily obtain \eqref{small1} and \eqref{small2} simply by estimating $\nl2{u_0}\leq\nh1{u_0}$.
\medskip
We state now our result on the ``uniform local existence of solutions''.
\begin{theorem}\label{local}
There exist two constants $\varepsilon=\varepsilon(\Omega)$ and $K=K(\Omega)$ depending only on $\Omega$ such that if $u_0\in H^3(\Omega)$ is divergence free and vanishes on the boundary and
\begin{equation}\label{hyplocal}
\nh1{u_0}\leq \varepsilon\nu\alpha^{-\frac14}\qquad\text{and}\qquad \nh3{u_0}\leq\varepsilon\nu\alpha^{-\frac54}
\end{equation}
then there exists a $H^3$ solution $u\in L^\infty(0,T;H^3(\Omega))$ of \eqref{maineq}-\eqref{3} which is defined at least up to the time
\begin{equation*}
T=\frac{\nu^3}{K(\nh1{u_0}+\sqrt\alpha\nh2{u_0})^4}.
\end{equation*}
\end{theorem}
Let us emphasize again that in the above theorem the time of existence of the solution depends on the $H^2$ norm of the initial data, but not on the $H^3$ norm or the $W^{2,p}$, $p>3$, norm as in the previous results on local existence of solutions.
Another important thing to note is that both the smallness conditions for $\nh1{u_0}$ and $\nh3{u_0}$ involve negative powers of $\alpha$. So they disappear when taking the limit $\alpha\to0$. More precisely, we have the following corollary.
\begin{corollary}
Suppose that $u_0\in H^3(\Omega)$ is divergence free and vanishes on the boundary. Let $T_\alpha$ be the maximal time of existence of the $H^3$ solution $u$ of \eqref{maineq}-\eqref{3}, \textit{i.e.}
$$u\in L^\infty_{loc}([0,T_\alpha);H^3(\Omega))\setminus L^\infty(0,T_\alpha;H^3(\Omega)).$$
Then
\begin{equation*}
\liminf_{\alpha\to0}T_\alpha\geq \frac{\nu^3}{K\nh1{u_0}^4}.
\end{equation*}
\end{corollary}
Combining this corollary with Theorem \ref{teorema} yields, under the assumptions of the corollary, the convergence of the solutions of \eqref{maineq} towards the solutions of the Navier-Stokes equations up to the time $T= \frac{\nu^3}{K\nh1{u_0}^4}$.
\section{Global existence of solutions}
\label{globalsect}
In this section we show Theorem \ref{globalthm}. In the sequel, we will denote by $C$ a generic constant which depends only on $\Omega$ and whose value can change from one line to another. We will use the standard notation for the $H^m$ norms
\begin{equation*}
\|u\|_{H^m}=\Bigl(\sum_{\beta\in\mathbb{N}^3,|\beta|\leq m}\int_\Omega|\partial^\beta u(x)|^2\,dx\Bigr)^{\frac12}.
\end{equation*}
We denote by $\mathbb{P}$ the Leray projector, \textit{i.e.} the $L^2$ orthogonal projector on the subspace of divergence free and tangent to the boundary vector fields.
\medskip
We will prove that there exists a sufficiently small constant $\varepsilon_1=\varepsilon_1(\Omega)$ depending only on $\Omega$ such that if
\begin{equation}\label{cond}
\lip u\leq\frac\nu{2\alpha}\qquad\text{and}\qquad \nl3u\leq\varepsilon_1 \nu
\end{equation}
on some time interval $[0,T]$, then the two quantities above are even smaller:
\begin{equation}\label{condimpl}
\lip u\leq\frac\nu{4\alpha}\qquad\text{and}\qquad \nl3u\leq\frac{\varepsilon_1 \nu}2
\end{equation}
on the same time interval $[0,T]$. By time continuity, this implies that \eqref{cond} never breaks down if it is verified at the initial time and all the estimates that follow hold true globally in time. Let us now observe that \eqref{cond} holds true at the initial time if $\varepsilon$ and $\varepsilon_1$ are sufficiently small independently of $\alpha$ and $\nu$. We use the Gagliardo-Nirenberg inequality and \eqref{small1} to write
\begin{equation*}
\nl3{u_0}\leq C\nl2{u_0}^{\frac12}\nh1{u_0}^{\frac12} \leq C\varepsilon\nu< \varepsilon_1\nu
\end{equation*}
provided that $C\varepsilon<\varepsilon_1$. This is a condition that we assume in what follows. Next, we use an interpolation inequality and relation \eqref{small2} to estimate
\begin{equation*}
\nh1{u_0}\leq \nl2{u_0}^{\frac12}\nh2{u_0}^{\frac12}\leq \varepsilon\nu\alpha^{-\frac14}.
\end{equation*}
From the Gagliardo-Nirenberg inequality and \eqref{small2} we deduce that
\begin{equation*}
\lip{u_0}\leq C\nh1{u_0}^{\frac14}\nh3{u_0}^{\frac34}\leq C\frac{\varepsilon\nu}\alpha< \frac\nu{2\alpha}
\end{equation*}
provided that $\varepsilon<\frac1{2C}$ which is another condition that we will assume in the sequel. We conclude from the above estimates that \eqref{cond} holds true at the initial time with strict inequality provided that $C\varepsilon<\varepsilon_1$ and $\varepsilon<\frac1{2C}$.
\medskip
We assume in the following that \eqref{cond} holds true. Because we need a precise dependence on $\alpha$ of the constants, we have to split the estimates for $u$ in three different parts: $H^1$ estimates, $H^2$ estimates and $H^3$ estimates. The estimates below should be viewed as \textit{a priori} estimates. They can be turned into rigorous estimates via a standard Galerkin approximation procedure.
\subsection*{$H^1$ estimates.} We multiply \eqref{maineq} by $u$ and integrate in space to obtain
\begin{equation*}
\frac12 \partial_t(\nl2u^2+\alpha\nl2{\nabla u}^2)+\nu\nl2{\nabla u}^2=0
\end{equation*}
so, after an integration in time,
\begin{equation}\label{h1est}
\nl2{u(t)}^2+\alpha\nl2{\nabla u(t)}^2+2\nu\int_0^t\nl2{\nabla u(s)}^2\,ds= \nl2{u_0}^2+\alpha\nl2{\nabla u_0}^2.
\end{equation}
\subsection*{$H^2$ estimates.} To perform $H^2$ estimates on \eqref{maineq} we need to write it under a different form. First we remove the pressure by applying the Leray projector $\mathbb{P}$. We obtain the following equivalent equation:
\begin{equation*}
\partial_t (u-\alpha\mathbb{P}\Delta u)-\nu\mathbb{P}\Delta u +\mathbb{P}\bigl[u\cdot\nabla(u-\alpha\Delta u)\bigr]+\mathbb{P}\bigl[\sum_j(u-\alpha\Delta u)_j\nabla u_j\bigr]=0.
\end{equation*}
We observe first that
\begin{equation*}
\mathbb{P} \bigl(\sum_j u_j\nabla u_j\bigr)=\frac12\mathbb{P}\nabla(|u|^2)=0.
\end{equation*}
Next, we have that
\begin{equation*}
u\cdot\nabla\nabla q+\sum_j\partial_j q\nabla u_j=\nabla(u\cdot\nabla q)
\end{equation*}
is a gradient so
\begin{equation*}
\mathbb{P}\Bigl( u\cdot\nabla\nabla q+\sum_j\partial_j q\nabla u_j\Bigr)=0
\end{equation*}
for any $q$. Recalling that $\mathbb{P}\Delta u-\Delta u$ is a gradient, we infer that \eqref{maineq} can be written under the following equivalent form:
\begin{equation*}
\partial_t (u-\alpha\mathbb{P}\Delta u)-\nu\mathbb{P}\Delta u +\mathbb{P}\bigl[u\cdot\nabla(u-\alpha\mathbb{P}\Delta u)\bigr]-\alpha\mathbb{P}\bigl[\sum_j(\mathbb{P}\Delta u)_j\nabla u_j\bigr]=0.
\end{equation*}
We multiply the above equation by $-\mathbb{P}\Delta u$ and integrate in space. We recall that $\mathbb{P}$ is a self-adjoint projector. We have that $\mathbb{P}^2=\mathbb{P}$ and $\mathbb{P} u=u$ because $u$ is divergence free and tangent to the boundary. We infer that
\begin{align*}
\frac12\partial_t(\nl2{\nabla u}^2+&\alpha\nl2{\mathbb{P}\Delta u}^2)+\nu\nl2{\mathbb{P}\Delta u}^2\\
&=\int_\Omega \mathbb{P}\bigl[u\cdot\nabla(u-\alpha\mathbb{P}\Delta u)\bigr]\cdot \mathbb{P}\Delta u
-\alpha\int_\Omega \mathbb{P}\bigl[\sum_j(\mathbb{P}\Delta u)_j\nabla u_j\bigr] \cdot \mathbb{P}\Delta u \\
&=\int_\Omega u\cdot\nabla(u-\alpha\mathbb{P}\Delta u)\cdot \mathbb{P}\Delta u
-\alpha\int_\Omega \sum_j(\mathbb{P}\Delta u)_j\nabla u_j \cdot \mathbb{P}\Delta u.
\end{align*}
Using the cancellation
\begin{equation*}
\int_\Omega u\cdot\nabla\mathbb{P}\Delta u\cdot \mathbb{P}\Delta u=0
\end{equation*}
we infer that
\begin{equation}\label{h2diffeq}
\frac12\partial_t(\nl2{\nabla u}^2+\alpha\nl2{\mathbb{P}\Delta u}^2)+\nu\nl2{\mathbb{P}\Delta u}^2
=\int_\Omega u\cdot\nabla u\cdot \mathbb{P}\Delta u
-\alpha\int_\Omega \sum_j(\mathbb{P}\Delta u)_j\nabla u_j \cdot \mathbb{P}\Delta u.
\end{equation}
We bound
\begin{equation}\label{h2diffeq2}
- \alpha\int_\Omega \sum_j(\mathbb{P}\Delta u)_j\nabla u_j \cdot \mathbb{P}\Delta u
\leq \alpha\nl2{\mathbb{P}\Delta u}^2\lip u
\leq \frac\nu2 \nl2{\mathbb{P}\Delta u}^2
\end{equation}
and
\begin{align*}
\int_\Omega u\cdot\nabla u\cdot \mathbb{P}\Delta u
&\leq \nl3u\nl6{\nabla u}\nl2{\mathbb{P}\Delta u}\\
&\leq C\nl3u\nh1{\nabla u} \nl2{\mathbb{P}\Delta u} \\
&\leq C\nl3u\nl2{\mathbb{P}\Delta u}^2\\
&\leq C\varepsilon_1\nu \nl2{\mathbb{P}\Delta u}^2\\
&\leq \frac\nu4\nl2{\mathbb{P}\Delta u}^2
\end{align*}
provided that $\varepsilon_1\leq1/4C$ which is the only condition we will impose on $\varepsilon_1$. We used above the Sobolev embedding $H^1\subset L^6$ and the classical regularity result for the stationary Stokes operator which claims that $\nl2{\mathbb{P}\Delta u}\simeq\nh2u$.
We infer from the previous relations that
\begin{equation*}
\partial_t(\nl2{\nabla u}^2+\alpha\nl2{\mathbb{P}\Delta u}^2)+\frac\nu2\nl2{\mathbb{P}\Delta u}^2 \leq0
\end{equation*}
so, after an integration in time,
\begin{equation}\label{h2est}
\nl2{\nabla u(t)}^2+\alpha\nl2{\mathbb{P}\Delta u(t)}^2+\frac\nu2\int_0^t\nl2{\mathbb{P}\Delta u(s)}^2\,ds \leq \nl2{\nabla u_0}^2+\alpha\nl2{\mathbb{P}\Delta u_0}^2.
\end{equation}
\subsection*{$H^3$ estimates.}
Let us introduce the notation
\begin{equation*}
\omega=\curl u,\qquad \omega_\alpha=\omega-\alpha\Delta \omega.
\end{equation*}
To perform the $H^3$ estimates, we apply the curl operator to \eqref{maineq}. We obtain the following PDE:
\begin{equation*}
\partial_t \omega_\alpha-\nu\Delta\omega+u\cdot\nabla\omega_\alpha-\omega_\alpha\cdot\nabla u=0.
\end{equation*}
We multiply by $\omega_\alpha$ and integrate in space to obtain
\begin{align*}
\frac12\partial_t\nl2\omega_\alpha^2-\nu\int_\Omega\Delta\omega\cdot\omega_\alpha
&=\int_\Omega\omega_\alpha\cdot\nabla u\cdot\omega_\alpha\\
&\leq\nl2\omega_\alpha^2\lip u\\
&\leq\frac\nu{2\alpha}\nl2\omega_\alpha^2.\notag
\end{align*}
Next, since $\omega_\alpha=\omega-\alpha\Delta\omega$ we can write
\begin{equation*}
-\nu\int_\Omega\Delta\omega\cdot\omega_\alpha
= -\frac\nu\alpha\int_\Omega(\omega-\omega_\alpha)\cdot\omega_\alpha
=\frac{3\nu}{4\alpha}\nl2\omega_\alpha^2+\frac\nu\alpha\bigl\|\frac\oma2-\omega\bigr\|^2_{L^2}-\frac\nu\alpha\nl2\omega^2.
\end{equation*}
We infer that
\begin{equation*}
\frac12\partial_t\nl2\omega_\alpha^2 +\frac{3\nu}{4\alpha}\nl2\omega_\alpha^2+\frac\nu\alpha\bigl\|\frac\oma2-\omega\bigr\|^2_{L^2}-\frac\nu\alpha\nl2\omega^2 \leq\frac\nu{2\alpha}\nl2\omega_\alpha^2
\end{equation*}
so
\begin{equation*}
\partial_t\nl2\omega_\alpha^2+\frac\nu{2\alpha}\nl2\omega_\alpha^2\leq \frac{2\nu}\alpha\nl2\omega^2= \frac{2\nu}\alpha\nl2{\nabla u}^2.
\end{equation*}
The Gronwall inequality implies that
\begin{align}
\nl2{\omega_\alpha(t)}^2
&\leq \nl2{\omega_\alpha(0)}^2e^{-t\frac\nu{2\alpha}}
+\frac{2\nu}\alpha\int_0^te^{(s-t)\frac\nu{2\alpha}} \nl2{\nabla u(s)}^2\,ds\notag\\
&\leq \nl2{\omega_\alpha(0)}^2e^{-t\frac\nu{2\alpha}}
+\frac{2\nu}\alpha\sup_{s\in[0,t]}\nl2{\nabla u(s)}^2\int_0^te^{(s-t)\frac\nu{2\alpha}} \,ds\notag\\
&\leq \nl2{\omega_\alpha(0)}^2
+4\sup_{s\in[0,t]}\nl2{\nabla u(s)}^2\label{h3diffeq2}\\
&\leq \nl2{\omega_\alpha(0)}^2
+4\nl2{\nabla u_0}^2+4\alpha\nl2{\mathbb{P}\Delta u_0}^2\notag
\end{align}
where we used the bound for $\nl2{\nabla u}$ given in \eqref{h2est}. Adding this relation to \eqref{h2est} implies that
\begin{equation}\label{finalbound}
F(t)\leq CF(0)
\end{equation}
where
\begin{equation*}
F(t)=\nl2{\nabla u(t)}+\sqrt\alpha\nl2{\mathbb{P}\Delta u(t)}+\nl2{\omega_\alpha(t)}.
\end{equation*}
The following lemma holds true:
\begin{lemma}\label{lemma1}
We have that
\begin{equation*}
F(t)\simeq \nh1{u(t)}+\alpha\nh3{u(t)}
\end{equation*}
with constants depending only on $\Omega$.
\end{lemma}
\begin{proof}
Because $\mathbb{P}$ is bounded in $L^2$ and $\omega_\alpha=\curl u-\alpha\Delta\curl u$ we immediately get the bound
\begin{equation*}
F\leq C(\nh1{u}+\sqrt\alpha\nh2{u}+\alpha\nh3{u}).
\end{equation*}
From the interpolation inequality $\nh2u\leq \nh1u^{\frac12}\nh3u^{\frac12}$ we deduce that $\sqrt\alpha\nh2u\leq \nh1u+\alpha\nh3u$ so
\begin{equation*}
F\leq C(\nh1{u}+\alpha\nh3{u}).
\end{equation*}
To prove the reverse bound we observe first by the Poincaré inequality that $\nl2{\nabla u}\simeq\nh1u$.
Therefore, it suffices to show that
\begin{equation*}
F\geq C\alpha\nh3u.
\end{equation*}
Observing that $\nl2{\nabla u}=\nl2\omega$ we can write
\begin{equation*}
\nl2\omega_\alpha=\nl2{\omega-\alpha\Delta\omega}\geq\alpha\nl2{\Delta \omega}-\nl2\omega =\alpha\nl2{\curl\Delta u}-\nl2{\nabla u}.
\end{equation*}
Next, we recall that $\Delta u-\mathbb{P}\Delta u$ is a gradient so $\curl\Delta u=\curl\mathbb{P}\Delta u$. We infer from the previous relations that
\begin{equation*}
F\geq \sqrt\alpha\nl2{\mathbb{P}\Delta u}+\alpha\nl2{\curl\mathbb{P}\Delta u}
\geq \alpha(\nl2{\mathbb{P}\Delta u}+\nl2{\curl\mathbb{P}\Delta u} )
\end{equation*}
where we used that $\alpha\leq1$. Since $\mathbb{P}\Delta u$ is divergence free and tangent to the boundary, one can apply \cite[Proposition 1.4]{foias_remarques_1978} to deduce that
\begin{equation*}
\nl2{\mathbb{P}\Delta u}+\nl2{\curl\mathbb{P}\Delta u}\geq C\nh1{\mathbb{P}\Delta u}.
\end{equation*}
Finally, the classical regularity result for the stationary Stokes operator says in particular that
\begin{equation*}
\nh1{\mathbb{P}\Delta u}\geq C\nh3u.
\end{equation*}
We infer that $F\geq C\alpha\nh3u$ and this completes the proof of the lemma.
\end{proof}
We go back to the proof of Theorem \ref{globalthm}. From the Gagliardo-Nirenberg estimate
\begin{equation*}
\lip u\leq C\nh1u^{\frac14}\nh3u^{\frac34},
\end{equation*}
from relation \eqref{finalbound} and from Lemma \ref{lemma1} we infer that
\begin{equation*}
\alpha\lip {u(t)}\leq C\alpha^{\frac14}F(t)\leq C\alpha^{\frac14}F(0)\leq C(\alpha^{\frac14}\nh1{u_0}+\alpha^{\frac54}\nh3{u_0}).
\end{equation*}
From the Gagliardo-Nirenberg and Poincaré inequalities and from relations \eqref{h1est} and \eqref{h2est} we can bound
\begin{equation*}
\nl3u\leq C\nl2u^{\frac12}\nh1u^{\frac12}\leq C\nl2u^{\frac12}\nl2{\nabla u}^{\frac12}\leq C (\nl2{u_0}^2+\alpha\nl2{\nabla u_0}^2)^{\frac14}( \nl2{\nabla u_0}^2+\alpha\nl2{\mathbb{P}\Delta u_0}^2)^{\frac14}.
\end{equation*}
We conclude that for \eqref{condimpl} to be verified it suffices to assume the following conditions on the initial data:
\begin{gather}
C(\alpha^{\frac14}\nh1{u_0}+\alpha^{\frac54}\nh3{u_0})\leq \frac\nu{4}\label{small11}\\
\intertext{and}
C (\nl2{u_0}^2+\alpha\nl2{\nabla u_0}^2)^{\frac14}( \nl2{\nabla u_0}^2+\alpha\nl2{\mathbb{P}\Delta u_0}^2)^{\frac14}\leq \frac{\varepsilon_1\nu}2. \label{small22}
\end{gather}
We prove now that if our smallness assumptions \eqref{small1} and \eqref{small2} hold true for a sufficiently small constant $\varepsilon$, then \eqref{small11} and \eqref{small22} are verified. Observe first that \eqref{small2} implies \eqref{small11}. Indeed, this follows immediately from the interpolation inequality $\nh1{u_0}\leq\nl2{u_0}^{\frac12}\nh2{u_0}^{\frac12}$. To prove \eqref{small22}, we notice that it is equivalent to
\begin{equation*}
\nl2{u_0}\nh1{u_0}+\alpha^{\frac12}\nh1{u_0}^2+\alpha^{\frac12}\nl2{u_0}\nh2{u_0}+\alpha\nh1{u_0}\nh2{u_0}\leq\frac{\varepsilon_1^2\nu^2}{C}
\end{equation*}
for some constant $C$. The necessary bounds for the first, the third and the fourth term on the left-hand side are included in the hypothesis \eqref{small1} and \eqref{small2} for $\varepsilon$ small enough. The bound for the second term shows up in \eqref{small11}, so it is already proved.
We conclude that if \eqref{small1} and \eqref{small2} hold true for a sufficiently small constant $\varepsilon$, then \eqref{cond} never fails to be true and the solution $u$ exists globally. This completes the proof of Theorem \ref{globalthm}.
\section{Uniform local existence of solutions}
We show in this section Theorem \ref{local}.
To prove the uniform local existence of solutions, we will adapt the estimates from the previous section in the following manner.
We search for a constant $M$ and a time $T$ such that if we assume that
\begin{equation}\label{concl0}
\nl2{\nabla u}\leq M\text{ and }\nl\infty{\nabla u}\leq\frac\nu{2\alpha}
\end{equation}
on some sub-interval $[0,T']\subset[0,T]$ then
\begin{equation}\label{concl}
\nl2{\nabla u}\leq \frac M2\text{ and }\nl\infty{\nabla u}\leq\frac\nu{4\alpha}
\end{equation}
on the same sub-interval $[0,T']$. By time continuity it follows that relation \eqref{concl} holds true on $[0,T]$ if it holds true at the initial time. As in the previous section one easily checks that the second condition from \eqref{concl0} is verified at time $t=0$ if $\varepsilon$ is sufficiently small. Indeed, the condition on the Lipschitz norm of $u_0$ is the same as in \eqref{cond} and in Section 3 we proved it using bounds for $\nh1{u_0}$ and $\nh3{u_0}$ which are the same as those from \eqref{hyplocal}. For the first bound in \eqref{concl0} to hold true at time $t=0$ we impose the following condition on the constant $M$ (recall that $M$ will be chosed later):
\begin{equation*}
\nl2{\nabla u_0}< M.
\end{equation*}
So let us assume that \eqref{concl0} holds true and let us prove relation \eqref{concl}. We adapt the estimates from the previous section.
The $H^1$ estimates are not needed anymore. The $H^2$ estimates must be modified. Relations \eqref{h2diffeq} and \eqref{h2diffeq2} remain valid and imply that
\begin{equation}\label{loch2}
\partial_t(\nl2{\nabla u}^2+\alpha\nl2{\mathbb{P}\Delta u}^2)+\nu\nl2{\mathbb{P}\Delta u}^2
\leq 2\int_\Omega u\cdot\nabla u\cdot \mathbb{P}\Delta u.
\end{equation}
The estimate of the right-hand side must be modified in the following manner
\begin{align*}
2\int_\Omega u\cdot\nabla u\cdot \mathbb{P}\Delta u
&\leq 2\nl\infty u\nl2{\nabla u}\nl2{\mathbb{P}\Delta u}\\
&\leq C\nh1u^{\frac32}\nh2u^{\frac12} \nl2{\mathbb{P}\Delta u} \\
&\leq C\nl2{\nabla u}^{\frac32} \nl2{\mathbb{P}\Delta u}^{\frac32}\\
&\leq \frac\nu2 \nl2{\mathbb{P}\Delta u}^2+\frac{C_1}{\nu^3}\nl2{\nabla u}^6.
\end{align*}
for some constant $C_1$. We used the Poincaré and Young inequalities, the Gagliardo-Nirenberg estimate $\nl\infty u\leq C\nh1u^{\frac12}\nh2u^{\frac12}$ and the relation $\nh2u\leq C\nl2{\mathbb{P}\Delta u}$. Using this estimate in \eqref{loch2} yields
\begin{equation*}
\partial_t(\nl2{\nabla u}^2+\alpha\nl2{\mathbb{P}\Delta u}^2)+\frac\nu2\nl2{\mathbb{P}\Delta u}^2
\leq \frac{C_1}{\nu^3}\nl2{\nabla u}^6.
\end{equation*}
Recalling that we assumed $\nl2{\nabla u}\leq M$ and integrating in time yields
\begin{equation*}
\nl2{\nabla u(t)}^2+\alpha\nl2{\mathbb{P}\Delta u(t)}^2\leq \nl2{\nabla u_0}^2+\alpha\nl2{\mathbb{P}\Delta u_0}^2+\frac{C_1tM^6}{\nu^3}
\end{equation*}
If we make the assumption that
\begin{equation}\label{boundm2}
\nl2{\nabla u_0}^2+\alpha\nl2{\mathbb{P}\Delta u_0}^2+\frac{C_1TM^6}{\nu^3}\leq \frac{M^2}4
\end{equation}
then we get that
\begin{equation}\label{17a}
\nl2{\nabla u(t)}^2+\alpha\nl2{\mathbb{P}\Delta u(t)}^2\leq \frac{M^2}4.
\end{equation}
In particular, we have that
\begin{equation*}
\nl2{\nabla u(t)}\leq \frac M2
\end{equation*}
which implies the first half of \eqref{concl}.
The $H^3$ estimates remain valid up to the relation \eqref{h3diffeq2} that we recall now
\begin{equation*}
\nl2{\omega_\alpha(t)}^2\leq\nl2{\omega_\alpha(0)}^2+4\sup_{s\in[0,t]}\nl2{\nabla u(s)}^2.
\end{equation*}
We can further bound $\nl2{\nabla u(s)}\leq M$ to obtain that
\begin{equation*}
\nl2{\omega_\alpha(t)}^2\leq\nl2{\omega_\alpha(0)}^2+4M^2.
\end{equation*}
Adding this estimate to \eqref{17a} yields
\begin{equation*}
F(t)\leq CF(0)+CM.
\end{equation*}
We observed in the previous section that $\lip u\leq C\alpha^{-\frac34}F$ so we deduce that
\begin{equation*}
\lip{u(t)}\leq C\alpha^{-\frac34}(F(0)+M)
\leq C_2 (\alpha^{-\frac34}\nh1{u_0}+\alpha^{\frac14}\nh3{u_0}+\alpha^{-\frac34}M)
\end{equation*}
for some constant $C_2$, where we used Lemma \ref{lemma1}.
We conclude that if we assume \eqref{boundm2} together with
\begin{equation}\label{boundm3}
C_2 (\alpha^{-\frac34}\nh1{u_0}+\alpha^{\frac14}\nh3{u_0}+\alpha^{-\frac34}M)\leq\frac\nu{4\alpha}
\end{equation}
then \eqref{concl} holds true.
To satisfy the condition \eqref{boundm2} it suffices to assume that
\begin{equation}\label{f1}
\nl2{\nabla u_0}\leq \frac M{2\sqrt3}\qquad\text{and}\qquad \sqrt\alpha\nl2{\mathbb{P}\Delta u_0}\leq \frac M{2\sqrt3}
\end{equation}
and to choose $T$ such that
\begin{equation*}
\frac{C_1TM^6}{\nu^3}=\frac{M^2}{12}
\end{equation*}
that is
\begin{equation}\label{f3}
T=\frac{\nu^3}{12C_1M^4}.
\end{equation}
On the other hand, to ensure that \eqref{boundm3} holds true it suffices to assume that
\begin{equation}\label{f2}
\nh1{u_0}\leq \frac\nu{12C_2\alpha^{\frac14}}, \qquad \nh3{u_0}\leq \frac\nu{12C_2\alpha^{\frac54}}\qquad\text{and}\qquad M\leq \frac\nu{12C_2\alpha^{\frac14}}.
\end{equation}
Clearly, to be able to find a constant $M$ such that \eqref{f1} and \eqref{f2} are satisfied we need to assume the following conditions on the initial data:
\begin{equation}\label{condi}
\nh1{u_0}\leq \varepsilon\nu\alpha^{-\frac14}, \qquad \nh2{u_0}\leq \varepsilon\nu\alpha^{-\frac34}\qquad\text{and}\qquad \nh3{u_0}\leq \varepsilon\nu\alpha^{-\frac54}
\end{equation}
for a sufficiently small $\varepsilon$.
In fact, the condition on the $H^2$ norm is not necessary since the interpolation inequality $\|\cdot\|_{H^2}\leq\|\cdot\|_{H^1}^{\frac12}\|\cdot\|_{H^3}^{\frac12}$ shows that it follows from the other two conditions.
Once \eqref{condi} is satisfied, we can choose
\begin{equation*}
M=C_3(\nh1{u_0}+\sqrt\alpha\nh2{u_0})
\end{equation*}
for some constant $C_3$. According to \eqref{f3}, this gives a time of existence of the solution of the form
\begin{equation*}
T=\frac{\nu^3}{12C_1C_3^4 (\nh1{u_0}+\sqrt\alpha\nh2{u_0})^4 }.
\end{equation*}
This completes the proof of Theorem \ref{local}.
\section*{Acknowledgments}
This work was supported by the LABEX MILYON (ANR-10-LABX-0070) of Université de Lyon, within the program "Investissements d'Avenir" (ANR-11-IDEX-0007) operated by the French National Research Agency (ANR).
|
\section{Introduction}
\elsA/ (\url{http://elsa.onera.fr/}) is a large
\sigle{CFD} software for research and industry, mainly used in
aerospace design, for both internal and external flow. It has been
described before \citep{Gazaix2002,Cambier2011}.
Here we will be interested in the extension of the \elsAPython/ user
interface to additional software, for performing simulations not only
at a few discrete specified workflow conditions (e.g. given Mach,
Reynolds \ldots) but on a domain of variation, i.e. a Design of
Experiment space (\DoE/), with several algorithmic layers added above
the \sigle{CFD} solver.
The first step is to automate the spanning of a (given) large number
of specified points covering the \DoE/. The second one is to add a
stabilization layer for prolongation through unstable \DoE/ zones
(e.g. when the flow conditions lead to separation). The third one is
to use the resulting ``stabilized spanning'' algorithm as a provider
of \emph{observable quantities} for a sparse polynomial interpolator,
driving the simulation and selecting the \DoE/ points to compute. If
successful, this composite algorithm provides (for the chosen
observable) an efficient global representation (response surface)
allowing further studies on the \DoE/, at low cost; for example
stochastic analysis using the Monte-Carlo method on the response
surface would be a fourth step.
The specific point of interest of our approach is that, for this
complex composite algorithm, we manage to keep the user interface
simple and uniform, using the same concepts as for the base interface.
Sections \ref{orient}-\ref{statdyn} explain these concepts; sections
\ref{doc}\&\ref{pygelsa} deal with documentation, error messages, and
the \sigle{GUI} (Graphical User Interface); finally, sections
\ref{datanet}-\ref{doe} deal with high-level operations and how we use
them for \DoE/ spanning and analysis.
\enlargethispage{1cm}
This paper includes a number of elements from the \sigle{CFD}
application domain, and software considerations, along with user
interface concepts; these concepts (and most of the software
implementation) would be applicable to many different domains. Not
included here are considerations for field-like data, for which the
\sigle{CFD} General Notation System standard (\sigle{CGNS})
is better suited\footnote{But \sigle{CGNS} is not a user
interface system, lacking even basic user-oriented checks}.
\section{Interface development orientation}
\label{orient}
\subsection{User interaction model}
In the development of the \elsAPython/ user interface, we strived to
reach an equilibrium between power and usability, both on the user
side (running simulations) and on the developer side (adding
functionality); also, we have tried to avoid a steep user learning
curve.
The \elsAPython/ user interface has been used since June 2000 with no
change in its base principles (see section \ref{base}), and thus maybe
it is not too far off its target demography. The basic interface is
readily extended with additional classes allowing external algorithmic
layers, see \ref{products}, to complement the \sigle{CFD} solver coded
in the \elsA/ kernel.
The \elsAPython/ interface is built on a \emph{declarative} user
interaction model \citep{declarative}, whereas similar software
\citep{Gerhold2008} uses an \emph{imperative} model
\citep{imperative}. The scripts describing the \elsA/ \sigle{CFD}
simulations use context sharing to avoid data duplication; and all the
tedious details which are possibly automated are kept hidden in the
behaviour of the description classes.
The Python language has been a great help in building successive
abstract layers. We generally try to go with the Principle of Least
Astonishment \citep{Seebach2001,Ronacher2011}, and to facilitate an
object-oriented and \emph{functional}
\citep{functional,Backus1978b,Hudak1989} programming style, here in
user scripts.
The (functional) Reverse Polish Lisp (\sigle{RPL}) language of some
Hewlett-Packard handheld calculators, with its three language levels
(User \sigle{RPL}, System \sigle{RPL}, compiled code) \citep{RPLwiki},
has been the inspiration for an intermediate language where, as for
System \sigle{RPL}, a faster checks-less interpreted code -- usually
generated on the fly from the User level -- may be directly accessed
by experts.
\medskip
In \elsAPython/, user interaction is mainly through argument-less
method calls; this is possible because the would-be arguments are
already known as attribute values of the method's owner object; these
values may originate:
\begin{enumerate}[a)]
\item from user specification, using the owner object's \comput{set()}
method, see \ref{methods_desc};
\item from context rules -- \comput{set()} terminator in
context-dependent default rules for the owner object -- possibly
involving dependencies on objects other than the method's owner,
see \ref{context_manage};
\item using the \comput{set()} method in the behaviour logic of other
objects, possibly also involving context rules (e.g. in coupled
problems), see \ref{sfddmd}.
\end{enumerate}
The user interaction model thus includes an explicit part (method
calls in the user script) and an implicit part (context rules and
other class-specific object behaviour), providing
automatically-defined values. The latter part is essential, allowing
for a lighter work burden and thus better efficiency and security in
problem solving. The potential danger, regarding user confidence, of
automatically-defined attribute values is alleviated by the
\emph{display} (\comput{view} method, see \ref{methods_cntx}) and
\emph{introspection} features of the interface (\comput{show\_origin}
method, see \ref{contdefs}).
\subsection{Base principles, user \& developer sides}
\label{base}
The base elements and principles used in the development of this
interface are:
\smallskip
\noindent\emph{For the user (U) side:}
\begin{enumerate}[U1.]
\item declarative model: the problem definition is made up of
definitions of the simulation parameters; building the adequate
solver (acting) is left to a \emph{factory}, lower down in the C++
\elsA/ kernel; this separation of roles shields the interface from
modifications in the kernel implementation details;
\item object-oriented model: the simulation parameters (elements of
the mathematical model) are here represented as attributes of
\emph{description} classes; the whole simulation \emph{script} (a
\emph{description}s container) is also a class instance;
\emph{script}s include no data, only \emph{description} creations
and a few operations, like \comput{<desc>.set(<attribute>,
<value>)}, \comput{<desc>.check()} or \comput{<desc>.compute()};
\item implicit hierarchy: the whole simulation data structure is a
shallow (four-level) tree of scripts (possibly nested),
descriptions, attributes, and values (see \ref{static_tree}); this
tree is not explicitly referenced in interface use\footnote{Unlike
a \sigle{CGNS} tree -- managed in \elsA/ through \comput{py\sigle{CGNS}}
-- which large depth is especially apparent when it is accessed
through the \comput{\sigle{CGNS}view} graphical interface.};
\item static checks (attribute name and value type/definition
domain) are performed by default on attributes , see
\ref{statdefs};
\item dynamic (contextual) checks are performed on demand to ensure
the coherency of the problem description; a
\emph{context}\footnote{Contextual (influence/dependence)
relations introduce a different graph than the above shallow
tree; this second graph, traversed by the \comput{check()}
method, see \ref{methods_cntx}, may be deep (have a large number
of levels), see \ref{dynamic}, and is not always a tree, see
\ref{context_manage}.} is an instance of either a
\emph{description} class or a \emph{script} class, both owning a
\comput{check()} method.
\item context-dependent (contextual) default values allow tailoring
the software to different trades, see \ref{contdefs}; the origin
rule for default values may be traced back through the
\comput{show\_origin()} method;
\item keep the user in charge: no automatic modification of his/her
input; conflicts may only be resolved by the user, upon
notification by the interface;
\item the interface may be used as a standalone checking tool, with
no \sigle{CFD} kernel needed;
\item the user interface scope is limited to scalar data: mesh and
field data (e.g. files) is referenced without contents checking.
\end{enumerate}
\pagebreak[2]
\noindent\emph{For the developer (D) side:}
\begin{enumerate}[D1.]
\item the \sigle{API} (Application Programming Interface) between
the user interface and the \elsA/ kernel is based on three basic
types -- \comput{float} (floating-point), \comput{int}
(integer) and \comput{string} -- and a few \emph{methods} and
\emph{functions}, see \ref{methfunc};
\item the evolving part of the interface is mainly described through
resource files, see section \ref{statdyn}, from which the
documentation skeleton ({\LaTeX} commands, see \ref{urm}) is
automatically generated in a coherent way;
\item functionality may be added through ``products'', see
\ref{products}, the Python code for which is simply dropped in the
adequate place, defining additional \emph{description} classes;
\item obsolete elements of the interface, see \ref{metadata}, may be
re-activated for version comparison.
\end{enumerate}
\subsection{Naming things}
\label{naming}
One important aspect of the interface -- for which there are no exact
rules and only a few general principles -- is the \emph{naming} of the
various interface elements (especially attributes and their values),
see \ref{naming}. This aspect is more in the realm of linguistics, and
must take into account the culture of the users.
Naming is recognized as a hard problem in software
\citep{Deissenboeck2006}, and is all the more important when dealing
with user interaction. In programming, users better understand long,
self-explanatory, names \citep{Binkley2009a}, but ``The nonsense words
did surprisingly well \ldots distinctive names are helpful even when
they are not meaningful'' \citep{Shneiderman1997}.
It seems then that there is no symmetry between understanding which
concept is behind a name and, inversely, remembering which name
corresponds to the sought concept. It is probably better not to use
long names which may differ only by one or two characters -- better
use shorter names where the differences stands out -- or to use names
with easily confused characters \citep{Kupferschmid2009}. Also, we
have chosen the underscore style, e.g. \comput{global\_timestep}, for
the user interface rather than ``camel casing'',
e.g. \comput{GlobalTimestep}, although the latter seems more popular
with programmers \citep{Binkley2009b}.
A balance must somehow be found, for users to be able both to read and
write scripts without reaching for the documentation at every
step. This part of the interface management is probably the most
involved with language considerations, and the most dependent on user
cultural background (e.g. when reusing variable names from equations).
\section{\elsAPython/ description language}
\label{elsAPython}
\elsAPython/ is built upon the Python language to create a standard of
creation and modification of description (attributes container) and
script (descriptions container) objects. Only the main elements of the
language will be described here, hoping they give a taste of the
chosen \emph{user interaction model}. Both for the text-mode and
graphical-mode interfaces, we have tried to respect the design
principles of \citep{Shneiderman1997}, with \citep{Clarke1986} also as
a (more abstract) background.
From then on, the \comput{<>} brackets will be used with general
meaning of ``realization'' (instance, value, \ldots) of the enclosed
symbol (class, attribute, \ldots).
\subsection{Description object creation}
Creating an instance of the \comput{<desc>} description class is performed by:
\comput{<name> = <desc>(name='<name>')}
where \comput{<name>} (without quotes) is the canonical Python
reference to the created object and \comput{'<name>'} a text
identifier using the same characters. This identifier is used both for
allowing forward references (to not yet created objects) and for
out-of-memory references, either between computational nodes without
shared memory or in databases (for accessing objects directly, not
through property search), see \ref{database}.
Notice: simple test scripts may forgo the \comput{<name>=} argument,
thanks to an automatic canonical name search feature, but this is very
costly for large scripts and is precluded in industrial cases.
\subsection{Script object creation}
The \comput{script} class is derived from the Python \comput{module}
class, to which context-oriented behaviour is added.
A \comput{<script>} (\comput{script} instance) is automatically
created in-memory when loading a script file (containing Python code)
from the \elsA/ command-line; this is the most common way a
\comput{<script>} object is created.
Internally, this creation is performed through the \elsAPython/
\comput{load()} function, of which the simplest call form is:
\comput{<script> = load(<script\_file\_name>)}
More scripts may be loaded from the main (\comput{root}) script,
\emph{script} objects may thus be nested.
\subsection{Global tree structure}
\label{static_tree}
Building a problem description in \elsAPython/ implicitly creates a
tree, with the enclosing (top-level) \emph{script} object as
\comput{root}:
\begin{itemize}
\setlength\itemsep{0ex}
\item \emph{script}s reference (include) \emph{description}
constructors (and other \emph{method}, or \emph{function}, calls),
but own no \emph{attributes};
\item a \emph{script} may reference (include) other \emph{script}s,
through \comput{load()} calls;
\item a \emph{description} may reference other \emph{description}s,
through \comput{attach()} calls, e.g. \comput{cfd1.attach(mod1,
num1)};
\item \emph{description}s own (contain) \emph{attribute}s;
\item \emph{attribute}s have \emph{value}s;
\item \emph{value}s may be \comput{<float>}, \comput{<int>},
\comput{<string>}, or \comput{None} (meaning ``no value'', for all
types).
\end{itemize}
Calls to the \comput{check()} method of \emph{script}s and
\emph{description}s (both being \emph{contexts}, see
\ref{methods_cntx}) will transitively traverse the referenced objects
(included \emph{script}s, included or attached \emph{description}s).
\subsection{Methods and functions}
\label{methfunc}
We describe below a few methods of description classes, see
\ref{methods_desc}, and of contexts, see \ref{methods_cntx};
additionally, the \comput{close()} function call will end the
processing of a script.
\subsubsection{Specific methods of description classes}
\label{methods_desc}
\begin{itemize}
\setlength\itemsep{0ex}
\item \comput{<desc>.set(<attr>, <valu>)}: define \comput{valu} as the
value of the \comput{attr} attribute of the \comput{<desc>}
description class instance;
\item \comput{<desc>.get(<attr>)}: return the value of the
\comput{attr} attribute of the \comput{<desc>} description class
instance;
\item \comput{<desc>.compute()}: start the algorithm managed by
the \comput{desc} description class.
\end{itemize}
\subsubsection{Methods of contexts (descriptions and scripts)}
\label{methods_cntx}
Description classes and script classes are both contexts: they all
inherit from the \comput{context} class, and thus share a number of
methods, which are shown here (in the simplest syntax).
\begin{itemize}
\setlength\itemsep{0ex}
\item \comput{<context>.check()}: perform contextual checking (using
influence, dependency and contextual default rules) on the
\comput{<context>} instance; this method is the defining feature of
a context;
\item \comput{<context>.view()}: display a compact (using
macro-attributes) representation of the \comput{<context>} instance,
masking all non-coherent attribute values;
\item \comput{<context>.copy(name=<name>)}: return a copy of the
\comput{<context>} instance, with the \comput{<name>} identifier;
\item \comput{<context>.dump()}: dump a compact representation of the
\comput{<context>} instance to a file; dumped scripts may not
include explicit control structures (loops and tests), which
should be managed by description classes.
\end{itemize}
\subsection{Abstraction \& flexibility: redirections}
\label{redirect}
For better abstraction and flexibility, some functions and methods
are redirections to lower levels:
\begin{itemize}
\setlength\itemsep{0ex}
\item function to \comput{root} script method;
\item script method to \comput{rootboot} description method.
\end{itemize}
as explained below in \ref{functions}, \ref{methods_scpt}. These
redirections are most powerful when coupling with external layers, see
\ref{products}.
\subsubsection{Functions redirected to \comput{root} script methods}
\label{functions}
The \comput{compute}, \comput{extract}, \comput{check} and
\comput{dump} functions are indirections to the same-named methods of
the top-level \comput{root} script object. Moreover, the
\cmdopt{check} and \cmdopt{dump} command-line options
(\sigle{CLO}) trigger the corresponding function call on an explicit
\comput{close()} call or on natural termination of the \comput{root}
script, while the \cmdopt{strict} option triggers a
\comput{check()} call before the \comput{compute()} call.
\subsubsection{Script methods redirected to the \comput{boot} description}
\label{methods_scpt}
The \comput{compute} (start solver) and \comput{extract} (build the
specified output representation) methods of the current script object
are indirections to the same-named methods of the current
\comput{boot} description object, see \ref{rootboot}.
\section{Static and dynamic behaviour}
\label{statdyn}
New description classes do appear from time to time, but the main
mechanism for the evolution of the \elsAPython/ interface is the
addition of new attributes to existing classes. This is performed by
augmenting two resource files, one for static definitions and the
other for dynamic (context-related) ones.
All resource files (data defining the class attributes and various
contextual rules) are structured using varying combinations of the
\comput{dictionary} and \comput{list} Python types; these
\emph{varying} combinations are allowed (without additional coding)
because Python is a dynamically typed language.
Rather than using a uniform formal schema, different \emph{ad hoc}
grammars are used here for each kind of definition, abusing the expert
(if dated) advice: ``As far as we were aware, we simply made up the
language as we went along.'' - John Backus, Developer of Fortran
(1957) and inventor of \sigle{BNF} (1959) \citep{Backus1978a}.
We thus conveniently ``forget'' the invention of the more formal
Backus Normal (later Naur) Form (\sigle{BNF}), and use a homebrewed
grammar; this non-formal character introduces a degree of
incompleteness, which has not introduced problems so far for the
\sigle{CFD} target application field; a few special cases of rule
definitions are treated using \comput{lambda} expressions (anonymous
functions for in-lining rule code), with also some regular expression
matching in lieu of plain comparison on \comput{<string>} values.
Using only the basic \comput{dictionary} and \comput{list} Python
types for these homebrewed grammar definitions makes the corresponding
resource files compact and readable, so that developers may actually
augment them; they are also very fast to parse. Again referencing
(albeit indirectly) John Backus: ``Because the customers of the 704
were primarily scientists and mathematicians, the language would focus
on allowing programmers to write their formulas in a reasonably
natural notation.'' \citep{Aiken2007}.
\subsection{Definitions for static behaviour}
\label{statdefs}
The definitions for static behaviour (i.e. excluding contextual
checks) are centralized in a global file, as a Python module. They
include for each attribute:
\begin{itemize}
\setlength\itemsep{0ex}
\item a short descriptive text;
\item a type (float, integer or string) for the attribute value;
\item one or several checking methods, further restricting the
definition domain (e.g. \comput{<float>} type but further restricted
to $\mathbb{R}^{*+}$);
\item a list of default value mechanisms, among: static (value),
dynamic (reference to rule), \comput{None};
\item optionally, restrictions to modifications of the attribute
value, e.g. interface only (not user).
\end{itemize}
Example:
\newline
\comput{'phymod':["""fluid model""",['S','I'], \{'euler':0,'nslam':1,'nstur':2\}, [CNTX\_DEFV,None]]}
These definitions are rendered in plain English (along with the
context-related items, see \ref{dynamic}) by the \comput{man()} function
of the interface, see \ref{intman}.
Moreover, a number of entries may be defined for each class, defining
additional metadata, see \ref{metadata}.
Part of the attributes definitions (for example inheritance,
see \ref{metadata}) is only finalized at runtime. It is then used to
define class singletons, managing the attribute definitions for all
the instances of a given class. This allows caching (across same-class
instances) the internal representation of the definitions.
\subsubsection{Attribute definition details}
\paragraph{Description chain}
The description chain for each attribute is used by the \comput{man()}
function, the popups in the \PyGelsA/ \sigle{GUI}, see section
\ref{pygelsa}, and for generating the skeleton of the User's Reference
Manual, see section \ref{doc}.
\paragraph{Type and definition domain method(s)}
The basic type (float, integer or string) of the values of each
attribute is complemented with a list of allowed of values, a
(registered) checking method for the definition domain
(e.g. ``strictly positive'' or more complex), or a combination of
both. When the type changes between the interface and the kernel, the
list of allowed values is replaced with a conversion dictionary.
\paragraph{Default values}
The ``default values'' item may be an explicit value, a reference to
the (possibly converted) kernel default value, a reference to a
(possibly non-existent) context-dependent (contextual) default value,
or a list of such items (e.g. contextual, then ``static'' safe
value). The single \comput{None} default value means that the
attribute has no default value; if moreover the attribute is defined
as always required, context checks will return \comput{False} unless
it has been defined by the user.
\subsubsection{Macro-attributes}
To give structure to the large number of attributes of some classes,
they may be grouped into ``macro-attributes'', meaning named lists of
attributes. Macro-attributes are list-valued and may be managed
through \comput{set()}, \comput{get()}, \comput{view()} and generally
all methods of the description classes meant for plain attributes
(their ``atoms'').
A macro-attribute may have several versions with differing lengths,
e.g. \comput{conservative} is always the name of the list of the
attributes for the conservative variables, whatever the current number
of equations. Specific versions (as declared in the resource file) are
internally named e.g. \comput{conservative*05},
\comput{conservative*06} \ldots but on the user side, for input and
output, plain \comput{conservative} is used. This feature is not
guaranteed to always provide as easy a grouping of related attributes,
but it is quite useful.
In the \sigle{GUI}, see section \ref{pygelsa}, the widgets for
macro-attributes appear as foldable groups of ``atom''
widgets. Folding is automatic on failed influence/dependency rules
(meaningless macro-attribute).
\subsubsection{Attributes metadata}
\label{metadata}
Additional attributes metadata are structured in class-wide lists for:
\begin{itemize}
\setlength\itemsep{0ex}
\item always-required attributes (whose value must be defined);
\item obsolete attribute and values, and possibly their current
replacement\newline (\cmdopt{allow\_obsolete} \sigle{CLO});
\item attribute and values which may be filtered out (\cmdopt{filter}
\sigle{CLO});
\item not yet documented attributes and values (\cmdopt{unlock} \sigle{CLO});
\item inheritance (with possible modifications) of attribute
definitions from other classes.
\end{itemize}
\subsection{Definitions for dynamic behaviour}
\label{dynamic}
The definitions for dynamic behaviour are the most important part of
the base interface, allowing for context management and thus more
abstract problem specification. Context management allows the software
to behave differently, according either to a few trade-specific
indications, see \ref{contdefs}, or to the current global solving
operator (in case of coupling), see \ref{sfddmd}.
\subsubsection{Context management}
\label{context_manage}
The definitions for dynamic behaviour (i.e. rules for context
management) are centralized in a global resource file, as a Python
module; the inference engine (for applying these rules) is defined in
a separate module. Dynamic behaviour is represented here by the
context-dependent features of the interface: influence and dependency
rules, and contextual default rules, completed by the traversal of
relations created through \comput{attach()} calls and possibly by the
hierarchical structure of scripts (through inclusion).
Each problem description thus introduces a realization of a dynamic
structure, whose nodes are linked by contextual influence and
dependence relations; this structure is a Directed Acyclic Graph
(\sigle{DAG}), and not a true tree, because a node may have more than
one parent (several influence rules may lead to the same attribute
being required). The ``Acyclic'' word is important (also valid for a
tree, of course), because we need to ensure that no rules in the
grammar lead to circular checking. The ``dynamic'' qualifier comes
both from the influence and dependence rules and from the
context-dependent default values, leading to (possibly) different
graph realizations for different user inputs, contrary to the static
layout of a \sigle{CGNS} tree.
Default values are sought when a rules terminator (an end rule)
specifies that an attribute value must be defined, and no such value
exists; \emph{contextual} default rules, see \ref{contdefs}, depend on
the current state of the descriptions context. The
\comput{<desc>.get\_or\_deft(<attr>)} method call returns the
best-effort result for the value of the \comput{<desc>.<attr>}
attribute -- using available context state (attribute values) and
rules -- or \comput{None} if no default value may be computed.
\begin{figure}[!h]
\centerline{\includegraphics[width=1.05\columnwidth]{tree_dep_inf_2016.pdf}}
\caption{\label{tree_dep_inf}Contextual \sigle{DAG} information flow,
through influence and dependency.}
\end{figure}
\begin{figure}[!h]
\centerline{\includegraphics[width=1.05\columnwidth]{local_tree}}
\caption{\label{local_tree}Local rules evaluation (one attribute).}
\end{figure}
Each rule is centered on one attribute, \figurename~\ref{local_tree},
ignoring for the moment the potential complexity which results from
traversing all the rules, \figurename~\ref{global_tree}. When this
traversal is performed for a specific case (with user-specified values
as initial conditions), a dynamic \sigle{DAG} based on contextual
relations (context \sigle{DAG}) is realized. This \sigle{DAG}
typically has many levels, quite differently from the flat and static
one defined by the class structure, which is used for data storage
(and documentation).
\begin{figure}[!h]
\centerline{\includegraphics[width=1.05\columnwidth]{global_tree}}
\caption{\label{global_tree}Global rules traversal (more than one attribute).}
\end{figure}
Optionally, meaningless values may be ``pruned'' from the context
(using the \comput{<desc>.check(prune=True)} call), hopefully making
it both complete (computable) -- using default values -- and coherent
(here meaning minimal).A simple pruning example would be removing the
choice of turbulence model if the simulation is declared laminar, as
opposed to a plain \comput{<desc>.check()} which would only flag the
combination as non-coherent.
Pruning may not be fully automated, because the rules system cannot
decide on which side of the rule the error lies. Thus, in the spirit
below, we will not make a decision to remove the value: "In the face
of ambiguity, refuse the temptation to guess." \citep{Peters2004}.
\smallskip
For example, the user may have kept the turbulent model definition
from a previous case, or forgotten to change the flow model from
laminar to turbulent when choosing the turbulence model; in this
occurrence, a \comput{check()} call on the involved \comput{model}
object would lead to a \comput{Warning}-level message, see
\ref{messages}.
When performing pruning, the decision is made according to dependency
rules, this is why in the above example the turbulent model choice
would be removed while the flow model (higher up the context
\sigle{DAG}) is left unchanged. On the other hand, if unneeded values
are defined and the \cmdopt{strict} \sigle{CLO} is used (or if
the corresponding internal option is set from the script), the
simulation will be aborted because in this condition all
\comput{Warning}s are bumped to the \comput{Error} level.
In propagating an initial context of user-defined values using the
various contextual rules, we will call ``down-propagation'' context
flow that is centrifugal (propagating away from the root of the
context \sigle{DAG}), and ``up-propagation'' context flow that is
centripetal, see \figurename~\ref{tree_dep_inf}. A \emph{context}
check may start at any point in the \sigle{DAG}, performing both
up-propagation (for dependency) and down-propagation (for influence);
this is useful for the \sigle{GUI} and for interactive user help (what
if).
Example rules will be provided for the \comput{model} class of
\elsAPython/, which gathers (as attributes) parameters related to the
physical model.
\subsubsection{Dependency rules}
A dependency rule defines the up-propagation of an attribute, and
specify that other (source) attributes must have specified values for
coherency, see \figurename~\ref{tree_dep_inf}, e.g.:
\newline
\comput{'visclaw': \{'phymod': ['nslam', 'nstur']\}}
\newline
meaning that the \comput{visclaw} attribute has meaning only for the
\comput{'nslam'} and \comput{'nstur'} (laminar and Navier-Stokes
respectively) fluid models, and not for the remaining \comput{'euler'}
model, see \ref{statdefs}.
\subsubsection{Influence rules}
An influence rule defines the down-propagation of an attribute value,
and specify that other (target) attributes must be defined for
coherency, see \figurename~\ref{tree_dep_inf}, e.g.:
\newline
\comput{'visclaw': \{'sutherland': ['suth\_const', ['suth\_muref', 'suth\_muref\_fct'], 'suth\_tref']\}}
\newline
meaning that the \comput{'sutherland'} value of the \comput{visclaw}
attribute requires that the \comput{suth\_const} and
\comput{suth\_tref} attributes be defined, together with one of
\comput{suth\_muref} and \comput{suth\_muref\_fct}.
So-called ``strong'' influence rules specify moreover that the value of
the target attribute must belong to a specified list, depending on the
value of the origin attribute, e.g.:
\newline
\comput{'user\_config': \{'limited': [\{'turbmod':['keps', 'komega']\}, 'easy']\}}
\newline
would be a strong rule specifying that when
\comput{user\_config='limited'} the only possible choices of
turbulence model are \comput{'keps'} and \comput{'komega'}, and that
the \comput{'easy} attribute value is required.
\subsubsection{Contextual default rules}
\label{contdefs}
Contextual default rules provide a mechanism for defining default
values of attributes when neither user-defined nor static defaults are
provided, e.g.:
\newline
\comput{'suth\_muref': \{1.78938e-5: \{'mixture': ['air'],'cfdpb.units':['si']\}\}}
\newline
meaning that the default value of the \comput{suth\_muref} attribute
is \comput{1.78938e-5} (in \sigle{SI} units), provided that the fluid
composition is defined as \comput{'air'} (from the single-element
\comput{['air']} list) and the problem is specified in \sigle{SI}
units.
The \sigle{DAG} may be dynamically extended through contextual default
rules, which are applied iteratively on a \comput{check()} call until
no new values are defined (or until the maximal iteration count is
reached). For each descriptions context state -- user-defined values,
``static'' default values, and current contextual defaults --
influence and dependency rules, together with ``always required''
rules, define which remaining attribute values must be defined to
render the context complete. Each newly defined value corresponds to a
state transition of this checking automaton.
For all description classes, trade-specific rules (possibly using
regular expressions) may be triggered using the \comput{user\_config}
attribute with arbitrary \comput{<string>} values.
All attribute values may be traced to their creator (kernel, user,
static default, contextual rule) using the \comput{show\_origin}
method. If the creator is a contextual rule, it is listed.
\subsubsection{``Horizon'' mechanism}
\label{horizon}
An original ``horizon'' mechanism has been developed
\citep{Lazareff2009}, replacing the definition of various user skill
levels with a movable by-attribute boundary across the context-defined
\sigle{DAG}; it still has to be documented and user-tested.
\section{Documentation \& error management}
\label{doc}
\subsection{Integrated documentation}
\label{intman}
The \comput{man()} function of the interface provides, for any element
of the \elsAPython/ interface (function, class, method, attribute), a
compact (and always up-to-date) documentation suitable for a returning
user:
\begin{verbatim}
man(check)
Name : check
Type : function
Description: Check status of root script object
man(view)
Name : view
Type : function
Description: Facade for current script's method
man(model.view)
Name : view
Type : instancemethod
Description: Filtering view for a description
man('phymod')
1) Attribute name: phymod
2) Class(es) : model
3) Description : fluid model
4) Allowed values: 'euler', 'nslam', nstur'
5) Rules :
5b) influence rules:
phymod = 'nslam' requires:
value(s) for visclaw & prandtl & trans_mod & ...
phymod = 'euler' requires:
phymod = 'nstur' requires:
value(s) for visclaw & cv & prandtl & ...
5c) context-dependent default values:
phymod = 'nstur' IF:
user_config = 'test::wing' | 'test::body' | ...
5d) absolute rules:
attribute value is always required
6) Default value(s): 'euler'
context-dependent default values in
'5c)', if any, are applied first
\end{verbatim}
Notice: some output lines have been truncated.
When the same attribute name is shared by several classes, the output
of \comput{man()} is factorized across identical definitions.
\subsection{User's Reference Manual updating}
\label{urm}
A \sigle{PDF} version of the User's Reference Manual (\sigle{URM}) is
built using the {\LaTeX} typographical software \citep{latex}. For
each new version of the \elsA/ software, the \sigle{URM} must be
updated with new attribute descriptions (and possibly new functions,
classes, methods, and additions to the specialized appendices).
The updating process is partly automated, but requires some human
intervention: the \comput{CheckDefs} tool included in the Python
interface code provides automatic generation of the {\LaTeX} source
code for the description of the attributes missing in the current
\sigle{URM} version, based on the contents of the resource
files. Merging in the new {\LaTeX} source is done manually -- using
the \comput{ediff} tool in the \comput{emacs} editor -- and coherency
of the manual with the interface definitions checked, still using
\comput{CheckDefs}.
\subsection{Exceptions (error management)}
The \elsAPython/ interface defines its own exception classes, which
use two severity levels:
\begin{itemize}
\setlength\itemsep{0ex}
\item \comput{WARNING}: non-fatal error;
\item \comput{ERROR}: fatal error.
\end{itemize}
Notice: the \cmdopt{strict} \sigle{CLO} upgrades all
\comput{WARNING}s to \comput{ERROR}s.
\subsection{Error messages}
\label{messages}
The structure of \elsAPython/ error messages is:
\begin{itemize}
\setlength\itemsep{0ex}
\item first line: general information about the error, including the
severity level;
\item a few lines describing in more detail the problem leading to the
error;
\item last line: the suggested correction.
\end{itemize}
It has to be noted that, although these messages on their own appear
to be quite explicit, users of our software will sometimes \emph{not
read} the message, instead ringing up \elsA/ software support.
This does not appear to be uncommon \citep{UX2011}, and could be
linked to user stress, which is what we tried to avoid in the first
place, or to a desire not to do a mental ``context change'' by trying
to actually solve the problem themselves: ``What they want is to pick
up the phone, make a call, and have someone tell them what to do.''
\citep{slash2010}.
\section{The \PyGelsA/ \sigle{GUI}}
\label{pygelsa}
The \PyGelsA/ \sigle{GUI} (Graphical User Interface) is built
on-the-fly using the ``static'' interface definitions, see
\ref{statdefs}; this means that no modification of the \PyGelsA/ code
is needed when updating the user interface for new versions. When
loaded with a script, the \sigle{GUI} then uses the ``dynamic''
definitions, see \ref{dynamic}, for coherency in the display of the
description objects.
This display stays coherent thanks to on-the-fly application of
context checks. As shown on \figurename~\ref{Create_cfdpb_stars},
attributes with missing required values are labeled in red, while
meaningful but user-folded macro-attributes are labeled in green.
\begin{figure}[!h]
\centerline{\includegraphics[width=1.05\columnwidth]{Create_cfdpb_stars}}
\caption{\label{Create_cfdpb_stars}The \PyGelsA/ \sigle{GUI}.}
\end{figure}
The \sigle{GUI} includes (bottom part) a console for text-driven
interaction, which is synchronized with the graphical window
above. The interface state may later be re-built, either from the
logfile or from the file dumped from a \comput{dump()} call performed
at any time during the \sigle{GUI} use.
\section{Database and network operations}
\label{datanet}
Script (and description) objects may be serialized (transformed into
bytecode) and either dumped to/loaded from a database, or accessed
through the network.
\subsection{Database operations}
\label{database}
The script database feature of the interface is built upon a standard
underlying database implementation, which may be as simple as a
``dictionary on disk''.
A script is an in-memory structure of description objects -- and
usually some operations-- built either from an explicit script file or
from code. These operations may be kept ``pending'' (not executed) if
the script is to be stored in a database.
Script databases provide the basic dump/load operations, through the
\comput{dump()} and \comput{load()} methods. A dumped and re-loaded
script will re-create its description objects, and will execute its
pending operations, if the environment (mesh files \ldots) is
adequate.
\begin{figure}[!h]
\centerline{\includegraphics[width=1.05\columnwidth]{ValDatabase}}
\caption{\label{ValDatabase}Database and network use.}
\end{figure}
These basic operations are complemented by a \comput{search()}
method. For the search to be useful and efficient, the user has to
declare a \comput{view}, defining which parameters (out of possibly
several hundred) are of interest for the current study. The
\comput{<scri>.catalog()} method is then called -- using this
\comput{view} -- at each \comput{dump()} to update the database's
catalog. The \comput{catalog()} call may also be used for tagging the
output data for the corresponding simulation, providing an authentic
traceability (e.g. for plots).
Adequate locking allows the database to be shared in read/write mode
between several processes, possibly on different machines, allowing
natural parallelism. The default behaviour is to use separate
databases for script definition (read-only) and for job execution
metadata (read/write) for better efficiency.
The database keeps track of not yet started (\comput{NYS}), running
(\comput{RUN}), and completed (\comput{CMP}) state of computational
jobs, which helps recovering from crashes, generally using a global
\comput{<database>.clean()} call (resetting all \comput{RUN} jobs to
\comput{NYS} state) before restart. This partial restarting has been
very efficient to reduce user stress, especially when running
simulations with more jobs than computing nodes and (clock) execution
times counted in days.
\subsection{Network operations}
\label{network}
Methods for the \comput{network} class include a simple server/client
(sockets) pair, the \comput{dump()}/\comput{load()} pair, and the
\comput{popen()} (get command output) and \comput{db\_search()} (find
in database) commands.
This functionality allows to access a script database through the
network, which is useful when no shared filesystem is available and
the underlying database is not networked, see
\figurename~\ref{ValDatabase}.
\section{Extensibility}
\label{extend}
\subsection{Products}
\label{products}
So-called ``products'' are additional (Python) software packages which
couple with \elsAPython/ (and possibly with other products) through
description classes. A developer may build a new \comput{product}
(based on the provided skeleton) and simply drop it in the
\comput{Products} directory. Communication between \elsA/ and the
added software is performed through an instance of this
\comput{product}-based description class, which contributes to the
global context and may get to steer the simulation in lieu of the
\sigle{CFD} solver, see section \ref{doe}.
The algorithm associated with the product may be coded in Python
(possibly referencing external packages with compiled libraries)
and/or use \elsA/ kernel functionality.
\subsection{\computn{root} and \computn{boot} objects}
\label{rootboot}
If a description class declares a \comput{compute()} method, an
instance of this class may grab control of the simulation from the
\sigle{CFD} algorithm, using the global context to gather all the
required data. This instance is then called the \emph{boot object}.
Scripts define a \comput{compute()} method -- invoked in the
(top-level \comput{root}) script through the \comput{compute()}
function -- which is an indirection to that of the current
\comput{boot} object; a script with a \comput{compute()} statement
will thus execute differently -- without explicit tests -- depending
on the current boot object, or on the explicit
\comput{<desc>.compute()} call, e.g. (see \ref{sfddmdspi}):
\begin{verbatim}
cfd1 = cfdpb(name='cfd1')
cfd1.set('sfd', 'active')
...
dmd1 = dmd(name='dmd1')
spr1 = sparse_poly(name='spr1')
...
# cfd1.compute() # single-point SFD simulation
# dmd1.compute() # single-point SFD/DMD simulation
compute() # SFD/DMD/SPI coupling on DoE, using spr1
\end{verbatim}
\noindent where \comput{\#} starts a comment, and
\comput{compute()} is redirected to \comput{spr1.compute()} because
\comput{spr1} is the last-created bootable object.
This logic provides for testing a complex coupling script at different
levels with minimal modification; it can also be implemented as:
\begin{verbatim}
...
slvrs = {0:cfd1, 1:dmd1, 2:spr1}
slvr_lev = 1 # use dmd1
slvrs[slvr_lev].compute()
\end{verbatim}
The coupling level may thus be specified as an integer value, without
any test in the user script.
Lastly, the \comput{slvrs[slvr\_lev].compute()} call above may be
replaced by \comput{set\_boot\_objt(<desc>); compute()}, where the
first statement may be seen as passing a \emph{token} to
\comput{<desc>}, giving it specific rights. This token may be moved
during the simulation, i.e. for collaborating coupling algorithms.
\subsection{\computn{provide()} method for blind creation}
\label{provide}
When several \comput{product}-like description objects
(e.g. \comput{prd1}, \comput{prd2}) use a common description object
(e.g. \comput{slave}), it may occur that it is built by one of them,
which should be the first appearing in the script. If for example
\comput{slave} is built by \comput{prd1.compute()}, built first, and
then used by \comput{prd2}, all is well. But if the \comput{boot}
object is switched from \comput{prd1} to \comput{prd2}, \comput{slave}
will not be built and will miss to both \comput{prd1} and
\comput{prd2}, as \comput{prd1.compute()} is not called anymore.
The solution we use, to avoid here an explicit test on the existence
(with possible creation) of \comput{slave} before each reference, is
the use of the \comput{provide()} method, which wraps the conditional
creation, and may be ``blindly'' called to always return the (new or
pre-existent) object.
\subsection{\computn{target\_lift} class for target lift computations}
\label{targetlift}
The ``target lift'' functionality allows to replace a standard
lifting-body computation, at fixed angle-of-attack (\AoA/), with the
computation of the list of \AoA/s corresponding to the given list of
lift values (within convergence bounds).
This is performed by replacing the standard \comput{compute()} call
with one directed to a new \comput{<target\_lift>} boot object, as in:
\begin{verbatim}
tcl1 = target_lift(name='tcl1')
tcl1.attach(lift)
alphas = compute([0.05, .10, .15])
\end{verbatim}
where \comput{lift} is an \comput{<extractor>} defining the
computation of the body's lift and \comput{alphas} is the \AoA/
output for the specified lift coefficients.
This was the first example of an extension to be finally coded as a
\comput{product}.
\subsection{\computn{variator} class for parametric studies}
\label{variator}
The \comput{variator} class allows building script variations from a
base version and a list of parameter perturbations, dumping them to a
database. Later on this database will be spanned, meaning that all the
database scripts will be loaded and pending operations, see
\ref{database}, executed. This is performed in sub-directories of a
user-chosen base directory, with all file paths automatically shifted
as required. Spanning thus provides the database with an
\emph{iterator}.
A step further, this iterator is made into a more general
\emph{automaton}: during spanning, a simulation may be restarted
(chained) from a source simulation, chosen using a user-defined
non-isotropic \DoE/ distance (in parameter space), possibly with
additional rules for restricting the source/target pairing. When
chaining simulations, the automaton may also add intermediate points
(linearization) to the iteration list, to respect a user-specified
maximal jump size in parametric space \citep{Lazareff2014}.
An external algorithm may also add arbitrary points to the list, see
\ref{sfddmdspi}.
\subsection{\computn{swarm} class for operational efficiency}
\label{swarm}
The \comput{swarm} class provides an abstraction for a group of
spanning simulations. For operational efficiency, automated load
management on a computational node is provided either as a maximal
count of swarm job instances or as a fraction of the node's power.
The values of the selected observable quantity, for all the
simulations in a swarm, are returned as \comput{<valu\_list> =
<swarm>.compute()}, as used for the \sigle{SPI} algorithm,
\ref{sfddmdspi}.
\section{Design of Experiment studies}
\label{doe}
Studies in Design of Experiment (\DoE/) space involve at least two
levels: the basic automaton performing the space spanning, and the
mathematical tools, applied both to each \DoE/ point and to the global
space. These tools may be passive (using specified \DoE/ points in a
given order), or they may use observable quantities to steer the
spanning automaton, possibly creating new \DoE/ points on-the-fly.
In \ref{sfddmd} and \ref{sfddmdspi} we will give two examples in the
field of differentiable dynamical systems (the geometrical study of
complex systems and their stability) \citep{Smale1967} applied to
\sigle{CFD}, first as a local study and then on a \DoE/. Usually the
system is defined by an equation like $\dot{x} = v(x)$. Here however
the operator of the system is unknown, and only accessible through
\sigle{CFD} observable quantities at computed points of the \DoE/, so
that as a first step we have to address the problem of \DoE/
discovery, see \ref{span}, to ensure that the main features of the
operator are accounted for.
\subsection{\DoE/ spanning and discovery}
\label{span}
\DoE/ discovery is an extended parametric study, see \ref{variator},
where the aim is to describe (find the main features of) the space
inside a closed boundary in parametric space, with possible refinement
of the initial set of \DoE/ points \citep{Lazareff2014}. Continuation
techniques (e.g. stabilization to cross \DoE/ areas with separated
flow) may be needed, see \ref{sfddmd}. As a first approach, \DoE/
refinement may be manual, but in \ref{sfddmdspi} we introduce an
automatic refinement algorithm through the \sigle{SPI} algorithm
\citep{Chkifa2014}, which greatly simplifies user interaction.
The observable quantity here may be the aerodynamic lift or drag, the
spectral radius of the global (physical + numerical) operator, or any
other quantity of interest.
\subsection{Coupling with the \sigle{SFD} and \sigle{DMD} algorithms}
\label{sfddmd}
Coupling between \elsA/, the \sigle{SFD} and \sigle{DMD} algorithms
has been performed to provide stabilized \sigle{CFD} solutions,
initially at a single \DoE/ point for the flow around a cylinder
\citep{Cunha2015}. The \sigle{SFD} algorithm may be either part of the
\sigle{CFD} solver or treated as a wrapper in the encapsulated version
\citep{Jordi2014}, with an \comput{sfd} description class. The
\sigle{DMD} algorithm is coded in Python, with a \comput{dmd}
description class. Some parameters of the \sigle{CFD} algorithm
(\comput{numerics} class attribute values) are dependent on the
\sigle{SFD} and \sigle{DMD} algorithms through contextual rules.
This work has since been extended to a (Mach, Reynolds) \DoE/ using
the \comput{variator} class. Success is still dependent on the
adequate adjustment of some parameters, at this time taken to be
constant across the \DoE/. For the time being, this aspect is fully
dependent on user expertise.
\subsection{Coupling with the \sigle{SFD}, \sigle{DMD}, and \sigle{SPI} algorithms}
\label{sfddmdspi}
The most complex application to date of the extensibility of
\elsA/ through the \elsAPython/ interface is the algorithm
described above in \ref{sfddmd}, complemented by Sparse Polynomial
Interpolation (\sigle{SPI}) \citep{Chkifa2014} for \DoE/ discovery (to
be published).
Here the spanning automaton is managed by the \sigle{SPI} algorithm,
starting with nodes only at the \DoE/ summits and progressively
enriching the representation with inner points according to a
Clenshaw-Curtis distribution.
The \sigle{CFD} results for the chosen observable quantity at each
refinement level of the \sigle{SPI} algorithm are computed as a swarm,
see \ref{swarm}.
With a reasonable degree of convergence of the \sigle{SPI} algorithm,
we hope that this algorithm will lead to an adequate automatic
discovery of the \DoE/ features.
\section*{Conclusion}
The basic foundation of the \elsAPython/ interface, which has been
used since the creation of the \elsA/ software, has long been used for
the target lift extension, see \ref{targetlift}. Recently more complex
extensions have been introduced for \DoE/ studies.
This descriptive interface and associated functionality provides a
high level of abstraction for coupling mathematical algorithmic layers
with the \sigle{CFD} solver, reducing the tedious part of user's work
and augmenting efficiency in research and applications.
This approach, allowing for collaborative computing, seems to be well
suited for projects coupling several simulation kernels.
\section*{Perspectives}
A rule-based implementation of the currently kernel-side factory, see
\ref{base} would remove the potential fragility of the current one,
where tests are sometimes nested for several levels. A rule only has
to be written once, while inner tests may have to be repeated for each
branch of the upper level(s). This would benefit to user-side checking
by ensuring completeness and coherency of the rules with the actual
solver implementation.
The more immediate new development will be stochastic analysis, using
the Monte-Carlo method on the response surface of \ref{sfddmdspi},
introducing the first 4th-level layer algorithm above the \elsA/
interface (5th-level above the \sigle{CFD} kernel).
\tableofcontents
\section*{Acknowledgments}
This work has received partial financial support from the Agence
Nationale de la Recherche (\sigle{ANR}
\url{http://www.agence-nationalerecherche.fr/}), first through the
Carnot institutes network contract referenced as n\textsuperscript{o}
07 CARNOT 00801, and then as part of the \sigle{UFO} (Uncertainty For
Optimization) project referenced as ANR-11-MONU-0008.
\medskip
This paper is dedicated to my late colleague and friend Michel Gazaix,
who died unexpectedly on August 8, 2016. He was the one, as software
architect, who invited me into the early \elsA/ project to design and
build its user interface, and I am still indebted to him for this
challenge.
\section*{Bibliography}
\RaggedRight
\bibliographystyle{elsarticle-harv}
|
\section{Introduction}
Variational analysis has been recognized as a broad spectrum of mathematical theory that has grown in connection with the study of problems on optimization, equilibrium, control and stability of linear and nonlinear systems, and its focus is mainly on optimization of functions relative to various constraints and on sensitivity or stability of optimization-related problems with respect to perturbation. Since nonsmooth optimization problems by nonsmooth functions, sets with nonsmooth boundaries or set-valued mappings frequently appear in variational theory and its application, nonsmooth analysis in variational analysis has played an important role in such aspects of mathematical programming and optimization (cf. \cite{2,7,8,17,18} and references therein). Over the past several decades, the first-order nonsmooth analysis has been extensively and systemically studied by many authors in both finite-dimensional and infinite-dimensional spaces, and also fruitfully applied to many aspects of applied mathematics such as first-order optimality conditions, sensitivity analysis, constrained optimization, equilibrium problems with nonsmooth data and optimal control (cf. \cite{AF,2,7,8,16,19}). However, the literature in dealing with second-order nonsmooth analysis is not too much relative to the first-order analysis. We refer readers to books \cite{2,16,19} for the application of second-order generalized differential constructions to optimization, sensitivity and related problems. Given a nonsmooth function defined on a Hilbert space, we mainly study three types of generalized second-order derivatives in this paper: {\it second-order lower Dini-directional derivative}, {\it second-order mixed graphical derivative} and {\it second-order mixed proximal subdifferential} (see Section 3). Then we use these second-order derivatives to consider second-order optimality conditions and investigate their equivalent interrelationship.
Second-order optimality conditions have played important roles in mathematical programming and have been extensively studied by many authors (cf. \cite{4, 5,6,EM,13,PR,24,27} and references therein). Recently Eberhard and Wenczel \cite{11} and Eberhard and Mordukhovich \cite{EM} discussed three different types of second-order optimality conditions which are based on generalized second-order directional derivative, graphical derivative of proximal subdifferential and second-order proximal subdifferential defined via coderivative of proximal subdifferential. The equivalence among these optimality conditions for paraconcave functions is also proved. Using these three types of second-order derivatives aforementioned, we are inspired by \cite{EM,11} to continue studying second-order optimality conditions (with some minor modifications) in a Hilbert space, and mainly study the interrelationship among them. It is also proved that the equivalence among these optimality conditions for paraconcave functions is still valid in the Hilbert space setting. As applications, we use these second-order optimality conditions to investigate strict local minimizers of order two for extended real-valued nonsmooth functions in the Hilbert space.
The notion of strict minimizer of order two for a nonsmooth function has been proved to be useful in optimization and relates closely with the convergence of numerical procedures. Hestenes \cite{12} considered this notion and used it to prove sufficient optimality conditions. Cromme \cite{10} and Auslender \cite{1} studied this notion in connection with convergence of numerical procedures and provided stability conditions. Studniarski \cite{22} used first and second order lower Dini-directional derivatives to study the local strict minimizer of order two and established necessary and sufficient second-order optimality conditions. Along the line given in \cite{22}, Ward \cite{23} investigated another derivatives and tangent cones to study local strict minimizer of order two and optimality conditions. This notion has also been generalized in the senses of weak sharp minima and $\phi$-minima and was extended to vector optimization problems and set-valued mappings (see \cite{9, 14} and references therein).
Note that Eberhard and Wenczel \cite{11} discussed strict local minimizer of order two in a finite-dimension space and provided its characterizations in terms of second-order optimality conditions. Along this line, as one main goal of this paper, we apply the second-order optimality conditions to the strict local minimizer of order two and aim to establish its characterizations in the Hilbert space. These characterizations reduce to the existing ones when restricted to the finite-dimensional space.
The paper is organized as follows. In Section 2, we will give some definitions and preliminaries used in this paper. Our notation is basically standard and conventional in the area of variational analysis. Section 3 is devoted to three types of second-order derivatives of a nonsmooth function and their important properties. In Section 4, by using these second-order derivatives, we mainly study several kinds of second-order optimality conditions in a Hilbert space and present results on their equivalence interrelationship. In Section 5, we first present a counterexample to show that the existing theorem on strict local minimizers of order two given in finite-dimensional space is not valid for the Hilbert space case (see Example 5.1), and then apply these second-order optimality conditions to characterizing the strict local minimizers for this case. The conclusion of this paper is presented in Section 6.
\section{Preliminaries}
Let $H$ be a {\em Hilbert space} equipped with the {\em inner product}
$\langle\cdot, \cdot\rangle$ and the corresponding {\em norm}
$\|\cdot\|$, respectively. Denote by $B_H$ and $S_H$ the {\em unit closed ball} and the {\em unit sphere} of $H$, respectively. For $x\in
H$ and $\delta>0$, let $B(x, \delta)$ denote the {\em open ball}
with center $x$ and radius $\delta$.
Given a multifunction $F: H \rightrightarrows H$, the symbol
\begin{equation*}
\begin{array}r
\mathop{\rm Limsup}\limits_{y\rightarrow x}F(y):=\Big\{\zeta\in H: \exists \ {\rm sequences} \ x_n\rightarrow x \ {\rm and} \ \zeta_n\stackrel{w}\longrightarrow \zeta \ {\rm with}\ \\
\zeta_n\in F(x_n) \ {\rm for\ all \ } n\in \mathbb{N} \Big\}
\end{array}
\end{equation*}
signifies the {\it sequential Painlev\'{e}-Kuratowski outer/upper limit} of $F(x)$ as $y\rightarrow x$, where $
\zeta_n\stackrel{w}\longrightarrow \zeta$ means $\{\zeta_n\}$ converges weakly to $\zeta$.
Given a set $A\subset H$, we denote by $\delta_A(\cdot)$ the {\it indicator function} of $A$ which is defined as $\delta_A(x)=0$ if $x\in A$ and $\delta_A(x)=+\infty$ if $x\not\in A$. We denote by $\overline A$ and $\overline A^w$ the {\em norm closure}
and the {\em weak closure} of $A$ in the norm topology and the weak topology
respectively, and denote by ${\rm aff}A$ the {\em affine hull} of
$A$. Let ${\rm ri}A$, ${\rm qri}A$ and ${\rm sqri}A$ denote the {\em relative
interior}, the {\em quasi-relative interior} and the {\em strong quasi-relative interior}, respectively, which are
defined by
\begin{eqnarray*}
{\rm ri}A:&=&\big\{x\in A: \exists\ \delta>0\ {\rm such \ that}\ B(x, \delta)\cap {\rm aff}A\subset A\big\},\\
{\rm qri}A:&=&\big\{x\in A: \overline{\rm cone}(A-x)\,\,{\rm is\,\,a\,\,subspace}\big\},\\
{\rm sqri}A:&=&\big\{x\in A: {\rm cone}(A-x)\,\,{\rm is\,\,a\,\,closed\,\,subspace}\big\}
\end{eqnarray*}
where cone$(A-x)$ denotes the cone generated by $A-x$ and $\overline{\rm cone}(A-x)$ denotes the closure of cone$(A-x)$. Clearly ${\rm sqri}A\subset {\rm qri}A$. When $H$ is
finite-dimensional, these three types of relative interior coincide; that is ${\rm sqri}A={\rm qri}A={\rm ri}A$. Readers are invited to refer to \cite{3} for more details on these interior concepts.
Let $S$ be a nonempty closed subset of $H$. Recall from \cite{AF} that the {\it contingent cone} and the {\it weak contingent cone} of $S$ at $x\in S$, denoted by $T(S, x)$ and $T^w(S, x)$ respectively, are defined by
\begin{eqnarray*}
T(S, x):=\big\{h\in H: \exists\ t_n\rightarrow 0^+\ {\rm and} \ h_n\rightarrow h\ {\it s.t.} \ x+t_nh_n\in S \ \forall n\in \mathbb{N}\big\},\, \ \\
T^w(S, x):=\big\{h\in H: \exists\ t_n\rightarrow 0^+\ {\rm and} \ h_n\stackrel{w}\longrightarrow h\ {\it s.t.}\ x+t_nh_n\in S\ \forall n\in \mathbb{N}\big\}.
\end{eqnarray*}
When $H$ is a finite-dimensional space, both contingent cone and weak contingent cone coincide.
For any point $z\in H$, the distance between $z$ and $S$ is given by
$$
d(z, S):=\inf\{\|z-s\|\,:\,s\in S\}.
$$
Let $x\in S$. Recall from \cite{8} that the {\it proximal normal cone} of $S$ at $x$, denoted by $N^p(S, x)$, is defined as
\begin{equation}
N^p(S,x):=\{\zeta\in H: \exists \ t>0\ {\rm such\ that} \ d(x+t\zeta, S)=t\|\zeta\|\}.
\end{equation}
It is known and easy to verify that $\zeta\in N^p(S,x)$ if and only if there exists $\sigma\in (0, +\infty)$ such that
\begin{equation}
\langle\zeta, s-x\rangle\leq \sigma\|s-x\|^2\ \ {\rm for\ all} \ s\in S.
\end{equation}
Let $\hat N(S, x)$ denote the {\it Fr\'{e}chet normal cone} of $S$ at $x$; that is,
$$
\hat N(S, x):=\left\{\zeta\in H: \limsup_{y\stackrel{S}\longrightarrow x}\frac{\langle \zeta, y-x\rangle}{\|y-x\|}\leq 0\right\}
$$
where $y\stackrel{S}\longrightarrow x$ means $y\rightarrow x$ and $y\in S$. Since a Hilbert space is reflexive, it is easy to verify that
\begin{equation}\label{2.3}
\hat N(S, x)=\big(T^w(S, x)\big)^{\circ}
\end{equation}
where $\big(T^w(S, x)\big)^{\circ}$ is the {\it dual cone} of $T^w(S, x)$ which is defined by
$$
\big(T^w(S, x)\big)^{\circ}:=\{v\in H: \langle v, h\rangle\leq 0\ {\rm for\ all} \ h\in T^w(S, x)\}.
$$
The {\it Mordukhovich(limiting/basic) normal cone} of $S$ at $x$, denoted by $N(S, x)$, is defined as
\begin{equation}
N(S, x):=\mathop{\rm Limsup}_{y\stackrel{S}\longrightarrow x}N^p(S, y).
\end{equation}
Thus, $\zeta\in N(S, x)$ if and only if there exists a sequence $\{(x_n, \zeta_n)\}$ in $S\times H$ such that $x_n\rightarrow x$, $\zeta_n\stackrel{w}\longrightarrow \zeta$ and $\zeta_n\in N^p(S, x_n)$ for each $n\in \mathbb{N}$.
Let $f:H\rightarrow
\mathbb{R}\cup\{+\infty\}$ be an extended real-valued and lower semicontinuous function. We denote
$$
{\rm dom}f:=\{y\in H: f(y)<+\infty\}\,\,\,{\rm and}\,\,\,{\rm epi}f:=\{(x, \alpha)\in H\times \mathbb{R}: f(x)\leq \alpha\}
$$
the {\it domain} and the {\it epigraph} of $f$, respectively. Let $x\in{\rm dom}f$. Recall that the {\it proximal subdifferential} of $f$ at $x$, denoted by $\partial_pf(x)$, is defined by
\begin{equation}
\partial_pf(x):=\{\zeta\in H: (\zeta, -1)\in N^p({\rm epi}f, (x, f(x)))\}.
\end{equation}
It is known from \cite{8} that $\zeta\in \partial_pf(x)$ if and only if there exist $\sigma, \delta\in (0, +\infty)$ such that
\begin{equation}
f(y)\geq f(x)+\langle\zeta,y-x\rangle-\frac{\sigma}{2}\|y-x\|^2\ \ {\rm for\ all} \
y\in B(x,\delta).
\end{equation}
The {\it Mordukhovich(limiting/basic) subdifferential} of $f$ at $x$ is defined as
\begin{equation}
\partial f(x):=\{\zeta\in H: (\zeta, -1)\in N({\rm epi}f, (x, f(x)))
\}.
\end{equation}
It is proved in \cite{16, 17} that
$$
\partial f(x)=\mathop{\rm Limsup}_{y\stackrel{f}\rightarrow x}\partial_pf(y)
$$
where $y\stackrel{f}\longrightarrow x$
means $y\rightarrow x$ and $f(y)\rightarrow f(x)$. Therefore $\zeta\in\partial f(x)$ if and only if there exist $x_n\stackrel{f}\longrightarrow x$ and $\zeta_n\stackrel{w}\longrightarrow\zeta$ such that $\zeta_n\in\partial_pf(x_n)$ for each $n\in\mathbb{\mathbb{N}}$.
When $f$ is convex, the proximal subdifferential and the limiting subdifferential of $f$ at $x\in{\rm dom}f$ coincide and both reduce to the subdifferential in the sense of convex analysis, that is
$$
\partial_p f(x)=\partial f(x)=\{\zeta\in H: \langle \zeta, y-x\rangle\leq f(y)-f(x)\ \ {\rm for\ all} \ y\in H\}.
$$
Readers are invited to consult \cite{2,3,7,8,16,19} for more details on these various normal cones and subdifferentials.\\
The following concepts of {\it paraconcavity and paraconvexity} are used in our analysis.
For an extended-real-valued function $\varphi:H\rightarrow \mathbb{R}\cup\{+\infty\}$, recall from \cite{11} that $\varphi$ is said to be {\it paraconvex}, if there exists $\lambda\in (0, +\infty)$ such that $\varphi+\frac{1}{2\lambda}\|\cdot\|^2$ is
convex on $H$ and $\varphi$ is said to be {\it locally paraconvex} around $\bar
x\in{\rm dom}(\varphi)$, if there exist $\delta, \lambda\in (0, +\infty)$ such that $\varphi+\frac{1}{2\lambda}\|\cdot\|^2$ is
convex relative to $B(\bar x, \delta)$. The function $\varphi$ is said to
be {\it locally paraconcave}, if $-\varphi$ is locally paraconvex.
\setcounter{equation}{0}
\section{Second-order derivatives of an extended real-valued function}
In this section, we consider several types of second-order derivatives of a nonsmooth function; namely, second-order lower Dini-directional derivative of a function, second-order mixed graphical derivative and second-order mixed proximal subdifferential of a function, and then study some properties of these second-order derivatives which will be used in our analysis.
Let $f:H\rightarrow \mathbb{R}\cup\{+\infty\}$ be a proper lower semicontinuous function. We denote by
$$
{\rm gph}(\partial_pf):=\{(x, u)\in H\times H: u\in \partial_pf(x)\}
$$
the {\it graph} of proximal subdifferential $\partial_pf$. In this paper, taking into account the application to second-order optimality conditions in the Hilbert space, we first study the following {\it mixed contingent cone} of ${\rm
gph}(\partial_pf)$ and its associated polar.\\
{\it Let $(\bar x, p)\in
{\rm gph}(\partial_pf)$. The {\it mixed contingent cone} of ${\rm
gph}(\partial_pf)$ at $(\bar x, p)$, denoted by $T_M({\rm
gph}(\partial_pf), (\bar x, p))$, is defined as follows:}
\begin{equation}\label{3.1}
\begin{array}r
(h, z)\in T_M({\rm gph}(\partial_pf), (\bar x, p)) \Leftrightarrow
\exists\, t_n\rightarrow 0^+,\,\,h_n\rightarrow h\,\,{\it and}\,\,
z_n\stackrel{w}\longrightarrow z \ {\it such\,\,that}\\
(\bar x+t_nh_n, p+t_nz_n)\in {\rm gph}(\partial_pf)\ \ {\it for\ all} \ n\in \mathbb{N}.
\end{array}
\end{equation}
By the definition, the following inclusions are trivial:
\begin{equation}\label{3.2}
T({\rm gph}(\partial_pf), (\bar x, p))\subset T_M({\rm gph}(\partial_pf), (\bar x, p))\subset T^w({\rm gph}(\partial_pf), (\bar x, p)).
\end{equation}
We denote by $D^2f(\bar x, p)(h)$ and $D_M^2f(\bar x, p)(h)$ the {\it second-order graphical derivative} and the {\it second-order mixed graphical derivative} of $f$ at $(\bar x, p)$ in the direction
$h\in H$, respectively which are defined as
\begin{eqnarray*}
D^2f(\bar x, p)(h):&=&\big\{z\in H: (h, z)\in T({\rm gph}(\partial_pf), (\bar x, p))\big\}\\
D_M^2f(\bar x, p)(h):&=&\big\{z\in H: (h, z)\in T_M({\rm gph}(\partial_pf), (\bar x, p))\big\}.
\end{eqnarray*}
We denote by $\hat{\partial}^2f(\bar x, p)(h)$ and $\partial_M^2f(\bar x, p)(h)$ the {\it second-order proximal subdifferential} and the {\it second-order mixed proximal subdifferential} of $f$ at $(\bar x, p)$ in the direction $h\in H$, respectively and they are defined by
\begin{eqnarray*}\label{4.10}
\hat{\partial}^2f(\bar x, p)(h):&=&\big\{z\in H: (z, -h)\in \hat{N}({\rm gph}(\partial_pf), (\bar x, p))\big\}\\
\partial_M^2f(\bar x, p)(h):&=&\big\{z\in H: (z, -h)\in N_M({\rm gph}(\partial_pf), (\bar x, p))\big\}
\end{eqnarray*}
where $N_M({\rm gph}(\partial_pf), (\bar x, p))$ is the dual cone of $T_M({\rm gph}(\partial_pf), (\bar x, p))$; that is
\begin{equation}\label{3.3}
N_M({\rm gph}(\partial_pf), (\bar x, p)):=\big(T_M({\rm gph}(\partial_pf), (\bar x, p))\big)^{\circ}.
\end{equation}
By \eqref{2.3}, \eqref{3.2} and \eqref{3.3}, one can easily verify that
\begin{equation}\label{3.4}
\hat N({\rm gph}(\partial_pf), (\bar x, p))\subset N_M({\rm gph}(\partial_pf), (\bar x, p)).
\end{equation}
When $H$ is finite-dimensional, for any $h\in H$, one has
\begin{equation}\label{3.5a}
D^2f(\bar x, p)(h)=D_M^2f(\bar x, p)(h)\ \ {\rm and}\ \ \hat{\partial}^2f(\bar x, p)(h)=\partial_M^2f(\bar x, p)(h)
\end{equation}
since $T_M({\rm gph}(\partial_pf), (\bar x, p))$ coincides with $T({\rm gph}(\partial_pf), (\bar x, p))$ in this case.
Recall that the {\it second-order lower Dini-directional derivative} of $f$ at $\bar x$
for $p\in\partial_pf(\bar x)$ along the direction $h\in H$ is defined as
\begin{equation}\label{3.5}
f_-''(\bar x, p, h):=\liminf_{h'\rightarrow h, t\downarrow 0}\Delta_2f(\bar x, p, t,
h')
\end{equation}
where
$$
\Delta_2f(\bar x, p, t,
u):=\frac{f(\bar x+tu)-f(\bar x)-t\langle p, u\rangle}{\frac{1}{2}t^2},\ \ \forall (t, u)\in (0, +\infty)\times H.
$$
Applying \cite[Theorem 19]{11}, for all $t>0$, one has
\begin{equation}\label{3.6}
\partial_p\big(\frac{1}{2}\Delta_2f(\bar x, p, t, \cdot)\big)(w)=\frac{1}{t}(\partial_pf(\bar x+tw)-p).
\end{equation}
The following proposition provides some properties on the
second-order lower Dini-directional derivative $f_-''(\bar x, p, \cdot)$.
\begin{pro}
Let $f:H\rightarrow \mathbb{R}\cup\{+\infty\}$ be a proper lower
semicontinuous function and $\bar x\in{\rm dom}f$ with $p\in\partial_pf(\bar x)$. Then
{\rm (i)} $f_-''(\bar
x, p, \cdot)$ is lower semicontinuous.
{\rm (ii)} Suppose that $f$
is locally paraconcave around $\bar x$. Then $f_-''(\bar x, p,
\cdot)$ is also locally paraconcave around $\bar x$.
{\rm (iii)} Suppose that $f$
is paraconcave. Then $f_-''(\bar x, p,
\cdot)$ is also paraconcave.
\end{pro}
\begin{proof}
(i) Let $u\in H$ and take any $u_n\rightarrow u$. For each $n\in\mathbb{N}$, by the definition of $f_-''(\bar x, p, u_n)$, there exist $w_n\in B(u_n, \frac{1}{n})$ and $t_n\in (0, \frac{1}{n})$ such that
$$
\frac{2(f(\bar x+t_nw_n)-f(\bar x)-t_n\langle p, w_n\rangle)}{t^2_n}-\frac{1}{n}<f_-''(\bar x, p, u_n).
$$
This implies that
\begin{eqnarray*}
\liminf_{n\rightarrow \infty}f_-''(\bar x, p, u_n)&\geq&\liminf_{n\rightarrow\infty}\big(\frac{2(f(\bar x+t_nw_n)-f(\bar x)-t_n\langle p, w_n\rangle)}{t^2_n}-\frac{1}{n}\big)\\
&\geq&\liminf_{u'\rightarrow u, t\downarrow 0}\frac{2(f(\bar x+tu')-f(\bar x)-t\langle p, u'\rangle)}{t^2}\\
&=&f_-''(\bar x, p, u).
\end{eqnarray*}
Hence $f_-''(\bar x, p, \cdot)$ is lower semicontinuous at $u$.
(ii) Suppose that there exist $\delta, \lambda>0$ such that $f-\frac{1}{2\lambda}\|\cdot\|^2$ is concave on $B(\bar x, \delta)$. We next prove that $f_-''(\bar x, p, \cdot)-\frac{1}{\lambda}\|\cdot\|^2$ is concave on $B(\bar x, \delta)$.
Let $u, v\in B(\bar x, \delta)$ and $\mu\in [0, 1]$, and take any $z\rightarrow \mu u+(1-\mu)v$ and $t\rightarrow 0$. Define $u':=u+z-(\mu u+(1-\mu)v)$ and $v':=v+z-(\mu u+(1-\mu)v)$. Then
$$z=\mu u'+(1-\mu)v'\ \ {\rm and}\ \ (u', v')\rightarrow (u, v)\ \ {\rm as}\ \ z\rightarrow \mu u+(1-\mu)v.
$$
Note that
\begin{equation*}
\begin{array}l
\ \ \ \frac{2(f(\bar x+tz)-f(\bar x)-t\langle p, z\rangle)}{t^2}-\frac{1}{\lambda}\|\mu u+(1-\mu)v\|^2\\
=\frac{2(f(\bar x+tz)-\frac{1}{2\lambda}\|\bar x+tz\|^2-f(\bar x)-t\langle p, z\rangle+\frac{1}{2\lambda}\|\bar x+tz\|^2)}{t^2}-\frac{1}{\lambda}\|\mu u+(1-\mu)v\|^2 \\
\geq\frac{2\mu(f(\bar x+tu')-\frac{1}{2\lambda}\|\bar x+tu'\|^2)}{t^2}+\frac{2(1-\mu)(f(\bar x+tv')-\frac{1}{2\lambda}\|\bar x+tv'\|^2)}{t^2}+\\
\ \ \ \frac{\frac{1}{2\lambda}\|\bar x+t(\mu u'+(1-\mu)v')\|^2-f(\bar x)-t\langle p, z\rangle}{t^2}-\frac{1}{\lambda}\|\mu u+(1-\mu)v\|^2\\
=\frac{2\mu(f(\bar x+tu')-f(\bar x)-t\langle p, u'\rangle)}{t^2}+\frac{2(1-\mu)(f(\bar x+tv')-f(\bar x)-t\langle p, v'\rangle)}{t^2}-\\
\ \ \ \frac{\frac{\mu(1-\mu)}{2\lambda}\|\bar x+tu'-(\bar x+tv')\|^2}{t^2}-\frac{1}{\lambda}\|\mu u+(1-\mu)v\|^2\\
=\frac{2\mu(f(\bar x+tu')-f(\bar x)-t\langle p, u'\rangle)}{t^2}+\frac{2(1-\mu)(f(\bar x+tv')-f(\bar x)-t\langle p, v'\rangle)}{t^2}-\\
\ \ \ \frac{\mu(1-\mu)}{2\lambda}\|u'-v'\|^2-\frac{1}{\lambda}\|\mu u+(1-\mu)v\|^2.
\end{array}
\end{equation*}
By taking lower limits, one has
\begin{eqnarray*}
&&\liminf_{z\rightarrow\mu u+(1-\mu)v, t\downarrow 0}\frac{2(f(\bar x+tz)-f(\bar x)-t\langle p, z\rangle)}{t^2}-\frac{1}{\lambda}\|\mu u+(1-\mu)v\|^2\\
&\geq&\mu\liminf_{u'\rightarrow u, t\downarrow 0}\frac{2\big(f(\bar x+tu')-f(\bar x)-t\langle p, u'\rangle\big)}{t^2}+\\
&&(1-\mu)\liminf_{u'\rightarrow u, t\downarrow 0}\frac{2\big(f(\bar x+tv')-f(\bar x)-t\langle p, v'\rangle\big)}{t^2}-\\
&&\frac{\mu(1-\mu)}{\lambda}\|u-v\|^2
-\frac{1}{\lambda}\|\mu u+(1-\mu)v\|^2\\
&\geq&\mu(f_-''(\bar x, p, u)-\frac{1}{\lambda}\|u\|^2)+(1-\mu)(f_-''(\bar x, p, v)-\frac{1}{\lambda}\|v\|^2)\\
&&+\frac{\mu}{\lambda}\|u\|^2+\frac{1-\mu}{\lambda}\|v\|^2-\frac{\mu(1-\mu)}{\lambda}\|u-v\|^2
-\frac{1}{\lambda}\|\mu u+(1-\mu)v\|^2\\
&=&\mu(f_-''(\bar x, p, u)-\frac{1}{\lambda}\|u\|^2)+(1-\mu)(f_-''(\bar x, p, v)-\frac{1}{\lambda}\|v\|^2).
\end{eqnarray*}
This implies that $f_-''(\bar x, p, \cdot)-\frac{1}{\lambda}\|\cdot\|^2$ is concave on $B(\bar x, \delta)$.
Note that (iii) follows from (ii) and thu the proof is completed.
\end{proof}
For a twice epi-differentiable function $f$ defined on a finite-dimensional space, it is shown in \cite[Corollary 8.47]{19} that second-order epi-derivative of $f$ closely relates to protoderivative of limiting subdifferential $\partial f$. The authors \cite{28} also studied the relationship between the second-order epi-derivative and the protoderivative in a Hilbert space. In order to study the application of twice epi-differentiability in this paper, we first recall the following two important concepts of set convergence.
For a sequence $\{C_n: n\in \mathbb{N}\}$
of closed subsets in $H$,
$\liminf_{n\rightarrow\infty}C_n$ denotes the set of all limit points of sequences $\{x_n\}$ with $x_n\in C_n$ for all $n\in\mathbb{N}$, and $\limsup_{n\rightarrow\infty}C_n$ denotes the set of all cluster points of such sequences.
Recall that $\{C_n: n\in \mathbb{N}\}$ is said to be {\it Painlev\'{e}-Kuratowski convergent} to a subset $C$ of $H$, if
$$
C=\liminf\limits_{n\rightarrow\infty}C_n=\limsup\limits_{n\rightarrow\infty}C_n.
$$
and that $\{C_n: n\in \mathbb{N}\}$ is said to be {\it Mosco convergent} to a subset $C$ of $H$, if
$$
C=\liminf\limits_{n\rightarrow\infty}C_n={\rm w}\mbox{-}\limsup\limits_{n\rightarrow\infty}C_n,
$$
where ${\rm w}\mbox{-}\limsup\limits_{n\rightarrow \infty}C_n$ is the set of all weak cluster points of sequences from the sets $C_n$, that is, $x\in {\rm w}\mbox{-}\limsup\limits_{n\rightarrow \infty}C_n$ if and only if there exists a sequence $\{x_n\}$ such that $x_n\in C_n$ for all $n\in\mathbb{N}$ and a subsequence of $\{x_n\}$ converges to $x$ with respect to the weak topology.
Let $f, f_n: H\rightarrow \mathbb{R}\cup\{+\infty\}(n=1,2,\cdots)$ be proper
lower semicontinuous functions. We say that $\{f_n\}$ is {\it Mosco} (resp.{\it Painlev\'{e}-Kuratowski}) epi-convergent to $f$, if ${\rm epi}(f_n)$ is Mosco (resp. Painlev\'{e}-Kuratowski) convergent to ${\rm epi}(f)$; in the Mosco epi-convergent (resp. Painlev\'{e}-Kuratowski epi-convergent) case we write
$$
f={\rm M}\mbox{-}\lim_{n\rightarrow \infty}f_n ({\rm resp.}\ f={\rm PK}\mbox{-}\lim_{n\rightarrow \infty}f_n).
$$
Recall that $f: H\rightarrow \mathbb{R}\cup\{+\infty\}$ is said to be {\it twice epi-differentiable} at $\bar x$ relative to $p\in \partial_pf(\bar x)$ in the sense of Mosco (resp. Painlev\'{e}-Kuratowski), if the second-order difference quotient functions $\Delta_2f(\bar x, p, t, \cdot)$ are Mosco epi-convergent (resp. Painlev\'{e}-Kuratowski epi-convergent) to a proper function as $t\rightarrow 0^+$; that is $f$ is twice epi-differentiable at $\bar x$ relative to $p\in\partial_p f(\bar x)$ in the sense of Mosco (resp. Painlev\'{e}-Kuratowski) if and only if for any sequence $\{t_n\}$ in $(0, +\infty)$ convergent to 0, the function sequence $\{\Delta_2f(\bar x, p, t_n,
\cdot)\}$ is Mosco epi-convergent (resp. Painlev\'{e}-Kuratowski epi-convergent) to the same proper function. The Mosco epi-limit of these second-order difference quotient functions is called {\it second-order epi-derivative} of $f$ at $\bar x$ relative to $p$ and is denoted by $f_-''(\bar x; p)(\cdot)$. In this case, one can easily verify that
\begin{equation}\label{3.8a}
f_{-}''(\bar x;p)(h)=f_{-}''(\bar x,p, h)\ \ \forall h\in H.
\end{equation}
\noindent{\bf Remark 3.1.} From the definition, it is known that twice epi-differentiability of a function in the sense of Mosco is stronger than that in the sense of Painlev\'{e}-Kuratowski in the Hilbert space. When the space is finite-dimensional, both concepts coincide and reduce to the corresponding notion of twice epi-differentiability (cf. \cite[Definition 22]{11}). Unless otherwise stated, the twice epi-differentiability of a function studied in this paper is in the sense of Mosco.\\
The following theorem is a key tool in proving main results of this paper. Readers are invited to consult \cite[Theorem 4.2]{28} for more details and its proof.\\
\noindent{\bf Theorem A.} {\it Let $f, f_n: H\rightarrow \mathbb{R}\cup\{+\infty\}(n=1,2,\cdots)$ be proper
lower semicontinuous functions such that $f={\rm M}\mbox{-}\lim\limits_{n\rightarrow \infty}f_n$. Then for any $p\in \partial f(x)$ there exist sequences $\{(x_n, p_n)\}$ in $H\times H$ and a strictly increasing sequence $\{n_k\}$ in $\mathbb{N}$ such that}
$$
(x_{n_k}, f_{n_k}(x_{n_k}))\longrightarrow (x, f(x)), \ \ p_{n_k}\stackrel{w}\longrightarrow p\ \ and\ \ p_{n_k}\in \partial_pf_{n_k}(x_{n_k})\ {\it for\ all} \ k\in\mathbb{N}.
$$
The following proposition refers to epi-convergence of functions and epigraph of second-order lower Dini-directional derivative.
\begin{pro}
Let $f:H\rightarrow \mathbb{R}\cup\{+\infty\}$ be a proper lower
semicontinuous function and $\bar x\in{\rm dom}f$ with $p\in\partial_pf(\bar x)$. Then
\begin{equation}\label{3.9a}
\liminf_{t\rightarrow 0^+}{\rm epi}\big(\frac{1}{2}\Delta_2f(\bar x,
p, t, \cdot)\big)\subset {\rm epi}\big(\frac{1}{2}f_-''(\bar x, p, \cdot)\big)
\end{equation}
and
\begin{equation}\label{3.9b}
{\rm epi}\big(\frac{1}{2}f_-''(\bar x, p, \cdot)\big)\subset \limsup_{t\rightarrow 0^+}{\rm epi}\big(\frac{1}{2}\Delta_2f(\bar x,
p, t, \cdot)\big).
\end{equation}
\end{pro}
\begin{proof}
Let $(h, r)\in\liminf\limits_{t\rightarrow 0^+}{\rm epi}\big(\frac{1}{2}\Delta_2f(\bar x,
p, t, \cdot)\big)$. Then for any $t_k\rightarrow 0^+$, there exists $(h_k, r_k)\in {\rm epi}\big(\frac{1}{2}\Delta_2f(\bar x,
p, t_k, \cdot)\big) (k\in \mathbb{N})$ such that $(h_k, r_k)\rightarrow (h, r)$. This implies that
$$
\frac{1}{2} f_-''(\bar x,
p, h)\leq\liminf_{k\rightarrow \infty}\frac{1}{2}\Delta_2f(\bar x,
p, t_k, h_k)\leq\liminf_{k\rightarrow \infty}r_k=r.
$$
Hence $(h, r)\in{\rm epi}\big(\frac{1}{2}f_-''(\bar x, p, \cdot)\big)$ and consequently \eqref{3.9a} holds.
Next, let $(h, r)\in{\rm epi}\big(\frac{1}{2}f_-''(\bar x, p, \cdot)\big)$. By \eqref{3.5}, there are sequences $h_k\rightarrow h$ and $t_k\rightarrow 0^+$ such that $\frac{1}{2}\Delta_2f(\bar x,
p, t_k, h_k)\rightarrow \frac{1}{2}f_-''(\bar x, p, h)$. For each $k\in \mathbb{N}$, let
$$
r_k:=\frac{1}{2}\Delta_2f(\bar x,
p, t_k, h_k)+r-\frac{1}{2}f_-''(\bar x, p, h).
$$
Then $(h_k, r_k)\in {\rm epi}\big(\frac{1}{2}\Delta_2f(\bar x,
p, t_k, \cdot)\big)$ and $(h_k, r_k, t_k)\rightarrow (h, r, 0^+)$. From this, one has
$$
(h, r)\in \limsup_{t\rightarrow 0^+}{\rm epi}\big(\frac{1}{2}\Delta_2f(\bar x,
p, t, \cdot)\big).
$$
Thus \eqref{3.9b} holds. The proof is completed.
\end{proof}
\setcounter{equation}{0}
\section{Second-order optimality conditions}
This section is devoted to the study of second-order optimality conditions defined by three generalized second-order derivatives in the Hilbert space and the equivalence interrelationship among them. We begin with three types of second-order optimality conditions of an extended real-valued nonsmooth function on finite-dimension space studied by Eberhard and Wenczel \cite{11} (cf. \cite[Definition 56]{11} and \cite[Definition 6.1]{EM}).\\
\noindent{\bf Definition 4.1.} {\it Let $f:\mathbb{R}^m\rightarrow \mathbb{R}\cup\{+\infty\}$ be a proper lower semicontinuous function and assume that the first-order condition $0\in\partial_pf(\bar x)$ holds.
1. We say that $f$ satisfies the second-order condition of the first kind at $\bar x$, if there exists $\beta\in(0,+ \infty)$ such that $f_-''(\bar x, 0, h)\geq\beta$ for all $h\in S_{\mathbb{R}^m}$.
2. We say that $f$ satisfies the second-order condition of the second kind at $\bar x$, if there exists $\beta\in(0,+ \infty)$ satisfying for all $h\in{\rm dom}D^2f(\bar x, 0)\cap S_{\mathbb{R}^m}$, there is $z\in D^2f(\bar x, 0)(h)$ such that $\langle z, h\rangle\geq\beta$.
3. We say that $f$ satisfies the second-order condition of the third kind at $\bar x$, if there exists $\beta\in(0,+ \infty)$ such that for any $h\in S_{\mathbb{R}^m}$ and any $z\in\hat{\partial}^2f(\bar x, 0)(h)$, one has $\langle z, h\rangle\geq\beta$.}\\
It is known that Eberhard and Wenczel \cite{11} mainly investigate the close interrelationship among these optimality conditions and proved the following result on the equivalence among these optimality conditions in Definition 4.1 for paraconcave functions (cf. \cite[Theorem 6.3]{EM} and \cite[Theorem 66]{11}).\\
\noindent{\bf Theorem B.} {\it Let $f:\mathbb{R}^m\rightarrow \mathbb{R}\cup\{+\infty\}$ be a prox-bounded and lower semicontinuous function with $0\in\partial_pf(\bar x)$. Suppose that $f$ is finite and there exists $c>0$ such that $f-\frac{c}{2}\|\cdot\|^2$ and $f_-''(\bar x, 0, \cdot)-c\|\cdot\|^2$ are concave. Then all second-order optimality conditions are equivalent. Moreover, the same $\beta$ value may be used in each condition.}\\
As one part of main work in this paper, it is natural to study the original forms of second-order optimality conditions in the Hilbert space. However, the existing implication among these optimality conditions given in Definition 4.1 may not be valid for the case of Hilbert space (comparing Proposition 4.2 below with \cite[Proposition 45]{11}), and thus it is necessary to make some minor modification to these optimality conditions. Motivated by this observation, we consider the following second-order optimality conditions in the Hilbert space.\\
\noindent{\bf Definition 4.2.} {\it Let $f:H\rightarrow \mathbb{R}\cup\{+\infty\}$ be a lower semicontinuous function and assume that the first-order optimality
condition $0\in\partial_pf(\bar x)$ holds.
{\rm (i)} We say that $f$ satisfies the second-order optimality condition of the first kind at $\bar x$, if there exists $\beta>0$ such that $ f_-''(\bar x, 0, h)\geq \beta$ for all $h\in S_H$.
{\rm (ii)} We say that $f$ satisfies the second-order optimality condition of the second kind at $\bar x$, if there exists $\beta>0$ such that for all $h\in{\rm dom}D_M^2f(\bar x, 0)\cap S_H$, there is $z\in D^2_Mf(\bar x, 0)(h)$ such that $
\langle z, h\rangle\geq\beta$.
{\rm (iii)} We say that $f$ satisfies the second-order optimality condition of the third kind at $\bar x$, if there exists $\beta>0$ such that for all $h\in S_H\cap{\rm dom}\partial^2_Mf(\bar x, 0)$ and $z\in \partial_M^2f(\bar x, 0)(h)$, one has $\langle z, h\rangle\geq \beta$.}\\
\noindent{\bf{Remark 4.1.}} Note that mixed contingent cone and contingent cone coincide in the finite-dimensional space setting. Hence when restricted to the finite-dimensional space, second-order optimality conditions in Definition 4.2 reduce to those studied in \cite{11,EM} as Definition 4.1.\\
Now, we pay main attention to the equivalence interrelationship among these second-order optimality conditions in Definition 4.2. We first provide the following proposition whose proof mainly relies on Theorem A aforementioned in Section 3. This proposition is one key tool to prove main results in this section.
\begin{pro}
Let $f:H\rightarrow\mathbb{R}\cup\{+\infty\}$ be a proper lower semicontinuous function
and $\bar x\in{\rm dom}f$ with $p\in \partial_pf(\bar x)$. Suppose that $f$ is twice
epi-differentiable at $\bar x$ for $p$. Then
\begin{equation}\label{4.1a}
\frac{1}{2}\partial f''_-(\bar x, p, \cdot)(h)\subset D_M^2f(\bar x, p)(h)\ \ \forall \ h\in H
\end{equation}
and
\begin{equation}\label{4.11}
f_-''(\bar x, p, h)\leq \sup\Big\{\langle w, h\rangle: w\in D^2_Mf(\bar x, p)(h)\Big\}
\end{equation}
holds for all $h\in{\rm dom}\partial f''_-(\bar x, p, \cdot)$.
Assume further that $f$ is a paraconcave function and continuous at $\bar x$. Then
\begin{equation}\label{4-2a}
{\rm dom} D^2_Mf(\bar x, p)={\rm dom}\partial f_-''(\bar x, p, \cdot)=H
\end{equation}
and
\begin{equation}\label{4.12}
f_-''(\bar x, p, h)\leq \sup\Big\{\langle w, h\rangle: w\in D^2_Mf(\bar x, p)(h)\Big\} \ \ \forall h\in H.
\end{equation}
\end{pro}
\begin{proof}
Let $h\in{\rm dom}\partial f''_-(\bar x, p, \cdot)$ and $z\in \frac{1}{2}\partial f''_-(\bar x, p, \cdot)(h)$. Since $f$ is twice epi-differentiable at $\bar x$ for $p$, by \eqref{3.8a} and Theorem A, there exist sequences $t_n\rightarrow 0^+$, $h_n\rightarrow h$ and
$z_n\stackrel{w}\longrightarrow z$ such that
$$
\Delta_2f(\bar x, p, t_n,
h_n)\rightarrow
f_-''(\bar x, p, h)\ {\rm and}\ 2z_n\in\partial_p \Delta_2f(\bar x, p, t_n,
\cdot)(h_n).
$$
By virtue of \eqref{3.6}, one has
$$
z_n\in\partial_p\big(\frac{1}{2}\Delta_2f(\bar x, t_n, p, \cdot)\big)(h_n)=\frac{1}{t_n}(\partial_pf(\bar x+t_nh_n)-p).
$$
This implies that $(x+t_nh_n, p+t_nz_n)\in{\rm gph}(\partial_pf)$ and consequently it follows from \eqref{3.1} that $(h, z)\in T_M({\rm gph}(\partial_pf), (\bar x, p))$. Thus $z\in D_M^2f(\bar x,p)(h)$ and \eqref{4.1a} holds.
Noting that $f''_-(\bar x, p, \cdot)$ is 2-positively homogeneous and $2z\in \partial f''_-(\bar x, p, \cdot)(h)$, it follows from \cite[Theorem 3.1]{YW1} that $f''_-(\bar x, p, h)=\langle z, h\rangle $ and thus \eqref{4.11} holds.
Assume that $f$ is a paraconcave function and continuous at $\bar x$. We can take $\lambda>0$
such that $g(u):=f(u)-\frac{1}{2\lambda}\|u\|^2$ is concave. Using Proposition 3.1 and Corollary 3.1, one has that $f_-''(\bar x, p, \cdot)-\frac{1}{\lambda}\|\cdot\|^2$ is also concave. By computing, one has ${\rm dom}\partial_pf_-''(\bar
x, p, \cdot)={\rm dom}\partial_pg_-''(\bar x, q, \cdot)$ and
$$
g_-''(\bar x, q, u)=f_-''(\bar x, p, u)-\frac{1}{\lambda}\|u\|^2 \ \ \forall u\in H,
$$
where $q:=p-\frac{1}{\lambda}\bar x\in \partial_pg(\bar x)$. Thus, $g_-''(\bar x, q, \cdot)$ is concave. We claim that
\begin{eqnarray}\label{4.13}
{\rm dom}f_-''(\bar x, p, \cdot)={\rm dom}g_-''(\bar x, q, \cdot)=H.
\end{eqnarray}
Let $h\in H$. From $q\in \partial_pg(\bar x)$, there exist $r_0,\delta_0>0$ such that
\begin{equation}\label{4-5}
g(x)\geq g(\bar x)+\langle q, x-\bar x\rangle-\frac{r_0}{2}\|x-\bar x\|^2\ \ \forall x\in B(\bar x, \delta_0).
\end{equation}
Since $g$ is concave and $f$ is continuous at $\bar x$, it follows that $-g$ is continuous at $\bar x$ and $\partial_p(-g)(\bar x)\not=\emptyset$. Then we can choose $\zeta\in\partial_p(-g)(\bar x)$ and it follows from the convexity of $-g$ that
\begin{equation}\label{4-6}
\langle-\zeta,x-\bar x\rangle\geq g(x)-g(\bar x)\ \ \forall x\in H.
\end{equation}
By \eqref{4-5} and \eqref{4-6}, for any $v\in H$ and any $t>0$ sufficiently small, one has
\begin{eqnarray*}
\langle-\zeta, tv\rangle\geq g(\bar x+tv)-g(\bar x)\geq \langle q, tv\rangle-\frac{r_0}{2}\|tv\|^2.
\end{eqnarray*}
From this, one can verify that $q=-\zeta$ and thus
$$
\langle q, tv\rangle\geq g(\bar x+tv)-g(\bar x)\ \ \forall t>0\ {\rm and}\ \forall v\in H.
$$
This implies that
\begin{eqnarray}\label{4.14}
g_-''(\bar x, q, h)=\liminf_{h'\rightarrow h, t\downarrow 0}\frac{2(g(\bar x+th')-g(\bar x)-\langle q, th'\rangle)}{t^2}\leq 0.
\end{eqnarray}
Noting that $p\in \partial_pf(\bar x)$, there are $r, \delta>0$ such that
$$
f(x)\geq f(\bar x)+\langle p, x-\bar x\rangle-\frac{r}{2}\|x-\bar x\|^2\,\,\,\,\forall x\in B(\bar x, \delta).
$$
Thus,
\begin{eqnarray}\label{4.15}
\ \ \ \ f_-''(\bar x, p, h)=\liminf_{h'\rightarrow h, t\downarrow 0}\frac{2(f(\bar x+th')-f(\bar x)-\langle p, th'\rangle)}{t^2}\geq -r\|h\|^2.
\end{eqnarray}
Since $g_-''(\bar x, q, h)=f_-''(\bar x, p, h)-\frac{1}{\lambda}\|h\|^2$, it follows from \eqref{4.14} and \eqref{4.15} that
$$
-\infty< f_-''(\bar x, p, h) <+\infty,
$$
and consequently $g_-''(\bar x, q, h)\in \mathbb{R}$. Hence \eqref{4.13} holds.
We next prove that
\begin{equation}\label{4.9c}
{\rm dom}\partial f_-''(\bar x, p, \cdot)=H.
\end{equation}
Granting this, it follows that \eqref{4-2a} holds and \eqref{4.12} holds by \eqref{4.11}.
Let $h\in {\rm dom}\partial f_-''(\bar x, p, \cdot)$. Since $f_-''(\bar x, p, \cdot)$ is lower semicontinuous (by Proposition 3.1) and $h\in H={\rm dom}f_-''(\bar x, p, \cdot)$, by virtue of Density Theorem (cf. \cite[Theorem 3.1]{8}), there exists $(h_k, z_k)\in H\times H$ such that $2z_k\in \partial_pf_-''(\bar x, p, h_k)$ and $h_k\rightarrow h$. Now, using the concavity of $g_-''(\bar x, q, \cdot)$, one has that $g_-''(\bar x, q, \cdot)$ is locally Lipschtzian, and so is $f_-''(\bar x, p, \cdot)$. Then, by virtue of \cite[Theorem 7.3]{8}, there exists $L:=L(h)>0$ such that when $k$ is sufficiently large, one has $\|z_k\|\leq L$. By applying \cite[Corollary 2.8.9]{15}, $\{z_k\}$ has a weakly convergent subsequence $\{z_{k_i}\}$ and so we can assume that $z_{k_i}\stackrel{w}\longrightarrow z\in H$ as $i\rightarrow \infty$. This implies that $2z\in \partial f_-''(\bar x, p, \cdot)(h)$ as $h_{k_i}\rightarrow h (i\rightarrow \infty)$ and $2z_{k_i}\in \partial_pf_-''(\bar x, p, \cdot)(h_{k_i})$. Hence $h\in {\rm dom}\partial f_-''(\bar x, p, \cdot)$. The proof is completed.
\end{proof}
\noindent{\bf{Remark 4.2}} The proof of Proposition
4.1 is inspired by the idea from the proof of \cite[Proposition 45]{11} and Proposition 4.1 is an extension and improvement of \cite[Proposition 45]{11} since for the case of twice epi-differentiable function, the assumption that ``{\it$h\mapsto f_-''(\bar x, p, h)$ is paraconcave}" is dropped from Proposition 4.1 and the conclusion ``{\it ${\rm dom}D_M^2f(\bar x, p)=H$}" is stronger than that ``{\it${\rm dom}D^2f(\bar x, p)$ is dense in $\mathbb{R}^m$}" in \cite[Proposition 45]{11} ($D_M^2f(\bar x, p)$ coincides with $D^2f(\bar x, p)$ in $\mathbb{R}^m$).\\
The following proposition follows from Proposition 4.1.
\begin{pro}
Let $f:H\rightarrow \mathbb{R}\cup\{+\infty\}$ be a lower
semicontinuous function and $\bar x\in{\rm dom}f$ with $p\in \partial_pf(\bar x)$. Suppose that $f$ is twice epi-differentiable at $\bar x$ for $p$. Then
\begin{equation}\label{4.16}
\sup\big\{\langle w, h\rangle: w\in \partial^2_Mf(\bar x, p)(h)\big\}\leq f_-''(\bar x, p, h)
\end{equation}
and
\begin{equation}\label{4.10a}
\partial^2_Mf(\bar x, p)(h)\subset\{w\in H: \langle w, h\rangle\leq f_-''(\bar x, p, h)\}
\end{equation}
hold for all $h\in
{\rm dom}\partial f_-''(\bar x, p, \cdot)$.
\end{pro}
\begin{proof} Let $h\in {\rm dom}\partial f_-''(\bar x, p, \cdot)$ and choose $z\in H$ such that $z\in \frac{1}{2}\partial f_-''(\bar x, p, \cdot)(h)$. Then, by \cite[Theorem 3.1]{YW1}, one has
\begin{equation}\label{4.11a}
f_-''(\bar x, p, h)=\langle h, z\rangle.
\end{equation}
Let $w\in \partial^2_Mf(\bar x, p)(h)$. Then
\begin{eqnarray}\label{4.12a}
\langle w, y\rangle\leq\langle h, v\rangle\ \ \forall (y, v)\in T_M({\rm gph}(\partial_pf), (\bar x, p)).
\end{eqnarray}
Note that $z\in \frac{1}{2}\partial f_-''(\bar x, p, \cdot)(h)$ and thus $z\in D^2_Mf(\bar x, p)(h)$ by Proposition 4.1. Using \eqref{4.11a} and \eqref{4.12a}, one has
$$
\langle w, h\rangle\leq\langle h, z\rangle=f_-''(\bar x,p,h).
$$
This means that \eqref{4.16} holds and so does \eqref{4.10a}. The proof is completed.
\end{proof}
Combining Proposition 4.1 with Proposition 4.2, we have the following theorem.
\begin{them}
Let $f:H\rightarrow \mathbb{R}\cup\{+\infty\}$ be a proper lower
semicontinuous and paraconcave function and $\bar x\in{\rm dom}f$ with $0\in \partial_pf(\bar x)$. Suppose that $f$ is continuous at $\bar x$ and twice epi-differentiable at $\bar x$ for $0$. Then the second-order optimality condition of the first kind implies that of the second kind. Furthermore, if $f$ satisfies the second-order
optimality condition of the third kind at $\bar x$, then there exists $\beta>0$ such that
\begin{eqnarray}\label{4.17}
f_-''(\bar x, 0, h)\geq \beta
\end{eqnarray}
holds for all $h\in S_H\cap {\rm dom}\partial^2_Mf(\bar x, 0)$.
\end{them}
\begin{proof}
Suppose that there exists $\beta>0$ such that the second-order optimality condition of the first kind holds. Let $\beta_1\in (0, \beta)$ and $h\in {\rm dom}D^2_Mf(\bar x, 0)\cap S_H$. Since $f$ is paraconcave, it follows from Proposition 4.1 and optimality condition of the first kind that
$$
\sup\Big\{\langle w, h\rangle: w\in D^2_Mf(\bar x, 0)(h)\Big\}\geq f_-''(\bar x, 0, h)\geq \beta>\beta_1.
$$
Then there exists $z\in D^2_M(\partial_pf)(\bar x, 0)(h)$ such that $\langle z, h\rangle\geq\beta_1$. Hence $f$ satisfies the second-order optimality condition of the second kind at $\bar x$ with constant $\beta_1$.
Suppose that there exists $\beta>0$ such that the second-order
optimality condition of the third kind holds. Let $h\in S_H\cap {\rm dom}\partial^2_Mf(\bar x, 0)$. Since $f$ is paraconcave, by Proposition 4.1, one has $h\in {\rm dom}\partial f_-''(\bar x, p, \cdot)$ and it follows from Proposition 4.2 and the second-order optimality condition of the third kind that
$$
f_-''(\bar x, 0, h)\geq \sup\big\{\langle w, h\rangle: w\in \partial^2_Mf(\bar x, 0)(h)\big\}\geq\beta.
$$
The proof is completed.
\end{proof}
Next, we study the duality between second-order optimality conditions of the second and the third kinds. In order to deal with
it, we denote the following linear mapping by ${\rm Proj}_h(h',
z'):=h'$ and ${\rm Proj}_z(h', z'):=z'$. For a subset $S\subset
H\times H$, let
$$
{\rm Proj}_hS:=\{h\in H: \exists (h, z)\in S\} \ \ {\rm and} \ \
{\rm Proj}_zS:=\{z\in H: \exists (h, z)\in S\}.
$$
Then, when $0\in
\partial_pf(\bar x)$, we have
$$
{\rm dom}\partial^2_Mf(\bar x, 0)=-{\rm
Proj}_z N_M({\rm gph}(\partial f), (\bar x, 0))
$$
and
$$
{\rm dom}D^2_Mf(\bar x, 0)={\rm Proj}_hT_M({\rm gph}(\partial
f), (\bar x, 0)).
$$
\begin{pro} Let $T$ be a closed convex cone in $H\times H$ and $T(h):=\{z\in H: (h, z)\in T\}$, and let $N:=T^{\circ}$.
Consider the following statements:
{\rm (i)} there exists $\beta>0$ such that $\langle h, z\rangle\leq -\beta$ for all $(h, z)\in N$ with $\|z\|=1$;
{\rm (ii)} there exists $\beta_1>0$ such that for each $h\in S_H$ with $h\in {\rm sqri(Proj}_hT)$, there exists $(h, z)\in T$ such that $\langle h, z\rangle\geq \beta_1$.
Then ${\rm (i)}\Rightarrow{\rm (ii)}$. Furthermore, we assume that ${\rm qri}N\not=\emptyset$ and
\begin{equation}\label{4.18}
{\rm qri}({\rm Proj}_zN)\subset{\rm sqri(Proj}_hT).
\end{equation}
Then ${\rm (ii)}\Rightarrow{\rm (i)}$.
\end{pro}
\begin{proof} ${\rm (i)}\Rightarrow{\rm (ii)}$:
Let $\beta_1\in (0, \beta)$ and $h\in S_H\cap {\rm sqri(Proj}_hT)$. Define $v(h):=\sup\{\langle h, z\rangle: (h, z)\in T\}$. If $v(h)=+\infty$, then the conclusion holds. Next, we assume that $v(h)<+\infty$. Then, by computing, we have
\begin{equation}\label{4.19}
-v(h)=\inf_{(\hat h, \hat z)}\{\delta_T(\hat h, \hat z)+(\delta_{\{h\}}(\hat h)-\langle \hat h, \hat z\rangle)\}.
\end{equation}
Let $f(\hat h, \hat z):=\delta_T(\hat h, \hat z)$ and $g(\hat h, \hat z):=\delta_{\{h\}}(\hat h)-\langle \hat h, \hat z\rangle$. Then it is easy to verify that $g$ is convex. Noting that $h\in{\rm sqri(Proj}_hT)$, it follows that
$$
(0, 0)\in{\rm sqri(dom}f-{\rm dom}g).
$$
Applying the Fenchel duality in infinite-dimensional spaces(cf. \cite{3, 20}), one has
\begin{eqnarray*}
-v(h)&=&\sup_{(h^*, z^*)}\{-g^*(h^*, z^*)-f^*(-h^*, -z^*))\}\\
&=&\sup_{(h^*, z^*)}\{-\langle h, h^*\rangle-\delta_{\{-h\}}(z^*)-\delta_{T^{\circ}}(-h^*, -z^*)\}\\
&=&\sup_{h^*}\{-\langle h, h^*\rangle-\delta_{T^{\circ}}(-h^*, h)\}.
\end{eqnarray*}
Hence
$$
v(h)=\inf_{h^*}\{\langle h, h^*\rangle+\delta_{T^{\circ}}(-h^*, h)\}=\inf_{\{h^*: (h^*, h)\in T^{\circ}\}}\{\langle h, -h^*\rangle\}.
$$
Since $v(h)<+\infty$, then $\{h^*: (h^*, h)\in T^{\circ}\}\not=\emptyset$. For each $h^*\in H$ with $(h^*, h)\in T^{\circ}=N$, by the assumption, one has $\langle h, -h^*\rangle\geq \beta$ and consequently $v(h)\geq \beta>\beta_1$. Thus there exists $z\in T(h)$ such that $\langle h, z\rangle\geq \beta_1$.
${\rm (ii)}\Rightarrow{\rm (i)}$: Let $(h^*, h)\in N$ with $\|h\|=1$. By \cite[Proposition 2.5]{3}, there exists $(h^*_n, h_n)\in {\rm qri}N$ such that $(h^*_n, h_n)\rightarrow (h^*, h)$ and $\|h^*_n\|\rightarrow \|h\|=1$. Let $\hat{h}_n:=\frac{h_n}{\|h_n\|}$. Noting that $h_n\in{\rm qri}({\rm Proj}_zN)$ and $0\in {\rm Proj}_zN$, it follows that $\hat h_n\in {\rm qri}({\rm Proj}_zN)$ as $N$ and ${\rm Proj}_zN$ are cones. By the assumption, we have $\hat h_n\in {\rm sqri(Proj}_hT)$. Applying the Fenchel duality again, one has
\begin{equation}\label{4.20}
\inf_{\{h^*: (h^*, \hat h_n)\in T^{\circ}\}}\langle -h^*, \hat h_n\rangle=v(\hat h_n)\geq \beta_1.
\end{equation}
Since $(\frac{h^*_n}{\|h_n\|}, \hat h_n)\in N$ (thanks to $(h^*_n, h_n)\in N$), it follows from \eqref{4.20} that
$$
\langle -\frac{h^*_n}{\|h_n\|}, \hat h_n\rangle\geq\inf_{\{h^*: (h^*, \hat h_n)\in T^{\circ}\}}\langle -h^*, \hat h_n\rangle\geq\beta_1.
$$
Taking limits as $n\rightarrow \infty$, we have $\langle -h^*, h\rangle\geq\beta_1$. Hence $\langle h^*, h\rangle\leq-\beta_1$ and consequently (i) holds for $\beta=\beta_1$. The proof is completed.
\end{proof}
By using Proposition 4.3, the following theorem is immediate.
\begin{them}
Let $f:H\rightarrow \mathbb{R}\cup\{+\infty\}$ be a proper lower
semicontinuous function and $\bar x\in{\rm dom}f$ with $0\in \partial_pf(\bar x)$. Suppose that ${\rm qri}( N_M({\rm gph}(\partial_p f), (\bar x, 0)))\not=\emptyset$ and
\begin{equation}\label{4.22}
{\rm qri}\big({\rm dom}\partial^2_Mf(\bar x, 0)\big)\subset -{\rm sqri(dom}D^2_Mf(\bar x, 0)).
\end{equation}
Then the second-order optimality condition of the second kind implies that of the third kind.
\end{them}
\begin{proof}
Let $T:=\overline{\rm co}(T_M({\rm gph}(\partial_pf), (\bar x, 0)))$ and $N:=T^{\circ}$.
Then
$$N=N_M({\rm gph}(\partial_pf), (\bar x, 0)),
$$
and for each $h\in{\rm dom}D^2_Mf(\bar x, 0)$, one
has
$$
\sup\big\{\langle h, z\rangle: (h, z)\in T_M({\rm gph}(\partial_pf), (\bar x, 0))\big\}\leq v(h),
$$
where $v(h)$ is defined as in the proof of Proposition 4.3. Using the proof of (ii)$\Rightarrow$(i) in Proposition
4.3, one can prove Theorem 4.2. The proof is completed.
\end{proof}
The following theorem, as one main result in this paper, establishes the equivalence between these second-order optimality conditions for paraconcave and twice epi-differentiable functions in the Hilbert space.
\begin{them}
Let $f:H\rightarrow \mathbb{R}\cup\{+\infty\}$ be a proper lower
semicontinuous and paraconcave function and $\bar x\in{\rm dom}f$ with $0\in \partial_pf(\bar x)$. Suppose that $f$ is continuous at $\bar x$ and is twice epi-differentiable at $\bar x$ for $0$, ${\rm qri}( N_M({\rm gph}(\partial_p f), (\bar x, 0)))\not=\emptyset$ and ${\rm Proj_z}N_M({\rm gph}(\partial_p f), (\bar x, 0))=H$. Then all second-order optimality conditions are equivalent.
\end{them}
\begin{proof} By Theorem 4.1, one can get that the first kind implies the second kind.
Using Proposition 4.1, one has ${\rm dom}D^2_Mf(\bar x, 0)=H$ and thus
$${\rm sqri(dom}D^2_Mf(\bar x, 0))=H.
$$
This implies that \eqref{4.22} holds trivially and it follows from Theorem 4.2 that the second kind implies the third kind.
Since ${\rm dom}\partial_M^2f(\bar x, 0)=-{\rm Proj_z}N_M({\rm gph}(\partial_p f), (\bar x, 0))=H$, then the third kind implies the first kind by virtue of Theorems 4.1. The proof is completed.
\end{proof}
\noindent{\bf Remark 4.3.} For twice epi-differentiable functions, Theorem 4.3 is an extension of Theorem B from the finite-dimensional space to the Hilbert space setting under some mild assumptions. When restricted to the case of finite-dimensional spaces, the quasi-relative interior reduces to the relative interior and thus the assumption ${\rm qri}(N_M({\rm gph}(\partial_p f), (\bar x, 0)))\not=\emptyset$ holds trivially as $N_M({\rm gph}(\partial_p f), (\bar x, 0))$ is convex. With regard to assumption ${\rm Proj_z}N_M({\rm gph}(\partial_p f), (\bar x, 0))=H$, even in finite-dimensional space, very few is known about ${\rm dom}\partial_M^2f(\bar x, 0)$ and the inner estimate for ${\rm dom}\partial_M^2f(\bar x, 0)$ is currently lacking, both making analysis of optimality condition involving ${\rm dom}\partial_M^2f(\bar x, 0)$ difficult. Therefore, it is necessary to add this assumption in the analysis of second-order optimality conditions.
\setcounter{equation}{0}
\section{Applications to strict local minimizers of order two}
In this section, we apply main results on second-order optimality conditions obtained in Section 4 to strict local minimizers of order in the Hilbert space and aim to provide its necessary and/or sufficient conditions. We begin with the definition of strict local minimizer of order two.\\
\noindent{\bf Definition 5.1} {\it
Let $f: H\rightarrow\mathbb{R}\cup\{+\infty\}$ be a proper lower semicontinuous function. We say that $\bar x\in H$ is a strict local minimizer of order two for $f$, if there exist constants $\beta, \delta\in (0, +\infty)$ such that
\begin{equation}\label{5.1}
f(x)\geq f(\bar x)+\frac{\beta}{2}\|x-\bar x\|^2
\end{equation}
holds for all $x\in B(\bar x, \delta)$.}\\
The following theorem is a known and key characterization for strict local minimizers of order two in finite-dimension space. This theorem is established via second-order lower Dini-directional derivative. Readers are invited to consult \cite[Lemma 58]{11} and \cite[Proposition 3.3]{23} for more details.\\
\noindent{\bf Theorem C.} {\it Let $f:\mathbb{R}^m\rightarrow \mathbb{R}\cup\{+\infty\}$ be a proper lower semicontinuous function and $\bar x\in {\rm dom}f$. Assume that the first-order optimality condition $0\in \partial_pf(\bar x)$ holds. Then the following statements are equivalent:
{\rm (i)} $\bar x$ is a strict local minimizer of order two for $f$;
{\rm (ii)} there exists $\beta>0$ such that
\begin{equation}\label{4.1}
f_-''(\bar x, 0, h)\geq \beta\,\,\,\,\forall h\in \mathbb{R}^m\,\,{\it with}\,\,\|h\|=1;
\end{equation}
{\rm (iii)} $
f_-''(\bar x, 0, h)>0$ holds for all $h\in \mathbb{R}^m\,\,{\it with}\,\,\|h\|=1.
$
}\\
Clearly it is shown from Theorem C that the second-order optimality condition of the first kind is necessary and sufficient for strict local minimizer of order two in the finite-dimension space setting. Further, Eberhard and Wenczel \cite{11} provide some conditions under which the second-order optimality conditions of the second and the third kinds are also necessary and sufficient for the existence of strict local minimizers of order two. Naturally, one question arisen here is whether or not the same results as in Theorem C are still valid for the case of Hilbert space. Unfortunately, the following example shows that the answer to this question is negative.\\
\noindent{\bf Example 5.1.} Let $H:=l^2$, and $e_k=(0,\cdots,1,0,\cdots)$ for each natural number $k$. We define a function $f: H\rightarrow \mathbb{R}\cup \{+\infty\}$ as:
$$
f(x)=\left \{
\begin{array}l
\,\,\frac{1}{k^3}\,\,, \,\,x=\frac{1}{k}e_k, k=1,2\cdots;\\
\,\,\,\,\,0\,\,\,,\, x=0;\\
+\infty, \ {\rm otherwise}.
\end{array}
\right.
$$
One can easily verify that $\bar x=0$ is a global minimizer of $f$ and $f_-''(0, 0, h)=+\infty$ for all $h\in S_H$. However, if we take $x_k:=\frac{1}{k}e_k$ for all
$k\in \mathbb{N}$, then $x_k\rightarrow 0$ and
$f(x_k)=\frac{1}{k}\cdot\frac{1}{k^2}$. This implies that
$$
\frac{f(x_k)-f(0)}{\|x_k-0\|^2}=\frac{1}{k}\rightarrow 0.
$$
Hence $\bar x=0$ is not the strict local minimizer of order two for $f$ even though $f$ satisfies the second-order optimality condition of the first kind at $\bar x$. \\
Next, we focus on characterizations for strict local minimizers of order two in a Hilbert space. To this aim, we consider the following notion.\\
\noindent{\bf Definition 5.2} {\it
Let $f:H\rightarrow \mathbb{R}\cup\{+\infty\}$ be a proper lower semicontinuous function, $\beta\in (0, +\infty)$, $p\in \partial_p f(\bar x)$ and let $A$ be a nonempty set of $H$. We say that $f_-''(\bar x, p, h)\geq \beta$ holds uniformly with respect to $h\in A$ if for any $\varepsilon>0$ there exists $\delta>0$ such that
\begin{equation}\label{4.4a}
\frac{f(\bar x+th)-f(\bar x)-t\langle p, h\rangle}{\frac{1}{2}t^2}\geq \beta-\varepsilon
\end{equation}
holds for all $t\in (0, \delta)$ and $h\in A+\delta B_H$.}\\
\begin{pro}
Let $f:H\rightarrow \mathbb{R}\cup\{+\infty\}$ be a proper lower semicontinuous function, $p\in \partial_p f(\bar x)$ and let $A$ be a compact set of $H$. Then the following statements are equivalent:
{\rm (i)} there exists $\beta>0$ such that $f_-''(\bar x, p, h)\geq \beta$ holds uniformly with respect to $h\in A$;
{\rm (ii)} there exists $\beta>0$ such that
\begin{equation}\label{4.5a}
f_-''(\bar x, p, h)\geq \beta\ \ {\it for\ all} \ h\in A;
\end{equation}
{\rm (iii)} $f_-''(\bar x, p, h)>0$ holds for all $h\in A$.
\end{pro}
\begin{proof}
(i)$\Rightarrow$ (ii): The implication follows from \eqref{4.4a} and definition of $f_-''(\bar x, 0, h)$.
(ii)$\Rightarrow$ (i): Let $h\in A$ and $\varepsilon>0$. By \eqref{4.5a}, there exists $\delta_h>0$ such that
\begin{equation}\label{4.6a}
\frac{f(\bar x+th')-f(\bar x)-t\langle p, h'\rangle}{\frac{1}{2}t^2}\geq \beta-\varepsilon\ \ \forall t\in (0, \delta_h)\ {\rm and}\ \forall h'\in B(h, \delta_h).
\end{equation}
Noting that $A$ is compact, there exist $h_1,\cdots, h_n\in A$ such that
\begin{equation}\label{4.7a}
A\subset \bigcup_{i=1}^nB(h_i, \frac{\delta_{h_i}}{2}).
\end{equation}
Let $\delta:=\min\{\frac{\delta_{h_1}}{2},\cdots,\frac{\delta_{h_n}}{2}\}$ and take arbitrary $t\in (0, \delta)$, $h'\in A+\delta B_H$. By virtue of \eqref{4.7a}, there exists $j\in \{1, \cdots, n\}$ such that
$$
h'\in B(h_j, \frac{\delta_{h_j}}{2})+\delta B_H\subset B(h_j, \delta_{h_j}).
$$
This and \eqref{4.6a} imply that
$$
\frac{f(\bar x+th')-f(\bar x)-t\langle p, h'\rangle}{\frac{1}{2}t^2}\geq \beta-\varepsilon.
$$
Thus (ii) holds.
(iii)$\Rightarrow$ (ii): By Proposition 3.1, one has that $f_-''(\bar x, p, \cdot)$ is lower semicontinuous, and there exists $\hat h\in A$ such that
$$
\min_{h\in A}f_-''(\bar x, p, h)=f_-''(\bar x, p, \hat h)>0
$$
(thanks to the compactness of $A$). Then (ii) follows by choosing $\beta:=f_-''(\bar x, p, \hat h)>0$. Since (ii) implies (iii) trivially, the proof is completed.
\end{proof}
The following theorem provides characterizations for strict local minimizer of order two in the Hilbert space setting.
\begin{them}
Let $f: H\rightarrow \mathbb{R}\cup \{+\infty\}$ be a proper lower semicontinuous function and $\bar x\in{\rm dom}f$. Then the following statements are equivalent:
{\rm (i)} $\bar x$ is a strict local minimizer of order two for $f$.
{\rm (ii)} $0\in \partial_pf(\bar x)$ and there exists $\beta>0$ such that $f_-''(\bar x, 0, h)\geq \beta$ holds uniformly with respect to $h\in S_H$.
{\rm (iii)} The following inequality holds:
\begin{equation*}
\liminf_{t\downarrow 0}\Big(\inf_{h\in S_H}\frac{2(f(\bar x+th)-f(\bar x))}{t^2}\Big)>0.
\end{equation*}
\end{them}
\begin{proof} (i)$\Rightarrow$(ii): Since $\bar x$ is a strict local minimizer of order two for $f$, there exist $\beta, \delta>0$ such that
\begin{equation}\label{4a}
f(x)\geq f(\bar x)+\frac{\beta}{2}\|x-\bar x\|^2\ \ \forall x\in B(\bar x, \delta).
\end{equation}
This implies that $0\in\partial_pf(\bar x)$. Let $\varepsilon>0$ and take $\delta_1\in (0, \delta)$ such that $\delta_1<1$, $\delta_1(1+\delta)<\delta$ and $\beta(1-\delta_1)^2>\beta-\varepsilon$. Then for any $t\in(0, \delta_1)$ and $h\in S_H+\delta_1B_H$, one has $1-\delta_1\leq\|h\|\leq 1+\delta_1$ and it follow from \eqref{4a} that
\begin{eqnarray*}
\frac{f(\bar x+th)-f(\bar x)}{\frac{1}{2}t^2}\geq\beta\|h\|^2\geq\beta(1-\delta_1)^2>\beta-\varepsilon.
\end{eqnarray*}
Thus $f_-''(\bar x, 0, h)\geq\beta$ holds uniformly with respect to $h\in S_H$ and (ii) holds.
(ii)$\Rightarrow$(i): Let $\varepsilon\in (0, \beta)$. By Definition 5.2, there exists $\delta>0$ such that
\begin{equation}\label{2ab}
\frac{f(\bar x+th)-f(\bar x)}{\frac{1}{2}t^2}\geq\beta-\varepsilon\ \ \forall t\in(0, \delta)\ {\rm and}\ \forall h\in S_H+\delta B_H.
\end{equation}
Then for any $x\in B(\bar x, \delta)\backslash\{\bar x\}$, by \eqref{2ab}, one has
\begin{eqnarray*}
f(x)-f(\bar x)=f\big(\bar x+\|x-\bar x\|\cdot\frac{x-\bar x}{\|x-\bar x\|}\big)-f(\bar x)\geq\frac{\beta-\varepsilon}{2}\|x-\bar x\|^2.
\end{eqnarray*}
This implies that
$$
f(x)\geq f(\bar x)+\frac{\beta-\varepsilon}{2}\|x-\bar x\|^2\ \ \forall x\in B(\bar x, \delta).
$$
Hence $\bar x$ is a strict local minimizer of order two for $f$.
(i)$\Rightarrow$(iii): Let $\beta, \delta>0$ be such that \eqref{4a} hold. Then
$$
\inf_{h\in S_H}\frac{2(f(\bar x+th)-f(\bar x))}{t^2}\geq\beta \ \ \forall t\in (0, \delta)
$$
and consequently
$$
\liminf_{t\downarrow 0}\Big(\inf_{h\in S_H}\frac{2(f(\bar x+th)-f(\bar x))}{t^2}\Big)\geq\beta>0.
$$
This means that (iii) holds.
(iii)$\Rightarrow$(i): Define
$$
\beta:=\liminf_{t\downarrow 0}\Big(\inf_{h\in S_H}\frac{2(f(\bar x+th)-f(\bar x))}{t^2}\Big)>0
$$
and let $\varepsilon\in (0, \beta)$. Then there exists $\delta>0$ such that
\begin{equation}\label{3a}
\inf_{h\in S_H}\frac{2(f(\bar x+th)-f(\bar x))}{t^2}\geq\beta-\varepsilon\ \ \forall t\in (0, \delta).
\end{equation}
Let $x\in B(\bar x, \delta)\backslash\{\bar x\}$. By \eqref{3a}, one has
\begin{eqnarray*}
f(x)-f(\bar x)=f\big(\bar x+\|x-\bar x\|\cdot\frac{x-\bar x}{\|x-\bar x\|}\big)-f(\bar x)\geq\frac{\beta-\varepsilon}{2}\|x-\bar x\|^2.
\end{eqnarray*}
Hence $\bar x$ is a strict local minimizer of order two for $f$. The proof is complete.
\end{proof}
\noindent{\bf Remark 5.1.} We are now back to the Example 5.1. It is shown that $\bar x=0\in l^2$ is not the strict local minimizer of order two for $f$ appearing in Example 5.1. Further, if we take $t_k=\frac{1}{k}$ for each $k$, by computing, one has
$$
\inf_{h\in S_H}\frac{2(f(\bar x+t_kh)-f(\bar x))}{t_k^2}=\frac{2f(t_ke_k)}{t_k^2}=\frac{2}{k}.
$$
This means that
$$
\liminf_{t\downarrow 0}\Big(\inf_{h\in S_H}\frac{2(f(\bar x+th)-f(\bar x))}{t^2}\Big)\leq \lim_{k\rightarrow \infty}\Big(\inf_{h\in S_H}\frac{2(f(\bar x+t_kh)-f(\bar x))}{t_k^2}\Big)=0.
$$
Using Theorem 5.1, it follows that $\bar x=0\in l^2$ is not the strict local minimizer of order two for $f$.\\
Since the unit sphere of finite-dimensional space is compact, the following corollary is immediate from Proposition 5.1 and Theorem 5.1. This result also shows that Theorem C can be obtained from Theorem 5.1.
\begin{coro}
Let $H$ be a finite-dimensional space, $f:H\rightarrow \mathbb{R}\cup\{+\infty\}$ be a proper lower semicontinuous function and let $\bar x\in {\rm dom}f$. Then the following statements are equivalent:
{\rm (i)} $\bar x$ is a strict local minimizer of order two for $f$;
{\rm (ii)} $0\in \partial_pf(\bar x)$ and there exists $\beta>0$ such that $f_-''(\bar x, 0, h)\geq \beta$ holds uniformly with respect to $h\in S_H$;
{\rm (iii)} $0\in \partial_pf(\bar x)$ and there exists $\beta>0$ such that
$f_-''(\bar x, 0, h)\geq \beta$ holds for all $h\in S_H$;
{\rm (iv)} $0\in \partial_pf(\bar x)$ and $f_-''(\bar x, 0, h)>0$ holds for all $h\in S_H$.
\end{coro}
The following corollary, immediate from Theorem 5.1 and (i)$\Rightarrow$(ii) in Proposition 5.1, shows that the second-order optimality condition of the first kind is necessary for strict local minimizers of order two in the Hilbert space setting.
\begin{coro}
Let $f: H\rightarrow \mathbb{R}\cup\{+\infty\}$ be a lower semicontinuous function
and $\bar x\in{\rm dom}f$. If $\bar x$ is a strict local
minimizer of order two for $f$, then $0\in
\partial_pf(\bar x)$ and $f$ satisfies the second-order optimality condition of the first kind at $\bar x$.
\end{coro}
The following theorem, as one main result of this paper, is obtained from Corollary 5.2 and Theorems 4.1-4.3.
\begin{them}
Let $f:H\rightarrow \mathbb{R}\cup\{+\infty\}$ be a proper lower
semicontinuous and paraconcave function and $\bar x\in{\rm dom}f$ with $0\in \partial_pf(\bar x)$. Suppose that $f$ is continuous at $\bar x$ and twice epi-differentiable at $\bar x$ for $0$, and ${\rm qri}(N_M({\rm gph}(\partial_p f), (\bar x, 0)))\not=\emptyset$. If $\bar x$ is a strict
local minimizer of order two of $f$, then $f$ satisfies all three types of second-order optimality conditions at $\bar x$.
\end{them}
\section{Conclusions}
This paper is devoted to second-order optimality conditions as well as applications in the Hilbert space. Three types of second-order derivatives of nonsmooth functions are considered to discuss these second-order optimality conditions. Their equivalence for paraconcave functions are also proved. As applications, these optimality conditions are used to study the strict local minimizer of order two of nonsmooth functions and provide its necessary and/or sufficient conditions. The work in this paper generalizes and extends the study of second-order optimality conditions from finite-dimensional space to the Hilbert space setting.\\
\noindent{\bf Acknowledgement.} The authors are grateful to Professor Xi Yin Zheng for his helpful suggestions on Definition 5.2 and Theorem 5.1.
|
\section{Introduction}{\label{intro}
Lyman-$\alpha$ blobs (LABs) are extended Ly\,$\alpha$ nebulae with luminosities of
$L_{\rm Ly\,\alpha}=10^{42-44}$\,erg\,s$^{-1}$, populating the Universe at $z\gtrsim2$. They are often
selected using optical narrow-band filters that isolate redshifted Ly\,$\alpha$ emission
\citep[e.g.][]{ktk03,myh04,myh11,dbs05,ooe09,yzt09,yze10}. LABs are $20-200$\,kpc in
size and show a bewildering range of properties. They can be associated with Lyman
Break Galaxies \citep{sph95}, visible and obscured AGN, starburst sub-mm galaxies
and passively evolving red galaxies \citep[e.g.][]{fwc01,csw04,myh04,gal09,wyh09}.
LABs are landmarks of ongoing massive galaxy formation \citep{myh06,pkd08}, yet the
ionizing sources in many of them remain mysterious. Our understanding of these processes
would greatly benefit from studying the physical conditions in LABs. However, this is
difficult as (1) cosmological surface brightness dimming reduces the flux densities by
factors $\gtrsim100$, (2) the Ly\,$\alpha$ line is resonant, and (3) non-resonant optical
lines are redshifted into and beyond the near-infrared atmospheric passbands.
The resonant character of Ly\,$\alpha$ causes two main problems with the interpretation of
Ly\,$\alpha$ data.
First, Ly\,$\alpha$ photons scatter efficiently in space and frequency when propagating
through a moving medium. Three-dimensional radiative transfer calculations
\citep[e.g.][]{mer02,vsm06,kzd10} reveal a great variety of double-peaked Ly\,$\alpha$ line
profiles emerging for various static and kinematic source/halo configurations
\citep[for a one-dimensional analytic description in a static medium see][]{neu90}.
It is difficult at best to infer the gas kinematics and the Ly\,$\alpha$ production sites from
Ly\,$\alpha$ imaging and spectroscopy alone. A multi-wavelength perspective is required,
including optically thin lines such as [\ion{O}{III}] and H\,$\alpha$
\citep[e.g.][]{sso08,wbg10,yzj11,yzj14,mcm14,znf15,svs15}.
The second problem is that we need to understand the processes that govern how many
Ly\,$\alpha$ photons manage to escape, so that they become observable at all. Dust, neutral
hydrogen, metallicity and gas outflows control this escape fraction. The latter can
range from less than 1 per cent to more than 50 per cent \citep[][and references
therein]{ymg15}. Intrinsically, the H\,$\alpha$/Ly\,$\alpha$ line ratio is fixed for a photo-ionized
nebula in equilibrium; the H\,$\alpha$ line can then be used to estimate the escape fraction
and the total amount of Ly\,$\alpha$ produced. However, it is only for a small redshift window
of $z=1.9-2.4$ that \textit{both} lines are observable from the ground. Worse,
AGN variability may change the escape fraction further, by orders of magnitude, due
to delayed Ly\,$\alpha$ escape \citep{rsf10,xwf11}.
Probably the largest mystery with LABs is their frequent lack of ionizing sources; some LABs
show no continuum counterparts at all. The lack of accessible diagnostic lines has prevented
consistent conclusions on many occasions, and various processes have been suggested that
could power LABs. For example, LABs are preferentially found in denser areas and filaments
\citep{sso08,yze10,ebs11,myh11}, where LABs easily accrete cold neutral hydrogen from the
cosmic web \citep{hsq00,dil09,gds10}. This is a requirement by $\Lambda$CDM galaxy formation
models. The Ly\,$\alpha$ emission arises because of collisional excitation of hydrogen (virial
temperature of $10^{4-5}$\,K) when it sinks into the dark matter haloes. It can contribute to
an LAB's ionization over $10-30$ per cent of the Hubble time \citep{dil09}. Some calculations
including self-shielding and realistic gas phases indicate that cold accretion alone could be
insufficient to explain the Ly\,$\alpha$ fluxes of luminous LABs \citep{fkd10}. On the other hand,
\citet{rob12} and \citet{cez13} find cold streams to be rather powerful, reproducing the
size--luminosity function of observed LABs. The challenges in modelling the cold streams are
mirrored on the observational side. One such gravitationally powered LAB \citep{nfm06} is
questioned by \citet{pmb15}, who discovered an embedded obscured AGN and argue that the original
data speak \textit{against} cold accretion.
Alternatively, LABs can be shock-ionized by starburst-driven superwinds \citep{tas00}, and
photoionized by obscured AGN or starbursts \citep[e.g.][]{cls2001,gal09}. Starbursts
alone cannot explain Ly\,$\alpha$ equivalent widths (EW) higher than about 240\,\AA$\;$
\citep{mar02,sso08}; LABs often exceed this value. Another possibility is centrally produced
Ly\,$\alpha$, resonantly scattered by neutral hydrogen in the circum-galactic medium
\citep{las07,sbs11}. This leads to a characteristic polarization signal and can thus be
distinguished from photo-ionization and shock heating which produce Ly\,$\alpha$ \textit{in situ}
\citep{hss11,hvv13}. However, \citet{tvb14} find that similar polarization signals may arise
in cold streams as well.
Evidence for obscured AGN in some LABs has been found in infrared and sub-mm data
\citep{bzs04,gal09,myh11,ond13,pmb15}, and for other LABs they have been postulated. For example,
\citet{myh04} have found 35 LABs, some of which likely powered by superwinds and others by
cooling flows \citep{myh06}. About 30 per cent lack UV continuum counterparts. Associated visible
AGN are uncommon in this sample, and obscured star formation has been ruled out \citep{tmi13}.
\citet{gal09} have shown that 24 out of 29 of these LABs remain undetected in a 400\,ks
\textit{Chandra} exposure even after statistical stacking. They have suggested buried AGN and
star-bursts as power sources instead of cold accretion. This, however, requires particular
combinations of geometrical and radiative transfer effects to explain the substantial escape of
Ly\,$\alpha$, while simultaneously preserving the thick obscuration along the line-of-sight
\citep[see also][]{sas00}. While this certainly holds for some of these LABs, it seems unlikely
to be the case for all of them.
In this paper we investigate the effects of episodic AGN duty cycles (\textit{flickering}) on the
UV, optical and mid-infrared (MIR) properties of LABs, forming optical \textit{ionization echoes}
and MIR \textit{thermal echoes}. Ionization echoes have been reported before, mostly at lower redshift
and in smaller and less luminous nebulae \citep{sev10,kls12,kcb12,sdh13,ssk13,kmb15}. We show that
transient AGN naturally explain the ionization deficits in LABs. Our analysis is based on the
\textit{Green Bean} galaxies \citep[GBs;][hereafter S13]{sdh13} at $z\sim0.3$, hosting luminous
extended emission line regions (EELRs). We show that these EELRs are indeed LABs, and that they also
host recently faded AGN. GBs form the most impressive ionization and thermal echoes currently known.
This paper is structured as follows. In Section 2 we present an overview of the GBs, our
\textit{Chandra} X-ray data, archival \textit{GALEX} data, and our Gemini/GMOS optical
observations. The data are analysed in Sections 3, 4 and 5, respectively. In Section
6 we discuss the evidence for AGN flickering, and its effect on the UV, MIR and optical
properties of LABs. In Section 7 we discuss the LAB size--luminosity function and the evolution
of the LAB comoving density. Our summary and conclusions are presented in Section 8. Details
about individual targets are given in Appendix \ref{targetnotes}. We assume a flat cosmology
with $\Omega_m=0.27$, $\Omega_\Lambda=0.73$ and $H_0=70\,$km\,s$^{-1}$$\,{\rm Mpc}^{-1}$.
\section{\label{sectionobs}Sample selection, observations and data reduction}
\subsection{Identifying Green Bean galaxies}
\subsubsection{\label{GPvsGB}Differences between Green Peas and Green Beans}
Green Peas \citep[GPs,][]{css09} are compact galaxies with strong [\ion{O}{III}] emission lines
that are redshifted into $r$-band. GPs have been discovered in SDSS $gri$ images because
of their green colour. Amongst the 112 spectroscopically confirmed GPs (out of $\sim40000$ known)
are 80 star-forming galaxies with high specific star formation rates, 13 composite objects
revealing both AGN and star formation, 9 Seyfert-1s and 10 Seyfert-2s. Whilst the fraction of star
forming GPs has been studied in detail \citep{apv10,apv12,igt11,pvm12,haw12,jao13,hsm15,ymg15},
the AGN fraction has remained largely unexplored.
Green Beans \citepalias[GBs,][]{sdh13} are much larger and more luminous than GPs. Their spectra
are dominated by narrow lines with high EWs (e.g. $950$\,\AA$\;$ for [\ion{O}{III}] in
J2240$-$0927), and they are (or were) powered by radio-quiet/weak type-2 quasars. Some of the GBs
are probably extreme versions of Seyfert-2 GPs, whereas other GBs have different formation
histories and/or ionization sources. We investigate these aspects in Section \ref{opticalresults}.
Exploring further links between GPs and GBs is beyond the scope of this paper, also because very
few observations exist for the AGN fraction amongst the GPs.
\subsubsection{GB sample selection and completeness}\label{contamination}
GBs are highly unusual, yet they were not identified earlier despite their brightness, size and
colour. This is because (1) GBs are extremely rare, and (2) the SDSS $ugriz$ colour space
occupied by the GBs is contaminated to 95 per cent by artefacts. The 17 GBs known to date were
found by an automatic SQL query mining the SDSS-DR8 photometric data base (14500 deg$^2$). The
query consisted of broad-band colour criteria and a lower size threshold close to the resolution
limit of SDSS. The genuine GBs were isolated from the artefacts by visual inspection of SDSS post
stamp images. Details about the selection, the original SQL filter and the spectroscopic
verification can be found in \citetalias{sdh13}.
Wide and deep imaging surveys with higher resolution than SDSS are ideal to find GBs.
In Fig. \ref{GB_colorspace} we plot the $g-r$ and $r-i$ colours of GBs against those of galaxies in
the Canada-France-Hawaii Telescope Lens Survey \citep[CFHTLenS, 158\,deg$^2$;][]{hek12,ehm13}. In the CFHTLenS
catalogue we retained only bright ($r<20.5$) and large (half flux diameter more than 1\myarcsec1, like GBs)
galaxies. We also rejected objects near bright stars and within filter ghosts (their {\tt MASK} parameter).
The GBs are separated by a large margin from all other galaxies in this colour space. Only one object
(CFHTLenS ID \#W4m0m0\_29728) initially remained in the space occupied by GBs. We removed it as its
$r$-band photometry was falsified by a close pass of minor planet 704 \textit{Interamnia} (mag 11.3)
on 2006 June 07.
One can select GBs using
\begin{equation}
g-r\,>0.8
\end{equation}
and
\begin{equation}
g-r\,>1.5\,(r-i\,)+0.9\,.
\end{equation}
These criteria are based on \textit{SExtractor} \citep{ber06} {\tt MAG\_AUTO} parameters. They do not
change when using aperture magnitudes ({\tt MAG\_APER}) with diameters of 1\myarcsec5
(isolating the nuclei) and 4\myarcsec5 (including most of the EELR flux).
Figure \ref{GB_colorspace} also shows a significant redshift dependence because H\,$\alpha$ moves from
$i$-band into $z$-band for $z\gtrsim0.295$. This decreases $r-i$ and increases $i-z$. A further
discriminator is $u-g$ because of [\ion{O}{II}]$\lambda$3726,29 falling into $g$-band for
$0.06\lesssim z\lesssim0.48$. Our original SDSS selection criteria therefore also included $u$- and
$z$-band photometry \citepalias{sdh13}.
\begin{figure}
\includegraphics[width=1.0\hsize]{GB_colorspace.pdf}
\caption{\label{GB_colorspace}{Gemini/GMOS broad-band colours of GBs compared with bright galaxies
from the 158 square degrees of CFHTLenS.}}
\end{figure}
The GB sample is fairly complete over the SDSS-DR8 footprint and the $z=0.12-0.36$ redshift
range, with three caveats. First, some SDSS data have lower quality leading to an excess
of ``green'' artefacts, seemingly related to poor seeing. Any GBs in these unusable survey areas
were missed. This does not bias the sample as it is simply a matter of slightly lower sky coverage.
Second, due to our lower size threshold, smaller GBs might not be recovered from some areas due to
seeing variations. Most likely there is a smooth transition between large GPs and small GBs. We are
not concerned by this incompleteness as we are interested in the most extended sources, only.
Third, if an EELR coincides with a luminous elliptical galaxy ($M_i\gtrsim-23$ mag), then the
[\ion{O}{III}] EW might not be high enough to distinguish the object and it would be overlooked
(see also Section \ref{densityevolutionclusters}). Lastly, we mention that our search entirely
misses low-z LABs that do not emit in [\ion{O}{III}] (should they exist at $z\sim0.3$). This, however,
does not count against the completeness of our initial goal of identifying strong [\ion{O}{III}]
emitters.
Not all GBs have been discovered yet as SDSS covers a third of the sky, only. We expect
$20-30$ more GBs that are still awaiting their discovery at southern declinations.
\subsubsection{\label{previousdiscoveries}Previous discoveries of Green Bean galaxies}
GBs have a surface density of $1.1\times10^{-3}$\,deg$^{-2}$, falling off the grid of smaller
surveys and random observations. There are three exceptions, though. First, J2240$-$0927
($z=0.326$) was a chance discovery in a CFHT
wide-field data set \citep[program ID: 2008BO01;][]{shk11}. Only because of this
coincidence did we learn about the existence of GBs and initiate our survey. Second,
J0113+0106 ($z=0.281$) was selected automatically for SDSS spectroscopic follow-up to construct
a flux-limited $u$-band sample (SDSS3 target flags {\tt U\_EXTRA2} and {\tt U\_PRIORITY}).
Third, J1155$-$0147 ($z=0.306$; the brightest and largest GB), was picked
up independently by the Quasar Equatorial Survey Team \citep[QUEST;][]{sny98,rma04}. It
is the only GB in the QUEST survey area, a 2.4\,deg wide equatorial strip drift scanned
for emission line objects. \textit{Chandra} images were taken in 2003
(PI: Coppi; \textit{Chandra} Proposal Number 03700891; title: \textit{The X-ray Emission Of High
Luminosity Emission Line Galaxies: Quasar-2s And The Starburst-AGN Connection}). We did not find
any publications of these data, nor about J1155$-$0147 itself.
\begin{table*}
\caption{\label{genprops}
General properties of the GBs and their optical data. Column 1 lists the full names, which we abbreviate
in the main text to the first four digits in RA and Dec. Columns 2 and 3 contain the decimal sky coordinates.
The spectroscopic redshift is given in column 4. Columns 5, 6 and 7 contain the SDSS $r$-band AB magnitude
and the $g-r$ and $r-i$ colours, respectively. Column 8 lists the $r$-band image seeing, translated to a physical
resolution (at the respective source redshift) in column 9.}
\begin{tabular}{lrrccrrcc}
\noalign{\smallskip}
\hline
\noalign{\smallskip}
Name & $\alpha_{2000.0}$ & $\delta_{2000.0}$ & z & $r$ & $g-r$ & $r-i$ & Seeing & Resolution\\
& [deg] & [deg] & & [mag] & [mag] & [mag] & & [kpc]\\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
SDSS J002016.44$-$053126.6 & 5.06852 & -5.52405 & 0.334 & 18.3 & $1.19$ & $-0.17$ & 0\myarcsec59 & 2.8\\
SDSS J002434.90$+$325842.7 & 6.14543 & 32.97852 & 0.293 & 18.2 & $1.25$ & $-0.00$ & 0\myarcsec69 & 3.0\\
SDSS J011133.31$+$225359.1 & 17.88879 & 22.89976 & 0.319 & 19.1 & $1.26$ & $-0.21$ & 0\myarcsec59 & 2.8\\
SDSS J011341.11$+$010608.5 & 18.42129 & 1.10237 & 0.281 & 18.5 & $1.32$ & $0.08$ & 0\myarcsec77 & 3.3\\
SDSS J015930.84$+$270302.2 & 29.87851 & 27.05062 & 0.278 & 18.9 & $1.16$ & $0.11$ & 0\myarcsec57 & 2.4\\
SDSS J115544.59$-$014739.9 & 178.93580 & -1.79443 & 0.306 & 17.9 & $0.91$ & $-0.22$ & 0\myarcsec70 & 3.2\\
SDSS J134709.12$+$545310.9 & 206.78802 & 54.88637 & 0.332 & 18.7 & $1.17$ & $-0.33$ & 0\myarcsec37 & 1.8\\
SDSS J135155.48$+$081608.4 & 207.98117 & 8.26900 & 0.306 & 19.0 & $1.09$ & $-0.19$ & 0\myarcsec71 & 3.2\\
SDSS J144110.95$+$251700.1 & 220.29561 & 25.28337 & 0.192 & 18.5 & $1.09$ & $0.04$ & 0\myarcsec52 & 1.7\\
SDSS J145533.69$+$044643.2 & 223.89036 & 4.77866 & 0.334 & 18.5 & $0.97$ & $-0.03$ & 0\myarcsec55 & 2.7\\
SDSS J150420.68$+$343958.2 & 226.08615 & 34.66618 & 0.294 & 18.7 & $1.35$ & $0.05$ & 0\myarcsec37 & 1.6\\
SDSS J150517.63$+$194444.8 & 226.32347 & 19.74578 & 0.341 & 17.9 & $1.28$ & $-0.78$ & 0\myarcsec52 & 2.5\\
SDSS J205058.08$+$055012.8 & 312.74198 & 5.83688 & 0.301 & 18.6 & $0.99$ & $-0.18$ & 0\myarcsec77 & 3.5\\
SDSS J213542.85$-$031408.8 & 323.92855 & -3.23577 & 0.246 & 19.2 & $0.97$ & $-0.10$ & 0\myarcsec72 & 2.8\\
SDSS J220216.71$+$230903.1 & 330.56961 & 23.15086 & 0.258 & 18.9 & $1.44$ & $0.28$ & 0\myarcsec53 & 2.1\\
SDSS J224024.11$-$092748.1 & 340.10044 & -9.46335 & 0.326 & 18.3 & $1.14$ & $-0.11$ & 0\myarcsec69 & 3.3\\
SDSS J230829.37$+$330310.5 & 347.12239 & 33.05291 & 0.284 & 19.1 & $1.27$ & $0.10$ & 0\myarcsec67 & 2.9\\
\hline
\end{tabular}
\end{table*}
\subsection{\label{GMOSobs}Optical imaging with Gemini/GMOS}
We obtained $gri$ broad-band imaging of all 17 GBs. J2240$-$0927 was discovered earlier in deep
CFHT data (Section \ref{previousdiscoveries}). The remaining GBs
were observed with GMOS-N and GMOS-S at the 8-m Gemini Telescopes in Hawaii and Chile, respectively
(program IDs GS-2013A-Q-48, GN-2014B-Q-78, GN-2015A-DD-3, GN-2015A-FT-23). $20$ minutes
exposure time per filter were sufficient as the targets are bright ($r\sim18$ mag).
The GMOS data were obtained in grey time, under clear and thin cirrus conditions with good seeing.
Several of our targets were re-observed in $griz$ bands at the 4-m Southern Astrophysical Research
Telescope (SOAR, Chile) during the science verification runs of the SOAR Adaptive Module (SAM)
and the SAM Imager \citep[SAMI;][]{tts10,tts12}. SAMI has a 3\myarcmin1 field of view. SAM corrects
for ground layer turbulence using three natural guide stars and one UV laser guide star, achieving
a homogeneous PSF across the field and at optical wavelengths. Unfortunately, the seeing during
these nights was dominated by turbulence in the upper atmosphere and SAM's ground layer correction
could not yield an improvement over the GMOS data. The only exception are the $r$-band data of
J0113+0106 for which the corrected seeing is 0\myarcsec62 while the DIMM seeing was stable between
1\myarcsec0$-$1\myarcsec1; the GMOS image seeing for this object is 0\myarcsec80. We use the SAMI
images to characterize the morphology of J0113+0106.
Image processing was done with \textsc{THELI} \citep{sch13,esd05} using standard procedures.
Photometric zeropoints were tied to SDSS field magnitudes, correcting for non-photometric
conditions. The physical resolution of the coadded images is between $1.6-3.5$\,kpc, depending on
seeing and source redshift. Table \ref{genprops} summarizes the optical characteristics. The fully
calibrated, coadded $gri$ {\tt FITS} images are publicly
available\footnote{{\tt https://zenodo.org/record/56059}}.
\subsection{\label{GMOSobsspec}Optical spectroscopy with Gemini/GMOS}
We conducted a sparse redshift survey around 13 GBs to study their environment, using GMOS
poor weather programs (GN-2015A-Q-99, GS-2015A-Q-99; thin cirrus, seeing $\gtrsim$1\myarcsec2,
bright moon). We used the 1\myarcsec5 long-slit with the B600 grating, tuning the central
wavelength to the 4000\,\AA$\;$ break at the GBs' redshifts. Main redshift
indicators are \ion{Ca}{H+K}, [\ion{O}{II}], [\ion{O}{III}] and the Balmer series. $3\times600$\,s
exposures were used with $4\times4$ detector binning. Target selection was heterogeneous and
incomplete. We aimed at galaxies whose angular diameters, magnitudes and colours suggest similar
redshifts as the GBs. Red sequence galaxies were preferred if present. High priority was given
to galaxies in the immediate vicinity of the GBs, in particular if merger signatures such as tidal
tails, extended haloes, and warps are visible. Position angles were chosen to maximize the number
of objects (up to 7) on the slit. Up to three slit positions were observed per target area, and
a total of 52 redshifts were obtained. Results are presented in Section \ref{longslit} and
Table \ref{redshiftlist}.
\subsection{\label{Lickobsspec}Optical spectroscopy with Lick/Kast}
Four of the GBs from \citetalias{sdh13} were observed with the Kast
double-beam spectrograph at the 3-m Shane telescope of Lick Observatory to determine
their redshifts. A dichroic beamsplitter divided the beams at 4600\,\AA. The blue arm
used a grism setting spanning the $3390-4720$\,\AA\;range with a dispersion of
0.65\,\AA\;pixel$^{-1}$ and a resolution of 3.3\,\AA\;FWHM. The red spectra covered
$5010-7840$\,\AA\;at 2.4\,\AA\;pixel$^{-1}$ and resolution of 6.1\,\AA\;FWHM. The
$2^{\prime\prime}$ slit was oriented on a position angle chosen for each object to maximize
the line flux included (and the angular span for any kinematic information).
Individual 30-minute exposures were obtained on 12 and 13 March 2013 UT. Reduction used the
{\tt IRAF} long-slit tasks. A flux calibration was provided by observation of the standard stars
G191B2B and BD +26 2606 with the same grism and grating settings each night.
Two of the targets, J1347+5453 ($z=0.332$) and J1504+3439 ($z=0.294$),
belong to the sample studied in this paper because (1) their redshifted [\ion{O}{III}]
line falls into the $r$-band, and (2) with log([\ion{O}{III}]/H$\beta$)=1.00 and 0.94
they are also highly ionized as all other GBs. The other two, J1721+6322 and
J1913+6211, are also highly ionized, yet their redshifts of $z=0.544$ and $z=0.552$ are
above our upper redshift cut-off.
\subsection{\label{Chandraobs}Chandra X-ray imaging of 9 GBs}
We selected 9 GBs for follow-up with \textit{Chandra}, adding to the archival data of J1155$-$0147.
The target sample is comprised of GBs with different morphology, [\ion{O}{III}]
line structure, and [\ion{O}{III}] vs. mid-IR excess. The latter criterion was chosen to
include AGN at different stages of the fading process.
We used the aim point on ACIS/S3 for greater soft response and spectral resolution. As our
sources are faint we used the VFAINT mode, and pileup is well below 1 per cent according to
{\tt PIMMS}. The full emission region of each galaxy fit on the single CCD, and on-chip
background measurements were sufficient. No other bright X-ray sources are present in the
fields. The setup for the archival observations of J1155$-$0147 was similar.
\begin{table*}
\caption{\label{targetlist2}
MIR and X-ray properties. We list the names in columns 1, the X-ray fluxes, count rates and
fractional difference hardness ratios ($HR$) in columns $2-4$. Columns 5 and 6 contain the exposure time
and \textit{Chandra} dataset IDs, respectively.}
\begin{tabular}{lrrrrrr}
\noalign{\smallskip}
\hline
\noalign{\smallskip}
Name & $F_{22\mu{\rm m}}$ & $F_{0.3-8\,{\rm keV}}^{X}$ & $R_{\rm obs}$ &
$HR$ & $T_{\rm exp}^{\rm Chandra}$ & \textit{Chandra}\\
& [mJy] & [erg\,s$^{-1}$\,cm$^{-2}$] & [s$^{-1}$] & & [ks] & dataset ID\\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
SDSS J002016.44$-$053126.6 & 11.7 & $2.43\times10^{-14}$ & $2.57\times10^{-3}$ & $0.137\pm0.015$ & 30 & 16100\\
SDSS J002434.90$+$325842.7 & 25.4 & $1.19\times10^{-14}$ & $1.25\times10^{-3}$ & $-0.083\pm0.040$ & 20 & 16101\\
SDSS J011133.31$+$225359.1 & 23.1 & $-$ & $-$ & $-$ & $-$ & $-$ \\
SDSS J011341.11$+$010608.5 & 39.7 & $8.29\times10^{-14}$ & $8.98\times10^{-3}$ & $0.715\pm0.010$ & 15 & 16102\\
SDSS J015930.84$+$270302.2 & 18.1 & $1.95\times10^{-15}$ & $2.20\times10^{-4}$ & low & 30 & 16107\\
SDSS J115544.59$-$014739.9 & 16.9 & $1.18\times10^{-13}$ & $2.01\times10^{-2}$ & $0.351\pm0.002$ & 30 & 3140\\
SDSS J134709.12$+$545310.9 & 5.3 & $-$ & $-$ & $-$ & $-$ & $-$ \\
SDSS J135155.48$+$081608.4 & 25.7 & $-$ & $-$ & $-$ & $-$ & $-$ \\
SDSS J144110.95$+$251700.1 & 19.6 & $1.75\times10^{-14}$ & $1.83\times10^{-3}$ & $0.082\pm0.019$ & 30 & 16108\\
SDSS J145533.69$+$044643.2 & 20.4 & $7.51\times10^{-15}$ & $7.95\times10^{-4}$ & $-0.866\pm0.088$ & 20 & 16103\\
SDSS J150420.68$+$343958.2 & 7.6 & $-$ & $-$ & $-$ & $-$ & $-$ \\
SDSS J150517.63$+$194444.8 & 49.9 & $3.62\times10^{-14}$ & $3.86\times10^{-3}$ & $0.316\pm0.020$ & 15 & 16104\\
SDSS J205058.08$+$055012.8 & 49.5 & $4.70\times10^{-14}$ & $5.00\times10^{-3}$ & $0.719\pm0.016$ & 15 & 16106\\
SDSS J213542.85$-$031408.8 & 3.2 & $-$ & $-$ & $-$ & $-$ & $-$ \\
SDSS J220216.71$+$230903.1 & 24.8 & $-$ & $-$ & $-$ & $-$ & $-$ \\
SDSS J224024.11$-$092748.1 & 37.4 & $9.76\times10^{-15}$ & $1.04\times10^{-3}$ & $-0.077\pm0.067$ & 15 & 16105\\
SDSS J230829.37$+$330310.5 & 13.9 & $-$ & $-$ & $-$ & $-$ & $-$ \\
\hline
\end{tabular}
\end{table*}
Count rates were estimated with the AGN MIR X-ray correlation of
\citet{iut12} using SWIFT/BAT and AKARI. By using the hard $14-195$\,keV band, \citet{iut12}
avoid complications by absorption at softer energies. We used their offset between the
WISE and AKARI bandpasses, and estimated the GBs' intrinsic $14-195$\,keV luminosities from WISE
22\,$\mu$m data. We then inferred the \textit{Chandra}/ACIS-S $0.3-8$\,keV count rate using {\tt PIMMS},
assuming a power law with photon index $\Gamma=1.9$ for the unabsorbed AGN spectrum, and a
column density of $N_{\rm H}=10^{23}$\,cm$^{-2}$. Our choices for $\Gamma$ and $N_{\rm H}$
were based on our analysis of the J1155$-$0147 data (Section \ref{j1155chandra}).
We chose \textit{Chandra}/ACIS exposure times of $15-30$\,ks, aiming at a total of $\sim1000$
counts per target. This would secure a successful spectral analysis for the entire sample.
In the limiting case of Compton thick absorption ($N_{\rm H} = 1.5\times10^{24}$\,cm$^{-2}$),
without reflection the continuum count rate would be $\sim0.002$\,s$^{-1}$. In these instances
we would still detect an absorbed AGN in the strong Fe K$\alpha$ line. Note that even a low
column density of $N_{\rm H} = 10^{22}$\,cm$^{-2}$\; is sufficient to account for the non-detection
of all targets by ROSAT (which is sensitive at soft energies only).
Observations were carried out in \textit{Chandra} cycle 15, and the event files were processed in
CIAO following standard procedures. We corrected the WCS of the final X-ray maps by about half a
\textit{Chandra} pixel. The offset was calculated from the mean displacement observed between other
X-ray detected AGN in the field and their counterparts in the optical GMOS images. The GBs were
excluded from this calculation to avoid biasing by any true offset of their AGN with respect to
the peak of the optical emission. Our X-ray measurements are summarized in Table \ref{targetlist2}.
\begin{figure}
\includegraphics[width=1.0\hsize]{galex_effarea.pdf}
\caption{\label{galex_effarea}{The \textit{GALEX} FUV effective area, reproduced from the
\textit{GALEX} online documentation. The GBs' redshifted Ly\,$\alpha$ wavelengths fall into the
region of highest sensitivity.}}
\end{figure}
\subsection{\label{galexobs}GALEX observations}
\textit{GALEX} data are perfectly suited to detect redshifted Ly\,$\alpha$ from the GBs in the FUV
channel ($1340-1800$\,\AA; Fig. \ref{galex_effarea}). In the course of the various \textit{GALEX} surveys,
5 GBs were observed in the FUV with exposure times of $1200-2800$\,s, and 10 GBs with exposure
times of $60-300$\,s. Spectroscopic data were not taken. 14 out of 15 GBs are detected in the FUV,
with ${\rm S/N}=2-21$. For two GBs no FUV data are available, but they are visible in the NUV
($1750-2800$\,\AA). Only J0111+2253 is not detected in the FUV nor the NUV.
Flux measurements were taken from the \textit{GALEX} DR6 catalogue query page.
If multiple measurements of the same source were available, then we used the one with
the longest exposure time. The only exception is J1455$+$0446, which is marginally blended
in the \textit{GALEX} data with a large foreground galaxy and not available as a separate catalogue entry.
We downloaded the calibrated FUV and NUV images and measured the fluxes in a 10\myarcsec5 wide
circular aperture, cleanly separating J1455$+$0446 from its neighbour. The background signal and
measurement errors were estimated by placing the same aperture at 10 randomly chosen
nearby blank positions. The FUV and NUV spectral flux densities, exposure times,
galactic reddening and estimated Ly\,$\alpha$ luminosities are listed in Table \ref{targetUVprop}.
\section{Analysis of the X-ray data}
\subsection{Absence of kpc-scale AGN binaries}
Many GBs are interacting and/or merging (Sects. \ref{mergerrates} and
\ref{morphologies}), and could perhaps host binary AGN. Mergers boost the accretion rates of
supermassive black holes (SMBHs) by funnelling more gas toward the centres. This also holds
for binary AGN as shown by \citet{lss12}, who find that the log([\ion{O}{III}]) luminosity
increases by $0.7\pm0.1$ in AGN binaries when their separation decreases from 100 to 5\,kpc.
Therefore, the GBs' high [\ion{O}{III}] luminosities make binary AGN at least plausible. The
fraction of binaries with separations of tens of kpc amongst optically selected AGN is small
\citep[3.6 per cent;][]{lss11}, yet it could be enhanced in GBs. The GBs' complex line profiles
\citetext{\citetalias{sdh13}, \citealp{dst15}}, though, are much more likely caused by gas
kinematics \citep[e.g.][]{slg11,cgs12,cog14,ass15}.
We find that the nuclei in GBs are X-ray point sources. If binary AGN are present, then their
separations must be smaller than $1.7-1.9$\,kpc (about one \textit{Chandra} ACIS pixel),
and/or the secondary AGN is below our detection limit. Nonetheless, the advanced merger states
make it worthwhile to search for sub-kpc binaries at other wavelengths.
\subsection{Offsets between X-ray and [OIII] peaks}{\label{xrayoffsets}}
The positions of the X-ray peaks are fully consistent with the positions of the optical
peaks in the $r$-band images, i.e. the location of highest [\ion{O}{III}] brightness. The only
exception is J1505+1944, where the X-ray peak is offset by 0\myarcsec5 (2.4\,kpc) to the
West. Interestingly, the [\ion{O}{III}] nebula in J1505+1944 fragments in East-West direction.
The brightest [\ion{O}{III}] part could be powered by a shock or be part of an outflow.
Alternatively, it could harbour a second SMBH that is either deeply buried, or faded from our
view recently while its ionizing radiation is still propagating outwards; dynamic data are not
yet available for this system.
\subsection{Diffuse X-ray emission}
The X-ray contours are extended for 60 per cent of the targets
(J0024, J0159, J1155, J1455, J1055, J2240). While this is weakly significant for most
targets individually (caused by just $1-3$ extra counts), in all cases the extended
X-ray flux traces the most luminous parts of the [\ion{O}{III}] gas. We think this is
caused by photoionized emission from the gas. For the remaining 40 per cent, any diffuse
emission is below our detection threshold.
\subsection{\label{fadingagn}Compton-thick or intrinsically weak?}
Our observations yielded much lower count rates than anticipated, to the point where spectral
fitting became meaningless ($7-80$ total counts). Therefore, we did not obtain power law
indices, column densities and model fluxes apart from J1155$-$0147.
Figure \ref{ichikawa12} shows the MIR X-ray relation of \citet{iut12}. Based on the observed
\textit{Chandra} $0.3-8.0$\,keV count rates and a power law index of $\Gamma=1.9$, we calculate
the expected X-ray fluxes for four different intrinsic column densities,
${\rm log}\,(N_{\rm H}\,\rm{cm}^{2})=22.0$, 23.0, 24.0 and 24.3; error bars account for an
uncertainty of 0.1 in $\Gamma$. The result can be
interpreted in two ways: Either, most GBs are nearly or fully Compton-thick (Section \ref{argcthick}),
or they have faded recently and quickly, quicker than the typical response time of the
dusty tori's MIR emission (Section \ref{argcfading}).
\begin{figure}
\includegraphics[width=1.0\hsize]{ichikawa12.pdf}
\caption{\label{ichikawa12}{AGN thermal echoes. The black dots show the
AGN MIR X-ray relation studied by \citet{iut12}, the shaded area
represents its intrinsic scattering. Overlaid are the expected
X-ray luminosities of 10 GBs, based on their observed \textit{Chandra}
$0.3-8.0$\,keV fluxes, a power law and 4 different values of intrinsic
absorption. Horizontal error bars are smaller than the symbol size, vertical
error bars include the \textit{Chandra} measurement error and a fiducial uncertainty
of $\Delta\Gamma=0.1$ for the photon index. The GBs follow the observed MIR X-ray
relation only if Compton-thick; since we can exclude Compton-thick obscuration
for the sample, the GBs' X-ray fluxes must have faded quicker than their MIR
fluxes, causing thermal echoes. The typical MIR response time-scale to
X-ray fading is $10-1000$ years.}}
\end{figure}
\subsubsection{\label{argcthick}GBs cannot be Compton-thick as a sample}
The low count rates could be explained if, on average, GBs are obscured with
log$\,(N_{\rm H}\,\rm{cm}^{2})=24.1$, implying a Compton-thick fraction $f_{\rm CTK}\sim0.9$.
The latter is unusually high; for comparison, \citet{rms99}, \citet{gmp05} and \cite{msb09}
find $f_{\rm CTK}=0.4-0.5$ for optically and X-ray selected Seyfert-2s at $z\leq0.035$,
and \citet{bdm99} report $f_{\rm CTK}=0.23-0.30$.
It is well-known that the fraction of absorbed AGN with log$\,(N_{\rm H}\,\rm{cm}^{2})>22$ decreases with
increasing X-ray luminosity \citep{has08,uah14,mbb14}. How luminous are the AGN in GBs, and what
fraction of Compton-thick sources should we expect? Compared to the type-2 samples of \citet{rzs08}
and \citet{maf13}, GBs rank amongst the most [\ion{O}{III}] luminous type-2 AGN known, and should
harbour AGN of high bolometric luminosity \citep{bdm99,hkb04,lbm09}. For e.g. J2240$-$0927,
\citetalias{sdh13} measure an extinction corrected $L_{[\ion{O}{III}]}=(5.7\pm0.9)\times10^{43}$\,erg\,s$^{-1}$,
translating to $L_{\rm bol}\sim2.3\times10^{46}$\,erg\,s$^{-1}$\; following \citet{lbm09}. We should therefore
expect low values for $f_{\rm CTK}$.
Just how low can be estimated from \citet{uah14} and their fig. 13, showing the fractions of
moderately absorbed AGN (log$\,(N_{\rm H}\,\rm{cm}^{2})=22-24$) and Compton-thick AGN
(log$\,(N_{\rm H}\,\rm{cm}^{2})=24-26$).
Amongst type-2 AGN, $f_{\rm CTK}=0.3$ and 0.05 for the lower and higher X-ray luminous sources,
respectively. Both values are in strong disagreement with $f_{\rm CTK}=0.9$ for the GBs.
However, these statistical arguments alone are insufficient to reject the hypothesis that
nearly all GBs are Compton-thick. After all, GBs were discovered only recently and have not
been studied before. The selection function of GBs (essentially, $r$-band excess caused by
[\ion{O}{III}]) favours the selection of optically absorbed AGN: if unabsorbed type-1 AGN were
present amongst the GBs, then their continuum contribution to the broad-band photometry would
reduce the $r$-band excess and they would not be selected. Therefore, some obscuration amongst
GBs is expected, but they do not have to be exclusively Compton-thick.
\subsubsection{\label{argcfading}GBs are intrinsically X-ray weak}
If GBs were indeed Compton-thick, then we would still detect the fluorescent K$\alpha$ line
\citep{krk87}. However, this line is largely absent in our sample, favouring intrinsically weak
AGN over heavy obscuration. The only GB for which we detect K$\alpha$ is J1155$-$0147, which is
sufficiently bright to allow for spectral modelling. This is a moderately obscured source
(Section \ref{j1155chandra}).
Another indicator for intrinsically weak AGN comes from the fractional difference hardness ratio,
\begin{equation}
HR = \frac{H-S}{H+S}\;.
\end{equation}
Here, $S$ and $H$ are the counts in the soft ($0.3-2.0$\,keV) and hard ($2.0-8.0$\,keV) bands,
respectively. We observe a moderate sample mean of $\langle HR\,\rangle=0.14\pm0.48$ (Table
\ref{targetlist2}), meaning that the AGN cannot be deeply buried as a group.
\begin{table*}
\caption{\label{targetUVprop}
UV properties of the GBs, including the observed monochromatic FUV and NUV spectral flux
densities (uncorrected for galactic extinction), the colour excess for the respective
line of sight, an extinction-corrected estimate of the Ly\,$\alpha$ luminosity assuming no continuum,
and the \textit{GALEX} integration times. See text for details.}
\begin{tabular}{lrrrrrr}
\noalign{\smallskip}
\hline
\noalign{\smallskip}
Name & $f_{\nu}^{\;\rm FUV}$ & $f_{\nu}^{\;\rm NUV}$ & $E(B-V)$ & $L_{\rm Ly\alpha}$ & $T_{\rm exp}^{\;\rm FUV}$ & $T_{\rm exp}^{\;\rm NUV}$ \\
& [$\mu$Jy] & [$\mu$Jy] & [mag] & [$10^{43}$\,erg\,s$^{-1}$] & [s] & [s]\\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
SDSS J002016.44$-$053126.6 & $ 9.5\pm2.3$ & $ 9.0\pm2.4$ & 0.030 & $2.42\pm0.59$ & 206 & 206 \\
SDSS J002434.90$+$325842.7 & $ 5.9\pm2.2$ & $15.1\pm1.8$ & 0.051 & $1.12\pm0.42$ & 247 & 501 \\
SDSS J011133.31$+$225359.1 & undetected & undetected & 0.034 & $\lesssim0.31$ & 110 & 110 \\
SDSS J011341.11$+$010608.5 & $15.86\pm0.50$ & $15.44\pm0.50$ & 0.028 & $2.29\pm0.07$ & 2743 & 7999 \\
SDSS J015930.84$+$270302.2 & $ 7.1\pm2.4$ & $ 9.9\pm2.2$ & 0.056 & $1.23\pm0.42$ & 186 & 186 \\
SDSS J115544.59$-$014739.9 & $40.8\pm1.9$ & $41.2\pm1.1$ & 0.019 & $6.64\pm0.31$ & 2768 & 2768 \\
SDSS J134709.12$+$545310.9 & $31.0\pm5.3$ & $14.3\pm2.4$ & 0.010 & $6.47\pm1.11$ & 190 & 190 \\
SDSS J135155.48$+$081608.4 & $18.1\pm5.2$ & $ 7.6\pm2.5$ & 0.020 & $2.97\pm0.86$ & 106 & 106 \\
SDSS J144110.95$+$251700.1 & $29.5\pm8.5$ & $16.6\pm0.9$ & 0.023 & $1.29\pm0.37$ & 61 & 1690 \\
SDSS J145533.69$+$044643.2 & $19.8\pm1.6$ & $ 7.4\pm1.5$ & 0.033 & $5.14\pm0.42$ & 1650 & 1650 \\
SDSS J150420.68$+$343958.2 & $ 2.7\pm1.1$ & $10.1\pm0.8$ & 0.012 & $0.38\pm0.16$ & 306 & 2275 \\
SDSS J150517.63$+$194444.8 & $22.7\pm3.5$ & $25.4\pm2.6$ & 0.033 & $6.84\pm1.06$ & 234 & 234 \\
SDSS J205058.08$+$055012.8 & $ 6.6\pm2.5$ & $12.2\pm1.3$ & 0.088 & $1.77\pm0.68$ & 169 & 1616 \\
SDSS J213542.85$-$031408.8 & $18.0\pm1.3$ & $15.9\pm1.3$ & 0.033 & $1.58\pm0.11$ & 1561 & 1561 \\
SDSS J220216.71$+$230903.1 & no data & $11.4\pm2.8$ & 0.072 & $-$ & $-$ & 173 \\
SDSS J224024.11$-$092748.1 & $21.2\pm1.6$ & $14.3\pm1.0$ & 0.052 & $5.43\pm0.41$ & 1578 & 1578 \\
SDSS J230829.37$+$330310.5 & no data & $ 5.7\pm2.5$ & 0.073 & $-$ & $-$ & 158 \\
\hline
\end{tabular}
\end{table*}
\section{Analysis of the GALEX data}\label{GALEXresults}
In this Section we estimate the Ly\,$\alpha$ luminosities of the GBs using \textit{GALEX} FUV and NUV broad-band
imaging data (Table \ref{targetUVprop}). In the absence of \textit{GALEX} spectra, we must estimate continuum
contributions to the FUV, which could be mistaken for Ly\,$\alpha$ emission. We do not have sufficient ancillary
data available to perform this for all GBs in our sample. Nonetheless, in four cases we can do this,
and we show that continuum emission contributes at most a few 10 per cent to the FUV flux. As the
properties of the GBs are similar, we argue that our conclusions hold for the sample as a whole.
If the continuum contribution was 25 per cent, then 85 per cent (53 per cent) of the GBs have
Ly\,$\alpha$ luminosities in excess of $1\times10^{43}$ ($2\times10^{43}$)\,erg\,s$^{-1}$\;with Ly\,$\alpha$ EWs
of up to $1000$\,\AA. We conclude that we have indeed found LABs at low redshift,
17 years after their initial discovery at high redshift (see also Fig. \ref{labstats}).
\subsection{\label{estlya}Estimating the Ly\,$\alpha$ luminosities}
We correct the FUV spectral flux densities for galactic extinction using the \citet{scf11}
tables and a $R_v=3.1$ dust model. The correction factors range between 1.08 (J1347+5453) and
1.97 (J2050+0550), and are calculated for the redshifted Ly\,$\alpha$ wavelengths assuming that most
of the FUV flux is caused by this line. Other bright lines such as CIV$\lambda$1549 are redshifted
beyond the \textit{GALEX} FUV bandpass even for the lowest redshift in our sample ($z=0.192$, J1441+2517).
We must account for the relative response function of \textit{GALEX} when estimating the total
Ly\,$\alpha$ flux from the FUV broad-band data. The bandpass-averaged observed monochromatic spectral
flux density, $f_\nu^{\;\rm obs}$, is calculated from the redshifted source spectrum, $f_\nu(\nu)$,
as
\begin{equation}
\label{fnuobs}
f_\nu^{\;\rm obs} = \frac{\int{f_\nu(\nu)\,T(\nu)\,{\rm d}\nu}}{\int{T(\nu)\,{\rm d}\nu}}\;.
\end{equation}
Here, $T(\nu)$ is the unnormalized relative system throughput which we interpolate from the
FUV effective area (Fig. \ref{galex_effarea}).
We approximate the spectrum as the sum of a constant continuum and some line profile. The
continuum is parametrized as a fraction $c$ of the observed spectral flux density,
$f_\nu^{\;\rm obs}$, and the spectrum is written as
\begin{equation}
f_\nu(\nu) = c\,f_\nu^{\;\rm obs}\,+\,f_\nu^{\;\rm line}(\nu)\,.
\end{equation}
Insert this into equation (\ref{fnuobs}) and we have
\begin{equation}
(1-c)\,f_\nu^{\;\rm obs} = \frac{\int{f_\nu^{\;\rm line}(\nu)\,T(\nu)\,{\rm d}\nu}}{\int{T(\nu)\,{\rm d}\nu}}\;.
\end{equation}
The Ly\,$\alpha$ line width is just a few \AA$\;$ even for a velocity dispersion of 1000\,km\,s$^{-1}$.
The \textit{GALEX} FUV response can be considered constant over such small a wavelength range.
We model the line profile as a Dirac delta function, normalized to yield the total line flux
density, $F_{{\rm Ly}\alpha}$, when integrated over frequency:
\begin{equation}
(1-c)\,f_\nu^{\;\rm obs} = \frac{\int{F_{{\rm Ly}\alpha}\;\delta_{\rm D}(\nu-\nu_{{\rm Ly}\alpha})\,T(\nu)\,{\rm d}\nu}}{\int{T(\nu)\,{\rm d}\nu}}=
F_{{\rm Ly}\alpha}\;\frac{T(\nu_{{\rm Ly}\alpha})}{\int{T(\nu)\,{\rm d}\nu}}\;.
\end{equation}
Here, $\nu_{{\rm Ly}\alpha}$ is the frequency of the redshifted Ly\,$\alpha$ line. We solve for $F_{{\rm Ly}\alpha}$
and derive the Ly\,$\alpha$ luminosity using the luminosity distance. In Table \ref{targetUVprop} we list the
Ly\,$\alpha$ luminosities assuming no continuum ($c=0$). If a fraction $c$ of the FUV flux is in
the continuum, then the true line flux will be $(1-c)$ times the tabulated value.
Possible continuum sources are stars, the nebular continuum, and scattered AGN light:
\begin{equation}
c=c^{\,\rm stellar}+c^{\,\rm nebular}+c^{\,\rm scatter}\,.
\end{equation}
We discuss each of these terms below.
\subsection{Young and old stars must be considered}
Young hot stars contribute to the UV continuum. GBs are gas rich and often found in
mergers (Section \ref{mergerrates}), a combination known to boost star formation. AGN feedback
may also trigger star formation by shock-inducing cloud collapse \citep[e.g.][]{sdd13,sil13}.
However, our images also bear evidence for strong AGN-driven outflows, which may quench star
formation by removing gas \citep[e.g.][]{gdg14}. High values of log([\ion{O}{III}]/H\,$\beta)\sim1$
show that star formation plays a minor role at least for the optical line emission \citepalias{sdh13}.
Old stars with high surface temperatures also contribute to the UV continuum. This includes
binary stars \citep{hpl07}, low-mass helium-burning stars in the horizontal branch
\citep[e.g.][]{cyl11,rjo12} and evolved post-AGB stars \citep[e.g.][]{cog10}. These types are
thought to cause the UV excess (UVX) observed in elliptical galaxies, in particular bluewards of
2000\,\AA$\;$ \citep[UV upturn; for a review see][]{crw99}.
Which stellar populations are present in GBs? The red colours of the host galaxies (e.g. J1347+5453,
J1504+3439) and of tidal stellar debris (J0024+3258, J0111+2253) are consistent with the prevalence
of older stars. In most other cases the host galaxies are too compact to determine reliable colours
in the presence of the nebular emission. Four GBs (J1155$-$0147, J1505+1944, J2050+0550, J2202+2309)
are in groups or clusters with masses of at most a few $10^{13}\;M_{\odot}$ (e.g. Section
\ref{j11550147target}). Dynamical friction \citep{cha43,nus99} and merging is efficient in such
low velocity environments, as witnessed by the presence of red sequence galaxies. It is plausible
that these four GBs are also red sequence galaxies as they share the same environment with
their neighbours.
We have to assume that both young and old stars contribute to the FUV continuum. We can estimate this
for J2240$-$0927, for which we have useful spectra (Section \ref{starlight}), and for three other GBs
where sufficient red sequence galaxies and \textit{GALEX} data are available (Section \ref{fuvJ1155}).
\begin{figure*}
\includegraphics[width=1.0\hsize]{j2241-0927_starlight.pdf}
\caption{\label{j2241-0927_starlight}{Continuum restframe VLT/XSHOOTER spectrum of J2240$-$0927
(black, emission lines and nebular continuum removed). Overplotted are three representative
composite SSP models, used to predict the luminosities in the (blueshifted) \textit{GALEX} FUV
and NUV bandpasses (shaded areas). Comparison with an old stellar base population (red line) reveals
excess flux below 3800\,\AA, requiring additional young stars. The SSP models are divergent below
2500\,\AA$\;$for which we do not have observational data. Also shown is the nebular continuum
for three electron temperatures and $n_{\rm e}=100$\,cm$^{-3}$. The XSHOOTER spectrum was centred on
the nucleus, and the extraction area is shown in the lower right inset. The latter displays the
reconstructed 2D images (nonlinear stretch) from the GMOS-S data cube, near and on the H$\beta$
line, and the H$\beta$ equivalent width.}}
\end{figure*}
\subsection{\label{starlight}Constraining the stellar FUV/NUV flux for J2240$-$0927 using SED fitting}
\citet{dst15} have presented a 3D spectroscopic study of J2240$-$0927. The continuum maps reveal a
compact spherical galaxy with $\lesssim9$\,kpc diameter. Our VLT/XSHOOTER spectrum \citepalias{sdh13}
of the nucleus covers the $2500-18800$\,\AA$\;$ restframe range. It is corrected for galactic
extinction using a $R_V=3.1$ dust model and the extinction maps of \citet{scf11}, de-redshifted,
and we subtract the nebular continuum for a 15000\,K hydrogen helium gas mix in ionization equilibrium
(see also Section \ref{nebcontinuum}). We then fit a combination of simple stellar populations (SSPs)
with the {\textsc STARLIGHT} code \citep{msc06}, and determine the FUV and NUV fluxes permitted by the
models.
We use the evolutionary models of \citet{brc03}, the SSPs of \citet{mar05}, which include the effects
of thermally pulsating AGB stars (TP-AGB), \citet{mnb09}, the binary SSPs of \citet{hpl07}, and
lastly the SSPs of \citet{cog10}, which better describe the UV properties of the horizontal branch
and post-AGB stars. A comparison can be found in \citet{cog10}.
We run {\textsc STARLIGHT} on a grid with 90 different configurations (SSPs, dust and extinction
models, hard and soft convergence criteria, wavelength ranges). Figure \ref{j2241-0927_starlight}
shows the spectrum together with a choice of three composite SSP models, displaying the
full range of predicted FUV/NUV fluxes including extreme cases. We summarize our findings:
First, none of the fits is superior. The continuum levels and slopes are reproduced
by all fits for $\lambda\gtrsim2800$\,\AA, whereas absorption features (CaH+K, G-band, Mg,
NaD, \ion{Ca}{II} triplet, etc) are recovered with lesser accuracy. Inclusion of different dust
models and extinction laws do not change the fits significantly.
Second, we observe absorption bands near 8500\,\AA$\;$ and 9500\,\AA, attributed to
TP-AGB stars, CN and CO bands, and they are better described by the MA05 models. All fits reproduce
the spectrum well between $13000-19000$\,\AA$\;$ (not shown in Fig. \ref{j2241-0927_starlight}).
Third, a reference model consisting of old populations with ages $\geq1$ Gyr only (red line),
reveals excess flux below 3800\,\AA. Consequently, all models require
the presence of a younger population with age of a few to a few ten Myrs. Depending on the
{\textsc STARLIGHT} setup, $10-30$ per cent of the bolometric luminosity are caused by young
stars. This contribution drops to 2 per cent when we exclude wavelengths shorter than
3800\,\AA$\;$ from the fit.
Fourth, all models diverge below 2500\,\AA. This is caused by the strong metallicity and age
dependence of the UV upturn which is unconstrained by our data.
Finally, we shift the various composite SEDs to the redshift of J2240$-$0927, add back the
reddening, and calculate bandpass-averaged FUV/NUV spectral flux densities for comparison with
the \textit{GALEX} data. We apply aperture correction factors, as \textit{GALEX} integrated the entire
galaxy light, whereas XSHOOTER observed through a 0\myarcsec9 wide slit. To this end we use the 3D
GMOS-S spectroscopic cube of \citet{dst15}, which covers the full spatial extent of J2240$-$0927. We
integrate the light over the reconstructed IFU image, once over the full field, and once within the XSHOOTER
aperture. Seeing corrections are unnecessary because both data sets were taken with a seeing of
0\myarcsec5$-$0\myarcsec6. For the stellar continuum (taken near H$\beta$) and the nebular emission (taken
on H$\beta$) we determine aperture correction factors of $2.10\pm0.04$ and $2.65\pm0.05$, respectively.
The nebular correction factor is larger because the continuum light is much more concentrated (see the
inset in Fig. \ref{j2241-0927_starlight}).
\subsubsection{Results of the SED fitting}
The observed \textit{GALEX} FUV flux density of J2240$-$0927 is $21.2\pm1.6$\,$\mu$Jy (Table \ref{targetUVprop}).
We find the stellar model fluxes to range between $2.7-12.1$\,$\mu$Jy (most extreme values, $13-57$ per cent
contribution). Deeper observations below restframe wavelengths of 2500\,\AA$\;$(observed 3400\,\AA) are
required to better discriminate between the models. The flux calibration and S/N of our XSHOOTER observations
(${\rm S/N}\sim1$ after $32\times$ spectral binning) are too poor for this purpose. Given these data alone,
the most plausible contribution is $\sim5-8$\,$\mu$Jy, i.e.
\begin{equation}
c^{\,{\rm stellar}} = 0.2-0.4\;\; ({\rm J2240-0927}).
\end{equation}
For the NUV, we find stellar model flux densities between $10.1-16.7$\,$\mu$Jy, compared to an observed
value of $14.28\pm0.95$\,$\mu$Jy. Consequently, the model that produces the lowest FUV contribution of
13 per cent accounts for 70 per cent of the observed NUV flux, ruling out models that contribute
more than about $20$ per cent to the FUV. Including this constraint from the NUV data, we update
\begin{equation}
c^{\,{\rm stellar}} \lesssim 0.1-0.2\;\; ({\rm J2240-0927}).
\end{equation}
\subsection{\label{fuvJ1155}Constraining the stellar FUV flux in 3 GBs from red sequence galaxies}
J1155$-$0147, J1505$+$1944 and J2050$+$0550 are in spectroscopically confirmed galaxy groups with
a red sequence. We derive mean stellar FUV to $i$-band flux ratios,
$\langle f_\nu^{\,\rm FUV} / f_\nu^{\,i}\rangle$, for the red sequence members. Assuming that the GBs
are also red sequence members, we use their $i$-band magnitude to estimate their stellar FUV
flux (note that H\,$\alpha$ is redshifted beyond $i$-band in all three cases).
We place apertures over the red sequence members in the $i$-band image and measure their $i$-band
spectral flux density. The apertures are then transformed to the \textit{GALEX} FUV image accounting
for the larger plate scale and PSF, and the measurement is repeated. The red sequence members are not
detected individually by \textit{GALEX}. Integrating over all apertures, we derive
$\langle f_\nu^{\,\rm FUV} / f_\nu^{\,i}\rangle = (5.0\pm6.5)\times10^{-3}$, $(-2.7\pm2.0)\times10^{-3}$,
and $(1.07\pm0.62)\times10^{-2}$, respectively, for these three systems. This describes the total FUV
contribution from young and old stars. Comparison with the observed FUV flux densities yields
\begin{eqnarray}
c^{\,{\rm stellar}} & = & 0.027\pm0.035\;\; ({\rm J1155-0147})\\
c^{\,{\rm stellar}} & < & 0.018\;\; ({\rm J1505+1944;\;95\,\%\;confidence)}\\
c^{\,{\rm stellar}} & = & 0.11\pm0.07\;\; ({\rm J2050+0550}).
\end{eqnarray}
The value for J1505$+$1944 is an upper limit. These contributions are lower than or equal to
what we have found for J2240$-$0927 using SED fitting. All GB host galaxies have similar
$i$-band magnitudes ($18.8\pm0.4$ mag). Stars, therefore, cannot explain their high FUV fluxes
(accounting for a few per cent, at most a few 10 per cent of the FUV flux).
\subsection{Nebular continuum for J2240-0927}\label{nebcontinuum}
The nebular continuum also contributes to the UV. We model it using our custom-made software
{\tt NEBULAR} (Schirmer 2016, submitted), which is publicly
available\footnote{{\tt https://zenodo.org/record/55843}}.
In particular, we use a mixed hydrogen helium plasma in ionization equilibrium, with a helium
abundance (by parts) of 0.1.
The continuum of the nebular spectrum is comprised of free-bound recombination emission from
\ion{H}{I}, \ion{He}{I} and \ion{He}{II}, free-free emission from electrons scattering at charged ions,
and the two-photon continuum.
The two-photon continuum far exceeds the free-bound emissivity below 2000\,\AA. It arises in
hydrogenic ions from the decay of the 2$^2S$ level to the 1$^2S$ level by simultaneous emission
of two photons (a single photon decay is prohibited by the dipole selection rules). The energy of
the two photons adds up to the Ly\,$\alpha$ energy. The two-photon spectrum has a natural upper cut-off at
the Ly$\alpha$ frequency, and peaks at half the Ly$\alpha$ frequency when expressed in frequency units.
We approximate the two-photon spectrum following \citet{nus84}. The 2$^2S$ level is
increasingly de-populated by collisions for electron densities $n_{\rm e}>1000$\,cm$^{-3}$ \citep{pes64},
reducing the two-photon continuum. This process can be ignored in the low-density gas
\citep{dst15} of the GBs.
At optical wavelengths, the nebular continuum is faint and dominated by the stellar continuum.
Fotunately, its amplitude is fixed with respect to the intensity of the Balmer lines at a
given electron temperature and density. Using {\tt NEBULAR}, we derive H$\beta$ equivalent
widths of 1370, 770 and 650\,\AA$\;$over the nebular continuum, for electron temperatures of
10000, 15000 and 20000\,K, respectively (and $n_{\rm e}=100$\,cm$^{-3}$). Using the total observed
H$\beta$ flux from our GMOS-S data
cube ($7.5\times10^{-15}$\,erg\,s$^{-1}$), we find the following: In the FUV, the nebular continuum
contributes 0.5, 1.6 and 2.1\,$\mu$Jy for $T_{\rm e}=10000$, 15000 and 20000\,K, respectively, i.e.
2, 8 and 10 per cent of the observed total FUV flux. \cite{dst15} have shown that the typical
gas temperature in J2240$-$0927 is around 13800\,K, and 15500\,K if the hotter nuclear region
is included as well.
As can be seen in Fig. \ref{j2241-0927_starlight}, the (redshifted) nebular continuum peaks in the
\textit{GALEX} NUV channel (mostly because of the strong two-photon spectrum). Consequently, we determine
much higher NUV flux densities of 5.4, 9.0 and 10.3\,$\mu$Jy for $T_{\rm e}=10000$, 15000 and 20000\,K,
respectively (38, 63 and 73 per cent of the observed total NUV flux).
\subsubsection{Results for the nebular continuum}
The nebular continuum contributes $2-10$ per cent to the FUV flux of J2240$-$0927, and $38-73$ per cent
of the observed NUV flux. It is much better constrained than the stellar contribution
from SED fitting, because the H$\beta$ line is detected with high S/N and the nebular continuum is
fixed to the H$\beta$ flux. The stellar FUV/NUV continuum is much more uncertain as it is mostly
unconstrained by observational data below 2500\AA. For J2240$-$0927, we have
\begin{eqnarray}
c^{\,{\rm nebular}} & = & 0.02\;\; ({\rm T=10000\,K})\\
c^{\,{\rm nebular}} & = & 0.08\;\; ({\rm T=15000\,K})\\
c^{\,{\rm nebular}} & = & 0.10\;\; ({\rm T=20000\,K}).
\end{eqnarray}
The most conservative estimates from the nebular continuum and the stellar continuum easily account
for the entire \textit{GALEX} NUV flux. In the FUV, on the other hand, the largest conceivable combination
yields about 30 per cent, and the remainder must be attributed to Ly\,$\alpha$ .
\subsection{Scattered AGN light}
Another source of UV continuum is light from the AGN accretion disk, scattered in areas that have an
unobscured view of the nucleus \citep[see][for examples of scattering in the near UV]{pdr93,zss05}.
Without FUV polarization measurements we cannot constrain this effect directly. The absence of
scattered broad lines in the shallow optical spectra of \citetalias{sdh13} indicates that this effect
is insignificant for the sample as a whole. We have also shown above for J2240$-$0927 that the stellar
and the nebular continuum fully account for the NUV observations, leaving little to no headroom
for additional scattered light (Sects. \ref{starlight} and \ref{nebcontinuum}). Therefore,
\begin{equation}
c^{\,\rm scatter}\approx 0\,.
\end{equation}
\section{Analysis of the optical data}\label{opticalresults}
In this Section we describe global characteristics of the GBs. Notes about individual targets
are given in Appendix \ref{targetnotes}.
\subsection{\label{longslit}Environment}
We obtained 52 spectroscopic redshifts (Table \ref{redshiftlist}) of selected field galaxies
to determine the local environment of 13 GBs. The selection function is described in Section
\ref{GMOSobsspec}.
The majority of the GBs live in low-density areas. 35 per cent are isolated, and for
25 per cent we can currently not say whether they are isolated as well, or have $1-3$ possible
companion galaxies. $15-20$ per cent are located in sparse groups with low concentration and
perhaps $3-5$ members. The remaining 25 per cent are found in richer groups of galaxies
with $M_{200}=(1-6)\times10^{13}\; M_{\odot}$ and well-defined red sequences (see Table
\ref{targetlist} and Appendix \ref{targetnotes}).
This is in stark contrast with LABs at high redshift, which are preferentially found in filaments
and clusters. Possibly, at $z=0.3$ the cold accretion streams have been exhausted, and low-z LABs
are mostly formed and ionized by AGN. If a GB is located in an apparent
group or cluster, then it is found near the centre of the distribution of galaxies. Particularly
noteworthy is J1155$-$0147, dominating the group with its size and luminosity. This is the only
GB whose morphology could match a cold accretion stream.
\subsection{\label{mergerrates}Merger rates}
$50-65$ per cent of the GBs interact or merge as evidenced by extended warped stellar haloes,
tidal stellar streams and close companions. Some of the companions show tidal distortions
(e.g. J2240$-$0927), others appear spherically compact, undisturbed and just embedded in
the gas (e.g. J0020$-$0531). Only 15 per cent of the GBs reveal seemingly tidally undisturbed host
galaxies (J0159+2703, J1347+5453, J1504+3439; it is possible that some signs of tidal tails and
interactions have been missed due to their low surface brightness). In all other cases, the bright
EELR prevents a clear view of the hosts, or the hosts are obviously interacting with their
companions. \cite{ylz13} have shown with hydrodynamical simulations and radiative transfer
calculations that binary galaxy mergers will produce LABs with Ly\,$\alpha$ luminosities of $10^{42-44}$\,erg\,s$^{-1}$\;
and typical sizes of $10-50$\,kpc (like GBs), albeit at $z=3-7$. The Ly\,$\alpha$ emission
in these model mergers is mostly produced by intense star formation and gravitational
cooling, whereas in GBs the main power sources are AGN. This is a another indication
of a strong redshift evolution of LABs. We discuss this in Section \ref{discussion}.
\subsection{\label{morphologies}Morphologies of the host galaxies}
The host galaxies of five GBs are easy to classify because of the EELRs' low EWs and the hosts'
large diameters. In J0159+2703 we find a large, 46\,kpc face-on barred spiral galaxy. J1347+5453
is an edge-on disk with 21\,kpc diameter and an axis-ratio of at least 5:1 (the minor axis is
not spatially resolved). The $i$-band data reveal a bright bulge or unresolved nucleus.
J1504+3439 is an elliptical galaxy with a major axis of 37\,kpc. J2202+2309 is a luminous
elliptical near the centre of a galaxy cluster. It can be traced over at least $40\times25$
kpc and is embedded in a common halo with two other ellipticals of similar size and luminosity.
The system could form the future brightest cluster galaxy (BCG) of this structure. J2308+3303
is comprised of a 6\,kpc bright nucleus surrounded by a face-on featureless disk with
$22\times20$\,kpc diameter.
The classification of most other hosts is hampered by the low spatial resolution and
strong line emission in the $gri$ filters. They appear to be compact with major axes of
$8-18$\,kpc (Table \ref{targetlist}). Colours of tidal stellar streams suggest older stellar
populations, but that does not exclude ongoing star formation. Perhaps the most bizarre object
is J1455+0446, consisting of a 40\,kpc large jumbled mix of ionized gas and stars as judged
by its large colour variations. The bright nucleus is found at the edge of the system.
Continuum images from 3D spectroscopy, and $K$-band images of relatively line-free regions
of the spectrum would greatly help the classification.
\subsection{Morphologies of the emission line regions}
The emission line regions in the GBs extend over several 10\,kpc. In the absence of kinematic data,
the spatial image resolution of $2-3$\,kpc allows for some constraints on the formation of the GBs.
Most compelling is the bewildering range of morphologies arising from the combination of various
intrinsic shapes and viewing angles.
\citet{hhg13} have spectroscopically determined the size of [\ion{O}{III}] narrow-line regions (NLRs)
around luminous type-2 quasars, measuring within an $1\times10^{-15}$\,erg\,s$^{-1}$\,cm$^{-2}$\,arcsec$^{-2}$
isophote. They have found typical sizes of $6-8$ kpc, with an upper limit of $10-20$\,kpc.
We do not have spectroscopic data at hand for a direct comparison with their results; however, within
the same $r$-band surface brightness, and along the minor axis, we find typical sizes of $12-25$\,kpc.
At these radii the flux is dominated by [\ion{O}{III}] emission, not by $r$-band continuum, thus a
comparison of our measurements and those of \citet{hhg13} are still meaningful. \citet{hhg13} have
argued that their size limit is caused by the unavailability of gas at larger radii to be ionized.
Likely, this is the reason why our sample differs so much: it was selected because of its extreme
broad-band colours, caused by very gas-rich systems.
\citet{lzg13} and \citet{ham14} have also studied the properties of [\ion{O}{III}] NLRs around luminous
radio-quiet type-2 quasars. The sizes of the [\ion{O}{III}] nebulae in GBs are
consistent or somewhat larger compared to their results; no corrections are made for methodology.
Both authors find mostly circular or moderately elliptical, smooth morphologies for the outflows. Irregular
morphologies are commonly coupled with radio excess. This is in stark contrast with the nebulae observed
in GBs. Most are highly asymmetric, irregular and patchy, apart from J1351+0816 and J2050+0550, which
reveal rather smooth spheres. Again, this could be a selection effect: we found 17 objects in SDSS with
extreme broad-band photometry, whereas \citet{ham14} chose 16 AGN out of a parent sample of 24000 SDSS
AGN. It is entirely possible that some GBs could have ended up in the sample of \citet{ham14};
however, as we have mentioned in Section \ref{contamination}, the SDSS colour space occupied by GBs is
too contaminated for automatic source extraction.
For some targets we give simple estimates about the duration of an AGN burst, and/or the time it
must have occurred in the past. For simplicity, we assume a single, average outflow velocity of
$v=1000$\,km\,s$^{-1}$\;and an inclination angle of $\theta=90$ degrees, i.e. the outflows are moving
perpendicular to the line of sight. Therefore, time estimates must be scaled by
$1000\;{\rm km\;s^{-1}}\;v_{\rm obs}^{\;-1}\;{\rm sin}^{-1}\,\theta_{\rm obs}$
to obtain the true values.
\subsubsection{\label{agnoutflows}AGN driven outflows}
$65-75$ per cent of the EELRs have AGN outflow signatures, such as collimated beams or symmetric ejecta
in opposite directions (Table \ref{targetlist}). The outflows
are usually launched by the injection of thermal energy into the surrounding gas during an AGN
burst. The heated gas expands and sweeps up (and shocks) colder material along its path. Such
outflows have been well studied both observationally as well as theoretically; a detailed account
of these efforts is beyond the scope of our work. We compare our findings to the simulations of
\cite{gab14}, who typically find unipolar outflows with wide opening angles. Accordingly, denser
gas on one side of the nucleus may fully stop an outflow and reradiate its energy, while the
outflow may escape on the other side through a thinner interstellar medium. One object in our sample,
J0111+2253, fits this picture well. It displays a strong unipolar outflow emerging on one side of
the nucleus where it is also wide; a weaker second outflow (or ionized material) is seen at a
30 deg angle, and no outflow is found on the opposite side.
However, J0111+2253 appears to be the exception. For example, we observe bipolar outflows in J0024+3258,
J0113+0106 and J1347+5453 that are well focused near the geometric centre of the host galaxy or their
nuclei. The southern outflow in J0024+3258 even appears collimated over 15\,kpc. J1347+5453 is a posterchild
bipolar outflow, launched from the nucleus of a spiral galaxy perpendicular to the edge-on disk.
Some of the outflows must have been sustained over a prolonged time because their gas is
continuously distributed all the way to the nucleus. Differential velocities in the outflow
will enhance this effect. In case of J0024+3258, the burst would have lasted $8-11$ Myr assuming no
velocity dispersion within the stream. Such long (and Eddington-limited) accretion phases are also found
by \citet{gab14}. Higher resolution images are needed to detect discontinuities in that
outflow. J0113+0106, on the other hand, appears to have experienced a powerful event $\sim5$
Myr ago producing two superbubbles $5-8$\,kpc in size on either side of the nucleus. The bubbles are
$1.5-2$ times larger than the seeing disk and therefore not well resolved. If the observed distances
of the gas from the nucleus are caused by differential gas velocities, then this event could have been
much shorter than 1 Myr. Ionized material at larger distances shows that this recent burst was preceded
by another one, perhaps $20-30$ Myr ago. Recurrent events likely occurred as well in J1347+5453, J1441+2517,
and J1504+3439.
Other systems have a more multipolar character with outflows in different directions.
This could be caused by variable gas densities near the nucleus which may partially stop an outflow or
divide it (as in the northwestern outflow in J1347+5453, and in J0111+2253).
\subsubsection{Cloud systems}
Another typical feature are single or multiple regions of gas, apparently detached from the nucleus.
We refer to them as \textit{clouds}. This could be tidally stripped gas contributed by gas-rich mergers
and now passing through the AGN's ionization cones \citep[like in \textit{Hanny's Voorwerp},
see][]{lsk09,rgj10,kls12}. Typically, these clouds have a relatively smooth appearance and a physical
size of $5-15$\,kpc. Examples are J1441+2517, J1504+3439, J1505+1944, J2050+0550, and most spectacular
in J2240$-$0927 \citep[][]{dst15}.
Alternatively, the clouds were ejected during one or more previous bursts, and then disconnected from
the nucleus and now reside in the galaxies' haloes. Currently, this disconnection could be happening in
J0024+3258, whose northern outflow appears to be still feeding such a cloud, and in J1347+5453, whose
southeastern outflow has a similar structure. In both cases the clouds are significantly misaligned with
the feeding stream, as if they experienced tidal dragging or other interactions with the intergalactic
medium (see also Section \ref{warps}).
\subsubsection{\label{warps}Warps}
Several EELRs show warps and other symmetric and asymmetric deformations that could be caused by
various mechanisms. The gas in J0020$-$0531 resembles a spiral with two widely opened arms that become
thinner with increasing nuclear distance. This could be differential orbital motion, tidal
interaction with two embedded compact ellipticals, or a continuous change in outflow direction. In
J0024+3258, J0113+0106, and J1347+5453 it appears that the ejection direction has changed during or
between bursts, or that the gas has been shaped by interactions with the surrounding halo and/or a
radio jet. GBs are mostly radio quiet or radio weak \citepalias{sdh13}, and thus jet interaction is
unlikely. Radio data from the VLA FIRST survey \citep{wbh97} have insufficient resolution
and depth for further investigation.
Alternatively, the warps could be caused by a change in the ionization cone's opening angle and
strength because of intervening or sublimating dust. Spin precession of the SMBH and its
accretion disk could also play a role. Typical precession periods of $10^{3-7}$ years fully overlap
with the duration of AGN bursts (Section \ref{discSec2}), and the precession cones' half opening angles
range from $1-70\,^{\circ}$ \citep[see][]{luz05}.
\subsubsection{Smooth spheres}
The EELRs in J1351+0816 and J2050+0550 are featureless spheres in our data (J2050+0550 is accompanied
by an ionized cloud, see above). All systems for which we observe outflow signatures are highly
structured, suggesting that a different process has formed these spheres. J2050+0550 is in a
cluster of galaxies and could be in an advanced merger state, engulfed in gas that is now ionized
by the AGN. [\ion{O}{III}] is detected at least to a radius of 20\,kpc by our field redshift survey.
J1351+0816, on the other hand, is isolated in the field. Our redshift survey detects [\ion{O}{III}]
emission out to a radius of at least 48\,kpc. Some process must have transported the gas to these
distances. Unfortunately, the depth and resolution of our spectral data are insufficient to obtain
kinematics and further constraints.
\subsubsection{Peculiar systems}
Three GBs are set apart from the rest by their distinct nebular morphologies. First, J1455+0446 appears
totally disrupted by a merger. Second, J1504+3409 is reminiscent of the
\textit{Voorwerpjes}, ionization echoes found at low redshift \citep{kcb12}. It has several ionized
clouds and bubbles superimposed on the body of a larger \textit{elliptical} galaxy, which distinguishes
it from the Voorwerpjes (mostly spirals and irregulars).
Third, and most interesting, is J1155$-$0147. This is the brightest and also intrinsically most luminous
object in our sample (both in terms of FUV/Ly\,$\alpha$ and [\ion{O}{III}]). It is also the largest object
in terms of area, and second largest in terms of linear diameter (second to J0113+0106). Curiously, it is
also located at the geometric centre of a relatively compact group. The ionized nebula is richly substructered,
fragmenting into smaller clouds. A detailed description is given in Appendix \ref{targetnotes}. Possibly,
J1155$-$0147 has formed by accretion from the intracluster medium, and its Ly\,$\alpha$ emission is a mix of AGN
photoionization and gravitational cooling radiation.
\section{Discussion -- AGN variability and LABs}\label{discussion}
The impact of variable AGN on the appearance of LABs has not yet been studied in detail. In Section
\ref{discSec1} we review the literature, and in Section \ref{discSec2} we present
theoretical and observational evidence for significant episodic AGN phases. We discuss the effects
of AGN variability on the Ly\,$\alpha$, MIR and optical properties in Sects. \ref{discSecLYA},
\ref{discSecMIR}, and \ref{discSecOPT}, respectively.
\subsection{\label{discSec1}Earlier considerations about variability}
AGN variability as an explanation for the ionization deficits in LABs has not been a serious
contender in the light of cold accretion, shocks, starbursts, obscured AGN, and resonant
scattering. Nonetheless, it has been mentioned early on: \citet{sas00} have emphasized the
absence of strong radio and continuum sources in a luminous LAB and noted the possibility
of a ``\textit{dead radio galaxy}'', albeit without elaborating the idea further. Later,
\citet{kwc09} have stated in their summary that ``\textit{Among the proposed explanations
for Ly\,$\alpha$ blobs ... [is] photoionization by active nuclei which may be obscured or transient}''.
The discovery of \textit{Hanny's Voorwerp}, the prototypical quasar ionization echo, has been
published soon thereafter by \citet{lsk09}.
\citet{ond13} have found that LABs with $L_{\rm Ly\,\alpha}\gtrsim5\times10^{43}$\,erg\,s$^{-1}$\; almost always harbour
a luminous (obscured) quasar. Given that the AGN duty cycle is much shorter than that of cold
accretion, they have argued that the high incidence of obscured quasars in these LABs implies a
substantial contribution to the ionization of the gas; the latter could still be provided
by cold accretion streams. Furthermore, given the discovery of ionization echoes, they have stated
that ``\textit{[...] episodic AGN activity may need to be considered as well when interpreting
high-redshift LABs}.
\subsection{\label{discSec2}Evidence for AGN flickering}
Cosmological simulations require recurrent periods of rapid black hole growth, setting black hole
scaling relations and unleashing strong feedback \citep{svg15}. Simulations resolving the gas dynamics
on sub-kpc scales confirm these sharp intermittent bursts of AGN activity, followed by rapid shutdowns,
on time-scales of $\sim10^5$ years \citep{hoq10,noc11,gdg14}. Mergers, disk bar instabilities, and clumpy
accretion may boost the quasar-modes further \citep[e.g.][]{bdt11,bjf12}.
These predictions have been verified observationally by discoveries of ionization
echoes at $z=0.05-0.35$ \citep{sev10,kls12,kcb12,sdh13,ssk13,kmb15}. AGN must undergo several $100-1000$
of these duty cycles (``\textit{flickering}'') to build up their mass \citep{skb15}. Independent
evidence for flickering, albeit on longer time-scales of $\gtrsim1-10$\,Myrs, has been reported by
\citet{kit08} and \citet{ful11} studying the transverse proximity effect in the hydrogen and
helium Ly\,$\alpha$ forest of selected quasars, respectively \citep[see also][]{khm16}.
\subsection{\label{discSecLYA}AGN duty cycles and delayed Ly\,$\alpha$ response}
What does AGN flickering mean for LABs? The escape of Ly\,$\alpha$ photons is delayed because of
resonant scattering, and the Ly\,$\alpha$ flux will lag behind the light curve of the ionizing
source. The effect increases with the optical depth $\tau$, in particular if the ionizing
source is a central AGN. The mean optical depth at the Ly\,$\alpha$ line centre is
\begin{equation}
\tau_0=\pi^{-1/2}\,1.04\times10^7\,\left(\frac{T}{10^4\,{\rm K}}\right)^{-1/2}\,\left(\frac{N_{\rm HI}}{10^{20}\,{\rm cm^2}}\right),
\end{equation}
\citep[e.g.][]{neu90,rsf10}, where $N_{\rm HI}$ is the column number density of neutral hydrogen.
Accordingly, LABs with typical temperatures of $\sim10^4$\,K can have a great range of optical
depths, $\tau=10^{2-4}$ \citep[e.g.][]{dhs06}, or even higher. At much higher temperatures hydrogen
is mostly ionized and optically thin to Ly\,$\alpha$ .
The spatial transfer of resonantly scattered Ly\,$\alpha$ photons does not follow a pure Brownian random
walk because frequency scattering moves photons out of resonance, and therefore they propagate faster.
For these purposes the photon frequency is commonly parametrized as $x=(\nu-\nu_0)/\Delta\nu_{\rm D}$,
measuring the frequency deviation from the resonance frequency, $\nu_0$, in units of the Doppler
broadening, $\Delta\nu_{\rm D}$. The two maxima of the double-peaked Ly\,$\alpha$ line profile occur
at $|x|\gtrsim2$ \citep[e.g.][]{rsf10}.
\citet{rsf10} and \citet{xwf11} have studied the Ly\,$\alpha$ response of spherical Damped Ly\,$\alpha$ haloes
(DLAs; constant hydrogen density and temperature) to an ionizing flash of finite duration. The
effect of dust on the escape times is negligible \citep{yrs11}. We summarize their results,
pertinent to our work, as follows:
\begin{enumerate}
\item{Typical Ly\,$\alpha$ escape times for $10-100$ kpc DLAs are $10-100$ times longer than in the
absence of resonant scattering; the delay scales roughly with $10\,\tau$. \textit{The Ly\,$\alpha$ response
is a very damped version of the light curve of the ionizing source, and the Ly\,$\alpha$ peak brightness
might be reached long after the central source has switched off.}}
\item{Photons with $|x|<4$ are \textit{effectively trapped in an optically thick halo and stored for a
long time}, approximately proportional to $\tau$.}
\item{Photons with $|x|<2$ are \textit{thermalised about 10 times sooner than their typical escape time,
meaning they have lost all memory about the location, spectrum and time variability of the source.}}
\end{enumerate}
The analyses of \citet{rsf10} and \citet{xwf11} were performed for spherical DLAs with high
$\tau\geq10^{6-7}$. LABs are complex objects as witnessed by their vast range of morphologies,
both for our low-z LABs as well as those at high redshift. AGN outflows may clear escape paths
for the Ly\,$\alpha$ photons through the neutral hydrogen, whereas other regions maintain a high optical
thickness. Regardless, the typical double peaked Ly\,$\alpha$ line profiles show that the effects of
resonant scattering are ubiquitous and paramount in LABs. Therefore, the three main results
listed above still apply. A more differentiated analysis of these effects on LABs is desirable,
and will be presented in Malhotra et al. (2016; in prep.).
We conclude that the observed Ly\,$\alpha$ fluxes effectively decorrelate from typical AGN flickering.
LABs with high Ly$\alpha$ luminosities do not require currently powerful AGN. The LABs could
simply be gradually releasing stored photons from earlier high states, while the AGN actually
is in a low state. Conversely, a previously dormant AGN could experience several duty cycles,
stocking up the halo with Ly\,$\alpha$ photons well before any Ly$\alpha$ manages to escapes.
\citet{hwr15} report such Ly\,$\alpha$ deficient radio-quiet quasars, but do not consider variability
as an explanation.
To give an order-of-magnitude calculation: \citet{noc11} find that AGN spend perhaps 1 per cent
of their time in quasar mode. If the storage time of an optically thick LAB is 10 times longer than the
typical burst duration of $\sim10^5$ years \citep{skb15}, then, statistically, in 90 per cent
of the cases we would \textit{not} detect a quasar in X-rays in a randomly selected sample of
the most luminous LABs (if the AGN is below our detection threshold while being in the low state).
This explains at least some of the non-detections that have been attributed to
heavy obscuration \citep[e.g.][]{gal09,ond13}.
\subsection{\label{discSecMIR}AGN duty cycles and delayed MIR response (thermal echoes)}
\subsubsection{\label{argmirx}GBs must have faded recently to violate the MIR X-ray relation}
The X-ray data require the GBs to be intrinsically weak, violating the MIR X-ray relation
(Fig. \ref{ichikawa12}). For a fixed column density of $N_{\rm H}=10^{23}$\,cm$^{-2}$\;,
the GB sample is a factor of $\sim30$ fainter than expected from the MIR X-ray relation.
This discrepancy increases to a factor of $80$ for $N_{\rm H}=10^{22}$\,cm$^{-2}$.
The violation is naturally explained by rapid fading of the AGN. Information about a change
in nuclear luminosity will take $t_{\,\rm lag}=r_{\,\rm sub}/c$ years to reach the dusty torus at
its sublimation radius, $r_{\,\rm sub}$. This time lag is between ten to a few hundred years for
typical tori, and is the \textit{minimum} time for the torus to start a response in the MIR.
The actual shape of the response, and the time needed by the torus to reach a new thermal
equilibrium, depend on the torus' radial dust distribution. The directly illuminated surfaces
at $r_{\rm sub}$ react quickly \citep{nsi08}, whereas shielded and indirectly illuminated parts
further in the back are delayed. \citet{hok11} have analysed the MIR (and NIR) response of
various dusty torus configurations to a discrete pulse from the nucleus with $0.5\,t_{\,\rm lag}$
length. They find that \textit{compact} tori reach their peak brightness quickly, coincident
with the arrival time of the end of the pulse at the sublimation radius at $1.5\,t_{\,\rm lag}$.
At $2.5\,t_{\,\rm lag}$ the MIR luminosity has already dropped again to 50 per cent of the peak
flux. For \textit{thick} tori the peak brightness will be reached at a much later time,
$\gtrsim10\,t_{\,\rm lag}$, and the full MIR response can easily be delayed by $10^3$ years
or more (\textit{thermal echoes}).
The MIR X-ray relation is easily violated by a transient AGN, in particular for ``slow'' tori.
An individual AGN with high MIR luminosity and low or absent X-ray flux is not necessarily
deeply obscured; it might just be fading.
We conclude that the GBs, as a group, have faded by several factors $10-100$ and quicker than
the response times of their dusty tori; more accurate evaluations of $N_{\rm H}$ require deeper
X-ray observations, or observations extending to higher X-ray energies, and have commenced for
some GBs already. Depending on the response times, this change occurred over $10-1000$ years. The
[\ion{O}{III}] excess with respect to the MIR emission reported in \citetalias{sdh13} implies
another drop in luminosity by a factor of $5-50$ over the light crossing time of the optical
EELR ($10^{4-5}$ years). Taking the thermal echoes and ionization echoes together, the AGN in GBs
likely faded by $3-4$ orders of magnitude over the last $10^{4-5}$ years.
\subsubsection{Strong thermal echoes are underrepresented in the MIR X-ray relation}
The MIR X-ray relation is based on observational data, and as such individual geometric
properties, anisotropic shielding as well as AGN variability contribute to its intrinsic
scatter. The influence of strong variability, however, is small. \citet{hoq10} have shown
in their sub-kpc simulations that even for active systems with large time-averaged accretion
rates, the instantaneous inflow rates are modest most of the time. The black hole mass is
typically built up during many short duty cycles with rapid switch on/off times. The AGN
thus spend very little time in the transition phases. \citet{noc11} show in their 2D
simulations that the AGN duty cycles (defined as the fraction of time above a certain
Eddington ratio) are short, typically $10^{-3}-10^{-2}$ or less for an Eddington ratio of
0.1. They do not, however, elaborate on the fraction of time the AGN spend on
\textit{switching} from high to low accretion states (when -- and shortly thereafter --
we would observe them as echoes). \citet{skb15} use observations of low-z ionization echoes
to argue that the time used to switch states is about ten times shorter than the time spent
in the high state.
Taking these results together, in a randomly selected sample of AGN, a fraction of
$\lesssim10^{-4}-10^{-2}$ is expected to be in a significant transient or echo state. The
small numbers of known ionization echoes suggest an even lower occurrence, both at redshifts
$z<0.1$ \citep[\textit{Voorwerpjes,}][]{kcb12,kmb15} and at $z\sim0.3$ (this work). However,
this is a lower limit as these objects need sufficient gas within the ionization cones to
work as an echo screen in first place; otherwise we cannot recognize them as echoes.
We conclude that the intrinsic scatter observed in the MIR X-ray relation (constructed from 127
sources) is caused mostly by intrinsic properties rather than AGN flickering. The MIR X-ray
relation is not an adequate tool to infer properties for AGN that are suspected to be transient.
\subsection{\label{discSecOPT}AGN duty cycles and instantaneous optical response}
The hydrogen recombination time-scale is $\tau_r=(n_{\rm e}\alpha_{\rm B})^{-1}$, where $n_{\rm e}$
is the electron density and $\alpha_{\rm B}$ the recombination coefficient for ``Case B''
\citep{osf06}. For typical densities of $n_{\rm e}=50-200$\,cm$^{-3}$ \citetext{\citetalias{sdh13},
\citealp{dst15}} $\tau_r\sim500$ years in the denser parts of the GBs, and $\gtrsim2000$ years
for the lower densities further out in the nebula. These are short compared to the typical
light crossing times of the resolution elements in our data (a few kpc, Table \ref{genprops}),
unless the density becomes very low $(n_{\rm e}\sim1$\,cm$^{-3})$. The recombination time-scale of
O$^{++}$ is about one order of magnitude shorter than that of hydrogen \citep{bir87}. Therefore,
the response of the GBs' [\ion{O}{III}] and H\,$\alpha$ lines to a sudden change of the ionization
parameter can be considered instantaneous, which we do throughout this paper.
\subsubsection{Reconstructing historic X-ray light curves}
On a side note, their quick optical response makes ionization echoes suitable to reconstruct
historic X-ray light curves. If the echoes' physical extent exceeds 10 kpc as in GBs, then
the reconstructed time line would reach $10^{4-5}$ years into the past, directly
testing AGN accretion models.
This requires that the ionization parameter and the ionizing spectrum can be inferred locally
and with good accuracy. The hardness of the ionizing spectrum can change locally e.g. due to
anisotropic shielding of the nucleus (ionization cones) and local star bursts. In addition,
the 3D cloud must be de-projected, translating angular separations into true physical distances
(light travel times) to the nucleus. Such a de-projection is facilitated using Doppler mapping
and extinction maps to break the foreground-background degeneracy. Additionally, differential
decay times of various optical lines would help \citep{bir87}. Nonetheless, this task is
formidable. Such a reconstruction of the X-ray light curve has yet to be demonstrated. GBs are
ideal for this purpose as they are well resolved and offer high flux densities suitable
for 3D spectroscopy.
\section{Discussion -- Evolution of LABs}
\subsection{\label{discussion1}The LAB size--luminosity function}
Figure \ref{labstats} shows the size--luminosity function for our low-z LABs and some high-z LABs.
For this plot we assume that 75 per cent of the \textit{GALEX} FUV flux is caused by Ly\,$\alpha$
(Sects. \ref{starlight}$-$\ref{fuvJ1155}). The GBs' Ly\,$\alpha$ luminosities overlap well with those of
the high-z LABs, whereas the GBs' ([\ion{O}{III}]) nebulae appear more compact than high-z LABs.
We evaluate the validity of this comparison below.
\subsubsection{Size estimates}
\citet{myh04} and \citet{yze10} list isophotal surface areas which, for our comparison, we converted
to physical diameters assuming circular shapes. These diameters are biased toward smaller values
when compared to major elliptical diameters used by \citet{ebs11}, \citet{pdj13} and also by us.
No correction was made for this effect.
Can [\ion{O}{III}] be used to estimate the Ly\,$\alpha$ extents? \citet{dst15} show for one GB that
[\ion{O}{III}] and H$\alpha$ occupy similar volumes, and thus the [\ion{O}{III}] emission will
provide a lower limit to the Ly\,$\alpha$ extent because of resonant scattering. Long-slit observations
near J1351+0816, J2050+0550, and J2202$+$2309 reveal [\ion{O}{III}] emission out to large radii,
suggesting that the nebulae are up to 4 times larger than measured in our images (see Appendix
\ref{targetnotes}).
\subsubsection{Survey depths}
\citet{sso06} and \citet{yze10} find that different survey depths significantly affect size
estimates for LABs. \citet{myh04,myh11}, \citet{yze10}, and \citet{ebs11} use $1\sigma$ surface
brightness limits of 2.2, 5.5, 1.5 and $1.8\times10^{-18}$\,erg\,s$^{-1}$\,cm$^{-2}$\,arcsec$^{-2}$, targeting $z=3.1$,
2.3, 2.3 and 3.1, respectively. We did not correct sizes for differential survey depths
\citep[see][for a discussion]{sbs11}.
How does our survey depth compare to theirs? Our limiting $r$-band isophotes are
$2\sigma$ above the sky noise (Table \ref{targetlist}, $\sim2.3\times10^{-16}$\,erg\,s$^{-1}$\,cm$^{-2}$\,arcsec$^{-2}$).
Redshifted to $z=3.1$, this becomes $2.3\times10^{-18}$\,erg\,s$^{-1}$\,cm$^{-2}$\,arcsec$^{-2}$ if all flux was caused
by a single emission line, and thus our depth is comparable. However, this is built on the assumption
that the [\ion{O}{III}] surface brightness is an unbiased estimator of the Ly\,$\alpha$ surface brightness.
Line ratios of Ly\,$\alpha$/[\ion{O}{III}]$\,\sim0.1-8$ for other Ly\,$\alpha$ emitters \citep{kww02,mfr11,ond13,mrr14}
show that this will not hold up in general.
We conclude that low-z LABs have similar Ly\,$\alpha$ extents as high-z LABs, yet direct Ly\,$\alpha$ imaging is
required for an unbiased view. While the Ly\,$\alpha$ luminosities of low- and high-z LABs match well, the
high-z Universe is capable of producing more powerful LABs. This could e.g. be caused by cooling
flows, either alone or in addition to AGN ionization.
\begin{figure}
\includegraphics[width=1.0\hsize]{labstats.pdf}
\caption{\label{labstats}{Qualitative comparison of LAB sizes and luminosities. The data
have different selection functions and size definitions (see text). Our sample is shown
by the red dots. Therein, the lower and upper ends of the vertical error bars represent
fractions of 50 and 100 per cent of the FUV flux being caused by Ly\,$\alpha$, the data points
are drawn at 75 per cent.}}
\end{figure}
\begin{figure*}
\includegraphics[width=1.0\hsize]{lab_comoving_dens.pdf}
\caption{\label{lab_comoving_dens}{Left: Comoving density of LABs
($L_{{\rm Ly}\alpha}\geq10^{43}$\,erg\,s$^{-1}$, $D\geq20$\,kpc) as a function of redshift. Solid dots
represent blind surveys and open dots cluster areas. Horizontal lines show the redshift range
covered, vertical lines are the Poisson errors. Arrows represent upper limits. Symbol colours
display the survey comoving volumes (see also the right panel). The pink lines show how GBs
evolve for different X-ray luminosities if linked to the AGN density evolution (Section \ref{diffevolution}). Right: Comoving
density as a function of the surveys' comoving volume. There should be no dependence on volume
if the surveys are independent of the environment. The dashed line shows the best fit to the
blind surveys at $z=2-4$ (larger filled dots). We kept the colour coding for easier comparison
with the left panel.}}
\end{figure*}
\subsection{Evolution of the LAB comoving density}
With our complete sample of GBs, selected from a comoving volume of 3.9 Gpc$^3$ ($z=0.12-0.36$ over
14500 deg$^2$ of SDSS), we can for the first time pin down the LAB comoving density, $\rho_c$,
and its evolution at low redshift. Figure \ref{lab_comoving_dens} displays $\rho_c$ for the GBs and
for some high-z LABs from the literature. The symbol colour encodes the comoving volumes
probed by the surveys, spanning seven orders of magnitude between $10^{5-12}$ Mpc$^3$. The left panel
in Fig. \ref{lab_comoving_dens} reveals a range of five orders of magnitude in density between surveys.
Highest densities are found for (proto-)cluster structures at $z=2-3$ (open symbols), whereas blind
surveys (solid dots) yield lower densities. We describe the evolution as a power law,
$\rho_c\propto(1+z)^n$. Our main results are summarized as follows:
\begin{itemize}
\item{We confirm the previously reported strong evolution below $z=2$; LABs mostly
disappear before $z=0.3$. The slope $n$ remains uncertain. It could be as low as $n=1.7$
over $z=4$ to $z=0.3$, or as high as $n=6$ between $z=0.85$ and $z=0.3$.}
\item{High-z and low-z LAB populations are fundamentally different. Most likely, cold accretion
streams exhaust sometime between $z=2$ and $z=0.3$. At $z=0.3$, LABs are mostly powered
by AGN, and these follow a flatter evolution all the way from $z=4$.}
\item{There is an expected strong dependence of the density on the survey volume,
$\rho_c\propto V_c^{\,-1.6}$. At least Gpc volumes should be probed to appreciate the cosmic
large scale structure and cosmic variance, otherwise density measurements from different
surveys are difficult to compare.}
\end{itemize}
The following unpublished information has also been used for our analysis: \citet{bbb13} have
confirmed that all spectroscopically verified LABs in their sample have $L_{\rm Ly\,\alpha}\geq10^{43}$\,erg\,s$^{-1}$.
\citet{wbc14} have confirmed that no other resolved Ly\,$\alpha$ emitters than the one presented in
\citet{bcw12} are present in their survey. Finally, \citet{myh11} have provided the full list of
LABs including those with diameters less than 100\,kpc. We combine their cluster field (SSA22Sb1)
and the adjacent fields (SSA22Sb2-7) into one data point and consider it a cluster field
(``Matsuda+11 (SSA22)''); we refer to their comparison field survey as ``Matsuda+11 (blind)''.
\subsubsection{Selection effects and biases}
There is currently no definition for LABs \citep[see e.g.][]{sbs11}. For this analysis we require
a minimum diameter of $D\geq20$\,kpc and a minimum luminosity of $L_{\rm Ly\,\alpha}\geq10^{43}$\,erg\,s$^{-1}$. The latter
is a safeguard against completeness issues, biasing us against LABs with low Ly\,$\alpha$ escape fractions.
13 of the 17 GBs are retained within these limits. Most surveys used for our comparison are sensitive
to several $10^{42}$\,erg\,s$^{-1}$, and thus should have picked up most LABs with $L_{\rm Ly\,\alpha}\geq10^{43}$\,erg\,s$^{-1}$\;
in their search volumes. The broad-band survey of \citet{pdj12,pdj13} is an exception, as only the
brightest and largest LABs are detected. We then count how many LABs in the comparison surveys fulfil
our criteria, and we recompute the comoving volumes and densities according to our cosmology. For some
surveys only a sub-sample of the LAB candidates was followed up spectroscopically, and the total number
is extrapolated from the fraction of verified sources. We adopt these correction factors whenever
available \citep[e.g.][]{pdj13,bbb13}.
Constraining the density evolution is complicated by heterogeneous surveys and selection
functions: \citet{myh11}, \citet{sso08} and \citet{yzt09} use narrow-band imaging to select Ly\,$\alpha$
directly, and \citet{pdj12} use broad-band imaging. Spectroscopy is chosen by \citet{kwc09},
\citet{bcw12} and \citet{wbc14}, whereas \citet{bbb13} select dusty Ly\,$\alpha$ emitters using MIR
data. Lastly, our LABs were selected based on their [\ion{O}{III}] strength in $r$-band.
We did not correct for any such selection effects. Further biases arise due to cosmic variance
\citep{yzt09} and because of the original purpose of a survey (clusters, empty fields, etc).
We analyse the density evolution separately for blind surveys and cluster surveys.
Lastly, LABs with different ionization sources (cold accretion, AGN, etc.) evolve differently and
should be studied separately. However, determining the power source(s) of individual LABs is
notoriously difficult (Section \ref{intro}). We ignore this issue for the moment and address it in
Section \ref{diffevolution}.
\subsubsection{\label{clusterdensity}Representative surveys require $10^8$Mpc$^3$ volumes}
Ideally, measurements of the comoving density are independent of the survey volume. If objects
such as LABs are rare, and preferentially found in certain environments such as clusters, then
the search volume must be large to become representative. The colour coding in the left panel
of Fig. \ref{lab_comoving_dens} shows that this is not the case for most LAB surveys. We illustrate this
in the right panel of Fig. \ref{lab_comoving_dens}, plotting the density as a function of volume. Based
on the blind surveys at $z=2-4$, excluding surveys with upper and lower limits and the work of \citet{bbb13},
we find $\rho_c\propto V_c^{\,-1.6\pm0.5}$ (dashed line). In particular narrow-band surveys are observationally
expensive when it comes to probing large volumes, as they cut out thin redshift slices, only.
Most current search volumes do not appreciate the full cosmic environment required to produce LABs. How
big should these volumes be? LABs with $L_{\rm Ly\,\alpha}\geq10^{43}$\,erg\,s$^{-1}$\; and $D\geq20$\,kpc are rare objects at
$z\sim2-4$. They are preferentially found in massive structures with masses of up to $10^{15}$ $M_\odot$
\citep{ptf04,myh11,ebs11,kyi15}, which will evolve into the richest clusters known today. What is
the volume density of such massive clusters? \citet{eeh01} have compiled a statistically complete sample
of the most massive X-ray selected clusters between $z=0.3-0.7$, for which we calculate an average weak
lensing mass of $(1.3\pm0.4)\times10^{15}$ $M_\odot$ based on the data of \citet{alk14}. We derive the
clusters' comoving density as $3.8\pm0.4$ Gpc$^{-3}$. To be \textit{fully} representative of the environment,
an LAB search volume should therefore contain at least one Gpc$^3$. However, much smaller volumes of 0.1
Gpc$^3$ should be sufficient when studying the LAB density in the field, e.g. when probing the fast
evolution between $z=0.5-1.5$ (Section \ref{densityevolutionfield}).
Our subsequent analysis of the LAB density evolution is limited by this effect. We apply approximate
corrections when meaningful.
\subsubsection{Previous evidence for rapid density evolution}
\citet{bcw12} have searched 0.7 deg$^2$ of the Chandra Deep Field South for Ly\,$\alpha$ emitters between
$z=0.67-1.16$ using \textit{GALEX} grism images. They have found only one extended source at
$z=0.977$ with $L_{\rm Ly\,\alpha}=7.2\times10^{42}$\,erg\,s$^{-1}$\; and $D=120$\,kpc. This was the lowest redshift LAB
known so far. For comparison, J1155$-$0147 ($z=0.306$) in our sample is nine times more luminous
(Table \ref{targetUVprop}) while having $D=65$\,kpc (in [\ion{O}{III}]). \citet{wbc14}
have extended the survey of \citet{bcw12} to four times the area, but have not found any other
resolved Ly\,$\alpha$ emitters.
To study the evolution of LABs in clusters, \citet{kwc09} have conducted a \textit{GALEX} grism search
between $z=0.64-1.25$ toward two massive clusters. Numerous LABs \citep[like in][]{myh04} should
have been found in the absence of evolution, yet none was discovered. \citet{kwc09} have concluded
that the LAB density in clusters must be evolving at least as fast as $n=3.0-3.5$ between
$z\sim0.8$ and $z=2-3$.
\subsubsection{\label{densityevolutionfield}LAB density evolution in the field}
For the moment, we marginalize the nature and powering sources of LABs, and we neglect the work of
\citet{bbb13} (see Section \ref{bridge}). Our survey is a blind survey at $z=0.3$. In comparison with
the blind surveys of \citet{sso08}, \citet{yzt09,yze10} and \citet{myh11} we find $n=2.3\pm1.0$ for
$z=0.3-4$, using an unweighted fit. Comparing against \citet{yzt09,yze10} only,
the evolution accelerates to $n=3.6\pm1.2$ between $z=2$ and $z=0.3$. Comparing our result to the
upper limit of \citet{wbc14}, only, the evolution could have been as fast as $n\sim4$ between
$z=0.85$ and $z=0.3$.
Unfortunately, the sampling between $z=0.4-1.5$ is sparse, and it is this interval
of cosmic time when $99-99.9$ per cent of the LABs disappear.
\subsubsection{\label{densityevolutionclusters}LAB density evolution in clusters}
At high redshift, LABs are preferentially found in massive structures with masses of up to
$10^{15}$\,$M_\odot$. These systems evolve into the most massive clusters known at low redshift.
However, the majority of our low-z LABs are found in isolation or small groups. Only four
of them (J1155$-$0147, J1505+1944, J2050+0550, J2202+2309) are in ``denser'' environments with
masses of up to a few $10^{13}$\,$M_{\odot}$.
Have we possibly missed LABs in massive clusters due to selection effects? We do detect luminous
EELRs if superimposed on brighter ellipticals (e.g. J1504+3439 and J2202+2309 with $M_i\sim-22.5$
mag). However, we would not detect them on top of BCGs in the most massive clusters, as
the [\ion{O}{III}] EW would become too small to distinguish the galaxies in broad-band colour
space (Fig. \ref{GB_colorspace}). Some BCGs in massive clusters host emission-line nebula,
powered by AGN outflows, starbursts or cooling flows from the IGM. For a few of them Ly\,$\alpha$
luminosities are known. \citet{hue92} find $L_{\rm Ly\alpha}=(1-5)\times10^{41}$\,erg\,s$^{-1}$\; for
clusters at $z=0.02-0.06$, including NGC 1275A, the BCG of the Perseus cluster. This is a
factor of $\sim100$ below the typical Ly\,$\alpha$ luminosities of GBs. Going to $z=0.06-0.29$,
\citet{obm04,oqo10} find $L_{\rm Ly\alpha}=(0.3-2.8)\times10^{43}$\,erg\,s$^{-1}$\; for 9 massive clusters
with a median of $0.56\times10^{43}$\,erg\,s$^{-1}$\; that is at the bottom of the GB Ly\,$\alpha$ luminosity
function. Abell 1835 ($z=0.253$) is the most luminous system in their sample, reaching
about half the luminosity of the most luminous GBs. Such objects would not normally be considered
as LABs because of their strong continuum.
Given our non-detection, we conclude that LABs are largely absent in clusters with
$M_{200}\gtrsim10^{14}$\,$M_{\odot}$ (unless they are superimposed on the most massive BCGs). This
corresponds to an upper limit of $\rho_c\sim2.5\times10^{-10}$\,Mpc$^{-3}$ for clusters at $z\sim0.3$,
integrated over our entire survey volume. However, we cannot compare this \textit{directly} with
the densities for clusters at $z=2-3$ \citep{ptf04,myh11,ebs11} due to the environmental dependency
of the latter studies caused by their small search volumes. Statistically, the density of clusters
with masses $\gtrsim10^{15}$\,$M_{\odot}$ at $z=0.3-0.7$ is $3.8$ Gpc$^{-3}$ (Section \ref{clusterdensity}).
We expect 15 such clusters in our search volume. If at most one of them contains a LAB, then we have
a conservative upper limit of $\rho_c<3.5\times10^{-8}$ Mpc$^{-3}$ over the same search volume
as \citet{myh11}. This yields a slope of $n>2.8\pm0.6$, compared to $n>3.0-3.5$ obtained by
\citet{kwc09}.
We conclude that, like in empty fields, the conditions that support LABs in dense environments at
$z\sim2$ do not hold anymore at low redshift. This implies that cold accretion streams are important
at high redshift when pristine gas is still abundant in the IGM, whereas they must have been nearly
depleted at low redshift.
\subsubsection{\label{diffevolution}Differential evolution of LABs, or how many GBs do exist at
high redshift?}
We have shown that (1) only LABs in low density environments survive redshift evolution, (2) that low-z
LABs are powered by AGN (Section \ref{agnoutflows}), and (3) that cold accretion streams have likely
ceased at $z\sim0.3$. Consequently, the low-z LAB population is profoundly different from the high-z
population, for which accretion streams abound. If the low-z population of AGN-powered LABs (i.e. the GBs)
exists independently of the high-z population, how many GBs would we expect at
high redshift? In the following consideration we ignore the possibility that high-z LABs, powered
by both AGN and cold accretion, could become pure AGN-powered LABs at low redshift. We also ignore
AGN flickering which moves individual AGN across luminosity bins. And we assume that the occurrence
of an AGN-related GB phase is independent of redshift.
If the GB phase is a (short-lived) phenomenon common amongst active galaxies, then the GB volume
density is tied to the evolution of the AGN volume density. The latter is well known
\citep[e.g.][]{rgw13,mhs15} for different AGN X-ray luminosity bins. We predict the GB volume
density using Fig. 6 in \cite{mhs15} for redshifts $z=1-4$. The predictions are shown by the pink
lines emerging from our data point in the left panel of Fig. \ref{lab_comoving_dens}. With the exception
of the most luminous bin, it is evident that this evolution is much flatter than the evolution of
``typical'' high-z LABs.
If GBs adhere to the lowest luminosity bin, log$(L_x)=42-43$, then with respect to $z=0.3$ their
density would be unchanged at $z=1$ and dropped by a factor of 10 at $z=3$. For the second lowest
bin, log$(L_x)=43-44$, the density would remain constant within a factor of 2 out to $z=3$. For the
two brightest bins, the density would increase significantly by factors $3-10$ at $z=1$. If GBs
were represented by the brightest bin only, then their density would increase further by a factor
50 at $z=2-3$.
Now we can estimate how many high-z GBs have been found by previous LAB surveys. For
the volume probed by \citet{yzt09}, we expect 0.001, 0.01, 0.06 and 0.4 GBs for the four bins of
increasing X-ray luminosity, respectively. For the blind survey part of \citet{myh11}
we expect 0.0006, 0.005, 0.05, and 0.3 GBs. However, these surveys have found 4 and 21 LABs, respectively,
that match our selection criteria ($D\geq20$\,kpc, $L_{\rm Ly\,\alpha}\geq10^{43}$\,erg\,s$^{-1}$). For the broad-band survey
of \citet{pdj13} we find 0.02, 0.1, 1, and 6 GBs, averaging over the survey's large redshift range, and
based on a sub-sample of GBs with $L_{\rm Ly\,\alpha}\geq5\times10^{43}$\,erg\,s$^{-1}$\; detectable by this survey.
\citet{pdj13} extrapolate that 18 such LABs exist in their volume. We conclude that it is unlikely that
GBs have been detected in any current high-z survey, unless the survey volume exceeds $10^7$\,Mpc$^3$
and GBs belong to the intrinsically most luminous AGN (currently they are in a low state).
\subsubsection{\label{bridge}Comparison with \citet{bbb13}}
Now we can also better interpret the comoving density of \cite{bbb13}, with $1.1\times10^{-9}$ Mpc$^{-3}$
$100-10000$ times below that of other measurements at $z=2-4$ (Fig. \ref{lab_comoving_dens}). The MIR
approach chosen by \citet{bbb13} selects LABs with dusty AGN, most of them between $z=1.6-2.8$. X-ray
observations of a representative large sub-sample are not yet available. \citet{sla14} have observed 3 of
the sources (one of which an LAB). They argue that high MIR fluxes and low X-ray counts favour
Compton-thick obscuration over intrinsically weak sources; variability is not considered (see also Section
\ref{fadingagn}). Larger samples are required to characterize this population of LAB/AGN.
The left panel of Fig. \ref{lab_comoving_dens} shows that GBs would evolve to similar volume densities
at $z\sim2$ for log$(L_X)=42-44$. It is plausible that at least some of the LABs described by
\citet{bbb13} and the GBs belong to the same physical population. Bridge, C.R. report that the Ly\,$\alpha$
emission of their LABs is rather filamentary based on the extent of the emission measured at
various position angles (priv. comm.). This would also be the case for our GBs, a large fraction of which
reveals elongated outflows. Another commonality is that both populations are unusually red in the MIR,
although for GBs this is based on the W3 and W4 passbands \citepalias{sdh13} compared to W1 and W2 for
\citet{bbb13}.
\subsubsection{\label{overzier}Meeting the postulation of \citet{ond13}}
\citet{ond13} argue that most luminous LABs with $L_{\rm Ly\,\alpha}\gtrsim5\times10^{43}$\,erg\,s$^{-1}$\;harbour
obscured AGN, and that these AGN are responsible for most of the ionization in the LABs. They consider
that ``\textit{If the luminous LAB phenomenon is associated with powerful AGNs, we would naively expect
that the population of LABs should extend toward much lower redshifts [...]}''. Since the colder gas-rich
conditions of $z\sim2$ proto-clusters, in which the most luminous LABs are found, are replaced by
a hot intracluster medium at $z\lesssim1$, \citet{ond13} suspect that
``\textit{[...] luminous LABs at low redshift may still be found in the more typical, lower density,
and gas-rich environments of actively accreting galaxies and AGNs.}''.
This is a spot-on prediction of the properties of our low-z LABs: 85 per cent have Ly\,$\alpha$ luminosities
in excess of $10^{43}$\,erg\,s$^{-1}$, 75 per cent are found in isolation or with a small number of companion
galaxies, the remaining 25 per cent are found in low mass groups with low concentration.
And lastly, all of our LABs harbour AGN.
\subsection{\label{incompleteness}Are there type-1 Green Beans, and Green Beans with currently active AGN?}
Our optical broad-band selection criteria (Section \ref{contamination}) favour absorbed (type-2)
AGN; unobscured AGN would reduce the observed EW of [\ion{O}{III}] by means of their high continuum
fluxes, making the galaxies integrated colours inconspicuous. An unobscured AGN would also partially
or fully outshine the EELRs, and the latter would be difficult to recognize without PSF subtraction.
The observed obscuration of an AGN strongly depends on the observer's viewing angle \citep{ant93}, and
therefore unobscured (type-1) GBs must exist as well.
It is also well-known that the fraction of obscured AGN decreases with increasing X-ray
luminosity \citep[e.g.][]{mbb14}. Our work, and that of e.g. \citet{hoq10,noc11,gdg14,skb15} show
that high amplitude X-ray flickering is an inherent property of AGN. This suggests that intrinsic
obscuration is also variable and roughly contemporal with the burst duration. The details
are beyond the scope of our paper. We expect that unobscured low-z LABs (or type-1 GB analogs)
must exist in our SDSS search volume, perhaps as many as described in this work.
The following question must be seen in the same context: How many GBs exist with their AGN still
being in the high state, and how would they look like? \citet{bir87} have shown that [\ion{O}{III}]
responds nearly instantaneously to a change of ionizing radiation, at least when considering volumes
with kpc scales such as in GBs. That means that the life-time of an observable GB phase is tightly
linked with the life-time of the AGN bursts; a strong ionization echo is only observed as long as
the ionizing radiation from a previous high state is still escaping the galaxy. Therefore, the ratio
of galaxies that appear as echoes (GBs) and their progenitors (with the AGN still in a high state)
is similar to the ratio of the AGN fading time-scale and the duration of the AGN high state. Both
occur on scales of thousands to hundreds of thousands of years
\citep[this work, and][]{hoq10,noc11,gdg14,skb15}. It is plausible that there might be $1-100$
times as many GB ``progenitors'' as GBs themselves. As argued above, the optical appearance of the
progenitors (and their obscuration) is likely to change \citep[see also the \textit{changing look
quasars}][]{lcm15,rcr16}; we expect galaxies with brighter nuclei, making it more difficult to
recognize their EELRs.
Imaging surveys with greater angular resolution than SDSS (e.g. LSST) would simplify the
identification for unobscured GBs and their progeenitors. They are likely already contained
in public quasar catalogues because of their high fluxes ($r\lesssim18$ mag).
\section{Summary and conclusions}
Green Bean galaxies (GBs, $z\sim0.3$) are spectacular and the most powerful emission-line objects
known in the nearby Universe. Their extended emission-line regions (EELRs) measure between
$20-70$\,kpc, and they are ionized by radio-weak type-2 quasars. GBs are extremely rare with a
surface density
of $1.1\times10^{-3}$ deg$^{-2}$; only 17 of them are known. They were selected photometrically from
the 14500 deg$^2$ footprint of SDSS-DR8. In \citetalias{sdh13} we have
suspected that the EELRs are ionization echoes, i.e. the AGN have faded recently and more quickly
than the EELRs' typical light crossing times. GBs have not been investigated further apart from
the single case study of \citet{dst15},
In this paper we have presented multi-wavelength observations to understand the unusual nature of
GBs and to view them in the context of galaxy evolution. With \textit{Chandra} we have probed the current
activity of the AGN and the intrinsic obscuration along the line of sight toward 10 GBs. With
\textit{GALEX} archival images we have estimated the far-UV properties of 15 GBs, and we have obtained
high resolution optical images in $gri$ filters for all objects.
\subsection{Main results}
\begin{enumerate}
\item{\textit{Chandra} has revealed low counts, moderate hardness ratios and weak or absent
K$\alpha$ lines. This implies that these AGN are intrinsically weak rather than Compton-thick.
The EELRs' high [\ion{O}{III}] luminosities require recent and rapid fading of the AGN, confirming
them as ionization echoes (Section \ref{fadingagn}).}
\item{Strongly variable AGN do not follow the MIR X-ray relation. The MIR response of a dusty torus
can be delayed by up to $\sim10^3$ years \citep{hok11}, forming a thermal echo. Accordingly, the AGN
in the GBs must have faded by several factors $10-100$ over the last $100-1000$ years (Section
\ref{argmirx}). Combining the thermal and ionization echoes, the AGN must have faded by $3-4$ orders
of magnitude within the last $10^{4-5}$ years. This rate is similar to that observed in the
\textit{changing look} quasars \citep[][also attributed to change of accretion rates]{lcm15,rcr16};
however, in GBs it could be sustained over much longer periods of time.}
\item{\textit{GALEX} FUV images require that at least 85 per cent of the GBs have Ly\,$\alpha$ luminosities
in excess of $10^{43}$\,erg\,s$^{-1}$. They form Lyman-$\alpha$ blobs or LABs (Section \ref{GALEXresults} and
\ref{discussion1}). We have proven that LABs still exist in the Universe $4-7$ billion
years later than previously known. Ultimately, the Ly\,$\alpha$ emission has to be confirmed with
FUV spectroscopy using \textit{HST} or \textit{Astrosat} \citep{hut14}.}
\item{We propose rapid duty cycles (AGN flickering) as a natural explanation for the mysterious
ionization deficits observed in LABs. Resonant Ly\,$\alpha$ photons are efficiently stored
in LABs and only gradually released, \textit{decorrelating} from AGN variability on scales of
up to $10^6$ years depending on the optical depth. An AGN may undergo several duty cycles before
Ly\,$\alpha$ escapes; a luminous LAB does not require a \textit{currently} powerful AGN, independent of
obscuration (Section \ref{discSecLYA}). This does not mean that we can relinquish e.g. cold accretion,
star formation and shocks as ionization sources. It means that multi-wavelength observations are
required to identify the ionizing source(s) of individual LABs.}
\item{Low-z LABs live mostly in isolation or in low density environments (Section \ref{longslit}),
whereas high-z LABs are preferentially found in massive structures. Sometime between $z=2$ and $z=0.3$
these structures must lose their ability to form and sustain LABs, probably because cold accretion
streams have depleted. AGN survive, and may continue powering LABs at low redshift (Section
\ref{agnoutflows}).}
\item{Our comoving volume is 3.9 Gpc$^3$, $100-1000$ times larger than other LAB surveys, second only
to \cite{bbb13}. The density at $z\sim0.3$ is $\rho_c=3.3\pm0.9$\,Gpc$^{-3}$ for LABs with
$L_{\rm Ly\,\alpha}\geq10^{43}$\,erg\,s$^{-1}$\; and $D\geq20$\,kpc. The density evolves with $\rho_c\propto(1+z)^{3-4}$ for
both clusters and in the field between $z=0.3$ and $z=2$. During this time, $99-99.9$ per cent of the
LABs disappear. A more accurate determination of the evolution requires (1) better sampling between
$z=0.5-1.5$, and (2) volumes of $0.1-1$\,Gpc$^3$ to overcome cosmic variance and
to be independent of environment. Otherwise, densities are not directly comparable (Section
\ref{clusterdensity}).}
\item{LABs with different ionization sources evolve differently. Gravitationally powered
LABs do not survive redshift evolution as the cold accretion streams are depleting. We find only
one LAB, J1155$-$0147 ($z=0.306$), that could still be powered by cold streams in addition to an
AGN. The density of LABs powered solely by AGN is evolving much slower, if at all, depending on the
AGN's intrinsic X-ray luminosity (Sects. \ref{diffevolution} and \ref{bridge}).}
\end{enumerate}
\subsection{Conclusion and future observations}
LABs should be considered as efficient Ly\,$\alpha$ photon stores, albeit with a badly maintained inventory.
They trap Ly\,$\alpha$ photons for a time much longer than their light crossing time. When the thermalised
photons are eventually released in a gradual manner, all memory about the location, spectrum and
time variability of their source has been lost. Typical storage times are short compared to the life
times of cooling flows, star bursts, and shock ionization populating the store with photons. However,
the storage times become long compared to episodic AGN bursts and their transition from high to
low states (and vice versa). Even if optical and MIR observations point to the most powerful AGN, the
absence of the latter in X-rays does not necessarily mean Compton-thick obscuration.
The AGN timescale
is much shorter than the typical galaxy time-scale and the galaxies response time to AGN activity;
care must be taken when seeking or applying relations between AGN and other ``galactic'' observables,
in particular if the AGN are suspected to be transient \citep[see also][]{hma14}.
The GBs and the associated LABs published in this paper are \textit{much} easier to study than their
high-z counterparts, as low fluxes, redshift incompleteness, and physical resolution are not a problem.
They are perfectly suited to study AGN feedback, large-scale outflows, quasar duty cycles, mode switching,
and the Ly\,$\alpha$ escape fraction, the latter being controlled by an interplay of geometry and velocity field,
metallicity, hydrogen density, dust obscuration and AGN variability.
Our own \textit{NUSTAR} observations have commenced in cycle 2 for the two X-ray brightest GBs,
J1155$-$0147 and J0113+0106, to accurately determine the obscuration and the actual shut-down
``depth'' of their AGN. Time has also been awarded for an initial survey with the
\textit{Hubble Space Telescope}, to determine the FUV properties of the same two GBs, and
J2240$-$0927 \citep{dst15}, using ACS/SBC imaging and spectroscopy.
\section*{Acknowledgements}
I (MS) thank my wife Karianne and my children Hendrik and Jakob for their patience over the last
four years. Carrie Bridge, Gary Ferland, Pascale Hibon, Kohei Ichikawa, Lia Sartori, Mark Schartmann, Peter Schneider,
Isak Wold, Yujin Yang, and the anonymous referee helped to improve this paper with discussions and
comments. Carrie Bridge, Yuichi Matsuda and Isak Wold shared unpublished data for which I am very
grateful.
The authors wish to recognize and acknowledge the very significant cultural role and reverence
that the summit of Mauna Kea has always had within the indigenous Hawaiian community. We are most
fortunate to have the opportunity to conduct observations from this mountain.
Author contributions: MS obtained all data, performed the scientific analysis and wrote the
manuscript; SM pointed out the delayed Ly\,$\alpha$ escape times with respect to the hydrogen recombination
time-scale; ultimately, that led to the solution of the LAB ionization deficit problem; NAL reduced
the \textit{Chandra} data and extracted the X-ray fluxes; HF first suggested the possibility that
GBs might be low-z LAB analogs; MS, RLD, HF, WCK, and PT discussed AGN ionization echoes in depth;
VNB, AP and WCK provided missing redshifts for two GBs using the Lick observatory; JEHT double-checked
the flux calibration of the IFU data.
Support for this work was provided by the National Aeronautics and Space Administration
through \textit{Chandra} Award Number GO4-15110X (PI: M. Schirmer) issued by the
\textit{Chandra} X-ray Observatory Center, which is operated by the Smithsonian Astrophysical
Observatory for and on behalf of the National Aeronautics Space Administration under contract
NAS8-03060. VNB gratefully acknowledges assistance from a National Science Foundation (NSF)
Research at Undergraduate Institutions (RUI) grant AST-1312296. Note that findings and conclusions
do not necessarily represent views of the NSF.
The scientific results reported in this article are based in part on observations made by
the \textit{Chandra} X-ray Observatory, and on data obtained from the \textit{Chandra} Data
Archive. We also made use of the software provided by the \textit{Chandra} X-ray Center (CXC)
in the application packages CIAO.
Based on observations obtained at the Gemini Observatory, which is operated by the Association
of Universities for Research in Astronomy, Inc., under a cooperative agreement with the NSF on
behalf of the Gemini partnership: the National Science Foundation (United States), the National
Research Council (Canada), CONICYT (Chile), Ministerio de Ciencia, Tecnolog\'{i}a e Innovaci\'{o}n
Productiva (Argentina), and Minist\'{e}rio da Ci\^{e}ncia, Tecnologia e Inova\c{c}\~{a}o (Brazil).
Based on observations made with the NASA Galaxy Evolution Explorer.
\textit{GALEX} is a NASA Small Explorer launched in 2003 April. We gratefully acknowledge
NASA's support for the construction, operation, and science analysis for the \textit{GALEX}
mission, developed in cooperation with the Centre National d'Etudes Spatiales of France and
the Korean Ministry of Science and Education.
Based on data products from the Wide-field Infrared Survey Explorer,
which is a joint project of the University of California, Los Angeles, and the Jet Propulsion
Laboratory/California Institute of Technology, funded by the National Aeronautics and Space
Administration.
Based on observations obtained with MegaPrime/MegaCam, a joint project of CFHT and CEA/DAPNIA,
at the Canada-France-Hawaii Telescope (CFHT) which is operated by the National Research Council
(NRC) of Canada, the Institut National des Sciences de l'Univers of the Centre National de la
Recherche Scientifique of France, and the University of Hawaii.
Based on observations made with the European Southern Observatory under Program
287.B-5008, Chile.
Based on observations obtained at the Southern Astrophysical Research (SOAR) telescope, which
is a joint project of the Minist\'{e}rio da Ci\^{e}ncia, Tecnologia, e Inova\c{c}\~{a}o (MCTI)
da Rep\'{u}blica Federativa do Brasil, the U.S. National Optical Astronomy Observatory (NOAO),
the University of North Carolina at Chapel Hill (UNC), and Michigan State University (MSU).
This research has also made use of NASA's Astrophysics Data System Bibliographic Services;
the NASA/IPAC Extragalactic Database (NED) which is operated by the Jet Propulsion Laboratory,
California Institute of Technology, under contract with the National Aeronautics and Space
Administration; the Python {\tt matplotlib} package \citep{hun07}.
\bibliographystyle{mnras}
|
\section{Computing the logical channel and decoding}
\label{s:logical_channel}
Here we provide more detail on how to compute the logical channel for a round of error correction. Note that, via the Choi-Jamiolkowski isomorphism, a qubit channel $\mathcal{E}$ is completely described by the $4\times4$ process matrix
\begin{equation}
C_{ij}=\mbox{Tr}([P_i\otimes P_j][(\mathcal{E}\otimes I) (\ketbra{\Psi^{+}}{\Psi^{+}})])\,,
\end{equation}
where $P_i$ for $i=0,1,2,3$ are the identity and Pauli X, Y and Z matrices respectively. This represents the state obtained when the channel is applied to the first qubit of a Bell state $\ket{\Psi^{+}}=\ket{00}+\ket{11}$, when expressed in the Pauli basis. Here we will show how to compute $C_{ij}$ for the logical channel in the case of surface-code error correction, which will depend on the noise map, the syndrome and the decoder.
The simulated error correction process can be decomposed into three parts
\begin{equation}
\mathcal{E}=\mathcal{D}_s\circ \mathcal{R}_s \circ \mathcal{N}\,,
\end{equation}
where $\mathcal{N}$ is the physical noise map acting on the $N$ qubits of the code, $\mathcal{R}_s$ is the recovery map and $\mathcal{D}_s$ is the decoder correction. The recovery map returns the noisy state to the code space (without performing any classical processing of the syndrome) and can further be decomposed into $\mathcal{R}_s(\rho)=T_s\Pi_s\rho \Pi_s T_s$, where $\Pi_s$ is a projector onto the subspace corresponding to the syndrome $s$ and $T_s$ is a Pauli operator which returns a state in the image of $\Pi_s$ to the codespace. We have defined $T_s$ as a product of Paulis that connects the flipped $X$ checks to the top boundary and flipped $Z$ checks to the left boundary. The decoder correction $\mathcal{D}_s$ is the correction selected from $\{\overline{I}, \overline{X}, \overline{Y}, \overline{Z}\}$ by the decoding algorithm to minimise the logical error.
For now, assume that the decoder does nothing so that $\mathcal{E}=\mathcal{R}_s \circ \mathcal{N}$. The logical channel $\mathcal{E}$ is a map from the codespace to itself, so it is effectively a single-qubit map.
The process matrix describing the transformation of the logical information given the syndrome $s$ is thus
\begin{equation}
C_{ij}=\mbox{Tr}([P_i\otimes P_j][((\mathcal{R}_s \circ \mathcal{N})\otimes I) (\ketbra{\Psi^+}{\Psi^+})])\,,
\end{equation}
where $\ket{\Psi^+}$ is a Bell state of the form $\ket{\Psi^+}=\ket{0}_L\ket{0}_a+\ket{1}_L\ket{1}_a$, where the first qubit is encoded into the surface code and the second qubit is an unencoded ancilla qubit that is assumed to be noise free.
Using the cyclic property of the trace and the fact that $T_s \Pi_s = \Pi_0 T_s$, where $\Pi_0$ is the projection onto the code space, we can write the above expression more explicitly as
\begin{equation}
C_{ij}=\mbox{Tr}([T_s P_i T_s \Pi_s\otimes P_j][ (\mathcal{N}\otimes I)(\ketbra{\Psi^+}{\Psi^+})])\,.
\label{e:process_matrix}
\end{equation}
In the main text we have shown how to represent the Bell state as a tensor network, and how this tensor-network description changes as local noise $\mathcal{N}$, and syndrome projectors $\Pi_s$ are applied to the state. Using this, the trace over the physical indices only
\begin{equation}
A_i=\mbox{Tr}_L([T_s P_i T_s \Pi_s\otimes I][ (\mathcal{N}\otimes I)(\ketbra{\Psi^+}{\Psi^+})])
\label{e:partial_contraction}
\end{equation}
can be represented as a square lattice tensor network, as in \ref{f:tn_evo}d), and can be contracted using the method described in the text. Here $\mbox{Tr}_L$ indicates that the trace is taken over the physical indices of the logical qubit of $\ket{\Psi^+}$ and the ancilla qubit is left uncontracted, and untouched by the other operators in the trace. Therefore the expression in Eq. \eqref{e:partial_contraction} has two uncontracted indices (the ancilla indices), i.e. it is a $2\times2$ matrix. We use these free indices to compute the target quantities $C_{ij}$ in \eqref{e:process_matrix} via $C_{ij}=\mbox{Tr}(P_j A_i)$.
Computing $C_{ij}$ in this way would seem to require computing four full square lattice tensor contractions: one for each $A_i$, i.e. each row of $C_{ij}$. However, as explained in Sec. \ref{s:optimizations} a lot of the intermediate contractions can be reused for different $i$, allowing the full matrix $C_{ij}$ to be computed using only two full lattice contractions. The overall complexity of this calculation is $\mathcal{O}(LW4^W)$ if performed exactly, or $\mathcal{O}(LW\chi^3)$ using the approximate algorithm.
Having computed the effective logical channel $\mathcal{R}_s\circ\mathcal{N}$, a decoder correction $\mathcal{D}_s$ can be incorporated simply by composing the logical channel with the chosen Pauli operator.
Here we have shown how to compute the logical channel for a given syndrome and noise model. In the following section we will describe a decoding algorithm based on the above calculations.
\subsection{Exact decoding}
In the previous section we showed how to compute the logical channel $\mathcal{E}$ for a given noise model and syndrome. This same calculation allows us to perform exact decoding i.e. to select a optimal correction for a given syndrome. Assume that we are given a syndrome $s$ and a noise model $\mathcal{N}$. We can compute $\mathcal{R}_s \circ \mathcal{N}$ using the method described above. Then computing the optimal correction simply amounts to choosing the $\mathcal{D}_s$ from $\{I, X, Y, Z\}$ that minimises the distance of the effective channel from the identity
\begin{equation}
d({D}_s\circ \mathcal{R}_s \circ \mathcal{N},I)\,,
\end{equation}
where $d$ can be any distance between operators e.g. diamond distance, fidelity, or 2-norm distance. In this work we used the 2-norm distance due to its easy computability. If $\mathcal{R}_s \circ \mathcal{N}$ is calculated exactly, this decoding algorithm is optimal, in the sense that it exactly chooses the correction that minimises the distance of the logical channel from the identity. It involves the exact calculation of $\mathcal{R}_s \circ \mathcal{N}$ and is therefore exponential, as described above. However, if $\mathcal{R}_s \circ \mathcal{N}$ is already calculated, the remaining calculations involve manipulations of small matrices and take constant time. Thus choosing an optimal correction requires negligible extra work compared to exactly simulating error correction, and therefore we have used this above optimal decoding in all of our exact simulations. When the approximate contraction algorithm is used, the decoder functions almost identically, except that it uses an approximation of $\mathcal{R}_s \circ \mathcal{N}$, rather than the exact channel. In this case, due to possible errors in the approximation, the decoding is not guaranteed to be optimal. However it would be interesting to see whether such a decoding algorithm is sufficiently accurate for practical purposes.
\section{Computing the error threshold}
\label{s:thresholds}
In Fig. \ref{f:thresholds} we include plots obtained from our simulation data. The simulations were run for depolarizing, amplitude damping, and systematic rotation noise models and the lattice dimensions and the noise strength were varied. For each data point, a large number of samples (between $1.3\times 10^4$ and $1.3 \times 10^5$) were obtained. We have used the decoding algorithm described in the previous section. A threshold was identified as the point below which increasing the lattice size resulted in a decrease in the logical error rate. For amplitude damping we have also plotted data obtained using an approximate contraction algorithm to compute the logical channel for the largest two lattice sizes.
\begin{figure}[h]
\centering
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[width=\textwidth]{threshold_dp.pdf}
\end{subfigure}%
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[width=\textwidth]{threshold_ad.pdf}
\end{subfigure}%
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[width=\textwidth]{threshold_rotation.pdf}
\end{subfigure}
\caption{Logical error rate (defined as the diamond distance of the logical channel from the identity) vs. noise strength for various noise models and lattice sizes.}
\label{f:thresholds}
\end{figure}
For depolarizing and amplitude damping, the thresholds are clear from the graphs and are listed in Table \ref{t:thresholds}. We remark that, for amplitude damping, the crossing point for the approximate data is identical to that obtained using exact data. On the other hand, the z-rotation did not have a clear threshold. The plot is symmetric about the point $\theta=0.25\pi$, where the code performs most poorly. Increasing the lattice size appeared to decrease the logical error rate for a large range of $\theta$, however at certain points the error suppression (e.g. between $0.15\pi$ and $0.225\pi$) was minimal. Therefore we could not confidently pin-down a threshold for systematic rotation.
\section{Optimizations}
\label{s:optimizations}
A number of optimizations were required for our exact algorithm to handle the code sizes studied in this work. In this section, we will briefly describe the main optimizations used.
\medskip
\noindent{\em Reducing the inter-column bond dimension --- } Recall that the exact tensor contraction algorithm we use is exponential, which results from the fact that amount of memory required to store the tensor associated to a column is exponential in the number of inter-column bonds. Every check projector that is applied in the algorithm adds bonds to the tensor network. Therefore we can obtain substantial savings if we can reduce the number of check projectors that are applied. In certain calculations, we do not need to apply all of the checks that we might naively expect. For instance, when computing the logical channel $\mathcal{E}$ as detailed in Sec. \ref{s:logical_channel}, we need to apply the recovery map $\mathcal{R}_s$ to the state. The definition of $\mathcal{R}_s$ involves two layers of checks $\Pi_s$, one to the left and one to the right of the input. However, by making use of the cyclic property of the trace, one of the projectors can be absorbed into the other thereby reducing the number of bonds in the network.
Also, when simulating $z$-rotation, only $x$-checks ever need to be applied. The $z$-checks commute with the noise, and therefore act as the identity on any state in the code space. This is why it was possible to simulate larger codes for $z$-rotation compared with other noise models.
Finally, we can reduce the number of inter-column bonds substantially by exploiting the invariance of the check-projector tensors in Eq. \eqref{e:tensor_projector} under permutation of the particles, and always orienting the three bonds of the check such that there is only one inter-column bond per check. This optimised layout is illustrated for the $\ket{0}_L$ state in Fig.~\ref{fig:TN}b).
\begin{figure}
\centering
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[width=\textwidth]{check_diagram.pdf}
\caption{}
\end{subfigure}\hspace{1cm}%
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[width=\textwidth]{surface_code_tn.pdf}
\caption{}
\end{subfigure}
\caption{a) Distance-5 surface code layout with qubits on the vertices of a rectangular lattice. Orange faces indicate $x$-checks, while white faces indicate $z$-checks. b) A tensor network representing an encoded $\ket{0}_L$ state on this layout. Each box is a $Q^{+}$ tensor, which is applied either to a physical $\ket{0}$ state or to the output leg of another $Q^{+}$ tensor. We have coloured the bottom layer of checks red and the top layer black to make them easier to distinguish. Virtual indices all lie in the plane of the page, and physical indices are perpendicular to it. The bond layout is optimised for the exact contraction scheme we have used.}
\label{fig:TN}
\end{figure}
\medskip
\noindent{\em Reusing tensor contractions --- } In many cases, when computing the contraction of the square lattice tensor network, the entire network does not need to be contracted. When sampling syndrome measurements, for instance, a check measurement does not affect the tensors in columns to the left and right of the check. The contractions in these unaffected regions can therefore be reused for every check measurement in a column. Similar savings can be obtained when computing different elements of the logical channel process matrix $C_{ij}$, since the application of a logical $Z$ only affects a single column of the tensor network.
\medskip
\noindent{\em Optimized layout --- } We have used an optimized surface-code layout which was introduced in \cite{bombin_optimal_2007}. This layout achieves a given code distance with fewer physical qubits than the standard layout. It consists of four-qubit $x$ and $z$ checks arranged in a checkerboard pattern, as well as a number of two-qubit checks along the boundaries. We have illustrated this layout in Fig. \ref{fig:TN}a).
\end{document}
|
\section{Introduction}
Let $X=\{X(t)\mid t\in \R^d\}$ be a real Gaussian random field defined on a probability space $(\Omega, \mathcal{F},\mathbb{P})$. The excursion set of $X$ for the level $u\in\R$ is the random set
\begin{align*}
X^{-1}([u,\infty))=\{t\in \R^d\mid X(t)\geq u\},
\end{align*}
whose properties are an active area of research, cf. \cite{book:AdlerTaylor}, \cite{book:AzaisWschebor}, \cite{paper:LachiezeRey}, \cite{paper:AdlerMoldavskayaSamorodnitsky} among others. As a stochastic model, random fields have many applications, for instance in human brain mapping (\cite{paper:CaoWorsley}), astrophysics (\cite{book:LiddleLyth}) and optics (\cite{paper:BerryDennis}).
To gain a deeper understanding of random excursion sets, several geometric characteristics can be used. In this paper we generalize results for the Euler-Poincar\'e characteristic to the so-called Lipschitz-Killing curvatures $\LK_m$, which are given for $m=0,\ldots ,d-1$ and $M\subset \R^d$ closed with nonempty interior, $\mathcal{C}^2$ boundary and induced Riemannian structure by
\begin{align*}
\LK_{m}(M)= \frac{1}{\omega_{d-m}}\int_{\partial M} \on{detr}_{d-1-m}(S_{E_d}(E_i,E_j))_{i,j=1}^{d-1} \,d\mathcal{H}^{d-1},
\end{align*}
where $(E_i)_{i=1\,\ldots,d-1}$ denotes an orthonormal frame field on $\partial M$, $E_d$ denotes the inward normal, $S$ denotes the scalar second fundamental form, $\on{detr}_{d-1-m}(A)$ denotes the sum over all $(d-1-m)\times (d-1-m)$ principal minors of $A$, the constant $\omega_{d-m}$ denotes the surface area of the $(d-m-1)$-dimensional unit sphere $S^{d-m-1}$ and $\mathcal{H}^{d-1}$ denotes the $(d-1)$\nobreakdash-\hspace{0pt}dimensional Hausdorff measure. For further details see \cite [(10.7.6)]{book:AdlerTaylor}, and in particular, \cite [Section 10.7]{book:AdlerTaylor} for the more complex framework of Whitney stratified spaces considered in this paper. For special choices of $m$ the Lipschitz-Killing curvatures describe simple geometric features of the set like the volume ($m=d$), half the surface area ($m=d-1$) and the Euler-Poincar\'e characteristic ($m=0$).
The aim of this work is to establish a central limit theorem for the standardized $m$\nobreakdash-\hspace{0pt}th Lipschitz-Killing curvature of the intersection of an excursion set for the level $u$ of a stationary isotropic Gaussian random field with an open ball $B^d_N$ of radius $N$, as $N$ goes to infinity, that is
\begin{align*}
\frac{\LK_{m}\left (B^d_N\cap X^{-1}([u,\infty))\right)-\E\left[\LK_{m}\left (B^d_N\cap X^{-1}([u,\infty))\right) \right]}{(\LK_d(B^d_N))^{\frac{1}{2}}} \underset{N\to \infty}{\overset{\mathcal{D}}{\longrightarrow}} \mathcal{N}(0,\sigma^2_m)
\end{align*}
for some $\sigma^2_m\geq0$, where a lower bound for $\sigma_m^2$ is given in Lemma \ref{lem:lowerBound}. The present paper generalizes the work of \cite{paper:EstradeLeon}, where such a CLT is established for $m=0$. The case $m=d-1$ and $d=2$ is treated in \cite{paper:KratzLeon}. For the case $m=d$ of the volume, the central limit theorem holds under weaker requirements than Gaussianity, for instance, for quasi-associated random fields, PA- or NA-random fields, Max- or $\alpha$-stable fields, cf. the survey \cite{review:Spodarev} and the references therein. For this reason we concentrate on the cases $m=0,\ldots,d-1$ in this work.
We pursue the following strategy of proof. First, we apply the Crofton formula from integral geometry to express the $m$-th Lipschitz-Killing curvature of a sufficiently regular set $M\subset \R^d$ as an integral average of the Euler-Poincar\'e characteristics of the intersections of $M$ with affine $(d-m)$-flats, where the integration is with respect to the suitably normalized motion invariant measure $\mu$ over the affine Grassmannian $A(d,d-m)$ of all affine $(d-m)$-flats of $\R^d$ (cf. \cite[Thm. 13.1.1]{book:AdlerTaylor}).
An application to $M=B^d_N\cap X^{-1}([u,\infty))$ leads to the investigation of the Euler-Poincar\'e characteristic of the intersection of the Gaussian excursion with a lower dimensional ball in an affine subspace. By Morse Theory (cf. \cite [Corollary 9.3.5]{book:AdlerTaylor}), this characteristic can be expressed as a difference of counting variables.
Inspired by the ideas of \cite{paper:EstradeLeon}, we use a refinement of the approach in \cite{paper:EstradeLeon} to control the dependence of the counting variables on the affine flat. That is, we use Rice's formulas, cf. \cite [Chapter 6]{book:AzaisWschebor}, \cite [Section 11.2]{book:AdlerTaylor}, in the affine flat to obtain a Hermite expansion of the $m$-th Lipschitz-Killing curvature via an approximation argument. This Hermite expansion leads to a representation of $\LK_{m}\left (B^d_N\cap X^{-1}([u,\infty))\right)$ by stochastic integrals, to which we apply results from the theory of normal approximation based on Stein's method and Malliavin calculus as described in \cite{book:NourdinPeccati}.
The basic tool of our approach, the Wiener chaos expansion, was already prominent in the works of \cite{paper:ChambersSlud}, \cite{paper:Slud} and \cite{paper:KratzLeon1997}, to mention just a few. This access to normal approximations is very popular and is used in various settings similar to ours, cf. \cite{paper:AdlerNaitzat}, who show a central limit theorem for the Euler integration of random functions, \cite{paper:CammarotaMarinucci}, who investigate Gaussian excursions on the $2$-sphere or \cite{paper:Nicolaescu}, who studies critical points of random Fourier series on the $m$-dimensional torus.
Although less explicit, the results of this paper might be compared with recent progress in the second order analysis of the Boolean model, another fundamental model of stochastic geometry, cf. \cite{paper:HugLastSchulte}, \cite{book:LastPenrose}. This progress is largely based on the Malliavin calculus for general Poisson processes.
\section{Main Theorem}
We impose the following conditions on a given real random field $X=\{X(t)\mid t\in \R^d\}$.
\begin{enumerate}[label= (A\arabic*)]
\item \label{as:dif} $X$ is a centered, stationary, isotropic Gaussian field. The trajectories are almost surely of class $\mathcal{C}^3$. The abbreviation $\Co^X(t):=\E\left[ X(t)X(0)\right]$, $t\in \R^d$, denotes the covariance function of $X$, which satisfies $\on{Cov}^X(0)=1$ and $-D^2 \on{Cov}^X(0)=I_d$.
\item \label{as:nondegenerate} For $0\neq t\in \R^d$ the covariance matrix of the vector
\begin{align*}
\left(X(t),\left(\frac{\partial}{\partial t_i} X(t)\right)_{i=1}^d,\left(\frac{\partial^2}{\partial t_i \partial t_j }X(t)\right)_{1\leq i\leq j\leq d},\left(\frac{\partial}{\partial t_i} X(0)\right)_{i=1}^d\right)
\end{align*}
has full rank.
\item\label{as:covariance} The mapping defined by
\begin{align*}
\psi(t):=\max\left \{\left |\frac{\partial^{k}}{\partial t_{j_1}\ldots \partial t_{j_k}}\on{Cov}^X(t)\right |:k\in\{0,\ldots, 4\},1\leq j_1,\ldots,j_k \leq d\right\}
\end{align*}
for $t\in \R^d$, satisfies
\begin{align*}
\psi(t)\stackrel{\|t\|\to \infty}{\longrightarrow}0\text{ and } \psi \in L^1(\R^d).
\end{align*}
\end{enumerate}
We heavily rely on \ref{as:dif} in several places, for instance in the proof of Lemma \ref{lem:decom} and in the calculations in the appendix for Lemma \ref{lem:hilfL^2}. If \ref{as:nondegenerate} holds, the conditions on the covariance from \ref{as:dif} are always satisfied after normalizing the Gaussian field. We believe that it is enough to assume $\mathcal{C}^2$ regularity and an integrability condition on $\Co^X$, cf. \cite{paper:EstradeFournier}, but stick to the $\mathcal{C}^3$ assumption to smoothen the computations of the appendix. Under the differentiability assumptions of \ref{as:dif}, the condition \ref{as:nondegenerate} ensures that the paths of $X$ are almost surely Morse functions and allows us to perform calculations involving Gaussian regressions. Condition \ref{as:covariance} implies that the conditions for a central limit theorem are satisfied.
Note that from \ref{as:covariance} we obtain that $\psi \in L^q(\R^d)$, $q\in \N$, and moreover that $X$ admits a continuous spectral density, cf. \cite [Theorem 2.§12.3 (Inversion Formula)]{book:Shiryaev}. Furthermore the mapping defined by
\begin{align*}
\widetilde\psi(t):=\sup\left \{\left |\frac{\partial^{k}}{\partial v_1\ldots \partial v_k}\on{Cov}^X(t)\right |:k\in\{0,\ldots, 4\},v_1,\ldots,v_k\in S^{d-1}\right\}, \quad t\in \R^d
\end{align*}
satisfies $\widetilde\psi(t)\leq d^2\psi(t)$, for $t\in \R^d$, and therefore is also in $L^q(\R^d)$, $q\in \N$.
Let $u\in \R$ be the level of the considered excursion set and denote by $B^d_N\subset \R^d$ the open ball with radius $N\in \N$ centered at the origin.
We proof the following central limit theorem.
\begin{theorem}\label{thm:mainTheorem}
Let $X$ be a real Gaussian field on $\R^d$, which satisfies the assumptions (A1)--(A3) and let $m\in\{0,\ldots,d-1\}$. Then the $m$-th Lipschitz-Killing curvature $\LK_m$ of the excursion set for the level $u\in \R$ satisfies
\begin{align*}
\frac{\LK_{m}\left (B^d_N\cap X^{-1}([u,\infty))\right)-\E\left[\LK_{m}\left (B^d_N\cap X^{-1}([u,\infty))\right) \right]}{\mathcal{H}^d(B^d_N)^{\frac{1}{2}}} \stackrel{\mathcal{D}}{\longrightarrow} \mathcal{N}(0,\sigma^2_m)
\end{align*}
for $N\to \infty$ and some $\sigma^2_m\geq0$.
\end{theorem}
A lower bound for the asymptotic variance $\sigma_m^2$ is shown in Lemma \ref{lem:lowerBound}.
\section{Approximation of Lipschitz-Killing curvatures}
We fix the following notation. Let $f\colon\R^d \to \R$ be a mapping of class $\mathcal{C}^2$. We denote for $t\in \R^d$ by $\nabla f(t)$ and $D^2 f(t)$ the gradient and the $d\times d$-matrix $(\frac{\partial ^2}{\partial t_i \partial t_j} f)_{1\leq i,j \leq d}$ of second derivatives of $f$, respectively. For $F\in A(d,d-m)$ we denote by $F^\circ$ the directional space of $F$, which is an element in the Grassmannian $G(d,d-m)$ of $(d-m)$-dimensional linear subspaces of $\R^d$. The motion invariant measure $\nu$ on $G(d,d-m)$ is normalized such that $\nu(G(d,d-m))=\binom{d}{d-m}\frac{\omega_{d}}{\omega_{m}\omega_{d-m}}$. We denote by $v(F):=(v_1,\ldots,v_{d-m})$ an orthonormal basis of $F^\circ$ and define the gradient of $f$ in $F$ as the vector field given by
\begin{align*}
\nabla(f|_F)(t):=\sum_{i=1}^{d-m} \frac{\partial}{\partial v_i} f(t) v_i,
\end{align*}
for $t\in F$, where $\frac{\partial}{\partial v_i}$ denotes the directional derivative in direction $v_i$. The second derivative of $f$ in the affine flat $F$ and in point $t\in F$ is defined as the linear mapping on $F^\circ$ given by the $d\times d$-matrix
\begin{align*}
D^2(f|_F)(t) := \left (\begin{array}{@{}c|c|c@{}}
v_1 & \cdots & v_{d-m} \\
\end{array}\right ) \left(\frac{\partial^2}{\partial v_i\partial v_j}f(t)\right)_{1\leq i,j\leq d-m}\left (\begin{array}{@{}c|c|c@{}}
v_1 & \cdots & v_{d-m} \\
\end{array}\right )^\top.
\end{align*}
We note that these definitions coincide with the Riemannian ones, using for $F$ the coordinate map $\varphi\colon F\to \R^{d-m}$ given by $v\mapsto (v_1| \ldots |v_{d-m})^\top v$ and therefore do not depend on the choice of $v(F)$. Moreover, we define
\begin{align} \label{def:nablaInF}
\nabla_{v(F)} f \colon \R^d \to \R^{d-m},\quad t \mapsto \left ( \frac{\partial}{\partial v_i} f(t)\right)_{i=1}^{d-m},
\end{align}
whose components are the coefficients of $\nabla(f|_F)$ in the basis $v(F)$, as well as
\begin{align}\label{def:D2InF}
D^2_{v(F)} f \colon \R^d \to \R^{(d-m)\times (d-m)},\quad t\mapsto \left(\frac{\partial^2}{\partial v_i \partial v_j}f(t)\right)_{i,j=1}^{d-m}.
\end{align}
Using standard results from \cite{book:AdlerTaylor}, we now derive a more practical representation of the $m$-th Lipschitz-Killing curvature $\LK_m$ of the excursion set in $B^d_N$.
We define $\kappa_m:=\mathcal{H}^m(B^m_1)$, $m\in \N$, and consider for $\varepsilon >0$ the mapping
\begin{align*}
\delta_\varepsilon \colon \R^{d}\to\R, \quad x\mapsto \frac{1}{\varepsilon^{d-m}\kappa_{d-m}}\ind_{B^{d}_\varepsilon}(x),
\end{align*}
which is a Dirac sequence for $\varepsilon \to 0$ on every $(d-m)$-dimensional linear subspace $E$ of $\R^d$, that is, for each continuous mapping $f\colon E\to \R$, we have
\begin{align*}
\lim_{\varepsilon \to 0} \int_{E} \delta_\varepsilon(x) f(x) \,dx= f(0).
\end{align*}
We apply the Crofton formula in \cite [Thm. 13.1.1]{book:AdlerTaylor}, to obtain
\begin{align*}
\LK_m\left (B^d_N\cap X^{-1}([u,\infty))\right)&=\LK_m\left (\overline{B^d_N}\cap X^{-1}([u,\infty))\right )-\LK_m\left (S^{d-1}_N\cap X^{-1}([u,\infty))\right)\\
&=\int_{A(d,d-m)}\LK_0\left (\overline{B^d_N}\cap X^{-1}([u,\infty))\cap F\right ) \,\mu(dF)\\
&\quad-\int_{A(d,d-m)}\LK_0\left (S^{d-1}_N\cap X^{-1}([u,\infty))\cap F\right) \,\mu(dF)\\
&=\int_{A(d,d-m)}\LK_0\left (B^d_N\cap X^{-1}([u,\infty))\cap F\right) \,\mu(dF).
\end{align*}
By the assumptions made, we know that the trajectories of $X$ are almost surely Morse functions on $B^d_N\cap F$, for $\mu$ almost all $F$, cf. \cite [Definition 9.3.1]{book:AdlerTaylor} and Lemma \ref{lem:pre}. Therefore, restricting the integration to a suitable subset $A'\subset A(d,d-m)$ as provided by Lemma \ref{lem:pre}, we can apply \cite [Cor. 9.3.5]{book:AdlerTaylor} to the above integrand, to see that
\begin{align*}
\LK_0&(B^d_N\cap X^{-1}([u,\infty))\cap F) \\
&= \#\{t\in B^d_N\cap F:X(t)\geq u, \nabla (X|_F)(t)=0,\iota_{-X,B^d_N\cap F}(t) \text{ even}\}\\
&\quad- \#\{t\in B^d_N\cap F:X(t)\geq u, \nabla (X|_F)(t)=0,\iota_{-X,B^d_N\cap F}(t) \text{ odd}\},
\end{align*}
where $\iota$ denotes the tangential Morse index, cf. \cite [(9.1.2)]{book:AdlerTaylor}. Later computations will benefit from a more general definition in which we define the latter random variable for a bounded, convex window $W\subset \R^d$, and thus define
\begin{align}\label{def:zeta}
\zeta_{m,W}&:=\int_{A(d,d-m)}\#\{t\in W\cap F:X(t)\geq u, \nabla (X|_F)(t)=0,\iota_{-X,W\cap F}(t) \text{ even}\}\notag\\
&\quad- \#\{t\in W\cap F:X(t)\geq u, \nabla (X|_F)(t)=0,\iota_{-X,W\cap F}(t) \text{ odd}\}\,\mu(dF).
\end{align}
Motivated by the use of a Dirac sequence to approximate these counting variables, cf. \cite [Lemma 11.2.10]{book:AdlerTaylor}, we introduce the approximation
\begin{align}\label{def:zetaEpsilon}
\zeta^\varepsilon_{m,W} := (-1)^{d-m}\int_{A(d,d-m)} \int_{W\cap F} \delta_\varepsilon(\nabla (X|_F)(t))\ind\{X(t)\geq u\}\det(D^2(X|_F)(t)) \,dt\, \mu(dF)
\end{align}
and now specify the quality of this approximation. We first need the following Lemma, whose proof is postponed to the appendix:
\begin{lemma}\label{lem:hilfL^2}
Let $D\subset\R^d$ be compact, assume \ref{as:dif} and \ref{as:nondegenerate} and let $W\subset \R^d$ be convex and bounded. Then the following is true:
\begin{enumerate}[label={\normalfont(\roman*)}]
\item There is a constant $c>0$, depending on $X$, $d$, $m$, and $W$, such that for $F\in A(d,d-m)$ and $y\in F^\circ\cap D $
\begin{align*}
\E\left[\# \{t\in W \cap F \colon \nabla (X|_F)(t)=y\}^2 \right]< c.
\end{align*}
\item For all $F\in A(d,d-m)$ the mapping
\begin{align*}
y\mapsto \E\left[ \# \{t\in W \cap F \colon \nabla (X|_F)(t)=y\}^2\right]
\end{align*}
is continuous on $F^\circ \cap D$.
\item For all $F\in A(d,d-m)$, we have
\begin{align*}
\xi_W(F&,\varepsilon)\overset{L^2(\mathbb{P})}{\underset{\varepsilon \to 0}{\longrightarrow}} \xi_W(F),
\end{align*}
where
\begin{align*}
\xi_W(F,\varepsilon)&:=(-1)^{d-m}\int_{W\cap F} \delta_\varepsilon(\nabla (X|_F)(t))\ind\{X(t)\geq u\}\det(D^2(X|_F)(t))\, dt,\\
\xi_W(F)&:=\#\{t\in W\cap F:X(t)\geq u, \nabla (X|_F)(t)=0,\iota_{-X,W\cap F}(t) \mathrm{\, even}\}\\
&\quad- \#\{t\in W\cap F:X(t)\geq u, \nabla (X|_F)(t)=0,\iota_{-X,W\cap F}(t) \mathrm{\, odd}\}.
\end{align*}
\end{enumerate}
\end{lemma}
We now show that the approximation $\zeta_{m,W}^\varepsilon$ is indeed an approximation of the variable of interest $\zeta_{m,W}$.
\begin{lemma}\label{lem:L^2approx}
Let $(X_t)_{t\in\R^d}$ be a real-valued Gaussian field satisfying \ref{as:dif} and \ref{as:nondegenerate} and let $W\subset\R^d$ be convex and bounded. Then
\begin{align*}
\zeta^\varepsilon_{m,W} \stackrel{L^2(\mathbb{P})}{\longrightarrow} \zeta_{m,W}
\end{align*}
for $\varepsilon \to 0$, where $\zeta_{m,W}$ and $\zeta_{m,W}^\varepsilon$ are defined by \eqref{def:zeta} and \eqref{def:zetaEpsilon}, respectively.
\end{lemma}
\begin{proof}
By Jensen's inequality and Fubini's theorem
\begin{align*}
\E\left[ \left (\zeta_{m,W}-\zeta^\varepsilon_{m,W}\right)^2\right] &\leq c \E\left[ \int_{A(d,d-m)}\left ( \xi_W(F)-\xi_W(F,\varepsilon)\right)^2\,\mu(dF)\right]\\
&=c \int_{A(d,d-m)}\E\left[\left ( \xi_W(F)-\xi_W(F,\varepsilon)\right)^2\right]\, \mu(dF),
\end{align*}
where $c=\mu(\{F:F\cap W\neq \emptyset\})\leq {d \brack d-m } \on{diam}(W)^m\kappa_m$, cf. \cite [(6.3.12)]{book:AdlerTaylor} for the definition of the flag coefficients.
Thus, if we justify changing the order of the limits $\lim_{\varepsilon \to 0}$ and $\int_{A(d,d-m)}$, we are done by Lemma \ref{lem:hilfL^2} (iii). In order to apply the dominated convergence theorem, we bound the integrand by an integrable function, not depending on $\varepsilon$. Observe that
\begin{align*}
\E\left[\left ( \xi_W(F)-\xi_W(F,\varepsilon)\right)^2\right]&\leq 2 \E\left[ \#\{t\in W\cap F : \nabla(X|_F)(t)=0\}^2\right] \\
&\quad +2 \E\left[ \left(\int_{W\cap F}\delta_\varepsilon(\nabla (X|_F)(t))|\det(D^2(X|_F)(t))| \,dt\right)^2\right].
\end{align*}
For the first term Lemma \ref{lem:hilfL^2} (i) yields
\begin{align*}
\E\left[ \#\{t\in W \cap F : \nabla(X|_F)(t)=0\}^2\right] \leq c \ind\{F\cap W \neq \emptyset \},
\end{align*}
where $c>0$ is a constant depending on $X$,$d$, $m$ and $W$. For the second term, we apply the coarea formula to $\nabla(X|_F)$, cf. \cite [Theorem 3.2.12]{book:Federer}, then Jensen's inequality to the measure $\ind\{y\in F^\circ\}\delta_\varepsilon(y) \mathcal{H}^{d-m}(dy)$ followed by Fubini's theorem, to obtain
\begin{align*}
\E&\left[ \left(\int_{W\cap F}\delta_\varepsilon(\nabla (X|_F)(t))|\det(D^2(X|_F)(t))| \,dt\right)^2\right] \\
&\leq \int_{F^\circ} \E\left[\#\{t\in W\cap F : \nabla(X|_F)(t)=y\}^2\right]\delta_\varepsilon(y)\,dy.
\end{align*}
Again by Lemma \ref{lem:hilfL^2} (i), we can bound this for all $\varepsilon \leq1$ by the expression
\begin{align*}
c \int_{F^\circ} \delta_{\varepsilon}(y) \,dy \ind\{F\cap W \neq \emptyset \} = c\ind\{F\cap W \neq \emptyset \}.
\end{align*}
Both bounds are independent of $\varepsilon$ and integrable with respect to $\mu$, which shows the assertion. \qedhere
\end{proof}
Before we move on with the main proof, we show the following lemma to obtain
a more concrete representation of $\zeta_{m,W}^\varepsilon$. We note that the special choice of the orthonormal basis $v(F)$ of $F^\circ$, for $F\in A(d,d-m)$, is irrelevant.
\begin{lemma} \label{lem:repcoord}
Let $\varepsilon >0$, $W\subset \R^d$ be convex and bounded and assume \ref{as:dif}. Then
\begin{align*}
\zeta_{m,W}^\varepsilon= (-1)^{d-m}\int_{G(d,d-m)}\int_{W}&\delta_\varepsilon(\nabla_{v(F)}X(t)) \ind\{X(t)\geq u\} \det\left(D^2_{v(F)}X(t)\right)\,dt\, \nu(dF),
\end{align*}
where $\nabla_{v(F)}$ and $D^2_{v(F)}$ are defined in \eqref{def:nablaInF} and \eqref{def:D2InF}, respectively.
\end{lemma}
\begin{proof}
Recall, that by definition $\nabla(f|_F)(t)=\sum_{i=1}^{d-m} \frac{\partial}{\partial v_i} f(t) v_i$ and therefore the rotation invariance of $\delta_\varepsilon$ yields
\begin{align*}
\delta_\varepsilon(\nabla(X|_F))= \frac{1}{\varepsilon^{d-m}\kappa_{d-m}} \ind_{B^d_\varepsilon}(\nabla(X|_F))=\delta_\varepsilon(\nabla_{v(F)}X).
\end{align*}
Also by definition $D^2(X|_F)(t) = \left (\begin{array}{@{}c|c|c@{}}
v_1 & \cdots & v_{d-m} \\
\end{array}\right ) \left(\frac{\partial^2}{\partial v_i\partial v_j}X(t)\right)_{1\leq i,j\leq d-m}\left (\begin{array}{@{}c|c|c@{}}
v_1 & \cdots & v_{d-m} \\
\end{array}\right )^\top$
so that, as a linear mapping from $F^\circ$ into $F^\circ$ it has the transformation matrix $\left(\frac{\partial^2}{\partial v_i\partial v_j}X(t)\right)_{i,j=1}^{d-m}$ with respect to the chosen basis, and therefore we have
\begin{align*}
\det(D^2(X|_F))= \det\left(D^2_{v(F)}X\right).
\end{align*}
This yields with definition \eqref{def:zetaEpsilon}
\begin{align*}
\zeta_{m,W}^\varepsilon= (-1)^{d-m}\int_{A(d,d-m)}\int_{W\cap F}&\delta_\varepsilon(\nabla_{v(F)}X(t))\ind\{X(t)\geq u\} \det\left(D^2_{v(F)}X(t)\right)\,dt\, \mu(dF)
\end{align*}
and we conclude by an application of Fubini's Theorem
\begin{align*}
\zeta_{m,W}^\varepsilon&= (-1)^{d-m}\int_{G(d,d-m)}\int_{L^\perp}\int_{W\cap (L+y)}\delta_\varepsilon(\nabla_{v(L+y)}X(t))\ind\{X(t)\geq u\} \\
&\hspace{4.8cm}\times \det\left(D^2_{v(L+y)}X(t)\right)\mathcal{H}^{d-m}(dt)\,\mathcal{H}^m(dy) \,\nu(dL)\\
&=(-1)^{d-m}\int_{G(d,d-m)}\int_{L^\perp}\int_{L}\ind\{t+y\in W\}\delta_\varepsilon(\nabla_{v(L)}X(t+y))\ind\{X(t+y)\geq u\} \\
&\hspace{4.8cm}\times \det\left(D^2_{v(L)}X(t+y)\right)\,\mathcal{H}^{d-m}(dt)\,\mathcal{H}^m(dy) \,\nu(dL)\\
&=(-1)^{d-m}\int_{G(d,d-m)}\int_{W}\delta_\varepsilon(\nabla_{v(F)}X(t))\ind\{X(t)\geq u\} \det\left(D^2_{v(F)}X(t)\right)\,\mathcal{H}^{d}(dt) \,\nu(dF).\qedhere
\end{align*}
\end{proof}
\section{Hermite type expansion}
From now on, let the field $X$ satisfy the assumptions \ref{as:dif}--\ref{as:covariance}.
We begin this section by defining for $D:=d-m+(d-m)(d-m+1)/2+1$ the $\R^{D}$-valued Gaussian random field $(\mathcal{X}^F_t\colon \Omega \to \R^D\mid(F,t)\in G(d,d-m)\times\R^d)$ by
\begin{align*}
\mathcal{X}^F(t):=\left(\nabla_{v(F)}X(t),\left(\frac{\partial^2}{\partial v_i \partial v_j} X(t)\right)_{1\leq i \leq j \leq d-m}, X(t)\right)
\end{align*}
and denote by $\Sigma$ the covariance matrix of $\mathcal{X}^F(t)$, $(F,t)\in G(d,d-m)\times\R^d$. We note that the definition depends on the choice of $v(F)$, but considering Lemma \ref{lem:repcoord}, this does not matter. We formulate the following lemma.
\begin{lemma}\label{lem:decom}
The matrix $\Sigma$ is independent of $t\in \R^d$ and $F\in G(d,d-m)$. Moreover, we have $\Sigma=\Lambda\Lambda^\top$, where $\Lambda\in\on{GL}_D(\R)$ is given by $\Lambda=\begin{pmatrix}
I_{d-m \times d-m} & 0 \\
0 & \Lambda_2 \\
\end{pmatrix} $,
for some lower triangular matrix $\Lambda_2 \in GL_{D-(d-m)}(\R)$.
\end{lemma}
\begin{proof}
By assumption \ref{as:dif} on the random field $X$, we obtain from \cite [(5.5.3), (5.7.3)]{book:AdlerTaylor} and isotropy
\begin{align}\label{eq:covFirstDer}
&\E\left[\frac{\partial}{\partial v_i}X(t)\frac{\partial}{\partial v_j}X(t)\right]=\E\left[\frac{\partial}{\partial t_i}X(0)\frac{\partial}{\partial t_j}X(0)\right]=\delta_{ij},\\
&\E\left[\frac{\partial}{\partial v_i}X(t)\frac{\partial^2}{\partial v_k\partial v_l}X(t)\right]= \E\left[\frac{\partial}{\partial t_i}X(0)\frac{\partial^2}{\partial t_k\partial t_l}X(0)\right]=0,\notag\\
&\E\left[\frac{\partial}{\partial v_i}X(t)X(t)\right]=\E\left[\frac{\partial}{\partial t_i}X(0)X(0)\right]=0\notag,
\end{align}
as well as
\begin{align*}
&\E\left[\frac{\partial^2}{\partial v_i\partial v_j}X(t)\frac{\partial^2}{\partial v_k\partial v_l}X(t)\right]= \E\left[\frac{\partial^2}{\partial t_i\partial t_j}X(0)\frac{\partial^2}{\partial t_k\partial t_l}X(0)\right],\\
&\E\left[\frac{\partial^2}{\partial v_i\partial v_j}X(t)X(t)\right]=\E\left[\frac{\partial^2}{\partial t_i\partial t_j}X(0)X(0)\right],\\
&\E\left[X(t)X(t)\right]=\E\left[X(0)X(0)\right].
\end{align*}
Assumption \ref{as:nondegenerate} yields that $\Sigma$ is positive definite. Hence the well-known Cholesky decomposition, cf. \cite [Fact 8.9.37]{book:Bernstein}, yields the assertion. \qedhere
\end{proof}
Using $\Lambda$, we define the decorrelated process
\begin{align}\label{def:Y}
Y^F(t):=\Lambda^{-1}\mathcal{X}^F(t), \quad t\in \R^d, F\in G(d,d-m).
\end{align}
For fixed $t\in \R^d$ and $F\in G(d,d-m)$, the random vector $Y^F(t)$ is standard normal, i.e. $Y^F(t)\sim \mathcal{N}(0,I_{D\times D})$. However, note that for different $t,s\in \R^d$ the vectors $Y^F(t)$ and $Y^F(s)$ are in general not independent. In what follows we shall be using the stationarity
\begin{align*}
\left (Y^F(t),Y^{F'}(t')\right )=\left ( Y^F(t+h),Y^{F'}(t'+h) \right),
\end{align*}
where $t,t',h \in \R^d$ and $F,F'\in G(d,d-m)$. Indeed, we have for suitable mappings $f^F$ and $f^{F'}$ that
\begin{align*}
(Y^F&(t),Y^{F'}(t'))\\
&=(f^F(\nabla X(t),D^2X(t),X(t)),f^{F'}(\nabla X(t'),D^2X(t'),X(t')))\\
&\stackrel{\mathcal{D}}{=}(f^F(\nabla X(t+h),D^2X(t+h),X(t+h)),f^{F'}(\nabla X(t'+h),D^2X(t'+h),X(t'+h)))\\
&=(Y^F(t+h),Y^{F'}(t'+h)).
\end{align*}
We now define the mapping $G_\varepsilon\colon\R^{d-m}\times \R^{(d-m)(d-m+1)/2+1} \to \R $, where we use the notation $(x)_{i_1,\ldots,i_k}:=(x_{i_1},\ldots, x_{i_k})$, by
\begin{align*}
G_{\varepsilon}(x,y):= (-1)^{d-m}\delta_{\varepsilon}(x) \det\Big(\big(\Lambda_2y\big)_{1,\ldots,(d-m)(d-m+1)/2} \Big)\ind\{\big(\Lambda_2y\big)_{(d-m)(d-m+1)/2+1}\geq u\},
\end{align*}
so that, by Lemma \ref{lem:repcoord}, we can rewrite the random variable $\zeta^\varepsilon_{m,W}$ as
\begin{align*}
\zeta^\varepsilon _{m,W} = \int_{G(d,d-m)}\int_{W}G_\varepsilon(Y^F(t))\, dt\, \mu(dF).
\end{align*}
In the above definition the vector $\big(\Lambda_2y\big)_{1,\ldots,(d-m)(d-m+1)/2}$
is identified with the symmetric $(d-m)\times(d-m)$-matrix, whose diagonal and upper diagonal entries are given by $\left(\Lambda_2y\right)_{1,\ldots,(d-m)(d-m+1)/2}$, according to the way one identifies $\left(\frac{\partial^2}{\partial v_i \partial v_j} X(t)\right)_{1\leq i \leq j \leq d-m}$ with a vector.
Moreover the mapping $G_\varepsilon$ is an element of $L^2(\R^D,\phi_D\lambda^D)$, where $\phi_D$ denotes the density of a $D$-dimensional standard normal distribution and $\lambda^D$ the $D$-dimensional Lebesgue measure, and therefore can be expanded in the orthonormal basis $\{n!^{-1/2}\widetilde H_n:n\in \N^D\}$, where $\widetilde H_n:=\otimes_{i=1}^DH_{n_i}$ and $H_{k}(x):=(-1)^ke^{\frac{x^2}{2}}\frac{\partial^k}{\partial x^k}e^{-\frac{x^2}{2}}$, $k\in \N\setminus \{0\}$ and $H_0=1$, cf. \cite [Proposition 1.4.2 (iv)]{book:NourdinPeccati}, \cite [Example E.9]{book:Janson}. Thus we obtain
\begin{align}\label{eq:L^2approxG}
G_\varepsilon = \sum_{q=0}^\infty\sum_{n\in \N^D,|n|=q} c(G_\varepsilon,n)\widetilde H_n,
\end{align}
in $L^2(\phi_D\lambda^D)$, where
\begin{align}\label{def:c(Gvarepsilon,n)}
c(G_\varepsilon,n)&:=n!^{-1}\int_{\R^D}G_\varepsilon(x)\widetilde H_n(x) \phi_D(x)\, dx\notag\\
&=\frac{(-1)^{d-m}}{\prod_{i=1}^{d} n_i!}\int_{\R^{d-m}}\delta_\varepsilon(x)\prod_{i=1}^{d-m} H_{n_i}(x) \phi_{d-m}(x)\,dx\int_{\R^{D-(d-m)}}\ind\{(\Lambda_2y)_{D-(d-m)}\geq u\} \notag\\
&\quad \times\det\big((\Lambda_2y)_{1,\ldots,(d-m)(d-m+1)/2} \big)\prod_{i=d-m+1}^D H_{n_i}(y) \phi_{D-(d-m)}(y)\,dy.
\end{align}
It is this expansion, which helps to establish an expansion of the random variable $\zeta^\varepsilon_{m,W}$, as is shown in the next lemma.
\begin{lemma}\label{lem:l^2approx}
Let $\varepsilon>0$ and $W\subset \R^d$ be bounded and convex. Then
\begin{align*}
\zeta^\varepsilon_{m,W}=\sum_{q\geq 0} \sum_{n\in \N^D,|n|=q} \int_{G(d,d-m)}c(G_{\varepsilon},n)\int_{W}\widetilde H_n(Y^F(t))\,dt\, \nu(dF),
\end{align*}
where the convergence is in $L^2(\mathbb{P})$.
\end{lemma}
\begin{proof}
The right side is in $L^2(\mathbb{P})$ since it is a Cauchy sequence, which can be seen by Jensen's inequality and \eqref{eq:L^2approxG}. Recall that by Lemma \ref{lem:repcoord}
\begin{align*}
\zeta^\varepsilon_{m,W}&=\int_{G(d,d-m)}\int_{W}G_\varepsilon(Y^F(t))\,dt\, \nu(dF),
\end{align*}
thus for $Q\in\N$ we have that
\begin{align*}
\E\bigg[ &\bigg( \zeta_{m,W}^\varepsilon- \int_{G(d,d-m)}\int_{W}\sum_{q= 0}^Q \sum_{n\in \N^D,|n|=q}c(G_\varepsilon,n)\widetilde H_n(Y^F(t))\,dt\, \nu(dF)\bigg)^2\bigg]\\
&=\E\bigg[ \bigg( \int_{G(d,d-m)}\int_{W}G_\varepsilon(Y^F(t))-\sum_{q= 0}^Q \sum_{n\in \N^D,|n|=q} c(G_\varepsilon,n)\widetilde H_n(Y^F(t))\,dt\, \nu(dF)\bigg)^2\bigg]\\
&\leq c\E\bigg[ \int_{G(d,d-m)}\int_{W}\bigg (G_\varepsilon(Y^F(t))-\sum_{q= 0}^Q \sum_{n\in \N^D,|n|=q}c(G_\varepsilon,n)\widetilde H_n(Y^F(t))\bigg)^2 \,dt\, \nu(dF)\bigg],
\end{align*}
where we used Jensen's inequality in the last step and $c=\binom{d}{d-m}\frac{\omega_{d}}{\omega_{m}\omega_{d-m}}\mathcal{H}^d(W)$. By Fubini's theorem the latter term equals
\begin{align*}
&c \int_{G(d,d-m)}\int_{W}\E\bigg[\bigg(G_\varepsilon(Y^F(t))-\sum_{q= 0}^Q \sum_{n\in \N^D,|n|=q}c(G_\varepsilon,n)\widetilde H_n(Y^F(t))\bigg)^2\bigg] \,dt\, \nu(dF)\\
&=c \int_{G(d,d-m)}\int_{W}\int_{\R^D}\bigg(G_\varepsilon(x)-\sum_{q= 0}^Q \sum_{n\in \N^D,|n|=q}c(G_\varepsilon,n)\widetilde H_n(x)\bigg)^2 \phi_D(x) \,dx \,dt\, \nu(dF)\\
&=c^2\int_{\R^D}\bigg(G_\varepsilon(x)-\sum_{q= 0}^Q \sum_{n\in \N^D,|n|=q}c(G_\varepsilon,n)\widetilde H_n(x)\bigg)^2 \phi_D(x) \,dx.
\end{align*}
Hence, by (\ref{eq:L^2approxG}), we conclude
\begin{align*}
c^2\int_{\R^D}\bigg(G_\varepsilon(x)-\sum_{q= 0}^Q \sum_{n\in \N^D,|n|=q}c(G_\varepsilon,n)\widetilde H_n(x)\bigg)^2 \phi_D(x)\, dx \stackrel{Q\to \infty}{\longrightarrow }0,
\end{align*}
which shows the assertion. \qedhere
\end{proof}
The following lemma is a special case of \cite [Lemma 3.2]{paper:Taqqu}. We give a prove for completeness.
\begin{lemma} \label{lem:generalisedMehlersFormula}
Let $F,F'\in G(d,d-m), t,t'\in \R^d$ and $n,n'\in \N^D$. Then
\begin{align*}
\E[\widetilde H_n(Y^F(t))\widetilde H_{n'}(Y^{F'}(t'))]=\sum_{\substack{d\in \N^{D\times D}, \\ \sum_{i=1}^D d_{ij}=n_j\,,\,\sum_{j=1}^D d_{i j}=n'_i}} n!n'! \prod_{1\leq i,j \leq D}\frac{\E\left[ Y_i^F(t)Y_j^{F'}(t')\right]^{d_{ij}}}{d_{ij}!}
\end{align*}
for $|n|=|n'|$ and for $|n|\neq |n'|$
\begin{align*}
\E[\widetilde H_n(Y^F(t))\widetilde H_{n'}(Y^{F'}(t'))]=0.
\end{align*}
\end{lemma}
\begin{proof}
We first proof the following: Let $V,W$ be two $D$-dimensional random vectors where $(V,W)\sim \mathcal{N}_{2D}\left(0,\begin{pmatrix}
I_{D} & (\E\left[ V_iW_j\right])_{1\leq i,j\leq D} \\
(\E\left[W_i V_j\right])_{1\leq i,j\leq D} & I_{D} \\
\end{pmatrix}
\right)$. Then
\begin{align*}
\E\left[ \widetilde H_n(V)\widetilde H_{n'}(W)\right]=\ind\{|n|=|n'|\}\sum_{\substack{d\in \N^{D\times D}\\ \sum_{i=1}^D d_{ij}=n_j\,,\,\sum_{j=1}^D d_{i j}=n'_i}} n!n'! \prod_{1\leq i,j \leq D}\frac{\E\left[ V_iW_j\right]^{d_{ij}}}{d_{ij}!}.
\end{align*}
Observe that via the moment generating function of a multivariate normal distribution, we obtain for $t\in \R^{2D}$
\begin{align}
\E&\left[ \prod_{i=1}^D\exp(t_iV_i-\frac{1}{2}t_i^2)\prod_{i=D+1}^{2D}\exp(t_iW_{i-D}-\frac{1}{2}t_i^2)\right] =\exp\left(\sum_{i,j=1}^D t_i t_{D+j}\E\left[ V_iW_j\right]\right).\label{eq:VW}
\end{align}
We use the identity $\exp(tx-1/2t^2)=\sum_{q=0}^\infty t^q/q! H_q(x)$ to see the equality of the left side in (\ref{eq:VW}) to
\begin{align*}
\sum_{n_1,\ldots ,n_D,n'_1,\ldots,n'_{D}=0}^\infty \frac{t_1^{n_1}\ldots t_D^{n_D}t_1^{n'_1}\ldots t^{n'_D}_D}{n!n'!}\E\left[ \widetilde H_{n}(V)\widetilde H_{n'}(W)\right],
\end{align*}
where we used \cite [Lemma 3.1]{paper:Taqqu} to change the order of summation and expectation. The right side in (\ref{eq:VW}) equals
\begin{align*}
&\sum_{r=0}^\infty \frac{1}{r!}\left(\sum_{i,j=1}^Dt_it_{D+j}\E\left[ V_iW_j\right]\right)^r
\\&=\sum_{r=0}^\infty\sum_{d\in\N^{D\times D}, \sum_{i,j=1}^{D} d_{ij}=r} \prod_{1\leq i,j\leq D}\frac{1}{d_{ij}!}(t_it_{D+j})^{d_{ij}}\E\left[ V_iW_j\right]^{d_{ij}}\\
&=\sum_{r=0}^\infty\sum_{d\in\N^{D\times D}, \sum_{i,j=1}^{D} d_{ij}=r} \prod_{1\leq i,j\leq D}\left(\frac{\E\left[ V_iW_j\right]^{d_{ij}}}{d_{ij}!}\right) t_1^{\sum_{k=1}^D d_{k1}}\ldots t_D^{\sum_{k=1}^D d_{kD}}t_{D+1}^{\sum_{k=1}^D d_{1k}}\ldots t_{2D}^{\sum_{k=1}^D d_{Dk}},
\end{align*}
by the multinomial theorem in the first line. Note that the sum over the exponents of the variables $t_1,\ldots,t_D$ equals the one over the exponents of variables $t_{D+1},\ldots ,t_{2D}$, i.e.
$\sum_{i=1}^D \sum_{j=1}^Dd_{ji}= \sum_{i=1}^D \sum_{j=1}^Dd_{ij} = r.$
Hence by comparing the coefficients, we obtain for $|n|\neq |n'|$
\begin{align*}
\E\left[ \widetilde H_{n}(V)\widetilde H_{n'}(W)\right] =0,
\end{align*}
and furthermore for $|n|=|n'|$, the monomial of degree $(n,n')$ corresponds to $r=\frac{1}{2}(|n|+|n'|)$ and can therefore be found in a unique term of the sum over $r$, which yields the assertion.
To conclude the lemma, note that the process $(Y^F(t))_{(F,t)\in G(d,d-m)\times \R^d}$ is Gaussian and the vector $Y^F(t)$ is standard normal for fixed $(F,t)\in G(d,d-m)\times \R^d$. \qedhere
\end{proof}
Using the last lemmata, we can now give a Hermite type expansion of the $m$-th Lipschitz-Killing curvature of the excursion set in the ball of radius N, namely $\zeta_{m,N}$. We first define
\begin{align}\label{def:coefc}
c(n)&:=(2\pi)^{-(d-m)/2}\prod_{i=1}^{d-m}\frac{H_{n_i}(0)}{n_i!}\frac{(-1)^{d-m}}{\prod_{i=d-m+1}^Dn_i!} \int_{\R^{D-(d-m)}}\det\big((\Lambda_2y)_{1,\ldots,(d-m)(d-m+1)/2} \big) \notag\\
&\quad\times\ind\{(\Lambda_2y)_{(d-m)(d-m+1)/2+1}\geq u\}\prod_{i=d-m+1}^D H_{n_i}(y) \phi_{D-(d-m)}(y)\,dy.
\end{align}
Since $\frac{1}{\prod_{i=1}^{d-m} n_i!}\int_{\R^{d-m}}\delta_\varepsilon(x)\prod_{i=1}^{d-m} H_{n_i}(x) \phi_{d-m}(x)\,dx \stackrel{\varepsilon\to 0}{\longrightarrow} (2\pi)^{-(d-m)/2}\prod_{i=1}^{d-m}\frac{H_{n_i}(0)}{ n_i!}$ we obtain $c(n)=\lim_{\varepsilon\to 0} c(G_\varepsilon,n)$. The coefficient $c(\cdot)$ is the coefficient in the expansion of $\zeta_{m,W}$ as we see in the next lemma. Note that, the following expansion is orthogonal due to the last lemma.
\begin{theorem}\label{thm:L^2Approx}
Let $W\subset \R^d$ be convex and bounded. Then
\begin{align}
\zeta_{m,W}\stackrel{L^2(\mathbb{P})}{=}\sum_{q\geq 0} \sum_{n\in \N^D,|n|=q} \int_{G(d,d-m)}c(n)\int_{W}\widetilde H_n(Y^F(t))\,dt\, \nu(dF). \label{eq:L^2Approx}
\end{align}
\end{theorem}
\begin{proof}
We show that $(\sum_{q=0}^Q\sum_{|n|=q}\int_{G(d,d-m)}c(n)\int_{W}\widetilde H_n(Y^F(t))dt \nu(dF))_{Q\in \N}$ is a Cauchy sequence, so that the right side of the asserted equation is in $L^2(\mathbb{P})$. For $Q_1<Q_2 \in \N$ we have
\begin{align*}
\E[&\big( \sum_{q=Q_1}^{Q_2}\sum_{|n|=q}\int_{G(d,d-m)}c(n)\int_{W}\widetilde H_n(Y^F(t))\,dt\, \nu(dF) \big)^2]
\\&\leq\liminf_{\varepsilon \to 0}\E[\big( \sum_{q=Q_1}^{Q_2}\sum_{|n|=q}\int_{G(d,d-m)}c(G_\varepsilon,n)\int_{W}\widetilde H_n(Y^F(t))\,dt \,\nu(dF) \big)^2]
\end{align*}
by Fatou's lemma. The orthogonality of Lemma \ref{lem:generalisedMehlersFormula} yields equality to
\begin{align}\label{eq:CauchySeq}
\liminf_{\varepsilon \to 0}&\sum_{q=Q_1}^{Q_2}\E[\big( \sum_{|n|=q}\int_{G(d,d-m)}c(G_\varepsilon,n)\int_{W}\widetilde H_n(Y^F(t))\,dt \,\nu(dF)\big)^2]
\notag\\&\leq\sum_{q=Q_1}^{\infty}\liminf_{\varepsilon \to 0}\E[\big( \sum_{|n|=q}\int_{G(d,d-m)}c(G_\varepsilon,n)\int_{W}\widetilde H_n(Y^F(t))\,dt\, \nu(dF)\big)^2],
\end{align}
where we added positive terms in the second line. Note that, in order to use the orthogonality we need Fubini's theorem, which is applicable as a consequence of \cite [Lemma 3.1]{paper:Taqqu}.
By Fatou's lemma and the Pythagorean identity the latter is bounded from above by
\begin{align*}
\liminf_{\varepsilon \to 0}\E[\big(\sum_{q=0}^{\infty} \sum_{|n|=q}\int_{G(d,d-m)}c(G_\varepsilon,n)\int_{W}\widetilde H_n(Y^F(t))\,dt\, \nu(dF)\big)^2]
&=\liminf_{\varepsilon \to 0} \E[(\zeta_{m,W}^\varepsilon)^2]\\
&=\E[(\zeta_{m,W})^2]<\infty,
\end{align*}
where we have used Lemma \ref{lem:l^2approx} and finally Lemma \ref{lem:hilfL^2} (i). Thus \eqref{eq:CauchySeq} is the tail of a convergent series, which yields that the sequence is Cauchy.
Now define $\widetilde I_q:=\sum_{n\in\N^D,|n|=q}\int_{G(d,d-m)}c(n)\int_{W}\widetilde H_n(Y^F(t))\,dt \,\nu(dF)$ and write $\pi^Q(f)$ for the projection onto the first $Q$ chaos in $L^2(\mathbb{P})$ and likewise $\pi_Q(f)$ for the projection onto the chaos greater than $Q$, $Q\in \N$, $f\in L^2(\mathbb{P})$. To show the asserted equality, observe that
\begin{align*}
\|\zeta_{m,W}-\sum_{q=0}^\infty \widetilde I_{q}\|_{L^2}&\leq \|\pi_Q(\zeta_{m,W})-\sum_{q=Q}^\infty\widetilde I_{q}\|_{L^2}+\|\pi^Q(\zeta_{m,W}-\zeta^\varepsilon_{m,W})\|_{L^2}+ \|\pi^Q(\zeta^\varepsilon_{m,W})-\sum_{q=0}^Q\widetilde I_{q}\|_{L^2}\\
&\leq \|\pi_Q(\zeta_{m,W})\|_{L^2} + \|\sum_{q=Q}^\infty\widetilde I_{q}\|_{L^2} + \|\zeta_{m,W}-\zeta^\varepsilon_{m,W}\|_{L^2}+ \|\pi^Q(\zeta^\varepsilon_{m,W})-\sum_{q=0}^Q\widetilde I_{q}\|_{L^2}.
\end{align*}
The first two terms tend to 0 for $Q\to \infty$, since both functions belong to $L^2(\mathbb{P})$, as does the third one for $\varepsilon \to 0$, due to Lemma \ref{lem:L^2approx}. For the last one we have
\begin{align*}
\|\pi^Q(\zeta^\varepsilon_{m,W})-\sum_{q=0}^Q\widetilde I_{q}\|_{L^2}&=\E[\big(\sum_{q=0}^Q\sum_{|n|=q}\int_{G(d,d-m)}c(G_\varepsilon,n)\int_{W}\widetilde H_n(Y^F(t))\,dt \,\nu(dF)\\
&\quad-\sum_{q=0}^Q\sum_{|n|=q}\int_{G(d,d-m)}\lim_{\varepsilon \to 0}c(G_\varepsilon,n)\int_{W}\widetilde H_n(Y^F(t))\,dt \,\nu(dF) \big)^2],
\end{align*}
which equals
\begin{align*}
&\sum_{q,q'=0}^Q\sum_{|n|=q}\sum_{|n'|=q'}\left (c(G_\varepsilon,n)-\lim_{\varepsilon \to 0}c(G_\varepsilon,n)\right)\left (c(G_\varepsilon,n')-\lim_{\varepsilon \to 0}c(G_\varepsilon,n')\right)\\
&\quad \times \E[\int_{G(d,d-m)}\int_{W}\widetilde H_n(Y^F(t))\,dt\, \nu(dF)\int_{G(d,d-m)}\int_{W}\widetilde H_{n'}(Y^F(t))\,dt\, \nu(dF)].
\end{align*}
The assertion follows by first taking the limit $\varepsilon \to 0$ and then $Q\to \infty$. \qedhere
\end{proof}
\section{Embedding into an isonormal process}
We now embed the Gaussian field $(Y^F_t\colon \Omega \to \R^D \mid (F,t)\in G(d,d-m)\times \R^d)$ into an isonormal process. By standard theory, for instance in \cite [Section 5.4]{book:AdlerTaylor}, we obtain for $s,t\in \R^d$
\begin{align*}
\on{Cov}^X(s,t)=\on{Cov}^X(s-t)=\int_{\R^d}e^{i\langle s-t, \lambda \rangle} \,f\lambda^d(d\lambda),
\end{align*}
where $f\lambda^d$ denotes the spectral measure of $X$ and $f$ the spectral density. Recall that the spectral density exists due to \ref{as:covariance}. Moreover we obtain
\begin{align}\label{eq:covDif}
\E&\left[ \frac{\partial^k}{\partial v_{1}\ldots\partial v_{k}} X (t)\frac{\partial^l}{\partial v'_{1}\ldots\partial v'_{l}} X (s)\right]=(-1)^l \int_{\R^d}\frac{\partial^{(k+l)}}{\partial v_{1}\ldots\partial v_{k}\partial v'_{1}\ldots\partial v'_{l}}\left(e^{i\langle \cdot, \lambda \rangle}\right)(s-t)\,f\lambda^d(d\lambda),
\end{align}
where $k,l\in \{0,1,2\}$, $v_1,\ldots, v_{k},v'_1,\dots, v'_l\in S^{d-1}$. We define the real Hilbert space \begin{align*}
\mathfrak{H}:=\{h\colon \R^d\to \C\mid h(-x)=\overline{h(x)}\}
\end{align*}
equipped with the scalarproduct $\langle f,g\rangle_{L^2(f\lambda^d)}:=\int_{\R^d}f(\lambda)\overline{g(\lambda)}\,f\lambda^d(d\lambda)$, which is real since the functions are Hermitian and $f\lambda^d$ is symmetric. By \cite [Prop. 2.1.1]{book:NourdinPeccati}, we know that there exists an isonormal process $W$ on $\mathfrak{H}$, so that for $f,g\in \mathfrak{H}$
\begin{align}\label{eq:isoProcess}
\E\left[ W(f)W(g)\right]=\langle f,g\rangle_{L^2(f\lambda^d)}.
\end{align}
Moreover we define for $F\in G(d,d-m)$ and $j=1,\ldots,D$ the mapping
\begin{align*}
\varphi^F_{t,j}\colon \R^d \to \C, \lambda \mapsto \sum_{k=1}^D \Lambda_{jk}^{-1}\nu_{k}^F(\lambda)e^{i\langle t, \lambda \rangle} \in \mathfrak{H},
\end{align*}
where
\begin{align*}
\nu^F\colon \R^d\to \C^D, \lambda \mapsto \big((i\langle v_l,\lambda\rangle)_{1\leq l \leq d-m},(-\langle v_l,\lambda\rangle \langle v_s,\lambda\rangle)_{1\leq l\leq s\leq d-m},1\big )
\end{align*}
and $v_1,\ldots, v_{d-m}$ denotes the chosen orthonormal basis of $F$. Note that $\nu_k^F(\lambda)e^{i\langle \cdot, \lambda \rangle}$ is the directional derivative of $e^{i\langle \cdot, \lambda \rangle}$ of the same order and in the same direction as the derivative of $X$ in the $k$-th component of $\mathcal{X}^F$.
Then we obtain
\begin{align*}
Y^\cdot(\cdot\cdot)\stackrel{\mathcal{D}}{=} \left(W(\varphi^\cdot_{\cdot\cdot,1}),\ldots, W(\varphi^\cdot_{\cdot\cdot,D})\right)
\end{align*}
as processes on $G(d,d-m)\times \R^d$. To see this, it suffices to show that their covariance structures coincide, since both processes are centered Gaussian processes. By the definition of $Y$, cf. \eqref{def:Y}, and \eqref{eq:covDif}
\begin{align}\label{eq:covY}
\E[Y^F_i(t)Y^{F'}_j(t')]&= \sum_{r,s=1}^{D}\Lambda_{ir}^{-1}\Lambda_{js}^{-1} \E[\mathcal{X}_{r}^F(t)\mathcal{X}_{s}^{F'}(t')]\notag \\
&= \sum_{r,s=1}^{D}\Lambda_{ir}^{-1}\Lambda_{js}^{-1}\int_{\R^d} \nu^F_r(\lambda)e^{i\langle t, \lambda \rangle}\overline{\nu^{F'}_s(\lambda)e^{i\langle t', \lambda \rangle}} \,f\lambda^d(d\lambda)\notag \\
&=\langle \varphi^F_{t,i},\varphi^{F'}_{t',j}\rangle_{L^2(f\lambda^d)},
\end{align}
for $(F,t),(F',t')\in G(d,d-m)\times \R^d$ and $i,j\in\{1,\ldots, D\}$. By \eqref{eq:isoProcess} we obtain
\begin{align*}
\langle \varphi^F_{t,i},\varphi^{F'}_{t',j}\rangle_{L^2(f\lambda^d)} =\E[W(\varphi^F_{t,i})W(\varphi^{F'}_{t',j})]
\end{align*}
and therefore the assertion. Moreover, observe that
\begin{align*}
\langle \varphi^F_{t,i},\varphi^F_{t,j}\rangle_{L^2(f\lambda^d)} =\E[Y^F_i(t)Y^F_j(t)]=\delta_{ij},
\end{align*}
for $i,j=1,\ldots,D$ and $(F,t)\in G(d,d-m)\times\R^d$. Hence \cite [Theorem 13.25]{book:Kallenberg} implies the second equality in
\begin{align*}
\prod_{i=1}^D H_{n_i}(Y^\cdot_i(\cdot\cdot))\stackrel{\mathcal{D}}{=}\prod_{i=1}^DH_{n_i}(W(\varphi^\cdot_{\cdot\cdot,i}))=I_q(\varphi_{\cdot\cdot,1}^{\cdot\otimes n_1}\otimes \ldots \otimes \varphi_{\cdot\cdot,D}^{\cdot\otimes n_D}),
\end{align*}
where $q,D\in \N$ and $n\in \N^D$ such that $|n|=q$.
The last equation and Theorem \ref{thm:L^2Approx} with the choice $W=B^d_N$ yield
\begin{align*}
\frac{\zeta_{m,B^d_N}-\E[\zeta_{m,B^d_N}]}{(N^{d}\kappa_{d})^{1/2}}\stackrel{\mathcal{D}}{=}& \sum_{q=1}^\infty\frac{1}{(N^{d}\kappa_{d})^{1/2}}\sum_{n\in\N^D,|n|=q}\int_{G(d,d-m)} c(n)\int_{B^d_N} I_{q}(\varphi_{t,1}^{F\otimes n_1}\otimes \ldots \otimes \varphi_{t,D}^{F\otimes n_D}) \,dt \,\nu(dF),
\end{align*}
where the right side converges in $L^2(\mathbb{P})$.
We now symmetrise the arguments of the stochastic integral. To this end define for $q,D\in \N$ and $n\in \N^D$ with $|n|=q$ the set
\begin{align*}
\A_n:=\{k\in\{1,\ldots,D\}^q:\sum_{j=1}^q\ind_{\{i\}}(k_j)=n_i,\forall i=1,\ldots,D\}
\end{align*}
of multiindices, which contain the number $i$ exactly $n_i$ times. Note that for $k\in \A_n$ all permutations of $k$ are also in $\A_n$, and moreover, these sets form a partition of the set $\{1,\ldots,D\}^q$, i.e. $\{1,\ldots,D\}^q=\dot\cup_{n\in\N^D,|n|=q}\A_n$. We further define for $k\in \{1,\ldots,D\}^q$
\begin{align*}
b(k):= \sum_{n\in\N^D,|n|=q}\ind\{k\in \A_n\}\frac{c(n)}{|\A_n|},
\end{align*}
which is symmetric in the components of $k$.
Since the Wiener-It\^o integrals are invariant with respect to permutations, we obtain for $n\in \N^D$ with $|n|=q$
\begin{align*}
I_q(\varphi_{t,1}^{F\otimes n_1}\otimes \ldots \otimes \varphi_{t,D}^{F\otimes n_D}) = \frac{1}{|\A_n|}\sum_{k\in \A_n}I_q(\varphi^F_{t,k_1}\otimes \ldots \otimes \varphi^F_{t,k_q})
\end{align*}
and thus
\begin{align*}
\sum_{n\in\N^D,|n|=q} c(n)I_q(\varphi_{t,1}^{F\otimes n_1}\otimes \ldots \otimes \varphi_{t,D}^{F\otimes n_D})&= \sum_{n\in\N^D,|n|=q}\sum_{k\in\A_n}\frac{c(n)}{|\A_n|}I_q(\varphi_{t,k_1}^F\otimes \ldots \otimes \varphi^F_{t,k_q})\\
&=\sum_{k\in\{1,\ldots,D\}^q}b(k)I_q(\varphi^F_{t,k_1}\otimes \ldots \otimes \varphi^F_{t,k_q}).
\end{align*}
Hence by Fubini's theorem for Wiener-It\^o integrals, we finally obtain a representation for the standardized $\zeta_{m,B^d_N}$, which is amenable to the theory described in \cite{book:NourdinPeccati}, i.e.
\begin{align*}
\frac{\zeta_{m,B^d_N}-\E[\zeta_{m,B^d_N}]}{(N^{d}\kappa_{d})^{1/2}}\stackrel{\mathcal{D}}{=}\sum_{q=1}^\infty I_q(g_{N,q}),
\end{align*}
where
\begin{align}\label{def:gNq}
g_{N,q}:=\frac{1}{(N^{d}\kappa_{d})^{1/2}}\sum_{k\in\{1,\ldots,D\}^q}\int_{G(d,d-m)} b(k)\int_{B^d_N} \varphi^F_{t,k_1}\otimes \ldots \otimes \varphi^F_{t,k_q} \,dt\, \nu(dF)
\end{align}
is symmetric, since the coefficients $b(\cdot)$ are.
\section{Proof of the main theorem}
We now apply Theorem 6.3.1 in \cite{book:NourdinPeccati}, which yields, once we have checked the required conditions, the main theorem of this paper. We repeat it here for completeness and note that in the monograph \cite{book:NourdinPeccati} condition (iv) is stated slightly differently but the proof given there remains the same.
\begin{theorem}[Theorem 6.3.1 in \cite{book:NourdinPeccati}]\label{thm:NourdinPeccati}
Let $F_N \in L^2(\mathbb{P})$, for $N\in \N$, such that $\mathbb{E}[F_N]=0$. Then there exist functions $g_{N,q}\in \mathfrak{H}^{\odot q}$, for $N,q\in \N$, such that $F_N=\sum_{q\geq 1} I_q(g_{N,q})$.
Suppose that the following conditions
\begin{enumerate}[label={\normalfont(\roman*)}]
\item For fixed $q\geq 1$ there exists $\sigma^2_q\geq 0$ such that $q!\|g_{N,q}\|^2_{\mathfrak{H}^{\otimes q}} \stackrel{N \rightarrow \infty}{\longrightarrow }\sigma_q^2$,
\item $\sigma^2:=\sum_{q\geq 1} \sigma_q^2 <\infty$,
\item For all $q \geq 2$ and $r=1,\ldots,q-1$ we have $\|g_{N,q}\otimes_{r}g_{N,q}\|_{\mathfrak{H}^{\otimes(2q-2r)}}\stackrel{N \rightarrow \infty}{\longrightarrow } 0$,
\item $\lim_{Q\to \infty} \limsup_{N\to \infty} \sum_{q=Q+1}^\infty q! \|g_{N,q}\|^2_{\mathfrak{H}^{\otimes q}}=0$
\end{enumerate}
are true. Then $F_n \stackrel{\mathcal{D}}{\longrightarrow } \mathcal{N}(0,\sigma^2)$.
\end{theorem}
Before we verify the conditions, we need to prove the following auxiliary lemma, which will be needed for condition (ii) and (iv).
\begin{lemma}\label{lem:auxBounds}
There exists $c>0$ depending on the covariance of $X$, $d$ and $m$ such that
\begin{enumerate}[label={\normalfont(\roman*)}]
\item \label{lem:ub} $\sum_{n\in\N^D,|n|=q}c(n)^2 n! \leq c q^D$ for $q\geq 1$.
\item \label{lem:uniBound} $\sup\limits_{W\subset [0,1)^d \mathrm{\, convex}} \E\left[ (\sum_{q=1}^\infty \sum_{n\in \N^D, |n|=q}c(n)\int_{G(d,d-m)}\int_{W}\widetilde H_n(Y^F(t))dt \nu(dF))^2\right] \leq c.$
\end{enumerate}
\end{lemma}
\begin{proof}
In the following the constant $c>0$ may be changing from line to line.
Recall \eqref{def:coefc}
\begin{align*}
&c(n)=(2\pi)^{-(d-m)/2}\prod_{i=1}^{d-m}\frac{H_{n_i}(0)}{n_i!}\frac{(-1)^{d-m}}{\prod_{i=d-m+1}^Dn_i!} \\
&\quad\times\underbrace{\int_{\R^{D-d+m}}\det\big((\Lambda_2y)_{1,\ldots,\frac{(d-m)(d-m+1)}{2}} \big)\ind\{(\Lambda_2y)_{D-d+m}\geq u\}\prod_{i=d-m+1}^D H_{n_i}(y) \phi_{D-d+m}(y)\,dy.}_{:=Z(n)}
\end{align*}
Proposition 3 in \cite{paper:Imkeller} yields $\prod_{i=1}^{d-m}\frac{|H_{n_i}(0)|}{\sqrt{n_i!}}\leq c$, for a constant $c>0$, and thus
\begin{align*}
\left((2\pi)^{-(d-m)/2}\prod_{i=1}^{d-m}\frac{H_{n_i}(0)}{n_i!}\right)^2 \leq \frac{c}{ \prod_{i=1}^{d-m}n_i!}.
\end{align*}
By H\"older's inequality, we obtain
\begin{align*}
Z(n)^2&\leq \int_{\R^{D-d+m}}\det\big( (\Lambda_2y)_{1,\ldots,\frac{(d-m)(d-m+1)}{2}} \big)^2\ind\{(\Lambda_2y)_{D-d+m}\geq u\}\phi_{D-d+m}(y)\, dy\\
&\quad\times\int_{\R^{D-d+m}}\left(\prod_{i=d-m+1}^D H_{n_i}(y)\right)^2 \phi_{D-d+m}(y)\,dy\\
&=c\prod_{i=d-m+1}^D n_i!.
\end{align*}
The last two inequalities yield for $q\geq 1$
\begin{equation*}
\sum_{n\in \N^D,|n|=q}c(n)^2n! \leq c \sum_{n\in \N^D,|n|=q}1 \leq c \sum_{0\leq n_1,\ldots,n_D\leq q}1 \leq c(q+1)^D\leq cq^D,
\end{equation*}
which shows the first assertion. We now proof \ref{lem:uniBound}. By Theorem \ref{thm:L^2Approx}
\begin{align*}
\sum_{q=0}^\infty& \sum_{n\in \N^D, |n|=q}c(n)\int_{G(d,d-m)}\int_{W}\widetilde H_n(Y^F(t))\,dt\, \nu(dF) \\
&=\int_{A(d,d-m)} \#\{t\in W\cap F \mid X(t)\geq u, \nabla (X|_F)(t)=0,\iota_{-X,W\cap F}(t )\text{ even}\}\\
&\hspace{1.5cm} -\#\{t\in W\cap F \mid X(t)\geq u, \nabla (X|_F)(t)=0,\iota_{-X,W\cap F}(t )\text{ odd}\}\, \mu(dF),
\end{align*}
whose second moment is a upper bound for the expectation in \ref{lem:uniBound}. The latter can be bounded by
\begin{align*}
\int_{A(d,d-m)} \#\{t\in W\cap F \mid \nabla (X|_F)(t)=0\}\, \mu(dF).
\end{align*}
By Jensen's inequality and Fubini's theorem
\begin{align*}
&\E\left[ (\int_{A(d,d-m)} \#\{t\in W\cap F \mid \nabla (X|_F)(t)=0\}\, \mu(dF))^2 \right]\\
\quad&\leq \mu(\{F:F\cap W\neq \emptyset\}) \int_{A(d,d-m)}\E\left[\#\{t\in W\cap F \mid \nabla (X|_F)(t)=0\}^2 \right]\,\mu(dF).
\end{align*}
Now, by \eqref{eq:rice} and \eqref{eq:ricefac} we bound the integrand by
\begin{align*}
c\mathcal{H}^{d-m}(W^F_{v(F)}) + &\int_{W^F_{v(F)}-W^F_{v(F)}} \E\left[ |\det D^2_{v(F)}(X)(\rho_{v(F)}^F(t))\det D^2_{v(F)}(X)(\rho_{v(F)}^F(0))| \mid \mathcal{E}(F,t,0) \right]\\
&\times p_{\nabla_{v(F)} (X)(\rho_{v(F)}^F(t))\nabla_{v(F)} (X)(\rho_{v(F)}^F(0))}(0,0)\mathcal{H}^{d-m}(W^F_{v(F)}\cap (W^F_{v(F)}-t)) \,dt,
\end{align*}
where $c\geq0$ is a constant depending on $X$, $d$ and $m$. Taking $N:=d^{1/2}$ in Lemmata \ref{lem:boundDensity} and \ref{lem:boundExp} we obtain for the second summand the upper bound
\begin{align*}
c\int_{W^F_{v(F)}-W^F_{v(F)}} \|t\|^{-(d-m)+2}\mathcal{H}^{d-m}(W^F_{v(F)}\cap (W^F_{v(F)}-t))\,dt.
\end{align*}
Since $W^F_{v(F)}\subset B^{d-m}_{d^{1/2}}$ and $W^F_{v(F)}-W^F_{v(F)}\subset B^{d-m}_{d^{1/2}}$, we conclude
\begin{align*}
\E\left[\#\{t\in W\cap F \mid \nabla (X|_F)(t)=0\}^2 \right] \leq c\mathcal{H}^{d-m}(B^{d-m}_{d^{1/2}})\left(1+\int_{B^{d-m}_{d^{1/2}}}\|t\|^{-(d-m)+2}\,dt\right).
\end{align*}
Hence
\begin{align*}
\E&\left[ (\int_{A(d,d-m)} \#\{t\in W\cap F \mid \nabla (X|_F)(t)=0\}\, \mu(dF))^2 \right]\\
&\quad\leq c\mu(\{F:F\cap B^d_{d^{1/2}}\neq \emptyset\})^2\mathcal{H}^{d-m}(B^{d-m}_{d^{1/2}}) \left(1+\int_{B^{d-m}_{d^{1/2}}}\|t\|^{-(d-m)+2}\,dt\right).\qedhere
\end{align*}
\end{proof}
In the following we verify the conditions of Theorem \ref{thm:NourdinPeccati}.
{\textbf{Condition (i):}} We calculate the norm of $g_{N,q}$, cf. \eqref{def:gNq}, by an application of Fubini's theorem and obtain
\begin{align*}
&q!\| g_{N,q}\|^{2}_{\mathfrak{H}^{\otimes q}}\\
&= \int_{\R^{dq}}\frac{q!}{N^d\kappa_d}\sum_{k,l\in\{1,\ldots, D\}^q}b(k)b(l)\int_{G(d,d-m)}\int_{G(d,d-m)}\int_{B^d_N}\int_{B^d_N}\varphi^{F}_{t,k_1}\otimes \ldots \otimes \varphi^{F}_{t,k_q}(\lambda_1,\dots,\lambda_q)\\
&\quad\times\overline{\varphi^{F'}_{t',l_1}}\otimes \ldots \otimes \overline{\varphi^{F'}_{t',l_q}}(\lambda_1,\ldots,\lambda_q) \,dt\,dt'\,\nu(dF)\,\nu(dF')\,(f\lambda^d)^q(d(\lambda_1,\ldots,\lambda_q))\\
&=\frac{q!}{N^d\kappa_d}\sum_{k,l\in\{1,\ldots, D\}^q}b(k)b(l)\\
&\quad\times \int_{G(d,d-m)}\int_{G(d,d-m)}\int_{B^d_N}\int_{B^d_N}\prod_{i=1}^q \int_{\R^d}\varphi^{F}_{t,k_i}(\lambda)\overline{\varphi^{F'}_{t',l_i}}(\lambda)\,f \lambda^d(d\lambda) \,dt\, dt'\, \nu(dF)\,\nu(dF').
\end{align*}
Recalling $\int_{\R^d}\varphi^{F}_{t,k_i}(\lambda)\overline{\varphi^{F'}_{t',l_i}}(\lambda)\,f\lambda^d(d\lambda) =\E\left[Y^F_{k_i}(t)Y^{F'}_{l_i}(t')\right]$ in \eqref{eq:covY},
the above equals
\begin{align*}
\frac{q!}{N^d\kappa_d}&\sum_{k,l\in\{1,\ldots, D\}^q}b(k)b(l)\int_{G(d,d-m)}\int_{G(d,d-m)}\int_{B^d_N}\int_{B^d_N}\prod_{i=1}^q \E\left[Y^F_{k_i}(t)Y^{F'}_{l_i}(t')\right] \,dt \,dt' \,\nu(dF)\,\nu(dF').
\end{align*}
By stationarity in the first and Fubini's theorem in the second line
\begin{align*}
&\frac{1}{N^d\kappa_d}\int_{G(d,d-m)}\int_{G(d,d-m)}\int_{B^d_N}\int_{B^d_N} \prod_{i=1}^q\E\left[Y^F_{k_i}(t)Y^{F'}_{l_i}(t')\right] \,dt\, dt'\,\nu(dF)\,\nu(dF')\\
&=\frac{1}{N^d\kappa_d}\int_{G(d,d-m)}\int_{G(d,d-m)}\int_{B^d_N}\int_{B^d_N-t'} \prod_{i=1}^q\E\left[Y^F_{k_i}(t+t')Y^{F'}_{l_i}(t')\right] \,dt\, dt'\,\nu(dF)\,\nu(dF')\\
&=\int_{B^d_{2N}}\int_{G(d,d-m)}\int_{G(d,d-m)} \prod_{i=1}^q\E\left[Y^F_{k_i}(t)Y^{F'}_{l_i}(0)\right]\frac{\mathcal{H}^d((B^d_N-t)\cap B^d_N)}{N^d\kappa_d}\,\nu(dF)\,\nu(dF')\,dt.
\end{align*}
By the definition of $Y$, cf. \eqref{def:Y}, we have the following equality for the covariance matrix of $Y$ $\left(\E\left[Y_i^F(t)Y_j^{F'}(0)\right]\right)_{1\leq i,j\leq D}=\Lambda^{-1}\left(\E\left[\mathcal{X}^F_i(t)\mathcal{X}^{F'}_j(0)\right]\right)_{1\leq i,j\leq D}\Lambda^{-\top}$ and therefore by assumption \ref{as:covariance} there exists a constant $c\geq 0 $ such that
\begin{align*}
\sup_{1\leq i,j\leq D}\left|\E\left[Y_i^F(t)Y_j^{F'}(0)\right]\right| \leq c\widetilde\psi(t),
\end{align*}
which is an integrable upper bound. By the dominated convergence theorem
\begin{align}\label{eq:asymptoticVariance}
q!&\| g_{N,q}\|^{2}_{\mathfrak{H}^{\otimes q}}\notag\\
&=\frac{q!}{N^d\kappa_d}\sum_{k,l\in\{1,\ldots, D\}^q}b(k)b(l)\int_{B^d_{2N}}\int_{G(d,d-m)}\int_{G(d,d-m)}\mathcal{H}^d((B^d_N-t)\cap B^d_N)\notag\\
&\hspace{2cm}\times\prod_{i=1}^q \E\left[Y^F_{k_i}(t)Y^{F'}_{l_i}(0)\right] \,\nu(dF)\,\nu(dF')\,dt
\notag\\& \stackrel{N\to \infty}{\longrightarrow} q!\sum_{k,l\in\{1,\ldots, D\}^q}b(k)b(l)\int_{\R^d}\int_{G(d,d-m)}\int_{G(d,d-m)}\prod_{i=1}^q \E\left[Y^F_{k_i}(t)Y^{F'}_{l_i}(0)\right]\,\nu(dF)\,\nu(dF')\,dt,
\end{align}
where we define the limit as $\sigma^2_{m,q}$. Note that we implicitly used $\mathcal{H}^d((B^d_N-t)\cap B^d_N)/\mathcal{H}^d(B^d_N)\to 1$ for $N\to \infty$ and $t\in \R^d$. To see this consider the discussion following equation (3.22) in \cite{paper:HugLastSchulte}.
\textbf{Condition (ii):} We observe that
\begin{align*}
\sum_{q=1}^\infty &\lim_{N\to \infty} q!\|g_{N,q}\|^2= \sum_{q=1}^\infty\lim_{N\to \infty} \E\left[ I_q(g_{N,q})^2\right]
\end{align*}
by the It\^o isometry. Reversing some of the earlier manipulations the latter equals
\begin{align*}
\sum_{q=1}^\infty\lim_{N\to \infty}\frac{1}{\mathcal{H}^d(B^d_N)}\E\left[\left (\sum_{n\in\N^D,|n|=q}\int_{G(d,d-m)}c(n)\int_{B^d_N}\widetilde H_n(Y^F(t))\,dt\, \nu(dF)\right)^2\right].
\end{align*}
Fatou's Lemma and orthogonality yield the upper bound
\begin{align*}
\liminf_{N\to \infty}\frac{1}{\mathcal{H}^d(B^d_N)}\E\left[\left (\sum_{q=1}^\infty\sum_{n\in\N^D,|n|=q}\int_{G(d,d-m)}c(n)\int_{B^d_N}\widetilde H_n(Y^F(t))\,dt\, \nu(dF)\right)^2\right].
\end{align*}
Partitioning the space $\R^d$ into translates of the unit cube $[0,1)^d$ the latter without the limit inferior equals
\begin{align*}
\frac{1}{\mathcal{H}^d(B^d_N)}\sum_{z_1,z_2\in \Z^d\cap B^d_{N+d}} &\E\left[ \sum_{q=1}^\infty\sum_{|n|=q}\int_{G(d,d-m)}c(n)\int_{[0,1)^d+z_1}\ind\{t\in B^d_N\} \widetilde H_n( Y^F(t))\,dt\,\nu(dF)\right.\\
&\left.\times \sum_{q=1}^\infty\sum_{|n|=q}\int_{G(d,d-m)}c(n)\int_{[0,1)^d+z_2}\ind\{t\in B^d_N\} \widetilde H_n( Y^F(t))\,dt\,\nu(dF)\right].
\end{align*}
We define $\tau^{F,F'}(t):=\max\{\max_{i}\sum_{k=1}^D|\E\left[ Y^F_i(0)Y^{F'}_k(t)\right]|, \max_{k} \sum_{i=1}^D |\E\left[ Y^F_i(0)Y^{F'}_k(t)\right]| \}$ for $t\in \R^d$, $F,F'\in G(d,d-m)$, and note that due to \ref{as:covariance} there is a constant $c>0$ so that $\tau^{F,F'}(t)\leq c\psi(t)$. Moreover \ref{as:covariance} implies that for $\rho\in (0,1)$ and $\rho<1/c$ there is a constant $s>0$ such that
\begin{align*}
\psi(t)\leq \rho, \text{ for } |t|\geq s.
\end{align*}
Using $s$, we split the above summation into one over $I_1^N:=\{(z_1,z_2)\in (\Z^d\cap B^d_{N+d})^2\mid \|z_1-z_2\|_\infty \geq s+1 \}$ and $I_2^N:=\{(z_1,z_2)\in (\Z^d\cap B^d_{N+d})^2\mid \|z_1-z_2\|_\infty \leq s \}$. By Fubini's theorem, orthogonality and stationarity the first sum equals
\begin{align}\label{eq:sum1Arc}
&\sum_{(z_1,z_2)\in I_1^N} \sum_{q=1}^\infty\int_{G(d,d-m)}\int_{G(d,d-m)} \int_{(-1,1)^d}\E\left[ \sum_{|n|=q}c(n) \widetilde H_n(Y^{F}(t+z_1))\sum_{|n|=q}c(n)\widetilde H_n(Y^{F'}(z_2))\right]\notag\\
&\quad\quad\times \mathcal{H}^d([0,1)^d\cap[0,1)^d-t\cap B^d_N-z_2\cap B^d_N-t-z_1)\,dt\,\nu(dF)\,\nu(dF')\mathcal{H}^d(B^d_N)^{-1}.
\end{align}
Now, we use Lemma 1 in \cite{paper:Arcones}, which reads
\begin{lemma}[Arcones 1994]\label{lem:Arcones}
Let $V,W$ be two centered $d$-dimensional Gaussian random vectors such that $\E[V_iV_j]=\E[W_iW_j]=\delta_{ij}$ and let $h\colon \R^d\to \R$ have Hermite rank $r\in \N$ (i.e. $r=\inf\{k\in \N : \exists l_j \text{ such that }\sum_{j=1}^dl_j=k\text{ and }\E[(h(N)-\E[h(N)])\widetilde H_l(N)]\neq 0\}$ where $N\sim\mathcal{N}_d(0,I)$). Define $\tau:=\max\{\max_{1\leq j \leq d}\sum_{k=1}^d |\E[V_jW_k]|,\max_{1\leq k \leq d}\sum_{j=1}^d |\E[V_jW_k]|\}$, which is assumed to be less than 1. Then we have
\begin{align*}
|\E[(h(V)-\E[h(V)])(h(W)-\E[h(W)])]|\leq \tau ^r \E[h(V)^2].
\end{align*}
\end{lemma}
To apply this Lemma, we choose $V=Y^{F'}(z_2)$, $W=Y^F(t+z_1)$ and $h_q\colon \R^D\to \R$ given by $h_q(x):= \sum_{n\in \N^D, |n|=q} c(n) \widetilde H_n(x)$.
Then we have $r=q$ and $\tau^{F,F'} (t+z_1-z_2) \leq c \psi(t+z_1-z_2)<1$ for $t\in (-1,1)^d$ and $z_1,z_2\in I_1^N$. Moreover we have
\begin{align*}
\E\left[ h_q(Y^{F'}(z_2))^2\right]&=\sum_{n,n'\in\N^D,|n|=|n'|=q} c(n')c(n) \E[\widetilde H_n(Y^{F'}(0))\widetilde H_{n'}(Y^{F'}(0))]\\
&=\sum_{n,n'\in\N^D,|n|=|n'|=q} c(n')c(n) \prod_{i=1}^D\E[H_{n_i}(Y^{F'}_i(0))H_{n'_i}(Y^{F'}_i(0))]
\\&=\sum_{n\in\N^D,|n|=q} c(n)^2n!
\end{align*}
and for $q\geq 1$ we obtain $\E\left[ h_q(Y^{F'}(z_2))\right]=\E\left[ h_q(Y^{F}(t+z_1))\right]=0.$
Thus we bound \eqref{eq:sum1Arc} by
\begin{align*}
K\mathcal{H}^d(B^d_N)^{-1}\sum_{(z_1,z_2)\in I_1^N}\sum_{q=1}^\infty\int_{(-1,1)^d}c^q\psi(t+z_1-z_2)^q\,dt \sum_{n\in\N^D,|n|=q} c(n)^2n!,
\end{align*}
where $K\geq 0$ depends on $d$ and $m$. Lemma \ref{lem:auxBounds}\ref{lem:ub} and $\psi(t+z_1-z_2)^q\leq \rho^{q-1}\psi(t+z_1-z_2)$ yield
\begin{align*}
K/\rho\mathcal{H}^d(B^d_N)^{-1} \sum_{(z_1,z_2)\in I_1^N} \int_{(-1,1)^d}\psi(t+z_1-z_2)\,dt \sum_{q=1}^\infty q^D (c\rho)^q,
\end{align*}
as an upper bound, where $K\geq 0$ depends on the covariance of $X$, $d$ and $m$. By the estimate $\sum_{(z_1,z_2)\in I_1^N} \int_{(-1,1)^d}\psi(t+z_1-z_2)\,dt\leq 2^d\#\{\Z^d\cap B^d_{N+d}\}\int_{\R^d}\psi(t)\,dt$ we obtain the upper bound
\begin{align*}
\frac{2^dK}{\rho} \frac{\#\{\Z^d\cap B^d_{N+d}\}}{\mathcal{H}^d(B^d_N)}\int_{\R^d}\psi(t)\,dt \sum_{q=1}^\infty q^D(c\rho)^q.
\end{align*}
The latter is finite since $\liminf_{N\to \infty}\frac{\#\{\Z^d\cap B^d_{N+d}\}}{\mathcal{H}^d(B^d_N)}=1$ and the series converges.
We now analyse the sum over $I_2^N$ and start by using the inequality $ab\leq a^2+b^2$, $a,b\in \R$, to obtain the upper bound
\begin{align*}
\frac{2(2s+1)^d}{\mathcal{H}^d(B^d_N)}\sum_{z\in \Z^d\cap B^d_{N+d}}\hspace{-3pt}\E\left[ \left(\sum_{q=1}^\infty \sum_{|n|=q}\int_{G(d,d-m)}c(n)\int_{[0,1)^d+z}\ind\{t\in B^d_N\}\widetilde H_n(Y^F(t))\,dt \,\nu(dF)\right)^2\right].
\end{align*}
By stationarity the latter can be bounded by
\begin{align*}
2(2s+1)^d\frac{\#\{\Z^d\cap B^d_{N+d}\}}{\mathcal{H}^d(B^d_N)}\sup_{\substack{W\subset [0,1)^d\\ \mathrm{\, convex}}}\E\left[ \left(\sum_{q=1}^\infty \sum_{|n|=q}\int_{G(d,d-m)}c(n)\int_{W}\widetilde H_n(Y^F(t))\,dt \,\nu(dF)\right)^2\right],
\end{align*}
whose limit inferior is finite, since the supremum is finite by Lemma \ref{lem:auxBounds}\ref{lem:uniBound}.
\textbf{Condition (iii):}
The $r$-th contraction of $g_{N,q}$, cf. \eqref{def:gNq}, with itself is given by
\begin{align*}
&g_{N,q}\otimes_r g_{N,q}(a_1,\ldots,a_{2q-2r})\\
&=\int_{\R^{dr}}\frac{1}{N^d\kappa_d}\sum_{k\in\{1,\ldots,D\}^q}b(k)\int_{G(d,d-m)}\int_{B^d_N} \varphi^F_{t,k_1}(\lambda_1)\ldots\varphi^F_{t,k_r}(\lambda_r)
\\&\quad\times \varphi^F_{t,k_{r+1}}(a_1)\ldots \varphi^F_{t,k_q}(a_{q-r})\,dt \,\nu(dF) \sum_{l\in \{1,\ldots,D\}^q}b(l)\int_{G(d,d-m)}\int_{B^d_N} \overline{\varphi^{F'}_{t',l_1}}(\lambda_1)\ldots\overline{\varphi^{F'}_{t',l_r}}(\lambda_r)
\\&\quad\times\varphi^{F'}_{t',l_{r+1}}(a_{q-r+1})\ldots \varphi^{F'}_{t',l_q}(a_{2q-2r})\,dt' \,\nu(dF')\,(f\lambda^d)^r(d(\lambda_1,\ldots,\lambda_r)),
\end{align*}
for $a_1,\ldots, a_{2q-2r}\in \R^d$. By Fubini's theorem and \eqref{eq:covY} the above equals
\begin{align*}
\frac{1}{N^d\kappa_d}&\sum_{k,l\in \{1,\ldots,D\}^q} b(k)b(l)\int_{G(d,d-m)} \int_{G(d,d-m)}\int_{B^d_N} \int_{B^d_N} \prod_{i=1}^r \int_{\R^d}\varphi^F_{t,k_i}(\lambda)\overline{\varphi^{F'}_{t',l_i}}(\lambda)\,f\lambda^d(d\lambda)
\\&\quad\times \prod_{i=r+1}^q\varphi^F_{t,k_i}(a_{i-r})\varphi^{F'}_{t',l_i}(a_{q-2r+i})\,dt \,dt'\,\nu(dF)\,\nu(dF')
\\&=\frac{1}{N^d\kappa_d}\sum_{k,l\in \{1,\ldots,D\}^q} b(k)b(l)\int_{G(d,d-m)} \int_{G(d,d-m)}\int_{B^d_N} \int_{B^d_N} \prod_{i=1}^r \E[Y^F_{k_i}(t)Y^{F'}_{l_i}(t')]
\\&\quad\times \prod_{i=r+1}^q\varphi^F_{t,k_i}(a_{i-r})\varphi^{F'}_{t',l_i}(a_{q-2r+i})\,dt \,dt'\,\nu(dF)\,\nu(dF').
\end{align*}
Thus we obtain for the norm
\begin{align*}
&\|g_{N,q}\otimes_r g_{N,q}\|_{\mathfrak{H}^{\otimes(2q-2r)}}^2
\\&=\int_{\R^{d(2q-2r)}}\frac{1}{N^d\kappa_d}\sum_{k,l\in \{1,\ldots,D\}^q}b(k)b(l)\int_{G(d,d-m)} \int_{G(d,d-m)}\int_{B^d_N} \int_{B^d_N} \prod_{i=1}^r \E[Y^F_{k_i}(t)Y^{F'}_{l_i}(t')]
\\&\quad\times \prod_{i=r+1}^q\varphi^F_{t,k_i}(a_{i-r})\varphi^{F'}_{t',l_i}(a_{q-2r+i})\,dt \,dt'\,\nu(dF)\,\nu(dF')
\\&\quad\times \frac{1}{N^d\kappa_d}\sum_{k,l\in \{1,\ldots,D\}^q} b(k)b(l)\int_{G(d,d-m)} \int_{G(d,d-m)}\int_{B^d_N} \int_{B^d_N} \prod_{i=1}^r \E[Y^F_{k_i}(t)Y^{F'}_{l_i}(t')]
\\&\quad\times \prod_{i=r+1}^q\overline{\varphi^F_{t,k_i}}(a_{i-r})\overline{\varphi^{F'}_{t',l_i}}(a_{q-2r+i})\,dt \,dt'\,\nu(dF)\,\nu(dF')\,(f\lambda^d)^{2q-2r}(d(a_1,\ldots,a_{2q-2r})).
\end{align*}
And again Fubini's theorem yields equality to
\begin{align*}
\frac{1}{N^{2d}\kappa_d^2}&\sum_{k,l,k',l'\in \{1,\ldots,D\}^q}b(k)b(l)b(k')b(l')\idotsint_{(G(d,d-m))^4}\idotsint_{(B^d_N)^4}
\\&\quad\times \prod_{i=r+1}^q\int_{\R^d}\varphi^{F_1}_{t_1,k_i}(\lambda) \overline{\varphi^{F_3}_{t_3,k'_i}}(\lambda)\, f\lambda^d(d\lambda) \int_{\R^d}\varphi^{F_2}_{t_2,l_i}(\lambda)\overline{\varphi^{F_4}_{t_4,l'_i}} (\lambda)\,f\lambda^d(d\lambda)
\\&\quad\times\prod_{i=1}^r\E[Y^{F_1}_{k_i}(t_1)Y_{l_i}^{F_2}(t_2)] \E[Y^{F_3}_{k'_i}(t_3)Y_{l'_i}^{F_4}(t_4)] \,dt_1\ldots dt_4 \,\nu(dF_1)\ldots \nu(dF_4),
\end{align*}
which by \eqref{eq:covY} equals
\begin{align*}
\frac{1}{N^{2d}\kappa_d^2}&\sum_{k,l,k',l'\in \{1,\ldots,D\}^q}b(k)b(l)b(k')b(l')
\\&\quad\times\idotsint_{(G(d,d-m))^4}\idotsint_{(B^d_N)^4} \prod_{i=1}^r\E[Y^{F_1}_{k_i}(t_1)Y_{l_i}^{F_2}(t_2)]\E[Y^{F_3}_{k'_i}(t_3)Y_{l'_i}^{F_4}(t_4)]
\\&\quad\times\prod_{i=r+1}^q\E[Y^{F_1}_{k_i}(t_1)Y^{F_3}_{k'_i}(t_3)] \E[Y^{F_2}_{l_i}(t_2)Y^{F_4}_{l'_i}(t_4)]\,dt_1\ldots dt_4 \,\nu(dF_1)\ldots \nu(dF_4).
\end{align*}
By \ref{as:covariance} and stationarity there exists a constant $c>0$ such that for all $t\in \R^d$ and $F,F'\in G(d,d-m)$
\begin{align*}
\sup_{1\leq i, j \leq D}\left | \E\left[Y_i^{F}(t)Y_j^{F'}(s)\right]\right | \leq c \psi(t-s)
\end{align*}
and hence
\begin{align*}
\|g_{N,q}&\otimes_r g_{N,q}\|_{\mathfrak{H}^{\otimes(2q-2r)}}^2
\\&\leq \frac{c^{2q}}{N^{2d}\kappa_d^2}\sum_{k,l,k',l'\in \{1,\ldots,D\}^q}b(k)b(l)b(k')b(l') \idotsint_{(G(d,d-m))^4}\idotsint_{(B^d_N)^4} \psi(t_1-t_2)^r
\\&\quad\times\psi(t_3-t_4)^{r}\psi(t_1-t_3)^{q-r}\psi(t_2-t_4)^{q-r}\,dt_1\ldots dt_4 \,\nu(dF_1)\ldots \nu(dF_4)
\\&=c^{2q}\sum_{k,l,k',l'\in \{1,\ldots,D\}^q}b(k)b(l)b(k')b(l')\\
&\quad\times\underbrace{\frac{1}{N^{2d}\kappa_d^2}\idotsint_{(B^d_N)^4} \psi(t_1-t_2)^r\psi(t_3-t_4)^{r}\psi(t_1-t_3)^{q-r}\psi(t_2-t_4)^{q-r}\,dt_1\ldots dt_4 }_{:=Z(N)}.
\end{align*}
Using the inequality $a^rb^{q-r}\leq a^q+b^q$ for $a=\psi(t_3-t_4)$ and $b=\psi(t_1-t_3)$, we obtain
\begin{align*}
Z(N)&\leq \frac{1}{N^{2d}\kappa_d^2}\idotsint_{(B^d_N)^4} \psi(t_1-t_2)^r\psi(t_3-t_4)^q\psi(t_2-t_4)^{q-r}\,dt_1\ldots dt_4
\\&\quad +\frac{1}{N^{2d}\kappa_d^2}\idotsint_{(B^d_N)^4} \psi(t_1-t_2)^r\psi(t_1-t_3)^q\psi(t_2-t_4)^{q-r}\,dt_1\ldots dt_4.
\end{align*}
By \ref{as:covariance} we have $\infty > c_n:= \int_{\R^d}\psi^n(x)\,dx \geq \int_{B^d_N}\psi^n(x)\,dx$, for $n \in \N$, and therefore obtain the following upper bound for the first summand
\begin{align*}
\frac{1}{N^{2d}\kappa_d^2}&\idotsint_{(B^d_N)^3} \psi(t_1-t_2)^r\psi(t_2-t_4)^{q-r}\int_{\R^d}\psi(t_3-t_4)^q \,dt_3\,dt_1\,dt_2\, dt_4
\\&\leq\frac{c_q}{N^{2d}\kappa_d^2}\int_{B^d_N}\int_{B^d_N}\psi(t_1-t_2)^r \int_{\R^d}\psi(t_2-t_4)^{q-r} \,dt_4 \,dt_1\,dt_2.
\end{align*}
Repeating this argument yields the upper bound
\begin{align*}
\frac{c_qc_{q-r}}{N^{2d}\kappa_d^2}\int_{B^d_N}\int_{\R^d}\psi(t_1-t_2)^r \,dt_1\,dt_2=\frac{c_qc_{q-r}c_r}{N^{2d}\kappa_d^2} N^d\kappa_d.
\end{align*}
Note that $\frac{c_qc_{q-r}c_r}{N^{d}\kappa_d} \stackrel{N\to \infty}{\longrightarrow} 0$. Proceeding analogously for the second summand yields
\begin{align*}
\|g_{N,q}&\otimes_r g_{N,q}\|_{\mathfrak{H}^{\otimes(2q-2r)}}^2 \leq c^{2q}\sum_{k,l,k',l'\in \{1,\ldots,D\}^q}b(k)b(l)b(k')b(l') Z(N) \stackrel{N\to \infty}{\longrightarrow} 0.
\end{align*}
{\textbf{Condition (iv):}} By orthogonality
\begin{align*}
\sum_{q=Q+1}^\infty q!\|g_{N,q}\|^2 = \mathcal{H}^d(B^d_N)^{-1}\E\left[ \left(\sum_{q=Q+1}^\infty \sum_{|n|=q}\int_{G(d,d-m)}c(n)\int_{B^d_N}\widetilde H_n(Y^F(t))\,dt \,\nu(dF)\right)^2\right].
\end{align*}
The same computations as in the verification of condition (ii) bound the latter by
\begin{align*}
&\frac{2^dK}{\rho} \frac{\#\{\Z^d\cap B^d_{N+d}\}}{\mathcal{H}^d(B^d_N)}\int_{\R^d}\psi(t)dt \sum_{q=Q+1}^\infty q^D(c\rho)^q + \frac{2(2s+1)^d}{\mathcal{H}^d(B^d_N)}\\
&\quad\times\sum_{z\in \Z^d\cap B^d_{N+d}}\E\left[ \left(\sum_{q=Q+1}^\infty \sum_{|n|=q}\int_{G(d,d-m)}c(n)\int_{[0,1)^d}\ind\{t+z\in B^d_N\}\widetilde H_n(Y^F(t))\,dt \,\nu(dF)\right)^2\right].
\end{align*}
In the limit $N\to\infty$ and then $Q\to \infty$ the first summand vanishes, since the series is the tail of a convergent series. The second summand needs more attention. We first split the summation into one over the indices $I_1^N:=\{z\in \Z^d\cap B^d_{N+d}\mid z+[0,1)^d\subset B^d_N\}$ and one over $I_2^N:=\{z\in \Z^d\cap B^d_{N+d}\mid z+[0,1)^d\cap (B^d_N)^\mathsf{c}\neq \emptyset\}$. The sum over $I_1^N$ can be bounded by
\begin{align*}
2(2s+1)^d\E\left[ \left(\sum_{q=Q+1}^\infty \sum_{|n|=q}\int_{G(d,d-m)}c(n)\int_{[0,1)^d}\widetilde H_n(Y^F(t))\,dt \,\nu(dF)\right)^2\right]
\end{align*}
since $\#I_1^N\mathcal{H}^d(B^d_N)^{-1}\leq 1$. Hence the latter tends to zero for $Q\to \infty$ by Lemma \ref{thm:L^2Approx}. We bound the summation over $I_2^N$ by
\begin{align*}
\frac{2(2s+1)^d\#I_2^N}{\mathcal{H}^d(B^d_N)} \sup_{W \subset [0,1)^d\mathrm{\, convex}}\E\left[ \left(\sum_{q=1}^\infty \sum_{|n|=q}\int_{G(d,d-m)}c(n)\int_{W}\widetilde H_n(Y^F(t))\,dt \,\nu(dF)\right)^2\right].
\end{align*}
Lemma \ref{lem:auxBounds}\ref{lem:uniBound} yields the upper bound
\begin{align*}
c2(2s+1)^d \#I_2^N \mathcal{H}^d(B^d_N)^{-1},
\end{align*}
which vanishes for $N\to \infty$. This shows the assertion.
\section{A lower bound for the asymptotic variance}
We follow the lines of \cite [Lemma 2.2]{paper:EstradeLeon} and give a lower bound for the asymptotic variance.
\begin{lemma}\label{lem:lowerBound}
Let $X$ be a real Gaussian field on $\R^d$, which satisfies the assumptions (A1)--(A3). We then have for $m\in \{0,\ldots,d-1\}$ and $\sigma_m^2$ given as in Theorem \ref{thm:mainTheorem} that
\begin{align*}
\sigma^2_m \geq {d \brack d-m }^2(2\pi)^mf(0)H_{d-m}(u)^2\phi(u)^2.
\end{align*}
\end{lemma}
\begin{proof}
Recall that according to Theorem \ref{thm:NourdinPeccati} the asymptotic variance is given by $\sum_{q\geq 1}\sigma^2_{m,q}$, where $\sigma_{m,q}^2$ is defined as the limit in condition (i) of that theorem. Hence, we obtain a lower bound for the asymptotic variance by computing $\sigma_{m,1}^2$. By (\ref{eq:asymptoticVariance})
\begin{align*}
\sigma_{m,1}^2=\sum_{k,l\in\{1,\ldots, D\}}b(k)b(l)\int_{\R^d}\int_{G(d,d-m)}\int_{G(d,d-m)} \E\left[Y^F_{k}(t)Y^{F'}_{l}(0)\right]\nu(dF)\nu(dF')\,dt,
\end{align*}
where the coefficients $b(\cdot)$ are given by
\begin{align*}
b(k)=\sum_{n \in \N^D,|n|=1} \ind\{k\in \A_{n}\} \frac{c(n)}{|\A_{n}|}.
\end{align*}
The sets $\A_{n}$ consist of only one element, namely the number of the component of $n$, which contains the $1$. Thus if we write $e_{i}\in \R^{D}$ for the vector, whose components are $0$ except for the $i$-th component, which is $1$, we obtain
\begin{align*}
b(k)=c(e_k).
\end{align*}
By the definition of the coefficients $c(\cdot)$, cf. (\ref{def:coefc}), we see that $c(e_k)=0$ for $k=1,\ldots,d-m$ and therefore obtain
\begin{align*}
\sigma_{m,1}^2=\sum_{k,l=d-m}^D c(e_k)c(e_l)\int_{\R^d}\int_{G(d,d-m)}\int_{G(d,d-m)} \E\left[Y^F_{k}(t)Y^{F'}_{l}(0)\right]\,\nu(dF)\,\nu(dF')\,dt.
\end{align*}
We now show that
\begin{align*}
\int_{\R^d}\E\left[\mathcal{X}^F_{k}(t)\mathcal{X}^{F'}_{l}(0)\right] \,dt = (2\pi)^d f(0)\delta_{k,l}(D,D),
\end{align*}
for $F,F'\in A(d,d-m)$ and $k,l=1,\ldots, D$.
Consider the case $(k,l)=(D,D)$. Then the equality
$\E\left[\mathcal{X}^F_{D}(t)\mathcal{X}^{F'}_{D}(0)\right] = \E\left[X(t)X(0)\right] = (2\pi)^{d/2}\mathcal{F}(f)(t)$ holds,
where $\mathcal{F}$ denotes the Fourier transformation. By \ref{as:covariance} the spectral density $f$ is continuous and $\E\left[X(t)X(0)\right]$ is integrable, which yields that $\int_{\R^d}\E\left[X(t)X(0)\right] dt=(2\pi)^df(0)$, via the Fourier cotransformation. In the cases, where $(k,l)\neq (D,D)$, at least one of the factors $\mathcal{X}^F_{D}$ or $\mathcal{X}^{F'}_{D}$ is at least one directional derivative of the field $X$, say in direction $u\in S^{d-1}$. This yields that $\E\left[\mathcal{X}^F_{k}(t)\mathcal{X}^{F'}_{l}(0)\right]$ equals, up to a power of $-1$, the function $\frac{\partial}{\partial u} g$, where $g$ is either the covariance function or a derivative of it. Thus by Fubini's theorem, we conclude that
\begin{align*}
\int_{\R^d}\E\left[\mathcal{X}^F_{k}(t)\mathcal{X}^{F'}_{l}(0)\right] dt &= \int_{\R}\ldots \int_{\R} \frac{\partial}{\partial u} g(t_1,\ldots,t_d) \,dt_1 \ldots dt_d \\
&=\sum_{i=1}^d u^{(i)} \int_{\R}\ldots \int_{\R} \frac{\partial}{\partial t_i} g(t_1,\ldots,t_d) \,dt_i\, dt_{1} \ldots\overline{dt_i}\ldots dt_d \\
&= \sum_{i=1}^d u^{(i)} \int_{\R}\ldots \int_{\R} \left .g(t_1,\ldots,t_d)\right |_{t_i=-\infty}^\infty \,dt_{1} \ldots\overline{dt_i}\ldots dt_d\\
&=0,
\end{align*}
where we used in the last line assumption \ref{as:covariance}, that is, the covariance function and its derivatives tend to 0 for $\|t\|\to \infty$.
The definition of $Y$, cf. \eqref{def:Y}, implies
\begin{align*}
\E\left[Y^F_{k}(t)Y^{F'}_{l}(0)\right]= \sum_{r=1}^D\sum_{s=1}^D\Lambda_{l,r}^{-1}\Lambda_{k,s}^{-1} \E\left[\mathcal{X}^F_{r}(t)\mathcal{X}^{F'}_{s}(0)\right],
\end{align*}
which yields
\begin{align*}
\int_{\R^d} \E\left[Y^F_{k}(t)Y^{F'}_{l}(0)\right]\, dt = \Lambda_{l,D}^{-1}\Lambda_{k,D}^{-1} (2\pi)^df(0)
\end{align*}
and we conclude with Fubini's theorem, that
\begin{align*}
\sigma_{m,1}^2={d \brack d-m }^2\sum_{k,l=d-m}^D c(e_k)c(e_l)\Lambda_{l,D}^{-1}\Lambda_{k,D}^{-1} (2\pi)^df(0) = {d \brack d-m }^2c(e_D)^2(\Lambda^{-1}_{D,D})^2(2\pi)^df(0),
\end{align*}
where the last equality holds since $\Lambda$ is lower triangular.
In order to calculate the coefficients $c(e_D)$, we have to analyse the covariance structure of $\mathcal{X}^F$. We first write the $K+1:=(d-m)(d-m+1)/2+1$ last coordinates of $\mathcal{X}^F$ in the order
\begin{align*}
\left( \left (\frac{\partial^2}{\partial v_i\partial v_j} X \right)_{1\leq i < j\leq d-m}, \left(\frac{\partial^2}{\partial v_i\partial v_i} X\right)_{i=1}^{d-m} , X \right).
\end{align*}
Thus, using stationarity, isotropy and $\Co^X(0)=1$, the covariance matrix of this vector at $0$ is given by
\begin{align*}
\begin{pmatrix}
\Co\left(\left(\frac{\partial^2}{\partial t_i\partial t_j} X\right )_{ i < j}\right) & \Co\left(\left (\frac{\partial^2}{\partial t_i\partial t_j} X \right)_{i< j}, \left(\frac{\partial^2}{\partial t_i\partial t_i} X\right)_{i=1}^{d-m}\right) & 0 \\
\Co\left(\left(\frac{\partial^2}{\partial t_i\partial t_i} X\right)_{i=1}^{d-m},\left (\frac{\partial^2}{\partial t_i\partial t_j} X \right)_{i < j} \right)& \Co\left(\left(\frac{\partial^2}{\partial t_i\partial t_i} X\right)_{i=1}^{d-m}\right) & -1 \\
0 & -1 & 1 \\
\end{pmatrix},
\end{align*}
which equals the product $\Lambda_2\Lambda_2^\top$, where $\Lambda_2\in\R^{K+1}$ is a lower triangular matrix, given in Lemma \ref{lem:decom}. We choose the matrix $L\in \R^{K\times K}$, the vector $l\in \R^{K}$ and $\alpha>0$ such that
\begin{align*}
\Lambda_2=\begin{pmatrix}
L & 0 \\
l^\top& \alpha \\
\end{pmatrix}.
\end{align*}
Then the relation $\|l\|^2+\alpha^2 =1$ holds as well as $Ll=( 0_{1\times K}, -1_{1\times d-m}).$ With this specific representation of $\Lambda_2$, we have
\begin{align*}
c(e_D)&=(2\pi)^{-(d-m)/2}(-1)^{d-m}\int_{\R^{K}\times \R} \det(Ly)\ind\{\langle l,y\rangle +\alpha z \geq u\}z\phi_{K}(y)\phi(z) \,d(y,z)\\
&=-(2\pi)^{-(d-m)/2}(-1)^{d-m}\int_{\R^{K}\times \R} \det(Ly)\ind\{\langle l,y\rangle +\alpha z \geq u\}\phi_{K}(y)\phi'(z) \,d(y,z)\\
&=(2\pi)^{-(d-m)/2}(-1)^{d-m}\int_{\R^{K}} \det(Ly)\phi_{K}(y)\phi\left(\alpha^{-1}(u-\langle l,y\rangle)\right) \,dy,
\end{align*}
where we used that $z\phi(z)=-\phi'(z)$ in the first line and Fubini's theorem in the second. Using the Hermite expansion of $y\mapsto \det(Ly)$ given in \cite [Lemma A.2]{paper:EstradeLeon}, we obtain
\begin{align*}
c(e_D)&=(2\pi)^{-(d-m)/2}(-1)^{d-m} \sum_{\substack{m\in \N^{K},|m|=d-m}}\beta_m\int_{\R^K}\widetilde{H}_m(y)\phi_{K}(y)\phi\left(\alpha^{-1}(u-\langle l,y\rangle)\right) \,dy\\
&=(2\pi)^{-(d-m)/2} \sum_{\substack{m\in \N^{K},|m|=d-m}}\beta_m \int_{\R^K}D^m\phi_{K}(y)\phi\left(\alpha^{-1}(u-\langle l,y\rangle)\right) \,dy,
\end{align*}
where $D^m\phi$ denotes $\frac{\partial^{|m|}}{\partial t_{m_1}\ldots \partial t_{m_{K}}}\phi$ and $\beta_m$ are real coefficients. Following the argument in \cite{paper:EstradeLeon}, we define $h\colon \R^K\to \R, x\mapsto \phi(\alpha^{-1}\langle l,y\rangle)$ and choose $l^\prime$ such that $\langle l , l^\prime\rangle=1$. We then obtain
\begin{align*}
\int_{\R^K}D^m\phi_{K}(y)\phi\left(\alpha^{-1}(u-\langle l,y\rangle)\right) \,dy = (h\ast D^m\phi_K)(ul^\prime)=D^m(h \ast \phi_K)(ul^\prime).
\end{align*}
By \cite [Remark A.4]{paper:EstradeLeon}, which reads $(h \ast \phi_K)(y)= \alpha\phi(\langle l,y\rangle)$ for $y\in \R^K$, we obtain
\begin{align*}
D^m(h \ast \phi_K)(y)=\alpha l^{(m)} \phi^{(d-m)}(\langle l , y\rangle)=(-1)^{d-m}\alpha l^{(m)} H_{d-m}(\langle l,y\rangle)\phi(\langle l,y\rangle).
\end{align*}
Thus by \cite [Lemma A.2]{paper:EstradeLeon} in the second equality
\begin{align*}
c(e_D)&= (2\pi)^{-(d-m)/2} \sum_{\substack{m\in \N^{K}\\|m|=d-m}}\beta_m l^{(m)} (-1)^{d-m}\alpha H_{d-m}(u)\phi(u) \\
&= (2\pi)^{-(d-m)/2}\det(Ll)(-1)^{d-m}\alpha H_{d-m}(u)\phi(u).
\end{align*}
Note that the $K$-dimensional vector $Ll$ corresponds to the symmetric $(d-m)\times (d-m)$\nobreakdash-\hspace{0pt}matrix, whose nondiagonal entries are given by the first $(d-m)(d-m-1)/2$ entries of $Ll$ and whose diagonal is given by the $d-m$ last entries of $Ll$, thus $\det(Ll)=(-1)^{d-m}$. Hence we obtain
\begin{align*}
c(e_D)=(2\pi)^{-(d-m)/2}\alpha H_{d-m}(u)\phi(u)
\end{align*}
and therefore conclude as asserted
\begin{equation*}
\sigma_{m,1}^2={d \brack d-m }^2(2\pi)^{m}f(0)H_{d-m}(u)^2\phi(u)^2.\qedhere
\end{equation*}
\end{proof}
|
\section{Experimental}
The optical properties of AFM tips are studied using a custom-built confocal microscope with a supercontinuum laser source for dark-field scattering spectroscopy (Fig.~\ref{fig:simple_optics_layout}). Both illumination and collection share the optical axis of a 0.8\,NA IR objective. Supercontinuum laser light is filtered into a ring and incident on a tip at 0.6--0.8\,NA while light scattered by the tip is confocally collected from the central laser focus using an iris to restrict the collection NA below 0.6. Broadband polarising beamsplitters are used to simultaneously measure spectra which are linearly polarised both along the tip axis (axial) and perpendicular to the tip axis (transverse).
A scanning hyperspectral imaging technique is applied to determine the local optical response at the tip apex. Tips are raster scanned under the laser spot and the dark field scattering from the confocal sampling volume measured at each point, forming a hyperspectral data cube. Images are formed at each wavelength contained in the cube, with each image pixel digitised into 1044 wavelengths between 400--1200\,nm. Measured spectra are normalised to a spectrum of flat metal of the same material to show only structural effects. Image slices at individual wavelengths or wavelength bands are then readily constructed to display localised spectral features.
Fast image acquisition is made possible by the high brightness supercontinuum laser source (100 $\mu$W.$\mu$m$^{-2}$) and cooled benchtop spectrometers, enabling 10\,ms integration times (5 mins per image).
Within plasmonics, this approach to hyperspectral imaging has been used to identify distributed plasmon modes in aggregated AuNP colloids \cite{herrmann2013} and to image SPPs \cite{bashevoy2007hyperspectral} but has yet to be applied to tips. By using this technique, radiative plasmons can be spatially identified with a resolution around 250\,nm.
\begin{figure}[t]
\centering
\includegraphics[width=0.98\columnwidth]{tip_sems}
\caption{SEM images of (a) sharp Au AFM tip, (b) Au-coated spherical AFM tip (Nanotools), and (c) electrochemically-deposited AuNP-on-Pt AFM tip.}
\label{fig:tip_sems}
\end{figure}
To investigate the radiative plasmonic properties of nanostructured tips, hyperspectral images are taken of both standard (sharp) and spherical-tipped Au AFM tips. Spherical tips are either 300\,nm diameter, 50\,nm Au-coated NanoTools B150 AFM probes or electrochemically-deposited AuNP-on-Pt AFM probes, fabricated in-house \cite{sanders2014} (shown in Fig.~\ref{fig:tip_sems}). Fabricated tips are pre-treated where possible prior to use with ambient air plasma and/or piranha solution to remove organic surface residue and, in some cases, smooth out surface roughness.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\columnwidth]{hs_images_and_tip_spectra.pdf}
\caption{
Hyperspectral images of (a) sharp Au tip, (b) Au-coated spherical tip (Nanotools), and (c) electrochemically-deposited AuNP-on-Pt tip. Collected light is polarised along tip axis, colour maps all have same normalisation. Scale bar is $600\,$nm.
(d,e) Scattering spectra of both sharp and spherical metal tips, extracted from hyperspectral images around the apex region, in (d) axial and (e) transverse polarisations.
(f) Integrated SERS background from sharp and spherical Au tips. Scattering spectrum of spherical Au tip apex shown shaded.
}
\label{fig:hyperspectral_images}
\end{figure}
\begin{figure*}
\centering
\includegraphics{numerical_simulations}
\vspace{-5pt}
\caption{(a) Numerically simulated near-field apex spectra of spherical Au and AuNP-on-Pt tips with (b,c) near-field maps of the main resonance in each, as highlighted by circles in (a). Simulated tips have a 300\,nm spherical radii, 120\,nm neck widths, \SI{20}{\degree} opening angles and \SI{1.88}{\micro\metre} lengths to best match typical experimental tip geometries and avoid truncation artefacts. Tips are illuminated by plane waves orientated along the tip axis. (d) Interpolated field enhancement map with superimposed resonant wavelengths, as the neck width varies from a spherical to a sharp tip. Tips have a 250\,nm apex diameter, \SI{1.88}{\micro\metre} length, and \SI{10}{\degree} opening angle.}
\label{fig:numerical_simulations}
\end{figure*}
Comparisons between spherical- and sharp-tipped Au probes using hyperspectral image slices (Fig.~\ref{fig:hyperspectral_images}) shows that spherical tips exhibit a characteristic red (600--700\,nm) scatter, separated from the bulk tip. No similar localised scattering is seen in the visible spectrum with sharp Au tips, which have a ten-fold weaker optical response and appear similar to non-plasmonic Pt tips. This delocalised apex scatter can also be directly seen in dark-field microscopy images (Fig.~\ref{fig:simple_optics_layout}b).
The AuNP-on-Pt structure behaves very similarly to the Au-coated spherical tip (which has diamond-like-carbon inside), likely because the 50\,nm coating thickness is greater than the skin depth \cite{stockman2011, huber2014}.
As we show below, differences in plasmon resonances arise due to the Au-Pt and Au-Au neck boundaries.
Integrating spectra around each tip better shows the 600--700\,nm scattering resonance from spherical Au tips (Fig.~\ref{fig:hyperspectral_images}d,e), which are reliably present in all spherical-tipped AFM probes, both vacuum-processed and electrochemically deposited. We attribute these to localised surface plasmon excitation, while electron microscopy confirms this resonance correlates only with spherical Au tip shapes. The response of sharp Au tips shows no similar plasmonic features, while the slow rise in scattering towards the NIR is consistent with lightning rod scattering \cite{zhang2009}.
Broadband tuneable SERS measurements \cite{lombardi2016} confirm that the optical scattering resonance seen in spherical Au tips is indeed caused by radiative plasmon excitation. The trapped plasmon fields enhance optical processes on the surface such as surface-enhanced Raman scattering (SERS) and here we use the SERS background \cite{lombardi2016,hugall2015} as a reporter of the plasmonic near-field strength.
SERS background spectra are integrated across a range of excitation wavelengths between 500 and 700\,nm, spaced 10\,nm apart, to extract any scattering resonances. The resulting spectrum (Fig.~\ref{fig:hyperspectral_images}f) shows a distinct peak around the spherical Au tip scattering resonance, while no such resonance is seen for sharp Au tips.
Further confirmation stems from direct observation of plasmon coupling between spherical tips, as has been previously reported \cite{savage2012}.
Plasmon resonances in spherical AuNP tips correspond to \emph{radiative} antenna-like modes, similar to those in plasmonic nanoparticles, that efficiently couple far-field light into strong collective free electron oscillations without the need for SPP momentum matching. As with nanoparticles, the signature of these plasmons is an optical resonance indicating their large dipole moment (Fig.~\ref{fig:hyperspectral_images}d). Such radiative plasmons only form if multipolar surface charge oscillations are supported, requiring a structure with multiple metal-dielectric interfaces. Since spherical metallic tips possess a neck behind the tip, they can support NP plasmonics. Sharp tips do not have this back surface, hence cannot support radiative plasmon resonances, although the single metal-dielectric surface supports launching of evanescent SPPs and a strong lightning rod component.
Simulated near-field spectra (using the boundary element method) around the apex of 300\,nm spherical Au and AuNP-on-Pt tips with 120\,nm neck diameters ($d_{\mathrm{neck}}=0.4d_{\mathrm{sphere}}$) are shown in Fig.~\ref{fig:numerical_simulations}a. Tips are simulated with a length of \SI{1.88}{\micro\metre} to avoid truncation artefacts which are commonly seen in tip simulations and erroneously suggest plasmonic performance even in sharp tips. Strong modes appear along the tip axis for all spherical tips between 550--700\,nm, as in experiments with peak wavelengths that match our hyperspectral results.
Near-field maps corresponding to the main resonance in each tip (Fig.~\ref{fig:numerical_simulations}b,c) show dipole-like resonances with the neck spatially splitting the underside of each mode, mixing it with quadrupolar modes and shifting it towards the blue.
In order to directly compare the \emph{plasmonic} behaviour of spherical and sharp Au tips independent of lightning rod contributions, the neck width is incrementally increased. This allows us to study structures which smoothly transition from a nanoparticle attached to the apex of a sharp Au tip, into a rounded tip geometry, without the apex radius ever changing. The field enhancement and peak positions extracted from this morphology transition (Fig.~\ref{fig:numerical_simulations}d) show resonances insensitive to the neck width until $d_{\mathrm{neck}}>0.8d_{\mathrm{sphere}}$, explaining the robustness of observed spherical tip plasmons between different tip morphologies. However a steady decrease in the field enhancement is observed once $d_{\mathrm{neck}}>0.4d_{\mathrm{sphere}}$, decreasing faster once $d_{\mathrm{neck}}>0.8d_{\mathrm{sphere}}$. This supports the claim that sharp tips cannot sustain antenna-like plasmons and that the majority of enhancement is from lightning rod effects. We note that the lateral spatial localisation of the field approaches $0.3 d_{\mathrm{sphere}}$ independent of this neck diameter.
These results demonstrate the importance of considering which plasmons might exist in a particular experiment and nanostructure geometry, and that it is vital to characterise nanostructures prior to their application.
Apex nanostructuring can controllably introduce radiative plasmons into the tip geometry, lifting the evanescent illumination restriction of sharp tips and permitting use of a wider range of microscope configurations. While the lightning rod effect will always contribute to the field enhancement and favour sharp tips, exploiting resonant plasmonic enhancement in a carefully optimised spherical tip can further improve the near-field enhancement. The spherical tip geometry and materials shown here are optimised for use with the typically-used 633\,nm laser wavelengths.
Demonstrated interactions between spherical tip plasmons \cite{savage2012} also suggests coupling with an image charge in a planar surface is possible and could be used in nanometric tip-surface gaps to further localise the field on resonance with near infrared lasers. Exploiting radiative tip plasmons in this manner bridges the gap between SERS and conventional TERS, forming a spatially-mappable version of the highly successful nanoparticle-on-mirror geometry \cite{mertens2013, taylor2014}. These systems repeatedly produce Raman enhancements of up to \num{e7} with nanometric mode volumes, much like tips, and demonstrate that plasmonic gaps can exhibit comparatively large field enhancements without relying only on the lightning rod effect.
Secondly, without prior knowledge of the tip-system spectral response it is difficult to properly interpret any measurements, such as TERS spectra. Improved tip characterisation is crucial to understanding variations in TERS spectra. Standard, wide-field microscopy/spectroscopy is not a particularly effective tool for optically characterising tips. Instead, confocal hyperspectral imaging provides a viable method for mapping the local scattering response while broadband tuneable SERS offers a unique way of optically characterising the near-field. Incorporating these techniques into existing microscopes is relatively simple and will greatly improve the reliability of tip-based near-field microscopy.
\begin{acknowledgments}
The authors thank EPSRC grants EP/G060649/1 and EP/L027151/1, and ERC grant LINASS 320503 for funding and NanoTools for their services providing Au-coated spherical AFM tips. RWB thanks Queens' College and the Royal Commission for the Exhibition of 1851 for financial support.
\end{acknowledgments}
|
\section{Introduction }
\label{sec:intro}
Soon after the discovery of QCD \cite{Fritzsch:1973pi}, and following
the realization that QCD exhibits asymptotic freedom \cite{Gross:1973id,Politzer:1973fx},
is was recognized that QCD predicts a new high temperature phase of
weakly interacting quarks and gluons, termed the Quark Gluon Plasma
\cite{Collins:1974ky,Cabibbo:1975ig,Shuryak:1977ut}. The existence
of a new phase was confirmed in the first calculations using the lattice
formulation of QCD, initially for pure $SU(2)$ gauge theory \cite{Creutz:1980zw,Creutz:1979dw}.
Over the years, as Lattice QCD methods have been refined to allow
for continuum extrapolated calculations with dynamical quarks at the
physical masses, it has been found that the transition from hadrons
to partons at vanishing net baryon density is an analytic cross over
\cite{Aoki:2006we}. At the same time many model calculations suggested
that at vanishing temperature but large baryon density there might
be a first order transition \cite{Stephanov:2004wx}. This first order
phase transition will end at a critical point, the location of which
is not really constrained by any model calculations let alone Lattice
QCD, which, due to the fermion sign problem can only explore regions
of small net-baryon chemical potential, $\mu_{B}/T\lesssim1$.
The potential presence of a first order phase co-existence region
together with a critical point has motivated a dedicated experimental
program at RHIC, the so-called RHIC Beam Energy Scan (BES). Experimentally,
regions of higher baryon density can be reached by lowering the beam
energy where some of the projectile and target baryons are stopped
at mid-rapidity. The study of fluctuations play an important role
in the quest to experimentally explore the QCD phase diagram. Both,
the second order phase transition associated with a critical point
and the first order transition give rise to characteristic fluctuation
pattern. Of course the system produced in a heavy ion collision is
of finite size and evolves in time which smoothens the singular structures
associated with phase transitions. However, fluctuation measurements
are still helpful in this case, because, as we shall discuss below,
fluctuations are related to derivatives of the free energy. For example
cumulants of the baryon number are given by derivatives with respect to
the baryon chemical potential, $\mu_{B}$, etc. Therefore, the measurement
of cumulants of a sufficient high order will allow to explore experimentally
if there are any ``wiggles'' in the free energy, which may be associated
with some phase changes.
In addition to thermal fluctuations there are many other sources and
types of fluctuations. On the most fundamental level there are quantum
fluctuations, which arise if we measure several non-commuting observables.
In heavy-ion collisions, we encounter fluctuations and correlations
related to the initial state of the system, fluctuations reflecting
the subsequent evolution of the system, and trivial fluctuations
induced by the experimental measurement process. Initial state fluctuations
are driven by, e.g., inhomogeneities in the initial energy and baryon number deposition.
These fluctuations are quite substantial, and are reflected, for example,
in higher harmonics of the radial flow field.
In this contribution, we will concentrate on thermal fluctuations,
which, away from some phase transitions, are typically small, suppressed
by $1/\sqrt{N}$ where $N$ is the average number of particles in
the volume considered. We will also be concerned with fluctuations
originating in the measurement. These need to be understood, controlled
and subtracted in order to access the dynamical fluctuations which
tell as about the properties of the system.
In experiment fluctuations are most effectively studied by measuring
so-called event-by-event (E-by-E) fluctuations, where a given observable
is measured on an event-by-event basis and its fluctuations are studied
for the ensemble of events. Alternatively, one may analyze the appropriate
multi-particle correlations measured over the same region in phase
space \cite{Bialas:1999tv}.
This contribution is organized as follows. We will first provide a
short review on thermal fluctuations and how they can be addressed,
e.g., by Lattice QCD. We will then discuss various corrections which
need to be applied to the data and (model) calculations. We will close
with a discussion of the recent preliminary measurement of net-proton
cumulants by the STAR collaboration. Finally, we wish to dedicate this
contribution to Andrzej Bialas on the occasion of his
$80^{\mathrm{th}}$ birthday.
\section{Fluctuations of a thermal system}
\label{sec:2}
The system created in a ultra-relativistic heavy ion collision reaches,
to a very good approximation, thermal equilibrium
(see e.g. \cite{Braun-Munzinger:2015hba} for a recent review).
Thermal fluctuations are typically characterized
by the appropriate cumulants of the partition function or, equivalently,
by equal time correlation functions which in turn correspond to the
space-like (static) responses of the system. In the following we will
concentrate on cumulants of conserved charges, such as baryon number
and electric charge. To this end, we will work within the grand-canonical
ensemble, where the system is in contact with an energy and ``charge''
reservoir. Consequently, the energy and the various charges are only
conserved on the average with their mean values being controlled by
the temperature and the various chemical potentials. As far as heavy
ion reactions are concerned, the grand canonical ensemble appears
to be a good choice as long as one considers a sufficiently small
subsystem of the entire final state. In addition, as discussed e.g.
in \cite{Andronic:2014zha}, the final state hadron yields are very
well described by a grand canonical thermal system of hadrons.
Fluctuations of conserved charges are characterized by the cumulants
of susceptibilities of that charge. Given the partition function of
the system with conserved charges $Q_{i}$
\begin{align}
Z={\rm Tr}\left[\exp\left(-\frac{H-\sum_{i}\mu_{i}Q_{i}}{T}\right)\right]
\end{align}
the susceptibilities are defined as the derivatives with respect to
the appropriate chemical potentials. In case of three flavor QCD the
conserved charges are the baryon number, strangeness and electric
charge, $\left(B,S,Q\right)$, and we have
\begin{align}
\chi_{n_{B},n_{S},n_{Q}}^{B,S,Q}\equiv\frac{1}{VT^{3}}\frac{\partial^{n_{B}}}{\partial\hat{\mu}_{B}^{n_{B}}}\frac{\partial^{n_{S}}}{\partial\hat{\mu}_{S}^{n_{s}}}\frac{\partial^{n_{Q}}}{\partial\hat{\mu}_{Q}^{n_{Q}}}\ln Z,\label{eq:fluct:susz_genreal}
\end{align}
where $\hat{\mu}_{i}=\mu_{i}/T$ is the reduced chemical potential for charge
$i$. Since the pressure is given by $P=(T/V)\ln(Z)$, the above susceptibilities
also control its Taylor expansion for small values of the various
chemical potentials. For example
\begin{align}
\frac{P\left(T,\mu_{B}\right)}{T^{4}}=\frac{P\left(T,\mu_{B}=0\right)}{T^{4}}+\sum_{n}c_{n}\,\left(\mu_{B}/T\right)^{n},\label{fluct:eq:pressure_mu}
\end{align}
with the expansion coefficients given by $c_{n}=\frac{\chi_{n}^{B}}{n!}$.
Such a Taylor expansion is employed in order to determine the QCD
equation of state for small chemical potentials \cite{Allton:2002zi,Gavai:2008zr,Borsanyi:2012cr}
from lattice QCD, since the Fermion sign problem does not allow for
a direct calculation. Let us next discuss two examples which illustrate
how the study of fluctuations and correlations provide insight into
the structure of QCD matter.
\subsection*{Net Charge Fluctuations }
\label{sec:fluct:charge}
One example are the fluctuations of the net electric charge. In Refs.
\cite{Jeon:2000wg,Asakawa:2000wh} it has been realized that the electric
charge $q$ of particles contributes in square to the fluctuations
of the net-charge,
\begin{eqnarray}
\ave{\left(\delta Q\right)^{2}}=q^{2}\ave{\left(\delta N\right)^{2}}=q^{2}\ave N,
\end{eqnarray}
where in the last step we assumed the particles to be uncorrelated.
Thus, cumulants of the net-charge are sensitive to the fractional
charge of quarks in a quark gluon plasma. To remove the dependence
on the system size it is convenient to scale the charge variance by
another extensive quantity, such as the entropy,
\begin{align}
R & =\frac{\ave{\left(\delta Q\right)^{2}}}{S}\label{R-def}
\end{align}
A simple estimate using Boltzmann statistics gives \cite{Jeon:2000wg,Jeon:2003gk}
\begin{align}
R_{QGP}=\frac{1}{24}
\end{align}
for a two flavor quark-gluon plasma whereas for a gas of massless
pions we obtain
\begin{align}
R_{\pi}=\frac{1}{6}.
\end{align}
In other words, due to the fractional charges of the quarks and the
increased entropy due to the presence of gluons, the charge
fluctuations per entropy in a QGP is roughly a factor four smaller
than that in a pion gas at the same temperature. In reality the hadronic
phase contains more particles than pions, and, taking into account
hadronic resonances, the charge variance per entropy is reduced by
about 30\% which leaves about a factor three difference between a
hadronic system and a QGP. Also, a system of constituent quarks,
without any thermal gluons leads to a ratio of charge fluctuation to
entropy similar to a hadron resonance gas \cite{Bialas:2002hm}.
Finally, it is worth pointing out that similar arguments
have been utilized to identify the fractional charges in a quantum
Hall system as well as the double charge of cooper pairs in measurements
of shot noise \cite{shot_quantum_hall,shot_cooper}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.55\textwidth]{charge_fluct_lattice}
\end{center}
\caption{Net-charge variance per entropy, $R$, as a function of temperature
from 2+1 flavor lattice QCD with physical quark masses. The red-dashed
line indicate the uncertainty. Results for $\protect\ave{\left(\delta Q\right)^{2}}$
are from \cite{Borsanyi:2011sw} and the entropy is extracted from
\cite{Borsanyi:2010cj}. The dashed horizontal lines indicate the
results for a massless pion gas, a hadron gas as well as a non-interacting
QGP with three flavors of massless quarks. }
\label{fig:fluct:charge_fluct_lattice}
\end{figure}
Since the above ratio, Eq. (\ref{R-def}), contains only well defined
thermal observables, it can be determined using lattice QCD methods,
thus accounting for all possible correlations, the presence of strange
quarks etc. In Fig.~\ref{fig:fluct:charge_fluct_lattice} we show
lattice QCD results for the net-charge variance per entropy based
on the calculations for the net-charge variance from \cite{Borsanyi:2011sw}
and for the entropy density from \cite{Borsanyi:2010cj}. We also
show the results for a free pion gas and a QGP with three flavors
of massless quarks, both using the proper quantum statistics, as
well as that for a hadron resonance gas. We see that the hadron resonance
gas agrees well with the lattice results for temperatures up to, $T\lesssim160{\rm \, MeV}$,
which is close to the pseudo-critical or cross-over temperature of $T_{pc}=154\pm8{\rm \, MeV}$.
For temperatures in the range of $160{\rm \, MeV}\lesssim T\lesssim250{\rm \, MeV}$
the lattice calculations are in between the resonance gas prediction
and that of a non-interacting QGP, indicating that some of the correlations
leading to resonance formation are still present in the system. With
few exceptions, this trend is seen for most quantities which have
been calculated on the lattice, such as energy density, cumulant ratios,
etc. Good agreement with the hadron resonance gas up to the cross-over
temperature, followed by a rather smooth transition to a free QGP
which takes place over a temperature interval of approximately $\Delta T\sim100{\rm \, MeV}$,
where the correlations slowly disappear. As we will show in the next
example, some of these correlations, namely those between the various
quark-flavors, can be explored explicitly by studying so called mixed
flavor or ``off-diagonal'' cumulants.
\subsection*{Correlations between quark flavors }
Let us start by considering the co-variance between strangeness and
baryon number, $\ave{\delta B\delta S}\sim\chi_{1,1}^{B,S}$. To illustrate
the sensitivity of this co-variance to correlations among quarks,
let us again compare a non-interacting QGP with a non-interacting
hadron resonance gas (HRG). In the QGP strangeness is carried exclusively
by baryons, namely the strange quarks, whereas in a HRG strangeness
can also reside in strange mesons. Therefore, baryon number and strangeness
are more strongly correlated in a QGP than in a hadron gas, at least
at low baryon number chemical potential, where mesons dominate. To
quantify this observation, Ref.~\cite{Koch:2005vg} proposed the
following quantity
\begin{align}
C_{BS}\equiv-3\frac{\ave{\delta B\delta S}}{\ave{\delta S^{2}}}=1+\frac{\ave{\delta u\,\delta s}+\ave{\delta d\,\delta s}}{\ave{\delta s^{2}}},\label{eq:fluct:CBS}
\end{align}
where we have expressed $C_{BS}$ also in terms of quark degrees of
freedom, noting that the baryon number of a quark is $1/3$ and the
strangeness of a s-quark is negative one. Here $(u,d,s)$ represent
the net-number of up, down and strange quarks, i.e., the difference
between up and anti-up quarks etc. For a non-interacting QGP, $\ave{\delta u\,\delta s}=\ave{\delta d\,\delta s}=0$,
so that $C_{BS}=1$. For a gas of kaons and anti-kaons, on the other
hand, where a light (up or down) quark is always correlated with a
strange anti-quark (kaons) or vice versa (anti-kaons) $\ave{\delta u\,\delta s}<0$,
resulting in $C_{BS}<1$. Strange baryons, on the other hand, correlate
light quarks with strange quarks or light anti-quarks with strange
anti-quarks, so that $\ave{\delta u\,\delta s}>0$. Therefore, for
sufficiently large values of the baryon number chemical potential,
$C_{BS}>1$ for a hadron gas, whereas for a non-interacting QGP $C_{BS}=1$
for all values of the chemical potential \cite{Koch:2005vg}. Since
$C_{BS}$ can be expressed in terms of susceptibilities, $C_{BS}=-3\frac{\chi^{B,S}_{11}}{\chi^{S}_{2}}$,
it can and has been calculated on the lattice with physical quark
masses by two groups \cite{Borsanyi:2011sw,Bazavov:2012jq}. Both
calculations agree with each other, and both report a small, but significant
difference between the lattice results and that from the hadron resonance
gas. In \cite{Bazavov:2014xya} it has been argued that this discrepancy
may be removed by allowing for additional strange hadrons, which are
not in the tables of the Particle Data Group (PDG) \cite{Agashe:2014kda},
but are predicted by various quark models. This is shown in Fig.~\ref{fig:fluct:cbs},
where the lattice QCD results are compared with a hadron resonance
gas based on all the hadrons in the Review of Particles \cite{Agashe:2014kda}
(dotted line) and a hadron gas with additional strange hadrons (full
line). Whether or not this turns out to be the correct explanation,
this comparison demonstrates that these cumulant ratios are a sensitive
probe of the relevant microscopic degrees of freedom.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.7\textwidth]{my_cbs_karsch}
\end{center}
\caption{Lattice QCD results for $-\frac{\chi^{B,S}_{11}}{\chi^{S}_{2}}=\frac{1}{3}C_{BS}$
together with results from hadron resonance gas with and without extra
strange mesons. Figure adapted from \cite{Bazavov:2014xya}.}
\label{fig:fluct:cbs}
\end{figure}
To summarize, the above examples illustrate how cumulants of conserved
charges can be utilized to extract useful information about the correlations
and relevant degrees of freedom of QCD matter. Since they are amenable
to lattice QCD methods, the insights derived from such studies are
rather model independent.
\section{Measuring Cumulants}
\label{sec:3}
\subsection{Some general considerations}
Given the wealth of information which can be extracted from cumulants
of conserved charges and the fact that they can be determined model
independently, it would be desirable to measure these cumulants in
heavy ion collisions. However, a heavy ion collision is a highly dynamical
process whereas lattice QCD deals with a static system in global equilibrium.
In addition, real experiments have limitations in acceptance etc.,
which are difficult to map onto a lattice QCD calculation. Consequently
a direct comparison of experiment with lattice QCD results for fluctuation
observables is a non-trivial task. In the following we will discuss
various issues which need to be understood and addressed in order
for such a comparison to be meaningful.
\begin{itemize}
\item \textbf{Dynamical evolution:} So far our discussion assumed that the
system is static and in global thermal equilibrium. However, even
if fluid dynamics is applicable the system is at best in local thermal
equilibrium. The difference between local and global thermal equilibrium
is an important aspect of the evolution of fluctuations of conserved
charges, because the amount of conserved charge in a given co-moving
volume can only change by diffusion, and the rate of diffusion is
limited by causality \cite{Shuryak:2000pd}. This observation is central
to the use of the variable $R$ defined in Eq.~(\ref{R-def}) to
detect the presence of quark gluon plasma. If we consider a sufficiently
large rapidity window $\Delta y$ then the value of $R$ is frozen
during the QGP phase, and cannot change in the subsequent hadronic
stage. Of course, if $\Delta y$ is chosen too large, then $R$ never
equilibrates, and reflects properties of the initial state. This observation
can be made more quantitative using the theory of fluctuating hydrodynamics.
However, so far most theoretical studies have focused on schematic
models, see, for example \cite{Kitazawa:2013bta}.
\item \textbf{Global charge conservation:} Obviously, baryon number, electric
charge and strangeness are conserved globally, i.e., if we detected
all particles, none of the conserved charges would fluctuate. In contrast,
lattice QCD calculations are carried out in the grand canonical ensemble,
which allows for exchange of conserved charges with the heat bath.
Consequently, charges are conserved only on the average and, thus,
do fluctuate due to the exchange with the heat bath. These exchanges
and thus the fluctuations depend on the correlations between particles
and, as demonstrated above, on the magnitude of the charges of the
individual particles. Therefore, in order to compare with lattice
QCD, one has to mimic a grand canonical ensemble in experiment. This
can be done by analyzing only a subset of the particles in the final
state. However, even in this case, corrections due to global charge
conservation are present. These corrections increase with the order
of the cumulant \cite{Bzdak:2012an} and need to be taken into account
as discussed in \cite{Kitazawa:2013bta,Koch:2001zn,Aziz:2004qu,Kitazawa:2012at}.
\item \textbf{Finite acceptance:} All real experiments do have a finite
acceptance, i.e., they are not able to cover all of phase space. In
addition, most experiments are unable to detect neutrons, which do
carry baryon number. However, due to rapid isospin exchange processes,
the lack of neutron detection may be modeled by a binomial
distribution \cite{Kitazawa:2012at}. While it is desirable to study
only a subset of particles, in order to mimic a grand canonical ensemble,
it is mandatory to have sufficient coverage in phase space in order
to capture all correlations.
\item \textbf{Efficiency corrections:} A real world experiment detects a
given particle only with a probability, commonly referred to as efficiency
$\epsilon$, which is smaller than one, $\epsilon<1$. However, this
does not imply that in every event the same fraction of produced particles
is detected. Rather, the number of measured particles fluctuates even
if the number of produced particles does not. In other words the finite
detection efficiency gives rise to fluctuations, which need to be
removed or unfolded before comparing with any theoretical calculation.
If the efficiency follows a binomial distribution, analytic formulas
for the relation between measured and true cumulants can be derived
\cite{Bzdak:2012ab,Bzdak:2013pha,Luo:2014rea}. Those have been applied
to the most recent analysis by the STAR collaboration.
\item \textbf{Dynamical fluctuations:} A heavy ion collision is a highly
dynamical process and the initial conditions as well as the time evolution
may easily give rise to additional fluctuations. Especially at lower
energies, $\sqrt{s}\lesssim30{\rm \, GeV}$, the incoming nuclei are stopped
effectively and deposit baryon number and electric charge in the
mid-rapidity region. Clearly the amount of baryon number deposited will
vary from event to event, resulting in fluctuations of the baryon
number at mid-rapidity, which are not necessarily the same as those
of a thermal system. This potential source of background needs to
be understood and removed, especially at low energies where one uses
higher cumulants of the net proton distribution in order to find signals
for a possible QCD critical point. Not only does the number of baryon
and charges fluctuate due to the collision dynamics, so does the size
of the system. And while ratios of cumulants do not depend on the
average system size (unless the system is at a second order phase
transition), they are affected by event by event fluctuation
of the system size. This has been studied in \cite{Skokov:2012ds}
and it was found that only for the very most central collisions these
fluctuations are suppressed. Alternatively, one can device observables,
which are not sensitive to size fluctuation \cite{Jeon:2003gk,Koch:2008ia,Gazdzicki:2013ana,Sangaline:2015bma}.
\end{itemize}
The first three points deserve some additional discussion, as they
pose contradictory demands on the measurement \cite{Koch:2008ia}.
In order to minimize corrections from global charge conservation one
wants to keep the acceptance window $\Delta$, say in rapidity, as
small as possible. On the other hand, in order to capture the physics,
the acceptance window needs to be sufficiently wide in order to catch
the correlation among the particles. Therefore, if $\sigma$ is the
correlation length in rapidity and $\Delta_{charge}$ the range over
which all the charges are distributed, then $\Delta/\Delta_{charge}\ll1$
in order to minimize the effects of charge conservation, and $\sigma/\Delta\ll1$
in order to capture the physics.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.6\textwidth]{schematic_correlation}
\end{center}
\caption{Observed scaled variance as a function of the acceptance window in
units of the correlation length. The full (black) line corresponds
to an infinite system where global charge conservation can be ignored.
The long-dashed (blue) and short-dashed (red) lines correspond to the
situation where the charge is conserved within $(-10\sigma,10\sigma)$
and $(-5\sigma,5\sigma)$, respectively.}
\label{fig:fluct:schematic_correlation}
\end{figure}
To illustrate this point, let consider the following schematic model.
Let us define a two-particle correlation function (in rapidity $y$)
\begin{align}
\ave{n(y_{1})\left(n(y_{2})-\delta\left(y_{1}-y_{2}\right)\right)}=\ave{n\left(y_{1}\right)}\ave{n\left(y_{2}\right)}\left(1+C\left(y_{1},y_{2}\right)\right)
\end{align}
with $\ave{n(y)}=\rho=const.$ Then the (acceptance dependent) scaled
variance of the particle number is given by
\begin{align}
\frac{\ave{(\delta N)^{2}}}{\ave N}=1+\frac{\rho}{\Delta}\int_{-\Delta/2}^{\Delta/2}dy_{1}\,dy_{2}\,C\left(y_{1},y_{2}\right)\label{fluct:eq:schematic_correlate}
\end{align}
where the acceptance in rapidity is given by $-\Delta/2<y<\Delta/2$.
Using a simple Gaussian for the correlation function
\begin{align}
C\left(y_{1},y_{2}\right)=\frac{C_{0}}{\rho}\exp\left(-\frac{\left(y_{1}-y_{2}\right)^{2}}{2\sigma}\right)
\end{align}
in Fig.~\ref{fig:fluct:schematic_correlation} we show the scaled
variance as a function of the size of the acceptance window in units
of the correlation length $\Delta/\sigma$. The black line is simply
the expression of Eq.~(\ref{fluct:eq:schematic_correlate}), where
we have ignored any effects due to global charge conservation, i.e.,
$\Delta_{charge}\rightarrow\infty$. The red and blue dashed lines
represent the situation where the total charge is distributed over
a range of $\Delta_{charge}/\sigma\le5$ and $\Delta_{charge}/\sigma\le10$,
respectively. Here we used the leading order formulas of \cite{Bleicher:2000ek}
to account for charge conservation noting that a more sophisticated
treatment a la \cite{Sakaida:2014pya} would not change the picture
qualitatively. Lattice QCD and model calculations, on the other hand,
would give the asymptotic value indicated by the dashed gray line,
which we have chosen to be $\frac{\ave{(\delta N)^{2}}}{\ave N}=1.5$.
The obvious lesson from this exercise is that a comparison of a measurement
at one single acceptance window $\Delta$ with any model calculation
is rather meaningless. Instead, one needs to measure the cumulants
for various values of $\Delta$, and remove the effect of charge conservation.
If the subsequent results trend towards an asymptotic value for large
$\Delta$, it is this value which should be compared with model and
lattice calculations. Such a program has been carried by the ALICE
collaboration in order to extract the aforementioned charge fluctuations
\cite{Abelev:2012pv}. In this context it is worth mentioning that
recent comparisons of measured cumulant ratios with Lattice QCD to
extract the freeze out conditions \cite{Karsch:2012wm,Borsanyi:2013hza}
are based on measured cumulants which have not been extrapolated as
described above. Thus these results need to be taken with some care.
Before we turn to the proton cumulants, let us make a few additional
remarks concerning efficiency corrections, as they do play a significant
role in the recent STAR data \cite{Luo:2015ewa}. As can be seen in
the left panel of Fig. \ref{fig:mult_eps}, finite detection efficiency,
$\epsilon<1$, affects the observed cumulants considerably, and, thus,
needs to be corrected or unfolded. As discussed in Refs. \cite{Bzdak:2012ab,Bzdak:2013pha,Luo:2014rea}
such an unfolding can be done analytically if the probability for
detection of a particle follows a binomial distribution. However,
there is no a priori reason why the response of a complicated detector
should follow a binomial distribution. For example, in most experiments
the efficiency depends on the particle multiplicity, which would not
be the case for a binomial distribution where the binomial probability,
i.e, the efficiency, is constant, independent of the number of Bernoulli
trials. In Ref. \cite{Bzdak:2016qdc} the effect of a multiplicity
dependent efficiency and various other corrections have been explored.
In the right panel of Fig. \ref{fig:mult_eps} we show the resulting
cumulant ratios $K_{n}/K_{2}$ assuming that the efficiency depends
linearly on the multiplicity $M$
\[
\epsilon\left(M\right)=\epsilon_{0}+\epsilon'\left(M-\ave M\right).
\]
Already a rather weak multiplicity dependence gives rise to correction
of order $50\%$ for $K_{4}/K_{2}$.
\begin{figure}
\begin{center}
\includegraphics[width=0.45\textwidth]{c4c2_1} \hspace{1cm}
\includegraphics[width=0.45\textwidth]{KnK2}
\end{center}
\caption{Left panel: Observed cumulant ratio as a function of binomial probability
$p$. The lines from top to bottom correspond to true cumulant ratios
of $K_{4}/K_{2}=5,\,1,\,0,\,-1,\,-5$. Figure adapted from \cite{Bzdak:2012ab}.
Right panel: Effect of multiplicity dependent efficiency on various
cumulant ratios. Deviations form $K_{n}/K_{2}=1$ indicate the effect
of unfolding based on the binomial distribution with constant efficiency,
$\epsilon_{0}$.
For reference, the STAR data at $7.7\protect{\rm \, GeV}$ exhibit a multiplicity dependence
corresponding to $\epsilon'\simeq -5\times 10^{-4}$ \cite{Luo:2015ewa}. Figure
adapted from \cite{Bzdak:2016qdc}.}
\label{fig:mult_eps}
\end{figure}
The multiplicity dependence of the efficiency is just one example
for a non-binomial behavior of the detection probability. There are
certainly others and some examples are discussed in \cite{Bzdak:2016qdc}.
Therefore, the only way to assure that detector effects are probably
accounted for is for individual experiments to simulate and understand
the response of the detector and carry out the necessary unfolding.
That such an exercise is necessary should be obvious from the above
examples.
\subsection{Proton cumulants}
Let us next turn to the net-proton cumulants. It has been suggested
that higher order baryon-number cumulants are particularly sensitive
to the presence of a critical point in the QCD phase-diagram \cite{Stephanov:2008qz}.
Since it is difficult to detect neutrons, this let to a series of
measurements of net-proton cumulants at various energies \cite{Luo:2015ewa,Adamczyk:2013dal}.
As shown in \cite{Kitazawa:2012at}, given rapid, pion-driven, isospin
exchange, the absence of neutrons can be rather reliably modeled by
a binomial process, with binomial probability $p\simeq0.5$. This,
on the other hand, implies that in addition to the detection efficiencies
one also needs to unfold the absence of neutrons, or, in other words,
detection of protons with an efficiency of 0.8 corresponds to detection
of baryons with an effective efficiency of 0.4. As a result, the sensitivity
to the correct magnitude of the true cumulants gets considerably reduced
as can be seen in the left panel of Fig. \ref{fig:mult_eps}.
\begin{figure}
\begin{center}
\includegraphics[width=1.0\textwidth]{star_data}
\end{center}
\caption{Preliminary data by the STAR collaboration for the energy and rapidity
dependence of the $K_{4}/K_{2}$ cumulant ratio. Figures adapted from
\cite{Luo:2015ewa,Luo:2015doi}.}
\label{fig:star_k4}
\end{figure}
Finally let us discuss the preliminary results for the $K_{4}/K_{2}$
cumulant ratio for net-protons obtained by the STAR collaboration.
In Fig. \ref{fig:star_k4} we show both the dependence on the beam
energy (left panel) and on the width of the rapidity window (right
panel) for the lowest beam energy of $\sqrt{s_{NN}}=7.7{\rm \, GeV}$. Also
shown in the left panel are results from UrQMD calculations. These
exhibit a decreasing cumulant ratio with decreasing beam energy, which
is due to baryon number conservation \cite{Bzdak:2012an}. This behavior
is in stark contrast with the measured cumulant ratio, which shows
a steep rise towards lower energies. It is noteworthy, that this rise
only occurs $after$ corrections for efficiency based on a binomial
efficiency distribution have been applied \cite{Luo:2015ewa}. Obviously,
these preliminary data are very intriguing, especially since most
``trivial'' effects tend to reduce the cumulant ratios, such as
the aforementioned baryon number conservation. However, in light of
our discussion, it will be important that the validity of the binomial
efficiency distribution is verified by a detailed analysis of the
STAR detector response.
The dependence on the size of the rapidity window, shown in the right
panel of Fig. \ref{fig:star_k4} is also quite interesting. The cumulant
ratio keeps increasing up to the maximum available range of $\Delta y=1$.
Following our simple model consideration, this implies the the underlying
rapidity correlations are rather long range. Typically long range
rapidity correlations are associated with early times in the collision.
Although this correspondence is somewhat washed out at lower beam
energies, it raises the question if the observed signal may be due
to some initial state effects such as impact parameter/volume or stopping
fluctuations.
\section{Discussion}
In this contribution we have discussed fluctuations of conserved charges
and their utility for the exploration of QCD matter. In particular
we have concentrated on various cumulants which have the advantage
of being accessible to Lattice QCD calculations. Alternatively one
may study the underlying correlations, as suggested by Bialas et al.
\cite{Bialas:1999tv}. These may actually provide more physical insight
into the dynamics at play. If only one particle species is being considered,
such as e.g. protons, one can relate the cumulants and the correlation
functions \cite{Ling:2015yau,bzdak2016}. For example the two-particle
correlation function is simply given by the first and second order
cumulants, $K_{1,}\,K_{2}$,
\[
C_{2}=K_{2}-K_{1}.
\]
However, once net-protons, i.e. protons \emph{and} anti-protons, are
being considered, no direct relations between the correlation functions
and the cumulants exist. Instead additional (factorial) moments are
required, which can be measured but not be calculated in Lattice QCD.
To conclude, fluctuations are a powerful tool to explore the structure
of QCD matter. They provide insight into the relevant degrees of freedom,
their correlations, and possible phase structures. The measurement
of fluctuations requires some care. First the detector response needs
to be well understood and removed by a proper unfolding procedure.
Furthermore, since a heavy ion collision is a highly dynamical process,
fluctuations induced by the initial state or by the dynamical evolution
need to be understood before a comparison with model or Lattice QCD
calculations is possible. Preliminary data on net-proton cumulants
from the STAR collaboration show intriguing features, especially at
the lowest energies. To which extend they constitute our first glimpse
at structures in the QCD phase diagram can only be found out if all
these uncertainties are fully understood.
\section{Acknowledgments}
This work is supported by the Director, Office of Energy Research,
Office of High Energy and Nuclear Physics, Divisions of Nuclear Physics,
of the U.S. Department of Energy under Contract No. DE-AC02-05CH11231, and
by the Polish Ministry of Science and Higher Education (MNiSW),
by founding from the Foundation for Polish Science,
and by the National Science Centre (Narodowe Centrum Nauki),
Grant No. DEC-2014/15/B/ST2/00175 and in part by DEC-2013/09/B/ST2/00497.
|
\section{Introduction}
It is well established that massive galaxies form most of their stars
at early times, and the more massive, the earlier (e.g., \citealt{renz06,shap11}).
Hyper-luminous infrared galaxies (HyLIRGs), or galaxies with L$_{IR}(8-1000\mu$m$) >10^{13}$ L$_\odot$ \citep{sand96}
at high redshift discovered in wide field submm surveys,
play an important role in the study of the early
formation of massive galaxies, corresponding to perhaps the dominant
star formation episode in the formation of massive elliptical galaxies
(e.g., \citealt{case14}). Star formation rates over 1000 M$_\odot$
year$^{-1}$, occurring for timescales up to 10$^8$ years, can form the
majority of stars in a large elliptical (e.g., \citealt{nara15}). These
systems are typically highly dust-obscured, and best studied at IR
through radio wavelengths.
A second important finding in galaxy evolution is the correlation
between the masses of supermassive black holes (SMBHs) and their host spheroidal
galaxies (e.g., \citealt{korm13}). While the exact nature of this correlation
remains under investigation, including its redshift evolution (e.g., \citealt{walt04,wang13,will15,kimb15}),
and certainly counter examples exist (e.g., \citealt{vand12}), the general
implication of such a correlation would be that there exists some
sort of co-evolution of supermassive black holes and their host galaxies
Submm galaxies (SMGs) and other ultra-luminous infrared galaxies (ULIRGs), powered by either star
formation or active galactic nucleus (AGN), have
clustering properties that imply they reside in the densest cosmic
environments (proto-clusters) in the early Universe (e.g., \citealt{blai02,chap09,capa11}).
The mechanism driving the extreme star
formation rates remains uncertain, or at least multivariate. Some
systems are clearly in the process of a major gas rich merger, in
which nuclear starbursts are triggered by tidal torques driving gas to
the galaxy centers (e.g., \citealt{enge10,tacc08,riec11L31}). However, some
SMGs show clear evidence for smoothly rotating disk galaxies, with
little indication of a major disturbance (e.g., \citealt{hodg12,kari16}).
There is some evidence that the most extreme luminosity systems,
in particular powerful AGN in HyLIRGs (quasars and powerful radio
galaxies), are preferentially involved in active, gas rich major
mergers leading to compact, nuclear starbursts (e.g., \citealt{riec08,riec11L32,mile08,ivis12}).
These merging systems may indicate a major accretion event in the
formation of the SMBH, coeval with the major star formation episode of
the host galaxy.
Wide field surveys have identified thousands of extreme starbursts in
the early Universe, and the statistical properties and demographics
are reasonably well determined (e.g., \citealt{case12}). Study of such systems are
now turning to the detailed physical processes driving extreme
starbursts in the early Universe, enabled by the advent of sensitive,
wide-band interferometers such as the Atacama Large Millimeter/submillimeter Array (ALMA),
the Very Large Array (VLA), and the NOrthern Millimeter Extended Array (NOEMA). These
facilities allow for deep, very high resolution imaging of the dust,
gas, star formation, and dynamics in extreme starbursts, unhindered by
dust obscuration. Key questions can now be addressed, such as: What
is the relationship between gas mass and star formation (i.e., the `star formation
law')? What are the interstellar medium (ISM) physical conditions that drive
the extreme star formation? What dominates the gas supply? What role
does the local environment play (proto-cluster, group harassment)?
What is the role of feedback in mediating galaxy formation, driven by
either AGN and/or starbursts?
The BRI1202--0725 ($z\sim$4.7) and BRI1335--0417 ($z\sim$4.4) systems were
among the first HyLIRG systems discovered at very high redshift
(\citealt{irwi91,mcma94,omonA96}), and they
remain two of the brightest unlensed submm sources known at $z > 4$.
These two systems are the archetypes for coeval extreme starbursts
and luminous AGN within 1.4 Gyr of the Big Bang. We have undertaken
an extensive study of these systems, using ALMA, the VLA, NOEMA, and
other telescopes, to determine the dominant physical processes driving
the extreme starbursts and their evolution.
In this paper, we present our latest VLA observations of the CO($J=2\rightarrow1$)
emission from these two systems, at a resolution as high as $0.15''=1\,$kpc. Imaging of low order CO emission is crucial to understand the
distribution and dynamics of the cool molecular gas fueling star
formation in the systems. We will assume ($\Omega_{\Lambda}$,$\Omega_m$,h)=(0.682,0.308,0.678) \citep{plan15} throughout. At this distance, 1 arcsecond corresponds to 6.63$\,$kpc at $z$=4.69 (BRI1202--0725) and 6.82$\,$kpc at $z$=4.41 (BRI1335--0417).
\section{The Sources}
\subsection{BRI1202--0725}
BRI1202--0725 ($z=4.69$) is perhaps the richest physical laboratory to study major gas rich mergers in the early Universe. The system includes two HyLIRGs separated by $4''$: a prototypical, highly dust obscured SMG, and a luminous optical quasar (QSO) in an extreme starburst galaxy. We will refer to the QSO host galaxy as the QSO in what follows. Two Ly-$\alpha$ emitting galaxies (LAEs) are also seen in the system (\citealt{omonA96,hu96}). The 340$\,$GHz emission of \citet{cari13} and the \textit{HST}/ACS F775W filter image of \citet{deca14} are shown in an overview of the field (Fig. \ref{finder}).
\begin{figure}[h]
\centering
\includegraphics[scale=0.6,clip=true]{f1.eps}
\caption{The BRI1202--0725 field. Contours show 340$\,$GHz continuum from \citet{cari13}. Contours are a geometric progression of 2, starting at $\pm4\times0.1\,$mJy beam$^{-1}$, and the beam is $1.2''\times0.8''$ at PA -7$^{\circ}$. The background shows the \textit{HST}/ACS F775W filter image of \citet{deca14}. Positions of each object from 340$\,$GHz are marked, expect for Ly$\alpha$-1, whose optical position is marked.}
\label{finder}
\end{figure}
\subsubsection{SMG \& QSO}
The first detected component of the system, the QSO, was originally discovered as the highest redshift source in a color-selected (UKST $\mathrm{B_J}$, R, and I plates) survey for high redshift QSOs \citep{irwi91} using the Automated Plate Measuring (APM) machine. This discovery in the optical was followed with a detection of FIR emission from thermal dust \citep{mcma94} with a $12''$ beam. Optical spectra show that it features a central black hole with mass $\sim10^9\,$M$_{\odot}$, as estimated from the MgII linewidth \citep{carn13}.
The SMG companion was first discovered in the 1.25$\,$mm observations of \citet{omonN96}. These $\sim2''$ resolution images revealed mm continuum and CO($J=5\rightarrow4$) emission in the SMG and QSO, showing that they are a pair of distinct galaxies \citep{cari02}. Spectral energy distribution (SED) fits of \citet{iono06} yield FIR luminosities of 1.2$\times10^{13}\,$L$_{\odot}$ and 3.8$\times10^{13}\,$L$_{\odot}$ for the SMG and QSO, respectively, indicating that both galaxies are HyLIRGS independently.
The field has been well studied in the submm (\citealt{isaa94,benf99,wagg12,cari13}), mm (\citealt{mcma94,omonA96}), and radio continuum (\citealt{yun00,cari02,momj05,wagg14}). The SEDs of these observations are well fit by combined synchrotron, free-free, and dust emission models for star forming galaxies. The dust masses are M$_D\sim10^9\,$M$_{\odot}$ (\citealt{mcma94,isaa94,hu96,benf99}) and the dust temperatures in each object are T$_D\sim 70\,$K (\citealt{isaa94,leec01,wagg14}). Very Long Baseline Interferometry (VLBI) radio continuum observations yield a non-thermal brightness temperature consistent with a nuclear starburst in the QSO and the SMG \citep{momj05}.
CO observations have successfully detected $J=7\rightarrow6$ (\citealt{omonN96,salo12}), $J=5\rightarrow4$ (\citealt{omonN96,ohta96,yun99,salo12}), $J=4\rightarrow3$ \citep{omonA96}, $J=2\rightarrow1$ (\citealt{kawa99,cari02,wagg14}), and $J=1\rightarrow0$ (\citealt{riec06}). The QSO CO emission has a smaller velocity width than the SMG. The different CO transitions show relative line strengths that imply high excitation, essentially thermal excitation, to $J=4\rightarrow3$ \citep{riec08}, $J=5\rightarrow4$ \citep{wagg14}, or up to $J=8\rightarrow7$ \citep{salo12}.
Assuming a CO luminosity to gas mass conversion factor typical of ULIRGs ($\alpha_{CO}=0.8\,$M$_{\odot}$K$^{-1}$ km$^{-1}$ s pc$^{-2}$; \citealt{bola13}), the molecular gas masses are M$_{gas}\sim10^{11}\,$M$_{\odot}$ (\citealt{mcma94,ohta96,kawa99,cari02,salo12,carn13}). In particular, \citet{wagg14} find gas masses of $(7.0\pm0.6)\times10^{10}\,$M$_{\odot}$ for the SMG and $(4.8\pm0.4)\times10^{10}\,$M$_{\odot}$ for the QSO. The SED fits of \citet{wagg14} give star formation rates (SFRs) of $(4\pm2)\times10^3\,$M$_{\odot}$ year$^{-1}$ for the SMG and for the QSO. These values give short gas depletion timescales of $18\pm9\,$Myr for the SMG and $12\pm6\,$Myr for the QSO and placing them at the low-timescale end, although within the scatter, of values seen in extreme starbursts in the early Universe (\citealt{hodg15,kenn12}).
Imaging has been performed in the atomic fine structure lines, including [CI] \citep{salo12}, [NII] \citep{deca14}, and [CII] (\citealt{iono06,wagg12,cari13,carn13}). The [CII] imaging in particular shows an ordered velocity gradient across the SMG, suggesting rotation on scales of a few kpc \citep{cari13}. If this is the case, the dynamical mass is M$_{dyn}\sim 4\times 10^{11}\,$M$_{\odot}$. The velocity field of the QSO is less ordered, but [CII] observations suggest a small-scale outflow \citep{wagg12,cari13,carn13}.
\subsubsection{Ly-$\alpha$ Emitters}
The Ly-$\alpha$ emission in the BRI1202--0725 field shows a broad line from the QSO, and two additional sources: a detected extension 2.3$''$ to the north of the QSO \citep{peti96,hu96,will14}, called Ly$\alpha$-1, and emission from a galaxy $\sim3''$ to the southwest of the QSO, called Ly$\alpha$-2 \citep{hu97,will14}. Optical continuum emission from both Ly$\alpha$-1 (\citealt{font96,hu96,font98}) and Ly$\alpha$-2 (\citealt{will14}) is also detected. Ly$\alpha$-1 shows a rest frame \textit{B} magnitude comparable to a L* galaxy (\citealt{hu96}).
A study of the permitted UV atomic emission lines and the forbidden FIR fine structure lines from Ly$\alpha$-1 suggest that the QSO is unlikely to be the heating source for the ionized gas, and the line ratios are consistent with heating by local star formation \citep{will14}. Ly$\alpha$-1 may then be shielded by an obscuring torus around the QSO. This face-on geometry relative to us is consistent with the fact that we observe the broad line region of the QSO in strong UV lines, even though the host galaxy has a large dust mass.
The [CII] emission shows Ly$\alpha$-1 may be in a complete bridge of atomic gas between the QSO and SMG \citep{cari13}. Such a bridge would be the natural consequence of strong gravitational interaction during the merging of two gas rich galaxies, although the detection of the bridge remains tentative. Using a stellar population synthesis model, \citet{font96} found a stellar mass of M$_{stellar}\sim10^9\,$M$_{\odot}$. The SFR of Ly$\alpha$-1 ($\sim15\,$M$_{\odot}$ year$^{-1}$) was originally estimated using a similar synthesis model \citep{font98}, while \citet{cari13} found $\sim19\,$M$_{\odot}$ year$^{-1}$ using a $\mathrm{L_{[CII]}}$ conversion.
Evidence suggests that Ly$\alpha$-2 may be somehow associated with an outflow from the QSO. In particular, an extension from the QSO towards Ly$\alpha$-2 has been seen in CO($J=5\rightarrow4$) \citep{salo12}, and emission between the QSO and Ly$\alpha$-2 appears in the [CII] channel maps of \citet{cari13}. The star formation rate, estimated from the dust continuum emission, is $\sim170\,$M$_{\odot}$ year$^{-1}$ \citep{cari13}.
Ly-$\alpha$ emission from Ly$\alpha$-1 and Ly$\alpha$-2 shows full width at half maximum (FWHM) $\sim$1400 and 1200$\,$km s$^{-1}$ \citep{will14}, which are much broader than their [CII] emission (FWHM$\sim$50 and 340 km s$^{-1}$; \citealt{cari13}).
While \citet{ohta00} detected [OII] in Ly$\alpha$-1 and \citet{deca14} found [NII] in Ly$\alpha$-2, deep Very Large Telescope (VLT) observations did not detect UV lines in either source (no NV, SiIV, CIV, HeII; \citealt{will14}). These non-detections suggest that the LAEs originate from star formation, not QSO photoionization, and trace a turbulent ISM/star formation environment affected by the QSO outflow.
\subsection{BRI1335--0417 ($z$=4.4)}
Also discovered in the APM survey \citep{irwi91}, BRI1335--0417 ($z$=4.4) is likely a late-stage, gas rich, ``wet'' merger, with one highly disturbed galaxy associated with an AGN \citep{riec08}. The molecular gas shows a dominant southern core, associated with the QSO host galaxy, and a $\sim1''$ extension towards a weaker northern peak \citep{cari02,riec08}. It is also classified as a HyLIRG due to its FIR luminosity of 3.1$\times10^{13}$L$_{\odot}$ \citep{cari02}. The relation of $5100\,$\AA$\,$ luminosity and H$\beta$ linewidth to black hole mass predicts M$_{BH}\sim10^{10}\,$M$_{\odot}$ \citep{shie06}. Similarly to BRI1202--0725, the rest frame radio through FIR SED of the system is consistent with a star forming galaxy \citep{wagg14}. SED modeling suggests a total dust mass of M$_D\sim10^{8-9}M_{\odot}$ (\citealt{omonA96,benf99,riec08}) and a dust temperature T$_D\sim50$K (\citealt{benf99,cari99}).
CO observations have revealed both $J=2\rightarrow1$ (\citealt{cari99,cari02,riec08,wagg14}) and $J=5\rightarrow4$ \citep{guil97}. Again, the CO spectral line energy distribution (SLED) implies very high excitation and thermalized line strengths up to $J=5\rightarrow4$ (\citealt{wagg14,guil97}). The standard ULIRG L$'_{CO}$ to molecular gas mass conversion factor gives a molecular gas mass of $\sim10^{11}$M$_{\odot}$ (\citealt{guil97,cari99,cari02,riec08,wagg14}).
[CII] emission was detected in this source with the Atacama Pathfinder EXperiment telescope (APEX) \citep{wagg10}.
The SED fits of \citet{wagg14} give an SFR of $(5\pm1)\times10^3\,$M$_{\odot}$ year$^{-1}$. When combined with their gas mass estimate of $(5.8\pm0.5)\times10^{10}\,$M$_{\odot}$, they find a low gas depletion timescale of $12\pm3\,$Myr, placing this source in the starbursting area of a Kennicutt-Schmidt diagram.
VLBI radio continuum observations did not detect any compact radio source, consistent with the idea that the radio continuum emission is due to star formation, not a compact AGN \citep{momj07}.
\section{Observations}
Observations were carried out between September 2013 and February 2014
with the VLA of the NRAO
in the B-- and BnA--configurations, as well as during the move time between the two configurations. Each observing
session was 2 hours, consisting of time on the target
(BRI1202--0725/BRI1335--0417), complex gain calibrator
(J1229+0203/J1338--0432), and 3C286, which was used for bandpass, delay, and flux calibration. Table \ref{obstime} summarizes the observing sessions for each
target.
\begin{deluxetable}{lccc}
\tablecolumns{4}
\tablewidth{0pt}
\tablecaption{VLA Observations \label{obstime}}
\tablehead{ \colhead{Source} & \colhead{Date} & \colhead{Configuration} & \colhead{Time [hr]}}
\startdata
BRI1202--0725&2013 Oct 28 & B&2 \\
BRI1202--0725&2014 Jan 23 & B$\rightarrow$BnA&2\\
BRI1202--0725&2014 Jan 30 & BnA&2 \\
BRI1202--0725&2014 Feb 5 & BnA &2\\
BRI1202--0725&2014 Feb 9 & BnA &2\\
BRI1202--0725&2014 Feb 10 & BnA&2 \\ \hline
BRI1335--0417&2013 Sep 29 & B&$2\times2$ \\
BRI1335--0417&2014 Jan 18 & B&2 \\
BRI1335--0417&2014 Jan 19 & B&$2\times2$ \\
BRI1335--0417&2014 Jan 21 & B&$2\times2$ \\
BRI1335--0417&2014 Jan 23 & B$\rightarrow$BnA&$2\times2$ \\
BRI1335--0417&2014 Jan 24 & B$\rightarrow$BnA&2 \\
BRI1335--0417&2014 Feb 5 & BnA&2 \\
BRI1335--0417&2014 Feb 8 & BnA&$2\times2$ \\
BRI1335--0417&2014 Feb 9 & BnA&2 \\
BRI1335--0417&2014 Feb 10 & BnA&2 \\\hline
\enddata
\end{deluxetable}
In the following, we define zero velocity of the CO($J=2\rightarrow1$)
line as its frequency at CO redshifts of \citet{wagg14} for the
BRI1202--0725 SMG ($z=4.6915$), BRI1202--0725 QSO ($z=4.6942$), and
BRI1335--0417 ($z=4.4065$). This shifts the CO($J=2\rightarrow1$) line
($\nu_{rest}=230.538$ $\,$GHz) to observed frequency of $\nu_{obs}=40.5057\,$GHz,
40.4865$\,$GHz, and 42.6409$\,$GHz, respectively.
We used the Q-band receiver and the four pairs of the 3-bit samplers on each VLA antenna to deliver a total of 8$\,$GHz ($4\times2\,$GHz; $40\,$GHz to $48\,$GHz), with each $2\,$GHz being in right and left circular polarization. During correlation, each $2\,$GHz was further split into 16 spectral windows (SPWs) of width $128\,$MHz, and each SPW was split to 64 $2\,$MHz wide channels. To correct for the poor sensitivity of the SPWs at their edges, two tunings were observed with a 16$\,$MHz offset (39.960 to 48.008$\,$GHz and 39.976 to 48.024$\,$GHz) to mitigate the ranges of poor sensitivity in the instrument's frequency response.
Data were calibrated using the Common Astronomy
Software Applications (CASA) and altered versions of its version 4.1.0 calibration pipeline. First, the
data were processed by a pipeline that was optimized for spectral line
data. This optimization entailed the removal of Hanning smoothing and the flagging of only the one end channel
of each SPW, rather than the end three channels. In addition, the two (B$\rightarrow$BnA) or three (BnA)
northernmost antennas were flagged to create a circular beam.
The data were then inspected for troublesome antennas or
frequency ranges, which were manually flagged. The flagged measurement
set was then passed through a final, calibration-only pipeline that did not flag any additional data. The
six measurement sets of BRI1202--0725 and fifteen measurement sets of
BRI1335--0417 were used to create naturally weighted
images. Additionally, D--configuration VLA data of both
objects from \citet{wagg14} was used. The absolute flux scale uncertainty was $\sim3\%$ \citep{perl13}. Analysis was performed with both
Astronomical Image Processing System (AIPS) and CASA.
Images were created with the CASA task \textit{clean}, using natural
weighting. Continuum images were made by combining the data of all
observation sessions using multi-frequency synthesis (MFS), excluding
channels that contained line emission. Spectral line cubes were made
by applying radial \textit{uv}-tapers to the B--configuration data,
creating images with different synthesized beam sizes to investigate
limits to the CO source sizes. We also explored combining the
D-- and B--configuration \textit{uv}-data before imaging. However, the synthesized beam
solid angles differed by two orders of magnitude, and the resulting
combined synthesized beam had very broad wings. We explored
extensively different visibility weighting schemes in order to restore
jointly any large and small scale structure, while retaining adequate
sensitivity. Unfortunately, for faint sources such as these, we found
that it was difficult to differentiate conclusively between real
extended structure and apparent structure caused by the broad wings of
the combined synthesized beam. Hence, we have taken a conservative
approach to determining source sizes by comparing D--configuration spectra with
B--configuration spectra generated using gradual tapers of the visibilities.
\section{Results}
\subsection{Continuum}
We first consider the 44$\,$GHz continuum emission.
In both sources, the SED fitting of broad band continuum emission implies
that the 44$\,$GHz observed emission (250$\,$GHz rest frame), is thermal emission
from cold dust \citep{wagg14}.
Fig. \ref{cont1}a shows the VLA B--configuration
44GHz continuum image of the SMG in BRI1202--0725, at a
resolution of 0.21$''\times0.15''$ at PA = -24$^{\circ}$. A two--dimensional Gaussian fit
to the emission seen in this image yields a size of (0.18$'' \pm0.05'')\times(0.14''\pm0.08''$) at PA = 2$^{\circ}$, or $(1.2\pm0.3)\times (0.9\pm 0.5)\,$kpc$^2$.
The peak and integrated flux densities of the SMG fit are
$23\pm6\,\mu$Jy beam$^{-1}$ and $41\pm6\,\mu$Jy, respectively. The D--configuration
analysis at $2''$ by \citet{wagg14} found $51\pm 6\,\mu$Jy.
These nearly equal integrated flux densities suggest that the majority of the continuum emission originates from the central $1\,$kpc, not from diffuse structures.
For the QSO host galaxy in 1202--0725, the continuum emission is weak (see Fig. \ref{cont1}b),
and we can only determine an upper limit to the source size of
$<(1.4 \times 1.0)\,$kpc$^2$. From Gaussian fitting, the peak and
integrated flux densities of the QSO are $14\pm6\,\mu$Jy beam$^{-1}$
and $16\pm5\,\mu$Jy. \citet{wagg14} D--configuration analysis found $24\pm
6\,\mu$Jy.
\citet{momj05} observed the synchrotron emission of BRI1202--0725 using VLBI at
1.4$\,$GHz. The SMG was clearly resolved, and larger than, but comparable to
the sizes observed in the 44GHz dust continuum ($1.91\times1.10\,$kpc$^2$).
\begin{figure}[h]
\centering
\includegraphics[scale=0.4,clip=true]{f2a.eps}
\includegraphics[scale=0.4,clip=true]{f2b.eps}
\caption{Naturally weighted VLA B--configuration 44$\,$GHz continuum images of the SMG (a) and QSO (b) in BRI1202--0725. Contours begin at $\pm2\sigma$, where $1\sigma= 5.6\,\mu$Jy beam$^{-1}$, and are in steps of $1\sigma$. Crosses represent source locations from CO($J=2\rightarrow1$) emission. Restoring beam $(0.21''\times0.15'')$ at position angle $-24^{\circ}$ shown as an ellipse in the bottom left corner. The greyscale is in units of $\mu$Jy beam$^{-1}$.}
\label{cont1}
\end{figure}
BRI1335--0417 is detected in continuum emission with a peak at the 4$\sigma$
level (Fig. \ref{cont2}). A two--dimensional Gaussian fit
to the source yields a size of $(0.28''\pm0.06'')\times(0.2''\pm0.1'')$ at PA = 15$^{\circ}$, or $(1.9\pm 0.4)\times(1.2\pm0.7)\,$kpc$^2$, a peak flux density of $10\pm3\,\mu$Jy beam$^{-1}$,
and an integrated flux density of $24\pm 5\,\mu$Jy. \citet{wagg14}
reported a D--configuration flux density of $40\pm7\,\mu$Jy, which is
$\sim3\sigma$ greater than our value. A VLBI observation of BRI1335--0417 at 1.4$\,$GHz yielded a size of $1.29\times0.77\,$kpc$^2$ \citep{momj07}.
\begin{figure}[h]
\centering
\includegraphics[scale=0.4,clip=true]{f3.eps}
\caption{Naturally weighted VLA B--configuration 44$\,$GHz continuum image of BRI1335--0417. Contours begin at $\pm2\sigma$, where $1\sigma= 3.1\,\mu$Jy beam$^{-1}$, and are in steps of $1\sigma$. Cross represents source location from CO($J=2\rightarrow1$). Restoring beam $(0.22''\times0.16'')$ at position angle $-22^{\circ}$ shown as an ellipse in the bottom left corner. The greyscale is in units of $\mu$Jy beam$^{-1}$.}
\label{cont2}
\end{figure}
We do not detect BRI1202--0725 Ly$\alpha$-2 at 44$\,$GHz, implying a lower size limit of $>0.2''$, or $>1.3\,$kpc.
\subsection{Peak Spectra Size Constraints}
In order to get a firm handle on the CO source size limits, we compare
B--configuration CO spectra obtained using different \textit{uv}--tapers, corresponding
to different spatial resolutions, to much lower spatial resolution D--configuration spectra ($\sim 2''$). In each case, the B--configuration spectrum was
extracted at the peak of the velocity integrated emission (moment zero image) for the given
resolution.
We applied circular Gaussian \textit{uv}--tapers to our B--configuration data
in units of kilo--wavelengths ($\lambda\sim 7.5\,$mm for BRI1202--0725
\& $\lambda \sim 7.1\,$mm for BRI1335--0417). We included tapers of
FWHM 750k$\lambda$ (hereafter B750), 450k$\lambda$ (B450), and
250k$\lambda$ (B250). Note that while these tapers increased the
effective beam size of the configuration, these are still much
smaller than the $2''$ resolution of the D--configuration whose maximum baseline corresponds to $\sim140\,$k$\lambda$. The beam
sizes of the tapered and untapered data are listed in Table
\ref{north1}. Note that for tapers lower than 250k$\lambda$, the noise
increases dramatically due to the lack of short spacings.
All spectral line image cubes were created with natural weighting (after tapering), and
1$\,$GHz total bandwidth, approximately centered on the CO($J=2\rightarrow1$) line. Each channel was 5$\,$MHz wide, or $\sim35\,$km
s$^{-1}$. Fig. \ref{Nspec} compares spectra taken at the
emission peak of the untapered B--configuration, tapered
B--configurations (B750, B450, and B250), and D--configuration data
for the SMG and QSO of BRI1202--0725. The results of fitting
single 1-D Gaussians to spectra of each dataset taken at the B750 emission peaks are also
listed in Table \ref{north1}.
\begin{figure}[h]
\centering
\includegraphics[scale=0.5,clip=true]{f4av2.eps}
\includegraphics[scale=0.5,clip=true]{f4b.eps}
\caption{Spectra taken at emission peaks of the SMG (a) and QSO (b) of BRI1202--0725. These are extracted from the following images: untapered B--configuration (B), untapered D--configuration (D), B--configuration tapered to 450k$\lambda$ (B450), and B--configuration tapered to 250k$\lambda$ (B250).}
\label{Nspec}
\end{figure}
\begin{deluxetable}{lccccccc}
\tablecolumns{8}
\tablewidth{0pt}
\tabletypesize{\scriptsize}
\tablecaption{CO(2$\rightarrow$1) emission peak spectrum fits and image properties \label{north1}}
\tablehead{
& & \colhead{Amplitude} & \colhead{Center} & \colhead{$\Delta$v$_{\tiny{FWHM}}$} & \colhead{Line Integral} & \colhead{Restoring Beam} & \colhead{$\sigma$}\\
\colhead{Source} & \colhead{Image} & & & & & & \\
& & \colhead{[mJy beam$^{-1}$]} & \colhead{[km s$^{-1}$]} & \colhead{[km s$^{-1}$]} & \colhead{[mJy beam$^{-1}$ km s$^{-1}$]} & \colhead{[$''\times ''$]} & \colhead{[mJy beam$^{-1}$]}}
\startdata
BRI1202--0725 SMG & B & $0.23 \pm 0.02$ & $115 \pm 35$&$688 \pm 81$&$168 \pm 18$& 0.22$\times$0.16 &0.17\\
BRI1202--0725 SMG & B750 & $0.28 \pm 0.03$ &$ 113 \pm 29 $&$ 685 \pm 69 $&$ 206 \pm 19$& 0.29$\times$0.21 &0.18\\
BRI1202--0725 SMG & B450 & $0.31 \pm 0.03$&$99 \pm 35$&$724 \pm 83$&$236 \pm 25$& 0.38$\times$0.29 &0.20\\
BRI1202--0725 SMG & B250 &$ 0.37 \pm 0.05 $&$ 55 \pm 46 $&$ 727 \pm 108 $&$ 286 \pm 40$& 0.58$\times$0.44 &0.26\\
BRI1202--0725 SMG & D &$ 0.38 \pm 0.03 $&$ 9 \pm 37 $&$ 1052 \pm 88 $&$ 421 \pm 33$& 2.57$\times$1.73 & 0.18\\ \hline
BRI1202--0725 QSO & B & $0.21 \pm 0.03$ & $-96 \pm 29$&$422 \pm 69$&$96 \pm 15$& 0.22$\times$0.16 &0.17\\
BRI1202--0725 QSO & B750 & $0.37 \pm 0.04$ &$ -72 \pm 14 $&$ 262 \pm 33 $&$ 102 \pm 12$& 0.29$\times$0.21 &0.18\\
BRI1202--0725 QSO & B450 & $0.50 \pm 0.04$&$-69 \pm 8$&$196 \pm 19$&$105 \pm 9$& 0.38$\times$0.29 &0.20\\
BRI1202--0725 QSO & B250 &$ 0.70 \pm 0.06 $&$ -64 \pm 8 $&$ 174 \pm 18 $&$ 128 \pm 13$& 0.58$\times$0.44 &0.26\\
BRI1202--0725 QSO & D &$ 0.87 \pm 0.06 $&$ -14 \pm 11 $&$ 365 \pm 27 $&$ 336 \pm 23$& 2.57$\times$1.73 & 0.18\\ \hline \hline
BRI1335--0417 N & B & $0.18 \pm 0.03$ & $46 \pm 31$&$346 \pm 73$&$67 \pm 13$&0.23$\times$ 0.15& 0.13\\
BRI1335--0417 N & B750 & $0.27 \pm 0.03$ &$77 \pm 25 $&$400 \pm 59$&$114 \pm 16$&0.30$\times$0.21&0.13\\
BRI1335--0417 N & B450 & $0.34 \pm 0.04$&$94 \pm 21$&$412 \pm 49$&$150 \pm 17$& 0.36$\times$0.29&0.15\\
BRI1335--0417 N & B250 &$0.49 \pm 0.04$&$ 103 \pm 15 $&$ 424 \pm 35$&$ 222 \pm 17$& 0.54$\times$0.46&0.19\\
BRI1335--0417 N & D &$ 1.14 \pm 0.05 $&$ 40 \pm 6 $&$ 320 \pm 15 $&$ 387 \pm 17 $&2.54$\times$1.95&0.29\\ \hline
BRI1335--0417 S & B & $0.30 \pm 0.02$ & $ 58\pm 16$&$486 \pm 37$&$153 \pm 11$&0.23$\times$ 0.15& 0.13\\
BRI1335--0417 S & B750 & $0.33 \pm 0.02$ &$ 35 \pm 13 $&$ 537 \pm 31 $&$ 190 \pm 10$&0.30$\times$0.21&0.13\\
BRI1335--0417 S & B450 & $ 0.38\pm 0.02$&$36 \pm 14$&$556 \pm 33$&$226 \pm13$& 0.36$\times$0.29&0.15\\
BRI1335--0417 S & B250 &$ 0.51 \pm 0.03 $&$ 41 \pm 15 $&$ 503 \pm 36 $&$ 273 \pm 18$& 0.54$\times$0.46&0.19\\
BRI1335--0417 S & D &$ 1.15 \pm 0.05$&$ 33\pm 7 $&$ 331 \pm 16 $&$ 404 \pm 18$&2.54$\times$1.95&0.29\\ \hline
\enddata
\end{deluxetable}
The first set of plots in Fig. \ref{Nspec} (a) shows that, for the
BRI1202--0725 SMG, the untapered B--configuration captures less that
half of the D--configuration line flux, and falls particularly short at negative
velocity. The 250k$\lambda$ taper spectrum is comparable to that of
the D--configuration, within the noise, indicating a source size approximately equal to the beam size. From Table \ref{north1}, this
places the scale of the BRI1202--0725 SMG at around that of the
restoring beam of B250, namely $\sim0.5''$ ($\sim 3.3\,$kpc).
The second set of plots (b) shows that the full B--configuration resolution
spectrum of the QSO recovers very little of the CO emission, and
essentially zero emission on the high positive velocity wing of the line. The
B250 taper captures most of the negative velocity part of the emission, but
still misses most of the D--configuration emission above zero velocity,
implying an extent $> 0.5''$ ($> 3.3\,$kpc).
\begin{figure}[h]
\centering
\includegraphics[scale=0.5,clip=true]{f5.eps}
\caption{Spectra taken at the emission peak of BRI1335--0417. These are extracted from the following images: untapered B--configuration (B), untapered D--configuration (D), and B--configuration tapered to 250k$\lambda$ (B250).}
\label{ooooo}
\end{figure}
BRI1335--0417 is an extended CO source, as was already
demonstrated in \citet{riec08}. This is seen clearly in the
comparison of the full B--configuration resolution spectrum and the D--configuration (Fig. \ref{ooooo}).
Even for the B250 taper, we are only recovering about half of the
integrated D--configuration emission. From this, we conclude that the extent of the emission
is $>> 0.5''$ ($>> 3.4\,$kpc).
\subsection{Moment Zero Images}
Our spectral analysis has demonstrated clearly that the CO emission
is spatially extended in all three sources. We now turn
to the imaging analysis for further details, keeping in mind that
the signal to noise ratio for some of the sources is low, and hence
detailed imaging is difficult. We employ the B750 images, which retain
some brightness sensitivity, but also provide reasonable resolution ($\sim 0.25''$).
Moment zero (velocity integrated CO flux) images were generated by applying the
CASA task \textit{immoments} to the tapered image cubes. The velocity ranges for
integration were determined by examining the width (Full Width at Zero Intensity; FWZI) of the CO lines in each dataset. This extent was
-357 to 606$\,$km s$^{-1}$ (40.43 to 40.56$\,$GHz) for the BRI1202--0725 SMG, -277 to 241$\,$km s$^{-1}$ (40.46 to 40.53$\,$GHz)
for the BRI1202--0725 QSO, and -281 to 392$\,$km s$^{-1}$ (42.595 to 42.690$\,$GHz) for BRI1335--0417.
For the SMG in BRI1202--0725, the moment zero map in Fig. \ref{mom}a
show an east-west extension at the $3\sigma$ level, with a maximum
extent of $\sim 0.6''$. This extension is admittedly of marginal significance, but it is certainly consistent with the
direction of the velocity gradient seen in [CII] images
taken with ALMA (\citealt{wagg12,cari13,carn13}).
For the QSO in BRI1202--0725, the B750 moment zero image (Fig. \ref{mom}b) shows a marginal (3$\sigma$) detection, with only a fraction of the integrated flux density seen in the D--configuration, implying that the B750 image resolves out
most of the emission. It should be noted that fits to each of the two components of the B750 data show that the emission is equally shared between them ($0.12\pm0.03\,$Jy km s$^{-1}$ for the west peak and $0.13\pm0.03\,$Jy km s $^{-1}$ for the east peak). For completeness, a single Gaussian fit to the QSO emission yields
$(3.1\pm0.3)\times(1.1\pm0.5)\,$kpc$^2$ at PA = 81$\pm3^{\circ}$, with a peak of $0.09\pm0.03\,$Jy beam$^{-1}$
km s$^{-1}$ and a total of $0.24\pm0.03\,$Jy km s $^{-1}$. This integrated flux density is
$75\%$ that of the D--configuration data ($0.32\pm0.01\,$Jy km s $^{-1}$; \citealt{wagg14}).
\begin{figure}[h]
\centering
\includegraphics[scale=0.4,clip=true]{f6a.eps}
\includegraphics[scale=0.4,clip=true]{f6b.eps}
\caption{Moment zero images of BRI1202--0725 made with the B--configuration data tapered to 750k$\lambda$ (B750), averaged over the width of the CO($J=2\rightarrow1$) line as seen in B750. Contours begin at $\pm2\sigma$ and are in steps of $1\sigma$. 44$\,$GHz continuum positions of this work shown by crosses (size of $\sim 0.2''$ corresponds to positional uncertainty). The restoring beam $(0.31''\times0.23'')$ at position angle $-17^{\circ}$ is shown as an ellipse in the lower left corner. a) Image of the SMG, $1\sigma= 32.3\,$mJy beam$^{-1}$ km s$^{-1}$. b) Image of the QSO, $1\sigma=28.8\,$mJy beam$^{-1}$ km s$^{-1}$. The greyscale is in units of mJy beam$^{-1}$ km s$^{-1}$.}
\label{mom}
\end{figure}
Moment zero images of BRI1335--0417 (Fig. \ref{1335mom0}) show
clearly the extended emission to the north and the ever present dominant southern source. We have generated
moment zero images at three different resolutions, showing the emergence
of diffuse structure to the north. Fitting Gaussians to the southern and northern sources separately in the B750 image yields sizes of $(2.2\pm0.1)\times(1.8\pm0.2)\,$kpc$^2$ at PA = $(26\pm20)^{\circ}$ and $(2.1\pm0.1)\times(0.8\pm0.2)\,$kpc$^2$ at PA = $(71\pm5)^{\circ}$, respectively. A fit to the brighter southern source gives a larger integrated flux density of ($0.36\pm0.02)\,$Jy km s$^{-1}$ versus ($0.21\pm0.02)\,$Jy km s$^{-1}$ for the northern source. It also features a higher fit peak flux ($0.15\pm0.02)\,$Jy beam$^{-1}$ km s$^{-1}$ versus ($0.11\pm0.02)\,$Jy beam$^{-1}$ km s$^{-1}$ for the northern source. A fit to the D--configuration data returns an integrated flux density of $(0.62\pm0.03)\,$Jy km s$^{-1}$, which is comparable to the two above B750 fits combined ($0.57\pm0.04\,$Jy km s$^{-1}$). This suggests that the majority of the CO emission emanates from these two compact sources, not from diffuse gas.
\begin{figure}[h]
\centering
\includegraphics[scale=0.28,clip=true]{f7a.eps}
\includegraphics[scale=0.28,clip=true]{f7b.eps}
\includegraphics[scale=0.28,clip=true]{f7c.eps}
\caption{Moment zero images of BRI1335--0417 made with tapered and untapered B--configuration data, integrated over the width of the CO($J=2\rightarrow1$) line. Location of continuum emission from this work shown by cross (size of $\sim0.2''$ corresponds to positional uncertainty based on our beam size). Contours begin at $\pm2\sigma$ and are in steps of $1\sigma$. Restoring beams shown in lower right corner. The greyscale is in units of mJy beam$^{-1}$ km s$^{-1}$.
Left: B--configuration tapered to 450k$\lambda$, $1\sigma=27\,$mJy beam$^{-1}$ km s$^{-1}$. Restoring beam is $(0.38''\times0.29'')$ at position angle $-15^{\circ}$.
Middle: B--configuration tapered to 750k$\lambda$, $1\sigma=24\,$mJy beam$^{-1}$ km s$^{-1}$. Restoring beam is $(0.30''\times0.21'')$ at position angle $-20^{\circ}$.
Right: Untapered B--configuration, $1\sigma=23\,$mJy beam$^{-1}$ km s$^{-1}$. Restoring beam is $(0.23''\times0.15'')$ at position angle $-27^{\circ}$.}
\label{1335mom0}
\end{figure}
\section{Analysis}
\subsection{BRI1202--0725: Ly$\alpha$-2 CO Emission}
We next investigate CO emission from Ly$\alpha$-2 to the
southwest of the QSO using the D--configuration observations (Fig. \ref{1202twospec} and \ref{food}). Velocities are relative to the QSO redshift.
Ly$\alpha$-2 is separated by $2.5''$ from the QSO, corresponding to less than the 20\%
point of the synthesized beam. Note that no CO emission was detected from Ly$\alpha$-1 in any dataset.
Fig. \ref{1202twospec} shows the CO($J=2\rightarrow1$) spectrum of the QSO, again,
as well as one taken at the position of Ly$\alpha$-2. Two velocity ranges are
indicated: one corresponding to the main QSO line (-500 to 450$\,$km s$^{-1}$),
and a second at higher positive velocity (450 to 1240$\,$km s$^{-1}$).
Note that the main QSO line has no emission over most of this high positive
velocity range, hence sidelobes from the QSO emission are not an issue in terms of contaminating
the Ly$\alpha$-2 spectrum. There appears to be emission over this high positive
velocity range at the position of Ly$\alpha$-2. The low end of this
velocity range also corresponds to the [CII] emission seen from Ly$\alpha$-2 \citep{wagg12}.
Unfortunately, the ALMA [CII] observations covered only the lower
velocities. The full width at tenth of maximum (FWTM) of the [NII] observation of \citet{deca14} (magenta line) is comparable to our upper velocity bin.
Fig. \ref{food} shows an image of the integrated emission in the high
positive velocity range. A clear source appears at the
position of Ly$\alpha$-2, with a 6$\sigma$ significance. The
integrated line flux density over this velocity range is
$0.19\pm0.03\,$Jy km $s^{-1}$, which is significantly larger than the
limit of $<0.06\,$Jy km $s^{-1}$ from \citet{wagg14}. Adopting
bounding conversion factors of
$\alpha_{CO}=0.8$M$_{\odot}$K$^{-1}$ km$^{-1}$ s pc$^{-2}$ (for
starbursting galaxies) and $\alpha_{CO}=4$M$_{\odot}$K$^{-1}$
km$^{-1}$ s pc$^{-2}$ (for the Milky Way) yields a gas mass range of
($3.2-16)\times10^{10}\,$M$_{\odot}$.
Briefly, we consider why Ly$\alpha$-2 CO(J=$2\rightarrow1$)
emission was not reported in the original paper of \citet{wagg14}.
First, Wagg et al. were focused principally on the BRI1202--0725 SMG and AGN host
galaxies. And second, the velocity selection of Ly$\alpha$-2 was based on
the truncated [CII] line of \citet{cari13}. Recent [NII] observations
by \citet{deca14} show that the full profile of the object is
much broader. We have increased the signal-to-noise by opening up the search
space in velocity to accommodate a broader line.
We have looked for the emission from Ly$\alpha$-2 CO in the B--configuration data
over the broader velocity range. We do not detect emission at
full resolution, of B750, implying a $3\sigma$ level of $0.1\,$Jy beam$^{-1}$ km s$^{-1}$. These results give a
rough lower limit to the source size of $>0.3"$. Conversely, fitting to the D array
image of the velocity integrated CO emission leads to an upper limit to the
size of $<3''$. We may then state a conservative CO extent of $2-20\,$kpc.
\begin{figure}[h]
\centering
\includegraphics[scale=0.5,clip=true]{f8.eps}
\caption{Comparison of D--configuration peak spectra of the BRI1202--0725 QSO and Ly$\alpha$-2. Solid vertical lines divide channels of Ly$\alpha$-2 emission from those of QSO emission. Green and magenta lines in lower panel shows centroid and FWTM of [CII] from \citet{cari13} and [NII] emission from \citet{deca14}, respectively. Note that the width of the [CII] line is cutoff at the observed band edge, as shown by dotted vertical line.}
\label{1202twospec}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[scale=0.4,clip=true]{f9a.eps}
\includegraphics[scale=0.39,clip=true]{f9b.eps}
\caption{Moment zero images of the BRI1202--0725 field from D--configuration data, integrated over the QSO and Ly$\alpha$-2 linewidths.
a) The field integrated over the QSO linewidth (-500 to 450$\,$km s$^{-1}$; 40.4209 to 40.5491$\,$GHz). The crosses represent (from north to south) the 44$\,$GHz continuum positions of the SMG and QSO from this work and the \citet{cari13} submillimeter positions of Ly$\alpha$-2 (cross sizes are arbitrary). Contours begin at $\pm2\sigma$ and are in steps of $1\sigma= 27\,$mJy beam $^{-1}$ km s$^{-1}$. Restoring beams shown as an ellipse in the lower right corner. The greyscale is in units of mJy beam$^{-1}$ km s$^{-1}$. The restoring beam is $(2.74''\times1.86'')$ at position angle $14^{\circ}$.
b) The field integrated over the Ly$\alpha$-2 linewidth (450 to 1240$\,$km s $^{-1}$; 40.5491 to 40.6563$\,$GHz). Contours begin at $\pm2\sigma$ and are in steps of $2\sigma= 30\,$mJy beam $^{-1}$ km s$^{-1}$. Positions of 340$\,$GHz detections of \citet{cari13} are marked. The color shows the \textit{HST}/ACS F775W filter image of \citet{deca14}}
\label{food}
\end{figure}
\subsection{Star Formation Surface Densities}
We have SFRs for each object \citep{wagg14}, but require star
formation areas to calculate surface densities. Current measurements
are only limits or marginal measurements. \citet{carn13} present
unresolved ALMA 340$\,$GHz size estimates of $<22\,$kpc$^2$ for the
BRI1202--0725 SMG and QSO. Our measured 44$\,$GHz size for the
BRI1202--0725 SMG was $0.9\pm0.5\,$kpc$^2$, which is an order of
magnitude smaller. Another constraint is the 1.4$\,$GHz size of the
BRI1202--0725 SMG from \citet{momj05},
$(0.3\pm0.2)\,$kpc$^2$. Similarly, we found that BRI1335--0417 showed
a 44$\,$GHz size of $1.8\pm1.2\,$kpc$^2$, while \citet{momj07} found a
1.4$\,$GHz size of $(0.4\pm0.2)\,$kpc$^2$. These 1.4GHz
synchrotron-emitting areas are smaller than, but within
$\sim1.2\sigma$ of our $44\,$GHz regions. While our 44$\,$GHz size
estimates feature large uncertainties, they also show integrated flux
densities that are greater than their peak surface brightnesses
($41\pm6\,\mu$Jy and $23\pm6\,\mu$Jy beam$^{-1}$ for BRI1202--0725
SMG, $24\pm5\,\mu$Jy and $10\pm3\,\mu$Jy beam$^{-1}$ for
BRI1335--0417). The greater integrated flux density vs. peak surface
brightness suggests that the sources are indeed resolved. However, in
both cases, these integrated flux densities from the B array imaging
are still (marginally) smaller than those seen in the D array data:
$51 \pm 6$ $\mu$Jy and $40 \pm 7$ $\mu$Jy, respectively, suggesting
additional diffuse emission. Given the low signal-to-noise of these
observations, we adopt for the sake of calculation the 44GHz sizes as
as a guide, keeping in mind that these are likely lower limits, with
upper limits set by ALMA dust continuum imaging.
The BRI1202--0725 SMG shows a density of
$\Sigma_{SFR}=(2\pm1)\times10^3\,$M$_{\odot}$ yr$^{-1}$ kpc$^{-2}$,
while BRI1335--0417 shows
$\Sigma_{SFR}=(3\pm2)\times10^3\,$M$_{\odot}$ yr$^{-1}$
kpc$^{-2}$. The continuum image of the BRI1202--0725 QSO has too low a
signal to noise ratio to determine a size. The SFR surface density of
Ly$\alpha$-2 may be estimated using a SFR=$170\,$M$_{\odot}$ yr$^{-1}$
\citep{cari13} and a 44$\,$GHz size limit of $>1.4''\sim9\,$kpc, giving
$<121\,$M$_{\odot}$ year$^{-1}$ kpc$^{-2}$.
The fraction of the luminosity at which gravitational collapse is balanced by radiation pressure, or Eddington fraction, of these sources can be estimated using these surface densities. \citet{thom05} develop the limit of SFR surface density $\Sigma_{SFR}>1000\,$M$_{\odot}$ year$^{-1}$ kpc$^{-2}$ for super-Eddington luminosity, assuming a star formation efficiency and opacity. By this criterion, both the BRI1202--0725 SMG and BRI1335--0417 are super-Eddington. An alternate estimate may be found using a luminosity-to-mass limit of $>$500$\,$L$_{\odot}$/M$_{\odot}$ \citep{scov12}. Approximate values of $10^{13}\,$L$_{\odot}$ and $10^{11}\,$M$_{\odot}$ give Eddington fractions of $\sim20\%$. See Appendix A for an additional Eddington criterion which returns $<1\%$ for both sources.
While the scatter is significant between these methods, the first suggests that both the SMG and BRI1335--0417 are radiating near their Eddington limit. It should be noted that the Eddington fraction is dependent on assumed geometry (disk/sphere), L'$_{CO}$ to M$_{H_2}$ conversion factor, opacity, and total object mass, among other variables.
\subsection{Gas Surface Densities}
For the sources that we have both 44$\,$GHz continuum and CO($J=2\rightarrow1$) size estimates (BRI1202--0725 SMG and BRI1335--0417), it is evident that the CO extent is greater. This continuum, which is $\sim250\,$GHz (rest frame), is a tracer of cold dust, and thus indirect star formation. The more compact size of the continuum emitting region compared to CO emission, which traces the molecular gas, suggests a central, excited region surrounded by a gaseous reservoir.
We estimate the molecular gas masses using three conversion factors from L$'_{CO}$ to M$_{H_2}$. However, all three factors use the luminosity of the $J=1-0$ transition of CO, not the observed $J=2-1$ transition. The ratios of these luminosities are given by \citet{cariw13} as r$_{21}$=L'$_{CO(2-1)}$/L'$_{CO(1-0)}$=0.85, 0.99, 0.97, and 0.5 for SMGs, QSOs, color selected galaxies (CSGs), and the Milky Way, respectively. The application of the SMG ratio to the BRI1202--0725 SMG and the QSO ratio to the BRI1202--0725 QSO are obvious. Less so is the correct correction for BRI1335--0417, which shows evidence of an AGN but no compact radio emission. We will use the SMG correction factor, due to its extended starburst region. To account for the unknown nature of BRI1202--0725 Ly$\alpha$-2, we will use the range of factors from Milky Way-type objects to CSGs (0.5-0.97).
The first gas mass estimate uses the standard conversion for low $z$ ULIRGs, $\alpha_{CO}=0.8\,$M$_{\odot}$K$^{-1}$ km$^{-1}$ s pc$^{-2}$. We also use the flexible factor of \citet{nara12}:
\begin{equation}
\alpha_{CO}=\frac{min[6.3,10.7\times W_{CO}^{-0.32}]}{Z'^{0.65}}
\label{eqnDN}
\end{equation}
where $Z'=Z/Z_{\odot}$ is the normalized metallicity and $W_{CO}$ is the line intensity, which we will define as $L'_{CO}/A_{CO}$, where $A_{CO}$ is the area of our CO($J=2-1$) emission and $L'_{CO}$ is our $L'_{CO(2-1)}/r_{21}$. To explore the metallicity dependence of the gas mass, we used solar metallicity ($Z'=1$) and the extremely low metallicity of I Zw 18 ($Z'=0.02$; \citealt{lequ79}). Note that the latter value is unphysically low for areas showing this level of stellar development, and it is only included as an illustrative boundary.
\begin{deluxetable}{lcccc}
\tablecolumns{5}
\tablewidth{0pt}
\tabletypesize{\scriptsize}
\tablecaption{Source Properties \label{allthat}}
\tablehead{
\colhead{} & \colhead{BRI1202--0725 SMG} & \colhead{BRI1202--0725 QSO} & \colhead{BRI1335--0417} & \colhead{BRI1202--0725 Ly$\alpha$-2}}
\startdata
L'$_{CO(2-1)}$ [$10^{11}\,$K km s$^{-1}$ pc$^2$] & $1.1\pm0.2$ & $0.51\pm0.06$ & $1.09\pm0.08$ & $0.40\pm0.06$ \\
$\alpha_{CO}(Z'=1)$ [M$_{\odot}$ K$^{-1}$ km$^{-1}$ s pc$^{-2}$] & $0.41\pm0.05$ & $0.46\pm0.07$ & $0.40\pm0.02$ & $0.41-2.2$ \\
$\alpha_{CO}(Z'=0.02)$ [M$_{\odot}$ K$^{-1}$ km$^{-1}$ s pc$^{-2}$] & $5.2\pm0.6$ & $5.8\pm0.9$ & $5.1\pm0.2$ & $5.3-28$ \\
M$_{\alpha=0.8}$ [$10^{11}\,$M$_{\odot}$] & $1.0\pm0.2$ & $0.41\pm0.05$ & $1.03\pm0.07$ & $0.3-0.7$ \\
M$_{Z'=1}$ [$10^{10}\,$M$_{\odot}$] & $5\pm1$ & $2.3\pm0.5$ & $5.1\pm0.4$ & $1.5-3.3$ \\
M$_{Z'=0.02}$ [$10^{11}\,$M$_{\odot}$] & $7\pm1$ & $3.0\pm0.6$ & $6.5\pm0.5$ & $1.9-4.2$ \\
A$_{CO(2-1)}$ [kpc$^2$] & $5\pm2$ & $3\pm1$ & $4.4\pm0.5$ & $3-300$ \\
$\Sigma_{H_2,\alpha=0.8}$ [$10^4\,$M$_{\odot}$ pc$^{-2}$] & $2.1\pm0.8$ & $1.5\pm0.7$ & $2.3\pm0.3$ & $0.009-2$ \\
$\Sigma_{H_2,Z'=1}$ [$10^4\,$M$_{\odot}$ pc$^{-2}$] & $1.1\pm0.4$ & $0.9\pm0.4$ & $1.2\pm0.2$ & $0.005-1$ \\
$\Sigma_{H_2,Z'=0.02}$ [$10^5\,$M$_{\odot}$ pc$^{-2}$] & $1.4\pm0.6$ & $1.1\pm0.6$ & $1.5\pm0.2$ & $0.006-1$ \\
SFR [$10^3\,$M$_{\odot}$ year$^{-1}$]\tablenotemark{a} & $4\pm2$ & $4\pm2$ & $5\pm1$ & 0.17 \\
A$_{44\,\mathrm{GHz}}$ [kpc$^2$] & $0.9\pm0.5$ & - & $1.8\pm1.1$ & $>1.4$ \\
$\Sigma_{SFR}$ [$10^3\,$M$_{\odot}$ year$^{-1}$ kpc$^{-2}$] & $4\pm3$ & - & $3\pm2$ & $<120$ \\
\hline
\enddata
\tablenotetext{a}{\citet{wagg14}}
\end{deluxetable}
Our molecular gas masses and resulting surface densities are listed in Table \ref{allthat}. Using our CO emission extents (Section 4.3) and the molecular gas masses, we find gas surface densities $\Sigma_{H_2}\sim10^{4-5}\,$M$_{\odot}$ pc$^{-2}$ for the three HyLIRGs. These are larger than the average gas surface density reported by \citet{grev05}, who observed 11 SMGs in $z=1.0-3.4$ and found $\Sigma_{H_2}=(2.4\pm1.3)\times 10^3\,$M$_{\odot}$ pc$^{-2}$. Our values are more comparable to the estimates of \citet{dadd09}, who find $\Sigma_{H_2}=(1.6,1.0)\times 10^4\,$M$_{\odot}$ pc$^{-2}$ for GN20 and GN20.2a.
The placement of these objects on a Kennicutt-Schmidt plot (Fig. \ref{KSPLOT}) shows that they are separate from the main sequence. If an unrealistic metallicity is assumed, they may have comparable gas depletion timescales to local starbursts or $z\sim2$ SMGs. A limit on source metallicity can be be found by assuming $\alpha=0.8$ and inverting equation \ref{eqnDN} to solve for Z', yielding Z'$_{BRI1202-0725 SMG}=0.38\pm0.06$ and Z'$_{BRI1335-0417}=0.34\pm0.02$.
Finally, we may use our CO size limits and gas mass estimate for Ly$\alpha$-2 (Section 5.1) to find $\Sigma_{H_2}=10^{1.7-5.0}\,$M$_{\odot}$ pc$^{-2}$. This range includes our uncertainties in the CO size, the $\alpha_{CO}$ conversion factor, and the $r_{21}$ ratio. Due to its unconstrained size, it may match well with either starbursts or normal galaxies in the Kennicutt-Schmidt diagram (Fig. \ref{KSPLOT}).
\begin{figure}[h]
\centering
\includegraphics[scale=0.5,clip=true]{f10v2.eps}
\caption{Kennicutt-Schmidt diagram including our sources and those from literature. BRI1202--0725 QSO is not shown due to its unresolved 44$\,$GHz extent and resultant lack of $\Sigma_{SFR}$. Literature sources include normal galaxies (\citealt{kenn98,kenn07}, circumnuclear starburst areas (\citealt{kenn98}), $z\sim2$ SMGs (\citealt{genz10,both10}), $z\sim1$ SFGs (\citealt{freu13}), and $z\sim1-3$ SFGs of \citealt{tacc13}). Positions of our sources are shown using $\alpha_{CO}=0.8\,$M$_{\odot}$K$^{-1}$ km$^{-1}$ s pc$^{-2}$ (no marker), the adaptive factor of \citet{nara12} using $Z'=1$ (circles) and $Z'=0.02$ (diamonds).}
\label{KSPLOT}
\end{figure}
\begin{comment} STACEY
\subsection{Diagnostic Plot}
An additional method of deducing the physical conditions of these objects
is to use the diagnostic plot of \citet{stac10}. This plot of L$_{[CII]}$/L$_{FIR}$
versus L$_{CO}$/L$_{FIR}$ allows us to estimate both the matter density (n[=]cm$^{-3}$)
and strength of the FUV radiation field (G$_0$=F$_{FUV}$/($1.6\times10^{-3}\,$erg cm$^{-2}$ s$^{-1}$)).
The placement of our sources is compared to densities and radiation fields of simulated
photodissociation regions (PDRs), resulting in limits in Table \ref{diag}. The densities
of the BRI1202--0725 SMG and Ly$\alpha$-2 were similar to that of BRI1335--0417, while the
BRI1202--0725 QSO was less dense. The radiation field of BRI1202--0725 Ly$\alpha$-2 was
markedly less than that of the three HyLIRGs. Overall, these objects are higher density
and show lower intensity FUV fields than the other objects in the sample.
\begin{deluxetable}{lccccc}
\tablecolumns{6}
\tablewidth{0pt}
\tabletypesize{\scriptsize}
\tablecaption{Diagnostic plot data. \label{diag}}
\tablehead{
\colhead{Source} & \colhead{L$_{[CII]}$} & \colhead{L$_{CO(1-0)}$} & \colhead{L$_{FIR}$} & \colhead{n} & \colhead{G$_0$}\\
& \colhead{[10$^9\,$L$_{\odot}$]} & \colhead{[$10^7$L$_{\odot}$]} & \colhead{[10$^{13}$L$_{\odot}$]} & \colhead{[cm$^{-3}$]} & }
\startdata
BRI1202--0725 SMG & $13$\tablenotemark{a} & $5.1$ & $1.5$\tablenotemark{d} & $10^{5-6}$ & $10^{2-3}$\\
BRI1202--0725 QSO & $4.5$\tablenotemark{a} & $2.0$ & $1.8$\tablenotemark{d} & $10^{6-7}$ & $10^{2-3}$\\
BRI1335--0417 & $16$\tablenotemark{b} & $5.0$ & $3.1$\tablenotemark{e} & $10^{5-6}$ & $10^{2-3}$\\
BRI1202--0725 Ly$\alpha$-2 & $0.85$\tablenotemark{c} & $1.6$ & $0.17$\tablenotemark{f} & $10^{5-6}$ & $10^{1-2}$\\ \hline
\enddata
\tablecomments{$^\mathrm{a}$\citet{wagg12} $^\mathrm{b}$\citet{wagg10} $^\mathrm{c}$\citet{cari13} $^\mathrm{d}$\citet{carn13} $^\mathrm{e}$\citet{cari02} $^\mathrm{f}$\citet{will14}}
\end{deluxetable}
\end{comment}
\section{Discussion}
We have imaged the archetypal HyLIRGs BRI1202--0725 and BRI1335--0417
with the VLA B--configuration in the rest-frame 250$\,$GHz continuum
and CO($J=2\rightarrow1$) at high resolution. These observations allow us to determine
sizes for the 44GHz continuum and the CO emission
on scales down to $\sim 1$kpc.
The 44$\,$GHz continuum emission in all three sources appears extended on a scale
of 1 to 2$\,$kpc, although only marginally so for the QSO host in
BRI1202--0725. Based on radio through FIR SED fitting,
the observed 44$\,$GHz continuum emission is thermal emission from
cold dust \citep{wagg14}. For the BRI1202--0725 SMG, the 44$\,$GHz size roughly agrees with the nonthermal 1.4$\,$GHz size from the VLBI observations of \citet{momj05}, while the 44$\,$GHz size of BRI1335--0417 is about twice as large as the VLBI 1.4$\,$GHz extent of \citet{momj07}. Assuming these 44$\,$GHz continuum extents correspond to that of the starbursting regions, we derive SFR surface densities of $\Sigma_{SFR}=(4\pm3)\times10^3\,$M$_{\odot}$ yr$^{-1}$ kpc$^{-2}$ for the BRI1202--0725 SMG and $\Sigma_{SFR}=(3\pm2)\,$M$_{\odot}$ yr$^{-1}$ kpc$^{-2}$ for BRI1335--0417. While estimates of the Eddington fractions of these sources vary, evidence from one test suggests that both of these objects are possibly radiating above their stable limit.
Using both the standard ULIRG conversion factor for CO luminosity to H$_2$ mass and the more flexible factor of \citet{nara12}, we derive surface densities for the gas mass for these systems, based on the CO($J=2\rightarrow1$) source sizes. Using metallicities of 1$\,$Z$_{\odot}$ and 0.02$\,$Z$_{\odot}$, the values range from $10^{4-5}$ M$_\odot$ pc$^{-2}$. When plotted in a Kennicutt-Schmidt diagram, the low metallicity assumption places the HyLIRGs closer to the trend at lower SFR, but all three assumptions mark the HyLIRGs as strong starbursts, i.e. well above main sequence galaxies. A possible limit on object metallicity was found by setting the flexible conversion factor to that of ULIRGs, yielding $Z\sim Z_{\odot}/3$.
A tentative east-west linear structure is seen in the CO($J=2\rightarrow1$) image of the SMG in BRI1202--0725, possibly a disk of size $\sim 0.8''$. This molecular disk is consistent with the east-west velocity gradient seen in the [CII] ALMA observations \citep{carn13}.
For the QSO host galaxy in BRI1202--0725, the CO($J=2\rightarrow1$) emission is clearly
extended, although the signal-to-noise of the emission in the B--configuration
provides only a lower limit to the size of $0.5'' (\sim 3.3\,$kpc).
For BRI1335--0417, we confirm the very extended CO emission to the north of the
QSO host galaxy, as seen in \citet{riec08}. This extended
emission has a narrower velocity dispersion than for the main galaxy.
The negligible velocity offset with respect
to the main galaxy and the lower velocity dispersion of this extended
gas suggests that it may be a remnant tidal feature from a previous
major merger of gas rich galaxies. This is based on radio observations,
as past optical and submm observations lacked the necessary resolution to
distinguish the northern extension.
In all three sources, the extent of the 44$\,$GHz continuum emission
appears smaller than that of the CO($J=2\rightarrow1$) line emission,
or the area of active star formation is smaller than its fuel
supply. This suggests a varying level of excitation and star formation
activity across each source. The variation can be seen in
BRI1335--0417, which shows a southern core in 44$\,$GHz emission, but
no northern extension. Since both are seen in CO emission, this
suggests that the northern area is not starbursting. This size
discrepancy between low order CO and star formation has been noted for
other high redshift SFGs (star forming galaxies; e.g., as compiled by
\citealt{spil15}). Alternatively, the different sizes could simply be
a result of low signal-to-noise continuum observations not
detecting the lower optical depth outer sections of the star forming
regions. Higher resolution, more sensitive ALMA observations of the
dust continuum emission are planned. These should answer this question
of the relative distributions of gas and star formation.
We also detect CO($J=2\rightarrow1$) emission from Ly$\alpha$-2 in the BRI1202--0725
system. The total gas mass is $(3.2\pm0.5)\times (\alpha_{CO}/0.8)\times10^{10}\,$M$_{\odot}$, and the gas depletion timescale is $2\times 10^8\times (\alpha_{CO}/0.8)\,$yr,
where $\alpha_{CO}$ has units of M$_{\odot}$K$^{-1}$ km$^{-1}$ s pc$^{-2}$. Even assuming a low value of $\alpha_{CO}$ (compared to $\alpha_{CO}\sim4$ for normal galaxies; \citealt{bola13}), this galaxy
has a gas depletion timescale comparable to main sequence galaxies at
low and high redshift (e.g., $7\times10^8\,$years for $z=1-3$ MS SFGs, \citealt{tacc13}), and not the extremely short timescales
applicable to starbursts.
\citet{klam04} suggested that the extreme aspects of the BRI1202--0725 system might relate to star formation induced by a strong radio jet from the QSO. While no such radio jet has yet been seen, the presence of Ly$\alpha$-2 in the direction of the ouflow from the QSO seen in [CII] \citep{cari13}, is circumstantially suggestive of a hydrodynamic interaction as well as gravitational.
We have used our CO($J=2-1$) luminosities in conjunction with previous [CII] luminosities (\citealt{wagg12,wagg10,cari13}) and FIR luminosities (\citealt{carn13,cari02,will14}) to constrain the density and FUV radiation field of photodissociation regions (PDRs) in each source with the diagnostic plot of \citet{stac10}. Assuming thermalized ratios (L$_{CO(2-1)}$/L$_{CO(1-0)}$=4), this showed that PDRs in the three HyLIRGs showed similar densities and radiation fields to local ULIRGs and $z>2.3$ objects. When comparing the density and FUV level of PDRs in BRI1202-0725Ly$\alpha$-2 to other objects, an ambiguity arises. They are either similar to normal galaxies in level of FUV emission but are more dense or are similar to ULIRGs in density but feature weaker radiation fields. Current size constraints do not allow for concrete conclusions here.
Our data confirm that these HyLIRGs are extreme starbursts with short depletion timescales, as shown in the Kennicutt-Schmidt Relation. They are consistent with models that show SMG formation in short time periods \citep{nara15}, although they are at slightly higher redshift than expected.
The data for these two HyLIRG systems suggest that they represent two
different stages in the evolution of extreme starbursts driven by
major gas rich mergers in the early Universe. BRI1202--0725 appears to be a
relatively early stage merger, with a number of distinct galaxies
still observed in stars and gas. There are clear indications of
strong gravitational interaction between the galaxies likely driving
the extreme starbursts, as well as possible evidence for a strong
QSO driven outflow assisting in quenching the star formation in the
QSO host. BRI1335--0417 appears to be a later stage merger, with just one
galaxy seen (the QSO host), plus what may be tidal remnants of the
merger seen in the extended cold gas.
The high luminosities, disrupted morphologies, evidence of gravitational interaction, and short gas depletion timescales of these objects suggest that they represent a transient, but highly star forming phase of early galaxy evolution.
\vspace{2mm}
G. C. J. is grateful for support from NRAO through the Grote Reber Doctoral Fellowship Program. R. G. M. acknowledges the UK Science and Technology Facilities Council (STFC). K. O. acknowledges the Kavli Institute Fellowship at the Kavli Institute for Cosmology in the University of Cambridge supported by the Kavli Foundation. The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc. We thank all those involved in the VLA project for making these observations possible (project code 13A-012).
|
\section{Introduction}
Quantum-limited detectors and amplifiers are important modules for practical quantum information architectures. High-efficiency signal processing implemented with these systems has enabled single shot readout \cite{UCSBJeffrey2014} and real-time feedback control \cite{DelftRiste2013} of quantum bits in recent times. Realizing quantum-limited detection is intimately tied to the minimality of the mode space of amplification, as each additional mode introduced into the system potentially brings along its associated noise, the minimum being the quantum noise or zero-point fluctuations of the mode. Parametric systems achieve quantum-limited amplification by splitting a pump photon(s) between two channels, the (desired) signal and the (auxiliary) idler, a process which leads to the well-known quantum limit of half-a-photon of added noise at the signal frequency \cite{Caves1982}. Such amplification, in general, is described as a scattering between input and output signal and idler field amplitudes $a_{\rm sig,idl}$,
\begin{eqnarray*}
\left(\begin{array}{c}
a_{\rm sig}\\
a_{\rm idl}
\end{array}\right)^{\rm out}
=
\left(\begin{array}{cc}
\sqrt{\mathcal{G}+1} & \sqrt{\mathcal{G}} \\
\sqrt{\mathcal{G}} & \sqrt{\mathcal{G} + 1}
\end{array}\right)
\left(\begin{array}{c}
a_{\rm sig}\\
a_{\rm idl}
\end{array}\right)^{\rm in}.
\end{eqnarray*}
where $\mathcal{G} > 1$ denotes the gain of the amplifier. As evident, such scattering is symmetric between signal and idler. Breaking this symmetry and realizing directional amplification is of immediate relevance to multiple quantum information processing (QIP) platforms, as it would (i) ensure unidirectional information transfer, (ii) prevent any noise impinging on the output port from getting amplified and re-directed to the signal-source (such as qubits), and (iii) significantly simplify measurement chains by eliminating the need of bulky components such as circulators and isolators, ultimately paving the way towards fully-integrated QIP.
\par
It is worth noting that ideal directionality, while remaining strictly confined to two modes,
\begin{eqnarray*}
\left(\begin{array}{c}
a_{\rm sig}\\
a_{\rm idl}
\end{array}\right)^{\rm out}
=
\left(\begin{array}{cc}
0 & 0\\
\sqrt{\mathcal{G}} & \sqrt{\mathcal{G} + 1}
\end{array}\right)
\left(\begin{array}{c}
a_{\rm sig}\\
a_{\rm idl}
\end{array}\right)^{\rm in},
\end{eqnarray*}
is forbidden by the requirement of symplectic structure of scattering \footnote{Symplectic symmetry follows from the bosonic commutation relations and the associated conserved quantity is the mode space of amplification.}. Thus the challenge is to realize nonreciprocal signal transfer and amplification while introducing a minimum number of additional modes and preserve the quantum-limited operation of the device. Given the application potential of such systems, recent years have witnessed a strong surge in theoretical \cite{Koch2010, Kamal2011, Kamal2012, Ranzani2014, Metelmann2015, Kerckhoff2015} and experimental \cite{Abdo2013, Tzuang2014, Fleury2014, Sliwa2015} efforts that have aimed to realize quantum-limited nonreciprocity at acoustic, microwave and optical frequencies.
\par
In this work, we analyze nonreciprocal photon dynamics in a framework that emphasizes minimality of the mode space and parametric pumping --- a feature especially desirable for hardware-efficient and scalable implementations of such detection protocols. Considering a generic system of parametrically-coupled three harmonic modes, which is the natural next step in increased mode complexity, we show how a two-pump biharmonic drive of the form
\begin{equation}
G(t) = G_{\omega_{p}} \cos(\omega_{p}t) + G_{2\omega_{p}} \cos(2\omega_{p} t + \alpha),
\label{EqBipump}
\end{equation}
suffices to implement various kinds of nonreciprocal couplings. Such biharmonic drives (${\alpha \neq n\pi}$) are an economical way to realize time-asymmetric driving. This has been exploited to various ends previously, such as realizing noise-induced ratchet dynamics in Brownian motors \cite{Hanggi2005}, directed diffusion of cold atoms in optical lattices \cite{Schiavoni2003}, manipulation of fluxon transport in annular Josephson junctions \cite{Ustinov2004} and asymmetric driving of Landau-Zener-St\"{u}ckelberg-Majorana (LZSM) interferences in double quantum dots \cite{Forster2015} and superconducting qubits \cite{Gustavsson2013}. An additional advantage associated with using biharmonic drives is their autonomous generation in nonlinear optical crystals \cite{SHG} and Josephson junctions in the voltage state \cite{Kamal2014}.
\par
The paper is organized as follows: in section \ref{sec_Raman} we describe directional phase-preserving amplification realizable in a new class of amplifiers, which we call biharmonic Raman amplifiers. We present calculations for both unresolved and resolved sideband regimes, and compare the available directionality with each under inclusion of relevant frequency-dependent non-resonant corrections. In section \ref{sec_phasesensitive} we describe directional phase-sensitive amplification with a biharmonically-pumped three-mode system. In section \ref{sec_discussion}, we discuss generic behavior and tradeoffs concerning gain, bandwidth and directionality. We also establish the connection of biharmonic pumping schemes to recently proposed dissipation engineering frameworks \cite{Metelmann2015}, and show that the general recipe of balancing dissipative and coherent interactions for implementing nonreciprocity simply maps to tuning the amplitude ratio ($G_{2 \omega_{p}}/G_{\omega_{p}}$) and phase difference ($\alpha$) of the two
harmonics. We conclude with a summary of our results in section \ref{sec_conclusions}. Additional details are included in appendices \ref{app_SRA} and \ref{app_circ}.
\section{Directional phase-preserving amplification: Biharmonic Raman Amplifiers}
\label{sec_Raman}
Phase-preserving amplification refers to equal amplification of both quadratures of a photonic field; this process maintains the phase information of the amplified signal in the quadrature space. Our general scheme to realize a directional phase-preserving amplifier is best understood in the framework of stimulated Raman scattering. It involves using a pump tone blue-detuned from the lower sideband resulting in Stokes scattering of the pump photons and red-detuned from the upper sideband leading to anti-Stokes scattering [cf. Fig.~\ref{Fig.:Raman}]. While the single-pump Stokes and anti-Stokes scattering is symmetric, it has been shown previously that the addition of the second pump harmonic induces asymmetry in these conversion processes \cite{Kamal2014}. Here we elaborate how such a process leads to directional amplification in the reduced subspace of the two sidebands. The basic idea relies on balancing (or `interfering') an indirect interaction between the two sidebands mediated by the first pump harmonic (through a third auxiliary mode), with a direct coherent interaction between the sidebands mediated by the second pump harmonic. The indirect interaction mediated by the auxiliary mode models a dissipative interaction as discussed later in Sec. \ref{sec_discussion}.
\par
We discuss two different regimes of these amplifiers:
\begin{enumerate}[(i)]
\setlength{\itemsep}{0pt}
\item Unresolved sideband amplification, i.e. $\omega_{a} \ll \kappa_{b}$, where $\kappa_{b}$ denotes the linewidth of the high frequency oscillator at $\omega_{b}$
and $\omega_a$ corresponds to the resonant frequency of the auxiliary mode [cf. Fig.~\ref{Fig.:Raman}(a)]. This regime corresponds to \emph{degenerate phase-preserving amplification} since both input and output channels are accessed through a single mode \cite{DevoretRoy2016}.
\item Resolved sideband amplification, i.e. $\omega_{a} \gg \kappa_{b}$. This regime corresponds to \emph{non-degenerate phase-preserving amplification} since input and output channels are accessed via two distinct modes
[cf. Fig.~\ref{Fig.:Raman}(b) ].
\end{enumerate}
\subsection{Unresolved sideband (USB) amplification}
\begin{figure}
\centering\includegraphics[width=\columnwidth]{Figure1.pdf}
\caption{Frequency landscape of Raman amplifiers in (a) unresolved sideband regime ($\kappa_{a} \ll \omega_{a} < \kappa_{b}$), and (b) resolved sideband regime ($\kappa_{a} \ll \kappa_{1,2} < \omega_{a}$). In both the cases, the presence of an additional tone at $2\omega_{p}$ leads to directional amplification between $\omega_{-} (\omega_{2})$ and $\omega_{+} (\omega_{1})$. }
\label{Fig.:Raman}
\end{figure}
We start with a generic Hamiltonian describing three harmonic modes coupled via time-varying (pairwise) interactions of the form,
\begin{align} \label{Eq.:3modeHamiltonian}
\hat{\mathcal{H}}_{s} =&
G_{1} (t) \left[\hat a + \hat a^{\dag} \right] \left(\hat d_1 + \hat d_1^{\dag} \right)
+ G_2(t) \left(\hat d_1 + \hat d_1^{\dag} \right) \left(\hat d_2 + \hat d_2^{\dag} \right)
\nonumber \\ &
+ G_3(t) \left[\hat a + \hat a^{\dag} \right] \left(\hat d_2 + \hat d_2^{\dag} \right),
\end{align}
where $\hat a (\hat d_{1,2}) $ denotes the photon annihilation operator of the mode with frequency $\omega_a (\omega_{1,2})$. The modulations $G_{j}(t); (j \in 1; 2; 3)$ include the first and/or the second harmonic of an external pump at frequency $\omega_{p}$, cf. Eq.~(\ref{EqBipump}).
\par
We first consider the scheme depicted in Fig.~\ref{Fig.:Raman}(a). This system, in principle, is a two-mode system and can be realized using two parametrically coupled oscillators with frequencies $\omega_{a,b}$. In the unresolved sideband limit, where $\omega_{a}/\kappa_{b} \ll1$, we can consider the low-frequency mode at $\omega_{a}$ and the two sidebands at $\omega_{\pm} = \omega_{b} \pm \omega_{a}$ forming an effective three-mode system. Treating the sidebands as independent modes $\hat{b}_{\pm}$, we can map the system to the general three mode interaction Hamiltonian of Eq.~\ref{Eq.:3modeHamiltonian} via the correspondence
\begin{align}\label{Eq.MappingUSB}
\hat d_1 \ \rightarrow \ \hat b_{+} = \hat b e^{ i\omega_a t },
\hspace{0.5cm}
\hat d_2 \ \rightarrow \ \hat b_{-} = \hat b e^{-i\omega_a t },
\end{align}
In the presence of a biharmonic drive of the form indicated in Eq.~(\ref{EqBipump}) and with $\omega_p = \omega_b$, the first pump harmonic induces Stokes (anti-Stokes) scattering to the lower (upper) sideband, while the second pump mediates an amplifying interaction between the two sidebands since $2\omega_{p} =\omega_{+} + \omega_{-}$. This leads to an effective mixing Hamiltonian in the interaction picture, with respect to the free Hamitonian, as
\begin{align}\label{EqHamil}
\hat{\mathcal{H}}_{s} = & \; G_{1} \left( \hat{a} e^{-i \omega_{a} t} + \hat{a}^{\dagger} e^{i \omega_{a}t} \right)
\left(\hat{b} + \hat{b}^{\dagger} \right)
\nonumber\\ & \qquad +
\frac{G_{2}}{2} \left( \hat{b}\hat{b} \ e^{i\alpha} + \hat{b}^{\dagger}\hat{b}^{\dagger} \ e^{-i\alpha} \right).
\end{align}
The first line in Eq.~(\ref{EqHamil}) describes up- and downconversion processes between the low frequency mode $\hat a$ and the sidebands of the high frequency mode at $\omega_{\pm} = \omega_b \pm \omega_a$. The second line corresponds to an additional mixing pathway between these sidebands. A combination of both these interactions results in asymmetric frequency conversion between auxiliary mode and sidebands \cite{Kamal2014}. Moreover, it allows for directional amplification within the sidebands. It is worthwhile to note here, that for $G_2 = 0$ and $\omega_P = \omega_b$, i.e., a monochromatic driving on resonance with the high frequency mode, the system realizes an effective two-mode phase-preserving amplifier which is not directional, but still has the interesting property of having no gain-bandwidth limitation.
The monochromatic driving scheme is closely related to a kind of dissipative amplifier introduced recently \cite{Metelmann2014} (see appendix \ref{app_SRA} for further details).
\par
Setting the coupling strengths of the two oscillators to the external input/output ports as $\kappa_{a,b}$, we can use standard input-output theory \cite{IOT} to derive the Heisenberg-Langevin equations describing the dynamics of our system. To highlight the relevant features of this setup, we start by focusing on the zero-frequency case. Thus the low frequency mode is on resonance and has the stationary solution,
\begin{align}
\hat a[0] =&
- i \frac{2 G_1}{\kappa_a} \left( \hat b[ \omega_a] +\hat b^{\dag}[ \omega_a] \right),
\label{Eqao}
\end{align}
where, we neglect any noise contribution driving the low frequency oscillator for simplicity. We see that the mode $\hat a$ is coupled to the two sideband lying at $\pm \omega_a$ in this rotated frame. Correspondingly, the sidebands couple to $\hat a[0]$ and $\hat a^{\dag}[2 \omega_a]$. For $\omega_a \gg \kappa_a$, the processes mixing Stokes and anti-Stokes with $\hat a^{\dag}[2 \omega_a]$ can be ignored under a rotating wave approximation in the $\hat{b}_{\pm}$ basis \footnote{This is equivalent to treating the two sidebands as independent modes.}. In this limit, we can use the stationary solution for $\hat a[0]$ to obtain the corresponding solutions for the sidebands
\begin{align} \label{Eq.:EoMstationarySidebandNDPA}
\chi_{+}^{-1} \hat b[ \omega_a] =& - \frac{2}{\sqrt{\kappa_b}} \hat b_{\rm in}[ \omega_a]
- \left[ \mathcal C + i \frac{2 G_2 }{\kappa_b} e^{-i \alpha} \right] \hat b^{\dag}[ \omega_a],
\nonumber \\
\chi_{-}^{-1} \hat b^{\dag}[ \omega_a] =& - \frac{2}{\sqrt{\kappa_b}} \hat b_{\rm in}^{\dag}[\omega_a]
+ \left[ \mathcal C + i \frac{2 G_2 }{\kappa_b} e^{i \alpha} \right] \hat b[\omega_a],
\end{align}
with the modified susceptibilities $ \chi_{\pm}^{-1} = \left[1 \pm \mathcal C -i \frac{2 \omega_a}{\kappa_b} \right] $ and the cooperativity $\mathcal C = \frac{4 G_1^2}{\kappa_a \kappa_b}$. The operators $\hat b_{\rm in}^{(\dag)} $ describe any input impinging at the sideband frequencies of the b-mode, i.e., an input signal one wishes to amplify or just thermal and vacuum fluctuations. Unlike the case of a single pump ($G_{2} =0$), there are differences in how each sideband couples to the other as reflected by the asymmetries in the terms within square brackets of Eq.~(\ref{Eq.:EoMstationarySidebandNDPA}). Moreover, this asymmetry is tunable with phase $ \alpha$ and the strength $G_{2}$ of the second pump. It is straightforward to see that for
\begin{align}\label{Eq.:DirCondNDPA}
G_2 = \frac{\kappa_b}{2} \mathcal C\;,
\hspace{0.5cm}
\alpha = - \frac{\pi}{2},
\end{align}
$ \hat b[\omega_a]$ decouples from the reduced system of the two sidebands. This realizes a directional interaction, as we now have the situation that $\hat b^{\dag}[\omega_a]$ is influenced by $\hat b [\omega_a]$ but not vice versa.
\par
In order to calculate the full nonreciprocal scattering matrix of the system, we use the standard input-output relation ${\hat o_{\rm out} = \hat o_{\rm in} + \sqrt{\kappa_o} \hat o, \ (o \in a,b)}$. Hereby we include the fluctuations impinging on the low frequency mode ($\hat a_{\rm in}$) as well, which we had neglected earlier. Then the zero-frequency scattering matrix in the basis $\mathbf{\hat D}[0] = [\hat a^{\phantom{\dag}}[0], \hat b^{\phantom{\dag}}[\omega_a],\hat b^{\dag}[\omega_a]]^{\rm T}$ becomes (for $\omega_{a}/\kappa_{b} \ll1$)
\begin{equation} \label{Eq.:SmatrixSidebandNDPA}
\mathbf{s}[0] =
\left(
\begin{array}{ccc}
\displaystyle -1
& \displaystyle
\frac{ i 2 \sqrt{\mathcal C}}
{1 - \mathcal C}
& \displaystyle
\frac{ i 2 \sqrt{\mathcal C}}
{1 - \mathcal C}
\\[0.2cm]
\displaystyle \frac{ i 2 \sqrt{\mathcal C}}
{1 + \mathcal C}
& \displaystyle -\frac{1 - \mathcal C}
{1 + \mathcal C}
& \displaystyle 0
\\[0.2cm]
\displaystyle \frac{ -i 2 \sqrt{\mathcal C}}
{1 + \mathcal C }
& \displaystyle - \frac{ 4 \mathcal C}
{1- \mathcal C^2}
& \displaystyle - \frac{1 + \mathcal C}
{1 - \mathcal C}
\\
\end{array}
\right),
\end{equation}
where $\mathbf{\hat D}_{\rm out}[0] = \mathbf{s}[0] \mathbf{\hat D}_{\rm in}[0]$. The diagonal elements of this matrix correspond to the reflection coefficients, the off-diagonal elements $s_{21},s_{31}$ describe upconversion from the low frequency to the sideband frequencies while $s_{12},s_{13}$ describe the corresponding down-conversion. The important elements are $s_{23}$ and $s_{32}$ which describe the amplification between the two sidebands. An input signal at the upper sideband shows up amplified at the lower sideband, i.e., it gets down-converted in frequency, while any input on the lower sideband will never show up at the higher sideband as $s_{23} = 0$. Note, that for $\alpha = \frac{\pi}{2}$ the situation is reversed, i.e., a signal is up-converted between the sidebands and amplified, but this leads to unwanted amplification of the reflected input signal too.
\par
The zero-frequency gain can be read off from Eq.~(\ref{Eq.:SmatrixSidebandNDPA}) as
\begin{align}\label{Eq.:ZeroFrequGain}
\mathcal G_{0} \equiv \left| s_{32} [0]\right|^2 =& \frac{ 16 \mathcal C^2 }{ \left[ \mathcal C^2 - 1 \right]^2},
\end{align}
which increases as $\mathcal C \rightarrow 1^{-}$; as usual stability requires $\mathcal C < 1 $. From the scattering matrix in Eq.~(\ref{Eq.:SmatrixSidebandNDPA}) we see that in the large gain limit the output at the upper sideband contains noise stemming from the low frequency auxiliary mode alone; also the reflection at the input vanishes ($s_{22} =0$), a feature desirable for applications such as qubit readout. Moreover, this amplification process is quantum-limited, as can be seen by calculating the added noise,
\begin{align}
\bar n_{\rm add} =& \frac{1}{2} +\bar n_{b}^{T}
+ \left(2\bar{n}_{b}^{T} + \bar n_{a}^{T} + \frac{3}{2} \right) \frac{1}{\mathcal G_0} + \mathcal{O}\left(\frac{1}{\mathcal{G}_{0}^{2}}\right).
\end{align}
In the regime of large gain ($\mathcal G_0$) and zero-temperature baths ($ n_{a}^{T} = n_{b}^{T} \approx 0$), we obtain the quantum limit of $\bar n_{\rm add}= 1/2$.
\subsubsection*{Frequency dependence of directional gain}
Finally, we take a look at the expressions for the gain and the reverse gain as a function of frequency. We still consider the situation where we have an input signal at the upper sideband which gets amplified and completely down-converted to the lower sideband under the directionality condition Eq.~(\ref{Eq.:DirCondNDPA}). The finite frequency gain is ($\omega_a \ll \kappa_a$)
\begin{align}\label{Eq.:GainSidebandNDPA}
\mathcal G[\omega] =& \frac{ 16 \mathcal C^2 \left[1 + \frac{\omega^2}{\kappa_a^2} \right] \left[ 1 + \frac{4 \omega^2}{\kappa_a^2} \right]^{-1}}
{\left[ \left[\mathcal C - 1 \right]^2 + \frac{4\left(\omega + \omega _a \right)^2}{ \kappa _b^2} \right]
\left[ \left[\mathcal C + 1 \right]^2 + \frac{4\left(\omega + \omega _a \right)^2}{ \kappa _b^2} \right]},
\end{align}
where we keep the ratio of $\omega_a/\kappa_b$ unspecified, i.e., we do not restrict ourselves to unresolved sideband regime. From the full expression for the frequency-dependent gain, we see that the gain profile shows a peak at ${\omega = - \omega_a}$ which corresponds to the resonance frequency $\omega_b$ in this rotated frame [cf. Fig.~\ref{Fig:SmatNDPA}].
\begin{figure}[t!]
\centering\includegraphics[width=0.95\columnwidth]{Figure2.pdf}
\caption{Frequency dependence for the scattering matrix elements of the unresolved sideband amplification.
The dashed-gray line depicts the gain $|s_{32}[\omega]|$ under RWA, cf. Eq.~(\ref{Eq.:GainSidebandNDPA}).
The non-RWA results for forward and reverse gains are plotted as the solid-black and the dotted-blue lines
respectively. The scattering matrix elements at the sideband resonances $\omega_{b}\pm\omega_a$ describe
downconversion to the lower sideband for an input signal injected at the upper sideband.
Parameters used in the calculation are $\mathcal C = 0.9$, $\omega_a/\kappa_b = 0.1$ , $Q_a = 10$.}
\label{Fig:SmatNDPA}
\end{figure}
However, the reverse gain only vanishes at the lower sideband $(\omega = 0)$,
\begin{align}\label{Eq.:RevGainSidebandNDPA}
\bar{\mathcal G}[\omega] \equiv \left|s_{23}[\omega] \right|^2
= \frac{ \frac{ \omega^2}{\kappa_a^2}}{ \left(1 + \frac{\omega^2}{\kappa_a^2} \right)} \mathcal G[\omega],
\end{align}
which describes the up-conversion of possible inputs, i.e., thermal or vacuum fluctuations from the lower sideband. The ideal situation corresponds to a vanishing of this reverse gain over a wide frequency bandwidth. The directionality bandwidth scales with $\kappa_a$, as evident from Eq.~(\ref{Eq.:RevGainSidebandNDPA}). However, in order to treat both sidebands independently we need to have $\kappa_{a} \ll \omega_{a}$, i.e. a high-quality factor for low frequency $\hat{a}$ mode. One would think that having this mode at high frequency may do the trick and a large directionality bandwidth could, in principle, be maintained. However, just having a large $\omega_a$ is not sufficient, as for $\omega_a \rightarrow \infty $ the gain vanishes as well. The relevant quantity here is the ratio $\omega_a/\kappa_b$; this becomes obvious if we consider the gain at resonance [cf. Eq.~(\ref{Eq.:GainSidebandNDPA})] for $\mathcal C \rightarrow 1^{-}$,
\begin{align}
\mathcal G[0] =& \frac{1}{\frac{ \omega_a^2}{ \kappa_b^2} \left( 1 + \frac{ \omega_a^2}{ \kappa_b^2} \right)} \approx \frac{ \kappa_b^2}{\omega_a^2}.
\label{Eq.ZerogainUSB}
\end{align}
The gain saturates and the maximal gain value scales with $(\kappa_b/\omega_a)^2$ for $\mathcal C \rightarrow 1^{-}$. Thus, the unresolved sideband regime is an important ingredient to obtain any gain at all in this scheme.
\subsection{Resolved sideband (RSB) amplification}
\label{Sec.:3modeNDPA}
Our analysis in the previous section showed that the parameter hierarchy $\kappa_a \ll \omega_a < \kappa_b$ is crucial to realize a directional phase-preserving amplification between the sidebands. The restriction to the unresolved sideband regime, however, constrains both the forward gain and directionality of such a biharmonic Raman amplifier. For instance a $20$~dB of gain would already require a ratio $\kappa_b/\omega_a \simeq 10$; though achievable in opto/electro-mechanical setups, this limits the application potential of such a scheme in superconducting setups employing microwave frequencies. In this section, we show how operating in the resolved sideband regime alleviates these difficulties. We now extend our system to include three independent oscillators with resonance frequencies $\omega_a \ll \omega_2 < \omega_1$. As shown in Fig.~\ref{Fig.:Raman}(b), the independent oscillator modes at frequencies $\omega_{1,2}$ play the role of the sidebands of the USB amplifier case. Choosing the driving frequencies $\omega_{P,i},(n \in (1,2,3)) $ of the time-dependent couplings $G_{j}(t)$ in Eq.~(\ref{Eq.:3modeHamiltonian}) as
\begin{subequations}
\begin{align}
\omega_{P,1} =& \ \omega_1 - \omega_a \equiv \omega_0,
\\
\omega_{P,2} = & \ \omega_2 + \omega_a \equiv \omega_0,
\\
\omega_{P,3} = & \ \omega_1 + \omega_2 \equiv 2 \omega_0,
\end{align}
\end{subequations}
makes the following interactions resonant in the system, just as in the unresolved sideband regime:
\begin{enumerate}[(i)]
\setlength{\itemsep}{0pt}
\setlength{\parskip}{0pt}
\item the first harmonic at $\omega_0$ mediates a hopping interaction between the auxiliary mode $\hat a$ and mode $\hat d_1$
\item the first harmonic at $\omega_0$ mediates an amplifier interaction between the auxiliary mode $\hat a$ and mode $\hat d_2$
\item the second harmonic at $2\omega_0$ mediates an amplifier interaction between the modes $\hat d_2$ and $\hat d_1$.
\end{enumerate}
As before, we work in an interaction picture with respect to the oscillators' free Hamiltonian and obtain
\begin{align}
\hat{\mathcal{H}} =& G_1 \left( \hat a \hat d_1^{\dag} + \hat a^{\dag} \hat d_2^{\dag} \right)
+ G_2 \hat d_1^{\dag} \hat d_2^{\dag} e^{- i \alpha} + \hat{\mathcal{H}}_{\rm CR} + h.c.,
\label{Eq.:Ham3mode}
\end{align}
where $ \hat{\mathcal{H}}_{\rm CR}$ contains the counter-rotating terms.
\par
We consider the same situation as before, i.e., the directional amplification of an input signal at the upper mode with frequency $\omega_1$ which shows up at the output of the mode at frequency $(\omega_2)$. Using input-output theory and utilizing the directionality condition of Eq.~(\ref{Eq.:DirCondNDPA}), we obtain the same zero-frequency scattering matrix as in Eq.~(\ref{Eq.:SmatrixSidebandNDPA}) but now in the basis $\mathbf{\hat D}[0] = [\hat a^{\phantom{\dag}}[0], \hat d_{1}^{\phantom{\dag}}[0],\hat d_{2}^{\dag}[0]]^{\rm T}$. The finite frequency gain for RSB amplifier is given as (neglecting $ \hat{\mathcal{H}}_{\rm CR}$)
\begin{align}\label{Eq.:Gain3modeNDPA}
\mathcal G[\omega] =&
\frac{16 \mathcal C ^2 \left(1+ \frac{\omega^2}{\kappa_a^2}\right)}
{\left[1+ \frac{4 \omega^2}{\kappa_a^2}\right]
\left[ \left(\mathcal C - 1\right)^2 + \frac{4\omega^2}{\kappa_b^2} \right]
\left[ \left(\mathcal C + 1\right)^2 + \frac{4\omega^2}{\kappa_b^2} \right] },
\end{align}
where, for simplicity, we have assumed equal decay rates for both the oscillator modes 1 and 2, $\kappa_{1} =\kappa_{2} \equiv \kappa_{b}$. The reverse gain of RSB amplifier is the same as that given in Eq.~(\ref{Eq.:RevGainSidebandNDPA}) with $\mathcal{G} [\omega]$ now given by Eq.~(\ref{Eq.:Gain3modeNDPA}). Thus we have the same situation as in the USB case; the reverse gain vanishes on resonance and the directionality bandwidth increases with $\kappa_a$. The important difference, however, is that the gain in Eq.~(\ref{Eq.:Gain3modeNDPA}) does not depend on the ratio $\omega_a/\kappa_b$, in contrast to Eq.~(\ref{Eq.:GainSidebandNDPA}). Thus, there is no saturation of the gain as encountered in the unresolved sideband regime [cf. Eq.~(\ref{Eq.ZerogainUSB})].
\par
We discuss another mode of operation of this system, namely a nonreciprocal photon transmission without gain or a frequency circulator, in appendix \ref{app_circ}.
\subsection{Influence of counter-rotating terms: USB versus RSB}
All the calculations in previous sections were done neglecting the counter-rotating terms. These terms oscillate with twice of one of the systems frequencies, i.e., $e^{i 2\omega_{j} t}, (j \in a,1,2,0)$; since the smallest frequency is $\omega_a$, the terms oscillating at $2\omega_{a}$ are the most relevant counter-rotating terms. Under this assumption, we describe the counter-rotation Hamiltonian as
\begin{align}
\hat{\mathcal{H}}_{\rm CR} \simeq&
G_1 \left( \hat a^{\dag} \hat d_1^{\dag} e^{ i 2 \omega_a t } + \hat a \hat d_2^{\dag} e^{-i 2 \omega_a t } \right) + {\rm h.c.},
\label{Eq.:Hamilcounter}
\end{align}
\begin{figure}[t!]
\centering
\includegraphics[width=\columnwidth]{Figure3.pdf}
\caption{Comparison of USB and RSB biharmonic Raman amplifiers. Parameters are $\mathcal C = 0.9$ and $\kappa_b = 100 \kappa_a$. The upper (lower) graph depicts the forward (reverse) gain at resonance, as a function of quality factor of the auxiliary mode $Q_{a} =\omega_{a}/\kappa_{a}$. In the USB case, the gain first improves as $Q_{a}$ increases and then decreases and reaches the limit $\kappa_b^2/\omega_a^2$ (grey dashed line) for large $Q_{a}$. The RSB amplifier, on the other hand, suffers from no gain saturation and it operates at the expected maximal gain value, i.e., $\mathcal G \sim \mathcal G_0 \approx 360$ for high $Q_{a}$. Further, a comparison of the RWA and non-RWA results shows that $\omega_a > \kappa_a$ is necessary to suppress reverse gain in both the USB and RSB schemes (under RWA, the reverse gain is always zero).}
\label{Fig.:ComparisonRaman}
\end{figure}
\par
Fig.~\ref{Fig.:ComparisonRaman} compares the forward and reverse gains as a function of $Q_{a}$, calculated including the effect of Eq.~(\ref{Eq.:Hamilcounter}) for both the USB- and RSB types of biharmonic Raman amplifiers. It is clear that the counter-rotating terms lower the forward gain of both amplifiers unless filtered out by a sufficiently high-$Q_{a}$ for the auxiliary mode,$\omega_a/\kappa_a \gg 1$. Furthermore, though (reduced) forward gain persists in low-$Q_{a}$ regime, the reverse gain vanishes and directionality is restored only in the high-$Q_{a}$ limit.
\par
While the behaviors of the RSB amplifier and the USB amplifier coincide in the low-$Q_{a}$ regime, the gain for the USB amplifier decreases strongly with increasing $Q_{a}$. The saturation of the gain sets in as $\omega_{a}/\kappa_{a} \rightarrow \kappa_{b}/\kappa_{a}$ [cf. Eq.~(13)]. This significantly limits the useful bandwidth over which it can be exploited as a \emph{directional} phase-preserving amplifier. Operating in the RSB regime alleviates this problem and drastically increases the bandwidth over which the amplifier is directional, though it still requires a modest $Q_{a}$ for the auxiliary mode.
\section{Directional phase-sensitive amplification}
\label{sec_phasesensitive}
We now present a biharmonically-pumped three-mode scheme that realizes directional phase-sensitive amplification, i.e. only one of the input quadratures gets amplified and appears at the output. Originally proposed in Ref. \cite{Metelmann2015}, the key idea here is to realize a quantum non-demolition (QND) interaction between the input and output modes. In the most general case, at least six driving tones are required to mediate the requisite interactions. In this section, we show how such an interaction can be realized using a biharmonic tone, with only a single constraint on the auxiliary mode frequency $\omega_{a}$ being degenerate with one of the input or output modes. To this end, we consider the most general three-mode Hamiltonian of Eq.~(\ref{Eq.:3modeHamiltonian}) under the pumping conditions,
\begin{subequations}
\begin{align}
& \omega_{1} + \omega_{a} \equiv 2 \omega_{0}, \quad \omega_{1} - \omega_{a} = \omega_{0}\\
& \omega_{1} + \omega_{2} \equiv 2 \omega_{0}, \quad \omega_{1} - \omega_{2} = \omega_{0}\\
& \omega_{2} + \omega_{a} \equiv \omega_{0}, \quad \omega_{2} - \omega_{a} = 0,
\end{align}
\end{subequations}
which selects the following interactions to be resonant:
\begin{enumerate}[(i)]
\setlength{\itemsep}{0pt}
\setlength{\parskip}{0pt}
\item the first harmonic at $\omega_0$ mediates a hopping interaction while the second harmonic mediates an amplifying interaction between the auxiliary mode $\hat a$ and mode $\hat d_1$
\item the first harmonic at $\omega_0$ mediates a hopping interaction while the second harmonic mediates an amplifying interaction between modes $\hat{d}_{1}$ and $\hat d_2$
\item the first harmonic at $\omega_0$ mediates an amplifying interaction between the the auxiliary mode $\hat a$ and mode $\hat d_2$
\end{enumerate}
This pumping scheme can be easily realized for any combination of mode frequencies of the form $\omega_{a} = \omega_{2} = (1/3) \omega_{1}$. Note that the interaction between the auxiliary mode and mode 2 can only be of the amplifying type in this case since these modes are designed to be at the same frequency.
\begin{figure}[t!]
\centering
\includegraphics[width=0.8\columnwidth]{Figure4.pdf}
\caption{The direct interaction between the modes 1 and 2 (solid-blue) alone corresponds to an information transfer where the information of the X-quadrature of each mode is dumped into the P-quadrature of the respective opposite mode. The coupling to the auxiliary mode breaks this reciprocal process by mediating the same transfer process between the modes (solid-cyan). Balancing these process leads to perfect decoupling of mode 1 from mode 2. The dashed lines show the feedback loops which are unavoidable in this minimal biharmonic driving scheme, however, they are not damaging for the whole scheme to work.}
\label{Fig.:DPSA}
\end{figure}
This leads to following interaction Hamiltonian (under RWA),
\begin{eqnarray}
\hat{\mathcal{H}} = G \hat{X}_{1}\hat{X}_{2} - G_{3} [\hat{X}_{2} \hat{V} + \hat{P}_{2}\hat{U}] + G \hat{X}_{1} \hat{U},
\label{EqHPSAquad}
\end{eqnarray}
with $\hat{X}_{i} = (\hat{d}_{i} + \hat{d}_{i}^{\dagger})/\sqrt{2}, \hat{P}_{i} = -i(\hat{d}_{i} - \hat{d}_{i}^{\dagger})/\sqrt{2}, \; (i \in 1,2)$ being the quadratures associated with input-output modes, and $(\hat{U}, \hat{V})$ the quadratures associated with the auxiliary mode. Here we have chosen $G_{1} = G_{2} \equiv G/2$ and phase difference between $G$ and $G_{3}$ to be $\pi/2$ [cf. Eq.~(\ref{Eq.:3modeHamiltonian})]. Eq.~(\ref{EqHPSAquad}) shows that $\hat{X}_{1}$ is a QND observable and is, therefore, preserved from quadrature mixing as desired for phase-sensitive amplification. In the optimal case $\hat{X}_{2}$ would be as well a QND-observable \cite{Metelmann2015}; this can be accomplished by balancing out the term $\hat{P}_{2}\hat{U}$ either through a static coupling \cite{Abdo2013} between modes $\hat{a}$ and $\hat{d_{2}}$, or by lifting their degeneracy at the expense of introducing additional pumps. Figure~\ref{Fig.:DPSA} illustrates the information transfer mediated by the Hamiltonian in Eq.~(\ref{EqHPSAquad}). The information in the X-quadratures of each mode is transferred to the respective P-quadratures of the other mode in two ways: via a direct transfer and via a transfer over the auxiliary mode. Balancing these interactions as before allows us to realize desired unidirectional information transfer between selected quadratures. This can be easily seen from the zero-frequency scattering matrix calculated from the coupled equations of motion for the quadratures, which after the elimination of auxiliary mode reads
\begin{equation} \label{Eq.:SmatrixSidebandDPA}
\mathbf{s}[0] =
\left(
\begin{array}{cccc}
\displaystyle -1
& \displaystyle
0
& \displaystyle 0
& \displaystyle 0
\\[0.2cm]
\displaystyle 0
& \displaystyle -1
& \displaystyle 0
& \displaystyle 0
\\[0.2cm]
\displaystyle 0
& \displaystyle 0
& \displaystyle - \frac{\kappa +\kappa_{a}}{\kappa -\kappa_{a}}
& \displaystyle 0
\\[0.2cm]
\displaystyle \frac{8G}{\kappa -\kappa_{a}}
& \displaystyle 0
& \displaystyle 0
& \displaystyle - \frac{\kappa +\kappa_{a}}{\kappa -\kappa_{a}}
\\
\end{array}
\right),
\end{equation}
where $\mathbf{\hat D}_{\rm out} = \mathbf{s} \mathbf{\hat D}_{\rm in} + \mathbf{\xi}$,
$\mathbf{\hat D}_{\rm i \in ( \rm {in, out} )} = \left[\hat{X}_{1}^{\rm i}, \hat{P}_{1}^{\rm i}, \hat{X}_{2}^{\rm i}, \hat{P}_{2}^{\rm i}\right]^T$,
$\kappa_{1} = \kappa_{2} \equiv \kappa$, and $\mathbf{\xi}$ denotes the noise contribution from the auxiliary mode. Here we have set the interaction strength $G_{3}=\kappa_{a}/2$ to impose unidirectional coupling from mode 1 to mode 2. This results in directional phase-sensitive amplification: the $\hat{P}_{2}^{\rm out}$ contains the amplified input quadrature $\hat{X}_{1}^{\rm in}$, while no input on either of the cavity-2 quadratures shows up at the output of cavity 1. The amplitude gain scales with $G$ and stability requires $\kappa_{a} \ll \kappa$ as the paramp interaction between the auxiliary mode and cavity-2 mode introduces anti-damping of the latter. Moreover, the added noise
\begin{eqnarray}
\bar{n}_{\rm add} = \left(\frac{\kappa + \kappa_{a}}{8 G}\right)^{2}\left(\bar{n}_{2}^{T} + \frac{1}{2} \right) + \frac{\kappa_{a}\kappa}{16 G^{2}} \left(\bar{n}_{a}^{T} + \frac{1}{2} \right)
\end{eqnarray}
goes to zero in the high-gain limit $G \gg \kappa,\kappa_{a}$ [cf. Fig.~\ref{Fig.:DPSAchar}(c)], as desired for ideal phase-sensitive amplification. The expressions for frequency-dependent forward gain and reverse gain,
\begin{eqnarray}
\mathcal{G} [\omega] &=& |s_{P_{2} \leftarrow X_{1}}[\omega]|^{2} \nonumber\\
& =& \frac{16 \mathcal{C}_{a} (1 + \frac{\omega^{2}}{\kappa_{a}^{2}})}{\left(1+ \frac{4 \omega^{2}}{\kappa_{a}^{2}}\right)\left(\left[\frac{4 \omega^{2}}{\kappa_{a}^{2}}+ \frac{\kappa}{\kappa_{a}} - 1\right]^{2} + \frac{4\omega^{2}}{\kappa_{a}^{2}}\left(1+\frac{\kappa}{\kappa_{a}}\right)^{2}\right)}, \nonumber\\
\label{Eq.:FwdGainPSA}\\
\mathcal{\bar{G}} [\omega] &=& |s_{P_{1} \leftarrow X_{2}}[\omega]|^{2} = \frac{\frac{\omega^{2}}{\kappa_{a}^{2}}}{1 + \frac{\omega^{2}}{\kappa_{a}^{2}}}\mathcal{G} [\omega],
\label{Eq.:RevGainPSA}
\end{eqnarray}
where $\mathcal{C}_{a} = 4 G^{2}/\kappa_{a}^{2}$. Crucially, the anti-damping does not scale with the gain (though we need $\kappa_{a} < \kappa$) leading to no limitation on the gain-bandwidth product for this system.
\par
\begin{figure}[ht]
\centering
\includegraphics[width=\columnwidth]{Figure5.pdf}
\caption{Characteristics for the directional phase-sensitive amplifier scheme, for three different ratios of $\kappa_{a}/\kappa$ = 0.1 (blue), 0.5 (red), 0.9 (cyan). For each plot, the calculation assumed directionality condition namely $G_3 = \kappa_a/2$. (a) Frequency-dependent forward and reverse gains as a function of $Q_a = \omega_a/\kappa_a$ for the auxiliary mode. The dashed gray line shows the gain without the counter rotating terms which is independent from $Q_a$, here $\mathcal G[0]^{\textup{RWA}}$ = 26 dB at resonance, while the corresponding reverse gain is always zero. We see that for moderate quality factors the reverse gain is highly suppressed, while the gain approaches the RWA result; hence we chose $\omega_a/\kappa_a = 10$ for the remaining graphs (b-d).
(b) Reverse gain evaluated at half of the amplification bandwidth, i.e., at $\omega = \Delta/2$ with $\Delta$ being the full-width at half-maximum of the forward gain profile. (c) Added noise quanta versus zero frequency gain for zero temperature baths. (d) Amplification bandwidth expressed in units of linewidth of auxiliary mode, $\Delta/\kappa_{a}$. Crucially, the bandwidth does not decrease while increasing the gain, hence we have no gain-bandwidth limitation.
}
\label{Fig.:DPSAchar}
\end{figure}%
Fig.~\ref{Fig.:DPSAchar} depicts the relevant figures of merit for the directional PSA, calculated including the relevant next sideband contributions, i.e., counter rotating terms associated with $ \omega_0 = 2 \omega_b = 2\omega_a$ up to first order. Fig.~\ref{Fig.:DPSAchar}(a) shows that the auxiliary mode $Q_{a}$ needs to be sufficiently high in order to obtain useful directionality; this coincides with the results found for phase-preserving amplification with biharmonic Raman amplifiers (cf. Fig.~\ref{Fig.:ComparisonRaman}). Furthermore, as shown by Figs.~\ref{Fig.:DPSAchar}(c,d), for a given $Q_{a}$ having too large a $\kappa_{a}$ is also detrimental from the point of view of bandwidth and noise properties of such an amplifier. On the other hand, having a too small $\kappa_{a}/\kappa$ ratio is unfavorable for directionality, as evident from the calculation of the reverse gain.
Our calculations show that the reverse gain is strongly suppressed for a large $\kappa_{a}/\kappa$. Note that this ratio is always limited to less than unity due to stability considerations; this constraint arises due to the feedback of the quadrature mixing term $\hat{P}_{2}\hat{U}$ in Eq.~(\ref{EqHPSAquad}). The desirable hierarchy of different frequency scales for stable directional operation, thus becomes $\kappa_{a} < \kappa < \omega_{a}$.
\section{Discussion}
\label{sec_discussion}
\subsection{Gain versus directionality bandwidth}
\label{sec_tradeoff}
The constraint of a constant gain-bandwidth product for biharmonic Raman amplifiers, discussed in Sec.~\ref{sec_Raman}, can be calculated using the expression for forward gain [cf. Eq.~(\ref{Eq.:Gain3modeNDPA})]. The forward gain is highest at resonance of the selected mode and decreases with increase in detuning as
\begin{eqnarray}
\mathcal{G} [\omega] \underset{\mathcal{C}\rightarrow1^{-}}{=} \mathcal{G}_{0} \frac{f(\vartheta_{a})}{1 + \mathcal{G}_{0}\vartheta_{b}^{2}},
\label{Eq.:Gainvsdetuning}
\end{eqnarray}
where $\vartheta_{a} = \omega /\kappa_{a}, \vartheta_{b} = (\omega + \omega_{a})/\kappa_{b}$ for the USB amplifier and $\vartheta_{a,b} = \omega/\kappa_{a,b}$ for the RSB amplifier denote the respective reduced detunings. Here we have retained only leading order terms in $\vartheta_{b}$. Also $f(\vartheta_{a})$ denotes a polynomial function of $\vartheta_{a}$ which does not depend on the gain $\mathcal{G}_{0}$; hence it does not affect gain-bandwidth product and can be taken as unity for $\vartheta_{a} \ll 1$. Eq.~(\ref{Eq.:Gainvsdetuning}) allows us to write the instantaneous amplification bandwidth as
\begin{eqnarray}
\Delta (\mathcal{G}_{0}) \approx \frac{2\kappa_{b}}{\mathcal{G}_{0}^{1/2}}.
\end{eqnarray}
This trade-off between maximum useful gain of an amplifier and the instantaneous bandwidth over which such a gain is realizable is universal in most parametric amplifier systems \cite{ArchanaJPC, Eichler2014}. The directional phase-sensitive amplifier, on the other hand, shows no gain-bandwidth tradeoff as shown in Fig.~\ref{Fig.:DPSAchar}(d) [cf. Eq.~(\ref{Eq.:FwdGainPSA})].
\par
The directionality of biharmonic amplifiers is also frequency-dependent; in order to quantify this, we introduce a directionality parameter [cf. Eq.~(\ref{Eq.:RevGainSidebandNDPA})]
\begin{eqnarray}
d[\omega] \equiv 1 - \frac{\bar{\mathcal{G}}[\omega]}{\mathcal{G}[\omega]} = \frac{ 1 }{1 + \vartheta_{a}^{2}},
\label{Eq.:dparameter}
\end{eqnarray}
with $d =0$ corresponding to usual reciprocal or symmetric amplification and $d= 1$ corresponding to perfect nonreciprocity. This allows us to define the directionality bandwidth as
\begin{eqnarray}
\Delta_{d} = 2\kappa_{a}.
\label{Eq.:dbandwidth}
\end{eqnarray}
Using arguments similar to those for the gain-bandwidth tradeoff, it is straightforward to see that $\Delta_{d}/2$ denotes the detuning from resonance at which the directionality parameter reduces to $d=0.5$ --- a value that corresponds to a 3 dB isolation between forward gain $\mathcal{G}[\omega]$ and reverse gain $\bar{\mathcal{G}}[\omega]$. Crucially, \emph{directionality bandwidth $\Delta_{d}$ is independent of amplification bandwidth $\Delta$}, a behavior generic to these three-mode directional amplifiers. This is further borne out by following observations:
\begin{enumerate}[(i)]
\setlength{\itemsep}{0pt}
\setlength{\parskip}{0pt}
\item Eq.~(\ref{Eq.:dparameter}), and by consequence Eq.~(\ref{Eq.:dbandwidth}), hold true for both phase-preserving and phase-sensitive operations which show qualitatively different gain-bandwidth behavior
\item Unlike $\Delta$, $\Delta_{d}$ is not limited by the gain $\mathcal{G}_{0}$ or by the linewidth $\kappa_{b}$ of the amplified or deamplified mode, and strictly scales with the linewidth of the auxiliary mode alone.
\end{enumerate}
\par
Therefore, in order to have directionality over a large bandwidth, it is essential for the auxiliary mode to have a proportionately large linewidth. However, as indicated by calculations including counter-rotating terms, too large $\kappa_{a}$ ($Q_{a} \rightarrow 0$) can cost net achievable directionality in these systems [Figs.~\ref{Fig.:ComparisonRaman} and \ref{Fig.:DPSAchar}(a)]. This is clear from Fig.~\ref{Fig.:Deltad} where we show an example calculation of the effect of counter-rotating terms in biharmonic Raman amplifiers, and the tradeoff they impose between the net achievable directionality $d$ and directionality bandwidth $\Delta_{d}$.
In Fig.~\ref{Fig.:Deltad} the directionality parameter is evaluated on resonance ($\omega=0$). For the case of phase-preserving amplification we find the scaling
\begin{align}
d_{\rm CR}[0] =& \ \frac{ 1 }{ 1 + \frac{1}{ 64 Q_a^2} } ,
\end{align}
hence, for large $Q_a$ we have $d \simeq 1$.
Note, that a directionality parameter of unity can only be achieved at resonance ($\omega=0$), the maximum attainable value of $d$ decreases quadratically with detuning, as per Eq.~(\ref{Eq.:dparameter}).
In addition, effects of non-RWA corrections are slightly more pronounced at finite detunings.
\par
Thus, while it is desirable to have the auxiliary mode in the steady state for stable device operation, it cannot be a very low-Q waveguide or a resistive load. A useful framework to distinguish the effects of auxiliary mode dynamics is to view this mode as an engineered reservoir, as elaborated in the following section.
\begin{figure}[t!]
\centering
\includegraphics[width=1.0\columnwidth]{Figure6.pdf}
\caption{Variation of directionality parameter at resonance $d[0]$ (solid cyan) and directionality bandwidth $\Delta_{d}$ (dashed red), with the quality factor $Q_{a}$ of the auxiliary mode for (a) biharmonic Raman amplifiers, (b) directional phase-sensitive amplifier. Both plots were calculated for a forward gain of $\mathcal{G} = 20$ dB.}
\label{Fig.:Deltad}
\end{figure}
\subsection{Connection to dissipation engineering}
\label{sec_dissengg}
It was recently shown that any factorisable coherent interaction can be rendered directional by balancing it with the corresponding dissipative interaction \cite{Metelmann2015}.
Dissipation is, therefore, the crucial element to obtain any directionality at all. This holds true as well for all biharmonic amplifier schemes proposed in this work.
In this section we briefly sketch how this manifest itself, using the particular example of biharmonic Raman amplifiers.
\par
We start out from the Hamiltonian in Eq.~(\ref{Eq.:Ham3mode}), describing a hopping (paramp) interaction between the upper (lower) sideband mode and the low frequency mode,
as well as mixing between the two sidebands mediated by the second harmonic. The auxiliary mode $\hat{a}$ can be considered as the engineered reservoir that provides us with the desired dissipative interaction. Elimination of this mode leads to the following coupled equations for the remaining modes 1 and 2 as
\begin{align}\label{Eq.:CouplingBRA}
\hat d_{1}[\omega] \sim& - i \left[ \widetilde G_2[\omega] - i \frac{\Gamma[\omega]}{2} \right] \hat d_2^{\dag}[\omega] ,
\nonumber \\
\hat d_{2}^{\dag}[\omega] \sim& + i \left[ \widetilde G_2^{\ast}[\omega] - i \frac{\Gamma[\omega]}{2} \right] \hat d_{1}[\omega].
\end{align}
Here we consider the full frequency dependence of the auxiliary mode, and accordingly define the couplings
\begin{align}
\widetilde G_2[\omega] &= G_2 e^{- i \alpha} + \frac{ \frac{ \omega}{\kappa_a} \Gamma_{0} }{1 + \frac{4\omega^2}{\kappa_{a}^2} },
\hspace{0.5cm}
\Gamma [\omega] = \frac{\Gamma_{0}}{1 + \frac{4\omega^2}{\kappa_{a}^2} },
\end{align}
with $\Gamma_{0} = 4G_1^2/\kappa_a $. The first terms in Eqs.~(\ref{Eq.:CouplingBRA}) corresponds to a coherent interaction, they could be derived from an effective Hamiltonian of the form $\mathcal H_{\rm coh} = \widetilde G_2[\omega] \hat d_1^{\dag} \hat d_2^{\dag} + h.c.$. The same does not hold true for the second coupling terms, they can never be derived from a coherent Hamiltonian. These terms correspond to a dissipative interaction mediated by the auxiliary mode, i.e., they could be derived from an effective non-local dissipator
$\Gamma [\omega] \mathcal L [\hat d_1 + \hat d_2^{\dag} ] $ in a master equation \footnote{We use the definition $\mathcal L [\hat o] \hat \rho = \hat \hat o \hat \rho \hat o^{\dag} - 1/2 \hat o^{\dag}\hat o \hat \rho - 1/2 \hat \rho \hat o^{\dag} \hat o $.}.
\par
The general condition of balancing a coherent interaction with its dissipative counterpart reported in Ref. \cite{Metelmann2015} translates into simply tuning the amplitude and the phase of the coherent coupling,
\begin{align}
\big| \widetilde G_2[\omega] \big| = \frac{ \Gamma [\omega]}{2},
\hspace{0.5cm}
\arg( \widetilde G_2[\omega]) = \pm \frac{\pi}{2}.
\end{align}
Applying these conditions to Eq.~(\ref{Eq.:CouplingBRA}) renders the coupling between the two modes directional. This selective decoupling would not be possible without the dissipative interaction. In principle, the system could be rendered directional for every frequency with the above conditions. However, in an experiment the amplitude $G_{2}$ and the phase $\alpha$ will be fixed and the system is rendered completely directional at a single frequency, e.g., in the present frame this would be at resonance or $\omega = 0$. The frequency range around that frequency over which the reverse gain is suppressed is then determined by $\kappa_a$, i.e., the inverse memory time of the engineered reservoir as explained in the previous section. If we assume that this memory time is vanishingly small, i.e., that the auxiliary mode $\hat{a}$ is strongly damped, we can treat the mode as a Markovian reservoir and the whole system can be modeled via a Lindblad master equation of the form
\begin{align}\label{Eq.:LdissNDPA}
\frac{d}{dt} \hat \rho = -i \left[\mathcal H_{\rm coh} , \hat \rho\right] +
\Gamma_0 \mathcal L \left[\hat d_1 + \hat d_2^{\dag} \right] \hat \rho.
\end{align}
Here the non-local dissipator describes a Markovian reservoir which absorbs excitation from mode $\hat d_1$ and emits it into $\hat d_2^{\dag}$, with a rate $\Gamma_0$. In the overdamped case the master equation sufficiently describes the system. However, for arbitrary damping $\kappa_a$ one has to only include the non-Markovian effects due to a finite life-time of the reservoir, i.e., the low-frequency mode.
\section{Conclusions}
\label{sec_conclusions}
We have studied different modalities of directional quantum-limited amplification realizable in a three-mode system pumped with a biharmonic pump. For optimal amplitude and phase difference between the two harmonics and appropriate choice of mode frequencies, such a system provides the minimal implementation of nonreciprocal photon transmission and amplification.
We explicitly present the schemes for a directional phase-preserving (both degenerate and non-degenerate) and phase-sensitive amplification.
The generality and minimality of our proposals should make it suitable for implementation in multiple platforms, such as superconducting qubits and opto- or electro-mechanical systems. Using pump harmonics can be particularly desirable in optical systems, where supplemented by second/sub harmonic generation, it can drastically reduce the resource overhead in nonreciprocal optical platforms.
\par
We also evaluate full frequency-dependent forward and reverse gains, and the available bandwidth with each scheme. Our results show that there is a universal separation of parameters determining directionality and amplification bandwidths. In particular, the directionality bandwidth increases directly with the linewidth of the dissipative/auxiliary mode alone. On the other hand, the net magnitude of directionality is predicated on a modestly high quality factor for the engineered reservoir mode, required to suppress the deleterious counter-rotating contributions.
\section{Acknowledgements}
\label{sec_thanks}
The authors wish to thank Aashish Clerk, Michel Devoret, Leonardo Ranzani and John Teufel for useful discussions.
|
\section{Introduction}
The detection of metal pollution in white dwarf atmospheres provides strong evidence that 25-50\,per\,cent of white dwarfs host remnants of planetary systems \citep{zuckermanetal03-1, koesteretal14-1, koester+kepler15-1}. The survival of planets through the post main-sequence evolution of their host star is further supported by the detection of more than 35 dusty debris discs at white dwarfs with metal pollution \citep{kilicetal06-1, farihietal08-1, farihietal09-1, juraetal09-1, debesetal11-2, hoardetal13-1, bergforsetal14-1, rocchettoetal15-1}, and is also predicted by theoretical studies \citep{villaver+livio07-1, veras+gaensicke15-1}.
Debris discs around white dwarfs are thought to be produced by the tidal disruption of asteroids (or comets) scattered onto a highly eccentric orbit by planets in the system \citep{grahametal90-1, jura03-1}. While recent simulations have begun to explore the formation and evolution of these discs; from the initial disruption of the planetesimal and the shrinking of its orbit \citep{debesetal12-1,verasetal14-1, verasetal15-1}, to the dynamics and interactions of the debris within the disc \citep{rafikov11-2, metzgeretal12-1}, many aspects remain poorly understood.
Over the last decade, gaseous debris discs have been detected around eight white dwarfs, all of which also host circumstellar dust and have metal polluted photospheres \citep{gaensickeetal06-3, gaensickeetal07-1, gaensickeetal08-1, gaensicke11-1, farihietal12-1, melisetal12-1, wilsonetal14-1, guoetal15-1}. These systems were identified by the detection of Ca\,{\textsc{ii}} 8498.02\,\AA, 8542.09\,\AA, 8662.14\,\AA\ emission lines (henceforth the Ca\,{\textsc{ii}} triplet, air wavelengths are given).
\begin{figure*}
\centerline{\includegraphics[width=2\columnwidth]{fig1.eps}}
\caption{\label{f-1043_plot} X-Shooter spectra of SDSS\,J1043+0855 (grey, obtained in January and May 2011) together with a model atmosphere fit (blue) using the atmospheric parameters listed in Table\,\ref{t-wds}. The strongest emission and absorption lines have been labelled (note that the Fe\,{\textsc{ii}} feature is in emission).}
\end{figure*}
\begin{table*}
\centering
\caption{Log of observations of SDSS\,J1043+0855. $^{a}$ Different exposure times for the individual X-Shooter arms (UVB / VIS). We did not use data collected by the NIR arm of X-Shooter as the signal-to-noise ratio was too poor. \label{t-dates}}
\begin{tabular}{rrrrr}
\hline
Date & Telescope/Instrument & Wavelength Range [\AA] & Resolution [\AA] & Exposure Time [s]\,$^{a}$ \\
\hline
2003 April 05 & SDSS & 3800 -- 9200 & 2.9 & 2900\\
2007 February 03 & WHT/ISIS & 7400 -- 9200 & 2.0 & 4800\\
2009 February 16 & WHT/ISIS & 6000 -- 8900 & 3.7 & 3950\\
2010 April 22 & WHT/ISIS & 8100 -- 8850 & 1.1 & 7200\\
2011 January 29 & VLT/X-Shooter & 2990 -- 10400 & 1.12 & 2950 / 2840\\
2011 May 30 & VLT/X-Shooter & 2990 -- 10400 & 1.15 & 2950 / 2840\\
2012 January 03 & SDSS & 3602 -- 10353 & 3.2 & 2702\\
\hline
\end{tabular}
\end{table*}
So far only four of these eight gas disc systems have multi-epoch spectroscopy over time scales of years, three of which show variability in the Ca\,{\textsc{ii}} emission of the gaseous disc; either as significant changes in the morphology of the Ca\,{\textsc{ii}} triplet \citep{wilsonetal15-1, manseretal16-1}, or as a decrease in strength of the lines \citep{wilsonetal14-1}. It is thought that the gaseous components of the debris discs in these systems are tracers of dynamic activity \citep{gaensickeetal08-1, wilsonetal14-1, wilsonetal15-1, manseretal16-1}.
The gaseous disc around SDSS\,J104341.53+085558.2 (henceforth SDSS\,J1043+0855) was discovered by \cite{gaensickeetal07-1} via the detection of Ca\,{\textsc{ii}} triplet emission. An infrared excess was detected by \cite{melisetal10-1} and \cite{brinkworthetal12-1}, confirming the presence of a dusty disc in SDSS\,J1043+0855. Here, we report nine years of spectroscopic observations of SDSS\,J1043+0855 that reveal a change in the morphology of the Ca\,{\textsc{ii}} triplet, similar to those seen in other gaseous disc systems (SDSS\,J122859.93+104032.9 and SDSS\,J084539.17+225728.0, henceforth SDSS\,J1228+1040 and SDSS\,J0845+2257, \citealt{wilsonetal15-1, manseretal16-1}). We also present the accretion rates of the debris onto the white dwarf and the metal abundances of the debris.
\section{Observations}
\label{sec:observations}
We obtained optical spectroscopy of SDSS\,J1043+0855 from 2003 to 2012 with several instruments: X-Shooter \citep{vernetetal11-1} on the ESO Very Large Telescope (VLT); the 2.5\,m Sloan Digital Sky Survey telescope (SDSS, data retrieved from DR7 and DR9, \citealt{gunnetal06-1, abazajianetal09-1, eisensteinetal11-1, ahnetal14-1, smeeetal13-1}); and the Intermediate dispersion Spectrograph and Imaging System (ISIS) on the William Herschel Telescope (WHT). A log of the observations is summarised in Table\,\ref{t-dates}. The X-Shooter data were reduced within the \texttt{\textsc{reflex}}\,\footnote{Documentation and software for \texttt{\textsc{reflex}} can be obtained from http://www.eso.org/sci/software/reflex/} reduction work flow using the standard settings and optimising the slit integration limits \citep{freudlingetal13-1}. The sky spectrum of each observation was used to determine the spectral resolution. The first ISIS spectrum of SDSS\,J1043+0855 was reported in \cite{ gaensickeetal07-1}, the additional ISIS spectra were obtained with a similar setup, and were reduced in the same fashion (see \citealt{farihietal12-1} and \citealt{wilsonetal14-1} for additional details). We removed the telluric lines present in the VIS arm of the X-Shooter spectra using the X-Shooter Spectral Library (XSL) provided by \cite{chenetal14-1}, and applying the method outlined in \cite{manseretal16-1}.
\begin{figure}
\centerline{\includegraphics[width=0.95\columnwidth]{fig2.eps}}
\caption{\label{f-1043_Ca_vel} Time series spectroscopy of the Ca\,{\textsc{ii}} triplet of SDSS\,J1043+0855 spanning nine years. The spectra are normalised and offset in steps of one from the 2012 January observation.}
\end{figure}
\section{Evolution of the calcium emission profile}
\subsection{Double peaked emission from a disc}
The characteristic emission profile from a gaseous disc with a radially symmetric intensity distribution and circular orbit is a symmetric, double-peaked emission profile (see Figure 1 of \citealt{horne+marsh86-1}). The width of the profile arises from the wide range of velocities projected along the line of sight and can reveal details about the structure and geometry of the gaseous disc. The largest velocities correspond to material at the inner disc radius, which can hence be determined from the full width at zero intensity, i.e. the point at which the emission drops to the continuum level, and with knowledge of the inclination, $i$, using $R\,\sin^{2}\,i = R_{\textrm{obs}}$, where $R$ and $R_{\textrm{obs}}$ denote the actual and observed radii respectively. We define the maximum red/blue-shifted velocities to represent the red/blue inner edges of the disc. The outer radius can be estimated from the peak separation in the double peaked profile.
Any departure from a radially symmetric, circular disc will manifest itself as asymmetries in the double peaked emission profile. An eccentric disc of uniform intensity, for example, would generate an asymmetric double-peaked profile if viewed along the semi-minor axis, as there is an asymmetry in the velocity distribution of the material, as well as in the total red and blue-shifted light emitted from the disc. We use this insight below to discuss the changes in the line profiles observed at SDSS\,J1043+0855.
\subsection{Variation of the Calcium emission lines}
The Ca\,{\textsc{ii}} triplet in SDSS\,J1043+0855 changes in morphology over a time scale of nine years (Figure\,\ref{f-1043_Ca_vel}). All three components of the triplet vary in the same manner and are henceforth referred to in singular.
The initial three spectra are noisy and of low resolution, showing a broad line profile. In contrast, finer features appear in the higher resolution 2010 WHT spectrum, namely a sharp red-shifted peak and a gradual drop off to blueward wavelengths, revealing a clear asymmetry in the red and blue inner edges of the disc. This asymmetry decreases in the 2011 January spectrum, and vanishes in the 2011 May spectrum. The second SDSS spectrum obtained in 2012 shows no sign of any sharp departures from symmetry, although it is of relatively low spectral resolution. In Section\,\ref{sec:discussion} we discuss the similarities, and the possible origin, of the observed variations with those seen in other systems.
\subsection{Inclination of the disc}
In the 2011 January and May spectra, the `valley' in between the two peaks of the emission profiles almost reaches the continuum level, which suggests that the debris disc is optically thick and seen at a large inclination. \cite{horne+marsh86-1} have shown that in an optically thick accretion disc, line emission is more likely to escape along paths of largest velocity gradient provided by Keplerian shear flow, which is at a minimum for purely tangential or radial emission through the disc. For an observer looking through the disc at an inclination of 90$^{\circ}$ (edge on), material travelling perpendicular to the line of sight (emission in the `valley') will be emitting along the radial direction of the disc and will therefore be suppressed.
While the formulation described in \cite{horne+marsh86-1} was developed for circular orbits, it was expanded to include eccentric orbits in order to model the emission profile of SDSS\,J1228+1040 obtained in 2006 \citep{gaensickeetal06-3}, and we give additional details on this extension in Appendix\,\ref{appendix-1}. We fit the emission profiles of the 2011 January and May spectra of SDSS\,J1043+0855 in the same manner (see Figure\,\ref{f-fit}), and obtain an inclination of $i\simeq$\,76$^{\circ}$ and 72$^{\circ}$ respectively, leading to an average inclination of $i=$\,74$^{\circ}$. The statistical uncertainty obtained for the two inclinations was $\pm$\,0.5$^{\circ}$, which is probably an underestimate, and a more realistic uncertainty is taken to be $\simeq$\,$\pm$\,5$^{\circ}$.
\begin{figure}
\centerline{\includegraphics[width=1\columnwidth]{plotmodel.eps}}
\caption{\label{f-fit} Model fit to the normalised 8542.09\,\AA\ component of the Ca\,{\textsc{ii}} triplet for the 2011 January and May X-shooter spectra, from which we obtain an inclination of $i \simeq 74^{\circ}$ for the disc at SDSS\,J1043+0855. The model applies the formulation described in \protect\cite{horne+marsh86-1} which has been expanded to include eccentric orbits (see Appendix\,\ref{appendix-1} for more details).}
\end{figure}
\section{Metal abundances in the photosphere of SDSS\,J1043+0855}
\label{sec:abund}
We fitted the two SDSS spectra with DA (hydrogen dominated) white dwarf model atmospheres using the methods described in \cite{gaensickeetal12-1} and \cite{koesteretal14-1}, and find $T_{\textrm{eff}}$ = 17879\,$\pm$\,195\,K, and $\log$\,$g$ = 8.124\,$\pm$\,0.033, corresponding to $M_{\textsc{wd}}$ = 0.693\,$\pm$\,0.020\,\mbox{$\mathrm{M}_{\odot}$}\ and $R_{\textsc{wd}}$ = 0.0120\,$\pm$\,0.0003\,\mbox{$\mathrm{R}_{\odot}$}\ (see Table\,\ref{t-wds}). Using the initial-to-final mass relation of \cite{casewelletal09-1}, \cite{kaliraietal08-1}, \cite{catalanetal08-1}, and \cite{williamsetal09-1}, we estimate the mass of the white dwarf progenitor to be 2.76\,$\pm$\,0.04\,\mbox{$\mathrm{M}_{\odot}$}.
We note that while the atmospheric parameters derived here are consistent with previous measurements \citep{eisensteinetal06-1, gaensickeetal07-1, tremblayetal11-2, kleinmanetal13-1} the $ugriz$ photometry suggests a lower effective temperature, even when allowing for the maximum reddening along the line of sight obtained from \cite{schlafly+finkbeiner11-1}. This discrepancy is likely due to a higher amount of extinction, either by a denser patch in the interstellar medium that remains unresolved in the dust maps, or by circumstellar dust. We speculate that there could be additional dust in the system, which may not be associated with the dusty component of the debris disc at SDSS\,J1043+0855 \citep{melisetal10-1, brinkworthetal12-1}, and could possibly be the cause of the observed reddening.
\begin{table*}
\centering
\caption{Metal polluted white dwarfs with circumstellar gas detected in emission (e) or absorption (a), and evidence for photometric or spectroscopic variability (v). System parameters and accretion rates are given with errors where known. Values derived or updated in this paper are set in italics.
$^{1}$ \protect\cite{dufouretal12-1},
$^{2}$ \protect\cite{wilsonetal15-1},
$^{3}$ \protect\cite{farihietal12-1},
$^{4}$ \protect\cite{xu+jura14-1},
$^{5}$ \protect\cite{gaensickeetal07-1},
$^{6}$ \protect\cite{melisetal10-1},
$^{7}$ This paper,
$^{8}$ \protect\cite{guoetal15-1},
$^{9}$ \protect\cite{gaensickeetal06-3},
$^{10}$ \protect\cite{gaensickeetal12-1},
$^{11}$ \protect\cite{koesteretal14-1},
$^{12}$ \protect\cite{manseretal16-1},
$^{13}$ \protect\cite{koesteretal05-2},
$^{14}$ \protect\cite{vossetal07-1},
$^{15}$ \protect\cite{melisetal12-1},
$^{16}$ \protect\cite{wilsonetal14-1},
$^{17}$ \protect\cite{vennes+kawka13-1},
$^{18}$ \protect\cite{koester+wilken06-1},
$^{19}$ \protect\cite{debesetal12-1},
$^{20}$ \protect\cite{vanderburgetal15-1},
$^{21}$ \protect\cite{xuetal16-1}. \label{t-wds}}
\begin{tabular}{lrllllllr}
\hline
Name & Type & $\log\,g$ & $T_{\textrm{eff}}$ & $M_{\mathrm{WD}}$ & $\tau_{\textrm{cool}}$ & $\dot{M}$ & Features & ref \\
& & (g\,cm$^{-2}$) & (K) & (\mbox{$\mathrm{M}_{\odot}$}) & (Myr) & ($\times\,10^{8}$\,g\,s$^{-1}$) & & \\
\hline
SDSS\,J0738+1835 & DB & 8.4 (0.2) & 13950 (100) & 0.841 (0.131) & 477 (160) & $1300$ & e & 1 \\
SDSS\,J0845+2257 & DB & 8.18 (0.20) & 19780 (250) & \textit{0.73 (0.11)} & \textit{122 (44)} & $160$ & e, v & 2 \\
SDSS\,J0959--0200 & DA & 8.06 (0.03) & 13280 (20) & 0.64 (0.02) & \textit{324 (17)} & $0.32$ & e, v & 3, 4 \\
SDSS\,J1043+0855 & DA & \textit{8.124 (0.033)} & \textit{17879 (195)} & \textit{0.693 (0.020)} & \textit{153 (10)} & \textit{(2.5 - 12)} & e, v & 5, 6, 7 \\
WD\,1144+0529 & DA & 7.74 (0.03) & 23027 (219) & 0.49 (0.03) & 21.2 (1.9) & - & e & 8 \\
SDSS\,J1228+1040 & DA & 8.150 (0.089) & 20713 (281) & 0.705 (0.051) & 100 (5) & $5.6$ & e, a, v & 6, 9, 10, 11, 12 \\
HE\,1349--2305 & DA & 8.133 & 18173 & 0.673 & 149.4 & $1.3$ & e & 13, 14, 15\\
SDSS\,J1617+1620 & DA & 8.11 (0.08) & 13520 (200) & 0.68 (0.05) & 350 (50) & $(6.4$ - $7.8)$ & e, v & 16 \\
\hline
PG\,0843+516 & DA & 7.902 (0.089) & 22412 (304) & 0.577 (0.047) & \textit{42 (4)} & $10.2$ & a & 11 \\
WD\,1054--226 & DA & 8.04 (0.03) & 7903 (16) & - & \textit{1255 (92)} & - & a & 17 \\
WD\,1124--293 & DA & 8.1 & 9700 & 0.66 & \textit{843} & $1.3$ & a & 18, 19 \\
WD\,1145+017 & DB & - & 15900 (500) & - & 175 (75) & $430$ & a, v & 20, 21\\
\hline
\end{tabular}
\end{table*}
\begin{table*}
\centering
\caption{Diffusion timescales, $\tau_{\mathrm{diff}}$, and average accretion fluxes, $\dot{M}$, for the metals detected in the photosphere of SDSS\,J1043+0855. Photospheric abundances by number are given with respect to hydrogen. $^{a}$ We infer a total accretion rate based on the mass fluxes assuming a bulk Earth composition and the respective mass fractions for each element (from \citealt{allegreetal01-1}). \label{t-abund}}
\begin{tabular}{lrrrrr}
\hline
Element & $\log\,[\mathrm{Z}/\mathrm{H}]$ & $\tau_{\mathrm{diff}}$ [h] & $\dot{M}$ [g\,s$^{-1}$] & Bulk Earth mass fraction [per\,cent] $^{a}$ & Inferred total $\dot{M}$ [g\,s$^{-1}$] \\
\hline
8 O & $< -4.00$ & 65.0 & $< 2.0 \times 10^{9}$ & 32.4 & $< 6.2\times10^9$ \\
12 Mg & $-5.15$ (0.15) & 26.5 & $4.0 \times 10^{7}$ & 15.8 & $2.5\times10^8$\\
14 Si & $-4.80$ (0.15) & 13.8 & $2.0 \times 10^{8}$ & 17.1 & $1.2\times10^{9}$\\
20 Ca & $-6.00$ (0.20) & 18.6 & $1.3 \times 10^{7}$ & 1.6 & $8.0\times10^{8}$\\
26 Fe & $< -4.70$ & 11.1 & $< 6.3 \times 10^{8}$ & 28.8 & $< 2.2\times10^{9}$ \\
\hline
\end{tabular}
\end{table*}
Using the system parameters, the metal absorption lines detected in the spectra of SDSS\,J1043+0855 were modelled (see Figure\,\ref{f-absorp_lines}) to measure abundances relative to hydrogen which are given in Table\,\ref{t-abund}, along with the diffusion time scales and accretion fluxes. The abundance by number of Mg, Ca, and an upper limit for Fe relative to Si were found to be $\log$\,(Mg/Si) = $-0.64\,\pm\,0.21$, $\log$\,(Ca/Si) = $-1.33\,\pm\,0.24$, and $\log$\,(Fe/Si) $\leqslant 0.19$. From Figure 7 of \cite{jura+young14-1}, the ratios of Mg and Fe to Si are broadly consistent with those found for the crust of the Earth, and imply the accreted object is processed rather than having a 'chondritic' composition \citep{zuckermanetal07-1, xuetal13-1, xuetal14-1}. This is also supported by the relatively low number abundance of Mg with respect to Ca, $\log$\,(Mg/Ca) = $0.70\,\pm\,0.25$, when compared to a sample of 60 externally polluted white dwarfs (see Figure 1 of \citealt{jura+xu13-1}). Similarly low $\log$\,(Mg/Ca) ratios have also observed at other white dwarfs that are thought to accrete crust material, such as NLTT 43806 and NLTT\,19868 \citep{zuckermanetal11-1,kawka+vennes16-1}.
We also note that the Ca\,K 3934\,\AA\ feature shows clear emission in the averaged, normalised X-Shooter spectra (Figure\,\ref{f-absorp_lines}). Similar emission is seen at SDSS\,J1228+1040, which also revealed a difference in morphology between the Ca\,H\,\&\,K profiles and the Ca\,{\textsc{ii}} triplet, although the signal-to-noise of the current SDSS\,1043+0855 spectra is not high enough to determine this \citep{manseretal16-1}.
We estimate the total accretion rate for each element from the mass flow rates given in Table\,\ref{t-abund}, assuming a bulk Earth composition and the respective mass fractions of each element (from \,\citealt{allegreetal01-1}). The resulting range from Mg, Si, and Ca, $\dot{M}_{\mathrm{Total}} = 2.5 \times 10^8 - 1.2 \times 10^9$\,gs$^{-1}$, reflects the uncertainty in the bulk abundances of the planetary debris, but agrees well with the distribution of accretion rates inferred for 19 other metal-polluted DA white dwarfs with dusty discs \citep{bergforsetal14-1}. For reference, the total accretion rate of SDSS\,J1228+1040 (where accretion fluxes for all major elements were measured from \textit{HST}/COS ultraviolet spectra) is $5.6\times10^{8}\,\textrm{g\,s}^{-1}$ \citep{gaensickeetal12-1}.
\begin{figure}
\centerline{\includegraphics[width=1\columnwidth]{fig3.eps}}
\caption{\label{f-absorp_lines} Model fits (blue) to absorption lines present in the combined X-Shooter spectra of SDSS\,J1043+0855 (black). The Ca\,{\textsc{k}} feature also contains a clearly detected emission profile from the gaseous disc.}
\end{figure}
\begin{table}
\centering
\caption{Equivalent width measurements of the Ca\,{\textsc{ii}} triplet in SDSS\,J1043+0855. The errors given here are purely statistical. \label{t-ewtrip}}
\begin{tabular}{lr}
\hline
Date & Equivalent width [\AA] \\
\hline
2003--04 & -27 (3)\\
2007--02 & -22 (1)\\
2009--02 & -15 (1)\\
2010--04 & -13 (1)\\
2011--01 & -18 (1)\\
2011--05 & -20 (1)\\
2012--01 & -19 (1)\\
\hline
\end{tabular}
\end{table}
\section{Discussion}
\label{sec:discussion}
Up until the last decade, debris discs around white dwarfs have appeared static in nature, with no significant detections of variability in the properties of the disc itself or in the strength of the absorption lines in metal polluted systems. Due to the short diffusion time scales at the majority of these systems \citep{koesteretal14-1}, any change in the strength of the absorption lines would imply a change in the accretion rate onto the white dwarf. \cite{vonhippel+thompson07-1} claimed changes in the equivalent width of the photospheric Ca\,{\textsc{ii}}\,K line over a time scale of days at GD\,29--38, however additional observations obtained by \cite{debes+lopez-morales08-1} did not confirm such variation, and they concluded further data were required to determine the possible variable nature of the accretion onto the white dwarf.
In recent years, the spectroscopic monitoring of gaseous discs has revealed variability that gives us insight into their formation and dynamics. Table\,\ref{t-wds} lists the stellar parameters of the eight published gas disc systems. Other than SDSS\,J1043+0855, there are four gaseous disc systems with multi-epoch spectroscopy: SDSS\,J161717.04+162022.4 (\citealt{wilsonetal14-1}, henceforth SDSS\,J1617+1620), SDSS\,J0845+2257 \citep{gaensickeetal08-1, wilsonetal15-1}, SDSS\,J1228+1040 \citep{manseretal16-1}, and SDSS\,J073842.56+183509.6 (\citealt{gaensicke11-1, dufouretal12-1}, henceforth SDSS\,J0738+1835).
SDSS\,J0845+2257, SDSS\,J1228+1040, and SDSS\,J1617+1620 all show variations on a time scale of years, but follow two distinct types of evolution. In SDSS\,J1617+1620 the Ca\,{\textsc{ii}} triplet emission gradually decreased in strength over a time scale of eight years while not undergoing noticeable changes in the line profile shape \citep{wilsonetal14-1}. In contrast, the changes seen in SDSS\,J0845+2257 and SDSS\,J1228+1040 are of a morphological nature, analogous to the changes we present here for SDSS\,J1043+0855. Table\,\ref{t-ewtrip} lists the equivalent widths (subject to systematic uncertainties related to the method used in continuum fitting, as well as the statistical uncertainties given in Table\,\ref{t-ewtrip}) of the Ca\,{\textsc{ii}} emission lines in SDSS\,J1043+0855, which do not show any long term decay of the equivalent width of the Ca\,{\textsc{ii}} triplet such as seen at SDSS\,J1617+1620. Only SDSS\,J0738+1835 has displayed no changes in the shape and strength of the Ca\,{\textsc{ii}} triplet over a period of six years, although only three epochs are available, with two of them spaced only a year apart.
\cite{manseretal16-1} showed that the variable Ca\,{\textsc{ii}} triplet line profiles of SDSS\,J1228+1040 could be interpreted as the emission from a fixed intensity pattern that precesses over a time scale of decades, possibly indicating a young debris disc that still has eccentric orbits and has not fully circularised. General relativistic precession will cause the debris to precess with a radially dependent period, causing orbits to cross one another and inducing collisions which produces the observed gaseous component to the debris disc.
While the evolution of the emission from SDSS\,J1043+0855 appears to be remarkably similar to SDSS\,J1228+1040 and SDSS\,J0845+2257, the data have a lower signal to noise and have fewer epochs, and thus, while it is likely that the same physical mechanism is responsible for the evolution of the line profiles observed in all three systems, regular spectroscopic monitoring of all gas discs is necessary to develop a more detailed understanding of the dynamical processes present in planetary debris discs around white dwarfs.
Variability of debris discs is not only limited to the Ca\,{\textsc{ii}} triplet line profile. The dusty disc around SDSS\,J0959--0200 was observed to significantly decrease in infrared flux by \cite{xu+jura14-1}, who propose two mechanisms by which the disc could be depleted; a recent planetesimal impact on the disc, or instability near the inner edge. We suggest an additional scenario of a vertically extended cloud of dust, generated from an asteroid colliding with a pre-existing disc \citep{jura08-1}. Such an optically thin cloud would temporarily add to the infrared emission of the optically thick disc, but the overall infrared emission from the system would decrease as the dust cloud settled into the disc.
In Table\,\ref{t-wds} we also include four additional systems where circumstellar absorption of gaseous material has been detected around the host white dwarf, including WD\,1145+017, which is orbited by highly-dynamic debris, transiting the white dwarf with periods of $\simeq$\,4.5\,hr \citep{vanderburgetal15-1, xuetal16-1, gaensickeetal16-1, rappaportetal16-1}. WD\,1145+017 is unequivocally a highly dynamical and evolving system with a planetesimal currently undergoing disruption, and also hosts circumstellar gas absorption \citep{xuetal16-1}. Curiously, the detection of absorption due to circumstellar gas does not correlate with the presence of Ca\,{\textsc{ii}} triplet emission: SDSS\,1228+1040 is so far the only system in which both have been detected \citep{gaensickeetal12-1}. We note that circumstellar gas has been detected also around a number of hot and young white dwarfs, their origin is probably diverse in nature and not unambiguously associated with evolved planetary systems \citep{dickinsonetal12-2, barstowetal14-1}.
\section{Conclusions}
\label{sec:conc}
We report here the morphological variability of the Ca\,{\textsc{ii}} triplet in SDSS\,J1043+0855 on a time scale of nine years. The evolution of the Ca\,{\textsc{ii}} triplet reported here is similar to that of two other systems, SDSS\,J1228+1040 and SDSS\,J0845+2257.
We have also analysed the optical spectra of SDSS\,J1043+0855 to determine its stellar parameters and the photospheric metal abundance. The Mg/Si and (upper limit to the) Fe/Si ratios of the planetary debris that has been accreted onto the white dwarf are broadly consistent with those of the crust of the Earth.
The recent detection of the 'real time' disruption of a planetesimal at WD\,1145+017, along with the dynamical evolution seen at the gaseous discs SDSS\,J1043+0855, SDSS\,J1228+1040, SDSS\,J1617+1620, and SDSS\,J0845+2257 reveals that variability at planetary systems around white dwarfs is more common than was initially thought. Additional spectroscopic and photometric monitoring of all the gaseous discs known so far is key to developing a more detailed understanding of the dynamical processes present in planetary debris discs at white dwarfs.
\section*{Acknowledgements}
The research leading to these results has received funding from the European Research Council under the European Union's Seventh Framework Programme (FP/2007-2013) / ERC Grant Agreement n.\,320964 (WDTracer). We would like to thank Yan-Ping Chen and Scott Trager for sharing their X-Shooter telluric template library. We thank the anonymous referee for a timely and constructive report.
Based on observations made with ESO Telescopes at the La Silla Paranal Observatory under programme IDs: 087.D-0139 and 386.C-0218. This work has made use of observations from the SDSS-III, funding for which has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Science Foundation, and the U.S. Department of Energy Office of Science. The SDSS-III web site is http://www.sdss3.org/. Based on observations made with the WHT operated on the island of La Palma by the Isaac Newton Group in the Spanish Observatorio del Roque de los Muchachos of the Instituto de Astrofísica de Canarias.
\bibliographystyle{mnras}
|
\section{Introduction}
\noindent
Many questions in classical statistics involve characterization problems,
which usually are instances of the following very general question:
\begin{problem}
Let ${\mathnormal X}_1, {\mathnormal X}_2,\dots, {\mathnormal X}_n$ be independent random variables with common
unknown distribution function $F$, and $T :=
T({\mathnormal X}_1, {\mathnormal X}_2,\dots, {\mathnormal X}_n)$ a statistic, based on
${\mathnormal X}_1, {\mathnormal X}_2,\dots, {\mathnormal X}_n$, with distribution function $G$.
Can $F$ be recovered from $G$?
\end{problem}
Problems of this kind are the central leitmotiv of the fundamental work of
Kagan, Linnik and Rao \cite{KaganLinnikRao:1973}.
In the present paper we solve the free version of the following problem,
which is still open in classical probability and might be called
$\chi^2$-conjecture,
see \cite[p.~466]{KaganLinnikRao:1973}:
\begin{Con}
\label{conj:KaganLinnikRao}
If ${\mathnormal X}_1, {\mathnormal X}_2,\dots, {\mathnormal X}_n$ are non-degenerate, independently and identically distributed classical random
variables with finite non-zero variance $\sigma^2$, then a necessary and
sufficient condition for ${\mathnormal X}_1$ to be normal is that
$\sum_{i=1}^n({\mathnormal X}_i-\overline{{\mathnormal X}})^2/\sigma^2$ be distributed as classical
chi-square distribution with $n -1$ degrees of freedom.
\end{Con}
The classical $\chi^2$-conjecture was studied previously by several authors.
The first result is due to Ruben \cite{Ruben:1974},
who proved the conjecture under the assumption that either $n=2$ or
${\mathnormal X}_1$ is symmetric.
It is not known whether the symmetry hypothesis
can be dropped for $n\geq 3$.
In a later paper \cite{Ruben:1975} Ruben used combinatorial tools
to show that the symmetry condition can be dropped
provided the sum of squares of the sample observations about the sample
mean, divided by $\sigma^2$, is distributed as chi-square for two distinct
sample sizes $m\neq n$ and $m,n\geq 2 $.
The proof given by Ruben is based on the cumulants of the sample variance
and is somewhat complicated.
Shortly later a simpler and more direct proof
based on the moments of the sample variance
was presented by Bondesson \cite{Bondesson:1977}.
The original problem was solved recently by Golikova and Kruglov
\cite{GolikovaKruglov:2015}
under the additional assumption that ${\mathnormal X}_1, {\mathnormal X}_2,\dots, {\mathnormal X}_n$
are independent infinitely divisible random variables.
The following related characterization problem
was solved by Kagan and Letac~\cite{KaganLetac:1999}:
Let
${\mathnormal X}_1 ,{\mathnormal X}_2 ,\dots, {\mathnormal X}_n$
be independent and identically distributed random variables
and assume that the distribution of the quadratic statistic
$\sum_{i=1}^n({\mathnormal X}_i-\overline{{\mathnormal X}}+a_i)^2$ depends only on $\sum_{i=1}^na_i^2$.
Then each ${\mathnormal X}_i$ have distribution $N(0, \sigma)$.
In the present paper we answer analogous questions in free probability.
Free probability and free convolution
was introduced by Voiculescu in \cite{Voiculescu:1985}
as a tool to study the von Neumann algebras of free groups.
Free probability is now an established field of research with deep connections
to combinatorics, random matrix theory, representation theory
and many analogies to classical probability.
Let us restrict our discussion to two specific ones, which are relevant
to the problems discussed in the present paper.
On the analytic side the Bercovici-Pata bijection
\cite{BercoviciPata:1999} provides a one-to-one correspondence
between infinitely divisible measures with respect to classical and
free convolution. For example, the analogue of the normal law is played
by Wigner's semicircle distribution which features as the limit law
in the free central limit theorem.
On the combinatorial side we will make heavy use of
free cumulants introduced by Speicher~\cite{Speicher:1994}.
Roughly speaking, any result about classical cumulants can be
translated to free probability by replacing the lattice of set partitions
by the lattice of noncrossing partitions.
Our standard reference for free cumulants is the book~\cite{NicaSpeicher:2006}.
We are concerned here with free analogues of characterization theorems
in the spirit of \cite {KaganLinnikRao:1973}.
The study of free analogues of classical theorems has witnessed
increasing interest during the last decade, see, e.g.,
\cite{BozejkoBryc:2006,Ejsmont:2013,HiwatashiKurodaNagisaYoshida:1999,Lehner:2004,SzpojankowskiWesolowski:2014,Szpojankowski:2014,Szpojankowski:2015}.
Many properties of free random variables are analogous to those of their
classical counterparts, in particular when they are picked according
to the Bercovici-Pata bijection. There are, however, exceptions, mostly
due to the failure of Marcinkiewicz' and Cram\'er's theorems in free probability.
In particular, Bercovici and Voiculescu \cite{BercoviciVoiculescu:1995}
showed that there exist free random variables
with a finite number of nonvanishing free cumulants which are not semicircular,
see \cite{ChistyakovGoetze:2011} for a characterization of such distributions.
This class of distributions appears in some (but not all)
free characterization problems which are analogues of classical
characterizations of the normal law,
cf.~\cite{Lehner:2003,ChistyakovGoetze:2011}.
In the present paper we show that
Conjecture~\ref{conj:KaganLinnikRao} also falls in this class of problems
and instead of Wigner laws we obtain the class of \emph{odd} laws,
i.e., laws with vanishing even cumulants.
Such laws do not exist in classical probability, but
can be constructed in free probability using the results of
\cite{ChistyakovGoetze:2011}. On the way we encounter a remarkable
cancellation phenomenon: odd cumulants do not contribute to the distributions
of certain quadratic statistics.
The paper is organized as follows. In section 2 we review basic free probability and the statement of the main result. Next in the subsection 2.2 we quote complementary facts, lemmas and indications. In the third section we prove our main results.
Finally, in section 4
we look more closely at the relation between the sample variance, the free commutator, $R-$cyclic matrices and free infinite divisibility.
\section{Free probability and statement of the main result}
\subsection{Basic Notation and Terminology}
A tracial noncommutative probability space is a pair $(\mathcal{A},\tau)$
where $\mathcal{A}$ is a von Neumann algebra, and $\tau:\mathcal{A} \to
{\mathbb C}$ is a normal, faithful, tracial state, i.e., $\tau$ is linear and
continuous in the weak* topology, $\tau({\mathnormal X} \Y)=\tau(\Y {\mathnormal X})$, $\tau(\mathrm{I})=1$, $\tau({\mathnormal X}\X^*)\geq 0$ and $\tau({\mathnormal X} {\mathnormal X}^{*}) = 0$ implies ${\mathnormal X} = 0$ for all ${\mathnormal X},\Y \in \mathcal{A}$.
The (usually taken to be self-adjoint) elements ${\mathnormal X}\in{\mathcal{A}}_{sa}$
are called (noncommutative) random variables.
Given a noncommutative random variable ${\mathnormal X}\in{\mathcal{A}}_{sa}$,
the distribution of ${\mathnormal X}$ in the state $\tau$ is the unique probability measure
$\mu_{\mathnormal X}$ on ${\mathbb R}$ (given by the spectral theorem) which reproduces its moments, i.e.,
$\tau( X^n)=\int_{{\mathbb R}} \lambda^n\,d\mu_{\mathnormal X}(\lambda)$
for $n\in{\mathbb N}$. This definition can be extended to self-adjoint possibly
unbounded operators ${\mathnormal X}$ affiliated to $\mathcal{A}$ by requiring that
$\tau( f({\mathnormal X}))=\int_{{\mathbb R}} f(\lambda)\,d\mu_{\mathnormal X}(\lambda)$
for any bounded Borel function $f$ on ${\mathbb R}$.
The set of affiliated operators is denoted by $\widetilde{\mathcal{A}}_{sa}$.
\subsection{Free Independence, Free Convolution and Free infinite Divisibility}
A family of von Neumann subalgebras $\left(\mathcal{A}_i\right)_{i\in I}$
of $\mathcal{A}$
are called \emph{free}
if $\tau({\mathnormal X}_{1} \dots {\mathnormal X}_{n} ) = 0$ whenever $\tau({\mathnormal X}_{j} ) = 0$ for all
$j = 1,\dots, n$ and ${\mathnormal X}_{j} \in \mathcal{A}_{i(j)}$ for some indices $i(1)\neq i(2)\neq \dots \neq i(n)$.
Random variables ${\mathnormal X}_{1},\dots ,{\mathnormal X}_{n} $ are freely independent (free) if the
subalgebras they generate are free.
Free random variables can be constructed using the reduced free product
of von Neumann algebras \cite{Voiculescu:1985}.
For more details about free convolutions and free probability theory, the reader can consult \cite{NicaSpeicher:2006,VoiculescuDykemaNica:1992}.
It can be shown that the joint distribution of free random variables $X_i$
is uniquely determined by the distributions of the individual random variables
$X_i$ and therefore the operation of \emph{free convolution} is well defined:
Let $\mu$ and $\nu$ be probability measures on ${\mathbb R}$, and
${\mathnormal X},\Y$ self-adjoint free random variables with respective distributions
$\mu$ and $\nu$,
The distribution of ${\mathnormal X}+\Y$ is called the free additive convolution of $\mu$
and $\nu$ and is denoted by $\mu \boxplus \nu$.
In analogy with classical probability,
a probability measure $\mu$ on ${\mathbb R}$ is said to be
\emph{freely infinitely divisible} (or FID for short)
if for each $n \in \{1, 2, 3, \dots \}$ there exists a probability measure
$\mu_n$ such that
$\mu=\underbrace{ \mu_n\boxplus\dots\boxplus\mu_n}_{n-times}$.
\subsection{The Cauchy-Stieltjes Transform and Free Convolution}
The analytic approach to free convolution uses the Cauchy transform
\begin{align}
\label{eq:hm:ball}
G_\mu(z)=\int_{{\mathbb R}}\frac{1}{z-y}\mu(dy).
\end{align}
of a probability measure $\mu$. It is analytic on the upper half
plane ${\mathbb C}^+=\{x+iy|x,y\in {\mathbb R}, y>0\}$
and takes values in the closed lower half plane
${\mathbb C}^-\cup{\mathbb R}$.
The Cauchy transform has an inverse at a neighbourhood of infinity
which has the form
$$
G_\mu^{-1}(z) = \frac{1}{z} + R_\mu(z)
$$
where $R_\mu(z)$ is analytic in a neighbourhood of zero and is called
\emph{$R$-transform}.
Then free convolution is defined (see~\cite{Voiculescu:1986})
via the identity
\begin{align} \label{freeconv}
R_{\mu\boxplus\nu}=R_\mu+R_\nu.
\end{align}
The coefficients of the $R$-transform
\begin{align} \label{rtr}
R_{{\mathnormal X}}(z)=\sum_{n=0}^{\infty}\,K_{n+1}({\mathnormal X})\,z^n.
\end{align}
are called \emph{free cumulants} of the random variable ${\mathnormal X}$.
The Cauchy transform is related to the
moment generating function $M_{{\mathnormal X}}$ as follows:
\begin{align}\label{mgf}
M_{{\mathnormal X}}(z)=\sum_{n=0}^{\infty}\,\tau({\mathnormal X}^n)\,z^n = \frac{1}{z}G_{\mathnormal X}\left(\frac{1}{z}\right).
\end{align}
\subsection{Some probability distributions}
Let us now recall basic properties of some specific probability distributions
which play prominent roles in the present paper.
\subsubsection{Wigner semicircular distribution}
A non-commutative random variable ${\mathnormal X}$ is said to be free normal variable
(i.e. have Wigner semicircular distribution) if the Cauchy-Stieltjes transform
is given by the formula
\begin{eqnarray}
G_{\mu}(z)=\frac{z -\sqrt{z ^2 - 4}}{2} , \label{eq:GtransformataMixner}
\end{eqnarray}
where $|z|$ is big enough, where the branch of the analytic square root should be determined by the condition
that $\Im(z)>0\Rightarrow \Im(G_\mu(z))\leqslant 0$ (see \cite{SaitohYoshida:2001}).
Equation (\ref{eq:GtransformataMixner}) describes the family of distributions with mean zero and variance one (see \cite{Ejsmont:2012,SaitohYoshida:2001}).
This measure has density
$$\frac{\sqrt{4-x^2}}{2\pi},$$
on $-2 \leq x \leq 2 $.
The Wigner semicircular distribution have cumulants $
K_i=0
$
for $i>2$.
\subsubsection{Free Poisson distribution}
A non-commutative random variable ${\mathnormal X}$ is said to be free-Poisson variable if it has Marchenko-Pastur (or free-Poisson) distribution $\nu=\nu(\lambda,\alpha)$ defined by the formula
\begin{align} \label{MPdist}
\nu=\max\{0,\,1-\lambda\}\,\delta_0+ \tilde{\nu},
\end{align}
where $\lambda\ge 0$ and the measure $\tilde{\nu}$, supported on the interval $(\alpha(1-\sqrt{\lambda})^2,\,\alpha(1+\sqrt{\lambda})^2)$, $\alpha>0$ has the density (with respect to the Lebesgue measure)
$$
\tilde{\nu}(dx)=\frac{1}{2\pi\alpha x}\,\sqrt{4\lambda\alpha^2-(x-\alpha(1+\lambda))^2}\,dx.
$$
The parameters $\lambda$ and $\alpha$ are called the rate and the jump size, respectively.
It is worth to note that a non-commutative variable with Marchenko-Pastur distribution arises also as a limit in law (in non-commutative sense) of variables with distributions $((1-\frac{\lambda}{N})\delta_0+\frac{\lambda}{N}\delta_{\alpha})^{\boxplus N}$ as $N\to\infty$, see \cite{NicaSpeicher:2006}. Therefore, such variables are often called free-Poisson. It is easy to see that if $X$ is free-Poisson, $\nu(\lambda,\alpha)$, then $K_n({\mathnormal X})=\alpha^n\lambda$. Therefore its
$R$-transform has the form
$$R(z)=\frac{\lambda\alpha}{1-\alpha z}.$$
\subsubsection{Free chi-square distribution}
Let ${\mathnormal X}_1,\dots,{\mathnormal X}_n$ be free identically distributed random variables from the Wigner semicircular distribution with non-zero variance $\sigma^2$ and mean zero, and $\delta=\sum_{i=1}^n m_i^2$ $(m_i \in {\mathbb R})$. We call the distribution of the random variable $\sum_{i=1}^n ({\mathnormal X}_i+m_i)^2$ the free chi-square distribution with $n$ degrees of freedom
and noncentrality parameter $\delta$, and we denote this distribution $\chi^2(n,\sigma,\delta)$ (a first version of this definition was introduced in \cite{HiwatashiKurodaNagisaYoshida:1999}).
In terms of $R$-transforms, a random variable $Y$ has distribution
$\chi^2(n,\sigma,\delta)$ if and only if
\begin{align} R_Y(z)=\frac{n\sigma^2}{1-\sigma^2z}+\frac{\delta}{(1-2z)^2}. \label{eq:rtransormatachisq}\end{align}
If $\delta=0$, the free chi-square distribution is called central, otherwise non-central and then we will write $\chi^2(n,\sigma)$ and from \eqref{eq:rtransormatachisq} we see that
$\chi^2(n,\sigma)$ has the Marchenko-Pastur distribution $\nu(n,\sigma^2)$. Moreover, we will use the notation $\chi^2(n):=\chi^2(n,1)$.
It was shown in \cite{HiwatashiKurodaNagisaYoshida:1999}
that these distributions form a semigroup, namely
$\chi^2(n_1,\sigma,\delta_1)\boxplus\chi^2(n_2,\sigma,\delta_2)=
\chi^2(n_1+n_2,\sigma,\delta_1+\delta_2)$.
\subsubsection{Even elements}
We call an element ${\mathnormal X}\in \mathcal{A}$ even if all its odd moments vanish, i.e. $\tau({\mathnormal X}^{2i+1})=0$ for all $i\geq 0.$
It is immediately seen that the vanishing of all odd moments is
equivalent to the vanishing of all odd cumulants, i.e., $K_{2i+1}({\mathnormal X})=0$
and thus the even cumulants contain the complete information about the distribution of an even element. The sequence
$\alpha_n=K_{2n}({\mathnormal X})$ is called the \emph{determining sequence} of $X$.
\subsubsection{Odd elements}
We call an element ${\mathnormal X}\in \mathcal{A}$ \emph{odd} if $K_2({\mathnormal X})>0$ and all its
even free cumulants of order higher than two vanish,
i.e. if $K_{2i}({\mathnormal X}) = 0$ for all $i \geq 2$.
The basic example of such a law is Wigner's semicircular distribution.
The classical analogue of odd elements only include the normal distribution
because otherwise we could construct a normal random variable
which is the sum of independent non-normal random variables
(see below for the free case).
This contradicts Cram\'er's decomposition theorem.
However the free analogues of Marcinkiewicz' and Cram\'er's theorems fail.
Bercovici and Voiculescu~\cite{BercoviciVoiculescu:1995} showed that
there exist probability distributions $\mu_\epsilon$ with free cumulants
$K_1({\mathnormal X})=0$, $K_2({\mathnormal X})=1$, $K_3({\mathnormal X})=\epsilon$ and $K_i({\mathnormal X})=0$
for $i\geq 4 $ if $\epsilon$ is small enough.
This is an odd element and thus an explicit counterexample to the free analogue
of Marcinkiewicz' theorem. To invalidate Cram\'er's theorem,
take free copies ${\mathnormal X}_1$ and ${\mathnormal X}_2$ of random variables with distribution $\mu_\epsilon$, then the difference ${\mathnormal X}_1-{\mathnormal X}_2$ is semicircular.
Chistyakov and G\"otze \cite{ChistyakovGoetze:2011} gave a detailed
description of laws with finitely many free cumulants of arbitrary order.
Thus an abundance of odd laws exists.
\subsection{The main result}
The main result of this paper is the following characterization of odd elements
in terms of the sample variance. The proof of this theorem is given in
Section 3.
\textbf{The sample variance} of a finite sequence of random variables $X_i$
is the quadratic form
\begin{equation}
S^2_n
=\frac{1}{n}\sum_{i=1}^n({\mathnormal X}_i-\overline{{\mathnormal X}})^2
=\frac{1}{n}\Big(1-\frac{1}{n}\Big)\sum_{i=1}^n{\mathnormal X}_i^2-\frac{1}{n}\sum_{i,j=1, \textrm{ }i\neq j}^n{\mathnormal X}_i{\mathnormal X}_j
=\frac{1}{n^2}\sum_{1\leq i < j \leq n} ({\mathnormal X}_i-{\mathnormal X}_j)^2.
\label{eq:SampleVariance}
\end{equation}
However in order to simplify notation in the present paper we chose to
consider and call ``sample variance'' the rescaled quadratic form $Q_n=n S^2_n
=\sum_{i=1}^n({\mathnormal X}_i-\overline{{\mathnormal X}})^2
$.
Our main result resolves the free analogue of $\chi^2$-conjecture.
\begin{theo}
Let ${\mathnormal X}_1, {\mathnormal X}_2,\dots, {\mathnormal X}_n \in \mathcal{A}_{sa}$ be free copies of
a random variable $X$ with finite non-zero variance $\sigma^2$.
Then ${\mathnormal Q}_n$
is distributed according to $\chi^2(n-1,\sigma)$ if and only if ${\mathnormal X}$ is odd.
\label{twr:OddElement}
\end{theo}
Depending on the point of view it can be interpreted both as a positive
and a negative solution.
Taking into account the failure of Marcinkiewicz theorem this confirms
the free analogue of the $\chi^2$-conjecture in the broad sense.
If we suppose in addition that the distribution is even,
then the above theorem gives a positive answer to the free analogue of
Ruben's first theorem \cite{Ruben:1974}.
\begin{prop}
Let ${\mathnormal X}_1, {\mathnormal X}_2,\dots, {\mathnormal X}_n $ be free identically distributed random variables with finite non-zero variance $\sigma^2$, and assume that the
distribution of ${\mathnormal X}_1$ is symmetric.
Then ${\mathnormal Q}_n$
is distributed as $\chi^2(n-1,\sigma)$ if and only if ${\mathnormal X}_1$ has
Wigner semicircular law.
\label{twr:1}
\end{prop}
On the other hand, the free analogue of Ruben's second theorem
\cite{Ruben:1975} (see also \cite{Bondesson:1977}) does not hold:
\begin{prop}
Let ${\mathnormal X}_1, {\mathnormal X}_2,\dots$ denote free independently and identically distributed random variables with finite non-zero variance $\sigma^2$. Let $m, n$ denote
distinct integers not less than 2.
Then for ${\mathnormal Q}_n/\sigma^2$ and ${\mathnormal Q}_m/\sigma^2$
to be distributed as $\chi^2(n-1)$ and $\chi^2(m-1)$, respectively,
it is not necessary that ${\mathnormal X}_1$ is semicircular.
\label{twr:2}
\end{prop}
\begin{proof}[Proof of Proposition~\ref{twr:1}]
If ${\mathnormal Q}_n$
is distributed as $\chi^2(n-1,\sigma)$ then ${\mathnormal X}_1$ is odd, but taking into account that ${\mathnormal X}_1$ is symmetric we have that its odd central moments vanish, and therefore its odd cumulants higher than the first vanish, so ${\mathnormal X}_1$ has Wigner semicircular distribution.
\\
\textit{Proof of Proposition~\ref{twr:2}.}
Assume that $K_1({\mathnormal X}_1) = 0$, $K_2({\mathnormal X}_1) = \sigma^2$, $K_3({\mathnormal X}_1) = \epsilon$ and $K_i({\mathnormal X}_1)=0$ for $i\geq 4$ where $\epsilon$ is small enough. By Theorem \ref{twr:OddElement} we see that ${\mathnormal Q}_n$ and ${\mathnormal Q}_m$ have $\chi^2(n-1,\sigma)$ and $\chi^2(m-1,\sigma)$ distribution respectively.
\end{proof}
\begin{Rem}
In this paper we assume that the involved random variables are bounded,
that is ${\mathnormal X}_i \in \mathcal{A}$, as was common practice for a long time.
Recently however unbounded random variables, i.e., operators affiliated with
the von Neumann algebra in question, came into the focus of research.
This happened in particular in connection with certain characterization
problems, see, e.g.,
\cite{ChistyakovGoetzeLehner:2011,EjsmontFranzSzpojankowski:2015,
ChistyakovGoetzeLehner:2016}.
It follows from the following result Chistyakov and Goetze
that for the characterization problems pertinent to the present paper
the question of boundedness is unessential.
\end{Rem}
\begin{lemm}[{ \cite[Lemma~3.10]{ChistyakovGoetze:2011}}]
\label{lemm:bounded}
Assume that $\mu=\mu_1\boxplus\mu_2$, where $\mu$ has compact support.
Then $\mu_1$ and $\mu_2$ have compact support as well.
\end{lemm}
In terms of operators this means that
if ${\mathnormal X},\Y \in \widetilde{\mathcal{A}}_{sa}$ are free random variables
affiliated with $\mathcal{A}$ and
such that ${\mathnormal X}+\Y\in \mathcal{A}_{sa}$, i.e., ${\mathnormal X}+\Y$ is bounded,
then ${\mathnormal X},\Y\in \mathcal{A}_{sa}$.
Now we will show that Theorem \ref{twr:OddElement}
is true under weaker conditions.
\begin{cor}
Let ${\mathnormal X}_1, {\mathnormal X}_2,\dots, {\mathnormal X}_n \in \widetilde{\mathcal{A}}_{sa}$
be selfadjoint free random variables and assume
${\mathnormal Q}_n=\frac{1}{n}\sum_{i<j}(X_i-X_j)^2$ is bounded. Then all $X_i$ are bounded.
\end{cor}
\begin{proof}
Since ${\mathnormal X}_i$ are self-adjoint and
$n{\mathnormal Q}_n=\sum_{1\leq i \leq j \leq n} ({\mathnormal X}_i-{\mathnormal X}_j)^2$ is bounded,
it follows that $({\mathnormal X}_i-{\mathnormal X}_j)^2$ is bounded
and hence also ${\mathnormal X}_i-{\mathnormal X}_j$.
By Lemma \ref{lemm:bounded} we deduce that all ${\mathnormal X}_i$ are bounded.
\end{proof}
The proof of Ruben's theorem \cite{Ruben:1974} heavily relies on the symmetry
of random variables. Is it possible to drop the hypothesis that the
random variables are symmetric? Golikova and Kruglov \cite{GolikovaKruglov:2015}
give a partial answer to this question -- instead of symmetry of ${\mathnormal X}_1$ they assume infinite divisibility. The following is a free version of their result
which characterizes the classical normal law by the sample variance.
We drop the assumption that ${\mathnormal X}_i$ have the same distribution, because with this assumption the result follows directly from Theorem \ref{twr:OddElement} (we cannot use the Bercovici-Pata bijection to prove it because it does not map classical chi-square to free chi-square distributions).
\begin{prop}
Let ${\mathnormal X}_1, {\mathnormal X}_2,\dots {\mathnormal X}_n$ denote free independent, freely infinitely divisible random variables with mean $\tau({\mathnormal X}_1)=\tau({\mathnormal X}_2)= \dots =\tau({\mathnormal X}_n)$ and $\Var({\mathnormal X}_1)=\Var({\mathnormal X}_2)= \dots =\Var({\mathnormal X}_n)=1$.
Then if ${\mathnormal Q}_n$
is distributed as free $\chi^2(n-1)$ if and only if ${\mathnormal X}_1, . . . , {\mathnormal X}_n$ are
identically distributed Wigner semicircular random variables.
\label{twr:kruglow}
\end{prop}
We conclude with a free version of a the following result of
Kagan and Letac \cite{KaganLetac:1999}:
Fix an integer $n\geq 3$ and let ${\mathnormal X}_1, {\mathnormal X}_2,\dots, {\mathnormal X}_n$
be independent identically distributed random variables.
Consider the linear subspace $E=\mathbf{1}^\perp$ of Euclidean space ${\mathbb R}^n$ , i.e.,
the hyperplane $E=\{(a_1, a_2,\cdots,a_n ) : a_1+a_2+\cdots+a_n =0\}$.
Then the following characterizations hold:
\begin{enumerate}[(i)]
\item If the distribution of the $E$-valued random variable
$$V=({\mathnormal X}_1-\overline{{\mathnormal X}},\dots,{\mathnormal X}_n-\overline{{\mathnormal X}})$$
is invariant under all rotations of the Euclidean space $E$, then the ${\mathnormal X}_i$'s
are normally distributed.
\item If the distribution of the random variable
$$\sum_{i=1}^n({\mathnormal X}_i-\overline{{\mathnormal X}}+a_i)^2$$
does not change as the real parameters $a_i$ vary on a sphere (i.e.,
the euclidean length $\norm{a}^2=a_1^2+a^2_2+\cdots+a_n^2 $ remains constant),
then the ${\mathnormal X}_i$'s are normally distributed.
\end{enumerate}
A key ingredient of the proof of these classical results is played by
\emph{Marcinkiewicz' theorem}.
As we discussed above, Marcinkiewicz' theorem has no analogue in free probability
and we will use different methods to prove the following free version of
\cite{KaganLetac:1999}.
This method also works in classical probability if we assume that all moment exists.
\begin{prop}
Let $n$ be a fixed integer $n\geq 3$. Let ${\mathnormal X}_1, {\mathnormal X}_2,\dots, {\mathnormal X}_n$ be free identically distributed random variables.
\begin{enumerate}
\item If for all $a \in E\subset {\mathbb R}^n$
the distribution of the random variable
$$\sum_{i=1}^n({\mathnormal X}_i-\overline{{\mathnormal X}}+a_i)^2$$
depends only on $\norm{a}^2=a_1^2+a^2_2+\cdots+a_n^2 $,
then the ${\mathnormal X}_i$'s obey the semicircle law.
\item If the distribution of the $E$-valued random variable
$$V=({\mathnormal X}_1-\overline{{\mathnormal X}},\dots,{\mathnormal X}_n-\overline{{\mathnormal X}})$$
is invariant under all rotations of the Euclidean space $E$, then the ${\mathnormal X}_i$'s
obey the semicircle law.
\end{enumerate}
\label{twr:OdpowidenikKaganLetac}
\end{prop}
\subsection{Noncrossing Partitions}
Let $S$ be finite subset of ${\mathbb N}$.
A partition of $S$ is a set of mutually disjoint subsets
(also called \emph{blocks}) $B_1,B_2,\dots,B_k\subseteq S$
whose union is $S$. Any partition $\pi$ defines an equivalence relation on $S$,
denotes $\sim_\pi$, such that the equivalence classes are the blocks $\pi$.
That is, $i\sim_\pi j$ if $i$ and $j$ belong to the same block of $\pi$.
A partition $\pi$ is called \emph{noncrossing}
if different blocks do not interlace, i.e., there is no quadruple of
elements $i<j<k<l$ such that $i\sim_\pi k$ and $j\sim_\pi l$
but $i\not\sim_\pi j$.
The set of non-crossing partitions of $S$ is denoted by $\NC(S)$,
in the case where $S=[n]:=\{1, \dots , n\}$ we write
$\NC(n):=\NC([n])$.
$\NC(n)$ is a poset under refinement order, where we say $\pi\leq \rho$ if
every block of $\pi$ is contained in a block of $\rho$.
It turns out that $\NC(n)$ is in fact a lattice,
see \cite[Lecture~9]{NicaSpeicher:2006}.
The maximal element of $\NC(n)$ under this order
is denoted by $\hat{1}_{n}$. It is the partition consisting
of only one block. On the other hand the minimal element $\hat{0}_n$
is the unique
partition where every block is a singleton.
Sometimes it is convenient to visualize partitions as diagrams, for example
$\hat{1}_n=\input{sp1n}$
and $\hat{0}_n=\input{sp0n}$.
\subsection{Some Special Notations}
We will be concerned with certain special classes of noncrossing partitions.
If $n$ is even we denote by $\NCeven(n)$
the subset of even noncrossing partitions,
where we say that a partition is even if all its blocks have even cardinality.
Even more specific we denote by $\NC_2(n)$ is the set of all noncrossing pair
partitions, i.e., partitions where every block has size $2$.
Two specific minimal pair partitions will play a particularly important role,
namely
$\onetwo{r}=\input{minpartition2}\in\NC(2r)$, which is a kind of blow up
of $\hat{1}_r$ and its shift $\pispecial{r}=\input{minpartition1}\in\NC(2r)$.
In the proof of Theorem~\ref{twr:OddElement} we will use telescoping argument
and put a filtration on $\NC(n)$ by avoiding certain blocks. For this
purpose we introduce the following notation.
For a subset $B\subseteq {\mathbb N}$ let
$\NC^{B}(n):=\{\pi\in \NC(n): B\in\pi\}$, i.e., the collection
of noncrossing partitions which contain $B$ as a block.
On the other hand,
for a family $B_1,B_2,\dots,B_m\subseteq {\mathbb N}$ of subsets
let
$\NC_{B_1,\dots,B_m}(S):=\{\pi\in \NC(S):\pi\cap\{B_1,\dots,B_m\}=\emptyset\}$,
i.e., the collection of noncrossing
partition which do not contain any $B_i$ as a block.
Finally, combining the two notations we define
$\NC^{B}_{B_1,\dots,B_m}(n):=\NC^{B}(n)\cap\NC_{B_1,\dots,B_m}(n)$.
\subsection{Kreweras Complements}
Kreweras \cite{Kreweras:1972} discovered an interesting antiisomorphism
of the lattice $\NC(n)$, now called the \emph{Kreweras complementation map},
of which we will need two variants.
Given a noncrossing partition $\pi$ of $\{1,2,\dots,n\}$,
the \emph{left Kreweras complement} $\Krewl\pi$ is the maximal
noncrossing partition of the ordered set $\{\bar{1},\bar{2},\dots,\bar{n}\}$
such that $\pi\cup\Krewl\pi$ is a noncrossing partition of
the interlaced set
$\{\bar{1},1,\bar{2},2,\dots,\bar{n},n\}$.
Similarly, the \emph{right Kreweras complement} $\Krewr\pi$ is the maximal
noncrossing partition of the ordered set $\{\bar{1},\bar{2},\dots,\bar{n}\}$
such that $\pi\cup\Krewr\pi$ is a noncrossing partition of
the interlaced set
$\{1,\bar{1},2,\bar{2},\dots,n,\bar{n}\}$.
It is then clear that $\Krewr\circ\Krewl=\mathrm{id}$ and it can be shown
that
\begin{equation}
\label{eq:cardKrew}
\abs{\Krewr\pi}=\abs{\Krewl\pi}=n+1-\abs\pi
.
\end{equation}
Finally we define the \emph{extended Kreweras complement} $\Krewlr\pi$
to be the maximal
noncrossing partition of the ordered set $\{\bar{0},\bar{1},\dots,\bar{n}\}$
such that $\pi\cup\Krewlr\pi$ is a noncrossing partition of
the interlaced set
$\{\bar{0},1,\bar{1},2,\bar{2},\dots,n,\bar{n}\}$.
The extended Kreweras complement is always irreducible,
i.e., $\bar{0}$ and $\bar{n}$ are in the same block of $\Krewlr\pi$.
In fact it is obtained by joining $\bar{0}$ to the last block of $\Krewr\pi$,
i.e., the block containing $\bar{n}$,
or by joining $n+1$ to the first block of $\Krewl\pi$.
The following observation is useful for recursive proofs involving
the Kreweras complement(s).
\begin{lemm}
\label{lemm:kreweras}
Let $\pi\in\NC(n)$ and $B=\{j_1,j_2,\dots,j_p=n\}$ be its last block.
Let $\pi_1,\pi_2,\dots,\pi_p$ be the restrictions of $\pi$ to the
maximal intervals of $\{1,2,\dots,n\}\setminus B$ as shown in
the following picture:
$$
\input{kreweras1}
$$
Then the left Kreweras complement of $\pi$ is the concatenation
of the extended Kreweras complements of the subpartitions $\pi_j$:
$$
\Krewl\pi = \Krewlr{\pi_1}\,\Krewlr{\pi_2}\dotsm\Krewlr{\pi_p}.
$$
\end{lemm}
\subsection{Free Cumulants}
Let ${\mathbb C} \langle {\mathnormal X}_{1},\dots ,{\mathnormal X}_{n} \rangle$ denote
the non-commutative ring of polynomials in variables ${\mathnormal X}_{1},\dots
,{\mathnormal X}_{n}\in\mathcal{A} $.
The free cumulants are multilinear maps $K_r : \mathcal{A}^r\to{\mathbb C}$
defined implicitly by the relation (connecting them with mixed moments)
\begin{align}
\tau({\mathnormal X}_{1}{\mathnormal X}_{2}\dots {\mathnormal X}_{n}) = \sum_{\pi \in \NC(n)}K_{\pi}({\mathnormal X}_{1},{\mathnormal X}_{2},\dots ,{\mathnormal X}_{n}),\label{eq:DefinicjaKumulant}
\end{align}
where
\begin{align}
K_{\pi}({\mathnormal X}_{1},{\mathnormal X}_{2},\dots ,{\mathnormal X}_{n}):=\Pi_{B \in \pi}K_{|B|}({\mathnormal X}_{i}:i \in B) \label{eq:DefinicjaProduktuKumulant}
\end{align}
and $\NC(n)$ is the set of all non-crossing partitions of $\{1, 2,\dots, n \}$
(see \cite{NicaSpeicher:2006}). Sometimes we will write $K_{r}({\mathnormal X})=K_{r}(
{\mathnormal X},\dots ,{\mathnormal X} )$.
Free cumulants provide the most important technical tool to investigate
free random variables.
This is due to the basic property of \emph{vanishing of mixed cumulants}.
By this we mean the fact that
$$
K_r(X_1,X_2,\dots,X_n)=0
$$
for any family of random variables $X_1,X_2,\dots,X_n$ which can be partitioned
into two free subsets.
For free sequences this can be reformulated as follows.
Let $(X_i)_{i\in{\mathbb N}}$ be a sequence of free random variables and
$h:[r]\to{\mathbb N}$ a map. We denote by $\ker h$ the set partition
which is induced by the equivalence relation
$$
k\sim_{\ker h} l
\iff
h(k)=h(l)
.
$$
Similarly, for a multiindex $\underline{i} = i_1i_2\dots i_n$ we denote its
kernel
$\ker\underline{i}$ by the relation $k\sim l$ if $i_k=i_l$.
Using this notation, we have that
\begin{equation}
\label{eq:kerh>=pi}
K_\pi(X_{h(1)},X_{h(2)},\dots,X_{h(r)})=0
\text{ unless $\ker h\geq \pi$.}
\end{equation}
Our main technical tool is the free version,
due to Krawczyk and Speicher~\cite{KrawczykSpeicher:2000} (see also \cite[Theorem 11.12]{NicaSpeicher:2006}), of the classical formula of
James/Leonov and Shiryaev \cite{James:1958,LeonovShiryaev:1959} which
expresses cumulants of products in terms of individual cumulants.
\begin{theo}
\label{thm:krawczyk}
Let
$r,n \in {\mathbb N}$ and $ i_1 < i_2 < \dots < i_r = n$ be given and let
$$\rho=\{(1,\dots,i_1),\dots,(i_{r-1}+1,\dots,i_r)\}\in \NC(n)$$
be the induced interval partition.
Consider now random variables ${\mathnormal X}_1,\dots,{\mathnormal X}_n\in\mathcal{A}$.
Then the free cumulant of the products can be expanded
as follows:
\begin{align}
K_r({\mathnormal X}_1\dots {\mathnormal X}_{i_1},\dots,{\mathnormal X}_{i_{r-1}+1}\dots{\mathnormal X}_n)=\sum_{\substack{\pi\in \NC(n) \\ \pi\vee\rho=\hat{1}_{n}} }K_\pi({{\mathnormal X}_1,\dots,{\mathnormal X}_n}).
\end{align} \label{twr:produktargumentow}
\end{theo}
In the special case of products of pairs of free elements this yields
the following formula for multiplicative free convolution.
\begin{theo}[{\cite[Theorem~14.4]{NicaSpeicher:2006}}]
\label{thm:multfreeconvolution}
Let $\{X_1,X_2,\dots,X_r\}$ and $\{Y_1,Y_2,\dots,Y_r\}$ be two
mutually free sets of random variables, then
$$
K_r(X_1Y_1,X_2Y_2,\dots,X_rY_r)=\sum_{\pi\in\NC(r)} K_\pi(X_1,X_2,\dots,X_r)\,
K_{\Krewr{\pi}}(Y_1,Y_2,\dots,Y_r)
$$
\end{theo}
This motivates the following definition.
\begin{defi}[{\cite[Ch.~17]{NicaSpeicher:2006}}]
\label{def:boxedconvolution}
Let $(a_n)_{n\geq0}$ a sequence and $\pi\in\NC(n)$ a noncrossing partition.
As in \eqref{eq:DefinicjaProduktuKumulant} we denote by
$a_\pi$ the product
$$
a_\pi=\prod_{B\in\pi} a_{\abs{B}}
.
$$
Given two sequences $(a_n)_{n\geq0}$ and $(b_n)_{n\geq0}$ with
respective generating formal power series
$f(z)=\sum_{n=1}^\infty a_nz^n$ and $g(z)=\sum_{n=1}^\infty b_nz^n$
we define their \emph{boxed convolution} as the formal series
$f\boxstar g(z)=\sum_{n=1}^\infty c_nz^n$ with coefficients
$$
c_n = \sum_{\pi\in\NC(n)} a_\pi b_{\Krewr{\pi}}
= \sum_{\pi\in\NC(n)} a_{\Krewl\pi} b_{\pi}
.
$$
\end{defi}
As examples, consider the univariate case
of Theorem~\ref{thm:multfreeconvolution}, which can be rewritten
as $R_{XY}(z) = R_X\boxstar R_Y(z)$,
or \cite[Proposition~11.25]{NicaSpeicher:2006}, which states
that the $R$-transform of an even element $X$ can be written as
$$
R_{X^2}(z) = \alpha\boxstar \zeta(z)
$$
where
$\alpha(z) = \sum_{n=1}^\infty K_{2n}(X)z^n$ is the determining series of $X$
and $\zeta(z)=\sum_{n=1}^\infty z^n$ is the so-called \emph{Zeta-series}.
Combinatorially this means
\begin{equation}
\label{eq:KX2}
K_r(X^2) = \sum_{\pi\in\NC(r)}\alpha_\pi = \sum_{\pi\in\NC(r)}\prod_{B\in\pi} K_{2\abs{B}}(X).
\end{equation}
\noindent
The next result follows from \cite[Proposition 2.2]{HiwatashiKurodaNagisaYoshida:1999}; see Corollary~\ref{cor:KrQn} below for a generalization.
\begin{prop}
\label{twr:Lehner:2003}
Let ${\mathnormal X}_1,{\mathnormal X}_2,\dots,{\mathnormal X}_n$ be free identically distributed Wigner semicircular random variables with mean zero and variance $\sigma^2$. Then the cumulants of ${\mathnormal Q}_n$ are given as follows:
\begin{align}
K_r({\mathnormal Q}_n) &=(n-1)\,\sigma^{2r}.
\end{align}
\end{prop}
The following lemma connects Theorem~\ref{thm:krawczyk}
with Definition~\ref{def:boxedconvolution} and is the key to the main result.
Its proof is contained in the proof of Proposition~11.25 in
the book \cite{NicaSpeicher:2006}.
\begin{lemm}
\label{lemm:lematoparzystych}
Let $r\in\mathbb{N}$ and $\pi \in \NCeven(2r)$,
then $\pi \vee \onetwo{r}=\hat{1}_{2r}$ if and only if $\pi\geq \pispecial{r}$,
i.e., $1$ and $2r$ lie in the same block of $\pi$ and elements $2i$ and $2i+1$
also lie in the same block of $\pi$ for $i\in[r-1]$.
Consequently
$$
\{\pi : \pi \vee \onetwo{r}=\hat{1}_{2r}\}\cap \NCeven(2r)
= [\pispecial{r}, \hat{1}_{2r} ] ,
$$
is a lattice isomorphic to $\NC(r)$.
\end{lemm}
\begin{cor} \label{corr:dwupartycje}
There is
only one non-crossing pair partition $\pi$ such that $\pi \vee
\onetwo{r}=\hat{1}_{2r} $,
namely $\pispecial{r}=\input{minpartition1}=\{(1,2r),(2,3),\dots,(2r-2,2r-1)\}$.
\end{cor}
\begin{defi}
Let $B_1,\dots,B_m $ be subsets of ${\mathbb N}$
and random variables ${\mathnormal X}_1,\dots,{\mathnormal X}_n\in\mathcal{A},$ be given.
Then for an interval partition $\rho=\{(1,\dots,i_1),(i_1+1,\dots,i_2),\dots,
(i_{r-1}+1,\dots,i_r)\}$
we define the partial cumulant functional
\begin{align}
K^\rho_{B_1,\dots,B_m}({\mathnormal X}_1,{\mathnormal X}_2,\dots,{\mathnormal X}_n)=\sum_{\substack{\pi\in \NC_{B_1,\dots,B_m} (n)\\ \pi\vee\rho=\hat{1}_{n}} }K_\pi({{\mathnormal X}_1,\dots,{\mathnormal X}_n}),
\end{align}
Usually we will abuse notation and abbreviate this expression as
\begin{align}
K^r_{B_1,\dots,B_m}({\mathnormal X}_1\dots {\mathnormal X}_{i_1},\dots,{\mathnormal X}_{i_{r-1}+1}\dots{\mathnormal X}_{i_r})=\sum_{\substack{\pi\in \NC_{B_1,\dots,B_m} (n)\\ \pi\vee \rho=\hat{1}_{n}} }K_\pi({{\mathnormal X}_1,\dots,{\mathnormal X}_n}).
\end{align}
\end{defi}
\begin{lemm}
\label{lem:centeredstatistic}
Let $P=P(X_1,X_2,\dots,X_n)$ be a polynomial
of degree at most two in noncommuting variables
$X_1,X_2,\dots,X_n$. Then $\tau(P({\mathnormal X}_1,{\mathnormal X}_2,\dots,{\mathnormal X}_n))=0$
for every i.i.d.\ free family $X_i$
if and only if
$$
\sum_{\sigma\in\SG_n} P(X_{\sigma(1)},X_{\sigma(2)},\dots,X_{\sigma(n)})=0
.
$$
\end{lemm}
\begin{proof}
Clearly by symmetry the second condition is stronger than the first condition.
In order to show that it is also necessary, we first note that
by a simple scaling argument we may assume without loss of generality
that the polynomial in consideration is homogeneous.
Clearly such a polynomial cannot have a constant term and
we start with a linear polynomial $P=\sum_{i=1}^n \alpha_i X_i$.
By evaluating a distribution with nonzero first moment it follows
that $\sum_{i=1}^n \alpha_i=0$. But then we have
$$
\sum_{\sigma\in\SG_n}P(X_{\sigma(1)},X_{\sigma(2)},\dots,X_{\sigma(n)})
=\sum_{i=1}^n\alpha_i(n-1)!\sum_{k=1}^n X_k=0
.
$$
Let us now turn to a homogeneous polynomial of second order
$$
P = \sum_{i,j=1}^n \alpha_{ij}X_iX_j
.
$$
Evaluating at a distribution with first moment $\mu_1$ and second moment $\mu_2$
we obtain
$$
\sum_{i\ne j}\alpha_{ij}\mu_1^2 + \sum_{i}\alpha_{ii}\mu_2 = 0
$$
and it follows that
$$
\sum_{i\ne j}\alpha_{ij}
=
\sum_i \alpha_{ii}
=0
.
$$
Now consider the symmetrization
\begin{align*}
\sum_{\sigma\in\SG_n}P(X_{\sigma(1)},X_{\sigma(2)},\dots,X_{\sigma(n)})
&= \sum_{i,j}\sum_\sigma X_{\sigma(i)}X_{\sigma(j)} \\
&= \sum_{i\ne j}\alpha_{ij}(n-2)!\sum_{k\ne l} X_kX_l
+ \sum_i \alpha_{ii} (n-1)!\sum_k X_k^2
\\
&=0
.
\end{align*}
\end{proof}
\begin{Rem}
\begin{enumerate}
\item
Our typical example of a centered linear statistic will be
${\mathnormal X}_{i}-\overline{{\mathnormal X}}$.
\item
The example $P=X_1X_2X_1-X_1^2X_2$ shows that in the present formulation
the lemma cannot be extended beyond degree $2$.
\end{enumerate}
\end{Rem}
In the following a polynomial $P(X_1,X_2,\dots,X_n)$ in noncommuting variables
is called \emph{symmetric} if it is invariant under permutations, i.e.,
$P(X_{\sigma(1)},X_{\sigma(2)},\dots,X_{\sigma(n)})=P(X_1,X_2,\dots,X_n)$ for any
permutation $\sigma\in\SG_n$. For a linear form $L=\sum_{i=1}^n \alpha_i X_i$
we denote the permuted form by $L_\sigma=\sum_{i=1}^n \alpha_iX_{\sigma(i)}$.
\begin{lemm}
\label{lemm:sumyobcietelematgeneral}
Let ${\mathnormal X}_1, {\mathnormal X}_2,\dots, {\mathnormal X}_n \in \mathcal{A}_{sa}$ be free identically distributed random variables, $L=\sum_{i=1}^n \alpha_i {\mathnormal X}_i$ a linear form such that $\tau(L)=0$
and $P_j=P_j(X_1,X_2,\dots,X_n)$ symmetric polynomials for $j\in\{1,2,\dots,r\}\setminus\{k\}$.
Then
$$K^r_{B_1,\dots,B_m}(P_1,\dots,P_{k-1},L,P_{k+1},\dots,P_r)=0.$$
\end{lemm}
\begin{proof}
Let us first observe that for $i\neq j$ we have
$$K^r_{B_1,\dots,B_m}(P_1,\dots,P_{k-1},{\mathnormal X}_{i},P_{k+1},\dots,P_r)=K^r_{B_1,\dots,B_m}(P_1,\dots,P_{k-1},{\mathnormal X}_{j},P_{k+1},\dots,P_r),$$
for all $i,j\in[n],$ and $r\geq 1$. This follows from the argument that ${\mathnormal X}_i$ are free i.i.d. and $P_j$ are symmetric polynomials in the $n$ variables ${\mathnormal X}_1,\dots,{\mathnormal X}_n.$
It follows by multilinearity that
$$K^r_{B_1,\dots,B_m}(P_1,\dots,P_{k-1},L,P_{k+1},\dots,P_r)=K^r_{B_1,\dots,B_m}(P_1,\dots,P_{k-1},L_\sigma,P_{k+1},\dots,P_r),$$ for every permutation $\sigma\in\SG_n$
and taking the average, we have
$$
K^r_{B_1,\dots,B_m}(P_1,\dots,P_{k-1},L,P_{k+1},\dots,P_r)=
\frac{1}{n!}\sum_{\sigma\in\SG_n}
K^r_{B_1,\dots,B_m}(P_1,\dots,P_{k-1},L_\sigma,P_{k+1},\dots,P_r)
=0,
$$
again by multilinearity and taking into account Lemma~\ref{lem:centeredstatistic}.
\end{proof}
\begin{cor}
\label{lemm:sumyobcietelemat}
Let ${\mathnormal X}_1, {\mathnormal X}_2,\dots, {\mathnormal X}_n \in \mathcal{A}_{sa}$ be free identically distributed random variables then
$$K^r_{B_1,\dots,B_m}({\mathnormal Q}_n,\dots,{\mathnormal Q}_n,{\mathnormal X}_{i}-\overline{{\mathnormal X}},{\mathnormal Q}_n,\dots,{\mathnormal Q}_n)=0.$$
\end{cor}
\section{Proof of the main theorem}
Continuing Lemma~\ref{lem:centeredstatistic} we
establish a curious cancellation result for symmetrized squares of centered
linear statistics. A similar phenomenon was observed by
Nica and Speicher \cite[Theorem 1.2]{NicaSpeicher:1998} in the case
of the free commutator. We postpone the investigation of a possible common
pattern between these phenomena to future work.
\begin{lemm}
\label{lem:oddcancellation}
Let ${\mathnormal X}_1, {\mathnormal X}_2,\dots, {\mathnormal X}_n$ be free identically distributed copies
of a random variable $X$
and
$L=\sum_{i=1}^n\alpha_i X_i$ a linear form with $\tau(L)=0$.
Then the distribution of the quadratic statistic
$$
P=\sum_{\sigma\in \SG_n} L_\sigma^2,
$$
does not depend on the odd cumulants of ${\mathnormal X}$.
\end{lemm}
\begin{proof
We show by induction that the cumulants of $P$ can be expressed in
terms of the even cumulants of $X$.
First we apply the product formula of Theorem~\ref{thm:krawczyk} and obtain
\begin{align*}
K_r(P) &= \sum_{\sigma_1,\dots,\sigma_r\in\SG_n}
K_r(L_{\sigma_1}^2,L_{\sigma_2}^2,\dots,L_{\sigma_r}^2)\\
&= \sum_{\sigma_1,\dots,\sigma_r\in\SG_n}
\sum_{\substack{\pi\in \NC(2r)\\ \pi\vee \onetwo{r}=\hat{1}_{2r}}}
K_\pi(L_{\sigma_1},L_{\sigma_1},
L_{\sigma_2},L_{\sigma_2},
\dots,
L_{\sigma_r},L_{\sigma_r})\\
&= \sum_{\substack{\pi\in \NC(2r)\\ \pi\vee \onetwo{r}=\hat{1}_{2r}}}
\tilde{K}_\pi(L)
\end{align*}
where for $\pi\in\NC(2r)$ we write
\begin{equation}
\label{eq:Ktilde}
\tilde{K}_\pi(L)
=\sum_{\sigma_1,\sigma_2,\dots,\sigma_r\in\SG_n}
K_\pi(L_{\sigma_1},L_{\sigma_1},
L_{\sigma_2},L_{\sigma_2},
\dots,
L_{\sigma_r},L_{\sigma_r}).
\end{equation}
We claim that in this decomposition
the contributions of non-even partitions cancel each other.
To see this, we proceed by induction and use
Lemma~\ref{lemm:sumyobcietelematgeneral}.
Let $B_1,B_2,\dots,B_m$ be an enumeration of all odd subsets of $[2r]$,
then we can split off the sum \eqref{eq:Ktilde} the even part
and decompose the rest in a ``telescope'' fashion
as
\begin{equation}
\label{eq:telescope}
K_r(P) =
\sum_{\substack{\pi\in \NCeven(2r)\\ \pi\vee\onetwo{r}=\hat{1}_{2r}}}
\tilde{K}_\pi(L)
+\sum_{k=1}^m
\sum_{\substack{\pi\in \NC^{B_k}_{B_1,B_2,\dots,B_{k-1}}(2r)\\ \pi\vee\onetwo{r}=\hat{1}_{2r}}}
\tilde{K}_\pi(L)
.
\end{equation}
The last formula is obtained using
the following decomposition
$$\NC(2r)\setminus \NCeven(2r)=NC^{B_1}(2r) \cup NC^{B_2}_{B_1}(2r)\cup NC^{B_3}_{B_1,B_2}(2r) \cup\dots \cup NC^{B_m}_{B_1,\dots,B_{m-1}}(2r).$$
Directly from the definition we have $NC^{B_i}_{B_1,\dots,B_{i-1}}(2r)\cap
NC^{B_j}_{B_1,\dots,B_{j-1}}(2r)=\emptyset,$ for $i\neq j$. We will show the
inclusion
$$\NC(2r)\setminus \NCeven(2r)\subseteq NC^{B_1}(2r) \cup NC^{B_2}_{B_1}(2r)\cup NC^{B_3}_{B_1,B_2}(2r) \cup\dots \cup NC^{B_m}_{B_1,\dots,B_{m-1}}(2r)$$
only because the converse inclusion is obvious.
Given $\pi\in \NC(2r)\setminus \NCeven(2r)$, let $k$ be the smallest index
such that $B_k\in\pi$ then $\pi\in NC^{B_{k}}_{B_1,\dots,B_{k-1}}(2r)$ for the blocks
$B_1,B_2, \dots,B_{k-1}$ do not appear in $\pi$.
It remains to show that each individual sum
\begin{equation}
\label{eq:sumNCBk}
\sum_{\substack{\pi\in \NC^{B_k}_{B_1,B_2,\dots,B_{k-1}}(2r)\\ \pi\vee\onetwo{r}=\hat{1}_{2r}}}
\tilde{K}_\pi(L)
\end{equation}
vanishes.
Every $\pi$ in this sum contains the odd block $B_k$ and splits
the complement $[2r]\setminus B_k$ into intervals
$I_1,I_2,\dots,I_l$, interpreted in a circular manner,
see Fig.~\ref{fig:oddinterval}. Then
at least one of these intervals must be odd.
To simplify the discussion we may assume that either $I_1$
is odd and $2r\in B_k$ or $I_l$ is odd and $1\in B_k$;
this may always be achieved by applying an even rotation, which does
not change the values of the cumulants because of traciality.
We are now in one of the situations depicted in Fig.~\ref{fig:oddinterval}.
\begin{figure}
\input{oddinterval}
\caption{Two types of partitions with an odd block}
\label{fig:oddinterval}
\end{figure}
We concentrate on the first case, i.e., $I_1=\{1,2,\dots,j_1-1\}$
and $B_k=\{j_1,j_2,\dots,j_l\}$ where
$j_1$ is even, and $j_l=2r$.
Observe that every partition
$\pi\in \NC^{B_k}_{B_1,B_2,\dots,B_{k-1}}(2r)$ is the concatenation
of some noncrossing partition $\pi'\in
\NC_{B_1,B_2,\dots,B_{k-1}}(j_1-1)$
and $\pi''\in \NC^{B_k}_{B_1,B_2,\dots,B_{k-1}}(\{j_1,\dots,2r\})$.
Now $\pi\vee \onetwo{r}=\hat{1}_{2r}$ if and only if
$\pi'\vee
\input{oddoneright}
=\hat{1}_{j_1-1}
$
and
$\pi''\vee
\input{oddoneleft}
=\hat{1}_{\{j_1,\dots,2r\}}$.
Thus we may unfold \eqref{eq:Ktilde} and factor the sum \eqref{eq:sumNCBk}
to obtain
\begin{multline*}
\sum_{\substack{\pi\in \NC^{B_k}_{B_1,B_2,\dots,B_{k-1}}(2r)\\ \pi\vee\onetwo{r}=\hat{1}_{2r}}}
\tilde{K}_\pi(L)
= \sum_{\substack{\pi'\in \NC_{B_1,B_2,\dots,B_{k-1}}(j_1-1)
\\ \pi'\vee
\,
{
\setlength{\unitlength}{0.7\unitlength}
\tiny
\input{oddoneright}
}
\,
=\hat{1}_{j_1-1}}}
\tilde{K}_{\pi'}(L)
\sum_{\substack{\pi''\in \NC^{B_k}_{B_1,B_2,\dots,B_{k-1}}(\{j_1,\dots,2r\}) \\
\\ \pi''\vee
\,
{
\setlength{\unitlength}{0.7\unitlength}
\tiny
\input{oddoneleft}
}
\,
=\hat{1}_{\{j_1,\dots,2r\}}}}
\tilde{K}_{\pi''}(L)
\\
\begin{aligned}
& = \sum_{\sigma_1,\sigma_2,\dots,\sigma_r\in\SG_n}\sum_{\substack{\pi'\in \NC_{B_1,B_2,\dots,B_{k-1}}(j_1-1)
\\ \pi'\vee
\,
{
\setlength{\unitlength}{0.7\unitlength}
\tiny
\input{oddoneright}
}
\,
=\hat{1}_{j_1-1}}}
K_{\pi'}(L_{\sigma_1}, L_{\sigma_1},\dots,L_{\sigma_{j_1'-1}},L_{\sigma_{j_1'-1}},L_{\sigma_{j_1'}})
\\
&\phantom{=xxxxxxxxxx}\times
\sum_{\substack{\pi''\in \NC^{B_k}_{B_1,B_2,\dots,B_{k-1}}(\{j_1,\dots,2r\}) \\
\\ \pi''\vee
\,
{
\setlength{\unitlength}{0.7\unitlength}
\tiny
\input{oddoneleft}
}
\,
=\hat{1}_{\{j_1,\dots,2r\}}}}
K_{\pi''}(L_{\sigma_{j_1'}}, L_{\sigma_{j_1'+1}},\dots,L_{\sigma_r},L_{\sigma_r})
\\
&=
\sum_{\sigma_{j_1'},\sigma_{j_1'+1},\dots,\sigma_r\in\SG_n}
K^{j_1'}_{B_1,B_2,\dots,B_{k-1}}(P,P,\dots,P,L_{\sigma_{j_1'}})
\\
&\phantom{=xxxxxxxxxx}\times
\sum_{\substack{\pi''\in \NC^{B_k}_{B_1,B_2,\dots,B_{k-1}}(\{j_1,\dots,2r\}) \\
\\ \pi''\vee
\,
{
\setlength{\unitlength}{0.7\unitlength}
\tiny
\input{oddoneleft}
}
\,
=\hat{1}_{\{j_1,\dots,2r\}}}}
K_{\pi''}(L_{\sigma_{j_1'}}, L_{\sigma_{j_1'+1}},\dots,L_{\sigma_r},L_{\sigma_r})
\end{aligned}
.
\end{multline*}
And by Lemma~\ref{lemm:sumyobcietelematgeneral} the factor
$$
K^{j_1'}_{B_1,B_2,\dots,B_{k-1}}(P,P,\dots,P,L_\sigma),
$$
vanishes for every $\sigma$.
Here we use the notation $j'_{i}=j_{\lceil i/2 \rceil}$
where $\lceil \cdot \rceil$ is the ceiling function
which rounds up to the nearest integer.
\end{proof}
\begin{Rem}
\begin{enumerate}
\item
Note that in the case of sample variance we have to assume identical distribution
of the involved random variables for the cancellation phenomenon to take place;
in the case of the free commutator this requirement is not necessary.
\item
The argument put forward in the previous proof
is not valid in classical probability except in the case where
$B_k$ is an interval block.
For example if $r=3$ and $B=\{1,3,6\}$
then the block $B$ alone ensures that $\pi\vee \input{sp112233}
= \input{sp111111}$
and thus
$$
\sum_{
\substack{
\pi\in \SP^{B}(6)\\
\pi\vee
\input{sp112233}
= \input{sp111111}
}}
K_\pi(T_1,T_1,T_2,T_2,T_3,T_3)
=K_3(T_1,T_2,T_3)\,\tau( T_1T_2T_3),
$$
where the sum runs over all set partitions.
\end{enumerate}
\end{Rem}
\begin{proof}[Proof of Theorem~\ref{twr:OddElement}]
Let ${\mathnormal X}_1,\dots,{\mathnormal X}_n$ be free copies of a fixed random variable $X$.
We apply Lemma~\ref{lem:oddcancellation} for $L=X_1-\overline{X}$ and
$P=\sum_{\sigma\in\SG_n} L_\sigma^2=(n-1)!\,Q_n$
to conclude from \eqref{eq:telescope} that
$$
K_r(Q_n) =
\left(\frac{1}{(n-1)!}\right)^r
\sum_{\substack{\pi\in \NCeven(2r)\\ \pi\vee\onetwo{r}=\hat{1}_{2r}}}
\tilde{K}_\pi(L).
$$
Now the fact that $X_i$ are free identically distributed implies that in the sum
\eqref{eq:Ktilde}
every term either vanishes or is a multiple of $K_\pi(X)$.
Therefore we may write
\begin{equation}
\label{eq:Pom1DowodTwrGlo}
K_r({\mathnormal Q}_n) =
\sum_{\pi\in \NCeven_0(2r)}
c_n(\pi)\,K_\pi(X)
\end{equation}
where $ \NCeven_0(2r) =\{\pi\in\NCeven(2r) \mid \pi \vee \onetwo{r}= \hat{1}_{2r}\}$ and $c_n(\pi)\in {\mathbb R}$.
We will derive an explicit formula for these coefficients
in Section~\ref{sec:Rcyclic} below.
Using identity \eqref{eq:Pom1DowodTwrGlo} we will show that all even cumulants of higher order vanish,
i.e., $K_{2i}(X)=0$ for $i\geq 2$.
First let us compute the parameters $c_n(\pi)$ in the extreme cases
$\pi=\pispecial{r}$
and
$\pi=\hat{1}_{2r}$.
For $\pispecial{r}=\input{minpartition1}$ only the
second cumulant contributes to $c_n(\pispecial{r})$ and
the value of the latter does not change if we replace
$X$ with a centered semicircular variable of variance $\sigma^2$.
In this case Corollary~\ref{corr:dwupartycje} implies that
$$
K_r(Q_n) = c_n(\input{minpartition1})\,\sigma^{2r},
$$
and from Proposition~\ref{twr:Lehner:2003} we infer that
$c_n(\input{minpartition1})=n-1$.
To compute the value of $c_n(\hat{1}_{2r})$
it is convenient to switch to tensor notation and to identify the multilinear
cumulant functional $K_r:\mathcal{A}^n\to {\mathbb C}$ with
its linear extension $K_r:\mathcal{A}^{\otimes n}\to {\mathbb C}$.
Let us now assume without loss of generality that $\sigma=1$.
We have to evaluate
\begin{align*}
\tilde{K}_{\hat{1}_{2r}}(L)
&= \sum_{i_1,i_2,\dots,i_r=1}^n K_{2r}(X_{i_1}-\overline{X},X_{i_1}-\overline{X},X_{i_2}-\overline{X},X_{i_2}-\overline{X},\dots,X_{i_r}-\overline{X},X_{i_r}-\overline{X})
\\
&= K_{2r}
\biggl(
\biggl(
\sum_{i=1}^n (X_i-\overline{X})\otimes(X_i-\overline{X})
\biggr)^{\otimes r}
\biggr)
\\
&= K_{2r}
\biggl(
\biggl(
\sum_{i=1}^n X_i\otimes X_i-n\overline{X}\otimes\overline{X}
\biggr)^{\otimes r}
\biggr).
\\
\end{align*}
Expanding this power yields cumulants of the form
\begin{align*}
K_{2r}\Bigl(
\Bigl(
\sum_{i=1}^n X_i\otimes X_i
\Bigr)^{\otimes k}
\otimes
\Bigl(
-n\overline{X}\otimes\overline{X}
\Bigr)^{\otimes(r-k)}
\Bigr)
&= n
K_{2r}\Bigl(
\Bigl(
X_1\otimes X_1
\Bigr)^{\otimes k}
\otimes
\Bigl(
-\frac{1}{n} X_1\otimes X_1
\Bigr)^{\otimes(r-k)}
\Bigr)
\\
&=
n
\left(
-\frac{1}{n}
\right)^{r-k}
K_{2r}(X),
\end{align*}
and in total
$$
\tilde{K}_{\hat{1}_{2r}}(L)
= n\sum_{k=0}^r
\binom{r}{k}
\left(
-\frac{1}{n}
\right)^{r-k}
K_{2r}(X)
=
n
\left(
1-\frac{1}{n}
\right)^r
K_{2r}(X) .
$$
Next, to evaluate even cumulants, equate the $r$-th cumulants of ${\mathnormal Q}_n$ and $\chi^2(n-1)$, i.e.,
$$K_r({\mathnormal Q}_n)=K_r(\chi^2(n-1))=n-1.$$
Denote $\NCeven_0'(2r)=\NCeven_0(2r) \setminus \{
\input{minpartition1}
,
\hat{1}_{2r}
\}$,
then we have
\begin{equation}
\begin{aligned}
n-1 &=\sum_{\pi\in \NCeven_0(2r)}c_n(\pi) \,K_\pi(X)
\\ &= \sum_{\pi\in \NCeven_0'(2r)}c_n(\pi)\,K_\pi(X)
+\sum_{ {\substack{\pi\in \NC_{2}(2r)\\ \pi\vee\onetwo{r}=\hat{1}_{2r}}}}c_n(\pi)
\,K_\pi(X) +
c_n(\hat{1}_{2r})\,K_{2r}(X)
\\ &=
\sum_{\pi\in \NCeven_0'(2r)}c_n(\pi)\,K_\pi(X)
+ c_n(\input{minpartition1})
+\frac{(n-1)^{r}}{n^{r-1}}\,K_{2r}(X)
\\ &=
\sum_{\pi\in \NCeven_0'(2r)}c_n(\pi)\,K_\pi(X)
+n-1
+\frac{(n-1)^{r}}{n^{r-1}}\,K_{2r}(X).
\end{aligned}
\end{equation}
This yields
\begin{equation}
\label{eq:PomDowod1}
\sum_{\pi\in \NCeven_0'(2r)}c_n(\pi)\,K_\pi(X)+\frac{(n-1)^{r}}{n^{r-1}}\,K_{2r}(X)=0.
\end{equation}
and the blocks of any $\pi\in\NCeven_0'(2r)$ have size strictly smaller
than $2r$, it follows by induction that
$K_{2r}(X)=0$ for $r\geq 2$.
Conversely,
suppose that ${\mathnormal X}_i$'s are odd, then from Lemma~\ref{lem:oddcancellation}
we get
\begin{align*}
K_r({\mathnormal Q}_n)
&=\sum_{\pi\in \NCeven_0(2r)}
c_n(\pi)\,K_\pi(X)
=\sum_{\substack{\pi\in \NC_{2}(2r)\\ \pi\vee\onetwo{r}=\hat{1}_{2r}}}c_n(\pi)\,K_\pi(X)
\\
&= c_{n}(
\input{minpartition1}
) =n-1.
\end{align*}
\end{proof}
\noindent \textit{Proof of Proposition \ref{twr:kruglow}}.
Recall that as a consequence of the free L\'evy-Khinchin formula
(see for example \cite[Theorem~13.16]{NicaSpeicher:2006})
the random variable ${\mathnormal X}_i$ is freely infinitely divisible if and only if
$$K_{n+2}({\mathnormal X}_i)=\int_{{\mathbb R}}x^n d\rho_i(x),$$
for some positive finite measure $\rho_i(x)$ on ${\mathbb R}$.
For the semicircular distribution the measure is $\rho_i=\delta_0$ and it
suffices to show that $\int x^2d\rho_i(x)=K_4(X_i)=0$.
Now if $\tau(Y_1)=\tau(Y_2)=\tau(Y_3)=\tau(Y_4)=0$
then the product formula from Theorem~\ref{thm:krawczyk} implies
\begin{equation}
\label{eq:krawczykQ}
K_2(Y_1Y_2,Y_3Y_4)=K_2(Y_1,Y_4)\,K_2(Y_2,Y_3)+K_4(Y_1,Y_2,Y_3,Y_4)
.
\end{equation}
We will apply this to $Q_n$, so let us first compute the cumulants
which will appear after evaluation of \eqref{eq:krawczykQ}.
By assumption $K_2(X_i)=1$ for all $i$ and therefore the covariances are
$$
K_2({\mathnormal X}_{i}-\overline{{\mathnormal X}},{\mathnormal X}_{j}-\overline{{\mathnormal X}})=
\begin{cases}
\frac{n-1}{n} &\text{if $i=j$,} \\
-\frac{1}{n} &\text{if $i\neq j$.}
\end{cases}
$$
It remains to consider cumulants of order 4.
First,
\begin{align*}
\sum_{i=1}^n K_4(X_i-\overline{X})
&= \sum_{i=1}^n
\biggl(
\left(
1-\frac{1}{n}
\right)^4
K_4(X_i)
+
\sum_{l\ne i}
\left(
-\frac{1}{n}
\right)^4
K_4(X_l)
\biggr)
\\
&=
\left(
\left(
1-\frac{1}{n}
\right)^4
+
\frac{n-1}{n^4}
\right)
\sum_{i=1}^n K_4(X_i),
\end{align*}
second,
\begin{multline*}
\sum_{\substack{i,j=1\\ i\ne j}}^n
K_4(X_i-\overline{X},X_i-\overline{X},X_j-\overline{X},X_j-\overline{X})
\\
\begin{aligned}
&=
\sum_{\substack{i,j=1\\ i\ne j}}^n
\biggl(
\left(
1-\frac{1}{n}
\right)^2
\left(
-\frac{1}{n}
\right)^2
(
K_4(X_i)
+
K_4(X_j)
)
+
\sum_{l\ne i,j}
\left(
-\frac{1}{n}
\right)^4
K_4(X_l)
\biggr)
\\
&=
\left(
2(n-1)
\left(
1-\frac{1}{n}
\right)^2
\left(
\frac{1}{n}
\right)^2
+
\frac{(n-1)(n-2)}{n^4}
\right)
\sum_{i=1}^n
K_4(X_i),
\end{aligned}
\end{multline*}
and thus
$$
\sum_{\substack{i,j=1}}^n
K_4(X_i-\overline{X},X_i-\overline{X},X_j-\overline{X},X_j-\overline{X})
=
\frac{(n-1)^2}{n^2}
\sum_{i=1}^n
K_4(X_i).
$$
Using these formulas
we now proceed to \eqref{eq:krawczykQ}
and obtain
\begin{align*}
n-1
&= K_2({\mathnormal Q}_n,{\mathnormal Q}_n)
\\
&=\sum_{i,j=1}^n K_2(({\mathnormal X}_i-\overline{{\mathnormal X}})^2,({\mathnormal X}_j-\overline{{\mathnormal X}})^2)
\\
&=\sum_{i,j=1}^n K_2({\mathnormal X}_i-\overline{{\mathnormal X}},{\mathnormal X}_j-\overline{{\mathnormal X}})\,
K_2({\mathnormal X}_j-\overline{{\mathnormal X}},{\mathnormal X}_i-\overline{{\mathnormal X}})
\\ &\phantom{===}
+
\sum_{i,j=1}^n K_4({\mathnormal X}_i-\overline{{\mathnormal X}},{\mathnormal X}_i-\overline{{\mathnormal X}},
{\mathnormal X}_j-\overline{{\mathnormal X}},{\mathnormal X}_j-\overline{{\mathnormal X}})
\\
&=\sum_{i=1}^n [K_2(({\mathnormal X}_i-\overline{{\mathnormal X}}),({\mathnormal X}_i-\overline{{\mathnormal X}}))]^2+\sum_{i=1,j=1, i\neq j}^n[K_2({\mathnormal X}_i-\overline{{\mathnormal X}},{\mathnormal X}_j-\overline{{\mathnormal X}})]^2
\\
&\phantom=+\sum_{i=1}^n K_4({\mathnormal X}_i-\overline{{\mathnormal X}})+\sum_{i=1,j=1, i\neq j}^nK_4({\mathnormal X}_i-\overline{{\mathnormal X}},{\mathnormal X}_i-\overline{{\mathnormal X}},{\mathnormal X}_j-\overline{{\mathnormal X}},{\mathnormal X}_j-\overline{{\mathnormal X}}) \\
&= \frac{(n-1)^2}{n}+\frac{n-1}{n}+\frac{(n-1)^2}{n^2}\sum_{i=1}^nK_4({\mathnormal X}_i).
\end{align*}
So we see that $\sum_{i=1}^nK_4({\mathnormal X}_i)=\sum_{i=1}^n\int_{{\mathbb R}}x^2 d\rho_i(x)=0$ and thus $\rho_i(x)=\delta_0(x)$.
Note that the above measure is the free L\'evy measure of the semicircle distribution with mean zero, and variance
one. \begin{flushright} $\square$ \end{flushright}
\begin{proof}[Proof of Proposition \ref{twr:OdpowidenikKaganLetac}]
Part (1).
We write $a=\norm{a}\theta=\norm{a}(\theta_1,\dots,\theta_n)$ where $\theta$ belongs to the unit sphere
of $E$, i.e., $\sum\theta_i=0$.
Thus for $\norm{a}>0,$ $a_i=\norm{a}\theta_i$ and $r\geq 2$ we have
\begin{equation}
\label{eq:kumualntydonieskonczonosci}
\begin{aligned}
\frac{K_r(\sum_{i=1}^n({\mathnormal X}_i-\overline{{\mathnormal X}}+a_i)^2)}{\norm{a}^r}
&=K_r\Big(\sum_{i=1}^n({\mathnormal X}_i-\overline{{\mathnormal X}})^2/\norm{a}
-\sum_{i=1}^n2({\mathnormal X}_i-\overline{{\mathnormal X}})\,\theta_i+\norm{a}\Big)\\
&=K_r\Big(\sum_{i=1}^n({\mathnormal X}_i-\overline{{\mathnormal X}})^2/\norm{a}-\sum_{i=1}^n2({\mathnormal X}_i-\overline{{\mathnormal X}})\,\theta_i\Big)
\\
&=K_r\Big(\sum_{i=1}^n({\mathnormal X}_i-\overline{{\mathnormal X}})^2/\norm{a}-\sum_{i=1}^n2{\mathnormal X}_i\theta_i\Big).
\end{aligned}
\end{equation}
By the hypothesis the left hand side of \eqref{eq:kumualntydonieskonczonosci} does not depend on $\theta$, and thus the limit on the right hand side
\begin{align*}
\lim_{\norm{a}\to + \infty}\frac{K_r(\sum_{i=1}^n({\mathnormal X}_i-\overline{{\mathnormal X}}+a_i)^2)}{\norm{a}^r}&=K_r\Big(-\sum_{i=1}^n2{\mathnormal X}_i\theta_i\Big),
\end{align*}
does not depend on $\theta$ either.
Now freeness implies that
\begin{align}
S_r(\theta_1,\dots,\theta_n):=K_r\Big(\sum_{i=1}^n{\mathnormal X}_i\theta_i\Big)=\sum_{i=1}^n\theta_i^rK_r({\mathnormal X}_i)=\Big(\sum_{i=1}^n\theta_i^r\Big)K_r,
\end{align}
is a constant function on the unit sphere
of the space $E$. Thus we see $S_2(\theta_1,\dots,\theta_n)=K_2$ and $S_r$ for $r\geq 3$ is constant function on the unit sphere
of the space $E$ if and only $K_r=0$ for $r\geq 3$.
We now show part $(2)$ of Proposition \ref{twr:OdpowidenikKaganLetac}. It's easy to observe that for $\theta\in E$ we have
\begin{align}
K_r\Big(\sum_{i=1}^n({\mathnormal X}_i-\overline{{\mathnormal X}})\,\theta_i\Big)=\sum_{i=1}^n\theta_i^rK_r({\mathnormal X}_i)=\Big(\sum_{i=1}^n\theta_i^r\Big)K_r,
\end{align}
is constant function on the unit sphere
of the space $E$ and from the above consideration we get the statement.
\end{proof}
\begin{Rem}
The preceding proof is also valid in the general setting of
\cite{Lehner:2003}.
\end{Rem}
\section{$R$-cyclic matrices and free infinite divisibility of quadratic forms}
\label{sec:Rcyclic}
In this section we show how the sample variance can be combined with
the concept of $R$-cyclicity and give a recipe for computing the coefficients
$c_n(\pi)$ in equation \eqref{eq:Pom1DowodTwrGlo}. In particular we also show
that sample variance preserves free infinite divisibility.
\subsection{$R$-cyclic matrices and the distribution of sample variance}
The concept of $R$-cyclicity was introduced by Nica, Shlyakhtenko and Speicher
\cite{NicaSpeicherShlyakhtenko:2002}. Our aim is now to exhibit its relation
to the sample variance and other quadratic forms.
We show that the theory of $R$-cyclic matrices
can be used to compute the distribution
of the sample variance and give a formula for the cumulants of the sample
variance in terms of the even cumulants, which generalizes and unifies
two types of results, namely cumulants of squares of even elements
\cite[Proposition~11.25]{NicaSpeicher:2006} and cumulants of quadratic
forms in gaussian random variables \cite[Proposition~4.4]{Lehner:2003}.
Here we consider matrices over a non-commutative probability space.
Let $(\mathcal{A},\tau)$ be a non-commutative probability
space, and let $n$ be a positive integer.
The algebra $M_n(\mathcal{A})$
of $n\times n$ matrices over $\mathcal{A}$ is a noncommutative probability space
with canonical expectation functional
$$\tau_n(A)=\frac{1}{n}\sum_{i=1}^n\tau(a_{i,i}),$$
where $A=[a_{i,j}]_{i,j=1 }^n$ is a matrix in $M_n(\mathcal{A})$. Then $(M_n(\mathcal{A}),\tau_n)$ is itself a non-commutative probability space.
The definition of $R$-cyclicity is in terms of
the joint $R$-transform of the entries of the matrix: one requires that
only the cyclic non-crossing cumulants of the entries are allowed to be
different from $0$, see Definition~\ref{def:Rcyclic} below.
Equivalently, it was shown in \cite[Theorem~8.2]{NicaSpeicherShlyakhtenko:2002}
that $R$-cyclicity is equivalent to the fact that $A$ is free from $M_n({\mathbb C})$
with amalgamation over the algebra $\mathcal{D}_n$ of scalar diagonal matrices
with respect to the conditional expectation
\begin{equation}
\label{eq:ED}
\begin{aligned}
\ED:M_n(\mathcal{A})&\to M_n({\mathbb C})\\
A &\mapsto \sum_{i=1}^n E_i\tau^{(n)}(A)E_i,
\end{aligned}
\end{equation}
where by $E_i$ we denote the projection matrix onto the $i$-th unit vector
and $\tau^{(n)}(A)_{ij}=\tau(a_{i,j})$ is the entry-wise trace.
\begin{defi}\label{def:Rcyclic}
Let $(M_n(\mathcal{A}),\tau_n)$ and $n$ be as above,
then a matrix ${\boldsymbol{X}}=[{\mathnormal X}_{i,j}]_{i,j=1 }^n\in M_n(\mathcal{A})$.
is said to be \emph{$R$-cyclic} if
for every $r\geq 1$ and for every choice of indices
$1\leq i_1,j_1,\dots,i_r,j_r\leq n $ the cumulant
$$
K_r({\mathnormal X}_{i_1,j_1},{\mathnormal X}_{i_2,j_2},\dots,{\mathnormal X}_{i_r,j_r})=0,$$
unless the indices are cyclic in the sense
that $ j_1=i_2,j_2=i_3,\dots,j_{r-1}=i_r,j_r=i_1$.
In this case the formal noncommutative power series
\begin{equation}
\label{eq:determiningseries}
f_{{\boldsymbol{X}}}(z_1,z_2,\dots,z_n)
=\sum_{r=1}^\infty\sum_{i_1,i_2,\dots,i_r=1}^n
K_r({\mathnormal X}_{i_r,i_1},{\mathnormal X}_{i_1,i_2},\dots,{\mathnormal X}_{i_{r-1},i_r})
\,
z_{i_1}z_{i_2}\cdots z_{i_r},
\end{equation}
is called the \emph{determining series} of the entries of ${\boldsymbol{X}}$.
\end{defi}
The concept of $R$-cyclicity generalizes the concept of $R$-diagonality
\cite[Ex.~20.5]{NicaSpeicher:2006} in the sense that $X$ is $R$-diagonal
if and only if the matrix
$
\left[
\begin{smallmatrix}
0&X\\
X^*&0
\end{smallmatrix}
\right]
$
is $R$-cyclic.
\begin{lemm}
\label{lem:kreweras}
For scalar matrices $A\in M_n({\mathbb C})$ we have
\begin{enumerate}[(i)]
\item \label{it:kreweras1}
$$
\sum_{i=1}^n E_iA_1E_iA_2\dotsm E_iA_rE_i
= \ED(A_1)\ED(A_2)\dotsm \ED(A_r).
$$
\item \label{it:kreweras2}
Let $\pi\in\NC(r)$, then
$$
\sum_{\ker\underline{i}\geq\pi} \ED (A_1E_{i_1}A_2E_{i_2}\dotsm A_rE_{i_r}A_{r+1})
=\ED[\Krewlr\pi] (A_1,A_2,\dots,A_{r+1}).
$$
\end{enumerate}
\end{lemm}
\begin{proof}
Part~\eqref{it:kreweras1} follows immediately from the expansion
$$
\sum_{k=1}^n E_kA_1E_kA_2\dotsm A_rE_k
= \sum_{k=1}^n E_{k} a^{(1)}_{k,k}a^{(2)}_{k,k}\dotsm a^{(r)}_{k,k}.
$$
To see part~\eqref{it:kreweras2} we single out the last block
of $\pi$ (i.e., the block containing $r$, see Lemma~\ref{lemm:kreweras}),
say
$B=\{j_1<j_2<\dotsm <j_p=r\}$,
and group the remaining blocks into
subpartitions, empty partitions allowed,
say $\pi_1\in\NC([1,j_1-1])$,
$\pi_2\in\NC([j_1+1,j_2-1]),\dots,\pi_p\in\NC([j_{p-1}+1,j_p-1])$.
Then we have
$$
\ED\biggl(
\sum_{\ker\underline{i}\geq\pi} (A_1E_{i_1}A_2E_{i_2}\dotsm E_{i_r} A_{r+1})
\biggr)
=
\ED\biggl(
\sum_i
A_1'E_iA_2'E_i\dots A_p'E_i A_{r+1}
\biggr),
$$
where
$$
A_k'= \sum_{\ker\underline{i}\geq \pi_k}
A_{j_{k-1}+1}E_{i_1}
A_{j_{k-1}+2}E_{i_2}
\dotsm
A_{j_{k}}.
$$
By part~\eqref{it:kreweras1} this is
$$
\ED(A_1'\ED(A_2') \dots\ED(A_p') A_{r+1}),
$$
and by induction this is
$$
\ED(A_1' \ED[\Krewlr{\pi_2}](A_{j_1+1},\dots,A_{j_2})
\dotsm
\ED[\Krewlr{\pi_p}](A_{j_{p-1}+1},\dots,A_r) A_{r+1})
= \ED[\Krewlr\pi](A_1,A_2,\dots,A_{r+1}),
$$
where we used Lemma~\ref{lemm:kreweras}.
\end{proof}
\begin{prop}
\label{prop:Rcyclic}
Let ${\mathnormal X}_1, {\mathnormal X}_2,\dots, {\mathnormal X}_n\in \mathcal{A}$ be a free family of even
random variables and $A=[a_{i,j}]_{i,j=1}^n\in M_n({\mathbb C})$ a scalar matrix.
Then
the Hadamard product matrix
\begin{equation}
\label{eq:ZAX}
\mathbf{Z}=A\circ{\boldsymbol{X}}= [a_{i,j}X_iX_j]_{i,j=1 }^n=
\begin{bmatrix}
a_{1,1}{\mathnormal X}_1^2 & a_{1,2}{\mathnormal X}_1{\mathnormal X}_2 & \dots & a_{1,n}{\mathnormal X}_1{\mathnormal X}_n \\
a_{2,1}{\mathnormal X}_2{\mathnormal X}_1 & a_{2,2}{\mathnormal X}_2^2 & \dots & a_{2,n}{\mathnormal X}_2{\mathnormal X}_n
\\
\multicolumn{4}{c}{\dotfill}\\
a_{n,1}{\mathnormal X}_n{\mathnormal X}_1 & a_{n,2}{\mathnormal X}_n{\mathnormal X}_2 & \dots & a_{n,n}{\mathnormal X}_n^2
\end{bmatrix},
\end{equation}
is $R$-cyclic.
\end{prop}
\begin{proof}
We make use of the product formula of Theorem~\ref{thm:krawczyk}
and Lemma~\ref{lemm:lematoparzystych} to compute
\begin{equation}
\label{eq:KrHx1Xh2}
\begin{aligned}
K_r(X_{i_1}X_{i_2},
X_{i_3}X_{i_4},
\dots,
X_{i_{2r-1}}X_{i_{2r}})
&= \sum_{\substack{\pi\in\NCeven(2r)\\
\pi\vee \onetwo{r} = \hat{1}_{2r}}}
K_\pi(X_{i_{1}},X_{i_{2}},\dots,X_{i_{2r}})
\\
&= \sum_{\substack{\pi\in\NCeven(2r)\\
\pi\geq \pispecial{r}}}
K_\pi(X_{i_{1}},X_{i_{2}},\dots,X_{i_{2r}})
\end{aligned}
\end{equation}
and by \eqref{eq:kerh>=pi}
these mixed cumulants vanish unless $\ker \underline{i}\geq\pispecial{r}$,
i.e., unless $i_{1}=i_{2r}$ and $i_{2j}=i_{2j+1}$ for all $j$,
which exactly means $R$-cyclicity. It is easy to see that the same holds
for $Z_{i,j}=a_{i,j}X_iX_j$.
\end{proof}
\begin{Rem}
In some sense Proposition~\ref{prop:Rcyclic}
is a generalization of the fact
\cite[Theorem~20.6]{NicaSpeicher:2006}
that the product of two free even selfadjoint elements is $R$-diagonal.
This fact is indeed a consequence if we put
$A=
\left[
\begin{smallmatrix}
0&1\\
1&0
\end{smallmatrix}
\right]
$
in the preceding proposition.
In fact it was shown in \cite{HaagerupLarsen:1998} that
every $R$-diagonal element can be written as a product
of two free even selfadjoint elements.
It is an interesting question what would be a natural factorization
of $R$-cyclic matrices. While it is necessary for a
matrix to be $R$-cyclic that its entries form $R$-diagonal pairs,
example \cite[Ex.~20.6]{NicaSpeicher:2006} shows that
the representation \eqref{eq:ZAX} in the preceding
proposition does not cover all $R$-cyclic matrices.
\end{Rem}
\begin{prop} \label{prop:CykliczneVariancja}
Let ${\mathnormal X}_1, {\mathnormal X}_2,\dots, {\mathnormal X}_n\in \mathcal{A}$ be a free family of even random variables, ${\boldsymbol{X}}= [X_iX_j]_{i,j=1 }^n$ and $A=[a_{i,j}]_{i,j=1}^n\in M_n({\mathbb C})$ a scalar matrix.
\begin{enumerate}[(i)]
\item \label{it:cyclic2}
The determining series of the entries of the $R$-cyclic matrix
${\boldsymbol{Z}}=A\circ{\boldsymbol{X}}$ and the $R$-transform of the
quadratic form $T_n=\sum_{i,j}^na_{i,j}{\mathnormal X}_i{\mathnormal X}_j$ are related by
\begin{equation}
\label{eq:cyclic2}
f_{A\circ{\boldsymbol{X}}}(z,\dots,z)=\mathcal{R}_{T_n}(z),
\end{equation}
where $\mathcal{R}_{T_n}(z)=zR_{T_n}(z)$.
\item \label{it:cyclic3}
The cumulants of $T_n$ are given by
\begin{align} \label{eq:kumulantsamplevariancenotiid}
\nonumber &K_r(T_n)\\&=\sum_{i_1,\dots,i_r\in[n]}
\;\mathrm{Tr}\,(AE_{i_1}AE_{i_2}\dots AE_{i_r})\,
\sum_{\substack{ \pi\in \NCeven(2r)\\
\pi \vee \onetwo{r}=\hat{1}_{2r}}}
K_\pi(X_{i_r},X_{i_1},X_{i_1},X_{i_2},\dots,X_{i_{r-1}},X_{i_r}).
\end{align}
\item If we assume in addition that $X_i$ are identically distributed
the previous formula simplifies to the following convolution-like expression
\begin{equation} \label{eq:kumulantsamplevariance}
K_r(T_n)=\sum_{ \pi\in \NC(r)}
\;\mathrm{Tr}\,(\ED[\Krewl{\pi}](A)) \prod_{B\in\pi}K_{2|B|}(X).
\end{equation}
\end{enumerate}
\end{prop}
\begin{proof}
From the definition of $T_n$ we see that
\begin{align*}
K_r(T_n)
&= \sum_{i_1,i_2,\dots,i_{2r}\in[n]}
K_r(Z_{i_{1},i_{2}},
Z_{i_{3},i_{4}},
\dots,
Z_{i_{2r-1},i_{2r}})
\\
&= \sum_{\substack{
i_1,i_2,\dots,i_{2r} \in[n] \\
\ker \underline{i}\geq\pispecial{r} }}
\sum_{\substack{\pi\in\NCeven(2r)\\ \pi\geq \pispecial{r}}}
a_{i_{1},i_{2}} a_{i_{3},i_{4}} \dotsm a_{i_{2r-1},i_{2r}}
K_\pi(X_{i_{1}},X_{i_{2}},\dots,X_{i_{2r}}),
\intertext{where we used \eqref{eq:KrHx1Xh2}.
Having eliminated the zero contributions
we can apply Lemma~\ref{lemm:lematoparzystych} in the reverse direction and obtain}
&= \sum_{\substack{
i_1,i_2,\dots,i_{2r} \in[n] \\
\ker \underline{i}\geq\pispecial{r} }}
K_r(Z_{i_{1},i_{2}},
Z_{i_{3},i_{4}},
\dots,
Z_{i_{2r-1},i_{2r}})
\\
&= \sum_{i_1,i_2,\dots,i_r}
K_r(Z_{i_r,i_1},Z_{i_1,i_2},\dots,Z_{i_{r-1},i_r}),
\intertext{
which after comparison with \eqref{eq:determiningseries} yields
\eqref{eq:cyclic2}.
We now expand further and obtain}
&= \sum_{i_1,i_2,\dots,i_r\in[n]}
a_{i_r,i_1} a_{i_1,i_2} \dotsm a_{i_{r-1,i_r}}
K_r(X_{i_r}X_{i_1},X_{i_1}X_{i_2},\dots,X_{i_{r-1}}X_{i_r})
\\
&=\sum_{i_1,\dots,i_r\in[n]}
\;\mathrm{Tr}\,(AE_{i_1}AE_{i_2}\dots AE_{i_r})\,
\sum_{\substack{ \pi\in \NCeven(2r)\\
\pi \vee \onetwo{r}=\hat{1}_{2r}}}
K_\pi(X_{i_r},X_{i_1},X_{i_1},X_{i_2},\dots,X_{i_{r-1}},X_{i_r}),
\\
\intertext{which yields \eqref{eq:kumulantsamplevariancenotiid}. Now denoting by $\hat\pi$ the image of $\pi\in\NC(r)$ under the bijection introduced in Lemma~\ref{lemm:lematoparzystych} we can rewrite this as}
&=\sum_{ \pi\in \NC(r)}
\left(\sum_{\ker\underline{i}\geq \pi}\;\mathrm{Tr}\,(AE_{i_1}AE_{i_2}\dots AE_{i_r})\right)
K_{\hat\pi}(X)
.
\end{align*}
Finally we infer \eqref{eq:kumulantsamplevariance} from Lemma~\ref{lem:kreweras}.
\end{proof}
\begin{Rem}
It was observed in \cite[Rem.~4.1]{NicaSpeicherShlyakhtenko:2002} that
$R$-cyclicity is preserved under Hadamard products with constant matrices.
Moreover inspecting the preceding proof one can easily see the that
for an arbitrary $R$-cyclic matrix ${\boldsymbol{X}}=[X_{i,j}]$ and any scalar matrix
$A=[a_{i,j}]$
the determining series of the Hadamard product $A\circ {\boldsymbol{X}}=[a_{i,j}X_{i,j}]$
is given by
$$
f_{A\circ {\boldsymbol{X}}}(z_1,z_2,\dots,z_n)
= \;\mathrm{Tr}\,(f_{\boldsymbol{X}}(AE_1\otimes z_1,
AE_2\otimes z_2,
\dots,
AE_n\otimes z_n))
.
$$
\end{Rem}
In fact we have proved the following slightly more general statement.
\begin{theo}
Let $X_i$ be free copies of an even random variable $X$,
${\boldsymbol{X}}=[X_iX_j]_{i,j=1}^n$ be the matrix of products as above (which is $R$-cyclic)
and let $A_1,A_2,\dots,A_r \in M_n({\mathbb C})$ be
arbitrary scalar matrices. Then $(A_1\circ {\boldsymbol{X}},A_2\circ {\boldsymbol{X}},\dots,A_r\circ {\boldsymbol{X}})$
is an $R$-cyclic family and the joint cumulant of $T_k=\sum_{ij}a^{(k)}_{ij}X_iX_j$ is
$$
K_r(T_1,T_2,\dots,T_r)=\sum_{ \pi\in \NC(r)}
\;\mathrm{Tr}\,(\ED[\Krewl{\pi}](A_1,A_2,\dots,A_r))
\prod_{B\in\pi}K_{2|B|}(X).
$$
\end{theo}
It was shown in \cite[Section~8]{NicaSpeicherShlyakhtenko:2002} that
$R$-cyclicity of a matrix is equivalent to freeness
from the algebra of constant matrices $M_n({\mathbb C})$ with amalgamation
over the commutative subalgebra $\mathcal{D}_n$ of constant diagonal matrices.
Moreover, the cyclic scalar cumulants can be interpreted as entries
of the $\mathcal{D}_n$-valued cumulants as follows.
\begin{prop}[{\cite[Theorem~7.2]{NicaSpeicherShlyakhtenko:2002}}]
\label{prop:NSS72}
Let $({\boldsymbol{X}}_i)\subseteq M_n(\mathcal{A})$ be an $R$-cyclic family over
some noncommutative probability space $(\mathcal{A},\tau)$ and
denote by $K_r^{\mathcal{D}}$ the operator valued cumulant functionals
with respect to the conditional expecation \eqref{eq:ED}.
Then for any $\Lambda_1,\Lambda_2,\dots,\Lambda_{r-1}\in\mathcal{D}_n$
we have
\begin{multline*}
K_r^{\mathcal{D}}({\boldsymbol{X}}_1\Lambda_1,{\boldsymbol{X}}_2\Lambda_2,\dots,{\boldsymbol{X}}_{r-1}\Lambda_{r-1},{\boldsymbol{X}}_r)
\\
= \sum_{i_1,i_2,\dots,i_r=1}^n
\lambda_{i_1}^{(1)} \lambda_{i_2}^{(2)}\dotsm \lambda_{i_{r-1}}^{(r-1)}
K_r(X_{i_r,i_1}^{(1)},X_{i_1,i_2}^{(2)},\dots,X_{i_{r-2},i_{r-1}}^{(r-1)},X_{i_{r-1},i_r}^{(r)})
E_{i_r}.
\end{multline*}
\end{prop}
In our context this leads to an operator valued boxed convolution
in the sense of \cite[Definition~2.1.6]{Speicher:1998} as follows.
\begin{prop}
Let ${\mathnormal X}_1, {\mathnormal X}_2,\dots, {\mathnormal X}_n\in \mathcal{A}$ be free even copies
of a random variable $X$ and
let ${\boldsymbol{X}}=[X_iX_j]_{i,j=1}^n$.
Then for any scalar matrices $A_1,A_2,\dots,A_r\in M_n({\mathbb C})$ and
$\Lambda_1,\Lambda_2,\dots,\Lambda_{r-1}\in\mathcal{D}_n$ we have
\begin{multline*}
K_r^{\mathcal{D}}(A_1\circ{\boldsymbol{X}}\Lambda_1,A_2\circ{\boldsymbol{X}}\Lambda_2,\dots,A_{r-1}\circ{\boldsymbol{X}}\Lambda_{r-1},A_r\circ{\boldsymbol{X}})
\\
= \sum_{\pi\in \NC(r)}
\ED[\Krewl{\pi}](A_1\Lambda_1,A_2\Lambda_2,\dots,A_{r-1}\Lambda_{r-1},A_r)
\prod_{B\in\pi}K_{2|B|}(X).
\end{multline*}
\end{prop}
\begin{proof}
We use Proposition~\ref{prop:NSS72} and expand
\begin{align*}
\MoveEqLeft{
K_r^{\mathcal{D}}(A_1\circ{\boldsymbol{X}}\Lambda_1,A_2\circ{\boldsymbol{X}}\Lambda_2,\dots,A_{r-1}\circ{\boldsymbol{X}}\Lambda_{r-1},A_r\circ{\boldsymbol{X}}) }&
\\
&= \sum_{i_1,i_2,\dots,i_r=1}^n
a_{i_r,i_1}^{(1)}
\lambda_{i_1}^{(1)}
a_{i_1,i_2}^{(2)}
\lambda_{i_2}^{(2)}
\dotsm
a_{i_{r-1},i_r}^{(r)}
K_r(X_{i_r}X_{i_1},X_{i_1}X_{i_2},\dots,X_{i_{r-1}}X_{i_r}) E_{i_r}
\\
&= \sum_{i_1,i_2,\dots,i_r=1}^n
a_{i_r,i_1}^{(1)}
\lambda_{i_1}^{(1)}
a_{i_1,i_2}^{(2)}
\lambda_{i_2}^{(2)}
\dotsm
a_{i_{r-1},i_r}^{(r)}
\sum_{\substack{\pi\in\NC(2r)\\\pi\geq \pispecial{r}}}
K_\pi(X_{i_r},X_{i_1},X_{i_1},X_{i_2},\dots,X_{i_{r-1}},X_{i_r}) E_{i_r}
\\
&= \sum_{\substack{\pi\in\NC(2r)\\\pi\geq \pispecial{r}}}
\sum_{\ker\underline{i}\geq\pi}
a_{i_r,i_1}^{(1)}
\lambda_{i_1}^{(1)}
a_{i_1,i_2}^{(2)}
\lambda_{i_2}^{(2)}
\dotsm
a_{i_{r-1},i_r}^{(r)}
E_{i_r}
K_{\pi}(X)
\\
&= \sum_{\pi\in\NC(r)}
\ED[\Krewl{\pi}](A_1\Lambda_1,A_2\Lambda_2,\dots,A_{r-1}\Lambda_{r-1},A_r)
K_{\hat\pi}(X),
\end{align*}
where $\hat\pi$ is defined in the proof of Proposition \ref{prop:CykliczneVariancja}.
\end{proof}
\begin{Rem}
In fact it is easy to see that the matrix
$\Xi=\diag(X_1,X_2,\dots,X_n)$ is free from $M_n({\mathbb C})$ with
amalgamation over $\mathcal{D}_n$ as well
\cite[Example~2.3]{NicaSpeicherShlyakhtenko:2002}.
We have shown above that $A\circ{\boldsymbol{X}}=\Xi A\Xi$ has the same property
although $A$ has not.
\end{Rem}
As a final corollary we obtain the following formula
for the cumulants of the sample variance.
\begin{cor}
\label{cor:KrQn}
Let ${\mathnormal X}_1, {\mathnormal X}_2,\dots, {\mathnormal X}_n$ be free copies
of a random variable $X$ and $Q_n=nS_n^2$ the rescaled sample variance
defined in \eqref{eq:SampleVariance}.
Let
$\tilde{X}$ be the symmetrization of $X$, i.e., a formal random variable with
even distribution and cumulants $K_{2r}(\tilde X)=K_{2r}(X)$.
Then
\begin{equation}
K_r({\mathnormal Q}_n)= (n-1)K_r(Z^2),
\end{equation}
where
$Z=\sqrt{\frac{n}{n-1}}P\tilde{X}P$ is the free compression of
the symmetrization
$\tilde{X}$ of $X$ by a projection $P$ of trace $\tau(P)=\frac{n-1}{n}$.
\end{cor}
\begin{proof}
By Lemma~\ref{lem:oddcancellation} the distribution of $Q_n$ does not
change if we drop the odd cumulants and replace $X$
by its symmetrization $\tilde{X}$.
The symmetrization $\tilde{X}$ being even, it follows from
Proposition~\ref{prop:CykliczneVariancja} that
the information about the distribution of the sample variance is contained in
the $R$-cyclic matrix
$A\circ \tilde{{\boldsymbol{X}}}=[a_{ij}\tilde{X}_i\tilde{X}_j]_{i,j=1}^n$,
where $A=I-\frac{1}{n}\mathbf{1}$ and $\mathbf{1}$ is the $n\times n$ matrix all of
whose entries are $1$.
This matrix is idempotent with $\ED(A)=(1-1/n)I$ and therefore
$\ED[\pi](A)=(1-\frac{1}{n})^{\abs\pi}I$
for every $\pi\in\NC(r)$. We insert this
into~\eqref{eq:kumulantsamplevariance} and
the cumulants of $Q_n$ evaluate to
\begin{align*} %
K_r(Q_n)
&= n
\sum_{\pi\in\NC(r)}
\left(
1-\frac{1}{n}
\right)^{\abs{\Krewl\pi}}
\prod_{B\in\pi}K_{2\abs{B}}(\tilde{X})
\intertext{This in turn by \eqref{eq:cardKrew} is equal to}
&= n
\left(
1-\frac{1}{n}
\right)^{r+1}
\sum_{\pi\in\NC(r)}
\prod_{B\in\pi}\frac{n}{n-1}K_{2\abs{B}}(\tilde{X}).
\end{align*}
\begin{align*} %
&= (n-1)
\sum_{\pi\in\NC(r)}
\prod_{B\in\pi}\frac{n}{n-1}
K_{2\abs{B}}\Bigg( \sqrt{
1-\frac{1}{n}
}\tilde{X}\Bigg).
\end{align*}
In view of~\eqref{eq:KX2} this is the same as the cumulant $K_r(Z^2)$ where
$Z$ is an even random variable with cumulants
$$
K_r(Z)=\frac{n}{n-1}K_r\Bigg( \sqrt{
1-\frac{1}{n}
}\tilde{X}\Bigg)
.
$$
Such a random variable can be modeled as a free compression
$$
Z=\frac{n}{n-1}P \sqrt{
1-\frac{1}{n}
}\tilde{X}P=\sqrt{
\frac{n}{n-1}
}P \tilde{X}P,
$$
with $\tau(P)=\frac{n-1}{n}$, see
\cite[Corollary~14.13]{NicaSpeicher:2006}.
\end{proof}
\begin{Rem}
In the paper \cite{NicaSpeicher:1996} of Nica and Speicher cited above,
it was shown that for every probability measure $\mu$ there
is a convolution semigroup $\{\mu^{\boxplus t}\mid t\geq 1\}$.
Denote $\psi(\mu) = \inf\{t \mid \mu^{\boxplus t}\text{ exists}\}$.
This can be seen as some kind of ``measure of free non-infinite divisibility''
in the sense that $\mu$ is freely infinitely divisible if and only if
$\psi(\mu)=0$. It is related to the
free divisibility indicator $\phi(\mu)$ of \cite{BelinschiNica:2008}
by the inequality $\psi(\mu)\leq 1- \phi(\mu)$.
If $\tilde{X}$ exists,
the preceding proof shows that $\psi(Z)\leq \frac{n-1}{n}\psi(\tilde{X})$
and in particular, if $\tilde{X}$ is $\boxplus$-infinitely divisible,
then so is $Z$.
It then follows from
\cite[Theorem 6.1]{ArizmendiHasebeSakuma:2013}
that $Z^2$ is freely infinitely divisible as well and consequently also $Q_n$.
However if $X$ is not freely infinitely divisible,
the symmetrization $\tilde{X}$ constructed in Corollary~\ref{cor:KrQn}
in general cannot be realized as an operator,
see \cite[Remark~12~(2)]{NicaSpeicher:1998}.
\end{Rem}
We show in the final section that any quadratic form in free even
random variables preserves free infinite divisibility.
\subsection{Preservation of free infinite divisibility}
It is shown in \cite{ArizmendiHasebeSakuma:2013} that the free commutator of
freely infinitely
divisible random variables is also freely infinitely divisible
and the authors ask whether there
are other noncommutative polynomials which preserve free infinite divisibility.
We show here that for self-adjoint operators this is the case
for any quadratic form in free random variables
whose distribution does not depend on the odd cumulants
of the original distribution.
This includes the free commutator and free sample variance.
In the proof below we will use compound free Poisson distributions $\mu$
with rate $\lambda$ and jump distribution $\nu$ which is
the unique probability distribution with free cumulants
$K_n(\mu)=\lambda m_n(\nu)$.
Compound free Poisson distributions are freely infinitely divisible,
and moreover, any freely infinitely
divisible probability measure is the weak limit distribution of a sequence of
compound free Poisson random variables, see
\cite[Proposition A.2]{ArizmendiHasebeSakuma:2013}.
\begin{prop} \label{prop:NieskonczonaPodzielnosc}
Let ${\mathnormal X}_1, {\mathnormal X}_2,\dots, {\mathnormal X}_n\in \mathcal{A}_{sa}$ be a free family
of even freely infinitely divisible random variables.
Let $A=[a_{i,j}]_{i,j=1}^n\in M_n({\mathbb C})$ be a selfadjoint matrix,
then the distribution of the quadratic form
$T_n=\sum_{i,j}^na_{i,j}{\mathnormal X}_i{\mathnormal X}_j$ is also freely infinitely divisible.
\end{prop}
\begin{proof}
Suppose first that each $X_i$ is a symmetric compound free Poisson variable
with rate $\lambda_i$ and jump distribution $\nu_i$. Let $\Y_i$ be free random
variables with compound free Poisson distribution of rate $\lambda_i$ and jump
distribution $\nu_i^2$, respectively, i.e., with cumulants given by
$K_{r}(\Y_i)=K_{2r}({\mathnormal X}_i)$. Since the support of the jump distribution
$\nu_i^2$ is contained in the positive real axis, it is the spectral
measure of some positive random variable $Z_i$ and follows from
\cite[Proposition~12.18]{NicaSpeicher:2006} that we can represent $Y_i$ as
a free compression $Y_i=SZ_iS$ by a free semicircular element $S$ and in
particular $Y_i$ is positive as well.
Using the equation \eqref{eq:kumulantsamplevariancenotiid} we have
\begin{align*}
K_r(T_n)&=
\sum_{i_1,\dots,i_r\in[n]}
\;\mathrm{Tr}\,(AE_{i_1}AE_{i_2}\dots AE_{i_r})\,
\sum_{\substack{ \pi\in \NCeven(2r)\\
\pi \vee \onetwo{r}=\hat{1}_{2r}}}
K_\pi(X_{i_r},X_{i_1},X_{i_1},X_{i_2},\dots,X_{i_{r-1}},X_{i_r});
& \intertext{Now the bijection introduced in Lemma~\ref{lemm:lematoparzystych}
implies that $K_{r}(\Y_i)=K_{2r}({\mathnormal X}_i)$ and thus the above is equal to}
&= \sum_{ i_1,\dots,i_r\in[n]}\;\mathrm{Tr}\,(AE_{i_1}AE_{i_2}\dots AE_{i_r})\sum_{\substack{ \pi\in \NC(r) \\ \pi\leq \ker\underline{i}
} } K_\pi(Y_{i_1},Y_{i_2},\dots,Y_{i_r})
\\
&= \sum_{ i_1,\dots,i_r\in[n]}\;\mathrm{Tr}\,(AE_{i_1}AE_{i_2}\dots AE_{i_r})\sum_{\pi\in \NC(r) } K_\pi(Y_{i_1},Y_{i_2},\dots,Y_{i_r})
\\
&= \sum_{ i_1,\dots,i_r\in[n]}
\;\mathrm{Tr}\,(AE_{i_1}AE_{i_2}\dots AE_{i_r})
\,
\tau \Big(\prod_{j=1}^r\Y_{i_j}\Big)=n\times \;\mathrm{Tr}\,_n\otimes \tau\left[ \Big(\sum_{i=1}^nAE_i\otimes \Y_{i}\Big)^r\right].
\end{align*}
Hence the cumulant sequence of $T_n$ is the moment sequence of $(A\otimes I)Y$
in the noncommutative probability space $M_n({\mathbb C})\otimes \mathcal{A}$, with state
$\;\mathrm{Tr}\,_n\otimes \tau$ where ${\mathbf{Y}}=\sum_{i=1}^nE_i\otimes \Y_{i}$.
This operator is not self-adjoint, yet the following
arguments show that it is indeed a positive definite moment sequence.
We have seen above that all $Y_i$ are positive and it follows that ${\mathbf{Y}}$ is
positive as well, thus the sequence
$$
\;\mathrm{Tr}\,\otimes\tau( ((A\otimes I){\mathbf{Y}} )^r)
=\;\mathrm{Tr}\,\otimes\tau( ({\mathbf{Y}}^{1/2}(A\otimes I){\mathbf{Y}}^{1/2} )^r)
$$
is indeed the moment sequence of a self-adjoint random variable.
Suppose now that $X_i$ has a more general symmetric distribution $\mu_i$.
Then the argument of the proof of Proposition A.2. in \cite{ArizmendiHasebeSakuma:2013}
shows that $\mu_i$ can be approximated by symmetric compound free Poisson variables,
say $\mu_i=\lim_{k\to \infty}\mu_{i,k}$.
It follows from the above argument that the distribution $T_n$ can be approximated by
freely infinitely divisible distributions and since $ID(\boxplus)$ is closed under
convergence in distribution, $T_n$ is freely infinitely divisible as well.
\end{proof}
Putting together Lemma~\ref{lem:oddcancellation} and Proposition~\ref{prop:NieskonczonaPodzielnosc} we obtain the following corollary.
\begin{cor}
Let ${\mathnormal X}_1, {\mathnormal X}_2,\dots, {\mathnormal X}_n\in \mathcal{A}_{sa}$ be a free family
of freely infinitely divisible random variables.
Let $P$ be a selfadjoint symmectric polynomial of degree $2$
in noncommuting variables such that the distribution of
the random variable $Y=P(X_1,X_2,\dots,X_n)$ does not depend on the odd cumulants.
Then the distribution of $Y$ is freely infinitely divisible as well.
In particular,
the commutator $i(X_1X_2-X_2X_1)$ of two freely infinitely divisible
random variables is freely infinitely divisible and
the same is true of the sample variance of a free identically distributed
family of freely infinitely divisible random variables.
\end{cor}
\section{Concluding Remarks}
In the present paper we have shown that the sample variance
shares the following properties with the free commutator:
\begin{enumerate}
\item Odd cumulants do not contribute to the distribution.
\item Free infinite divisibility is preserved.
\end{enumerate}
This phenomenon raises the following problems and conjectures,
some of which will be investigated in forthcoming papers.
\begin{problem}
\label{prob:oddpoly}
Characterize the class of selfadjoint polynomials
$P\in{\mathbb C}\langle X_1,X_2,\dots,X_n\rangle$
in noncommuting variables $X_1,X_2,\dots,X_n$ with the property
that the distribution of $P({\mathnormal X}_1,\dots, {\mathnormal X}_n)$ does not depend on the
odd cumulants of $X$ whenever ${\mathnormal X}_1, {\mathnormal X}_2,\dots, {\mathnormal X}_n$
are free copies of a fixed random variable $X$.
\end{problem}
\begin{Con}
Whenever a homogeneous polynomial $P$ has the properties described
in Problem~\ref{prob:oddpoly}
and $X_1,X_2,\dots,X_n$ are free copies of a freely infinitely divisible
random variable $X$, then $P({\mathnormal X}_1,\dots, {\mathnormal X}_n)$ is
freely infinitely divisible as well.
\end{Con}
\emph{Acknowledgments}.
The authors would like to thank Marek Bo\.zejko and
Roland Speicher for several discussions and helpful comments.
We are very grateful to Takahiro Hasebe for many comments and in particular
for pointing out a gap in the proof of Proposition~\ref{prop:NieskonczonaPodzielnosc}.
The first author also thanks Abram Kagan for a very interesting discussion about Ruben's
problem during AMISTAT 2015 in Prague.
Finally we thank the referee for a careful reading of the manuscript and
numerous minor corrections.
The work was partially supported by grant number 2014/15/B/ST1/00064 from the \textit{Narodowe Centrum Nauki},
Project No P 25510-N26 of the Austrian Science Fund (FWF) and
travel grant PL 08/2016 of the oead.
|
\section{Introduction}
The goal of this paper is to show that certain functions which take multiple matrices as inputs can be reduced to a combination of unitarily invariant functions operating on disjoint subsets of the input matrices.
As such, the original functions will be invariant with respect to unitary conjugation among the subsets of input matrices.
We will achieve these functions by summing over a finite group of unitary matrices.
This group is guaranteed to exist whenever the underlying space is a commutative ring (simply by satisfying the ring conditions) and so all formulas will continue to hold when integrated against any measure which is invariant under the action of this group.
In particular, this will be true for any Haar measure defined over unitary matrices (where the property of being ``unitary'' is dictated by the underlying ring operations).
As such, they generalize results in \cite{mypolys, polys} that were used in the development of a finite version of free probability.
Formulas of this type were also used in \cite{IF4} in combination with the ``method of interlacing polynomials'' (developed in \cite{IF1, IF2}) to show the existence of Ramanujan graphs of all sizes and degrees.
These formulas can be viewed as quadratures formulas, polarization formulas, or statements about zonal spherical polynomials.
Will go out of our way to keep all proofs as elementary as possible.
The presentation is designed to highlight the generality of the results; in particular, it should be noted that nothing done in this article requires the existence of mulitiplicative inverses of any form.\footnote{
We will exercise one exception to the ``no dividing'' rule with respect to writing factorials --- rather than introduce additional notation (the Pochhammer symbol, for example), we will leave constants of the type $m!/n!$ where $m \geq n$ in a ``divided'' form.
In each such case, the resulting division will result in a positive integer and is used more as a notational convenience than as an actual scalar quantity (see the discussion at the end of Section~\ref{sec:rings}).}
\subsection{Organization}
In Section~\ref{sec:prelims}, we will introduce/review the notations, terminology and definitions we will use.
This will include defining two subgroups of matrices that we call $\mathcal{Q}$ and $\mathcal{P}$.
We also state our main technical lemma (Lemma~\ref{lem:minors}).
Section~\ref{sec:lemmas} contains our main lemmas regarding the symmetries of $\mathcal{Q}$ and $\mathcal{P}$ and then in Section~\ref{sec:apps}, we use these lemmas to prove identities involving some determinantal functions that exhibit these symmetries.
In Section~\ref{sec:cp}, we give applications of the previous sections to obtain formulas for characteristic polynomials of the type used in \cite{mypolys, polys}.
Finally, in Section~\ref{sec:end}, we conclude with an open problem.
\section{Definitions and such}\label{sec:prelims}
\subfile{new_prelims.tex}
\section{Symmetries of $\mathcal{Q}$ and $\mathcal{P}$}\label{sec:lemmas}
\subfile{lemmas.tex}
\section{Applying symmetries to determinants}\label{sec:apps}
\subfile{applications.tex}
\section{Characteristic Polynomials}\label{sec:cp}
\subfile{new_characteristic_polynomials.tex}
\section{Conclusion}\label{sec:end}
The results in this paper generalize results from \cite{mypolys, polys} in two ways: firstly, in the types of determinantal constructs that can be considered (more matrices are allowed) and secondly, in the underlying domain of the matrices that can be considered.
One lemma that could has potential for being extended is Lemma~\ref{lem:cancel2}.
We suspect that a formula can be obtained for four sets in general (as was the case for Lemma~\ref{lem:cancel}):
\begin{problem}
Find a formula for
\[
\sum_{P \in \mathcal{P}_n} [P]_{S, T} [P^{-1}]_{U, V}
\]
that holds for all $S, T, U, V \in \binom{[n]}{k}$.
\end{problem}
\subsection{Rings}\label{sec:rings}
We will use $\mathbbm{K}$ throughout to denote a commutative ring with multiplicative identity $\mathbbm{1}$ and additive identity $\emptyset$.
As is customary, we will denote the additive inverse of $\mathbbm{1}$ as $(-\mathbbm{1})$.
In the event that $\mathbbm{1} = (-\mathbbm{1})$, we will call the ring {\em boolean}.
For an assertion $X$, we define the function $\true{X}$ as
\[
\true{X}
=\begin{cases}
\mathbbm{1} & \text{ if $X$ is true } \\
\emptyset & \text{ if $X$ is false }.
\end{cases}
\]
We denote the ring of univariate polynomials over $\mathbbm{K}$ as $\mathbbm{K}[x]$ ($x$ being a formal variable).
Let $S_n$ be the symmetric group on $n$ elements.
The mapping $\sigma \mapsto P_\sigma$ where $P_\sigma$ is the $n \times n$ matrix with
\[
P_{\sigma}(i, j)
= \true{\sigma(i) = j}
=
\begin{cases}
\mathbbm{1} & \text{if } \sigma(i) = j \\
\emptyset & \text{otherwise}
\end{cases}
\]
is an injective homomorphism from $S_n$ into $\mathbbm{K}^{n \times n}$; we will denote the image of this mapping as $\mathcal{P}_n$ and refer to members of $\mathcal{P}$ as {\em permutation matrices}.
For a set $S \subset [n]$, the mapping $S \mapsto Q_S$ where $Q_S$ is the $n \times n$ diagonal matrix with
\[
Q_S(i, i)
=
\begin{cases}
-\mathbbm{1} & \text{if } i \in S \\
\mathbbm{1} & \text{if } i \notin S \\
\end{cases}
\]
is an injective homomorphism from $\mathbbm{F}_2^n$ into $\mathbbm{K}^{n \times n}$; we call the image of this mapping $\mathcal{Q}_n$, and refer to its members as {\em sign matrices}.
We remark that the only matrix in $\mathcal{P}_n \cap \mathcal{Q}_n$ is the identity matrix $I$, and that $\mathcal{Q}_n = \{ I \}$ if $\mathbbm{K}$ is boolean.
Also, it should be clear (via the associated homomorphisms) that each element of $\mathcal{P}_n$ and $\mathcal{Q}_n$ is invertible.
One can check that the group generated by $\mathcal{P}_n$ and $\mathcal{Q}_n$ (via multiplication) is the largest subalgebra that can be guaranteed to be invertible (for general $\mathbbm{K}$).
This group is isomorphic to the Coxeter group $B_n$ (or, dually, $C_n$), the group of symmetries of the $n$-cube (or, dually, the $n$-crosspolytope).
It is also known as the hyperoctahedral group.
Note that the only explicit ring elements we will use are $\{ (-\mathbbm{1}, \emptyset, \mathbbm{1}) \}$.
All other ring elements will appear implicitly as the elements of matrices.
If other constants appear (like positive integers), they are not to be considered ring elements.
Rather for an element $k \in \mathbbm{K}$ and a nonnegative integer $n$, the following interpretations should be used
\begin{enumerate}
\item $n k$: adding $n$ copies of $k$ (using ring addition)
\item $k^n$: multiplying $n$ copies of $k$ (using ring multiplication).
\end{enumerate}
We direct the reader to \cite{lang} for definitions related to rings.
\subsection{Determinants}
Since $\mathbbm{K}$ is commutative, we can take the usual definition of determinant for matrices in $\mathbbm{K}^{n \times n}$:
\begin{equation}
\label{def:det}
\mydet{A} = \sum_{\sigma \in S_n} (-\mathbbm{1})^{|\sigma|} \prod_i A(i, \sigma(i))
\end{equation}
where $S_n$ denotes the symmetric group on $n$ elements and $|\sigma|$ denotes the parity of $\sigma$.
Note that all additions and multiplications are considered to be with respect to the ring operations.
For sets $S, T \in \binom{[n]}{k}$, we recall that the $S, T$ {\em minor} of $A$ is defined as
\[
[A]_{S, T} =
\begin{cases}
\mydet{\{A(i,j)\}_{i\in S, j\in T}} & \text{for } k > 0 \\
1 & \text{for } k = 0.
\end{cases}
\]
Here, the notation
\[
\mydet{\{A(i,j)\}_{i\in S, j\in T}}
\]
is used to denote the determinant of the submatrix of $A$ with rows indexed by $S$ and columns indexed by $T$.
For $A \in \mathbbm{K}^{n \times n}$, we also define
\begin{equation}
\label{def:A^k}
[A]^{(k)} = \sum_{S \in \binom{[n]}{k}} [A]_{S, S}
\end{equation}
(with $[A]^{(0)} = 1$).
We show in Lemma~\ref{lem:cp} that $[A]^{(k)}$ is equal to the coefficient of $x^{n-k}$ in the polynomial
\[
\mydet{x I + A} \in \mathbbm{K}[x].
\]
Thus, to the extent that it makes sense for a given $\mathbbm{K}$ (the reals, for example), $[A]^{(k)}$ can be viewed as the $k$th elementary symmetric polynomial evaluated at the $n$ eigenvalues of $A$.
Our main technical tools will be the following decompositions:
\begin{lemma}
\label{lem:minors}
For $A \in \mathbbm{K}^{m \times n}$ and $B \in \mathbbm{K}^{n \times p}$, we have
\begin{equation}
\label{eq:mult}
[AB]_{S, T} = \sum_{U \in \binom{[n]}{k}} [A]_{S, U} [B]_{U, T}
\end{equation}
for all $S \in \binom{[m]}{k}$ and $T \in \binom{[p]}{k}$, and
\begin{equation}
\label{eq:add}
\mydet{A + B} = \sum_{S, T \subseteq [n]}(-\mathbbm{1})^{\|S + T\|_{1}} [A]_{S, T} [B]_{\overline{S}, \overline{T}}
\end{equation}
for all $A, B \in K^{n \times n}$.
\end{lemma}
\begin{proof}
The first identity is a direct application of the well known Binet--Cauchy theorem, while the second is merely a grouping of terms in (\ref{def:det}), see \cite{hornjohnson}.
\end{proof}
One technical issue that will arise concerning (\ref{eq:add}) is that the correctness of (\ref{eq:add}) depends upon the ground set being $[n]$ (otherwise the $\| S + T \|_{1}$ term would be altered).
In order to apply it to quantities like $[A + B]_{X, X}$ where $X$ is some subset of $[n]$, we will need to keep track of rows both in the frame of $X$ and in the frame of the larger matrix.
This motivates the introduction of the induced set constructs defined in (\ref{eq:setset}).
\end{document} |
\section[Introduction]{Introduction}
The structure of network glasses is often described by continuous random network (CNR)
model. In this model, building units form a random network where short-range order is
preserved similar to that in crystals but translational long-range order is absent
due mainly to distorted bond angles \cite{rosenhain1927structure,zachariasen1932atomic,
wright2013eighty}. Such structures have been generally studied by models
\cite{gupta1990topologically} and diffraction experiments \cite{warren1992x}
which have provided invaluable information on short-range and medium-range
order, mostly in the form of pair distribution functions (PDFs)
\cite{wright1994neutron,elliott1991medium,thorpe2012properties,
treacy2005fluctuation}.
One challenge in using diffraction data is that this only provides average properties
such that the structure cannot be reconstructed uniquely. Meanwhile,
Scanning Probe Microscopy (SPM) and Electron Microscopy (EM) techniques have radically
shortened the resolution limit and recently true atomic resolution images of silica
bilayers and other two-dimensional (2D) amorphous surfaces have become
available \cite{lichtenstein2012atomic,huang2012direct}. However, high
resolution imaging of bulk amorphous materials remains elusive \cite{burgler1999atomic}.
These new results on 2D glasses have opened up numerous opportunities to
study the structure of glasses using actual atomic
coordinates. Recent work on 2D glasses includes modeling of silica
bilayers \cite{wilson2013modeling,wilson2015modelling}, ring
distribution \cite{kumar2014}, medium-range order \cite{buchner2016building},
suitable boundary conditions to recover missing
constraints in the surface \cite{theran2015anchored} and the refinement of experimental
samples \cite{sadjadi17refining}. Rigidity theory has also uncovered a
connection between 2D glasses and jammed disk
packings \cite{thorpe1983continuous,ellenbroek2015rigidity}.
The remarkable images of vitreous bilayer silica (SiO$_{2}$) unveil a ring
structure which is the characteristic of covalent glasses. But similar underlying
structure also can be found in various amorphous materials such as amorphous graphene
\cite{kotakoski2011point,kumar2012amorphous,buchner2014topological}. In fact,
these atomic materials are members of a larger class of materials
(many with larger length scales) collectively
known as \textit{cellular networks}. Examples are foams and grains \cite{sadoc%
2013foams}, biological tissues \cite{mombach1993two}, metallurgical aggregates,
geographical structures, crack networks \cite{korneta1998topological}, ecological
territories, Voronoi tessellations \cite{weaire1984soap, stavans1993evolution}
and even the universe at large scale \cite{aragon2014universe} and
fractals \cite{schliecker2001scaling}. Given wide range of length scales,
formation mechanisms and physical properties, cellular networks have been subject
of many studies \cite{gibson1999cellular,schliecker2002structure}. Despite the
topological resemblance between 2D amorphous systems and other cellular networks,
one should note that these materials are microscopic systems with a very
different nature of bonds and forces and hence they can shed light on new
properties of cellular networks, in particular those related to geometry.
\begin{figure}[b!]
\includegraphics[width=4.0cm,height=4.0cm]{fig1.pdf}\quad
\caption{\label{fig:randnet} A piece of a two dimensional cellular network generated by
bond-switching algorithm from a honeycomb lattice. Rings are colored based
on their size. On the bottom left corner, a group of six-fold rings can be
seen which also happens in experimental samples and is a feature of
amorphous materials, due to statistical correlations. A central six fold ring
has been left uncolored and shells of rings will be found around this. Any
ring can be used as a central location.}
\end{figure}
These glassy networks are almost entirely $3-$coordinated networks, i.e.,
each vertex is connected to three other vertices through edges which form the
boundary of polygonal rings (Fig. \ref{fig:randnet}). In the case of
amorphous graphene - vertices represent carbon atoms. In silica bilayer, rings
are formed by connecting silicon atoms while intervening oxygen atoms
are omitted.
These glassy networks, to some extent, are random and their study requires a
statistical approach but experimental samples of amorphous materials
are relatively small \cite{klemm2016preparation}. Additionally, the small size of
many samples does not permit the study of ring correlations at larger distances
with good statistics. In this work, we employ large computer
models to study correlations among the rings.
In the literature, the focus has been on the correlation among adjacent
rings where well-known Aboav-Weaire's law captures the tendency of smaller
and larger rings to be adjacent. This paper studies various correlation
functions out to large topological and geometrical distances and generalizes the
Aboav-Weaire's law.
\begin{figure}[t!]
\includegraphics[scale=0.27]{fig2a.pdf}\quad
\includegraphics[scale=0.27]{fig2b.pdf}
\caption{\label{fig:dis} (a): Partitioning of the random network in Fig.
\ref{fig:randnet} into
topological shells. The shells grow roughly in circular shapes.
This piece also has a triplet inclusion in the forth (blue) shell where a
5-ring is isolated from the fifth (purple) shell. (b) Although shells are
roughly circular, no circle can sweep all rings within a single shell; hence
ring distributions with topological and geometrical definitions are different.}
\end{figure}
\section{Shell analysis and correlations}\label{sec:framework}
We define an $n-$ring as a ring with $n$ adjacent rings. The ring distribution
of a network with a total of $N$ rings is characterized by $p(n)$, the fraction of
$n-$rings, its mean $\braket{n} = \sum_{n} n p(n)$, and the second moment about
the center $\mu = \braket{n^2} - \braket{n}^2$. According to Euler's theorem,
mean ring size for a network with periodic boundary conditions (PBCs) is exactly
$\braket{n} = 6$ \footnote{The mean ring size in the finite experimental
samples is slightly less since the surface sites are
under-coordinated. Although for sufficiently large
systems, boundary effects are negligible.}. The ensemble
average of a quantity $x$ is defined as $\braket{x} = \sum_{n} p(n) x$.
To overcome the finite size effect in the experimental samples,
we use computer-generated models under PBCs with $\sim$100000 vertices
($\sim$50000 rings) generated from an initially honeycomb lattice using
bond-switching algorithm. Here,
a bond between two nearest neighbor sites is selected and replaced by a dual bond
at right angle and local topology is reconstructed to maintain the three-fold
coordination everywhere~\cite{wooten1985computer,stone1986theoretical}. Although,
experimental samples contain rings with size $4$ to $9$, but fraction of rings
with sizes other than $5$ to $7$ are statistically quite rare \cite{kumar2014}.
We studied two networks one with only $5$ to $7$ rings and one with $5$ to $8$
but no essential difference was observed. Therefore we report results
of the network with $5$ to $8-$fold rings with the following ring
distribution: $p(5) = 0.262$, $p(6) = 0.494$,
$p(7) = 0.227$, $p(8) = 0.0172$ and $\mu = 0.558$.
Nevertheless, the measures of this paper are general and can be applied
to all glassy and cellular networks.
The correlation among rings is usually defined over
a \textit{topological distance} $t$. The topological distance between two rings
is defined as the minimum number of bonds should be traversed to connect two
rings. This distance is the equivalent of distance of two nodes in the dual graph
(when each ring is represented by a node) of Fig. \ref{fig:randnet}.
The distance of a ring from itself is zero ($t=0$). All rings
which have one common side with a given central ring are located at $t=1$
(first shell). Adjacent rings to the first shell, excluding the central ring,
are at $t=2$ (second shell). This process can be continued to find
shells at any topological distance similar to Fig. \ref{fig:dis}. A ring at shell
$t$ is adjacent to at least one ring at shell $t-1$ and usually adjacent to
at least one ring in shell $t+1$, otherwise this ring is
trapped and forms a \textit{triplet inclusion} (Fig. \ref{fig:dis}).
This definition naturally divides/partitions the network into concentric shells
around any given ring. Therefore, all properties of the network are studied as a
function of the topological distance and the size of the central ring
\cite{fortes1993average,mason2015geometric}, as first pointed out by Aste et
al \cite{aste1996statistical, aste1996one}.
A shell at distance $t$ from an $n-$ring is characterized by
three numbers: number of $n'$-rings $N_t(n,n')$; total number of rings (shell
size) $K_t(n)$, and total number of sides (edges) $M_t(n)$. These quantities
are related as follows:
\begin{align}
K_t(n) &= \sum_{n'} N_t(n,n'), \label{eq:shellsize} \\
M_t(n) &= \sum_{n'} n' N_t(n,n'). \label{eq:edgenumber}
\end{align}
Since these equations are linear, they are also valid for the averaged values over
all $n-$rings. More importantly, note that $N_t(n,n')$ is not symmetric in respect
to $n$ and $n'$. This reflects the fact that local order of the rings
is strongly dependent on the size of the central ring.
Specially,
$N_t(n,n')$ should not be confused by the number of
$n-n'$ pairs at topological distance $t$:
\begin{equation}\label{eq:symfun}
N p(n) N_t(n,n') = N p(n') N_t(n',n),
\end{equation}
which by definition is symmetric.
This symmetry can relate the ensemble average
of the number of sides (Eq. \ref{eq:edgenumber}) to the ensemble average of shell
size (Eq. \ref{eq:shellsize}) at any topological distance:
\begin{align}\label{eq:gwsr}
\braket{M_t} &= \sum_{n} p(n)M_t(n)
= \sum_{n}\sum_{n'}p(n) n' N_t(n,n') \nonumber \\
&= \sum_{n'} n' p(n') K_t(n')
= \braket{n K_t}.
\end{align}
This relation is the generalized Weaire sum rule which was originally
proposed for the first shell where it
takes the form $\braket{M_1}=\braket{n^2} = \braket{n}^2 + \mu$
\cite{weaire1974some,lambert1983order}. Note that the first shell is the
only shell that $K$ is exactly determined $[ \,K_1(n)=n] \,$ but Eq. \ref{eq:gwsr}
surprisingly encapsulates all the statistical variation in the local ring
distribution in a simple form.
\begin{figure}[t]
\centering
\includegraphics[width=5.8cm,height=5.8cm]{fig3.pdf}
\caption{
\label{fig:shellsize}
Dependence of the number of rings $K_t(n)$ on topological distance $t$ and
size of the central ring $n$. $K_t(n)$ grows linearly for $t \ge 4$. Solid lines
are fitted lines to the last three points. Points are offset for clarity with
$6n$.}
\end{figure}
The space-filling nature of rings in the network requires that
$K_t(n)$ scales linearly with $t$ in the absence of correlation. This means that the
growth rate of the shell size is a constant number independent of the size of
the central ring. Although, geometrical constraints on the polygonal tiling of
the plane does not allow a complete independence from the central ring simply
because shell closure around a larger ring requires more rings. As a result,
the intercept of $K_t(n)$ remains a function of $n$. Therefore
we expect that:
\begin{equation}\label{eq:linform}
K_t(n) = A t + B(n),
\end{equation}
for $t \ge \xi$, where $\xi$ is the ring correlation length.
In a hexagonal lattice, the growth rate $A$ is 6 but as Fig. \ref{fig:dis}
shows, in a random network, shells grow roughly in circular form and simple
geometrical arguments predict that the growth rate should be $2\pi$.
However, because rings meet each other at random orientations and the shell
surface is rough, the actual growth rate is usually greater than $2\pi$ and
$A$ can be a measure of this roughness \cite{oguey2001roughness}.
Figure \ref{fig:shellsize} shows the number of the rings in the shells around
different central rings. The linear behavior of the shell size is observed in
various systems and is present in 2D glass, as expected. However, in 2D glasses
$A=7.31\pm0.1$ which is much smaller compared to the reported
values for Voronoi tessellation ($11.0\pm0.2$) and soap ($9.45\pm0.1$)
\cite{aste1996statistical}, probably due to the bond bending interactions
which result in the high symmetry (close to the maximum area forgiven edge
lengths) of the rings in the 2D glass \cite{kumar2014}.
Another useful quantity is the topological charge of an $n$-ring defined as
$6-n$. Since the mean ring size in the network is $6$, equivalently total charge
of the network is zero. However a piece of the network can contain any
amount of charge depending on the local ring distribution. Hence, topological
charge is a useful quantity that monitors the local deviation from the bulk
properties. In particular, the topological charge of a shell $q_t(n)$ can
be defined as the sum of the charge of its rings:
\begin{equation}\label{eq:shellcharge}
q_t(n) = \sum_{n'} (\braket{n}-n') N_t(n,n') = \braket{n} K_t(n) - M_t(n).
\end{equation}
\begin{figure}[t]
\centering
\includegraphics[width=5.8cm,height=5.8cm]{fig4.pdf}
\caption{\label{fig:charge}
Shell charge $q_t(n)$ vs. topological distance $t$. The shell charge
settles to a constant non-zero number for $t~\ge~4$. The dashed line shows the
asymptotic offset $-0.4$.}
\end{figure}
From short- and medium-range order, it is expected that rings around a given
ring are distributed such that the charge of the central ring is screened by
the charge of the neighboring shells and for $t>\xi$, the ring distribution is
similar to the bulk (charge per
shell is zero). But as Eq. \ref{eq:linform} shows, the shell size is a function
of $n$ for any distance and therefore rings are counted with different weights
in calculating the charge per shell. In fact, Eqs. \ref{eq:gwsr},
\ref{eq:linform} and \ref{eq:shellcharge} readily
yield an asymptotic value for the shell charge for $t>\xi$:
\begin{equation}\label{eq:asymcharge}
\braket{q_t} = \braket{(\braket{n}-n)K_t}
\approx p(5)B(5)-p(7) B(7),
\end{equation}
which is exact for a network with $n=5,6,7$ and approximately correct
as long as fraction of the other rings is negligible. Therefore $\braket{nK_t}$
does not factorize and statistically, there is a tendency to have
larger rings in a shell $[ \,\braket{q_t}<0$ since $B(7)>B(5)] \,$ .
The results of calculating the charge per shell is shown in Fig.~\ref{fig:charge}.
For $t=1$, the total shell charge has an opposite sign to the charge of the
central ring to screen the charge but for $t>1$ screening does not happen and the
charge per shell reaches a non-zero constant value, conjectured in Eq. \ref{eq:asymcharge}.
It is interesting to note that although the charge of $5$- and $7$-rings have
the same magnitude, the strength of screening for these two is considerably
different in the first shell. This shows that geometry has a strong effect on the
ring distribution. Note that hexagons have short-range correlations ($\xi=1$)
but other rings are correlated up to $\xi=3$ (medium-range correlation)
with different strengths.
Topological charge gives a rather complete picture of correlations in the
shell structure, but the most studied measure of correlations in the literature
is the mean ring size in the first shell around a central ring, through the
well-known Aboav-Weaire law that a ring with large size tends to have smaller
rings in its neighborhood and vice versa \cite{chiu1995aboav,
mason2012geometric}. Mathematically, the mean ring size $m_1(n)$ around a ring with
$n$ neighbors can be written (to a very good approximation)
as \cite{aboav1970arrangement,weaire1974some}:
\begin{equation}\label{eq:awl}
n m_1(n) = \braket{n}^2 + \mu + \braket{n} (1-\alpha)(n-\braket{n}),
\end{equation}
where $\alpha$ is a fitting parameter which depends on the specific network.
Usually a network is characterized by
$(\mu,\alpha)$. The meaning of $\alpha$ is not clear but it has been argued that
it is a metrical quantity \cite{aboav1984arrangement} or the average excess
curvature \cite{mason2012geometric} but these definitions only work in special
cases. In our network, $\alpha\approx0.23$ which is somewhat smaller than values
extracted from experiments \cite{kumar2014} showing computer generated
models still need further refinement.
\begin{figure}[t]
\centering
\includegraphics[width=5.8cm,height=5.8cm]{fig5.pdf}
\caption{
\label{fig:chargegeo}
The topological charge $q_r(n)$ per shell is plotted against the geometrical
distance $r$. The shell charge approaches zero for distances about three
rings away. This figure should be compared to Fig. \ref{fig:charge}. The two
dashed lines represent the geometrical distance corresponding to the minimum
and maximum values for the first shell with $t=1$. Curves are offset for clarity
where horizontal solid lines show the expected asymptotic values of zero.}
\end{figure}
We would like to extend Aboav-Weaire law to longer distances to study
correlation of a ring with the shells around it. The above form can be
used to propose a generalized Aboav-Weaire law as:
\begin{equation}\label{eq:nmn}
n m_t(n) = \braket{n}^2 + \mu_t + \braket{n} (1-\alpha_t)(n-\braket{n}),
\end{equation}
where for $t=1$ we recover Eq. \ref{eq:awl} with $\mu_{1}=\mu$.
A similar argument presented to derive Eq. \ref{eq:asymcharge} can be used
to find an asymptotic value for $m_t(n)$. At sufficiently long distances, the
ring distribution in the shells is independent of the size of the central
ring and $\braket{M_t} \approx \braket{m_t K_t} =
\braket{m_t} \braket{K_t}$, therefore for $t>\xi$:
\begin{equation}\label{eq:asvm}
\braket{m_t} = \frac{\braket{M_t}}{\braket{K_t}} =
\frac{\braket{nK_t}}{\braket{K_t}} = 6 - \frac{\braket{q_{\infty}}}{\braket{K_t}}.
\end{equation}
While we expect $\alpha_{\infty}=0$ but we showed, $\braket{q_{\infty}}<0$,
so the asymptotic value of $m_{\infty}(n)$ is larger than the bulk value $6$.
For this reason, $m_t(n)$ approaches 6 as $t^{-1}$ (since $K_t(n) \sim t$)
which is sometimes interpreted as a long-range correlation \cite{
wang2012generalization,wang2014long}. However this should be regarded as an
artifact because the shells are defined in such a way (topologically) which
results(unfortunately) in the topological charge never going to zero, even at
very large distances, and in fact approaching a constant as shown here. This
is due to the non-circular nature of the shells, and can be avoided if the
shells are chosen in such a way as to make them more nearly circular.
Unfortunately this is not possible with a purely topological definition, and
so we are forced to adopt a \textit{geometrical definition} for the ring-shell
correlations.
Figure \ref{fig:dis} shows the difference between topological and geometrical
distance. Despite the fact that shells found by topological distance are roughly
circular, it is not possible to find a single circle which contains all the
rings in the shell, therefore ring distributions etc. are different in the two cases.
The \textit{geometrical } distance $r$ between two rings is defined as the Euclidean
distance between their centroids. Therefore, instead of using the discrete integer
distance $t$, the quantities $q$ and $m$ are written as a function of a
continuous distance $r$:
\begin{align}\label{}
q_r(n) &= 6 K_r(n) - M_r(n), \\
n m_r(n) &= \braket{n}^2 + \mu_r + \braket{n} (1-\alpha_r)(n-\braket{n}).
\end{align}
\begin{figure}[t]
\centering
\includegraphics[width=5.8cm,height=5.8cm]{fig6.pdf}
\caption{\label{fig:aboav}
Plot of two coefficients in the generalized Aboav-Weaire law,
$\alpha_r$ and $\mu_r$ with their topological counterparts,
$\alpha_t$ and $\mu_t$. Geometrical definitions show that
correlation quickly decays to zero while pseudo-correlations
in the topological case last over a long-range for $\mu_t$. The geometrical
distances are chosen so the geometrical and topological distances agree
for the first shell.}
\end{figure}
Since $r$ is continuous, a binning procedure is used to compare with the previous
results using topological distance. Small bins are used with a windowing
procedure where the width of the window mimics unity in topological distance.
Results for the charge are shown in Fig. \ref{fig:chargegeo}. It is evident
that correlations last about $3$ shells and are quite short-ranged with the
charge going to $0$ over the same range, as expected. Therefore this definition
of a shell using geometrical distance is more useful. Because of the different
size of the rings, e.g., distance between a $5-6$ pair is greater than a $7-8$
pair so a range of geometrical distances corresponds to a single topological
distance. To compare the two distances, we rescale the geometrical distance
by the average distance between adjacent rings, which is defined to be unity.
Fig. \ref{fig:chargegeo}
shows this for the first neighbors with two dashed lines. Within this
window, all four curves show a common trend: a maximum followed by a
minimum. The former corresponds to $5-$rings (positively charge) and
the latter to $7-$ and $8-$rings (negatively charged). The point in the middle
corresponds to neutral $6-$rings. The horizontal axis is normalized such
that these three points line up for all curves. According to Aboav-Weaire
law, smaller rings surround a larger ring; the pronounced minimum
of $q_r(5)$ due to $7-$ and $8-$rings and the pronounced maximum of $q_r(7)$
and $q_r(8)$ due to $5-$rings admit this law. In the case of $q_r(6)$,
minimum and maximum have the same amplitude due to uniform distribution
of the rings around hexagons hence their weak correlations with other
rings.
It is also constructive to look at Aboav-Weaire law using geometrical
distance. In this case, we expect that both $\alpha_{r}$ and $\mu_{r}$ decay
rapidly to zero in accordance with the absence of correlations for large
$r$. This is confirmed in Fig. \ref{fig:aboav} which clearly for distances
larger than $3$, the mean ring size is essentially exactly $6$.
This confirms our assertion that ring correlations in glassy networks are
either short-range or medium-range and using geometrical distance in the
calculations of topological charge and mean ring size resolves the issue
of excess topological charge in the shells found by topological distance
which is shown by the long-tail of $\mu_2$ in Fig. \ref{fig:aboav}.
Fig. \ref{fig:nmn} shows linearity of the generalized Aboav-Weaire
law for the third neighbors. The plot shows that $nm(n)$ is indeed a linear
function of $n$ but because of pseudo-correlations, the average ring size using
topological distance is slightly larger than expected for geometrical
distance, where the mean ring size is 6 for three-fold coordinated
networks.
\begin{figure}[t]
\centering
\includegraphics[width=5.8cm,height=5.8cm]{fig7.pdf}
\caption{
\label{fig:nmn}
Plot of weighted mean ring size $nm(n)$ versus ring size $n$ for
the third neighbors using both geometrical and topological distance. This plot
shows that the mean ring size for all shells follows the generalized
Aboav-Weaire law (Eq. \ref{eq:nmn}). Note that the topological definition
leads to a slightly larger mean ring size.}
\end{figure}
Although the topological charge and Aboav-Weaire law are useful tools to quantify
correlations, they only measure correlations between a ring and shells.
The ring-ring correlation function is perhaps a better measure of
correlations especially since, as it was shown, definition of shells using
the topological distance do introduce some artifacts such as excess charge.
To find out the correlation between two single rings, we need to derive an
expression for the probability $p_t(n,n')$ of finding a pair of $n, n'$ rings
with distance $t$. For a given $n-$ring, the number of $n'-$rings
at distance $t$ is $N_t(n,n')$ while on average a typical shell has
$\braket{K_t}$ rings. Therefore the probability of having a pair of rings
is \cite{szeto1998topological}:
\begin{equation}\label{probability}
p_t(n,n') = \frac{p(n)N_t(n,n')}{\braket{K_t}}.
\end{equation}
This equation is important as it relates ring distributions of the shell
structure to of the network (For $t=1$, this equation reduces to the
correlation function defined in Ref. \cite{le1993correlations}).
If the rings were independent, this probability is simply product of
the individual probabilities but we showed the ring distribution
of a shell is different from the bulk and rings are topologically
dependent even for large $t$. This motivates us to define the
probability of having an $n-$ring at shell $t$ (independent of the
central ring) as:
\begin{equation}\label{shellprob}
p_t(n)=\sum_{n'} p_t(n',n) = p(n)\frac{K_t(n)}{\braket{K_t}},
\end{equation}
which can be derived using Eqs. \ref{eq:shellsize} and \ref{eq:symfun}.
The probability of having $n-$ring is proportional to the average shell
size around $n-$fold rings and the ensemble averaged shell size.
We define correlation function between two $n$ and $n'$ sided rings as:
\begin{equation}\label{eq:rrc}
C_t(n,n') = p_t(n,n') - p_t(n)p_t(n')
\end{equation}
Figure \ref{ringringcor} shows the results for the above correlation function.
This clearly shows the medium-range order of the rings except for hexagons
where correlations are weak and short-range. In contrast with the results
in Ref. \cite{szeto1998topological}, hexagon-hexagon is short-range
and only non-zero for adjacent cells ($t=1$) which is a signature
of microcrystal regions in the network (see Fig. \ref{fig:randnet}). If
we had used $p(n)p(n')$ instead of $p_t(n)p_t(n')$, ring-ring correlation
shows a long-range behavior due to topological effect
\cite{oguey2011long,miri2007topological} but Eq. \ref{shellprob}
correctly captures the nature of correlations in the random network.
\section{Discussion and Conclusion}
We have shown that correlations between rings in glassy networks can be treated
best if geometrical rather than topological distances between rings are used.
Using topological distances, which would be preferable, unfortunately leads to
spurious long range correlations as the topological charge for each shell around
a central ring does not approach zero at large distances, due to the non-circular
nature of the shells. These issues are absent if the geometrical distances between
the centers of rings are used. We find in this case that correlations only extend
out to about third neighbor rings, and can be described by a generalized
Aboav-Weaire law. These studies have been done on a very large computer-generated
network with periodic boundary conditions~\cite{wooten1985computer,stone1986theoretical}.
Experimental samples of bilayer of
vitreous silica are currently too small to allow for the study of longer range
correlations, but the main conclusion of the paper that geometrical rather
than topological distances should be used is expected to hold. Future studies
comparing experimental and computer-generated networks (both three-coordinated
with similar ring distributions) should help explain why different values
of $\alpha$ are obtained in these two cases~\cite{kumar2014}.
\begin{figure*}[t]
\centering
\includegraphics[width=11cm,height=11cm]{fig8.pdf}
\caption{\label{ringringcor}
Ring-ring correlation $C_t(n,n')$ versus topological distance $t$.
The correlations are short or medium range depending on the size of the
interacting rings. Although hexagons are weakly correlated with their
neighbor rings, other rings show a high degree of correlations up to
three rings away. Very similar results are obtained using geometrical
distances. Note that correlations are symmetric so that $5-6$ is
the same as $6-5$ etc. where panel (a) is for five-fold rings, panel (b)
six-fold rings, panel (c) seven-fold rings, and panel (d) eight-fold rings.}
\end{figure*}
\begin{acknowledgments}
We should thank Avishek Kumar for providing the computer-generated networks,
and David Sherrington and Mark Wilson for useful ongoing discussions. MS was
partially supported by the Arizona State University Graduate and Professional
Student Association's JumpStart Grant Program. MS was aided in this work by
the training and other support offered by the Software Carpentry project.
Support through NSF grant \# DMS 1564468 is gratefully acknowledged.
\end{acknowledgments}
|
\section{Introduction}
Pulsars are believed to be rapidly rotating neutron stars (NSs) that can emit
continuous gravitational wave (GW) radiation if their mass distributions are
asymmetric~\cite{1979PhRvD..20..351Z}. Observations from first-generation GW
detectors have placed upper limits on the amplitude of these GWs from the known
galactic millisecond pulsars. This in turn allows constraints to be placed on
the ellipticities of these NSs~\citep{2010ApJ...713..671A}. With the advanced
detector era having recently begun with Advanced
LIGO~\cite{2015CQGra..32g4001L} in operation and Advanced
Virgo~\cite{2015CQGra..32b4001A}, and KAGRA~\cite{Aso:2013} close behind, we
will soon be able to make observations of these sources with significantly
increased sensitivity.
For each pulsar with known sky location and assumed GW phasing (as inferred
from arrival times of its radio pulses), time and frequency-domain
matched-filtering
approaches~\cite{2005PhRvD..72j2002D,Pitkin:2011cl,1998PhRvD..58f3001J,2014CQGra..31p5014A,2014ApJ...785..119A}
are commonly applied. The former has been used within the LIGO-Virgo
Collaboration for the known pulsar searches and applies a Bayesian
marginalization strategy to the unknown system
parameters~\cite{2005PhRvD..72j2002D}. The latter, frequency-domain approach,
known as the $\mathcal{F}$-statistic ~\cite{1998PhRvD..58f3001J} performs an
analytical maximization of the likelihood over the unknown parameters of each
pulsar and it is this method that we make use of for the remainder of this
paper.
Combining sources to improve detection probability is an attractive
approach to weak signal detection (e.g.~detecting NS ellipticity from
analysis of the GW stochastic background~\cite{2014PhRvD..89l3008T} and
detecting gravitational wave memory using binary black hole
mergers~\cite{2016arXiv160501415L}). Since GW detectors currently study
${\sim}200$ known pulsars, the existing detection strategy for this relatively
large ensemble can be viewed as trying to detect each one separately, and then
waiting for the first detection to appear. This is certainly the most obvious
strategy to take, but not obviously the most optimal. Cutler and Schutz (CS)~\cite{2005PhRvD..72f3006C} proposed an alternative: first sum the
$\mathcal{F}$ statistic from each pulsar, and then use that sum as a new
detection statistic. In this initial study, CS used an equal weight for all the
pulsars to be combined. One issue with this approach is that including pulsars
which are likely to emit relatively weak GWs decreases the signal-to-noise
ratio (SNR) of the combined statistic. As indicated in their paper, the SNR of
the combined statistic decreases if the detection ensemble includes weak
sources where the squared SNR is less than half of the average squared SNR for
all observed pulsars. Therefore, to more efficiently detect GWs from an
ensemble of all known pulsars, it seems sensible to investigate the effects of
giving nonequal weights to the pulsars within the ensemble.
In this paper, we generalize the idea proposed by CS, by considering the
prior distribution of GW strengths from the pulsars within the ensemble. After
a brief introduction to pulsar GW emission and the $\mathcal{F}$-statistic, we
apply the general theory of hypothesis testing, and obtain a
Neyman-Pearson criterion for detecting GWs from an ensemble of pulsars. This
leads to an optimal detection statistic, which in idealized situations (i.e.,
when our prior knowledge of the signal and our model for the noise are an
accurate representation of reality) provides the highest detection probability
with a given false-alarm probability. As we show, this statistic can in
some cases be approximated by linearly combining $\mathcal{F}$-statistic values
from the ensemble of pulsars with appropriate weights.
We assume that the ellipticities of pulsars follow a common (yet unknown)
intrinsic distribution and that the orientation of their rotation axes is
isotropically distributed. We then draw on our knowledge of their sky location,
distance from the Earth, and their rotation frequency to construct prior
distributions on the expected GW amplitudes from our known pulsars. Since the
intrinsic ellipticity distribution remains unknown, we model it as a simple
exponential distribution, but perform tests using both exponential and Gaussian
distributions.
This paper is organized as follows. In Sec.~\ref{review}, we briefly review
the form of GW emission from individual pulsars and the
$\mathcal{F}$ statistic; in Sec.~\ref{f_statistic}, we introduce the optimal
statistic for a general ensemble of pulsars and discuss how it may apply to a
set of pulsars in idealized situations; in Sec.~\ref{sim}, we test our
statistic on two possible intrinsic distributions of pulsar ellipticity. We
summarize our main conclusions in Sec.~\ref{res}.
\section{Brief review of GW from known pulsars and the $\mathcal{F}$ statistic}
\label{review}
In this section, we give a brief overview of the signal model and
maximum-likelihood detection statistic for a single pulsar.
\subsection{Gravitational waveform}
For a single GW detector, the signal strain as a function of time, $h(t)$, from
\begin{align}
h(t) =\frac{16\pi^2\epsilon I f^2}{d}
\left[
\alpha_+ \tilde F_+(t)
+\alpha_\times \tilde F_\times(t)\right] {\rm cos}[\Phi(t)+\Phi_0]\,,
\end{align}
with
\begin{align}
\alpha_+&= \frac{1+\rm cos^2\iota}{2} \rm cos 2\psi +
\rm cos\iota\,\sin2\psi \,,\\
\alpha_\times&=-\frac{1+\rm cos^2\iota}{2} \rm sin 2\psi +
\rm cos\iota\,\rm cos 2\psi.
\end{align}
Here we have assumed the pulsar, at distance $d$ from the Earth, to be an
triaxial ellipsoid rotating at frequency $f$ around one of its minor axes,
which stays constant in orientation. In~\cite{1998PhRvD..58f3001J} this is the
case when the angle between the total angular momentum vector of the star and
the star's axis of symmetry is $\pi/2$.
The pulsar is nearly spherical, with a moment of inertial $I$ around its
rotation axes, $\epsilon$ is its ellipticity, given by
\begin{equation}
\epsilon=\frac{I_1-I_2}{I}
\end{equation}
with $I_1$ and $I_2$ being the two moments of inertia around the two principal axes
that are orthogonal to the rotation axis. The above four quantities
$(d,f,\epsilon,I)$ define the strength of the source as received at the
detector.
In addition, $\tilde F_{+,\times}(t)$ are the (time-dependent, due to Earth's
rotation) antenna patterns of the detector toward a source at the sky location
of the pulsar, while $\Phi(t)$ defines the GW phase evolution inferred from its
radio (or x-ray) pulsations. -both are considered known. For the type of
emission we are considering, GW radiation will be emitted at twice the rotation
frequency, $2f$, with additional modulations due to the orbital motion of the
pulsar and the motion of the detector due to the Earth's rotation and orbit.
Finally, we have the polarization angle $\psi$ , the inclination angle $\iota$
that describes the pulsar's orientation, and $\Phi_0$ an additional unknown GW
reference phase, all of which we consider as unknown.
In terms of notation, our $\iota$ and $\psi$ are the same as used
in~\cite{1998PhRvD..58f3001J}, while $\tilde F_+(t)$ and $\tilde F_\times(t)$
are respectively equivalent to $a(t)$ and $b(t)$ of~\cite{1998PhRvD..58f3001J}
where we have assumed that the angle between the two interferometer arms equals
$\pi/2$.
\subsection{The single-pulsar $\mathcal{F}$ statistic}
\label{sec:fstat}
Under the assumption that the measured strain is a combination of a GW signal
and additive detector noise $n$, with a single-sided noise special density
$S_{h}(f)$, the "near optimal" statistic is given by the so-called
$\mathcal{F}$ statistic, derived by Jaranowski, Krolak and
Schutz~\cite{1998PhRvD..58f3001J}. For point hypotheses with no uncertain model
parameters the maximum-likelihood approach of the $\mathcal{F}$-statistic is
optimal in the Neyman-Pearson sense whereby the detection probability $P_{\rm
DE}$ is maximized at fixed false-alarm probability $P_{\rm FA}$. However, even
for individual pulsar detection the signal model does include additional
unknown model parameters in which case the truly optimal approach is Bayesian
and requires marginalisation over those parameters~\cite{2008arXiv0804.1161S}.
Our investigation makes use of the $\mathcal{F}$-statistic as our input data
and hence by association also suffers from a lack of total optimality. However,
as shown in~\cite{2009CQGra..26t4013P} the reduction in sensitivity of the
$\mathcal{F}$-statistic over the fully optimal approach is slight.
For an observation time $T_{\rm obs}$, the $\mathcal{F}$-statistic satisfies a
$\chi^2$ distribution with 4 degrees of freedom (4-D) and has a noncentrality
parameter equal to the squared optimal SNR $\rho^2$, defined by
\begin{equation}
\label{eqrho}
\rho^2 =\frac{256\pi^4 \epsilon^2 I^2 f^4 \mathcal{K} }{ d^2}\frac{T_{\rm obs}}{S_h(2f)}
\end{equation}
(note that $2f$ is approximately the gravitational wave frequency) with
\begin{align}
\mathcal{K}&=\sum_{p,q=+,\times}\alpha_p\alpha_q F_{pq}\,,\nonumber\\
F_{pq} & =\frac{1}{T_{\rm obs}}\int_{0}^{T_{\rm obs}}\tilde F_p(t) \tilde F_q(t) dt\,.
\end{align}
The unknown quantities defining the optimal SNR are the ellipticity $\epsilon$
and the geometrical factors contained within $\alpha_{+,\times}$ describing
the GW polarization and orientation of the pulsar. Note that $F_{++}\neq
F_{\times\times}$ and that averaging over many sidereal days leads to
$F_{+\times} \rightarrow 0$ and so such terms can be ignored.
For $\alpha_{+,\times}$ we shall assume that $\iota$ and $\psi$ are distributed
according to a random orientation of the pulsar's rotation axis. In this case,
points with coordinates $(\alpha_+,\alpha_\times)$ are distributed on the two-dimensional
plane axisymmetrically around the origin, with modulus
\begin{equation}
\zeta \equiv \alpha_+^2+\alpha_\times^2 = \frac{1+6\,\rm cos^2\iota+\rm cos^4\iota}{4}
\end{equation}
and $\rm cos\,\iota$ uniformly distributed between $-1$ and $+1$. We can write
\begin{align}
\label{eqK}
\mathcal{K} =\zeta
\bigg[
F_{++}{\rm cos}^2(2\tilde\psi) +
F_{\times\times}{\rm sin}^2(2\tilde\psi)
\bigg]
\end{align}
with $\tilde\psi$ related to $\psi$ by an offset,
\begin{equation}
4\tilde\psi = 4\psi - \arctan \frac{4\rm cos\iota(1+\rm cos^2\iota)}{\rm sin^4\iota}
\end{equation}
hence uniformly distributed between 0 and $2\pi$. In this paper, we
simply generate an ensemble of binaries using uniformly distributed
$\rm cos\,\iota$ and uniformly distributed $\psi$. The average of $\mathcal{K}$
over this ensemble is given by
\begin{equation}
\langle \mathcal{K} \rangle = \frac{\langle \zeta \rangle}{2} (F_{++} +F_{\times\times}) =\frac{2}{5}(F_{++} +F_{\times\times})
\end{equation}
It was shown by CS that for the detection of a single pulsar in a network of
$M$ detectors, the $\mathcal{F}$-statistic still satisfies a 4-D $\chi^2$
distribution with a noncentrality parameter $\rho_{\text{net}}^2 = \sum_i^M
\rho_i^2$, where $\rho_i^2$ is the optimal single detector SNR as defined in
Eq.~\ref{eqrho}.
\subsection{Scaling of detectability with observation time}
\label{subsecthreshold}
In our idealized treatment with Gaussian noise, the significance of
detection only depends on the noncentrality parameter $\rho^2$, which is
proportional to the observation time $T_{\rm obs}$. For this reason, the $T_{\rm
obs}$ required for a detection with a particular confidence level is
inversely proportional to $\mathcal{K}$ and $\epsilon^2$, or
\begin{equation}
T_{\rm det} = \frac{S_h(2f) d^2}{256\pi^4\epsilon^2 I^2 f^4 \mathcal{K}}\rho_*^2
\end{equation}
with $\rho_*$ being a threshold (or a sensitivity level for $\rho$)
determined by the desired false-alarm probability ($P_{\rm FA}$) and detection
probability ($P_{\rm DE}$), as we discuss below.
Let us follow a frequentist approach of hypothesis testing. Suppose $X$ is
our detection statistic, which is either a 4-D $\chi^2$ distribution or a 4-D
noncentral $\chi^2$ distribution with noncentrality parameter $\rho^2$. Let
us first impose a detection threshold $X_{\rm th}$ on $X$, so that $P[X>X_{\rm
th}|\rho^2=0]=P_{\rm FA}$, which leads to
\begin{equation}
\left(1+\frac{X_{\rm th}}{2}\right)e^{-\frac{X_{\rm th}}{2}}=P_{\rm FA}\,.
\end{equation}
where the threshold $X_{\rm th}$ is determined implicitly from $P_{\rm FA}$.
If $X$ now has a nonzero $\rho^2$, its probability of overcoming the threshold
becomes the detection probability, or
\begin{equation}
\label{Pde}
P_{\rm DE} = P[X>X_{\rm th}|\rho^2]\,.
\end{equation}
The threshold $\rho_*^2$ is determined by requiring that when
$\rho^2\ge\rho_*^2$, Eq.~\eqref{Pde} provides a significant $P_{\rm DE}$.
\section{The detection statistic of multiple pulsars}
\label{f_statistic}
In this section we extend the single-pulsar analysis approach of
Sec.~\ref{subsecthreshold} to apply to the detection of GWs from an ensemble of
pulsars.
\subsection{General theory}
To formulate how we might detect a combination of $n$ nearby sources, let us
consider the general problem of distinguishing the distribution of $n$ random
variables, $(X_1,\ldots,X_n)$ , between two probability densities $p_A$ and
$p_B$. Suppose we have a region $\mathcal{V}$, and we claim $A$ if
$(X_1,\ldots,X_n)\in \mathcal{V}$, and $B$ otherwise. In the context of GW
detection $A$ is without signal, while $B$ is detection. In this way, our
false-alarm probability is
\begin{equation}
P_{\rm FA} = \int_{\bar{\mathcal{V}}} p_A(x_1,\ldots,x_n) dx_1\ldots dx_n,
\end{equation}
where $\bar{\mathcal{V}}$ represents not being within the region $\mathcal{V}$,
and our detection probability is
\begin{equation}
P_{\rm DE} = \int_{\bar{\mathcal{V}}} p_B(x_1,\ldots,x_n) dx_1\ldots dx_n.
\end{equation}
We then have to find the region $\mathcal{V}$ for which $P_{\rm DE}$ is
maximized given $P_{\rm FA}$. It is possible to find that the boundary of
$\mathcal{V}$ should be given by
\begin{equation}
\label{eqbound1}
\mu p_A (x_1,\ldots,x_n)= p_B (x_1,\ldots,x_n)\,.
\end{equation}
This is an implicit formula: given different values of the Lagrange multiplier
$\mu$, we arrive at regions that have particular pairs of ($P_{\rm
FA},P_{\rm DE}$). For each pair, the detection probability is the maximum
possible value given $P_{\rm FA}$. Operationally, the boundaries of all these
$\mathcal{V}$'s are given by surfaces specified by Eq.~\eqref{eqbound1}. In
other words, for data $X_{1,\ldots,n}$, if we define the likelihood ratio
\begin{equation}
\mathcal L =\frac{p_B(X_1,\ldots,X_n)}{p_A(X_1,\ldots,X_n)}
\end{equation}
as a detection statistic, and by imposing a threshold, we obtain the best
$P_{\rm DE}$ with given $P_{\rm FA}$.
If we have various versions of $B$ parameterized by a set of parameters
$\boldsymbol{\theta}$, we can further average over these possibilities with
their prior probability distributions $w(\bg{\theta})$, such that
\begin{equation}
P_{\rm DE} = \int d\bg{\theta} w(\bg{\theta} ) \int_{\bar{\mathcal{V}}} p_B(x_1,\ldots,x_n;\bg{\theta} ) dx_1\ldots dx_n.
\end{equation}
This simply arrives at modified boundaries of $\mathcal{V}$ given by
\begin{equation}
\label{eqbound}
\mu p_A (x_1,\ldots,x_n)= \int w(\bg{\theta} ) p_B (x_1,\ldots,x_n;\bg{\theta} ) d\bg{\theta} \,
\end{equation}
meaning that
\begin{equation}
\mathcal L =\frac{\displaystyle\int w(\bg{\theta} )p_B(X_1,\ldots,X_n;\bg{\theta} )d\bg{\theta} }{p_A(X_1,\ldots,X_n)}.
\end{equation}
This is in fact the same as the marginal likelihood ratio ( the Bayes
factor in a Bayesian approach) for obtaining the data $X_{1,\ldots,n}$ ---
therefore we have simply established the optimality of the Neyman-Pearson
approach in our case.
\subsection{Multiple pulsars}
In the detection of multiple pulsars, let us consider $A$ to be $n$ independent
4-D $\chi^2$ distributions, and $B$ to be $n$ independent 4-D noncentral
$\chi^2$ distributions, with noncentrality parameter $\lambda_1$, \ldots,
$\lambda_n$ (for simplicity, we use $\lambda$ rather than the optimal SNR
$\rho^2$). Recall that for a $k$-D noncentral $\chi^2$ distribution, we have
\begin{equation}
p_{(k,\lambda)}(x) = \frac{1}{2}e^{-(x+\lambda)/2}
\left(\frac{x}{\lambda}\right)^{k/4-1/2} I_{k/2-1}(\sqrt{\lambda x})\,,\quad x>0\,.
\end{equation}
where $I$ is the modified Bessel function of the first kind. We can then write
\begin{equation}
p_A (x_1,\ldots,x_n) = p_{(4,0)}(x_1) \cdots p_{(4,n)}(x_n)
\end{equation}
and
\begin{equation}
p_B (x_1,\ldots,x_n) = p_{(4,\lambda_1)}(x_1) \cdots p_{(4,\lambda_n)}(x_n)
\end{equation}
Following Eq.~\ref{eqbound1}, for fixed values of $\lambda_1$, \ldots
$\lambda_n$, we have
\begin{equation}
\prod_{j=1}^n \frac{ 2e^{-\lambda_j/2} I_1\left(\sqrt{\lambda_j x_j}\right)}{\sqrt{\lambda_j x_j}} =\mu
\end{equation}
as optimal boundaries of $\mathcal{V}$ which can also be written as
\begin{equation}
\sum_j \log \left[\frac{I_1\left(\sqrt{\lambda_j x_j}\right)}{\sqrt{\lambda_j x_j}}\right] = \mbox{const}\,.
\end{equation}
This shows how signals should be combined resulting in our combined detection
statistic
\begin{equation}
\label{optfixed}
\mathcal L^{\rm opt}_{\rm fix}= \sum_j \log
\left[\frac{I_1\left(\sqrt{\lambda_j X_j}\right)}{\sqrt{\lambda_j X_j}}\right]
\end{equation}
where $X_{1,\ldots,n}$ are the $n$ observables.
If each $\lambda$ depends on a set parameters $\boldsymbol{\theta}$, and for
each $j$ there is a corresponding prior distribution $w_j(\bg{\theta})$, then
from Eq.~\eqref{eqbound}, we can write
\begin{equation}
\label{eqoptglobal}
\mathcal L^{\rm opt}= \sum_j \log\left[ \int w_j(\bg{\theta}) e^{-\lambda(\bg{\theta})/2} \frac{I_1\left(\sqrt{\lambda(\bg{\theta}) X_j}\right) }{\sqrt{\lambda(\bg{\theta}) X_j}}d\bg{\theta}\right].
\end{equation}
As a sanity check, if $w_j(\lambda)=\delta(\lambda-\bar\lambda_j)$, we recover
the previous result.
\subsection{Special case: exponential distribution}
\label{subsec:exp}
We can further simplify the construction of the optimal statistic, simply
and arbitrarily assuming that each $\lambda_j$ value is drawn from an
exponential distribution, or
\begin{equation}
w_j (\lambda) = \frac{1}{\bar\lambda_j}e^{-\lambda/\bar\lambda_j} \,,\quad \lambda>0\, ,
\end{equation}
where $\bar{\lambda}$ is the mean value of the prior distribution on $\lambda$
for each pulsar. In this case, we obtain the following closed-form
expression,
\begin{equation}
\label{lexpopt}
\mathcal{L}_{\rm exp}^{\rm opt} = \sum_j \log\left[\frac{e^{Y_j}-1}{Y_j}\right]
\end{equation}
with
\begin{equation}
Y_j = \frac{\bar\lambda_j}{\bar\lambda_j+2}\frac{X_j}{2}
\end{equation}
This is quite interesting: those sources with $\bar\lambda_j \gg 2$ (already
quite detectable individually), should be combined with a similar weight, while
those much less than unity should be combined according to the expectation
value of the noncentrality parameter, $\bar{\lambda}_j$. The latter case is
discussed further below.
\subsection{Special case: weak-signal limit}
A different way to obtain an optimal statistic is to directly assume that we
should linearly combine the $\mathcal{F}$-statistic according to
\begin{equation}\label{lin_opt}
\mathcal{L}^{\rm wopt}_{\rm lin} =\sum_j \alpha_j X_j\,,
\end{equation}
and optimize the ``signal-to-noise ratio'', which is given by the increase of
$\langle\mathcal{L}^{\rm wopt}_{\rm lin}\rangle$ due to nonzero $\lambda_j$
divided by the variance of $\mathcal{L}^{\rm wopt}_{\rm lin}$ in the absence of
signal. This leads to
\begin{align}
\alpha_j & \propto \bar{\lambda}_j \nonumber \\
&=\frac{f_j^4 \langle \mathcal{K}_j \rangle}{d_j^2 S_{h}(2f_j)},
\end{align}
where the second line is valid for the known pulsars case, if we assume the intrinsic parameter of pulsars follows the same distribution (see the discussion in Sec. \ref{sim_known} ).
This can be derived from the optimal statistic, if we assume that we are
interested in the low signal amplitude limit where the $\lambda_j$ are small.
In this case, we can Taylor expand Eq.~\eqref{eqoptglobal} and obtain, at
leading order
\begin{equation}
\mathcal{L}^{\rm wopt}_{\rm lin} \approx \sum_j \frac{(X_j-2)\bar{\lambda}_j}{4}
\end{equation}
which is equivalent to using
\begin{equation}
\mathcal{L}^{\rm wopt}_{\rm lin} \approx \sum_j \alpha_j X_j ,
\end{equation}
which is also consistent with Eq.~\eqref{lexpopt} when $\bar\lambda_j$ is small.
This implies that if we could tolerate a high false-alarm probability by
setting our threshold low, it is plausible that combining the observables
proportional to the (prior) expectation value of noncentrality parameters
would be optimal. However, as shown in Sec.~\ref{sim}, in the situations we
encounter, this approximation is not quite valid.
\subsection{Comparison with individual pulsar detection}
Before we compare our strategy with existing strategies that do not combine
signals from multiple pulsars, let us first clarify what it means to ``not
combine signals''. A careful examination provides two possible variants.
\subsubsection{Assigning equal false-alarm probability to each pulsar}
\label{subequalth}
The first approach regards treating each pulsar as truly independent, and by
setting the same false-alarm probability for each pulsar --- even though each
pulsar is not equally likely to provide detection. In this procedure, we
therefore set the same threshold $X^{\rm th}$ for each pulsar, requiring
\begin{equation}\label{fa_individual}
1- P^n(X<X^{\rm th}|\rho^2=0)= P_{\rm FA}
\end{equation}
and leading to the following total detection probability
\begin{equation}
P_{\rm DE}=1 -\prod_{j=1}^n P(X<X^{\rm th}|\lambda_j)
\end{equation}
of detecting at least one pulsar within this ensemble.
\subsubsection{Assigning false-alarm probability according to signal strength}
\label{subvaryth}
This is clearly problematic since we have potentially $\sim 200$ pulsars ---
assigning the same false-alarm value to pulsars with dramatically different
potential signal strength is clearly wasteful. If a different threshold is set
for each pulsar, in such a way that the detection probability of an ensemble is
maximum, we then require
\begin{align}
\label{muthreshold}
\mu = &\int w_j(\bg{\theta})\frac{p_{(4,\lambda_j)}}{p_{(4,0)}} d\bg{\theta} \nonumber\\
=&\int w_j(\bg{\theta})\frac{e^{-\lambda(\bg{\theta})/2} I_1\left(\sqrt{\lambda(\bg{\theta}) X_j^{\rm th}}\right)}{\sqrt{\lambda(\bg{\theta}) X_j^{\rm th}}} d\bg{\theta}
\end{align}
where $\mu$ is a constant independent of $j$. As we vary $\mu$, we obtain a
varying set of $X_j^{\rm th}$ that would provide us with the optimal thresholds
for each $X_j$, such that the total detection probability of detecting a GW
signal within this ensemble is maximum given the false-alarm probability [as
defined in Eq.~\ref{fa_individual} with different $X^{\rm th}$ ].
\begin{figure}
\centerline{\includegraphics[width=0.5\textwidth]{com_lambda_pde_2016}}
\caption{(Color Online.) Detection probability for the model given by
Eq.~\eqref{lambdasample}, using the optimal combined statistic (com-opt) and
individual thresholds (ind-subopt for equal thresholds, see
Sec.~\ref{subequalth}, and ind-opt for optimal thresholding, see
Sec.~\ref{subvaryth}). We have fixed $P_{\rm
FA}=0.01$.\label{fig:fixedlambda}}
\end{figure}
\section{Monte Carlo Simulations of Simple Models}
\label{sim}
In this section, we perform numerical investigations of two simple models. In
particular, we study the case of constant $\lambda_j$ first, then the
case where the $\lambda_j$ values follow exponential distributions. This provides important basic understanding before we move on to the known pulsars.
\subsection{Constant $\lambda_j$ }
In this section, we perform Monte Carlo simulations for signals with
fixed $\lambda_j$ -- the simplest case. We shall compare four strategies: (i)
imposing a constant threshold on all $X_j$ [Sec.~\ref{subequalth}], (ii)
imposing a variable threshold on $X_j$, according to Eq.~\eqref{muthreshold}, (iii) using a linear-combination statistic
\begin{equation}
\label{eqlincom}
\mathcal{L}_{\rm lin} = \sum_{j=1}^n\alpha_j^\beta X_j
\end{equation}
with various values of $\beta$, and (iv) using the optimal statistic, according to Eq.~\eqref{optfixed}.
\begin{figure}
\centerline{\includegraphics[width=0.5\textwidth]{com_beta_pde_2016}}
\caption{(Color Online.) Detection probability of linear combination
statistics [Eq.~\eqref{eqlincom}] compared to the optimal detection
probability. Different traces correspond to $\lambda_0$ ranging from 3 to 10,
and we have fixed $P_{\rm FA}=0.01$. \label{fig:beta}}
\end{figure}
We have chosen
\begin{equation}
\label{lambdasample}
\alpha_j=\lambda_j = \frac{\lambda_0}{j},\quad j=1,\ldots,8,
\end{equation}
which is designed to simulate an ensemble of sources that are distributed
on a two-dimensional plane. If, within each disk with radius $r$, the number of sources
is proportional to $r^2$, then for the $N$-th source, its distance should
be ${\sim}\sqrt{N}$; therefore the noncentrality parameter should be ${\sim}1/N$.
As we vary $\lambda_0$ from 3 to 15, and fixing $P_{\rm FA}=0.01$, we compare
the detection probability. As is shown by Fig.~\ref{fig:fixedlambda}, the
optimal strategy is substantially better than strategies (i) and (ii). In
particular, in order for (i) and (ii) to achieve 50\% detection probability,
the noncentrality parameter must be a factor of $\sim 2$ stronger.
In Fig.~\ref{fig:beta}, we investigate the performance of the
linear-combination statistics. For $P_{\rm FA}=0.01$, we plot $P_{\rm DE}$ as
a function of the index $\beta$. It seems here that $\beta\sim 0.5$ performs
slightly better than $\beta \sim 1$, although the optimal $\beta$ value depends
on $\lambda_0$, and is located somewhere between 0.5 and 0.8.
\subsection{Exponential distributions for $\lambda_j$}
Let us now consider $\lambda_j$ values that have simple exponential prior
distributions for which we have analytical formulas derived in Sec.
~\ref{subsec:exp}. This is also important because we can test whether
having the correct prior information in constructing the detection statistic
can significantly affect detection efficiency. In particular, while the
optimal statistic seems highly dependent on the prior distribution of
$\lambda$, the linear statistic $\mathcal{L}_{\rm lin}$ is robust against a
rescaling of the distributions of all $\lambda_j$.
\begin{figure}
\includegraphics[width=0.5\textwidth]{lambda_pde_2016}
\caption{Detection probability for models with $\lambda_j$
following exponential distributions with mean value given by
Eq.~\ref{lambdasample}. We have fixed $P_{\rm FA}=0.01$. Shown here are from
the optimal statistic (com-opt), optimal statistic scaling prior distributions
by 10 (com-opt-10p) and by 1/10 (com-opt-01p), linear-combination statistic
with $\beta=1/2$ (com-lin), individual pulsar detection with optimal
thresholding on each $X_j$ (ind-opt) and individual pulsar detection with
common threshold (ind-subopt).
\label{fig8exp}}
\end{figure}
Again, to be concrete, we chose to have $\lambda_j$'s follow exponential
distributions, with mean values given by Eq.~\ref{lambdasample}. The detection
probability with $P_{\rm FA}=0.01$ is shown in Fig.~\ref{fig8exp} for
$\lambda_0$ ranging from 3 to 30. Here, we see again that the optimal statistic
is substantially better than individually detecting the pulsars --- while a
more strategic thresholding allows some improvement.
In this case, we can see the potential benefits of the linear statistic: when
the wrong prior distributions are used (with $\bar\lambda_j \rightarrow
10\bar\lambda_j$ and $\bar\lambda_j \rightarrow \bar\lambda_j/10$) to compute
the optimal statistic, the detection efficiency drops to a level worse than
using the linear statistic, which is independent of an overall rescaling of
all $\lambda_j$ values.
\begin{figure}
\includegraphics[width=0.5\textwidth]{n_lambda_2016}
\caption{Values of signal strength ($\lambda_0$) at which each
detection strategy can yield detection probability of 50\% (red symbols) and
90\% (blue symbols), respectively, as a function of the number of sources, $N$.
We have used the optimal statistic (stars), linear-combination statistic with
$\beta=0.5$ (square), uniform threshold for all sources (hollow circles) and
optimal thresholding (solid circles). Increase of $\lambda$ with $N$ in the
uniform threshold case indicates that including more sources introduces
contamination from weaker sources. We assumed exponential distributions for
$\lambda_j$ in this plot and have fixed $P_{\rm FA}=0.01$.
\label{fig:scaling}}
\end{figure}
\subsection{Scaling with the number of sources}
Let us now consider how the detection probabilities of the various schemes
scale with the number of sources. We do this by simply extending
Eq.~\eqref{lambdasample} to include a variable number of sources N.
In Fig.~\ref{fig:scaling}, we can see that as the number of sources increases
the detection probability of the optimal and linear combination statistics with
$\beta=0.5$ also increases. The detection probability of individual
pulsars using a common threshold decreases, while the individual detection with
optimal thresholding also keeps increasing, but stops increasing at a
relatively low number of sources. This can be explained as being due to the combined
statistics' ability to incorporate weaker sources without sacrificing
sensitivity.
Numerically, we can see that a substantially larger signal strength has to be
present for the individual detection strategies. In addition, we emphasize
that the linear-combination statistic, here shown to be very close to being
optimal, is independent from an overall rescaling of the distribution of
$\lambda_j$'s. The optimal thresholding, on the other hand, does depend on the
particular model of $\lambda$.
\section{Monte Carlo Simulations for Known Pulsars}\label{sim_known}
We now discuss the case of detecting GWs from multiple known pulsars. We
start by describing the known and unknown aspects of these sources, and
then present the setup and conclusions of our numerical simulations.
\subsection{Known pulsars: prior distributions for $\lambda_j$.}
For the case of multiple pulsars, the noncentrality parameter $\lambda_j$ for
each pulsar in a single detector is simply equal to $\rho_j^2$, as given by
Eq.~\eqref{eqrho}. We now discuss in detail all factors contributing to
our prior knowledge of $\rho_j^2$.
The ellipticity $\epsilon_j$ crucially defines the level of quadrupole
deformation of the NS. At present, we have only theoretical constraints based
on the internal structure of NSs, which span a wide range, and observational
upper limits from from previous GW searches, which span the range $\sim
10^{-7}-10^{-2}$~\citep{2010ApJ...713..671A}. Our baseline assumption is that
the $\epsilon$ of all pulsars follows a common (yet unknown) distribution; this
could be motivated as arising from the belief that all these eccentricities
were generated by the same physical mechanism. We note that it is plausible for
Advanced LIGO to detect at the level of $\overline{\epsilon} \sim
\mbox{few}\times 10^{-8}$.
The geometrical factor $\mathcal{K}$ depends on the inclination angle $\iota$,
polarization angle $\tilde\psi$, and antenna patterns $F_{++}$ and
$F_{\times\times}$, see Sec.~\ref{sec:fstat}. We assume no knowledge
concerning the orientation of the pulsar, therefore uniformly distributing
$\rm cos\iota$ between $-1$ and $+1$, and uniformly distributing $\tilde\psi$
between $0$ and $2\pi$. As for $F_{++}$ and $F_{\times\times}$, they further
depend on the geographical location and orientation of the detector, as well as
the source's declination angle (the right ascension dependence is averaged away
after many sidereal days observation).
As noted by CS, for the network of $M$ detectors case, the noncentrality
parameter of each pulsar is simply
\begin{equation}
\lambda_{j}^{\rm net} =\sum_i^M \lambda_{ji},
\end{equation}
\subsection{Simulations and results}
Our simulations assume one year of observation using the network of Advanced
LIGO and Virgo at design sensitivity~\footnote{LIGO Document
T1200307-v4}~\footnote{LIGO Document T1300121-v1}. The positions and
orientations of the detectors are taken from Table 1
of~\cite{1998PhRvD..58f3001J}. The known pulsar parameters (distance, sky
location and frequency) used to compute $\lambda_{j}^{\rm net}$ are taken from
the 195 known pulsars analysed in the initial detection era, and we assume that
the moment of inertia $I=10^{38} \, \rm{kg \,m^2}$~\citep{2014ApJ...785..119A}.
To provide a proof-of-principle of our proposed method, we assume (i)
$\epsilon^2$ values follow exponential distributions with two different rate
parameters $4\times10^{-16}$ and $9\times10^{-16}$, and (ii) $\epsilon$ values
follow normal distributions with mean values of $1.5\times10^{-8}$, and
$2\times10^{-8}$ with standard deviations equal to half of their respective
mean values.
We have performed simulations to test the detection efficiency of our proposed
robust statistic ${\mathcal L_{\rm lin}}$ (Eq.~\ref{eqlincom}, see discussion
in~\ref{sim}) via the receiver operating characteristic (ROC) curve, which is a
parametric plot of the probability of false alarm versus the probability of
detection.
The ROC curve is constructed using $10^5$ simulations of $X_j$ with
noncentrality parameter $\lambda_{j}^{\rm net}$ and $10^5$ noise only
simulations. With the assumption of $\epsilon^2$ following the same
distribution, the $\alpha_{j}^{\rm net}$ is used in
Eq.~\ref{eqlincom} to compute every simulated $\mathcal L_{\rm lin}$ in place of $\alpha_{j}$ for the $M$ detectors case, defined as
\begin{equation}
\alpha_{j}^{\rm net} =\sum_i^M \alpha_{ji}=\sum_i^M \frac{f_j^4 \langle \mathcal{K}_j \rangle}{d_j^2 S_{hi}(2f_j)}.
\end{equation}
We compare the detection efficiencies of ensemble based
strategies including the weighted-combination ($\beta=0.5$) and
equal-combination ($\beta=0$, the CS case) method, with the individual
pulsar detection strategy including the expected brightest [the largest value
of $\frac{f^4 \langle \mathcal{K} \rangle}{d^2 S_h(2f)}$], measured brightest
(the maximum $\rho_j^2$ in each simulation) case.
\begin{figure*}[t]
\begin{center}
\begin{tabular}{cc}
\includegraphics[angle=0,width=3in]{2015_s200_roc_a1_e_dis_exp_2e8_randndist_0_subW.pdf}
\includegraphics[angle=0,width=3in]{2015_s200_roc_a1_e_dis_exp_2e8_randndist_0_subE.pdf}
\end{tabular}
\end{center}
\caption{ROC curves for different detection methods. WA (EA), W3 (E3), W6
(E6), W50 (E50), N1 and M1 correspond to weighted (equal) combinations of all,
the expected brightest three, the expected brightest six, the expected
brightest fifty, the expected brightest and the measured brightest
pulsar(s) respectively. The $\epsilon^2$ in injections were drawn from} an exponential
distribution with rate parameter $2\times10^{-16}$.
\label{roc_moresource_exp}.
\end{figure*}
\begin{figure*}[t].pdf \begin{center}
\begin{tabular}{cc}
\includegraphics[angle=0,width=3in]{2015_s200_roc_a1_e_dis_nor_15e8_randndist_0_subW.pdf}
\includegraphics[angle=0,width=3in]{2015_s200_roc_a1_e_dis_nor_15e8_randndist_0_subE.pdf}
\end{tabular}
\end{center}
\caption{Same as Fig.~\ref{roc_moresource_exp}
but with $\epsilon$ following the Gaussian distribution with mean value of
$1.5\times10^{-8}$, and half of mean values as its variance value.
\label{roc_moresource_nor}}.
\end{figure*}
The results are first presented in terms of answers to the following two
questions: Will collecting more pulsars return higher $P_{\rm DE}$ values than
an individual detection? How many sources should be combined to obtain the
maximum $P_{\rm DE}$ at given $P_{\rm FA}$? As shown in
Figs.~\ref{roc_moresource_exp} and~\ref{roc_moresource_nor} for various
ellipticity distributions, the more sources are combined, the higher $P_{\rm
DE}$ is for our proposed robust statistic ${\mathcal L_{\rm lin}}$ with
$\beta=0.5$, although combining the weakest part of the population (e.g. the
weakest 50 sources) will not greatly contribute to $P_{\rm DE}$. As expected,
we find that the $P_{\rm DE}$ increases when combining the first few high
amplitude sources, and then decreases for the equal-weight method ($\beta=0$)
as more and more weak sources are added to the combination. These results are
consistent with the simple test in Sec.~\ref{sim}. Since we do not know the
true values of all pulsar parameters, it is interesting to ask whether the
measured brightest source or the expected brightest source would be more
detectable than any other ensemble of sources.
Since the weighted-combination method is optimized for the whole population of
GW signals, neither the measured brightest source nor the expected brightest
one is more detectable than the whole population. This is not the case for the
equal-weight combination method (see Figs.~\ref{roc_moresource_exp} and
\ref{roc_moresource_nor}).
As shown in Fig.~\ref{roc}, our proposed weighted-combination method includes
the known information of all sources and detectors, therefore combining all
sources should yield a higher $P_{\rm DE}$ compared to other methods. In the
case of $\overline{\epsilon^2}= 2\times10^{-16}$, given $P_{\rm FA}=0.0001$,
the $P_{\rm DE}{\sim}0.9$ for the weighted-combination method is a factor of
${\sim}2$ to $4$ more sensitive than other methods (see top-left panel of
Fig.~\ref{roc}). The improved performance of the weighted-combination method
over other methods appears to be independent of the ellipticity distribution
types and distribution parameter values used in our simulations.
The typical pulsar distance measurement error is ${\sim}20\%$ but could be up
to a factor of 2-3 larger~\cite[e.g.][]{1993ApJ...411..674T}. To test the
robustness of our proposed method, we test our sensitivity to distance
uncertainty by drawing our pulsar distances from Gaussian distributions with
mean values equal to the best estimated distance and with a standard deviation
equal to $20\%$ of the mean. As shown in Fig.~\ref{roc_d}, the distance
uncertainties do not change the general performance of all methods: our
proposed weighted-combination method still is the most efficient method and
improves the $P_{\rm DE}$ by a factor of $\sim 1.5$ to $4$ compared to
different methods and given $P_{\rm FA}=0.0001$. The level of improvement
decreases when the GW signals become stronger.
\begin{figure*}
\begin{center}
\begin{tabular}{cc}
\includegraphics[angle=0,width=3in]{2015_s200_roc_a1_e_dis_exp_2e8_randndist_0.pdf}
\includegraphics[angle=0,width=3in]{2015_s200_roc_a1_e_dis_nor_15e8_randndist_0.pdf}\\
\includegraphics[angle=0,width=3in]{2015_s200_roc_a1_e_dis_exp_4e8_randndist_0.pdf}
\includegraphics[angle=0,width=3in]{2015_s200_roc_a1_e_dis_nor_2e8_randndist_0.pdf}\\
\end{tabular}
\end{center}
\caption{ROC curves of the weighted-combination (WA), equal-combination (EA),
the expected brightest (N1) and the the measured brightest (M1) detection
methods for $10^5$ signal injection simulations and $10^5$ noise only
simulations. WA and EA correspond to weighted and equal combinations of all
pulsars, respectively. The intrinsic parameter in injections was drawn from two
distributions: (i) $\epsilon^2$ follows an exponential distribution with
rate parameter $2\times10^{-16}$ (top left panel) and $4\times10^{-16}$
(bottom-left panel), and (ii) $\epsilon$ follows a Gaussian distribution
with mean value of $1.5\times10^{-8}$ (top-right) and $2\times10^{-8}$
(bottom right) with standard deviations equal to half of the mean value.
Detection probability ratios (dashed lines) are shown in each panel.}
\label{roc}.
\end{figure*}
\begin{figure*}
\begin{center}
\begin{tabular}{cc}
\includegraphics[angle=0,width=3in]{2015_s200_roc_a1_e_dis_exp_2e8_randndist_2.pdf}
\includegraphics[angle=0,width=3in]{2015_s200_roc_a1_e_dis_nor_15e8_randndist_2.pdf}
\\
\end{tabular}
\end{center}
\caption{The same as for the top panels of Fig.~\ref{roc} but taking into
account the measured distance error. The measured distance error effect is
included by drawing "true" pulsar distances from a normal distribution with
mean value equal to their current best estimates and a standard deviation of
$20\%$ of those mean values.\label{roc_d}}.
\end{figure*}
%
\section{Discussion}\label{res}
We have proposed a novel weighted-combination detection statistic for GWs from
an ensemble of known pulsars. The aim of this approach is to improve the
detection efficiency of GWs over that of individual pulsar detection based on
the ${\mathcal F}$-statistic applied to single pulsars. The general argument
behind the combination detection strategy is that a group of sources should be
more detectable than an individual one if they share certain characteristics.
We have shown that our general optimal statistic for the weighted combination
of GW signals outperforms all other approaches.
We have shown that to more efficiently detect GW signals emitted from a
ensemble of pulsars, each source within the ensemble could be assigned a
different detection statistic threshold based on the expected signal strength.
Furthermore, by assuming that the SNRs of all sources are constant or follow
exponential distributions, we have shown that the linearly weighted-combination
statistic is very close to being optimal and is robust to the choice of prior
SNR distributions. These analytic and simple Monte Carlo test predictions are
consistent with results obtained from simulations of known pulsars.
We have also used the ROC function to determine the sensitivity of a range of
possible search strategies where the detection probability between
approaches is compared as a function of false-alarm probability. To
demonstrate the performance of the new weighted-combination detection method
for the Advanced detectors era, we have compared the detection efficiency
of the linearly weighted-combination method versus the equal-combination and individual
detection method. We have done this by simulating GW signals emitted from the 195 known pulsars
within the sensitive frequency band of Advanced LIGO and Virgo. We assume that the
intrinsic pulsar parameter ellipticity $\epsilon^2$ follows a common
distribution in these simulations. The true form of the ellipticity
distribution and its associated parameters are unknown. We have chosen to use
both exponential and Gaussian distributions with mean values corresponding to
ellipticities $\epsilon \sim 10^{-8}$, a value consistent with the initial GW era
nondetection of pulsar signals and a possible advanced era detection. In
general, the combination methods return better detection efficiency than a
method that simply considers the closest or brightest pulsar. Being
consistent with results of simple Monte Carlo tests, the most efficient
method in simulations for known pulsars involves combining all known pulsars
with weights $\propto \overline{\rho}$, the expected value of the optimal SNR
of each pulsar. For the specific case where $\overline{\epsilon} \sim
1.5\times 10^{-8}$, for one year observation of the Advanced detector network,
we find that $P_{\rm DE}{\sim}0.95$ given $P_{\rm FA}=0.0001$. In this case, the
improvement by our proposed combined method could be up to a factor of $\sim 4$
compared with other methods. These results are consistent with the case of
taking into account the measurement errors of pulsar distances.
An important feature of the proposed combination method is that it is very
flexible. Using the new method it is simple to include more observed pulsars
or updated source information (e.g. distance or orientation parameters) ,
without recalculating any individual detection ${\mathcal F}$-statistic
values. However, we would expect that a fully Bayesian approach for combining
all known pulsars may be more sensitive albeit at an increased computational
cost.
The flagship known pulsar analysis within the GW community is a Bayesian
approach~\cite{2005PhRvD..72j2002D,Pitkin:2011cl,2009CQGra..26t4013P,2014CQGra..31f5002W}.
We note that it is likely that a comprehensive Bayesian approach to combining
all known pulsars into a single analysis may produce a truly optimal result.
Besides all of the information discussed above, one could also consider the
uncertainty of the major assumption (model) of this work: that all pulsars'
ellipticity values follow a common but unknown distribution. A hierarchical Bayesian approach would allow us to naturally investigate the true priors governing the distribution. In this case the form of the prior would be represented as a possible model and the parameters governing that distribution would be the ``hype'' parameters of that model. We could also apply Bayesian model selection to distinguish between different prior distributions e.g. exponential vs Gaussian or power law, etc... However, it is unclear how constraining such an analysis would be and we hope to tackle this problem in future studies.
Beyond the detection of GWs emitted by a ensemble of pulsars, the posterior probability of
all parameters could be output from a Bayesian approach. In future studies we
hope to investigate such a Bayesian application to the detection of GWs from
the ensemble of known pulsars.
\acknowledgments
We would like to acknowledge valuable input from our anonymous referee, M. Pitkin and G. Woan, whose input has greatly
improved the manuscript. XF
acknowledges financial support from National Natural Science Foundation of
China (grant No.~11303009 and 11673008). XF is a Newton Fellow supported by the Royal
Society. YC is supported by NSF grants PHY-1404569 and C.~M. is
supported by a Glasgow University Lord Kelvin Adam Smith Fellowship and the
Science and Technology Research Council (STFC) grant No. ST/ L000946/1.
|
\section{Introduction}
Dark matter constitutes the most abundant type of matter in our universe and its density is now experimentally well-determined $\Omega_{CDM}h^2= 0.1199 \pm 0.0027$ \cite{cdm}. Worldwide efforts to constrain its nature and interactions have led the community to a puzzling situation where null results coexist with direct detection experiments that find high significance excesses \cite{dama}. In particular, in the low mass region of DM candidates i.e. $m_{\chi}<1\; \rm GeV/c^2$, cosmological, astrophysical and collider constraints seem to be the most important, see for example a discussion in \cite{lin}. Direct detection searches of thermalized galactic DM are mostly based on nuclear recoils on selected targets. In this scenario, LDM particles with masses much smaller than that of the nucleon, $m_\chi \ll m_N$, can only provide energies $\sim$ eV which are below the $\sim$keV threshold for conventional terrestrial searches. If one, instead, considers LDM scattering off bound electrons, energy transfer can cause excitation or even ionization and thus seems promising for exploring the phase space in a complementary way in the near future \cite{ele}.
DM hitting terrestrial targets is expected to have low velocities $v_\chi \sim 10^{-3}$ (we use $c=\hbar=1$ units) as the gravitational boost is small for the Earth i.e. its Lorentz factor $\gamma_{\varoplus} =1/\sqrt{1-\beta^2_\varoplus} \sim 1$ with $\beta_\varoplus=\sqrt{\frac{2GM_\varoplus}{R_\varoplus}}$. However, for compact objects, such as neutron stars (NSs) with masses $M_{NS}\simeq 1.5 M_\odot$ and radius $R_{NS}\simeq 12$ km, $\gamma_{NS}\sim 1.26$ or $v_\chi \sim 0.7$ and thus provides a mechanism to boost particles to higher velocities or, accordingly, test the same length scales with smaller projectile masses. In particular, the outer crusts
in NSs are formed by periodically arranged nuclei with typical densities ranging from $\rho \simeq 2 \,10^6-4\,10^{11}$ $\rm g/cm^3$. In the single-nucleus description \cite{chamel}, a series of nuclei with increasing baryonic number, $A$, from Fe to Kr form a lattice before neutrons start to leak out of nuclei. At these high densities, electrons form a degenerate Fermi sea. At even larger densities and up to nuclear saturation density, around $\rho_0\simeq n_0 m_N \simeq 2.4\,10^{14}$ $\rm g/cm^3$, a number of different nuclear structures called {\it pasta} phases appear \cite{pasta}.
In this work we study the effect of LDM scattering in the production of quantized lattice vibrations (phonons) in the outer NS crust. Later, we will discuss how this result can impact subsequent quantities of interest, such as the ion thermal conductivity, that are relevant for computing the cooling behavior of NSs. Phonons are quantized vibrational modes characterized by a momentum $\vec{k}$ and polarization vector $\vec{\epsilon}_{\lambda}$ appearing in a nuclear periodic system \cite{ziman}. They can have a number of different sources. They can be excited due to non-zero temperature $T$ in the medium. The Debye temperature allows us to evaluate the importance of the ion motion quantization. For a bcc lattice \cite{carr}, for example, $T_D\simeq 0.45 T_p$, being $T_p=\omega_p /k_B=\sqrt{\frac{4 \pi n_AZ^2e^2}{k_B^2 m_A}}$ the plasma temperature associated to a medium of ions with number density $n_A$, baryonic number $A$, electric charge $Ze$ and mass $m_A$. $k_B$ is the Boltzmann constant.
At low temperatures $T<T_D$, the quantization becomes increasingly important and the thermal phonons produced are typically acoustic modes, following a linear dispersion relation $\omega_{{k},\lambda}=c_{l, \lambda}|{\vec k}|$, where $c_l=\frac{\omega_p/3}{ (6\pi^2 n_A)^{1/3}}$ is the sound speed.
In addition, phonon production can be caused by an external scattering agent, for example, standard model neutrinos. In this respect, weak probes such as cosmological neutrinos with densities $n_\nu \sim 116$ $\rm cm^{-3}$ per flavor have been shown to provide small phonon production rates in a crystal target \cite{ferreras}. Due to the tiny mass of the neutrino, the experimental signature of this effect seems however hard to confirm.
The main interest in the astrophysical context we discuss in this contribution follows as phonon excitation in a periodic system, such as the outer crust of a NS, can affect the thermal transport coefficients in the star. The potential modification of transport properties of heat/energy in the external layers in NSs is crucial to possibly identifying relevant distortions in the cooling behavior of these astrophysical objects in rich LDM environments.
The structure of this contribution is as follows. In section II, we present the effective field theory Lagrangian model using dark matter-nucleon contact interactions via scalar and vector couplings in a relativistic framework and compute the single phonon excitation rate, discussing sources of uncertainty. Later, in section III we compute the thermal conductivity in the outer crust with LDM contributions comparing the results to the standard thermal value and discussing possible astrophysical consequences. Finally, in Section IV we give our conclusions.
\section{Light Dark Matter scattering and phonon excitation rate}
In this work, we consider the interaction of an incoming fermionic DM particle, $\chi$, scattering quasi-elastically with a nucleus in the outer NS crust lattice via scalar and vector couplings \cite{haxton, cermeno} composed of $Z$ protons $(p)$ and $A-Z$ neutrons $(n)$
\begin{equation}
\mathcal{L_I}=\sum_{N=n,p}g_{s,N}\chi\overline{\chi} N \overline{N}+g_{v,N}\chi\gamma^{\mu}\overline{\chi} N\gamma_{\mu} \overline{N},
\label{int_l}
\end{equation}
where $g_{s,N}$ $(g_{v,N})$ are the effective scalar (vector) couplings of the DM particle to the nucleon ($N$) field. We will focus on a weakly interacting candidate (WIMP) with mass in the sub-GeV range. This interaction is equivalent to considering a Fermi four-fermion interaction model, where the effective couplings of mass dimension $(-2)$ for these operators are obtained by integrating out the propagator of a generic $\phi$ mediator with mass $M_\phi$. Let us mention that, indeed, more effective operators for Dirac LDM candidates are possible, see for example, Table I in \cite{tait} for leading coupling contributions to Standard Model fermions. However, in order to keep our description concise for the sake of clarity, we will restrict here to the spin-independent interaction model used in previous works \cite{cermeno}.
Motivated by the need to compare bounds from colliders to direct detection, we describe interactions of DM with quarks $q=u,d$ and averaging in terms of nucleon fields we can write for the vector case $g_{v,N}/M_\phi^2\sim 1/{\Lambda_v}^2$ and $g_{s,N}/M_\phi^2\sim m_q/{\Lambda_s}^3$ where ${\Lambda_v}$ ( ${\Lambda_s}$) is the suppression mass scale for the vector (scalar) case, assuming the effective couplings are of order $O(1)$ and can be absorbed into ${\Lambda_{s,v}}$ \cite{ci}. Using bounds from CMS and ATLAS \cite{limits} we set ${\Lambda_v}\gtrsim 1$ TeV and ${\Lambda_s}\gtrsim 100$ GeV.
We denote $p'^{\mu}_N=(E'_N, \vec{p'_N})$ and $p^{\mu}_N=(E_N, \vec{p_N})$ as the four momentum for the outgoing and incoming nucleon, and $p'^{\mu}_\chi=(E'_\chi, \vec{p'_\chi})$ and $p^{\mu}_\chi=(E_\chi, \vec{p_\chi})$ those analogous for the LDM particle, respectively. Momentum transfer is denoted by $q^{\mu}=p'^{\mu}_\chi-p^{\mu}_\chi$.
Generically, given an interaction potential $\mathcal {V}$ felt by an interacting DM particle when approaching a nucleus in the periodic lattice, the single phonon excitation rate {\it per mode} can be obtained using the Fermi golden rule, $R_{\vec{k},\lambda}={2\pi}\delta (E_f-E_i)|\langle f |\mathcal {V}|i\rangle |^2$ where $i$ and $f$ are the initial and final states considered and $\delta (E_f-E_i)$ assures energy conservation. Given the fact that incoming (outgoing) LDM particles suffer a very moderate perturbation from the plane wave state, we will describe its incoming (outgoing) quantum state as $\Psi_{\vec{p_{\chi}}}(\vec{r})=\frac{1}{\sqrt{V}}e^{i\vec{p_{\chi}}\vec{r}}$ with $V$ the volume of the system.
The interaction potential felt by the LDM particle is the sum \cite{ferreras} over lattice sites
$\mathcal{V}(\vec{r})=\sum_{j}v(\vec{r}-\vec{r_j})$ that we describe for the sake of simplicity as impenetrable point-like spheres $v(\vec{r}-\vec{r_j})=\delta^3(\vec{r}-\vec{r_j})\; v_0$. We, nevertheless, comment on corrections to this picture later in the manuscript.
Using the Born approximation, the scattering amplitude for an incident $\chi$ particle can be written as
\begin{equation}
f(\vec{p_\chi},\vec{p'_\chi})\simeq -\frac{m_{\chi}}{2\pi}\int e^{i(\vec{p_\chi}-\vec{p'_\chi})\vec{r'}}v(\vec{r'})d^3\vec{r'},
\label{born1}
\end{equation}
and from its squared value, the differential cross section in the center of mass frame, $\frac{d\sigma}{d\Omega}|_{CM}=|f(\vec{p_{\chi}},\vec{p'_{\chi}})|^2$. The validity of the Born approximation is provided by the finite-range potential $\mathcal{V}(\vec{r})$ so the condition $|(\vec{p_{\chi}}-\vec{p'_{\chi}}).\,\vec{r'}|\ll1$ is fulfilled, being $|\vec{r'}|$ a typical target size. The effective interaction potential can be obtained from the squared interaction matrix element as calculated from the Lagrangian in eq.(\ref{int_l}) as $\frac{d\sigma}{d\Omega}|_{CM}=\frac{|\mathcal{\overline{M}}_{\chi N}|^2}{64\pi^2 s}$. First, we compute the scattering amplitude $|\mathcal{\overline{M}}_{\chi N}|^2$ being $s=(p_N+p_\chi)^2$ the Mandelstam variable. Adding the contribution over proton and neutron amplitudes coherently, we can obtain the LDM particle-nucleus differential cross-section and then integrate to find a relation between the total cross-section $\sigma_{\chi A} \simeq 4\pi a^2$ or, equivalently, the effective potential from eq.(\ref{born1}), and the scattering length, $a$, at low incident energies. We obtain $v(\vec{r})=\frac{2\pi a}{m_{\chi}}\delta (\vec{r})$. Besides, we have used a normalization of the delta function as $\int_{VT} \delta(x) d^4x=1$. From the Lagrangian in eq.(\ref{int_l}) the spin-averaged scattering amplitude \cite{cermeno} reads
\begin{equation}
\begin{aligned}
|\mathcal{\overline{M}}_{\chi \rm N}|^2 =& 4g_{s,N}^2 [({p_N}{p'_N}+m_N^2)({p_\chi}{p'_\chi}+m_{\chi}^2)] + 8 g_{v,N}^2 [2m_N^2m_{\chi}^2 -\\
& m_N^2p'_\chi p_\chi-m_{\chi}^2 p_N p'_\chi+(p'_\chi p'_N)(p_N p_\chi)+(p'_Np_\chi)(p_Np'_\chi)]+ \\
& 8 g_{s,N} g_{v,N} [m_N m_{\chi}(p_N+p'_N)(p_\chi+p'_\chi)].
\end{aligned}
\end{equation}
\normalsize
Due to the mildly relativistic nature of nucleons inside the nucleus, energy and momentum will lie close to the Fermi surface values $E_{FN}, |\vec{p}_{FN}|$ and $ |\vec{p}_{N}|^{2} \sim |\vec{p'}_{N}|^{2} \sim |\vec{p}_{FN}|^{2} \ll m^2_N$. We will approximate the product $p'_N p_N=E_N E'_N-|\vec{p_N}||\vec{p_N}'|cos\; \theta_{\vec{p_N},\vec{p_N}'}\simeq E_{FN}^2$. On the other hand, for the more relativistic DM particle products $p_\chi p'_\chi=E_\chi E'_\chi-|\vec{p_\chi}||\vec{p'_\chi}| cos \; \theta_{\vec{p_\chi},\vec{p'_\chi}}=E_{\chi}^2-|\vec{p_{\chi}}|^2cos \; \theta_{\chi}=m_{\chi}^2+|p_{\chi}|^2(1-cos \; \theta_{\chi})$ where we use $\theta_{\vec{p_\chi},\vec{p_\chi}'}\equiv \theta_{\chi}$.
The density dependence will be retained using a parametrization of the nuclear Fermi momentum $|\vec{p}_{FN}| \sim(3 \pi^2 n_0 Y_N)^{1/3}$ and the nuclear fractions $Y_p=Z/A$, $Y_n=(A-Z)/A$. If we now average over angular variables,
\begin{equation}
\begin{aligned}
\int_{-1}^{1} 2\pi d( cos \; \theta_{\chi}) |\mathcal{\overline{M}_{\chi \rm N}}|^2 =& 16\pi g_{s,N}^2 [(2m_N^2+|\vec{p}_{FN}|^2)(2m_\chi^2+|\vec{p_{\chi}}|^2)]+ \\
& 32\pi g_{v,N}^2 [2E_{\chi}^2E_{FN}^2-m^2_N|\vec{p_{\chi}}|^2-m_{\chi}^2|\vec{p}_{FN}|^2] +\\
& 128 \pi g_{s,N}g_{v,N}[m_Nm_{\chi}E_{FN}E_{\chi}].
\end{aligned}
\label{eq1}
\end{equation}
\normalsize
In the nucleus, we can use the previous expression, eq. (\ref{eq1}), to find the coherent contribution of the $A$ nucleons in a similar way to what is done in direct detection \cite{ddetec},
\begin{equation}
\int_{-1}^{1} 2\pi d ( cos \; \theta_{\chi}) |\mathcal{\overline{M}_{\chi A}}|^2 \simeq {m^2_A} \left (\frac{Z}{m_p}\sqrt{|\tilde{\mathcal{M}_p}|^2}+\frac{(A-Z)}{m_n}\sqrt{|\tilde{\mathcal{M}_n}|^2}\right) ^2,
\end{equation}
with $\int_{-1}^{1} 2\pi d( cos \; \theta_{\chi}) |\mathcal{\overline{M}_{\chi \rm N}}|^2 \equiv |\tilde{\mathcal{{M}_{\rm N}}}|$. The Mandelstam variable $s=(p_A+p_\chi)^2=m_{\chi}^2+m_A^2+2E_AE_{\chi}-2\vec{p_A}\vec{p_{\chi}}$ can be approximated as $s\simeq (m_\chi + m_A)^2$, neglecting the mildly relativistic nuclei momenta. Thus we can express the cross-section in the center of mass frame as
\begin{equation}
\sigma_{A,\chi}= 4 \pi a^2= {m^2_A} \frac{\left(\frac{ Z}{m_p}\sqrt{|\tilde{\mathcal{M}_p}|^2}+\frac{(A-Z)}{m_n}\sqrt{|\tilde{\mathcal{M}_n}|^2}\right) ^2}{16 \pi (m_\chi+ m_A)^2}.
\label{se}
\end{equation}
From a zero-order momentum expansion, we recover the usual expression for direct detection spin independent cross-section at low energies \cite{gluscev} for each coupling $\sigma_{A,\chi}\rightarrow \frac {\mu^2_{\chi A}}{\pi} (Z g_{s,p}+ (A-Z)g_{s,n})^2$ where $\mu_{\chi A}=\frac{m_\chi m_A}{m_\chi+ m_A}$ is the reduced $\chi-A$ mass. Note at this point that the $\sim A^2$ enhancement in the obtained cross-section remains as the coherence condition $\lambda \ge R_{A}$ is fulfilled, being $R_A$ the nuclear radius and $\lambda=h/|\vec{p_\chi}|$ the De Broglie wavelength. In addition, the contribution of the nuclear lattice will be described by the summations extended over the lattice sites or, equivalently, by the inclusion of the structure factor $S(q)\sim |\sum_{j} e^{-i\vec{q}\vec{r_j}}|^2$, in the full phonon excitation rate expression as will be shown later in the manuscript. Some studies have included form factors $F^2(q)$ to correct a point-like nucleus nature approach \cite{gluscev}, however since we will be focusing on $q\rightarrow 0$ limit we will consider them as unity for the sake of simplicity. In what follows we will refer to $\vec{p'}\equiv \vec{p'_\chi}$ and $\vec{p} \equiv \vec{p_\chi}$. The single-phonon excitation time rate from the ground state now reads
\begin{equation}
R^{(0)}_{{k},\lambda}=\frac{4\pi^2a^2}{V^2m_{\chi}^2} \delta (E_{\vec{p'}}+\omega_{{k},\lambda}-E_{\vec{p}})\; 2\pi \; |\sum_{j} \langle 1, \vec{k} \lambda |e^{-i\vec{q}\vec{r_j}}|0 \rangle|^2,
\label{rate2}
\end{equation}
where $\vec{r}_j=\vec{x}^{(0)}_j+\vec{u}_j$ with $\vec{x}^{(0)}_j$ the lattice point and $\vec{u}_j$ the displacement vector \cite{Ashcroft}. We must note at this point that the previous expression includes the squared modulus of the Fourier transform of the periodically arranged lattice sites including thus the usual description in terms of the structure factors $S(q)$. This function provides information on the spatial distribution through a correlation function and presents maxima at the crystal nuclear positions. The contribution of this factor to the global cross-section describes coherent scattering from all of the different nuclei as discussed in \cite{horo}. There, the effect of efficient low-energy scattered WIMPS from the interior of the stelar DM distribution was mentioned as an additional factor to prevent DM escaping from the NS once inside. In this way it thus constitutes a mechanism for {\it trapping} DM, besides the deep gravitational potential felt by these sub-GeV mass particles.
Beyond this point, we will consider an isotropic medium and since the Born approximation $|\vec{q}.{\vec r'}|\ll1$ holds, it is most likely that acoustic modes are excited. It follows that
\begin{equation}
-i\vec{q}\sum_{j} e^{-i\vec{q}\vec{x}^{(0)}_j} \langle 1, \vec{k} \lambda |\vec{u}_j|0 \rangle = -i n_A \delta^{(3)}(\vec{k}-\vec{q}) \sqrt{\frac{|\vec{k}|}{2m_Ac_l }},
\end{equation}
where we have used the continuum limit $\sum_j \rightarrow n_A \int d^3x$ and the fact that all $\vec{k}$ have a polarization vector that verifies $\vec{\epsilon_l} // \vec{k}$ and the other two vectors are perpendicular to $\vec{k}$.
Finally eq.\eqref{rate2} can be written as
\begin{equation}
R^{(0)}_{{k}}=\frac{4\pi^2a^2}{m_{\chi}^2V} \delta (E_{\vec{p'}}+\omega_{{k},\lambda}-E_{\vec{p}})\; 2\pi n^2_A\frac{ \delta^{(3)}(\vec{k}-\vec{q}) |\vec{k}|}{2m_Ac_l}.
\end{equation}
At this point we must consider the peculiarities of the incoming LDM phase space distribution $f_{\chi}(\vec{p})$ as it will impact the averaged final phonon excitation rate. Typically, for the Sun or the Earth the uncertainties have different sources including orbital speed of the Sun, escape velocity from the DM halo and the form of the phase space distribution itself. About the latter and in local searches, direct and indirect detection are affected in different manner. For example, direct detection is sensitive to DM with high velocities \cite{dir} while for indirect detection the low-velocity part of the distribution is tested \cite{ind, choi} .
A popular choice is obtained using an approximation based on an isotropic sphere with density profile $\rho_{DM}(r) \propto r^{-2}$ of collisionless particles, i.e. a Maxwell-Boltzmann type with a local mass density $\rho_{LDM} = 0.3$ $\rm GeV$ $\rm cm^{-3}$. Uncertainties on the knowledge of the distribution function must be carefully considered as this impacts accuracy when translating event rates to constraints on particle physics models of DM.
In the case we analyse here of a more compact object, it is the high velocity part of the distribution that is tested, as typical values for boosted root-mean-squared velocities are ${<v^2>}\sim {2 G M_{NS}/R_{NS}}\sim (0.6)^2$.
For these relativistic regimes one must use the Maxwell-J\"uttner distribution \cite{juttner} function and, more properly, take into account the space time curvature due to gravitational field created by the NS \cite{Kremer}
\begin{equation}
f_{\chi}(\vec{p})=\frac{n_\chi \mu }{4 \pi m_\chi^3 K_2(\mu)} e^{-\mu \sqrt{1+g_1(r)\frac{|\vec{p}|^2}{m_\chi^2}}},
\label{f}
\end{equation}
where $\mu=\frac{m_\chi}{k_BT}$ and $K_2(\mu)$ is the modified Bessel function of second kind defined as $K_n(\mu)=\left( \frac{\mu}{2}\right)^n \frac{\Gamma(1/2)}{\Gamma(n+1/2)} \int_0^\infty e^{-\mu y}(y^2-1)^{n-1/2}dy$. The isotropic Schwarzschild metric for the gravitational field created by the NS source is \cite{Kremer} $ds^2=g_0(r)(dx^0)^2-g_1(r) \delta_{ij} x^i x^j$, $i,j=1,2,3$.
Note that in the close vicinity of the NS where we will be interested in assesing our quatities of interest, $r\sim R_{NS}$, and follows that $g_1(R)=\left( 1+ \frac{GM_{NS}}{2R_{NS}}\right)^4\sim 1.42$, $g_0(R)=\left( 1- \frac{GM_{NS}}{2R_{NS}}\right)^2/\left( 1+ \frac{GM_{NS}}{2R_{NS}}\right)^2\sim 0.69$. The distortion from the flat space with a Minkowski metric efectively sets $g_0(r), g_1(r) \neq 1$ as expected. Furthermore, if we obtain from the above the root-mean-squared $\sqrt{<v^2>}\sim 0.6$ this implies $\mu \approx 6.7$ \cite{Hakim, Cercignani}.
The normalization condition is such that the particle four-flow $J^\alpha$ can be defined and taking the $\alpha=0$ component we obtain $\int d^3\vec{p}f_{\chi}(\vec{p})\sqrt{-g}/g_0=J^0=n_{\chi}/\sqrt{g_0}$ with $\sqrt{-g}=\sqrt{g_0 g^3_1}$. $n_{\chi}$ is the DM number density near the NS. Note that at non-relativistic velocities and flat space we do recover the Maxwell-Boltzmann distribution as expected. Further, we consider all outgoing $\chi$ states are allowed as the net number will be tiny as compared to ordinary matter. The phonon excitation time rate must be weighted with the momenta of the {\it local} $\chi$ phase space that, as mentioned, is shifted to the relativistic values
\begin{eqnarray}
R^{(0)}_{{k}} \nonumber & = & \frac{4\pi^3n_A^2V }{m_{\chi}^2m_Ac_l} \int \frac{d^3\vec{p} \,f_{\chi}(\vec{p})}{(2 \pi)^3} \int \frac{d^3\vec{p'}}{(2 \pi)^3}\delta^{(3)}(\vec{k}-\vec{q})\; \delta (E_{\vec{p'}}+\omega_{\vec{k},\lambda}-E_{\vec{p}})|\vec{k}|a^2.\\
\label{rate31}
\end{eqnarray}
\normalsize
Computing the zeros of the delta function and expressing the incoming momentum as $|\vec{p}_{0}|= \sqrt{\gamma^2-1}m_{\chi}$ we obtain an interval of kinematically allowed $|\vec{k}|$ values $0\leq |\vec{k}|\leq 2m_{\chi} \left( \frac{c_l\gamma }{(c_l^2-1)}+\frac{\sqrt{\gamma^2-1}}{|c_l^2-1|}\right)$ and the eq.\eqref{rate31} takes the form,
\begin{equation}
R^{(0)}_{{k}}=\frac{8\pi^4n^2_A V}{(2 \pi)^6 m_{\chi}^2m_Ac_l} \int_{0}^{\infty} |\vec{p}|^2d|\vec{p}| f_{\chi}(\vec{p})\frac{|\gamma m_{\chi}-|\vec{k}|c_l|}{m_{\chi}\sqrt{\gamma^2-1}}a^2.
\end{equation}
In Fig.(\ref{fig1}) we show the single phonon excitation rate (per unit volume) from the ground state and averaged over $\chi$ phase space as a function of density in the outer crust using the single-nucleus table from \cite{chamel}. Curves plotted with solid, dashed and dash-dotted lines correspond to excitation of phonons with $|\vec{k}|\rightarrow 0$ for $m_\chi=500, 100$ and $5$ MeV and $n_\chi/n_{0,\chi}=10$. We also plot for the sake of comparison the specific excitation rate at $|\vec{k}|\rightarrow 0$, $R_{\nu 0}$, for neutrinos with masses $m_{\nu}=0.1, 1$ eV with dotted and doble-dashed lines, respectively. Note, however, that in this later case, there is a strong momentum dependence that declines rapidly. We can fit this behavior for $m_{\nu}=0.1$ eV as $ R^0_\nu(|\vec{k}|)=R_{\nu 0} e^{ \left(\frac{ -1754|\vec{k}|}{1 \,\rm eV}\right)}$ and for $m_{\nu}=1$ eV as $ R^0_\nu(|\vec{k}|)=R_{\nu 0} e^{-\left( \frac{ 2561.3|\vec{k}|}{1 \,\rm eV}\right)}$.
We have verified that since, typically, the speed of the thermalized LDM particles far from the star is essentially $v_\chi \sim v_\infty\sim 10^{-3}$, when hitting the NS it has already acquired a boosted energy. Using an estimate based on a monochromatic value $E_{\chi}=\gamma_{NS} m_{\chi}$, $\gamma_{NS}=1.26$ we can straightforwardly integrate and obtain the analytical result
\begin{equation}
R^{(0)}_{{k}}=\frac{n_\chi n^2_A V}{4 (2 \pi)^3 m_{\chi}^3m_Ac_l}\frac{|\gamma_{NS} m_{\chi}-|\vec{k}|c_l|}{\sqrt{\gamma^2_{NS}-1}}a^2,
\label{ro}
\end{equation}
that under-predicts the exact result by $\sim 20\%$. As deduced from the previous expression eq.(\ref{ro}) the rate is indeed constant as a function of momentum as the inequality $\gamma_{NS} m_{\chi} \ll |\vec{k}|c_l|$ is fulfilled. It seems that the contribution of the phase space distribution of LDM may also have strong impact on the results, as it happens for the Sun or Earth.
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth, angle=0,scale=1.25]{new_figure_1.eps}
\caption{Averaged single phonon excitation rate per unit volume as a function of density in the outer crust. DM particle masses $m_\chi=500, 100$ and $5$ MeV are used and $n_\chi/n_{0,\chi}=10$. Neutrino contribution at $|\vec{k}|\rightarrow 0$, $R_{\nu 0}$, is also shown for $m_{\nu}=0.1, 1$ eV. See text for details. }
\label{fig1}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth, angle=0,scale=1.25]{new_fig2.eps}
\caption{Phonon thermal conductivity as a function of density (in units of $10^{10}$ $\rm g/cm^3$) for temperatures $T=5\,10^7$ K (blue), $5\,10^8$ K (red) and $m_\chi=100$ MeV. Dash-dotted and dashed lines depict the impact of a LDM density $n_\chi/n_{0\chi}=10, 100$. Solid lines are the standard thermal result with no DM for each case. See text for details.}
\label{Fig2}
\end{figure}
\section{Astrophysical impact on thermal conductivity}
Phonon production can be crucial for determining further transport properties, in particular, thermal conductivity in an ion-electron system such as that in the outer NS crust. As an important contribution to the total ion conductivity, $\kappa_i$, partial ion conductivities due to ion-ion, $\kappa_{ii}\equiv \kappa_{ph}$, and ion-electron collisions, $\kappa_{ie}$, must be added \cite{chugunov} under the prescription $\kappa^{-1}_i=\kappa^{-1}_{ii}+\kappa^{-1}_{ie}$. Standard mechanisms to produce lattice vibrations include thermal excitations, as analyzed in detail in previous works \cite{pot, baiko}. In a NS, the outer crust can be modelled under the one-component-plasma description. This low density solid phase can be classified according to the Coulomb coupling parameter $\Gamma=Z^2e^2/a k_B T$ where $a=(4 \pi n_A/3)^{1/3}$ is the ion sphere radius. It is already known that typically for $\Gamma \ge \Gamma_m\simeq 175$, or below melting temperature $T<T_m$, single-ion systems crystallize \cite{gamma}.
There are a number of processes that can affect thermal conductivity in the medium. The so-called U-processes \cite{ziman} are responsible for modifying the electron conductivity such that for high temperatures, $T>T_U$, electrons move almost freely. Assuming a bcc lattice, $T_U\simeq 0.07T_D$. Thus in the scenario depicted here, the temperature range must be $T_U <T<T_D<T_m$ for each density considered. According to kinetic theory, the thermal conductivity can be written in the form \cite{ziman}
\begin{equation}
\kappa_{ii}=\frac{1}{3} k_B C_A n_A c_l L_{ph},
\end{equation}
where $C_A=9 \left(\frac{T}{T_D}\right)^3 \int_0^{T_D/T} \frac{x^4 e^x dx}{(e^x -1)^2}$ is the phonon (dimensionless) heat capacity {\it per ion}, $L_{ph}$ is an effective phonon mean free path that includes all scattering processes considered: U-processes and impurity (I) scattering processes (both dissipative) and the phonon normal (N) scattering which are non dissipative $L^{-1}_{ph}=L^{-1}_U+L^{-1}_I+L^{-1}_N$. Typically the thermal conductivity is related to the thermal phonon number at temperature $T$, $L_{ph}\sim 1/N_{0,k \lambda}$ where $N_{0,k \lambda}=(e^{\omega_{k\lambda}/k_B T}-1)^{-1}$. The contribution from DM can be obtained by the net number of phonons that results from the competition of thermal and scattering excitation and stimulated emission \cite{ferreras} in a 4-volume $\delta V \delta t$ using the averaged rate per unit volume and weighting with the incoming distribution providing the frequencies of different values of momenta we obtain
\begin{equation}
N_{k\lambda}\simeq N_{0,k\lambda}+ R^{(0)}_k \delta V \delta t - \int \frac{d^3\vec{p}}{n_\chi} \,f_{\chi}(\vec{p}) {\tilde R^{(0)}_k} N_{0,k\lambda} e^{(\omega_{k,\lambda}+ {\vec k}.{\vec v})/ K_\chi} \delta V \delta t,
\end{equation}
where $K_\chi=(\gamma-1)m_\chi$ is the $\chi$ kinetic energy and ${\tilde R^{(0)}_k}$ is the single phonon excitation rate for each particular momentum value (not averaged over incoming $\chi$ momenta). Since the source (NS) is in relative motion to the LDM flux, there is a Doppler shift characterized by the source velocity ${v}\equiv v_{NS}\sim 10^{-2}$ i.e galactic NS drift velocity.
The distribution of NS in our galaxy peak at distances $\langle r \rangle_{max} \lesssim 4$ Kpc \cite{lorimer} where the DM density is enhanced with respect to the solar neighborhood value $n_{0,\chi}\simeq 0.3$ $\rm GeV/cm^3$ thus we will consider $n_\chi\simeq (10, 100)n_{0,\chi}$ as prescribed by popular galactic DM distribution profiles.
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth, angle=0,scale=1.25]{new_figure3.eps}
\caption{Phonon thermal conductivity as a function of density (in units of $10^{10}$ $\rm g/cm^3$) at $T=10^8$ K and $m_\chi=65$ MeV. Solid, dot-dashed and dashed lines correspond to cases with no DM, $n_\chi/n_{0,\chi}=10,\,100$. Perpendicular electron thermal conductivity is also shown for $B=10^{14}, 10^{15}$ G.}
\label{Fig3}
\end{figure}
In Fig. (\ref{Fig2}) we show the phonon thermal conductivity as a function of density (in units of $10^{10}$ $\rm g/cm^3$) at $T=5\,10^7 \,$ K, $5\,10^8$ K typical for the base of the crust, for $m_\chi=100$ MeV. Solid lines are the standard thermal result with no DM. Dash-dotted and dashed lines correspond to $n_\chi/n_{0,\chi}=10, \,100$, respectively.
We see that at the largest LDM local densities considered, there is a clear enhancement over the thermal result well inside the outer crust. This corresponds to the site where the DM-induced effects have the most influence \cite{azorin} as this is the most massive part of the outer crust. Below these densities, there is a negligible change though. At lower $T$, the effect of a perturbation over the thermal phonon population is more important. Enhanced (decreased) conductivities at moderate LDM densities are due to a net reduction (increase) of the number of phonons in the lattice as a result of cancellation of modes. As a representative scenario, we have taken $|\vec{k}|=0.01/a$ at each density since we have verified that this choice verifies the kinematical restrictions on $|\vec{k}|$ when performing the averages over phase space distribution as discussed in section II. Besides, rates are mostly constant at low $|\vec{k}|$. Note that in standard calculations \cite{chugunov}, there is no momentum dependence as they replace the frequency mode $\omega_{k\lambda}$ by a constant threshold. We must bear in mind that this result must be compiled with a realistic impurity fraction so that conductivity remains finite. We have considered $L_I\sim 5a$ \cite{chugunov}.
In order to understand the significance of our result in the dense stellar context in Fig.(\ref{Fig3}) we show the phonon thermal conductivity as a function of density (in units of $10^{10}$ $\rm g/cm^3$) at $T=10^8$ K and $m_\chi=65$ MeV for $|\vec{k}|=0.01/a$. Solid, dot-dashed and dashed lines correspond to cases with no DM, $n_\chi/n_{0,\chi}=10,\,100$, respectively. Electron thermal conductivity is also shown for magnetized realistic scenarios in the perpendicular direction to a magnetic field $B$ of strength $B=10^{14}$ G (dotted) and $B=10^{15}$ G (doble dotted). Ions are mostly not affected by the presence of a magnetic field. The parallel direction electronic contribution is not depicted here since it is typically much larger $k_{e \parallel}\sim 10^{17}-10^{19}$ $\rm erg\,cm^{-1} \,s^{-1}\, K^{-1}$. Since we perform averages over the $\chi$ phase space we again use $|\vec{k}|=0.01/a$. On the plot we can see that the electronic contribution in the perpendicular direction falls below the enhanced $n_\chi/n_{0,\chi}=100$ DM value for densities $\gtrsim 3.5\,10^{11}$ $\rm g/cm^3$. Note that the low value chosen for $|\vec{k}|$ in this plot is to be understood as a compromise value, larger $|\vec{k}|$ values would imply the impossibility of exciting phonons from low-momenta incoming LDM. Since the global conductivity is $\kappa=\kappa_e+\kappa_{ph}$, the obtained result is expected to contribute to the reduction of the difference in heat conduction in both directions and thus to the isotropization of the NS surface temperature pattern as seen in \cite{azorin} for standard physics. Temperatures would be smoothly driven towards more isothermal profiles for latitudes among pole and equator. It is already known \cite{kamin} that the outer crust plays an important role in regulating the relation among temperature in the base of it and the surface. The detailed calculation of this implication for surface temperatures remains, however, for future work.
\section{Conclusions}
In conclusion, we have derived for the first time the single phonon excitation rate in the outer NS crust for relativistic LDM particles in the sub-GeV mass range. We have found that this rate is constant with the phonon momentum and much larger than for cosmological neutrinos at finite $|\vec{k}|$. A non-negligible correction to the local phonon excitation rate of $\sim 20\%$ is obtained when full relativistic phase distribution functions are considered for the incoming $\chi$ particles with respect to a monochromatic approximation, that under-predicts the result.
As an astrophysical consequence of the previous, we have calculated the ion thermal conductivity in the dense and hot outer envelope founding that it can be largely enhanced at LDM densities in the maximum of the NS galactic distribution $n_\chi\sim 100 n_{0,\chi}$ due to a net modification of the acoustic phonon population. This effect is non negligible at densities beyond $\sim 3.5\, 10^{11}$ $\rm g/cm^3$ in the base of the outer crust at the level of standard ion-electron or thermal effects \cite{chugunov, pot}. We do not expect the degenerate electron contribution to largely modify this result as this would mildly screen nuclear charge in the lattice however it remains to be further studied.
Although a detailed study of the quantitative effect in the surface temperature pattern remains to be undertaken, it is expected that for magnetized NSs the LDM-enhanced global enhancement of the perpendicular thermal conductivity allows a reduction of the difference of heat transport among parallel and perpendicular directions to the magnetid field. Based on previous works only including standard thermal contributions we expect that, as a natural consequence, the surface temperature profile would be more isotropic yielding flatter profiles for intermediate latitudes and remains to be calculated in a future contribution.
\section{Acknowledgments}
We acknowledge useful comments from J. Pons and C. Albertus. This research has been partially supported by MULTIDARK and FIS2015-65140 MINECO projects and at IAP by the ERC project 267117 (DARK) hosted by Universit\'e Pierre et Marie Curie - Paris 6 and at JHU by NSF grant OIA-1124403. M. Cerme\~no is supported by a fellowship from the University of Salamanca.
|
\section{Introduction}
An \emph{Opinion Game} is a mathematical model of the evolution of the
opinions of agents within a population, from creation of an idea to
its diffusion through the population and finally the ultimate state of
the system. Opinion games (or opinion dynamics models) have been
studied since the middle of the 20th century. They often are used to
study the dynamics of generic opinions that flow within a population.
By creating a formal mathematical model we can study detailed
hypotheses about transmission mechanisms and the impact of population
structure on opinion formation. Clearly such models are idealized and
fail to capture all the nuances and fuzziness of individual human
thought, but they give at least a disciplined glimpse into an aspect
of how agents interact. A formal analytic approach also admits the
ability to establish theorems about opinion dynamics, allowing one to
derive conclusions with a rigorous chain of formal reasoning
supporting them. This differs from approaches that take a purely
simulation-based approach in which conclusions are derived from
statistical properties of the resultant data instead of the underlying
dynamical rules dictating interactions. By building a solid foundation
for the opinion dynamics we get connection between model structure and
statistical properties.
Models typically vary in the assumptions that they make about the
nature of individual interactions and the overall goal of the
population in the limit of an arbitrary number of interactions. In
some cases, models are used to study specific classes of opinions that
are created and flow within a population (such as linguistic
games\cite{Baronchelli20101,kay2003, Baronchelli2010, Puglisi2008}).
In other cases, the model seeks to explore
equilibrium states in which not all agents necessarily agree, but
reach opinion states that no longer change\cite{FRENCH1956}. The
generalized version of these games can be split into two categories. A
binary (or more generally, discrete) opinion
model~\cite{Follmer,Ding2010,Galam1991,Sznajd2000,Dietrich2003,Martins2007,GalamDiscrete}
requires agents to negotiate an opinion on a topic that has two
possibilities and therefore, the final state of the system divides
agents into two distinct groups, see \cite{Stauffer2003},
\cite{Biswas2009}. The alternative is a system in which the opinion
space is continuous~\cite{Weisbuch2002, Fortunato, Deffuant2000,
Hegselmann2002}. For example ``\textit{How good do you think the New
York Times is as a media outlet?}'' does not have a binary answer --
you might strongly believe it is a strictly good or bad source of
information, or you may have a mixed opinion somewhere in between. An
opinion game model allows us to study the state reached after a set of
individuals interacts an arbitrary number of times. This final state
may be consensus or there may emerge stable sub-populations
(\emph{clusters}) in which members of each group reach consensus, but
difference in opinion exist across these groups. The concept of
opinion can be extended to higher dimensions, allowing for more than
two extreme positions to be taken on a topic (e.g., ``what is your
favorite color'').
\subsection{Understanding social dynamics}
The dynamics of an opinion game are dictated by the micro-dynamics
that occur in individual-to-individual interactions between agents
connected by some network topology. At the micro-scale we would
represent the value of each opinion held by each individual, and how
opinions of two individual can change and evolve via direct
interaction. An example would be the one in which both agents have the
same power over each other and therefore, after an interaction they
would learn equally from each other and move relative to their
interaction partner by the same amount either in the direction of
increased agreement or disagreement or do not change whatsoever. For
example, repulsion could occur when two agent's opinion are too far
apart, especially when they talk about sensitive topics such as
religion or politics. In some situations we may wish to model
interactions where a difference of opinion between two individuals is
insufficiently large to cause either attractive or repulsive behavior
in which no change occurs. In yet another situation we may model
dynamics in which opinions change only when they are sufficiently
different in a repulsive fashion, but with no attractive behavior when
they are near by. We will see how these can be modeled in
section~\ref{sec:interaction-potentials} via an interaction rule that
seeks to minimize an interaction energy via \emph{potential
functions}.
\subsubsection{Asymmetry of influence}
In the real world, it is frequently the case that agents do not have
equivalent influence over each other. Such asymmetric relationships
are commonplace: teacher/student, parent/child, expert/amateur, and so
on. This can be modeled by applying weights, or \emph{influence
power}, to the relationship between agents. In this case the
movements of the agents are not the same. For instance, if two agents
are talking about chemistry and one is a chemist and the other is a
student, then we can add an influence factor to the interaction and
make the chemist to have more influence on the student (expert
power). The result of such weighted influence would cause the student
to experience significantly higher changes of opinion than the teacher
during an interaction (which is part of the learning process), while
allowing the teacher to admit some small level of change as a part of
their interaction to maintain consistency in what they teach while
allowing for some level of flexibility in response to their
environment.
In an extreme case, we can consider the situation in which one agent
is absolutely dominant and acts purely as a \emph{speaker} and the
other is acting purely as a \emph{hearer} such that the opinion is
transmitted unidirectionally. Consequently after an interaction,
speaker would not change her opinion and hearer is the only one who
changes. This allows for the modeling of information sources such as
news or propaganda outlets that act in a purely influencing role
within the population. It is worth mentioning that influence of
agents over each other might change over time and evolve, we refer
enthusiastic reader to~\cite{Friedkin2011}. Expert power is one
mechanism by which asymmetry is introduced into an opinion model and
is applied on the interaction \textit{between} two individuals.
\subsubsection{Tendency for individualism}
The dynamics that emerge often fail to model phenomena
that appear in real social systems due to decisions made solely by
individuals. We can add more ingredients such as adaptive noise to
agents' decisions to capture these decisions. This noise can be a
function of the collective opinion within the population or cluster
that an individual finds themselves a member of. For example,
individuals that find themselves in a population that all hold the
same opinions may have a desire to express their individuality by
making an individual internal decision to change their opinion
slightly to differentiate themselves. Exactly when and how this
desire for individuality will emerge and balance with a desire to
conform and form coherent like-minded groups is question open to
investigation.
Different real world phenomena emerge due to this desire for
individualism such as fads and trends, as well as the spontaneous
emergence of new trends. In the fashion world, when a new style
becomes popular there is a period in which individuals adapt their
opinions about fashion towards this popular style. Eventually when
this style becomes widespread and relatively commonplace, some
individuals suddenly express a desire to be unique and begin avoiding
this style. The more popular the style becomes, the more
powerful would be tendency of peers to try something new and leave the
cluster. This force is referred to as a ``\textit{disintegrating
force}'' or ``\textit{tendency for individualism}'' or
``\textit{tendency for uniqueness}''\cite{Mas2010}. Many formal models
of opinion dynamics introduced in the 20th century are based on the
assumption that such disintegrating forces are not present, which
yields interesting but relatively synthetic idealized opinion
dynamics and equilibria.
\subsubsection{Coupling of topics}
It is rarely the case that opinions on specific topics exist in
isolation. An opinion about one topic frequently influences the
opinion an individual holds on other topics. These range from the
trivial (e.g., a preference for musical style influencing ones like or
dislike of a given musician in a given genre), to those that reflect
complex social influences (e.g., an opinion about religious belief
influencing opinions about reproductive rights). Mathematically we can
treat opinions that change together as being \emph{coupled}, where
there exists some functional that defines how the state of two or more
coupled opinions co-evolve. In systems such as linguistic games to
determine color terminology, coupling rarely has an impact, but to
model social topics that have overlap at some level, coupling is an
unavoidable aspect of capturing the complexity and nuance of human
behavior. There are some works \cite{Li,Fortunato,Touri} in which
several topics are being discussed in the system, but the topics are
independent and coupling is not present. A recent work~\cite{Parsegov2015} studies coupling of interdependent topics for the Friedkin-Johnsen (FJ) model. A dynamical system view is
presented in \cite{Touri} along with an upper bound for convergence
time of the Hegselmann-Krause model~\cite{Hegselmann2002,blondel}.
\subsubsection{Interconnection topology}
Geographic and social factors dictate the likelihood of any two
individuals interacting. This can be encoded via an interaction
network. Given a population of individuals, the likelihood of any
pair of individuals interacting is strongly influenced by their social
and geographic network. Agents who are friends are more likely to
interact than agents who are not, as are agents who are nearby versus
those that are physically separated by a large distance. In our
model, we impose a graph topology on all individuals who can interact
in which the connectivity of the graph dictates which pairwise
interactions are possible. For example, a tightly knit community of a
small number of individuals likely can be treated as a fully connected
graph in which all individuals may interact with each other. On the
other hand, a set of individuals in widely distributed cities who do
not interact with each other but read the same newspaper can be
modeled as a star graph. Realistic interaction networks can be
derived from social networking sites such as Twitter or Facebook,
allowing for connectivity graphs that mimic interactions that are
encountered in the real world.
\subsection{Contributions}
The model we present in this paper makes a number of noteworthy contributions to the study of opinion dynamics building upon decades of related research (Detailed in Sec.~\ref{sec:relatedwork}).
\begin{itemize}
\item The treatment of opinion evolution as a result of minimization
of a (potentially nonlinear) energy functional based on the
difference of opinion on a topic between interacting
individuals. (Sec.~\ref{sec:interaction-potentials})
\item Demonstration that this model is a generalization of a family of
previously published models from the literature by embedding them within
the framework that we present. (Sec.~\ref{sec:embedding})
\item Coupling of topics based on kernel-based coupling functions, to
model the influence of explicit communication on one or more topics
on a broader set of opinions held on topics not communicated.
(Sec.~\ref{Topic-coupling})
\item A study of the dynamical systems properties of our model to
understand the existence of regions of the opinion space that
represent basins of attraction, fixed points and
sensitivity. (Sec.~\ref{sec:dynamics})
\end{itemize}
A set of results of computational experiments are provided
demonstrating some of the dynamics one can observe in these models in
addition to the analytical results that we derive. We close by posing
a set of directions of potential future research based on our model
and our early explorations with it.
\section{Related research}
\label{sec:relatedwork}
Models of opinion dynamics have been studied over a long period of
time with various formal methods and goals. In this section we
provide a short review of noteworthy works in this history and discuss
how our work is novel relative to prior work. This review is not
intended to be exhaustive.
\subsection{French model: early formal modeling of opinion dynamics}
One of the earliest works in this area appears in
1956~\cite{FRENCH1956}, in which French considers agents as particles
in a physical system which can induce forces on each other referred to
as \emph{interagental power}. Interagental power can be characterized
in five ways: attraction power, expert power, reward power, legitimate
power and coercive power. Let $v_1$ and $v_2$ be two agents in the
network, then attraction power of $v_1$ over $v_2$ is how much $v_2$
likes $v_1$. Agents tend to listen to those whom they like and
respect. Expert power is based on $v_2$'s understanding of $v_1$'s
knowledge. The student/professor example is the case in which a
student believes their teacher knows more on a given subject, and
therefore is influenced by their teacher. Reward power is based on
the right of an individual to offer tangible rewards of any kind to
somebody for doing what is wanted or expected of them. For example,
parent might give a gift to a child for getting a good GPA in a year,
or a boss can promote one of their employees. A negative example of
reward power could be taking away a child's toy for not performing well
in school. Coercive power is based on $v_1$'s ability to impose
sanctions on $v_2$, it uses the threat of force to gain compliance
from another. The idea is that someone is forced to do something that
he/she does not desire to do. For instance in a court room the judge
has utmost absolute power over a defendant, or a boss has power over
their employee. And finally legitimate power is based on how much $v_2$
believes $v_1$ has a right to prescribe their opinion. It comes from an
elected or appointed position of authority and is based on the social
norm which requires agents to be obedient to those who hold superior
positions social structure.
At a given time, each individual is forced by the cumulative force
applied by all of their neighbors and their own resistance to move
towards an equilibrium point where the resultant force is zero. The
result is linear model that determines behavior of the system and
individuals. In this model, the tendency of individuals is to move
towards a weighted mean of their opinions at each step. French also
explores the role of connectivity between individuals by applying the
model on different types of directed graphs, where direction of an
edge denotes a unidirectional (``speaks to'') relation.
\subsection{DeGroot's averaging model and social influence evolution}
The model described by DeGroot\cite{DeGroot1974} is a simplified
version of the model we describe in this paper. The core concept of
this model is that for a population of $n$ individuals, their set of
opinions on one topic at any given point in time can be represented as
a vector $\mathbf{y}^{(t)} \in \mathbb{R}^n$. An $n$-by-$n$
row-stochastic matrix of influence weights, $\mathbf{P}$, is provided
in which the element $\mathbf{P}_{ij}$ corresponds to the amount of
influence that individual $j$ has on individual $i$ when updating
their opinion. The diagonal elements $\mathbf{P}_{ii}$ represent each
individuals opinion of their own degree of influence within the
population, which can be interpreted as their confidence in their own
opinion relative to their peers. The update process is simply
matrix-vector multiplication:
\begin{equation}
\label{eq:DegrootEq}
\mathbf{y}^{(t+1)} = \mathbf{P} \mathbf{y}^{(t)}
\end{equation}
In this model, given a set of initial opinions the result will
converge to the single weighted average of the initial opinions. As
shown in Figure~(\ref{fig:degrootevolution}), for a full graph the convergence
is very regular for all individuals. When a random graph is used, the
system still ends at an average value, but the individual trajectory
to this final value is more interesting (not necessarily monotonic).
Our model subsumes this behavior, and supports behavior that DeGroot's
model cannot capture. For example, DeGroot's model implies that all
individuals will seek consensus - no matter how far apart the starting
opinions are, they will come together. Our model captures the case
where sufficiently different opinions on a topic may result in an
increase in disagreement (e.g., polarization of opinions) in which the
final outcome is not a single value but two subpopulations that reach
two distinct opposing opinions. We illustrate this in our
experimental results shown in Sec.~\ref{experiments}. As is clear
from Eq.~(\ref{eq:DegrootEq}), each node updates their opinion based on
their own and their neighbors opinions. An extension model in which each
agent incorporates their initial opinion throughout all iterations
was developed by Friedkin and Johnsen~\cite{Friedkin1999,FriedkinBook1,FriedkinBook2}.
\begin{figure}[httb!]
\centering
\includegraphics[width=.35\linewidth]{degroot.png}
\caption{Evolution of the DeGroot model from a set of initial opinions distributed over $[0,1]$ to consensus.}
\label{fig:degrootevolution}
\end{figure}
Moreover, an application of DeGroot's model, the DeGroot-Friedkin,
model has recently been studied with respect to the evolution of
social influence in work by Jia, et al.~\cite{jia2015}. In French's
model the influences are constant over
time. In~\cite{jia2015,Lorenz2005} authors model a new version of the
game where influences of agents evolve depending on the outcome of the
previous issue's result.
\subsection{The Bounded Confidence Model}
There are several different models for explaining and modeling the
reasons and dynamics of phenomena such as the reason why agents tend
to maintain their differences while they learn and become more alike
by interacting. Why does the process of becoming more alike stop
before a complete convergence to homogeneity of humans attributes? In
1997, Axelrod injected the idea of homophily in \cite{Axelrod} to the
opinion game which is referred to as \textit{``bounded confidence"}
model (BCM). In this model, individuals talk to agents who are
similar to them and consequently become even more similar and agents
whose opinions are too different do not interact. The pairwise
interaction was developed later on by Deffuant~\cite{Deffuant2000} and
a synchronized version is investigated by~\cite{Krause2000,
dittmer2001consensus,blondel,Lorenz2007}. The influence of network
topology on BCM dynamics is investigated in~\cite{Weisbuch2004}.
In this model at each time step a random pair of neighbors is chosen
and the agents change their opinions if the difference of opinions are
less than a threshold. BCM succeeded in explaining the formation of
clusters, i.e. maintaining the differences between groups and not
coming to a total consensus, to some extent, but it was fragile and
sensitive.
BCM generates opinion clustering in the context of discrete,
categorical opinions rather than continuous opinions. Moreover, it is
sensitive to ``interaction noise'' in the sense that even if with
small probability agents talk to others whose opinions are not
similar, i.e. talk to agents from different clusters, then we would
see the emergence of a consensus monoculture or polarized final
state. Therefore, BCM cannot fully explain coexistence of several
different clusters in the system.
This is where Durkheim's theory \cite{Durkheim} comes into
play. Durkheim argues that the integrating forces that bind
society together and binds an individual to others is that each
individual adopts a linear combination of others' opinions. M{\"a}s in
\cite{Mas2010} adopts Durkheim's theory and at a given time $t$,
updates opinion of individual $i$ using weighted average of the
difference between opinion of agent $i$ and the rest of
population. In this manner, the system would not be sensitive to
interaction noise and therefore tendency for individualism can be
added to the system so that it would not hurt coexistence of several
clusters in the society.
Krause~\cite{Hegselmann2002} explains analytically and by simulations
what can be said about the final state of a game given its initial
profile for existing models in which opinion of each agent evolves as
weighted average of other agents in the network. Different cases are
considered where the weights are held constant or evolve over
time. Negative ties, differentiation from and xenophobia toward those
who are different, and its polarization effect is explained in
~\cite{Flache2011,Proskurnikov,Altafini}. Moreover, Anahita
in~\cite{BCMDiverse} investigates the heterogeneous system, i.e the
model in which each node has their own confidence interval.
\subsection{Distributed behavior control and flocking models}
The set of opinions held by an individual on a set of topics can be
thought of as nothing more than the state within that individual with
respect to a set of variables. The association with these states and
variables is largely an interpretation of this abstraction in order to
study a specific system. Looking at other systems in which a set of
agents with state interact, we can find interesting related work not
originating in the study of opinion models in the context of
distributed control problems in robotics. In~\cite{reif99} the
problem of distributed behavior control in autonomous robots is
investigated using a method similar to the energy potential model used
in our model. This potential model for robotic control was introduced
even earlier by Khatib~\cite{khatib86} for obstacle avoidance
purposes. In particular, the use of potential fields to model
interactions removes the need for any centralized controller or state.
Decision making can occur solely via pairwise interactions between
robots and updates within each individual robot based on these
pairwise interactions and knowledge about the interaction potential
field. There is a striking similarity in the models of flocking
behavior (e.g., boids~\cite{reynolds86}) to those in opinion dynamics
in which a desire for population consensus (cohesion) is balanced with
individual attraction or repulsion depending on proximity. Similar
noise injection mechanisms to the social tendency towards
individualism can also be found in flocking models to avoid having the
system fall into a steady state.
\section{Model Definition}
\begin{tcolorbox}
\begin{center}
\textbf{Notation used throughout the paper}
\end{center}
\begin{multicols}{2}
[
]
\begin{itemize}
\item $\psi$ potential function.
\item $\nabla \psi$ gradient of function $\psi$.
\item $\tau$ parameter for some potential functions like (tip of the) tent, (bend point of) BCM potential.
\item $\mathbb{O}$ opinion space.
\item $\mathbb{T}$ topic space, if topic space is finite, $|\mathbb{T}| = n$.
\item $o_i^{(t)}$ opinion of agent $i$ at time $t$ if there is only one topic in the system.
\item $o_i^{(t)}(s_k)$ opinion of agent $i$ at time $t$ about topic $s_k$.
\item $\dot{o}$ derivative of $o$ with respect to time.
\item $\Delta o_i^{(t+1)}(s_k) = o_i^{(t+1)}(s_k) - o_i^{(t)}(s_k) $.
\item $d_{ij}^{(t)}(s) = o_i^{(t)}(s) - o_j^{(t)}(s)$.
Difference between opinion of agents $i$ and $j$ about topic $s$ at a time $t$.
\columnbreak
\item $N(\mu, \sigma)$ normal distribution with mean $\mu$ and standard deviation $\sigma$.
\item $\alpha$ learning rate.
\item $\sigma(\mathbf{A})$ spectrum of matrix $\mathbf{A}$.
\item $\mathbf{J}$ Jacobian matrix.
\item $\mathbf{C}$ coupling matrix.
\item $c_{ij}$ inverse coupling strength of topic $i$ over topic $j$.
\item $G = (V,E)$ graph $G$ with set of nodes $V$ and set of edges $E$.
\item $n(i)$ set of neighbors of agent $i$.
\item $||.||$ Euclidean norm.
\item vectors and matrices will be bold letters.
\item $\mu_s$ mean of polarization counts
\item $\mu_p$ mean of stabilization time
\end{itemize}
\end{multicols}
\end{tcolorbox}
\subsection{Topic and opinion spaces}
First we will introduce definitions and associated notation that we will be using throughout this paper.
\theoremstyle{definition}
\begin{definition}{}
The set of topics being negotiated in the model constitute the \textit{topic space}, denoted by $\mathbb{T}$. The topic space is effectively an index set, and for $n$ topics we can assume $\mathbb{T} = \mathbb{Z}_n$, and for an uncountable set of
topics ordered on real line we have $\mathbb{T} :=[0,L]$.
\end{definition}
\theoremstyle{definition}
\begin{definition}{}
The set of all possible (numerical) opinions, denoted by $\mathbb{O}$,
is called \textit{opinion space}, and will be a subset of an m-dimensional hypercube
$ \mathcal{C}^m$, for some $m$.
\end{definition}
\begin{remark}[]
In this paper we focus on the case in which
$\mathbb{O} = \mathcal{C}^1= [0,1]$ and this leads to the
opinion state space $\mathbb{O}^N$ where $N$ is the population of the network.
\end{remark}
For a given topic we assign an opinion value to it from
the opinion space. For example if the question is
``How good is the New York Times as a news source?'' the opinion
space would be $\mathbb{O} = [0,1]$.
The interpretation of the endpoints is arbitrary, but we
will adopt the convention when discussing opinions
where zero corresponds to an absolutely negative
opinion about a topic and one corresponds to an
absolutely positive opinion. For more complex topics
the opinion space is defined by either the $d-1$
dimensional simplex where $d$ corresponds to
the number of extreme opinions that the topic
supports and an opinion would be any point
within the simplex, or d-dimensional hypercube
where each dimension is a topic's attribute. \\
For example, if we look at the the color space
defined by three primary colors - Red, Green
and Blue - which is called the color triangle, then
the color space (opinion space) would be a
triangle and favorite color of an individual is
a single point lying within the triangle.
This opinion space can be naturally embedded
inside a hypercube of dimension at least 3.
On the other hand, we can take the RGB cube as our opinion space in
which each dimension corresponds to a hue with the intensity of each
hue varying between 0 and 1. The favorite color of an individual is a
vector in $\mathbb{R}^3$ lying inside the cube. The choice of
specific opinion space used to encode specific topics is a problem
dependent choice.
\theoremstyle{definition}
\begin{definition}{}
Opinion of an individual $i$ about topic $x$ at given time $t$ is
denoted by $o_i^{(t)}(x)$. $o_i: \mathbb{T} \rightarrow \mathcal{C}^m$
is a function of both time and topic.
\end{definition}
We may drop each of the subscripts/superscripts when
the meaning is clear from the context.
\subsection{Interactions}
Individuals that interact to share and update their opinions are
connected via an \textit{interaction network} represented by a
graph. Let $G = ( V , E ) $ represent the network under consideration
where $V$ is the set of nodes and $E$ is the set of edges. The set of
all neighbors of node $i \in V$ is denoted by $n(i).$ In this
experimental results presented in this paper we only consider the fully
connected network. One might consider different graphs such as star
graph, ring, random graph, etc. and study the effect of graph
measures, such as different centralities, on the opinion game.
Let $d_{ij}^{(t)}(x):= o_i^{(t)}(x) -o_j^{(t)}(x)$,
$\Delta o_i^{(t+1)}(x) := o_i^{(t+1)}(x) - o_i^{(t)}(x) $
for arbitrary topic $x$, and $w_{ij}$ to be the amount
of influence of node $i$ on node $j$. $w_{ji}$ and $w_{ij}$
are not necessarily the same as the influence that two individuals
have on each other is not necessarily symmetric (as in the case of a teacher/student or parent/child relationship).
\theoremstyle{definition}
\begin{definition}{}
Let $G = ( V , E ) $ be a given network. A sub-graph of
G, $\hat{G} = ( \hat{V} , \hat{E} ) $, where $\hat{V} \subset V$ and $\hat{E} \subset E$,
is called a $\epsilon-$cluster if $\forall i \in \hat{V} \: \exists j \in \hat{V} s.t. \: |o_i - o_j | < \epsilon$.
\end{definition}
\subsection{Units of time}
\label{sec:time-units}
The model relates the state of opinions across the population
to an abstract notion of time in order to study the evolution and
dynamics of this opinion state. As such we must carefully
define what constitutes a unit of time and how this choice has
an impact on the evolution of the model. To start, we establish
that there exists a maximum speed at which information can be
propagated through the population in a single unit of time.
In this model where all individuals are connected by an interaction
network the minimum distance unit is a single edge in the network.
We adopt the convention then that the maximum distance that
information is allowed to flow is at most one edge from the origination
point of the information.
A single time unit therefore allows any valid set of interactions
such that this restriction on information flow distance holds.
Formally, evolution of the system in a time unit entails a subset
of individuals $\hat{V} \subset V$ and subset of edges representing
pairwise interactions $\hat{E} \subset E$. The restriction on information
flow holds when for the subgraph of $G$, $\hat{G}=(\hat{V},\hat{E})$,
there does not exist any path in $\hat{G}$ of length greater than 1.
Additional restrictions may also be imposed if interactions are directed
(e.g., in the case of a hearer and speaker where only the hearer updates their state),
such as the in-degree of all vertices in $\hat{V}$ being restricted to at most 1.
\subsection{Interaction energy potentials}
\label{sec:interaction-potentials}
In our model the strength of an interaction is dictated by an
interaction \textit{potential function} $\psi$. While, in this paper
we focus on those potentials which depend only on differences of
opinions, our model does not require this in general. As a result, our
model has most of the previous models in opinion dynamics as special
cases. This is illustrated in~\ref{sec:embedding} where we show that the
Bounded Confidence model, the DeGroot model and the French model are
special cases of our model.
Examples of potential function are illustrated in
Fig.~(\ref{fig:example-potential}). The potential function determines
the degree to which individuals must react in adjusting their own
opinion depending upon the difference of their opinion
$d_{ij}^{(t)}(x)$ when interacting with another individual on topic
$x$. The interaction rules that update opinions of individuals encode
this minimization by pushing individual opinions in the opposite
direction of the gradient of the interaction potential ($-\nabla
\psi$).
For example in the BCM, which is a special case of our model,
individuals would learn from each other when their difference
is less than a threshold and they do not interact otherwise.
This potential function is illustrated in Fig.~(\ref{fig:BCMpoten})
with a threshold of $\tau$.
The tent function illustrated in Fig.~(\ref{fig:tentfig1}) represents
the case where the individuals whose opinions difference is
less than $\tau$ attract each other by becoming closer and
repel by becoming further apart if their difference is more than
$\tau$. In the case of the skewed flat top tent potential, shown
in Fig.~(\ref{fig:skewtentfig1}), if the difference is between $\tau_l$
and $\tau_r$ then individuals are indifferent and will not change
their position. The flat top tent potential is similar to that of the
bounded confidence model, except that it not only forces
individuals to come closer together when they are sufficiently
similar but also causes them to repel when they are sufficiently different.
\begin{figure}[httb!]
\begin{subfigure}{.35\textwidth}
\centering
\includegraphics[width=1\linewidth]{Tent.png}
\caption{tent potential}
\label{fig:tentfig1}
\end{subfigure}%
\begin{subfigure}{.35\textwidth}
\centering
\includegraphics[width=1\linewidth]{FlatTopTent.png}
\caption{skewed flat top tent potential}
\label{fig:skewtentfig1}
\end{subfigure}%
\begin{subfigure}{.35\textwidth}
\centering
\includegraphics[width=1\linewidth]{GaussianFinal.png}
\caption{Gaussian Potential}
\label{fig:sfig2}
\end{subfigure}
\begin{subfigure}{.35\textwidth}
\centering
\includegraphics[width=1\linewidth]{BCM.png}
\caption{BCM Potential Function}
\label{fig:BCMpoten}
\end{subfigure}
\hspace{2.2in}
\begin{subfigure}{.35\textwidth}
\centering
\includegraphics[width=1\linewidth]{FrenchPotential.png}
\label{fig:FrenchPoten}
\caption{A Simple Potential Function}
\end{subfigure}
\caption{Potential function examples}
\label{fig:example-potential}
\end{figure}
Given a pairwise update rule based on a potential function
that dictates the attractive or repulsive relationship between
two opinions, we must define an update rule for the entire population.
Two options exist: one in which each step of the model involves
only one pair of individuals, and one in which a maximal cohort
is updated under a constraint dictated by the interaction network
such that no individual makes more than one opinion update in
a single step. In all cases the restrictions on information flow
distance discussed in Sec.~\ref{sec:time-units} must be obeyed.
We will also discuss dynamics of the continuous game where
the time is continuous and agents are continuously interacting.
\subsection{Individual and population update rules}
The model proceeds as a sequence of update steps in which
one or more individuals update their opinion state based on a
set of interactions. Update rules can be either applied one-at-a-time,
in which random pairs of individuals interact, or in a concurrent
population wide fashion. We start with the simplest case and build the model up.
The continuous game updates, of course, are done synchronously, like a physical system
with positively or negatively charged particles that are within reach
of each other and exert force on one another.
\subsubsection{One-at-a-time updates - Derivation of interaction update rule}
We let pairs of individuals interact and update their opinions.
Suppose there is one topic, $x$, in the system and two individuals
$i$ and $j$ are interacting at a time $t$. Let $\psi: \mathbb{R} \rightarrow \mathbb{R}$
be a bounded potential energy function which measures the energy in the opinion interaction:
\begin{equation}
Q_{ij}:= \psi(|d_{ij}|) = \psi(|o_i - o_j|)
\end{equation}
In order to reduce the energy of interaction, we have to move in
the opposite direction of gradient of the energy function:
\begin{equation}
\frac{\Delta d_{ij}}{\Delta t} = -\alpha \frac{\partial}{\partial d}\psi(|d|)
\end{equation}
where $\alpha$ is a parameter that increases/decreases rate
of change of opinions. We refer to $\alpha$ as the ``\textit{learning rate}''.
This would give us the update rule as follows:
\begin{equation}
\frac{ \Delta d_{ij}} {\Delta t} = \frac{ d_{ij}^{(t+1)} - d_{ij}^{(t)}} {1} = -\alpha \psi'(|d_{ij}|) \: \frac{d_{ij}^{(t)}}{|d_{ij}^{(t)}|}
\end{equation}
consequently,
\begin{equation}
d_{ij}^{(t+1)} = d_{ij}^{(t)} -\alpha \psi'(|d_{ij}|) \: \frac{d_{ij}^{(t)}}{|d_{ij}^{(t)}|}
\end{equation}
If we want to nudge both agents' opinion by the same amount we must have:
\begin{equation}\label{eq:equalNudge}
\left\{
\begin{array}{lr}
o_i^{(t+1)} - o_i^{(t)} &= -\frac{\alpha}{2} \psi'(|d_{ij}^{(t)}|) \: \frac{d_{ij}^{(t)}}{|d_{ij}^{(t)}|} \\
-(o_j^{(t+1)} - o_j^{(t)}) &= -\frac{\alpha}{2} \psi'(|d_{ij}^{(t)}|) \: \frac{d_{ij}^{(t)}}{|d_{ij}^{(t)}|}
\end{array}
\right.
\end{equation}
And therefore the update rule from time $t$ to $t+1$ is given by,
\begin{equation} \label{eq:updateNoWeights}
\left\{
\begin{array}{lr}
o_i^{(t+1)} &= o_i^{(t)} -\frac{\alpha}{2} \: \psi'(|d_{ij}^{(t)}|) \: \frac{d_{ij}^{(t)}}{|d_{ij}^{(t)}|} \vspace{.1in}\\
o_j^{(t+1)} &= o_j^{(t)} +\frac{\alpha}{2} \psi'(|d_{ij}^{(t)}|) \: \frac{d_{ij}^{(t)}}{|d_{ij}^{(t)}|}
\end{array}
\right.
\end{equation}
If the two agents have different influences on each other, then
the opinions change by different amounts. We can dictate this
fact by inserting the influence weights into Eq.\eqref{eq:updateNoWeights} as follows:
\begin{equation} \label{eq:WeightedUpdateRule}
\left\{
\begin{array}{lr}
o_i^{(t+1)} &= o_i^{(t)} - \frac{\alpha}{2} w_{ji} \: \psi'(|d_{ij}^{(t)}|) \: \frac{d_{ij}^{(t)}}{|d_{ij}^{(t)}|} \vspace{.1in}\\
o_j^{(t+1)} &= o_j^{(t)} + \frac{\alpha}{2} w_{ij} \: \psi'(|d_{ij}^{(t)}|) \: \frac{d_{ij}^{(t)}}{|d_{ij}^{(t)}|}
\end{array}
\right.
\end{equation}
Note that $o_i^{(t)}(x)$ is function of both time and the topic. So,
if there are finite number of topics in the system, then the total
energy between the two given individuals $i$ and $j$ would be the sum
of all energies of opinion interactions. The tendency for
individualism could be added to the model as well at this point. We
will use the noise model introduced in \cite{Mas2010}. By adding this
term to Eq.~\eqref{eq:WeightedUpdateRule}, we get:
\begin{equation}
\left\{
\begin{array}{lr}
o_i^{(t+1)} &= o_i^{(t)} - \frac{\alpha}{2} w_{ji} \: \psi'(|d_{ij}^{(t)}|) \; \frac{d_{ij}^{(t)}}{|d_{ij}^{(t)}|} + \xi_i(t) \vspace{.1in}\\
o_j^{(t+1)} &= o_j^{(t)} + \frac{\alpha}{2} w_{ij} \: \psi'(|d_{ij}^{(t)}|) \; \frac{d_{ij}^{(t)}}{|d_{ij}^{(t)}|}+\xi_j(t)
\end{array}
\right.
\end{equation}
where $\xi_i(t)$ is a random (sample) value from a distribution $\xi_i(t) \sim N(0,\sigma_i(t))$ associated with individual $i$ at time $t$ with mean zero and variance defined by:
\begin{equation}
\label{eq:adaptive-noise}
\sigma_i(t) = s\sum_{j \in N(i)} e^{-d_{ij}(t)}
\end{equation}
where the parameter $s$ is used to manipulate the strength
of disintegrating forces. This tendency for uniqueness increases
when the $d_{ij}$'s are small, i.e. there is high uniformity.
If $o_i^{(t+1)}$ becomes more than 1 or less than zero we
should set it to 1 or zero, respectively. This can be achieved
by applying a clamping function:
\begin{equation*}
clamp(x) =
\begin{cases} 0 & \text{if $x < 0$,} \\
x &\text{if $0 \leq x \leq 1$,} \\
1 &\text{if $x > 1$.}
\end{cases}
\end{equation*}
\subsubsection{ All neighbor interaction }
At this point we have derived a rule for updating opinions after a
single interaction. Now we can use it for the model in which at each
time step, each agent updates their opinion according to all of their
neighbors' opinion, i.e. at each time step a agent would have
interaction with all of their neighbors, and their opinion in time $t+1$
depends on all of her neighbor's opinion in time $t$. Let the node $i$
have $k$ neighbors, then the updating rule is given by:
\begin{equation}\label{eq:synchronizedUpdate}
o_i^{(t+1)} = o_i^{(t)} + \sum_{j=1}^k -\frac{\alpha}{2} \: \psi'(|d_{ij}|)\; \frac{d_{ij}}{|d_{ij}|}
\end{equation}
The influence weights can be added accordingly. The influence
weights that individual $i$ assigns to their neighbors and themselves
(resistance force) has to add up to 1, so the weight matrix $w$ is
stochastic:
$$\sum_{j \in N(i)} w_{ji} = 1$$
This is the simplest case in which there is only one potential
function for all interaction. We can also have a potential function
for each pair or each agent could have her own potential function.
\subsubsection{Bidirectional vs unidirectional information exchange}
In the previous sections we assume that information flow
is bi-directional in an interaction such that both individuals
share their opinion state and both update it based on what
they receive. It is also possible to model interactions as
uni-directional such that there is a well defined speaker
that shares information to a receiving hearer but not vice
versa. This distinction can be encoded in the network
through the use of directed versus undirected edges.
It must be noted that a single undirected edge between
individuals $i$ and $j$ is not equivalent to two directed
vertices in opposing directions ($i$ to $j$ and $j$ to $i$),
as each edge corresponds to a single communication event.
A similar effect can be accomplished with undirected graphs through
the use of zero weights. An individual $i$ that influences others will
have weights $w_{i*}$ that are not necessarily zero, but if $i$ is not
influenced at all by others then $\forall j \neq i$, $w_{ji} = 0$.
The use of weights in this manner extends to graphless models in which
connectivity is based on spatial proximity where no notion of
directedness or edge connectivity exists. We do not cover this mechanism
for individual connectivity in this paper.
\subsubsection{ Embedding other models into our model}
\label{sec:embedding}
To show BCM is special case of our model, let the potential function be
\begin{equation}
\psi(x) = \begin{cases}
x^2 & 0 \le x\leq \tau \\
\tau^2 & \tau < x\leq 1 \\
\end{cases}
\end{equation}
Then we have
\begin{equation}
\psi'(x) = \begin{cases}
2x & 0 \le x\leq \tau \\
0 & \tau < x\leq 1 \\
\end{cases}
\end{equation}
Hence, using \eqref{eq:updateNoWeights} if the two agents are
within confidence interval of each other we get:
\begin{equation} \label{eq:BCMspecial}
\left\{
\begin{array}{lr}
o_i^{(t+1)} &= o_i^{(t)} -\frac{\alpha}{2} \: (2|d_{ij}|) \: \frac{d_{ij}^{(t)}}{|d_{ij}^{(t)}|} = o_i^{(t)} - \alpha \: d_{ij}^{(t)} \vspace{.1in}\\
o_j^{(t+1)} &= o_j^{(t)} +\frac{\alpha}{2} (2|d_{ij}|) \: \frac{d_{ij}^{(t)}}{|d_{ij}^{(t)}|} = o_j^{(t)} + \alpha \: d_{ij}^{(t)}
\end{array}
\right.
\end{equation}
which is identical to the equation in section 2.1 of~\cite{Deffuant2000}.
This is the pairwise interaction and it is easy to build the synchronized version.\\
We saw in the DeGroot model we have $\mathbf{y}^{(t+1)} = \mathbf{P} \mathbf{y^{(t)}}$.
The following conditions will show the DeGroot model is a special
case of our model. Consider synchronized version of the game with
the following conditions. Define potentials
\begin{equation}
\psi_{ij} = \psi(o_i,o_j) = \begin{cases}
- p_{ij} o_i o_j & j \ne i \\
-\frac{1}{2} (p_{ii}-1) o_i^2 & j = i \\
\end{cases}
\end{equation}
then,
\begin{equation}
\frac{\partial}{o_i} \psi_{ij} = \begin{cases}
-p_{ij} o_j & j \ne i \\
-(p_{ii}-1) o_i & j = i \\
\end{cases}
\end{equation}
Now, the DeGroot model says $\mathbf{y}^{(t+1)} = \mathbf{P} \mathbf{y}^{(t)}$
which is the same as \linebreak $\mathbf{y}^{(t+1)} - \mathbf{y}^{(t)} = (\mathbf{P}-\mathbf{I}_N) \mathbf{y}^{(t)}$
where $\mathbf{I}_N$ is identity matrix of size $N$. Since we are letting
self interaction, defined by $\psi_{ii}$ for individual $i$ we have the
following, whose right hand side will still give us a gradient descent
method: (Let $\alpha = 2$)
\[
\begin{array}{lllllll}
o_i^{(t+1)} - o_i^{(t)} &=& - \psi'_{i1}(o_i^{(t)},o_1^{(t)}) - \psi'_{i2}(o_i^{(t)},o_2^{(t)}) - \cdots - \psi'_{ii}(o_i^{(t)},o_i^{(t)}) - \cdots - \psi'_{iN}(o_i^{(t)},o_N^{(t)}) \\
&=& p_{i1}o_1^{(t)} + p_{i2}o_2^{(t)} + \cdots + ( p_{ii} - 1 ) o_i^{(t)} + \cdots p_{iN}o_N^{(t)}
\end{array}
\]
and therefore,
\[
\begin{array}{lllllll}
o_i^{(t+1)} &=& - \psi'_{i1}(o_i^{(t)},o_1^{(t)}) - \psi'_{i2}(o_i^{(t)},o_2^{(t)}) - \cdots - \psi'_{ii}(o_i^{(t)},o_i^{(t)}) - \cdots - \psi'_{iN}(o_i^{(t)},o_N^{(t)}) \\
&=& p_{i1}o_1^{(t)} + p_{i2}o_2^{(t)} + \cdots + p_{ii} o_i^{(t)} + \cdots p_{iN}o_N^{(t)}
\end{array}
\]
which is the $i^{th}$ row of the DeGroot model matrix.\\
The interesting dynamics of French model~\cite{FRENCH1956} which
works with digraphs comes from the topology of the network .
Let $i$ be a node with indegree $k$ and define $N(i) = \{i_1, i_2, \cdots , i_k\}$
be the neighbors of agent $i$ who have power over it. Define the potential functions
\begin{equation}
\psi_{ij} = \psi(o_i,o_{i_j}) = \begin{cases}
- \frac{o_i o_{i_j}}{k+1} & 1 \le i_j \le k \\
\frac{k}{2(k+1)} o_i^2 & i_j = i \\
\end{cases}
\end{equation}
where $o_{i_j}^{(t)}$ is opinion of $j^{th}$ neighbor of $i$
which has power over $i$, i.e. there is an incoming edge
from $i_j$ agent to $i$, and the case $i_j = i $ is the
self-resistance element of French model, then we have
\begin{equation}
\frac{\partial}{\partial o_i} \psi_{ij} = \begin{cases}
- \frac{ o_{i_j} } { k+1 } & 1 \le i_j \le k \\
\frac{k}{(k+1)} o_i & i_j = i \\
\end{cases}
\end{equation}
and therefore, moving in the opposite of gradient of the potential function with synchronous update gives us the following step size
\[ o_i^{(t+1)} - o_i^{(t)}= \frac{1}{k+1} ( o_{i_1}^{(t)} + o_{i_2}^{(t)} + \cdots + o_{i_k}^{(t)} ) - \frac{k}{k+1} o_i^{(t)}\]
which yields the French update rule:
\[ o_i^{(t+1)} = \frac{1}{k+1} ( o_i^{(t)} + o_{i_1}^{(t)} + o_{i_2}^{(t)} + \cdots + o_{i_k}^{(t)} )\]
\subsection{Topic coupling}
\label{Topic-coupling}
In the real world, topics are not independent of each other. For example,
there might be good reasons that if someone changes their opinion
about education then they would also change their mind about social
topics in which education is a factor.
\theoremstyle{definition}
\begin{definition}{}
Change of opinion about topic $s_k$ as a result of change of opinion about topic $s_\ell$ is called \emph{coupling}.
\end{definition}
The coupling is done in two ways. The first case is the discrete case
where there are a finite number of topics in the system. In a game the
two agents can talk about some of the topics and then the coupling is
performed after the interaction is complete. In the second model
there are uncountable number of topics, at a given time step two given
individuals could talk about all topics at the same time to the extent
they please, i.e. they have the option of not talking about some of
the topics or revealing all information about a given topic in a game,
and that is captured by the so called \textit{conversation filter}
function. This generalized model consist of three steps. In this
first phase, each opinion is changed only with the force enforced by
negotiation about the given topic and in the second phase the coupling
is done between topics within each individual prior to any other
interactions. In the coupling phase some opinions might be pushed out
of the opinion space they belong to, hence, in the third phase we
would push the out of place opinions back in place.
\subsubsection{Discrete coupling}
\theoremstyle{definition}
\begin{definition}{}
Given a topic space $\mathbb{T} = \{ s_1, s_2, \cdots, s_n \}$ and let
$\mathbf{C}$ be the \emph{coupling matrix} for the system. For two
given topics $s_{\ell}$ and $s_k$, we define the \textit{inverse
coupling strength}(or \textit{inverse coupling coefficient} or
\textit{inverse correlation strength}), denoted by $c_{\ell k}$,
to be the parameter that determines how much the two topics are
influenced by each other. The coefficient $c_{\ell k}$ dictates how the
opinion $o(s_k)$ will change in response to a change in opinion
$o(s_{\ell})$ that occurs due to an interaction.
$c_{\ell k}$ does not equal $c_{k\ell}$ necessarily.
\end{definition}
A positive coupling coefficient would mean the two topics have direct
correlation and a negative coefficient means the two topics have an
inverse correlation. Topics are by definition completely correlated
with themselves, so $c_{s_is_i}=1$ for all topics $s_i$.
To derive the update rule for coupled topics, suppose that individual
$i$ has their opinion changed about a topic after an interaction. This
change is due to an external force (negotiation force) resulting from
the opinion held by the other individual on the subject discussed
during the interaction. Let $f:= 1 \cdot \Delta o_i(s_{\ell}) $ and
let topic $s_{\ell}$ to be related to topic $s_k$ with coefficient
$c_{\ell k}$. Then the same force would be applied to opinion of the
agent about subject $s_k$ and therefore:
\[f = 1 \cdot \Delta o_i (s_{\ell}) = c_{\ell k} \; \Delta o_i (s_{k}) \]
which gives us the update rule for the opinions that are correlated to the topic $s_{\ell}$:
\[ o_i^{(t+1)}(s_k) = o_i^{(t)}(s_k) + \frac{1}{c_{\ell k}} \; \Delta o_i^{(t)}(s_\ell). \]
\subsubsection{Continuous Kernel-based coupling}
In the previous section we modeled a discrete version of coupling where
there are finite number of topics in the system and each pair of topics are
coupled with their specific coupling coefficient. Here we would model the generalized
continuous topic space where the coupling coefficients are determined
via a two variable function and each pair of agents have the option of
revealing all of their information about all topics in a single game.
So, in a single time step, the two agents have the choice of talking
about as many topics as they please, to whatever extent they prefer.
For example, two agents can talk about their favorite food and favorite
chef in an interaction, but not their favorite president.
This is implemented via a conversation \textit{filter} $f_{ij}(x,t) \in [0,1]$
which can evolve over time. $f_{ij}(x,t)=0$ means they would not
talk about topic $x$, $f_{ij}(x,t)=1$ would mean they reveal all of
their information about the given topic. It will not effect the opinions that are revealed, but it determines the weight that each revealed opinion contributes to the interaction energy.\\
Let $\mathbb{T} := [0,L]$ be the topic space and $\mathcal{S}$ be the set
of measurable functions \linebreak $f : [0,L] \rightarrow [0,1]$.
Suppose further that $\psi: \mathbb{R} \rightarrow \mathbb{R}$
is bounded. Then $o_i \in \mathcal{S}$, $\forall i \in V$.
We define the interaction energy between nodes $i$ and $j$ over topic space by:
\[ Q_{ij} = \int_0^L \psi(||o_i^{(t)}(x) - o_j^{(t)}(x) ||) \; f_{ij}^{(t)}(x)dx \]
Minimizing this energy would lead us to an update rule for the agents at topic $x$.
After the conversation, each individual has updated her opinion
about any given topic via interaction and then it is time to do the
coupling in the topic network. In order to do this, we measure the
energy between two topics for node $i$ as follows
(we can have different potential energy $\phi$ for every pair of
topics $x$ and $y$, hence we use the notation $\phi(||o_i(x) - o_i(y)||,x,y)$):
\[ U_i[x,y] = \int_0^L \int_0^L \phi(||o_i(x) - o_i(y) ||, x, y) \; k(x,y) dx dy \]
Where the kernel $k(x,y)$ determines the connection strength between topics $x$ and $y$.\\
To obtain the update rule for the two individuals we take the variational derivative of functional $U$.\\
Let \[ U_i[x,y] = \int_0^L \int_0^L \phi(||o_i(x) - o_i(y) ||, x, y) \; k(x,y) dx dy \]
In ordinary calculus $\Delta f$ is computed (as a function of $x$) when $x$ changes
by $\Delta x$, here, $\delta U$ is computed when $o_i$ changes by $\delta o_i$.
{\small
\begin{eqnarray*}
\frac{d}{d \epsilon} U_i[o(x)\hspace{-1.75em}&& + \epsilon h(x) , o(y) ]\Big|_{\epsilon=0} \\ &&= \left. \bigg[\int_0^L \int_0^L \partial_1 \phi(||o(x) +\epsilon h(x) - o(y) ||, x, y) \Big( \frac{o(x)+\epsilon h(x) - o(y)}{ || o(x)+\epsilon h(x) - o(y)|| }\cdot \: h(x)\Big) k(x,y) dx dy\bigg] \right|_{\epsilon=0}\\
&&=\int_0^L \int_0^L \frac{\partial}{\partial o_i}\phi(|| o(x) - o(y) ||, x, y) \; \Big( \frac{o(x) - o(y)}{ || o(x) - o(y) || } \cdot h(x) \Big) \; k(x,y) dx dy\\
&&=\int_0^L \bigg( \int_0^L \partial_1\phi(||o(x) - o(y)||, x, y) \frac{o(x) - o(y)}{ || o(x) - o(y) || } \; k(x,y) dy \bigg)\cdot h(x) dx\\
\end{eqnarray*}
}
Therefore, the derivative which gives the updating rule would be:
\[ \frac{\delta U}{\delta o_i(x)} = \int_0^L \partial_1 \phi(||o_i(x) - o_i(y) ||, x, y) \: \frac{o_i(x) - o_i(y)}{ || o_i(x) - o_i(y) || } \; k(x,y) dy \]
which yields to:
\begin{equation}
\label{eq:Coupiling-Topics}
o_i^{(t+1)}(x) = o_i^{(t)}(x) - \alpha \int_0^L \partial_1 \phi(||o_i^{(t)}(x) - o_i^{(t)}(y) ||, x, y) \Big( \frac{o_i^{(t)}(x) - o_i^{(t)}(y) }{ || o_i^{(t)}(x) - o_i^{(t)}(y) || } \Big) \; k(x,y) \: dy
\end{equation}
For the reason explained below, we break the time step into two
parts, $(t,\: t+1/2)$ and \linebreak $(t+1/2, \: t+1)$. Therefore, the
left side of Eq.~\eqref{eq:Coupiling-Topics} is modified to
$o_i^{(\tilde{t})}(x)$, where $\tilde{t} = t + 1/2$.
\section{Dynamics of the opinion game}\label{sec:dynamics}
The family of opinion game models we have introduced is rich and
flexible. While the special case of symmetric interactions yields
gradient flows, the full range of this model is much broader. But even
in the gradient flow case, dynamics need not be simple: a sufficiently
complex potential can lead to very complex dynamical behavior, at
least computationally.
In this section we begin the exploration of model dynamics. Much of
the section focuses on the behavior in the continuous, synchronous
limit for systems in which the interaction potentials are symmetric
and identical. But even in these simpler cases we find behavior that
is not completely trivial.
We first study the behavior of a two agent system, each agent having a
different interaction potential. This is simple enough that the
behavior can be understood in detail.
Moving to three agent systems in which all potentials are the same
symmetric tent function, we find that in addition to consensus
clusters, there are entire neighborhoods of fixed points. For a
deterministic interaction order, the discrete, asynchronous game can
see subsets of these regions as fixed regions in which the system
cycles endlessly. Even when the interaction order is random, these
regions of fixed points can lead to slow drifting random walks in
opinion space. The results from these studies are immediately relevant
for systems of $N = 3k$ agents for any $k$.
We then turn to potentials that are no longer tent potentials to find
non-trivial stable fixed points. These bell shaped potentials show us
that even in the case of gradient systems (which is what we have when
all potentials are the same) there can be interesting equilibrium
states.
While our explorations in this section are admittedly just the
beginning of a more comprehensive study of the model family, what we
show here already suggests some richness in behavior that encourages
us to push ahead with the other, more complex members of this family
of models.
\subsection{Two Agent Systems}
\label{TwoAgentsSystem}
In this section, we consider the dynamics of a two agent system:
\begin{equation}
\label{eq:DifferenceVectorField}
\begin{array}{llr}
\dot{o}_1 &=& f_1(o_1,o_2) = -\frac{\partial\psi_1(|o_1-o_2|)}{\partial o_1}\\
\dot{o}_2 &=& f_2(o_1,o_2) = -\frac{\partial\psi_2(|o_1-o_2|)}{\partial o_2}
\end{array}
\end{equation}
The situation is simple enough that we can understand everything in
detail. Two observations enable us to fairly easily unravel the
dynamics of any two agent system:
\begin{enumerate}
\item the vector field in the 2-dimensional opinion state space,
$\{(o_1,o_2) \;|\; o_1,o_2\in [0,1] )\}$, is constant along $o_1 -
o_2 = c$ lines.
\item the dynamics are completely determined by the positions of the
maxima and minima of the potential functions.
\end{enumerate}
To illustrate this, we look at a couple of systems in which each
individual's interaction potential is a simple tent function, though
with different peak positions. We are able to give a complete picture
of the simple dynamics in each case.
What about the case when the potentials are not functions only of the
difference $|o_1-o_2|$? I.e. when
\begin{equation}
\label{eq:VectorField}
\begin{array}{llr}
\dot{o}_1 &=& f_1(o_1,o_2) = -\frac{\partial\psi_1(o_1,o_2)}{\partial o_1}\\
\dot{o}_2 &=& f_2(o_1,o_2) = -\frac{\partial\psi_2(o_1,o_2)}{\partial o_2}
\end{array}
\end{equation}
While this question is mostly left to future papers, we close the
section with a simple sensitivity result for the case of general
potentials.
\subsubsection{Two tents}
\label{sec:twotents}
In Fig.~(\ref{fig:EqualTipsSubFigure}) we see the case of two tent
potentials with their tips at $\tau_1$ and $\tau_2$ where $\tau_1 \leq
\tau_2$, and $1-\tau_2 = \tau_1$. In this simple case, the 2
dimensional state space is divided into regions in which the
vector field generated by the partial derivatives of the potentials are
constant. There are three cases:
\begin{enumerate}
\item When $ o_1 - o_2 < \tau_1$, the agent's opinions move
towards each other and the slope of vectors is $\frac{\Delta
o_2}{\Delta o_1} = - \frac{\tau_1}{\tau_2}$.
\item If $ \tau_1 < o_1 -
o_2 < \tau_2$, then, agent 1 moves away with agent two in pursuit,
$\frac{\Delta o_2}{\Delta o_1} = \frac{1 - \tau_1}{\tau_2} =
1$.
\item Finally, if $ o_1 - o_2 > \tau_2$ they repulse each other
towards polarization.
\end{enumerate}
We denote the region in which the opinions are
attracted towards each other by A, the region where one opinion chases
the other we denote by C, and R is used to label the region where they
repulse each other.
\begin{figure}[httb!]
\hspace{.6in}
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=1\linewidth]{equalTips.png}
\caption{ $1 - \tau_2 = \tau_1$}
\label{fig:EqualTipsSubFigure}
\end{subfigure}%
\hspace{.2in}
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=1\linewidth]{ChaseArea.png}
\caption{$1 - \tau_2 < \tau_1$}
\label{fig:ChaseAreaSubFigure}
\end{subfigure}%
\caption{Two agents vector fields}
\label{fig:Two AgentsVectorFields}
\end{figure}
In the Fig.~(\ref{fig:ChaseAreaSubFigure}) we show the case of two tent
potentials with their tips at $\tau_1$ and $\tau_2$ where $\tau_1 \leq
\tau_2$ and $1-\tau_2<\tau_1$. Again we use A, C, and R for the
attraction, chase and repulsion regions, but now the chase region is
divided into 2 sub-regions, $\text{C}_a$ and $\text{C}_r$, depending on
whether the opinion state eventually enters the attraction region or
the repulsion region.
\begin{enumerate}
\item In region A, the slope of vectors is $m = \Delta o_2 / \Delta
o_1 = -\tau_1 / \tau_2 < -1$, directed towards consensus. All
points on the line segment that connects $p = (a,a)$ to $q=
(a+\cfrac{\tau_1\tau_2}{\tau_1+\tau_2}, \: a -
\cfrac{\tau_1^2}{\tau_1+\tau_2})$ flow towards to $p = (a, \: a)$.
\item We subdivide region C into $\text{C}_a$ and $\text{C}_r$ regions:
\begin{enumerate}
\item The line $o_2 = \frac{1-\tau_1}{\tau_1}o_1 + \tau_2$ divides C
into two regions.
\item Above that line the chase eventually ends with in region A.
\item Below that line the chase eventually ends in region R.
\end{enumerate}
\item In the region R $o_1 - o_2 > \tau_2$, the opinions repulse and
$o_1$ converges to $1$ and $o_2$ to $0$.
\end{enumerate}
\begin{remark}
We are ignoring the fact that the tent system has undefined gradient
at the tent tips, though it is not hard to deduce the results if
smoothed out versions are used instead.
\end{remark}
\subsubsection{Two Smooth Potentials}
\label{sec:twolip}
When $\psi_1 = \psi_2$ or when $\psi_1$ and $\psi_2$ are piecewise
linear, the trajectories in the opinion state space are piecewise
linear and the sensitivity (separation of orbits) is very simple to
understand: mostly orbits that start close, remain close to each
other. When there are two different potentials with non-constant
derivatives, e.g. two different Gaussian-like potentials, the vector
field becomes slightly more interesting, though it is still constant
on $o_1-o_2 = c$ lines. This again results in a system whose
sensitivity to initial conditions are quite straightforward to
understand. (The simple structure of the lines of unstable fixed
points gives us the structure of the regions in which small
perturbations result in large differences in final equilibrium
positions.) When we relax the assumption that the potentials are
functions of the difference $o_1 - o_2$ alone, we can get more
interesting behavior. In this section, we close with an example of the
kind of result we can get when, instead of assuming $\psi_1$ and
$\psi_2$ are functions of $|o_1-o_2|$ alone, we assume that $\psi_1$
and $\psi_2$ are smooth.
\begin{prop}\label{driftOffProp}
Let $G$ be a network with two agents with smooth, compactly
supported potential functions, $\psi_1$ and $\psi_2$, generating a
vector field, as in Eq.~(\ref{eq:VectorField}), with Lipschitz
constant $L$. Let $p = (o_1^{(0)}, \: o_2^{(0)} )$ be the initial
opinion state of the system, and $\hat{p} = (\hat{o}_1^{(0)}, \:
\hat{o}_2^{(0)} ) $ be a perturbation of $p$. Then a bound on the
difference between the states at time $T$ is given by \[||p^T -
\hat{p}^T|| \leq || p^{(0)} - \hat{p} ^{(0)} || e^{TL^*}\] where
$L^* \sim O(L)$.
\begin{proof}
Let $\psi_1$ and $\psi_2$ be smooth, compactly supported potential
functions of nodes 1 and 2 respectively. The Lipschitz constant of
the vector field $(f_1,f_2)$ in Eq.~(\ref{eq:VectorField}) is
$L=\sqrt{L_1^2+L_2^2}$ where $L_i$ is Lipschitz constant of each
$f_i$. (The partial derivatives of $\psi_1$ and $\psi_2$ are
Lipschitz since they are continuously differentiable on a compact
set.) Define $p^{(0)} = [o_1^{(0)} \quad o_2^{(0)}]$ and $\hat{p}^{(0)} =
[\hat o_1^{(0)} \quad \hat o_2^{(0)}]$ be two initial conditions to $n$
steps of Runge-Kutta method $(A,b^T,c)$, using step size $h \le
h_0$, where $h_0L\rho(|A|) < 1$, and let $p_n$ and $\hat{p}_n$ be the
corresponding output values, then by \textbf{Lemma 319A} of
\cite{Butcher} we have:
\begin{equation}
\label{RKbound}
|| p_n - \hat{p}_n || \le (1+hL^*)^n || p_0 - \hat{p}_0 ||
\end{equation}
where $L^* = L|b^T|(I-h_0L|A|)^{-1} \mathbf{1} $ . Letting $h =
\frac{T}{n}$ we have
\begin{equation}
\dfrac{|| p_n - \hat{p}_n ||}{|| p_0 - \hat{p}_0 || } \le e^{TL^*}
\end{equation}
Since, as $n\rightarrow\infty$, $p_n \rightarrow p^T$ and $\hat p_n
\rightarrow \hat{p}^T$, this last inequality gives us the result.
\end{proof}
\end{prop}
\begin{exmp} We begin by defining $\gamma = o_1 - o_2$ and the Gaussian potential $\psi_{\mu,\sigma}(\gamma) = \frac{1}{\sigma
\sqrt{2\pi}}e^{-\frac{(\gamma - \mu)^2}{2\sigma^2}}$. Then
\begin{equation}\psi_{\mu,\sigma}''(\gamma) = \frac{1}{\sigma^3 \sqrt{2\pi}} e^{-\frac{(\gamma - \mu)^2}{2\sigma^2}} [\frac{(\gamma-\mu)^2}{\sigma^2} - 1]\end{equation}
We now choose $\psi_1 = \psi_{0.7,1}$ and $\psi_2 = \psi_{0.3,2}$. We
compute $L_1 = \sup_{|\gamma|\leq 1}|\psi''_1(\gamma)| =
|\frac{-.3992}{ \sqrt{2\pi}}|$ at $\gamma = 0$ and $L_2 =
\sup_{|\gamma|\leq 1}|\psi''_2(\gamma)|=|\frac{-.8254}{2^3
\sqrt{2\pi}}|$ at $\gamma = 1$ and consequently $L = 0.164529$ and
hence, $L^* = 0.16$. As an example let $O^{(0)} = [0.2 \quad .1]$ be
an initial state and two perturbations of it, $O_t^{(0)} = O^{(0)} +
10^{-8} \bm{v}$ and $O_p^{(0)} = O^{(0)} + 10^{-8} \bm{w}$, where
$O_t^{(0)}$ is perturbation in direction of tangent line to vector
field at the point $[0.2 \quad 0.1]$ and $O_p^{(0)}$ is perturbation
in the direction perpendicular to the tangent line. $\bm{v}$ and
$\bm{w}$ are unit vectors. Applying RK41 with step size of $h = 0.1$
and final time $T = 30$ we take $n = T/h = 300$ steps to get to the
points $O^{(300)}$, $O_t^{(300)}$ and $O_p^{(300)}$. Let $\bm{v}_1 =
O_t^{(0)} - O^{(0)}$, $\bm{\hat{v}}_1 = O_t^{(300)} - O^{(300)}$,
$\bm{w}_1 = O_p^{(0)} - O^{(0)}$ and $\bm{\hat{w}}_1 = O_p^{(300)} -
O^{(300)}$, then we have \[\cfrac{||\bm{v}_1||}{||\bm{\hat{v}}_1||} =
0.0363, \quad \cfrac{||\bm{w}_1||}{||\bm{\hat{w}}_1||} = 1.145 \]
while by Eq.~(\ref{RKbound}) we have
\[\dfrac{|| O_n - \hat O_n ||}{ || O_0 - \hat O_0 ||} \le (1 +
0.0165)^{300} = 135.58,\] showing that in this case, the bounds,
though correct, are very pessimistic.
\end{exmp}
\begin{remark}
Note that in the previous example, even though the potential
functions are no compactly supported in the unit square, they still
have the property that all derivatives are bounded, which is all the
we actually used in the previous proposition.
\end{remark}
\subsection{Dynamics of three agents}
\subsubsection{Regions of Fixed Points for Tent Potentials}
In this section we let $G = (V,E)$ be a a fully connected network with
$|E| = 3$ and study the continuous dynamical system generated by those
three interacting agents. All the interaction potentials are assumed
to be identical. The total energy
of such a gradient system is given by:
\begin{equation}\label{eq:totalEnergyDef}
\Psi_{123} = \psi_{12}(|o_1 - o_2|) + \psi_{23}(|o_2 - o_3|) + \psi_{31}(|o_3 - o_1|)
\end{equation}
where $\psi_{ij}$ is the potential assigned to the edge
between nodes $i$ and $j$. Each interaction gives us
\begin{equation}\label{eq:equalNudgeCont}
\text{$i$-$j$ interaction contribution to gradient vector field }\left\{
\begin{array}{lr}
\dot{o}_i(i,j) &= - \psi_{ij}'(|d_{ij}^{(t)}|) \: \frac{d_{ij}^{(t)}}{|d_{ij}^{(t)}|} \\
\dot{o}_j(i,j) &= - \psi_{ij}'(|d_{ij}^{(t)}|) \: \frac{-d_{ij}^{(t)}}{|d_{ij}^{(t)}|} \\
\end{array}
\right.
\end{equation}
where $d_{ij} = o_i-o_j$.
Along the diagonal lines in the coordinate subspace corresponding to
individuals $i$ and $j$, the difference
$o_i - o_j$ is constant and since both agents' behavior is enforced by
the same potential, $\psi_{ij}$, the direction of movement is either
$(1,\:-1)$ or $(-1,\:1)$. See Figure~(\ref{fig:TwoagentSpace}).
\begin{figure}[httb!]
\centering
\includegraphics[width=.6\linewidth]{TwoPersonSpace.png}
\caption{Coordinate subspace corresponding to
individuals $i$ and $j$.}
\label{fig:TwoagentSpace}
\end{figure}
By Eq.~(\ref{eq:equalNudgeCont}) we have:
\begin{equation}
\nabla \psi_{12}(|o_2 - o_1|) = (a,-a,0) , \hspace{.1in} \nabla \psi_{13}(|o_3 - o_1|) = (-b, 0, b) , \hspace{.1in} \nabla \psi_{23}(|o_3 - o_2|) = (0 , c , -c )
\end{equation}
where $a$, $b$ and $c$ can be either positive or
negative and are given by the right side of Eq.~(\ref{eq:equalNudgeCont}).
Combining these, we get that the gradient vector field generated by our potential energy is given by:
\begin{equation} \label{eq:totalGrad}
\begin{array}{lllllll}
\dot{\vv{\mathbf{o}}} &=& \nabla \Psi_{123} \\
&=& \nabla \psi_{12}(|o_1 - o_2|) + \nabla \psi_{13}(|o_1 - o_3|) + \nabla \psi_{32}(|o_3 - o_2|)\\
&=& ( a ,\: -a,\: 0) + ( 0 ,\: b,\: -b) + ( -c ,\: 0,\: c) \\
&=& ( a - c, \: b - a, \: c - b).
\end{array}
\end{equation}
Equation~(\ref{eq:totalGrad}) shows we will get fixed points if $a = b = c$.
\bigskip
Denoting the opinion state space by $\mathcal{C}_3 = \Bbb{O}^3 = [0,1]^3$,
we define \[\mathcal{D}_3 = \{(o_1, o_2, o_3)| o_i = c \in [0,1], 1 \le i
\le 3 \}\] to be the diagonal of $\mathcal{C}_3$ corresponding to
consensus states, denote the set of extreme points where the system is
polarized by \[\mathcal{E}_3 = \{ (1,0,0), (0,1,0), (0,0,1), (1,1,0),
(1,0,1), (0,1,1) \}.\]
The next theorem shows that when the common potential is a tent
function, there are regions of neutral fixed points in the system that
are neither consensus points nor polarization points.
\begin{theorem}\label{fixedPointRegion}
Let $G_3 = (V, E)$ be a fully connected network of three nodes. Let
all interaction potentials be the same tent potential function, with
its tip at $\tau = 0.5$ and maximum height $h$. Then there are
subsets of $\mathcal{C}_3\setminus (\mathcal{D}_3 \cup \mathcal{E}_3)$
with nonempty interior, which are made up of neutral fixed points.
\end{theorem}
\begin{proof} Without loss of generality, we will consider the case
$o_1 < o_2 < o_3$, other cases follows similarly. Let the point $p =
(o_1, o_2, o_3)\in\mathcal{C}_3$ such that $d_{12} = o_2 - o_1 =
\tau - \epsilon = o_3 - o_2 = d_{23}$ and $d_{13} = o_3 - o_1 = 2
(\tau - \epsilon) > \tau$. Since the potential function has the same
slope in magnitude on both sides of $\tau = 0.5$, by
~(\ref{eq:totalGrad}) we have:
\begin{equation}
\begin{array}{lllllll}
\dot{\vv{\mathbf{o}}} &=& \nabla \Psi_{123} \\
&=& ( \alpha h ,\: -\alpha h,\: 0) + ( 0 ,\: \alpha h,\: -\alpha h) + ( -\alpha h ,\: 0,\: \alpha h) \\
&=& ( 0, \: 0, \: 0)
\end{array}
\end{equation}
Moreover, the point $P$ is not an isolated
fixed point. Notice that the point $p$ can
be perturbed to $\hat{p} = p + \zeta (1, 1, 1)$
and still be a fixed point as long as
$ \tau < \hat{d}_{13} < 1 $ and $ 0 < \hat{d}_{12}, \hat{d}_{23} < \tau$,
where $\hat{d}_{ij} = | \hat{o}_i - \hat{o}_j| $.
\end{proof}
\begin{exmp} Let $G_3$ be a fully connected
network with potential tent functions,
$\tau = 0.5$ and $h = 1$, for each edge.
Let $(o_1,\: o_2,\: o_3) = (0.2, \: 0.5, \: 0.8)$.
Then each agent is pushed and pulled by
the same force, $ a = b = c = 1$ above, by
the other two agents, and therefore they
do not move, hence, we have a fixed
point and indeed a region of fixed points.
\end{exmp}
In the Thm~(\ref{fixedPointRegion}) we
considered only the case $o_1 < o_2 < o_3$.
There are five other possible cases, and all
these six regions are mutually exclusive.
The regions are given in Sec. (\ref{sec:Appendix})
and it is shown that their interior are mutually disjoint and
afterward we show that there must be
fixed points inside these regions.
\bigskip
These fixed points are not specific to three
agent systems. In fact, we can immediately extend the results to the case in
which there are $N = 3k$ agents.
\begin{corr}\label{TheCorollary} Let $G = (V,E)$ be a fully connected graph with $|V|= N =
3k$ for some $k \in \mathbb{N}$ and let all the agents have the same
tent potential function $\psi$ with peak at $0.5$. Then there the
are subsets of
$\mathcal{C}_N\setminus(\mathcal{E}_N\cup\mathcal{D}_N)$ with
non-empty interior which are comprised of fixed points.
\begin{proof}
Let $p = (o_1, \: o_2, \: o_3)$ be a
neutrally stable fixed point in the $R_1$ derived
above. Moreover, let $S_i$ be a set of $k$ agents in the
$\delta_i$ neighborhood of $o_i$, $ i \in \{1, \: 2, \: 3 \}$.
Furthermore, let
\begin{align*}
\max\{S_2\} - \min\{S_1\} < 0.5 - \epsilon \\
\max\{S_3\} - \min\{S_2\} < 0.5 - \epsilon \\
\min\{S_3\} - \max\{S_1\} > 0.5 + \epsilon
\end{align*}
Then since the potential function has constant
and the same slope size on the
whole region except $ (0.5 - \epsilon, 0.5 + \epsilon)$,
the force applied to nodes in $S_i$ by nodes in $S_j$
and nodes in $S_k$- $i \neq j \neq k $- would cancel out
and the only force working on nodes in the $S_i$
is applied by nodes in the $S_i$, and therefore,
they all would come to consensus at a point
inside the $\delta_i$ neighborhood of $o_i$,
denote it by ${o_i^*}$ and therefore,
$p^* = (o_1^*, \: o_2^*, \: o_3^*)$ would be a fixed point inside the $R_1$.
\end{proof}
\end{corr}
\bigskip
\begin{remark}
In the case of this potential $\Psi_{123}$ we know that when we are
not in at a fixed point, the rate of energy decrease is bounded
below, so we are guaranteed to converge to a fixed point in time
less than or equal to $\frac{\sup_{\mathcal{C}_3} \Psi_{123}}{\alpha h}$.
\end{remark}
Next, we look at bell-shaped potentials and the fixed
points they generate using a simple example potential.
\subsubsection{Bell shaped potential}
Take a symmetric potential function about $x = 0.5$
with zero slope at $x = 0$ and $x = 1$. We can obtain
this by taking a smooth function between zero and
$0.25$ and then rotate, reflect and shifting it to make
up the function, like Fig.~(\ref{fig:rotatedPotential}),
which corresponds to the function by Eq. (\ref{eq:RotateFunct}).
\begin{equation} \label{eq:RotateFunct}
\psi(x) = \left\{
\begin{array}{lr}
\vspace{.1in}
x^2 & \hspace{.1in} 0 \leq x \leq 0.25\\
\vspace{.1in}
- ( x - \frac{1}{2} )^2 + \frac{1}{8} & \hspace{.1in} 0.25 < x < 0.75 \\
( x - 1)^2 & 0.75 \leq x \leq 1\\
\end{array}
\right.
\end{equation}
\begin{figure}[httb!]
\centering
\includegraphics[width=.4\linewidth]{RotatedPotential.png}
\caption{ Potential function by rotations, reflections and shifts of $x^2$}
\label{fig:rotatedPotential}
\end{figure}
Next we show a symmetric bell-shaped potential function has lines of neutrally stable fixed points.
\begin{theorem}\label{Bell-Shaped-Fixed-Points}
Let $G = (V, E)$ be a fully connected network of 3
agents and let each agent to have the same
bell-shaped potential $\psi$. Then there exist
six line segments of fixed points. Moreover,
these regions can be turned into unstable
fixed point regions by perturbation of the potential function.
\begin{proof}
Let $p = (o_1^*, \: o_2^*, \: o_3^*)$ be a given fixed point in
$(\mathcal{D} \cup \mathcal{E})'$ where $o_1^* < o_2^* < o_3^*$.
Moreover, let $ f_{ij} = \frac{\alpha}{2} \psi'_{ij}$ be the force imposed
by agent $i$ on agent $j$ where $\psi'_{ij} = \psi'(d_{ij}) $.
Since $p$ is a fixed point the forces $f_{ij}$ in Fig.~(\ref{fig:BellForces})
have the same magnitude. Since $f_{21} = f_{12}$ we must have
$f_{12} = f_{32}$ in order for agent two to not move. Therefore,
$d_{12} = d_{23} = d \in (0.25, \: 0.5)$ and $d_{13} = 2d \in (0.5, \: 0.75)$.
Note that $d_{12}$ and $d_{23}$ cannot be in
$(0, \: 0.25)$ because otherwise,
$d_{13} \in (0, \: 0.5)$ and there would
not exist a repelling force. Since
$2d \in (0.5, 0.75)$ we have to have $d \in (0.25, \: 0.375)$.
\begin{figure}[httb!]
\centering
\includegraphics[width=.6\linewidth]{BellForces.png}
\caption{}
\label{fig:BellForces}
\end{figure}
Starting from agent 1, we want to have $F_1 = f_{21} + f_{31} = 0$.
Since the potential function is symmetric about $x = 0.5$
there exist the points $d \in (0.25, \:0.5)$
such that $\psi'(d) = -\psi'(2d)$ and
consequently $F_1 = 0$. The same is true for $F_2$ and $F_3$.
Neutrality of $p$ is shown below.
In our system we have $\dot{\vv{\mathbf{o}}} = \nabla \psi_{123}$.
Denote different components as follows:
\begin{equation} \label{eq:VFfunction1}
\left\{ \begin{array}{lr}
\vspace{.1in}
\dot{o}_1 &= f(o_1,o_2,o_3)\\
\vspace{.1in}
\dot{o}_2 &= g(o_1,o_2,o_3)\\
\dot{o}_3 &= h(o_1,o_2,o_3)\\
\end{array}
\right .
\end{equation}
Let $p = (o_1^* , \: o_2^* , \: o_3^*)$ be a fixed
point of the three-agent interactions. To show
these are unstable fixed points we would use
linearization of the potential about the fixed
point. Let $u = o_1 - o_1^*$, $v = o_2 - o_2^*$
and $w = o_3 - o_3^*$ be small perturbation
of the fixed point. To see whether the
perturbation grows or decays we have
to derive differential equations for $u$, $v$ and $w$:
\begin{equation}
\begin{array}{lllllll}
\dot{u} = \dot{o_1} &=& f(o_1^* + u, o_2^* + v, o_3^* + w) \\
&=& f(o_1^*, o_2^*, o_3^*) + f_{o_1} u + f_{o_3} v + f_{o_3} w + \frac{1}{2!} \begin{bmatrix}
u & v & w
\end{bmatrix} H(o_1, o_2 ,o_3 )\begin{bmatrix}
u \\
v \\
w\\
\end{bmatrix} + \cdots\\
&=& f_{o_1} u + f_{o_2} v + f_{o_3} w + O( u^2, v^2, w^2, uv, wv, uw) \\
\end{array}
\end{equation}
Where $H(o_1, o_2 ,o_3) = \begin{bmatrix}
f_{o_1o_1} & f_{o_1o_2} & f_{o_1o_3} \\
f_{o_2o_1} & f_{o_2o_2} & f_{o_2o_3} \\
f_{o_3o_1} & f_{o_3o_2} & f_{o_3o_3}
\end{bmatrix}$, called the Hessian matrix and $f_{o_i}$ is derivative of $f$ with
\vspace{.1in}
respect to $o_i$ evaluated at $( o_1^* , o_2^*, o_3^*)$. The same can be obtained for $v$ and $w$. And perturbation evolves according to
\begin{equation}\label{eq:LinearizedEq}
\begin{array}{lllllll}
\begin{bmatrix}
\dot{u} \\
\dot{v} \\
\dot{w}\\
\end{bmatrix} &=& \begin{bmatrix}
f_{o_1} & f_{o_2} & f_{o_3} \\
g_{o_1} & g_{o_2} & g_{o_3} \\
h_{o_1} & h_{o_2} & h_{o_3}
\end{bmatrix}_{(o_1^*, o_2^*, o_3^*)} \begin{bmatrix}
u \\
v \\
w\\
\end{bmatrix} + O( u^2, v^2, w^2, uv, wv, uw)\\
&=& J(o_1^*, o_2^*, o_3^*)\begin{bmatrix}
u \\
v \\
w\\
\end{bmatrix} + O( u^2, v^2, w^2, uv, wv, uw)\\
\end{array}
\end{equation}
which by dropping the quadratic terms we get the linearized system.
Dropping the learning rate $\alpha$, we have:
\begin{equation}
\begin{array}{lllllll}
\dot{o} = \nabla \Psi_{123} &=& \nabla \psi_{12} + \nabla \psi_{32} + \nabla \psi_{31} \\ &=& (\psi'_{12}(|o_2 - o_1|) \frac{o_2 - o_1}{|o_2 - o_1|}, \: - \psi'_{12}(|o_2 - o_1|) \frac{o_2 - o_1}{|o_2 - o_1|}, \: 0) + \\
& & (0, \: \psi'_{32}(|o_3 - o_2|) \frac{o_3 - o_2}{|o_3 - o_2|}, \: -\psi'_{32}(|o_3 - o_2|) \frac{o_3 - o_2}{|o_3 - o_2|}) + \\
& & ( -\psi'_{31}(|o_3 - o_1|) \frac{o_3 - o_1}{|o_3 - o_1|} , \: 0, \: \psi'_{31}(|o_3 - o_1|) \frac{o_3 - o_1}{|o_3 - o_1|}) \\
& &\\
\vspace{.1in}
&=&
\begin{bmatrix}
\psi'_{12}(|o_2 - o_1|) \frac{o_2 - o_1}{|o_2 - o_1|} -\psi'_{31}(|o_3 - o_1|) \frac{o_3 - o_1}{|o_3 - o_1|} \\
\psi'_{32}(|o_3 - o_2|) \frac{o_3 - o_2}{|o_3 - o_2|} - \psi'_{12}(|o_2 - o_1|) \frac{o_2 - o_1}{|o_2 - o_1|} \\
\psi'_{31}(|o_3 - o_1|) \frac{o_3 - o_1}{|o_3 - o_1|} -\psi'_{32}(|o_3 - o_2|) \frac{o_3 - o_2}{|o_3 - o_2|}\\
\end{bmatrix}^T\\
&=& \begin{bmatrix}
f(o_1, o_2, o_3) \\
g(o_1, o_2, o_3) \\
h(o_1, o_2, o_3) \\
\end{bmatrix}^T
\end{array}
\end{equation}
Therefore, the Jacobian matrix of~(\ref{eq:LinearizedEq}) is given by
\[\small{
\mathbf{J} = \begin{bmatrix}
\psi''_{13}(|o_3-o_1|) - \psi''_{12}(|o_2-o_1|) & \psi''_{12}(|o_2-o_1|) & -\psi''_{13}(|o_3 -o_1|) \\
\psi''_{12}(|o_2-o_1|) & -\psi''_{23}(|o_3 - o_2|) - \psi''_{12}(|o_2-o_1|) & \psi''_{23}(|o_3 - o_2|) \\
-\psi''_{13}(|o_3 - o_1|) & \psi''_{23}(|o_3 - o_2|) & \psi''_{13}(|o_3 - o_1|) - \psi''_{23}(|o_3 - o_2|)
\end{bmatrix}
}
\]
Since $d_{12} = d_{32}$ the Jacobian has the following structure:
\[
\mathbf{J} = \begin{bmatrix}
b - a & a & -b \\
a & -2a & a \\
-b & a & b-a
\end{bmatrix}
\]
The Jacobian matrix is a real symmetric matrix,
and therefore, the eigenvalues are real, and
therefore, centers or spirals do not exist.
Since there is a line of fixed points, we
cannot have three dimensions all going
to the fixed points. That can also be seen
from the structure of the Jacobian matrix.
Due to the structure of the matrix, determinant
of $J$ is zero, therefore, one of the eigenvalues
is zero, which indicates fixed points are not isolated.
Moreover, in a real symmetric matrix, number
of negative eigenvalues is the same as number of negative pivots.
The Jacobian matrix $J$ given above always have two negative eigenvalues which can
be easily seen by looking at pivots signs.
However, we can perturb
the potential function given by Eq.~(\ref{eq:RotateFunct})
using non-analytical functions at the given $d_{12}, d_{13}, d_{23}$, so that the slope is kept
the same, but the second derivative is changed so that the
Jacobian matrix has one positive eigenvalue, one negative
and a zero eigenvalue. (Two positive eigenvalues is impossible!)
We proved the theorem for $o_1 < o_2 < o_3$, the other five cases
follows similarly.
\end{proof}
\end{theorem}
\begin{exmp}
Let the potential function of the three agents be given by Eq.~(\ref{eq:RotateFunct}).
The fixed points in $(\mathcal{D} \cup \mathcal{E})'$ are of the form $p = (o_1^*, \: o_2^*, \: o_3^*) = ( o - \frac{1}{3}, \: o, \: o + \frac{1}{3})$.
The line segment generated by $o$ where $o - \frac{1}{3} \geq 0 $ and $o + \frac{1}{3} \leq 1$ is the line segment of neutrally stable fixed points since:
\[F_1 = f_{21} + f_{31} = \frac{\alpha}{2} \left( \frac{1}{3} - \frac{1}{3} \right) = 0 = F_2 = F_3 \]
For example, if $p = (1/6, 1/2, 5/6)$, then
$J(o_1^*, o_3^*, o_3^*)= \frac{1}{3}\begin{bmatrix}
-2 & 1 & 1 \\
1 & -2 & 1 \\
1 & 1 & -2
\end{bmatrix}.$
Eigenvalues of the Jacobian are $\sigma(\mathbf{J}) = \{-3, 0 \}$.
\end{exmp}
\begin{remark}
Note that there are also fixed points of the forms
$p = (a, \: a, \: a + 0.5)$, \linebreak $q = (a, a+\: 0.5, \: a)$
and $r = (a + 0.5, \: a, \: a)$ due to having derivative zero
at end points and in the middle of the potential.
\end{remark}
\begin{remark}
The bell shaped potential above is just one potential in a family of
potentials determined by strictly increasing smooth functions
$f:[0,0.25] \rightarrow \mathbb{R}^+ \cup \{0\}$ such that \linebreak $f(0) = f'(0) = 0$. The fact
that the above potential generates a stable line of fixed points
means that small enough perturbations of the bell shaped potential
retain the non-consensus, non-polarized fixed point. Furthermore,
using these bell-shaped potentials, we can obtain fixed points for
systems with $N = 3k$ agents, as was done with the tent potential in
Corollary~\ref{TheCorollary}.
\end{remark}
\subsection{Discrete game properties}
\begin{prop}\label{LiapunovDance}
Let $G_3$ be a fully connected network of three nodes and let $\psi$
be a tent potential function centered at $\tau = 0.5$, with maximum
$h$ and a learning rate $\alpha$. Let $p = (\hat{o}_1, \:
\hat{o}_2, \: \hat{o}_3)$ be a given point in the unit cube so that:
\begin{enumerate}
\item $0 < d^{(0)}_{12}, d^{(0)}_{23} < 0.5$
\item $d^{(0)}_{13} > 0.5$
\item $\alpha h < \min \{ 0.5 - d^{(0)}_{23}, \frac{d^{(0)}_{13} -
0.5}{2} \}$ and
\item the deterministic, asynchronous game order is
$(1,2)\rightarrow(2,3)\rightarrow(3,1)\rightarrow(1,2)\rightarrow\cdots$
\end{enumerate}
Then $p$ and all the points in an open neighborhood of p are
neutrally stable in the sense that $p^{(0)} = p^{(3)} = p^{(6)} =
\cdots = p^{(3k)} = \cdots$. I.e. all points in that neighborhood
define periodic orbits of with period 3.
\end{prop}
\begin{proof}
Since the step size $\alpha h < 0.5 - d^{(0)}_{23}$,
after node 1 playing with node 2, we have $d^{(1)}_{23} < 0.5$,
and therefore, nodes 2 and 3 attract each other, and node 2
is back on its starting point. And since $\alpha h < \frac{d^{(0)}_{13} - 0.5}{2}$,
we have $d^{(2)}_{13} > 0.5$, and therefore nodes 1 and 3 repel
and they both are back on their initial positions. The loop would
goes forever. Not only $p$ generates this periodic state, but also
a neighbor of $p$, $(\hat{o}_2 + \eta_1, \: \hat{o}_2 + \eta_2 , \: \hat{o}_3 + \eta_3 )$ ,
in which the conditions on the $d^{(0)}_{ij}$ and $\alpha h$
are satisfied would be a periodic path.
\end{proof}
\begin{remark}
The discrete, asynchronous, deterministic game will approximate the
continuous synchronous game as long as we stay away from the
boundaries of the 6 neutral regions: how close we can get depends on
how small the learning rate is. If the step sizes in the game we
are playing cause us to step across the boundary of the neutral
region, then we converge to either consensus or polarization,
depending on which boundary we step across. A game in which the
play order is not deterministic, but is rather random, would
generate a random walk that would eventually escape the neutral
region.
\end{remark}
\section{Experimental Results}
\label{experiments}
In this section we present experiments examining the behavior of
networks whose parameters, such as population or topology or
interaction potential parameters, are fixed except one. Unless
otherwise is stated, the stopping condition is the window-convergence
(introduced below). In these preliminary studies, we are interested
in understanding how changing a parameter changes the average
stabilization time and the population of the polarization/consensus
clusters.
\begin{remark}
In the experiments we use the update rule $o_i^{(t+1)} =
o_i^{(t)} - \frac{\alpha} {2} \psi'(|d_{ij}|) d_{ij}$ rather than
$o_i^{(t+1)} = o_i^{(t)} - \frac{\alpha} {2} \psi'(|d_{ij}|)
\frac{d_{ij}}{|d_{ij}|} $ to prevent over shooting. This might be
considered as an adaptive learning rate.
\end{remark}
\subsection{Convergence and stability}
One of the primary goals of an opinion dynamics model is to understand
the state of the system in the limit of an arbitrary number of time
steps. Does the system reach a stable steady state, oscillate between
a finite number of deterministic states, or is the outcome stochastic
or unstable? In the case where multiple final states can be reached,
what is the probability of any given state being reached relative to
others? To reason about this we must define the concepts of
convergence and stability with respect to the opinion
game. Convergence relates to a slowing of change within the system,
where we would say that the system has converged to a steady state if
the opinion state across the entire system has stopped
changing. \emph{Consensus} is a special case of convergence in which
the system has converged and the opinion state across the individuals
is one in which every individual agrees with the others. This is the
final state of the DeGroot averaging model. Our model admits converged
states in which consensus is not present (e.g., reaching a final state
in which opinions are split between two sub-populations). Stability
relates to the sensitivity of the system to changes when it has
reached a steady state. A stable system will be able to tolerate some
degree of change while converged, while an unstable system may enter
into a non-steady state when a change occurs and ultimately end up in
a new converged final state.
\theoremstyle{definition}
\subsubsection{Definition of Window-convergence}
In order to decide when a system has converged, one could just wait
until nothing changes in a sequence of interactions involving all
pairs, but this is usually not a practical approach. Instead, we will use a
notion of convergence (or pseudo-convergence) in which we continuously
observe changes in some \emph{moving window in time} and stop when
some criterion has been met.
\bigskip
Let $\omega$ be the length of that time interval over which we monitor
opinion changes. Assume there is one topic we are tracking over
time. Let $E_s$ be a matrix of agents opinions from time $t = 1$ to
time $t = s$, where the $k^{th}$ column contains the state vector of
the opinions of the $N$ agents on the topic at time $k$. In this
approach to a stopping criterion, we consider only the last $\omega$
columns of $E_s$, which we denote by $S_s$': I.e. $S_s =
E(1:N,s-\omega+1:s)$. Define the $N$ dimensional column vector $F_s$:
\[F_s(i) = \max\{\max_{j=1,...,\omega} S_s(i,j) - \min_{j=1,...,\omega}
S_s(i,j) - \eta, 0\}.\]
\bigskip
\noindent We stop when F is the zero vector.
\subsection{Effect of Initial Opinion}
In begin to see how initial opinions effect stabilization time, we
generated initial opinions using normal distributions with a mean of
$0.5$ and a range of variances. In this case, the game was played by
picking pairs at random and calculating changes in opinion for each
node/agent based on the variational model determined by a tent function
with peak at $0.5$.
Define $\mu_t$ to be the mean time to stability, $\sigma_t$ the
variance in the stabilization times, and $\mu_p$ to be the mean
fraction of final states that are polarized (as opposed to converging
to consensus). The following figures plot (a) initial opinion variance
versus the mean and standard deviation of stabilization time, $\mu_t$
and $\sigma_t$, and (b) initial opinion variance versus the mean
polarized fraction, $\mu_p$.
For a given $\sigma$, 100 initial opinions were drawn from the normal
distribution with mean 0.5 and standard deviation $\sigma$,
$N(0.5,\sigma)$. For every initial opinion, the experiment was run 100
times to begin to average out the effect of the random order in which
the game was played.
\begin{figure}[httb!]
\centering
\begin{subfigure}{.35\textwidth}
\includegraphics[width=1\linewidth]{ProperAvgStdTime.png}
\caption{Effect of initial state on stabilization time}
\label{fig:stepPolFull}
\end{subfigure}%
\hspace{.5in}
\begin{subfigure}{.35\textwidth}
\includegraphics[width=1\linewidth]{initialPolPlot.png}
\caption{Effect of initial state on polarization count}
\label{fig:stepTimeFull}
\end{subfigure}%
\caption{Initial state effects}
\label{fig:InitialEffect}
\end{figure}
\subsection{Effect of Tent Tip}
\label{full-experiments-vary-potential}
The experiments in this section are done with a fully connected
interaction graph and a tent potential function where $\tau$, the peak
of the tent, moves between 0 and 1. Three experiments were run, each
with a different strategy for picking initial conditions:
\begin{enumerate}
\item \emph{Step initials}, in which the opinions were spread out
between zero and one equally, with the same step size between them,
\item \emph{Uniform initials}, in which initial opinions were sampled from
the uniform distribution on $[0,1]$, and
\item \emph{Normal initials}, in which initial opinions were sampled
from a normal distribution. The desired
number of samples were drawn from a normal distribution with mean 0 and
standard deviation 1, and then shifted and scaled so that the samples
all lie within $[0,1]$.
\end{enumerate}
In each of these experiments there were 20 nodes/agents were involved and
for each initial condition, the game is played to convergence 1000 times.
\subsubsection{Step Initials}
Table~\ref{table:stepInitPot} and Figure~\ref{fig:StepPeak}\subref{fig:stepPolFull} show how polarization varies as the tent peak
$\tau$ moves from zero to 1. Because there is no randomness in the initial
conditions, each experiment is run 1000 times for each position of the
tent peak. This allows the effects of the randomness in the order of game
negotiations to be averaged out.
Observe that we get a probability of polarization of $0.5$ when
$\tau \approx 0.63$. Because $\tau=0$ implies always-polarization and $\tau=1$
implies always-consensus, we might be tempted to expect that at $\tau=0.5$ we
would have a probability of polarization of $0.5$. But this is a
result of identifying the domain of the potential with the opinion
space, even though this is not correct. The potential is a function of
opinion differences, not the opinions themselves. A deeper look at
this reveals that the important factors determining equilibrium states
are volumes of attractors in the opinion state spaces which are
controlled by the location of the tip of the tent. This is part of our ongoing
work to be included in a subsequent paper to follow this one.
\begin{figure}[httb!]
\centering
\begin{subfigure}{.35\textwidth}
\includegraphics[width=1\linewidth]{stepPolPeak.png}
\caption{Polarization mean}
\label{fig:StepPolPeak}
\end{subfigure}%
\hspace{.6in}
\begin{subfigure}{.35\textwidth}
\includegraphics[width=1\linewidth]{stepTimePeak.png}
\caption{Mean of stabilization time}
\label{fig:StepTimePeak}
\end{subfigure}%
\caption{Effect of tent peak position $\tau$}
\label{fig:StepPeak}
\end{figure}
\begin{table}[!htb]
\caption{Step initial opinions with different tent potentials with $\tau \in [0.5, 0.7]$.}
\begin{center}
\begin{tabular}{| l | l | l | l | l | l| l| l| l | l | l | p{1cm} |}
\hline
$\tau$ & 0.5 & 0.58 & .6 & 0.61 & 0.62 & 0.63 & 0.637 & 0.65 & 0.66 & 0.67 & .7 \\ \hline
Polarization count & 100 & 90 & 89 & 75 & 71 & 57 & 38 & 42 & 20 & 15 & 0 \\ \hline
Stabilization time & 92 & 116 & 127 & 136 & 146 & 149 & 152 & 163 & 164 & 160 & 141 \\ \hline
\end{tabular}
\end{center}
\label{table:stepInitPot}
\end{table}
\subsubsection{Uniform initials}
In this experiment we sampled the initial opinions from a uniform
distribution over $[0,1]$ 100 times, and for each sample the
simulation is run 1000 times. As a result, both the order of
negotiations and effect of initial sampling are taken into effect in
the resulting statistics. The same is done for initial opinions
sampled from normal distribution. Thus, for each value of $\tau$, 100,000
experiments are run. In Table~\ref{table:unifPotTable} we report the
normalized fraction of those experiments that resulted in
polarization by dividing the total number of polarizations by 1000.
\begin{figure}[httb!]
\centering
\begin{subfigure}{.35\textwidth}
\includegraphics[width=1\linewidth]{uniformPolPeak.png}
\caption{Polarization mean}
\label{fig:UniformPolPeak}
\end{subfigure}%
\hspace{.6in}
\begin{subfigure}{.35\textwidth}
\includegraphics[width=1\linewidth]{uniformTimePeak.png}
\caption{Mean of stabilization time}
\label{fig:UniformTimePeak}
\end{subfigure}%
\caption{Effect of $\tau$ on uniform initials}
\label{fig:UniformPeak}
\end{figure}
\begin{table}[!htb]
\caption{Uniform initial opinions with different tent potentials with $\tau \in [0.3,0.8]$.}
\begin{center}
\begin{tabular}{| l | l | l | l | l | l| l| l| l | l | l | p{1cm} |}
\hline
$\tau$ & .3 & .4 & .5 & .55 & .58 & .6 & .65 & .67 & .68 & .7 & .8 \\ \hline
$\mu_p$ & 100 & 99.81 & 95.27 & 81.98 & 70.87 & 59 & 31.69 & 22.08 & 10.3 & 6.82 & 0 \\ \hline
$\mu_s$ & 110.73 & 109.29 & 104.52 & 120.75 & 130.39 & 133.76 & 143.66 & 147.84 & 149.48 & 147.20 & 139.18 \\ \hline
\end{tabular}
\end{center}
\label{table:unifPotTable}
\end{table}
\subsubsection{Normal initials}
The initial 20 opinions here are generated by sampling the standard
normal distribution ($\mu =0$, $\sigma = 1$), and then
shifted and scaled so that they fit in the interval $[0,1]$.
\begin{figure}[httb!]
\centering
\begin{subfigure}{.35\textwidth}
\includegraphics[width=1\linewidth]{normalPolPeak.png}
\caption{Polarization mean}
\label{fig:NormalPolPeak}
\end{subfigure}%
\hspace{.6in}
\begin{subfigure}{.35\textwidth}
\includegraphics[width=1\linewidth]{normalTimePeak.png}
\caption{Mean of stabilization time}
\label{fig:NormalTimePeak}
\end{subfigure}%
\caption{Effect of $\tau$ on normal initials}
\label{fig:NormalPeak}
\end{figure}
\begin{table}[!htb]
\caption{Normal initial opinions with different tent potentials with $\tau \in [0.3,0.7]$.}
\begin{center}
\begin{tabular}{| l | l | l | l | l | l| l| l| l | l | l | p{1cm} |}
\hline
$\tau$ & .3 & .5 & .53 & .55 & .58 & .59 & .6 & .62 & .63 & .65 & .7 \\ \hline
$\mu_p$ & 100 & 87.57 & 81.02 & 63.04 & 55.31 & 50.02 & 35.45 & 32.63 & 22.32 & 18.41 & 0 \\ \hline
$\mu_s$ & 114.91 & 134.27 & 131.36 & 134.39 & 129.25 & 131.26 & 129.24 & 131.15 & 131.14 & 131.85 & 124.61 \\ \hline
\end{tabular}
\end{center}
\label{table:normalPotTable}
\end{table}
In Figure~\ref{fig:PeakVsAll} it is clear that as the peak of the tent
function increases, all experiments with different samples from
different distributions show similar behavior.
In general, one can first consider the continuous flow generated by
the vector field the potentials generate and then consider the random
walk behavior that results when we (essentially) compute the gradients
asynchronously. These two factors: geometry of state space (determined
by $\tau$) and random walk behavior (from random negotiation order),
give rise these observed results. Because the value of $\tau$
changes the volumes of the attractors of the consensus clusters and
the polarization clusters, a full explanation of these figures will
wait for the next paper.
\begin{figure}[httb!]
\centering
\includegraphics[width=.35\linewidth]{ProperPolCountsPeak.png}
\caption{Polarization mean vs potential $\tau$ parameter}
\label{fig:PeakVsAll}
\end{figure}
\subsection{Effect of Learning Rate}
In this section everything is fixed except the learning rate, which we
vary to see how it effects the final state of the system and the
stabilization time. In Figure~\ref{fig:LR-effect} initial opinions and
order of negotiations/conversations are the same, only the learning
rate is varied. As we can see, the choice of learning rate can cause the
long term equilibrium state to change from a single consensus cluster to
two polarized clusters. There remains work to be performed to understand
how sensitive an initial opinion state is to such effects as a function
of the potential, learning rate, and distribution of initial opinion
values.
\begin{figure}[httb!]
\centering
\begin{subfigure}{.35\textwidth}
\includegraphics[width=1\linewidth]{twoTenthLR.png}
\caption{$\alpha = 0.2$ }
\label{fig:LRS}
\end{subfigure}%
\hspace{.5in}
\begin{subfigure}{.35\textwidth}
\includegraphics[width=1\linewidth]{ThreeTenthLR.png}
\caption{$\alpha = 0.3$ }
\label{fig:LRL}
\end{subfigure}%
\caption{Learning rate}
\label{fig:LR-effect}
\end{figure}
\subsection{Coupling Experiment}
In this section we take a very brief peek at the effect of coupling
strength in a system with two topics and a fully connected network of
20 agents. In these 6 experiments -- one each for 6 different coupling
strengths, each time step is comprised of 10 disjoint (but random)
interactions. The same set of interactions is used for each coupling
strength. The experiments are terminated after 1000 time steps.
Without coupling, topic 1 polarizes and topic 2 converges to
consensus. Introducing a small amount of (symmetric) coupling between
the topics for all individuals, we see that the consensus state for
topic 2 is destabilized. It is worth noting that there exist a small number
of individuals that fail to reach a polarized state when coupling occurs.
The exact cause of this behavior is the subject of future work, but we
hypothesize that this is due to a small number of individuals starting
with an initial opinion state with the following properties. For topic 1,
they start with an opinion that leads to the consensus state with a set of
other individuals. On topic 2, since everyone moves to consensus there exists
only one cluster in equilibrium. When coupling is enabled, the topic 2
consensus state is perturbed and polarization occurs. In the case of the
individuals who oscillate over the long term, they likely are in a situation
where their opinion on topic 2 polarizes with the \emph{opposite} set of
individuals than they cluster with on topic 1. As such, they experience a
constant tug from each population towards the opposing polarized state and
cannot reach and stay at the polarized opinion state for either topic.
\begin{figure}[htp]
\begin{adjustbox}{varwidth=\textwidth,fbox,center}
\vspace{0.25cm}
\begin{subfigure}{.32\textwidth}
\vspace{0.25cm}
\includegraphics[width=1\linewidth]{T1-coeff-1.png}
\label{fig:NormalPolPeak2}
\end{subfigure}%
\begin{subfigure}{.32\textwidth}
\vspace{0.25cm}
\includegraphics[width=1\linewidth]{T1-coeff-1_2.png}
\label{fig:NormalPolPeak3}
\end{subfigure}%
\begin{subfigure}{.32\textwidth}
\vspace{0.25cm}
\includegraphics[width=1\linewidth]{T1-coeff-1_3.png}
\label{fig:NormalPolPeak4}
\end{subfigure
\begin{subfigure}{.32\textwidth}
\includegraphics[width=1\linewidth]{T1-coeff-1_4.png}
\label{fig:NormalPolPeak5}
\end{subfigure}%
\begin{subfigure}{.32\textwidth}
\includegraphics[width=1\linewidth]{T1-coeff-1_5.png}
\label{fig:NormalPolPeak6}
\end{subfigure}%
\begin{subfigure}{.32\textwidth}
\includegraphics[width=1\linewidth]{T1-coeff-0.png}
\label{fig:NormalPolPeak7}
\end{subfigure}%
\end{adjustbox}
\caption{Different coupling coefficients - topic 1}
\label{topic1Coup}
\end{figure}
\begin{figure}[htp]
\begin{adjustbox}{varwidth=\textwidth,fbox,center}
\begin{subfigure}{.32\textwidth}
\vspace{0.25cm}
\includegraphics[width=1\linewidth]{T2-coeff-1.png}
\label{fig:NormalPolPeak8}
\end{subfigure}%
\begin{subfigure}{.32\textwidth}
\vspace{0.25cm}
\includegraphics[width=1\linewidth]{T2-coeff-1_2.png}
\label{fig:NormalPolPeak9}
\end{subfigure}%
\begin{subfigure}{.32\textwidth}
\vspace{0.25cm}
\includegraphics[width=1\linewidth]{T2-coeff-1_3.png}
\label{fig:NormalPolPeak10}
\end{subfigure
\begin{subfigure}{.32\textwidth}
\includegraphics[width=1\linewidth]{T2-coeff-1_4.png}
\label{fig:NormalPolPeak11}
\end{subfigure}%
\begin{subfigure}{.32\textwidth}
\includegraphics[width=1\linewidth]{T2-coeff-1_5.png}
\label{fig:NormalPolPeak12}
\end{subfigure}%
\begin{subfigure}{.32\textwidth}
\includegraphics[width=1\linewidth]{T2-coeff-0.png}
\label{fig:NormalPolPeak13}
\end{subfigure}%
\end{adjustbox}
\caption{Different coupling coefficients - topic 2}
\label{topic2Coup}
\end{figure}
\section{Conclusions and future work}
The work presented in this paper introduces a potential based model
with inter-topic coupling. As defined, the model is relatively general as
we show by embedding other models from the literature in our framework.
There remain a large number of questions that can be studied based on this
work. We will provide a few noteworthy questions that we identified while
performing this research. First, a number of recent publications have studied
the impact of network properties (e.g., centrality measures) on dynamical
systems that are dependent on the network. Furthermore, it is known that
real-world systems that are best represented by a network are often not static
and instead exhibit time-varying properties - both in connectivity as well
as parameters such as edge weights.
As mentioned in places within the paper, we have chosen specific
update criteria for modeling single agent pairs interacting and
sharing opinions. Other update methods can be created that impose
differing levels of synchronous behavior, differing subpopulations,
and so on. These will likely have an effect on long term behavior of
the model and should be studied. One of the more interesting initial
results that we are studying in more detail is the cross-over point
shown in Fig.~(\ref{fig:UniformPolPeak}). Why do we observe the 50/50
polarization/consensus split occur when the tent $\tau$ parameter is
approximately 0.6? Can this value be predicted analytically from the
update rules, potential function, and model parameters?
A number of questions can be posed about the long term evolution of
the model. We can study more about dynamics of the continuous system
such as determining the conditions under which consensus/polarization
occurs. Determining a lower bound on convergence time would be another
interesting question. In particular, we would like to understand the
long term state reached when individuals oscillate due to coupling of
opinions that disagree with larger subpopulations - does this
oscillation run indefinitely, or does it damp out and eventually move
to a polarized state? Many of parameters of the models can also
be made dynamic. For example, we may allow coupling or inter-agent
weights to deviate from some equilibrium state to represent transient
phenomena (e.g., conflicts).
We believe that this work represents a noteworthy accomplishment in
opinion dynamics research for two reasons: it provides a flexible framework
for exploring variant models based on a common core, and by adopting a common
framework we are able to then use theorems about one model to reason about
others. The common model framework will allow such translation of theorems
and properties from one model to another, shedding light on models that would
be difficult to analyze directly.
\section{Appendix}\label{sec:Appendix}
Here we will show how regions of fixed points look like assuming each person or each edge has a tent potential function assigned to it with $\tau = 0.5$. The regions are given by:
\begin{itemize}
\item \:\: $R_1 = \{0 < o_2 - o_1 < \tau\} \cap \{0 < o_3 - o_2 < \tau\} \cap \{ \tau < o_3 - o_1 \leq 1 \}$, where $o_1 < o_2 < o_3$.
\item \:\: $R_2 = \{0 < o_1 - o_2 < \tau\} \cap \{0 < o_3 - o_1 < \tau\} \cap \{ \tau < o_3 - o_2 \leq 1 \}$, where $o_2 < o_1 < o_3$.
\item \:\: $R_3 = \{0 < o_3 - o_1 < \tau\} \cap \{0 < o_2 - o_3 < \tau\} \cap \{ \tau < o_2 - o_1 \leq 1 \}$, where $o_1 < o_3 < o_2$.
\item \:\: $R_4 = \{0 < o_3 - o_2 < \tau\} \cap \{0 < o_1 - o_3 < \tau\} \cap \{ \tau < o_1 - o_2 \leq 1 \}$, where $o_2 < o_3 < o_1$.
\item \:\: $R_5 = \{0 < o_1 - o_3 < \tau\} \cap \{0 < o_2 - o_1 < \tau\} \cap \{ \tau < o_2 - o_3 \leq 1 \}$, where $o_3 < o_1 < o_2$.
\item \:\: $R_6 = \{0 < o_2 - o_3 < \tau\} \cap \{0 < o_1 - o_2 < \tau\} \cap \{ \tau < o_1 - o_3 \leq 1 \}$, where $o_3 < o_2 < o_1$.
\end{itemize}
Note that in the first two cases, for $R_1$ and $R_2$ ,the condition $ o_1 < o_2$ and $o_2 < o_1$ proves that $R_1 \cap R_2 =$ \O. Similarly, $R_i \cap R_j =$ \O, $\: 1 \leq i,j \leq 6$.\\
In the first case we have \[R_1 = \{0 < o_2 - o_1 < \tau \} \cap \{0 < o_3 - o_2 < \tau \} \cap \{\tau < o_3 - o_1 \leq 1 \} \]
The first part of this region would give us two planes given by
\begin{equation}
\label{eq:plane1}
P1 \equiv o_2 - o_1 = 0 \:\: \textbf{and} \:\: P2 \equiv o_2 - o_1 = \tau
\end{equation}
This region is drawn in Fig. (\ref{fig:FirstPartofR}). And of course there are the two boundary planes, the cube walls $o_1 = 0$ and $ o_2 = 1 $, which we do not mention it, because the unit cube is the whole space we are existing in, but we would not forget about them. The planes given by Eq.~(\ref{eq:plane1}) is intersected by the two planes from the second region given by:
\begin{equation}
\label{plane2}
P3 \equiv o_3 - o_2 = 0 \:\: \textbf{and} \:\: P4 \equiv o_3 - o_2 = \tau
\end{equation}
drawn in Fig.~(\ref{fig:SecondPartofR}).
\begin{figure}[httb!]
\centering
\begin{subfigure}{.35\textwidth}
\includegraphics[width=1\linewidth]{0_y-x_tau.png}
\caption{First Part of $R_1$}
\label{fig:FirstPartofR}
\end{subfigure}%
\hspace{.5in}
\begin{subfigure}{.35\textwidth}
\includegraphics[width=1\linewidth]{0_z-y_tau.png}
\caption{Second Part of $R_1$}
\label{fig:SecondPartofR}.
\end{subfigure}%
\caption{First two parts of $R_1$}
\label{fig:FirstTwoParts}
\end{figure}
Therefore, $P1$ intersects $P3$ in a line $L_{13}$ and $P4$ in a line $L_{14}$, and the same happens for $P2$.
The normal vector of $P1$ and $P2$ is $\mathbf{n_1} = \mathbf{n_2} =\begin{bmatrix} -1 & 1 & 0 \end{bmatrix}$, and we also have $\mathbf{n_3} = \mathbf{n_4} = \begin{bmatrix} 0 & -1 & 1 \end{bmatrix}$.
Hence, the direction of $L_{13}$, $L_{14}$, $L_{23}$ and $L_{24}$ all are the same and is given by \[\mathbf{v} = \mathbf{n_1} \times \mathbf{n_3} =\begin{bmatrix} 1 & 1 & 1 \end{bmatrix} = \mathbf{n_1} \times \mathbf{n_4} = \mathbf{n_2} \times \mathbf{n_3} = \mathbf{n_2} \times \mathbf{n_4}\]
The line $L_{13}$ is given by
\begin{equation}
\label{border1}
(o_1, \: o_2, \: o_3) = t \:(1, \:1 , \: 1)
\end{equation}
and $L_{14}$ is given by
\begin{equation}
\label{border2}
(o_1, \: o_2, \: o_3) = (0,0, \: \tau) + t \: (1, \:1 , \: 1)
\end{equation}
the lines $L_{23}$ and $L_{24}$ are given below respectively:
\begin{equation}
\label{border3}
(o_1, \: o_2, \: o_3) = (0, \: \tau, \: \tau) + t \: (1, \:1 , \: 1)
\end{equation}
\begin{equation}
\label{border4}
(o_1, \: o_2, \: o_3) = (0,\: \tau , \: 2\tau) + t \: (1, \:1 , \: 1)
\end{equation}
So, the area surrounded by the first two parts of $R_1$ is given by the four planes given by Eq.~(\ref{eq:plane1}) and Eq.(\ref{plane2}) , borders are given by (\ref{border1}) - (\ref{border4}) and the surfaces of the unit cube. (Fig. \ref{fig:FirstTwoPartsIntersection})
\begin{figure}[httb!]
\centering
\includegraphics[width=.5\linewidth]{firstTwoIntersection.png}
\caption{First two parts intersection}
\label{fig:FirstTwoPartsIntersection}
\end{figure}
This area is cut by two other planes from the third part of defining $R_1$, $\{\tau < o_3 - o_1 \leq 1\}.$
\begin{equation}
\label{plane56}
P5 \equiv o_3 - o_1 = \tau \:\: \textbf{and} \:\:P6 \equiv o_3 - o_1 = 1
\end{equation}
Other than the cube's surfaces, the box has 4 sides inside the cube. Let's find the intersection of $P5$ with those 4 sides of the box. We know
\[\mathbf{n_1} = \mathbf{n_2} = \begin{bmatrix} -1 & 1 & 0 \end{bmatrix},\: \: \mathbf{n_3} = \mathbf{n_4} = \begin{bmatrix} 0 & -1 & 1 \end{bmatrix} \: \text{and} \: \:\mathbf{n_5} =\begin{bmatrix} -1 & 0 & 1 \end{bmatrix} \]
Consequently,
\[\mathbf{n_1} \times \mathbf{n_5} = \begin{bmatrix} 1 & 1 & 1 \end{bmatrix} \:\:\text{and }\mathbf{n_3} \times \mathbf{n_5} = \begin{bmatrix} -1 & -1 & -1 \end{bmatrix} \]
and we get:
\[\left\{ \begin{array}{lllllll}
L_{15} &\equiv& \: (0,\: 0,\: \tau) &+& t \: (1, \:1, \:1)\\
L_{25} &\equiv& \: (0,\: \tau,\: \tau) &+& t \: (1, \:1, \:1)\\
L_{35} &\equiv& \: (-\tau,\: 0,\: 0) &+& t \: (1, \:1, \:1)\\
L_{45} &\equiv& \: (0, \: 0,\: \tau) &+& t \: (1, \:1, \:1)\\
\end{array} \right.\]
Hence, the region is given by:
\[\left\{ \begin{array}{lllllll}
P1 \text{ from \:} L_{13} &\text{to}& L_{14}\\
P2 \text{ from \:} L_{23} &\text{to}& L_{24}\\
P3 \text{ from \:} L_{13} &\text{to}& L_{23}\\
P4 \text{ from \:} L_{14} &\text{to}& L_{24}\\
\end{array} \right.\]
This box is cut by the fifth plane, $o_3 - o_1 = \tau $. Observe that $L_{14} = L_{15} = L_{45}$ and $L_{23} = L_{25} = L_{35}$. The new box has five sides to it of which two are the unit cube walls, and the other three are P5, P2 and P4. (Fig. \ref{fig:R1})
\begin{figure}[httb!]
\centering
\begin{subfigure}{.4\textwidth}
\includegraphics[width=1\linewidth]{R111.png}
\caption{ $R_1$-Angle One}
\label{fig:R1AngleOne}
\end{subfigure}%
\hspace{.5in}
\begin{subfigure}{.4\textwidth}
\includegraphics[width=1\linewidth]{R122.png}
\caption{$R_1$-Angle Two}
\label{fig:R1AngleTwo}
\end{subfigure}%
\caption{ The $R_1$ region}
\label{fig:R1}
\end{figure}
\pagebreak
\bibliographystyle{plain}
|
\section*{\leftline{\large\bf #1}}}
\def\th#1{\noindent{\bf #1}\bgroup\it}
\def\endth{\egroup\par}
\title[The Classical Inequalities]{
Vector Lattices and $f$-Algebras: The Classical Inequalities}
\author{G. Buskes}
\address{Department of Mathematics, University of Mississippi, University,
MS 38677, USA}
\email{<EMAIL>}
\author{C. Schwanke}
\address{Unit for BMI, North-West University, Private Bag X6001, Potchefstroom,
2520, South Africa}
\email{<EMAIL>}
\date{\today}
\subjclass[2010]{46A40}
\keywords{vector lattice, $f$-algebra, Cauchy-Schwarz inequality, H\"older inequality, Minkowski inequality}
\begin{abstract}
We present some of the classical inequalities in analysis in the context of Archimedean (real or complex) vector lattices and $f$-algebras. In particular, we prove an identity for sesquilinear maps from the Cartesian square of a vector space to a geometric mean closed Archimedean vector lattice, from which a Cauchy-Schwarz inequality follows. A reformulation of this result for sesquilinear maps with a geometric mean closed semiprime Archimedean $f$-algebra as codomain is also given. In addition, a sufficient and necessary condition for equality is presented. We also prove a H\"older inequality for weighted geometric mean closed Archimedean $\Phi$-algebras, substantially improving results by Boulabiar and Toumi. As a consequence, a Minkowski inequality for weighted geometric mean closed Archimedean $\Phi$-algebras is obtained.
\end{abstract}
\maketitle
\section{Introduction}\label{S:intro}
Rich connections between the theory of Archimedean vector lattices and the classical inequalities in analysis, though hitherto little explored, were implied in \cite{BerHui} and the subsequent developments in \cite{Bo2,BusvR4,Kus2,Tou}. In particular, \cite{BusvR4} presents a relationship between the Cauchy-Schwarz inequality and the theory of multilinear maps on vector lattices, built on the analogy between disjointness in vector lattices and orthogonality in inner product spaces. The ideas in \cite{BusvR4} led to the construction of powers of vector lattices (see \cite{BoBus,BusvR2}), a theory that was recently extended to the complex vector lattice environment in \cite{BusSch2}. This paper follows the complex theme of \cite{BusSch2} and in fact contains results that are valid for both real vector lattices and complex vector lattices. We also conjoin the Cauchy-Schwarz, H\"older, and Minkowski inequalities with the theory of geometric mean closed Archimedean vector lattices, as found in \cite{Az,AzBoBus}.
We first discuss these results more closely.
In \cite[Corollary 4]{BusvR4}, the first author and van Rooij extend the classical Cauchy-Schwarz inequality as follows. If $V$ is a real vector space and $A$ is an Archimedean almost $f$-algebra then for every bilinear map $T\colon V\times V\rightarrow A$ such that
\begin{itemize}
\item[(1)] $T(v,v)\geq 0\ (v\in V)$, and
\item[(2)] $T(u,v)=T(v,u)\ (u,v\in V)$
\end{itemize}
we have
\[
T(u,v)^2\leq T(u,u)T(v,v)\ (u,v\in V).
\]
This is the classical Cauchy-Schwarz inequality when $A=\mathbb{R}$, which is equivalent to
\begin{equation}\label{(1)}
|T(u,v)|\leq\bigl(T(u,u)T(v,v)\bigr)^{1/2}=2^{-1}\inf\{\theta T(u,u)+\theta^{-1}T(v,v):\theta\in(0,\infty)\}
\end{equation}
for $u,v\in V$.
The proof of \cite[Corollary 4]{BusvR4} easily adapts to a natural complex analogue for sesquilinear maps, though the condition for equality in the classical Cauchy-Schwarz inequality (see e.g. \cite[page 3]{Con}) does not hold in this more general context (see Example~\ref{E:noclaeqco}).
In Theorem~\ref{T:CSI} of this paper, we extend both the real and complex versions of the classical Cauchy-Schwarz inequality by replacing the codomain of the sesquilinear maps with an Archimedean (real or complex) vector lattice that is closed under the infimum in \eqref{(1)} above. We also prove a convenient formula for the difference between the two sides of the Cauchy-Schwarz inequality and use it to generalize the known condition for equality in the classical case. In Corollary~\ref{C:CSI}, we obtain a Cauchy-Schwarz inequality as well as a condition for equality for sesquilinear maps with values in a semiprime Archimedean $f$-algebra that is closed under the infimum in \eqref{(1)}.
Theorem~\ref{T:HI} of this paper proves a H\"older inequality for positive linear maps between Archimedean $\Phi$-algebras that are closed under certain weighted renditions of \eqref{(1)}. Our H\"older inequality generalizes \cite[Theorem 5, Corollary 6]{Bo2} and \cite[Theorem 3.12]{Tou} by (1) weakening the assumption of uniform completeness, (2) including irrational exponents via explicit formulas without restricting the codomain to the real numbers, (3) providing a result for several variables, and (4) enabling the domain and codomain of the positive linear maps in question to be either both real $\Phi$-algebras or both complex $\Phi$-algebras.
We remark that Theorem~\ref{T:HI} is itself a consequence of Proposition~\ref{P:Mali&Me}, which in turn generalizes a reformulation of the classical H\"older inequality by Maligranda \cite[(HI$_{1}$)]{Mali}. Indeed, Proposition~\ref{P:Mali&Me} is a reinterpretation of (HI$_{1}$) for Archimedean vector lattices that simultaneously extends (HI$_{1}$) to several variables. We add that Kusraev \cite[Theorem 4.2]{Kus2} independently developed his own version of (HI$_{1}$) in the setting of uniformly complete Archimedean vector lattices. Our Proposition~\ref{P:Mali&Me} contains \cite[Theorem 4.2]{Kus2}. Noting that Proposition~\ref{P:Mali&Me} relies primarily on the Archimedean vector lattice functional calculus, it (contrary to \cite[Theorem 4.2]{Kus2}) depends only on (at most) the countable Axiom of Choice.
Finally, we employ the H\"older inequality of Theorem~\ref{T:HI} to prove a Minkowski inequality in Theorem~\ref{T:MI}.
We proceed with some preliminaries.
\section{Preliminaries}\label{S:prelims}
We refer the reader to \cite{AB,LuxZan1,Zan2} for any unexplained terminology regarding vector lattices and $f$-algebras. Throughout, $\mathbb{R}$ is used for the real numbers, $\mathbb{C}$ denotes the complex numbers, $\mathbb{K}$ stands for either $\mathbb{R}$ or $\mathbb{C}$, and the symbol for the set of strictly positive integers is $\mathbb{N}$.
An Archimedean real vector lattice $E$ is said to be \textit{square mean closed} (see \cite[page 482]{AzBoBus}) if $\sup\{ (\cos\theta)f+(\sin\theta)g:\theta\in[0,2\pi]\}$
exists in $E$ for every $f,g\in E$, and in this case we write
\[
f\boxplus g=\sup\{ (\cos\theta)f+(\sin\theta)g:\theta\in[0,2\pi]\}\ (f,g\in E).
\]
The notion of square mean closedness in vector lattices dates back to a 1973 paper by de Schipper, under the term \textit{property (E)} (see \cite[page 356]{dS}). We adopt de Schipper's definition of an Archimedean complex vector lattice but use the terminology found in \cite{AzBoBus}.
Throughout, $V+iV$ denotes the commonly used vector space complexification of a real vector space $V$. An \textit{Archimedean complex vector lattice} is a complex vector space of the form $E+iE$, where $E$ is a square mean closed Archimedean real vector lattice \cite[pages 356--357]{dS}.
An Archimedean real vector lattice will also be called an \textit{Archimedean vector lattice over $\mathbb{R}$}, and an Archimedean complex vector lattice will additionally be referred to as an \textit{Archimedean vector lattice over $\mathbb{C}$}. An \textit{Archimedean vector lattice over} $\mathbb{K}$ is a vector space that is either an Archimedean vector lattice over $\mathbb{R}$ or an Archimedean vector lattice over $\mathbb{C}$. Equivalently, an Archimedean vector lattice over $\mathbb{K}$ is a vector space over $\mathbb{K}$ that is equipped with an Archimedean modulus, as defined axiomatically by Mittelmeyer and Wolff in \cite[Definition 1.1]{MW}.
Given an Archimedean vector lattice $E+iE$ over $\mathbb{C}$, we write $\text{Re}(f+ig)=f$ and $\text{Im}(f+ig)=g\ (f,g\in E)$. For convenience, we write $\text{Re}(f)=f$ and $\text{Im}(f)=0\ (f\in E)$ when $E$ is an Archimedean vector lattice over $\mathbb{R}$. Lemma 1.2, Corollary 1.4, Proposition 1.5, and Theorem 2.2 of \cite{MW} together imply that the Archimedean modulus on an Archimedean vector lattice $E$ over $\mathbb{K}$ is given by the formula
\begin{align*}
|f|=\sup\{\text{Re}(\lambda f):\lambda\in\mathbb{K},|\lambda|=1\}\ (f\in E).
\end{align*}
In particular, for an Archimedean vector lattice $E$ over $\mathbb{R}$ we have $|f|=f\vee(-f)\ (f\in E)$, while $|f+ig|=f\boxplus g\ (f,g\in E)$ holds in any Archimedean vector lattice $E+iE$ over $\mathbb{C}$.
For an Archimedean vector lattice $E$ over $\mathbb{K}$, we define the \textit{positive cone} $E^{+}$ of $E$ by $E^{+}=\{ f\in E:|f|=f\}$, while the real vector lattice $E_{\rho}=\{ f-g:f,g\in E^{+}\}$ is called the \textit{real part} of $E$. With this notation, $E=E_{\rho}$ for every Archimedean vector lattice $E$ over $\mathbb{R}$, whereas $E=E_{\rho}+iE_{\rho}$ whenever $E$ is an Archimedean vector lattice over $\mathbb{C}$.
We say that an Archimedean vector lattice $E$ over $\mathbb{K}$ is \textit{uniformly complete} (respectively, \textit{Dedekind complete}) if $E_{\rho}$ is uniformly complete (respectively, Dedekind complete). We denote the Dedekind completion of an Archimedean vector lattice $E$ over $\mathbb{R}$ by $E^{\delta}$. Given an Archimedean vector lattice $E$ over $\mathbb{C}$, we define $E^{\delta}=(E_{\rho})^{\delta}+i(E_{\rho})^{\delta}$ and note that $E^{\delta}$ is also an Archimedean vector lattice over $\mathbb{C}$. Indeed, every Dedekind complete Archimedean vector lattice over $\mathbb{R}$ is uniformly complete \cite[Lemma 39.2, Theorem 39.4]{LuxZan1}, and every uniformly complete Archimedean vector lattice over $\mathbb{R}$ is square mean closed \cite[Section 2]{BeuHuidP}.
An Archimedean vector lattice $E$ over $\mathbb{R}$ is said to be \textit{geometric mean closed} (see \cite[page 486]{AzBoBus}) if $\inf\{\theta f+\theta^{-1}g:\theta\in(0,\infty)\}$ exists in $E$ for every $f,g\in E^{+}$, and in this case we write
\[
f\boxtimes g=2^{-1}\inf\{\theta f+\theta^{-1}g:\theta\in(0,\infty)\}\ (f,g\in E^{+}).
\]
We define an Archimedean vector lattice over $\mathbb{K}$ to be \textit{geometric mean closed} (respectively, \textit{square mean closed}) if $A_{\rho}$ is geometric mean closed (respectively, square mean closed). Thus every Archimedean vector lattice over $\mathbb{C}$ is square mean closed.
We next provide some basic information regarding Archimedean $f$-algebras that will be needed throughout this paper.
The multiplication on an Archimedean $f$-algebra $A$ canonically extends to a multiplication on $A+iA$. We call an Archimedean vector lattice $A$ over $\mathbb{K}$ an \textit{Archimedean $f$-algebra over $\mathbb{K}$} if $A_{\rho}$ is an $f$-algebra. If in addition $A$ has a multiplicative identity then we say that $A$ is an \textit{Archimedean $\Phi$-algebra over $\mathbb{K}$}. It was proved in \cite[Corollary 10.4]{dP} (also see \cite[Theorem 142.5]{Zan2}) that every Archimedean $\Phi$-algebra over $\mathbb{R}$ (and therefore every Archimedean $\Phi$-algebra over $\mathbb{K}$) is semiprime.
The multiplication on an Archimedean $f$-algebra over $\mathbb{K}$ will be denoted by juxtaposition throughout. For Archimedean $f$-algebras $A$ and $B$ over $\mathbb{K}$, we say that a map $T\colon A\rightarrow B$ is \textit{multiplicative} if $T(ab)=T(a)T(b)\ (a,b\in A)$.
Let $A$ be an Archimedean $f$-algebra over $\mathbb{K}$, and let $n\in\mathbb{N}$ and $a\in A^{+}$. If there exists a unique element $r$ of $A^{+}$ such that $r^{n}=a$, we write $r=a^{1/n}$ and say that $a^{1/n}$ exists. If $A$ is an Archimedean semiprime $f$-algebra and $a,r\in A^{+}$ satisfy $r^{n}=a$ then $r=a^{1/n}$ \cite[Proposition 2(ii)]{BeuHui}. Given $m,n\in\mathbb{N}$, we write $a^{m/n}=(a^{m})^{1/n}$, provided $(a^{m})^{1/n}$ exists.
Every uniformly complete semiprime Archimedean $f$-algebra $A$ over $\mathbb{R}$ is geometric mean closed (see \cite[Theorem 2.21]{Az}) and
\begin{equation}\label{(2)}
f\boxtimes g=(fg)^{1/2}\ (f,g\in A^{+}).
\end{equation}
The formula \eqref{(2)} also holds in the weaker condition when $A$ is geometric mean closed. In fact, the proof of \eqref{(2)} in \cite[Theorem 2.21]{Az} does not require uniform completeness, as illustrated in the next proposition. Since \cite{Az} is not widely accessible, we reproduce the proof of \cite[Theorem 2.21]{Az}, while (trivially) extending this theorem to include complex vector lattices.
\begin{proposition}\label{P:gmfa}
Let $A$ be a semiprime Archimedean $f$-algebra over $\mathbb{K}$. If $A$ is geometric mean closed then $f\boxtimes g=(fg)^{1/2}\ (f,g\in A^{+})$.
\end{proposition}
\begin{proof}
Evidently, $A_{\rho}$ is a semiprime Archimedean $f$-algebra over $\mathbb{R}$. Let $f,g\in A^{+}$, and let $C$ be the $f$-subalgebra of $A_{\rho}$ generated by the elements $f,g,f\boxtimes g$. Suppose that $\phi\colon C\rightarrow\mathbb{R}$ is a nonzero multiplicative vector lattice homomorphism. Using \cite[Proposition 2.20]{Az} or \cite[Corollary 3.13]{BusSch} (first equality), we obtain
\begin{align*}
\phi(f\boxtimes g)=\phi(f)\boxtimes\phi(g)=\bigl(\phi(f)\phi(g)\bigr)^{1/2}=\bigl(\phi(fg)\bigr)^{1/2}.
\end{align*}
Therefore,
\begin{align*}
\phi\bigl((f\boxtimes g)^{2}\bigr)=\bigl(\phi(f\boxtimes g)\bigr)^{2}=\phi(fg).
\end{align*}
Since the set of all nonzero multiplicative vector lattice homomorphisms from $C$ into $\mathbb{R}$ separates the points of $C$ (see \cite[Corollary 2.7]{BusdPvR}), we have $(f\boxtimes g)^{2}=fg$.
\end{proof}
We conclude this section with some basic terminology regarding Archimedean vector lattices over $\mathbb{K}$.
Given an Archimedean vector lattice $E$ over $\mathbb{C}$, we define the complex conjugate
\begin{align*}
\overline{f+ig}=f-ig\ (f,g\in E_{\rho}).
\end{align*}
Since every Archimedean vector lattice over $\mathbb{R}$ canonically embeds into an Archimedean vector lattice over $\mathbb{C}$ (see \cite[Theorem 3.3]{BusSch2}), the previous definition also makes sense in Archimedean vector lattices over $\mathbb{R}$ (via such an embedding). If $E$ is an Archimedean vector lattice over $\mathbb{K}$ then the familiar identities $\text{Re}(f)=2^{-1}(f+\bar{f})$ and $\text{Im}(f)=(2i)^{-1}(f-\bar{f})$ are valid for every $f\in E$.
Let $V$ be a vector space over $\mathbb{K}$, and suppose that $F$ is an Archimedean vector lattice over $\mathbb{K}$. A map $T\colon V\times V\rightarrow F$ is called \textit{positive semidefinite} if $T(v,v)\geq 0$ for every $v\in V$. If $T(u,v)=\overline{T(v,u)}$ for each $u,v\in V$ then $T$ is said to be \textit{conjugate symmetric}. We say that $T$ is \textit{sesquilinear} if
\begin{itemize}
\item[(1)] $T(\alpha u_{1}+\beta u_{2},v)=\alpha T(u_{1},v)+\beta T(u_{2},v)\ (\alpha,\beta\in\mathbb{K}, u_{1},u_{2},v\in V)$, and
\item[(2)] $T(u,\alpha v_{1}+\beta v_{2})=\bar{\alpha}T(u,v_{1})+\bar{\beta}T(u,v_{2})\ (\alpha,\beta\in\mathbb{K}, u,v_{1},v_{2}\in V)$.
\end{itemize}
\section{A Cauchy-Schwarz Inequality}\label{S:CSI}
We prove a Cauchy-Schwarz inequality for sesquilinear maps with a geometric mean closed Archimedean vector lattice over $\mathbb{K}$ as codomain (Theorem~\ref{T:CSI}) and with a geometric mean closed semiprime Archimedean $f$-algebra over $\mathbb{K}$ as codomain (Corollary~\ref{C:CSI}). A necessary and sufficient condition for equality in Theorem~\ref{T:CSI} and Corollary~\ref{C:CSI} is given via an explicit formula for the difference between the two sides in the Cauchy-Schwarz inequality of Theorem 3.1 mentioned above. However, Example~\ref{E:noclaeqco} illustrates that the condition for equality in the classical Cauchy-Schwarz inequality fails in Theorem~\ref{T:CSI} and Corollary~\ref{C:CSI}.
We proceed to the main result of this section.
\begin{theorem}\label{T:CSI} \textnormal{\textbf{(Cauchy-Schwarz Inequality)}}
Let $V$ be a vector space over $\mathbb{K}$, and suppose that $F$ is a geometric mean closed Archimedean vector lattice over $\mathbb{K}$. If a map $T\colon V\times V\rightarrow F$ is positive semidefinite, conjugate symmetric, and sesquilinear then
\begin{itemize}
\item[(1)] $\underset{z\in\mathbb{K}\setminus\{ 0\}}{\inf}\{|z|^{-1}T(zu-v,zu-v)\}$ exists in $F\ (u,v\in V)$,
\item[(2)] $|T(u,v)|=T(u,u)\boxtimes T(v,v)-2^{-1}\underset{z\in\mathbb{K}\setminus\{ 0\}}{\inf}\{|z|^{-1}T(zu-v,zu-v)\}\ (u,v\in V)$,
\item[(3)] $|T(u,v)|\leq T(u,u)\boxtimes T(v,v)\ (u,v\in V)$, and
\item[(4)] $|T(u,v)|=T(u,u)\boxtimes T(v,v)$ if and only if $\underset{z\in\mathbb{K}\setminus\{ 0\}}{\inf}\{|z|^{-1}T(zu-v,zu-v)\}=0$.
\end{itemize}
\end{theorem}
\begin{proof}
Suppose that $T\colon V\times V\rightarrow F$ is a positive semidefinite, conjugate symmetric, sesquilinear map. Consider $T$ as a map from $V\times V$ to $F^{\delta}$. Let $u,v\in V$, and put $\theta\in (0,\infty)$. Using the sesquilinearity of $T$ and the identity $\text{Re}(f)=2^{-1}(f+\bar{f})$ for $f\in F^{\delta}$, we obtain
\begin{align*}
T(\theta u-v,u-\theta^{-1}v)&=T(\theta u,u)+T(v,\theta^{-1}v)-T(\theta u,\theta^{-1}v)-T(v,u)\\
&=\theta T(u,u)+\theta^{-1}T(v,v)-2\text{Re}\bigl(T(u,v)\bigr).
\end{align*}
Therefore,
\begin{align*}
\text{Re}\bigl(T(u,v)\bigr)=2^{-1}\bigl(\theta T(u,u)+\theta^{-1}T(v,v)\bigr)-(2\theta)^{-1}T(\theta u-v,\theta u-v).
\end{align*}
In particular, for every $\lambda\in S=\{\lambda\in\mathbb{K}:|\lambda|=1\}$ we have
\begin{align*}
\text{Re}\bigl(\lambda T(u,v)\bigr)=\text{Re}\bigl(T(\lambda u,v)\bigr)=2^{-1}\bigl(\theta T(u,u)+\theta^{-1}T(v,v)\bigr)-(2\theta)^{-1}T(\theta\lambda u-v,\theta\lambda u-v).
\end{align*}
Thus we obtain
\begin{align*}
|T(u,v)|&=\underset{\lambda\in S}{\sup}\bigl\{\text{Re}\bigl(\lambda T(u,v)\bigr)\bigr\}\\
&=2^{-1}\underset{\lambda\in S}{\sup}\bigl\{\theta T(u,u)+\theta^{-1}T(v,v)-\theta^{-1}T(\theta\lambda u-v,\theta\lambda u-v)\bigr\},
\end{align*}
where the suprema above are in $F^{\delta}$. Since $-\theta^{-1}T(\theta\lambda u-v,\theta\lambda u-v)\leq 0\ (\lambda\in S)$, it follows that $\underset{\lambda\in S}{\sup}\bigl\{-\theta^{-1}T(\theta\lambda u-v,\theta\lambda u-v)\bigr\}$ exists in $F^{\delta}$. Moreover,
\begin{align*}
\underset{\lambda\in S}{\sup}&\bigl\{\theta T(u,u)+\theta^{-1}T(v,v)-\theta^{-1}T(\theta\lambda u-v,\theta\lambda u-v)\bigr\}\\
&=\theta T(u,u)+\theta^{-1}T(v,v)+\underset{\lambda\in S}{\sup}\bigl\{-\theta^{-1}T(\theta\lambda u-v,\theta\lambda u-v)\bigr\}\\
&=\theta T(u,u)+\theta^{-1}T(v,v)-\underset{\lambda\in S}{\inf}\bigl\{ \theta^{-1}T(\theta\lambda u-v,\theta\lambda u-v)\bigr\}.
\end{align*}
Therefore,
\begin{align*}
2^{-1}\bigl(\theta T(u,u)+\theta^{-1}T(v,v)\bigr)=|T(u,v)|+2^{-1}\underset{\lambda\in S}{\inf}\bigl\{\theta^{-1}T(\theta\lambda u-v,\theta\lambda u-v)\bigr\}.
\end{align*}
Since $\underset{\lambda\in S}{\inf}\bigl\{\theta^{-1}T(\theta\lambda u-v,\theta\lambda u-v)\bigr\}\geq 0\ (\theta\in(0,\infty))$, it follows that
\begin{align*}
\underset{\theta\in(0,\infty)}{\inf}\Bigl\{\underset{\lambda\in S}{\inf}\{\theta^{-1}T(\theta\lambda u-v,\theta\lambda u-v)\}\Bigr\}
\end{align*}
exists in $F^{\delta}$. Then with the infima still in $F^{\delta}$,
\begin{align*}
T(u,u)\boxtimes T(v,v)&=\underset{\theta\in(0,\infty)}{\inf}\Bigl\{|T(u,v)|+2^{-1}\underset{\lambda\in S}{\inf}\{\theta^{-1}T(\theta\lambda u-v,\theta\lambda u-v)\}\Bigr\}\\
&=|T(u,v)|+2^{-1}\underset{\theta\in(0,\infty)}{\inf}\Bigl\{\underset{\lambda\in S}{\inf}\{\theta^{-1}T(\theta\lambda u-v,\theta\lambda u-v)\}\Bigr\}\\
&=|T(u,v)|+2^{-1}\underset{\lambda\in S,\theta\in(0,\infty)}{\inf}\{\theta^{-1}T(\theta\lambda u-v,\theta\lambda u-v)\}\\
&=|T(u,v)|+2^{-1}\underset{z\in\mathbb{K}\setminus\{ 0\}}{\inf}\{|z|^{-1}T(zu-v,zu-v)\}.
\end{align*}
Thus, now in $F$, the equality
\begin{align*}
2^{-1}\underset{z\in\mathbb{K}\setminus\{ 0\}}{\inf}\{|z|^{-1}T(zu-v,zu-v)\}=\bigl(T(u,u)\boxtimes T(v,v)-|T(u,v)|\bigr)\
\end{align*}
authenticates statement (1) of this theorem and implies (in $F$)
\begin{align*}
|T(u,v)|=T(u,u)\boxtimes T(v,v)-2^{-1}\underset{z\in\mathbb{K}\setminus\{ 0\}}{\inf}\{|z|^{-1}T(zu-v,zu-v)\}.
\end{align*}
Therefore, part (2) is verified. Statements (3) and (4) immediately follow from (2).
\end{proof}
As a consequence of Theorem~\ref{T:CSI} and Proposition~\ref{P:gmfa}, we obtain the following.
\begin{corollary}\label{C:CSI}
Let $V$ be a vector space over $\mathbb{K}$, and suppose that $A$ is a geometric mean closed semiprime Archimedean $f$-algebra over $\mathbb{K}$. If $T\colon V\times V\rightarrow A$ is a positive semidefinite, conjugate symmetric, sesquilinear map then
\begin{itemize}
\item[(1)] $\underset{z\in\mathbb{K}\setminus\{ 0\}}{\inf}\{|z|^{-1}T(zu-v,zu-v)\}$ exists in $A\ (u,v\in V)$,
\item[(2)] $|T(u,v)|^{2}=\Bigl(\bigl(T(u,u)T(v,v)\bigr)^{1/2}-2^{-1}\underset{z\in\mathbb{K}\setminus\{ 0\}}{\inf}\{|z|^{-1}T(zu-v,zu-v)\}\Bigr)^{2}\ (u,v\in V)$,
\item[(3)] $|T(u,v)|^{2}\leq T(u,u)T(v,v)\ (u,v\in V)$, and
\item[(4)] $|T(u,v)|^{2}=T(u,u)T(v,v)$ if and only if $\underset{z\in\mathbb{K}\setminus\{ 0\}}{\inf}\{|z|^{-1}T(zu-v,zu-v)\}=0$.
\end{itemize}
\end{corollary}
Statement (3) in Corollary~\ref{C:CSI} follows from part (2) of the corollary as well as \cite[Proposition 2(iii)]{BeuHui}. For $\mathbb{K}=\mathbb{R}$, this statement is contained in \cite[Corollary 4]{BusvR4}. Similarly, the given statement is contained in the complex analogue of \cite[Corollary 4]{BusvR4} (mentioned in the introduction) when $\mathbb{K}=\mathbb{C}$. Part (2) of Corollary~\ref{C:CSI}, however, depends on the uniqueness of square roots, which implies the semiprime property in Archimedean almost $f$-algebras. Indeed, let $A$ be an Archimedean almost $f$-algebra and suppose $a^{2}=b^{2}$ implies $a=b\ (a,b\in A^{+})$. Let $a$ be a nilpotent in $A$. Then $a^{3}=0$ (see \cite[Theorem 3.2]{BerHui2}), and thus $a^{4}=0$. Using the uniqueness of square roots twice, we obtain $a=0$. Finally, every semiprime almost $f$-algebra is automatically an $f$-algebra (\cite[Theorem 1.11(i)]{BerHui2}).
The special case where $A=\mathbb{K}$ in the inequality of Corollary~\ref{C:CSI} is the classical Cauchy-Schwarz inequality. Thus we know in this special case that $|T(u,v)|^{2}=T(u,u)T(v,v)$ if and only if there exist $\alpha,\beta\in\mathbb{K}$, not both zero, such that $T(\beta u+\alpha v,\beta u+\alpha v)=0$ (see, e.g., \cite[page 3]{Con}). This criterion no longer holds for Theorem~\ref{T:CSI} nor Corollary~\ref{C:CSI}.
\begin{example}\label{E:noclaeqco}
Define $T\colon \mathbb{K}^{2}\times\mathbb{K}^{2}\rightarrow\mathbb{K}^{2}$ by
\begin{align*}
T\bigl((z_{1},z_{2}),(w_{1},w_{2})\bigr)=(z_{1}\bar{w_{1}},z_{2}\bar{w_{2}})\ \bigl((z_{1},z_{2}),(w_{1},w_{2})\in\mathbb{K}^{2}\bigr).
\end{align*}
Since $\mathbb{C}^{2}=\mathbb{R}^{2}+i\mathbb{R}^{2}$, we see that $\mathbb{K}^{2}$ is a geometric mean closed semiprime Archimedean $f$-algebra over $\mathbb{K}$ with respect to the coordinatewise vector space operations, coordinatewise ordering, and coordinatewise multiplication. Also, $T$ is a positive semidefinite, conjugate symmetric, sesquilinear map. Note that
\[
|T\bigl((1,0),(0,1)\bigr)|^{2}=T\bigl((1,0),(1,0)\bigr)T\bigl((0,1),(0,1)\bigr).
\]
Suppose there exist $\alpha,\beta\in\mathbb{K}$, not both zero, for which
\[
T\bigl(\beta(1,0)+\alpha(0,1),\beta(1,0)+\alpha(0,1)\bigr)=0.
\]
Then $(|\beta|^{2},|\alpha|^{2})=(0,0)$, a contradiction.
\end{example}
\section{A H\"older Inequality}\label{S:HI}
We prove a H\"older inequality for positive linear maps between weighted geometric mean closed Archimedean $\Phi$-algebras over $\mathbb{K}$ in this section, extending \cite[Theorem 5, Corollary 6]{Bo2} by Boulabiar and \cite[Theorem 3.12]{Tou} by Toumi. We begin with some definitions.
Let $A$ be an Archimedean $f$-algebra over $\mathbb{K}$, and suppose that $n\in\mathbb{N}$. As usual, we write $\prod\limits_{k=1}^{n}a_{k}=a_{1}\cdots a_{n}$ for $a_{1},\dots,a_{n}\in A$.
For every $r_{1},\dots,r_{n}\in(0,1)$ such that $\sum\limits_{k=1}^{n}r_{k}=1$, we define a \textit{weighted geometric mean} $\gamma_{r_{1},\dots,r_{n}}:\mathbb{R}^{n}\rightarrow\mathbb{R}$ by
\begin{align*}
\gamma_{r_{1},\dots,r_{n}}(x_{1},\dots,x_{n})=\prod\limits_{k=1}^{n}|x_{k}|^{r_{k}}\ (x_{1},\dots,x_{n}\in\mathbb{R}).
\end{align*}
The weighted geometric means are concave on $(\mathbb{R}^{+})^{n}$ as well as continuous and positively homogeneous on $\mathbb{R}^{n}$. Moreover, for each $r_{1},\dots,r_{n}\in(0,1)$ with $\sum\limits_{k=1}^{n}r_{k}=1$, it follows from \cite[Lemma 3.6(iii)]{BusSch} that
\begin{align*}
\gamma_{r_{1},\dots,r_{n}}(x_{1},\dots,x_{n})=\inf\Bigl\{\sum\limits_{k=1}^{n}r_{k}\theta_{k}x_{k}:\theta_{k}\in(0,\infty),\ \prod\limits_{k=1}^{n}\theta_{k}^{r_{k}}=1\Bigr\}\ (x_{1},\dots,x_{n}\in\mathbb{R}^{+}).
\end{align*}
An Archimedean vector lattice $E$ over $\mathbb{K}$ is said to be \textit{weighted geometric mean closed} if $\inf\Bigl\{\sum\limits_{k=1}^{n}r_{k}\theta_{k}|f_{k}|:\theta_{k}\in(0,\infty),\ \prod\limits_{k=1}^{n}\theta_{k}^{r_{k}}=1\Bigr\}$ exists in $E$ for every $f_{1},\dots,f_{n}\in E$ and every $r_{1},\dots,r_{n}\in(0,1)$ with $\sum\limits_{k=1}^{n}r_{k}=1$. In this case, we write
\begin{align*}
\bigtriangleup\limits_{k=1}^{n}(f_{k},r_{k})=\inf\Bigl\{\sum\limits_{k=1}^{n}r_{k}\theta_{k}|f_{k}|:\theta_{k}\in(0,\infty),\ \prod\limits_{k=1}^{n}\theta_{k}^{r_{k}}=1\Bigr\}\ (f_{1},\dots,f_{n}\in E).
\end{align*}
Let $(X,\mathcal{M},\mu)$ be a measure space, and let $p\in(1,\infty)$. It follows from \cite[(HI$_{1}$)]{Mali} by Maligranda that $|f|^{1/p}|g|^{1-1/p}\in L_{1}(X,\mu)$ for $f,g\in L_{1}(X,\mu)$, and
\begin{align*}
||\ |f|^{1/p}|g|^{1-1/p}\ ||_{1}\leq||f||_{1}^{1/p}||g||_{1}^{1-1/p}.
\end{align*}
Following Maligranda's proof, we redevelop and extend \cite[(HI$_{1}$)]{Mali} to a multivariate version in the setting of positive operators between vector lattices.
\begin{proposition}\label{P:Mali&Me}
Let $E$ and $F$ be weighted geometric mean closed Archimedean vector lattices over $\mathbb{K}$, and suppose that $r_{1},\dots,r_{n}\in(0,1)$ satisfy $\sum\limits_{k=1}^{n}r_{k}=1$.
\begin{itemize}
\item[(1)] For each positive linear map $T\colon E\rightarrow F$,
\begin{align*}
T\Bigl(\bigtriangleup\limits_{k=1}^{n}(f_{k},r_{k})\Bigr)\leq\bigtriangleup\limits_{k=1}^{n}\bigl(T(|f_{k}|),r_{k}\bigr)\ (f_{1},\dots,f_{n}\in E).
\end{align*}
\item[(2)] If $T\colon E\rightarrow F$ is a linear map then $T$ is a vector lattice homomorphism if and only if
\begin{align*}
T\Bigl(\bigtriangleup\limits_{k=1}^{n}(f_{k},r_{k})\Bigr)=\bigtriangleup\limits_{k=1}^{n}\bigl(T(f_{k}),r_{k}\bigr)\ (f_{1},\dots,f_{n}\in E).
\end{align*}
\item[(3)] If $G$ is a (not necessarily weighted geometric mean closed) vector sublattice of $E$, $T\colon G\rightarrow F$ is a vector lattice homomorphism, and $f_{1},\dots,f_{n},\bigtriangleup\limits_{k=1}^{n}(f_{k},r_{k})\in G$ then
\[
T\Bigl(\bigtriangleup\limits_{k=1}^{n}(f_{k},r_{k})\Bigr)=\bigtriangleup\limits_{k=1}^{n}\bigl(T(f_{k}),r_{k}\bigr).
\]
\end{itemize}
\end{proposition}
\begin{proof}
(1) Assume $T\colon E\rightarrow F$ is a positive linear map. Let $f_{1},\dots,f_{n}\in E$, and suppose $\theta_{1},\dots,\theta_{n}\in(0,\infty)$ are such that $\prod\limits_{k=1}^{n}\theta_{k}^{r_{k}}=1$. From the positivity and linearity of $T$ we have
\begin{align*}
T\Bigl(\bigtriangleup\limits_{k=1}^{n}(f_{k},r_{k})\Bigr)\leq T\Bigl(\sum\limits_{k=1}^{n}r_{k}\theta_{k}|f_{k}|\Bigr)=\sum\limits_{k=1}^{n}r_{k}\theta_{k}T(|f_{k}|).
\end{align*}
Then
\begin{align*}
T\Bigl(\bigtriangleup\limits_{k=1}^{n}(f_{k},r_{k})\Bigr)\leq\inf\Bigl\{\sum\limits_{k=1}^{n}r_{k}\theta_{k}T(|f_{k}|):\theta_{k}\in(0,\infty),\ \prod\limits_{k=1}^{n}\theta_{k}^{r_{k}}=1\Bigr\}.
\end{align*}
(2) Suppose $T\colon E\rightarrow F$ is a vector lattice homomorphism. It follows from \cite[Theorem 3.7(2)]{BusSch} that $\gamma_{r_{1},\dots,r_{n}}(f_{1},\dots,f_{n})$, which is defined via functional calculus \cite[Definition 3.1]{BusdPvR}, exists in $E$ for every $f_{1},\dots,f_{n}\in E^{+}$ and
\[
\gamma_{r_{1},\dots,r_{n}}(f_{1},\dots,f_{n})=\bigtriangleup\limits_{k=1}^{n}(f_{k},r_{k})\ (f_{1},\dots,f_{n}\in E^{+}).
\]
It is readily checked using \cite[Definition 3.1]{BusdPvR} and the identity
\[
\gamma_{r_{1},\dots,r_{n}}(x_{1},\dots,x_{n})=\gamma_{r_{1},\dots,r_{n}}(|x_{1}|,\dots,|x_{n}|)\ (x_{1},\dots,x_{n}\in\mathbb{R})
\]
that $\gamma_{r_{1},\dots,r_{n}}(f_{1},\dots,f_{n})$ exists in $E$ for all $f_{1},\dots,f_{n}\in E$ and
\[
\gamma_{r_{1},\dots,r_{n}}(f_{1},\dots,f_{n})=\gamma_{r_{1},\dots,r_{n}}(|f_{1}|,\dots,|f_{n}|)\ (f_{1},\dots,f_{n}\in E).
\]
It follows that
\[
\gamma_{r_{1},\dots,r_{n}}(f_{1},\dots,f_{n})=\bigtriangleup\limits_{k=1}^{n}(f_{k},r_{k})\ (f_{1},\dots,f_{n}\in E).
\]
By \cite[Theorem 3.11]{BusSch} (second equality), we have for all $f_{1},\dots,f_{n}\in E$,
\begin{align*}
T\bigl(\bigtriangleup\limits_{k=1}^{n}(f_{k},r_{k})\bigr)&=T\bigl(\gamma_{r_{1},\dots,r_{n}}(f_{1},\dots,f_{n})\bigr)\\
&=\gamma_{r_{1},\dots,r_{n}}\bigl(T(f_{1}),\dots,T(f_{n})\bigr)=\bigtriangleup\limits_{k=1}^{n}\bigl(T(f_{k}),r_{k}\bigr).
\end{align*}
On the other hand, assume $T\colon E\rightarrow F$ is a linear map and
\begin{align*}
T\Bigl(\bigtriangleup\limits_{k=1}^{n}(f_{k},r_{k})\Bigr)=\bigtriangleup\limits_{k=1}^{n}\bigl(T(f_{k}),r_{k}\bigr)\ (f_{1},\dots,f_{n}\in E).
\end{align*}
From \cite[Theorem 3.11]{BusSch}, we conclude that
\begin{align*}
T\bigl(\gamma_{r_{1},\dots,r_{n}}(f_{1},\dots,f_{n})\bigr)=\gamma_{r_{1},\dots,r_{n}}\bigl(T(f_{1}),\dots,T(f_{n})\bigr)\ (f_{1},\dots,f_{n}\in E),
\end{align*}
and (since $\gamma_{r_{1},\dots,r_{n}}(x,\dots,x)=|x|\ (x\in\mathbb{R})$) that $T$ is a vector lattice homomorphism.
(3) Let $G$ be a vector sublattice of $E$, and let $T:G\rightarrow F$ be a vector lattice homomorphism. Let $\mathcal{D}$ be the collection of all weighted geometric means. That is, let
\[
\mathcal{D}=\left\{\gamma_{r_{1},\dots,r_{n}}:r_{1},\dots,r_{n}\in(0,1)\ \text{and}\ \sum\limits_{k=1}^{n}r_{k}=1\right\}.
\]
It follows from \cite[Theorem 3.7(2)]{BusSch} that $F$ is $\mathcal{D}$-complete (see \cite[Definition 3.2]{BusSch}). By \cite[Theorem 3.17]{BusSch}, $T$ uniquely extends to a vector lattice homomorphism $T^{\mathcal{D}}:G^{\mathcal{D}}\rightarrow F$, where $G^{\mathcal{D}}$ denotes the $\mathcal{D}$-completion of $G$ (see \cite[Definition 3.10]{BusSch}). Note that $G^{\mathcal{D}}$ is $\mathcal{D}$-complete by definition. Thus \cite[Theorem 3.7(2)]{BusSch} implies that $G^{\mathcal{D}}$ is weighted geometric mean closed. An appeal to (2) now verifies (3).
\end{proof}
Let $A$ be an Archimedean $\Phi$-algebra over $\mathbb{K}$. If $A$ is uniformly complete then $a^{1/n}$ exists in $A$ for every $a\in A^{+}$ and every $n\in\mathbb{N}$ (see \cite[Corollary 6]{BeuHui}). It follows that $a^{q}$ exists in $A$ for every $a\in A^{+}$ and every $q\in\mathbb{Q}\cap(0,\infty)$. The assumption of uniform completeness in \cite[Corollary 6]{BeuHui} can be weakened, which is the content of our next lemma.
\begin{lemma}\label{L:a^q}
Let $A$ be a weighted geometric mean closed Archimedean $\Phi$-algebra over $\mathbb{K}$. Then $a^{q}$ exists in $A$ for all $a\in A^{+}$ and $q\in\mathbb{Q}\cap(0,\infty)$. Furthermore, if $q_{1},\dots,q_{n}\in\mathbb{Q}\cap(0,1)$ are such that $\sum\limits_{k=1}^{n}q_{k}=1$ then for every $a_{1},\dots,a_{n}\in A^{+}$,
\begin{align*}
\prod\limits_{k=1}^{n}a_{k}^{q_{k}}=\bigtriangleup\limits_{k=1}^{n}(a_{k},q_{k})\in A.
\end{align*}
\end{lemma}
\begin{proof}
Denote the unit element of $A$ by $e$, and let $a\in A^{+}$. In order to prove that $a^{q}$ is defined in $A$ for every $q\in\mathbb{Q}\cap(0,\infty)$, it suffices to verify that $a^{1/n}$ exists in $A$ for all $n\in\mathbb{N}$. To this end, let $n\in\mathbb{N}\setminus\{1\}$. Let $C$ be the Archimedean $\Phi$-subalgebra of $A_{\rho}$ generated by the elements $a\in A^{+}$ and
\[
b=\inf\{n^{-1}\theta_{1}a+(1-n^{-1})\theta_{2} e:\theta_{1},\theta_{2}\in(0,\infty),\ \theta_{1}^{1/n}\theta_{2}^{1-1/n}=1\}\in A^{+}.
\]
Suppose that $\omega\colon C\rightarrow\mathbb{R}$ is a nonzero multiplicative vector lattice homomorphism. It follows that $\omega(e)=1$. Using Proposition~\ref{P:Mali&Me}(3) (third equality), we obtain
\begin{align*}
\omega(b^{n})&=\omega(b)^{n}\\
&=\big(\omega(\inf\{n^{-1}\theta_{1}a+(1-n^{-1})\theta_{2} e:\theta_{1},\theta_{2}\in(0,\infty),\ \theta_{1}^{1/n}\theta_{2}^{1-1/n}=1\})\bigr)^{n}\\
&=\big(\inf\{n^{-1}\theta_{1}\omega(a)+(1-n^{-1})\theta_{2} :\theta_{1},\theta_{2}\in(0,\infty),\ \theta_{1}^{1/n}\theta_{2}^{1-1/n}=1\})\bigr)^{n}\\
&=\Bigl(\gamma_{\frac{1}{n},\frac{n-1}{n}}\bigl(\omega(a),1\bigr)\Bigr)^{n}=\Bigl(\bigl(\omega(a)\bigr)^{1/n}\Bigr)^{n}=\omega(a).
\end{align*}
Since the set of all nonzero multiplicative vector lattice homomorphisms $\omega\colon C\rightarrow\mathbb{R}$ separates the points of $C$ (see \cite[Corollary 2.7]{BusdPvR}), we have in $A_{\rho}$ that $b^{n}=a$. But then $a^{1/n}=b$.
Finally, a similar proof verifies that
\begin{align*}
\prod\limits_{k=1}^{n}a_{k}^{q_{k}}=\bigtriangleup\limits_{k=1}^{n}(a_{k},q_{k})
\end{align*}
for every $a_{1},\dots,a_{n}\in A^{+}$ and all $q_{1},\dots,q_{n}\in\mathbb{Q}\cap(0,1)$ such that $\sum_{k=1}^{n}q_{k}=1$.
\end{proof}
We next use the proof of Lemma~\ref{L:a^q} as a guide to define strictly positive irrational powers of positive elements in weighted geometric mean closed Archimedean $\Phi$-algebras in an intrinsic manner that does not require representation theory dependent on more than the countable Axiom of Choice. For $r\in(0,\infty)$, define
\begin{align*}
\lfloor r\rfloor=\max\{n\in\mathbb{N}\cup\{0\}:n\leq r\}\ \text{and}\ \tilde{r}=r-\lfloor r\rfloor.
\end{align*}
\begin{definition}\label{D:a^r}
Suppose that $A$ is a weighted geometric mean closed Archimedean $\Phi$-algebra over $\mathbb{K}$. Let $e$ be the unit element of $A$. For $a\in A^{+}$ and $r\in(0,\infty)$, define
\begin{align*}
a^{r}=a^{\lfloor r\rfloor}\inf\{\tilde{r}\theta_{1}a+(1-\tilde{r})\theta_{2} e:\theta_{1},\theta_{2}\in(0,\infty),\ \theta_{1}^{\tilde{r}}\theta_{2}^{1-\tilde{r}}=1\},
\end{align*}
where $a^{\lfloor r\rfloor}$ is taken to equal $e$ in the case where $\lfloor r\rfloor=0$.
\end{definition}
By Lemma~\ref{L:a^q}, the above definition of strictly positive real exponents extends the natural definition of strictly positive rational exponents previously discussed. We next give an easy corollary of Proposition~\ref{P:Mali&Me}(3).
\begin{corollary}\label{C:T(a^r)=T(a)^r}
Let $A$ and $B$ be weighted geometric mean closed Archimedean $\Phi$-algebras over $\mathbb{K}$ with unit elements $e$ and $e'$, respectively. Suppose $C$ is a $\Phi$-subalgebra of $A$ and that $T\colon C\rightarrow B$ is a multiplicative vector lattice homomorphism such that $T(e)=e'$. Let $a\in A^{+}$ and $r\in(0,\infty)$. If $a,a^{r}\in C$ then $T(a^{r})=\bigl(T(a)\bigr)^{r}$.
\end{corollary}
The following lemma verifies some familiar (and needed for Theorems~\ref{T:HI} and \ref{T:MI}) arithmetical rules for positive real exponents in geometric mean closed Archimedean $\Phi$-algebras over $\mathbb{K}$.
\begin{lemma}\label{L:powerrules}
Let $p,q\in(0,\infty)$, and let $A$ be a weighted geometric mean closed Archimedean $\Phi$-algebra over $\mathbb{K}$. For each $a\in A^{+}$, the following hold.
\begin{itemize}
\item[(1)] $(a^{p})^{q}=a^{pq}$.
\item[(2)] $a^{p}a^{q}=a^{p+q}$.
\end{itemize}
\end{lemma}
\begin{proof}
We prove (1), leaving the similar proof of (2) to the reader. To this end, let $a\in A^{+}$. Let $C$ be the real $\Phi$-subalgebra of $A_{\rho}$ generated by $a, a^{\tilde{p}}, (a^{p})^{\tilde{q}}$, and $a^{\overset{\sim}{pq}}$, and note that $a, a^{p}, (a^{p})^{q}, a^{pq}\in C$. If $\omega\colon C\rightarrow\mathbb{R}$ is a nonzero multiplicative vector lattice homomorphism then $\omega$ is surjective and thus $\omega(e)=1$. Using Corollary~\ref{C:T(a^r)=T(a)^r}, we obtain
\begin{align*}
\omega\bigl((a^{p})^{q}\bigr)&=\bigl(\omega(a^{p})\bigr)^{q}=\bigl(\omega(a)\bigr)^{pq}=\omega(a^{pq}).
\end{align*}
Since the nonzero multiplicative vector lattice homomorphisms separate the points of $C$ (see \cite[Corollary 2.7]{BusdPvR}), we conclude that $(a^{p})^{q}=a^{pq}$.
\end{proof}
In light of Definition~\ref{D:a^r}, the second part of Lemma~\ref{L:a^q} can now be improved to include irrational exponents. The proof of Lemma~\ref{L:a^r} uses real-valued multiplicative vector lattice homomorphisms, similar to what is found in the proofs of Lemmas \ref{L:a^q} and \ref{L:powerrules}. Therefore, the proof is omitted.
\begin{lemma}\label{L:a^r}
Let $A$ be a weighted geometric mean closed Archimedean $\Phi$-algebra over $\mathbb{K}$. If $r_{1},\dots,r_{n}\in(0,1)$ are such that $\sum\limits_{k=1}^{n}r_{k}=1$ then for every $a_{1},\dots,a_{n}\in A^{+}$,
\begin{align*}
\prod\limits_{k=1}^{n}a_{k}^{r_{k}}=\bigtriangleup\limits_{k=1}^{n}(a_{k},r_{k}).
\end{align*}
\end{lemma}
We proceed with the main theorem of this section.
\begin{theorem}\label{T:HI} \textnormal{\textbf{(H\"older Inequality)}}
Let $p_{1},\dots,p_{n}\in(1,\infty)$ with $\sum\limits_{k=1}^{n}p_{k}^{-1}=1$. Assume $A$ is a weighted geometric mean closed Archimedean $\Phi$-algebra over $\mathbb{K}$.
\begin{itemize}
\item[(1)] If $B$ is also a weighted geometric mean closed Archimedean $\Phi$-algebra over $\mathbb{K}$ and $T\colon A\rightarrow B$ is a positive linear map then
\begin{align*}
T\Bigl(\prod\limits_{k=1}^{n}|a_{k}|\Bigr)\leq\prod\limits_{k=1}^{n}\bigl(T(|a_{k}|^{p_{k}})\bigr)^{1/p_{k}}\ (a_{1},\dots,a_{n}\in A).
\end{align*}
\item[(2)] If $B$ is a weighted geometric mean closed Archimedean vector lattice over $\mathbb{K}$ and $T\colon A\rightarrow B$ is a positive linear map then
\begin{align*}
T\Bigl(\prod\limits_{k=1}^{n}|a_{k}|\Bigr)\leq\bigtriangleup\limits_{k=1}^{n}\bigl(T(|a_{k}|^{p_{k}}),1/p_{k}\bigr)\ (a_{1},\dots,a_{n}\in A).
\end{align*}
\end{itemize}
\end{theorem}
\begin{proof}
We only prove (1) since the proof of (2) is similar. To this end, let $B$ be a weighted geometric mean closed Archimedean $\Phi$-algebra over $\mathbb{K}$, and suppose $T\colon A\rightarrow B$ is a positive linear map. Using Lemma~\ref{L:powerrules}(1) (first equality), Lemma~\ref{L:a^r} (second equality and last equality), and Proposition~\ref{P:Mali&Me}(1) (for the inequality), we have
\begin{align*}
T\Bigl(\prod\limits_{k=1}^{n}|a_{k}|\Bigr)&=T\Bigl(\prod\limits_{k=1}^{n}(|a_{k}|^{p_{k}})^{1/p_{k}}\Bigr)=T\Bigl(\bigtriangleup\limits_{k=1}^{n}(|a_{k}|^{p_{k}},1/p_{k})\Bigr)\\
&\leq\bigtriangleup\limits_{k=1}^{n}\bigl(T(|a_{k}|^{p_{k}}),1/p_{k}\bigr)=\prod\limits_{k=1}^{n}\bigl(T(|a_{k}|^{p_{k}})\bigr)^{1/p_{k}}.
\end{align*}
\end{proof}
\section{A Minkowski Inequality}\label{S:MI}
We employ the H\"older inequality in Theorem~\ref{T:HI}(1) to prove a Minkowski inequality in the setting of Archimedean $\Phi$-algebras over $\mathbb{K}$ in this section.
\begin{theorem}\label{T:MI} \textnormal{\textbf{(Minkowski Inequality)}}
Let $p\in(1,\infty)$. Suppose that $A$ and $B$ are both weighted geometric mean closed Archimedean $\Phi$-algebras over $\mathbb{K}$. For every positive linear map $T\colon A\rightarrow B$, we have
\begin{align*}
\Bigl(T\Bigl(\bigl|\sum\limits_{k=1}^{n}a_{k}\bigr|^{p}\Bigr)\Bigr)^{1/p}\leq\sum\limits_{k=1}^{n}\bigl(T(|a_{k}|^{p})\bigr)^{1/p}\ (a_{1},\dots,a_{n}\in A).
\end{align*}
\end{theorem}
\begin{proof}
We prove the result for $n=2$ and note that the rest of the proof follows from a standard induction argument. To this end, let $T\colon A\rightarrow B$ be a positive linear map, and assume that $a,b\in A$. Let $q\in(1,\infty)$ satisfy $q^{-1}+p^{-1}=1$. By Lemma~\ref{L:powerrules}(2) (first equality), Theorem~\ref{T:HI}(1) (second inequality), and Lemma~\ref{L:powerrules}(1) (third equality),
\begin{align*}
T(|a+b|^{p})&=T(|a+b|^{p-1}|a+b|)\\
&\leq T\bigl(|a+b|^{p-1}(|a|+|b|)\bigr)=T(|a+b|^{p-1}|a|)+T(|a+b|^{p-1}|b|)\\
&\leq T\bigl((|a+b|^{p-1})^{q}\bigr)^{1/q}T(|a|^{p})^{1/p}+T\bigl((|a+b|^{p-1})^{q}\bigr)^{1/q}T(|b|^{p})^{1/p}\\
&=T(|a+b|^{p})^{1/q}T(|a|^{p})^{1/p}+T(|a+b|^{p})^{1/q}T(|b|^{p})^{1/p}\\
&=T(|a+b|^{p})^{1/q}\bigl(T(|a|^{p})^{1/p}+T(|b|^{p})^{1/p}\bigr).
\end{align*}
Setting $f=T(|a+b|^{p})$ and $g=T(|a|^{p})^{1/p}+T(|b|^{p})^{1/p}$, we have $f\leq f^{1/q}g$. Next let $C$ be the vector $\Phi$-subalgebra of $B$ generated by $f,f^{1/p},f^{1/q}$, and $g$. Let $\omega\colon C\rightarrow\mathbb{R}$ be a nonzero multiplicative vector lattice homomorphism, so that $\omega(e)=1$, where $e$ is the unit element of $A$. Using Corollary~\ref{C:T(a^r)=T(a)^r}, we have
\[
\omega(f)\leq\omega(f^{1/q}g)=\omega(f^{1/q})\omega(g)=\omega(f)^{1/q}\omega(g).
\]
Thus if $\omega(f)\neq 0$ then $\omega(f)^{1/p}\leq\omega(g)$. Of course, $\omega(f)^{1/p}\leq\omega(g)$ also holds in the case that $\omega(f)=0$, since $g\in B^{+}$. By Corollary~\ref{C:T(a^r)=T(a)^r} again, $\omega(f^{1/p})\leq\omega(g)$. Since the collection of all nonzero multiplicative vector lattice homomorphisms separate the points of $C$ (see \cite[Corollary 2.7]{BusdPvR}), we conclude that $f^{1/p}\leq g$. Therefore, we obtain
\[
T(|a+b|^{p})^{1/p}\leq T(|a|^{p})^{1/p}+T(|b|^{p})^{1/p}.
\]
\end{proof}
|
\section{Introduction}
Since the formulation of AdS/CFT correspondence there has been a great effort in trying to understand how gravity could possibly emerge from the degrees of freedom of the dual field theory. In this context entanglement entropy has been a promising tool. Entanglement entropy has been extensively studied not only because it is an order parameter for quantum phase transitions \cite{amico}, but because the celebrated proposal for the computation of holographic entanglement entropy \cite{Ryu:2006bv} has given a geometric interpretation to a quantity that is intrinsically quantum mechanical. This geometric interpretation has helped in building further connections between the gauge and the gravity sides of the duality \cite{Bhattacharya:2014vja} \cite{Faulkner:2013ica} \cite{Lashkari:2014kda}.\\
\indent Since entanglement entropy is hard to calculate theoretically and difficult to measure experimentally, it may be useful to find other quantum information quantities that could be understood holographically.
One quantity that has been recently explored is the Quantum Information Metric (QIM). It is defined on an infinite-dimensional space of all the deformations induced by all possible operators away from the unperturbed theory. The authors of \cite{MIyaji:2015mia} and \cite{Bak:2015jxd} have focused only on deformations induced by a single marginal operator. In that particular case one can argue that the QIM can be constructed from the on shell action of a Janus type solution \cite{Bak:2003jk}. Since this solution is generally not available the authors of \cite{MIyaji:2015mia} have suggested that the Janus solution could be replaced by a probe brane. This prescription is limited to the case of deformation induced by marginal operators, in addition it reproduces results only qualitatively and up to an order one constant.
In this note we take a different approach to the study of the QIM, using perturbative techniques. We focus on deformations induced by scalar primaries. Since the QIM measures the distance between two infinitesimally separated states we are interested only in infinitesimal deformations. From the bulk point of view this means that the scalar field dual to the operator that induces the deformation on the CFT side is going to be considered as a perturbative excitation of an unperturbed background (dual to the CFT state that we are deforming). This allows us to extend previous results to deformations induced by any primary scalar operators (not necessarily marginal) and to explore configurations in which the Janus solution is not available. The spirit of this paper is to present the perturbative approach as a new tool for the holographic computation of the QIM.
The paper is organized as follows: after giving a quick introduction to the QIM (section \ref{QIM into}) we compute this quantity holographically in different set ups. Here is a summary of our main results:
\begin{itemize}
\item We construct the QIM for the ground state of a $d$-dimensional CFT deformed by a primary operator of dimension $\Delta>d/2+1$ (section \ref{CFT GS}). This extends the results available in the existing literature where the holographic computation was performed only in the case of a marginal deformation.
\item In section \ref{globalQIM} we put the CFT on a cylinder. We compute the QIM holographically for a marginal deformation in any dimension. We also give a concrete example of non marginal deformation in 1+1 dimensions.
\item In section \ref{TFD} we reproduce the QIM for the thermo field double state of a two dimensional CFT deformed by a marginal operator. This set up has been studied holographically using the aforementioned brane approximation \cite{MIyaji:2015mia}. This approximation captures the qualitative behavior of the QIM, however it fails to reproduce exactly the CFT result.
\item We generalize the bulk construction to a multi dimensional parameter space where the deformation is induced by marginal operators spanning a moduli space (section \ref{multi dim}).
\end{itemize}
We stress that all these computations are done in the bulk. Whenever the QIM can be computed on the CFT side we find exact agreement.
\section{Introduction to the Quantum Information Metric}\label{QIM into}
A quantity that finds application in condensed matter physics and information theory is fidelity \cite{fidelity}. For two generic quantum states $A$ and $B$ described by density matrices $\rho_A$ and $\rho_B$ we define the fidelity $F(\rho_A, \rho_B)$ by the following formula:
\begin{equation}
F(\rho_A, \rho_B)=\text{tr}\sqrt{{\rho_A}^{1/2} \rho_B {\rho_A}^{1/2}}
\end{equation}
where the trace is taken over the Hilbert space of all quantum states of the system.\\ \indent
When the states $A$ and $B$ are pure fidelity reduces to the absolute value of the overlap, i.e. if $\rho_A=\ket{\Psi_A}\bra{\Psi_A}$ and $\rho_B=\ket{\Psi_B}\bra{\Psi_B}$ we have that
\begin{equation}
F(\Psi_A,\Psi_B)=|\braket{\Psi_A|\Psi_B}|.
\end{equation}
We now consider a one parameter family of states, parametrized by $\lambda$, with corresponding density matrix $\rho_\lambda$ and we define the QIM $G_{\lambda \lambda}$ by considering fidelity between two states relative to infinitesimally close parameters, say $\lambda$ and $\lambda +\delta \lambda$, and expanding in $\delta \lambda$:
\begin{equation}
F(\rho_\lambda,\rho_{\lambda+\delta \lambda})=1- G_{\lambda \lambda} \delta \lambda^2+\mathcal{O}(\delta \lambda^3).
\end{equation}
We can generalize this concept to a multi dimensional parameter space with $\lambda=\{\lambda^a\}$ and $a=1,...,N$. The natural generalization is:
\begin{equation}
F(\rho_\lambda,\rho_{\lambda+\delta \lambda})=1- \sum _{a,b=1}^N G_{a b} \delta \lambda^a \delta \lambda^b+\mathcal{O}(\delta \lambda^3).
\end{equation}
Notice that the presence of a term linear in $\delta \lambda$ vanishes by unitarity.\\
\section{The QIM for the vacuum state of a CFT living on $\mathbb{R}^{d-1}\times \mathbb{R}$.}\label{CFT GS}
In this section we firstly review the computation of the QIM for a CFT in its ground state. We then explain how to compute the same quantity holographically using a perturbative approach.
\subsection{CFT construction}
Let us consider a $d$-dimensional CFT with Euclidean Lagrangian $\mathcal{L}_0$. We deform the theory by adding to $\mathcal{L}_0$ a term of the form $\delta \lambda \mathcal{O}(x)$, where $\mathcal{O}$ corresponds to a conformal primary operator of the original theory with conformal dimension $\Delta$ and $\delta \lambda$ is a coupling constant. In order to distinguish between quantities computed in the unperturbed theory with quantities in the deformed theory we use respectively the subscripts $0$ and $1$. \\ \indent
We are interested in computing the absolute value of the overlap between the ground states of the two theories $|\braket{\Psi_1|\Psi_0}|$ at second order in $\delta \lambda$. In order to do that we use a path integral formalism.\\ \indent
We start by considering the overlap between the ground state of the undeformed theory $\ket{\Psi_0}$ and a generic state $\ket{\tilde \varphi}$. In a path integral language the quantity $\braket{\tilde \varphi|\Psi_0}$ can be obtained by considering an Euclidean evolution from $\tau=-\infty$ to $\tau=0$ where the state $\ket{\tilde \varphi}$ is inserted. In equation:
\begin{equation}
\braket{\tilde{\varphi}|\Psi_0}=\frac{1}{\sqrt{Z_0}}\int_{\varphi(\tau=0)=\tilde{\varphi}} \mathcal{D}\varphi \exp\left(-\int_{-\infty}^{0} d \tau \int d^{d-1}x \mathcal{L}_0 \right),
\end{equation}
where $Z_0$ is the partition function of the unperturbed theory:
\begin{equation}
Z_0=\int \mathcal{D} \varphi \exp\left(-\int_{-\infty}^{\infty} d \tau \int d^{d-1} x \mathcal{L}_0\right).
\end{equation}
In a similar way we construct $\braket{\Psi_1|\tilde \varphi}$ by considering the Euclidean evolution from $\tau=0$, where the state $\ket{\tilde \varphi}$ is inserted, to $\tau=\infty$:
\begin{equation}
\braket{\Psi_1|\tilde \varphi}=\frac{1}{\sqrt{Z_1}}\int_{\varphi(\tau=0)=\tilde{\varphi}} \mathcal{D}\varphi \exp\left(-\int_{0}^{\infty} d \tau \int d^{d-1}x (\mathcal{L}_0+\delta \lambda \mathcal{O}) \right),
\end{equation}
with
\begin{equation}
Z_1=\int \mathcal{D} \varphi \exp\left(-\int_{-\infty}^{\infty} d \tau \int d^{d-1} x (\mathcal{L}_0+\delta \lambda \mathcal{O})\right)
\end{equation}
being the partition function of the deformed theory. Notice that in this case we have used the deformed Lagrangian $\mathcal{L}_1=\mathcal{L}_0+\delta \lambda \mathcal{O}$.\\ \indent
\begin{figure}
\centering
\begin{tikzpicture}
\filldraw[fill=blue!10!white, draw=black] (-1.5,0)--(1.5,0)--(4,1.5)--(1,1.5)--cycle;
\filldraw[fill=red!10!white, draw=black] (4.5,0)--(1.5,0)--(4,1.5)--(7,1.5)--cycle;
\node[anchor=south] at (4,1.5) {$\tau=0$};
\draw[->] (0,-0.3)--(2.5,-0.3);
\node[anchor=north] at (1.5,-0.5) {$\tau$};
\draw[->] (-2, 0)--(0.5,1.5);
\node[anchor=north] at (-1,1.5) {$\mathbb R^{d-1}$};
\node at (1.5,0.7) {$\mathcal{L}_0$};
\node at (4.2,0.7) {$\mathcal{L}_0+\delta \lambda\mathcal{O}$};
\end{tikzpicture}
\caption{Pictorial representation of the path integral construction used to build $\braket{\Psi_1 |\Psi_0}$. The Euclidean propagation is governed by the unperturbed Lagrangian $\mathcal{L}_0$ in the blue region, while in the red region we use the deformed Lagrangian $\mathcal{L}_0+\delta \lambda\mathcal{O}$. } \label{fig1}
\end{figure}
The overlap $\braket{\Psi_1|\Psi_0}$ can then be obtained as
\begin{eqnarray}\label{overlap}
\braket{\Psi_1|\Psi_0}&=&\int \mathcal{D} \tilde \varphi \braket{\Psi_1|\tilde \varphi} \braket{\tilde \varphi|\Psi_0}\nonumber\\
&=&\frac{\int \mathcal{D} \varphi \exp \left(-\int_{-\infty}^{0} d \tau \int d^{d-1}x \mathcal{L}_{0}- \int_{0}^{\infty} d \tau \int d^{d-1}x ( \mathcal{L}_0 +\delta \lambda \mathcal{O})\right)}{(Z_0 Z_{1})^{1/2}}.
\end{eqnarray}
This overlap is generally speaking ill defined, since the Lagrangian governing the Euclidean propagation changes discontinuously at $\tau=0$ and this introduces UV divergences. For this reason one should think of equation (\ref{overlap}) as formal. For all explicit computations we regularize the UV divergences in the formula for the overlap (\ref{overlap}) by replacing $\ket{\Psi_1}$ with
\begin{equation}
\ket{\Psi_1 (\epsilon)}=\frac{e^{-\epsilon H_0}\ket{\Psi_1}}{\left(\braket{\Psi_1|{e^{-2\epsilon H_0}|\Psi_1}} \right)^{1/2}},
\end{equation}
where $H_0$ is the Euclidean Hamiltonian of the unperturbed theory. $\epsilon$ should be thought as an UV cut-off, its physical meaning might seem obscure at this point, but it will later become clear that $\epsilon$ removes the region where the Lagrangian changes abruptly .\\
We then rewrite equation (\ref{overlap}) as an expectation value in the state $\ket{\Psi_0}$:
\begin{equation}\label{overlap2}
\braket{\Psi_1 (\epsilon)|\Psi_0}=\frac{\braket{\exp\left( -\int_{\epsilon}^{\infty} d \tau \int d^{d-1}x \delta \lambda \mathcal{O}(\tau, x)\right)}}{\braket{\exp\left( -(\int_{-\infty}^{-\epsilon}+\int_{\epsilon}^{\infty}) d \tau \int d^{d-1}x \delta \lambda \mathcal{O}(\tau, x)\right)}^{1/2}}.
\end{equation}
We can now expand the overlap (\ref{overlap2}) in powers of $\delta \lambda$. Taking into account that $\braket{\mathcal{O}}=0$ for an operator of non-zero dimension in the unperturbed theory and that the two point function of a primary operator enjoys the time reversal symmetry relation $\braket{\mathcal{O}(-\tau_1)\mathcal{O}(-\tau_2)}=\braket{\mathcal{O}(\tau_1)\mathcal{O}(\tau_2)}$, we get that:
\begin{equation}
|\braket{\Psi_1|\Psi_0}|=1-G_{\lambda \lambda} \delta \lambda^2+\mathcal{O}(\delta \lambda^3),
\end{equation}
where
\begin{equation}
G_{\lambda \lambda} =\frac{1}{2}\int d^{d-1}x_1 \int d^{d-1}x_2\int_{-\infty}^{-\epsilon} d \tau_1 \int_{\epsilon}^{\infty}d \tau_2\braket{\mathcal{O}(\tau_1,x_1)\mathcal{O}(\tau_2,x_2)}
\end{equation}
is the QIM. Notice that, as anticipated before, $\epsilon$ effectively removes a slab centered at $\tau=0$.\\ \indent
Using the two point function for a primary operator \footnote{We left the normalization of the two point function unspecified. In the particular set-ups we study in the sequel we choose the normalization $\mathcal{C}$ such that the two point function computed from the bulk agrees with equation (\ref{2 point function}).}
\begin{equation}\label{2 point function}
\braket{\mathcal{O}(\tau_1,x_1)\mathcal{O}(\tau_2,x_2)}=\frac{\mathcal{C}}{((\tau_1-\tau_2)^2+(x_1-x_2)^2)^{\Delta}}
\end{equation}
if $d+1-2 \Delta<0$ we get:
\begin{equation}\label{QIMCFT}
G_{\lambda \lambda}=\mathcal{C} N_{d} V_{\mathbb{R}^{d-1}} {\epsilon}^{d+1-2\Delta},
\end{equation}
where
\begin{equation}
N_d=\frac{2^{d-1-2\Delta}\pi^{(d-1)/2}\Gamma(\Delta-d/2-1/2)}{(2 \Delta -d)\Gamma(\Delta)}.
\end{equation}
Note that if we had to deform the theory by a linear combination of two primary operators, i.e. $\mathcal{L}_1=\mathcal{L}_0+\delta \lambda_A \mathcal{O}_A+\delta \lambda_B \mathcal{O}_B$, normalized such that $\braket{\mathcal{O}_A \mathcal{O}_B}=0$, the QIM would be diagonal. We will expand the discussion on multi dimensional parameter space in section \ref{multi dim} where we study the QIM in the case of a deformation induced by a linear combination of marginal operators spanning a moduli space.
\subsection{Bulk computation}\label{bulk}
In this subsection we discuss the holographic dual of this setup.
The computation of the QIM on the gravity side has appeared in \cite{MIyaji:2015mia} and \cite{Bak:2015jxd} where $\mathcal{O}$ was taken to be exactly marginal. We develop a perturbative method that allows us to deal with any primary (provided \linebreak $\Delta>\frac{d+1}{2}$).
The basic idea is to look at the right hand side of equation (\ref{overlap}) and interpret is as a combination of partition functions. We have:
\begin{equation}
\braket{\Psi_1|\Psi_0}=\frac{Z_2}{(Z_1 Z_{0})^{1/2}},
\end{equation}
where $Z_0$ is the partition function of a pure CFT, $Z_1$ is the partition function of the deformed CFT and $Z_2$ is the partition function of a CFT that is deformed only for $\tau>0$.\\ \indent
We can evaluate these partition functions on the gravity side. In the large $N$ limit we can write $Z_k=\exp(-I_k)$ where $I_k$ is the on-shell action of the gravity solution dual to the corresponding field theory configuration ($k=0,1,2$). Since we consider the operator to have conformal dimension $\Delta$ the dual scalar field is going to have mass {$m^2=\Delta (\Delta-d)$}.\\ \indent
The action governing the bulk physics is
\begin{equation}\label{action}
I=-\frac{1}{\kappa^2}\int d^{d+1}x \sqrt{g} \left(\frac{1}{2} R-\frac{1}{2}\partial_\mu \Phi \partial^\mu \Phi-\frac{1}{2}m^2\Phi^2 +\frac{d(d-1)}{2 L^2} \right)+{I_{BND}},
\end{equation}
{where the last term has been introduced in order to guarantee that the variational principle is well posed}. The massive field is going to have a different profile in the three different cases of interest. In particular for the computation of $Z_0$ we notice that the massive field is turned off, the dual solution is pure AdS, then $Z_0=\exp(-I_{AdS})$.\\ \indent
The scalar field profile for $I_1$ and $I_2$ will depend on $\delta \lambda$. Since we are interested only in this quantities at order $\delta \lambda^2$ we can use a perturbative approach. We write the fields as\footnote{If the operator $\mathcal{O}$ is marginal the massless field should be taken to be $\Phi(x)=\lambda_0+\delta \lambda \tilde{\Phi}(x)$, where $\lambda_0$ is the coupling constant of the undeformed theory.}:
\begin{eqnarray}
\Phi(x)&=&\delta \lambda \tilde{\Phi}(x),\\
g_{\mu \nu}(x)&=& g_{\mu \nu}^0(x)+\delta \lambda^2 \tilde{g}_{\mu \nu}(x),
\end{eqnarray}
where $g_{\mu \nu}^0$ is the metric of pure $AdS_{d+1}$. Notice that the metric receives corrections at order $\delta \lambda^2$ since the scalar field enters quadratically in Einstein's equations.\\ \indent
We can now expand the on-shell action around the unperturbed solution
\begin{eqnarray}\label{EQ3}
\delta I&=& \delta \lambda \int \frac{\delta I}{\delta \Phi(x)} \bigg|_{g_0} \tilde{\Phi}(x) +\frac{1}{2}\delta \lambda^2 \int \frac{\delta^2 I}{\delta \Phi(x) \delta \Phi(y)}\bigg|_{ g_0} \tilde{\Phi}(x) \tilde{\Phi}(y)+\nonumber \\ & &+\delta \lambda^2 \int \frac{\delta I}{\delta g_{\mu\nu}(x)}\bigg|_{ g_0} {\tilde g_{\mu \nu}(x)}+\mathcal{O}(\delta\lambda^3).
\end{eqnarray}
Notice that the first and third terms vanish because the equations of motion of the background are satisfied. {Notice also that the boundary term of equation (\ref{action}) gets canceled by the boundary terms that arise from integration by parts when obtaining the first and third terms of equation (\ref{EQ3}). The second term of equation (\ref{EQ3}) should have been accompanied by the boundary term that arises when recasting the second variation of $I$ in that guise. We omit it, since we will reintegrate the second term back by parts to bring $\delta I$ in a form similar to the original one. It is now clear that} we are simply interested in computing the contribution of the scalar field probing the unperturbed background.
We can then write $I_k=I_{AdS}+\delta I_k$, with
\begin{equation}\label{EQ2}
\delta I_k=\frac{1}{2\kappa^2}\int d^{d+1}x \sqrt{g_0} \left(g^{\mu \nu}_0\partial_\mu \Phi_k \partial_\nu \Phi_k +m^2\Phi_k^2 \right).
\end{equation}
$\Phi_k$ is the solution of the equation of motion of the massive field with fixed background. $\Phi_k$ can be obtained easily by using the boundary to bulk propagator:
\begin{eqnarray}
\Phi_1 (z,\tau,x)&=&z^{d-\Delta} \delta \lambda\\
\Phi_2(z,\tau,x)&=& \delta \lambda z^{d-\Delta } \left(\frac{ \tau \Gamma \left(-\frac{d}{2}+\Delta +\frac{1}{2}\right) \, _2F_1\left(\frac{1}{2},-\frac{d}{2}+\Delta +\frac{1}{2};\frac{3}{2};-\frac{\tau ^2}{z^2}\right)}{\sqrt{\pi } z \Gamma \left(\Delta -\frac{d}{2}\right)}+\frac{1}{2}\right).
\label{field}\end{eqnarray}
We write the overlap as:
\begin{eqnarray}
\braket{\Psi_1|\Psi_0}&=&\frac{Z_{2}}{\sqrt{Z_1 Z_0}}=\exp \left(-I_{Ads}-\delta I_{2} + \frac{1}{2}(I_{Ads}+\delta I_{1}+I_{Ads})\right) \nonumber\\
&=& \exp \left( -\delta I_{2} + \frac{1}{2} \delta I_{1} \right).
\end{eqnarray}
\begin{figure}
\centering
\begin{tikzpicture}
\draw[color=blue] (-3,0)--(0,0);
\draw[color=red] (0,0)--(3,0);
\draw[color=blue!20!red] (0,0) ..controls (0.5,-0.25) and (1,-0.5).. (2.8,-0.6);
\draw[color=blue!30!red] (0,0) ..controls (0.5,-0.5) and (1,-1).. (2.2,-1.2);
\draw (0,0)[color=blue!40!red] ..controls (0.2,-0.7) and (1,-1.6).. (1.3,-1.8);
\draw[color=blue!50!red] (0,0) --(0,-2);
\draw[color=blue!60!red] (0,0) ..controls (-0.2,-0.7) and (-1,-1.6).. (-1.3,-1.8);
\draw[color=blue!70!red] (0,0) ..controls (-0.5,-0.5) and (-1,-1).. (-2.2,-1.2);
\draw[color=blue!80!red] (0,0) ..controls (-0.5,-0.25) and (-1,-0.5).. (-2.8,-0.6);
\node[anchor=south] at (0,0) {$\tau$=0};
\node[anchor=south] at (1.8,0) {$\rho=\infty$};
\node[anchor=south] at (-1.8,0) {$\rho=-\infty$};
\node[anchor=west] at (2,-1.2) {\textcolor{blue!30!red}{$AdS_{d}$}};
\node[anchor=east] at (-2,-1.2) {\textcolor{blue!70!red}{$AdS_{d}$}};
\end{tikzpicture}
\caption{Schematic representation of the $AdS_d$ slicing of $AdS_{d+1}$. Each colored line corresponds to a single $AdS_d$ slice located at a fixed value of the coordinate $\rho$.} \label{fig2}
\end{figure}
We now need to regularize the action $\delta I_k$ ($k=1,2$). We mentioned before that the background is Euclidean signature Poincar\'e $AdS_{d+1}$:
\begin{equation}
ds^2=L^2 \frac{dz^2+d\tau^2+\sum_{i}^{d-1} dx_i^2}{z^2},
\end{equation}
the boundary is located at $z=0$ and it is parametrized by $(\tau, x_i)$. We are going to work with {$AdS_{d+1}$ in $AdS_{d}$ slicing} by performing the following change of coordinates:
\begin{eqnarray}\label{changeofcoord}
\begin{split}
z=Z \sech \rho\\
\tau= Z \tanh \rho,
\end{split}
\end{eqnarray}
the metric becomes
\begin{equation}\label{slicing}
ds^2=\cosh^2 \rho\left(L^2 \frac{dZ^2+\sum_{i}^{d-1} dx_i^2}{Z^2}\right)+L^2 d \rho^2.
\end{equation}
There are two different ways to reach the boundary: either we take $Z \rightarrow 0$ keeping $\rho$ fixed or we take $\rho \rightarrow \pm \infty$ keeping $Z$ fixed. In the first limit we reach the boundary at $(\tau=0, x_i)$, while in second limit we reach the points $(\tau=\pm Z, x_i)$. The $AdS_d$ slicing of $AdS_{d+1}$ is schematically represented in figure \ref{fig2}.\\ \indent
We regularize the action by putting cut-offs at $\rho= \pm \rho_\infty$ and $Z=\epsilon$. We call the regularized manifold $\tilde{\mathcal{M}}$. {This regularization choice might seem odd, since it is not the standard regularization condition used in many AdS/CFT examples. A regularization procedure which uses two cut offs has been recently suggested in computations of different holographic quantities in the context of gravity duals of interface conformal field theories (ICFT) \cite{Gutperle:2016gfe}. Since our set up share the same symmetry as the bulk dual of an ICFT we follow this new regularization prescription. One could have adopted the usual regularization prescription by cutting off the $AdS$ volume at $z=\delta$ obtaining the same results.}
The action we are interested in is the on-shell action for a massive scalar on a background geometry. It can be written as a boundary contribution by integrating by parts and using the equation of motion. In particular the regularized action is:
\begin{eqnarray}
\delta I_k&=&\frac{1}{2 \kappa^2}\int_{\partial \tilde{\mathcal{M}}}\sqrt{\gamma_0} n_{\mu} g^{\mu \nu}_0 \Phi_k \partial_\nu \Phi_k,
\end{eqnarray}
with $n_{\mu}$ being the unit normal and $\gamma_0$ the determinant of the induced metric on the boundary.\\ \indent
We write $\Phi_k= \delta \lambda Z^{d-\Delta} f_k(\rho) $ where
\begin{eqnarray}
f_1(\rho)&=& (\sech \rho)^{d-\Delta} \nonumber \\
f_2(\rho)&=& (\sech \rho)^{d-\Delta } \left(\frac{ \sinh \rho \Gamma \left(-\frac{d}{2}+\Delta +\frac{1}{2}\right) \, _2F_1\left(\frac{1}{2},-\frac{d}{2}+\Delta +\frac{1}{2};\frac{3}{2};-\sinh^2 \rho \right)}{\sqrt{\pi } \Gamma \left(\Delta -\frac{d}{2}\right)}+\frac{1}{2} \right). \nonumber
\end{eqnarray}
The regularized action is then:
\begin{eqnarray}
\delta I_k &=& \frac{V_{\mathbb{R}^{d-1}}L^{d-1} \delta \lambda^2}{2 \kappa^2} \int_{-\rho_{\infty}}^{\rho_\infty} d\rho \left(\frac{\cosh \rho}{\epsilon}\right)^{d-1} \frac{ \cosh \rho}{\epsilon} \frac{\epsilon^2}{\cosh^2 \rho } (d-\Delta) \epsilon^{2d-2\Delta-1} f_k^2(\rho) \nonumber\\
&+&\frac{V_{\mathbb{R}^{d-1}}L^{d-1}\delta \lambda^2}{2 \kappa^2} \int_{\epsilon}^{\infty} dZ\left( \frac{\cosh \rho_\infty}{Z}\right)^{d} Z^{2(d-\Delta)} f_k(-\rho_{\infty}) \partial_\rho f_k(\rho)\bigg|_{-\rho_\infty} \nonumber \\
&+&\frac{V_{\mathbb{R}^{d-1}}L^{d-1}\delta \lambda^2}{2 \kappa^2} \int_{\epsilon}^{\infty} dZ\left( \frac{\cosh \rho_\infty}{Z}\right)^{d} Z^{2(d-\Delta)} f_k(\rho_{\infty}) \partial_\rho f_k(\rho)\bigg|_{\rho_\infty} \nonumber
\end{eqnarray}
Assuming $2\Delta>{d+1}$ we get the following result:
\begin{equation}
-\delta I_2+\frac{1}{2}\delta I_1=\frac{V_{\mathbb{R}^{d-1}}L^{d-1} \epsilon^{d-2 \Delta+1}}{2 \kappa^2} (J_a+J_b+J_c)\delta \lambda^2
\end{equation}
where
\begin{eqnarray}
J_a&=&\frac{1}{ (2 \Delta-d-1)} \left( \left(-f_2 \partial_\rho f_2 + f_1 \partial_\rho f_1 \right)\cosh^d \rho \right)\bigg|_{\rho_{\infty}},\\
J_b&=&(d-\Delta) \int_{0}^{\rho_\infty} d\rho \left(-f_2^2(\rho)+f_1^2(\rho)\right) \cosh^{d-2} \rho\\
J_c&=&(d-\Delta) \int_{-\rho_{\infty}}^{0} d\rho \left(-f_2^2(\rho)\right) \cosh^{d-2} \rho.
\end{eqnarray}
$J_a$ comes form the boundary term located at $\rho=\pm \rho_{\infty}$. Notice that $J_a$,$ J_b$ and $J_c $ have no divergence associated with $\rho_\infty$ and we are free to take the limit $\rho_{\infty}\rightarrow \infty $. $J_a$ can be computed explicitly:
\begin{equation}
J_a= \frac{-d \Gamma \left(-\frac{d}{2}+\Delta +\frac{1}{2}\right)}{2 (2 \Delta -d-1)\sqrt{\pi } \Gamma \left(-\frac{d}{2}+\Delta +1\right)}.
\end{equation}
The QIM is
\begin{equation}
G=-\frac{V_{\mathbb{R}^{d-1}}L^{d-1} \epsilon^{d-2 \Delta+1}}{2 \kappa^2} (J_a+J_b+J_c).
\end{equation}
This matches the CFT result (\ref{QIMCFT}) in the sense that we recover the same divergence structure, we do not compare the coefficient of the divergence because it is a non universal quantity. Notice that, as in the CFT computation, the divergence arises only from the location where we turned on the deformation.\\ \indent
Note that if the operator $\mathcal{O}$ is marginal, i.e. $\Delta=d$, $J_a$ and $J_b$ vanish and we can write the result explicitly as:
\begin{equation}
G=\frac{d \Gamma\left(\frac{d+1}{2}\right)V_{\mathbb{R}^{d-1}}L^{d-1} \epsilon^{-d+1}}{4\sqrt{\pi}(d-1)\Gamma(\frac{d}{2}+1) \kappa^2}.
\end{equation}
{It is important to remark that the quantity computed is the bare QIM. Considering bare quantities and regularizing them by imposing a cut-off is usual practice in many AdS/CFT computations. The understanding is that, since the QIM is obtained by path integral arguments, one could make use of the standard holographic renormalization. The divergences can be removed by adding local counter terms to the bulk action. See \cite{Skenderis:2002wp} for a review on the topic. Finally, it is worth to point out that even the bare QIM has its own significance. In fact it was the degree of the divergence of the bare QIM for a marginal deformation to suggest the brane approximation proposed in \cite{MIyaji:2015mia}. }
\section{The QIM for the vacuum state of a CFT living on $\mathbb{S}^{d-1}\times \mathbb{R}$. }\label{globalQIM}
In this section we discuss the QIM obtained by studying deformation of the vacuum state of a CFT living on a cylinder. By the same arguments of the previous section we can write the overlap between the CFT vacuum and the deformed vacuum as:
\begin{equation}
\braket{\Psi_1|\Psi_0}=\frac{Z_{2}}{\sqrt{Z_1 Z_0}},
\end{equation}
where $Z_{0}$ is the partition function of the CFT, $Z_1$ is the partition function of the deformed theory, obtained by deforming the action by a term of the form
\begin{equation}\label{def}
\int \delta \lambda \mathcal O,
\end{equation}
with $\delta \lambda$ begin a constant. Finally $Z_2$ is the partition function of the theory obtained by deforming the CFT action by the same term appearing in (\ref{def}) only for $\tau>0$. The situation is schematically represented in figure \ref{massivefig}.
\begin{figure}
\centering
\begin{tikzpicture}
\fill[left color=blue!50,right color=blue!50,middle color=blue!30,shading=axis,opacity=1] (1,0) -- (1,3) arc (360:180:1cm and 0.25cm) -- (-1,0) arc (180:360:1cm and 0.25cm);
\fill[left color=blue!50,right color=blue!50,middle color=blue!80!black,shading=axis,opacity=0.75] (0,3) circle (1cm and 0.25cm);
\draw[->] (-1.2,2.5)--(-1.2,3);
\node[left] at (-1.25,2.75) {$\tau$};
\end{tikzpicture}
\hspace{2cm}
\begin{tikzpicture}
\fill[left color=red!50,right color=red!50,middle color=red!30,shading=axis,opacity=1] (1,0) -- (1,3) arc (360:180:1cm and 0.25cm) -- (-1,0) arc (180:360:1cm and 0.25cm);
\fill[left color=red!50!,right color=red!50,middle color=red!50!black,shading=axis,opacity=1] (0,3) circle (1cm and 0.25cm);
\draw[->] (-1.2,2.5)--(-1.2,3);
\node[left] at (-1.25,2.75) {$\tau$};
\end{tikzpicture}
\hspace{2cm}
\begin{tikzpicture}
\fill[left color=red!50,right color=red!50,middle color=red!30,shading=axis,opacity=1] (1,0) -- (1,1.5) arc (360:180:1cm and 0.25cm) -- (-1,0) arc (180:360:1cm and 0.25cm);
\fill[left color=red!50!,right color=red!50,middle color=red!50!black,shading=axis,opacity=1] (0,1.5) circle (1cm and 0.25cm);
\fill[left color=blue!50,right color=blue!50,middle color=blue!30,shading=axis,opacity=1] (1,-1.5) -- (1,0) arc (360:180:1cm and 0.25cm) -- (-1,-1.5) arc (180:360:1cm and 0.25cm);
\node[left] at (-1,0) {$\tau=0$};
\draw[->] (-1.2,1)--(-1.2,1.5);
\node[left] at (-1.25,1.25) {$\tau$};
\end{tikzpicture}\caption{Schematic representations of the configurations corresponding to $Z_0$, $Z_1$ and $Z_2$. The undeformed theory is represented in blue while the red color represents a deformation of the theory induced by adding to the action a term of the form $\int \delta\lambda \mathcal O$.} \label{massivefig}
\end{figure}
We find the partition functions $Z_k$, $k=0,1,2$, holographically. In particular we have:
\begin{equation}\label{massive2}
\braket{\Psi_1|\Psi_0}= \exp \left( -\delta I_{2} + \frac{1}{2} \delta I_{1} \right),
\end{equation}
where $\delta I_k$ is the on-shell action of a scalar field dual to the operator $\mathcal O$, probing global $AdS$, whose metric is:
\begin{equation}\label{global}
ds^2=\frac{dr^2}{r^2+1}+(r^2+1) d\tau^2+r^2 g_{\mathbb S^{d-1}},
\end{equation}
with $r \in [0, \infty])$, $\tau \in (-\infty, \infty)$ and we have set the $AdS$ radius to one. The expression for $\delta I_k$ is:
\begin{equation}\label{massive1}
\delta I_k=\frac{1}{2 \kappa^2}\int_{\partial \tilde{\mathcal{M}}}\sqrt{\gamma_0} n_{\mu} g^{\mu \nu}_0 \Phi_k \partial_\nu \Phi_k,
\end{equation}
where $\Phi_1$ and $\Phi_2$ satisfy different boundary conditions, as prescribed the by standard AdS/CFT dictionary.
\subsection{Marginal Deformation}
We focus at first on a marginal perturbation. In this case the only relevant term in equation (\ref{massive2}) is $\delta I_2$. We then need to find the scalar field that obeys Laplace equation in global $AdS$ with the following boundary conditions:
\begin{equation}
\lim_{r\rightarrow \infty}\Phi_2(r,\tau)=\begin{cases}
\delta\lambda & \text{if }\tau>0 \\
0& \text{if }\tau<0.
\end{cases}
\end{equation}
This field configuration can be found easily by writing the metric (\ref{global}) in global $AdS_{d}$ slicing
\begin{equation}\label{globalslicing}
ds^2=d\rho^2+\left(\frac{\cosh \rho}{\sin T}\right)^2 \left(dT^2+g_{\mathbb S^{d-1}}\right),
\end{equation}
through the following change of coordinates:
\begin{eqnarray}
r=\frac{\cosh\rho}{\sinh T} & &\tau=\frac{1}{2}\log \frac{\cosh(T+\rho)}{\cosh(T-\rho)},
\end{eqnarray}
with $T>0, \rho \in \mathbb{R}$.
Since the deformation is marginal we can interpret the path integral of $Z_2$ as the path integral of an interface conformal field theory. The interface preserve the subgroup $SO(d,1)$ of the entire conformal group. This is exactly the isometry group of the $AdS_d$ slice. We then conclude that the bulk field $\Phi_2$ has to depend only on $\rho$. Under these assumptions Laplace equation becomes an ordinary differential equation, the solution is given by:
\begin{equation}\label{EQ1}
\Phi_2(\rho)=\frac{\delta\lambda \Gamma(\frac{1+d}{2})}{\sqrt \pi \Gamma(d/2)} \int_{-\infty}^{x} \frac{1}{\cosh^d x} dx.
\end{equation}
{We point out that the full Janus solution for a massless scalar on global $AdS$ is available in literature \cite{Bak:2016rpn}. One can check that it does reproduce equation (\ref{EQ1}) when the deformation parameter is taken to be infinitesimal.}
The corresponding on shell action is:
\begin{equation}
\delta I_2=\frac{L^{d-1} \delta \lambda^2 \Gamma\left(\frac{1+d}{2}\right)}{2 \kappa^2 \sqrt{\pi} \Gamma(d/2)} \Vol(\mathbb S^{d-1}) \int_{\text{cut off}}^{\infty} \frac{1}{\sinh^d T} dT.
\end{equation}
We put a cut off at $\sinh T=\epsilon$. We now change variable of integration $r=(\sinh T)^{-1}$. We get the following expression for the QIM:
\begin{equation}\label{globalG}
G=\frac{L^{d-1} \Gamma\left(\frac{1+d}{2}\right)}{2 \kappa^2 \sqrt{\pi} \Gamma(d/2)} \Vol(\mathbb S^{d-1}) \int_{0}^{1/\epsilon} \frac{r^{d-1}}{\sqrt{1+r^2}}dr.
\end{equation}
Notice that $G$ has a universal constant which is cut-off independent. For even $d$ the universal term is the $\mathcal O(\epsilon^0)$ term, while for $d$ odd it is the coefficient of the logarithmic term. This can be explained by the observation that the conformal symmetry is restricted to the interface, which is even dimensional for $d$ odd and vice versa. Note that we could have simply taken the form of the field given by equation (\ref{field}) and performed a change of coordinates to obtain the field of interest. This is because a plane with a planar interface can be conformally mapped into a cylinder with a spherical interface (see figure \ref{massivefig2}), thus the path integral formulation of $Z_2$ for the CFT living on the plane is conformally related to the set up on the cylinder.
\begin{figure}
\begin{tikzpicture}
\filldraw[fill=blue!40!white, draw=black] (-1.5,0)--(1.5,0)--(4,1.5)--(1,1.5)--cycle;
\filldraw[fill=red!40!white, draw=black] (4.5,0)--(1.5,0)--(4,1.5)--(7,1.5)--cycle;
\node[anchor=south] at (4,1.5) {$t=0$};
\draw[->] (0,-0.3)--(2.5,-0.3);
\node[anchor=north] at (1.5,-0.5) {$t$};
\draw[->] (-2, 0)--(0.5,1.5);
\node[anchor=north] at (-1,1.5) {$\mathbb R^{d-1}$};
\fill[green!80!black] (6+2,0.75)--(6.5+2,0.75)--(6.5+2,1)--(7.2+2,0.5)--(6.5+2,0)--(6.5+2,0.25)--(6+2,0.25)--(6+2,0.75);
\fill[left color=red!50,right color=red!50,middle color=red!30,shading=axis,opacity=1] (1+11.5,0+1) -- (1+11.5,1.5+1) arc (360:180:1cm and 0.25cm) -- (-1+11.5,0+1) arc (180:360:1cm and 0.25cm);
\fill[left color=red!50!,right color=red!50,middle color=red!50!black,shading=axis,opacity=1] (0+11.5,1.5+1) circle (1cm and 0.25cm);
\fill[left color=blue!50,right color=blue!50,middle color=blue!30,shading=axis,opacity=1] (1+11.5,-1.5+1) -- (1+11.5,0+1) arc (360:180:1cm and 0.25cm) -- (-1+11.5,-1.5+1) arc (180:360:1cm and 0.25cm);
\node[left] at (-1+11.5,0+1) {$\tau=0$};
\draw[->] (-1.2+11.5,1+1)--(-1.2+11.5,1+1.5);
\node[left] at (-1.25+11.5,1.25+1) {$\tau$};
\draw[->] (-1+11.5,-1.5+1-0.5) arc (180:300:1cm and 0.25cm);
\node at (12,-1) {$\mathbb S^{d-1}$};
\end{tikzpicture}\caption{A plane with a planar interface is conformally equivalent to a cylinder with a spherical interface placed at a fixed location on the non-compact direction.} \label{massivefig2}
\end{figure}
As a final remark note that equation agrees with the brane model of \cite{MIyaji:2015mia}. In particular since there is a universal term we could fix the brane tension as
\begin{equation}
n_d=\frac{L^{d-1} \Gamma\left(\frac{1+d}{2}\right)}{2 \kappa^2 \sqrt{\pi} \Gamma(d/2)},
\end{equation}
however it is unclear how universal this quantity is.
\subsection{Example of non marginal deformation in $d=2$}
The arguments that allowed us to find the field $\Phi_2$ do not work in the case of a deformation that is not marginal. This is due to the fact that a non marginal deformation does not preserve conformal symmetry on the interface. For that reason the general analysis is difficult to be carried out for generic $d$ and $\Delta$. We specify to $d=2$ and $\Delta=4$ to give a concrete example of the computation of $G$ in this set up.
In order to construct the fields $\Phi_2$ and $\Phi_1$ we map the boundary problem to a planar geometry where we can use the bulk to boundary propagator. The metric (\ref{global}) can be written in Poincar\'e coordinates by using the following change of coordinates:
\begin{eqnarray}\label{chagecoordglobal}
x&=&\frac{r}{\sqrt{1+r^2}} \cos \phi e^{\tau} \nonumber \\
y&=&\frac{r}{\sqrt{1+r^2}} \sin \phi e^{\tau} \\
z&= & \frac{1}{\sqrt{1+r^2}} e^{\tau}.\nonumber
\end{eqnarray}
Notice that this change of coordinates induces a conformal transformation on the boundary. Since the operator $\mathcal O$ is not marginal we need to change the boundary value of the source to compensate for the conformal transformation. In particular if the coupling constant of $\mathcal O$ on the cylinder is given by a certain function, say $\delta \lambda(\tau,\phi)$, once we go to flat space the new source will be given by $\delta \lambda(x,y) (x^2+y^2)^{(\Delta-d)/2}$. Once this is taken into account we can find the expressions for $\Phi_1$ and $\Phi_2$. In the coordinates of equation (\ref{globalslicing}) we have:
\begin{eqnarray}
\Phi_1&=& \delta\lambda \left( \cosh ^2(\rho ) \text{csch}^2(T)+1/2\right)\\
\Phi_2&=&\frac{\delta \lambda}{32} e^{\rho } \text{sech}^3(\rho ) \text{csch}^2(T) \left(e^{4 \rho }+e^{2 \rho } (\cosh (2 T)+4)+3 \cosh (2 T)+7\right).
\end{eqnarray}
We now proceed to the evaluation of the action (\ref{massive1}). We again introduce the following cut-offs: $\rho=\pm\rho_{\infty}$ and $T=\epsilon$. We then have:
\begin{eqnarray}
\frac{2 \kappa^2}{\pi}\delta I_k&=&\int_{\epsilon}^{\infty}\left(\frac{\cosh\rho_{\infty}}{\sinh T}\right)^2 \partial_\rho \Phi_k^2 \big|_{\rho=\rho_\infty} dT-\int_{\epsilon}^{\infty}\left(\frac{\cosh\rho_{\infty}}{\sinh T}\right)^2 \partial_\rho \Phi_k^2 \big|_{\rho=-\rho_\infty} dT \nonumber\\
& &-\int_{-\rho_\infty}^{r_\infty}\partial_T \Phi_k^2 \big|_{T=\epsilon} d\rho.
\end{eqnarray}
For $k=1$ the first two integrals are identical, for $k=2$ the second integral vanishes. Under this considerations one has {the following partial result}:
\begin{eqnarray}
\frac{2 \kappa^2}{\pi}\left(\delta I_2-\frac{1}{2}\delta I_1\right)&=&\int_{\epsilon}^{\infty} \left(\frac{\cosh \rho}{\sinh T}\right)^2 \left(\partial_\rho \Phi_2^2-\partial_\rho \Phi_1^2\right)\bigg|_{\rho=\rho_{\infty}} dT+\nonumber\\
& & +\int_{-\rho_{\infty}}^{\rho_{\infty}}\left(\frac{1}{2}\partial_T \Phi_1^2-\partial_T \Phi_2^2\right)\bigg|_{T=\epsilon}\bigg|d\rho.
\end{eqnarray}
At this point we can perform the integrals and safely take the limit $\rho_{\infty}\rightarrow \infty$. We obtain the following result for the {regularized} QIM:
\begin{equation}
G=\sum_{i=0}^{2}n_{2i+1}\epsilon^{-2i-1}-\frac{\pi}{24 \kappa^2},
\end{equation}
where the non universal constants $n_{5},n_{3}$ and $n_{1}$ have been left unspecified since they depend on the details of the regularization procedure. Notice that the leading divergence in the QIM is unchanged. As in the marginal case, we have now a universal contribution to the QIM.
\begin{equation}
{G_{\text{UNIV}}=-\frac{\pi}{24 \kappa^2}}.
\end{equation}
To stress the fact that the method presented in this paper is exact we compute the QIM on the field theory side and see if we find agreement.
The two point function for a primary operator of dimension $\Delta$ in the vacuum state of a CFT living on a cylinder is given by:
\begin{equation}
\braket{\mathcal O (\phi_1,\tau_1) \mathcal O (\phi_2, \tau_2)}=\frac{ 2^{-2 \Delta}\mathcal C}{\left(\sin^2\left(\frac{\phi_1-\phi_2}{2}\right)+\sinh^2\left(\frac{\tau_1-\tau_2}{2}\right)\right)^{\Delta}}.
\end{equation}
We use $\Delta=4$ and $\mathcal C=6/(\pi \kappa^2)$. This normalization is found by asking that $\braket{\int \delta\lambda \mathcal O}=\exp(-I_{\text{on shell}})$ for any $\delta \lambda$. We finally find {the following result for the regularized QIM}:
\begin{equation}
G=\frac{\pi}{8 \kappa^2}\left(\frac{1}{64 \epsilon ^5}-\frac{1}{24 \epsilon ^3}+\frac{1}{6 \epsilon }-\frac{1}{3}\right).
\end{equation}
{From this expression we can read the universal part of the QIM:
\begin{equation}
G_{\text{UNIV}}=-\frac{\pi}{24 \kappa^2},
\end{equation}}
which shows agreement between the bulk and the CFT computation.
\section{The QIM for a time dependent thermo field double state in $d=2$}{\label{TFD}}
In this section we are going to consider the QIM for a thermo field double state. We limit our study to a marginal deformation. We firstly present the CFT computation that was firstly derived in \cite{MIyaji:2015mia}, we then show how to obtain the same quantity holographically.
\subsection{CFT construction}
Generally speaking when the CFT is thermal the computation of the QIM becomes very involved. The reason is that there is not a pure state associated with a thermal CFT. In order to include temperature in our set-up we consider the thermo field double (TFD) construction for a CFT. A TFD state is a pure state obtained by taking a double copy of the original CFT:
\begin{equation}
\ket{TFD}\propto\sum_n e^{-\frac{\beta}{4}(H_A+H_B)}\ket{n}_A \ket{n}_B,
\end{equation}
where $A$ and $B$ label the two copies of the CFT. We choose the total Hamiltonian to be $H_{\text{tot}}=H_A+H_B$, the TFD state will evolve non trivially under time evolution. Let's now assume that the Hamiltonian depends on a parameter $\lambda$ which governs the perturbation by an exactly marginal operator $\mathcal{O}$. We are interested in studying the overlap between two TFD states corresponding to the perturbed and unperturbed Hamiltonian after we let them evolve for a real time $t$, i.e.:
\begin{equation}\label{QIMTFD}
|\braket{TFD_{\lambda+\delta \lambda}(\tilde{\tau})|TFD_{\lambda}(\tilde{\tau})}|=1-G_{\lambda \lambda}(\tilde{\tau}) \delta \lambda^2+...
\end{equation}
where $\tilde{\tau}=i t$.\\ \indent
We want to write this overlap as a path integral. The overlap ${}_A\bra{\varphi_1} {}_B\braket{\varphi_2|TFD}$ can be obtained as a path integral on an Euclidean time interval of length $\beta/2$ times the spatial manifold on which the theory lives (we will consider it to be the real line).\\ \indent
It follows that $\braket{TFD_{\lambda+\delta \lambda}(\tilde{\tau}=0)|TFD_{\lambda}(\tilde{\tau}=0)}$ is represented as a path integral in which the Euclidean propagation is governed by the Lagrangians $\mathcal{L}_{\lambda}$ and $\mathcal{L}_{\lambda+\delta \lambda}$ for the two halves of the thermal circle. The construction obtained by turning on the time evolution is represented in figure \ref{TFDpath}.
\begin{figure}
\centering
\begin{tikzpicture}
\draw[color=blue, thick] (0,0) arc (-30:210:1);
\draw[color=red, thick] (0,0) arc (-30:-150:1);
\node[anchor=south] at (-1,1.5) {\textcolor{blue}{$\mathcal{L}_\lambda$}};
\node[anchor=north] at (-1,-0.5) {\textcolor{red}{$\mathcal{L}_{\lambda+\delta \lambda}$}};
\node[anchor=east] at (-2,0) {$\tau= \beta/2+\tilde{\tau}$};
\node[anchor=west] at (0,0) {$\tau= -\tilde{\tau}$};
\end{tikzpicture}\caption{Path integral construction for the time dependent TFD state overlap. In this construction the Euclidean time $\tau$ is compactified on a circle of periodicity $\beta$. In order to build the overlap $\braket{TFD_{\lambda+\delta \lambda}(\tilde \tau)|TFD_{\lambda}(\tilde \tau)}$ we consider an Euclidean propagation governed by the Lagrangian $\mathcal{L}_\lambda$ for $- \tilde \tau<\tau<\beta/2+\tilde \tau $ (blue line) and by the Lagrangian $\mathcal{L}_{\lambda+\delta \lambda}$ for the remaining portion of the thermal circle (red line). } \label{TFDpath}
\end{figure}
\newline
\indent Once we have a path integral formulation for the overlap we can expand it in powers of $\delta \lambda$, at second order in $\delta \lambda$ we find
\begin{equation}
G_{\lambda \lambda}(\tilde{\tau})=\frac{1}{2} \int_{-\tilde{\tau}+\epsilon}^{\beta/2+\tilde{\tau}-\epsilon}d\tau_1 \int_{\beta/2+\tilde{\tau}+\epsilon}^{\beta-\tilde{\tau}-\epsilon} d\tau_2 \int dx_1 \int dx_2 \braket{\mathcal{O}(x_1,\tau_1)\mathcal{O}(x_2,\tau_2)}.
\end{equation}
The two point function for a primary operator in this geometry is fixed by conformal invariance. For a marginal operator $\mathcal{O}$ one finds:
\begin{equation}\label{QIMCFTTFD}
G_{\lambda \lambda}(\tilde{\tau})=\mathcal{C}\left( \frac{\pi V_{\mathbb{R}}}{8 \epsilon}+\frac{2 \pi^2 V_{\mathbb{R}}}{\beta^2}\left(\tilde{\tau}\cot\frac{4 \pi \tilde{\tau}}{\beta}-\frac{\beta}{4 \pi}\right)\right).
\end{equation}
Notice that in $d=2$ if we want the two point function on the CFT side to agree with the two point function on the bulk side we need to choose $\mathcal{C}=\frac{2 L}{\pi \kappa^2}$. {The universal part of the QIM is then:
\begin{equation}
G_{\lambda \lambda}(\tilde{\tau})_{\text{UNIV}}= \frac{V_{\mathbb{R}}}{2 \pi \kappa^2}(-1+2 \tilde{\tau} \cot(2 \tilde{\tau}) ).
\end{equation}}
\subsection{Bulk computation}
In order to find the QIM holographically we study a massless field probing a fixed background.
The unperturbed metric is given by
\begin{equation}\label{EBTZ}
ds^2= \frac{dZ^2}{(1-Z^2)Z^2} + \frac{1-Z^2}{Z^2} d \tau^2+ \frac{ d\phi^2}{Z^2},
\end{equation}
where we have fixed the AdS radius to be one and the inverse temperature $\beta=2 \pi$ for simplicity. The boundary conditions for the scalar field are
\begin{equation}
\Phi(Z\rightarrow 0, \tau, \phi)\equiv \tilde{\Phi}(\tau,\phi)=
\begin{cases}
\lambda & \text{ for}-\tilde{\tau}<\tau<\pi +\tilde{\tau} \\
\lambda+ \delta\lambda &\text{ otherwise}.
\end{cases}
\end{equation}
The metric is just pure AdS, under the following change of coordinates:
\begin{eqnarray}\label{chagecoord}
x&=&\sqrt{1-Z^2} \cos \tau e^{\phi} \nonumber \\
y&=&\sqrt{1-Z^2} \sin \tau e^{\phi} \\
z&= & Z e^{\phi}\nonumber
\end{eqnarray}
we can write it as Poincar\'e AdS
\begin{equation}
ds^2=\frac{\left(dx^2+dy^2+dz^2\right)}{z^2}.
\end{equation}
In order to find the massless field profile we can make use of the scalar propagator on Euclidean AdS:
\begin{equation}
\Phi(x,y,z)=\int \frac{z^2 \tilde{\Phi}(x',y')}{\pi (z^2+(x-x')^2+(y-y')^2)^2} dx' dy'.
\end{equation}
$\tilde{\Phi}$ in $x,y$ coordinates is represented in figure \ref{dilaton}.
\begin{figure}
\centering
\begin{tikzpicture}
\draw[red!30!white, fill=red!30!white] (0,0)--(1,2)--(2,2)--(2,-2)--(1,-2)--(0,0);
\draw[blue!30!white, fill=blue!30!white] (0,0)--(1,2)--(-2,2)--(-2,-2)--(1,-2)--(0,0);
\draw [thick, <->] (0,2) -- (0,0) -- (-2,0);
\draw [thick, -] (0,-2) -- (0,0) -- (2,0);
\draw[red!50!blue, thick] (1,-2)--(0,0)--(1,2);
\node[below right] at (-2,0) {y};
\node[below left] at (0,2) {x};
\node[blue] at (-1,0.3) {$\lambda$};
\node[red] at (1.3,0.3) {$\lambda+\delta \lambda$};
\draw[<-] (0.5,1) arc (60:90:1);
\node at (0.3, 1.3) {$\tilde{\tau}$};
\end{tikzpicture}
\caption{Boundary values for the scalar. Under the change of coordinates (\ref{chagecoord}) the blue and red regions of figure \ref{TFDpath} map respectively to the red wedge and its blue compliment represented here. } \label{dilaton}
\end{figure}
This integral can be solved, we get:
\begin{equation}
\Phi(x,y,z)=\lambda+\frac{\delta \lambda}{\pi}\left(h(x,y,z)+h(-x,y,z) - \pi/2+\tilde{\tau}\right)
\end{equation}
where
\begin{eqnarray}
h(x,y,z)&=& \frac{ (y+\tilde{t}x)\left(-\pi+2\arctan\left(\frac{\tilde{t} y - x}{\sqrt{(y+ \tilde{t} x)^2+ z^2(1+\tilde{t}^2 )}}\right)\right)}{4 \sqrt{(y+ \tilde{t} x)^2+ z^2(1+ \tilde{t}^2 )}}\\
\tilde{t}&=& \tan \tilde{\tau}.
\end{eqnarray}
The on-shell action for the massless field can be written as a boundary contribution:
\begin{eqnarray}
\delta I&=&\frac{1}{2 \kappa^2}\int_{\mathcal{M}} d^3 x \sqrt{g_0}g_0^{\mu \nu }\partial_{\mu} \Phi \partial_{\nu}\Phi\nonumber\\
&=& \frac{1}{2 \kappa^2}\int_{\partial\mathcal{M}} d^2 x \sqrt{\gamma_0} n_{\mu} g_0^{\mu \nu} \Phi \partial_\nu \Phi,
\end{eqnarray}
where $n_{\mu}$ is the unit normal and $\gamma_0$ is the determinant of the induced metric on the boundary.\\ \indent
Of course we need to specify the boundary on which the integral appearing in the second line is evaluated. We simply put a cut off at $z=\epsilon$ as natural in a $AdS/CFT$ context. We then have:
\begin{eqnarray}
\delta I&=&- \frac{1}{2 \kappa^2} \int dx dy \frac{1}{\epsilon}\Phi \partial_z \Phi \big|_{z=\epsilon}.
\end{eqnarray}
We now change variables of integration, using the $(Z,\tau, \phi)$ coordinates. Note that since we put a cut off at $z=\epsilon$ we have that $Z=\epsilon e^{-\phi}$, thus:
\begin{eqnarray}
x&=& \sqrt{e^{2 \phi}-\epsilon^2} \cos \tau\\
y&=& \sqrt{e^{2 \phi}-\epsilon^2} \sin \tau.
\end{eqnarray}
We obtain
\begin{eqnarray}
\delta I&=&-\frac{1}{2 \kappa^2}\int d\phi d\tau \lim_{\epsilon\rightarrow 0} \frac{1}{\epsilon }\Phi \partial_{z}\Phi \big|_{z=\epsilon}. \label{a}
\end{eqnarray}
We note that the $\tau$ integral will effectively run over $\tau\in [-\pi+\tilde \tau, -\tau]$. Since the integrand is even in $\tau$ we limit to $\tau \in [-\pi/2, -\tau] $, at this point we shift variable of integration $\tau'= \tau-\epsilon$. We can now take the $\epsilon$ limit
\begin{equation}
\partial_z \Phi (\tau'+\epsilon,\phi) \big|_{z=\epsilon}=\frac{\delta \lambda}{\pi} e^{-2 \phi} \mathcal F (\tau)\epsilon
\end{equation}
where
\begin{eqnarray}
\mathcal{F}(\tau)&=&\frac{\mathcal F_1(\tau) }{2 (\cos (2 T)-\cos (2 \tau ))^2}\\
\mathcal F_1(\tau)&=&2 \cos (2 \tau ) (\sin (2 T)+\pi \cos (2 T))+h_1(\tau)+h_1(-\tau)-\sin (4 T)-2 \pi \\
h_1(\tau) &= &4 \sin ^2(T-\tau ) \tan ^{-1}(\cot (\tau +T)).
\end{eqnarray}
The Jacobian gives a $e^{2 \phi}$, making the integrand $\phi$ independent. We find:
\begin{eqnarray}
\delta I &=&- \frac{\delta \lambda^2 V_{\mathbb{R}}}{ \pi \kappa^2}\int_{-\pi/2-\epsilon}^{-\tilde{\tau}-\epsilon} \mathcal{F}(\tau) \label{b}\\
&=& \frac{\delta \lambda^2 V_{\mathbb{R}}}{ \pi \kappa^2}\left(\frac{\pi}{2 \epsilon}+\frac{-1+2 \tilde \tau \cot (2 \tilde \tau)}{2}\right).
\end{eqnarray}
From the on shell action we obtain the following expression for the QIM:
\begin{equation}
G_{\lambda \lambda}=\frac{ V_\mathbb{R}}{2 \epsilon}+\frac{V_{\mathbb{R}}}{2 \pi \kappa^2}(-1+2 \tilde{\tau} \cot(2 \tilde{\tau}) ).
\end{equation}
\indent We notice that the divergence structure matches the CFT result (\ref{QIMCFTTFD}), in particular the universal term is the same, we do not compare the coefficient of the divergent term since it is not universal.
\section{The QIM for a multi dimensional parameter space}\label{multi dim}
In this section we show how to generalize the method used so far to the case of a multi dimensional parameter space.\\ \indent
We consider a CFT that has $N$ coupling constants $\lambda^a$ that couple to marginal operators. We change each coupling constant by an infinitesimal amount $\delta \lambda^a$. We are interested in studying the QIM in this set-up. We consider the absolute value of the overlap and expand it in $\delta \lambda$:
\begin{equation}
|\braket{\Psi_{\lambda+\delta \lambda}|\Psi_{\lambda}}|=1+G_{a b} \delta \lambda^a \delta\lambda^b+...
\end{equation}
As usual we write this overlap as a path integral where the value of each coupling constant is changed at $\tau=0$.\\ \indent
On the bulk side we then must have $N$ massless fields with nontrivial profile, dual to the operators $\mathcal{O}_a$. In this context it is quite natural to study systems in which the bulk physics is governed by
\begin{equation}
I=-\frac{1}{\kappa^2}\int d^{d+1}x \sqrt{g} \left(\frac{1}{2} R-\frac{1}{2} \mathcal{G}_{a b}(\Phi)\partial_\mu \Phi^a \partial^\mu \Phi^b +\frac{d(d-1)}{2 L^2} \right),
\end{equation}
the term in the action involving the scalars is a non linear sigma model which parameterize a moduli space with metric $\mathcal{G}_{a b}$.\\ \indent
The reason for this choice is that the constant values of $\Phi^a$ around which we are perturbing correspond to moduli, which need to be allowed to be arbitrary for marginal operators.\\ \indent
The equations of motion are
\begin{eqnarray}
&-2 \partial^\mu \left( \sqrt{g} \mathcal{G}_{a b}(\Phi) \partial_\mu \Phi^a \right)+\sqrt{g}\frac{\partial \mathcal{G}_{a c} (\Phi)}{\partial \Phi^b } \partial_\mu \Phi^a \partial^\mu \Phi^c=0 \label{scalars}&\\
&R_{\mu \nu}=\mathcal{G}_{a b}(\Phi)\partial_\mu \Phi^a \partial_\nu \Phi^b -\frac{d}{L^2}.&
\end{eqnarray}
We now consider a perturbative expansion for the fields $\Phi^a$. It is clear that the second term in equation (\ref{scalars}) is of order $\delta \lambda^a \delta \lambda^c$. This means that at first order in $\delta \lambda^a$ the scalars decouple from each other and they probe an unperturbed background. The derivation is then analogous to the one parameter case. The profiles for the scalars reduce to:
\begin{equation}
\Phi^a(\rho)=\lambda^a+\frac{\delta \lambda^a}{ I_d} \int_{-\infty}^{\rho}\frac{1}{\cosh^d (r)}dr,
\end{equation}
where $\rho$ is the coordinate that foliate $AdS_{d+1}$ in $AdS_d$ slices, as in equation (\ref{slicing}), and \linebreak $I_d=\int_{-\infty}^{\infty}\frac{1}{\cosh^d (r)}dr=\frac{2 \sqrt{\pi} \Gamma \left(\frac{d}{2}+1\right) }{d \Gamma (\frac{d+1}{2})}$.\\ \indent
The QIM is found to be
\begin{equation}
G_{a b}=\frac{d \Gamma\left(\frac{d+1}{2}\right)V_{\mathbb{R}^{d-1}}L^{d-1} \epsilon^{-d+1}}{4\sqrt{\pi}(d-1)\Gamma(\frac{d}{2}+1) \kappa^2} \mathcal{G}_{a b}(\lambda).
\end{equation}
We notice that the information metric inherits the same tensor structure and symmetry structure as the metric of the space where the scalars live.
\section{Conclusion}\label{conclusions}
\indent We have presented a new technique to compute the Quantum Information Metric, based on the observation that a perturbative approach is natural in this context. In particular we were able to find the QIM holographically for the vacuum state of a CFT deformed by a relevant operator. We have studied this set up for a CFT on a plane and on a cylinder, we noticed that the two set ups are related by a conformal transformation only if the deformation is induced by a marginal operator. We have also studied the case of a thermo field double state in two dimensions deformed by a marginal operator. Finally we have studied the case of a CFT ground state deformed by marginal operators spanning a moduli space. In all these cases the Janus solution is not available, the perturbative approach adopted allowed us to simplify the problem and to compute the QIM for these set-ups without the use of any approximations.
\section*{Acknowledgments}
I would like to thank Eric D'Hoker and Michael Gutperle for useful discussions and careful reading of the manuscript.
\providecommand{\href}[2]{#2}\begingroup\raggedright |
\section{Introduction}
In a point-to-point communication system with full feedback, the transmitter has noiseless access to all the previously received symbols. For discrete memoryless channels (DMCs), it turns out that this additional information does not increase capacity when codes of fixed blocklengths are used. Specifically, Shannon~\cite{ZeroErrorCapacityShanon} proved that the capacity with full-feedback fixed-blocklength codes is no larger than the one achievable in the no-feedback case. Dobrushin~\cite{doburshin1962anasymptotic} established a similar result for the reliability function of symmetric DMCs (the general case is, however, open). However, if the use of variable-length codes is permitted, the availability of full feedback turns out to be beneficial. Burnashev~\cite{burnashev1976data} derived the reliability function for the case when full feedback is available and variable-length feedback (VLF) codes are used, for all rates between zero and capacity. He showed that this reliability function is
\begin{IEEEeqnarray}{rCl}
E\parantheses{R}&=& C_1\parantheses{1-\frac{R}{C}},\qquad R\in\parantheses{0,C}.\label{expression_burnashev_reliability}
\end{IEEEeqnarray}
Here, $C$ is the channel capacity and $C_1$ denotes the maximum relative entropy between two arbitrary conditional output distributions. Note that~\eqref{expression_burnashev_reliability} is strictly larger than the reliability function for the no-feedback case. Burnashev's proof relies on the asymptotic analysis of an achievability and a converse bound on the maximum rate obtainable with VLF codes, for a given average blocklength and a fixed average error probability. Yamamoto and Itoh~\cite{YamamotoAchievability1979} gave an alternative proof of Burnashev's achievability bound, which relies on a two-phase scheme: a standard transmission phase where feedback is not used at the transmitter is followed by a confirmation phase where the transmitter uses feedback to confirm/contradict the decision of the receiver. Berlin~\etal~\cite{BerlinSimpleConverse2009} provided a stronger version of Burnashev's converse bound, whose proof parallels the two-phase scheme in~\cite{YamamotoAchievability1979}. A one-phase scheme that achieves~\eqref{expression_burnashev_reliability} was proposed in~\cite{JensenShannonDivergenceNaghshvar2015}.
Polyanskiy~\etal~\cite{polyanskiy2011feedback} obtained a nonasymptotic converse bound that improves on Burnashev's one~\cite{burnashev1976data}. In the same work, an achievability bound is provided, which is used to show that with VLF codes one can approach capacity faster than in the fixed-blocklength case. Specifically, the \emph{channel dispersion}~\cite[Eq. (221)]{polyanskiy2010channel} turns out to be zero. The achievability bound used in~\cite{polyanskiy2011feedback} to prove this result is actually based on variable-length stop-feedback (VLSF) codes. In the VLSF setup, the feedback link is used by the receiver only to send a single bit indicating to stop the transmission of the current message. This setup is of interest from a practical point of view, because it encompasses hybrid automatic repetition request (ARQ) schemes. Note that VLSF codes are a special case of VLF codes.
In this paper, we shall focus on the binary erasure channel (BEC) and seek nonasymptotic achievability and converse bounds, for both the VLF and the VLSF setups, which improve on the ones available in the literature. Note that the two-phase converse bounds~\cite[Thm. 1]{burnashev1976data},~\cite[Thm. 6]{polyanskiy2011feedback} require that all entries of the channel transition matrix of the DMC are strictly positive\footnote{This is required for~$C_1$ in~\eqref{expression_burnashev_reliability} to be finite.}---an assumption that does not hold for the BEC. Bounds on the maximum rate achievable over a BEC in the VLF setup are provided in~\cite[Thm. 7]{polyanskiy2011feedback}. The achievability bound is based on a simple scheme (also suggested in~\cite{ForneyExponentialBounds1968}) where each bit is repeated until it is received correctly. The converse bound can be seen as a variable-length analogue of Fano's inequality (see~\cite[Lemmas 1 and 2]{burnashev1976data}). This bound does not require $C_1$ to be finite.
The problem of constructing VLSF codes over a BEC reduces to problem of constructing rateless erasure codes. Thus, one can get achievability bounds on the maximum coding rate in the VLSF setup by analyzing the performance of family of rateless codes such as random linear fountain codes~\cite[Sec. 3]{mackay2005fountain}. The only converse bounds that are available for VLSF codes (stop feedback) hold also in the VLF setup (full feedback) to the best of the authors' knowledge. This is actually the case for both the maximum coding rate and the reliability function.
Our contributions in this paper are as follows:
\begin{itemize}[leftmargin=*]
\item We provide nonasymptotic converse and achievability bounds on the maximum coding rate of VLF codes over BECs, which improve upon the ones provided in~\cite[Thm. 7]{polyanskiy2011feedback}. Our converse bound relies on sequential hypothesis testing and is inspired by the meta-converse framework~\cite[Sec. III.E]{polyanskiy2010channel}; the achievability bound combines the simple repetition scheme used in~\cite[Thm. 7]{polyanskiy2011feedback} with variable-length Huffman coding. For the case of zero error probability, the achievability and converse bounds match.
\item For the VLSF setup, we provide nonasymptotic achievability bounds that improve on the one reported in~\cite[Thm. 3]{polyanskiy2011feedback}. The bounds are obtained by exploiting that, for a BEC, the decoder is able to identify the correct message whenever only a single codeword is compatible with the sequence of channel outputs received up to that point (a property noted previously in e.g.,~\cite{massay2007zeroerror}). The random coding argument used in one of the bounds utilizes linear codes. Hence, the resulting coding scheme can be seen as variable-length extension of random linear fountain codes.
\end{itemize}
\iflongversion\else Some proofs are omitted for space constraint; they can be found in~\cite{devassy2016nonasymptoticlong}. \fi
\subsubsection*{Notation} Uppercase curly letters denote sets. The~$n$-fold Cartesian product of a set~$\chinpspace$ is denoted by~$\chinpspace^n.$ Uppercase boldface letters denote random quantities and lightface letters denote deterministic quantities. The distribution of a random variable~$\chinp$ is denoted by~$P_\chinp.$ With~$\Bexpectation{\cdot}$ we denote expectation and with~$\Bcondexpectation{P}{\cdot}$ we stress that the expectation is with respect to the probability law~$P.$ The indicator function is denoted by~$\indicator{\cdot}$ and we use the symbol~$\binaryfinitefield$ to indicate the binary Galois field. With~$\chinp_m^n$ we denote the random vector with entries~$\parantheses{\chinp_m,\chinp_{m+1},\dots,\chinp_n}.$ Similarly,~$x_m^n$ stands for a deterministic vector with entries~$\parantheses{x_m,x_{m+1},\dots,x_n}.$ We shall often use the following function:
\begin{IEEEeqnarray}{rCl}
\lstar{x}&=&\floor{\log_2x}+2(1-2^{\floor{\log_2x}-\log_2x}),\qquad x\in\reals.\label{defnition_huffman_avg_length}
\end{IEEEeqnarray}
Here,~$\floor{\cdot}$ denotes the floor operator. Furthermore, we shall use~$\ceil{\cdot}$ to denote the ceil operator. We let~$\Bbernoullidist{p}$ denote a Bernoulli-distributed random variable with parameter~$p$ and~$\Bgeometricdist{p}$ a geometrically distributed random variable with parameter~$p.$ The binary entropy function~$\Bbinaryentropy{\cdot}$ is defined as follows:
\begin{IEEEeqnarray}{rCl}
\Bbinaryentropy{x}&=&-x\log_2 x - (1-x)\log_2(1-x),\iflongversion\qquad\else\quad\fi x\in\parantheses{0,1}.\IEEEeqnarraynumspace
\end{IEEEeqnarray}
\section{Definition}\label{section_system_model}
We consider a BEC with input alphabet~$\chinpspace=\curlybrac{0,1}$ and output alphabet~$\choutspace=\curlybrac{0,\erasure,1},$ where~$\erasure$ denotes an erasure. A VLF code for the BEC is defined as follows.
\begin{defn}\label{definition_vlf_codes}
\textit{(\cite[Def. 1]{polyanskiy2011feedback})} An~$(\latency,\messagecount,\errorprob)$--VLF code, where~$\latency$ is a positive real,~$\messagecount$ is a positive integer, and~$\errorprob\in\closedclosedinterval{0,1},$ consists~of:
\begin{enumerate}[leftmargin=*]
\item A random variable~$\commonrand,$ defined on a set~$\commonrandspace$ with\footnote{The bound on the cardinality of~$\commonrandspace$ given in~\cite{polyanskiy2011feedback} (i.e.,~$\abs{\commonrandspace}\leq 3$) can be improved by using the Fenchel-Eggleston theorem~\cite[p. 35]{eggleston1958convexity} in place of Caratheodory's theorem.}~$\abs{\commonrandspace}\leq 2,$ whose realization is revealed to the encoder and the decoder before the start of transmission. The random variable~$\commonrand$ acts as common randomness and enables the use of randomized encoding and decoding strategies.
\item A sequence of encoders~$f_n:\commonrandspace\times\messagespace\times\choutspace^{n-1}\functionto\chinpspace,n\geq 1$ that generate the channel inputs
\begin{IEEEeqnarray}{rCl}
\chinp_n &=& f_n\parantheses{\commonrand,\inpmessage,\chout_1^{n-1}}. \label{expression_encoder_bec_vlf}
\end{IEEEeqnarray}
Here,~$\inpmessage$ denotes the message, which is uniformly distributed on~$\messagespace=\curlybrac{1,2,\dots,\messagecount}.$ Note that the channel input at time~$n$ depends on all previous channel outputs (full feedback).
\item A sequence of decoders~$g_n:\commonrandspace\times\choutspace^n\functionto\messagespace$ that provide the estimate of~$\inpmessage$ at time~$n.$
\item A nonnegative integer-valued random variable~$\stoppingtime,$ which is a stopping time of the filtration
\begin{IEEEeqnarray}{rCl}
\mathcal{G}_n=\sigma\curlybrac{\commonrand,\chout_1^n} \label{expression_filtration_bec_vlf}
\end{IEEEeqnarray}\vspace{-9pt}
and satisfies
\begin{IEEEeqnarray}{rCl}
\Bexpectation{\stoppingtime} &\leq& \latency.\label{expression_stoppingtime_bec_vlf}
\end{IEEEeqnarray}
\item The final estimate~$\decoderoutput = g_\stoppingtime\parantheses{\commonrand,\chout_1^\stoppingtime}$ of~$\inpmessage,$ which satisfies the error-probability constraint
\begin{IEEEeqnarray}{rCl}
\probof{\decoderoutput\neq\inpmessage}&\leq&\errorprob.\label{expression_prob_error_constraint_bec_vlf}
\end{IEEEeqnarray}
\end{enumerate}
\end{defn}
The rate~$\rate$ of an~$(\latency,\messagecount,\errorprob)$--VLF code is defined as
\begin{IEEEeqnarray}{rCl}
\rate &=& \frac{\log_2\messagecount}{\Bexpectation{\stoppingtime}}.\label{expression_rate_bec_vlf}
\end{IEEEeqnarray}
Furthermore, we define the minimum average blocklength of VLF codes with~$\messagecount$ codewords and error probability not exceeding~$\errorprob$ as follows:
\begin{IEEEeqnarray}{rCl}
\vlffundlatencylimit{\messagecount}{\errorprob}&=&\min\curlybrac{\latency\ :\ \exists(\latency,\messagecount,\errorprob)\text{--VLF code}}.\label{expression_min_avg_latency_vlf}
\end{IEEEeqnarray}
VLSF codes are a special case of VLF codes. The peculiarity of VLSF codes is that the sequence of encoders is not allowed to depend on the past channel outputs, i.e.,
\begin{IEEEeqnarray}{rCl}
f_n:\commonrandspace\times\messagespace\functionto\chinpspace,\ n\geq 1.\label{expression_encoder_bec_vlsf}
\end{IEEEeqnarray}
In the VLSF case, the feedback link is used by the receiver only to inform the transmitter that the message has been decoded (stop/decision feedback).
\section{Existing Results for BEC}\label{section_existing_results}
In this section, we review the results available in literature on the minimum average blocklength~$\vlffundlatencylimit{\messagecount}{\errorprob}$ for the BEC. The following achievability bound is obtained by time-sharing between a scheme that drops the message to be transmitted without using the channel at all, and a scheme that repeats the channel input until it is received correctly.
\begin{thm}\label{theorem_achievability_BEC_yury_zero_prob}
\emph{(\cite[Thm. 7]{polyanskiy2011feedback})} For a BEC with erasure probability~$\becerasureprob,$ there exists an~$(\latency,\messagecount,\errorprob)$--VLF code with
\begin{IEEEeqnarray}{rCl}
\latency&\leq& \frac{\parantheses{1-\errorprob}\ceil{\log_2\messagecount}}{1-\becerasureprob}.\label{expression_achievability_BEC_yury_zero_prob}
\end{IEEEeqnarray}
\end{thm}
Next, we provide a converse bound.
\begin{thm}\label{theorem_converse_BEC_fano_zero_prob}
\emph{(\cite[Thm. 7]{polyanskiy2011feedback},\cite[Lemmas 1 and 2]{burnashev1976data})} For every~$(\latency,\messagecount,\errorprob)$--VLF code with~$0\leq\errorprob\leq 1-1/\messagecount$ operating over a BEC with erasure probability~$\becerasureprob,$ we have
\begin{IEEEeqnarray}{rCl}
\latency&\geq& \frac{(1-\errorprob)\log_2\messagecount-\Bbinaryentropy{\errorprob}}{1-\becerasureprob}.\label{expression_converse_BEC_fano_zero_prob}
\end{IEEEeqnarray}
\end{thm}
This converse bound can be obtained by constructing an appropriate martingale using the conditional entropy of the \emph{a posteriori} distribution of the message given the channel output. Note that the bounds~\eqref{expression_achievability_BEC_yury_zero_prob} and~\eqref{expression_converse_BEC_fano_zero_prob} coincide for~$\errorprob=0$ whenever the number of messages~$\messagecount$ is a power of~$2.$
\section{Novel Bounds for VLF Codes}\label{section_results_bec_vlf}
In this section, we present an achievability and a converse bound on~$\vlffundlatencylimit{\messagecount}{\errorprob}$ that improve upon the ones given in Theorems~\ref{theorem_achievability_BEC_yury_zero_prob} and~\ref{theorem_converse_BEC_fano_zero_prob}. The idea behind the achievability bound is to combine the scheme in Theorem~\ref{theorem_achievability_BEC_yury_zero_prob} with a Huffman code, whose purpose is to reduce the average blocklength when the number of messages is not a power of two. The converse bound relies on sequential hypothesis testing and is inspired by the meta-converse framework~\cite[Sec. III.E]{polyanskiy2010channel}. As we shall see, achievability and converse bounds are tight when~$\errorprob=0,$\emph{ for every} integer~$\messagecount.$ Our achievability bound is given in Theorem~\ref{theorem_achievability_bec_vlf} below.\ignorespaces
\begin{thm}\label{theorem_achievability_bec_vlf}
For a BEC with erasure probability~$\becerasureprob,$ there exists an~$(\latency,\messagecount,\errorprob)$--VLF code with
\begin{IEEEeqnarray}{rCl}
\latency&\leq& \frac{\parantheses{1-\errorprob}\lstar{\messagecount}}{1-\becerasureprob}\label{expression_vlf_achievability}
\end{IEEEeqnarray}
where~$\lstar{\cdot}$ is defined in~\eqref{defnition_huffman_avg_length}.
\end{thm}
\begin{IEEEproof}
\iflongversion
See Appendix~\ref{appendix_proof_achievability_bec_vlf}.
\else
See~\cite[App. A]{devassy2016nonasymptoticlong}.
\fi
\end{IEEEproof}
The converse bound is based on binary sequential hypothesis testing~\cite{wald1945sequential}. Let\footnote{We use the same notation as in~\cite[Ch. 3]{tartakovsky2014sequential}.}~$\parantheses{\decisionrule,\stoppingtime}$ denote a generic binary sequential hypothesis test between two stochastic processes~$\processP$ and~$\processQ.$ Here,~$\stoppingtime$ is a stopping time and~$\decisionrule$ is a decision rule ($0$ indicates that~$\processP$ is chosen and~$1$ that~$\processQ$ is chosen). Let~$\Btaurulespace{\processP}{\processQ}$ denote the set of all possible binary sequential hypothesis tests. We are interested in the minimum average number of samples~$\Bprocessminavglatency{\testpowerP}{\testpowerQ}{\processP}{\processQ}$ required by a binary sequential hypothesis test to identify~$\processP$ and~$\processQ$ correctly with probability at least~$\testpowerP$ and~$\testpowerQ,$ respectively. Formally,
\begin{IEEEeqnarray}{rCl}
\Bprocessminavglatency{\testpowerP}{\testpowerQ}{\processP}{\processQ} &=& \min_{\substack{\parantheses{\decisionrule,\stoppingtime}\in\Btaurulespace{\processP}{\processQ},\\ P\squarebrac{\decisionrule=0}\geq\testpowerP,\\ Q\squarebrac{\decisionrule=1}\geq\testpowerQ}} \Bcondexpectation{P}{\stoppingtime}.\label{defnition_minlatency_given_testpower}
\end{IEEEeqnarray}
In Lemma~\ref{lemma_metaconverse_bec_vlf} below we establish a connection between~$\Bprocessminavglatency{\testpowerP}{\testpowerQ}{\processP}{\processQ}$ and the parameters of a given VLF code. For the sake of generality, the lemma is formulated for the case of arbitrary DMCs (this requires a suitable generalization of the definition of VLF codes provided in Definition~\ref{definition_vlf_codes} to arbitrary DMCs).
\begin{lem}\label{lemma_metaconverse_bec_vlf}
Consider an~$(\latency,\messagecount,\errorprob)$--VLF code for the DMC~$P_{\chout\given\chinp}.$ Let~$\errorprobQ$ denote the probability of error when this code is used over the DMC~$Q_{\chout\given\chinp}.$ Let~$\vlfprocessPygivenx$ and~$\vlfprocessQygivenx$ be the probability distribution of the process~$\commonrand,\curlybrac{\parantheses{\chinp_n,\chout_n}}_{n=1}^\infty$ under~$P_{\chout\given\chinp}$ and~$Q_{\chout\given\chinp},$ respectively. The distributions of the stochastic processes depend on the chosen~$(\latency,\messagecount,\errorprob)$--VLF code through its encoder according to~\eqref{expression_encoder_bec_vlf}. We consider binary sequential hypothesis testing between the two processes, under the assumption that the realization of~$\commonrand$ is known to the test before processing~$(\chinp_1,\chout_1).$ We have
\begin{IEEEeqnarray}{rCl}
\latency\geq \Bprocessminavglatency{1-\errorprob}{\errorprobQ}{\vlfprocessPygivenx}{\vlfprocessQygivenx}.\IEEEeqnarraynumspace\label{expression_metaconverse_bec_vlf}
\end{IEEEeqnarray}
\end{lem}
\begin{IEEEproof}\ignorespaces
See Appendix~\ref{appendix_proof_metaconverse_bec_vlf}.\ignorespaces
\end{IEEEproof}
The bound~\eqref{expression_metaconverse_bec_vlf} can be viewed as the variable-length analogue of the meta-converse theorem~\cite[Thm. 26]{polyanskiy2010channel}. The meta-converse theorem links the average error probabilities resulting by using the same fixed-blocklength code over two different channels by means of binary hypothesis testing. Similarly, Lemma~\ref{lemma_metaconverse_bec_vlf} relates the average error probabilities and the average blocklengths resulting by using a given VLF code over two different channels by means of binary sequential hypothesis testing.
To obtain a converse bound from~\eqref{expression_metaconverse_bec_vlf}, we take~$Q_{\chout\given\chinp}=Q_{\chout},$ with~$Q_{\chout}$ being the capacity-achieving output distribution of the BEC. Then, we solve the minimization in~\eqref{defnition_minlatency_given_testpower} by using the sequential probability ratio rest (SPRT)~\cite{wald1945sequential} (see\iflongversion Appendix~\ref{appendix_wald_sprt}\else~\cite[App. C]{devassy2016nonasymptoticlong} \fi for a short review). This yields the following bound.
\begin{thm}\label{theorem_converse_bec_vlf}
Every~$(\latency,\messagecount,\errorprob)$--VLF code operating over a BEC with erasure probability~$\becerasureprob$ satisfies
\begin{IEEEeqnarray}{rCl}
\latency &\geq& \frac{\parantheses{1-\errorprob}\lstar{\messagecount(1-\errorprob)}}{1-\becerasureprob}.\label{expression_converse_bec_vlf}
\end{IEEEeqnarray}
\end{thm}
\begin{IEEEproof}
\iflongversion
See Appendix~\ref{appendix_theorem_converse_bec_vlf}.
\else
See~\cite[App. D]{devassy2016nonasymptoticlong}.
\fi
\end{IEEEproof}
We would like to emphasize that although~\eqref{expression_converse_bec_vlf} is tighter than the converse bound reported in~\cite[Thm. 6]{polyanskiy2011feedback}, a generalization of Theorem~\ref{theorem_converse_bec_vlf} to DMCs with finite~$C_1$ yields a converse bound that is in general looser than the ones reported in~\cite[Thm. 1]{burnashev1976data} and~\cite[Thm. 6]{polyanskiy2011feedback}. The peculiarity of the BEC is that the decoder is able to determine if its estimate is correct or not by assessing whether the estimated codeword is the only codeword compatible with the erasure pattern. This implies that a two-phase scheme with a confirmation from the encoder is not required.
Note that the right-hand sides of~\eqref{expression_vlf_achievability} and~\eqref{expression_converse_bec_vlf} coincide when~$\errorprob=0.$ This fact is collected in the following corollary.\ignorespaces
\begin{cor}\label{corollary_vlf_min_latency_zero_error}
The minimum average blocklength~$\vlffundlatencylimit{\messagecount}{0}$ of an~$(\latency,\messagecount,0)$--VLF code over a BEC with erasure probability~$\becerasureprob$ is given by\ignorespaces
\begin{IEEEeqnarray}{rCl}
\vlffundlatencylimit{\messagecount}{0} &=& \frac{\lstar{\messagecount}}{1-\becerasureprob}\label{expression_vlf_min_latency}
\end{IEEEeqnarray}\ignorespaces
where~$\lstar{\cdot}$ is defined in~\eqref{defnition_huffman_avg_length}.
\end{cor}
\section{Novel Bounds for VLSF Codes}\label{section_vlsf_results}
We now focus on the VLSF setup and provide achievability bounds for the case~$\errorprob=0.$ Achievability bounds for arbitrary~$\errorprob$ can be obtained by allowing the receiver to send a stop signal at time zero with probability~$\errorprob$ (see~\cite[Sec. III.D]{polyanskiy2011feedback}). The corresponding achievability bounds can be readily obtained from the ones presented in this section by multiplying them by~$\parantheses{1-\errorprob}.$
As already mentioned, in the BEC case the decoder can assess the correctness of its message estimate by verifying whether the codeword corresponding to the chosen message is the only one that is compatible with the received sequence. It is therefore natural to consider a decoder whose stopping time is given by\ignorespaces
\begin{IEEEeqnarray}{rCl}
\stoppingtime &=& \inf \curlybrac{n\geq 1: \probof{\inpmessage = \Rdecoderoutput_n \given \chout_1^n = \Rchout_1^n} = 1 } \label{expression_bec_vlsf_optimal_stopping_time}
\end{IEEEeqnarray}\ignorespaces
where~$\Rdecoderoutput_n$ denotes the message estimate at the decoder after~$n$ channel uses:
\begin{IEEEeqnarray}{rCl}
\Rdecoderoutput_n &=& \argmax_{w\in\messagespace} \probof{\inpmessage=w\given \chout_1^n = \Rchout_1^n}.\label{expression_bec_vlsf_optimal_decoder}
\end{IEEEeqnarray}
The decoding rule~\eqref{expression_bec_vlsf_optimal_stopping_time}--\eqref{expression_bec_vlsf_optimal_decoder} combined with random coding (independent and identically distributed (\iid)~$\Bbernoullidist{0.5}$ ensemble) yields the following achievability bound.\ignorespaces
\begin{thm}\label{theorem_bec_achievability_iid_no_union_bound}
For a BEC with erasure probability~$\becerasureprob,$ there exists an~$(\latency,\messagecount,0)$--VLSF code with
\begin{IEEEeqnarray}{rCl}
\latency &\leq& \frac{1}{1-\becerasureprob}\parantheses{1-\sum_{i=1}^{\messagecount-1}\binom{\messagecount-1}{i}\frac{\parantheses{-1}^i}{2^i-1}}.\label{expression_bec_achievability_iid_no_union_bound}
\end{IEEEeqnarray}
\end{thm}
\begin{IEEEproof}\ignorespaces
See Appendix~\ref{appendix_proof_bec_achievability_iid_no_union_bound}.\ignorespaces
\end{IEEEproof}
The achievability bound~\eqref{expression_bec_achievability_iid_no_union_bound} suffers from two pitfalls: (i) the bound is loose when~$\messagecount$ is small because the random coding ensemble contains few codebooks with abnormally large average blocklengths; (ii) since the bound requires the computation of differences of binomial coefficients, it becomes difficult to compute when~$\messagecount$ is larger than~$10^4.$ Next, we present a different achievability bound that addresses these two shortcomings. To tighten~\eqref{expression_bec_achievability_iid_no_union_bound} for small~$\messagecount$ we use an expurgation technique similar to the ones utilized by Shannon, Gallager, and Berlekamp~\cite[p. 529]{ShannonGallagher1967522}. Specifically, we view each codebook as a random matrix with~$\messagecount$ rows and an infinite number of columns and we assign the following probability distribution on the VLSF code ensemble: each column is drawn uniformly and independently from the set of binary vectors with~$\ceil{\messagecount/2}$ zeros. Furthermore, to obtain an expression that is computable for arbitrary values of~$\messagecount,$ we upper-bound the average blocklength of the expurgated ensemble using the union bound. The achievability bound thus obtained is given in the following theorem.\ignorespaces
\begin{thm}\label{theorem_bec_achiveability_vlsf_zero_error}
For a BEC with erasure probability~$\becerasureprob,$ there exists an~$(\latency,\messagecount,0)$--VLSF code with
\begin{IEEEeqnarray}{rCl}
\latency&\leq& \frac{1}{1-\becerasureprob}\parantheses{\floor{m}+1+\frac{\invprobofunequalbits^{m-\floor{m}}}{\invprobofunequalbits-1}} \label{expression_bec_achiveability_vlsf_zero_error}
\end{IEEEeqnarray}
where~$m=\log_\invprobofunequalbits\parantheses{\messagecount-1}$ and~$\invprobofunequalbits$ is related to the number of messages~$\messagecount$ as follows:
\begin{IEEEeqnarray}{rCl}
\invprobofunequalbits &=& 2+\frac{1}{\ceil{\messagecount/2}-1}.\label{expression_inv_prob}
\end{IEEEeqnarray}
\end{thm}
\begin{IEEEproof}
\iflongversion
See Appendix~\ref{appendix_proof_bec_achiveability_vlsf_zero_error}.
\else
See~\cite[App. G]{devassy2016nonasymptoticlong}.
\fi
\end{IEEEproof}
The parameter~$\invprobofunequalbits$ in~\eqref{expression_inv_prob} is the reciprocal of the probability that the first two bits in a random vector that is uniformly distributed over the set of vectors in~$\binaryfinitefield^{\messagecount}$ with~$\ceil{\messagecount/2}$ zeros, are equal.
For the case when the number of messages~$\messagecount$ is a power of~$2,$ one can obtain an achievability bound that is tighter than~\eqref{expression_bec_achiveability_vlsf_zero_error}, that does not require the union bound, and that is easily computable. The bound relies on a linear codebook ensemble in which the columns of the random generator matrix are distributed uniformly over the set of all nonzero vectors from~$\binaryfinitefield^{\log_2\mathopen{}\mathclose\messagecount}.$ Specifically, consider a received vector of length~$n$ and let~$\indexset_n$ be the set containing the indices of unerased symbols in the received vector. We are interested in the dimension of the subspace spanned by the columns of the generator matrix with index in~$\indexset_n.$ This dimension evolves as a Markov chain with a single absorbing state (the state that corresponds to maximum dimension~$\log_2\mathopen{}\mathclose\messagecount$). it follows that the average blocklength (averaged over the ensemble) coincides with the expected absorption time of the Markov chain, which follows a discrete phase-type distribution~\cite[Ch. 2]{neuts1981matrix}. This achievability scheme can be seen as a variable-length analogue of random linear fountain codes~\cite[Sec. 3]{mackay2005fountain}. The performance of the achievability scheme just described is characterized in the following theorem.\ignorespaces
\begin{thm}\label{theorem_bec_achievability_expurgated_linear}
For every integer~$\messagebitcount\geq1,$ there exists an~$(\latency,2^\messagebitcount,0)$--VLSF code for a BEC with erasure probability~$\becerasureprob$ with
\begin{IEEEeqnarray}{rCl}
\latency &\leq& \frac{1}{1-\becerasureprob}\parantheses{\messagebitcount+\sum_{i=1}^{\messagebitcount-1}\frac{2^{i}-1}{2^k-2^{i}}}.\label{expression_bec_achievability_expurgated_linear}
\end{IEEEeqnarray}
\end{thm}
\begin{IEEEproof}
\iflongversion
See Appendix~\ref{appendix_proof_bec_achievability_expurgated_linear}.
\else
See~\cite[App. H]{devassy2016nonasymptoticlong}.
\fi
\end{IEEEproof}
The achievability bounds~\eqref{expression_bec_achievability_iid_no_union_bound},~\eqref{expression_bec_achiveability_vlsf_zero_error}, and~\eqref{expression_bec_achievability_expurgated_linear} are plotted in Fig.~\ref{fig_vlsf_bec}. \ignorespaces
\begin{figure}
\centerline{\includegraphics{bec_vlsf_fig.pdf}}
\caption{Achievability and converse bounds for zero error VLSF codes over BEC with~$\becerasureprob=0.5.$ The converse bound is the one given in~\eqref{expression_vlf_min_latency} for VLF codes.}\label{fig_vlsf_bec}\vspace{-15pt}\ignorespaces
\end{figure}\ignorespaces
\iftwocol
\begin{figure*}[!ht]\ignorespaces
\begin{IEEEeqnarray}{rCl}
\IEEEeqnarraymulticol{3}{l}{\probof{\errorevent=0\given \chinp_1^n=\Rchinp_1^n,\chout_1^n=\Rchout_1^n,\commonrand=\Rcommonrand,\stoppingtime=n} }\nonumber\\
&=& \sum_{w\in\messagespace}\probof{\inpmessage=w\given \chinp_1^n=\Rchinp_1^n,\chout_1^n=\Rchout_1^n,\commonrand=\Rcommonrand}\probof{\decoderoutput=w\given \chout_1^n=\Rchout_1^n,\commonrand=\Rcommonrand,\stoppingtime=n}.\IEEEeqnarraynumspace\label{expression_errorevent_conditioned}
\end{IEEEeqnarray}\ignorespaces
\hrulefill\ignorespaces\vspace{-10pt}
\end{figure*}\ignorespaces
\fi\ignorespaces
As expected, the bound~\eqref{expression_bec_achiveability_vlsf_zero_error} is tighter than~\eqref{expression_bec_achievability_iid_no_union_bound} at small average blocklengths (because of expurgation) and looser at large average blocklengths (because of union bound). The achievability bound~\eqref{expression_bec_achievability_expurgated_linear} is tighter than~\eqref{expression_bec_achiveability_vlsf_zero_error} for all blocklengths and looser than~\eqref{expression_bec_achievability_iid_no_union_bound} for large average blocklengths. When~$\messagecount=2,$ the achievability bounds~\eqref{expression_bec_achiveability_vlsf_zero_error} and~\eqref{expression_bec_achievability_expurgated_linear} coincide with the converse bound for VLF codes given in~\eqref{expression_vlf_min_latency}. This holds because the scheme that achieves~\eqref{expression_vlf_min_latency} (repeat each bit until it is received correctly) can be implemented with stop feedback when\footnote{Recall that in the VLSF setup the decoder is allowed to send only one stop signal per message.}~$\messagecount=2.$ As~$\messagecount$ increases, the gap between the VLSF achievability bounds and the VLF converse bound increases and gets as large as~$23\%$ when~$\messagecount=8,$ before vanishing asymptotically as~$\messagecount\tendsto\infty.$ It remains to be seen whether this gap is fundamental.\ignorespaces
\begin{appendices}
\iflongversion
\section{Proof of Theorem~\ref{theorem_achievability_bec_vlf}}\label{appendix_proof_achievability_bec_vlf}
We use time sharing between a scheme that drops the message to be transmitted without using the channel at all, and a zero-error VLF code constructed as follows: we first generate a prefix-free Huffman code~\cite{huffman1952method} for the~$\messagecount$ equiprobable messages. To send a given message, we repeat each bit of the corresponding Huffman codeword until it is received correctly (note that this requires full feedback at the transmitter). The average blocklength of the resulting VLF code can be analyzed as follows. Let~$\huffmanavglength{\messagecount}$ denote the average code length of the Huffman code for the~$\messagecount$ equiprobable messages. The average blocklength of the VLF code resulting from our construction is given by
\begin{IEEEeqnarray}{rCl}
\latency&=&\frac{\huffmanavglength{\messagecount}}{1-\becerasureprob}.\label{expression_huffman_avg_latency}
\end{IEEEeqnarray}
Note now that the length of the Huffman codeword assigned to each message is either~$\floor{\log_2\messagecount}$ or~$\floor{\log_2\messagecount}+1$ (see~\cite[p. 598]{ahlswede2006general}). Specifically, the Huffman code assigns codewords of length~$\floor{\log_2\messagecount}$ to~$2^{\floor{\log_2\messagecount}+1}-\messagecount$ messages, and codewords of length~$\floor{\log_2\messagecount}+1$ to the remaining messages. This allows us to conclude that
\begin{IEEEeqnarray}{rCl}
\huffmanavglength{\messagecount}&=&\lstar{\messagecount}\label{expression_huffman_avg_codelength}
\end{IEEEeqnarray}
where~$\lstar{\cdot}$ is defined in~\eqref{defnition_huffman_avg_length}. We obtain~\eqref{expression_vlf_achievability} by allowing the transmitter to drop each codeword with probability~$\errorprob$ (see~\cite[Sec. III.D]{polyanskiy2011feedback}).
\fi
\section{Proof of Lemma~\ref{lemma_metaconverse_bec_vlf}}\label{appendix_proof_metaconverse_bec_vlf}
\iftwocol\ignorespaces
Consider the random variable~$\errorevent=\indicator{\inpmessage\neq\decoderoutput}.$ The conditional distribution of~$\errorevent$ given $\chinp_1^n=\Rchinp_1^n,\chout_1^n=\Rchout_1^n,\commonrand=\Rcommonrand,\stoppingtime=n,$ which is given in~\eqref{expression_errorevent_conditioned}, does not depend on whether the underlying channel is~$P_{\chout\given\chinp}$ or~$Q_{\chout\given\chinp}.$
Indeed in~\eqref{expression_errorevent_conditioned}, the first factor depends only on the encoder and second factor depends only on the decoder.
\else
Consider the random variable~$\errorevent=\indicator{\inpmessage\neq\decoderoutput}.$ The conditional distribution of~$\errorevent$ given $\chinp_1^n=\Rchinp_1^n,\chout_1^n=\Rchout_1^n,\commonrand=\Rcommonrand,\stoppingtime=n$ does not depend on whether the underlying channel is~$P_{\chout\given\chinp}$ or~$Q_{\chout\given\chinp}.$ Indeed,
\begin{IEEEeqnarray}{rCl}
\IEEEeqnarraymulticol{3}{l}{\probof{\errorevent=0\given \chinp_1^n=\Rchinp_1^n,\chout_1^n=\Rchout_1^n,\commonrand=\Rcommonrand,\stoppingtime=n} }\nonumber\\
&=& \sum_{w\in\messagespace}\probof{\inpmessage=w\given \chinp_1^n=\Rchinp_1^n,\chout_1^n=\Rchout_1^n,\commonrand=\Rcommonrand}\probof{\decoderoutput=w\given \chout_1^n=\Rchout_1^n,\commonrand=\Rcommonrand,\stoppingtime=n}\IEEEeqnarraynumspace\label{expression_errorevent_conditioned}
\end{IEEEeqnarray}
where the first factor depends only on the encoder and second factor depends only on the decoder.
\fi
Using the stopping time~$\stoppingtime$ associated to the given code and the family of probability kernels defined by the conditional distribution~\eqref{expression_errorevent_conditioned} we construct a binary sequential hypothesis test~$\parantheses{\decisionrule,\stoppingtime}$. By definition, we have that under~$P_{\chout\given\chinp},$
{\setlength{\belowdisplayskip}{0pt} \setlength{\belowdisplayshortskip}{0pt}
\begin{IEEEeqnarray}{rCl}
\probof{\decisionrule=0} &=& 1-\errorprob
\end{IEEEeqnarray} and under~$Q_{\chout\given\chinp}$
\begin{IEEEeqnarray}{rCl}
\probof{\decisionrule=0}&=& 1-\errorprobQ.
\end{IEEEeqnarray} Thus,
\begin{IEEEeqnarray}{rCl}
\latency &\geq&\Bcondexpectation{\vlfprocessPygivenx}{\stoppingtime}\\
&\geq&\Bprocessminavglatency{1-\errorprob}{\errorprobQ}{\vlfprocessPygivenx}{\vlfprocessQygivenx}.\IEEEeqnarraynumspace
\end{IEEEeqnarray}}\ignorespaces\vspace{-10pt}
\iflongversion
\section{Sequential Probability Ratio Test (SPRT)}\label{appendix_wald_sprt}\ignorespaces
In this appendix, we provide a brief overview of the sequential probability ratio test (SPRT)~\cite{wald1945sequential} and discuss its optimality. Let~$\parantheses{\decisionrule,\stoppingtime}$ denote a generic binary sequential hypothesis test between two stationary memoryless stochastic processes with marginal distribution~$P_{\chinp}$ and~$Q_\chinp.$ Here,~$\stoppingtime$ is a stopping time---a random variable denoting the number of samples taken before making a decision---and~$\decisionrule$ is a decision rule ($0$ indicating that~$P_{\chinp}$ is chosen and~$1$ that~$Q_{\chinp}$ is chosen). With~$\Bcondexpectation{P_\chinp}{\stoppingtime}$ and~$\Bcondexpectation{Q_\chinp}{\stoppingtime}$ we denote the average number of samples required under the hypothesis~$P_{\chinp}$ and~$Q_\chinp,$ respectively. The probability of correct decision under hypothesis~$P_{\chinp}$ and~$Q_\chinp$ are denoted by~$P_\chinp\parantheses{\decisionrule=0}$ and~$Q_\chinp\parantheses{\decisionrule=1},$ respectively.
\begin{defn}\label{defnition_wald_sprt}
Let the log-likelihood ratio (LLR) after~$n$ samples be recursively defined as follows:
\begin{IEEEeqnarray}{rCl}
S_0&=&0\\
S_n&=&S_{n-1}+\Blog{\frac{\infinitesimal P_{\chinp}}{\infinitesimal Q_{\chinp}}(\Rchinp_n)},\qquad n\geq 1
\end{IEEEeqnarray}
where~$\Rchinp_n$ is the~$n$th sample and~$\frac{\infinitesimal P_{\chinp}}{\infinitesimal Q_{\chinp}}(\cdot)$ denotes the Radon-Nikodym derivative. Let~$A_Q$ and~$A_P$ be two nonnegative scalars. The SPRT with stopping bounds~$A_Q$ and~$A_P$ is defined as follows:
\begin{IEEEeqnarray}{rCl}
\stoppingtime &=& \min\curlybrac{n\geq 0, S_n \notin \parantheses{-A_Q,A_P} } \label{expression_wald_sprt_stopping_rule}\\
\decisionrule &=& \left\{ \,
\begin{IEEEeqnarraybox}[][c]{l?l}
\IEEEstrut
1, & S_\stoppingtime\leq -A_Q \\
0, & S_\stoppingtime\geq A_P.
\IEEEstrut
\end{IEEEeqnarraybox}\right.\label{expression_wald_sprt_decision_rule}
\end{IEEEeqnarray}
\end{defn}
We denote the SPRT defined through~\eqref{expression_wald_sprt_stopping_rule} and~\eqref{expression_wald_sprt_decision_rule} by~$S(A_Q,A_P).$ Note that SPRT allows for the possibility that~$\stoppingtime=0$ (the test stops before processing the first sample). Specifically, the tests~$S(0,A_P)$ and~$S(A_Q,0)$ will stop at~$\stoppingtime=0$ (recall that~$S_0=0$) and declare the hypothesis to be~$Q_{\chinp}$ and~$P_\chinp,$ respectively. It will turn out convenient to denote with~$S(A_Q-,A_P)$ the test with stopping time given by
\begin{IEEEeqnarray}{rCl}
\stoppingtime &=& \min\curlybrac{n\geq 0, S_n \notin \closedopeninterval{-A_Q,A_P} }
\end{IEEEeqnarray}
and decision rule~\eqref{expression_wald_sprt_decision_rule}. Similarly, we use~$S(A_Q,A_P+)$ to denote the test with stopping time
\begin{IEEEeqnarray}{rCl}
\stoppingtime &=& \min\curlybrac{n\geq 0, S_n \notin \openclosedinterval{-A_Q,A_P} }
\end{IEEEeqnarray}
and decision rule~\eqref{expression_wald_sprt_decision_rule}. Finally, the test with stopping time
\begin{IEEEeqnarray}{rCl}
\stoppingtime &=& \min\curlybrac{n\geq 0, S_n \notin \closedclosedinterval{-A_Q,A_P} }
\end{IEEEeqnarray}
and decision rule~\eqref{expression_wald_sprt_decision_rule} is denoted by~$S(A_Q-,A_P+).$
A randomization between a finite collection of binary sequential hypothesis tests refers to a testing procedure where a test is randomly selected from the collection according to a given probability law. Such a testing procedure is also referred to as randomized test. Next, we provide an extension of SPRT that allows for randomization.
\begin{defn}\label{defnition_extended_sprt}
The \emph{extended} SPRT~\cite{Burkholder1963ptimumProperties} with stopping bounds~$A_Q,A_P$ is the test that chooses~$\decisionrule=1$ and~$\decisionrule=0$ when~$S_n<-A_Q$ and~$S_n>A_P,$ respectively, and requests the next sample when~$-A_Q<S_n<A_P.$ When~$S_n=-A_Q,$ a possibly randomized rule is adopted to decide whether to set~$\decisionrule=1$ or to request the next sample. Similarly, when~$S_n=A_P,$ a possibly randomized rule is adopted to decide whether to set~$\decisionrule=0$ or to request the next sample.
\end{defn}
As noted in~\cite[Rem. 2.1]{Burkholder1963ptimumProperties}, every randomized test obtained by randomizing between~$S(A_Q,A_P),$ $S(A_Q-,A_P),$ $S(A_Q,A_P+),$ and~$S(A_Q-,A_P+)$ is an extended SPRT with stopping bounds~$A_Q,A_P$. Next, we shall define the following \emph{optimum property}.
\begin{defn}\label{defnition_optimum_property}
A binary sequential hypothesis test~$(\decisionrule^*,\stoppingtime^*)$ is said to have optimum property~(OP) if~$\Bcondexpectation{P_\chinp}{\stoppingtime^*}$ and~$\Bcondexpectation{Q_\chinp}{\stoppingtime^*}$ are finite, and for every other test~$(\decisionrule,\stoppingtime)$ with finite~$\Bcondexpectation{P_\chinp}{\stoppingtime}$ and~$\Bcondexpectation{Q_\chinp}{\stoppingtime}$ the conditions
\begin{IEEEeqnarray}{rCl}
P_\chinp\parantheses{\decisionrule=0} &\geq&P_\chinp\parantheses{\decisionrule^*=0}\\
Q_\chinp\parantheses{\decisionrule=1} &\geq&Q_\chinp\parantheses{\decisionrule^*=1}
\end{IEEEeqnarray}
imply that
\begin{IEEEeqnarray}{rCl}
\Bcondexpectation{P_\chinp}{\stoppingtime^*}&\leq&\Bcondexpectation{P_\chinp}{\stoppingtime}\\
\Bcondexpectation{Q_\chinp}{\stoppingtime^*}&\leq&\Bcondexpectation{Q_\chinp}{\stoppingtime}.
\end{IEEEeqnarray}
\end{defn}
In~\cite{Wald1948OptimumCharacter} it is proven that every SPRT has OP. This result was later extended in~\cite[Cor. 2.1]{Burkholder1963ptimumProperties} to prove that every extended SPRT has OP as well. Since for every~$\testpowerP$ and~$\testpowerQ$ in~\eqref{defnition_minlatency_given_testpower} we can find an extended SPRT that satisfies~$P\squarebrac{\decisionrule=0}=\testpowerP$ and~$Q\squarebrac{\decisionrule=1}=\testpowerQ,$ we conclude that extended SPRT minimizes~\eqref{defnition_minlatency_given_testpower}. Note that randomization is in general required to guarantee that the test achieves~$P\squarebrac{\decisionrule=0}=\testpowerP$ and~$Q\squarebrac{\decisionrule=1}=\testpowerQ$ for every arbitrary pair of probabilities~$\testpowerP$ and~$\testpowerQ.$
\fi\iflongversion
\section{Proof of Theorem~\ref{theorem_converse_bec_vlf}}\label{appendix_theorem_converse_bec_vlf}
We use Lemma~\ref{lemma_metaconverse_bec_vlf} with~$Q_{\chout\given\chinp}=Q_{\chout},$ where~$Q_{\chout}$ is the capacity-achieving output distribution of the BEC, i.e.,
\begin{IEEEeqnarray}{rCl}
Q_{\chout}(0) &=& Q_{\chout}(1) = \frac{1-\becerasureprob}{2}\\
Q_{\chout}(\erasure) &=& \becerasureprob.
\end{IEEEeqnarray}
The error probability~$\errorprobQ$ of a given~$(\latency,\messagecount,\errorprob)$--VLF code over the channel~$Q_{\chout}$ can be evaluated as follows:
\begin{IEEEeqnarray}{rCl}
\IEEEeqnarraymulticol{3}{l}{1-\errorprobQ}\nonumber\\
&=&\sum_{\Rcommonrand\in\commonrandspace}\sum_{n=0}^\infty\sum_{\substack{\Rchinp_1^n\in\chinpspace^n \\ \Rchout_1^n\in\choutspace^n}}\sum_{w\in\messagespace}\frac{P_{\commonrand}\squarebrac{\Rcommonrand}\prod_{k=1}^n P_{\chinp_k \given \chout_1^{k-1},\inpmessage,\commonrand}\squarebrac{\Rchinp_k\given \Rchout_1^{k-1},w,\Rcommonrand} Q_{\chout_1^n}\squarebrac{\Rchout_1^n} P_{\decoderoutput,\stoppingtime \given \chout_1^n,\commonrand}\squarebrac{w,n\given \Rchout_1^n,\Rcommonrand}}{\messagecount}\nonumber\\
\label{expression_full_eps_q}
\\
&=&\sum_{\Rcommonrand\in\commonrandspace}\sum_{n=0}^\infty\sum_{\Rchout_1^n\in\choutspace^n}\sum_{w\in\messagespace}\frac{P_{\commonrand}\squarebrac{\Rcommonrand}Q_{\chout_1^n}\squarebrac{\Rchout_1^n}P_{\decoderoutput,\stoppingtime \given \chout_1^n,\commonrand}\squarebrac{w,n\given \Rchout_1^n,\Rcommonrand}}{\messagecount}\label{expression_xgone_eps_q}\\
&=&\sum_{\Rcommonrand\in\commonrandspace}\sum_{n=0}^\infty\frac{P_{\commonrand}\squarebrac{\Rcommonrand}P_{\stoppingtime\given \commonrand}\squarebrac{n\given \Rcommonrand}}{\messagecount}= \frac{1}{\messagecount}.\label{expression_full_eps_q_last_step}
\end{IEEEeqnarray}
To obtain~\eqref{expression_xgone_eps_q}, we used that
\begin{IEEEeqnarray}{rCl}
\sum_{\Rchinp_1^n\in\chinpspace^n}\prod_{k=1}^n f_k(x_k) &=& \prod_{k=1}^n\sum_{\Rchinp_k\in\chinpspace} f_k(x_k)
\end{IEEEeqnarray}
where~$f_k(\cdot)$ are arbitrary functions. Thus, we have
\begin{IEEEeqnarray}{rCl}
\errorprobQ &=& 1-1/\messagecount.
\end{IEEEeqnarray}
We now proceed to solve the minimization in~\eqref{defnition_minlatency_given_testpower} for the case~$\testpowerP=1-\errorprob,\testpowerQ=1-1/\messagecount,P=\vlfprocessPygivenx,Q=\vlfprocessQygivenx.$ Here,~$\vlfprocessQygivenx$ denotes the distribution of the stochastic process~$\commonrand,\curlybrac{\parantheses{\chinp_n,\chout_n}}_{n=1}^{\infty}$ under channel~$Q_\chout.$ Specifically,
\begin{IEEEeqnarray}{rCl}
Q_{\commonrand,\chinp_1^n,\chout_1^n}\squarebrac{\Rcommonrand,\Rchinp_1^n,\Rchout_1^n}&=& \sum_{\Rinpmessage\in\messagespace} P_\commonrand\squarebrac{\Rcommonrand}P_{\inpmessage\given\commonrand}\squarebrac{\Rinpmessage\given\Rcommonrand}\prod_{k=1}^{n}P_{\chinp_k\given\inpmessage,\commonrand,\chout_1^{k-1}}\squarebrac{\Rchinp_k\given\Rinpmessage,\Rcommonrand,\Rchout_1^{k-1}}Q_{\chout}\squarebrac{\Rchout_k}.\IEEEeqnarraynumspace
\end{IEEEeqnarray}
We shall show that the binary sequential hypothesis test that achieves the minimum in~\eqref{defnition_minlatency_given_testpower} is the extended SPRT reviewed in Appendix~\ref{appendix_wald_sprt}. To do so, we need to show that the LLR process (see Def.~\ref{defnition_wald_sprt} in Appendix~\ref{appendix_wald_sprt})
\begin{IEEEeqnarray}{c}
\curlybrac{\llr_n=\log{\frac{P_{\commonrand,\chinp_1^n,\chout_1^n}}{Q_{\commonrand,\chinp_1^n,\chout_1^n}}}}_{n=0}^{\infty}
\end{IEEEeqnarray}
is a process with~\iid increments~\cite[p. 157]{tartakovsky2014sequential}. Indeed, consider the following quantity:
\begin{IEEEeqnarray}{rCl}
\conditionalllr_n &=& \log\frac{P_{\chout_n\given\chinp_n}\squarebrac{\Rchout_n\given\Rchinp_n}}{Q_{\chout_n}\squarebrac{\Rchout_n}}.
\end{IEEEeqnarray}
One can verify that~$\llr_n$ can be expressed as follows:
\begin{IEEEeqnarray}{rCl}
\bS_0&=&0\\
\bS_n&=&\bS_{n-1}+\conditionalllr_n,\qquad n\geq 1.
\end{IEEEeqnarray}
Note now that under~$\vlfprocessPygivenx$ the distribution of $\conditionalllr_n$ is
\begin{IEEEeqnarray}{lCl}
\probof{\conditionalllr_n =\log 2}&=&1-\becerasureprob\label{expression_condionalllr_pmf_p_log2}\\
\probof{\conditionalllr_n =0}&=&\becerasureprob.\label{expression_condionalllr_pmf_p_0}
\end{IEEEeqnarray}
Furthermore, under~$\vlfprocessQygivenx$ the distribution of $\conditionalllr_n$ is
\begin{IEEEeqnarray}{lCl}
\probof{\conditionalllr_n =\log 2}&=&\frac{1-\becerasureprob}{2}\label{expression_condionalllr_pmf_q_log2}\\
\probof{\conditionalllr_n =0}&=&\becerasureprob\label{expression_condionalllr_pmf_q_0}\\
\probof{\conditionalllr_n =-\infty}&=&\frac{1-\becerasureprob}{2}.\label{expression_condionalllr_pmf_q_minusinfinity}
\end{IEEEeqnarray}
Moreover, the stochastic process~$\curlybrac{\conditionalllr_n}_{n=1}^{\infty}$ is~\iid under both distributions. This allows us to conclude that extended SPRT achieves the minimum in~\eqref{defnition_minlatency_given_testpower} (see~\cite[Sec. 3.2.3]{tartakovsky2014sequential}). Such a test will be a randomization between the following tests (see Appendix~\ref{appendix_wald_sprt} for a clarification on the notation used here):
\begin{IEEEeqnarray}{rCl}
S(0,m\log 2),S(0,m\log 2+),S(0-,m\log 2),S(0-,m\log 2+).
\end{IEEEeqnarray}
Here,~$m$ is a positive integer. Note that the tests~$S(0,m\log 2)$ and~$S(0,m\log 2+)$ stop at~$\stoppingtime=0$ and choose~$\vlfprocessQygivenx$. Furthermore, the test~$S(0-,m\log 2+)$ coincides with the test~$S(0-,(m+1)\log 2).$
The probability of correct decision~$\testpowerQ$ under hypothesis~$\vlfprocessQygivenx$ for the test~$S(0-,m\log 2)$ is given by
\begin{IEEEeqnarray}{rCl}
\testpowerQ&=&1-\parantheses{\frac{1-\becerasureprob}{2}}^m \sum_{n=0}^{\infty} \binom{n+m-1}{m-1} \becerasureprob^{n} \label{proof_step_bionmial_series}\\
&=& 1-2^{-m}. \label{proof_step_after_binomial_series}
\end{IEEEeqnarray}
In~\eqref{proof_step_bionmial_series}, we used that the probability of error under~$\vlfprocessQygivenx$ is the probability that a sequence of~\iid ternary random variables distributed according to~\eqref{expression_condionalllr_pmf_q_log2}-\eqref{expression_condionalllr_pmf_q_minusinfinity} has~$m$ entries equal to~$\log 2$ (one of them being in the last position) and all remaining entries equal to~$0.$ We obtain~\eqref{proof_step_after_binomial_series} by using that the sum in~\eqref{proof_step_bionmial_series} is the binomial series expansion of~$1/(1-\becerasureprob)^{m}.$ With similar steps, one can prove that the average number of samples under~$\vlfprocessPygivenx$ that are required for the test~$S(0-,m\log 2)$ to stop is
\begin{IEEEeqnarray}{rCl}
\Bcondexpectation{\vlfprocessPygivenx}{\stoppingtime} &=& \frac{m}{1-\becerasureprob}.
\end{IEEEeqnarray}
Finally, we obtain~\eqref{expression_converse_bec_vlf} by imposing that~$\testpowerP=1-\errorprob$ and~$\testpowerQ=1-1/\messagecount,$ and by solving for the integer~$m$ and the randomization probabilities.
\fi\iflongversion
\section{An Auxiliary Result}\label{appendix_auxiliary_results}
In this appendix, we provide a lemma that is used in the proof of Theorems~\ref{theorem_bec_achievability_iid_no_union_bound}--\ref{theorem_bec_achievability_expurgated_linear}. We consider the evaluation of the average blocklength, averaged over a given ensemble of VLSF codes that operate over a BEC with erasure probability~$\becerasureprob>0.$ The lemma allows us to relate this average blocklength to the one corresponding to the case~$\becerasureprob=0.$
\begin{lem}\label{lemma_reduce_to_zero_erasure}
Consider a BEC with erasure probability~$\becerasureprob>0$ and the ensemble of~$(\latency,\messagecount,0)$--VLSF codes constructed as follows: the stopping time and the decoder are defined as in~\eqref{expression_bec_vlsf_optimal_stopping_time} and~\eqref{expression_bec_vlsf_optimal_decoder}, respectively; the codebook of each code---a matrix with~$\messagecount$ rows and infinitely many columns---has columns independently generated from a given~$\messagecount$ dimensional probability distribution. Let~$\stoppingtime_0$ be the stopping time when~$\becerasureprob=0.$ Then
\begin{IEEEeqnarray}{rCl}
\Bexpectation{\stoppingtime} &=& \frac{\Bexpectation{\stoppingtime_0}}{1-\becerasureprob}\label{expression_reduce_to_zero_erasure}
\end{IEEEeqnarray}
where the expectation is over both channel and code ensemble.
\end{lem}
\begin{IEEEproof}
Let~$\stoppingtime$ be the random variable corresponding to the length of a sequence of output symbols for which the decoder stops. Let~$\stoppingtime_0$ be the number of unerased symbols in the sequence. The expected value of~$\stoppingtime_0$ averaged with respect to both code ensemble and channel law coincide with the average stopping time when~$\becerasureprob=0.$ Let now~$\unerasedposition_1,\unerasedposition_2,\dots,\unerasedposition_{\stoppingtime_0}$ be the position of the unerased bits in the sequence of output symbols. Let~$\erasurecount_1=\unerasedposition_1$ and $\erasurecount_n=\unerasedposition_n-\unerasedposition_{n-1},\ n=2,3,\dots,\stoppingtime_0.$ Then
\begin{IEEEeqnarray}{rCl}
\stoppingtime &=& \sum_{n=1}^{\stoppingtime_0} \erasurecount_n.
\end{IEEEeqnarray}
Note that~$\curlybrac{\erasurecount_n}_{n=1}^{\infty}$ are~\iid~$\Bgeometricdist{1-\becerasureprob}$--distributed. Using Wald's identity~\cite[Eq. (84)]{wald1944cumulativesum} we conclude that
\begin{IEEEeqnarray}{rCl}
\Bexpectation{\stoppingtime} &=& \Bexpectation{\stoppingtime_0} \Bexpectation{\bG_1}
\end{IEEEeqnarray}
from which~\eqref{expression_reduce_to_zero_erasure} follows.
\end{IEEEproof}\ignorespaces
\fi\ignorespaces
\section{Proof of Theorem~\ref{theorem_bec_achievability_iid_no_union_bound}}\label{appendix_proof_bec_achievability_iid_no_union_bound}
We consider the VLSF codebook ensemble specified by the set of all binary matrices with~$\messagecount$ rows and infinitely many columns. Furthermore, we assign a probability distribution on this ensemble by assuming each entry in the codebook being~\iid~$\Bbernoullidist{0.5}.$ Using the stopping time~\eqref{expression_bec_vlsf_optimal_stopping_time} and the decoder~\eqref{expression_bec_vlsf_optimal_decoder}, we can now create a VLSF code ensemble. By\iflongversion~Lemma~\ref{lemma_reduce_to_zero_erasure}\else~\cite[Lemma 10]{devassy2016nonasymptoticlong}\fi, we can write the ensemble average blocklength as
\begin{IEEEeqnarray}{rCl}
\Bexpectation{\stoppingtime}&=&\frac{\Bexpectation{\stoppingtime_0}}{1-\becerasureprob}\label{step_reduce_to_zero_erasure_iid_non_union_bound}
\end{IEEEeqnarray}
where $\stoppingtime_0$ is the stopping time when~$\becerasureprob=0.$ Let~$\messagebitsequalevent{w}{\chinp_1^n}$ be the event that the bits~$\chinp_1^n,$ which are distributed~\iid~$\Bbernoullidist{0.5},$ coincide with the first~$n$ bits of the codeword corresponding to message~$\Rinpmessage.$ Without loss of generality, we assume that message~$1$ is transmitted. The ensemble average of~$\stoppingtime_0$ is given by
\begin{IEEEeqnarray}{rCl}
\Bexpectation{\stoppingtime_0} &=&\iflongversion \sum_{n=0}^{\infty} \probof{\stoppingtime_0>n}\\
&=&\fi 1+\sum_{n=1}^{\infty} \probof{\bigcup_{w=2}^\messagecount \messagebitsequalevent{w}{\chinp_1^n}}\\
&=& \sum_{n=0}^{\infty} \parantheses{1-\parantheses{1-2^{-n}}^{\messagecount-1}}\\
&=& 1-\sum_{i=1}^{\messagecount-1}\binom{\messagecount-1}{i}\frac{\parantheses{-1}^i}{2^i-1}.\label{step_apply_non_union_bound_sum}
\end{IEEEeqnarray}
In~\eqref{step_apply_non_union_bound_sum}, we used the binomial theorem and the summation formula for geometric series. Substituting~\eqref{step_apply_non_union_bound_sum} into~\eqref{step_reduce_to_zero_erasure_iid_non_union_bound} we conclude that
\begin{IEEEeqnarray}{rCl}
\Bexpectation{\stoppingtime}&=&\frac{1}{1-\becerasureprob}\parantheses{1-\sum_{i=1}^{\messagecount-1}\binom{\messagecount-1}{i}\frac{\parantheses{-1}^i}{2^i-1}}.
\end{IEEEeqnarray}
Since there exists at least one VLSF code in the ensemble with average blocklength lower than the ensemble average blocklength, we conclude that~\eqref{expression_bec_achievability_iid_no_union_bound} must hold.
\iflongversion
\section{Proof of Theorem~\ref{theorem_bec_achiveability_vlsf_zero_error}}\label{appendix_proof_bec_achiveability_vlsf_zero_error}
We use the same steps as in the proof of Theorem~\ref{theorem_bec_achievability_iid_no_union_bound} except that the codebook ensemble is different. The columns are~\iid and uniformly distributed over the set of~$\messagecount$ dimensional vectors with~$\ceil{\messagecount/2}$ zeros. Using the same notation as in Appendix~\ref{appendix_proof_bec_achievability_iid_no_union_bound}, we have
\begin{IEEEeqnarray}{rCl}
\Bexpectation{\stoppingtime_0} &=& 1+\sum_{n=1}^{\infty} \probof{\bigcup_{w=2}^\messagecount \messagebitsequalevent{w}{\chinp_1^n}}\\
&\leq& \sum_{n=0}^{\infty} \min\parantheses{(\messagecount-1)\invprobofunequalbits^{-n},1}\label{step_apply_union_bound}\\
&=& \floor{m}+1+\frac{\invprobofunequalbits^{m-\floor{m}}}{\invprobofunequalbits-1}.\label{step_apply_geometric}
\end{IEEEeqnarray}
In~\eqref{step_apply_union_bound} we used the truncated union bound,~$\invprobofunequalbits$ is defined in~\eqref{expression_inv_prob},~$m=\log_\invprobofunequalbits\parantheses{\messagecount-1},$ and~\eqref{step_apply_geometric} follows from the summation formula for geometric series. Substituting~\eqref{step_apply_union_bound} into~\eqref{step_reduce_to_zero_erasure_iid_non_union_bound} we obtain~\eqref{expression_bec_achiveability_vlsf_zero_error}.
\fi\iflongversion
\section{Proof of Theorem~\ref{theorem_bec_achievability_expurgated_linear}}\label{appendix_proof_bec_achievability_expurgated_linear}
The proof follows again along the same lines as the proof of Theorem~\ref{theorem_bec_achievability_iid_no_union_bound}. This time, the ensemble contains only linear codes and is obtained as follows. The columns of the generator matrix are independent and uniformly distributed over the set of all nonzero vectors in~$\binaryfinitefield^\messagebitcount.$ Since the code is linear, the decoder stops when the columns of the generator matrix corresponding to unerased positions form a basis for~$\binaryfinitefield^\messagebitcount.$ Then,~$\Bexpectation{\stoppingtime_0}$ can be interpreted as the average number of columns that need to be collected to obtain a basis for~$\binaryfinitefield^\messagebitcount.$ The dimension of the subspace of the first~$n$ columns of the random generator matrix can be modeled as a Markov chain with a single absorbing state (the state corresponding to maximum dimension~$\messagebitcount$). The time to absorption for this Markov chain follows a discrete phase-type distribution~\cite[Ch. 2]{neuts1981matrix}. Its expectation can be shown to be
\begin{IEEEeqnarray}{rCl}
\Bexpectation{\stoppingtime_0}&=& \messagebitcount+\sum_{i=1}^{\messagebitcount-1}\frac{2^{i}-1}{2^k-2^{i}}. \label{expression_avg_latency_to_full_rank}
\end{IEEEeqnarray}
We obtain~\eqref{expression_bec_achievability_expurgated_linear} by substituting~\eqref{expression_avg_latency_to_full_rank} into~\eqref{step_reduce_to_zero_erasure_iid_non_union_bound}.
\fi
\end{appendices}
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Tunneling is a fundamental effect in wave mechanics, which allows for entering
classically inaccessible regions.
While textbooks focus on tunneling through potential barriers, tunneling
processes in nature often take place in the absence of such energetic barriers.
Instead one observes dynamical tunneling \cite{DavHel1981, KesSch2011}
between classically disjoint regions in phase space.
In generic Hamiltonian systems dynamical tunneling usually occurs between
regions of regular and chaotic motion.
For a typical phase space of a mixed regular--chaotic system see
Fig.~\ref{fig:RAT_Intro}(b).
\begin{figure}[b!]
\begin{center}
\includegraphics[]{fig1.eps}
\caption{(color online) (a) Regular-to-chaotic decay rate $\gamma_{0}$
versus $1/h$ for the standard map at $\K=3.4$. The numerically determined
rates (gray dots) are compared to (the sum of incoherent terms of) the
predictions of Eq.~\eqref{eq:GammaPrediction} ([red] triangles) and
Eq.~\eqref{eq:GammaPredictionAlt} ([magenta] squares). (b) Phase space with
regular orbits (lines) and a chaotic orbit (dots) including a 6:2 nonlinear
resonance chain. (c) Like (b) with an integrable approximation ([red] lines) on
top.}
\label{fig:RAT_Intro}
\end{center}
\end{figure}
In particular, while a classical particle cannot traverse from the regular
to the chaotic region, a wave can tunnel from the regular to the chaotic region.
This regular-to-chaotic tunneling process manifests itself impressively in
chaos-assisted tunneling \cite{LinBal1990, BohTomUll1993}.
Until today the importance of regular-to-chaotic tunneling has been
demonstrated in numerous experiments, including optical microcavities
\cite{ShiHarFukHenSasNar2010, ShiHarFukHenSunNar2011,
YanLeeMooLeeKimDaoLeeAn2010, KwaShiMooLeeYanAn2015, YiYuLeeKim2015,
YiYuKim2016}, microwave billiards
\cite{DemGraHeiHofRehRic2000, BaeKetLoeRobVidHoeKuhSto2008, DieGuhGutMisRic2014,
GehLoeShiBaeKetKuhSto2015}, and cold atom systems \cite{Hen2001,
SteOskRai2001}.
A recent success being the experimental verification
\cite{KwaShiMooLeeYanAn2015, GehLoeShiBaeKetKuhSto2015} that tiny nonlinear
resonance chains within the regular region, as shown in
Fig.~\ref{fig:RAT_Intro}(b), indeed drastically enhance tunneling as predicted
in Refs.~\cite{UzeNoiMar1983, Ozo1984, BroSchUll2001, BroSchUll2002}.
Furthermore, regular-to-chaotic tunneling is expected to play an important role
for atoms and molecules in strong fields, as discussed in
Refs.~\cite{ZakDelBuc1998, BucDelZak2002, WimSchEltBuc2006}.
Motivated by these applications regular-to-chaotic tunneling is also a field of
intense theoretical research \cite{ShuIke1995, ShuIke1998, PodNar2003,
PodNar2005, EltSch2005, WimSchEltBuc2006, SheFisGuaReb2006, BaeKetLoeSch2008,
BaeKetLoeRobVidHoeKuhSto2008, ShuIshIke2008, ShuIshIke2009a, ShuIshIke2009b,
BaeKetLoeWieHen2009, BaeKetLoe2010, LoeBaeKetSch2010, MerLoeBaeKetShu2013,
HanShuIke2015, KulWie2016}, which is mainly focused on periodically driven
model systems with one degree of freedom.
Here, a major achievement is the combination of (i) direct
\cite{BaeKetLoeSch2008, BaeKetLoe2010} and (ii) resonance-assisted
\cite{UzeNoiMar1983, Ozo1984, BroSchUll2001, BroSchUll2002} regular-to-chaotic
tunneling in a single prediction \cite{LoeBaeKetSch2010, SchMouUll2011}.
This prediction shows that as function of decreasing
effective Planck's constant $h$ one has two corresponding regimes:
(i) Regular states localize on a single quantizing torus.
In this regime, tunneling is determined by direct transitions from this regular
torus into the chaotic region \cite{BaeKetLoeSch2008, BaeKetLoe2010}
which can be evaluated semiclassically using complex paths
\cite{MerLoeBaeKetShu2013}.
(ii) For even smaller $h$ a regular state, while still mostly concentrated
on the main quantizing torus, acquires resonance-assisted contributions
on further quantizing tori \cite{UzeNoiMar1983, Ozo1984, BroSchUll2001,
BroSchUll2002} located more closely to the border of the regular region,
see Fig.~\ref{fig:Istates}(c) for an illustration.
This resonance-assisted contribution dominates
tunneling to the chaotic region \cite{EltSch2005, LoeBaeKetSch2010,
SchMouUll2011}.
Thus, one observes a resonance-assisted enhancement of regular-to-chaotic
tunneling.
For an example of this enhancement see Fig.~\ref{fig:RAT_Intro}(a).
Note, that for much smaller $h$ there is even a third regime for which
regular states may localize within the resonance chain. This regime is not
considered here.
Despite the above achievements, a semiclassical evaluation of resonance-assisted
tunneling in mixed regular--chaotic systems remains an open problem.
In particular, the state of the art predictions \cite{LoeBaeKetSch2010,
SchMouUll2011} defy a semiclassical evaluation using the techniques developed
for integrable systems~\cite{DeuMou2010, DeuMouSch2013}.
More specifically, so far (i) an integrable approximation of the regular region
which ignores resonance chains is used to predict the magnitude of direct
tunneling transitions from quantizing tori towards the chaotic region
\cite{BaeKetLoeSch2008, BaeKetLoe2010}.
Subsequently, (ii) resonance-assisted contributions are taken into account
by perturbatively solving \cite{BroSchUll2002} an additional pendulum
Hamiltonian which models the relevant resonance chain \cite{LoeBaeKetSch2010,
SchMouUll2011}.
However, only a perturbation-free prediction, based on a single integrable
approximation which includes the relevant resonance chain will allow for a
semiclassical evaluation of resonance-assisted regular-to-chaotic tunneling in
the spirit of Refs.~\cite{DeuMou2010, DeuMouSch2013}.
In this paper we derive such perturbation-free predictions of resonance-assisted
regular-to-chaotic tunneling.
They are based on a new class of integrable approximations $H_{r:s}$
\cite{KulLoeMerBaeKet2014} which include the dominant $r$:$s$ resonance, see
Fig.~\ref{fig:RAT_Intro}(c).
In particular, the eigenvalue equation
\begin{align}
\label{eq:HrsEigenvalue}
\widehat{H}_{r:s} \ket{m_{\text{\intindex}}} = E_m \ket{m_{\text{\intindex}}},
\end{align}
of such integrable approximations $H_{r:s}$ provides eigenstates $\ket{m_{\text{\intindex}}}$
which model the localization of regular states on the regular phase-space
region, explicitly including the resonance-assisted contributions on multiple
quantizing tori, in a non-perturbative way.
Using such states allows for extending the results of
Refs.~\cite{BaeKetLoeSch2008, BaeKetLoe2010} to the case of resonance-assisted
tunneling.
In particular, the decay rates $\gamma_m$ of metastable states which localize on
the regular phase-space region and decay via regular-to-chaotic tunneling can be
predicted according to
\begin{align}
\label{eq:GammaPrediction}
\gamma_{m} \approx \Gamma_{m}(t=1) := \big\|\widehat{P}_{\mathcal{L}} \widehat{\map} \ket{m_{\text{\intindex}}}
\big\|^2.
\end{align}
Here $\widehat{\map}$ is the time evolution operator and $\widehat{P}_{\mathcal{L}}$ is a projector onto a
leaky region $\mathcal{L}$ located in the chaotic part of phase space.
We further show that regular-to-chaotic decay rates can be predicted with
similar accuracy, when using a simplified formula which no longer contains the
time-evolution operator.
Instead it evaluates only the probability of the state $\ket{m_{\text{\intindex}}}$ on the
leaky region $\mathcal{L}$
\begin{align}
\label{eq:GammaPredictionAlt}
\gamma_{m} \approx \Gamma_{m}(t=0) := \big\|\widehat{P}_{\mathcal{L}} \ket{m_{\text{\intindex}}} \big\|^2.
\end{align}
Both perturbation-free predictions, Eqs.~\eqref{eq:GammaPrediction} and
\eqref{eq:GammaPredictionAlt}, give good results for the standard map,
see Fig.~\ref{fig:RAT_Intro}.
In that both predictions provide the foundation for future semiclassical
predictions of resonance-assisted regular-to-chaotic tunneling
\cite{FriMerLoeBaeKet}.
We remark that the prediction of Eq.~\eqref{eq:GammaPrediction} has previously
been evaluated semiclassically for integrable approximations without resonances
\cite{MerLoeBaeKetShu2013}, using the time-domain techniques of
Refs.~\cite{ShuIke1995, ShuIke1998, ShuIshIke2008, ShuIshIke2009a,
ShuIshIke2009b} giving predictions for direct regular-to-chaotic tunneling.
However, we believe that a future semiclassical prediction of resonance-assisted
regular-to-chaotic tunneling would be easier obtained from
Eq.~\eqref{eq:GammaPredictionAlt}, since it does not involve any time
evolution and thus allows for a semiclassical evaluation using the simpler
WKB-like techniques of Refs.~\cite{DeuMou2010, DeuMouSch2013}.
The paper is organized as follows:\
In Sec.~\ref{Sec:ExampleSystem} we introduce the standard map as a paradigmatic
Hamiltonian example system with a mixed phase space.
We further present regular-to-chaotic decay rates as a measure of
regular-to-chaotic tunneling and discuss their numerical evaluation.
In Sec.~\ref{Sec:RAT} we derive the predictions,
Eqs.~\eqref{eq:GammaPrediction} and \eqref{eq:GammaPredictionAlt}.
In Sec.~\ref{Sec:ResultsStandardMap} we illustrate how these predictions are
evaluated using the example of the standard map.
In Sec.~\ref{Sec:Results} we present our results and compare them to the
perturbative predictions of Refs.~\cite{LoeBaeKetSch2010, SchMouUll2011}.
In Sec.~\ref{Sec:Discussion} we discuss the main approximations and
limitations of our approach.
A summary and outlook is given in Sec.~\ref{Sec:Summary}.
\section{Example System}
\label{Sec:ExampleSystem}
In this paper we focus on periodically driven Hamiltonian systems with one
degree of freedom, which exhibit all generic features of a mixed phase space.
Classically, the stroboscopic map
\begin{align}
\label{eq:StroboscopicMap}
U: (q_n, p_n) \mapsto (q_{n+1}, p_{n+1}),
\end{align}
describes the evolution of positions and momenta, $(q,p)$, in phase space
from time $t=n$ to $t=n+1$ over one period of the external driving.
Quantum-mechanically, the time-evolution is given by the corresponding
unitary time-evolution operator $\widehat{\map}$.
In Sec.~\ref{Sec:StandardMap} we introduce the standard map as a paradigmatic
example of a periodically driven one-degree-of-freedom system with a mixed phase
space.
In Sec.~\ref{Sec:TunnelingRatesInTheStandardMap} we introduce regular-to-chaotic
decay rates $\gamma$, as the central object of our investigation.
Furthermore, we discuss their numerical computation.
Particular attention is paid to nonlinear resonance chains and their quantum
manifestations.
\subsection{Standard Map}
\label{Sec:StandardMap}
Classically, the standard map originates from a periodically kicked Hamiltonian
with one degree of freedom $H(q,p,t) = T(p) + V(q) \sum_{n\in\mathbb{N}}
\delta(n-t)$.
Here, $\delta(\cdot)$ is the Dirac delta function.
For the standard map $T(p)=p^2/2$ and $V(q) = \K/(2\pi)^2\cos{(2\pi q)}$, where
$\K$ is the kicking strength.
Its stroboscopic map $U$ \cite{Chi1979}, Eq.~\eqref{eq:StroboscopicMap}, in
its symmetrized version is given by
\begin{subequations}
\label{eq:SMap}
\begin{align}
q_{n+1} =&\, q_n + p_n + \frac{\K}{4\pi} \sin(2\pi q_n), \\
p_{n+1} =&\, p_n + \frac{\K}{4\pi} \sin(2\pi q_n) + \frac{\K}{4\pi}
\sin(2\pi q_{n+1}),
\end{align}
\end{subequations}
where $(q_n, p_n)$ represents a phase-space point in the middle of the $n$th
kick.
For convenience the standard map is considered on a torus $(q, p) \in [0,1[
\times [-0.5, 0.5[$ with periodic boundary conditions.
In this paper we mainly focus on kicking strength $\K=3.4$.
Here, the phase space exhibits a large regular region which is centered around
an elliptic fixed point, see Fig.~\ref{fig:RAT_Intro}(b).
As expected from the KAM theorem \cite{Kol1954, Arn1963, Arn1963b, Mos1962} the
regular region consists of one-dimensional invariant tori.
Along these tori orbits of regular motion rotate around the fixed point.
These tori are interspersed by nonlinear resonance chains, wherever $s$
rotations of a regular orbit match $r$ periods of the external driving
\cite{Bir1913, LicLie1983a, Chi1979}.
For example, the standard map at $\kappa=3.4$ has a dominant $r$:$s=6$:$2$
resonance, leading to the six regular sub-regions in
Fig.~\ref{fig:RAT_Intro}(b).
Note that we choose the numbers $r$ and $s$ in the ratio $r$:$s$ such that
$r$ is the number of sub-regions of the resonance.
The region of regular motion is embedded in a region of chaotic motion.
The quantum-mechanical analogue of the stroboscopic map $U$ is the unitary
time-evolution operator \cite{BerBalTabVor1979, HanBer1980, ChaShi1986,
KeaMezRob1999, DegGra2003b, Bae2003}
\begin{align}
\label{eq:Qmap}
\hspace*{-0.2cm} \widehat{\map} = \exp\left(-i\frac{V(\hat{q})}{2\hbar}\right)
\exp\left(-i\frac{T(\hat{p})}{\hbar} \right)
\exp\left(-i\frac{V(\hat{q})}{2\hbar}\right).
\end{align}
Here, $h = 2\pi\hbar$ is the effective Planck constant and $\hat{q}$ and $\hat{p}$
are the operators of position and momentum, respectively.
Similar to the classical case we consider $\widehat{\map}$ on a toric phase space, which
leads to grids in position and momentum space \cite{BerBalTabVor1979,
HanBer1980, ChaShi1986, KeaMezRob1999, DegGra2003b, Bae2003}
\begin{subequations}
\begin{align}
\label{eq:qnDef}
\overline{q}_n &= h (n+\theta_p),\quad \text{with}\quad
\overline{q}_{n}\in[0,1[\\
\label{eq:pnDef}
\overline{p}_n &= h (n+\theta_q),\quad \,\text{with}\quad
\overline{p}_{n}\in[-0.5,0.5[,
\end{align}
\end{subequations}
with $n\in \mathbb{N}$.
This implies that the inverse of the effective Planck constant is a natural
number $1/h=N \in \mathbb{N}$, giving the dimension of the Hilbert space.
For the standard map, we choose the Bloch phase $\theta_p=0$, while $\theta_q=0$
if $N$ is even and $\theta_q=0.5$ if $N$ is odd.
This gives the finite-dimensional time-evolution operator in position
representation
\begin{align}
\label{eq:StandardQmapOnTorus}
\braOpket{\overline{q}_n}{\widehat{\map}}{\overline{q}_k} \! &= \! \\ \nonumber
& \hspace*{-1.0cm}\frac{e^{-i\pi/4}}{\sqrt{N}}
\exp\!\left(i \,2\pi N \!
\left[- \frac{V(\overline{q}_n)}{2} +
\frac{(\overline{q}_n-\overline{q}_k)^2}{2}
- \frac{V(\overline{q}_k)}{2}\right]\!\right),
\end{align}
with $n,k = 0, \dots ,N-1$.
In the following it is fundamental that eigenstates of a
mixed regular--chaotic system can be
classified according to their semiclassical localization on the regular or
chaotic region, respectively.
More specifically, chaotic states spread across the chaotic region
\cite{Per1973, Ber1977b, Vor1979}, while regular states localize on a torus
$\tau_{m}$ of the regular region which has quantizing action \cite{Boh1913,
Boh1913b, Som1916}
\begin{align}
\label{eq:BohrSommerfeld}
J_{m} := \frac{1}{2\pi}\oint_{\tau_{m}} p(q)\: \text{d}q
= (m + 1/2) \hbar,
\end{align}
labeled by an index $m\in\mathbb{N}$.
In order to account for resonance-assisted tunneling it is further indispensable
to consider the finer structure of regular states.
In particular, it will be crucial that a regular state $m$ localizes not only on a
dominant quantizing torus $J_m$.
Instead, an $r$:$s$ resonance induces additional contributions on the tori
$J_{m+kr}$ with $k\in\mathbb{Z}$, see Refs.~\cite{UzeNoiMar1983, Ozo1984,
BroSchUll2001, BroSchUll2002, WisSarArrBenBor2011, Wis2014, WisSch2015} and
references therein.
\subsection{Regular-to-Chaotic Decay Rates in the Standard Map}
\label{Sec:TunnelingRatesInTheStandardMap}
In this section we introduce regular-to-chaotic decay rates $\gamma$ of an open
system for quantifying regular-to-chaotic tunneling.
Note that in closed systems chaos-assisted tunnel splittings
\cite{BohTomUll1993} are an often-used alternative \cite{BaeKetLoe2010}.
Our general approach for defining regular-to-chaotic decay rates proceeds in
three steps:\
(a) We introduce a leaky region $\mathcal{L}$ within the chaotic part of phase
space,
(b) we determine the \emph{decay rates} of its regular states, and
(c) we classify the corresponding decay rates as \emph{regular-to-chaotic
decay rates}.
Step (c) is justified because each regular state of the open system decays by
regular-to-chaotic tunneling towards the chaotic region and subsequently
entering the leaky region within the chaotic part of phase space.
More specifically, we proceed by (a) introducing a projector $\widehat{P}_{\mathcal{L}}$ which
absorbs probability on a phase-space region $\mathcal{L}$ within the chaotic
part of phase space.
Based on this projector and the unitary time-evolution operator $\widehat{\map}$ of the
closed system we define the time-evolution operator of the open system as
\begin{align}
\label{eq:UopenNum}
\widehat{\map}_{\text{o}} = (\hat{\mathbf{1}} - \widehat{P}_{\mathcal{L}})\hat{U}(\hat{\mathbf{1}} - \widehat{P}_{\mathcal{L}}).
\end{align}
(b) We solve its eigenvalue equation
\begin{equation}
\label{eq:UopenEigenvalueNum}
\widehat{\map}_{\text{o}}\ket{m} = \exp\left(i\phi_m-\frac{\gamma_m}{2}\right)\ket{m}.
\end{equation}
Here, $\ket{m}$ represents a metastable, right eigenvector of the sub-unitary
operator $\widehat{\map}_{\text{o}}$.
The corresponding eigenvalue is determined by an eigenphase $\phi_m$ and a
decay rate $\gamma_{m}$.
The latter describes the exponential decay of $\ket{m}$ in time.
(c) We assign to each regular state $\ket{m}$ a label $m$ according to its
dominant localization on the quantizing torus $J_m$ and refer to its decay
rate $\gamma_m$ as the regular-to-chaotic decay rate.
Specifically, for the standard map (a) we use
\begin{align}
\label{eq:LeakyRegion}
\!\!\mathcal{L} := \left\{(q,p) \; \left|\right. \;\; q <
q_l\;\;\;\text{or}\;\;\;q > q_r := 1- q_l \right\},
\end{align}
and define the projector
\begin{align}
\label{eq:Pabs}
\widehat{P}_{\mathcal{L}} \ket{q} = \chi(q)\ket{q} \quad \text{with }
\chi(q) = \left\{ \begin{array}{l l}
1 & \text{for } (q, \cdot) \in \mathcal{L}\\
0 & \text{for } (q, \cdot) \notin \mathcal{L}
\end{array}
\right. .
\end{align}
Here, we choose $q_l$ close to the regular--chaotic border.
This ensures that $\gamma_m$, which depends on the choice of the leaky region
$\mathcal{L}$, is dominated by tunneling from the regular towards the chaotic
region.
For a more detailed discussion see
Sec.~\ref{Sec:BeyondRTCTunneling}.
(b) We compute the finite-dimensional matrix representation of $\widehat{\map}_{\text{o}}$ for each
value of $1/h\in\mathbb{N}$.
To this end we set all those entries in Eq.~\eqref{eq:StandardQmapOnTorus}
equal to zero, for which either $\overline{q}_{n}$ or $\overline{q}_{k}$ are in
the leaky region $\mathcal{L}$.
We diagonalize the resulting $\widehat{\map}_{\text{o}}$ numerically.
(c) The regular-to-chaotic decay rates $\gamma_{m}$ are labeled according
to the dominant localization of $\ket{m}$ on the quantizing tori $J_m = \hbar
(m + 1/2)$.
We present the numerically obtained regular-to-chaotic decay rates
$\gamma_{0}$ of the standard map at $\kappa=3.4$ as a function of the inverse
effective Planck constant ([gray] dots) in Fig.~\ref{fig:RAT_Intro}(a).
The numerical results are consistent with the expectations due to
Refs.~\cite{LoeBaeKetSch2010, SchMouUll2011}:\
(i) For $1/ h \lesssim 35$ the state $\ket{0}$ localizes on the torus
$J_0$ such that the direct tunneling from $J_0$ to $\mathcal{L}$
dominates.
In this regime, $\gamma_0$ decreases exponentially for decreasing $h$ which
is a characteristic feature of direct transitions, see Ref.~\cite{HanOttAnt1984,
BaeKetLoeSch2008, BaeKetLoe2010, MerLoeBaeKetShu2013}.
(ii) In the regime $1/ h \gtrsim 35$ tunneling is enhanced by the $6$:$2$
resonance.
For $35 \lesssim 1/h \lesssim 80$ the resonance contribution of the state
$\ket{0}$ on $J_{6}$ is significant such that direct tunneling transition
from $J_{6}$ to $\mathcal{L}$ dominates $\gamma_m$.
This leads to a peak at $1/h=53$, where the state $\ket{0}$ has half its
weight on $J_{6}$.
Finally, for $1/h\gtrsim 80$ the resonance contribution of $\ket{0}$ on
$J_{12}$ is significant such that direct tunneling from $J_{12}$ to
$\mathcal{L}$ dominates the decay rate $\gamma_m$, with a peak at
$1/h=98$.
In Fig.~\ref{fig:Results}(a,c) we show similar numerical rates ([gray] dots)
for the standard map at $\K=2.9$ and $\K=3.5$ with a dominating $10$:$3$ and
$6$:$2$ resonance, respectively.
\section{Perturbation-Free Predictions of Resonance-Assisted Regular-to-chaotic
Tunneling}
\label{Sec:RAT}
In this section we derive the perturbation-free predictions for
resonance-assisted regular-to-chaotic decay rates.
In Sec.~\ref{Sec:DerivationWithTimeEvolution} we derive
Eq.~\eqref{eq:GammaPrediction} which uses the time-evolution operator.
In Sec.~\ref{Sec:DerivationWithoutTimeEvolution} we derive
Eq.~\eqref{eq:GammaPredictionAlt} which does not use the time-evolution
operator.
\subsection{Derivation of Eq.~\eqref{eq:GammaPrediction} with Time Evolution}
\label{Sec:DerivationWithTimeEvolution}
The starting point for deriving Eq.~\eqref{eq:GammaPrediction} is the definition
of the regular-to-chaotic decay rate $\gamma_m$ from the appropriate eigenvalue
problem.
We use the same definitions as for the numerical determination of
regular-to-chaotic decay rates, see Eqs.~\eqref{eq:UopenNum} and
\eqref{eq:UopenEigenvalueNum} of Sec.~\ref{Sec:TunnelingRatesInTheStandardMap}.
They are repeated for convenience, namely a general sub-unitary operator
\begin{align}
\label{eq:Uopen}
\widehat{\map}_{\text{o}} &:= (\hat{\mathbf{1}} - \widehat{P}_{\mathcal{L}})\hat{U}(\hat{\mathbf{1}} - \widehat{P}_{\mathcal{L}}),
\end{align}
and its eigenvalue equation
\begin{align}
\label{eq:UopenEigenvalue}
\widehat{\map}_{\text{o}}\ket{m} &= \exp\left(i\phi_m-\frac{\gamma_m}{2}\right)\ket{m}.
\end{align}
Here, the unitary operator $\widehat{\map}$ describes the time evolution of a mixed
regular--chaotic system over one unit of time.
Furthermore, $\widehat{P}_{\mathcal{L}}$ is a projection operator which absorbs probability on the
leaky region $\mathcal{L}$ within the chaotic part of phase space.
For decay rates of such systems, it can be shown, that the following formula
applies, see App.~\ref{App:Derivation} for details,
\begin{equation}
\label{eq:GammaPredOpen}
\gamma_{m} = - \log\left(1 - \big\|\widehat{P}_{\mathcal{L}} \widehat{\map} \!\ket{m}
\big\|^2\right)\!
\stackrel{\gamma_m \ll 1}{\approx} \!\big\|\widehat{P}_{\mathcal{L}} \widehat{\map} \!\ket{m} \big\|^2,
\end{equation}
i.\,e., a regular-to-chaotic decay rate $\gamma_m$ (for which $\gamma_m\ll1$) is
given by the probability transfer from the regular state $\ket{m}$ into the
leaky region $\mathcal{L}$ via the unitary time-evolution operator $\widehat{\map}$.
Equation~\eqref{eq:GammaPredOpen} is as such not useful, since it still contains
the unknown eigenvector $\ket{m}$.
In particular, it would require to solve Eq.~\eqref{eq:UopenEigenvalue} which defines
$\gamma_m$ in the first place.
Hence, we proceed in the spirit of Refs.~\cite{BaeKetLoeSch2008, BaeKetLoe2010},
i.\,e., we approximate $\ket{m}$ using the eigenstates $\ket{m_{\text{\intindex}}}$ of an
integrable approximation $H_{r:s}$, leading to our prediction
Eq.~\eqref{eq:GammaPrediction}.
The novel point of this paper is the use of an integrable approximation
$H_{r:s}$, which includes the dominant $r$:$s$ resonance.
This ensures that $\ket{m_{\text{\intindex}}}$ models not only the localization of $\ket{m}$
on the main quantizing torus $J_m$ but also accounts for the
resonance-assisted contributions on the tori $J_{m+kr}$.
Precisely this extends Eq.~\eqref{eq:GammaPrediction}, as previously used in
\cite{BaeKetLoeSch2008, BaeKetLoe2010} for direct tunneling, to the regime
of resonance-assisted regular-to-chaotic tunneling in a non-perturbative way.
\subsection{Derivation of Eq.~\eqref{eq:GammaPredictionAlt} without Time Evolution}
\label{Sec:DerivationWithoutTimeEvolution}
In this section we derive Eq.~\eqref{eq:GammaPredictionAlt}.
It predicts regular-to-chaotic decay rates from the localization of the mode
$\ket{m_{\text{\intindex}}}$ on the leaky region $\mathcal{L}$.
In contrast to Eq.~\eqref{eq:GammaPrediction} it does not use the
time-evolution operator.
In that, Eq.~\eqref{eq:GammaPredictionAlt} is an ideal starting point for future
semiclassical predictions of regular-to-chaotic decay rates
\cite{FriMerLoeBaeKet} in the spirit of Refs.~\cite{DeuMou2010, DeuMouSch2013}.
In particular, it avoids the complications which arise in a semiclassical
evaluation of Eq.~\eqref{eq:GammaPrediction} due to the time-evolution operator.
We further remark that predictions like Eq.~\eqref{eq:GammaPredictionAlt} are
common for open systems.
For regular-to-chaotic decay rates they have heuristically been used,
e.\,g. in Refs.~\cite{BroSchUll2002, PodNar2003, PodNar2005, SchMouUll2011}.
Here, the main purpose of deriving Eq.~\eqref{eq:GammaPredictionAlt} is to
explicitly point out the involved approximations.
The derivation starts from an alternative definition of the sub-unitary
time-evolution operator
\begin{align}
\label{eq:UopenAlt}
\widehat{\map}'_{\text{o}} := \widehat{\map} (\hat{\mathbf{1}} - \widehat{P}_{\mathcal{L}}),
\end{align}
which satisfies the eigenvalue equation
\begin{align}
\label{eq:UopenAltEigenvalue}
\widehat{\map}'_{\text{o}}\ket{m'} = \exp\left(i\phi_{m}-\frac{\gamma_{m}}{2}\right)\ket{m'}.
\end{align}
Compare with Eqs.~\eqref{eq:Uopen} and \eqref{eq:UopenEigenvalue}.
As shown in App.~\ref{App:Isospectrality} the operators $\widehat{\map}_{\text{o}}$ and
$\widehat{\map}'_{\text{o}}$ are isospectral.
Therefore, they exhibit the same eigenvalues, which give rise to the same
regular-to-chaotic decay rates $\gamma_m$.
Furthermore, the corresponding normalized right eigenvectors can be transformed
into each other, see App.~\ref{App:Isospectrality}.
We find,
\begin{align}
\label{eq:SameLocalization}
\ket{m} =
\frac{1}{\exp\left(i\phi_{m}-\frac{\gamma_{m}}{2}\right)}(\hat{\mathbf{1}}-\widehat{P}_{\mathcal{L}})\ket{m'},
\end{align}
which implies that $\ket{m}$ and $\ket{m'}$ localize on the quantizing tori
$J_{m+kr}$ of the regular region with equal probability (for $\gamma_m\ll1$).
On the other hand, $\ket{m'}$ is the time-evolved mode $\ket{m}$ according to
\begin{align}
\label{eq:TimeEvolvedMode}
\ket{m'} = \widehat{\map} \ket{m}.
\end{align}
Inserting Eq.~\eqref{eq:TimeEvolvedMode} into Eq.~\eqref{eq:GammaPredOpen} gives
\begin{align}
\label{eq:GammaPredOpenAlt}
\gamma_{m} = - \log\left(1 - \big\|\widehat{P}_{\mathcal{L}} \ket{m'} \big\|^2\right)
\stackrel{\gamma_{m}\ll1}{\approx} \big\|\widehat{P}_{\mathcal{L}} \ket{m'} \big\|^2,
\end{align}
which shows that a regular-to-chaotic decay rate $\gamma_m$ (for which
$\gamma_m\ll1$) is equivalent to the probability to find $\ket{m'}$ on the leaky
region $\mathcal{L}$.
Similar to Eq.~\eqref{eq:GammaPredOpen}, Eq.~\eqref{eq:GammaPredOpenAlt} is as such not
helpful, because it still contains the eigenvector $\ket{m'}$.
In particular, it would require to solve Eq.~\eqref{eq:UopenAltEigenvalue} which
defines $\gamma_m$ in the first place.
Hence, we approximate the mode $\ket{m'}$ using the more accessible eigenstates
$\ket{m_{\text{\intindex}}}$ of an integrable approximation $H_{r:s}$, leading to our
prediction Eq.~\eqref{eq:GammaPredictionAlt}.
Here, the key point is again the use of integrable approximations $H_{r:s}$
which includes the relevant $r$:$s$ resonance.
Therefore, $\ket{m_{\text{\intindex}}}$ models not only the localization of $\ket{m'}$ on
the main quantizing torus $J_m$ but also its resonance-assisted contributions
on the tori $J_{m+kr}$.
Precisely this allows for predicting resonance enhanced regular-to-chaotic
decay rates from Eq.~\eqref{eq:GammaPredictionAlt} in a non-perturbative
way.
An application of the predictions, Eqs.~\eqref{eq:GammaPrediction} and
\eqref{eq:GammaPredictionAlt}, for the standard map is demonstrated in
Sec.~\ref{Sec:ResultsStandardMap}.
The key approximation, i.\,e., modeling metastable regular states $\ket{m}$ (or
$\ket{m'}$) in terms of eigenstates $\ket{m_{\text{\intindex}}}$ of an integrable approximation
$H_{r:s}$, is discussed in Sec.~\ref{Sec:DiscussionStates}.
Moreover, a comparison of the non-perturbative predictions,
Eqs.~\eqref{eq:GammaPrediction} and \eqref{eq:GammaPredictionAlt}, to the
perturbative predictions of Refs.~\cite{LoeBaeKetSch2010,
SchMouUll2011} is given in Sec.~\ref{Sec:ResultsPerturbation}.
\section{Perturbation-free Prediction of Tunneling in the Standard Map}
\label{Sec:ResultsStandardMap}
In this section we illustrate our approach by applying it to the standard map.
In Sec.~\ref{Sec:DominantResonance}, we determine the $r$:$s$ resonance which
dominates tunneling.
In Sec.~\ref{Sec:IntegrableApproximation}, we set up an integrable approximation
including the nonlinear resonance chain using the iterative canonical
transformation method \cite{LoeLoeBaeKet2013, KulLoeMerBaeKet2014} as presented
in Ref.~\cite{KulLoeMerBaeKet2014}.
In Sec.~\ref{Sec:Quantization}, we quantize the integrable approximation and
determine its eigenstates $\ket{m_{\text{\intindex}}}$ from Eq.~\eqref{eq:HrsEigenvalue}.
Finally, the results will be discussed in the next section,
Sec.~\ref{Sec:Results}.
\subsection{Choosing the Relevant Resonance}
\label{Sec:DominantResonance}
In order to apply our prediction it is crucial to first identify the $r$:$s$
resonance which dominates the tunneling process.
A detailed discussion as to which resonance dominates tunneling in which regime,
can be found in Ref.~\cite{LoeBaeKetSch2010}.
Here, we focus on the $r$:$s$ resonance of lowest order~$r$, which
dominates the numerically and experimentally relevant regime where
$\gamma>10^{-15}$.
The area covered by the sub-regions of such a resonance can be very small, see
the inset of Fig.~\ref{fig:Results}(a).
Therefore, it is necessary to search for resonances systematically.
To this end we determine the frequencies of orbits within the regular region, as
described in Ref.~\cite{KulLoeMerBaeKet2014}.
We then identify the $r$:$s$ resonance of lowest order $r$, by searching for
the rational frequencies $2\pi s/r$ with smallest possible denominator.
Specifically, for the standard map parity implies that $r$ has to be an even
number in order to reflect the correct number of subregions forming the
resonance chain.
For the examples we consider in this paper we find a dominant
$10$:$3$ resonance for $\K=2.9$ and a dominant $6$:$2$ resonance
for both $\K=3.4$ and $\K=3.5$.
\subsection{Integrable Approximation of a Regular Region including a Resonance
Chain}
\label{Sec:IntegrableApproximation}
In order to determine an integrable approximation of the regular region which
includes the dominant $r$:$s$ resonance, we use the method introduced in
Ref.~\cite{KulLoeMerBaeKet2014}.
Here, we briefly summarize the key points.\
The integrable approximations $H_{r:s}(q, p)$ of
Ref.~\cite{KulLoeMerBaeKet2014} is generated in two steps.
First the normal-form Hamiltonian is defined as
\begin{subequations} \label{eq:HrsActAng-all}
\begin{align}
\label{eq:HrsActAng}
\hspace*{-0.4cm} \mathcal{H}_{r:s}(\theta, I) &= \mathcal{H}_{0}(I) +
2V_{r:s} \left(\frac{I}{I_{r:s}}\right)^{r/2}\!\!\!\cos(r\theta + \phi_0),\\
\label{eq:H0}
\mathcal{H}_{0}(I) &= \frac{(I-I_{r:s})^{2}}{2 M_{r:s}} +
\sum_{n=3}^{N_{\text{disp}}} h_{n}(I-I_{r:s})^{n}.
\end{align}
\end{subequations}
It contains the essential information on the regular region in the co-rotating
frame of the resonance.
This Hamiltonian is precisely the effective pendulum Hamiltonian used in
Ref.~\cite{LoeBaeKetSch2010, SchMouUll2011}.
Here, $\mathcal{H}_{0}(I)$ is a low order polynomial, chosen such that its
derivative fits the actions and frequencies of the regular region in the
co-rotating frame of the resonance.
The action of the resonant torus is $I_{r:s}$.
The parameters $M_{r:s}$ and $V_{r:s}$ are determined from the size of the
resonance regions in the mixed system as well as the stability of its central
orbit \cite{EltSch2005}.
Finally, $\phi_0$ is used to control the fix-point locations of the resonance
chain.
In a second step, a canonical transformation
\begin{align}
\label{eq:CanTrans}
\mathcal{T}: (\theta, I) \mapsto (q,p)
\end{align}
is used to adapt the tori of the effective pendulum Hamiltonian to the shape of
the regular region in $(q,p)$-space, giving the Hamilton function
\begin{align}
\label{eq:Hrs}
H_{r:s}(q,p) = \mathcal{H}_{r:s}(\mathcal{T}^{-1}(q,p)).
\end{align}
The transformation $\mathcal{T}$ is composed of:\
(i) a harmonic oscillator transformation to the fixed point of the regular
region $\mathcal{T}^{0}$, Eq.~\eqref{eq:CanTransInit}, which provides a rough
integrable approximation and
(ii) a series of canonical near-identity transformations $\mathcal{T}^{1}, ...,
\mathcal{T}^{N_{\mathcal{T}}}$, Eq.~\eqref{eq:CanTransIter}, which improve the
agreement between the shape of tori of the mixed system and the integrable
approximation.
Note that a successful prediction of decay rates requires an integrable
approximation which provides a smooth extrapolation of tori into the chaotic
region \cite{BaeKetLoeSch2008, BaeKetLoe2010}, see insets of
Fig.~\ref{fig:Results}.
This is ensured by using simple near-identity transformations
$\mathcal{T}^{1}, ..., \mathcal{T}^{N_{\mathcal{T}}}$, i.\,e., low orders $N_{q},
N_{p}$ in Eq.~\eqref{eq:CanTransIter}.
For further details the reader is referred to Ref.~\cite{KulLoeMerBaeKet2014}
and Appendix~\ref{App:IntegrableApproximation}, where it is described how the
integrable Hamiltonians for the standard map at $\K=2.9$, $\K=3.4$ and $\K=3.5$,
see insets of Fig.~\ref{fig:Results}, are generated.
\subsection{Quantization of the Integrable Approximation}
\label{Sec:Quantization}
In the following, we summarize the quantization procedure for the integrable
approximation.
The details are discussed in App.~\ref{App:IntegrableApproximationQuantum}.
In its final form, this quantization procedure is almost identical to the
approach presented in Ref.~\cite{LoeBaeKetSch2010}.
It consists of two steps:\
(Q1) The integrable approximation without resonance is used to construct states
which localize along a single quantizing torus of the regular region.
(Q2) The mixing of states, localizing along a single quantizing torus, is
described by solving the quantization of the effective pendulum Hamiltonian,
Eq.~\eqref{eq:HrsActAng-all}, introduced in Ref.~\cite{SchMouUll2011}.
Combining (Q1) and (Q2) gives the sought-after eigenstate $\ket{m_{\text{\intindex}}}$ of the
integrable approximation which includes the resonance.
More specifically:\
(Q1) We use the canonical transformation, Eq.~\eqref{eq:CanTrans}, in order to
define the function $I(q,p)$.
Its contours approximate the tori of the regular phase-space region, ignoring
the resonance chain.
It thus resembles the role of the integrable approximation, previously used in
Refs.~\cite{BaeKetLoeSch2008, BaeKetLoe2010, LoeBaeKetSch2010}.
The Weyl-quantization of this function on a phase-space torus gives a Hermitian
matrix
\begin{align}
\label{eq:WeylI}
&\braOpket{\overline{q}_n}{\hat{I}}{\overline{q}_m} = \frac{1}{2N}
\sum_{l=0}^{2N-1}
\exp\left(\frac{i}{\hbar}(\overline{q}_n\!-\overline{q}_m)\,\overline{p}_{\frac{
l } {2 }}
\right)\times \\
&\quad \quad\left[I\!\left(\frac{\overline{q}_n\!+\overline{q}_m}{2},
\overline{p}_{\frac{l}{2}}\right) + (-1)^l
I\!\left(\frac{\overline{q}_n \! +\overline{q}_m \!+ M_q}{2},
\overline{p}_{\frac{l}{2}}\right)\right]\!. \nonumber
\end{align}
Solving its eigenvalue equation gives states $\braket{\overline{q}_l}{I_n}$
which localize along a single contour of $I(q,p)$ with quantizing action $I_n
= \hbar(n+1/2)$.
These states model the localization of states along the tori of quantizing
action $J_n$ in the mixed system.
For an illustration see Fig.~\ref{fig:Istates}(a,b).
\begin{figure}[tb!]
\begin{center}
\includegraphics[]{fig2.eps}
\caption{(color online)
(a, b) Husimi representation of $\ket{I_n}$ for the standard map at $\K=3.4$
for (a) $n=0$ and (b) $n=6$ at $1/h=53$. Regular tori (gray lines) and
chaotic orbits (dots) illustrate the phase space. The quantizing tori of
$H_{r:s}$ for $V_{r:s}=0$ are shown by a thick (white) line.
(c) Approximate mode $\ket{m_{\text{\intindex}}}$ for $m=0$.}
\label{fig:Istates}
\end{center}
\end{figure}
(Q2) In the second step, we model the mixing of states
$\braket{\overline{q}_l}{I_n}$ due to the nonlinear resonance chain.
To this end, we follow Refs.~\cite{LoeBaeKetSch2010, SchMouUll2011} and
consider the quantization of the effective pendulum Hamiltonian,
Eq.~\eqref{eq:HrsActAng-all}, given by
\begin{align}
\label{eq:HrsActAngIBasis}
\braOpket{I_m}{\widehat{\mathcal{H}}_{r:s}}{I_n} &=
\mathcal{H}_0(I_n)\,\delta_{m,n} +
V_{r:s} \left(\frac{\hbar}{I_{r:s}}\right)^{r/2} \times \\
& \hspace*{-1.5cm} \left(e^{-i \phi_{0}}
\sqrt{\frac{n!}{(n-r)!}}\,\delta_{m,n-r} +
e^{i \phi_{0}}\sqrt{\frac{(n+r)!}{n!}}\,\delta_{m,n+r}\right). \nonumber
\end{align}
Solving this eigenvalue problem gives the sought-after state in the basis of
quantizing actions $\braket{I_n}{m_{\text{\intindex}}}$.
Note, that the matrix in Eq.~\eqref{eq:HrsActAngIBasis} couples basis states
$\ket{I_n}$ and $\ket{I_{n'}}$ only if $|n'-n|=kr$.
Thus, the coefficients $\braket{I_{n}}{m_{\text{\intindex}}}$ are non-zero, only if $n = m +
kr$.
This is called the selection rule of resonance-assisted tunneling.
Combining (Q1) and (Q2) results in the mode expansion
\begin{align}
\label{eq:ModeExpansion}
\braket{\overline{q}_l}{m_{\text{\intindex}}} = \sum_{k}\braket{\overline{q}_l}{I_{m +
kr}} \braket{I_{m+kr}}{m_{\text{\intindex}}}.
\end{align}
For an illustration of a state $\ket{m_{\text{\intindex}}}$ see Fig.~\ref{fig:Istates}(c).
Note that its Husimi-function exhibits exactly the morphology discussed in
Ref.~\cite{Wis2014}.
We now make a couple of remarks:\
(a) We use the above quantization procedure, rather than directly applying the
Weyl-rule to $H_{r:s}(q,p)$, Eq.~\eqref{eq:Hrs}, in order to explicitly enforce
the selection rule of resonance-assisted tunneling.
(b) The ad-hoc two step quantization scheme avoids the problem of defining the
quantum counterpart for the canonical transformations $\mathcal{T}^{1}, ...,
\mathcal{T}^{N_{\mathcal{T}}}$, Eq.~\eqref{eq:CanTransIter}, used in the classical
construction of the integrable approximation, see
App.~\ref{App:IntegrableApproximationQuantum} for details.
(c) The above quantization is almost identical to the procedure used in
Refs.~\cite{LoeBaeKetSch2010, SchMouUll2011}.
This allows for a direct comparison to the results of
Refs.~\cite{LoeBaeKetSch2010, SchMouUll2011}, see
Sec.~\ref{Sec:ResultsPerturbation}.
(d) The quantization procedure cannot determine the relative phase between
the terms in the mode expansion of Eq.~\eqref{eq:ModeExpansion}.
In order to understand the relative phase recall:\
(i) The coefficient vector $\braket{I_{m+kr}}{m_{\text{\intindex}}}$ is determined by solving
the eigenvalue problem of Eq.~\eqref{eq:HrsActAngIBasis}.
Hence, it is determined up to a global phase $\xi_{m}$.
(ii) The coefficient vectors $\braket{\overline{q}_l}{I_{m+kr}}$ are determined
by solving the eigenvalue problem of Eq.~\eqref{eq:WeylI}.
Hence, each coefficient vector is determined up to a global phase
$\varphi_{m+kr}$.
Therefore:
(i) Changing the phase of the coefficient vector $\braket{I_{m+kr}}{m_{\text{\intindex}}}$
in Eq.~\eqref{eq:ModeExpansion} changes the global phase of
$\braket{\overline{q}_l}{m_{\text{\intindex}}}$.
This has no consequences for predicting decay rates.
However, (ii) changing the phases $\varphi_{m+kr}$ of each coefficient vector
$\braket{\overline{q}_l}{I_{m+kr}}$, changes the relative phase of
contributions in Eq.~\eqref{eq:ModeExpansion}.
This changes the interference between the contributions to the sum in
Eq.~\eqref{eq:ModeExpansion} and affects the predicted decay rates.
So far the phase issue was avoided by neglecting interference terms in the
tunneling predictions \cite{LoeBaeKetSch2010, SchMouUll2011}.
For the symmetrized standard map, we propose to define the phases as follows:\
(i) Eq.~\eqref{eq:HrsActAngIBasis} gives a real symmetric matrix.
This allows for choosing real coefficients $\braket{I_{n}}{m_{\text{\intindex}}}$ such that
$\braket{I_{m}}{m_{\text{\intindex}}}>0$.
(ii) Eq.~\eqref{eq:WeylI} also gives a real symmetric matrix.
This allows for choosing real coefficients $\braket{\overline{q}_l}{I_{n}}$.
Choosing the sign of these coefficients is discussed in
App.~\ref{App:IntegrableApproximationQuantum}.
The main idea is to exploit the eigenstates of the harmonic oscillator which
approximates the central fixed point of the regular region.
For these harmonic oscillator states the relative phase is well-defined.
Then we choose the sign of $\braket{\overline{q}_l}{I_{n}}$ such that its
overlap with the corresponding eigenstate of the harmonic oscillator is
positive, Eq.~\eqref{eq:ModeAligning}.
\section{Results}
\label{Sec:Results}
We now apply the above procedure to the standard map at $\K=2.9$, $3.4$, and
$3.5$.
This gives eigenstates $\ket{m_{\text{\intindex}}}$ which we insert into our predictions,
Eqs.~\eqref{eq:GammaPrediction} and \eqref{eq:GammaPredictionAlt}.
The necessary time-evolution operator, used in Eq.~\eqref{eq:GammaPrediction},
is given by Eq.~\eqref{eq:StandardQmapOnTorus}.
The projector is defined by Eq.~\eqref{eq:Pabs} using $q_l=0.27, 0.26, 0.25$,
respectively.
The results are shown in Fig.~\ref{fig:Results}.
\begin{figure}[tb]
\begin{center}
\includegraphics[]{fig3.eps}
\caption{(color online) Decay rates for the standard map at (a) $\K=2.9$,
(b) $\K = 3.4$, and (c) $\K=3.5$ versus the inverse effective Planck constant
$1/h$. Numerically determined rates (dots) are compared to predicted rates,
using Eq.~\eqref{eq:GammaPrediction} ([red] triangles) and
Eq.~\eqref{eq:GammaPredictionAlt} ([magenta] squares). The insets show the
corresponding phase space with regular tori ([gray] lines) and chaotic orbits
(dots) with tori of the integrable approximation ([red] lines).}
\label{fig:Results}
\end{center}
\end{figure}
The numerically determined rates and the
predicted rates are overall in good qualitative agreement.
In both cases they deviate from the exact numerical rates by at most two orders
of magnitude.
In that the accuracy of the perturbation-free predictions,
Eq.~\eqref{eq:GammaPrediction} and Eq.~\eqref{eq:GammaPredictionAlt}, is
equivalent to perturbative predictions from Refs.~\cite{LoeBaeKetSch2010,
SchMouUll2011}.
This establishes Eqs.~\eqref{eq:GammaPrediction} and
\eqref{eq:GammaPredictionAlt} as state of the art perturbation-free predictions
of resonance-assisted regular-to-chaotic tunneling.
See Sec.~\ref{Sec:ResultsPerturbation} for a detailed comparison.
\subsection{Incoherent Predictions and Quantum Phase}
\label{Sec:ResultsIncoherent}
As discussed in Sec.~\ref{Sec:Quantization} our quantization scheme cannot
determine the relative phases between the contributions of
Eq.~\eqref{eq:ModeExpansion} for a system without time-reversal symmetry.
In the following, we discuss the consequences of such an undetermined phase for
the prediction of decay rates.
To this end we summarize our predictions, Eqs.~\eqref{eq:GammaPrediction} and
\eqref{eq:GammaPredictionAlt}, in the following compact form
\begin{align}
\label{eq:GammaPredictionCompact}
\Gamma_{m}(t) := \big\|\widehat{P}_{\mathcal{L}} \widehat{\map}^{t} \ket{m_{\text{\intindex}}} \big\|^2,
\end{align}
where $t=1$ denotes the prediction based on time-evolution and $t=0$ denotes
the prediction without time evolution.
Now we insert the mode expansion, Eq.~\eqref{eq:ModeExpansion}, and average
over the undetermined phases $\varphi_{m+kr}$ of the coefficient vectors
$\braket{\overline{q}_l}{I_{m+kr}}$.
This gives the incoherent prediction
\begin{align}
\label{eq:GammaPredictionCompactIncoherent}
\Gamma_{m}^{\text{inc}}(t) := \sum_{k} \Gamma_{m,m+kr}^{\text{diag}}(t)
\end{align}
where the diagonal term $\Gamma_{m,n}^{\text{diag}}(t)$ is the contribution of the
state $\ket{I_n}$ to the incoherent prediction as
\begin{align}
\label{eq:GammaPredictionDiagonal}
\Gamma_{m,n}^{\text{diag}}(t) := \left|\braket{I_{n}}{m_{\text{\intindex}}}\right|^{2}
\Gamma_{n}^{\text{d}}(t)
\end{align}
and
\begin{align}
\label{eq:GammaPredictionCompactDirekt}
\Gamma_{n}^{\text{d}}(t) := \big\|\widehat{P}_{\mathcal{L}} \widehat{\map}^{t} \ket{I_n} \big\|^2.
\end{align}
is the rate of direct regular-to-chaotic tunneling as previously introduced in
Refs.~\cite{BaeKetLoeSch2008, BaeKetLoe2010}.
The results based on Eq.~\eqref{eq:GammaPredictionCompactIncoherent} are
shown in Fig.~\ref{fig:ResultsIncoherent}.
\begin{figure}[tb]
\begin{center}
\includegraphics[]{fig4.eps}
\caption{(color online) Decay rates for the standard map at (a) $\K=2.9$,
(b) $\K = 3.4$, and $\K=3.5$ versus the inverse effective Planck constant
$1/h$.
Numerically determined rates (dots) are compared to rates, predicted from
incoherent terms according to Eq.~\eqref{eq:GammaPredictionCompactIncoherent},
with $\Gamma_{m}^{\text{inc}}(1)$ ([red] pluses) and $\Gamma_{m}^{\text{inc}}(0)$ ([magenta]
crosses).
We further show predictions according to Eq.~\eqref{eq:GammaPrediction}
([gray] triangles) and Eq.~\eqref{eq:GammaPredictionAlt} ([gray] squares).
\label{fig:ResultsIncoherent}}
\end{center}
\end{figure}
As expected the incoherent predictions,
Eq.~\eqref{eq:GammaPredictionCompactIncoherent}, and the full predictions,
Eq.~\eqref{eq:GammaPredictionCompact}, agree very well in the regime where a
single diagonal contribution dominates, i.\,e., in the regime of direct
tunneling as well as the peak region.
However, in between these regions there are always two diagonal contributions of
similar magnitude, which can interfere.
It is in these regions that we observe clear deviations between the predictions
of Eq.~\eqref{eq:GammaPredictionCompact} and the incoherent predictions of
Eq.~\eqref{eq:GammaPredictionCompactIncoherent}.
In particular, for $\K=3.4$ and $\K=3.5$ the prediction of
Eq.~\eqref{eq:GammaPredictionCompact} predicts destructive interference, while
the incoherent results describe the numerical rates much better.
These results highlight the relevance of the phase factor $\varphi_{m+kr}$ for
obtaining an accurate description of decay rates even between the
resonance-assisted tunneling peaks.
In previous studies of resonance-assisted tunneling in systems with a mixed
phase space \cite{LoeBaeKetSch2010} this phase factor has been ignored by
directly employing the incoherent predictions.
Hence, a satisfactory theoretical treatment of the phase factor $\varphi_{m+kr}$
does so far not exist.
Clearly, our current approach is also insufficient.
The precise reason is not clear to us.
We expect that exploiting the symmetry of the integrable approximation in order
to find a real representation of the approximate mode
$\braket{\overline{q}_l}{m_{\text{\intindex}}}$ is too naive.
In particular, because it is used for approximating the metastable state
$\braket{\overline{q}_l}{m}$ of the open standard map, which can never admit an
entirely real representation.
For a detailed discussion of this point see Sec.~\ref{Sec:ErrorAnalysis}.
Another possibility is that the phase factor in a non-integrable system is
beyond an integrable approximation.
\subsection{Perturbative Predictions}
\label{Sec:ResultsPerturbation}
In this section, we compare our results to the perturbative
predictions of Refs.~\cite{LoeBaeKetSch2010, SchMouUll2011}.
This perturbative prediction is obtained by approximating the coefficient
$\braket{I_{m+kr}}{m_{\text{\intindex}}}$ in the incoherent prediction
Eq.~\eqref{eq:GammaPredictionCompactIncoherent} by solving
Eq.~\eqref{eq:HrsActAngIBasis} perturbatively, \cite{SchMouUll2011},
\begin{align}
\label{eq:PerturbationCoefficients}
\braket{I_{m+kr}}{m_{\text{\intindex}}} \approx \mathcal{A}_{m,m+kr}^{(r:s)} := \prod_{l=1}^{k}
\frac{\braOpket{I_{m+lr}}{\widehat{\mathcal{H}}_{r:s}}{I_{m+(l-1)r}}}{\mathcal{H
} _ { 0 } (I_m) - \mathcal{H}_{0}(I_{m+kr})},
\end{align}
Note that $\mathcal{H}_{0}(I)$ is considered in the co-rotating frame.
This leads to
\begin{align}
\label{eq:GammaPredictionCompactPerturbation}
\Gamma_{m}^{\text{per}}(t) := \sum_{k} \left|\mathcal{A}_{m,m+kr}^{(r:s)}\right|^{2}
\Gamma_{m}^{\text{d}}(t).
\end{align}
A slight difference of the above expression as compared to
Ref.~\cite{LoeBaeKetSch2010, SchMouUll2011} is the use of the projector $\widehat{P}_{\mathcal{L}}$
rather than a projector on the whole chaotic region.
Thus our prediction eliminates a free parameter from the perturbative
predictions of Refs.~\cite{LoeBaeKetSch2010, SchMouUll2011}.
The results of the perturbative predictions are presented in
Fig.~\ref{fig:ResultsPerturbation}.
\begin{figure}[tb]
\begin{center}
\includegraphics[]{fig5.eps}
\caption{(color online) Decay rates for the standard map at (a) $\K=2.9$,
(b) $\K = 3.4$, and (c) $\K=3.5$ versus the inverse effective Planck constant
$1/h$.
Numerically determined rates (dots) are compared to rates, predicted
perturbatively according to Eq.~\eqref{eq:GammaPredictionCompactPerturbation}
with $\Gamma_{m}^{\text{per}}(1)$ ([red] pluses) and $\Gamma_{m}^{\text{per}}(0)$ ([magenta]
crosses).
We further show the prediction based on incoherent terms, according to
Eq.~\eqref{eq:GammaPredictionCompactIncoherent} with $\Gamma_{m}^{\text{inc}}(1)$
([gray] pluses) and $\Gamma_{m}^{\text{inc}}(0)$ ([gray] crosses).}
\label{fig:ResultsPerturbation}
\end{center}
\end{figure}
They agree with the prediction obtained from
Eq.~\eqref{eq:GammaPredictionCompactIncoherent}, with the slight difference
that the perturbative results deviate around the peak region.
We conclude this section with a short list of advantages and disadvantages of
the perturbation-free and perturbative predictions:\
(i) The perturbation-free framework, Eqs.~\eqref{eq:GammaPrediction} and
\eqref{eq:GammaPredictionAlt}, as well as their incoherent version,
Eq.~\eqref{eq:GammaPredictionCompactIncoherent}, predict numerical rates with
similar accuracy as the perturbative framework of Refs.~\cite{LoeBaeKetSch2010,
SchMouUll2011}.
(ii) One advantage of the perturbative prediction is the possibility to evaluate
the terms $\braket{I_{m}}{m_{\text{\intindex}}}$ analytically,
Eq.~\eqref{eq:PerturbationCoefficients}.
Yet, for practical use even the perturbative approach requires an integrable
approximation for predicting the direct rates $\Gamma_{m}^{\text{d}}$.
Hence, both predictions are equally challenging in their implementation.
(iii) Another advantage of the perturbative prediction is the possibility to
include multiple resonances into
Eq.~\eqref{eq:GammaPredictionCompactPerturbation}, which is not yet
possible for the perturbation-free predictions presented in this paper.
Note that this restriction is not too severe, because decay rates in the
experimentally and numerically accessible regimes ($\gamma>10^{-15}$) are
typically affected by a single resonance only.
Nevertheless, an extension of the perturbation-free results to the
multi-resonance regime is of theoretical interest and requires normal-form
Hamiltonians $\mathcal{H}_{r:s}$ which include multiple resonances.
(iv) The main advantage of the perturbation-free framework is that it provides
the foundation for deriving a future semiclassical prediction of
resonance-assisted regular-to-chaotic tunneling \cite{FriMerLoeBaeKet}.
\section{Discussion}
\label{Sec:Discussion}
In this section, we discuss several aspects of our results in detail.
In Sec.~\ref{Sec:BeyondRTCTunneling} we discuss the dependence of decay rates on
the choice of the leaky region.
In Sec.~\ref{Sec:DiscussionStates} we compare the metastable states $\ket{m}$
and $\ket{m'}$ to the eigenstate $\ket{m_{\text{\intindex}}}$ of an integrable approximation.
In Sec.~\ref{Sec:ErrorAnalysis} we analyze the approximation of $\ket{m}$ and
$\ket{m'}$ via $\ket{m_{\text{\intindex}}}$ more systematically.
In Sec.~\ref{Sec:PeakPosition} we comment on the predictability of peaks.
\subsection{Dependence of Decay Rates on the Leaky Region}
\label{Sec:BeyondRTCTunneling}
\begin{figure}[tb]
\begin{center}
\includegraphics[]{fig6.eps}
\caption{(color online) Numerically determined decay rates $\gamma_0$ of the
standard map at $\K=3.4$ versus the inverse effective Planck constant $1/h$
for $q_l=0.26$ ([gray] dots) and $q_l=0.1$ ([magenta] squares). (b, c) Phase
space with shaded areas showing the leaky regions corresponding to
$q_l$.
}
\label{fig:DiscussionLeakyRegionK34}
\end{center}
\end{figure}
\begin{figure}[tb]
\begin{center}
\includegraphics[]{fig7.eps}
\caption{(color online) Same as Fig.~\ref{fig:DiscussionLeakyRegionK34} for
$\K=2.9$ with $q_l=0.27$ ([gray] dots) and $q_l=0.1$ ([magenta] squares)}
\label{fig:DiscussionLeakyRegionK29}
\end{center}
\end{figure}
This paper focuses entirely on situations where the leaky region $\mathcal{L}$
is chosen close to the regular--chaotic border region.
However, in generic Hamiltonian systems like the standard map, the chaotic
region is interspersed with partial barriers \cite{KayMeiPer1984a,
KayMeiPer1984b}.
This leads to sticky motion in a hierarchical region surrounding the regular
region.
Furthermore, the chaotic component might be inhomogeneous and exhibit slow
classical transport.
In view of these classical phenomena, it is not surprising that the numerical
decay rates of the standard map, defined via Eqs.~\eqref{eq:UopenEigenvalue},
depend on the choice of the leaky region via the parameter $q_l$.
In order to illustrate this phenomenon, we show the numerically determined decay
rate $\gamma_{0}$ of the standard map for two choices of the leaky region
and two different $\K$ parameters in Figs.~\ref{fig:DiscussionLeakyRegionK34}
and \ref{fig:DiscussionLeakyRegionK29}, respectively.
In Fig.~\ref{fig:DiscussionLeakyRegionK34} we show results for the standard map
at $\K=3.4$.
Here, we compare (i) the regular-to-chaotic decay rates obtained for
$q_l=0.26$ (parameter used in this paper, [gray] dots) to (ii) decay rates
obtained for $q_l=0.1$ ([magenta] squares).
While the decay rates for $q_l=0.26$ exhibit a rather smooth behavior the decay
rates for $q_l=0.1$ clearly exhibit additional oscillations and some overall
suppression.
An even stronger deviation between regular-to-chaotic decay rates with
varying leaky regions is observed in Fig.~\ref{fig:DiscussionLeakyRegionK29}
for the standard map at $\K=2.9$.
Here, (i) the decay rates as obtained for $q_l=0.27$ (parameter used in this
paper, [gray] dots) are compared to (ii) the decay rates obtained for $q_l=0.1$
([magenta] squares).
In addition to oscillations, the decay rates for $q_l=0.1$ exhibit a clear
suppression of their average value.
The origin of these deviations is unclear.
The suppression of decay rates for leaky regions far from the regular--chaotic
border could be due to slow transport through an inhomogeneous chaotic region
from the regular--chaotic border towards the leaky region.
So far a quantitative prediction of decay rates with leaky regions far from the
regular--chaotic border remains an open problem.
While varying the leaky region $\mathcal{L}$ close to the regular--chaotic
border can be accounted for by our approach, predicting decay rates with leaky
region far from the regular--chaotic border is beyond our framework.
In particular, while we observe that the numerical decay rates stabilize when
pushing the leaky region away from the regular--chaotic border, the predicted
rates continue to decrease exponentially.
So far the best approach for dealing with this problem is to use an effective
prediction \cite{LoeBaeKetSch2010, SchMouUll2011}.
To this end one argues that the numerical decay rate would not change much upon
pushing the boundary of the leaky region $\mathcal{L}$ beyond some effectively
enlarged regular region $\mathcal{R}_{\text{eff}}$.
See Ref.~\cite{SchMouUll2011} for a discussion of $\mathcal{R}_{\text{eff}}$.
Successively one would approximate the projector onto the leaky region
$\mathcal{L}$ in our predictions by the projector onto the complement of the
effectively enlarged regular region $\mathcal{L}_{\text{eff}}$.
This would result in an effective prediction $\Gamma_{m}^{\text{eff}}$.
Yet, there are several problems with such effective predictions:\
(a) Even though there are semiclassical arguments to define the effectively
enlarged regular region in terms of partial barriers \cite{SchMouUll2011},
replacing the leaky region $\mathcal{L}$ with some effective region
$\mathcal{L}_{\text{eff}}$ introduces an effective parameter to the prediction.
(b) Throughout this paper, we used leaky regions $\mathcal{L}$ which were
almost tangential to the effectively enlarged regular regions discussed in
Ref.~\cite{SchMouUll2011}.
Hence, replacing the region $\mathcal{L}$ with the effective region
$\mathcal{L}_{\text{eff}}$ would not give results which are too far away from
the predictions discussed in this paper, i.\,e., even the effective predictions
$\Gamma_{m}^{\text{eff}}$ clearly deviates from numerically determined decay
rates with leaky regions far from the regular region.
(c) Even when using $\mathcal{L}_{\text{eff}}$ as a free fit parameter the
effective prediction $\Gamma_{m}^{\text{eff}}$ can at most capture the average
behavior of numerical decay rates with leaky region far from the
regular--chaotic border.
In particular, the oscillations observed for the numerical rates in
Figs.~\ref{fig:DiscussionLeakyRegionK34} and \ref{fig:DiscussionLeakyRegionK29}
which span up to four orders of magnitude cannot be accounted for even by an
effective theory.
Note that, accurately predicting decay rates based on
Eqs.~\eqref{eq:GammaPrediction} and \eqref{eq:GammaPredictionAlt}, even for
leaky regions far from the regular region, requires modes $\ket{m_{\text{\intindex}}}$ which
model the localization of $\ket{m}$ and $\ket{m'}$ even in the chaotic region.
We expect that this is beyond the framework of an integrable approximation.
\subsection{Metastable States and Integrable Eigenstates}
\label{Sec:DiscussionStates}
We now discuss the key approximation of our predictions.
To this end we compare the metastable states $\ket{m}$ and $\ket{m'}$
to the corresponding approximate state $\ket{m_{\text{\intindex}}}$,
which originates from an integrable approximation $H_{r:s}$ including
the relevant resonance.
We focus on a typical example using the states $m = m' = m_{\text{\intindex}} = 0$ of the
standard map at $\K=3.4$ with $1/h=55$ close to the first resonance peak in
Fig.~\ref{fig:RAT_Intro}(c).
The absolute squared values of the states in position representation are
shown in Fig.~\ref{fig:DiscussionStates}.
Here we compare (a) $\ket{m}$ to $\ket{m_{\text{\intindex}}}$, (b) $\ket{m'}$ to $\ket{m_{\text{\intindex}}}$,
and (c) $\ket{m'} = \widehat{\map}\ket{m}$ to $\widehat{\map}\ket{m_{\text{\intindex}}}$, depicting them by
(gray) dots and (magenta) squares, respectively.
\begin{figure}[tb]
\begin{center}
\includegraphics[]{fig8.eps}
\caption{(color online) For the standard map at $\K=3.4$ with $1/h = 55$
we compare the position representation of
(a) $|\langle q|m\rangle|^2$ to $|\langle q|m_{\text{\intindex}}\rangle|^2$,
(b) $|\langle q|m'\rangle|^2$ to $|\langle q|m_{\text{\intindex}}\rangle|^2$, and
(c) $|\langle q|\widehat{\map}|m\rangle|^2$ to $|\langle q|\widehat{\map}|m_{\text{\intindex}}\rangle|^2$,
depicting them by (gray) dots and (magenta) squares, respectively, for $m = m' =
m_{\text{\intindex}} = 0$. The dashed lines mark the positions $q_l$ and $1-q_l$ of the leaky
region, as given in the text.}
\label{fig:DiscussionStates}
\end{center}
\end{figure}
As a first conclusion we see that the metastable states are well approximated by
their integrable partners within the non-leaky region, i.\,e., the region
between the dashed lines in Figs.~\ref{fig:DiscussionStates}(a-c).
In particular, both the metastable states and their integrable approximations
exhibit the generic structure which is determined by the regular region and the
dominant $6$:$2$ resonance \cite{BroSchUll2001, BroSchUll2002}:\
(i) A main Gaussian-like hump at $q=0.5$ marks the main localization of the
modes on the torus $J_0$.
(ii) The decrease of the hump is interrupted at two side humps, which correspond
to the resonance-assisted contribution of each mode on the torus $J_6$.
From there, the Gaussian-like exponential decrease continues towards the leaky
region, which is outside the dashed lines in
Figs.~\ref{fig:DiscussionStates}(a-c).
As a second conclusion from Fig.~\ref{fig:DiscussionStates} we infer that beyond
the regular--chaotic border, i.\,e., within the leaky region the metastable
states deviate from their integrable counter parts.
Here, the integrable states continue to decrease exponentially.
In contrast, the state $\ket{m}$ vanishes, see Fig.~\ref{fig:DiscussionStates}(a),
while the state $\ket{m'}=\widehat{\map}\ket{m}$, Eq.~\eqref{eq:TimeEvolvedMode},
does not decrease much slower, see Fig.~\ref{fig:DiscussionStates}(b,c).
Finally, we emphasize that $\ket{m'}$ and $\ket{m_{\text{\intindex}}}$ agree for positions
close to the regular--chaotic border.
Furthermore, these contributions dominate the probability of $\ket{m'}$ and
$\ket{m_{\text{\intindex}}}$ on the leaky region.
Precisely this ensures that replacing $\ket{m'}$ in the exact prediction,
Eq.~\eqref{eq:GammaPredOpenAlt}, by $\ket{m_{\text{\intindex}}}$ results in a meaningful
prediction according to Eq.~\eqref{eq:GammaPredictionAlt}.
An analogous argument explains why replacing $\widehat{\map}\ket{m}$ in the exact result,
Eq.~\eqref{eq:GammaPredOpen}, by $\widehat{\map}\ket{m_{\text{\intindex}}}$ gives meaningful predictions
according to Eq.~\eqref{eq:GammaPrediction}.
\subsection{Error Analysis}
\label{Sec:ErrorAnalysis}
In this section, we investigate the approximation of the metastable states
$\ket{m}$ in the exact result Eq.~\eqref{eq:GammaPredOpen} via the mode
$\ket{m_{\text{\intindex}}}$ in Eq.~\eqref{eq:GammaPrediction} from the perspective of
Eq.~\eqref{eq:ModeExpansion}, i.\,e.,
(i) we investigate the basis states $\ket{I_n}$ and (ii) the expansion
coefficients $\braket{I_n}{m_{\text{\intindex}}}$.
We focus on the standard map at $\kappa=3.4$.
(i) In order to investigate our basis set $\ket{I_n}$, we expand the metastable
state $\ket{m}$ in this basis and insert this expansion into the exact
result~\eqref{eq:GammaPredOpen}.
This gives
\begin{align}
\label{eq:GammaPredOpenExpanded}
\gamma_{m} &= \sum_{n} \left|\braket{I_n}{m}\right|^{2} \big\|\widehat{P}_{\mathcal{L}} \widehat{\map}
\ket{I_n} \big\|^2 \\ \nonumber
&+ \sum_{n,n'} \braket{m}{I_{n'}}
\braOpket{I_{n'}}{\widehat{\map}^{\dagger}\widehat{P}_{\mathcal{L}}^2\widehat{\map}}{I_{n}}\braket{I_n}{m}.
\end{align}
Since the diagonal terms
\begin{align}
\label{eq:GammaOpenDiag}
\gamma_{m,n}^{\text{diag}} := \left|\braket{I_n}{m}\right|^{2}\big\|\widehat{P}_{\mathcal{L}} \widehat{\map}
\ket{I_n} \big\|^2 = \left|\braket{I_n}{m}\right|^{2}
\Gamma_{n}^{\text{d}}(1)\end{align}
provide a bound to the off-diagonal terms according to Cauchy's inequality
\begin{align}
\left|\braket{m}{I_{n'}}
\braOpket{I_{n'}}{\widehat{\map}^{\dagger}\widehat{P}_{\mathcal{L}}^2\widehat{\map}}{I_{n}}\braket{I_n}{m}\right| \le
\sqrt{\gamma_{m,n}^{\text{diag}} \gamma_{m,n'}^{\text{diag}}}
\end{align}
we can interpret them as a way to quantify the contribution of the $n$th basis
state $\ket{I_n}$ to the decay rate $\gamma_{m}$.
In that $\gamma_{m,n}^{\text{diag}}$ takes a similar role as the contribution
spectrum,
discussed in Ref.~\cite{HanShuIke2015}.
\begin{figure}[tb!]
\begin{center}
\includegraphics[]{fig9.eps}
\caption{(color online) Error analysis for the standard map at $\K=3.4$.
(a,b,c) The numerically determined rates ([gray] dots) and (d) the numerically
determined phases $\text{Arg}(\braket{I_n}{0})$ for $n=0,1,2$ (lines) are shown
versus the inverse effective Planck constant $1/h$.
(a) The contributions $\gamma_{0,n}^{\text{diag}}$, Eq.~\eqref{eq:GammaOpenDiag}, are
shown by lines.
(b) The reduced prediction, Eq.~\eqref{eq:GammaPredOpenExpanded} with
$n,n'\in\{0,6,12\}$ is shown by (red) triangles.
(c) The contributions $\gamma_{0,n}^{\text{diag}}$ of Eq.~\eqref{eq:GammaOpenDiag}
(lines), are compared to $\Gamma_{0,n}^{\text{diag}}(1)$ of
Eq.~\eqref{eq:GammaPredictionDiagonal} (markers) for $n=0,6,12$.
(d) The phases $\text{Arg}(\braket{I_n}{0})$ (lines) and
$\text{Arg}(\braket{I_n}{0_{\text{int}}})$ (markers) are compared for $n=0,6,12$.
[Phases close to zero are slightly shifted for $n=6,12$ for visibility.]
}
\label{fig:ErrorAnalysis}
\end{center}
\end{figure}
In Fig.~\ref{fig:ErrorAnalysis}(a), we consider all contributions
$\gamma_{0,n}^{\text{diag}}$ (lines) in comparison with the
decay rate $\gamma_{0}$ (dots) for the standard map at $\K=3.4$.
While most contributions are two to three orders of magnitude smaller than
$\gamma_{0}$, we find that the contributions $\gamma_{0,0}^{\text{diag}}$,
$\gamma_{0,6}^{\text{diag}}$, and $\gamma_{0,12}^{\text{diag}}$ dominate.
In order to further test whether the modes $\ket{I_{n}}$ with $n=0,6,12$ are
sufficient for describing $\gamma_{0}$ we sum the contributions $n,n'\in
\{0,6,12\}$ of the dominant terms in Eq.~\eqref{eq:GammaPredOpenExpanded}.
This gives the red curve of Fig.~\ref{fig:ErrorAnalysis}(b).
From this numerical observations we conclude that a reasonable description of
$\gamma_{0}$ can be extracted using an approximate mode exclusively
composed of states $\ket{I_n}$ with $n=0,6,12,...$, as used in
Eq.~\eqref{eq:ModeExpansion}.
However, it should be noted that the difference between $\gamma_{0}$ and its
reduced version, based on contributions $n,n'\in \{0,6,12\}$ in
Eq.~\eqref{eq:GammaPredOpenExpanded}, is already of the order of $\gamma_{0}$
itself.
See the region $70<1/h<100$ of Fig.~\ref{fig:ErrorAnalysis}(b) in
particular.
Hence, reducing the metastable state $\ket{m}$ to an approximate mode
$\ket{m_{\text{\intindex}}}$ using only basis states $\ket{I_{m+kr}}$ as in
Eq.~\eqref{eq:ModeExpansion} can at best provide a reasonable backbone for
describing the structure of $\gamma_{0}$.
On the other hand, for our example a prediction of $\gamma_{0}$ where the
remainder is smaller than the decay rate based on a reduced set of basis states
$\ket{I_n}$ is only possible when summing over many additional contributions,
even including $n\neq m+kr$.
The precise origin of such contributions $\gamma_{m,n}^{\text{diag}}$ with
$n\neq m+kr$ is currently under debate \cite{HanShuIke2015}:\
From the framework of resonance-assisted tunneling~\cite{BroSchUll2001,
BroSchUll2002, LoeBaeKetSch2010, SchMouUll2011}, we expect that the overlap
$\braket{I_n}{m}$ vanishes for $n\neq m+kr$.
Hence, one might argue that the contributions $\gamma_{m,n}^{\text{diag}}$ with $n\neq
m+kr$ arise in our example only because our basis $\ket{I_n}$ is insufficiently
accurate to decompose $\ket{m}$ according to the theoretical expectation of
resonance-assisted tunneling.
On the other hand, the authors of of Ref.~\cite{HanShuIke2015} observe
non-vanishing contributions $\braket{I_n}{m}$ also for $n\neq m+kr$ even for a
near-integrable situation, where an excellent integrable approximations exist.
They argue that non-vanishing $\braket{I_n}{m}$ should always occur and claim
their treatment is beyond the current framework of resonance-assisted tunneling.
Independent of the origin of the non-zero contributions
$\gamma_{m,n}^{\text{diag}}$ for $n\neq m+kr$, their theoretical description
is beyond the scope of this paper.
In our examples the irrelevance of these contributions is ensured by choosing
leaky regions close to the regular--chaotic border.
However, for leaky regions far from the regular--chaotic border the
contributions $\gamma_{m,n}^{\text{diag}}$ with $n\neq m+kr$ become relevant.
(ii) In the next step we evaluate the errors introduced by replacing the
expansion coefficients $\braket{I_{m+kr}}{m}$ by $\braket{I_{m+kr}}{m_{\text{\intindex}}}$ in
Eq.~\eqref{eq:GammaPredOpenExpanded}.
We focus on the corresponding diagonal contributions $\gamma_{m,m+kr}^{\text{diag}}$
and $\Gamma_{m,m+kr}^{\text{diag}}$, which represent the squared norm of the expansion
coefficients $\braket{I_{m+kr}}{m}$ and $\braket{I_{m+kr}}{m_{\text{\intindex}}}$ up to a
multiplication by the direct rate $\Gamma_{m+kr}^{\text{d}}(1)$.
See lines and symbols in Fig.~\ref{fig:ErrorAnalysis}(c), respectively.
From this data we conclude that the norm of $\braket{I_{m+kr}}{m_{\text{\intindex}}}$ provides
a reasonable approximations for the norm of the expansion coefficients
$\braket{I_{m+kr}}{m}$.
The deviations before each peak could be due to neglecting the higher order
action dependencies discussed in Ref.~\cite{SchMouUll2011} in the Hamilton
function of Eq.~\eqref{eq:HrsActAng-all}.
Furthermore, we expect that the slightly broader peaks in the numerical rates
$\gamma_{0,kr}^{\text{diag}}$ as compared to the sharper peaks of
$\Gamma_{0,kr}^{\text{diag}}(1)$ observed for the integrable approximation, are
related to the openness of the mixed system.
Finally, in Fig.~\ref{fig:ErrorAnalysis}(d) we compare the phases
$\text{Arg}(\braket{I_{m+kr}}{m})$ and $\text{Arg}(\braket{I_{m+kr}}{m_{\text{\intindex}}})$, for $m=0$ and
$k=0,1,2$, respectively.
Here, $\text{Arg}(\cdot)\in(-\pi,\pi]$ is the principal value of the complex argument
function.
Note that the global phase of $\ket{m}$ is fixed by setting
$\text{Arg}(\braket{I_{m}}{m})=0$.
The phases $\text{Arg}(\braket{I_{m+kr}}{m_{\text{\intindex}}})$ are fixed as described in
App.~\ref{App:IntegrableApproximationQuantum}.
While the phases of $\text{Arg}(\braket{I_{m+kr}}{m_{\text{\intindex}}})$ jump from $\pi$ to zero
upon traversing the peak for decreasing $1/h$ (change from destructive to
constructive interference) their counterparts for $\text{Arg}(\braket{I_{m+kr}}{m})$
seem to follow this jump in a smoothed out way.
Compare symbols and lines in Fig.~\ref{fig:ErrorAnalysis}(d).
We attribute this phase detuning to the openness of the system, i.\,e.:\
(a) The symmetries of the integrable approximation allow for choosing a real
representation of the coefficient $\braket{I_{m+kr}}{m_{\text{\intindex}}}$.
Its phase can thus only take values
$\text{Arg}(\braket{I_{m+kr}}{m_{\text{\intindex}}})\in\{0,\pi\}$.
In contrast (b) the mode $\ket{m}$ originates from an open system and thus the
coefficient $\braket{I_{m+kr}}{m}$ are usually complex such that
$\text{Arg}(\braket{I_{m+kr}}{m})$ might take any value.
While the deviation between the numerically determined phases $\text{Arg}( \braket{
I_{m+kr}}{m})$ and the theoretically predicted phases $\text{Arg}(\braket{ I_{m+kr}}{
m_{\text{\intindex}}})$ are seemingly small in Fig.~\ref{fig:ErrorAnalysis}(d), their deviation
has huge effects on the predicted decay rate, i.\,e.:\
(a) Eq.~\eqref{eq:GammaPrediction} predicts destructive interference of the
diagonal terms in the region before each peak.
This leads to strong deviations from the numerical decay rate, see
Fig.~\ref{fig:Results}(b).
On the other hand, (b) already the minimal detuning of
$\text{Arg}(\braket{I_{m+kr}}{m})$ from our prediction
$\text{Arg}(\braket{I_{m+kr}}{m_{\text{\intindex}}})$ is sufficient to lift the destructive
interference.
We assume that this explains why the incoherent prediction,
Eq.~\eqref{eq:GammaPredictionCompactIncoherent}, as illustrated in
Fig.~\ref{fig:ResultsIncoherent}, describe the numerical rates
much better than predictions according to, Eq.~\eqref{eq:GammaPrediction}, see
Fig.~\ref{fig:ResultsIncoherent}.
\subsection{Predictability of Peak Positions}
\label{Sec:PeakPosition}
Finally, we discuss the predictability of peak positions.
To this end we recall that $\mathcal{H}_{0}(I)$ in Eq.~\eqref{eq:H0}, is
determined by fitting its derivative to the numerically determined actions and
frequencies $(\bar{\omega}, \bar{J})$ of the regular phase-space region in
the co-rotating frame.
For an illustration see Fig.~\ref{fig:dispersion}.
\begin{figure}[tb!]
\begin{center}
\includegraphics[]{fig10.eps}
\caption{(color online) For the standard map at $\K=3.4$ we show (a) the fit
of $\mathcal{H}'_{0}(I)$ (line) to the actions and frequencies of the regular
region $(\bar{J},\bar{\omega})$ (crosses).
(b) The function $\mathcal{H}_{0}(I)$ is shown as a (red) line. The two (black)
dots show $(I_{m}, \mathcal{H}_{0}(I_{m}))$ and $(I_{m+kr},
\mathcal{H}_{0}(I_{m+kr}))$ at $1/h=53$.
(a, b) The dotted line shows the position of $\bar{J}_{\text{max}}$.
}
\label{fig:dispersion}
\end{center}
\end{figure}
In particular, the data of the mixed system has a maximal action
$\bar{J}_{\text{max}}$, see (gray) dotted line in Fig.~\ref{fig:dispersion}.
Hence, $\mathcal{H}_{0}$ can be well controlled in the regular region
$I<\bar{J}_{\text{max}}$.
However, for $I>\bar{J}_{\text{max}}$ the function $\mathcal{H}_{0}$ is only an
extrapolation to the chaotic region.
Furthermore, the integrable approximation predicts a peak for $\gamma_{m}$
\cite{BroSchUll2001, BroSchUll2002, LoeBaeKetSch2010, SchMouUll2011}, if
\begin{align}
\mathcal{H}_{0}(I_m) = \mathcal{H}_{0}(I_{m+kr}),
\end{align}
where $I_m=\hbar(m+1/2)$ and $I_{m+kr}=\hbar(m+kr+1/2)$.
This resonance conditions follows from Eq.~\eqref{eq:PerturbationCoefficients}.
However, for all examples presented in this paper the resonant torus
$I_{m+kr}$ is always located outside of the regular region, where
$\mathcal{H}_{0}(I)$ is only given by an extrapolation.
See Fig.~\ref{fig:Istates}(c) for an example of this situation.
The (black) dots in Fig.~\ref{fig:dispersion}(b) show the corresponding
situation for $\mathcal{H}_{0}(I)$.
In such a situation our approach cannot guarantee an accurate prediction of the
peak position.
Usually, this problem is not too severe and the extrapolation is good enough.
An example where this problem appears can be seen in the second peak of
Fig.~\ref{fig:Results}(a) where the peak of the numerical decay rates and the
predicted rates is shifted by $1/h=1$.
\section{Summary and Outlook}
\label{Sec:Summary}
In this paper we present two perturbation-free predictions of resonance-assisted
regular-to-chaotic decay rates, Eqs.~\eqref{eq:GammaPrediction} and
\eqref{eq:GammaPredictionAlt}.
Both predictions are based on eigenstates $\ket{m_{\text{\intindex}}}$ of an integrable
approximation $H_{r:s}$, Eq.~\eqref{eq:HrsEigenvalue}.
The key point is the use of an integrable approximation $H_{r:s}$ of the mixed
regular--chaotic system which includes the relevant nonlinear resonance chain.
Therefore $\ket{m_{\text{\intindex}}}$ models the localization of regular modes on the regular
region, including resonance-assisted contributions in a non-perturbative way.
This allows for extending the validity of Eq.~\eqref{eq:GammaPrediction},
previously used for direct tunneling in Refs.~\cite{BaeKetLoeSch2008,
BaeKetLoe2010}, to the regime of resonance-assisted tunneling.
Furthermore, we introduce a second prediction,
Eq.~\eqref{eq:GammaPredictionAlt}, which no longer requires the time-evolution
operator.
Instead it allows for predicting decay rates using the localization of the
approximate mode on the leaky region.
In that Eq.~\eqref{eq:GammaPredictionAlt} provides an excellent foundation for a
future semiclassical prediction of resonance-assisted regular-to-chaotic
decay rates \cite{FriMerLoeBaeKet} in the spirit of Refs.~\cite{DeuMou2010,
DeuMouSch2013}.
The validity of the presented approach is verified for the standard map, where
predicted and numerically determined regular-to-chaotic decay rates show
good agreement.
Finally, we list future challenges:\
(a) The presented approach is so far limited to periodically driven systems
with one degree of freedom.
An extension to autonomous or periodically driven systems with two or more
degrees of freedom is an interesting open problem.
(b) The perturbation-free approach applies to the experimentally and
numerically relevant regime, where a single resonance dominates
regular-to-chaotic tunneling.
Its extension to the semiclassical regime where multiple-resonances affect
tunneling is of theoretical interest.
(c) The suppression of decay rates due to partial barriers is so far treated by
choosing leaky regions close to the regular--chaotic region.
Explicitly predicting the additional suppression of decay rates due to slow chaotic
transport through inhomogeneous chaotic regions remains an open question.
\begin{acknowledgments}
We gratefully acknowledge fruitful discussions with
J{\'e}r{\'e}my Le Deunff,
Felix Fritzsch,
Yasutaka Hanada,
Hiromitsu Harada,
Kensuke Ikeda,
Martin K{\"o}rber,
Steffen L\"ock,
Amaury Mouchet,
Peter Schlagheck,
and
Akira Shudo.
We acknowledge support by the Deutsche Forschungsgemeinschaft (DFG) Grant No.\
BA 1973/4-1.
N.M.\ acknowledges successive support by JSPS (Japan) Grant No.\ PE 14701 and
Deutsche Forschungsgemeinschaft (DFG) Grant No.\ ME 4587/1-1.
\end{acknowledgments}
\begin{appendix}
\section{Derivation of Eq.~\eqref{eq:GammaPredOpen}}
\label{App:Derivation}
In this appendix we derive Eq.~\eqref{eq:GammaPredOpen} starting from
Eqs.~\eqref{eq:Uopen} and \eqref{eq:UopenEigenvalue}.
Taking the norm of the eigenvalue equation~\eqref{eq:UopenEigenvalue}
for a normalized state $\ket{m}$ one finds
\begin{align}
\exp{\left(-\gamma_{m}\right)}
&= \big\|\widehat{\map}_{\text{o}}\ket{m}\big\|^{2}
= \left\langle m \right| \widehat{\map}_{\text{o}}^{\dagger}\widehat{\map}_{\text{o}} \ket{m} \\
&\hspace*{-1.7cm}
= \left\langle m \right| (\hat{\mathbf{1}}-\widehat{P}_{\mathcal{L}})^{\dagger}
\widehat{\map}^{\dagger}(\hat{\mathbf{1}}-\widehat{P}_{\mathcal{L}})^{\dagger}(\hat{\mathbf{1}}-\widehat{P}_{\mathcal{L}})\widehat{\map}
(\hat{\mathbf{1}}-\widehat{P}_{\mathcal{L}})\ket{m}, \nonumber
\end{align}
where in the last step the definition of $\widehat{\map}_{\text{o}}$, Eq.~\eqref{eq:Uopen}, is used.
We simplify this expression using
\begin{align}
\label{eq:ProjectiveInvariance}
(\hat{\mathbf{1}}-\widehat{P}_{\mathcal{L}})\ket{m} = \ket{m},
\end{align}
which follows from Eqs.~\eqref{eq:Uopen} and \eqref{eq:UopenEigenvalue}, giving
\begin{align}
\exp{\left(-\gamma_{m}\right)}
&= \left\langle m \right|
\widehat{\map}^{\dagger}(\hat{\mathbf{1}}-\widehat{P}_{\mathcal{L}})^{\dagger}(\hat{\mathbf{1}}-\widehat{P}_{\mathcal{L}})\widehat{\map} \ket{m}.
\end{align}
Finally, exploiting the idempotence and hermiticity of the projector $\widehat{P}_{\mathcal{L}}$ gives
\begin{align}
\exp{\left(-\gamma_{m}\right)}
&= \left\langle m \right| \widehat{\map}^{\dagger}\widehat{\map}\ket{m}
- \left\langle m \right| \widehat{\map}^{\dagger} \widehat{P}_{\mathcal{L}} \widehat{\map} \ket{m} \nonumber\\
&= 1 - \|\widehat{P}_{\mathcal{L}}\widehat{\map}\ket{m}\|^{2},
\end{align}
where in the last step the unitarity of $\widehat{\map}$ is used. From
this follows the expression for regular-to-chaotic tunneling
rates, Eq.~\eqref{eq:GammaPredOpen}.
\section{Isospectrality}
\label{App:Isospectrality}
In this appendix, we demonstrate the isospectrality of the sub-unitary operators
$\widehat{\map}_{\text{o}}$ and $\widehat{\map}'_{\text{o}}$ as defined by Eqs.~\eqref{eq:Uopen} and
\eqref{eq:UopenAlt}, respectively.
Furthermore, we discuss the transformation relating their eigenmodes.
For convenience, we repeat the corresponding eigenvalue equations
\eqref{eq:UopenEigenvalue} and \eqref{eq:UopenAltEigenvalue}
\begin{align}
\label{eq:UopenEigenvalueAppendix}
\widehat{\map}_{\text{o}} \ket{m} &= \lambda_m \ket{m}, \\
\label{eq:UopenAltEigenvalueAppendix}
\widehat{\map}'_{\text{o}} \ket{m'} &= \lambda_{m}' \ket{m'},
\end{align}
where the eigenvalues have been denoted by $\lambda_m$ and
$\lambda_m'$.
We now demonstrate the isospectrality of $\widehat{\map}_{\text{o}}$ and $\widehat{\map}'_{\text{o}}$.
To this end we show:\
(a) For each eigenstate $\ket{m}$ of $\widehat{\map}_{\text{o}}$ with eigenvalue
$\lambda_m$, $\widehat{\map}\ket{m}$ is an eigenstate of $\widehat{\map}'_{\text{o}}$ with the same eigenvalue
$\lambda_m$
\begin{align}
\widehat{\map}'_{\text{o}}\widehat{\map}\ket{m}
&\stackrel{\eqref{eq:UopenAlt}}{=}\widehat{\map}(\hat{\mathbf{1}}-\widehat{P}_{\mathcal{L}})\widehat{\map}
\ket{m} \\
&\stackrel{\eqref{eq:ProjectiveInvariance}}{=} \widehat{\map}(\hat{\mathbf{1}}-\widehat{P}_{\mathcal{L}})\widehat{\map}
(\hat{\mathbf{1}}-\widehat{P}_{\mathcal{L}})
\ket{m} \\
&\stackrel{\eqref{eq:Uopen}}{=} \widehat{\map}\widehat{\map}_{\text{o}}\ket{m} \\
&\stackrel{\eqref{eq:UopenEigenvalueAppendix}}{=} \lambda_m \widehat{\map}
\ket{m}.
\end{align}
This further shows that the normalized eigenmode $\ket{m}$ of $\widehat{\map}_{\text{o}}$ with
eigenvalue $\lambda_m$ gives a normalized eigenmode $\ket{m'}$ of $\widehat{\map}'_{\text{o}}$
with eigenvalue $\lambda_m$ according to Eq.~\eqref{eq:TimeEvolvedMode}.
(b) For each eigenstate $\ket{m'}$ of $\widehat{\map}'_{\text{o}}$ with eigenvalue
$\lambda_m'$, the state $(\hat{\mathbf{1}}-\widehat{P}_{\mathcal{L}})\ket{m'}$ is an eigenstate of $\widehat{\map}_{\text{o}}$
with
the same eigenvalue $\lambda_m'$
\begin{align}
\widehat{\map}_{\text{o}} (\hat{\mathbf{1}}-\widehat{P}_{\mathcal{L}})\ket{m'}
&\stackrel{\eqref{eq:Uopen}}{=}(\hat{\mathbf{1}}-\widehat{P}_{\mathcal{L}})\widehat{\map}
(\hat{\mathbf{1}}-\widehat{P}_{\mathcal{L}})^2 \ket{m'} \\
&\;= (\hat{\mathbf{1}}-\widehat{P}_{\mathcal{L}})\widehat{\map} (\hat{\mathbf{1}}-\widehat{P}_{\mathcal{L}})\ket{m'} \\
&\stackrel{\eqref{eq:UopenAlt}}{=} (\hat{\mathbf{1}}-\widehat{P}_{\mathcal{L}})\widehat{\map}'_{\text{o}}\ket{m'} \\
&\stackrel{\eqref{eq:UopenAltEigenvalueAppendix}}{=} \lambda_m' (\hat{\mathbf{1}}-\widehat{P}_{\mathcal{L}})
\ket{m'}.
\end{align}
This further shows that for non-zero eigenvalue $\lambda_m'$ the normalized
eigenmode $\ket{m'}$ of $\widehat{\map}'_{\text{o}}$ gives a normalized eigenmode $\ket{m}$ of
$\widehat{\map}_{\text{o}}$ according to Eq.~\eqref{eq:SameLocalization}.
\section{Details of the Integrable Approximation}
\label{App:IntegrableApproximation}
In this appendix we summarize some technical aspects on the integrable
approximation.
Computational details of the classical integrable approximation as well as
slight changes as compared to Ref.~\cite{KulLoeMerBaeKet2014} are given in
Sec.~\ref{App:IntegrableApproximationClassical}.
Details of the quantization are discussed in
Sec.~\ref{App:IntegrableApproximationQuantum}.
\subsection{Details of the Classical Integrable Approximation}
\label{App:IntegrableApproximationClassical}
We now summarize the modifications of the algorithm described in
\cite{KulLoeMerBaeKet2014} in order to account for the symmetries of our system.
Then we give a list of relevant computational parameters.
\subsubsection{Symmetrization}
In agreement with Ref.~\cite{KulLoeMerBaeKet2014} the canonical transformation
$\mathcal{T}$, Eq.~\eqref{eq:CanTrans}, is composed of (i) an initial canonical
transformation
\begin{align}
\label{eq:CanTransInit}
\mathcal{T}^{0}: (\theta, I) \mapsto (Q,P)
\end{align}
which provides a rough integrable approximation of the regular phase-space
region and (ii) a series of canonical near-identity transformations
\begin{align}
\label{eq:CanTransIter}
\mathcal{T}' \equiv \mathcal{T}^{N_{\mathcal{T}}}\circ...\circ\mathcal{T}^{1}: (Q, P)
\mapsto (q,p),
\end{align}
which improve the agreement between the shape of the tori of the mixed system
and the integrable approximation.
In contrast to Ref.~\cite{KulLoeMerBaeKet2014} we use the symmetrized standard
map in this paper.
In order to account for this symmetry, we specify the canonical transformation,
Eq.~\eqref{eq:CanTransInit}, as
\begin{align}
\label{eq:CanTransInitHO}
\mathcal{T}^{0}:
\begin{pmatrix}
\theta \\ I
\end{pmatrix}
\mapsto
\begin{pmatrix}
Q \\ P
\end{pmatrix}
=
\begin{pmatrix}
q^{\star} + \sqrt{2I/\sigma}\cos(\theta) \\
p^{\star} - \sqrt{2I\sigma}\sin(\theta)
\end{pmatrix}
\end{align}
Here, $(q^{\star}, p^{\star}) = (0.5, 0.0)$ are the coordinates of the central
fixed point in the standard map.
The parameter $\sigma$ is determined from the stability matrix of the standard
map at $(q^{\star}, p^{\star})$
\begin{align}
\mathcal{M} =
\begin{pmatrix}
1-\frac{\K}{2} & 1 \\
- \K (1 - \frac{\K}{4}) & 1-\frac{\K}{2}
\end{pmatrix}
\end{align}
as \cite{LoeLoeBaeKet2013}
\begin{align}
\sigma^{2} =
\frac{\left|1+\frac{\K}{2}\right|-\left|1-\frac{\K}{2}\right|}{
\left|1+\frac{\K}{2}\right|+\left|1-\frac{\K}{2}\right|}.
\end{align}
Furthermore, the symmetry of the systems is imposed on the transformations
$\mathcal{T}^{1}, ..., \mathcal{T}^{N_{\mathcal{T}}}$, Eq.~\eqref{eq:CanTransIter},
by specifying their generating function as
\begin{align}
\label{eq:CanTransFamily}
F^{a}(q,p') &=\\ \nonumber
qp' + &\sum_{n=1}^{N_{q}}\sum_{m=1}^{N_{p}} a_{m,n}
\sin(2\pi n[q-q^{\star}]) \sin(2\pi m[p'-p^{\star}]),
\end{align}
rather than using the more general form of
Ref.~\cite[Eq.~(31)]{KulLoeMerBaeKet2014}.
\subsubsection{Algorithmic Overview}
(Ai) We determine the parameters $I_{r:s}$, $M_{r:s}$, $V_{r:s}$, $\phi_{0}$,
Eq.~\eqref{eq:HrsActAng-all} as described in Ref.~\cite{EltSch2005}.
(Aii) We determine $\mathcal{H}_{0}(I)$, Eq.~\eqref{eq:H0}, by fitting it to
$N_{\text{disp}}$ tuples of action and frequency $(\bar{J},\bar{\omega})$
describing the tori of the regular region in the co-rotating frame of the
resonance.
(Aiii) We determine the near-identity transformations of
Eq.~\eqref{eq:CanTransIter}.
Initially, this requires sampling of the regular region using $N_{\text{ang}}$
points along $N_{\text{tori}}$ tori.
The invertibility of the near-identity transformations in a certain
phase-space region is ensured by rescaling the coefficients $a_{m,n}\mapsto
\eta a_{m,n}$ in Eq.~\eqref{eq:CanTransFamily} using a damping factor $\eta$.
If $N_{q}$, $N_{p}$ in Eq.~\eqref{eq:CanTransFamily} are too large, the tori of
the integrable approximation form curls and tendrils in the chaotic region.
In that case the integrable approximation cannot predict decay rates.
We control this problem by choosing the largest possible parameters $N_{q}$,
$N_{p}$ for which the tori of the integrable approximation provide a smooth
extrapolation into the chaotic phase-space region.
After a finite amount of steps $N_{\mathcal{T}}$, the canonical transformations
do not improve the agreement between the regular region and the integrable
approximation.
At this point we terminate the algorithm.
\subsubsection{Computational Parameters}
\label{App:IntegrableApproximationListOfParameters}
In the following we list the important parameters of the integrable
approximation.
For $\K=2.9$ we use $I_{r:s}=0.009223$, $M_{r:s}=0.06243$, $V_{r:s}=-1.655 \cdot
10^{-7}$, and $\phi_{0}=\pi$.
For $\mathcal{H}_{0}$ in Eq.~\eqref{eq:H0} we used $N_{\text{disp}}=4$ and
fit its derivative to $N_{\text{disp}}=120$ tori of noble frequency.
We use $N_{\mathcal{T}}=40$ near-identity transformations,
Eq.~\eqref{eq:CanTransIter}, generated from Eq.~\eqref{eq:CanTransFamily} with
$N_{q}=N_{p}=2$ and coefficients rescaled by $\eta=0.05$.
The regular region was sampled using $N_{\text{ang}}=200$ points along
$N_{\text{tori}}=120$ tori, equidistantly distributed in action.
For $\K=3.4$ we use $I_{r:s}=0.01026$, $M_{r:s}=-0.047$, $V_{r:s}=-1.612 \cdot
10^{-5}$, and $\phi_{0}=0$.
For $\mathcal{H}_{0}$ in Eq.~\eqref{eq:H0} we use $N_{\text{disp}}=6$ and fit
its derivative to $N_{\text{disp}}=120$ tori, equidistantly distributed in
action.
We use $N_{\mathcal{T}}=15$ near-identity transformations,
Eq.~\eqref{eq:CanTransIter}, generated from Eq.~\eqref{eq:CanTransFamily} with
$N_{q}=N_{p}=2$ and coefficients rescaled by $\eta=0.25$.
The regular region is sampled using $N_{\text{ang}}=300$ points along
$N_{\text{tori}}=120$ tori, equidistantly distributed in action.
For $\K=3.5$ we use $I_{r:s}=0.01244$, $M_{r:s}=-0.048$, $V_{r:s}=-2.98 \cdot
10^{-5}$, and $\phi_{0}=0$.
For $\mathcal{H}_{0}$ in Eq.~\eqref{eq:H0} we use $N_{\text{disp}}=4$ and fit
its derivate to $N_{\text{disp}}=120$ tori, equidistantly distributed in
action.
We use $N_{\mathcal{T}}=15$ near-identity transformations,
Eq.~\eqref{eq:CanTransIter}, generated from Eq.~\eqref{eq:CanTransFamily} with
$N_{q}=N_{p}=2$ and coefficients rescaled by $\eta=0.25$.
The regular region is sampled using $N_{\text{ang}}=300$ points along
$N_{\text{tori}}=120$ tori, equidistantly distributed in action.
\subsubsection{Robustness}
After fixing all parameters as described above the final integrable
approximation might differ, depending on the sampling of the regular region.
In order to show that this does not affect the final prediction, we evaluate
Eq.~\eqref{eq:GammaPrediction}, for three integrable approximations which are
based on slightly different sets of sample points.
The result is illustrated in Fig.~\ref{fig:Robustness}.
\begin{figure}[tb!]
\begin{center}
\includegraphics[]{fig11.eps}
\caption{(color online) Decay rates $\gamma_{0}$ for the standard map at
$\K=3.4$ versus the inverse effective Planck constant.
Numerically determined rates ([gray] circles) are compared to predicted rates
according to Eq.~\eqref{eq:GammaPrediction} ([colored] symbols) based on three
slightly different integrable approximations.
}
\label{fig:Robustness}
\end{center}
\end{figure}
It shows that the prediction is clearly robust.
\subsection{Derivation of Quantization}
\label{App:IntegrableApproximationQuantum}
In the following we sketch the basic ideas leading to the quantization
procedure presented in Sec.~\ref{Sec:Quantization}.
To this end we first present the quantization of the Hamilton-function obtained
after the transformation $\mathcal{T}^{0}$, Eqs.~\eqref{eq:CanTransInit} and
\eqref{eq:CanTransInitHO}, in Sec.~\ref{App:IntegrableApproximationQuantumT0}.
In Sec.~\ref{App:IntegrableApproximationQuantumT} we present how we extend
these results to the full transformation $\mathcal{T}$.
\subsubsection{Quantization after $\mathcal{T}^{0}$}
\label{App:IntegrableApproximationQuantumT0}
To quantize the Hamilton-function $H_{r:s}^{(0)}(Q,P)$
obtained after the canonical transformation $\mathcal{T}^{0}$,
Eqs.~\eqref{eq:CanTransInit} and \eqref{eq:CanTransInitHO},
we follow Ref.~\cite{SchMouUll2011} by starting with the
transformed Hamilton-function
\begin{align}
H_{r:s}^{(0)}(Q,P) &= \mathcal{H}_{0}\left(\frac{Q^2 + P^2}{2}\right) +
\frac{V_{r:s}}{\left(2I_{r:s}\right)^{r/2}} \times \\ \nonumber
& \left[\exp(i\phi_0) \left(\sigma^{1/2}[Q-q^{\star}] -
i\frac{P}{\sigma^{1/2}}\right)^r\right. \\ \nonumber
&\left.+ \exp(-i\phi_0) \left(\sigma^{1/2}[Q-q^{\star}]
+ i\frac{P}{\sigma^{1/2}}\right)^r \right].
\end{align}
In order to quantize this function we replace the coordinates $(Q,P)$ by
operators
\begin{subequations}
\begin{align}
Q \mapsto \widehat{Q}\\
P \mapsto \widehat{P}
\end{align}
\end{subequations}
and demand the usual commutation relation
\begin{align}
\label{eq:CommutatrQP}
[\widehat{Q}, \widehat{P}] = i\hbar.
\end{align}
This allows for introducing the corresponding ladder operators as
\begin{subequations}
\label{eq:LadderoperatorQP}
\begin{align}
\widehat{a} &:=
\frac{1}{(2\hbar)^{1/2}}\left(\sigma^{1/2}[\widehat{Q}-q^{\star}] +
i\frac{\widehat{P}}{\sigma^{1/2}}\right)\\
\widehat{a}^{\dagger} &:=
\frac{1}{(2\hbar)^{1/2}}\left(\sigma^{1/2}[\widehat{Q}-q^{\star}] -
i\frac{\widehat{P}}{\sigma^{1/2}}\right)
\end{align}
\end{subequations}
which admit the commutator
\begin{align}
\label{eq:CommutatorLadderQP}
[\widehat{a}, \widehat{a}^{\dagger}] = 1,
\end{align}
such that we get the number operator
\begin{align}
\label{eq:numberOpQP}
\widehat{n} := \widehat{a}^{\dagger}\widehat{a}.
\end{align}
Based on these operators the quantization of $H_{r:s}^{(0)}$ takes the form
\cite{SchMouUll2011}
\begin{align}
\label{eq:HrsQP}
\widehat{H}_{r:s}^{(0)} &=
\mathcal{H}_{0}(\widehat{I}) \\ \nonumber
&+ V_{r:s}\left(\frac{\hbar}{I_{r:s}}\right)^{\frac{r}{2}}
\left[\widehat{a}^{\dagger^{r}}\exp(i\phi_0) +
\widehat{a}^{r}\exp(-i\phi_0)\right],
\end{align}
where
\begin{align}
\widehat{I} := \hbar(\widehat{n} + 1/2)
\end{align}
is the operator replacing the unperturbed action $I$.
Finally, in order to define the basis states, we identify them with the
eigenstates of the number operator leading to
\begin{align}
\label{eq:eigenvalueNumberOpQP}
\widehat{I}\Ket{I_n^{(0)}} &= I_n \Ket{I_n^{(0)}},
\end{align}
where the eigenvalues become quantizing actions $I_n=\hbar(n+1/2)$ and the basis
states $\Ket{I_n^{(0)}}$ fulfill
\begin{subequations}
\label{eq:basisstatesQP}
\begin{align}
\widehat{a}\Ket{I_0^{(0)}} &= 0\\
\Ket{I_n^{(0)}} &= \frac{1}{\sqrt{n!}}\widehat{a}^{\dagger^{n}}\Ket{I_0^{(0)}}.
\end{align}
\end{subequations}
With respect to this position basis $\Ket{I_n^{(0)}}$ become the eigenstates of
the harmonic oscillator
\begin{align}
\label{eq:basisstatesQP_Qrepr}
\!\!\!\BraKet{Q}{I_n^{(0)}} =
\left(\frac{\sigma}{\pi\hbar}\right)^{\frac{1}{4}} \!\!\!\!\frac{1}{\sqrt{2^n
n!}} \;\text{H}_{n}\!\left(\!\!\frac{Q}{\sqrt{\hbar/\sigma}}\!\!\right)
\exp{\!\left(\!\!-\frac{\sigma Q^2}{2\hbar}\right)},
\end{align}
where $\text{H}_{n}(\cdot)$ are the Hermite polynomials.
\subsubsection{Quantization after $\mathcal{T}$}
\label{App:IntegrableApproximationQuantumT}
Our final goal is of course to obtain the quantization of the Hamilton-function
$H_{r:s}(q,p)$ which is related to $H_{r:s}^{(0)}(Q,P)$ via the canonical
transformation $\mathcal{T}'$, Eq.~\eqref{eq:CanTransIter}.
In order to obtain its quantization we assume that $\mathcal{T}'$
quantum-mechanically corresponds to a unitary operator
$\widehat{U}_{\mathcal{T}'}$ which has the following properties:
\begin{subequations}
\label{eq:UT}
\begin{align}
\widehat{U}_{\mathcal{T}'}^{-1} &= \widehat{U}_{\mathcal{T}'^{-1}}\\
\widehat{Q}' &=
\widehat{U}_{\mathcal{T}'}\widehat{Q}\widehat{U}_{\mathcal{T}'}^{-1}\\
\widehat{P}' &=
\widehat{U}_{\mathcal{T}'}\widehat{P}\widehat{U}_{\mathcal{T}'}^{-1}
\end{align}
\end{subequations}
Such an operator exists at least within a semiclassical approximation
\cite{Bog1992}.
Note that $\widehat{Q}', \widehat{P}'$ represent the operators $\widehat{Q},
\widehat{P}$ within the final coordinate frame $(q,p)$.
However, they must not be confused with the operators $\widehat{q}, \widehat{p}$
which give rise to the position and momentum basis in the final coordinate
frame $(q,p)$.
In particular, while $\widehat{q}\ket{q}=q\ket{q}$, $\widehat{Q}\ket{q}\neq
q\ket{q}$.
Under the above assumption the transformed operators preserve the commutation
relation
\begin{align}
\label{eq:Commutatrqp}
[\widehat{Q}', \widehat{P}'] = i\hbar.
\end{align}
Furthermore, we get the transformed ladder operators as
\begin{subequations}
\label{eq:Ladderoperatorqp}
\begin{align}
\widehat{a}' &:=
\widehat{U}_{\mathcal{T}'}\widehat{a}\widehat{U}_{\mathcal{T}'}^{-1}\\
\widehat{a}'^{\dagger} &:=
\widehat{U}_{\mathcal{T}'}\widehat{a}^{\dagger}\widehat{U}_{\mathcal{T}'}^{-1}
\end{align}
\end{subequations}
which admit the same commutator
\begin{align}
\label{eq:CommutatorLadderqp}
[\widehat{a}', \widehat{a}'^{\dagger}] = 1,
\end{align}
such that we get the transformed number operator
\begin{align}
\label{eq:numberOpqp}
\widehat{n}' =
\widehat{U}_{\mathcal{T}'}\widehat{n}\widehat{U}_{\mathcal{T}'}^{-1},
\end{align}
and the transformed action operator
\begin{align}
\widehat{I}' := \hbar(\widehat{n}' + 1/2).
\end{align}
Based on these operators we can define the transformation of the quantization of
$H_{r:s}^{(0)}(Q,P)$ which we identify with the quantization of
$H_{r:s}(q,p)$.
It takes the form \cite{SchMouUll2011}
\begin{align}
\label{eq:Hrsqp}
\widehat{H}_{r:s} &=
\mathcal{H}_{0}(\widehat{I}') \\ \nonumber
&+ V_{r:s}\left(\frac{\hbar}{I_{r:s}}\right)^{\frac{r}{2}}
\left[\widehat{a}'^{\dagger^{r}}\exp(i\phi_0) +
\widehat{a}'^{r}\exp(-i\phi_0)\right].
\end{align}
Finally, in order to define the basis states $\ket{I_n}$, we identify them
with the eigenstates of the number operator $\widehat{n}'$, such that
\begin{align}
\label{eq:eigenvalueNumberOpqp}
\widehat{I}'\ket{I_n} &= I_n \ket{I_n}
\end{align}
with the basis states $\ket{I_n}$ which admit the property
\begin{subequations}
\label{eq:basisstatesqp}
\begin{align}
\widehat{a}'\ket{I_0} &= 0\\
\ket{I_n} &= \frac{1}{\sqrt{n!}}\widehat{a}'^{\dagger^{n}}\ket{I_0}.
\end{align}
\end{subequations}
Evaluating $\widehat{H}_{r:s}$, Eq.~\eqref{eq:Hrsqp} in the basis of
$\ket{I_n}$, based on Eqs.~\eqref{eq:basisstatesqp} gives the matrix
representation of Eq.~\eqref{eq:HrsActAngIBasis}.
Finally, for connecting $\widehat{H}_{r:s}$ and $\widehat{\map}$ we require the basis
states with respect to the basis $\ket{q}$.
To this end, one can show from the above equations that
\begin{align}
\label{eq:basisstates_transformed_qp}
\Ket{I_n} &= \widehat{U}_{\mathcal{T}'}\Ket{I_n^{(0)}},
\end{align}
such that
\begin{align}
\label{eq:QuantumCanonicalTrafoOfStates}
\BraKet{q}{I_n} = \int \text{d}Q \BraOpKet{q}{\widehat{U}_{\mathcal{T}'}}{Q}
\BraKet{Q}{I_n^{(0)}}.
\end{align}
In principle, the operator $\BraOpKet{q}{\widehat{U}_{\mathcal{T}'}}{q'}$ can be
evaluated semiclassically, using the techniques described in
Ref.~\cite{Bog1992}.
However, this does not give an analytical closed form result and its evaluation
is numerically extremely tedious.
Furthermore, $\widehat{U}_{\mathcal{T}'}$ is usually so close to an identity
transformation such that a semiclassical evaluation of
$\BraOpKet{q}{\widehat{U}_{\mathcal{T}'}}{Q}$ contains too many turning points.
Hence, we take an alternative approach, which is numerically feasible:\
(i) We recognize that the states $\ket{I_n}$ are the eigenstates of the operator
$\widehat{I}$, originating from the phase-space coordinate $I$.
(ii) We define the function $I(q,p)$ which is obtained after the full canonical
transformation $\mathcal{T}$.
(iii) We define the Weyl-quantization of this function on a phase-space torus
giving the hermitian matrix of Eq.~\eqref{eq:WeylI}.
(iv) We diagonalize this matrix numerically, yielding the states
$\BraKet{\overline{q}_l}{I_n}$.
Finally, obtaining the modes $\BraKet{\overline{q}_l}{I_n}$ from an eigenvalue
equation comes at the cost that their relative phase (usually ensured via
Eq.~\eqref{eq:basisstatesqp} or alternatively via Eqs.~\eqref{eq:basisstatesQP}
and \eqref{eq:basisstates_transformed_qp}) is lost.
For the standard map we try to restore this phase by exploiting the symmetry of
Eq.~\eqref{eq:WeylI}, which for our system becomes a real symmetric matrix.
In that we can ensure that the coefficient vector $\BraKet{\overline{q}_l}{I_n}$
can be chosen real.
Finally, we fix the sign of this coefficient vector, by aligning
it with the mode $\BraKet{Q}{I_n^{(0)}}$ defined via
Eq.~\eqref{eq:basisstatesQP_Qrepr}.
This means, we choose the sign of the coefficient vector
$\BraKet{\overline{q}_l}{I_n}$ such that the following relation is
fulfilled
\begin{align}
\label{eq:ModeAligning}
\sum_{n} \BraKet{I_n}{\overline{q}_l}
\left.\left[\BraKet{Q}{I_n^{(0)}}\right]\right|_{Q=\overline{q}_l} > 0.
\end{align}
This assumes that the unitary operator representing the quantum
canonical transformation in Eq.~\eqref{eq:QuantumCanonicalTrafoOfStates} is
sufficiently close to an identity transformation
$\widehat{U}_{\mathcal{T}'}\approx1$.
\end{appendix}
|
\section{Introduction and Motivation}
The discovery of neutrino mass and lepton mixing provides key evidence for
new physics beyond the Standard Model (SM) \cite{King:2015aea,King:2003jb,Altarelli:2010gt,King:2013eh,King:2014nza}.
The seesaw mechanism \cite{Minkowski:1977sc,GellMann:1980vs,Yanagida:1979as,Mohapatra:1979ia,Schechter:1980gr} is an attractive possibilty to account
for the origin of neutrino mass and lepton mixing in terms of right-handed neutrinos with large Majorana masses.
$SO(10)$ Grand Unified Theories (GUTs) \cite{Fritzsch:1974nn} predict such
right-handed neutrinos which appear along with SM matter fields in a single $\mathbf{16}$ multiplet. When the $SO(10)$ gauge group is broken to that of the SM, neutrino mass is an inevitable
consequence. In order to satisfy the constraint of gauge coupling unification, we shall here
assume low energy supersymmetry (SUSY) \cite{Martin:1997ns}. However to also account for gravity, one needs to
go beyond gauge theories, and here we shall focus on an $M$ theory version of string theory \cite{Witten:1995ex,Horava:1995qa}.
Recently we showed how $SO(10)$ SUSY GUTs could emerge from
$M$ Theory compactified on a $G_2$-manifold \cite{Acharya:2015oea}.
In this framework, discrete symmetry and Wilson lines \cite{Witten:2001bf} were used to prevent proton decay while maintaining gauge unification. In contrast to the $SU(5)$ version \cite{Acharya:2008zi,Acharya:2011te}, the Wilson line symmetry breaking mechanism
in $SO(10)$ requires additional matter at the TeV scale, with the quantum numbers of an extra
${\bf 16}_X$ plus $\overline{\bf 16}_X$ \cite{Acharya:2015oea}. In addition,
there were a number of unresolved issues in this approach, notably the mechanism for breaking the extra gauged $U(1)_X$ which accompanies the SM gauge group after the Wilson line symmetry breaking mechanism in $SO(10)$. This gauge group is the usual one in the maximal $SO(10)$ subgroup $SU(5)\times U(1)_X$ \footnote{The $U(1)_X$ is also commonly called $U(1)_{\chi}$ in the literature.},
where $SU(5)$ embeds the SM gauge group.
The key point is that, since Abelian Wilson line symmetry breaking preserves the rank of the gauge group,
the $U(1)_X$ gauge group needs to be broken by some other mechanism in the low energy effective field theory. Since right-handed
Majorana neutrino masses can only arise once the $U(1)_X$ is broken, the origin of neutrino mass
is therefore linked to this symmetry breaking.
In this paper we address the problem of $U(1)_X$ breaking and neutrino masses arising from
the $SO(10)$ $M$ theory, following the construction in \cite{Acharya:2015oea}, although our approach to solving these problems may be more general than the specific example studied.
To break the $U(1)_X$ gauge symmetry, we employ a (generalised) Kolda-Martin mechanism
\cite{Kolda:1995iw},
where higher order operators can break the symmetry, inducing vacuum expectation values (VEVs)
in the scalar right-handed neutrino components of both the
matter ${\bf 16}$ and the extra ${\bf 16}_X$, as well as their
conjugate partners. The subsequent induced R-parity violation \cite{Barbier:2004ez} provides additional sources
of neutrino mass, in addition to that arising from the seesaw mechanism
\cite{Minkowski:1977sc,GellMann:1980vs,Yanagida:1979as,Mohapatra:1979ia,Schechter:1980gr}. The resulting $11\times 11$ neutrino mass
matrix is analysed for one neutrino family (nominally the third family) and it is shown how a
phenomenologically acceptable neutrino mass can emerge. We defer any discussion of flavour mixing
to a possible future study of flavour from $M$ theory. Here we only show that symmetry breaking and
viable neutrino masses can arise within the framework of $M$ theory $SO(10)$, which is a highly non-trivial
result, given the constrained nature of $M$ theory constructions.
It is worth remarking that there are other alternative ways that have been proposed to study neutrino masses
in string theory, which are complementary to the approach followed here.
For example, it is possible to obtain large Majorana mass terms from instanton effects \cite{Acharya:2006ia,Blumenhagen:2006xt,Ibanez:2006da,Cvetic:2007ku,Buchmuller:2007zd}, large volume compactification \cite{Conlon:2007zza}, or orbifold compactfications of the heterotic string \cite{Buchmuller:2007zd}. However the origin of Majorana mass terms in $SO(10)$ has been non-trivial to realise from the string theory point of view. In GUTs all matter fields are unified in $\mathbf{16}$ multiplets whereas Higgs fields and triplet scalars are unified in $\mathbf{10}$. Since string theory does not predict
light particles in
representations larger than the adjoint, the traditional renormalisable terms involving $\mathbf{126}, \overline{\mathbf{126}}, \mathbf{210}$, e.g., $W \sim \mathbf{126}\:\mathbf{16}\:\mathbf{16}$, are not possible. The dominant higher order operators are quartic ones such as
$W = \overline{\mathbf{16}}\:\overline{\mathbf{16}}\:\mathbf{16}\:\mathbf{16}$.
Assuming that the supersymmetric partner of the right handed neutrino singlet gets a VEV, the Majorana mass is given by $M \sim \frac{\langle \widetilde{N} \rangle^2}{M_{PL}}$. However, the required values of neutrino mass imply $M > 10^{14}$ GeV, which gives $\langle \widetilde{N}\rangle \sim \sqrt{M m_{Pl}} \sim 10^{16}$ GeV.
The implementation of the seesaw mechanism \cite{Minkowski:1977sc,GellMann:1980vs,Yanagida:1979as,Mohapatra:1979ia,Schechter:1980gr} in other corners of string compactification
has also been discussed \cite{Faraggi:1990it,Faraggi:1993zh,Coriano:2003ui,Ghilencea:2002da}.
The layout of the remainder of the paper is as follows.
In section \ref{sec:review}, we will review the $SO(10)$ construction from $M$ Theory on $G_2$-manifolds, expanding the discussion in \cite{Acharya:2015oea}. In section \ref{sec:sym_breaking_section}, the mechanism for $U(1)_X$ breaking will be given. The neutrino mass matrix will be analysed in section \ref{neutrino_section}, and the numerical results presented in section \ref{sec:numerical}. Finally we conclude in section \ref{sec:conclusion}.
\section{SO(10) SUSY GUTS from $M$ Theory on $G_2$-manifolds} \label{sec:review}
$M$ Theory compactified on a $G_2$-manifold leads to a 4 dimensional theory with $\mathcal{N}=1$ SUSY, where gauge fields and chiral fermions are supported by different types of singularities in the compactified space \cite{Acharya:2001gy,Acharya:2004qe}. Yang-Mills fields are supported on three dimensional subspaces of the extra dimensions, along which there is an orbifold singularity, while chiral fermions will be further localised on conical singularities localised on these three dimensional spaces and interact with the gauge fields.
One of the key features of $M$ Theory compactified on $G_2$-manifolds without fluxes is that it provides a framework for generating hierarchies of mass scales. To understand the reason behind this notice that in $M$ Theory, the moduli fields, $s_i$, are paired with the axions, $a_i$, in order to form a complex scalar component of a superfield $\Phi_i$
\begin{equation}
\Phi_i = s_i + i a_i + \mbox{fermionic terms} \ .
\end{equation}
In the absence of fluxes, the axions enjoy an approximate shift-symmetry, which is remnant of the higher dimensional gauge symmetry, $a_i \to a_i + c_i $
where $c_i$ is an arbitrary constant. This Peccei-Quinn symmetry, in conjunction with holomorphicity of the superpotential, severely constrains the superpotential for the moduli. As such, terms which are polynomial in the moduli and matter fields are forbidden at tree-level in superpotential, appearing only in the K\"ahler potential.
In general
non-perturbative effects such as instantons
break the above shift symmetry, and generate a non-perturbative
superpotential involving moduli and matter. Interactions will be generated by membrane instantons, whose actions are given by exponentials of the moduli. As the moduli stabilise and acquire VEVs, these exponentials will turn out to be small, and the VEV of the hidden sector superpotential naturally leading to a generation of hierarchical masses at the GUT scale \cite{Acharya:2007rc}.
These ideas were used to construct the $G_2$-MSSM \cite{Acharya:2008zi,Acharya:2011te}, an $SU(5)$ SUSY GUT from $M$ Theory on a $G_2$ manifold with the MSSM spectrum. Here, we discuss an extention of the program to the $SO(10)$ GUT group \cite{Acharya:2015oea}, while referring to previous work on $G_2$ compactifications and consequent predictions for the parameters \cite{Acharya:2008hi,Acharya:2012tw}.
In the remainder of this section,
we focus on the $SO(10)$ SUSY GUT from $M$ Theory on $G_2$ manifolds which we proposed in \cite{Acharya:2015oea}.
The breaking patterns of an abelian Wilson line are the same as the ones of an adjoint Higgs. The simplest case of a surviving group that is the most resembling to the SM is
\begin{equation}
SO(10) \to SU(3)_c \times SU(2)_L \times U(1)_Y \times U(1)_X \ ,
\end{equation}
under which the branching rules of the GUT irreps read
\begin{align}
{\bf 10} : \ &H_u=({\bf 1},{\bf 2})_{\left(\frac{1}{2}, 2\right)} \oplus H_d=({\bf 1},{\bf 2})_{\left(-\frac{1}{2},-2\right)} \oplus D=({\bf 3},{\bf 1})_{\left(-\frac{1}{3},2\right)}\oplus \overline{D}=({\bf \overline{3}},{\bf 1})_{\left(\frac{1}{3},-2\right)}\ , \\
{\bf 16} : \ &L=( {\bf 1},{\bf 2})_{\left(-\frac{1}{2},3\right)}\oplus e^c=({\bf 1},{\bf 1})_{(1,-1)} \oplus N=({\bf 1},{\bf 1})_{(0,-5)}\oplus u^c = ({\bf \overline{3}},{\bf 1})_{\left(-\frac{2}{3},-1\right)}\oplus \nonumber \\
&\oplus d^c =({\bf \overline{3}},{\bf 1})_{\left(\frac{1}{3},3\right)}\oplus Q= ({\bf 3},{\bf 2})_{\left(\frac{1}{6},-1\right)} \ ,
\end{align}
and the subscripts are the charges under $U(1)_Y \times U(1)_X$, which are normalised as
$Q_Y = \sqrt{\frac{5}{3}}Q_1 , \ Q_X = \sqrt{40}\tilde Q_{X} $,
where $Q_1$, $\tilde Q_X$ are $SO(10)$ generators.
The Wilson line can be conveniently represented as
\begin{equation}
\mathcal W = \exp\left[\frac{i 2 \pi}{N}\left(a Q_Y + b Q_X\right)\right]=\sum^\infty_{m=0} \frac{1}{m!} \left(\frac{i 2\pi}{N}\right)^m\left(a Q_Y + b Q_X\right)^m \ ,
\end{equation}
where the coefficients $a$, $b$ are constrained by the requirement that $\mathcal{W}^N = 1$ and specify the parametrisation of the Wilson line. Under the linear transformation
\begin{align}
\frac{1}{2}a+2 b \to \alpha \ , \\
\frac{1}{3}a-2b \to \beta\ ,
\end{align}
its action on the fundamental irrep then reads
\begin{equation}\label{eq:W10}
\mathcal{W} 10 = \eta^\alpha H_u \oplus \eta^{-\alpha} H_d \oplus \eta^{-\beta} D \oplus \eta^{\beta} \overline{D} \ ,
\end{equation}
where $\eta$ is the $N$th root of unity.
Likewise the Wilson line matrix acts on the 16 irrep as
\begin{equation}\label{eq:W16}
\mathcal{W}16=\eta^{-\frac{3}{2}\beta} L \oplus \eta^{ \alpha +\frac{3}{2}\beta} e^c \oplus \eta^{-\alpha +
\frac{3}{2}\beta} N \oplus \eta^{-\alpha - \frac{1}{2}\beta} u^c \oplus \eta^{ \alpha - \frac{1}{2}\beta} d^c \oplus \eta^{\frac{1}{2}\beta} Q \ ,
\end{equation}
which could be simplified a bit further by replacing $\beta\to 2\beta$ without loss of generality, in order for the parameters to read as integers.
The effective discrete charges -- of different states on a chiral supermultiplet that absorbs Wilson line phases -- will be the overall charge of the discrete symmetry (common to all states belonging to the same GUT irrep) in addition to the Wilson line phases (different for each state inside the GUT irrep).
Having all the ingredients required to employ Witten's discrete symmetry proposal, we would like to have a consistent implementation of a well-motivated doublet-triplet splitting mechanism as it was done for $SU(5)$. Unfortunately the customary approach to the problem does not seem to work with $SO(10)$, as shown in \cite{Acharya:2015oea}. To understand this first notice that Witten's splitting mechanism can only work in order to split couplings between distinct GUT irreps. This is understood as $\mathcal W$ has the form of a gauge transformation of the surviving group and so it will never be able to split self bilinear couplings of a GUT irrep. For example, if one takes a $\textbf{10}$ with Wilson line phases to contain the MSSM Higgses, we can see from \cref{eq:W10} that both mass terms for the Higgses and coloured triplets are trivially allowed. We could consider that in order to split the Higgses, $H_u$ and $H_d$, from the coloured triplets -- $D$, $\overline D$ -- we would need to add another ${\bf 10}$, but it was shown that this cannot be achieved and so we are ultimately left with light coloured triplets.
In order to allow for light $D$, $\overline{D}$ we need to guarantee that they are sufficiently decoupled from matter to prevent proton-decay. To accomplish this, we can use the discrete symmetry to forbid certain couplings,
namely to {\it decouple $D$ and $\overline{D}$ from matter}.
Such couplings arise from the $SO(10)$ invariant operator $\mathbf{10}\ \mathbf{16}\ \mathbf{16}$, with $\mathbf{16}$ denoting the three $SO(10)$ multiplets, each containing a SM family plus right handed neutrino $N$. If $\mathbf{16}$ transforms as $\eta^{\kappa} \mathbf{16}$, the couplings and charge constraints are
\begin{align}
H_u \mathbf{16} \mathbf{16}\; : & \;2\kappa + \alpha + \omega = 0 \;\mbox{mod}\; N \\
H_d \mathbf{16} \mathbf{16}\; : & \;2\kappa - \alpha + \omega = 0 \;\mbox{mod}\; N \\
D \mathbf{16} \mathbf{16}\; : & \;2\kappa - \beta + \omega \neq 0 \;\mbox{mod}\; N \\
\overline{D} \mathbf{16} \mathbf{16} \; : & \;2\kappa + \beta + \omega \neq 0 \;\mbox{mod}\; N,
\end{align}
where we allow for up-type quark Yukawa couplings together with couplings to the right-handed neutrinos,
\begin{equation}
y_u^{ij}H_u^w \mathbf{16}_i \mathbf{16}_j \equiv y_u^{ij}H_u^w (Q_iu_j^c+L_iN_j +i\leftrightarrow j),
\label{yu}
\end{equation}
and similarly for down-type quarks and charged leptons.
The couplings forbidden at a renormalizable tree-level by the discrete symmetry are generically regenerated from K\"ahler interactions through the Giudice-Masiero mechanism \cite{Giudice:1988yz}. While this provides the Higgsinos a TeV scale $\mu$-term mass, it also originates effective trilinear couplings with an $\mathcal{O}(10^{-15})$ coefficient. As these are generic, we need to systematically study their physical implications at low energies, such as proton-decay, R-parity violation, and flavour mixing.
For proton decay, effective superpotential will be generate by the following K\"ahler potential
\begin{equation}
K \supset \frac{s}{m_{Pl}^2} \overline{D} d^c u^c + \frac{s}{m_{Pl}^2} D e^c u^c + \frac{s}{m_{Pl}^2} D Q Q + \frac{s}{m_{Pl}^2} \overline{D} Q L + \frac{s}{m_{Pl}^2} D N d^c + \mbox{h.c.} \ ,
\end{equation}
where we assume $\mathcal{O}(1)$ coefficients. As the moduli acquire non-vanishing VEVs, these become
\begin{eqnarray}
W_{eff} & \supset \lambda D Q Q + \lambda D e^c u^c + \lambda D N d^c\label{eq:proton} + \nonumber \\
& +\lambda \overline{D} d^c u^c+\lambda \overline{D} Q L,
\end{eqnarray}
where we considering all couplings to be similar and taking one family for illustrative purposes. Notice that contrary to $SU(5)$ case, there is no extra contribution from rotation of $L$ and $H_u$ as the bilinear term $\kappa L H_u$ is not allowed by gauge invariance.
We estimate the scalar triplet mediated proton decay rate to be
\begin{equation}
\Gamma_p \simeq \frac{\left| \lambda^2 \right|^2}{16 \pi^2}\frac{m_p^5}{m_D^4} \simeq \left( 10^{42} \;\mbox{yrs}\right)^{-1} ,
\end{equation}
where we took the mass of the colour triplets to be $m_D \simeq 10^3$ GeV.
Another limit for triplet scalar comes from the cosmological constraints on its decay. As we have seen from proton-decay operators, triplet scalars can decay into quarks. If they start to decay during the Big Bang Nucleosynthesis (BBN) then nucleons could be disassociated, spoiling the predictions for light element
abundances. We can estimate another limit on the triplet scalar mass by calculating its lifetime as it decay through the processes $D \rightarrow e^c u^c, Q Q, Q L, d^c u^c$, and we get
\begin{equation}
\Gamma \simeq \lambda^2 m_D \simeq ( 0.1 \;\mbox{sec})^{-1} ,
\end{equation}
which is approximately consistent with BBN constraint. They will also give interesting collider signatures due to their long-lived nature.
\subsection{The vector-like family splitting}
Because the presence of a light vector-like pair coloured triplets spoils unification, we need a workaround that will preserve unification while keeping the presented doublet-triplet problem solution. We achieve this by considering the presence of extra matter that would form a complete GUT irrep with the coloured triplets, and hence restore unification. Unification constraints requires heavy states with equivalent SM gauge numbers, say $d^c_X$ and $\overline{d^c}_X$, that have to be subtracted from the spectrum. This can be achieved by adding a vector-like family pair, ${\bf 16}_X \overline {\bf 16}_X$, and splitting its mass terms using Wilson line phases.
Furthermore, as the Wilson line breaking pattern is rank-preserving, we still need to break the extra abelian gauge factor $U(1)_X$. This can be achieved if a scalar component of the right-handed conjugated neutrino pair of an extra vector-like family ${\bf 16}_X$, $\overline{\bf{16}}_X$ acquires VEVs. On top of this, this VEV can generate a Majorana mass for the matter right-handed conjugated neutrinos, providing a crucial ingredient for a type I see-saw mechanism.
In order to preserve gauge coupling unification, we notice that the down-type quarks -- $d^c_X$, $\overline{d^c}_X$ -- have the same SM quantum numbers as the coloured triplet pair -- $D$, $\overline{D}$ -- coming from the ${\bf 10}$. We take ${\bf 16}_X$ to be localised along a Wilson line, and find that it
transforms under the discrete symmetry as
\begin{equation}
{\bf 16}_X \to \eta^x \left( \eta^{-3\gamma} L \oplus \eta^{ 3\gamma+\delta} e^c \oplus \eta^{3 \gamma - \delta} N \oplus \eta^{-\gamma-\delta} u^c \oplus \ \eta^{-\gamma +\delta} d^c \oplus \eta^{\gamma} Q \right) .
\end{equation}
On the other hand, we let ${\bf {\overline{16}}}_X$ transform without Wilson line phases, ${\bf {\overline{16}}}_X \to \eta^{\overline x}\, {\bf {\overline{16}}}_X$, and the condition for the mass term that will split the vector-like family is
\begin{equation}
\overline{d^c}_X d^c_X : x - \gamma + \delta + \overline{x} = 0 \mod N ,\
\end{equation}
whilst forbidding all the other self couplings that would arise from ${\bf 16}_X {\bf {\overline{16}}}_X$.
The $d^c_X$, $\overline{d^c}_X$ quarks will then be naturally endowed a GUT scale mass through membrane instantons, provided that the singularities supporting ${\bf 16}_X$, $\overline{\bf{16}}_X$ are close enough to each other in the compactified space. The remaining states of ${\bf 16}_X$, $\overline{\bf{16}}_X$ will have a $\mu$ term of order TeV through the Giudice-Masiero mechanism.
The coloured triplets -- $D$, $\overline{D}$ -- and the light components of ${\bf 16}_X$, $\overline{\bf{16}}_X$ will effectively account for a full vector-like family. The light spectrum is then the one of MSSM in addition to this vector-like family, which in turn preserves unification, with a larger unification coupling at the GUT scale.
\subsection{R-parity violation\label{sec:RPV}}
Despite the existence of an effective matter parity symmetry inside $SO(10)$, the presence of a vector-like family will lead to R-parity violating (RPV) interactions though the VEV of the $N_X$, $\overline{N}_X$ components in the presence of moduli generated interactions. Furthermore, as we will see in detail in Section \ref{sec:sym_breaking_section}, the scalar component of the matter conjugate right-handed neutrino, $N$, will also acquire a VEV. These VEVs break $SO(10)$ and will inevitably generate RPV.
These interactions will mediate proton-decay, enable the lightest supersymmetric particle (LSP) to decay, and generate extra contributions to neutrino masses. In our framework RPV is generic, not only arising from allowed superpotential terms but as well from K\"ahler interactions involving moduli fields.
The interactions that break R-parity can either be trilinear or bilinear (B-RPV), and have different origins in our framework. The first contribution we can find comes from the tree-level renormalizable superpotential allowed by the discrete symmetry. Since we will encounter $\langle N \rangle \neq 0$, this means that even in a minimal setup, there will be an R-RPV contribution from matter Dirac mass coupling
\begin{equation}
W \supset y_\nu N H_u L \ ,
\end{equation}
reading
\begin{equation}\label{eq:BRPVfromDirac}
W \supset y_\nu \langle N \rangle H_u L \ .
\end{equation}
Next we turn our attention to the K\"ahler potential, where interactions otherwise forbidden by the discrete symmetry might arise if there is a modulus with required charge. In such case, there is another contribution arising from the non-vanishing VEVs of $N_X$, $N$ $\overline N_X$ in conjugation with moduli VEVs. To see this, notice that in the K\"ahler potential there are generically interactions of the form
\begin{equation}
K \supset \frac{1}{m_{Pl}}N H_u L+\frac{s}{m_{Pl}^2} \overline N_X H_u L+\frac{s}{m_{Pl}^2} \overline N_X^\dagger H_u L +\mbox{ h.c.}\ ,
\end{equation}
where while the first term exists in zeroth order in moduli (otherwise there would be no neutrino Dirac mass in the superpotential), the last two are otherwise forbidden by the discrete symmetry, and $s$ denotes a generic modulus for each coupling. These terms will generate contributions to B-RPV as $N_X$, $N$ $\overline N_X$, $s$ acquire VEVs.
There are two types of contribution arising from the terms above. The first is generates through the Giudice-Masiero mechanism. As the moduli acquire VEVs, new holomorphic couplings will appear in the superpotential
\begin{equation}
W_{eff, 1} = \frac{m_{3/2}}{m_{Pl}} \langle N \rangle H_u L +0.1 \frac{m_{3/2}}{m_{Pl}} \langle N_X \rangle H_u L + 0.1 \frac{m_{3/2}}{m_{Pl}} \langle \overline N_X^\dagger \rangle H_u L \ ,
\end{equation}
where $m_{3/2} \simeq \mathcal{O}(10^{4})$ GeV, and since $s/m_{Pl} \simeq 0.1$ in $M$ Theory.
Notice that in principle we would also have a term in the K\"ahler potential involving $N$, but this can be found to be subleading in comparison to the term arising from the Dirac mass Eq. \eqref{eq:BRPVfromDirac}.
The second contribution arises if the F-terms of the fields $N_X$, $N$, $\overline N_X$ are non-vanishing. In this case, we expect the appearance of the contributions
\begin{equation}
W_{eff, 2} = \frac{\langle F_N \rangle}{m_{Pl}} H_u L+0.1\frac{ \langle F_{N_X} \rangle}{m_{Pl}} H_u L+0.1 \frac{ \langle F_{\overline{N}^\dagger_X}\rangle}{m_{Pl}} H_u L \ ,
\end{equation}
and its magnitude will depend on how much F-breaking provoked by our symmetry breaking mechanism. Here we are considering that the case where $\overline N_X^\dagger H_u L$ cannot exist in the K\"ahler potential in zeroth order in a modulus field.
Putting all together, the B-RPV interactions account to the B-RPV paramter
\begin{equation}
W \supset \kappa H_u L
\end{equation}
with
\begin{equation}
\kappa = \left(y_\nu+\frac{m_{3/2}}{m_{Pl}}\right) \langle N \rangle + 0.1 \frac{m_{3/2}}{m_{Pl}} \langle N_X \rangle + 0.1 \frac{m_{3/2}}{m_{Pl}} \langle \overline N^\dagger_X \rangle +\frac{\langle F_N \rangle}{m_{Pl}} + 0.1 \frac{\langle F_{N_X}\rangle}{m_{Pl}} + 0.1 \frac{ \langle F_{\overline{N}^\dagger_X}\rangle}{m_{Pl}} \ ,
\end{equation}
and the relative strength of each contribution is model detail dependent, namely on neutrino Yukawa textures, symmetry breaking details, and F-flatness deviation.
In a similar manner, trilinear RPV couplings will be generated when $N$, $N_X$, $\overline N_X$, $s$ acquire VEVs.
In order to systematically study this, we notice that the trilinear RPV couplings come from the term
\begin{equation}
\mathbf{16}\ \mathbf{16}\ \mathbf{16}\ \mathbf{16} , \
\mathbf{16}_X \mathbf{16}\ \mathbf{16}\ \mathbf{16} , \
\mathbf{\overline{16}}_X^\dagger \mathbf{16}\ \mathbf{16}\ \mathbf{16}
\end{equation}
as the scalar component of $N_X$, $N$ acquires non-vanishing VEVs. Notice that the last term lives in the K\"ahler potential.
These are made forbidden at tree-level using the discrete symmetry of the compactified space. However, just like the $\mu$ terms and the B-RPV terms shown above, these terms will in general be present in the K\"ahler potential and will effectively be generated as the moduli acquire VEVs.
This happens again through the Giudice-Masier mechanism and we will find
\begin{equation}
\mathcal{O} \left(\frac{m_{3/2}}{m_{Pl}^2} (\langle N \rangle + \langle N_X \rangle+\langle \overline{N}_X^\dagger \rangle)\right) \{L L e^c,\ LQ d^c,\ u^c d^c d^c\} ,
\end{equation}
where $m_{3/2}/m_{Pl}\simeq \mathcal{O}(10^{-14})$. The apparent suppression of trilinear RPV is understood as these terms can only be generated by non-renormalizable terms in an $SO(10)$ context.
Similarly to the B-RPV case, there will be further contributions if the F-terms of $N_X$, $N$, $\overline N_X$ are non-vanishing. Namely we find
\begin{equation}
\mathcal{O} \left(\frac{\langle F_N \rangle + \langle F_{N_X} \rangle + \langle F_{\overline{N}_X^\dagger}\rangle}{m_{Pl}^2}\right) \{L L e^c,\ LQ d^c,\ u^c d^c d^c\} ,
\end{equation}
and again we expect these to be sub-leading even if the F-terms are not vanishing.
We see then that the values of all RPV coupling are strictly related to the details of the breaking mechanism employed to break the extra $U(1)_X$. This will be studied in great detail in Section \ref{sec:sym_breaking_section}. Furthermore, the bilinear B-RPV term generates a contribution to the physical neutrino masses \cite{Banks:1995by,Barbier:2004ez}. The complete picture of neutrino masses, including B-RPV operators, will be discussed in Section \ref{neutrino_section}.
We can study now some direct effects of RPV in the dynamics of our class of models. Under the assumption that $\kappa \ll \mu$, performing a small rotation, of
$\mathcal{O}(\kappa/\mu)$, in $(H_d,L)$ space, the last term can be absorbed $\mu H_dH_u$.
As a consequence, the first two terms will be enhanced by the Yukawa couplings $y_eH_dLe^c$, etc., leading to
\begin{equation}
W \supset y_e \frac{\kappa}{\mu} L L e^c + y_d \frac{\kappa}{\mu} LQ d^c + \lambda \frac{v}{m_{Pl}} u^c d^c d^c , \label{eq:RPVrot}
\end{equation}
and we have dropped the $\mathcal{O}(1/m_{Pl})$ contributions to the first two terms since now the Yukawa rotated contributions are much larger.
Also, we kept the last term with the parametrization $v$ describing all contributions. These will be very small, for example in the case the VEVs are high-scale, $\langle N_X \rangle \simeq 10^{16}$ GeV, the trilinear RPV coupling strength is of $\mathcal{O}(10^{-16})$. A direct consequence of this result is that proton decay will be slow, even when the $\Delta L = 1$ terms are enhanced.
While the proton is relatively stable, the enhanced terms will provide a decay channel for the LSP, which is now unstable.
In the limit that we can take the final states to be massless, and considering that the LSP is a neutralino mainly composed of neutral gauginos, the LSP lifetime through the decay $\tilde \chi^0 \to d^cQL$ can be estimated from a tree-level diagram involving a virtual $\tilde{d}^c$ with mass $m_0$,\footnote{See, for example, the diagrams in \cite{Martin:1997ns}.}
\begin{equation}
\tau_{LSP} \simeq \left(3.9 \times 10^{-15} \right)\left(\frac{\mu}{g_w y_d \kappa}\right)^2\left( \frac{m_0}{10 \ \mbox{TeV}}\right)^4 \left( \frac{100 \ \mbox{GeV}}{m_{LSP}}\right)^5 \sec,
\end{equation}
where $g_w$ is a weak gauge coupling. The LSP lifetime is bounded to be either $\tau_{LSP} \lesssim 1$ sec or $\tau_{LSP} \gtrsim 10^{25}$ sec \cite{Banks:1995by,Dreiner:1997uz}, from Big Bang Nucleosythesis (BBN) and indirect Dark Matter (DM) experiments, respectively. If we take $m_{LSP} \simeq 100\mbox{ GeV}$, $m_0 \simeq 10 \mbox{ TeV}$, $y_d=y_b \simeq 10^{-2}$, $g_w \simeq 0.1$, we find that the VEV $v_X$ is constrained to be either
\begin{align}
\kappa & \gtrsim 6 \times 10^{-2} \mbox{ GeV} \label{eq:kLSPBBN} \\
\ {\rm or} \ \kappa &\lesssim 2 \times 10^{-14} \mbox{ GeV} ,
\end{align}
for a short- and long-lived LSP, respectively. In the above estimate we used the fact that the decay involving the bottom Yukawa is the largest contribution to the decay width.
We can use the above result to infer some parametric dependence on the scale of the $U(1)_X$ breaking. If we have the leading contribution to the B-RPV coupling to be $\kappa \simeq \langle N_X \rangle \lambda \Rightarrow \langle N_X \rangle \gtrsim 10^{12}$ GeV. In this case, the LSP is too short lived to be a good DM candidate, but decays quickly enough to not spoil BBN predictions. On the other-hand, a low-scale VEV is bound to be $\langle N_X \rangle \lesssim 1$ GeV in order to allow for a long-lived LSP. This would imply the abelian gauge boson associated with extra $U(1)_X$ to be light, $m_{Z^\prime} < \mathcal{O}(1)$ GeV. This last scenario is completely excluded from experimental searches.
The lack of a good DM candidate in the visible sector indicates us that DM is realised elsewhere. For instance, it has been recently suggested that in the context of String/$M$ Theory, the generic occurrence of hidden sectors could account for the required DM mechanics \cite{Acharya:2016fge}.
\subsection{The see-saw mechanism}
The relevance of the bounds on the rank-breaking VEV is only fully understood when studying the details of symmetry breaking mechanism and neutrino masses. For example, if we start with an $SO(10)$ invariant theory the Yukawas are unified for each family leading to at least one very heavy Dirac neutrino mass, $m_\nu^D$. However, if the right-handed conjugated neutrino has a heavy Majorana mass, then the physical left-handed neutrino mass will be small through a type I see-saw mechanism. In order to accomplish this, one has to allow the following terms in the superpotential
\begin{equation}
W \supset y_{\nu} H_u L N + M N N ,
\end{equation}
where $y_\nu$ are the neutrino Yukawas, $L$ the matter lepton doublets, $N$ the right-handed conjugated neutrino, and $M$ its Majorana mass, which we take $M \gg m_\nu^D = y_\nu \langle H_u \rangle$. With the above ingredients, a mostly left-handed light neutrino will have a physical mass
\begin{equation}
m^\nu_{phy} \simeq - \frac{(m_\nu^D)^2}{M} .
\end{equation}
One of the most appealing features of $SO(10)$ models is that each family is in a $\bf{16}$ which includes a natural candidate for the right-handed conjugated neutrino, the $N$. In order to employ a type I see-saw mechanism, we need to generate a Majorana mass term for the matter right-handed conjugated neutrino through the operator $W\supset \overline{\bf 16}_X\overline{\bf 16}_X {\bf 16}\: {\bf 16}$ \footnote{Given that in $M$ Theory one does not account for irreps larger than the adjoint, this is the lowest order term that can generate a right-handed neutrino Majorana mass.} leading to the operator
\begin{equation}
\frac{1}{m_{Pl}} \overline{N}_X \overline{N}_X N N\ ,
\end{equation}
from which the Majorana mass for the (CP conjugated) right-handed neutrino field $N$
is emerges as
\begin{equation}
M \simeq \frac{\langle \overline{N}_X \rangle^2}{m_{Pl}} \ .
\end{equation}
We can now relate the bounds on the value of the D-flat VEVs
$\langle \overline{N}_X \rangle = \langle N_X \rangle$ from both RPV and the
requirement of a realistic see-saw mechanism. Since the physical neutrino mass in type I see-saw mechanism is given by
\begin{equation}
m_\nu^{phy} \simeq \frac{(m_\nu ^D)^2}{M} \ ,
\end{equation}
assuming $m_\nu ^D \simeq \mathcal{O}(100\mbox{ GeV})$, and knowing that the upper bound on the neutrino masses $m_\nu^{phy} \lesssim 0.1$ eV, one finds
\begin{equation}
M \gtrsim 10^{14} \mbox{ GeV} \Rightarrow
\langle {N}_X \rangle \gtrsim 10^{16} \mbox{ GeV}.
\end{equation}
The above argument suggests that we need to break the $U(1)_X$ close to the GUT scale. Since the Wilson line breaking mechanism is rank-preserving, we need to look for an alternative solution. Although the neutral fermion mass matrix will be considerably more intricate, obscuring the relations and hierarchies amongst different contributions to the neutrino masses, the above estimate motivates
the need for a high-scale $U(1)_X$ breaking mechanism.
\subsection{Effective light families}
For a simple SUSY $SO(10)$ model where each family is unified into a single irrep with universal soft masses, it is well known that
electroweak symmetry is difficult to break \cite{Hall:1993gn,Rattazzi:1994bm,Murayama:1995fn,Baer:1999mc,Auto:2003ys,Baer:2009ie}. Since the two Higgs soft masses are unified at GUT scale and have similar beta function due to Yukawa unification, either both masses are positive at electroweak scale and symmetry is not broken or both masses are negative and the potential becomes unbounded from below. Another aspect of Yukawa unification problem lies in the fact that low energy spectrum of quarks and leptons requires some degree of tuning in parameter space when their RG runnings are considered.
The EWSB and Yukawa textures issues are naturally solved if each family is not contained in one single complete ${\bf 16}$, but is instead formed of states from different Ultra Violet (UV) complete ${\bf 16}$s. In order to implement this in our framework, first we assume the existence of multiple ${\bf 16}$ with independent and different Wilson Line phases, alongside the existence of multiple ${\overline{\bf 16}}$. Second, we employ Witten's proposal to turn on some vector-like masses such that three effective light ${\bf 16}$ survive. Since in $M$ Theory the strength of the Yukawa couplings is given by membrane instantons, and are therefore related to distances between the singularities supporting the respective superfields, by constructing effective families from different UV ${\bf 16}$s one can obtain different Yukawa couplings within each family.
Such solution can be achieved if one considers $M$ complete ${\bf \overline{16}}_j$ and $M+3$ complete ${\bf 16}_i$ UV irreps. Allowing for masses between different states of these UV irreps to appear, one has schematically the mass terms in the superpotential
\begin{equation}
\mathbf{16}_i \mu_{ij} \overline{\mathbf{16}}_j,
\end{equation}
but since $i=1,...,M$ while $j=1,...,M+3$ the mass matrix $\mu_{ji}$ can only have at most rank $M$ and hence there will be three linear combinations composing three ${\bf 16}$ that will remain massless. If these masses are truly $SO(10)$ invariant, i.e.
\begin{equation}
\mathbf{16}_i \mu_{ij} \overline{\mathbf{16}}_j = \mu_{ij} \left( Q_i \overline Q_j + L_i \overline L_j + \ldots\right),
\end{equation}
each effective light family will be $SO(10)$ invariant. Consequently each family will retain unified Yukawa textures, and so this does not solve our problem of splitting the Yukawa couplings within each family.
However, Witten's proposal endows our framework with a GUT breaking discrete symmetry which can be employed to ensure that the superpotential mass matrices between the UV states
\begin{equation}
\mu^Q_{ij} Q_i \overline Q_j +\mu^L_{ij} L_i \overline L_j + \ldots,
\end{equation}
are not the same, leading to different diagonalisations of $Q$, $L$, etc which in turn break the Yukawa $SO(10)$ invariance. In order to accomplish that, take for example that the ${\bf 16}_i$ absorb distinct and independent Wilson line phases, while $\overline{\bf 16}_j$ do not, i.e. the UV irreps will transform under the discrete symmetry as
\begin{align}
{\bf 16}_i &\to\eta^{m_i}\left(\eta^{-3\gamma_i} L_i \oplus \eta^{ 3\gamma_i+\delta_i} e_i^c \oplus \eta^{3 \gamma_i - \delta_i} N_i \oplus \eta^{-\gamma_i-\delta_i} u_i^c \oplus \eta^{-\gamma_i +\delta_i} d_i^c \oplus \eta^{\gamma_i} Q_i\right) \label{eq:16icharges}\\
\overline{\bf 16}_j &\to \eta^{\overline m_i}\overline{\bf 16}_j \label{eq:16bjcharges},
\end{align}
and look for solutions for the discrete charges where different states have different mass matrices. Since explicit examples can only be given by solving extensive modular linear systems, which are computationally prohibitive, a fully working example with three light-families is not provided.
\section{$U(1)_X$ Breaking scenarios and mechanisms} \label{sec:sym_breaking_section}
In this section we are interested in implementing a symmetry breaking mechanism for the extra $U(1)_X$ in which the breaking VEV is stabilised at high values, more or less close to the GUT scale. In order to do so, we will look into the D-flat direction of the potential that breaks the extra $U(1)_X$. It was shown \cite{Drees:1986vd,Kolda:1995iw} that in the D-flat direction, non-renormalisable operators can provide such scenario. In its simplest inception, the Kolda-Martin
mechanism \cite{Kolda:1995iw} relies on a vector-like pair which lowest order term allowed in the superpotential is non-renormalizable
\begin{equation}
W = \frac{c}{m_{Pl}} (\Phi \bar \Phi)^2
\end{equation}
and alongside the soft-term Lagrangian
\begin{equation}
-\mc{L}_{soft} = m_\Phi^2 |\Phi|^2 + m_{\bar\Phi}^2 |\bar{\Phi}|^2,
\end{equation}
it is immediate to find that along the D-flat direction the potential has a non-trivial minimum which fixes the VEVs at a high scale
\begin{equation}
\Phi^2 = \sqrt{-\frac{(m_\Phi^2+m_{\bar\Phi}^2)m_{Pl}^2}{12 c}} ,
\end{equation}
where if we take $m \simeq 10^4$ the VEVs are estimated at $\Phi \simeq 10^{11}\mbox{ GeV}$.
There are some caveats to this mechanism as presented above. First, there is significant F-breaking as $\langle F \rangle \simeq \mc{O}(10^{15})\mbox{ GeV}$. While this is not a problem if the vector-like family does not share gauge interactions with ordinary matter, in our case non-vanishing F-terms will originate undesirable interactions, c.f. Section \ref{sec:RPV}. We shall therefore focus on F-flat solutions.
Second, the mechanism is not complete in the absence of the full soft-terms Lagrangian, which has to include
\begin{equation}
-\mc{L}_{soft} \supset C \frac{1}{m_{Pl}} \Phi^2 \bar{\Phi} ^2 + \mbox{ h.c.} .
\end{equation}
As we estimate $C\simeq\mc{O}(m_{3/2})$ at the GUT scale from the SUGRA \cite{Brignole:1997dp}, at the VEV scale this term is competing with the non-renormalisable terms in the potential arising from the superpotential, and therefore cannot be ignored.
Finally the model presented differs from ours as $\mu$-terms are generically generated by moduli VEVs even if they are disallowed by the discrete symmetry of the compactified space.
In order to proceed, we turn to a more complete version of the mechanism. To do so, we include the $\mu$-term
\begin{equation}
W = \mu \Phi \bar{\Phi}+\frac{c}{m_{Pl}} (\Phi \bar \Phi)^2
\end{equation}
and the more complete soft Lagrangian,
\begin{equation}
-\mc{L}_{soft} = m_\Phi^2 |\Phi|^2 + m_{\bar\Phi}^2 |\bar{\Phi}|^2 - (B\mu \Phi \bar \Phi + \mbox{ h.c.})+ \frac{C}{m_{Pl}} \Phi^2 \bar{\Phi} ^2 + \mbox{ h.c.} .
\end{equation}
Due to the presence of the $\mu$-term, the F-term
\begin{equation}
F_{\Phi} = \mu \bar \Phi + \frac{2 c }{m_{Pl}} \Phi \bar \Phi ^2
\end{equation}
can be set to zero for two different field configurations
\begin{equation}
F_\Phi =0 \Rightarrow \left\{
\begin{array}{l}
\bar \Phi =0 \\
\Phi \bar \Phi = - \frac{\mu m_{Pl}}{2 c}
\end{array}
\right.
\end{equation}
and the non-trivial VEV can be estimated. Taking $\mu \simeq \mc{O}(10^3)$ GeV, this leads to $| \Phi | = 10^{10.5}$ GeV. This looks very similar to the original Kolda-Martin case, with the exception being that the F-term can vanish, and the parametric dependence on the VEV is now on $\mu$ instead of a soft-mass. In general there might be a non-SUSY preserving vacuum elsewhere in field space, but we will work under the assumption that the SUSY vacua discovered with this approach are at least stable enough to host phenomenologically viable models.
We wish to assess if we can minimise the potential in this SUSY-preserving field configuration. For that, we need to check if the above field configuration will also extremise the soft-term Lagrangian. To see this we take
\begin{equation}
- \partial_\Phi \mc{L}_{soft} = m^2_\Phi \Phi^* - B\mu \bar \Phi + \frac{2 C}{m_{Pl}} \Phi \bar \Phi^2 =0
\end{equation}
and, in the limit the VEVs are real, we find a trivial and a non-trivial solutions
\begin{equation}
- \partial_\Phi \mc{L}_{soft} = 0 \Rightarrow \left\{
\begin{array}{l}
\Phi =0 \\
\Phi ^2 = - \frac{(m_\Phi^2-B\mu) m_{Pl}}{2 C}
\end{array}
\right.
\end{equation}
and the second one seems very similar to the non-trivial configuration derived through the F-term. In fact, both conditions can be met. To see this, we re-parametrise the soft-terms by factoring out their dimensionful dependence on $m_{3/2}$
\begin{align}
B\mu &= m_{3/2} \mu b \\
C & = m_{3/2} \tilde c \\
m_\Phi & = m_{3/2} a ,
\end{align}
where $a$, $b$, $\tilde c$ are dimensionless, and from SUGRA formulae they are $\mc{O}(1)$ at the GUT scale. Of course they will evolve with the scale through RGE evolution, so they need not to be always of the same order. The condition that both the F-flatness and soft-term stabilisation are jointly achieved boils down to be a relation between parameters
\begin{equation}
\frac{\tilde c}{c} = \frac{2 a m_{\Phi} -\mu b}{\mu} ,
\end{equation}
which is generically valid.
In order for the above non-trivial VEV be a minimum, we need the trivial VEV solution to account for a maximum. This is to say that the mass matrix for the system $(\Phi, \bar \Phi^*)$ evaluated at the origin has a negative eigen-value. In our case this accounts for allowing its determinant to be negative
\begin{equation}
(|\mu|^2+m^2_\Phi)(|\mu|^2+m^2_{\bar \Phi})-B\mu^2<0.
\label{eq:CondTrivialMaximum}
\end{equation}
We notice as well that the above discussion can be immediately extended for the case that the lowest order non-renormalisable term allowed by the discrete symmetry
\begin{equation}
W \supset \frac{c}{m_{Pl}^{2n-3}}(\Phi \bar \Phi)^n \Rightarrow \Phi \simeq (\mu m_{Pl}^{2n-3})^{\frac{1}{2n-2}}
\end{equation}
happens for $n \geq 2$, and not only for $n=2$. Even so, the presented implementation of the Kolda-Martin mechanism only accounts for a vector-like pair of superfields, while in our case the system breaking the extra $U(1)_X$ is composed of $N$, $N_X$, $\overline N_X$ states.
Therefore, we want to find similar solutions starting with the superpotential
\begin{equation}
W = \mu^N_{Xm} N \overline N_X + \mu^{N}_X N_X \overline N_X + \frac{c_{2,2}}{m_{Pl}} (N \overline N_X)^2 + \frac{c_{n,k}}{m_{Pl}^{2n-3}} (N_X \overline N_X )^{n-k} (N \overline N_X)^k
\end{equation}
where $n\geq 2$ and $k<n$. The third term generates a Majorana mass for the matter right-handed conjugated neutrino, $N$. The full soft-term Lagrangian for this theory is
\begin{align}
-\mc{L}_{soft} = & m_{N}^2 |N|^2 + m_{N_X}^2 |N_X|^2 + m_{\overline N_X}^2 |\overline N_X|^2 - (B\mu^N_{Xm} N \overline N_X +\mbox{ h.c.}) - (B\mu^N_{X} N_X \overline N_X+\mbox{ h.c.}) \nonumber \\
&+ \left(\frac{C_{2,2}}{m_{Pl}}(N \overline N_X)^2+\mbox{ h.c.}\right) +\left(\frac{C_{n,k}}{m_{Pl}^{2n-3}} (N_X \overline N_X )^{n-k} (N \overline N_X)^k + \mbox{ h.c.}\right)
\end{align}
where again $C_{i,j}$ coefficients are $\mc{O}(m_{3/2})$ at the GUT scale.
The F-terms now read
\begin{align}
F_{N} & = \mu^N_{Xm} \overline N_X + \frac{2 c_{2,2}}{m_{Pl}} N \overline N_X^2 + \frac{ k c_{n,k}}{m_{Pl}^{2n-3}} N_X^{n-k} N ^{k-1}\overline N_X^n\\
F_{N_X} & = \mu^N_{X} \overline N_X + \frac{ (n-k) c_{n,k}}{m_{Pl}^{2n-3}} N_X^{n-k-1} N ^{k}\overline N_X^n\\
F_{\overline N_X} &= \mu^N_{Xm} N + \mu^N_{X} N_X + \frac{2 c_{2,2}}{m_{Pl}} N^2 \overline N_X + \frac{n c_{n,k}}{m_{Pl}^{2n-3}} N_X^{n-k} N^k \overline N_X^{n-1}
\end{align}
which have a significantly more challenging look than the simplified version presented above. Nonetheless, the same conclusions hold. The above F-terms become more tractable for the $k=0$ and $k=n-1$ cases. In these cases it is possible to get algebraic expressions for the VEVs estimates. For the $k=0$, the F-flatness conditions alone give us
\begin{align}
N \overline N_X &= - \frac{\mu^N_{Xm} m_{Pl}}{2 c_{2,2}} \\
N_X \overline N_X &= \left(- \frac{\mu^N_X m_{Pl}^{2n-3}}{n c_{n,o}}\right)^{\frac{1}{n-1}}
\end{align}
while for $k=n-1$, analogous expressions can be obtained
\begin{align}
| N \overline N_X | &\simeq ( \mu^N_{Xm} m_{Pl}^{2n-3} )^{\frac{1}{n-1}} \\
|N_X \overline N_X| &\simeq ((\mu^N_X)^{3-n} m_{Pl}^{3n-5})^{\frac{1}{n-1}}
\end{align}
where the approximations mean we dropped $\mc{O}(1)$ parameters and took all $\mu$-terms to be of the same order, which is expected.
In both cases, the ratio between the $N_X$ and $N$ VEV is follows the same dependency on $n$
\begin{equation}
\left|\frac{N_X}{N} \right| \simeq \left(\frac{m_{Pl}}{\mu}\right)^{\frac{n-2}{n-1}}\simeq \begin{cases}
1 & n=2 \\
10^{7.5} & n=3 \\
10^{10} & n = 4
\end{cases}
\end{equation}
where we $\mu$ is an $\mc{O}(\mu^N_{X},\mu^N_{Xm})$ parameter. This result shows that there is a hierarchy between $N_X$ and $N$ VEVs, which is very desirable as $N$ VEVs can generate large B-RPV couplings, c.f. Section \ref{sec:RPV}.
Just like before, we use the D-flat direction
\begin{equation}
\left|\frac{\overline N_X}{N_X}\right|^2=\left| \frac{N}{N_X}\right|^2 + 1,
\end{equation}
which sets the magnitude of the three VEVs. The results for $k=0$ and $k=n-1$ can be immediately estimated algebraically, in contrast to the other cases. The full result of SUSY preserving configurations can be seen in Table \ref{tab:KMSUSYVacua}. It is important to note that for $n=4$, the only viable scenario is for $k=0$, while for $n=3$ the $k=2$ is not viable as there are super-GUT VEVs. In the end we are only interested in the sensible cases, where the VEVs are below the GUT scale and therefore the mechanism is self-consistent.
\begin{table}
\begin{center}
{\renewcommand{\arraystretch}{1.5}
\begin{tabular}{|c|c||c|c|c|}
\hline\hline
$n$ & $k$ & $N$ (GeV) & $N_X$ (GeV) & $\overline N_X$ (GeV)\\ \hline\hline
\multirow{2}{*}{2} & 0 & $10^{10.5}$ & $10^{10.5}$ & $10^{10.5}$ \\ \cline{2-5}
& 1 & $10^{10.5}$ & $10^{10.5}$ & $10^{10.5}$ \\ \hline\hline
\multirow{3}{*}{3} & 0 & $10^{6.5}$ & $10^{14.25}$ & $10^{14.25}$ \\ \cline{2-5}
& 1 & $10^{10.2}$ & $10^{15.5}$ & $10^{15.5}$ \\ \cline{2-5}
& 2 & $10^{10.5}$ & $10^{18}$ & $10^{18}$ \\ \cline{2-5}\hline\hline
\multirow{4}{*}{4} & 0 & $10^{5.5}$ & $10^{15.5}$ & $10^{15.5}$ \\ \cline{2-5}
& 1 & $10^{10.1}$ & $10^{16.5}$ & $10^{16.5}$ \\ \cline{2-5}
& 2 & $10^{10.3}$ & $10^{18}$ & $10^{18}$ \\ \cline{2-5}
& 3 & $10^{10.5}$ & $10^{20.5}$ & $10^{20.5}$ \\ \cline{2-5}\hline\hline
\end{tabular} }
\end{center}
\label{tab:KMSUSYVacua}
\caption{Estimate of the magnitude of the VEVs in SUSY vacua for different implementations of the modified Kolda-Martin mechanism. In all cases the scalar component of the (CP conjugated)
right-handed neutrino field $N$ develops a VEV, breaking R-parity, in addition to the
$N_X$ and $\overline{N}_X$ VEVs.}
\end{table}
The SUSY configurations above are expected stabilise the soft-terms Lagrangian just before. The stabilisation conditions are
\begin{align*}
m_{\Phi_1}^2 \Phi_1^* - B\mu_1 \bar \Phi + \frac{2 C_{2,2}}{m_{Pl}}\Phi_1 \bar \Phi^2 &=0 \\
m_{\Phi_2}^2 \Phi_2^* - B\mu_2 \bar \Phi +\frac{ n C_{n,0}}{m_{Pl}^{2n-3}} \Phi_2^{n-1} \bar \Phi^n &=0\\
m_{\bar \Phi}^2 \bar \Phi^* - B\mu_1 \Phi_1 -B\mu_2 \Phi_2 + \frac{2 C_{2,2}}{m_{Pl}}\Phi_1^2 \bar \Phi +\frac{n C_{n,0}}{m_{Pl}^{2n-3}} \Phi_2^{n} \bar \Phi^{n-1}&=0
\end{align*}
and re-parametrising the dimensionful soft-terms just as before, the above conditions will resemble the F-flatness conditions in form and so they'll be jointly respected taken the parameters of the theory respect relations between them.
As before, the condition that the above extrema are minima is that the potential has a runaway direction around the origin. This is the same to say that, when close to the origin the potential takes the form
\begin{equation}
V \simeq {\bf N}^* \cdot M_{ N} \cdot {\bf N}
\end{equation}
with ${\bf N}=(N, N_X, \overline N_X^*)$, such that $M_{N}$ at least one negative eigenvalue to account for a run-away behaviour at the trivial extremum. Boundness of the potential in the D-flat direction is achieved by noticing that -- for each field direction -- at least a quadratic term from the non-renormalisable interactions becomes the leading contribution, while keeping a run-away behaviour at the origin.
\section{Neutrino-neutralino mass matrix}\label{neutrino_section}
The different breaking scenarios discussed in the previous section rely on different superpotential terms, which are either present or suppressed depending the discrete symmetry of the compactified $G_2$ space. Furthermore, the generic presence of a matter field VEV, $\langle N \rangle$, will generate B-RPV terms, as seen in Section \ref{sec:RPV}. In turn, these provide a new source of neutrino masses which has to be taken into account.
To be more precise we enumerate all the interactions that contribute to neutrino masses. First, we let the matter neutrino to have a Yukawa coupling at tree-level, of the form
\begin{equation}
W_{tree} \supset y_{\nu} N L H_u \ .
\end{equation}
Next we have to consider the non-renormalizable terms that employ the KM mechanism for each scenario. Alongside this, we also keep a term that can generate a Majorana mass for the matter right-handed conjugated neutrino, $N$. On top of these, we include a set of non-renormalizable terms involving the Higgses or $L$-type fields, in first order of $1/m_{Pl}$. The non-renormalizable terms that will affect the neutral fermion mass matrix are then
\begin{align}\label{eq:Wnonren}
W_{non.ren.} \supset & \frac{c_{2,2}}{m_{Pl}}\left(N N\right)(\overline N_X \overline N_X) + \frac{c_{n,k}}{m^{2n-3}_{Pl}} \left({N}_X \overline{N}_{{X}} \right)^{n-k}\left({N} \overline{N}_{{X}} \right)^{k} \nonumber \\
& + \frac{1}{m_{Pl}}\left(b_1 H_d H_u L \overline{L}_X + b_2 L L \overline L_X \overline L_X + b_3 H_d H_u L_X \overline{L}_X + b_4 L L_X \overline{L}_X \overline{L}_X \right. \nonumber\\
& +\left. b_5 L_X L_X \overline L_X \overline L_X + b_6 H_d H_u N \overline{N}_X + b_7 L \overline{L}_X N \overline{N}_X + b_8 L_X \overline{L}_X N \overline{N}_X \right.\nonumber\\
&+ \left. b_9 H_d H_u N_X \overline{N}_X + b_{10} L \overline L_X N_X \overline N_X +
b_{11} L_X \overline L_X N_X \overline N_X\right) .
\end{align}
The terms that are disallowed by discrete symmetry are generically re-generated as the moduli acquire VEVs. As such, the following K\"ahler potential terms will have an important contribution for neutrino masses
\begin{align}
K \supset & \frac{s}{m_{Pl}} \overline L_X L_X + \frac{s}{m_{Pl}} \overline L_X L + \frac{s}{m_{Pl}} \overline N_X N_X +\frac{s}{m_{Pl}} \overline N_X N +\frac{s}{m_{Pl}} \overline H_u H_d \nonumber \\
& + \frac{s}{m_{Pl}^2}N_X L_X H_u + \frac{s}{m_{Pl}^2}N L H_u + \frac{s}{m_{Pl}^2}N_X L H_u + \frac{s}{m_{Pl}^2}N L_X H_u + \frac{s}{m_{Pl}^2}\overline N_{X} \overline L_{X} H_d ,
\end{align}
where $s$ denotes a generic modulus fields that counterbalances the discrete charge. This modulus field needs not to be the same for each coupling. As the moduli acquire VEVs as they are stabilised, the above terms will generate the effective superpotential
\begin{align}
W_{eff} \supset & \mu^L_{XX} \overline{L}_X L_X + \mu^L_{Xm} \overline{L}_X L+\mu^N_{XX} \overline{N}_X N_X+\mu^N_{Xm} \overline{N}_X N+ \mu H_u H_d\nonumber\\
&+ \lambda_{\overline X \overline X} H_d \overline{L}_X \overline{N}_X + \lambda_\nu H_u L N + \lambda_{mX} H_u L N_X + \lambda_{Xm} H_u L_X N + \lambda_{XX} H_u L_X N_X
\end{align}
where the parameters can be estimated to lie inside the orders of magnitude
\begin{align}
\mu \simeq & m_{3/2}\frac{s }{m_{Pl}} \simeq \mathcal{O}(10^{3})\mbox{ GeV} \\
\lambda \simeq & m_{3/2}\frac{s }{m_{Pl}^2} \simeq \mathcal{O}(10^{-15}) .
\end{align}
Therefore, the total superpotential, which includes all the interactions that contribute to the neutral fermion mass matrix is give by
\begin{align}\label{eq:Wtotal}
W_{total} & \supset W_{tree} + W_{non.ren.}+ W_{eff} .
\end{align}
In our framework we have VEVs of the $N$-type fields that can be significantly large, depending on which implementation of the KM mechanism we assume. As such, B-RPV couplings, mixing Higgses superfields with $L$-type superfields, appear in the superpotential as
\begin{equation}
\kappa_m {H}_u L + \kappa_X {H}_u L_X +\kappa_{\overline{X}} {H}_d \overline L_{X}
\end{equation}
where the $\kappa$-parameters read
\begin{align}
\kappa_m &\simeq (y_{\nu} + \lambda_{\nu}) \langle N \rangle + \lambda_{mX} \langle N_X \rangle \label{eq:kappam
\\
\kappa_X &\simeq \lambda_{Xm} \langle N \rangle + \lambda_{XX} \langle N_X \rangle\label{eq:kappax
\\
\kappa_{\overline{X}} &\simeq \lambda_{\overline{X}\overline{X}}\langle \overline N_{X} \rangle\label{eq:kappaxbar
\end{align}
where we are dropping the $F$-terms contribution as the solutions for our KM mechanism presented in Section \ref{sec:sym_breaking_section} are aligned in the $D$ and $F$ directions. We also note that we are assuming no tree-level Yukawa couplings involving extra vector-like $N_X$, $\overline N_X$ for the KM scenarios.
Furthermore, the presence of B-RPV induces a sub-EWS VEV on the scalar components of the $\nu$-type fields. In our case, below the EWS, we expect all $\nu$-type scalars to acquire a non-vanishing VEV, generating a mixing between $N$-type fermions and Higgsinos through
\begin{equation}
\epsilon_m H^0_u N + \epsilon_X H^0_u N_X + \epsilon_{\overline{X}} H^0_d N_{\overline{X}}
\end{equation}
where the coefficients read
\begin{align}
\epsilon_m &\simeq (y_{\nu} + \lambda_{\nu} ) \langle \nu \rangle + \lambda_{mX} \langle \nu_X \rangle \\
\epsilon_X &\simeq \lambda_{Xm} \langle \nu \rangle + \lambda_{XX} \langle \nu_X \rangle \\
\epsilon_{\overline{X}} &\simeq \lambda_{\overline{X}\overline{X}} \langle \nu_{\overline{X}} \rangle
\end{align}
and, as expected, they have the same generic form as the $\kappa$-parameters since both set of parameters arise from trilinear, Yukawa, couplings in the superpotential.
Finally, as in the MSSM, the presence of VEVs will mix some fermions with gauginos through kinetic terms, namely the Higgsinos with $\tilde B_1$, $\tilde W^0$ due to the Higgses VEVs. In our case we also have $N$-type and $\nu$-type scalar VEVs, which will mix gauginos with matter fermions through kinetic terms. We have, for the $SU(2)$ states,
\begin{equation}
g' \widetilde{B} \langle \widetilde{\nu}_i\rangle \nu_i,\;\;\; g \widetilde{W}^0 \langle \widetilde{\nu}_i\rangle \nu_i,\;\;\; g'' \widetilde{B}_X \langle \widetilde{\nu}_i\rangle \nu_i
\end{equation}
while for the $N$-states, which are singlets under the SM gauge group, the mixing with the gaugino of the extra $U(1)_X$ gauge group
\begin{equation}
g'' \widetilde{B}_X \langle \widetilde{N}_i \rangle N_i
\end{equation}
where, in both expressions, we used the shorthand $g' = \sqrt{\frac{5}{3}} g_1$ and $g'' = \frac{1}{2\sqrt{10}}g_X$.
With all the above considerations, we can now construct the $11\times 11$ mass matrix for neutral fermions of our model. We define this matrix in the basis
\begin{equation}
\psi = (\widetilde{B},\widetilde{W}^0,\widetilde{B}_X, \widetilde{H}^0_d,\widetilde{H}^0_u,\nu,\nu_X,\nu_{\overline{X}},N,N_X,N_{\overline{X}}),
\end{equation}
and it has the schematic form
\begin{equation}
\mathbf{M}_{\chi-\nu} = \begin{pmatrix}
\mathbf{M}_{\chi^0}^{5\times5} & \mathbf{M}_{\chi\nu}^{5\times6} \\
(\mathbf{M}_{\chi\nu}^{5\times6})^T & \mathbf{M}_{\nu}^{6\times6}
\end{pmatrix}. \label{mass_matrix}
\end{equation}
The usually called neutralino part of the matrix includes only mass terms involving gauginos and Higgsinos, and its form is very similar to the MSSM, except we have an extended gauge group with one more $U(1)_X$ factor. It reads
\begin{equation}
\mathbf{M}_{\chi^0}^{5\times5} = \begin{pmatrix}
M_1& 0 & 0 & -\frac{1}{\sqrt{2}}g' v_d & \frac{1}{\sqrt{2}}g' v_u\\
0 & M_2 & 0 & \frac{1}{\sqrt{2}}g v_d & -\frac{1}{\sqrt{2}}g v_u\\
0 & 0 & M_X & -2\sqrt{2} g'' v_d & 2\sqrt{2} g'' v_u \\
-\frac{1}{\sqrt{2}}g' v_d & \frac{1}{\sqrt{2}}g v_d & -2\sqrt{2} g'' v_d & 0 & -\mu\\
\frac{1}{\sqrt{2}}g' v_u & -\frac{1}{\sqrt{2}}g v_u & 2\sqrt{2} g'' v_u & -\mu &0
\end{pmatrix}
\end{equation}
The next block is the one involving terms mixing the usual neutralino states with matter states. As such, they include B-RPV masses that mix matter with higgses. The matrix reads
\begin{equation}
\mathbf{M}_{\chi\nu}^{5\times6} = \begin{pmatrix}
-\frac{1}{\sqrt{2}}g' N & -\frac{1}{\sqrt{2}}g' N_X & \frac{1}{\sqrt{2}}g' \overline N_X & 0 & 0 & 0\\
\frac{1}{\sqrt{2}}g N & \frac{1}{\sqrt{2}}g N_X & -\frac{1}{\sqrt{2}}g \overline N_X & 0 & 0 & 0\\
3\sqrt{2}g'' N & 3\sqrt{2}g'' N_X & -3\sqrt{2}g'' \overline N_X & -5\sqrt{2}g'' {N} & -5\sqrt{2}g'' {N}_X & 5\sqrt{2}g'' \overline{N}_{{X}} \\
0 & 0 & \kappa_{\overline{X}} & 0 & 0 &\epsilon_{\overline{X}}\\
\kappa_m & \kappa_X & 0 & \epsilon_m & \epsilon_X & 0
\end{pmatrix}
\end{equation}
where, in order to de-clutter notation, we are taking the fields names as to represent the VEVs. We notice that the B-RPV couplings $\kappa$ and $\epsilon$ are superpotential terms, while the top three rows is generated by kinetic terms only.
The lower-right $6\times 6$ block is purely from the superpotential, and includes only the masses involving $\nu$-type and/or $N$-type fermions. To obtain the mass, one performs the usual SUSY rule for fermionic masses
\begin{equation}
\left(\mathbf{M}_{\nu}^{6\times6}\right)_{ij} = -\frac{1}{2}\frac{\partial^2}{\partial \psi_i \partial \psi_j} W_{total}
\end{equation}
where $i$, $j=\{\nu, \nu_X, \overline \nu_X, N, N_X, \overline N_X\}$.
This $6 \times 6$ matrix has three main blocks: the $\nu \nu$ block, $\nu N$ block, and $N N$ block. Schematically they are arranged, in our basis, as
\begin{equation}
\mathbf{M}_{\nu}^{6\times6} = -\frac{1}{2}
\left(
\begin{array}{c|c}
M_{\nu\nu} & M_{\nu N} \\
\hline
M_{\nu N}^T & M_{NN}
\end{array}
\right)
\end{equation}
The actual form of the matrix is obtained using the full superpotential in Eq. \eqref{eq:Wtotal}. Doing so, one gets the following sub-blocks. First we have the $\nu\nu$ block that has mixing between $\overline \nu_X$ and $\nu$, $\nu_X$. In the sub-basis $(\nu, \nu_X, \overline \nu_X)$ this reads
\begin{equation}\label{eq:Mnunu}
M_{\nu\nu}=
\begin{pmatrix}
0 & 0 & \frac{b_7 \overline N_X N}{m_{Pl}}+\frac{b_{10} \overline N_X N_X}{m_{Pl}}+\mu^L_{Xm} \\
& 0 & \frac{b_8 \overline N_X N}{m_{Pl}}+\frac{b_{11} \overline N_X N_X}{m_{Pl}}+\mu^L_X \\
& & 0
\end{pmatrix}
\end{equation}
where we dropped the terms $ \nu^2/m_{Pl}$, $v_{u/d}^2/m_{Pl}$ as they are irrelevant and to de-clutter, and since this block is symmetric we omit the lower left triangular part. But notice that the terms with coefficients $b_7$, $b_8$, $b_{10}$, $b_{11}$ can play an important role as they can generate heavy Dirac masses, depending on the KM mechanism.
Next we have the $\nu N$ block, where one can find the neutrino Dirac masses generated by the Higgses VEV at the EWS. Taking the rows to be along the basis $(\nu, \nu_X, \overline \nu_X)$, while the columns along $(N, N_X, \overline N_X)$, this block reads
\begin{equation}
M_{\nu N}=
\begin{pmatrix}
v_u y_\nu+\frac{b_7 \overline N_X \overline \nu_X}{m_{Pl}} & \frac{b_{10} \overline N_X \overline \nu_X}{m_{Pl}} & \frac{b_7 N \overline \nu_X}{m_{Pl}}+\frac{b_{10} N_X \overline \nu_X}{m_{Pl}} \\
\frac{b_8 \overline N_X \overline \nu_X}{m_{Pl}} & \frac{b_{11} \overline N_X \overline \nu_X}{m_{Pl}} & \frac{b_8 N \overline \nu_X}{m_{Pl}}+\frac{b_{11} N_X \overline \nu_X}{m_{Pl}} \\
\frac{b_7 \overline N_X \nu }{m_{Pl}}+\frac{b_8 \overline N_X \nu_X}{m_{Pl}} & \frac{b_{10} \overline N_X \nu }{m_{Pl}}+\frac{b_{11} \overline N_X \nu_X}{m_{Pl}} & \frac{b_7 N \nu }{m_{Pl}}+\frac{b_{10} N_X \nu }{m_{Pl}}+\frac{b_8 N \nu_X}{m_{Pl}}+\frac{b_{11} N_X
\nu_X}{m_{Pl}}
\end{pmatrix}
\end{equation}
where we dropped the sub-leading terms $v_{u/d} \lambda \simeq \mathcal{O}(10^{-12})$ GeV.
Finally we have the $N N$ block, that involves Dirac and Majorana masses generated through the first two terms in Equation \eqref{eq:Wnonren}. Ignoring the terms generated by Higgses and sneutrino VEVs, in the sub-basis $(N, N_X, \overline N_X)$ this block reads
{\tiny
\begin{align}
&M_{N N} = \\
& \begin{pmatrix}
\frac{c_{n,k} (k-1) k }{ m_{Pl}^{2n-3}} \overline N_X^{n} N_X^{n-k} N^{k-2}+\frac{2 c_{2,2}}{m_{Pl}} \overline N_X^2 & \frac{ c_{n,k} k (n-k) }{ m_{Pl}^{2n-3}} \overline N_X^{n} N^{k-1} N_X^{-k+n-1} & \frac{c_{n,k} k n}{m_{Pl}^{2n-3}} \overline N_X^{n-1} N_X^{n-k} N^{k-1}+\mu^N_{Xm}+\frac{4 c_{2,2}}{m_{Pl}} \overline N_X N \\
& \frac{ c_{n,k} (-k+n-1) (n-k)} {m_{Pl}^{2n-3}} \overline N_X^{n} N^k N_X^{n-k-2} & \frac{c_{n,k}n (n-k)}{m_{Pl}^{2n-3}} \overline N_X^{n-1} N_X^{-k+n-1} N^k+\mu^N_{XX} \\
& & \frac{c_{n,k}(n-1) n}{m_{Pl}^{2n-3}} \overline N_X^{n-2} N_X^{n-k} N^k+\frac{2 c_{2,2}}{m_{Pl}}N^2
\end{pmatrix} \nonumber
\end{align} }
where the orders of magnitude of each entry will largely depend on which KM scenario is being considered. The matrix is symmetric so only the upper diagonal entries are displayed.
\subsection{The mass matrix hierarchies}\label{sec:Mhierarchies}
Following the description of the mass matrix above, we will now try to infer the hierarchies between the entries of the matrix. First we notice that, regardless of the case (i.e. the allowed
Kolda-Martin operators), the biggest entry in the mass matrix is always in the Gaugino-$N$ mixing block. \footnote{The caveat to this statement is if we allow for an order 1 Neutrino Yukawa, in that case the $\kappa$ entry originated from $ y_\nu \langle N \rangle L H_u $, will have the same order of magnitude. But since the B-RPV coupling above does not involve $\tilde B_X$, $N$, $N_X$, or $\overline{N}_X$, the magnitude of this coupling does not change the following discussion. We will return to B-RPV couplings further below.} This result is understandable as we expect the breaking of the extra $U(1)_X$ to transform a chiral superfield and a massless vector superfield into a single massive vector superfield. The degrees of freedom add up correctly, and would mean that below the $U(1)_X$ breaking scale we can take $\tilde B_X$ and the linear combination of $N$-states that break the $U(1)_X$ to be integrated out jointly. The linear combination that breaks the extra $U(1)_X$ depends on the exact values of the VEVs, but we can highlight some characteristics and how the mass-matrix will look like after this is integrated out.
In order to single out the correct liner combination that breaks the extra $U(1)_X$, one can perform a rotation in the last three states -- $N$, $N_X$, $\overline N_X$ -- in order to retain only one mixing mass between these states and the $\tilde B_X$. In order to do so, in the limit the mass matrix is real, the rotation is
\[ U =
\begin{pmatrix}
1 & 0 & & & & \\
0 & 1 & \cdots& & & \\
& \vdots & \ddots & & & \\
& & &\cos (\theta ) & - \sin (\theta) \cos (\phi ) & \sin
(\theta ) \sin (\phi ) \\
& & &\sin (\theta ) & \cos (\theta ) \cos (\phi ) & -\cos
(\theta ) \sin (\phi ) \\
& & &0 & \sin (\phi ) & \cos (\phi )
\end{pmatrix}
\]
where the angles are determined by the strength of the mixing mass parameters. For instance, in the $n=2,k =0$ Kolda-Martin mechanism presented before, the VEVs of the scalar components of $N$, $N_X$, $\overline N_X$ are all of same order. In such case, taking $\theta \simeq 3 \pi/4$ and $\phi \simeq \arctan \sqrt{2}$ will leave only one state mixing with $\tilde B_X$. For the other Kolda-Martin implementations, the $N_X$, $\overline N_X$ VEVs are much larger than $N$ VEV and so we can take $\theta \simeq 0$ with $\phi \simeq 3\pi/4$ to accomplish the same.
The rotation above affects only the last three columns and rows. Since the matrix is unitary (orthogonal in the case the masses are real), the entries of last three columns of a given row will be mixed with at-most order 1 coefficients, and whilst there might be cancellations there will be no order of magnitude enhancements.
Once the rotation is performed one can then integrate out $\tilde B_X$ jointly with its Dirac partner. This in turn will affect all the remainder of the matrix. For example, the entry $i,j$ will receive a contribution from integrating out a Dirac mass at position $a,b$ of order
\[
- \frac{M_{i3}M_{bj}}{M_{3b}}
\]
with some order one coefficients from the rotation. In this case we are setting one of the indices to 3 as this is the position of $\tilde B_X$ in our basis. The remaining index, $b$, refers to the position of the linear combination that breaks the extra $U(1)_X$. If, for example, the breaking linear combination that breaks the extra $U(1)_X$ is mostly composed of $N_X$, $\overline N_X$ states, the main contribution to the $\nu$ Majorana mass is given by
\[
\frac{b_{10}}{m_{Pl}}\langle \nu \rangle \langle \overline \nu_X \rangle \ll 10^{-10} \mbox{ GeV}
\]
even if we let the respective coupling on, i.e. $b_{10} \simeq \mathcal{O}(1)$.
Therefore, after the above rotation and integrating out , the mass matrix remains schematically the same, but with the absence of $\tilde B_X$ and a linear combination composed of $N$, $N_X$, $\overline N_X$.
After integrating out the Dirac fermion originated by the breaking, one can see that the Majorana and Dirac masses -- generated at the $U(1)_X$ breaking scale -- involving only the surviving terms of the $N$, $N_X$, $\overline N_X$ system are the leading entries of the mass matrix. These are present in the bottom-right-most $2 \times 2$ block. These states will then be responsible for a type of see-saw mechanism involving the lighter $SU(2)$ doublet states $\nu$, $\nu_X$, $\overline \nu_X$, with EW scale Dirac mass terms. In order to make sense of this see-saw mechanism, the $\nu$-states need to be protected from too much mixing with the remaining gauginos and higgsinos, such that the lightest mass
eigenstate is dominantly composed of $\nu$. Actually
the mixing between the $\nu$-type states with gauginos is negligible since it is generated by $\nu$-type VEVs and are therefore sub-EWS. But the mixing with Higgsinos is parametrically dependent on $N$-type VEVs through B-RPV terms, the so called $\kappa$ mass parameters.
The $\kappa$ parameters defined in Equations \eqref{eq:kappam}, \eqref{eq:kappax} and \eqref{eq:kappax} can have other potentially undesirable consequences
as they can spoil Higgs physics. Take for example the matter B-RPV interaction, with
$\kappa_m$ significantly larger than any other mass involving $H_u$. If were to happens, then
$L$ and $H_u$ superfields would pair up to produce a heavy vector-like pair. Then $H_u$ would be much heavier than the EWS physics and would spoil Higgs physics, where $H_u$ and $H_d$ are identified as a vector-like pair. In order to preserve viable Higgs physics, we need all $\kappa$-parameters to be much smaller than the remaining masses appearing in the Higgs potential.
Finally, there is risk that $\nu$, $\nu_X$, $\overline \nu_X$ states will mix with each other too much. To see this consider the $3\times 3$ sub-block of the matrix as shown in Eq. \eqref{eq:Mnunu}. If all $b_i$ couplings are suppressed, this matrix will maximally mix $\nu$ and $\nu_X$ through the $\mu$-terms. But it is important to note that while most of the $b_i$ interactions will be generated by Higgses and $\nu$-type VEVs (making them naturally sub-leading even if they are allowed by discrete symmetry) there are two terms that can have important contributions
\begin{equation}
\frac{b_{10}}{m_{Pl}} N_X \overline N_X \nu \overline \nu_X , \ \frac{b_{11}}{m_{Pl}} N_X \overline N_X \nu_X \overline \nu_X,
\end{equation}
which for the KM cases can generate Dirac masses much greater than $\mu$-terms if the respective $b_i$ coefficients are unsuppressed. This can then provide a natural mechanism to split $\nu$ from $\nu_X$, $\overline \nu_X$, if the coupling $b_{10}$ is forbidden while $b_{11}$ is allowed. In this case, we define
\begin{equation}\label{eq:mu11}
\mu_{11} = \frac{b_{11}}{m_{Pl}} N_X \overline N_X
\end{equation}
and the leading entries for Eq. \eqref{eq:Mnunu} will take the form
\begin{equation}
\begin{pmatrix}
0 & 0 & \mu^L_{Xm} \\
0 & 0 & \mu_{11} \\
\mu^L_{Xm} & \mu_{11} & 0
\end{pmatrix} ,
\end{equation}
which will then lead to $\nu_X$, $\overline \nu_X$ to pair up and decouple from $\nu$.
\section{Numerical Results}\label{sec:numerical}
As the full mass matrix presents an intricate structure of relations and hierarchies between different states, it is ultimately impossible to obtain a simple and revealing analytic expression that describes how one should obtain good neutrino physics. Instead, we perform a numerical scan over space, ensuring that the above constraints are satisfied.
In so doing, we divided the analysis into different realisations of the Kolda-Martin mechanism, parametrised by different values of $(n,k)$, corresponding to the scenarios in
Table~\ref{tab:KMSUSYVacua}.
In all the cases, we considered a point of the parameter space to be good if the mass of the lightest eigenstate of the mass matrix, identified as a physical neutrino, has a mass in the range
\begin{equation}
[50, 100] \mbox{ meV},
\end{equation}
and in addition that the corresponding eigenstate is mostly composed of
the left-handed doublet component $\nu$ (i.e. the state arising from $(\nu \ e)^T$).
In order to do so, we compute the decomposition of the eigenstate in the original basis
\begin{equation}
| \nu_{light} \rangle = \alpha | \nu \rangle + \dots
\end{equation}
and impose $\alpha$ to be the largest of the coefficients. As discussed in the previous section, the prevalence of $\nu$ as the largest component of $\nu_{light}$ will depend greatly on the parameters of the mass matrix that mix different states, i.e. Dirac masses.
For definiteness, we shall also require that the second lightest mass eigenstate (essentially the lightest
non-neutrino-like neutralino) to be at least $100$ GeV.
For each example, we only allow the particular desired Kolda-Martin operator while preventing all tree-level Yukawas involving states of the extra vector-like family. Furthermore, unless otherwise stated we assume that all quadratic terms in Eq. \eqref{eq:Wnonren} involving large VEVs are turned off. As expected within the $M$ Theory framework, the disallowed tree-level couplings are regenerated through moduli VEVs, and so the respective coupling strength was set to be of $\mathcal{O}(10^{-15})$. Along the same line, the $\mu$-terms generated by moduli VEVs were set to $\mathcal{O}(1)$ TeV.
Below we will show our findings for the only promising cases, which are $(n,k)=(2,0),\ (2,1),\ (3,0)$. The other $(n,k)$ assignments either returned to little points or no viable correlation to enhance $\alpha$. This happens as for the $(3,1)$, $(4,0)$ cases, since $N_X, \ \overline N_X \simeq 10^{15.5}$ GeV, the B-RPV coupling is generically greater than $1$ GeV. As we will see below, the only viable regions of the parameter space coincide with a naturally suppressed B-RPV parameter.
\subsection{$\nu$ component of the lightest state}
From the discussion above, we expect the value of $\alpha$ to be correlated with some parameters of the theory. Namely, we expect $\alpha$ to be enhanced if $b_{11}$ is not suppressed and if the B-RPV coupling $\kappa_m$ is much smaller than any other mass involving Higgsinos.
Since any disallowed tree-level coupling can be regenerated through moduli VEVs with a $\lambda \simeq 10^{-15}$ suppression, we started our numerical study by looking at the behaviour of $\alpha$ as we let $b_{11}$ vary in the range
\begin{equation}
b_{11} \in [10^{-15},1] ,
\end{equation}
which, in conjugation with a non-vanishing $N_X$, $\overline N_X$ VEVs will lead to non-vanishing $\mu_{11}$ as defined in Eq. \eqref{eq:mu11}.
In order to assess the strength of the B-RPV term, $\kappa_m$, allowed in the regions of the parameter space that return good neutrinos, we also registered the value of $\kappa_m$ at each point which returned the mass inside the bounds stated.
\vspace{10mm}
\underline{\bf $(2,0)$ and $(2,1)$ cases}
\vspace{5mm}
For these two Kolda-Martin implementation cases, the three $(N,N_X,\overline N_X)$ VEVs are all of order $\mathcal{O}(10^{10.5})$ GeV. As such, we allowed these VEVs to take values around
\begin{equation}
N,\ N_X,\ \overline N_X \in [10^{9.5},10^{11.5}]\mbox{ GeV}
\end{equation}
to cover the range of expected values. Since with these values the mass matrix is very similar for both $(2,0)$ and $(2,1)$ cases, we present them together.
As a consequence of the values of the VEVs above, the $\mu_{11}$ Dirac mass between $\nu_X$, $\overline \nu_X$ , defined in Eq. \eqref{eq:mu11}, will take values spanning
\begin{equation}
\mu_{11} = b_{11} \frac{N_X \overline N_X}{m_{Pl}} = b_{11} [10,10^5]\mbox{ GeV}
\end{equation}
which means that, only for non-suppressed $b_{11}$ we expect
\begin{equation}
\mu_{11} > \mu^L_{Xm}
\end{equation}
as required to split $\nu$ from $\nu_X$, as discussed in Section \ref{sec:Mhierarchies}.
The above considerations indicate us that the mechanism to split $\nu$ from $\nu_X$ will only work for large values of $b_{11}$. This can be seen in Figures \ref{fig:(2,0)-Scatter-(alpha,b11)} and \ref{fig:(2,1)-Scatter-(alpha,b11)}, where a slight agglomeration of points around $(\alpha,b_{11})\simeq (1,1)$ can be identified.
\begin{figure}[h]
\centering
\subfigure[$(2,0)$ case\label{fig:(2,0)-Scatter-(alpha,b11)}]{\includegraphics[width=0.45\linewidth]{2,0-Scatter-alpha,b11.pdf}}
\subfigure[$(2,0)$ case\label{fig:(2,1)-Scatter-(alpha,b11)}]{\includegraphics[width=0.45\linewidth]{2,1-Scatter-alpha,b11.pdf}}
\caption{Scatter plots showing the amplitude $\alpha$ of the left-handed doublet state
$\nu$ in the lightest mass eigenstate
$\nu_{light}$ as $b_{11}$ varies for the $(2,0)$ and $(2,1)$ cases.
The points are fairly evenly distributed with a slight clustering near the desired value of $\alpha \approx 1$ for $b_{11}\approx 1$.}
\end{figure}
On the other hand, we find that the $\kappa_m$ parameter is mostly bounded to be smaller than $1$ GeV, as is shown in Figures \ref{fig:(2,0)-Scatter-(alpha,kappam)} and \ref{fig:(2,1)-Scatter-(alpha,kappam)}. Although such small values of $\kappa_m$ are welcome, the fact that there is no clear preference for $\kappa_m \gtrsim 10^{-2}$ GeV suggests this class of models is challenged by BBN constraints, c.f. Eq. \eqref{eq:kLSPBBN}.
\begin{figure}[h]
\centering
\subfigure[$(2,0)$ case\label{fig:(2,0)-Scatter-(alpha,kappam)}]{\includegraphics[width=0.45\linewidth]{2,0-Scatter-alpha,kappam.pdf}}
\subfigure[$(2,1)$ case\label{fig:(2,1)-Scatter-(alpha,kappam)}]{\includegraphics[width=0.45\linewidth]{2,1-Scatter-alpha,kappam.pdf}}
\caption{Scatter plots showing the amplitude $\alpha$ of the left-handed doublet state
$\nu$ in the lightest mass eigenstate
$\nu_{light}$ as $\kappa_m$ varies for the $(2,0)$ and $(2,1)$ cases.
The points are fairly evenly distributed with a slight clustering near the desired value of $\alpha \approx 1$. The horizontal dashed line represents the bound on the LSP lifetime, c.f. Eq. \eqref{eq:kLSPBBN}.}
\end{figure}
\vspace{10mm}
\underline{\bf $(3,0)$ case}
\vspace{5mm}
For the $(3,0)$ Kolda-Martin realisation, we found much promising results. Since the $N_X$, $\overline N_X$ VEVs are expected to be around $\mathcal{O}(10^{14.25})$ GeV, if we allow them to be in the range
\begin{equation}
N_X,\ \overline N_X \in [10^{13.25},10^{15.25}] \mbox{ GeV}
\end{equation}
we find
\begin{equation}
\mu_{11} \in b_{11} [10^{8.25},10^{12.25}]\mbox{ GeV}
\end{equation}
which implies that it is natural to achieve
\begin{equation}
\mu_{11} \gg \mu^L_{Xm}
\end{equation}
and consequently $\nu$ will decouple easily from the other $\nu$-type states.
The above expectations are confirmed by the numerical results, and the lightest state will be mostly composed of $\nu$ even for values of $b_{11}$ below $\mathcal{O}(1)$. This behaviour can be seen in Figure \ref{fig:(3,0)-Scatter-(alpha,b11)}.
\begin{figure}[h]
\centering
\subfigure[Scatter of $(\alpha,b_{11})$ plane\label{fig:(3,0)-Scatter-(alpha,b11)}]{\includegraphics[width=0.45\linewidth]{3,0-Scatter-alpha,b11.pdf}}
\subfigure[Scatter of $(\alpha,\kappa_m)$ plane\label{fig:(3,0)-Scatter-(alpha,kappam)}]{\includegraphics[width=0.45\linewidth]{3,0-Scatter-alpha,kappam.pdf}}
\caption{Scatter plots showing the amplitude $\alpha$ of the left-handed doublet state
$\nu$ in the lightest mass eigenstate
$\nu_{light}$ as $\kappa_m$ varies for the $(3,0)$ case.
The points are fairly evenly distributed except for a significant clustering near the desired value of $\alpha \approx 1$ for larger values of $b_{11}$. The horizontal dashed line represents the bound on the LSP lifetime, c.f. Eq. \eqref{eq:kLSPBBN}. The right panel shows that nearly all the points
satisfy $\kappa_m \gtrsim 10^{-2}$ GeV.}
\end{figure}
Interestingly, in the $(\alpha,\kappa_m)$ plane, shown in Figure \ref{fig:(3,0)-Scatter-(alpha,kappam)} we can see again that the mass matrix prefers $\kappa_m < 1$ GeV in order to reproduce a mostly-$\nu$ lightest state. This is a nice result which ensures that whenever we have good physical neutrinos, we also find sufficiently suppressed B-RPV.
Furthermore, all the good points also suggest $\kappa_m \gtrsim 10^{-2}$ GeV,
satisfying the requirement for successful BBN physics, c.f. Eq. \eqref{eq:kLSPBBN}.
\subsection{Matter Neutrino Yukawas and B-RPV couplings}
From the above analysis we learned that for the $(2,0)$, $(2,1)$ and $(3,0)$ cases we expect a non-suppressed $b_{11}$ to enhance the component of $\nu$ in the lightest state. As such, we will now consider this coupling to be of order 1 and re-run the analysis for these cases, with the goal being to assess what typical values $\kappa_m$ and $y_\nu$ should take for a successful implementation of the proposed Kolda-Martin mechanism.
\vspace{10mm}
\underline{\bf $(2,0)$ and $(2,1)$ cases}
\vspace{5mm}
In Figures \ref{fig:(2,0)-Histogram-Ynu} and \ref{fig:(2,1)-Histogram-Ynu} we see that the preferred points are those with $y_\nu \lesssim 10^{-10}$. This suggests that for theses cases, the see-saw mechanism does not take a great role in explaining the light neutrino masses.
\begin{figure}[h]
\centering
\subfigure[$(2,0)$ case\label{fig:(2,0)-Histogram-Ynu}]{\includegraphics[width=0.45\linewidth]{2,0-Histogram-Ynu.pdf}}
\subfigure[$(2,1)$ case\label{fig:(2,1)-Histogram-Ynu}]{\includegraphics[width=0.45\linewidth]{2,1-Histogram-Ynu.pdf}}
\caption{Histograms for the values of $y_\nu$ for the $(2,0)$ and $(2,1)$ cases with unsuppressed $b_{11}$}
\end{figure}
In Figures \ref{fig:(2,0)-Histogram-Kappam} and \ref{fig:(2,1)-Histogram-Kappam} we see that for these cases, the B-RPV parameter $\kappa_m$ is naturally very small. This result is easy to understand, considering the main contribution to $\kappa_m$ to be
\[
\kappa_m \simeq y_\nu v_m,
\]
and given the range of values that we are allowing the VEVs to take, $\kappa_m$ is expected to be small. Unfortunately, all points returning good neutrino physics also return $\kappa_m > 10^{-2}$ GeV, which means that these classes of models spoil BBN, c.f. \eqref{eq:kLSPBBN}.
Although not shown here one can also find that $\kappa_X$, $\kappa_{\overline X}$ parameters, which mix $L_X$, $\overline L_X$ with $H_u$, $H_d$ respectively, are also constrained to be smaller than 1 $GeV$.
\begin{figure}[h]
\centering
\subfigure[Histogram for values of $\kappa_m$\label{fig:(2,0)-Histogram-Kappam}]{\includegraphics[width=0.45\linewidth]{2,0-Histogram-Kappam.pdf}}
\subfigure[Histogram for values of $\kappa_m$\label{fig:(2,1)-Histogram-Kappam}]{\includegraphics[width=0.45\linewidth]{2,1-Histogram-Kappam.pdf}}
\caption{Histograms for the values of $y_\nu$ and $\kappa_m$ for the $(2,1)$ case with unsuppressed $b_{11}$. The vertical dashed line represents the bound on the LSP lifetime, c.f. Eq. \eqref{eq:kLSPBBN}.}
\end{figure}
\vspace{10mm}
\underline{\bf $(3,0)$ case}
\vspace{5mm}
For this realisation of the Kolda-Martin mechanism, the results are slightly different but in line with our expectations. In Figure \ref{fig:(3,0)-Histogram-Ynu} we can see that the matter Yukawa coupling is allowed to take values larger than in the previous case. This indicates that the see-saw mechanism is having an effect on reducing the contribution of the matter neutrino Dirac mass to the lightest eigenstate.
\begin{figure}[h]
\centering
\subfigure[Histogram for values of $y_\nu$\label{fig:(3,0)-Histogram-Ynu}]{\includegraphics[width=0.45\linewidth]{3,0-Histogram-Ynu.pdf}}
\subfigure[Histogram for values of $\kappa_m$. The vertical dashed line represents the bound on the LSP lifetime, c.f. Eq. \eqref{eq:kLSPBBN}.\label{fig:(3,0)-Histogram-Kappam}]{\includegraphics[width=0.45\linewidth]{3,0-Histogram-Kappam}.pdf}
\caption{Histograms for the values of $y_\nu$ and $\kappa_m$ for the $(3,0)$ case with unsuppressed $b_{11}$}
\end{figure}
In Figure \ref{fig:(3,0)-Histogram-Kappam} we see that $\kappa_m$ is bound to be smaller than $1$ GeV. The fact that $\kappa_m$ takes larger values for $(3,0)$ case than for the $n=2$ cases is easily understandable. The main contributions to $\kappa_m$ are
\begin{equation}
\kappa_m \simeq y_\nu N + \lambda N_X
\end{equation}
where the VEVs are expected as in Table \ref{tab:KMSUSYVacua}. These contributions are in general greater than those in $n=2$ cases, but they are still bounded to be smaller than $1$ GeV. This is fortunate, as $\kappa_m \gtrsim 10^{-2}$ GeV and hence this class of models retain the successful predictions of BBN, c.f. \eqref{eq:kLSPBBN}.
As before, although not shown here also finds that $\kappa_X$, $\kappa_{\overline X}$ parameter are also constrained to be smaller than 1 $GeV$.
\section{Conclusions and Discussion} \label{sec:conclusion}
In this paper we have studied the origin of neutrino mass from $SO(10)$ SUSY GUTs arising from
$M$ Theory compactified on a $G_2$-manifold. We have seen that this problem is linked to
the problem of $U(1)_X$ gauge symmetry breaking, which appears
in the $SU(5)\times U(1)_X$ subgroup of $SO(10)$, and remains unbroken by the Abelian Wilson
line breaking mechanism. In order to break the $U(1)_X$ gauge symmetry, we considered a
(generalised) Kolda-Martin mechanism.
Our results show that it is possible to break the $U(1)_X$ gauge symmetry
without further SUSY breaking while achieving high-scale VEVs that play a crucial role in
achieving the desired value of neutrino mass.
The subsequent induced R-parity violation provides an additional source
of neutrino mass, in addition to that arising from the seesaw mechanism from non-renormalisable terms. The resulting $11\times 11$ neutrino mass
matrix was analysed for one neutrino family and it was shown how a
phenomenologically acceptable neutrino mass can emerge. This happens easily for the $(n,k)=(3,0)$ case of the Kolda-Martin mechanism we developed. For this class of models, not only is the neutrino masses phenomenologically viable, but also the physical light neutrino eigenstate is almost entirely composed of the left-handed (weakly charged) state $\nu$ in the same doublet as the electron $(\nu, e)$, as desired.
Furthermore, our analysis showed that the B-RPV parameters, which play an important role in neutrino masses and low-energy dynamics, are in the required range, being smaller than $1$ GeV. Finally, we notice that contrary to the $n=2$ cases, the $n=3$ type of Kolda-Martin mechanism immediately preserves the successful predictions of BBN by allowing the LSP to decay quickly in early universe.
In conclusion, we have shown that $SO(10)$ SUSY GUTs from $M$ Theory on $G_2$ manifolds
provides a phenomenologically viable framework, in which the rank can be broken in the effective theory
below the compactification scale, leading to acceptable values of neutrino mass,
arising from a combination of the seesaw mechanism
and induced R-parity breaking contributions.
In principle the mechanism presented here could be extended to three neutrino families and eventually
could be incorporated into a complete theory of flavour, based on $M$ Theory $SO(10)$,
however such questions are beyond the scope of the present paper.
\section*{Acknowledgements}
The work of BSA is supported by UK STFC via the research grant ST/J002798/1.
SFK acknowledges support from the STFC Consolidated grant ST/L000296/1 and the
European Union Horizon 2020 research and innovation programme under the Marie
Sklodowska-Curie grant agreements InvisiblesPlus RISE No. 690575 and
Elusives ITN No. 674896. MCR acknowledges support from the FCT under
the grant SFRH/BD/84234/2012. CP is supported by the KCL NMS graduate school and ICTP Trieste.
The work of KB is supported by a KCL GTA studentship.
\bibliographystyle{JHEP}
|
\section{Introduction}
There have been hundreds or thousands of papers written about word embeddings and their applications, from Web search \cite{NMCC16} to parsing Curriculum Vitae \cite{resumes2015}. However, none of these papers have recognized how blatantly sexist the embeddings are and hence risk introducing biases of various types into real-world systems.
A word embedding that represent each word (or common phrase) $w$ as a $d$-dimensional {\em word vector} $\vec{w}\in \mathbb{R}^d$. Word embeddings, trained only on word co-occurrence in text corpora, serve as a dictionary of sorts for computer programs that would like to use word meaning. First, words with similar semantic meanings tend to have vectors that are close together. Second, the vector differences between words in embeddings have been shown to represent relationships between words ~\cite{rubenstein1965contextual, mikolov2013linguistic}. For example given an analogy puzzle, ``man is to king as woman is to $x$'' (denoted as \emph{man}:\emph{king} :: \emph{woman}:$x$), simple arithmetic of the embedding vectors finds that $x$=\emph{queen} is the best answer because:
\[
\overrightarrow{\text{man}}-\overrightarrow{\text{woman}}\approx \overrightarrow{\text{king}}-\overrightarrow{\text{queen}}
\]
Similarly, $x$=\emph{Japan} is returned for \emph{Paris}:\emph{France} ::
\emph{Tokyo}:$x$. It is surprising that a simple vector arithmetic can simultaneously capture a variety of relationships. It has also excited practitioners because such a tool could be useful across applications involving natural language. Indeed, they are being studied and used in a variety of downstream applications (e.g., document ranking~\cite{NMCC16}, sentiment analysis~\cite{IrsoyCardie14}, and question retrieval~\cite{LJBJTMM16}).
\begin{figure}
\begin{tabular}{lllc}
\multicolumn{3}{c}{\bf{Extreme \emph{she} occupations}}\\
{1. homemaker} & {2. nurse} & {3. receptionist} \\
{4. librarian} & {5. socialite} & {6. hairdresser} \\
{7. nanny} & {8. bookkeeper} & {9. stylist} \\
{10. housekeeper} & {11. interior designer} & {12. guidance counselor} \\[2ex]
\multicolumn{3}{c}{{\bf Extreme \emph{he} occupations}}\\
{1. maestro} & {2. skipper} & {3. protege} \\
{4. philosopher} & {5. captain} & {6. architect} \\
{7. financier} & {8. warrior} & {9. broadcaster} \\
{10. magician} & {11. figher pilot} & {12. boss} \\
\\\end{tabular}
\caption{\label{fig:occupation_words} The most extreme occupations as projected on to the {\em she$-$he} gender direction on g2vNEWS. Occupations such as {\em businesswoman}, where gender is suggested by the orthography, were excluded.}
\end{figure}
\begin{figure}
\begin{tabular}{lllc}
\multicolumn{3}{c}{\bf{Gender stereotype \emph{she}-\emph{he} analogies.}}\\
{sewing-carpentry} & {register-nurse-physician } & {housewife-shopkeeper } \\
{nurse-surgeon} & {interior designer-architect} & {softball-baseball } \\
{blond-burly} & {feminism-conservatism } & {cosmetics-pharmaceuticals} \\
{giggle-chuckle} & {vocalist-guitarist} & {petite-lanky } \\
{sassy-snappy} & {diva-superstar} & {charming-affable } \\
{volleyball-football} & {cupcakes-pizzas } & {hairdresser-barber } \\[2ex]
\multicolumn{3}{c}{{\bf Gender appropriate \emph{she}-\emph{he} analogies.}}\\
{queen-king} & {sister-brother} & {mother-father} \\
{waitress-waiter} & {ovarian cancer-prostate cancer} & {convent-monastery} \\
\\\end{tabular}
\caption{\textbf{Analogy examples}. Examples of automatically generated analogies for the pair \emph{she-he} using the procedure described in text. For example, the first analogy is interpreted as \emph{she}:\emph{sewing} :: \emph{he}:\emph{carpentry} in the original w2vNEWS embedding. Each automatically generated analogy is evaluated by 10 crowd-workers are to whether or not it reflects gender stereotype. Top: illustrative gender stereotypic analogies automatically generated from w2vNEWS, as rated by at least 5 of the 10 crowd-workers. Bottom: illustrative generated gender-appropriate analogies. }
\label{fig:analogy_examples}
\end{figure}
\begin{figure}
\begin{tabular}{lll}
{\bf {\em softball} extreme} &{\bf gender portion} &{\bf after debiasing} \\
1. pitcher &-1\% & 1. pitcher \\
2. bookkeeper &20\% &2. infielder \\
3. receptionist &67\% &3. major leaguer \\
4. registered nurse &29\% &4. bookkeeper \\
5. waitress &35\% &5. investigator \\ [2ex]
{\bf {\em football} extreme} &{\bf gender portion} &{\bf after debiasing} \\
1. footballer &2\% &1. footballer \\
2. businessman &31\% &2. cleric \\
3. pundit &10\% &3. vice chancellor \\
4. maestro &42\% &4. lecturer \\
5. cleric &2\% &5. midfielder \\
\end{tabular}
\caption{\textbf{Example of indirect bias}. The five most extreme occupations on the {\em softball-football} axis, which indirectly captures gender bias. For each occupation, the degree to which the association represents a gender bias is shown, as described in Section \ref{sec:indirect}.
\label{fig:occupation_associations2}
}
\end{figure}
However, the embeddings also pinpoint sexism implicit in text. For instance, it is also the case that:
\[
\overrightarrow{\text{man}}-\overrightarrow{\text{woman}}\approx \overrightarrow{\text{computer programmer}}-\overrightarrow{\text{homemaker}}.
\]
In other words, the same system that solved the above reasonable analogies will offensively answer ``man is to computer programmer as woman is to $x$'' with $x$=\textit{homemaker}. Similarly, it outputs that a {\em father} is to a {\em doctor} as a \textit{mother} is to a \textit{nurse}. The primary embedding studied in this paper is the popular publicly-available word2vec~\cite{MCCD13,MSCCD13} embedding trained on a corpus of Google News texts consisting of 3 million English words and terms into 300 dimensions, which we refer to here as the w2vNEWS. One might have hoped that the Google News embedding would exhibit little gender bias because many of its authors are professional journalists. We also analyze other publicly available embeddings trained via other algorithms and find similar biases.
In this paper, we will quantitatively demonstrate that word-embeddings contain biases in their geometry that reflect gender stereotypes present in broader society. Due to their wide-spread usage as basic features, word embeddings not only reflect such stereotypes but can also amplify them. This poses a significant risk and challenge for machine learning and its applications.
To illustrate bias amplification, consider bias present in the task of retrieving relevant web pages for a given query. In web search, one recent project has shown that, when carefully combined with existing approaches, word vectors have the potential to improve web page relevance results \cite{NMCC16}. As an example, suppose the search query is {\em cmu computer science phd student} for a computer science Ph.D.\ student at Carnegie Mellon University. Now, the directory\footnote{Graduate Research Assistants listed at \url{http://cs.cmu.edu/directory/csd}.} offers 127 nearly identical web pages for students --- these pages differ only in the names of the students. A word embedding's semantic knowledge can improve relevance by identifying, for examples, that the terms {\em graduate research assistant} and {\em phd student} are related. However, word embeddings also rank terms related to computer science closer to male names than female names (e.g., the embeddings give \emph{John}:\emph{computer programmer} :: \emph{Mary}:{\em homemaker}). The consequence is that, between two pages that differ only in the names \emph{Mary} and \emph{John}, the word embedding would influence the search engine to rank John's web page higher than Mary. In this hypothetical example, the usage of word embedding makes it even harder for women to be recognized as computer scientists and would contribute to widening the existing gender gap in computer science. While we focus on gender bias, specifically Female-Male (F-M) bias, the approach may be applied to other types of bias.
Uncovering gender stereotypes from text may seem like a trivial matter of counting pairs of words that occur together. However, such counts are often misleading \cite{gordon2013reporting}. For instance, the term {\em male nurse} is several times more frequent than {\em female nurse} (similarly {\em female quarterback} is many times more frequent than {\em male quarterback}). Hence, extracting associations from text, F-M or otherwise, is not simple, and ``first-order'' approaches would predict that the word {\em nurse} is more male than {\em quarterback}. More generally, Gordon and Van Durme show how {\em reporting bias} \cite{gordon2013reporting}, including the fact that common assumptions are often left unsaid, poses a challenge to extracting knowledge from raw text. Nonetheless, $\wvec{nurse}$ is closer to $\wvec{female}$ than $\wvec{male}$, suggesting that word embeddings may be capable of circumventing reporting bias in some cases. This happens because word embeddings are trained using second-order methods which require large amounts of data to extract associations and relationships about words.
The analogies generated from these embeddings spell out the bias implicit in the data on which they were trained. Hence, word embeddings may serve as a means to extract implicit gender associations from a large text corpus similar to how Implicit Association Tests \cite{greenwald1998measuring} detect automatic gender associations possessed by people, which often do not align with self reports.
To quantify bias, we compare a word embedding to the embeddings of a pair of gender-specific words. For instance, the fact that $\wvec{nurse}$ is close to $\wvec{woman}$ is not in itself necessarily biased (it is also somewhat close to $\wvec{man}$ -- all are humans), but the fact that these distances are unequal suggests bias. To make this rigorous, consider the distinction between {\em gender specific} words that are associated with a gender by definition, and the remaining {\em gender neutral} words. Standard examples of gender specific words include \emph{brother}, \emph{sister}, {\em businessman} and {\em businesswoman}.
The fact that $\wvec{brother}$ is closer to $\wvec{man}$ than to $\wvec{woman}$ is expected since they share the definitive feature of relating to males. We will use the gender specific words to learn a gender subspace in the embedding, and our debiasing algorithm removes the bias only from the gender neutral words while respecting the definitions of these gender specific words.
We refer to this type of bias, where there is an association between a gender neutral word and a clear gender pair as {\em direct bias}. We also consider a notion of {\em indirect bias},\footnote{The terminology indirect bias follows Pedreshi et al.~\cite{pedreshi2008discrimination} who distinguish {\em direct} versus {\em indirect} discrimination in rules of fair {\em classifiers}. Direct discrimination involves directly using sensitive features such as gender or race, whereas indirect discrimination involves using correlates that are not inherently based on sensitive features but that, intentionally or unintentionally, lead to disproportionate treatment nonetheless.} which manifests as associations between gender neutral words that are clearly arising from gender. For instance, the fact that the word {\em receptionist} is much closer to {\em softball} than {\em football} may arise from female associations with both {\em receptionist} and {\em softball}. Note that many pairs of male-biased (or female-biased) words have legitimate associations having nothing to do with gender. For instance, while the words {\em mathematician} and {\em geometry} both have a strong male bias, their similarity is justified by factors other than gender. More often than not, associations are combinations of gender and other factors that can be difficult to disentangle. Nonetheless, we can use the geometry of the word embedding to determine the degree to which those associations are based on gender.
\smallskip
\noindent
{\bf Aligning biases with stereotypes.} Stereotypes are biases that are widely held among a group of people. We show that the biases in the word embedding are in fact closely aligned with social conception of gender stereotype, as evaluated by U.S.-based crowd workers on Amazon's Mechanical Turk.\footnote{\url{http://mturk.com}} The crowd agreed that the biases reflected both in the location of vectors (e.g. $\wvec{doctor}$ closer to $\wvec{man}$ than to $\wvec{woman}$) as well as in analogies (e.g., \emph{he}:\emph{coward} :: \emph{she}:\emph{whore}) exhibit common gender stereotypes.
\smallskip
\noindent
{\bf Debiasing.}
Our goal is to reduce gender biases in the word embedding while preserving the useful properties of the embedding.
Surprisingly, not only does the embedding capture bias, but it also contains sufficient information to reduce this bias, as illustrated in \ref{fig:words}. We will leverage the fact that there exists a low dimensional subspace in the embedding that empirically captures much of the gender bias.
The goals of debiasing are:
\begin{enumerate}
\item Reduce bias:
\begin{enumerate}
\item Ensure that gender neutral words such as {\em nurse} are equidistant between gender pairs such as {\em he} and {\em she}.
\item Reduce gender associations that pervade the embedding even among gender neutral words.
\end{enumerate}
\item Maintain embedding utility:
\begin{enumerate}
\item Maintain meaningful non-gender-related associations between gender neutral words, including associations within stereotypical categories of words such as fashion-related words or words associated with football.
\item Correctly maintain definitional gender associations such as between {\em man} and {\em father}.
\end{enumerate}
\end{enumerate}
\paragraph{Paper outline.}
After discussing related literature, we give preliminaries necessary for understanding the paper in Section \ref{sec:defs}. Next we propose methods to identify the gender bias of an embedding and show that w2vNEWS exhibits bias which is aligned with common gender stereotypes (Section~\ref{sec:empiricalbias}). In Section~\ref{sec:bias}, we define several simple geometric properties associated with bias, and in particular discuss how to identify the gender subspace. Using these geometric properties, we introduce debiasing algorithms (Section~\ref{sec:debias}) and demonstrate their performance (Section~\ref{sec:debiasing_results}). Finally we conclude with additional discussions of related literature, other types of biases in the embedding and future works.
\section{Related work}
Related work can be divided into relevant literature on bias in language and bias in algorithms.
\subsection{Gender bias and stereotype in English}
It is important to quantify and understand bias in languages as such biases can reinforce the psychological status of different groups \cite{sapir1985selected}.
Gender bias in language has been studied over a number of decades in a variety of contexts (see, e.g., \cite{holmes2008handbook}) and we only highlight some of the findings here. Biases differ across people though commonalities can be detected. Implicit Association Tests \cite{greenwald1998measuring} have uncovered gender-word biases that people do not self-report and may not even be aware of. Common biases link female terms with liberal arts and family and male terms with science and careers \cite{nosek2002harvesting}. Bias is seen in word morphology, i.e., the fact that words such as {\em actor} are, by default, associated with the dominant class \cite{jakobson1990language}, and female versions of these words, e.g., {\em actress}, are marked. There is also an imbalance in the number of words with F-M with various associations. For instance, while there are more words referring to males, there are many more words that sexualize females than males \cite{stanley1977paradigmatic}.
Glick and Fiske \cite{glick1996ambivalent} introduce the notion of {\em benevolent sexism} in which women are perceived with positive traits such as helpful or intimacy-seeking. Despite its seemingly positive nature, benevolent sexism can be harmful, insulting, and discriminatory. In terms of words, female gender associations with any word, even a subjectively positive word such as {\em attractive}, can cause discrimination against women if it reduces their association with other words, such as {\em professional}.
Stereotypes, as mentioned, are biases that are widely held within a group. While gender bias of any kind is concerning, stereotypes are often easier to study due to their consistent nature. Stereotypes have commonalities across cultures, though there is some variation between cultures \cite{cuddy2015men}. {\em Complimentary stereotypes} are common between females and males, in which each gender is associated with strengths that are perceived to offset its own weaknesses and compliment the strengths of the other gender \cite{jost2005exposure}. These and compensatory stereotypes are used by people to justify the status quo.
Consistent biases have been studied within online contexts and specifically related to the contexts we study such as online news (e.g., \cite{ross2011women}), Web search (e.g., \cite{kay2015unequal}), and Wikipedia (e.g., \cite{wikipediaBias}). In Wikipedia, Wager et al.~\cite{wikipediaBias} found that, as suggested by prior work on gender bias in language \cite{finkbeiner2013}, articles about women more often emphasize their gender, their husbands and their husbands' jobs, and other topics discussed consistently less often than in articles about men. Regarding individual words, they find that certain words are predictive of gender, e.g., {\em husband} appears significantly more often in articles about women while {\em baseball} occurs more often in articles about men.
\subsection{Bias within algorithms}
A number of online systems have been shown to exhibit various biases, such as racial discrimination and gender bias in the ads presented to users \cite{sweeney2013discrimination,datta2015automated}. A recent study found that algorithms used to predict repeat offenders exhibit indirect racial biases \cite{angwin2016}. Different demographic and geographic groups also use different dialects and word-choices in social media \cite{diffusionlexical}. An implication of this effect is that language used by minority group might not be able to be processed by natural language tools that are trained on ``standard'' data-sets. Biases in the curation of machine learning data-sets have explored in \cite{Torralba12,annotationbias}.
Independent from our work, Schmidt \cite{schmidt2015} identified the bias present in word embeddings and proposed debiasing by entirely removing multiple gender dimensions, one for each gender pair. His goal and approach, similar but simpler than ours, was to entirely remove gender from the embedding. There is also an intense research agenda focused on improving the quality of word embeddings from different angles (e.g., \cite{levyGoldberg14,pennington2014glove,yogatamalearning,faruqui2015retrofitting}), and the difficulty of evaluating embedding quality (as compared to supervised learning) parallels the difficulty of defining bias in an embedding.
Within machine learning, a body of notable work has focused on ``fair'' binary classification in particular. A definition of fairness based on legal traditions is presented by Barocas and Selbst \cite{barocas2014big}. Approaches to modify classification algorithms to define and achieve various notions of fairness have been described in a number of works, see, e.g., \cite{barocas2014big, dwork2012fairness,feldman2015certifying} and a recent survey \cite{zliobaite2015survey}.
Feldman et al.~\cite{feldman2015certifying} distinguish classification algorithms that achieve fairness by modifying the underlying data from those that achieve fairness by modifying the classification algorithm. Our approach is more similar to the former. However, it is unclear how to apply any of these previous approaches without a clear classification task in hand, and the problem is exacerbated by indirect bias.
This prior work on algorithmic fairness is largely for supervised learning. Fair classification is defined based on the fact that algorithms were classifying a set of individuals using a set of features with a distinguished sensitive feature. In word embeddings, there are no clear individuals and no a priori defined classification problem. However, similar issues arise, such as direct and indirect bias~\cite{pedreshi2008discrimination}.
\section{Preliminaries}\label{sec:defs}
We first very briefly define an embedding and some terminology. An embedding consists of a unit vector $\vec{w}\in \mathbb{R}^d$, with $\|\vec{w}\|=1$, for each word (or term) $w\in W$. We assume there is a set of gender neutral words $N \subset W$, such as {\em flight attendant} or {\em shoes}, which, by definition, are not specific to any gender. We denote the size of a set $S$ by $|S|$. We also assume we are given a set of F-M gender pairs $P\subset W \times W$, such as
{\em she-he} or {\em mother-father} whose definitions differ mainly in gender. Section \ref{sec:gender_neutral} discusses how $N$ and $P$ can be found within the embedding itself, but until then we take them as given.
As is common, {\em similarity} between words $w_1$ and $w_2$ is measured by their inner product, $\vec{w}_1 \cdot \vec{w}_2.$ Finally, we will abuse terminology and refer to the embedding of a word and the word interchangeably. For example, the statement {\em cat} is more similar to {\em dog} than to {\em cow} means $\wvec{cat}\cdot \wvec{dog} \geq \wvec{cat}\cdot \wvec{cow}$. For arbitrary vectors $u$ and $v$, define:
$$\cos(u,v) = \frac{u \cdot v}{\|u\|\|v\|}.$$
This normalized similarity between vectors $u$ and $v$ is written as $\cos$ because it is the cosine of the angle between the two vectors.
Since words are normalized $\cos(\vec{w}_1,\vec{w}_2) = \vec{w}_1 \cdot \vec{w}_2$.
\smallskip
\noindent
{\bf Embedding.}
Unless otherwise stated, the embedding we refer to in this paper is the aforementioned w2vNEWS embedding, a $d=300$-dimensional word2vec~\cite{MCCD13,MSCCD13} embedding, which has proven to be immensely useful since it is high quality, publicly available, and easy to incorporate into any application. In particular, we downloaded the pre-trained embedding on the Google News corpus,\footnote{\url{https://code.google.com/archive/p/word2vec/}} and normalized each word to unit length as is common. Starting with the 50,000 most frequent words, we selected only lower-case words and phrases consisting of fewer than 20 lower-case characters (words with upper-case letters, digits, or punctuation were discarded). After this filtering, 26,377 words remained. While we focus on w2vNEWS, we show later that gender stereotypes are also present in other embedding data-sets.
\smallskip
\noindent
{\bf Crowd experiments.}
All human experiments were performed on the Amazon Mechanical Turk crowdsourcing platform. We selected for U.S.-based workers to maintain homogeneity and reproducibility to the extent possible with crowdsourcing. Two types of experiments were performed: ones where we solicited words from the crowd (to see if the embedding biases contain those of the crowd) and ones where we solicited ratings on words or analogies generated from our embedding (to see if the crowd's biases contain those from the embedding). These two types of experiments are analogous to experiments performed in rating results in information retrieval to evaluate precision and recall. When we speak of the majority of 10 crowd judgments, we mean those annotations made by 5 or more independent workers.
Since gender associations vary by culture and person, we ask for ratings of stereotypes rather than bias. In addition to possessing greater consistency than biases, people may feel more comfortable rating the stereotypes of their culture than discussing their own gender biases. The Appendix contains the questionnaires that were given to the crowd-workers to perform these tasks.
\section{Gender stereotypes in word embeddings}\label{sec:empiricalbias}
Our first task is to understand the biases present in the word-embedding (i.e. which words are closer to \emph{she} than to \emph{he}, etc.)
and the extent to which these geometric biases agree with human notion of gender stereotypes. We use two simple methods to approach this problem: 1) evaluate whether the embedding has stereotypes on occupation words and 2) evaluate whether the embedding produces analogies that are judged to reflect stereotypes by humans. The exploratory analysis of this section will motivate the more rigorous metrics used in the next two sections.
\paragraph{Occupational stereotypes.} Figure \ref{fig:occupation_words} lists the occupations that are closest to \emph{she} and to \emph{he} in the w2vNEWS embeddings. We asked the crowdworkers to evaluate whether an occupation is considered female-stereotypic, male-stereotypic, or neutral. Each occupation word was evaluated by ten crowd-workers as to whether or not it reflects gender stereotype. Hence, for each word we had a integer rating, on a scale of 0-10, of stereotypicality. The projection of the occupation words onto the \emph{she}-\emph{he} axis is strongly correlated with the stereotypicality estimates of these words (Spearman $\rho = 0.51$), suggesting that the geometric biases of embedding vectors is aligned with crowd judgment of gender stereotypes. We used occupation words here because they are easily interpretable by humans and often capture common gender stereotypes. Other word sets could be used for this task. Also note that we could have used other words, e.g. \emph{woman} and \emph{man}, as the gender-pair in the task. We chose \emph{she} and \emph{he} because they are frequent and do not have fewer alternative word senses (e.g., {\em man} can also refer to {\em mankind}).
We projected each of the occupations onto the \emph{she-he} direction in the w2vNEWS embedding as well as a different embedding generated by the GloVe algorithm on a web-crawl corpus \cite{pennington2014glove}. The results are highly consistent (Figure \ref{fig:other_embeddings}), suggesting that gender stereotypes is prevalent across different embeddings and is not an artifact of the particular training corpus or methodology of word2vec.
\begin{figure}
\includegraphics[width=2.7in]{cross_corpus_scatter.png}
\caption{\label{fig:other_embeddings}Comparing the bias of two different embeddings--the w2vNEWS and the GloVe web-crawl embedding. In each embedding, the occupation words are projected onto the \emph{she}-\emph{he} direction. Each dot corresponds to one occupation word; the gender bias of occupations is highly consistent across embeddings (Spearman $\rho = 0.81$). }
\end{figure}
\paragraph{Analogies exhibiting stereotypes.} Analogies are a useful way to both evaluate the quality of a word embedding and also its stereotypes. We first briefly describe how the embedding generate analogies and then discuss how we use analogies to quantify gender stereotype in the embedding. A more detailed discussion of our algorithm and prior analogy solvers is given in Appendix \ref{ap:analogies}.
In the standard analogy tasks, we are given three words, for example \emph{he, she, king}, and look for the 4th word to solve \emph{he} to \emph{king} is as \emph{she} to $x$. Here we modify the analogy task so that given two words, e.g. \emph{he, she}, we want to generate a pair of words, $x$ and $y$, such that \emph{he} to $x$ as \emph{she} to $y$ is a good analogy. This modification allows us to systematically generate pairs of words that the embedding believes it analogous to \emph{he, she} (or any other pair of seed words).
The input into our analogy generator is a seed pair of words $(a,b)$ determining a {\em seed direction} $\vec{a}-\vec{b}$ corresponding to the normalized difference between the two seed words. In the task below, we use $(a,b)=(\text{she}, \text{he})$. We then score all pairs of words $x, y$ by the following metric:
\begin{equation}
\label{eq:analogyS}
\mbox{S}_{(a,b)}(x, y) = \begin{cases}
\cos\left(\vec{a}-\vec{b}, \vec{x}-\vec{y}\right) & \text{if } \left\|\vec{x}-\vec{y}\right\| \leq \delta\\
0 & \text{otherwise}
\end{cases}
\end{equation}
where $\delta$ is a threshold for similarity. The intuition of the scoring metric is that we want a good analogy pair to be close to parallel to the seed direction while the two words are not too far apart in order to be semantically coherent. The parameter $\delta$ sets the threshold for semantic similarity. In all the experiments, we take $\delta=1$ as we find that this choice often works well in practice. Since all embeddings are normalized, this threshold corresponds to an angle $\leq \pi/3$, indicating that the two words are closer to each other than they are to the origin. In practice, it means that the two words forming the analogy are significantly closer together than two random embedding vectors.
Given the embedding and seed words, we output the top analogous pairs with the largest positive $S_{(a,b)}$ scores. To reduce redundancy, we do not output multiple analogies sharing the same word $x$.
Since analogies, stereotypes, and biases are heavily influenced by culture, we employed U.S. based crowd-workers to evaluate the analogies output by the analogy generating algorithm described above. For each analogy, we asked the workers two yes/no questions: (a) whether the pairing makes sense as an analogy, and (b) whether it reflects a gender stereotype. Every analogy is judged by 10 workers, and we used the number of workers that rated this pair as stereotyped to quantify the degree of bias of this analogy. Overall, 72 out of 150 analogies were rated as gender-appropriate by five or more crowd-workers, and 29 analogies were rated as exhibiting gender stereotype by five or more crowd-workers (Figure~\ref{fig:direct-result}). Examples of analogies generated from w2vNEWS that were rated as stereotypical are shown at the top of Figure~\ref{fig:analogy_examples}, and examples of analogies that make sense and are rated as gender-appropriate are shown at the bottom of Figure~\ref{fig:analogy_examples}. The full list of analogies and crowd ratings are in Appendix \ref{app:analogies}.
\paragraph{Indirect gender bias.} The direct bias analyzed above manifests in the relative similarities between gender-specific words and gender neutral words. Gender bias could also affect the relative geometry between gender neutral words themselves. To test this \emph{indirect} gender bias, we take pairs of words that are gender-neutral, for example \emph{softball} and \emph{football}. We project all the occupation words onto the $\wvec{softball} - \wvec{football}$ direction and looked at the extremes words, which are listed in Figure~\ref{fig:occupation_associations2}.
For instance, the fact that the words {\em bookkeeper} and {\em receptionist} are much closer to {\em softball} than {\em football} may result indirectly from female associations with {\em bookkeeper}, {\em receptionist} and {\em softball}. It's important to point out that
that many pairs of male-biased (or female-biased) words have legitimate associations having nothing to do with gender. For example, while both {\em footballer} and {\em football} have strong male biases, their similarity is justified by factors other than gender. In Section~\ref{sec:bias}, we define a metric to more rigorously quantify these indirect effects of gender bias.
\section{Geometry of Gender and Bias}\label{sec:bias}
In this section, we study the bias present in the embedding geometrically, identifying the gender direction and quantifying the bias independent of the extent to which it is aligned with the crowd bias. We develop metrics of direct and indirect bias that more rigorously quantify the observations of the previous section.
\subsection{Identifying the gender subspace}\label{sec:direction}
Language use is ``messy'' and therefore individual word pairs do not always behave as expected. For instance, the word {\em man} has several different usages: it may be used as an exclamation as in {\em oh man!} or to refer to people of either gender or as a verb, e.g., {\em man the station}. To more robustly estimate bias, we shall aggregate across multiple paired comparisons. By combining several directions, such as $\wvec{she}-\wvec{he}$ and $\wvec{woman} - \wvec{man}$, we identify a {\bf gender direction} $g\in \mathbb{R}^d$ that largely captures gender in the embedding. This direction helps us to quantify direct and indirect biases in words and associations.
\begin{figure}
\begin{tabular}{r@{\hskip -0.0in}lllr@{\hskip -0.0in}lll}
& & def. & stereo. & & & def. & stereo. \\
$\wvec{she}-$&$\wvec{he}$ & 92\% & 89\% & $\wvec{daughter}-$&$\wvec{son}$ & 93\% & 91\% \\
$\wvec{her}-$&$\wvec{his}$ & 84\% & 87\% & $\wvec{mother}-$&$\wvec{father}$ & 91\% & 85\% \\
$\wvec{woman}-$&$\wvec{man}$ & 90\% & 83\% & $\wvec{gal}-$&$\wvec{guy}$ & 85\% & 85\% \\
$\wvec{Mary}-$&$\wvec{John}$ & 75\% & 87\% & $\wvec{girl}-$&$\wvec{boy}$ & 90\% & 86\% \\
$\wvec{herself}-$&$\wvec{himself}$ & 93\% & 89\% & $\wvec{female}-$&$\wvec{male}$ & 84\% & 75\%
\end{tabular}
\caption{\label{fig:definitional_pairs}
Ten possible word pairs to define gender, ordered by word frequency, along with agreement with two sets of 100 words solicited from the crowd, one with definitional and and one with stereotypical gender associations. For each set of words, comprised of the most frequent 50 female and 50 male crowd suggestions, the accuracy is shown for the corresponding gender classifier based on which word is closer to a target word, e.g., the {\em she-he} classifier predicts a word is female if it is closer to {\em she} than {\em he}. With roughly 80-90\% accuracy, the gender pairs predict the gender of both stereotypes and definitionally gendered words solicited from the crowd. }
\end{figure}
In English as in many languages, there are numerous gender pair terms, and for each we can consider the difference between their embeddings. Before looking at the data, one might imagine that they all had roughly the same vector differences, as in the following caricature:
\begin{eqnarray*}
\wvec{grandmother} &=& \wvec{wise} + \wvec{gal}\\
\wvec{grandfather} &=& \wvec{wise} + \wvec{guy}\\
\wvec{grandmother} - \wvec{grandfather} &=& \wvec{gal} - \wvec{guy} = g
\end{eqnarray*}
However, gender pair differences are not parallel in practice, for multiple reasons. First, there are different biases associated with with different gender pairs. Second is polysemy, as mentioned, which in this case occurs due to the other use of {\em grandfather} as in {\em to grandfather a regulation}. Finally, randomness in the word counts in any finite sample will also lead to differences. Figure \ref{fig:definitional_pairs} illustrates ten possible gender pairs, $\bigl\{(x_i,y_i)\bigr\}_{i=1}^{10}$.
We experimentally verified that the pairs of vectors corresponding to these words do agree with the crowd concept of gender. On Amazon Mechanical Turk, we asked crowdworkers to generate two lists of words: one list corresponding to words that they think are gendered by definition ({\em waitress}, {\em menswear}) and a separate list corresponding to words that they believe captures gender stereotypes (e.g., {\em sewing}, {\em football}). From this we generated the most frequently suggested 50 male and 50 female words for each list to be used for a classification task. For each candidate pair, for example $\wvec{she}, \wvec{he}$, we say that it accurately classifies a crowd suggested female definition (or stereotype) word if that word vector is closer to $\wvec{she}$ than to $\wvec{he}$. Table~\ref{fig:definitional_pairs} reports the classification accuracy for definition and stereotype words for each gender pair. The accuracies are high, indicating that these pairs capture the intuitive notion of gender.
To identify the gender subspace, we took the ten gender pair difference vectors and computed its principal components (PCs). As Figure \ref{fig:PCA} shows, there is a single direction that explains the majority of variance in these vectors. The first eigenvalue is significantly larger than the rest. Note that, from the randomness in a finite sample of ten noisy vectors, one expects a decrease in eigenvalues. However, as also illustrated in \ref{fig:PCA}, the decrease one observes due to random sampling is much more gradual and uniform. Therefore we hypothesize that the top PC, denoted by the unit vector $g$, captures the gender subspace. In general, the gender subspace could be higher dimensional and all of our analysis and algorithms (described below) work with general subspaces.
\begin{figure}
\begin{tabular}{cc}
\includegraphics[width=2in]{pca.png} & \includegraphics[width=2in]{pca_random.png}
\end{tabular}
\caption{\label{fig:PCA} Left: the percentage of variance explained in the PCA of these vector differences (each difference normalized to be a unit vector). The top component explains significantly more variance than any other.
Right: for comparison, the corresponding percentages for random unit vectors (figure created by averaging over 1,000 draws of ten random unit vectors in 300 dimensions).
}
\end{figure}
\subsection{Direct bias}
To measure direct bias, we first identify words that should be gender-neutral for the application in question. How to generate this set of gender-neutral words is described in Section \ref{sec:gender_neutral}.
Given the gender neutral words, denoted by $N$, and the gender direction learned from above, $g$, we define the direct gender bias of an embedding to be
\[
\text{DirectBias}_c = \frac{1}{|N|} \sum_{w \in N} \left|\cos(\vec{w}, g) \right|^c
\]
where $c$ is a parameter that determines how \emph{strict} do we want to in measuring bias. If $c$ is 0, then $\left|\cos(\vec{w} - g) \right|^c = 0$ only if $\vec{w}$ has no overlap with $g$ and otherwise it is 1. Such strict measurement of bias might be desirable in settings such as the college admissions example from the Introduction, where it would be unacceptable for the embedding to introduce a slight preference for one candidate over another by gender. A more gradual bias would be setting $c = 1$. The presentation we have chosen favors simplicity -- it would be natural to extend our definitions to weight words by frequency. For example, in w2vNEWS, if we take $N$ to be the set of 327 occupations, then $\text{DirectBias}_1=0.08$, which confirms that many occupation words have substantial component along the gender direction.
\subsection{Indirect bias}\label{sec:indirect}
Unfortunately, the above definitions still do not capture indirect bias. To see this, imagine completely removing from the embedding both words in gender pairs (as well as words such as {\em beard} or {\em uterus} that are arguably gender-specific but which cannot be paired). There would still be indirect gender association in that a word that should be gender neutral, such as {\em receptionist}, is closer to {\em softball} than {\em football} (see Figure \ref{fig:occupation_associations2}). As discussed in the Introduction, it can be subtle to obtain the ground truth of the extent to which such similarities is due to gender.
The gender subspace $g$ that we have identified allows us to quantify the contribution of $g$ to the similarities between any pair of words. We can decompose a given word vector $w \in \mathbb{R}^d$ as $w = w_g + w_{\perp}$, where $w_g = (w \cdot g) g$ is the contribution from gender and $w_{\perp} = w - w_g$. Note that all the word vectors are normalized to have unit length. We define the gender component to the similarity between two word vectors $w$ and $v$ as
\[
\beta(w, v) = \left( w\cdot v - \frac{w_{\perp}\cdot v_{\perp}}{\|w_{\perp}\|_2 \|v_{\perp}\|_2} \right)\bigg/w\cdot v.
\]
The intuition behind this metric is as follow: $\frac{w_{\perp}\cdot v_{\perp}}{\|w_{\perp}\|_2 \|v_{\perp}\|_2}$ is the inner product between the two vectors if we project out the gender subspace and renormalize the vectors to be of unit length. The metric quantifies how much this inner product changes (as a fraction of the original inner product value) due to this operation of removing the gender subspace. Because of noise in the data, every vector has some non-zero component $w_{\perp}$ and $\beta$ is well-defined. Note that $\beta(w,w) = 0$, which is reasonable since the similarity of a word to itself should not depend on gender contribution. If $w_g = 0 = v_g$, then $\beta(w,v) = 0$; and if $w_{\perp} = 0 = v_{\perp}$, then $\beta(w, v) = 1$.
In Figure~\ref{fig:occupation_associations2}, as a case study, we examine the most extreme words on the $\wvec{softball} - \wvec{football}$ direction. The five most extreme words (i.e. words with the highest positive or the lowest negative projections onto $\wvec{softball} - \wvec{football}$) are shown in the table. Words such as \emph{receptionist}, \emph{waitress} and \emph{homemaker} are closer to \emph{softball} than \emph{football}, and the $\beta$'s between these words and \emph{softball} is substantial (67\%, 35\%, 38\%, respectively). This suggests that the apparent similarity in the embeddings of these words to $\wvec{softball}$ can be largely explained by gender biases in the embedding. Similarly, \emph{businessman} and \emph{maestro} are closer to \emph{football} and this can also be attributed largely to indirect gender bias, with $\beta$'s of 31\% and 42\%, respectively.
\begin{figure}
\includegraphics[width=\linewidth,trim={0.5in 0 0 0},clip]{fig1.pdf}
\caption{Selected words projected along two axes: $x$ is a projection onto the difference between the embeddings of the words {\em he} and {\em she}, and $y$ is a direction learned in the embedding that captures gender neutrality, with gender neutral words above the line and gender specific words below the line. Our hard debiasing algorithm removes the gender pair associations for gender neutral words. In this figure, the words above the horizontal line would all be collapsed to the vertical line.
\label{fig:words}}
\end{figure}
\section{Debiasing algorithms}\label{sec:debias}
The debiasing algorithms are defined in terms of sets of words rather than just pairs, for generality, so that we can consider other biases such as racial or religious biases. We also assume that we have a set of words to neutralize, which can come from a list or from the embedding as described in Section \ref{sec:gender_neutral}. (In many cases it may be easier to list the gender specific words not to neutralize as this set can be much smaller.)
The first step, called {\bf Identify gender subspace}, is to identify a direction (or, more generally, a subspace) of the embedding that captures the bias. For the second step, we define two options: {\bf Neutralize and Equalize} or {\bf Soften}. {\bf Neutralize} ensures that gender neutral words are zero in the gender subspace. {\bf Equalize} perfectly equalizes sets of words outside the subspace and thereby enforces the property that any neutral word is equidistant to all words in each equality set. For instance, if $\{\text{grandmother}, \text{grandfather}\}$ and $\{\text{guy}, \text{gal}\}$ were two equality sets, then after equalization {\em babysit} would be equidistant to {\em grandmother} and {\em grandfather} and also equidistant to {\em gal} and {\em guy}, but presumably closer to the grandparents and further from the {\em gal} and {\em guy}. This is suitable for applications where one does not want any such pair to display any bias with respect to neutral words.
The disadvantage of Equalize is that it removes certain distinctions that are valuable in certain applications. For instance, one may wish a language model to assign a higher probability to the phrase {\em to grandfather a regulation}) than {\em to grandmother a regulation} since {\em grandfather} has a meaning that {\em grandmother} does not -- equalizing the two removes this distinction. The Soften algorithm reduces the differences between these sets while maintaining as much similarity to the original embedding as possible, with a parameter that controls this trade-off.
To define the algorithms, it will be convenient to introduce some further notation. A subspace $B$ is defined by $k$ orthogonal unit vectors $B = \{b_1, \ldots, b_k\} \subset \mathbb{R}^d$. In the case $k=1$, the subspace is simply a direction. We denote the projection of a vector $v$ onto $B$ by,
$$v_B = \sum_{j=1}^k (v \cdot b_j) b_j.$$
This also means that $v-v_B$ is the projection onto the orthogonal subspace.
\smallskip
\noindent
{\bf Step 1: Identify gender subspace}. Inputs: word sets $W$, defining sets $D_1, D_2, \ldots, D_n \subset W$ as well as embedding $\bigr\{\vec{w} \in \mathbb{R}^d\bigl\}_{w \in W}$ and integer parameter $k \geq 1$. Let $$\mu_i := \sum_{w \in D_i} \vec{w}/|D_i|$$ be the means of the defining sets. Let the bias subspace $B$ be the first $k$ rows of $\mathrm{SVD}(\textbf{C})$ where $$\textbf{C} := \sum_{i=1}^n \sum_{w \in D_i} (\vec{w}-\mu_i)^T (\vec{w}-\mu_i)\bigl/|D_i|.$$
\smallskip
{\bf Step 2a: Hard de-biasing (neutralize and equalize)}. Additional inputs: words to neutralize $N\subseteq W$, family of equality sets $\mathcal{E} = \{E_1, E_2, \ldots, E_m\}$ where each $E_i \subseteq W$. For each word $w\in N$, let $\vec{w}$ be re-embedded to $$\vec{w} := (\vec{w} - \vec{w}_B)\bigl/\|\vec{w} - \vec{w}_B\|.$$
For each set $E\in \mathcal{E}$, let
\begin{eqnarray*}
\mu &:=& \sum_{w \in E} w/|E|\\
\nu &:=& \mu - \mu_B\\
\text{ For each }w \in E, ~~ \vec{w} &:=& \nu + \sqrt{1-\|\nu\|^2}\frac{\vec{w}_B-\mu_B}{\|\vec{w}_B-\mu_B\|}
\end{eqnarray*}
Finally, output the subspace $B$ and the new embedding $\bigr\{\vec{w} \in \mathbb{R}^d\bigl\}_{w \in W}$.
Equalize equates each set of words outside of $B$ to their simple average $\nu$ and then adjusts vectors so that they are unit length. It is perhaps easiest to understand by thinking separately of the two components $\vec{w}_B$ and $\vec{w}_{\perp B} = \vec{w}-\vec{w}_B$. The latter $\vec{w}_{\perp B}$ are all simply equated to their average. Within $B$, they are centered (moved to mean 0) and then scaled so that each $\vec{w}$ is unit length. To motivate why we center, beyond the fact that it is common in machine learning, consider the bias direction being the gender direction ($k=1$) and a gender pair such as $E=\{\text{male}, \text{female}\}$. As discussed, it so happens that both words are positive (female) in the gender direction, though {\em female} has a greater projection. One can only speculate as to why this is the case, e.g., perhaps the frequency of text such as {\em male nurse} or {\em male escort} or {\em she was assaulted by the male}. However, because {\em female} has a greater gender component, after centering the two will be symmetrically balanced across the origin. If instead, we simply scaled each vector's component in the bias direciton without centering, {\em male} and {\em female} would have exactly the same embedding and we would lose analogies such as {\em father:male :: mother:female}.
Before defining the Soften alternative step, we note that Neutralizing and Equalizing completely remove pair bias.
\begin{observation}
After Steps 1 and 2a, for any gender neutral word $w$ any equality set $E$, and any two words $e_1, e_2 \in E$, $\vec{w} \cdot \vec{e}_1 = w \cdot \vec{e}_2$ and $\|\vec{w}-\vec{e}_1\| = \|\vec{w}-\vec{e}_2\|$. Furthermore, if $\mathcal{E}=\bigl\{\{x,y\}|(x,y)\in P\bigr\}$ are the sets of pairs defining PairBias, then $\mathrm{PairBias}=0$.
\end{observation}
\begin{proof}
Step 1 ensures that $\vec{w}_B=0$, while step 2a ensures that $\vec{e}_1 - vec{e}_2$ lies entirely in $B$. Hence, their inner product is 0 and
$\vec{w} \cdot \vec{e}_1 = w \cdot \vec{e}_2$. Lastly, $\|\vec{w}-\vec{e}_1\| = \|\vec{w}-\vec{e}_2\|$ follows from the fact that $\|u_1-u_2\|^2 = 2-2u_1\cdot u_2$ for unit vectors $u_1, u_2$ and PairBias being 0 follows trivially from the definition of PairBias.
\end{proof}
{\bf Step 2b: Soft bias correction}.
Overloading the notation, we let $W \in \mathbb{R}^{d \times |vocab|}$ denote the matrix of all embedding vectors and $N$ denote the matrix of the embedding vectors corresponding to gender neutral words. $W$ and $N$ are learned from some corpus and are inputs to the algorithm. The desired debiasing transformation $T \in \mathbb{R}^{d \times d}$ is a linear transformation that seeks to preserve pairwise inner products between all the word vectors while minimizing the projection of the gender neutral words onto the gender subspace. This can be formalized as the following optimization problem
\[
\min_T \| (TW)^T(TW)-W^T W \|_F^2 + \lambda \|(TN)^T(TB)\|_F^2
\]
where $B$ is the gender subspace learned in Step 1 and $\lambda$ is a tuning parameter that balances the objective of preserving the original embedding inner products with the goal of reducing gender bias. For $\lambda$ large, $T$ would remove the projection onto $B$ from all the vectors in $N$, which corresponds exactly to Step 2a. In the experiment, we use $\lambda=0.2$. The optimization problem is a semi-definite program and can be solved efficiently. The output embedding is normalized to have unit length, $\hat{W} = \{Tw/\|Tw\|_2, w \in W\}$.
\section{Determining gender neutral words}\label{sec:gender_neutral}
For practical purposes, since there are many fewer gender specific words, it is more efficient to enumerate the set of gender specific words $S$ and take the gender neutral words to be the compliment, $N=W\setminus S$. Using dictionary definitions, we derive a subset $S_0$ of 218 words out of the words in w2vNEWS. Recall that this embedding is a subset of 26,377 words out of the full 3 million words in the embedding, as described in Section \ref{sec:defs}. This base list $S_0$ is given in Appendix \ref{ap:gender_neutral}. Note that the choice of words is subjective and ideally should be customized to the application at hand.
We generalize this list to the entire 3 million words in the Google News embedding using a linear classifier, resulting in the set $S$ of 6,449 gender-specific words. More specifically, we trained a linear Support Vector Machine (SVM) with the default regularization parameter of $C=1.0$. We then ran this classifier on the remaining words, taking $S=S_0\cup S_1$, where $S_1$ were the words labeled as gender specific by our classifier among the words in the entire embedding that were not in the 26,377 words of w2vNEWS.
Using 10-fold cross-validation to evaluate the accuracy of this process, we find an $F$-score of $.627 \pm .102$ based on stratified 10-fold cross-validation. The binary accuracy is well over 99\% due to the imbalanced nature of the classes. For another test of how accurately the embedding agrees with our base set of 218 words, we evaluate the class-balanced error by re-weighting the examples so that the positive and negative examples have equal weights, i.e., weighting each class inverse proportionally to the number of samples from that class. Here again, we use stratified 10-fold cross validation to evaluate the error. Within each fold, the regularization parameter was also chosen by 10-fold (nested) cross validation. The average (balanced) accuracy of the linear classifiers, across folds, was $95.12\% \pm 1.46\%$ with 95\% confidence.
Figure~\ref{fig:words} illustrates the results of the classifier for separating gender-specific words from gender-neutral words. To make the figure legible, we show a subset of the words. The $x$-axis correspond to projection of words onto the $\wvec{she} - \wvec{he}$ direction and the $y$-axis corresponds to the distance from the decision boundary of the trained SVM.
\section{Debiasing results}\label{sec:debiasing_results}
We evaluated our debiasing algorithms to ensure that they preserve the desirable properties of the original embedding while reducing both direct and indirect gender biases.
\begin{table}
\begin{tabular}{l|lll}
& RG & WS & analogy \\
\hline
\hline\\
Before & 62.3& 54.5 & 57.0 \\
Hard-debiased & 62.4 & 54.1 & 57.0\\
Soft-debiased & 62.4 & 54.2 & 56.8 \\
\end{tabular}
\caption{The columns show the performance of the original
w2vNEWS embedding (``before'') and the debiased w2vNEWS on the standard evaluation metrics measuring coherence and analogy-solving abilities: RG \cite{rubenstein1965contextual}, WS \cite{finkelstein2001placing}, MSR-analogy \cite{mikolov2013linguistic}. Higher is better. The results show that the performance does not degrade after debiasing.
Note that we use a subset of vocabulary in the experiments. Therefore,
the performances are lower than the previously published results.
\label{tab:variance}}
\end{table}
\paragraph{Direct Bias.}
First we used the same analogy generation task as before:
for both the hard-debiased and the soft-debiased embeddings, we automatically generated pairs of words that are analogous to
\emph{she-he} and asked crowd-workers to evaluate whether these pairs reflect
gender stereotypes. Figure \ref{fig:direct-result} shows the results.
On the initial w2vNEWS embedding, 19\% of the top 150 analogies were judged as
showing gender stereotypes by a majority of the ten workers.
After applying our hard debiasing algorithm, only 6\% of the new embedding were
judged as stereotypical. As an example, consider the analogy puzzle, \emph{he} to \emph{doctor} is as \emph{she} to $X$. The original embedding returns $X = $ \emph{nurse} while the hard-debiased embedding finds $X =$ \emph{physician}. Moreover the hard-debiasing algorithm preserved gender appropriate analogies such as \emph{she} to \emph{ovarian cancer} is as \emph{he} to \emph{prostate cancer}.
This demonstrates that the hard-debiasing has effectively reduced the gender stereotypes in the word embedding.
Figure \ref{fig:direct-result} also shows that the number of
appropriate analogies remains similar as in the original embedding after executing hard-debiasing. This
demonstrates that that the quality of the embeddings
is preserved. The details results are in Appendix \ref{app:analogies}. Soft-debiasing was less effective in removing gender bias.
To further confirms the quality of embeddings after debiasing,
we tested the debiased embedding on several standard benchmarks that measure whether related words have similar embeddings as well as how well the embedding performs in analogy tasks. Table \ref{tab:variance} shows the results on the original and the new embeddings and the transformation does not negatively impact the performance.
\paragraph{Indirect bias.} We also investigated how the strict debiasing
algorithm affects indirect gender bias. Because we do not have the ground
truth on the indirect effects of gender bias, it is challenging to quantify
the performance of the algorithm in this regard. However we do see promising
qualitative improvements, as shown in Figure~\ref{fig:occupation_associations2}
in the \emph{softball}, \emph{football} example. After applying the strict
debias algorithm, we repeated the experiment and show the most extreme words
in the $\wvec{softball}-\wvec{football}$ direction. The most extreme words
closer to \emph{softball} are now \emph{infielder} and \emph{major leaguer} in
addition to \emph{pitcher}, which are more relevant and do not exhibit
gender bias. Gender stereotypic associations such are \emph{receptionist},
\emph{waitress} and \emph{homemaker} are moved down the list. Similarly, words
that clearly show male bias, e.g. \emph{businessman}, are also no longer at
the top of the list. Note that the two most extreme words in the
$\wvec{softball} - \wvec{football}$ direction are \emph{pitcher} and
\emph{footballer}. The similarities between \emph{pitcher} and \emph{softball}
and between \emph{footballer} and \emph{football} comes from the actual
functions of these words and hence have little gender contribution. These two
words are essentially unchanged by the debiasing algorithm.
\begin{figure}
\begin{tabular}{cc}
\includegraphics[width=.45\textwidth]{bad.png} & \includegraphics[width=.45\textwidth]{good.png}
\end{tabular}
\caption{Number of stereotypical (Left) and appropriate (Right) analogies
generated by wordembeddings before and after debiasing. }
\label{fig:direct-result}
\end{figure}
\iffalse
\section{Other types of bias}\label{sec:racial}
This paper is about gender bias. Other forms of bias in word embeddings, such as racial bias, are challenges for future work. Not all biases are equally represented in language. For instance, while $x=homemaker$ is returned for man:computer programmer :: woman:$x$, the result for white:computer programmer :: black:}$y$ returns $y=$programmer or $y=$mechanical engineer, depending on which algorithm is used for computing analogies.
This is not because there are more black programmers than white programmers -- in fact the opposite is the case and hence it is quite reasonable to argue that. It is because gender is more transparent in English due to the fact that pronouns are gendered. That is, when speaking about a person, one generally identifies their gender by terms such as {\em she} or {\em he}, while the color of someone's skin is often not indicated. Furthermore, the words {\em black} and {\em white} refer to colors. Other terms such as {\em caucasian} and {\em African American} are much less frequent and therefore play less of a role in the embedding.
However, the racial discrimination that is highly prejudice
In this section, we briefly consider challenges and techniques for identifying such biases. We first took the words {\em black} and {\em white}.
We took the 18 black names and 18 white names from the work of Bertrand and Emily \cite{bertrand2004emily}. We took the averages of the black names and white names, and subtracted these vectors to get a vector approximating race. We then
([(-0.31401083, u'artiste'),
(-0.28365907, u'shopkeeper'),
(-0.27799177, u'taxi_driver'),
(-0.24146977, u'cab_driver'),
(-0.23398179, u'preacher'),
(-0.22104323, u'laborer'),
(-0.22067598, u'boxer'),
(-0.21806574, u'barber'),
(-0.19091426, u'cleric'),
(-0.18157765, u'gangster')],
[(0.099201627, u'manager'),
(0.10315713, u'adventurer'),
(0.10569732, u'screenwriter'),
(0.10582993, u'programmer'),
(0.10715814, u'naturalist'),
(0.1109848, u'planner'),
(0.11589377, u'adjunct_professor'),
(0.13659362, u'director'),
(0.13937929, u'architect'),
(0.14118233, u'consultant')])
\begin{figure}
\begin{tabular}{lllc}
\multicolumn{3}{c}{\bf{Extreme \emph{black} occupations}}\\
{1. homemaker} & {2. nurse} & {3. receptionist} \\
{4. librarian} & {5. socialite} & {6. hairdresser} \\
{7. nanny} & {8. bookkeeper} & {9. stylist} \\
{10. housekeeper} & {11. interior designer} & {12. guidance counselor} \\[2ex]
\multicolumn{3}{c}{{\bf Extreme \emph{white} occupations}}\\
{1. maestro} & {2. skipper} & {3. protege} \\
{4. philosopher} & {5. captain} & {6. architect} \\
{7. financier} & {8. warrior} & {9. broadcaster} \\
{10. magician} & {11. figher pilot} & {12. boss} \\
\\\end{tabular}
\caption{\label{fig:occupation_words} The most extreme occupations as projected on to the {\em she$-$he} gender direction on g2vNEWS. Occupations such as {\em businesswoman}, where gender is suggested by the orthography, were excluded.}
\end{figure}
([(-0.12514709, u'monk'),
(-0.084856719, u'graphic_designer'),
(-0.082347713, u'cinematographer'),
(-0.08089295, u'waiter'),
(-0.078203768, u'salesman'),
(-0.077655002, u'housekeeper'),
(-0.076808386, u'illustrator'),
(-0.071323149, u'protester'),
(-0.070619911, u'inventor'),
(-0.070567071, u'sheriff_deputy')],
[(0.10580911, u'alderman'),
(0.11328664, u'legislator'),
(0.12091675, u'marshal'),
(0.12097991, u'minister'),
(0.12368189, u'comic'),
(0.14253864, u'jeweler'),
(0.14631999, u'commissioner'),
(0.1527196, u'secretary'),
(0.16180006, u'councilman'),
(0.16603339, u'treasurer')])
\fi
\section{Discussion}
Word embeddings help us further our understanding of bias in language. We find a single direction that largely captures gender, that helps us capture associations between gender neutral words and gender as well as indirect inequality.The projection of gender neutral words on this direction enables us to quantify their degree of female- or male-bias.
To reduce the bias in an embedding, we change the embeddings of gender neutral words, by removing their gender associations. For instance, {\em nurse} is moved to to be equally male and female in the direction $g$. In addition, we find that gender-specific words have additional biases beyond $g$. For instance, {\em grandmother} and {\em grandfather} are both closer to {\em wisdom} than {\em gal} and {\em guy} are, which does not reflect a gender difference. On the other hand, the fact that {\em babysit} is so much closer to {\em grandmother} than {\em grandfather} (more than for other gender pairs) is a gender bias specific to {\em grandmother}. By equating {\em grandmother} and {\em grandfather} outside of gender, and since we've removed $g$ from {\em babysit}, both {\em grandmother} and {\em grandfather} and equally close to {\em babysit} after debiasing. By retaining the gender component for gender-specific words, we maintain analogies such as {\em she:grandmother :: he:grandfather}. Through empirical evaluations, we show that our hard-debiasing algorithm significantly reduces both direct and indirect gender bias while preserving the utility of the embedding. We have also developed a soft-embedding algorithm which balances reducing bias with preserving the original distances, and could be appropriate in specific settings.
One perspective on bias in word embeddings is that it merely reflects bias in society, and therefore one should attempt to debias society rather than word embeddings. However, by reducing the bias in today's computer systems (or at least not amplifying the bias), which is increasingly reliant on word embeddings, in a small way debiased word embeddings can hopefully contribute to reducing gender bias in society. At the very least, machine learning should not be used to inadvertently amplify these biases, as we have seen can naturally happen.
In specific applications, one might argue that gender biases in the embedding (e.g. \emph{computer programmer} is closer to \emph{he}) could capture useful statistics and that, in these special cases, the original biased embeddings could be used. However given the potential risk of having machine learning algorithms that amplify gender stereotypes and discriminations, we recommend that we should err on the side of neutrality and use the debiased embeddings provided here as much as possible.
In this paper, we focus on quantifying and reducing gender bias in word embeddings. Corpus of documents often contain other undesirable stereotypes and these can also be reflected in the embedding vectors. The same w2vNEWS also exhibits strong racial stereotype. For example, projecting all the occupation words onto the direction $\wvec{minorities} - \wvec{whites}$, we find that the most extreme occupations closer to \emph{whites} are \emph{parliamentarian}, \emph{advocate}, \emph{deputy}, \emph{chancellor}, \emph{legislator}, and \emph{lawyer}. In contrast, the most extreme occupations at the \emph{minorites} end are \emph{butler}, \emph{footballer}, \emph{socialite}, and \emph{crooner}. It is a subtle issue to understand the direct and indirect bias due to racial, ethnic and cultural stereotypes. An important direction of future work would be to quantify and remove these biases.
While we focus on English word embeddings, it is also an interesting direction to consider how the approach and findings here would apply to other languages, especially languages with grammatical gender where the definitions of most nouns carry a gender marker.
\bibliographystyle{abbrv}
|
\section{ INTRODUCTION }
We undertake a systematic survey of the global energetics of solar flares and
coronal mass ejections (CME) observed during the SDO era, which includes
all M and X-class flares during the first 3.5 years of the SDO mission,
covering some 400 flare events. This project embodies the most comprehensive
survey about various forms of energies that can be detected during flares, such
as the dissipated magnetic energy, the thermal energy, the nonthermal
energy, the radiative and conductive energy, and the kinetic energy
of associated CMEs. Two studies have been completed previously,
containing statistics on magnetic energies (Aschwanden Xu, and Jing 2014;
Paper I), and thermal energies (Aschwanden et al.~2015; Paper II).
In this study we focus on the third
part of this ``global flare energetics project", which entails the statistics
of nonthermal energies in hard X ray-producing electrons
that are observed in hard X-rays and gamma-rays, using data from the
Ramaty High-Energy Solar Spectroscopic Imager (RHESSI) spacecraft
(Lin et al.~2002).
The quantitative measurement of nonthermal energies in solar flares
allows us some tests of fundamental nature. One concept or working hypothesis
is that all primary energy input in solar flares is provided by dissipation
of free magnetic energy, for instance by a magnetic reconnection process,
which supplies energy for secondary processes, such as for acceleration
of charged particles and heating of flare plasma.
The accelerated (nonthermal) particles either escape
from the flare site into interplanetary space, or more likely precipitate
down to the chromosphere where they subsequently become thermalized and
radiate in hard X-rays and gamma rays, according to the thick-target
bremsstrahlung model (Brown 1971). In this picture we expect that the
total nonthermal energy $E_{\mathrm{nt}}$ (in electrons and ions)
produced in flares should not exceed the dissipated magnetic (free)
energy $E_{\mathrm{mag}}$, but on the other hand should yield an
upper limit on the thermal energy $E_{\mathrm{th}}$
inferred from the soft-X-ray and EUV-emitting plasma.
Alternative mechanisms to the thick-target model envision
thermal conduction fronts (e.g., Brown et al.~1979) or
direct heating processes (e.g., Duijveman et al.~1981).
In the previous
two papers we proved the inequality $E_{\mathrm{mag}} > E_{\mathrm{th}}$,
for which we found an energy conversion ratio of
$E_{\mathrm{th}}/E_{\mathrm{mag}} \approx 0.02-0.40$
(Paper II), which is about an order of magnitude higher than estimated
in a previous statistical study (Emslie et al.~2012), where an
{\sl ad hoc} value (30\%) of the ratio of the free magnetic energy to the
potential field energy was estimated. In this Paper III we investigate the expected
inequalities $E_{\mathrm{mag}} > E_{\mathrm{nt}} > E_{\mathrm{th}}$.
If these two inequalities are not fulfilled, it could be attributed
to insufficient accuracy of the energy measurements, or alternatively may question
the correctness of the associated low-energy cutoff model, the applied
magnetic reconnection models, or the efficiency of the electron
thick-target bremsstrahlung model.
Such an outcome would have important consequences in our understanding
of solar flare models and the related predictability of the most extreme
space weather events.
The measurement of nonthermal energies in solar flares requires
a spectral fit of the hard X-ray spectrum in the energy range of
$\varepsilon \approx 10-30$ keV (Aschwanden 2007),
from spectral data as they are available from
the HXRBS/SMM, BATSE/CGRO, or RHESSI instrument. Since the total
nonthermal energy contained in a flare requires integrations over
the temporal and spectral range, the largest uncertainty of this
quantity comes from the assumed low-energy cutoff,
because it cannot be directly measured due to the strong thermal
component that often dominates the spectrum at $\varepsilon \lower.4ex\hbox{$\;\buildrel <\over{\scriptstyle\sim}\;$} 20$ keV during
solar flares (for a review see Holman et al.~2011).
In a few cases, low-energy cutoffs of the nonthermal spectrum
could be determined by regularized inversion methods at $e_c=20-40$ keV
(Kasparova et al.~2005), $e_c \approx 20$ keV (Kontar and Brown 2006),
and $e_c=13-19$ keV (Kontar, Dickson, and Kasparova 2008).
For the 2002 July 23 flare, Holman et al.~(2003) deduced upper limits to
low-energy cutoffs by determining the highest values consistent with
acceptable spectral fits. Sui et al.~(2007) deduced the low-energy cutoff
in a flare from the combination of spectral fits and the time evolution of
the X-ray emission in multiple energy bands. Sui et al.~(2007) deduced
low-energy cutoffs for several flares with relatively weak thermal
components (``early impulsive flares'') from spectral fits, with
values ranging from $15-50$ keV. In the late peak of a multi-peaked flare,
Warmuth et al.~(2009) inferred low-energy cutoff values exceeding 100 keV,
but this unusually high value could possibly be explained also by
high-energy electrons that accumulate by trapping after the flare peak
(Aschwanden et al.~1997). Using a novel method of {\bf differentiating}
nonthermal
electrons by their time-of-flight delay from thermal electrons by their
thermal conduction time delay, a thermal-nonthermal crossover energy of
$e_c=18.0 \pm 3.4$ keV (or a range of $e_c = 10-28$ keV) was established
for the majority (68\%) of 65 analyzed flare events (Aschwanden 2007).
Statistical measurements of nonthermal flare energies
have been calculated from HXRBS/SMM data (Crosby et al.~1993),
or from RHESSI data (Hannah et al.~2008; Christe et al.~2008;
Emslie et al.~2012).
The low-energy cutoff was taken into account by assuming a
fixed energy cutoff of $e_c=25$ keV (Crosby et al.~1993),
a fixed spectral slope of $\gamma=-1.5$ below the thermal-nonthermal
cross-over energy $e_{\mathrm{co}}$ (Hannah et al.~2008),
or by adopting the largest energy $e_c$ that still produces a
goodness-of-fit with $\chi^2 \approx 1$ for the nonthermal
power law fit (Emslie et al.~2012). Low-energy cutoffs for
microflares were estimated in the range of $e_c \approx 9-16$ keV,
with a median of 12 keV (Hannah et al.~2008), using a numerical
integration code of Holman (2003).
The statistical study of
Emslie et al.~(2012) provides a comparison between nonthermal energies
$E_{\mathrm{nt}}$, thermal energies $E_{\mathrm{th}}$, and dissipated magnetic
energies $E_{\mathrm{mag}}$, yielding mean (logarithmic) ratios of
$E_{\mathrm{th}} \approx 0.005\ E_{\mathrm{mag}}$ and $E_{\mathrm{nt}}
\approx 0.03\ E_{\mathrm{mag}}$.
These results conform to the expected inequalities, but the magnetic
energies $E_{\mathrm{mag}}$ were actually not measured in the study of
Emslie et al.~(2012),
and most likely were overestimated by an order of magnitude (Paper I).
The dissipated magnetic energies $E_{\mathrm{mag}}$ were for the first time
quantitatively measured in Paper I, by automated tracing of coronal
flare loops from AIA/SDO images and by forward-fitting of a nonlinear
force-free magnetic field (NLFFF) model based on the vertical current
approximation (Aschwanden 2013, 2016).
The content of this paper consists of a theoretical model
to estimate the low-energy cutoff and the nonthermal energy (Section 2),
a description of the data analysis method (Section 3), the results
of the data analysis of 191 M and X-class flare events observed
with RHESSI (Section 4), a discussion of the results (Section 5),
and conclusions (Section 6).
\section{ THEORY }
\subsection{ Nonthermal Energy in Electrons }
The nonthermal energy in flare electrons is generally calculated
with the thick-target model (Brown 1971), which expresses the
hard X-ray photon spectrum by a convolution of the electron
injection spectrum with the Bethe-Heitler bremsstrahlung
cross-section. According to this model, the observed hard X-ray
photon spectrum $I(\varepsilon)$ observed at Earth can be approximated
by a power law function with slope $\gamma$ for the nonthermal
energies, while the spectral index generally changes at the lower
(thermal) energies. Thus, the nonthermal spectrum is defined as
(e.g., see textbook Aschwanden 2004; chapter 13),
\begin{equation}
I(\varepsilon) = A \ \varepsilon^{-\gamma}
\qquad ({\rm photons}\ {\rm cm}^{-2}\ {\rm s}^{-1}\ {\rm keV}^{-1}) \ ,
\end{equation}
which yields a thick-target (non-thermal) electron injection
spectrum $f_e(e)$,
\begin{equation}
f_e(e) = 2.68 \times 10^{33} \ b(\gamma) A
{e}^{-(\gamma+1)}
\qquad ({\rm electrons}\ {\rm keV}^{-1} \ {\rm s}^{-1}) \ ,
\end{equation}
which is a power law function also, but with a slope $\delta = \gamma+1$
that is steeper by one, and $b(\gamma)$ is an auxiliary function related
to the Beta function. The detailed shape of a nonthermal electron spectrum
that is affected by a low-energy cutoff is simulated in Holman (2003),
showing a gradual flattening at lower energies.
Note that we use the symbol $\varepsilon$ for photon
energies, while we use the symbol $e$ for electron energies.
The total power in nonthermal electrons above some cutoff
energy $e_c$, i.e., $P(e \ge e_c$), is
\begin{equation}
P(e \ge e_c)
= 4.3 \times 10^{24} \ {b(\gamma) \over (\gamma - 1)} \ A \
(e_c)^{-(\gamma-1)} \
\qquad ({\rm erg \ s^{-1})} \ .
\end{equation}
Thus, the three observables of the photon
flux $A$, the photon power law slope $\gamma$, and the low-energy
cutoff energy $e_c$ are required to calculate the
power during a selected flare time interval, which can be
calculated with the OSPEX package of the {\sl SolarSoftWare (SSW)} library
of the {\sl Interactive Data Language (IDL)} software (see RHESSI webpage
http://hesperia.gsfc.nasa.gov/ssw/packages/spex/doc/ospex$\_$explanation.html).
In order to calculate the total
nonthermal energy $E_{\mathrm{nt}}$ during an entire flare, we have to
integrate the power as a function of time,
\begin{equation}
E_{\mathrm{nt}} = \int_{t_{\mathrm{start}}}^{t_{\mathrm{end}}}
P(e > e_c(t), t) \ dt
\qquad ({\rm erg}) \ .
\end{equation}
While the photon fluxes $A(t)$ and the spectral slopes $\gamma(t)$ can
readily be measured from a time series of hard X-ray photon spectra (Eq.~1),
the largest uncertainty in the determination of the nonthermal energy
is the low-energy cutoff energy $e_c(t)$ between the thermal
and nonthermal hard X-ray components, typically expected in the range of
$\approx 10-30$ keV (see Table 3 in Aschwanden 2007).
In the following we outline two different
theoretical models of the low-energy cutoff that are applied
in this study.
\subsection{ Thermal-Nonthermal Cross-Over Energy }
The bremsstrahlung spectrum $I(\varepsilon)$ of a thermal plasma with
temperature $T$, as a function of the photon energy $\varepsilon = h\nu$,
setting the coronal electron density equal to the ion density $(n=n_i=n_e)$,
and neglecting factors of order unity (such as the Gaunt factor $g(\nu, T)$
in the approximation of the Bethe-Heitler bremsstrahlung cross-section),
and the ion charge number, $Z\approx 1$, is (Brown 1974; Dulk \& Dennis 1982),
\begin{equation}
I(\varepsilon) = I_0 \int
{\exp{(-{\varepsilon / k_B T})} \over T^{1/2}} {dEM(T) \over dT} \ dT \ ,
\end{equation}
where $I_0 \approx 8.1 \times 10^{-39}$ keV s$^{-1}$ cm$^{-2}$ keV$^{-1}$ and
$dEM(T)/dT$ specifies the {\sl differential emission measure}
$n^2 dV$ in the element of volume $dV$ corresponding to temperature range $dT$,
\begin{equation}
\left({dEM(T) \over dT}\right) dT = n^2(T) \ dV \ .
\end{equation}
Regardless, whether we define this differential emission measure (DEM)
distribution by an isothermal or by a multi-thermal plasma (Aschwanden 2007),
the thermal spectrum $I(\varepsilon)$ falls off similar to an exponential
function at an energy of $\varepsilon \lower.4ex\hbox{$\;\buildrel <\over{\scriptstyle\sim}\;$} 20$ keV (or up to $\lower.4ex\hbox{$\;\buildrel <\over{\scriptstyle\sim}\;$}
40$ keV in extremal cases), while the nonthermal
spectrum in the higher energy range of $\varepsilon \approx 20-100$ keV
can be approximated with a single (or broken) power law function (Eq.~1).
Because of the two different functional shapes, a cross-over energy
$\varepsilon_c$ can often be defined from the change in the spectral
slope between the thermal and the nonthermal spectral component.
The electron energy spectrum, however, can have a substantially lower
or higher cutoff energy (e.g., Holman 2003).
We represent the combined spectrum with the sum of the (exponential-like)
thermal and the (power law-like) nonthermal component, i.e.,
\begin{equation}
I(\varepsilon) =I_{\mathrm{th}}(\varepsilon)+I_{\mathrm{nt}}(\varepsilon)
= I_0 \int {\exp{(-{\varepsilon / k_B T})} \over T^{1/2}}
{dEM(T) \over dT} \ dT \ + \ A \ \varepsilon^{-\gamma} \ ,
\end{equation}
where the cross-over energy $\varepsilon_{\mathrm{co}}$ can be determined in the
(best-fit) model spectrum $I(\varepsilon)$ from the energy where the logarithmic
slope is steepest, i.e., from the maximum of $\partial \log I(\varepsilon)
/ \partial \log \varepsilon$.
\subsection{ Warm-Target Model }
A new theoretical model has recently been developed that allows us to
calculate the low-energy cutoff energy in the thick-target model
directly, by including the ``warming'' of the cold thick-target plasma
during the electron precipitation phase, when chromospheric
heating and evaporation sets in (Kontar et al.~2015).
Previous applications of the thick-target model generally assume
cold (chromospheric) temperatures in the electron precipitation site
(e.g., Holman et al.~2011, for a review).
The theoretical derivation of the warm-target model has been analytically
derived and tested with numerical simulations that include the effects of
collisional energy diffusion and thermalization of fast electrons
(Galloway et al.~2005; Goncharov et al.~2010; Jeffrey et al.~2014).
According to this model, the effective low-energy cutoff $e_c$ is
a function of the temperature $e_{\mathrm{th}}=k_B T_e$ of the warm-target
plasma and the power law slope $\delta = \gamma + 1$ of the
(nonthermal) electron flux,
\begin{equation}
e_c \approx (\xi + 2) \ k_B T_e = \delta \ k_B T_e \ .
\end{equation}
where $\xi = \gamma - 1$ is the power law slope of the
source-integrated mean electron flux spectrum ({\bf see Eqs.~8-10
in Kontar et al.~2015), and $T_e$ is the temperature of the
warm target, which is a mixture or the cold preflare plasma and
the heated evaporating plasma.}
Thus, for the temperature range of a medium-sized to a large
X-class flare, which spans $T_e \approx 10-25$ MK, the temperature in
energy units is $E_{\mathrm{th}} = k_B T_e \approx 0.9-2.1$ keV,
and for a range of
power law slopes of $\delta = 3-6$ (Dennis 1985; Kontar et al.~2011),
a range of $e_c \approx 3-13$ keV is predicted for the low-energy cutoffs
by this model.
Besides collisional heating of the warm chromospheric target,
electron beams and beam-driven Langmuir wave turbulence may affect
the low-energy cutoff additionally (Hannah et al.~2009).
Alternative analytical models on the low-energy cutoff can be derived
from a collisional time-of-flight model (Appendix A), from the
Rosner-Tucker-Vaiana heating/cooling balance model (Appendix B), and
from the runaway acceleration model (Appendix C).
\section{ DATA ANALYSIS METHOD }
From the same comprehensive catalog of 399 M and X-class flares observed with
SDO during 2010-2014, used in the first two studies of our global flare
energetics project, we will analyze all events that have been simultaneously
observed in hard X-rays and gamma-rays with RHESSI. The orbit of
RHESSI has a duty cycle of $\approx 50\%$, leading to a total of
191 events that have suitable time coverage. In the following we describe
the analysis of these events, which are also listed in Table 1
(labeled with identical identification numbers \#1$-$399 as used in
Papers I and II). We explain the various steps performed in our analysis
for three examples shown in Figs.~1-3.
\subsection{Spectral Modeling of RHESSI Data with OSPEX }
For the measurement of the nonthermal energy ($E_{\mathrm{nt}}$) of electrons
during solar flares we use the OSPEX (Object Spectral Executive) software,
which is an object-oriented interface for X-ray spectral analysis of
solar data, written by Richard Schwartz and others (see RHESSI website
{\sl http://hesperia.gsfc.nasa.gov/} for a documentation).
The OSPEX software allows the user to read RHESSI data,
to select and subtract a background, to select time intervals of interest,
to select a combination of photon flux model components,
and to fit those components to the spectrum in each selected time
interval. During the fitting process, the response matrix is
used to convert the photon model to the model counts that are fitted
to the observed counts. The OSPEX software deals also with changes of
attenuator states, decimation, pulse pile-up effects, albedo effects,
and provides procedures to calculate the nonthermal energy $(E_{\mathrm{nt}})$
(according to the thick-target model) and the thermal energy ($E_{\mathrm{th}}$)
down to energies of $\lower.4ex\hbox{$\;\buildrel >\over{\scriptstyle\sim}\;$} 3$ keV.
RHESSI complements spectral
information of the differential emission measure (DEM) distribution
at the high-temperature side ($T_e \lower.4ex\hbox{$\;\buildrel >\over{\scriptstyle\sim}\;$} 16$ MK)
(Caspi 2010; Caspi and Lin 2010; Caspi et al.~2014),
while AIA/SDO provides DEM information at the low-temperature side
($T_e \lower.4ex\hbox{$\;\buildrel <\over{\scriptstyle\sim}\;$} 16$ MK), as we determined in Paper II. For spectral modeling
we are using the two-component model {\sl vth+thick2$\_$vnorm}, which includes a
thermal component at low energies and a (broken) power law function at higher
(nonthermal) energies. In our spectral fits we are only interested in
the transition from the thermal to the nonthermal spectrum, which can be
expressed by an exponential-like plus a single power law function (Eq.~7),
and thus we use only the lower power law part of the two-component model
{\sl vth+thick2$\_$vnorm}, while the spectral slope in the upper part was set to
a constant ($\delta_2=4$). In addition we use
{\sl calc$\_$nontherm$\_$electron$\_$energy$\_$flux} of the OSPEX
package to calculate the nonthermal energy flux in the thick-target model.
\underbar{RHESSI Spectral Fitting Range Selection:}
In order to obtain a self-consistent measure of the nonthermal energy, which
varies considerably during the duration of a flare or among different flares,
we have to choose a spectral fitting range that covers a sufficient
part of both the thermal and nonthermal components. {\bf We choose
the maximum energy range $[\varepsilon_1,\varepsilon_2]$, bound by
$\varepsilon_1 = 6 ... 10$ keV and $\varepsilon_2 = 20 ... 50$ keV,
in which an acceptable (reduced) $\chi^2$-value ($\chi < 2.0$) is obtained
for the spectral fit.
The upper bound of the fitting range is mostly constrained by the photon
count statistics, which is often too noisy for energies at $\varepsilon_2
\lower.4ex\hbox{$\;\buildrel >\over{\scriptstyle\sim}\;$} 30$ keV during small flares (M-class here), given the
time steps of $\Delta t=20$ s chosen throughout.} The fitted energy ranges
cover also the range of cross-over energies (10-28 keV) found in
multi-thermal fitting of energy-dependent time delays (Aschwanden 2007).
As a general criticism, we have to be aware that the nonthermal
spectral component could in addition also be confused with a multi-thermal
component in the fitted spectral range of $\varepsilon \approx 10-30$ keV
(Aschwanden 2007), or with non-uniform ionization effects (Su, Holman, and
Dennis 2011), or with return-current losses (Holman 2012).
\medskip
\underbar{RHESSI Detector Selection:}
We used the standard option of OSPEX, where a spectral fit is calculated
from the combined counts of a selectable set of RHESSI subcollimaters.
RHESSI has 9 (subcollimator) detectors that had originially near-identical
sensitivities, but progressively deviate from each other as a result of
steady degradation over time due to radiation damage from charged particles.
Heating up the germanium restores the lost sensitivity and resolution,
and thus five annealing procedures have been applied to RHESSI so far
(second anneal at 2010 Mar 16 - May 1; third at 2012 Jan 17 - Febr 22;
forth at 2014 Jun 26 - Aug 13; and fifth at 2016 Feb 23 - Apr 23).
No science data are collected during the annealing periods. Based on
the performance of the individual detector sensitivities, it is
general practice to exclude the detectors 2 and 7 in spectral fits.
Furthermore, detectors 4 and 5 are considered as unreliable {\bf after
January 2012} (Richard Schwartz; private communication). Therefore, we select
the set of detectors [1, 3, 4, 5, 6, 8, 9] in spectral fits up to
the third anneal in January 2012 (events \# 1-126 in Table 1), and the set of
[1, 3, 6, 8, 9] after February 2012 (events \# 154-395 in Table 1).
Omitting detectors 4 and 5 in the latter set of 71 events yields
a total nonthermal energy that is by a factor of $q_{\mathrm{det}}=
E_{\mathrm{nt}}[1, 3, 6, 8, 9]/E_{\mathrm{nt}}[1, 3, 4, 5, 6, 8, 9]=1.3 \pm 0.5$
higher.
\medskip
\underbar{GOES Time Range and RHESSI Time Resolution:}
We download the GOES 1-8 \AA\ light curves $F_{\mathrm{GOES}}(t)$ and
calculate the time derivative as a proxy for the hard X-ray
time profile $F_{\mathrm{HXR}} \approx dF_{\mathrm{GOES}}/dt$, as shown in Figs.~1a ,2a, and 3a.
The start time $t_{\mathrm{start}}$, peak time $t_{\mathrm{peak}}$,
and end time $t_{\mathrm{end}}$ are defined from the NOAA/GOES catalog.
We compute consecutive spectra in time steps of $\Delta t=20$ s.
Note that RHESSI is a spinning spacecraft with a period of 4 s,
which does not cause any modulation effects for 20 s time integrations.
\medskip
\underbar{RHESSI Quick-Look Data:}
In a next step we inspect the RHESSI quick-look time profiles
(Figs.~1b, 2b, 3b), which show photon counts in
{\bf 5 different energy channels in the range of 6-300 keV}. Based on
these RHESSI time profiles we select time intervals for background
subtraction. {\bf Generally we select a time interval at flare start
as the background interval (in 90\%), and subtract this preflare spectrum
for the entire flare time interval. Only in a few cases (10\%)
where the preflare flux is higher than the postflare flux, we choose
a time interval at flare end for background subtraction.}
The RHESSI quick-look data show changes in the attenuator state
(e.g., Figs.2b, 3b),
which are automatically handled in most time intervals with the OSPEX
software, unless there is a change in the attenuator state during a
selected time interval itself, in which case this time interval is removed
from the spectral analysis. The quick-look data show occasionally data gaps
that are caused when RHESSI enters
spacecraft night in its near-Earth orbit. If the data gap does not
occur during the flare peak of hard X-ray emission, we still include
the event in the analysis, as long as the time interval of dominant
nonthermal HXR emission is covered (such as event \#219 in Fig.~2b).
\medskip
\underbar{OSPEX Spectral Fitting:}
For spectral fitting we perform first a semi-calibration
and store the detector response matrix (DRM), and then
run a spectral fit with the fit function {\sl vth+thick2$\_$vnorm}
using the OSPEX software, optimizing the following model fit
parameters (for each time interval $t$):
\begin{tabular}{ll}
$EM(t)$ & = Emission measure in units of $10^{49}$ cm$^{-3}$ \\
$T_e(t)$ & = plasma temperature in units of keV (1 keV=11.6 MK) \\
$A(t)$ & = photon flux at $\varepsilon = 50$ keV \\
$\delta(t)$ & = negative power law index of electron spectrum \\
$e_c(t)$ & = low-energy cutoff \\
\end{tabular}
Examples of spectral fits are shown in
Figs.~1c, 2c, and 3c, fitted at the time of the peak power
$P_{\mathrm{co}}(t)$ (indicated with red vertical lines in Figs.~1, 2, and 3).
The best-fit spectrum yields a cross-over energy $e_{\mathrm{co}}$ between
the thermal and nonthermal spectral component. Alternatively,
the warm-target model of Kontar et al.~(2015) yields a low-energy
cutoff value $e_{\mathrm{wt}}$. The fitted energy ranges
are listed in Table 1 and are indicated with dotted vertical lines
in Figs.~1g, 2g, and 3g. The goodness-of-fit is quantified with the
$\chi^2$-value criterion.
In case of bad fits of the $\chi^2$-values ($\chi
> 2$) {\bf we changed either the fitted energy range (in 13\%),
the selected interval for background subtraction (10\%),
or the fitted time range (5\%).}
\section{ RESULTS }
The numerical values of the main results of the low energy cutoffs
$e_c$ (which we label as $e_{\mathrm{co}}$ in the cross-over method,
and as $e_{\mathrm{wt}}$ in the warm-target method), and the nonthermal
energy $E_{\mathrm{nt}}$ for the analyzed 191 events are listed in
Table 1, while scatter plots and distributions are shown in Figs.~4-8.
\subsection{ Time Evolution of Flares }
Three examples of analyzed flare events are shown in Figs.~1, 2, and 3,
including one of the smallest events (Fig.~1: \#387, GOES M1.0 class),
{\bf an event with multi-peak characteristics} (Fig.~2; \#219, GOES M2.0 class),
and one of the largest events (Fig.~3; \#12, GOES X2.2 class). In all three
cases we show the time evolution of the most important fit
parameters in the various panels (d through j) of Figs.~1-3:
(d) the thermal emission measure $EM(t)$;
(e) the temperature evolution $T_e(t)$;
(f) the nonthermal photon flux $I_{\mathrm{nt}}(t)$ at 50 keV;
(g) the power law slope $\delta(t)$;
(h) the goodness-of-fit $\chi(t)$;
(i) the nonthermal power $P_{\mathrm{wt}}(t)$ using the low cutoff energy based on the
warm-target model (Section 2.3); and
(j) the low-energy cutoff $e_{\mathrm{wt}}(t)$ of the warm-target model.
{\bf In the examples shown in Figs.~1, 2 and 3 we see that the thermal
emission measure $EM(t)$ increases during the rise time of the GOES flux,
while the temperature $T_e(t)$ decreases, which indicates both density
increases due to chromospheric evaporation as well as subsequent
plasma cooling during the impulsive flare phase. Since multiple
heating and cooling cycles overlap during a flare, we see both effects
simultaneously. The cases shown in Figs.~1, 2 and 3 show also that
the nonthermal flux $I_{\mathrm{nt}}(t)$ (Figs.~1f, 2f, 3f) and the power
$P_{\mathrm{wt}}(t)$ (Figs.~1i, 2i, 3i) are correlated with the GOES
time derivative (Figs.~1a, 2a, 3a).}
\subsection{ Goodness-of-Fit }
The goodness of the spectral fits computed with the OSPEX code is
specified with the $\chi^2$-criterion, based on the least-square
difference between the theoretical spectral model (isothermal plus
power law nonthermal function) and the observed counts in the fitted
energy range [$\varepsilon_1$-$\varepsilon_2$]. The fitted energy
time interval (with a resolution of 1 keV) has about
$n_{\mathrm{bin}} \approx 30-10 = 20$ energy bins, while the model
has four ($n_{\mathrm{par}}=4)$ free parameters
$(EM, T_e, A_{50}, \delta)$, yielding a degree of freedom
$n_{\mathrm{free}}=n_{\mathrm{bin}}-n_{\mathrm{par}} \approx 20-4 = 16$.
In our spectral analysis of 191 flare events we performed spectral
fits, with an average of $n_t \approx 27$ time steps per event,
amounting to a total of $N_{\mathrm{spec}} \approx
191 \times 27 = 5157$ spectral fits. The values $\chi(t)$ of three
events are shown in Figs.~1h, 2h, and 3h. The median values of these
three events are $\chi=1.4, 1.0$, and 1.3. We obtained in all 191
events a median goodness-of-fit value of $\chi < 2$,
after adjustment of the fitted energy range if necessary.
The mean and standard deviations of the median $\chi^2$-values of all
191 events is $\chi = 1.2 \pm 0.4$, which indicates that the
fitted spectral model is adequate in the chosen fitted energy range.
Of course, if one particular model,
such as the two-component thermal-nonthermal model chosen here
(Eq.~7), is found to be consistent with the data according to an
acceptable goodness-of-fit criterion, it does not rule out
alternative models. For instance, the thermal component is often
modeled with an iso-thermal (single-temperature) spectrum, while
a multi-thermal power law function was found to fit the thermal
flare component in most flares equally well (Aschwanden 2007).
\subsection{ Temperature Definitions }
A representative value for the electron temperature during a flare
can be defined in various ways. In paper II we measured the peak
temperature $T_{\mathrm{AIA}}$ of the differential emission measure (DEM)
distribution at the peak time of the flare, as well as the
emission measure-weighted temperature $T_w$ (Eq.~13 in Paper II),
which approximately characterizes the ``centroid'' of the
(logarithmic) DEM function. The mean ratio of these two temperature
values was found to be $q_T=T_{\mathrm{AIA}}/T_w=0.31$ within
a standard deviation by a factor of $2.0$ (Fig.~4, left panel). The
emission measure-weighted temperature $T_w$ is generally found to be
higher, because near-symmetric DEM functions as a function of
the logarithmic temperature are highly asymmetric on a linear
temperature scale, with a centroid that is substantially higher than
the logarithmic centroid.
On the other hand, spectral fits of RHESSI data with an isothermal
component are known to have a strong bias towards the highest
temperatures occurring in a flare, because the fitted energy range
covers only the high-temperature tail of the DEM distribution function
(Battaglia et al.~2005; Ryan et al.~2014; Caspi et al.~2014).
A statistical study demonstrated that the high-temperature bias of
RHESSI by fitting in the photon energy range of $\varepsilon \approx
6-12$ keV amounts to a factor of $T_{R}/T_{\mathrm{AIA}}=1.9\pm1.0$
(Ryan et al.~2014).
Here we find that all RHESSI temperatures averaged during each flare
are found in a range of $T_{R}=16-40$ MK, which
is about equal to the emission measure-weighted temperature,
i.e., $T_{R}/T_w=0.90$ within a factor of 1.4 (Fig.~4, right panel).
The 1-$\sigma$ ranges (containing 67\% of the values) of the various temperature
definitions are $T_{\mathrm{AIA}} \approx 3-14$ MK, $T_w \approx 20-30$ MK, and
$T_{R} \approx 19-28$ MK.
So, we should keep these different temperature definitions in mind
when we calculate the low-energy cutoff $e_c(t)$ as a function of the
RHESSI temperature $T_{R}(t)$ (Eq.~8 for the warm-target model).
{\bf The most decisive parameter in the determination of the nonthermal energy
$E_{nt}$ is the low-energy cutoff $e_c$ (Eq.~4), which is directly proportional
to the temperature $T_e$ in the warm target (Eq.~8). What is the most likely
temperature of the warm target? The relevant temperature is a mixture of
pre-flare plasma temperatures and upflowing evaporating flare plasma.
In the absence of a sound model, we resort to the mean value of the
DEM peak temperatures determined in flaring active regions,
as determined with AIA in Paper II, yielding a mean value of $T_{AIA}=
8.8 \pm 6.0$ MK (Fig.~4 left panel), averaged over $N=380$ M and X-class
flare events. For the subset of 191 flare events observed with RHESSI,
this mean value is $T_e=8.6$ MK, or $k_B T_e = 0.74$ keV. Note that
a deviation of the plasma temperature by a factor of two will result
into a deviation in the determination of the nonthermal energy $E_{nt}$
by about an order of magnitude (using a power law with a typical
slope of $\gamma \approx 4$ in Eqs.~3 and 4).}
\subsection{ Nonthermal Energy Parameters }
The nonthermal energy in electrons, calculated as a time integral
$E_{\mathrm{nt}}$ (Eq.~4), using the low-energy cutoff according to the
warm thick-target model $e_{\mathrm{wt}}(t)$ (Section 2.3; Eq.~8),
or alternatively the thermal/nonther\-mal cross-over energy
$e_{\mathrm{co}}(t)$ (Section 2.2), is the main objective of this study.
Examples of the time evolution of the
nonthermal parameters [$A(t), \delta(t)$, $e_{\mathrm{co}}(t)$, $e_{\mathrm{wt}}(t)$]
and the resulting nonthermal energies $dE_{\mathrm{nt}}(t)$ are shown in
Figs.~1-3. In Fig.~5 we show statistical results of
these parameters. Investigating the dependence of these parameters
on the flare temperature $T_{R}$ we find that both the
low-energy cutoff energy $e_{\mathrm{wt}}$ {\bf (Fig.~5a) as well as
the nonthermal (warm-target) energy $E_{\mathrm{nt}}$ (Fig.~5b)
are {\bf uncorrelated} with the RHESSI temperature.
If we use the thermal-nonthermal cross-over method to estimate the
low-energy cutoff, we find a systematically higher value,
$e_{\mathrm{co}} \lower.4ex\hbox{$\;\buildrel >\over{\scriptstyle\sim}\;$} e_{\mathrm{wt}}$
(Fig.~5c). Consequently, the nonthermal energy
estimated with the cross-over method is systematically
lower than the nonthermal energy calculated with the warm-target model
(Fig.~5d). This result strongly depends on the assumption of the
warm-target temperature. Based on a mean temperature of $T_e=8.6$ MK
found in the active regions analyzed here,
we derive low-energy cutoff energies of $e_{\mathrm{wt}} =6.2\pm 1.6$ keV for
the warm-target model, which is significantly lower than the cross-over
energies $e_{\mathrm{co}}=21 \pm 6$ keV. If we adopt the warm-target
model, we conclude that the cross-over method over-estimates the
low-energy cutoff and under-estimates the nonthermal energies.}
\subsection{ Comparison of Magnetic, Nonthermal, and Thermal Energies }
In Fig.~6 we show scatterplots of the nonthermal energy $E_{\mathrm{nt}}$
measured here with other forms of previously determined energies, such as the
magnetic energy $E_{\mathrm{mag}}$ (Paper I) and the (total pre-impulsive and
post-impulsive) thermal energies $E_{\mathrm{th}}$ (Paper II).
{\bf The energy ratios are characterized with the means of the logarithmic
energies in the following. The ratios between the three forms of energies
are shown separately for the cross-over method in the left-hand panels
of Fig.~6, and for the warm-target model in the right-hand panels of Fig.~6.
The ratios between the nonthermal energies and the magnetically dissipated
energy is $E_{\mathrm{co}}/E_{\mathrm{mag}}=0.01$ for the cross-over method,
or $E_{\mathrm{wt}}/E_{\mathrm{mag}}=0.41$ for the warm-target model,
respectively. Thus, the warm-target model yields ratios that are closer
to unity, which is expected in terms of magnetic reconnection processes,
where most of the magnetic energy is converted into particle acceleration.
We find that the dissipated magnetic energy is sufficient to supply the
energy in nonthermal particles in 71\% for the warm-target model, or in
97\% for the cross-over model (Figs.~6a and 6b).
The ratios between the thermal energies and the magnetically dissipated
energy is $E_{\mathrm{th}}/E_{\mathrm{mag}}=0.08$ for both the
cross-over or the warm-target model (Fig.~6c and 6d).
We find that the dissipated magnetic energy is sufficient to supply the
thermal energy in 95\%.
Comparing the thermal with the nonthermal energies, we find a mean
a ratio of $E_{\mathrm{th}}/E_{\mathrm{wt}}=0.15$ for the warm-target
model, or $E_{\mathrm{th}}/E_{\mathrm{co}}=6.46$ for the cross-over
method. We find that the nonthermal energy is sufficient to supply the
thermal energy in 85\% for the warm-target model (Fig.~6f), but only in
29\% for the cross-over method. Thus, the warm-target model yields
values that are closer to the expectations of the standard
thick-target model, where the thermal energy is entirely produced
by the nonthermal energy of precipitating (nonthermal) electrons.
We show the comparison of nonthermal and thermal energies also in form of
cumulative size distributions in Fig.~7, for the subset of 75 flare for
which all three forms of (magnetic, thermal, nonthermal) energies
could be calculated. We find that the nonthermal energy is typically
an order of magnitude larger than the thermal energy in the statistical
average. Then we find that the nonthermal energy is smaller than the
magnetic energy, as expected for magnetic reconnection processes, for
smaller flares with energies of $E_{nt} < 3 \times 10^{32}$ erg.
However, we find the opposite results for larger flares, with the
nonthermal energy exceeding the magnetically dissipated energy,
for large events with $E_{nt} > 3 \times 10^{32}$ erg.
Since the uncertainties in nonthermal energies are about an
order of magnitude and the dissipated magnetic energy exceeds
the nonthermal energy in 71\% (Fig.~6b), we suspect that the
largest nonthermal energies are over-estimated, which would
indicate that a higher value of the low-energy cutoff or a
higher flare plasma temperature (than the mean active region
temperature $T_e=8.6$ MK used here) could ameliorate the
over-estimated nonthermal energies.}
We compare the occurrence frequency distributions of magnetic,
nonthermal, and thermal energies, as well as those
of the direct RHESSI observables: the peak counts $P$, total counts $C$,
and durations $D$ (Fig.~8). As a caveat, we have to be aware
that these values for $P$ and $C$ are obtained
from the online RHESSI flare catalog, and thus are not
well-calibrated, as they do not take attenuation or decimation
into account. Nevertheless, taking these raw values,
the magnetic and thermal energies have similar
power law slopes of $\alpha \approx 2.0$, while the nonthermal
energies have a slightly flatter slope of $\alpha_{\mathrm{nt}}=1.55 \pm 0.11$,
which can be compared with a previous study, where a power law slope
of $\alpha_{\mathrm{nt}}=1.53\pm0.02$ was found (Crosby et al.~1993).
The latter study is actually
based on larger statistics, containing 2878 flare events observed
with HXRBS/SMM during 1980-1982 (Crosby et al.~1993), but with a higher
assumed low-energy cutoff of $e_c > 25$ keV.
\section{ DISCUSSION }
\subsection{ Energy Partition in Flares }
While we determined the dissipated magnetic energies $E_{\mathrm{mag}}$ (Paper I;
called $E_{\mathrm{diss}}$ therein), thermal energies $E_{\mathrm{th}}$ (Paper II), and
the nonthermal energies $E_{\mathrm{nt}}$, we can ask now the question how the
energy partition from primary to secondary energy dissipation works
in solar flares. Many solar flare models are based on a magnetic
reconnection process, where a stressed non-potential magnetic
field becomes unstable and undergoes a reconfiguration towards a lower
magnetic energy state, releasing during this process some amount
$E_{\mathrm{mag}}=q_{\mathrm{diss}} E_{\mathrm{free}}$ of the magnetic free energy $E_{\mathrm{free}}$
(defined by the difference between the non-potential and the potential
energy, $E_{\mathrm{free}}=E_{\mathrm{np}}-E_p$). Excluding alternative energy sources,
we hypothesize that this dissipated magnetic energy $E_{\mathrm{mag}}$ is
considered to be the entire available primary energy input, while
other energy conversion processes represent secondary steps that
need to add up in the energy budget,
\begin{equation}
E_{\mathrm{mag}} = ( E_{\mathrm{nt}} + E_{\mathrm{cme}} + ... ) > E_{\mathrm{nt}} \ ,
\end{equation}
such as the nonthermal energy $E_{\mathrm{nt}}$ that goes into acceleration
of particles, or the energy $E_{\mathrm{cme}}$ to accelerate an accompanying
coronal mass ejection (CME). The nonthermal energy $E_{\mathrm{nt}}$ may be further
subdivided into energies in electrons $E_{\mathrm{nt,e}}$ and ions $E_{\mathrm{nt,i}}$,
\begin{equation}
E_{\mathrm{nt}} = ( E_{\mathrm{nt,e}} + E_{\mathrm{nt,i}} + ... ) > E_{\mathrm{nt,e}} \ ,
\end{equation}
while the CME energy $E_{\mathrm{cme}}$ consists of the
kinetic energy $E_{\mathrm{kin}}$ and the gravitational potential
energy $E_{\mathrm{grav}}$, and part of it may be converted into
acceleration of particles in the interplanetary CME shock ($E_{\mathrm{nt,cme}}$),
which are particularly present in solar energetic particle (SEP) events,
\begin{equation}
E_{\mathrm{cme}} = E_{\mathrm{kin}} +
E_{\mathrm{grav}} +
E_{\mathrm{nt,cme}} + ... \ .
\end{equation}
We have to be careful to avoid double-counting of secondary energies,
because there may be some tertiary energy conversion processes, such
as heating of chromospheric plasma according to the thick-target
bremsstrahlung model, $E_{\mathrm{th}}$, while upgoing nonthermal particles
escape into interplanetary space, carrying an energy of $E_{\mathrm{nt,esc}}$,
\begin{equation}
E_{\mathrm{nt}} = ( E_{\mathrm{th}} + E_{\mathrm{nt,esc}} + ... ) > E_{\mathrm{th}} \ .
\end{equation}
Since we have measured only three types of energies so far,
$E_{\mathrm{mag}}, E_{\mathrm{nt}}$, and $E_{\mathrm{nt}}$, we can test only the inequalities given
on the righthand-side of Eqs.~(9) and (12) at this point.
Based on the nonthermal energies in electrons
determined in this work we can answer the question
whether the so far measured magnetic energy is sufficient
to accelerate the electrons
observed in hard X-rays, i.e., $E_{\mathrm{mag}} > E_{\mathrm{nt}}$,
as expected for magnetic reconnection models.
Relying on the warm-target model we found that 41\% of the dissipated
magnetic energy (with a standard deviation of about an order of
magnitude) is converted into acceleration of nonthermal electrons,
or a total amount of $\approx 82\%$ for both electrons and ions in the
case of equipartition, while the rest is available to accelerate CMEs.
There are few statistical estimates of the flare energy budget in literature
(besides the work of Emslie et al.~2012; Warmuth and Mann 2016).
One early study quoted that the nonthermal energy in electrons $>20$ keV
contains 10-50\% of the total energy output for the August 1972 flares
(Lin and Hudson 1976; Hudson and Ryan 1995), which is consistent
with our result of 41\% within the measurement uncertainties.
Comparing the energy ranges determined in this global flare energetics
project with those obtained from 38 events in Emslie et al.~(2012),
we find higher amounts of nonthermal flare electron energies
in the statistical average,
covering the range of $E_{\mathrm{nt}} \approx (20-2000) \times 10^{30}$ erg
(Fig.~9), which is mostly accounted for by a lower value of the
low-energy cutoff predicted by the
warm-target model (Kontar et al.~2015) for some events, while
cutoff energies with the highest acceptable value of the
$\chi^2$ was used in Emslie et al.~(2012).
The magnetically dissipated energies appear to be over-estimated by
an order of magnitude (Fig.~9) in Emslie et al.~(2012), based on the ad hoc
assumption that the dissipated energy amounts to 30\% of the potential
field energy therein (Paper I). On the other hand, the thermal energies
appear to be underestimated by at least an order of magnitude (Fig.~9)
in Emslie et al.~(2012) due to the isothermal approximation,
as discussed in Paper II.
\subsection{Insufficiency of the Thick-Target Model?}
A second question we can answer is whether
the nonthermal energy in electrons is sufficient to
heat the flare plasma by the chromospheric evaporation process,
as expected in the thick-target model according to the Neupert effect
(Dennis and Zarro 1993), which requires $E_{\mathrm{nt}} > E_{\mathrm{th}}$.
{\bf Based on the warm-target model we found a mean (logarithmic) ratio of
$E_{\mathrm{th}}=0.15\ E_{\mathrm{nt}}$ (Fig.~6f). The fraction of flares
that have a thermal energy less than the nonthermal energy,
as expected in the standard thick-target model,
amounts in our analysis to $\approx 85\%$ for the warm-target method,
or $\approx 29\%$ for the cross-over model.
This means that the thick-target model
could be insufficient to supply enough energy to explain the thermal
energy produced by the chromospheric evaporation process in about
15\% of the flares for the warm-target model,
or in 71\% for the cross-over model.
Thus, the cross-over model would pose a series problem for the
thick-target model.}
The insufficiency of the thick-target model has been
addressed as a failure of the theoretical Neupert effect
(Veronig et al.~2005; Warmuth and Mann 2016),
which invokes testing of the correlation between the electron beam
power (from RHESSI) and the time derivative of the thermal energy
heating rate (from GOES). From such study it was concluded
that (1) fast electrons are {\sl not} the main source of soft X-ray
plasma supply and heating, (2) the beam low cutoff energy varies
with time, or (3) the theoretical Neupert effect is strongly
affected by the source geometry (Veronig et al.~2005).
If the thermally dominated flares cannot be fully explained by the
thick-target model, additional heating sources
besides precipitating electrons would be required. The most popular
alternative to the thick-target model is heating by thermal
conduction fronts (Brown et al.~1979; Batchelor et al.~1985;
Emslie and Brown 1980; Smith and Harmony 1982; Smith and Brown 1980;
Reep et al.~2016). Other forms of direct heating
(for an overview see chapter 16 in Aschwanden 2004) occur via
(1) resistive or Joule heating processes, such as anomalous
resistivity heating (Duijveman et al.~1981; Holman 1985; Tsuneta
1985), ion-acoustic waves (Rosner et al.~1978b), electron
ion-cyclotron waves (Hinata 1980), (2) slow-shock heating
(Cargill and Priest 1983; Hick and Priest 1989),
(3) electron beam heating by Coulomb collisional loss in the
corona (Fletcher 1995, 1996; Fletcher and Martens 1998),
(4) proton beam heating by kinetic Alfv\'en waves
(Voitenko 1995, 1996), or (5) inductive current heating
(Melrose 1995, 1997).
{\bf The thick-target model fails to explain the observed amount of
thermal energy only in a small number of flares for the warm-target
model, while it is a larger number of events for the
cross-over method. However, it is more likely that the cross-over
method over-estimates the low-energy cutoff, which under-estimates
the nonthermal energies, while the physics-based warm-target model
leads to higher nonthermal energies, in which case the problem
with the insufficiency of the thick-target model goes away.}
\subsection{ Nonthermal Low-Energy Cutoff in Flares }
We outlined two different methods to infer a low-energy
cutoff. The first method consists in measuring the cross-over
between the fitted thermal and nonthermal spectral components,
which yields an upper limit on the low-energy cutoff,
but a statistical test demonstrates that the obtained values
($e_{\mathrm{co}}=21 \pm 6$ keV) {\bf are significantly higher
than those obtained from the warm-target model
($e_{\mathrm{wt}}=6.2 \pm 1.6$ keV).}
There are pros and cons for each method. The cross-over method
requires a dominant thermal component, {\bf which is
not always detectable in the spectrum, in which case
the cross-over energy has a large uncertainty. The warm-target model
requires the measurement of the (warm) flare temperature, which
is measured at lower values from DEMs at EUV wavelengths than
from hard X-ray spectra observed with RHESSI.
Moreover, the spatial temperature
distribution is very inhomogeneous and the location with the
dominant temperature component relevant for the warm-target
collisional energy loss may be a mixture of colder pre-flare
plasma in active regions and heated evaporating flare plasma
at the location of instantaneous electron precipitation.
In summary, the value of the low-energy cutoff is strongly
dependent on the assumed warm-target temperature,
for which no physical model is established yet.}
In this study we investigated also the temporal evolution of the
low energy cutoff $e_c(t)$, for instance as shown in Fig.~1j, 2j, and 3j,
but we do not recognize a systematic pattern how the evolution
of this low-energy cutoff is related to other flare parameters.
\section{ CONCLUSIONS }
The energy partition study of Emslie et al.~(2012) was restricted to
38 large solar eruptive events (SEE). In a more comprehensive study
on the global flare energetics we choose a dataset that contains
the 400 largest (GOES M and X-class) flare events observed during
the first 3.5 years of the SDO era. Previously we determined the
dissipated magnetic energies $E_{\mathrm{mag}}$ in these flares based on
fitting the {\sl vertical-current approximation of a nonlinear force-free
field (NLFFF)} solution to the loop geometries detected in EUV
images from SDO/AIA, a new method that could be applied to 177 events
with a heliographic longitude of $\le 45^\circ$ (Paper I).
We also determined the thermal energy $E_{\mathrm{th}}$ in the soft X-ray
and EUV-emitting plasma during the flare peak times based on a
multi-temperature differential emission measure DEM forward-fitting
method to SDO/AIA image pixels with spatial synthesis,
which was applicable to 391 events
(Paper II). In the present study we determined the nonthermal energy
$E_{\mathrm{nt}}$ contained in accelerated electrons based on spectral
fits to RHESSI data using the OSPEX software, which was applicable
to 191 events. The major conclusions of the new results emerging
from this study are:
\begin{enumerate}
\item{The (logarithmic) mean energy ratio of the nonthermal
energy to the total magnetically dissipated flare energy is found to be
$E_{\mathrm{nt}}/E_{\mathrm{mag}}=0.41$, with a logarithmic standard
deviation corresponding to a factor of $\approx 8$, which yields
an uncertainty $\sigma/\sqrt{N}=0.41/\sqrt{191}=0.03$ for
the mean, i.e., $E_{\mathrm{nt}}/E_{\mathrm{mag}}=0.41\pm0.03$.
The majority ($\approx 85\%$) of the flare events fulfill the
inequality $E_{\mathrm{nt}}/E_{\mathrm{mag}} < 1$, which suggests
that magnetic energy dissipation (most likely by a magnetic
reconnection process) provides sufficient energy to accelerate
the nonthermal electrons detected by bremsstrahlung in hard X-rays.
Our results yield an order of magnitude higher electron acceleration
efficiency than previous estimates, i.e., $E_{\mathrm{nt}}/E_{\mathrm{mag}}
=0.03\pm 0.005$ (with $N=37$, Emslie et al.~2012).}
\item{The (logarithmic) mean of the thermal energy $E_{\mathrm{th}}$
to the nonthermal energy $E_{\mathrm{nt}}$ is found to be {\bf
$E_{\mathrm{th}}/E_{\mathrm{nt}}=0.15$, with a logarithmic
standard deviation corresponding to a factor of $\approx 7$.
The fraction of flares with a thermal energy being
smaller than the nonthermal energy, as expected in the thick-target
model, is found to be the case for $\approx 85\%$ only. Therefore,
the thick-target model is sufficient to explain the full amount of
thermal energy in most flares, in the framework of the warm-target
model. The cross-over method shows the opposite tendency,
but we suspect that the cross-over method over-estimates the
low energy cutoff and under-estimates the nonthermal energies.
Previous estimates yielded a similar ratio, i.e.,
$E_{\mathrm{th}}/E_{\mathrm{nt}}=0.15$ (Emslie et al.~2012).}}
\item{A corollary of the two previous conclusions is that the
thermal to magnetic energy ratio is $E_{\mathrm{th}}/E_{\mathrm{mag}}=0.08$. A total
of $95\%$ flares fulfils the inequality $E_{\mathrm{nt}}/E_{\mathrm{mag}} < 1$,
indicating that all thermal energy in flares is supplied by
magnetic energy. Previous estimates were a factor of 17 lower,
i.e., $E_{\mathrm{th}}/E_{\mathrm{mag}}=0.0045$ (Emslie et al.~2012),
which would imply a very inefficient magnetic to thermal energy conversion
process.}
\item{The largest uncertainty in the calculation of nonthermal
energies, the low-energy cutoff, is found to yield different values
for two used methods, i.e., $e_{\mathrm{wt}}=6.2 \pm 1.6$ keV
for the warm thick-target model, versus $e_{\mathrm{co}}=21 \pm 6$ keV
for the thermal/nonthermal cross-over method. The calculation of
the nonthermal energies is highly sensitive to the value of the
low-energy cutoff, which strongly depends on the assumed
(warm-target) temperature.}
\item{The flare temperature can be characterized with three different
definitions, for which we found the following ($67\%$-standard deviation)
ranges: $T_{\mathrm{AIA}} \approx 3-14$ MK for the AIA DEM peak temperature,
$T_w \approx 20-30$ MK for the emission measure-weighted temperatures, and
$T_{R} \approx 17-36$ MK for the RHESSI high-temperature DEM tails.
The median ratios are found to be $T_{\mathrm{AIA}}/T_w=0.31$ and
$T_{R}/T_w=0.90$.
{\bf The mean active region temperature evaluated from DEMs with AIA,
$T_e=8.6$ MK, is used to estimate the low-energy cutoff $e_c$ of the
nonthermal component according to the
warm-target model, i.e., $e_c \approx \delta (k_B T_{R})$.
The low-energy cutoff $e_c$ of the nonthermal spectrum has
a strong functional dependence on the temperature $T_{R}$.}}
\end{enumerate}
In summary, our measurements appear to confirm that the magnetically
dissipated energy is sufficient to explain thermal and nonthermal
energies in solar flares, which strongly supports the view that
magnetic reconnection processes are the primary energy source of flares.
The nonthermal energy, which represents the primary energy source of the
thick-target model, {\bf is sufficient to explain the full amount of
thermal energies in 71\% of the flares, according to the novel
warm-target model (Kontar et al.~2011). However, the derived
nonthermal energies are highly dependent on the
the assumed temperature in the warm-target plasma, for which
a sound physical model should be developed (see for instance
Appendix A and B), before it becomes a useful
tool to estimate the low-energy cutoff of nonthermal energy spectra.}
Future studies of this global flare energetics project
may also quantify additional forms of energies, such as the kinetic
energy in CMEs, and radiated energies in soft X-rays, EUV, and
white-light (bolometric luminosity).
\bigskip
\acknowledgements
We acknowledge useful comments from an anonymous referee
and discussions with Brian Dennis,
Gordon Emslie, Iain Hannah, Ryan Milligan, Linhui Sui, Daniel Ryan,
Richard Schwartz, Alexander Warmuth, and software support from
Kim Tolbert and Samuel Freeland.
This work was partially supported by NASA contract
NAS5-98033 of the RHESSI mission through University of California,
Berkeley (subcontract SA2241-26308PG), and
by NASA contract NNG 04EA00C of the SDO/AIA instrument.
AC and JMM were also supported by NASA grant NNX15AK26G.
\subsection*{ APPENDIX A : Collisional Time-of-Flight Model }
We can derive a collisional time-of-flight model for the thermal/non-thermal
cross-over energy that is complementary to the warm-target model of Kontar
et al.~(2015). For stochastic acceleration models, where particles
gain and lose energy randomly, the collisional deflection time yields
an upper time limit during which a particle can be efficiently accelerated.
The balance between acceleration and collisions can lead to the
formation of a kappa-distribution according to some solar flares models
(Bian et al.~2014). For solar flares, we can thus estimate the
cross-over energy between collisional and collisionless electrons
by setting the collisional deflection time $t_{\mathrm{defl}}$,
$$
t_{\mathrm{defl}} \approx 0.95 \times 10^8
\left( {e_{\mathrm{keV}}^{3/2} \over n_e} \right)
\left( { 20 \over \ln \Lambda} \right) \ ,
\eqno(A1)
$$
where $\ln \Lambda \approx 20$ is the Coulomb logarithm,
equal to the (relativistic) time-of-flight propagation time
between the coronal acceleration site and the chromospheric
thick-target energy loss site,
$$
t_{\mathrm{TOF}} = {L_{\mathrm{TOF}} \over v} = {L_{\mathrm{TOF}} \over \beta c} \ ,
\eqno(A2)
$$
where the relativistic speed $\beta = v/c$,
$$
\beta = \sqrt{ 1 - {1 \over \gamma^2} } \ ,
\eqno(A3)
$$
is related to the kinetic energy $e_{\mathrm{kin}}$ of the electron by
$$
e_{\mathrm{kin}} = m_e c^2 (\gamma - 1) = 511\ (\gamma - 1) \ {\rm [keV]} \ ,
\eqno(A4)
$$
where $\gamma$ represents here the relativistic Lorentz factor
(not to be confused with the spectral slope of the photon spectrum
used above, i.e., Eq.~1). So, setting these two time scales equal,
$$
t_{\mathrm{defl}} = t_{\mathrm{TOF}} \ ,
\eqno(A5)
$$
yields the relationship, using $\ln \Lambda \approx 20$,
$$
(\gamma - 1)^{3/2} (1 - {1 \over \gamma^2} )^{1/2}
= { L_{\mathrm{TOF}}\ n_e \over 0.95 \times 10^8 \times 511^{3/2}\ c } \ .
\eqno(A6)
$$
Using the low-relativistic approximation (for $\gamma \lower.4ex\hbox{$\;\buildrel >\over{\scriptstyle\sim}\;$} 1$),
$$
(\gamma - 1)^{3/2} (1 - {1 \over \gamma^2} )^{1/2} =
(\gamma - 1)^{3/2} {(\gamma-1)^{1/2} (\gamma+1)^{1/2} \over \gamma} =
{(\gamma - 1)^2 (\gamma+1)^{1/2} \over \gamma}
\approx (\gamma - 1)^2 \sqrt{2} \ ,
\eqno(A7)
$$
we obtain,
$$
(\gamma-1)^2 \ \sqrt{2} \approx 0.003 \times
\left( {L_{\mathrm{TOF}} \over 10^9 \ {\rm cm}} \right)
\left( {n_e \over 10^{11} \ {\rm cm}^{-3}} \right)
\quad {\rm [keV]} \ .
\eqno(A8)
$$
and by inserting $(\gamma-1) = e_{c}/511$ keV from Eq.~(A4) we find
the cross-over energy $e_c \approx e_{\mathrm{kin}}$
can be explicitly expressed as
$$
e_c \approx 24
\left( {L_{\mathrm{TOF}} \over 10^9 \ {\rm cm}} \right)^{1/2}
\left( {n_e \over 10^{11} \ {\rm cm}^{-3}} \right)^{1/2}
\quad {\rm [keV]} \ .
\eqno(A9)
$$
This expression requires the measurement of a mean length scale
$L_{\mathrm{TOF}}$ of flare loops and an average electron density $n_e$
where electrons propagate.
Turning the argument around predicts a time-of-flight distance
$L_{\mathrm{TOF}} \propto e_c^2/n_e$ as a function of the low-energy cutoff
$e_c$, which is a similar concept that has been applied to model
the size $L$ of the acceleration region as a function of the electron
energy $e$, i.e., $(L-L_0) \propto e^2/n_e$ (Guo et al.~2012a,b; 2013;
Xu, Emslie, and Hurford 2008).
\subsection*{ APPENDIX B : The Rosner-Tucker-Vaiana Model }
At the peak time of a flare, an energy balance between plasma
heating and cooling occurs at the turnover point of the temperature
maximum (Aschwanden and Tsiklauri 2009), which corresponds
to the scaling law of Rosner, Tucker, and Vaiana (1978a) that was
originally applied to steady-state heating of coronal loops,
where an energy balance between the heating rate and the conductive
and radiative cooling time is assumed. The RTV scaling law,
$T^3 \propto p L$, can be expressed in terms of the ideal gas pressure
$p = 3 n_e k_B T$, which yields for the loop apex temperature $T_{\mathrm{RTV}}$,
$$
T_{\mathrm{RTV}} = 0.0011 \ (n_e L_{\mathrm{RTV}})^{1/2} \ .
\eqno(B1)
$$
The loop half length and time-of-flight distance scale approximately
with the flare size, $L_{\mathrm{TOF}} \approx L_{\mathrm{RTV}} \approx L$.
Interestingly, the parameter combination $(n_e L)^{1/2}$ occurs
also in the expression for the collisional low-energy cutoff
(Eq.~A9), so that we can insert the RTV scaling law and obtain
an expression for the low-energy cutoff energy $e_c$ that depends
on the temperature $T_{\mathrm{RTV}}$ only,
$$
e_c \approx 25 \ (k_B T_{\mathrm{RTV}}) \quad {\rm [keV]} \ ,
\eqno(B2)
$$
which is similar to the result of the warm-target model (Eq.~8).
However, while the warm-target model is applied to the evaporating
upflowing flare plasma, which has temperatures of $T_e \approx 10-25$ MK,
the collisional deflection model should be applied to the temperature
of the cooler preflare loops, where the accelerated particles propagate
from the acceleration site to the thick-target site. These cooler
preflare loops may have typical coronal temperatures of
$T_{\mathrm{RTV}} \approx 5-6$ MK ($\approx 0.43-0.52$ keV) in active
regions (Hara et al.~1992), which predicts
then low-energy cutoff energies of $e_c=11-13$ keV.
If the time-of-flight distance $L_{\mathrm{TOF}}$ is corrected for
magnetic twist and the pitch angle of the electrons, the effective
time-of-flight distance is about $L_{\mathrm{TOF}} \lower.4ex\hbox{$\;\buildrel <\over{\scriptstyle\sim}\;$} 2 L$
(Aschwanden et al.~1996), which
increases the low-energy cutoff energy by a factor of $\sqrt{2}$,
predicting values of $e_c=15-18$ keV. Combining Eqs.~(8) and (B2), the
RTV model predicts a relationship between the preflare temperature
$T_{\mathrm{pre}}=T_{\mathrm{RTV}}$ and the (maximum)
flare temperature $T_{\mathrm{flare}}$,
$$
T_{\mathrm{pre}} \approx T_{\mathrm{flare}}
\left( {\delta \over 25} \right) \ ,
\eqno(B3)
$$
which yields $T_{\mathrm{pre}} \approx (0.12 - 0.24) \ T_{\mathrm{flare}}$
for a range of spectral slopes $\delta \approx 3-6$. Given the fact
that flare temperatures are typically found in the range of
$T_{flare} \approx 10-25$ MK, while preflare temperatures amount
to typical coronal temperatures in active regions, $T_{pre} \approx
1-4$ MK, we would expect indeed temperature ratios of
$T_{pre}/T_{flare} \approx 0.1-0.16$.
\subsection*{ APPENDIX C : The Runaway Acceleration Model }
Some particle acceleration models involve DC electric fields
that accelerate electrons and ions out of the bulk plasma.
Since the frictional drag on the electrons decreases with
increasing particle velocity ($\nu \propto v^{-3}$),
electrons in the initial thermal distribution with a high
enough velocity will not be confined to the bulk current,
but will be freely accelerated out of the thermal distribution
(Kuijpers et al.~1981; Holman 1985),
a process that is called runaway acceleration.
A thermal electron of velocity $v_e$ will run away if the
electric field strength is greater than the Dreicer field $E_D$,
$$
E_D = {m \over e} v_e \nu_e \ ,
\eqno(C1)
$$
where $m$ is the electron mass, $e$ the electron charge,
$v_e$ the electron velocity, and $\nu_e$ the electron collision
frequency. Since the square of the (non-relativistic) speed $v_e$
scales with the kinetic energy, $E_{\mathrm{kin}} = (1/2) m_e v_e^2$,
the critical runaway energy $E_{\mathrm{ra}}$ can be characterized by the
ratio of the critical velocity $v_e$ to the thermal speed $v_{\mathrm{th}}$,
$$
E_{\mathrm{ra}} = E_{\mathrm{th}} \left( { v_e \over v_{\mathrm{th}}} \right)^2 \ ,
\eqno(C2)
$$
We can associate this critical runaway energy $E_{\mathrm{ra}}$
with the low-energy cutoff $e_c$ and obtain again a relationship
that scales with the plasma temperature $T_e$ for a given critical
velocity ratio,
$$
e_c \approx E_{\mathrm{ra}} = k_B T_e \ \left( { v_e \over v_{\mathrm{th}}} \right)^2
\quad {\rm [keV]} \ .
\eqno(C3)
$$
Thus, for a typical velocity ratio of $(v_e/v_{\mathrm{th}}) \approx$ 2$-$3
and a plasma temperature range of $T_e \approx 5-6$ MK
$\approx 0.43-0.52$ keV in active regions, this model predicts a
range of $e_c \approx 1.7-8.3$ keV. Combining the relationships
of the warm-target model (Eq.~8) and the runaway acceleration model
(Eq.~C1) yields then a prediction for the nonthermal speed ratio
of the runaway electrons,
$$
\left( {v_e \over v_{\mathrm{th}}} \right) \approx \sqrt{\delta}
\approx (1.7 - 2.4) \ ,
\eqno(C4)
$$
which is consistent with solar parameters used in runway models
(Kuijpers et al.~1981; Holman 1985). Implications of runway
acceleration models for sub-Dreicer and super-Dreicer fields
are discussed also in Guo, Emslie and Piana (2013) and
Miller et al. (1997).
|
\section{Introduction and Experimental Setup}
Nuclear spin clocks, based on the detection of free spin precession of gaseous nuclear polarized $^{3}$He and $^{129}$Xe atoms with LT$_C$ SQUIDs as low-noise magnetic flux detectors are used as ultra-sensitive probe for nonmagnetic spin interactions, since the magnetic interaction (Zeeman term) drops out in the case of co-located spin samples (comagnetometry). Measurements of uninterrupted precession of one day can be achieved at the present stage of investigation due to long spin-coherence times. The principle of measurement is to search for sidereal variations of the precession frequency of co-located spin species while the Earth and hence the laboratory reference frame rotates with respect to distant stars. \\
In the context of the Standard-Model Extension (SME)\cite{Kostelecky,Colladay,Kostelecky2}, couplings of the neutron or proton spin $\vect{\sigma^{n, p}}$ to relic background fields $\vect{\tilde {b}^{n,p}}$ are discussed. The background fields have distinct directions in space and correspond to preferred spin directions. These couplings with the potential $V=\vect{\tilde {b}^{n,p}}\cdot \vect{\sigma^{n, p}} $ are purely non-magnetic, but change the energy levels of spins in a magnetic field, which can be detected by changes in the Larmor frequency of precessing spins.\\
To give a short overview of the setup (details are given in Ref.~\cite{Allmendinger}): The two polarized gas species (and N$_2$ as a buffer gas) are filled into a low-relaxation spherical glass cell with radius $R=5$~cm. Typically, the optimum conditions in terms of long transverse relaxation times $T_2^*$ and high Signal-to-Noise Ratio are met at a gas mixture with pressures of $p_{He}=3$~mbar, $p_{Xe}=5$~mbar, $p_{N2}=25$~mbar. The cell is positioned in a homogeneous static magnetic field of about 400~nT that is generated by Helmholtz coils mounted inside the strongly magnetically shielded room \textit{BMSR-2} at the \textit{Physikalisch-Technische Bundesanstalt} in Berlin.
At that field strength, the Larmor frequencies of $^{3}$He and $^{129}$Xe are about $\omega _{He}\approx 2\pi \cdot 13$~Hz and $\omega _{Xe}\approx 2\pi \cdot 4.7$~Hz, respectively. To measure these precession frequencies very precisely, low-noise low-temperature DC-SQUID gradiometers are used as magnetic flux detectors. Due to the very low field gradients in the order of pT/cm at the location of the cell, the transverse relaxation times reached $T_2^*=8.5$~h for $^{129}$Xe and up to $T_2^*=100$~h for $^{3}$He~\cite{Allmendinger}. The measured signal amplitudes at the beginning of the measurement were up to $A_\text{He}=20$~pT and $A_\text{Xe}=8$~pT for $^3$He and $^{129}$Xe, respectively. The noise level (combination of four gradiometers) was $\rho=3$~fT/$\sqrt{\text{Hz}}$. Due to the long spin-coherence time and the high initial Signal-to-Noise Ratio, the spin precession could be monitored for more than one day, which improves the sensitivity remarkably.
\section{Data Evaluation and Results}
To be sensitive to tiny nonmagnetic interactions, one has to consider the weighted difference of the respective Larmor frequencies of the co-located spin samples, or the corresponding time integral, the weighted phase difference, which are defined by
\begin{equation}
\label{eqn:weightedpd}
\Delta \omega =\omega_{\rm He} -\frac{\gamma_{\rm He} }{\gamma_{\rm Xe} }\omega _{\rm Xe}~~\text{and}~~\Delta \Phi =\Phi_{\rm He} -\frac{\gamma_{\rm He} }{\gamma _{\rm Xe}
}\Phi _{\rm Xe}~~.
\end{equation}
In doing so, magnetic field fluctuations are canceled, i.e. in principle $\Delta \omega =0$ and $\Delta \Phi=$const. if there are no further interactions. However, on a closer look, $\Delta \Phi $ is not constant in time, as
higher order effects have to be take into account. These can be parameterized by
\begin{eqnarray}
\Delta \Phi (t)&=&c_0 +c_1t+E_{\rm He} e^{-t/T_{\rm 2,He}^\ast
}+E_{\rm Xe} e^{-t/T_{\rm 2,Xe}^\ast } +F_{\rm He} e^{-2t/T_{\rm 2,He}^\ast }+F_{\rm Xe} e^{-2t/T_{\rm 2,Xe}^\ast }~~.
\label{eq2}
\end{eqnarray}
The linear contribution stems from Earth's rotation (i.e. the rotation of the SQUID detectors with respect to the precessing spins) and from chemical shift (diamagnetic shielding: the electron shells shield the nuclei against the external magnetic guiding field). The four exponential terms account for the Ramsey-Bloch-Siegert shift \cite{Bloch,Ramsey}. These effects are discussed in Ref.~\cite{Allmendinger}.
Finally, the function in Eq. (\ref{eq2}) together with the appropriate parameterization of the Lorentz-invariance-violating effect - in this case a sidereal modulation $\propto \tilde{b}\cdot\sin(\Omega_S\cdot t+\varphi_0)$ - is fitted to the combined weighted phase difference data of all measurement runs (7 in total). The resulting estimate on sidereal modulation is compatible with zero within the correlated and uncorrelated uncertainties and can be expressed as an upper limit on the magnitude of the hypothetical background field:
\begin{eqnarray}
\label{eqn:results}
\tilde{b}^{\text{n}}_{\bot}&<&6.7 \cdot 10^{-34}\text{ GeV} \text{ (68\% C.L.)~~.}
\end{eqnarray}
In Ref. \cite{Stadnik}, Y. Stadnik and V. Flambaum showed that the $^3$He-$^{129}$Xe comagnetometer is also sensitive to the proton interaction parameters of the SME \cite{Stadnik}. Based on our measurements the following values were derived:
\begin{eqnarray}
\label{eqn:protonresults}
\nonumber \tilde{b}^{\text{n}}_X+0.74\cdot\tilde{b}^{\text{p}}_X&=&(7.1\pm8.2)\cdot 10^{-34}\text{ GeV}\\
\tilde{b}^{\text{n}}_X+0.74\cdot\tilde{b}^{\text{p}}_Y&=& (5.0\pm 10.8) \cdot 10^{-34}\text{ GeV}~~~.
\end{eqnarray}
The corresponding upper limit of the equatorial component $\tilde{b}^{\text{p}}_{\bot}$ of the background tensor field interacting with the spin of the bound proton is
\begin{eqnarray}
\label{eqn:protonresults2}
\tilde{b}^{\text{p}}_{\bot}&<1.6 \cdot 10^{-33}\text{ GeV} &\text{(68\% Confidence Level)}~.
\end{eqnarray}
\section{Conclusion and Outlook}
Freely precessing gaseous, nuclear polarized $^{3}$He and $^{129}$Xe samples can be used as ultra-sensitive probe for nonmagnetic spin interactions, since the magnetic interaction (Zeeman term) drops out in the case of co-located spin samples. With a similar setup, upper limits on interactions mediated by axion-like particles were obtained~\cite{Tullney}.\\
The next step is to apply this method to search for a CP-violating permanent electric dipole moment (EDM) of $^{129}$Xe: A permanent EDM $\vect{d}$ of a fundamental or composite particle must be aligned parallel to the spin, as the spin is the only available vector for an eigenstate of the isolated particle. Thus, for a magnetic guiding field aligned along the z-direction, the Hamiltonian has the form $H=-\mu\cdot B_0-d \cdot E_z$, with the corresponding frequency shift $\delta\omega_\text{EDM}=\frac{2}{\hbar} d \cdot E_z$. By varying the z-component of the electric field $E_z$, the frequency shift is modulated correspondingly. As mentioned before, the principle of comagnetometry is applied to become insensitive to drifts of the magnetic guiding field. Subsequently, a non-zero EDM will manifest in a modulation of the weighted phase difference, and the corresponding value $d$ can be extracted.\\
The experimental setup has been changed to enable the measurement of the $^{129}$Xe EDM: The measurement cell has a cylindrical shape. The end planes are made of silicon and the lateral surface is composed of the low-relaxation GE-180 glass. The distance between the electrodes is 5~cm and the maximum voltage that can be applied is 12~kV. A small amount of SF$_6$ (a few mbar) is added to the gas mixture to suppress leakage currents. It is highly beneficial to maximize the coherent measurement time $T$, as the uncertainty in frequency determination (and thus the error on the EDM) decreases as $T^{-3/2}$ for white noise. One important mechanism that reduces the spin-coherence time is caused by magnetic field gradients across the measurement cell. Consequently, additional gradient coils have been integrated into the setup to compensate residual magnetic field gradients of the mu-metal shielding.
|
\section{Introduction}
In astrophysical context, the process in which diffuse gas or matter is accumulated around a compact
object under an influence of gravity is called accretion. The importance of accretion as the source
of steady or unsteady emission of radiation was first widely recognized in explaining
observations of binary systems, especially X-ray binaries. It is very important to understand
hydrodynamic properties of matter in the vicinity of a black hole
as the emitted radiation mainly depends on the density, velocity and temperature at each flow element at each instant.
Unique inner boundary condition which enables all infalling matter to cross the event horizon with the speed of light
$c$ (e.g., Chakrabarti, 1996) also makes all the accretion flows into black holes to be transonic in nature forcing
them to have the innermost sonic point in regions of strong gravity, i.e., just outside of the event horizon.
Due to centrifugal pressure, inflowing
matter slows down and piles up closer to the black hole forming a torus like structure at the inner part of the disc.
This torus like structure formed between the centrifugally supported shock and the innermost sonic point outside the
horizon is widely known as the Centrifugal pressure supported BOundary Layer or simply CENBOL
(Chakrabarti, 1989, hereafter C89; 1999; Molteni, Lanzafame \& Chakrabarti, 1994; MLC94).
In the limit of no radial velocity, the CENBOL would have been termed as a thick accretion
disk (Paczy\'nski \& Wiita, 1980) as shown by MLC94.
Earlier, a large number of numerical simulations of inviscid accretion flows around black holes,
have been presented by various research groups (e.g., Hawley, Wilson \& Smarr 1984;
Eggum, Coroniti, \& Katz 1985; Chakrabarti \& Molteni, 1993; MLC94; Ryu, Brown, Ostriker \& Loeb 1995;
Molteni, Sponholz \& Chakrabarti 1996; Molteni, Ryu \& Chakrabarti, 1996, hereafter, MRC96;
Ryu, Molteni \& Chakrabarti, 1997, hereafter RMC97; Igumenshchev, Abramowicz \& Narayan 2000;
Chakrabarti, Acharya \& Molteni, 2001, hereafter CAM01; Giri et al., 2010, hereafter GC10).
However, these simulations were performed assuming that the flow has an equatorial
symmetry, and therefore the flow behavior was studied only on one quadrant, i.e.,
first quadrant using the standard `reflection' boundary condition on the
equatorial plane. In MLC94 and GC10, the results of standing and oscillating shock
formations in inviscid flows are presented using Smoothed Particle Hydrodynamics (SPH)
method and finite difference method respectively. In GC10, several simulations have been
carried out choosing two conserved flow parameters (namely, specific energy ${\it E}$
and specific angular momentum $\lambda$) from the parameter space which provides complete
set of solutions of a black hole accretion flow (C89). In order to break the reflection
symmetry along equatorial plane, several simulations have been carried out by various groups
for both black hole accretion (Molteni et al., 2001, hereafter M01; Chakrabarti, Acharya \& Molteni, hereafter CAM01) as well as
stellar wind accretion onto stars (Fryxell \& Taam, 1988 ; Taam \& Fryxell, 1989; Matsuda et al., 1991, 1992).
In M01 and CAM01, using SPH, it was shown that an instability
can occur in the flow. They also demonstrated that although matter is supplied symmetrically, those instabilities may not remain
symmetric with respect to the equatorial plane. Furthermore, there is a strong interaction of the outgoing wind with the incoming flow
(M01). However, SPH is known to be dissipative in nature and it is not impossible that in energy conserving schemes one might
see that such oscillations are actually disrupting the flow altogether. We therefore extend the work of GC10 where energy
is accurately preserved by removing the reflection condition along the equatorial plane. By this procedure, we intend to give
answers to the following important questions: (a) Will the accretion flow be symmetric with
respect to the equatorial plane? And if so, under what conditions? (b) Will this two quadrant flow have
any effect on the formation of the so called `CENBOL'? This question is especially relevant
as the CENBOL acts as the Compton cloud (Chakrabarti \& Titarchuk, 1995, hereafter CT95)
while explaining the spectral and timing properties of black hole candidates.
(c) If the flow symmetry is absent then will the accretion flow remain stable
at all or the flow would be violent and disrupted? (d) What will be its effects
on outflows which are known to be produced on the CENBOL surface?
The plan of our paper is the following: in the next Section, we present the model
equations governing the flow. In Section 3, we describe the methodology for our
simulations using a grid based finite difference technique, called total variation diminishing
(TVD) method which was discovered by Harten (1981). In Section 4, we discuss our
simulation results and compare those with earlier simulations. Finally,
in Section 5, we make concluding remarks.
\section{Model Equations without Viscosity}
In our approach, we consider a two dimensional axisymmetric flow around a Schwarzschild black hole.
Instead of using general relativity, we use the pseudo-Newtonian potential first prescribed by
Paczy\'nski and Wiita (1980). This potential mimics the effects of general relativity very successfully (see, RCM97, GC10).
We use cylindrical polar coordinates ($r$, $\phi$ and $z$) for our calculations.
The mass, momentum and energy conservation equations in a compact form using non-dimensional
units are given in MRC96. We use the mass of the black hole
$M_{BH}$, the velocity of light $c$ and the Schwarzschild radius $r_g=2GM_{BH}/c^2$
as the units of the mass, velocity and distance respectively.
The equations governing the inviscid flow have been presented in Ryu et al. (1995), MRC96 and GC10 in great detail
and we do not repeat them here. In conservative form, the equations are given by,
$$
{\partial{\fam9 q}\over\partial t}+{1\over r}{\partial\left(r
{\fam9 F}_1\right)\over\partial r}+{\partial{\fam9 F}_2\over\partial r}
+{\partial{\fam9 G}\over\partial z} = {\fam9 S},
\eqno(1a)$$
where the state vector is
$${\fam9 q} = \left(\matrix{\rho\cr
\rho v_r\cr
\rho v_{\theta}\cr
\rho v_z\cr
E\cr}\right)_, $$
the flux functions are
$${\fam9 F}_1 = \left(\matrix{\rho v_r\cr
\rho v_r^2\cr
\rho v_{\theta}v_r\cr
\rho v_z v_r\cr
(E+p)v_r\cr}\right)\qquad
{\fam9 F}_2 = \left(\matrix{0\cr
p\cr
0\cr
0\cr
0\cr}\right)\qquad
{\fam9 G} = \left(\matrix{\rho v_z\cr
\rho v_r v_z\cr
\rho v_{\theta} v_z\cr
\rho v_z^2+p\cr
(E+p)v_z\cr}\right)_, \eqno(1b)$$
and the source function is
$$
{\textfont1 = \twelvei
\scriptfont1 = \twelvei \scriptscriptfont1 = \teni
\def\fam1{\fam1}
{\fam9 S} = \left(\matrix{0\cr
~~~\cr
{\rho v_{\theta}^2\over r}
-{\rho r\over2\left(\sqrt{r^2+z^2}-1\right)^2\sqrt{r^2+z^2}}\cr
~~~\cr
~~~\cr
-{\rho v_r v_{\theta}\over r}\cr
~~~\cr
-{\rho z\over2\left(\sqrt{r^2+z^2}-1\right)^2\sqrt{r^2+z^2}}\cr
~~~\cr
~~~\cr
-{\rho \left(rv_r+zv_z\right)\over
2\left(\sqrt{r^2+z^2}-1\right)^2\sqrt{r^2+z^2}}\cr}\right)_.}
\eqno(1c)$$
Here, expression for energy density $E$ (without potential energy) is given by,
$$E=p/(\gamma-1)+\rho(v_r^2+v_{\theta}^2+v_z^2)/2,$$ $\rho$ is the mass density,
$\gamma$ is the adiabatic index, $p$ is the pressure, $\rho$ is mass density,
$v_r$, $v_\theta$ and $v_z$ are the radial, azimuthal and vertical component of velocity
respectively. In case of an axisymmetric inviscid flow, the equation for azimuthal
component of the momentum simply signifies the conservation of specific angular momentum $\lambda$,
$$
\frac{d\lambda}{dt}=0.
$$
The general form of the equations of the flow in an inertial reference frame (Batchelor, 1967) is given by,
$$
\rho [{\partial {\bf v} \over \partial t} + {{\bf v} . {\nabla {\bf v}}}] =
- { \nabla p} + {\bf {F_b}} + {\nabla . {\bf {\tau}}}, \eqno(2)
$$
where, ${\bf v}$ is the flow velocity, ${\bf \tau}$ is the viscous stress tensor, and ${\bf {F_b}}$
represents body forces (per unit volume) acting on the fluid and ${\nabla}$ is the Del operator.
Typically, the body forces consist of only gravity forces, but may include other
types (such as electromagnetic forces). Since, here we are considering only inviscid and non-magnetic flow,
detailed discussion on viscous stress tensor is out of the scope of this paper.
To describe the gravitational field around a Schwarzschild Black hole, we use the Pseudo-Newtonian gravitational
field of a point mass $M_{BH}$ located at the centre in cylindrical coordinates $[r,\theta,z]$ described
by Paczy\'{n}ski \& Wiita (1980) potential,
$$
\phi(r,z) = -{GM_{BH}\over2(R-r_g)}, \eqno{(4)}
$$
where, $R=\sqrt{r^2+z^2}$.
\section{Methodology and simulation setup}
\begin{figure}
\begin{center}
\includegraphics[width=8cm]{fig1.pdf}
\caption{A schematic diagram of the system under consideration. Solid box is our computational
region having a dimension of $0 \leq r \leq 200$ and $-200 \leq z \leq 200$ on the $r-z$ plane.
Black hole is sitting at the origin of $r$ \& $z$ axis. Matter is injected from
the outer boundary and matter is sucked into the black hole. No reflection symmetry along
the equatorial plane has been assumed.}
\end{center}
\end{figure}
The computational box occupies two quadrants (first and fourth quadrant) of the r-z plane with $0 \leq r \leq 200$
and $-200 \leq z \leq 200$. Here, we use the reflection boundary condition only on z-axis to
obtain the solution in other quadrants. This is because in all our simulations we assume
the flow to be axisymmetric in nature as the black hole itself is strictly
axisymmetric and the inner boundary condition is expected to force the flow to be axisymmetric close to the black hole.
The incoming gas enters the computational box through the outer boundary (having a vertical cylinder like shape
in three-dimensions), located at $r_b = 200$. The flow is injected symmetrically both above and below the equatorial plane.
We supply radial velocity $v_r$, the sound speed $a$ (i.e., temperature) of the flow at boundary points
from the hybrid model and boundary values of density $\rho$ from standard vertical equilibrium solution (C89).
We scale the density in such a way that the incoming gas has the density of ${\rho}_{in} = 1.0$.
In order to mimic the horizon of the black hole, we place an absorbing inner boundary at $R = 2.5 r_g$,
inside which all the incoming matter is absorbed completely into the black hole. The inner sonic point
is formed around this radius anyway, so this choice of inner boundary does not affect the flow dynamics
upstream. In order to avoid singularities caused by `division by zero', we fill the grid with a
background matter of very low density ${\rho}_{bg} = 10^{-6}$ having a sound speed (or, temperature)
to be the same as that of the incoming matter. Hence, the incoming matter has a pressure $10^6$
times larger than that of the background matter. Of course, this initial matter is totally washed out
and replaced by the injected matter within a dynamical time scale.
Initially, the low density matter with which the grids are filled, is assumed to be static, i.e.,
the values of radial ($v_r$), rotational ($v_\phi$)
and azimuthal ($v_z$) components are all chosen to be zero for all
the grids except those on the outer boundary. Thus, the Mach number
is zero everywhere except on the outer boundary at the beginning of the simulation.
The calculations were performed with a very high resolution $512 \times 1023$ grids. Thus,
each r and z-grid has a size of $0.3906$ in units of the Schwarzschild radius.
Figure 1 shows schematically our system on the $r-z$ plane in a cylindrical co-ordinate system.
In this work, we are interested to see time dependence of CENBOL and instability in the
accretion disk around a black hole. So, the resolution which we have ($ \sim 0.4$ Schwarzschild radii)
is enough to catch these salient features.
All the simulation cases have been carried out assuming a stellar mass black hole $M_{BH}= 10{M_\odot}$.
The conversion of our time unit to physical unit is $2GM_{BH}/c^3$, and thus
the physical time for which the programme was run would scale with the mass of the black hole.
We typically find that the infall time from the outer to the inner boundary is about $\sim 0.5$s. This is
computed from the sum of $dr/<v_r>$ over the entire radial grid, $<v_r>$ being averaged over $20$ vertical grids.
We carry out simulations for several hundreds of dynamical time-scales in order to stay away from any transient effects.
\section{Results}
\begin{figure}
\begin{center}
\includegraphics[width=8cm]{fig2.pdf}
\caption{Density and velocity vector map to show quasi-periodic formation and
deformation of the CENBOL at (a) t = 21.36 s, (b) 21.84 s, (c) 22.08 s and (d) 24.95 s. Specific angular
momentum is considered to be $\lambda = 1.6$. In (a), high density region due to centrifugal supported boundary
layer (CENBOL) has a symmetric shape. In (b), symmetry is about to be broken due to vertical oscillation of the
perturbing mass. In (c), CENBOL is deformed but not destroyed. In (d), the CENBOL is restored back. }
\end{center}
\end{figure}
\begin{figure}
\hfill
\subfigure[]{\includegraphics[width=8.5cm]{fig3a.pdf}}
\hfill
\subfigure[]{\includegraphics[width=8.5cm]{fig3b.pdf}}
\hfill
\caption{Radial distribution of the (a) radial velocity component and
(b) radial Mach number ($v_r/a$) on the equatorial plane. Time is $t = 24.95$ s and specific
angular momentum is $\lambda = 1.6$. We clearly see the slowing down of matter at the centrifugal
barrier (a) and a supersonic to sub-sonic transition (b) forming a shock.}
\end{figure}
GC10 carried out all their simulations assuming a reflection boundary condition on the equatorial plane and injected matter
only in one quadrant. M01, in their work with cooling processes, has shown that certain instabilities in the
accretion disk set in even in SPH simulations where angular momentum is preserved more accurately (see, MRC96, RCM97).
Chakrabarti et al. (2004) have shown the quasi-periodic variabilities caused due to vertical
and horizontal oscillation of shock waves in a two dimensional axisymmetric flow using Smooth
Particle Hydrodynamics (SPH). Our result in this paper, obtained using a code
with finite difference method which employs an accurate Total Variation Diminishing (TVD) scheme,
clearly supports this work. Following C89, we injected the flow at the outer boundary with matter
in vertical equilibrium. The injection rate of the momentum density is kept uniform throughout the injected height
at the outer edge. We stop the simulations at $t=95$s (physical time). This time is more than two
hundred times the dynamical time of the flow. Solutions presented in Fig. 2(a-d) are at 45-50
times the dynamical time, long after transient effects ($\sim 1$s) die out. Thus the effects seen are real
and are expected to influence spectral and timing properties significantly.\\
C89 predicted that the standing shocks can form if $\lambda >1.525$. We find that, indeed, CENBOL is produced
when $\lambda>1.5$, the `discrepancy' being perhaps due to the presence of turbulence pressure (generated by
the centrifugally bounced outward flow colliding with the infalling gas in post-shock region)
which helps the formation of CENBOL even at a lower angular momentum.
\begin{figure}
\begin{center}
\includegraphics[width=9.0cm]{fig4.pdf}
\caption{Density and velocity vector plots of the accretion flow to show instability
within the flow. (a) and (c) show that the CENBOL is shifted above: flow density is higher in the upper quadrant
and (b) and (d) show that the CENBOL is shifted below: flow density is higher in the lower quadrant.
Plots are drawn at $t= 17.34, 22.57, 40.08$, and $43.2$ seconds respectively.
Specific angular momentum is considered to be $\lambda=1.7$.}
\end{center}
\end{figure}
In Fig. 3, we show the distribution of radial component of the velocity and the Mach number which is the ratio
between radial velocity ($<v_r>$) and sound speed ($<a>$). Velocity and sound speed are averaged over $10$ grids
located on both sides of the equatorial plane. The plot is drawn for $t \sim 24.95$ s. The radial velocity
suddenly drops at the shock location where the flow also becomes subsonic as evidenced by the Mach number distribution.
As the time progresses, the shock tends to oscillate back and forth and also vertically.\\
\begin{figure}
\subfigure[]{\includegraphics[width=7.0cm]{fig5a.pdf}}
\subfigure[]{\includegraphics[width=7.0cm]{fig5b.pdf}}
\caption{Time variation of the ratio between the total outflow rate
(${\dot M}_{out}$) and the total inflow rate (${\dot M}_{in}$)
showing an anti-correlated behaviour. When the outflow rate from the upper boundary is high,
the outflow rate in the lower boundary is low and vice versa.
Here (a) $\lambda = 1.7$ and (b) $\lambda = 1.8$. Dot-dashed curve represents the outflow inflow rate ratio in upper quadrant,
solid curve represents the ratio in lower quadrant, and dashed curve represents the mean value (time averaged)
of the outflow inflow rate ratio. We note that the mean outflow rate as well as the degree of deviation from the mean is higher when
angular momentum is higher.}
\end{figure}
We now carry out a simulation with a larger specific angular momentum. We choose $\lambda = 1.7$.
The centrifugal force increases the location of the shock and thus the size of the
CENBOL is increased. However we notice that after a transient state, the CENBOL takes part in a vertical oscillation around the
equatorial plane. We also observe that the outflowing wind is interacting with the incoming accreting matter creating
weaker oblique shocks near top-right and
bottom-right corners. In Fig. 4(a-d), we plot density and velocity vector maps of the flow at $t = 17.34$ s, $22.57$ s,
$39.13$ s, and $43.23$ s respectively. The CENBOL, though distinct, takes a complex shape.
It is also evident that the outflow oscillates, i.e., there are times when outflow from the upper
quadrant is large and at other times the outflow in the lower quadrant is large.
Disks in Fig. 4(a) and Fig. 4(c) have the CENBOL
shifted slightly to upper quadrant and the outflow rate from the upper boundary is found to be larger. However,
disks in Fig. 4(b) and Fig. 4(d) have the CENBOL shifted slightly towards the lower quadrant and the outflow from the lower
boundary is found to be larger.
\begin{figure}
\subfigure[(i)]{\includegraphics[width=8.3cm]{fig6a.pdf}}
\subfigure[(ii)]{\includegraphics[width=8.3cm]{fig6b.pdf}}
\caption{Time variation of shock location for two quadrant flow with $\lambda = 1.6$
angular momentum and their respective power density spectra. Group (i) show the shock location
variation in upper (a) and lower (b)
quadrants. Group (ii) show the power density spectra (PDS) of these locations. We see evidence of a prominent peak at
$\sim 0.16$Hz in both the cases and a harmonics at $\sim 0.32$Hz.
}
\end{figure}
\begin{figure}
\subfigure[(i)]{\includegraphics[width=8.3cm]{fig7a.pdf}}
\subfigure[(ii)]{\includegraphics[width=8.3cm]{fig7b.pdf}}
\caption{Same as in Fig. 6 for $\lambda = 1.7$. Note that the shock oscillates around a larger mean location, though the
symmetry in upper and lower quadrant is lost. The power density spectra have several peaks and the oscillation is more chaotic. }
\end{figure}
\begin{figure}
\subfigure[(i)]{\includegraphics[width=8.3cm]{fig8a.pdf}}
\subfigure[(ii)]{\includegraphics[width=8.3cm]{fig8b.pdf}}
\caption{Same as in Fig. 6 for $\lambda = 1.8$. Note that the shock oscillates around a mean location similar to what we observed
for $\lambda=1.7$, perhaps due to the post-shock turbulences. The symmetry
in upper and lower quadrant is lost. The power density spectra have several peaks and the oscillation is more chaotic. }
\end{figure}
\\In Fig. 5, we plot the ratio of the outflow rate to constant injected rate. The dot-dashed curve represents
the ratio of the total outflow rate (${\dot M}_{out}$) and the total inflow rate (${\dot M}_{in}$) in the
upper quadrant, solid curve represents the ratio in lower quadrant, and dashed curve represents the mean value (time averaged)
of the ratio. In Fig. 5(a) and Fig. 5(b), we
use $\lambda=1.7$ and $1.8$ respectively. We note that outflow rate as well as the mean rate increases with
specific angular momentum. This further establishes that the outflow is centrifugally driven. The overall
rate is found to be about $5-10$\% of the inflow rate. We also note that the deviation from the mean rises
as well, indicating that for higher $\lambda$, stronger vertical oscillation sets in. This was also seen in the density and velocity plots.
These plots show highly anti-correlated behaviour between the rates from the upper and the lower boundaries.
\section{Discussions and concluding remarks}
In this paper, we presented extensive time dependent numerical simulations
of two quadrant accretion flow around a black hole.
In earlier studies, such as in MLC94 and GC10, similar type of simulations were carried out in one quadrant using SPH and TVD schemes respectively. In these earlier simulations, our goal was to check if the shocks could be produced in the first place, and if yes, how does the puffed up post-shock region behave in reference of thick accretion flow.
It was found that indeed, since the post-shock flow is sub-sonic with low radial velocity and sub-Keplerian angular momentum, it does behave as a thick accretion disk, though without a cusp since radial velocity increases close to the inner edge. This post-shock region (CENBOL) was then used to Comptonize the low energy photons from a Keplerian disk.
In the present work, we asked ourselves if the CENBOL really remains symmetric with respect to the equatorial plane. For this we removed the reflection symmetry imposed forcefully in earlier simulations. We inject matter only in first and the
fourth quadrants. We find that for lower centrifugal force, the CENBOL remains symmetric, though a vertical oscillation sets it which becomes more and more violent as the specific angular momentum increases. This is superimposed with a horizontal oscillation. We also find that the outflow rate from each of the quadrants independently vary: The quadrant in which the CENBOL is tilted, also has the higher rate. Thus, the rates in the two quadrants are anti-correlated.
There are two important aspects which need major discussions: First, the oscillations seen here are in radial
and vertical directions. In earlier works, especially in Ryu et al. (1997), it was shown that when the Rankine-Hugoniot
relations are not satisfied, the shocks undergo radial oscillations. This is because a transonic accretion flow with a
significant amount of angular momentum, has two physical sonic points and the flow is required to have higher
entropy in order to pass through the innermost one (Chakrabarti, 1989). This means that even when the conditions
of steady shock formation is not satisfied, the flow will pass through the inner sonic point and the shock will be unsteady, moving radially, and searching for an acceptable solution. Another reason of radial oscillation was found by Molteni et al. (1996a) where it was seen that shocks start to oscillate only when the cooling time scale in the post-shock region roughly matches with the
infall time scale in that region. Since cooling process is absent in the present simulation, this latter explanation is
not relevant in the present context. In presence of viscosity, oscillation due to viscous overstability (Kato, 1978) is
well known. But ours is an inviscid flow. Hence this is also not the reason of the oscillation seen in our simulations. In case of slender tori, inertial modes are rapidly excited faster than the dynamical time scale (Blaes, 2006) and causes significant instability
(Horak, 2012). However, our post-shock region is advecting and not slender. Perturbations due to the inertial modes
are likely to be advected out of the disk when the flow passes through the innermost sonic point.
In the present context, we note that the disk instability is high only when the angular momentum is very large.
Two important physical processes are triggered by angular momentum: (i) infalling matter hits the centrifugal barrier
(defined by the location where the centrifugal forces is similar to the gravitational force) and bounces back near the
equatorial plane. This flow confronts the incoming matter and two turbulence cells of opposite vorticity are generated, one above and the other below the equatorial plane. (ii) Centrifugal
pressure driven winds are formed which also flow outwards (in between the centrifugal barrier and
the so-called funnel wall, see, Molteni et al. 1996b), confronting the incoming flow away from the
equatorial plane. This interface is therefore susceptible to Kelvin-Helmholtz instability. In all the cases we ran,
we found that for very low angular momentum, the wind does not form at all and thus this instability is absent. Higher the angular momentum, stronger is the shear instability between the incoming and outgoing components. When the amplitude of the fastest growing mode becomes non-linear, instabilities in the upper and lower halves join and push the entire disk on one side or the other. This is what we believe to be the cause of the vertical motion. Clearly, this requires a thorough study.
The simulations we carried out are inviscid and thus the oscillations are more violent as the angular momentum is not transported away. Similarly, we have not included radiative cooling, because of which, the flow is very hot inside CENBOL. It has already been suggested that the shock oscillation contributes to quasi-periodic oscillations (QPOs) of black hole candidates. We believe that the vertical oscillation of CENBOL may contribute to the variability classes as seen in objects such as GRS 1915+105. The vertical erratic movement of the CENBOL could be responsible for the varying Comptonized component as seen by a specific observer and this may give rise to a large number of variability classes even in so-called hard states, soft states or intermediate states.
Giri \& Chakrabarti (2013) and Giri et al. (2015) have shown that an injected sub-Keplerian flow can distribute the angular momentum and produce and sustain a Keplerian disk when viscosity is higher than a critical value, giving rise to the CT95 configuration of two component advective flow. However, this conclusion was drawn using simulations with the upper quadrant only. The conclusions drawn in the present paper with serve as the basis of of the next work with viscosity and radiative cooling. Most importantly, it would be clear if the resulting standard disk also exhibit such an oscillation. This will be discussed elsewhere.
|
\section{Introduction}
Conformal higher spin (CHS) theory \cite{Fradkin:1985am,Fradkin:1989md,Tseytlin:2002gz,Segal:2002gd,Bekaert:2010ky,Beccaria:2014jxa,Haehnel:2016mlb}
is a remarkable example of a formally consistent (gauge-invariant, though higher-derivative and thus non-unitary)
higher spin model that has a local action with a flat-space vacuum and one dimensionless coupling constant.
It is naturally associated with another consistent higher spin theory, {\em i.e.} 2-derivative massless higher spin theory
in AdS space of one dimension higher. Starting with a free complex scalar equation in 4 dimensions\footnote{In this paper we shall concentrate on the $d=4$ case but most of our discussion may be generalised to even $d>4$.}
$ \partial^2 \vp=0$ one gets an infinite tower of conserved traceless totally symmetric higher spin currents
$J_{s}= \vp^* \J_s \vp \ , \ \ \J_s \sim \partial^s + ...$, \ $s=1,2,...,\infty$ that generalise the spin 0
primary operator $J_0 = \vp^*\vp$.
The conserved charges generate an infinite dimensional symmetry algebra of the free scalar equation \cite{Eastwood:2002su,Vasiliev:2003ev}
that is associated to a collection of conformal Killing tensors. The CHS theory may be viewed as a gauge theory of this higher spin global symmetry.
A closely related approach is based on interpreting
CHS fields as "sources" for the currents $J_{s}$ that then inherit the linearised gauge invariances
$\delta h_{\m_1...\m_s} = \partial_{(\m_1 } \epsilon_{\m_2 ... \m_s)} + \eta_{(\m_1\m_2} \alpha_{\m_3...\m_s)}$
which generalise the usual reparametrisations and Weyl symmetry of conformal gravity.
Starting with free $U(N)$ complex scalar CFT, adding source terms $h_s J_s$ to the free action $ \partial \vp^* \partial \vp$
and integrating over $\vp$ one finds for the generating functional
of correlators of the currents $J_s$
\begin{equation} \label{1}
\G[h]= N \,\log \det \big( \partial^2 + \sum_s h_s\,\J_s\big)\ ,\qquad \ \ \ \ \quad \J_s \sim \partial^s\ .
\end{equation}
From the vectorial AdS/CFT point of view \cite{Klebanov:2002ja} the 4d currents $J_s$ are dual to massless higher spins in AdS$_5$
and the generating functional $\G[h]$ should then be equal to the on-shell value of the AdS action with $h_s$
being the boundary values of the higher spin fields in AdS.
One can then obtain a local gauge-invariant action for the CHS fields $h_s$ by identifying it with the
logarithmically UV divergent part of the "induced" action \rf{1} \cite{Tseytlin:2002gz,Segal:2002gd,Bekaert:2010ky}
\begin{align}
& S[h] \sim \,\log \det \big( \partial^2 + \sum_s h_s\,\J_s\big)\Big|_{\log \Lambda} \no\\
&\quad \sim
{ 1 \over {g}^2} \sum_s
\int d^4 x \Big( h_s \partial^{2s } h_s + \partial^{s_1+s_2+s_3 - 2 } h_{s_1} h_{s_2} h_{s_3}
+ \partial^{ s_1+s_2+s_3 + s_4 - 4 } h_{s_1} h_{s_2} h_{s_3} h_{s_4} + ... \Big) \ .
\label{2}
\end{align}
Here we introduced an arbitrary dimensionless coupling constant ${g}$ and
indicated symbolically the overall powers of derivatives
in the kinetic and interaction terms that follow from dimensional analysis.
Indeed, as the 4d scalar $\vp$ has mass dimension 1, the current $J_s$ has dimension $2+s$
and thus the corresponding source field $h_s$ must have the "shadow"-field dimension $\Delta_s= 2-s $
(i.e. 1 for vector field, 0 for conformal graviton, etc.). This then determines the derivative structure of \rf{2}.\footnote{
The
fact that the powers of derivatives are directly correlated with the values of the spins in the vertex
(which is a consequence of the underlying conformal invariance) is an important simplifying feature of this theory
compared to the AdS higher spin theory and a hypothetic 2-derivative massless higher spin theory in flat space
that both contain a dimensional parameter.}
In particular, the presence of $2s$ derivatives in the kinetic term in \rf{2} is consistent
with both the above linearised gauge invariance $\delta h_s = \partial \epsilon_{s-1} + \eta_2 \alpha_{s-2}$ and the locality
of the action.\footnote{The kinetic term should contain the transverse traceless spin $s$ projector
$\Pi_s$ that is given by products of $s$ factors of $\Pi^\m_\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J} = \delta^\m_\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J} - { \partial^\m \partial_\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}\over \partial^2}$
and thus $\Pi_s \partial^{2s}$ is local.}
In addition to the linearised gauge symmetry and the standard conformal symmetry
the CHS action should be invariant under the full infinite dimensional CHS symmetry
\cite{Segal:2002gd} whose global part is the symmetry of free scalar 4d Laplace equation.
This large symmetry should
provide strong constraints on the corresponding classical and quantum theory.
For example, for fixed spectrum of the CHS fields
the action \rf{2} should be essentially unique (modulo field redefinitions)\footnote{Starting instead
from a free scalar
CFT with a free spinor or free Maxwell vector CFT one gets a different spectrum of conserved currents
and thus a different "induced" CHS theory.}
and thus renormalisable.
In view of the conformal symmetry being gauged here it should actually be
UV finite,
provided the theory is quantum-consistent, {\em i.e.\ } there are no conformal and higher symmetry anomalies.
An indication of a hidden simplicity of the CHS theory is the vanishing of the regularised
total number of its degrees of freedom, or, equivalently, the triviality of the free (one-loop) partition function in flat space
\cite{Beccaria:2015vaa}. This partition function vanishes
also on 4-sphere implying the vanishing of the
Weyl anomaly a-coefficient \cite{Giombi:2013yva,Tseytlin:2013jya} (which is also in agreement with a triviality of
the 1-loop correction to the
massless HS partition function as required by the AdS/CFT \cite{Beccaria:2014xda}).\footnote{The definition of the
sum over spins requires
a particular prescription that should be consistent with the underlying symmetries
\cite{Giombi:2014iua,Beccaria:2015vaa}.}
Similar vanishing was also found (under some natural assumptions)
for the 1-loop Weyl anomaly
$c$-coefficient \cite{Tseytlin:2013jya,Giombi:2014iua,Beccaria:2014xda,Beccaria:2015vaa}.
As the Weyl symmetry is one of the CHS gauge
symmetries, the same anomaly cancellation may apply also to all algebraic CHS gauge symmetries.
The global part of the CHS symmetry should also strongly constrain other "observables", {\em e.g.},
the analog of the S-matrix involving exchanges of the CHS fields.
Indeed, it was found in \cite{Joung:2015eny} that starting with a free external scalar field coupled (via
the above current $\int d^4 x\ h_s J_s $ interaction) to free CHS fields with the action
$ \int d^4 x \sum_s h_s\, \Pi_s \, \partial^{2s } h_s $ and
computing the 4-scalar tree level scattering due to the exchange of the tower of
CHS fields one finds that while the individual spin $s$ exchange contributions are non-trivial, their sum over all $s=0,1,2,...$
vanishes.
This vanishing can be understood \cite{Joung:2015eny} as a consequence of
the CHS global symmetry of the coupled theory (in particular, the
"hypertranslations" $\delta \vp = \epsilon^{\m_1...\m_s} \partial_{\m_1} ... \partial_{\m_s} \vp$ and scale invariance).
The aim of the present paper is to show that this triviality of the 4-particle scattering amplitude is found also
when the external scalars are replaced by the CHS fields themselves
with cubic and quartic interactions given by \rf{2}. We shall consider a few particular examples of the CHS 4-particle scattering amplitudes (4-vector, 4-graviton, etc.)
and find that after summation over all exchanged conformal higher spins
the total amplitude vanishes. This cancellation is rather non-trivial and
like in the external scalar amplitude case \cite{Joung:2015eny} should again be
a consequence of the underlying higher spin global symmetry of the theory (and
should thus be a manifestation
of a "generalised" Coleman-Mandula theorem).
This suggests that the full "S-matrix" of the CHS theory should be trivial.\footnote{As we are dealing with a non-unitary higher derivative theory containing
an infinite number of fields some assumptions of the standard Coleman-Mandula theorem may not directly apply.
In particular, the definition of the scattering matrix for higher-derivative fields requires clarification, see below.}
To be able to compute scattering amplitudes of CHS states one needs first to determine the precise
structure of vertices in the "induced" action \rf{2}. For that one needs to find the logarithmically divergent
(or $1\over \varepsilon$ pole in dimensional regularisation) terms in the one-loop scalar loop diagrams with the two, three or four
current $J_s$ insertions. We shall choose the external $h_s$ legs to be in the transverse traceless gauge.\footnote{This
avoids, in particular, the explicit discussion of field redefinitions eliminating the traces.}
Having found the relevant terms in the action \rf{2} we will define the 4-particle scattering S-matrix
as the amputated tree-level Green's function ({\em i.e.}
the sum of the exchange term and contact vertex >\!----\!< + >\!<)
contracted with particular on-shell asymptotic states. For $s=1$ vector the asymptotic states
are the standard helicity $\pm1$ states, while in the $s>1$ case with the free equation (in TT gauge)
$ \partial^{2s} h_s=0$ describing total of
$s(s+1)$ dynamical degrees of freedom one may choose a special solution
corresponding, {\em e.g.}, to the standard massless
helicity $\pm s$ field.\footnote{For example, in the case of the
Weyl graviton one can always solve the linearised Bach equations by imposing the linearised Einstein equations.
One may also consider other special choices of solutions of $ \partial^4$ equations as asymptotic states.}
Before turning to the discussion of higher spin terms in \rf{2} let us first recall the structure of the non-linear
terms for the low ($s=0,1,2,$) spins only. Instead of starting with the scalar action involving only the linear coupling to the
background fields $h_0,h_1,h_2$ (which here we assume to be subject to TT condition and drop total derivatives)
\begin{equation} \label{3}
L= - \partial_\m \vp^* \partial^\m \vp + \sum_s h_s \vp^* \J_s \vp\te = \partial_\m \vp^* \partial^\m \vp
+ h_0 \vp^* \vp + {i} h^\m \vp^* \partial_\m \vp + { \textstyle{1 \over 2}} h^{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} \partial_\m \vp^* \partial_\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J} \vp + ... \end{equation}
let us start with the standard manifestly ($U(1)$, reparametrisation and Weyl) covariant
coupling of a complex scalar field to the background
metric $g_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} = \eta_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} + h'_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}}$, vector field $ h'_\m$
and a scalar $ h'_0$, {\em i.e.}
\begin{equation} \label{4}
I= \int d^4 x\te \sqrt g \Big[- g^{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} D_\m \vp^* D_\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J} \vp + (h_0'-{1\over 6} R) \vp^* \vp\Big] \ , \ \ \ \qquad
D_\m \vp = \partial_\m \vp + { i\over 2} h'_\m \vp \ . \end{equation}
The log UV divergent part of the resulting scalar determinant (cf. \rf{1})
is given by the standard covariant
Seeley coefficient expression
(we ignore unimportant overall constant related to coupling ${g}$ in \rf{2})\footnote{Note that this action may be interpreted as the bosonic sector of $\mathcal N=1$ conformal supergravity action with $h_0'$ playing the role of the auxiliary field.}
\begin{equation} \label{5}
S[h_0', h'_1,h'_2] = \int d^4 x \te \sqrt g \Big( h_0'^2 - {1 \over 24} F'^2_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} + { 1\over 60} C^2_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}\lambda} \def \r {\rho\r} \Big)\ , \end{equation}
where $F'_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}}= \partial_\m h'_\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J} - \partial_\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J} h'_\m $ and $C$ is the Weyl tensor for $g_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}}$.
Since the fields $h'_s$ in \rf{4} are related to $h_s$ in \rf{3} by a local non-linear redefinition
\begin{align}\label{6}\te
&\te h_0'= h_0 + {1\over 4} h_\m h^\m + {1\over 96} ( \partial_\lambda} \def \r {\rho h_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} \partial^\lambda} \def \r {\rho h^{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} + 2h_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} \Box h^{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}}
+ 2 \partial_\lambda} \def \r {\rho h_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} \partial^\m h^{\lambda} \def \r {\rho\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} ) +...\ ,
\\
&\no
\te h'_\m= h_\m + {1\over 2} h_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} h^\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J} +
{1\over 4} h_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} h^{\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}\lambda} \def \r {\rho} h_\lambda} \def \r {\rho + ... \ , \qquad
\ \ \ \ h'_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} = {1\over 2} h_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} + {1\over 4} h_{\m\lambda} \def \r {\rho} h^\lambda} \def \r {\rho_\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}
- {1\over 16} \eta_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} h^{\lambda} \def \r {\rho\r} h_{\lambda} \def \r {\rho\r} + ... \ , \end{align}
expanding \rf{5} we may thus read off
the cubic and quartic couplings of the original $h_0,h_\m,h_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}}$ fields
in \rf{2}.
In particular, using \rf{6}
we find that the scalar-vector sector of \rf{5} takes the form
\begin{equation} \label{7}
S[h_0, h_1] = \int d^4 x \te \Big[ ( h_0 + {1\over 4} h_\m h^\m)^2 - {1 \over 24} F_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}}^2 \Big] \ . \end{equation}
Thus the simplest cubic and quartic vertices are $011$ and $1111$.
We also conclude, in particular, that the contribution of the $h_0$ exchange to the
4-vector scattering amplitude
cancels against the 4-vector contact vertex. As there is no 3-vector coupling, the
full 4-vector tree-level amplitude should thus be given by the sum of all
exchanges of CHS fields with $s\geq 2$ and happens to vanish as we will find in section 3.
Similarly, the 112 vertex is related to the one in the Maxwell-Weyl theory, the
222 and 2222 vertices are related to those in the Weyl theory, etc.
Thus the contribution to the 4-graviton amplitude computed
from the $s=0,2$ exchanges and the 2222 vertex should be
the same as the 4-graviton amplitude in pure Weyl theory
(that happens to vanish). The contributions of all $s>2$ CHS exchanges vanishes separately as we
shall demonstrate in section 5.
This paper is organised as follows. In section 2 we shall present the results for
some cubic CHS vertices (relevant for the computation of spin 1 and spin 2 scattering amplitudes below)
from the UV singular parts of the scalar loop integrals, with some details relegated to Appendix \ref{A}.
In section 3 we shall compute the 4-vector scattering amplitude and demonstrate that
after summing over all CHS exchanges it vanishes.
We shall then observe in section 4 that the conformal higher spin exchange amplitudes should have the same general structure as the partial wave amplitudes in the representation of \cite{Jacob:1959at}.
In section 5 we shall find that the scattering amplitudes 22 $\to $ 22 and 11$\to $ 22
involving conformal gravitons do have this expected structure and they also vanish once one sums up all intermediate CHS exchanges.
Some concluding remarks will be made in section 6.
In Appendix \ref{B} we shall independently verify the vanishing of the 11$\to $ 11 amplitude at special
kinematics (backward scattering) and find that this vanishing appears to generalise to the case of jj $\to$ jj
scattering with all ${\rm j} =1,2,3,...$ supporting our
conjecture that the full 4-particle S-matrix in CHS theory should be trivial.
In Appendix \ref{C} we shall give the general derivation of the expression for the
CHS spin $s$ exchange contribution to the 11 $\to $ 11 amplitude.
\section{Vertices in induced conformal higher spin action}
To be able to compute the CHS scattering amplitudes we should
first determine the relevant cubic and quartic terms in the "induced" action \rf{2}.
We shall use the following notation for totally symmetric tensors: $J_{\m(s)} \equiv J_{\m_1...\m_s}$ and also
$ \partial_{\m(s)} \equiv \partial_{\m_1}... \partial_{\m_s}$.
Our starting point will be the complex scalar Lagrangian in external CHS background (see, {\em e.g.}, \cite{Bekaert:2010ky,Joung:2015eny}
and refs. there)
\begin{align}
\mathscr L &= -\partial_{\mu}\varphi^{*}\,\partial^{\mu}\varphi+
\sum_{s=0}^{\infty}
J_{\mu(s)}\,h^{\mu(s)} \ , \label{2.1}\\
J_{\mu(s)}(x) &={\te \frac{i^{s}\,2^{s}\,s!}{(2s)!}\,}\sum_{k=0}^{s}\te
\binom{s}{k}\binom{\frac{s+k-1}{2}}{s}\,G^{(k)}_{\mu(s)}\ , \label{2.2}\\
G^{(k)}_{\mu(s)} &= \Big[
(\partial-\partial')_{\mu(k)}(\partial+\partial')_{\mu(s-k)}\varphi(x)\,\varphi^{*}(x')
\Big]_{x=x'} \ , \label{2.3}
\end{align}
where the low-spin currents $J_{\mu(s)}$ are
\begin{equation}
\label{2.4}
\begin{split}
J &=\varphi\, \varphi^{*}, \qquad
J_{\mu} =\te \frac{i}{2}\,(\varphi^{*}\partial_{\mu}\varphi-\varphi\,\partial_{\mu}\varphi^{*}), \\
J_{\mu\nu} &=\te \frac{1}{6}\,\big[\partial_{\mu}\varphi\,\partial_{\nu}\varphi^{*}
+\partial_{\nu}\varphi\,\partial_{\mu}\varphi^{*} - { \textstyle{1 \over 2}} ( \varphi^{*} \partial_{\mu} \partial_{\nu}\varphi\,+\varphi\,\partial_{\mu} \partial_{\nu}
\varphi^{*}) \big]\ .
\end{split}
\end{equation}
The vertices in the CHS action \rf{2} may be thought of as originating from the coinciding-point limits
of the current correlators in the free scalar CFT \ $\langle J_{s_1}(x_1) ...J_{s_n}(x_n) \rangle\big|_{x_i \to x}$
and can be found in coordinate space using, e.g., differential regularisation \cite{Freedman:1991tk}.
Here
we shall use momentum space representation and
dimensional regularisation ($d=4-\varepsilon$) and define the classical CHS action $S[h]$ as in \rf{2}, {\em i.e.}
as the UV pole part of the one-loop scalar $\vp$ effective action:
\begin{equation} \label{02}
\Gamma[h]= \frac{1}{(4\,\pi)^{2}\ \varepsilon}\,S[h]+
\text{finite} \ .
\end{equation}
In general, the CHS action should contain an arbitrary dimensionless constant ${g}$ as in \rf{2} that
will then appear as ${g}^2$ factor in the resulting 4-particle tree-level amplitude; in what follows we shall ignore this universal overall
factor, {\em i.e.} set ${g}=1$.
We shall also assume that the background fields $h_s \equiv h_{\mu(s)}$ in \rf{2.1} are
transverse and traceless (TT) as this will be sufficient for the subsequent
computation of the on-shell scattering amplitudes.\footnote{In
contrast to usual massless Fronsdal HS fields (where one can only fix transverse or de Donder gauge off shell) for the
conformal higher spin fields the gauge symmetry involves both the differential and the algebraic symmetry allowing one to fix TT gauge, and this leads to substantial simplifications.}
Note that in this case we may integrate by parts to write the interaction terms as
$h^\m J_\m \to i h^\m \varphi^{*}\partial_{\mu}\varphi,
\ \ h^{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} J_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} \to {1\over 2} h^{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} \partial_\m \varphi^{*} \partial_{\nu}\varphi $ or $-{1\over 2} h^{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} \varphi^{*} \partial_\m \partial_{\nu}\varphi $, etc.
In general, for TT fields
the $h_s (p'-p) \vp^*(-p') \vp(p)$ vertex in momentum representation reduces simply to
\begin{equation} \label{2.5}
V_{\mu(s)} (p) =\te \frac{1}{s!}\, p_{\mu_{1}}\cdots p_{\mu_{s}} \ , \end{equation}
where $p$ is the momentum of the $\vp$ leg.
We can then compute the UV singular part of the scalar loop diagram with two $V_{\mu(s)} $ insertions
\begin{equation}\label{f1}
\vcenter{\hbox{\includegraphics[scale=0.45]{fig-kinetic.pdf}}}
\end{equation}
We then find that the kinetic term in the CHS action (restricted to TT fields) is given by\footnote{The relative normalisations of the first $s=0,1,2$ terms here are the same as in \rf{5},\rf{6}:
note that $h'_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}}=g_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} -\eta_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} $ in the manifestly covariant action \rf{4} is given by $h'_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}}= {1\over 2} h_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} + ...$
in terms of $h_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}}$ in \rf{2.1} so that $C^2_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}\lambda} \def \r {\rho\r} \to 2 R^2_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} + ... \to {1\over 2} h'^{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} \Box^2 h'_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} \to {1\over 8} h ^{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} \Box^2 h_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}}$.}
\begin{equation}
\label{2.6}
S_2[h]= \sum_{s=0,1,2,...} \te
\frac{1}{2^{s}\,(2s+1)!}\,\int d^{4}x\,
h_{\mu(s)}\,\Box^{s}\,h^{\mu(s)} \ .
\end{equation}
To determine the cubic $h_{s_1} h_{s_2} h_{s_3} $ couplings in the CHS action \rf{2} we are to
compute the UV singular part of the one-loop scalar diagram with three (spin $s_1$, $s_2$ and $s_3$) current vertex \rf{2.5} insertions
\begin{equation}
\label{f2}
\vcenter{\hbox{\includegraphics[scale=0.5]{fig-cubic.pdf}}}
\end{equation}
As each spin $s$ vertex involves $s$ derivatives, parity invariance implies that the resulting interaction
is non-zero only if
$s_1+s_2+s_3 =$ even.
One can also show (using that $h_s$ is subject to the TT condition and dimensional analysis)
that the 0-0-$s$ interaction vanishes for all $s$,
{\em i.e.} $S_3 [h_0, h_0, h_s] =0$.
For $s_1=s_2=1$ the interaction 1-1-$s$ is non-zero only if $s$ is even.
Written in coordinate space the corresponding cubic interaction in the CHS action \rf{2} (restricted again to TT fields)
is found to be (see Appendix \ref{A})\footnote{The last two terms involving $\Box h^{\mu}$, {\em i.e.} proportional to
the vector field equation of motion can be, in principle,
redefined away.}
\begin{align}
& S_3 [h_1, h_1, h_s] ={\te \frac{(-1)^{s/2}}{(s+2)!}}\int d^{4}x\Big[
\partial_{\rho(s)}h_{\mu}h^{\mu}h^{\rho(s)}
-2 h_{\mu}\, \partial^{\mu}\,\partial_{\rho(s-1)}\,h_{\nu}\, h^{\nu\rho(s-1)}\no \\
&\te - \partial_{\lambda}\partial^{\rho(s-2)}h^{\mu}\partial^{\lambda}h^{\nu}h_{\mu\nu\rho(s-2)}
-\frac{s}{2} \partial^{\rho(s-2)}\Box h^{\mu}h^{\nu}h_{\mu\nu\rho(s-2)}
-\frac{s}{2} \partial^{\rho(s-2)}h^{\mu}\Box h^{\nu}h_{\mu\nu\rho(s-2)}
\Big]\label{2.66}\ .
\end{align}
This vertex has total of $s$ derivatives in agreement with the general structure of the CHS action \rf{2}.
In particular,
\begin{equation}\label{2.7}
S_3 [h_1, h_1, h_0] ={\te \frac{1}{2}}\int d^{4}x\,
h_{\mu}h^{\mu}h_0 \
\end{equation}
is in agreement with \rf{5},\rf{7}.
We may also compute the 4-vector quartic vertex from the UV pole part of the diagram
\begin{equation}\label{f3}
\vcenter{\hbox{\includegraphics[scale=0.25]{fig-box-draft.pdf}}}
\end{equation}
getting, in agreement with \rf{7},
\begin{equation}\label{2.99}
S_4 [h_1, h_1, h_1,h_1]
= {\te \frac{1}{16}}\,\int d^{4}x (h_{\mu}h^{\mu})^{2} \ .
\end{equation}
The vector-vector-graviton coupling in \rf{2.66}
\begin{equation}\label{2.8}
S_3 [h_1, h_1, h_2] = {\te \frac{1}{24}}
\int d^{4}x\Big[
\partial_{\rho}h_{\mu}\,\partial_{\sigma}h^{\mu}h^{\rho\sigma}
-2 \partial_{\rho}h_{\mu}\, \partial^{\mu}\,h_{\nu}\, h^{\nu\rho} + \partial_{\r} h^{\mu}\partial^{\r}h^{\nu}h_{\mu\nu}
+ 2 h^{\mu}\,\Box h^{\nu}h_{\mu\nu}
\Big]
\end{equation}
is equivalent (for TT fields) to the standard graviton-Maxwell coupling in \rf{5}
provided one takes into account the redefinitions in \rf{6}.
Similar expressions are found when the vector vertices in \rf{f2} are replaced by the graviton ones, {\em i.e.}
for the case of the $s_1=s_2=2, \ s_3=s$ interaction term (see Appendix \ref{A}).
With $p_1,p_2$ being spin 2 momenta the resulting 2-2-$s$ interaction vertex contains $s+2$ powers of momentum
and reads
$
\text{V}_{\mu_{1}\mu_{2}, \nu_{1}\nu_{2}, \rho(s)} = V_{\mu_{1}\mu_{2}, \nu_{1}\nu_{2}, \rho(s)}
(p_{1}, p_{2})+V_{\mu_{1}\mu_{2}, \nu_{1}\nu_{2}, \rho(s)}
(p_{2}, p_{1}),
$
where\footnote{Here we drop terms
proportional to equations of motion for spin 2 states as we will be using this vertex to compute 2-2-2-2 scattering amplitude.
Note, however,
that for spin 2 exchange one is to use the 2-2-2 vertex that is symmetric in the three spin 2 legs with no on-shell condition assumed.
}
\begin{equation} \begin{split}\label{2.10}
&V_{\mu_{1}\mu_{2}, \nu_{1}\nu_{2}, \rho(s)}
(p_{1}, p_{2}) = {\te \frac{1}{8\,(s+4)!}\,}
\Big[
- \sum_{\mu\neq\mu',\nu\neq\nu'} \eta_{\mu\nu}\,p_{2\mu'}\,p_{1\nu'}\,(p_{1})_{\rho(s)}\\
&+2\,\sum_{\mu\neq \mu'}\,p_{2\mu'}\,p_{1\nu_{1}}\,p_{1\nu_{2}}\,\eta_{\mu\rho_{1}}p_{1\rho_{2}}\dots p_{1\rho_{s}}
-2\,\sum_{\nu\neq \nu'} p_{1\nu'}\,p_{2\mu_{1}}\,p_{2\mu_{2}}\,\eta_{\nu\rho_{1}}p_{1\rho_{2}}\dots p_{1\rho_{s}}\Big]\\
&{\te -\frac{p_{1}\cdot p_{2}}{16\,(s+4)!}} \,\Big\{ 2(
\eta_{\mu_{1}\nu_{1}}\eta_{\mu_{2}\nu_{2}}+ \eta_{\mu_{1}\nu_{2}}\eta_{\mu_{2}\nu_{1}} )\,(p_{1})_{\rho(s)}\\ &- 4
(\,p_{1\nu_{1}}\eta_{\mu_{1}\nu_{2}}\eta_{\mu_{2}\rho_{1}} - p_{2\mu_{1}} \eta_{\mu_{2}\nu_{1}}\eta_{\nu_{2}\rho_{1}} + {\rm sym}\ \mu_{1,2},\nu_{1,2} ) \,p_{1\rho_{2}}\cdots p_{1\rho_{s}}
\\
& + \Big[ 6\,(\eta_{\mu_{1}\rho_{1}}\eta_{\mu_{2}\rho_{2}}+\eta_{\mu_{1}\rho_{2}}\eta_{\mu_{2}\rho_{1}})\,
p_{1\nu_{1}}\,p_{1\nu_{2}}
+6\,
(\eta_{\nu_{1}\rho_{1}}\eta_{\nu_{2}\rho_{2}}+\eta_{\nu_{1}\rho_{2}}\eta_{\nu_{2}\rho_{1}})\,
p_{2\mu_{1}}\,p_{2\mu_{2}}\\
&- \sum_{\mu\neq\mu',\nu\neq\nu'}
4\,
(\eta_{\mu\rho_{1}}\eta_{\nu\rho_{2}}+\eta_{\mu\rho_{2}}\eta_{\nu\rho_{1}})\,
p_{2\mu'}\,p_{1\nu'}\Big]\,p_{1\rho_{3}}\cdots p_{1\rho_{s}}\Big\}\\
&
+ {\te \frac{(p_{1}\cdot p_{2})^{2}}{8\,(s+4)!} } \Big\{
\sum_{\mu\neq\mu', \nu\neq\nu'}
(\eta_{\mu\rho_{1}}\eta_{\nu\rho_{2}}\eta_{\mu'\nu'}+\eta_{\mu\rho_{2}}\eta_{\nu\rho_{1}}\eta_{\mu'\nu'})\,p_{1\rho_{3}}\cdots p_{1\rho_{s}}
\\
& - (
p_{1\nu_{1}}\, \eta_{\mu_{1}\rho_{1}}\eta_{\mu_{2}\rho_{2}}\eta_{\nu_{2}\rho_{3}}
-p_{2\mu_{1}}\,\eta_{\mu_{2}\rho_{1}}\eta_{\nu_{1}\rho_{2}}\eta_{\nu_{2}\rho_{3}} + {\rm sym}\ \rho_{1,2,3} )
\,p_{1\rho_{4}}\cdots p_{1\rho_{s}}\Big\}\\
&-{\te \frac{(p_{1}\cdot p_{2})^{3}}{32\,(s+4)!}\,}
(\eta_{\mu_{1}\rho_{1}}\eta_{\mu_{2}\rho_{2}}\eta_{\nu_{1}\rho_{3}}\eta_{\nu_{2}\rho_{4}}+
{\rm sym}\ \rho_{1,2,3,4})\,\,p_{1\rho_{5}}\cdots p_{1\rho_{s}}\ ,
\end{split}
\end{equation}
where ${\rm sym}$ stands for adding terms ensuring symmetry in $(\m_1,\m_2)$, $(\nu_1,\nu_2)$ and $(\r_1, ...,\r_s)$.
In particular, choosing $s=0$ we find that the 2-2-0 coupling term in the CHS action can be written as
\begin{equation} \label{2.111}
S_3[h_0,h_2,h_2] = { \te {1\over 48} } \int d^4x\ h_0 \big( \partial_\r h_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} \partial^\r h^{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} + 2 \partial_\r h_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}} \partial^\m h^{\r\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}}\big)
\ .
\end{equation}
One can trace the origin of this term to $h'^2_0$ term in \rf{5} and the redefinition \rf{6}
(in particular, it corresponds to cross-term $h_0 R$ with $R$ in \rf{4} expanded to quadratic order in $h'_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}}$).
The 1-0-$s$ vertex multiplying $h_{\mu}(p_{1}), \ h_0(p_{2})$ and $ h_{\rho(s)}(-p_{1}-p_{2})$
is non-zero when $s$ is odd and is found to be (where symmetrisation in $\r_i$ is assumed)
\begin{equation}
\label{2.11}
\begin{split}
\text{V}_{\mu,\rho(s)} &=\te \frac{2}{(s+1)!}\,
\eta_{\mu\rho_{1}}\,p_{\rho_{2}}
\cdots p_{\rho_s} \ .
\end{split}
\end{equation}
Here $p$ stands for either $p_1$ or $p_2$ (\rf{2.11} is symmetric
under $p_{1}\to p_{2}$ as the fields are assumed to be TT and $s$ is odd).
Similarly, the 2-0-$s$ vertex (non-vanishing for $s$=even) is given by
\begin{equation}\label{2.12}
\begin{split}
\text{V}_{\mu_{1}\mu_{2},\rho(s)}\te = &\te \frac{1}{(s+2)!}\,
\Big[-
(\eta_{\mu_{1}\rho_{1}}\,p_{1\, \mu_{2}}+\eta_{\mu_{2}\rho_{1}}p_{1\, \mu_{1}})\,p_{1\, \rho_{2}}
... p_{1\, \rho_{s}}\\ &\te \qquad \quad
-{ \textstyle{1 \over 2}} {p_{1}\cdot p_{2}}\,(
\eta_{\mu_{1}\rho_{1}}\eta_{\mu_{2}\rho_{2}}+\eta_{\mu_{2}\rho_{2}}\eta_{\mu_{1}\rho_{1}})\,
p_{1\, \rho_{3}}... p_{1\, \rho_{s}}
\Big]\ .
\end{split}
\end{equation}
In the case of 1-2-$s$
vertex (with $s$=odd) appearing multiplied by the TT fields
$h_{\mu_{1}\mu_{2}}(p_{1}),$ $ h_{\nu}(p_{2}),$ $ h_{\rho(s)}(-p_{1}-p_{2})$
we get
\begin{equation}
\label{2.13}
\begin{split}
&{\rm V}_{\mu_{1}\mu_{2},\nu,\rho(s)}(p_{1},p_{2}) =\te \frac{1}{(s+3)!}\Big\{
(\eta_{\mu_{1}\nu}p_{2\mu_{2}}+\eta_{\mu_{2}\nu}p_{2\mu_{1}}) p_{1\rho(s)}\\
&+(-\eta_{\mu_{1}\rho_{1}}p_{2\mu_{2}}p_{1\nu}
-\eta_{\mu_{2}\rho_{1}}p_{2\mu_{1}}p_{1\nu}
+2\eta_{\nu\rho_{1}}p_{2\mu_{1}}p_{2\mu_{2}})\,p_{1\rho_{2}}... p_{1\rho_{s}}\\
&-(p_{1}\cdot p_{2})\Big[
(\eta_{\mu_{1}\nu}\eta_{\mu_{2}\rho_{1}}+\eta_{\mu_{1}\rho_{1}}\eta_{\mu_{2}\nu})\,p_{1\rho_{2}}... p_{1\rho_{s}}\\
&+\Big( (\eta_{\mu_{1}\rho_{1}}\eta_{\nu\rho_{2}}+\eta_{\mu_{1}\rho_{2}}\eta_{\nu\rho_{1}})\,p_{2\mu_{2}}
-(\eta_{\mu_{1}\rho_{1}}\eta_{\mu_{2}\rho_{2}}+ \eta_{\mu_{2}\rho_{1}}\eta_{\mu_{1}\rho_{2}})\,
p_{1\nu}
\\
&+ (\eta_{\mu_{2}\rho_{1}}\eta_{\nu\rho_{2}}+\eta_{\mu_{2}\rho_{2}}\eta_{\nu\rho_{1}})
\,p_{2\mu_{1}}\Big)\,p_{1\rho_{3}}... p_{1\rho_{s}}\Big]\\
&\te +\frac{1}{3}(p_{1}\cdot p_{2})^{2}
(\eta_{\mu_{1}\rho_{1}}\eta_{\mu_{2}\rho_{2}}\eta_{\nu\rho_{3}}+{\rm sym} \
\rho_{1,2,3})\,p_{1\rho_{3}}... p_{1\rho_{s}}\Big\} \ .
\end{split}
\end{equation}
\section{Scattering
in CHS theory: 4-vector amplitude}
We can now use the interaction terms in the CHS action found in the previous section to compute some tree-level
scattering amplitudes. As the scalar $h_0$ is non-propagating, {\em i.e.} has zero on-shell value we will not discuss
analogs of scattering amplitudes with $h_0$ on external legs.
The vector $h_1$ has the standard Maxwell kinetic term, so
the definition of the corresponding 4-vector scattering amplitude is standard (the same as in the case of the
external scalar scattering in \cite{Joung:2015eny}): we consider physical helicity $\pm 1$ photon states
on external lines and include all exchanges with two 1-1-s vertices \rf{2.6} connected by
TT propagator for even-spin $s$ CHS field. The contribution of the $h_0$ exchange
due to 011 vertex \rf{2.7}
exactly cancels against the contact 4-vector vertex \rf{2.99} as follows from \rf{7}
so it remains to consider only the exchanges with $s=2,4,6,...$ fields on internal lines.
Before proceeding with spin 1 scattering let us note for the future discussion in sections 4 and 5
that as the CHS fields with $s>1$ in \rf{2} have higher-derivative kinetic terms,
the notion of S-matrix for $s>1$ external lines requires special definition.
Given the free spin $s>1 $ CHS equation in TT gauge $\Box^s h_s=0 $ one can always
choose a special solution $h_s=h^{(0)}_s $ satisfying $\Box h^{(0)}_s=0 $.
This equation has further on-shell gauge invariance allowing one to reduce the number of independent solutions
to just 2 of a standard 2-derivative massless particle. In what follows
we shall always consider only these special "physical" helicity $\pm s$
modes as the asymptotic states in the definition of
the CHS S-matrix.\footnote{Ideally,
one would like to start with a formulation of the CHS theory in terms of the set of fields
with ordinary (2-derivative) kinetic terms that exists at the quadratic level \cite{Metsaev:2007fq,Metsaev:2007rw}.
Unfortunately, an existence of such local action at the interacting level is an open
question for $s>2$.
}
Thus the asymptotic states will always be massless on-shell particles with
$p^2=0$ while the internal spin $s'$
propagator will be ${1\over p^{2s'}}$ times the TT projector.\footnote{The condition $p^2=0$ for the external lines
will help
to simplify the expressions for the
required cubic CHS vertices.}
\subsection{4-vector exchange amplitude}
Let us start with the 4-vector scattering amplitude and first set up the notation we will use.
We shall consider the scattering process
\begin{equation}\label{fig4}
\vcenter{\hbox{\includegraphics[scale=0.2]{fig-abcd-scattering-draft.pdf}}}
= \vcenter{\hbox{\includegraphics[scale=0.4]{fig-channels-draft.pdf}}}
\end{equation}
with
$(\lambda_{1},p_1), ( \lambda_{2},p_2)\to (\lambda_{3},p_3), (\lambda_{4},p_4)$, {\em i.e.} assume two momenta and helicities as incoming and two as outcoming with $\lambda} \def \r {\rho_i=\pm 1$ and $p^2_i=0$.
In the c.o.m. frame, we have for the momenta and polarisation vectors\footnote{The helicity $\pm 1$ polarisation vector for an { initial} state with
$p=(\omega, \omega\sin\theta, 0, \omega
\cos\theta)$ is $\varepsilon^{(\pm)}_{\mu}(p)=\mp{1 \over \sqrt 2} (0,\cos\theta,\pm i, -\sin\theta)$.
If the state is {final}, the polarisation vector is $(\varepsilon^{(\pm)}_{\mu}(p))^{*}$
(see, {\em e.g.}, \cite{Elvang:2013cua,Gleisberg:2003ue}).
}
\begin{equation}\label{3.1}
\begin{aligned}
p_{1} &= (\omega, 0, 0, \omega), &&
& \varepsilon_{1}(p_{1}) &= -{\te {1\ov \sqrt 2}} \lambda_{1}\,(0,1,i\,\lambda_{1},0) \\
p_{2} &= (\omega, 0, 0, -\omega), &&
&\varepsilon_{2}(p_{2}) &= -{\te {1\ov \sqrt 2}} \lambda_{2}\,(0,-1,i\,\lambda_{2},0) \\
p_{3} &= (\omega, \omega\,\sin\theta, 0, \omega\,\cos\theta), &&
&[\varepsilon_{3}(p_{3})]^{*} &= -{\te {1\ov \sqrt 2}} \lambda_{3}\,(0,\cos\theta,-i\,\lambda_{3},-\sin\theta) \\
p_{4} &= (\omega, -\omega\,\sin\theta, 0, -\omega\,\cos\theta), &&
&[\varepsilon_{4}(p_{4})]^{*} &= -{\te {1\ov \sqrt 2}} \lambda_{4}\,(0,-\cos\theta,-i\,\lambda_{4},\sin\theta) \\
\end{aligned}
\end{equation}
and the Mandelstam variables are
\begin{align}
&\qquad \textsf{\textbf s} =-(p_{1}+p_{2})^{2} = 4\omega^{2}, \qquad\qquad
\textsf{\textbf t} =-(p_{1}-p_{3})^{2} = -2\omega^{2}\,(1-\cos\theta),\no \\ &
\qquad
\textsf{\textbf u} =-(p_{1}-p_{4})^{2} = -2\omega^{2}\,(1+\cos\theta)\ , \qquad \qquad \textsf{\textbf s}+\textsf{\textbf t} + \textsf{\textbf u} =0 \ . \label{3.2}
\end{align}
The exchange diagrams involve two 11s vertices corresponding to $h_1(p)\, h_1(q)\, h_s(-p-q)$ from \rf{2.6}\footnote{Here we use that
$p^2=q^2=0$ for the external vector lines; $p=p_1, \, q=p_2$ in s-channel, etc.}
\begin{equation}
\label{3.3}
\begin{split}
&\text{V}_{\alpha,\beta,\rho(s)}(p, q) =\te \frac{1}{(s+2)!}\Big\{
\eta_{\alpha\beta}\big[\tfrac{1}{2}\,(p)_{\rho(s)}+\tfrac{1}{2}\,(q)_{\rho(s)}\big]\\
&-\tfrac{1}{2}\eta_{\alpha\rho_{1}}p_{\beta}p_{\rho_{2}}\dots p_{\rho_{s}}
+\tfrac{1}{2}\eta_{\beta\rho_{1}}q_{\alpha}p_{\rho_{2}}\dots p_{\rho_{s}}
-\tfrac{1}{2}\eta_{\beta\rho_{1}}q_{\alpha}q_{\rho_{2}}\dots q_{\rho_{s}}
+\tfrac{1}{2}\eta_{\alpha\rho_{1}}p_{\beta}q_{\rho_{2}}\dots q_{\rho_{s}}\\
&-\tfrac{1}{2}\,\eta_{\alpha\rho_{1}}\eta_{\beta\rho_{2}}
\,p_{\rho_{3}}\dots p_{\rho_{s}}\, p\cdot q
-\tfrac{1}{2}\eta_{\alpha\rho_{1}}\eta_{\beta\rho_{2}}
\,q_{\rho_{3}}\dots q_{\rho_{s}}\,p\cdot q
\Big\}.
\end{split}
\end{equation}
Here $h_s$ is assumed to be in TT gauge with the corresponding propagator (cf. \rf{2.6})
\begin{equation}\te \label{3.4}
D^{\alpha(s)}_{\beta(s) }(p) = \frac{2^{s-1}(2s+1)!}{(p^{2})^{s}}\ \Pi^{\alpha_{1}\cdots \alpha_{s}}_{\beta_{1}\cdots \beta_{s}}(p),
\end{equation}
where the TT projector $\Pi^{\alpha(s)}_{\beta(s) }$ is built out of products of
$\Pi^{\alpha}_{\beta}= \delta^{\alpha}_{\beta}-\frac{p^{\alpha}\,p_{\beta}}{p^{2}}$,
{\em e.g.,}\footnote{In general (cf. \rf{c3})
$$\Pi^{\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}(s)}_{\m(s)} = \sum_{l=0}^{[\frac{s}{2}]} {\rm a}_{s,l}\, \ M^{\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}(s-2l)}_{\m(s-2l)}\, N^{\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}(2l)}_{\m(2l)} \ , \ \ \ \qquad \qquad
{\rm a}_{s,l} = \frac{(-1)^l s! \, \G(s-l + {1\over 2} )}{2^{2l} (s-2l)! \, l! \,\G(s + {1 \over 2} ) }
$$
where
$M_{\m(p)}^{\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}(p)} = {\Pi^{(\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}_1}_{\m_1} \ldots \Pi^{\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}_p)}_{\m_p}}$ and
$
N_{\m(2q)}^{\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}(2q)} =\Pi_{(\m_1 \m_2} \ldots \Pi_{\m_{q-1} \m_q)} \, \Pi^{(\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}_1 \nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}_2} \ldots \Pi^{\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}_{q-1} \nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}_q)}
$.
}
\begin{align}
\label{3.9}
\Pi^{\alpha_{1}\alpha_{2}}_{\beta_{1}\beta_{2}} &= \te \Pi^{\alpha_{1}}_{(\beta_{1}}\,\Pi^{\alpha_{2}}_{\beta_{2})}-\frac{1}{3}\,
\Pi^{\alpha_{1}\alpha_{2}}\Pi_{\beta_{1}\beta_{2}}, \\
\Pi^{\alpha_{1}\alpha_{2}\alpha_{3}\alpha_{4}}_{\beta_{1}\beta_{2}\beta_{3}\beta_{4}} &\te =
\Pi^{(\alpha_{1}}_{(\beta_{1}}P^{\alpha_{2}}_{\beta_{2}}P^{\alpha_{3}}_{\beta_{3}}
\Pi^{\alpha_{4})}_{\beta_{4})}
-\frac{6}{7}\,P^{(\alpha_{1}\alpha_{2}}\Pi_{(\beta_{1}\beta_{2}}\Pi^{\alpha_{3}}_{\beta_{3}}\Pi^{\alpha_{4})}_{\beta_{4})}+\frac{3}{35}\,
\Pi^{(\alpha_{1}\alpha_{2}}\Pi^{\alpha_{3}\alpha_{4})}\Pi_{(\beta_{1}\beta_{2}}
\ \Pi_{\beta_{3}\beta_{4})}\ , \ \ {\rm etc.}\no
\end{align}
The resulting s, t, u-channel amplitudes are are given by\footnote{The factors of 2 in the vertices are due to the
symmetry of the external lines: for a Lagrangian term $\Phi^{n}=n!\frac{\Phi^{n}}{n!}$, the standard
Feynman rules imply the coefficient $n!$.}
\begin{equation}\label{3.10}
\begin{split}
A_\textsf{\textbf s}^{(s)} &= 2\,{\rm V}(p_{1}, p_{2})\cdot D(p_{1}+p_{2})\cdot 2\,{\rm V}(p_{3}, p_{4})\cdot \varepsilon_{1}\,
\varepsilon_{2}\,\varepsilon_{3}^{*}\,\varepsilon_{4}^{*},\\
A_\textsf{\textbf t}^{(s)} &= 2\,{\rm V}(p_{1}, p_{3})\cdot D(p_{1}-p_{3})\cdot 2\,{\rm V}(p_{2}, p_{4})\cdot \varepsilon_{1}\,
\varepsilon_{2}\,\varepsilon_{3}^{*}\,\varepsilon_{4}^{*},\\
A_\textsf{\textbf u}^{(s)} &= 2\,{\rm V}(p_{1}, p_{4})\cdot D(p_{1}-p_{4})\cdot 2\,{\rm V}(p_{2}, p_{3})\cdot \varepsilon_{1}\,
\varepsilon_{2}\,\varepsilon_{3}^{*}\,\varepsilon_{4}^{*} \ .
\end{split}
\end{equation}
Evaluating these amplitudes for various helicity choices we find that all amplitudes where helicity is not conserved
vanish\footnote{In standard terminology that means that only MHV amplitudes are non-zero. The same will be true for all amplitudes discussed below.}
and for the helicity-conserving cases $\pm\pm \to \pm \pm$ or the crossing-related cases $\pm \mp \to \pm \mp$ we get\footnote{The vanishing of
the s-channel exchange for the same helicity process may be related to the fact that
helicity is conserved in the 3-point
vertices where one has only the $\pm\mp$ combination (same happens for gravitational
interactions, cf. \cite{Datta:2004mr}).}
\begin{equation}\label{3.11}
\begin{split}
&\pm\pm\to \pm\pm: \ \qquad A_{\textsf{\textbf s}}^{(s)}= 0, \qquad
A_{\textsf{\textbf t}}^{(s)}= c_{s}\, \big(\frac{\textsf{\textbf s}}{\textsf{\textbf t}}\big)^{s}\,P_{s}\big(\frac{\textsf{\textbf t}}{\textsf{\textbf s}}\big),
\qquad
A_{\textsf{\textbf u}}^{(s)} = c_{s}\, \big(\frac{\textsf{\textbf s}}{\textsf{\textbf u}}\big)^{s}\,P_{s}\big(\frac{\textsf{\textbf u}}{\textsf{\textbf s}}\big), \\
&\pm\mp\to \pm\mp: \ \qquad A_{\textsf{\textbf s}}^{(s)}= c_{s}\, \big(\frac{\textsf{\textbf u}}{\textsf{\textbf s}}\big)^{s}\,P_{s}\big(\frac{\textsf{\textbf s}}{\textsf{\textbf u}}\big),
\qquad
A_{\textsf{\textbf t}}^{(s)} = c_{s}\, \big(\frac{\textsf{\textbf u}}{\textsf{\textbf t}}\big)^{s}\, P_{s}\big(\frac{\textsf{\textbf t}}{\textsf{\textbf u}}\big), \qquad
A_{\textsf{\textbf u}}^{(s)} = 0 \ .
\end{split}
\end{equation}
Like in the external scalar scattering case \cite{Joung:2015eny} the scale invariance of the CHS theory
and the fact that $h_1$ has canonical dimension 1 implies that the $d=4$ amplitude depends only on ratios of the
Mandelstam variables. We have isolated powers of these ratios containing the internal spin $s$ CHS propagator factor in each channel
({\em i.e.} $ \textsf{\textbf t}^{-s} $ in t-channel, etc.).
The remaining momentum dependence is given by the even degree $s-2$ polynomials $P_{s}(x)$
whose normalisation is fixed by the condition $P_{2}=1$ and $ P_{s>2}(-1)=1$
\begin{equation}
\label{3.12}
\begin{split}
P_{2}(x) &= 1, \qquad \qquad
P_{4}(x) =28+42\,x+15\,x^{2}, \\
P_{6}(x) &= 495+1320\,x+1260\,x^{2}+504\,x^{3}+70\,x^{4}, \\
P_{8}(x) &= 8008 + 30030\, x + 45045\, x^2 + 34320\, x^3 + 13860\, x^4 + 2772\, x^5 +
210\, x^6, \\
P_{10}(x) &= 125970+604656 \,x+1225224 \,x^2+1361360 \,x^3+900900 \,x^4+360360 \,x^5\\
&\ \ \ \ \ +84084
\,x^6+10296 \,x^7+495 \,x^8 \ .
\end{split}
\end{equation}
The overall numerical coefficients $c_s$ then are
\begin{equation}\label{3.13}\te
c_{2} = \frac{5}{12}, \qquad
c_{4} = \frac{1}{20}, \qquad
c_{6} = \frac{13}{840}, \qquad
c_{8} = \frac{17}{2520}, \qquad
c_{10} = \frac{7}{1980}.
\end{equation}
The expressions \rf{3.12},\rf{3.13} were found by direct computations for $s=2,...,10$
but admit a natural generalisation to any $s$.
The polynomials $P_{s}(x)$ may be expressed in terms of Jacobi polynomials $P_{n}^{(a,b)}(x)$
as ($s=2,4,6,\dots$)
\begin{equation}
\label{3.14}\te
P_{s}(x) = x^{s-2}\,P_{s-2}^{(4,0)}\big(\frac{x+2}{x}\big) \ ,
\end{equation}
while the simplest interpolating ansatz for $c_s$ is
\begin{equation} \label{3.15}\te
c_{s} = \frac{2\,(2\,s+1)}{(s-1)\,s\,(s+1)\,(s+2)} \ .
\end{equation}
We shall provide the general derivation of \rf{3.14},\rf{3.15} in Appendix \ref{C}.
Let us mention also some useful alternative forms of the polynomials $P_{s}$ in \rf{3.14}
\begin{align}
P_{s}(x) &\no = \sum_{j=0}^{s-2}\te (-1)^{j}\,\binom{s-2}{j}\,\binom{2s-j}{s+2}\,(1+x)^{s-2-j}\,x^{j} = \binom{2s}{s+2}\,(1+x)^{s-2}\,{}_{2}F_{1}\big(2-s,2-s,-2s,\frac{x}{x+1}\big) \no \\
&{\te \ = \frac{x^{2\,s+1}}{(s-2)!}\,\big(\frac{d}{dx}\big)^{s-2}
\frac{(1+x)^{s-2}}{x^{s+3}}
=} \sum_{j=2}^{s}\te \frac{1}{(j-2)!\,(j+2)!}\,\frac{(s+j)!}{(s-j)!}\,x^{s-j}
\label{3.155}
\\
& \te
\no
= \frac{1}{24}(s-1)\,s\,(s+1)\,(s+2)\,x^{s-2}\,{}_{2}F_{1}\big(2-s,s+3,5; -\frac{1}{x}\big).\notag
\end{align}
For comparison, in the case of the external massless
scalar scattering (coupled to CHS fields as in \rf{2.1},\rf{2.2} or \rf{2.5})
the s-channel $\varphi \varphi^* \to \varphi \varphi^*$ amplitude was given \cite{Joung:2015eny}
in terms of the Legendre polynomial ${\rm P}_s= P_{s}^{(0,0)}$
\begin{equation}
\label{555}
A^{(s)}_{\textsf{\textbf s} \, \vp \vp^* \to \vp \vp^*} = (s+ { \textstyle{1 \over 2}}) \ {P_{s}^{(0,0)}} (-1-2x) \ , \ \ \ \ \te \qquad x={\textsf{\textbf t}\over \textsf{\textbf s}} \ . \end{equation}
One can also consider the "mixed" scattering amplitude $\vp \vp^* \to 11$ of two external conformal scalars into two vectors
due to exchange of the tower of CHS fields.\footnote{Here we are assuming that one adds the action \rf{3} of one conformal scalar
coupled to CHS fields to the CHS action \rf{2} and then studies the S-matrix of the resulting theory.}
In this case the s-channel even spin s exchange amplitude is given by (cf. \rf{3.10})
\begin{equation}
A_{\textsf{\textbf s} \, \vp\vp^*\to 11}^{(s)} = {\rm V}_{\varphi\varphi^{*}s}(p_{1}, p_{2})\cdot D^{(s)}(p_{1}+p_{2})\cdot 2\,{\rm V}(p_{3}, p_{4})\cdot \varepsilon_{3}^{*}\,\varepsilon_{4}^{*}, \label{345}
\end{equation}
where ${\rm V}_{\varphi\varphi^{*}s}$ is the vertex in \rf{2.5} and ${\rm V}$ is the 11s vertex as in \rf{3.3},\rf{3.10}.
The resulting $\pm\pm$ amplitudes vanish while the helicity-preserving $\pm\mp$ ones may be written as (cf. \rf{3.11})
\begin{equation}
\label{328}
A^{(s)}_{\textsf{\textbf s} \, \vp\vp^*\to \pm1\mp1} = k_{s}\,\frac{\textsf{\textbf t}\,\textsf{\textbf u}}{\textsf{\textbf s}^{2}}\,Q_{s}\big(\frac{\textsf{\textbf t}}{\textsf{\textbf s}}\big)\ ,
\end{equation}
where we ignore the overall minus sign and assume that numerical coefficients are defined by normalising the order $s-2$ polynomial
$Q_s$ as $Q_{s}(-1)=1$. Explicitly, one finds $Q_{2} = 1, \
Q_{4} = \frac{1}{3}\,(3+14\,x+14\,x^{2}), ...$ and $k_1= \frac{5}{2} , \ k_{4} = \frac{9}{2}, ...$.
On the basis of $s=2, ...,10$ examples can guess the general $s$ expressions as
\begin{equation} \label{329}\te
Q_{s}(x) = \frac{2}{s\,(s-1)}\,P_{s-2}^{(2,2)}(-1-2\,x) \ , \ \ \ \ \qquad \ \ \qquad
k_{s} = s + { \textstyle{1 \over 2}} \ ,
\end{equation}
where $P_{s-2}^{(2,2)}$ is again the Jacobi polynomial (cf. \rf{3.14},\rf{3.15},\rf{555}).
\subsection{Summing over spins}
As we already mentioned above, the $s=0$ exchange contribution cancels against
the one of the 1111 vertex \rf{2.99}.
Thus to get the total amplitude it remains to sum
over all spin $s=2,4, ...$ exchanges.
Let us consider, {\em e.g.}, the $\pm\pm\to \pm\pm$ case in \rf{3.11} (the discussion of the
$\pm\mp\to \pm\mp$ case is similar)
where the sum over channels is
\begin{equation}
\label{3.16}
\pm\pm\to \pm\pm: \ \ \ \ \ \ A^{(s)} = c_{s}\,\Big[\big(\frac{\textsf{\textbf s}}{\textsf{\textbf t}}\big)^{s}\,
P_{s}\Big(\frac{\textsf{\textbf t}}{\textsf{\textbf s}}\Big)
+\big(\frac{\textsf{\textbf s}}{\textsf{\textbf u}}\big)^{s}\,P_{s}\Big(\frac{\textsf{\textbf u}}{\textsf{\textbf s}}\Big)\Big].
\end{equation}
Since $\textsf{\textbf u}=-\textsf{\textbf s}-\textsf{\textbf t}$ this may be written as a function of one variable $x \equiv {\textsf{\textbf t}\over \textsf{\textbf s}} $ as
\begin{equation}\label{3.17}
A^{(s)} = \sigma_{s}(x)+\sigma_{s}(-1-x), \qquad\qquad \sigma_{s}(x) \equiv
c_{s}\,x^{-s}\,P_{s}(x)\ . \end{equation}
We may compute the sum over $s$ by introducing first an extra regularisation parameter $z$ and defining
\begin{equation} \label{3.18}
\sigma (x) \equiv
\lim_{z\to 1} \sigma(x; z) \ , \ \ \ \ \ \qquad \sigma(x; z) \equiv \sum_{s=2,4,6,...}^{\infty} \sigma_{s}(x)\ z^{s-2} \ .
\end{equation}
Let us first omit the overall coefficient $c_s$ in $\sigma_{s}$ and consider the formal sum
over all (even and odd) $s=2,3,4,...$
\begin{equation}
K(x;z) \equiv \sum_{s=2}^{\infty} x^{-s}\,P_{s}(x)\,z^{s-2} \ . \label{3.19}
\end{equation}
This can be written in a closed form using the generating function for the Jacobi polynomials $P_{s-2}^{(4,0)}$
\cite{koekoek1996askey} as
\begin{equation}
\label{3.22}\te
K(x;z) = \frac{16}{x^{2}}\, \big[\sqrt{z^2-\frac{2 z (x+2)}{x}+1}\big]^{-1} \big[\sqrt{z^2-\frac{2 z
(x+2)}{x}+1}-z+1\big]^{-4}\ .
\end{equation}
Then using the fact that $c_s$ in \rf{3.15} admits the following representation
\begin{equation}\te
c_{s} = \frac{1}{s+2}-\frac{1}{s+1}+\frac{1}{s-1}-\frac{1}{s} \ , \label{3.23}
\end{equation}
we can compute $\sigma(x; z)$ by multiplying \rf{3.22} by a suitable power of $z$, integrating,
and then dividing by another appropriate power of $z$.
Finally, the sum over spins may be restricted to even $s$ only by
simply taking one half of the sum of the expressions with $z$ and with $-z$.
While the resulting expression is quite
cumbersome, its $z\to 1$ limit turns out to be finite and simple
\begin{equation}
\label{3.24}
\sigma(x)
=\te x (x+1)\, \log \frac{x+1}{x} - x -{ \textstyle{1 \over 2}} \ .
\end{equation}
As it is easy to check, this function satisfies the relation $\sigma(x) = - \sigma(-1-x) $
implying that the total summed-over-spins amplitude vanishes:
\begin{equation}\label{3.25}
A (x)= {\sum_{s=2,4,6,...}^{\infty}} A^{(s)}(x) = \sigma(x)+\sigma(-1-x)=0 \ .
\end{equation}
Here we formally assumed that $\sigma(x)$ is defined for any $x$ using analytic continuation.
In fact, this function is real for $x\in [-\infty,-1]\cup[0,\infty]$ while the argument of the amplitude in \rf{3.17} is
$
x ={ \textsf{\textbf t}\over \textsf{\textbf s}}=-\frac{1}{2}\,(1-\cos\theta)\in[-1,0].
$
In the latter "physical" interval one finds again that $A(x+i\,0)=0$ for any sign of the infinitesimal imaginary part.
In Appendix \ref{B} we provide an independent check of the vanishing of the amplitude \rf{3.25}
at the special kinematical point $\textsf{\textbf u}=0$ or $x=-1$ (or, equivalently, at $x=0$).
Another clarification is that in the above discussion we have excluded the special points $x=0,-1$
where the amplitude may have delta-function singularities as in the external scalar amplitude case \cite{Joung:2015eny}.
Indeed, as was shown in \cite{Joung:2015eny}, the sum of the Legendre polynomials in \rf{555}
is given by $\sum_{s=0}^\infty (s+{1\over 2} ) {\rm P}_s (x ) = \delta (x-1)$, so the total amplitude given by the sum of the s- and t-channels
is $\sim \delta({\textsf{\textbf u}\over \textsf{\textbf s}} ) + \delta({\textsf{\textbf u}\over \textsf{\textbf t}} ) $ which vanishes for real momenta.
Similar cancellation happens here as well as we show in Appendix \ref{B}.
\section{General structure of CHS exchange amplitudes }
To generalise the above vector scattering results to higher $s>1$ spin scattering case
it is useful first to discuss the structure of the CHS 4-particle amplitudes expected on the basis
of Lorentz and scale invariance. It turns out that
the appearance of the special Jacobi polynomials in \rf{3.14}, \rf{555} and \rf{329}
is not accidental and may be related to the partial wave expansion of the
$\lambda} \def \r {\rho_1,\lambda} \def \r {\rho_2 \to \lambda} \def \r {\rho_3,\lambda} \def \r {\rho_4$ transition amplitude discussed by Jacob and Wick \cite{Jacob:1959at} (see also \cite{eden1967high,collins}).
Considering the c.o.m. frame and using the completeness of states
relation one can represent generic scattering amplitude as a sum over on-shell states of a
massive particle with mass$=\sqrt \textsf{\textbf s}$ and spin $J$ \cite{Jacob:1959at}
\begin{align}\label{41}
&\quad A_{\{\lambda_i\} }(\textsf{\textbf s}, \theta) = R_{\{\lambda_i\}} (\theta)
\sum_{J\geq M}(J+\tfrac{1}{2})\, {\rm F}^{(J)}_{\{\lambda_i\}}(\textsf{\textbf s})
\te
\ P_{J-M}^{(|\lambda-\mu|, |\lambda+\mu|)}(\cos\theta)\ ,
\\
&\qquad \label{42}
\lambda=\lambda_{1}-\lambda_{2},\ \ \ \mu=\lambda_{3}-\lambda_{4}, \ \ \ \qquad
M=\max(|\lambda|, |\mu|) \ ,
\\ \label{43}
&\quad \te R_{\{\lambda_i\}}(\theta)= \big(\cos\frac{\theta}{2}\big)^{|\lambda+\mu|}\,\big(\sin\frac{\theta}{2}\big)^{|\lambda-\mu|}= \big(-\frac{\textsf{\textbf u}}{\textsf{\textbf s}}\big)^{\frac{1}{2}\,|\lambda+\mu|}\,\big(-\frac{\textsf{\textbf t}}{\textsf{\textbf s}}\big)^{\frac{1}{2}\,|\lambda-\mu|} \ .
\end{align}
Here ${\{\lambda_i\}}=(\lambda_{1}, \lambda_{2}; \lambda_{3}, \lambda_{4})$,\ \ $\cos \theta = 1 + 2 {\textsf{\textbf t}\over \textsf{\textbf s}}$ and $P^{(a,b)}_k$ is the Jacobi polynomial.
The latter originates from the expression for the spherical $d$-function ($N=\min(|\lambda|, |\mu|) $)
\begin{equation} d^J_{\lambda} \def \r {\rho\m}(\theta) = \te \sqrt{ (J+M)! (J-M)! \over (J+N)! (J-N)! } \ R_{\{\lambda_i\}}(\theta) \ P_{J-M}^{(|\lambda-\mu|, |\lambda+\mu|)}(\cos\theta)
\label{433}\ .
\end{equation}
We assume that the scattering particles are massless.
If the theory is scale-invariant (has no dimensional parameters)
the dependence of the coefficient functions ${\rm F}^{(J)}_{\{\lambda_i\}}$ on $\textsf{\textbf s}$ should be controlled only by dimensions $\Delta_i$
of the scattering fields,
\begin{equation} \label{44}
{\rm F}^{(J)}_{\{\lambda_i\}} (\textsf{\textbf s}) = F^{(J)}_{\{\lambda_i\}}\, \textsf{\textbf s} ^\Delta \ , \ \ \ \ \ \ \ \qquad \ \ \ \ \Delta\equiv { \textstyle{1 \over 2}} (4-\sum^4_{i =1}\Delta_i) \ . \end{equation}
Here $F^{(J)}_{\{\lambda_i\}}$ are some numerical coefficients encoding dynamical information
about a particular theory.
For example, $\Delta=0$ for scattering of dimension 1 massless scalars or vectors in 4d.
In the case of CHS scattering with asymptotic states chosen, as discussed in the previous section,
to be the standard massless spin $|\lambda} \def \r {\rho_i|$ particles the representation \rf{41},\rf{44} should again apply with
$\Delta= { \textstyle{1 \over 2}} (\sum_i |\lambda} \def \r {\rho_i|- 4)$ (the CHS field dimensions are $\Delta_i = 2 - |\lambda} \def \r {\rho_i|$).
Our key observation is that in the present context of conformal higher spin theory
the spin $J$ contribution
in \rf{41} should have the same structure as
the contribution of an intermediate CHS field exchange with $s=J$ in s-channel.
This should be a kinematical consequence of the fact that
a massive ($m^2=\textsf{\textbf s}$) intermediate spin $J$ state in \rf{41} may be described by a totally symmetric
field $\phi_{\m_1...\m_J}$ satisfying $(-\Box + m^2)\phi_{\m_1...\m_J}=0$ as well as the tracelessness and transversality
conditions (leaving only $2J+1$ states as physical degrees of freedom).
At the same time, the CHS scattering is also mediated by the TT field exchange with
the propagator in \rf{3.4}. The only formal difference is in the overall $\textsf{\textbf s}$-dependence that appears in ${\rm F}$
but in the CHS scattering case the latter is controlled by the scale invariance leading to \rf{44}.
This formal interpretation of the spin $J$ term in \rf{41} as the CHS spin $s=J$
exchange amplitude should directly apply only to the s-channel exchange: this is due to the selection
of $\textsf{\textbf s}$ variable as the c.o.m. frame mass parameter in \rf{41} and thus as
the variable that should appear in the propagator of the corresponding exchanged CHS field.
The total CHS amplitude given by the sum over all channels as in \rf{fig4},\rf{3.16} will also have the general form \rf{41}
when expanded in the Jacobi polynomials but the $J=s$ identification of the particular term in the sum
with the contribution of the CHS exchange will be valid only in a
particular channel (in s-channel or after renaming the kinematic variables and helicities also in t- and u-channels, see below).
Another remark is that this identification of the $J$-term in \rf{41} with the higher spin exchange does not apply
to the case of the 2-derivative massless higher spin scattering in flat space discussed in \cite{Ponomarev:2016jqk}. The reason is that
the massless spin $s$ propagator (taken, {\em e.g.}, in the de Donder gauge) is not traceless-transverse
and thus the massless higher spin particle exchange cannot be directly identified with a massive spin $J$ on-shell state
contribution in the sum in \rf{41}.\footnote{Indeed, the scattering
amplitude for four massless scalars exchanging the tower of massless higher spins
was given in \cite{Ponomarev:2016jqk} by the sum of the Chebyshev polynomials rather than the
Legendre polynomials
appearing \rf{555} in the case of the conformal higher spin exchange in \cite{Joung:2015eny}.
Interestingly, there is still a formal relation between 4-scalar scattering via massless higher spin exchange in $d+1$ dimensions and
via conformal higher spin exchange in $d$ dimensions suggesting possible AdS/CFT connection.
}
Let us now see how the previously discussed cases of the external scalar and vector scattering \rf{3.11},\rf{555},\rf{328}
via the CHS exchange may be related to \rf{41}.
In the case of the $\varphi\varphi^{*}\to \varphi\varphi^{*}$ scattering we have
$
\lambda_{i}=0, \ \lambda = \mu = 0, \ M=0,
$
and thus should expect, according to \rf{41},\rf{44}, to find the s-channel spin $J$ contribution to be
\begin{equation}
\label{4.8}
\begin{split}
A^{(J)} _{\textsf{\textbf s}\, 0,0; 0,0}(\textsf{\textbf s}, \cos\theta) =
(J+\tfrac{1}{2}) F^{(J)}_{0,0; 0,0} \,P_{J}^{(0,0)}(\cos\theta)\ .
\end{split}
\end{equation}
Comparing this with the s-channel result \rf{555} of the direct computation
using that $\cos\theta=
1+2\,\frac{\textsf{\textbf t}}{\textsf{\textbf s}} $
we conclude that
the two expression indeed match provided $s=J$ and
\begin{equation} \label{4411}
F^{(s)}_{0,0; 0,0}=1 \ . \end{equation}
For $\varphi\varphi^{*}\to {\pm}{\mp}$ in \rf{345},\rf{328} we have
$
\lambda} \def \r {\rho=\lambda_{1}=\lambda_{2}=0, \
\lambda_{3}=-\lambda_{4} = \pm 1, \ \mu = M= \pm 2,
$
and thus from \rf{41},\rf{44} should get
\begin{equation}
\label{411}
\begin{split}
A^{(J)}_{\textsf{\textbf s}\, 0,0; \pm 1,\mp 1}(\textsf{\textbf s}, \cos\theta) &=
(J+\tfrac{1}{2}) F^{(J)}_{0,0; \pm 1 , \mp 1} \,\frac{\textsf{\textbf u}\,\textsf{\textbf t}}{\textsf{\textbf s}^{2}}
\,P_{J-2}^{(2,2)}(\cos\theta) \ .
\end{split}
\end{equation}
Comparing this with \rf{328},\rf{329} we find perfect match
provided $s$ is identified with $J$ (which should be taken to be even) and\footnote{Note
that
the restriction that $J=s$ should be
even does not follow from \rf{41} and is an extra dynamical property of CHS theory (parity invariance of the original
scalar theory \rf{2.1} implying the absence 1-1-$s$ vertices with odd $s$).
For even $s$ the polynomial $P_{s-2}^{(2,2)}(x)$ is even.}
\begin{equation}\label{412}
F^{(s)}_{0,0;\pm \mp } = \frac{2}{s\,(s-1)}.
\end{equation}
In the case of ${\pm}1 {\pm}1\to {\pm}1 {\pm}1 $ scattering in \rf{3.11}
we have the two contributions of the t- and u- channels
that are to be analysed separately.
For example, considering the t-channel exchange to be able to compare it to \rf{41}
we should first re-interpret it as an s-channel exchange by relabelling the states and Mandelstam variables.
Explicitly, the t-channel scattering of original "X"-particles may be interpreted as s-channel scattering of
effective "Y"-particles, {\em i.e.}
$
X_{1}+\overline X_{3}\to X_{4}+\overline X_{2} $ is equivalent to
$Y_{1}+Y_{2}\to Y_{3}+Y_{4}.
$
For the Y-particles we then have
$
\lambda_{1}=-\lambda_{2}=\pm 1, \
\lambda_{3}=-\lambda_{4} = \pm 1, \
\lambda = \mu = 2, \ M=2
$
and thus from \rf{41},\rf{44} we should get
\begin{equation}\label{47}
\begin{split}
A^{(J)}_{\textsf{\textbf s} \, \pm 1, \mp 1; \mp 1, \pm 1}&=
(J+\tfrac{1}{2}) F^{(J)}_{\pm 1, \mp 1; \mp 1, \pm 1} \,\frac{\textsf{\textbf u}_{Y}^{2}}{\textsf{\textbf s}_{Y}^{2}}
\,P_{J-2}^{(0, 4)}(\cos\theta_{Y}),\qquad \cos\theta_Y = -1-2\,\frac{\textsf{\textbf u}_{Y}}{\textsf{\textbf s}_{Y}}\ .
\end{split}
\end{equation}
The Y-kinematics becomes the X-kinematics after
$\textsf{\textbf s}_{Y}\to \textsf{\textbf t}$, $\textsf{\textbf t}_{Y}\to \textsf{\textbf u}$, $\textsf{\textbf u}_{Y}\to \textsf{\textbf s}$. Thus for the t-channel exchange
of the X-particles we should get
\begin{equation}
\label{4.16}
\begin{split}
A^{(J)}_{\textsf{\textbf t} \, \pm 1, \pm 1; \pm 1, \pm 1}(\textsf{\textbf t}, \cos\theta)&=
(J+\tfrac{1}{2}) F^{(J)}_{\pm 1, \pm 1; \pm 1, \pm 1} \,\frac{\textsf{\textbf s}^{2}}{\textsf{\textbf t}^{2}}
\,P_{J-2}^{(0, 4)}(-1-2\,\frac{\textsf{\textbf s}}{\textsf{\textbf t}})\ .
\end{split}
\end{equation}
This matches the t-channel result in \rf{3.11},\rf{3.14},\rf{3.15} with $J=s$ since
\begin{equation}\label{49}
\big(\frac{\textsf{\textbf s}}{\textsf{\textbf t}}\big)^{s}\,\big(\frac{\textsf{\textbf t}}{\textsf{\textbf s}}\big)^{s-2}\,P_{s-2}^{(4,0)}\big(1+2\,\frac{\textsf{\textbf s}}{\textsf{\textbf t}}\big)
=
\frac{\textsf{\textbf s}^{2}}{\textsf{\textbf t}^{2}}\,P_{s-2}^{(0,4)}\big(-1-2\,\frac{\textsf{\textbf s}}{\textsf{\textbf t}}\big) \ ,
\end{equation}
provided also we choose
\begin{equation}\label{410}
F^{(s)}_{\pm 1, \pm 1; \pm 1, \pm 1} = {c_s \over s+ {1\over 2}} = \frac{4}{(s-1)\,s\,(s+1)\,(s+2)} \ .
\end{equation}
Guided by the above three examples \rf{4411},\rf{412} and \rf{410} we may conjecture the
general dependence of $F^{(J)}_{\{\lambda_i\}}$ in \rf{44}
on $J=s$ in the case of CHS exchange amplitudes to b
\footnote{One reason why this choice may be special
is the following property of the Jacobi polynomials:
\noindent
$
(\lambda+\mu)!\,\frac{(s-M)!}{(s+N)!}\,P_{s-M}^{(|\lambda-\mu|, \lambda+\mu)}(-1) = 1
$ where $\lambda} \def \r {\rho$ and $\m$ should be integer and $s-M$ should be even integer.
This implies the $s$-independence of the backward scattering amplitude in
the direct channel.}
\begin{equation}
\label{5.1}
F^{(s)}_{\{\lambda_i\}} = {\rm k}_{\lambda,\mu}\,\frac{(s-M)!}{(s+N)!}, \qquad\qquad N =\min(|\lambda|,|\mu|),\ M=\max(|\lambda|, |\mu|) \ .
\end{equation}
Then for $ {\rm k}_{\lambda,\mu}=1$ we indeed get
$F^{(s)}_{\{0\}}$ as in \rf{3.14}, $F^{(s)}_{0,0; \pm\mp}$ as in \rf{412}
and $F^{(s)}_{\{\pm 1\}}$ as in \rf{410} (where one should, as explained above, use $\lambda} \def \r {\rho=\m=2$
to match the t-channel result).
It turns out that
this ansatz \rf{5.1} applies also in all other cases discussed below.
\section{Scattering amplitudes with conformal gravitons }
Let us now turn to the discussion of conformal graviton scattering due to the exchange of the CHS fields.
The relevant 2-2-$s$ interaction vertex was given in \rf{2.10}.
As discussed in section 3, we shall be scattering only the "physical" massless spin 2 component
of the conformal spin 2 field, attaching the corresponding asymptotic states to the
amputated Green's functions.\footnote{Let us note also that
using the standard formulation of the CHS action \rf{2} one may also study
the scattering of other "ghost"-like modes described by the higher-derivative
CHS equations.
In general,
the
6 dynamical degrees of freedom of the Weyl graviton way be
described by the collection of the standard massless spin 2 field, massless vector, and massless
spin 2 ghost field states
(for a discussion of solutions of linearised Weyl gravity equations reproducing the
dynamical degrees of freedom
count \cite{Fradkin:1981iu,Fradkin:1981jc,Lee:1982cp} see \cite{Riegert:1984hf} and also Appendix C in \cite{Metsaev:2007fq}).
Explicitly, choosing the TT gauge ($h_m^m=0, \ \partial^mh_{mn} =0$)
we get the free conformal graviton equation as
$\Box^2 h_{mn}=0 $ which is solved by $ h_{mn} = h^{(1)}_{mn} +h^{(2)}_{mn}
= (a_{mn} + b_{mn} u_k x^k) e^{ip \cdot x} + c.c. $ where
$p^2=0 , \ u^2=-1 , \ u \cdot p \not=0\ , \ \ \ a^m_m = b^m_m =0 $.
Here
$h^{(1)}_{mn}$ represents the massless spin 2 and spin 1 modes and
$h^{(2)}_{mn} $ the ghost-like spin 2 mode (which grows in time and leads to negative energy contributions).
Using the Lorentz symmetry and the residual gauge freedom one may choose \cite{Riegert:1984hf}: \
$p^m=(p,0,0,p) , \ \ u^m=(1,0,0,0) \ $, \
$ a_{11}+a_{22}=b_{11}+b_{22}=0\ , \ \ \ a_{m3}=b_{m3}=b_{m0}=0 $
and then
the 2+2+2 dynamical d.o.f. are described by the helicity $\pm 2$ tensor
$(a_{11} \pm i a_{12} ) e^{ip \cdot x}$,
helicity $\pm 1$ vector
$(a_{01} \pm i a_{02} ) e^{ip \cdot x}$
and helicity $\pm 2$ ghost tensor
$(b_{11} \pm i b_{12} ) x^0 e^{ip \cdot x}$.
The spin 1 and ghost spin 2 become parts of massive spin 2 ghost
if one adds the $R$ term to Weyl action to get a diagonal mode decomposition.
At the level of the flat-space partition function of Weyl graviton the above 2+2+2
split corresponds to the following
decomposition \cite{Fradkin:1985am}:
$$Z_2= \Big[
\frac{(\det\Delta_{1})^{3} }{(\det \Delta_{2})^2}
\Big]^{1/2}
= \Big[
\frac{\det\Delta_{1\, \perp}}{\det\Delta_{2\, \perp}}
\Big]^{1/2} \Big[
\frac{\det\Delta_{0\, \perp}}{\det\Delta_{2\, \perp}}
\Big]^{1/2} = \Big[
\frac{\det\Delta_{1\, \perp}}{\det\Delta_{2\, \perp}}
\Big]^{1/2} \Big[
\frac{\det\Delta_{0\, \perp}}{\det\Delta_{1\, \perp}}
\Big]^{1/2} \Big[
\frac{\det\Delta_{1\, \perp}}{\det\Delta_{2\, \perp}}
\Big]^{1/2}
$$
Here $\Delta_s$ are 2-derivative Laplacians defined on traceless rank $s$ symmetric fields.
}
\subsection{$22\to 22$ scattering}
Let us first discuss what we should expect to get for the structure of the 22$\to$22 even spin $s\ge 4 $ exchange
on general symmetry grounds.
We shall assume that as in the 4-vector case the non-vanishing scattering amplitudes should be similar to
\rf{3.11} (where now $\pm$ will stand for $\pm 2$ helicities of the external massless graviton state). Thus
for the $++\to ++$ amplitude
we should have the contributions from the t- and u- channels.\footnote{We again assume two incoming and two outgoing momenta;
choosing all momenta as incoming this becomes the MHV $++--$ amplitude.}
Then repeating the analysis that in the vector case lead to \rf{4.16}
we conclude that for the t-channel exchange of CHS spin $J$ states
we should expect from \rf{41},\rf{44} to find for $ J\geq 4$
\begin{equation}
\label{6.2}
\begin{split}
A^{(J)}_{\textsf{\textbf t} \, \pm 2, \pm 2; \pm 2, \pm 2}(\textsf{\textbf t}, \cos\theta) =
(J+\tfrac{1}{2}) F^{(J)}_{\pm 2, \pm 2; \pm 2, \pm 2}\,\textsf{\textbf t}^{2} \,\frac{\textsf{\textbf s}^{4}}{\textsf{\textbf t}^{4}}
\,P_{J-4}^{(0, 8)}(-1-2\,\frac{\textsf{\textbf s}}{\textsf{\textbf t}})\ .
\end{split}
\end{equation}
Here $\textsf{\textbf t}^2$ factor reflects the fact that the conformal graviton has dimension 0 (cf. \rf{44}).
The total amplitude due to spin $s=J$ exchange should then be as in \rf{3.16} (cf. \rf{3.14})
\begin{align}\label{6.3}
\pm2\pm2 \to \pm2\pm2: \qquad A^{(s)} & = c_{s}\,\textsf{\textbf s}^{2}\,\big[\big(\frac{\textsf{\textbf s}}{\textsf{\textbf t}}\big)^{s-2}\,
P_{s}\big(\frac{\textsf{\textbf t}}{\textsf{\textbf s}}\big)
+\big(\frac{\textsf{\textbf s}}{\textsf{\textbf u}}\big)^{s-2}\,P_{s}\big(\frac{\textsf{\textbf u}}{\textsf{\textbf s}}\big)\big], \\
& \ \ \ P_{s}(x) = \te x^{s-2}\,P_{s-4}^{(8,0)}\big(\frac{x+2}{x}\big)\ . \label{63}
\end{align}
If we also assume the validity of the conjecture \rf{5.1} for the coefficients $F^{(J)}_{\{\lambda} \def \r {\rho_i\}}$
then we may expect also to get
\begin{equation}\label{64}\te
c_{s} ={\rm k}\ \frac{2s+1}{(s-3)(s-2)(s-1)s(s+1)(s+2)(s+3)(s+4)} \ ,
\end{equation}
where ${\rm k}$ is some $s$-independent numerical factor.
Remarkably, the direct computation based on the CHS action and carried out for several
even\footnote{Recall that the 2-2$s$ vertex \rf{2.10} vanishes for odd $s$.}
values of $s\geq 4 $
confirms the above expressions \rf{6.3},\rf{64} and fixes ${\rm k}$ in \rf{64} to be
\begin{equation} \te
{\rm k} = {9\over 8} \ . \label{65} \end{equation}
Similar result is found for the $\pm2\mp 2\to \pm2\mp2$ exchange (cf. \rf{3.11}).
The general derivation of \rf{6.3}--\rf{65} may be given using the same formalism as described for the 11 $\to$ 11 case in Appendix \ref{C}.
We can now sum the amplitude \rf{6.3} over all even $s=4,6,...$ using the same method as in the vector scattering case
\rf{3.17}--\rf{3.24}:
\begin{align} \label{66}
&\sum_{s=4,6,...}^\infty A^{(s)} (x) = \textsf{\textbf s}^2\big[ \sigma(x) + \sigma(-1-x) \big] \ , \ \ \ \ \ \qquad x\equiv \frac{\textsf{\textbf t}}{\textsf{\textbf s}}\ , \\
& \sigma(x) = \lim_{z\to 1} \sum_{s=4,6,...}^\infty c_{s} \, x^{2-s}\,P_{s}(x)\,z^{s-4} \ . \label{67}
\end{align}
After a rather involved computation using the generating function for the Jacobi polynomials in \rf{63}
we found that\footnote{Let us note a similarity in the structure of (\ref{66}) and (\ref{3.24}). This suggests that for
higher spin jj$\to$ jj scattering one may be able to guess
the expression for $\sigma(x)$ and then check that the coefficients in its expansion in a suitable set of
Jacobi polynomials reproduces the $c_{s}$ prefactor. Similar ideas have been
exploited in \cite{gustavsson2001some}.}
\begin{equation}
\label{6.6}
\sigma(x) = \te \frac{1}{4320}\Big[ 60\, (x+1)^3\, x^3\, \log \frac{x+1}{x}
-60\, x^5-150\, x^4-110\, x^3-15\, x^2+3\,
x-1\Big] \ .
\end{equation}
One can then check that the combination of the
$\sigma$ functions appearing in \rf{66} vanishes as in the vector exchange case \rf{3.25}
\begin{equation}\label{677}
\sigma(x)+\sigma(-1-x) = 0 \ ,
\end{equation}
{\em i.e.} the t- and u-channel contributions summed over $s=4,6,...$
cancel against each other.
To find the total $22\to 22$ amplitude
one is still to add (i) the contributions of the low-spin $s <4$ CHS exchanges
({\em i.e.} the exchange mediated by the non-propagating spin 0 field $h_0$
and the exchange of the spin 2 conformal graviton itself)
and also (ii) the contribution of the 2222 contact vertex
that is found from the UV singular part of the diagram \rf{f3} with four spin 2 current insertions
(with vertices in \rf{2.5} as the external legs are assumed to be TT).
We found the following expressions for the spin 0 exchanges with the cubic vertex in \rf{2.111}:
\begin{equation}\label{667}
\begin{split}
& \pm2 \pm2 \to \pm2\pm2: \ \ \qquad \te A_{\textsf{\textbf s}}^{(0)} = \frac{\textsf{\textbf s}^{2}}{4608}, \qquad
A_{\textsf{\textbf t}}^{(0)} = \frac{\textsf{\textbf t}^{2}\,\textsf{\textbf u}^{4}}{512\,\textsf{\textbf s}^{4}}, \qquad
A_{\textsf{\textbf u}}^{(0)} = \frac{\textsf{\textbf t}^{4}\,\textsf{\textbf u}^{2}}{512\,\textsf{\textbf s}^{4}}, \\
& \pm2 \mp2 \to \pm2\mp2: \ \ \qquad \te A_{\textsf{\textbf s}}^{(0)} = 0, \qquad\quad
A_{\textsf{\textbf t}}^{(0)} = \frac{\textsf{\textbf t}^{2}\,\textsf{\textbf u}^{4}}{512\,\textsf{\textbf s}^{4}}, \qquad
A_{\textsf{\textbf u}}^{(0)} = \frac{(\textsf{\textbf s}+3\textsf{\textbf t})^{2}\,\textsf{\textbf u}^{4}}{4608\,\textsf{\textbf s}^{4}} \ .
\end{split}
\end{equation}
The spin 2 exchanges (with the 2-2-2 vertices as in \rf{2.10})
are
\begin{align} \no
& \pm2 \pm2 \to \pm2\pm2: \ \ \qquad \te A_{\textsf{\textbf s} }^{(2)} = \frac{\textsf{\textbf s}^2+6\, \textsf{\textbf s}\, \textsf{\textbf t}+6\, \textsf{\textbf t}^2}{23040}, \quad
A_{\textsf{\textbf t} }^{(2)} =\frac{\textsf{\textbf u}^2 (2\, \textsf{\textbf s}^4-10\, \textsf{\textbf s}^3\, \textsf{\textbf t}
+33\, \textsf{\textbf s}^2\, \textsf{\textbf t}^2-24\, \textsf{\textbf s}\, \textsf{\textbf t}^3+
3\, \textsf{\textbf t}^4)}{7680\, \textsf{\textbf s}^4}\ , \\
&\qquad \qquad \qquad \qquad \qquad \qquad \qquad \label{678}\te A_{\textsf{\textbf u}}^{(2)} =\frac{\textsf{\textbf t}^2 \,(2\, \textsf{\textbf s}^4-10\, \textsf{\textbf s}^3\, \textsf{\textbf u}
+33\, \textsf{\textbf s}^2\, \textsf{\textbf u}^2-24\, \textsf{\textbf s}\, \textsf{\textbf u}^3+3\, \textsf{\textbf u}^4)}{7680\, \textsf{\textbf s}^4} , \\
& \pm2 \mp2 \to \pm2\mp2: \ \ \qquad \te A_{\textsf{\textbf s}}^{(2)} = 0, \quad \no
A_{\textsf{\textbf t} }^{(2)} =\frac{\textsf{\textbf u}^4 \,(2\, \textsf{\textbf s}^2+2\, \textsf{\textbf s}\, \textsf{\textbf t}+3\, \textsf{\textbf t}^2)}{7680\, \textsf{\textbf s}^4} , \quad
A_{\textsf{\textbf u}}^{(2)} =\frac{\textsf{\textbf u}^4 \,(10\, \textsf{\textbf s}^2+18\, \textsf{\textbf s}\, \textsf{\textbf u}+9\, \textsf{\textbf u}^2)}{23040\, \textsf{\textbf s}^4} \ .
\end{align}
The contributions of the 4-derivative 2222 contact vertex which is the $s=2$ analog of \rf{2.99}
are found to be
\begin{equation}\label{68}
\begin{split}
& \pm2 \pm2 \to \pm2\pm2: \ \ \qquad A_{ }^{\rm (cont)} \te = -\frac{\textsf{\textbf s}^6-\textsf{\textbf s}^5\, \textsf{\textbf t}+26\, \textsf{\textbf s}^4\, \textsf{\textbf t}^2+63\, \textsf{\textbf s}^3 \,\textsf{\textbf t}^3
+54\, \textsf{\textbf s}^2\, \textsf{\textbf t}^4+27\, \textsf{\textbf s} \,\textsf{\textbf t}^5+9 \,\textsf{\textbf t}^6}{1920\, \textsf{\textbf s}^4}\ , \\
& \pm2 \mp2 \to \pm2\mp2: \ \ \qquad A_{}^{\rm (cont)} \te = -\frac{\textsf{\textbf u}^{4}\, (\textsf{\textbf s}^2+3 \,\textsf{\textbf s} \,\textsf{\textbf t}+9\, \textsf{\textbf t}^2)}{1920\, \textsf{\textbf s}^4} \ .
\end{split}
\end{equation}
Remarkably, the sum of these three contributions vanishes for each of the helicity
choices:
\begin{equation}\label{69}
[A^{(0)}_{\textsf{\textbf s}} + A^{(0)}_{\textsf{\textbf t}} + A^{(0)}_{\textsf{\textbf u}}] + [A^{(2)}_{\textsf{\textbf s}} + A^{(2)}_{\textsf{\textbf t}}] + A^{(2)}_{\textsf{\textbf u}}
+A^{\rm (cont)} = 0 \ .
\end{equation}
Note that
this result is equivalent to the vanishing of the 4-graviton scattering
amplitude in the non-linear $C^2$ Weyl gravity theory.
Indeed, the linear scalar - CHS coupling action \rf{3} is equivalent to the covariant conformal scalar action \rf{4}
by a local field redefinition \rf{6}. As the latter action directly leads to the Weyl tensor squared
action as the "induced" one \rf{5},
and as the field redefinitions of $h_0$ and $h_2$ in the spin $\leq 2$ part of the
CHS action induced from \rf{3},\rf{2.1} should not change the graviton S-matrix, the latter
should be same as in the Weyl theory. In more detail,
adding the exchange of the non-propagating $h_0$ field produces (as it follows from \rf{2.111})
an extra 4-derivative 2222 contact vertex
contribution.
The remaining local redefinition of $h_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}}$ in \rf{6} may alter the 222 vertices by terms proportional to the linearised equations of motion ($\Box^2 h_{\m\nu} \def \m {\mu}\def \G {\Gamma} \def \J {{\cal J}\ } ^{\rm TT}=0$) and also change the quartic 2222 vertex, but
it cannot change the resulting on-shell
4-graviton scattering amplitude.
The vanishing of the tree-level 4-graviton amplitude in Weyl theory can be
deduced also
from the expressions in \cite{Dona:2015tra} for the massless graviton scattering in
$L= a R + b C^2$ theory by taking the limit $a\to 0$ in the final expression for the 4-graviton amplitude.
The propagator here is symbolically $\frac{1}{a p^{2} + b p^4}$ \cite{Stelle:1977ry}
(reducing to the Weyl graviton propagator for $a\to 0$ or to the Einstein propagator for $b\to 0$)
so as long as the asymptotic states are chosen to be massless helicity $\pm 2$ gravitons the resulting amplitude
interpolates smoothly between the standard Einstein
4-graviton one and zero in the Weyl theory.\footnote{Let us also mention that
the conformal graviton
amplitudes in flat space were computed in \cite{Dolan:2008gc} starting with
the twistor string theory of \cite{Berkovits:2004jj}. The latter should be related to "non-minimal" conformal supergravity
containing extra dimension 0 scalar coupling to Weyl squared term, $\phi \Box^2 \phi + (1+ k\, \phi + ... )C^2 + ...$.
The tree-level 4-graviton amplitude in such theory is given by the sum of the 4-graviton
amplitude in Weyl theory and the scalar exchange $\sim k^2 C^2 \Box^{-2} C^2$.
The non-zero result for the 4-graviton amplitude found in \cite{Dolan:2008gc} appears to be given just by this
scalar exchange, {\em i.e.} is consistent with the vanishing of the graviton amplitude in pure Weyl theory.
Similar result was found in \cite{Adamo:2012xe,Adamo:2013tja} by
taking the flat limit of the conformal graviton scattering amplitude in dS space
which is the same as the Einstein gravity one \cite{Maldacena:2011mk} times the cosmological constant factor.}
Remarkably, as we have just seen, the vanishing of tree-level 4-graviton amplitude in Weyl theory
generalises to the full CHS theory:
the results \rf{66},\rf{677} and \rf{69} combined together imply
that like the 11$\to$11 amplitude in \rf{3.25} (and also the
external conformal scalar amplitude \cite{Joung:2015eny}) the total 22$\to$22 conformal graviton scattering amplitude
in the CHS theory vanishes after all intermediate exchange
contributions are added together.
\subsection{$11\to 22$ scattering}
One may also consider some "mixed" 4-particle amplitudes involving both vectors and conformal gravitons.
The amplitudes with odd number of vectors vanish identically so one is to consider only 11$\to $ 22 case.
Here the two a priori non-trivial helicity choices are $\pm1\mp1\to \pm2 \mp2$ and $\pm1\pm1\to \pm2 \mp2$.
Let us first briefly mention also the expressions for the "mixed" amplitude
where two external conformal scalars $\vp$ in \rf{2.1}
scatter into two conformal gravitons.
As in the $\varphi\varphi^{*}\to 11$ case \rf{345} the non-vanishing helicity-preserving amplitude
$\vp\vp^* \to \pm 2\mp 2$ receives contributions from even spin $s\ge 4$ exchanges that should have the general structure
consistent again with $J=s$ term in \rf{41},\rf{44}:
\begin{equation}
\label{6.14}
\begin{split}
A^{(s)}_{\textsf{\textbf s}\ 0,0; \pm 2 ,\mp 2}
=\te (s+\tfrac{1}{2}) F^{(s)}_{0,0; \pm 2, \mp 2}\,\textsf{\textbf s} \,\big(\frac{\textsf{\textbf u}\,\textsf{\textbf t}}{\textsf{\textbf s}^{2}}\big)^{2}
\,P_{s-4}^{(4,4)} \big(-1-2\,\frac{\textsf{\textbf t}}{\textsf{\textbf s}}\big) \ .
\end{split}
\end{equation}
The explicit computation for $c_s =(s+\tfrac{1}{2}) F^{(s)}_{0,0; \pm 2, \mp 2}$ gives again the result consistent
with the ansatz \rf{5.1} (here $\lambda} \def \r {\rho=0, \ \mu=4, \ M=4, N=0$)
\begin{equation}\label{613}\te
c_{s} =(s+\tfrac{1}{2}) F^{(s)}_{0,0; \pm 2, \mp 2} = -\frac{3}{4}\,\frac{2s+1}{(s-3)(s-2)(s-1)s} \ .
\end{equation}
To get the full amplitude one is to add also the contributions of the $s=0, 2$ exchanges.
Turning to the $\pm1\mp1\to \pm2 \mp2$ amplitude, we find that the non-vanishing helicity-preserving even $s\geq 4$ exchange
amplitude in the s-channel has again the form as predicted by \rf{41},\rf{44}
\begin{align} \label{620}
&A^{(s)}_{\textsf{\textbf s} \ \ \pm1,\mp1; \pm2,\mp2} = c_s\, \textsf{\textbf s}\, \te
\frac{\textsf{\textbf t}\,\textsf{\textbf u}^{3}}{\textsf{\textbf s}^{4}}\,P_{s-4}^{(6,2)}(-1-2\,\frac{\textsf{\textbf t}}{\textsf{\textbf s}}) \ , \\
& \ \ \ c_s = (s+\tfrac{1}{2}) F^{(s)}_{\pm1,\pm 1; \pm 2, \mp 2} = \te \frac{3}{2}\,\frac{2s+1}{(s-3)(s-2)(s-1)s(s+1)(s+2)}\ , \label{621}
\end{align}
where $\lambda} \def \r {\rho=2, \mu=4, M=4, N=2$ so the expression for $c_s$ is again consistent with \rf{5.1}.
In the t-channel one finds (after an appropriate relabelling of helicities and kinematic variables)
that for odd $s\geq 3$ \footnote{Here we took into account that
for odd spin the momentum space propagator has an extra factor $(-1)^s=-1$, cf. \rf{2.6}.}
\begin{align} \label{6211}
&A^{(s)}_{\textsf{\textbf s} \ \ \pm1,\mp1; \pm2,\mp2} = c'_s\, \textsf{\textbf s}\, \te
\frac{\textsf{\textbf u}^{3}}{\textsf{\textbf s} \textsf{\textbf t}^{2}}\,P_{s-4}^{(6,0)}(-1-2\,\frac{\textsf{\textbf s}}{\textsf{\textbf t}}) \ , \\
& \ \ \ \ c'_s = \te \,\frac{2s+1}{(s-2)(s-1)s(s+1)(s+2)(s+3)}\ . \label{62111}
\end{align}
The u-channel contribution is zero.
The total s- plus t-channel contribution to the
amplitude from these higher spin exchanges is then
\begin{align}
&\qquad \qquad A_{s>2}= {\textsf{\textbf u}^3\over \textsf{\textbf s}^2} \bar A({\textsf{\textbf t}\over \textsf{\textbf s}}) \ , \qquad \qquad
\bar A(x)= x\,S(x)+x^{-2}\,T(x^{-1}) \ , \label{622} \ \\
&S(x) \equiv \sum_{s=4,6,8,...}^{\infty} c_s \,
\,P_{s-4}^{(6,2)}(-1-2\,x) \ ,\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ T(x) \equiv \sum_{s=3,5,7,...}^{\infty} c'_s \,
\,P_{s-4}^{(6,0)}(-1-2\,x) \ . \label{623}
\end{align}
The explicit evaluation of $S(x)$ and $T(x)$ for $-1<x<0$ gives
\begin{equation}\label{624}
\begin{split}\te
S(x) = -\frac{x^3+5 x^2+13 x-3}{96 (x+1)^5}-\frac{x \log (-x)}{8 (x+1)^6}, \qquad \qquad
T(x) = -\frac{(x-1) (x^2+8 x+1)}{96 (x+1)^5}-\frac{x^2 \log (-x)}{8
(x+1)^6}\ ,
\end{split}
\end{equation}
so that $ \bar A = -\frac{1}{96\,(x+1)}$. The resulting contribution of all higher $s>2$ spin exchanges to the total
amplitude is thus
\begin{equation}\label{625} \te
A_{s>2}= -\frac{1}{96}\,\frac{\textsf{\textbf u}^{3}}{\textsf{\textbf s}^{2}}\frac{1}{\frac{\textsf{\textbf t}}{\textsf{\textbf s}} +1 } = \frac{1}{96}
\frac{\textsf{\textbf u}^{2}}{\textsf{\textbf s}} \ .
\end{equation}
We are still to add possible contributions of low-spin $s=0,1,2$ exchanges and contact 1122 vertex.
The $h_0$ exchange is trivial as the $110$ vertex \rf{2.7} vanishes for the $\pm1,\mp1$ helicity choice.
The $h_2$ exchange is also found to vanish. The $h_1$ exchange gives the following non-zero
contributions in the t- and u-channels
\begin{equation}\te \pm1\mp1 \to \pm2\mp2: \ \ \ \ \qquad
A^{(1)}_{\textsf{\textbf t} } = \frac{\textsf{\textbf u}^{3}\,(2\textsf{\textbf s}+\textsf{\textbf t})}{192\,\textsf{\textbf s}^{3}}, \qquad \qquad
A^{(1)}_{\textsf{\textbf u} } = \frac{\textsf{\textbf t}\,\textsf{\textbf u}^{2}\,(\textsf{\textbf s}-\textsf{\textbf t})}{192\,\textsf{\textbf s}^{3}}
\ . \label{628}
\end{equation}
The 1122 contact term can be found by computing the UV singular part of the scalar loop diagram
with two spin 1 and two spin 2 current insertions
\begin{equation}\label{fig5}
\vcenter{\hbox{\includegraphics[scale=0.3]{fig-1122-contact-A-draft.pdf}
\hskip 1cm
\includegraphics[scale=0.3]{fig-1122-contact-B-draft.pdf}}}
\end{equation}
and then replacing the legs with physical polarisations.\footnote{Here the diagrams with the opposite loop orientation
give equal contributions.}
We get
$
I_{1} = \frac{1}{20}\,\frac{\textsf{\textbf t}\,\textsf{\textbf u}^{3}}{\textsf{\textbf s}^{3}}, \ \ I_{2} = -\frac{1}{60}\,\frac{\textsf{\textbf t}\,\textsf{\textbf u}^{3}}{\textsf{\textbf s}^{3}}$.
Including combinatorics factors the total contribution is $ (-1) {1\over (2!)^4} 2 (2I_1 +I_2)$, {\em i.e.}
\begin{equation}
\label{627}\te \pm1 \mp1 \to \pm2 \mp2: \ \ \ \ \qquad \qquad
A^{\rm (cont)}_{ } = -\frac{1}{96}\frac{\textsf{\textbf t}\textsf{\textbf u}^{3}}{\textsf{\textbf s}^{3}} \ .
\end{equation}
The total amplitude given by the sum of the contributions
of the higher-spin exchanges \rf{625}, spin 1 exchange \rf{628} and the 4-point contact vertex \rf{627}
is found to vanish
\begin{equation}\label{629}
\pm1\mp1 \to \pm2\mp2: \ \ \ \ \qquad A_{s>2} + \big[
A^{(1)}_{\textsf{\textbf t} } + A^{(1)}_{\textsf{\textbf u} }\big] + A^{(\rm cont)}=0 \ .
\end{equation}
Let us now consider the second non-trivial helicity amplitude $\pm1\pm1\to \pm2 \pm2$.
Here we find that the s-channel amplitude vanishes while
the higher odd $s\geq 3$ spin exchange contributions in t- and u-channels have the form
consistent with the general expectations \rf{41},\rf{44},\rf{5.1}
\begin{align} \te
\pm1\pm1\to \pm2\pm2\ : \qquad \ \ \ \ \te
&A_{\textsf{\textbf t}}^{(s)}=\te c_s \frac{\textsf{\textbf s}^{3}}{\textsf{\textbf t}^{2}}
\,P_{s-3}^{(0,6)}(-1-2\frac{\textsf{\textbf s}}{\textsf{\textbf t}}), \ \qquad
A_{\textsf{\textbf u}}^{(s)} = c_s \frac{\textsf{\textbf s}^{3}}{\textsf{\textbf u}^{2}}
\,P_{s-3}^{(0,6)}(-1-2\frac{\textsf{\textbf s}}{\textsf{\textbf u}})\ , \no \\
& \qquad \te c_s = -\frac{2s+1}{(s-2)(s-1)s(s+1)(s+2)(s+3)} \ , \ \ \ \ \ s=3,5,7,...\ \ . \label{640}
\end{align}
The sum over all odd spin $s\ge 3$ exchanges
can be done by observing that $
P_{s-3}^{(0,6)}(-1-2x) = P_{s-3}^{(6,0)}(1+2x)$ and that $c_s$ in \rf{640} is minus $c_s'$ in \rf{62111}.
One can then use the expression for $T(x)$ in \rf{623},\rf{624} to find that the sum
of t- and u-channel amplitudes in \rf{640} vanishes as a consequence of ($x = {\textsf{\textbf s}\over \textsf{\textbf t}}$)
\begin{equation}\te \label{642}
x^{2}\,T(-1-x)+(\frac{x}{1+x})^{2}\,T(-\frac{1}{1+x}) = 0\ .
\end{equation}
For the non-vanishing low-spin exchange and the 1122 contact term contributions here we get
\begin{align}
\pm1\pm1\to \pm2\pm2\ : \qquad & \te
A_{\textsf{\textbf s}}^{(0)}= -\frac{\textsf{\textbf s}}{128}, \qquad
A_{\textsf{\textbf t}}^{(1)}= \frac{\textsf{\textbf u}^{2}\,(\textsf{\textbf s}^{2}-6\,\textsf{\textbf s}\,\textsf{\textbf t}+2\textsf{\textbf t}^{2})}{128\,\textsf{\textbf s}^{3}}, \qquad
A_{\textsf{\textbf u}}^{(1)}= \frac{\textsf{\textbf t}^{2}\,(\textsf{\textbf s}^{2}-6\,\textsf{\textbf s}\,\textsf{\textbf u}+2\textsf{\textbf u}^{2})}{128\,\textsf{\textbf s}^{3}}, \no\\
& \label{644}\te
A^{\text{(cont)}} = -\frac{\textsf{\textbf t}\,\textsf{\textbf u}\,(\textsf{\textbf t}^{2}+3\,\textsf{\textbf t}\,\textsf{\textbf u}+\textsf{\textbf u}^{2})}{32\,\textsf{\textbf s}^{3}}.
\end{align}
They separately sum up to zero
\begin{equation}\label{645}
A_{\textsf{\textbf s}}^{(0)} + [A_{\textsf{\textbf t}}^{(1)} + A_{\textsf{\textbf t}}^{(1)}] + A^{\text{(cont)}} =0 \ ,
\end{equation}
so the total $\pm1\pm1\to \pm2\pm2$ amplitude is again zero.
We conclude that like the 11$\to$11 and 22$\to$22 amplitudes, the 11$\to$22 amplitudes also vanish.
It is thus natural to conjecture that all higher spin amplitudes in the CHS theory should also vanish.
In Appendix \ref{B} we provide a check of this conjecture by demonstrating that the exchange amplitude
for the scattering of four spin j CHS particles constructed using the general relations in \rf{41},\rf{44},\rf{5.1}
vanishes at the special kinematical point $\textsf{\textbf u}=0$ ({\em i.e.\ } for backward scattering).
\section{Concluding remarks}
In this paper we provided evidence that tree-level 4-particle scattering amplitudes for (the standard massless
modes of) conformal higher spin fields vanish after summing over all intermediate CHS exchanges.
The amplitudes vanish due to cancellation between
the summed up contributions of different scattering channels.
This is an indication that this cancellation may be a consequence of the underlying higher spin symmetry
which is an infinite dimensional extension of the usual conformal symmetry.
Indeed, the CHS theory inherits the global higher spin symmetry of the free scalar theory (which is also a
symmetry of the dual massless higher spin theory in AdS$_5$).
This symmetry acts on the scalar $\vp$ and the source fields $h_s$ in the coupled action \rf{2.1}
and thus becomes the symmetry of the local UV part of the induced action
\cite{Segal:2002gd,Bekaert:2010ky}.\footnote{It also acts on the
correlation functions of currents $J_s$ at separated points and, vice versa, requiring it to be a symmetry of these correlation functions implies that they should correspond to a free CFT \cite{Maldacena:2011jn,Boulanger:2013zza,Stanev:2013qra,Alba:2015upa}.}
As was shown in \cite{Joung:2015eny}, the vanishing of the conformal scalar 4-point amplitude in the coupled
scalar -- CHS theory can be understood as a consequence of a particular subset of transformations of the higher spin algebra -- the hypertranslations $\delta \vp = \epsilon^{\m(k)} \partial_{\mu(k)} \vp$ and the
rescalings.
Similar reasoning should apply also in the case of the scattering amplitudes with CHS fields on external lines considered in the present paper. The transformation of the CHS fields $h_s$
under the differential
part of the gauge symmetry is symbolically
$\delta h= \partial \epsilon + \epsilon \partial h $, i.e. like the usual diffeomorphisms it contains an inhomogeneous
and homogeneous parts with the latter mixing different spins. The global part of the algebra corresponds to $\epsilon$ being chosen as conformal Killing tensors. For example, for constant $\epsilon$
\begin{align}
& \delta h_0 = \sum_k \epsilon^{\m(k)}\partial_{\m(k)} h_0 \ , \ \ \ \ \ \qquad
\delta h^{\r} =\sum_k \big[ \epsilon^{\r\m(k)}\partial_{\m(k)} h_0 + \epsilon^{\m(k)} \partial_{\m(k)} h^{\r} \big] \ , \label{6.1} \\
&\delta h^{\r\sigma}= \sum_k \big[ \epsilon^{\r\sigma\m(k)}\partial_{\m(k)} h_0 + 2 \epsilon^{\m(k) (\r}\partial_{\m(k)} h^{\sigma)}
+ \epsilon^{\m(k)}\partial_{\m(k)} h^{\r\sigma}\big] \ , ... \label{6.22}
\end{align}
These transformations relate Green's functions with different types of legs $h_0, h_1, h_2, ...$.
In the case of the S-matrix where the non-propagating field $h_0$ does not appear on external lines
the transformation of $h_1$ under hypertranslations will be the same as of the conformal scalar in
\cite{Joung:2015eny} so that choosing $\epsilon^{\m_1...\m_k}= y^{\m_1 }... y^{\m_{k}}$ where $y^\m$ is an arbitrary constant vector we may then repeat the argument of \cite{Joung:2015eny} for the vanishing of the corresponding scattering amplitude.
Similar arguments should also apply to amplitudes involving conformal gravitons.
As the same higher spin algebra controls also the massless higher spin theory in AdS$_5$ \cite{Vasiliev:2003ev}
(of which the CHS theory is an effective 4d "shadow" or corresponds to the alternative choice of the boundary conditions
for higher spin fields \cite{Giombi:2013yva})
it would be interesting to know if there is an AdS related argument for
the vanishing of the CHS S-matrix.
One may also start with the CHS theory defined on AdS$_4$ or dS$_4$ and try to generalise the arguments of
\cite{Maldacena:2011mk,Adamo:2013tja} to argue that the corresponding S-matrix for massless higher spin modes of the CHS fields should be the same as in the corresponding
massless higher spin theory.
Taking the flat limit (i.e. the cosmological constant to zero)
may then lead to the conclusion that the S-matrix of the resulting hypothetical massless higher spin theory in flat space should also be trivial.
\iffa
\begin{itemize}
\item About the auxiliary field $h_{0}$, things may be better if we consider free fermions instead of free bosons. With only fermions the currents are still bosonic. Their general form
is $j_s \sim \overline\psi\partial^{s-1}\gamma\psi$ with dimension $\frac{3}{2}+s-1+\frac{3}{2} = s+2$ as in bosonic case. However, derivatives are different.
At spin 0, the two point function $\langle j_{0}j_{0}\rangle\sim p^{2}$ so we get a dynamical
$\phi \Box \phi$. This should be the usual remark about values of $\Delta$ for the singlet in
bosonic or fermionic case. For spin $s>0$, $\langle j_{s}j_{s}\rangle\sim p^{2s}$ as in bosonic case.
By the way, it seems quite unexplored the susy case where one starts with a free chiral field with both
$\phi$ and $\psi$. In this case we have also half integer currents.
\end{itemize}
\iffa
The definition of scattering amplitudes with external CHS fields of spin $s>1$
requires a dedicated discussion.
\begin{itemize}
\item a general problem is the violation of unitarity. Remarkably, this is a problem that has no
relevant -- {\em i.e.} stumbling blocks -- consequences in the perturbative treatment we are considering.
\item a technical problem is how to adapt the LSZ formula in presence of propagators
with strong IR singularities. We bypass this discussion and define on-shell amplitudes as coefficients in on-shell value of effective
action written as functional of in-fields that is solutions of free equations of motion.
{\em Useful references to be discussed/added: }
general scattering \cite{Jacob:1959at,Carmi:2011dt,Dolan:2011dv}
4-graviton scattering in Weyl gravity \cite{Dona:2015tra}
graviton helicity tensors ? \cite{Gross:1968in,Grisaru:1975bx}
no-go theorems \cite{Bekaert:2010hw,Taronna:2011kt}
CHS and twistors \cite{Adamo:2012xe,Adamo:2013tja,Haehnel:2016mlb}
\cite{eden1967high}
\fi
\acknowledgments
We would like to thank T. McLoughlin, R. Metsaev, D. Ponomarev and R. Roiban for useful discussions.
The work
of SN and AAT was supported by the STFC Consolidated grant ST/L00044X/1.
The work of AAT was also supported by the
ERC Advanced grant No.290456 and
by the Russian Science Foundation grant 14-42-00047 associated with Lebedev Institute.
\
\newpage
|
\section{Preliminaries and Notation}
At the joint birthday party of Bilbo and Frodo Baggins \cite{Tolkein},
the head table seats 144 which, we are told, is the sum eleventy-one
and thirty-three, the respective ages of the two hobbits. A
\emph{repdigit} is a positive integer all of whose (base-10) digits
are the same. The hobbits have provided us with an example of two
repdigits adding to a perfect square: $111+33 = 144$. In this paper
we find all examples of two repdigits summing to a square.
While Sloan (A010785) allows 1-digit numbers to be repdigits, in this paper
we will not, (since nothing is \emph{rep}eated.) In this way, we avoid examples
like $2+2=4$ and $7+9=16$. Also, if we add a repdigit consisting of an
even number of 9's to 1 (a 1-digit repdigit) the result is an even power
of 10, which is a perfect square. We prefer to avoid these cases and
assume all repdigits are greater than $10$.
For the sake of definiteness, if $a$ is a decimal digit, the repdigit
$aa\ldots a$, consisting of $m$ $a$'s will be denoted $a_m$. In this
notation, the hobbit example reads $1_3 + 3_2 = 12^2$. A little algebra
yields the formula:
$$a_m = aa\ldots a = a(11\ldots 1) = a\left(\frac{10^m-1}{9}\right).$$
The main result of the present paper is a complete classification
of perfect squares that are the sum of two repdigits.
\begin{theorem}
\label{main}
The only perfect squares that are sums of two repdigits are the following:
\begin{align*}
11^{2} &= 99+22 = 88+33 + 77+44 = 66+55\\
12^{2} &= 111+33\\
38^{2} &= 1111+333\\
211^{2} &= 44444 + 77.
\end{align*}
\end{theorem}
In Sections~\ref{mod1} and \ref{mod2} we use congruence conditions to
show that if $a_{m} + b_{n}$ is a perfect square then one of $m$ or
$n$ is $< 6$ and eliminate as many cases as possible. In
Section~\ref{ec}, we handle the remaining cases by using results about
integer points on elliptic curves and conclude the proof of
Theorem~\ref{main}. In Section~\ref{further} we raise some further
questions.
\section{A Couple of Pleasant Properties of Base-10}
\label{mod1}
Of course, once our result is proven in base-10, we would want to
generalize to other bases. However, $10-1$ is a perfect square, a
property of $10$ not true of most bases, so the
problem may be much harder in non-decimal bases. Another happy
accident is that there is a useful string of consecutive quadratic non-residues
modulo $10^6$, which is not to be expected using other bases. This
all comes out in our first theorem, where we show that in the case of
two repdigits adding to a square, at least one of the repdigits has
fewer than 6 digits:
\begin{theorem}
\label{lessthansix}
Let $a$ and $b$ be non-zero base-10 digits. If $a_m + b_n$ is a
perfect square, with $m \geq n \geq 2$, then $n<6$.
\end{theorem}
\begin{proof}
Suppose that $a_m+b_n = k^2$ for some integer $k$, and
that $m$ and $n$ are both at least 6. Then using
our formula above, we have
$$ a\left(\frac{10^m-1}{9}\right) + b\left(\frac{10^n-1}{9}\right) = k^2 $$
which simplifies to
\begin{equation}
a10^m +b10^n -(a+b) = (3k)^2.
\end{equation}
Since $m$ and $n \geq 6$, we can reduce this last equation modulo $10^6$
to get
$$-(a+b) \equiv (3k)^2 \pmod{10^6}.$$
Because $a$ and $b$ are non-zero base-10 digits, $-(a+b)$ must be an integer
between $-2$ and $-18$, inclusive. The congruence above insists that
$-(a+b)$ is a quadratic residue modulo $10^6$. But inspection of
Table A in the Appendix shows that this is impossible. That is, the
last column in the table consists of all X's, which means that none
of the numbers $-2, -3, -4, \ldots, -18$ is a quadratic residue
modulo $10^6$. We conclude that $n<6$.
\end{proof}
\section{Congruence Considerations}
\label{mod2}
The rest of this paper is devoted to showing that $m$ must also
be less than $6.$ So we now assume that $m\geq 6$ and start
eliminating cases. Our definition of \emph{repdigit} implies that
$m$ and $n$ are both at least $2$, so reducing Equation (1)
modulo $10^2$ gives us
$$-(a+b) \equiv (3k)^2 \pmod{10^2}.$$
Inspection of the second column of Table A shows us that $-(a+b)
\equiv -4, -11$, or $-16 \pmod{10^2}$, which divides our work into
three cases.
\vspace*{.1in}
\noindent{\bf Case 1:} $(a+b) = 4.$
\vspace*{.1in}
In this case the only possibilities are $(a,b) = (3,1), (2,2), \text{ or } (1,3)$.
Also note that, from Table A, $-4$ is a quadratic non-residue modulo $10^4$,
so we must have $n<4$ in this case. (Else, we reduce Equation (1) modulo
$10^4$ and get a contradiction.) Thus, the only possibilities are:
$$3_m+11, 2_m + 22, 1_m + 33, 3_m +111, 2_m + 222, \mbox{ and } 1_m+333.$$
Four of these six subcases can be eliminated by congruence considerations.
If we put the case $3_m+11$ into equation (1) and reduce modulo $10^6$, we have
$$3\cdot10^m -3 +10^2 - 1 \equiv 96 \equiv (3k)^2 \pmod{10^6}.$$
A solution to this congruence implies $96+d10^6 = x^2$ for some integers $d$
and $x$. Factoring $2^5$ from the left side gives us $2^5(3+2d \cdot 5^6) = x^2$.
The quantity in parentheses is odd, and so $x^2$ must be exactly divisible
by $2^5$, which is impossible. In other words, $96$ is not a quadratic
residue of $10^6$, so there is no solution. In the future, we'll just ask Maple
whether a number is a quadratic residue.
With a similar calculation, the three subcases $1_m+33$, $2_m +222$ and $1_m+333$
require $296$, $1996$ and $2996$, respectively, to be quadratic residues modulo $10^6$.
Maple says that all three of them are quadratic nonresidues, so we have eliminated
these three subcases. We will handle the remaining cases $2_m+22$ and $3_m+111$
later.
\vspace*{.1in}
\noindent{\bf Case 2:} $(a+b) = 11.$
\vspace*{.1in}
In this case the possibilities are $(a,b) = (2,9)$, $(3,8)$, $(4,7)$, $(5,6)$,
$(6,5)$, $(7,4)$, $(8,3)$, or $(9,2).$ Note that if $n\geq 3$, then modulo $10^3$, equation
(1) reduces to $-11 \equiv (3k)^2 \pmod{10^3}$, and that $-11$ is a quadratic nonresidue
of $10^3.$ Therefore $n=2$ is the only possibility. This leaves us with 8 subcases:
$2_m+99$, $3_m +88$, $4_m +77$, $5_m+66$, $6_m+55$, $7_m+44$, $8_m+33$, and $9_m+22$.
As we did in Case 1, we put each of these subcases in equation (1) and reduce
modulo $10^6$. Respectively, these congruences require $889$, $789$, $689$,
$589$, $489$, $389$, $289$ and $189$ to be quadratic residues modulo $10^6$.
Maple says that $789$, $589$, $389$ and $189$ are quadratic non-residues
module $10^6$, so we have eliminated the subcases $3_m+88$, $5_m+66$, $7_m+44$
and $9_m+22$. The other four subcases we save for Section~\ref{ec}.
\vspace*{.1in}
\noindent{\bf Case 3:} $(a+b) = 16.$
\vspace*{.1in}
In this case the possibilities are $(a,b) = (7,9)$, $(8,8)$ or
$(9,7)$. The $(8,8)$ subcase would have $8_m + 8_n = k^2$, in which
case $k$ would be even and we could divide the equation by $4$ to get
$2_m + 2_n = (k/2)^2$, which is one of our leftover subcases from Case
1. When $2_m +22$ is eliminated in Section~\ref{ec}, it will take
this subcase with it.
From Theorem 1, $n\leq 5$, so the possibilites are $9_m+77$, $9_m + 777$,
$9_m+7777$, $9_m+77777$, $7_m+99$, $7_m+999$, $7_m + 9999$, and $7_m + 99999.$
We again put each subcase into equation (1) and reduce modulo $10^6$. Respectively,
$700 - 16$, $7000-16$, $70000 - 16$, $700000 - 16$, $900-16$, $9000-16$,
$90000 - 16$ and $900000-16$ are required to be quadratic residues modulo $10^6$.
But only $700000-16$, $90000 - 16$ and $900000-16$ are. So we are left
with only $9_m + 77777$, $7_m+9999$, and $7_m+99999$ as possibilities in this case.
It turns out that $90000-16$ and $700000-16$ are not a quadratic residues modulo 7,
so we can also eliminate $9_m+77777$ and $7_m+9999$ from the list.
\section{Elliptic Curve Considerations}
\label{ec}
We have eliminated all cases except this short list: $2_m+22$, $3_m+111$,
$2_m+99$, $4_m+77$, $6_m+55$, $8_m+33$ and $7_m+99999$. Each of these cases
will be broken into three subcases depending on the character of $m$ modulo
3. In most cases, this will eliminate the case. In the rest, we will discover
the known solutions to the problem. We will rely on SAGE to find all the
integer points on the elliptic curves which arise.
We'll start with $8_m+33$. Suppose $m=3l$ for some integer $l$. Then
equation (1) becomes $8\cdot 10^{3l} + 300 - (8+3) = (3k)^2$. If we
set $y = 3k$ and $x=2\cdot 10^l$, the equation is $y^2 = x^3 +
289$.
According to Siegel's famous theorem \cite{Siegel}, this curve can
have only finitely many integral points. SAGE 6.7 implements an
algorithm due to Pethö, Zimmer, Gebel, and Herrmann \cite{PZGH}, which
will find all integral points on an elliptic curve.
SAGE says that the only integer points on this curve are $(-4, 15)$,
$(0,17)$ and $(68, 561)$. We can observe that none of the $x$-values
are of the form $2\cdot 10^l$, which eliminates this possibility.
Next suppose that $m=3l+1$ for some integer $l$. Equation (1), for this case,
reads $8\cdot 10^{3l+1} + 289 = (3k)^2.$ Multiply through
by $100$ to get $8\cdot 10^{3l+3} + 28900 = (30k)^2$. Set $x=2\cdot 10^{l+1}$
and $y=30k$ and we have $y^2 = x^3+28900$. SAGE says that the only
integer point on this curve is $(0,170)$ and since $0 \neq 2\cdot 10^{l+1}$,
we have eliminated this case.
Next suppose that $m=3l+2$ for some integer $l$. Equation (1), for this case
reads $8\cdot 10^{3l+2} + 289 = (3k)^2$. Multiply through by $10^4$ to get
$8\cdot 10^{3l+6}+289\cdot 100^2 = (300k)^2$. Set $y=300k$ and $x=2\cdot 10^{l+2}$
and we have $y^2 = x^3+2890000$. SAGE says that the only integer points on
this curve are $(-136,612)$, $(0,1700)$, $(200,3300)$ and $(425,8925)$. Looking
at the $x$-values, we see one that fits: $200 = 2\cdot10^{l+2}$. So $l = 0$
and $m=2$ and we're led to the solution $88+33 = 121$.
\vspace{.1in}
The other cases yield similarly. If $m=3l+r$, with $r=0,1,$ or $2$, then
we multiply equation (1) by $a^2 10^r$ to get
$$a^310^{3(l+r)} +a^2 10^r(b10^n-(a+b)) = (3a\cdot10^r k)^2.$$
Then we set $x=a10^{l+r}, y= 3a\cdot10^rk$ and $N=a^2 10^r(b10^n-(a+b))$ to
get the elliptic curve
$$y^2 = x^3 + N.$$
Then we ask SAGE to give us all
the integer points on the curve. If any of those points have $x$-coordinate
of the form $a \cdot 10^p$, then we might have a solution (and in each case,
it will turn out that $m<6$.) If not, the case
is eliminated. In the case $a=8$, which is already a cube, it is
sufficient to multiply equation (1) by $10^r$, as we did above.
Table B in the Appendix summarizes the calculations, where $r$ is the least
non-negative residue of $m$ modulo $3$. Whenever the $x$-coordinate of
an integer point matches the form $a10^{l+r}$, we put that number in bold
in the last column of the table. In each such case, $l=0$ or $1$, which
means that $m<6$.
\begin{proof}[Proof of Theorem~\ref{main}]
Therefore every case of the sum of two repdigits equaling a square must
have repdigits of 5 or fewer digits. There are only 45 repdigits of
2, 3, 4 or 5 digits, so that leaves us $45\cdot46/2 = 1035$ cases to check.
This is easily done and we find the complete list of solutions as in the
statement.
\end{proof}
\section{Further Observations and Questions}
\label{further}
The proof of Theorem 2 relies entirely on the fact that there is a
convenient string of consecutive quadratic non-residues modulo $10^6$
(due mostly to the fact that $10$ is even.) Finding strings of
consecutive quadratic residues and non-residues has been studied. (See
\cite{BH} and \cite{Wright}.) Certainly base-10 is special, but there
are some observations in other bases:
We noted above that there is a family of solutions $22+99 = 33+ 88 =
44+77 = 55+66 = 11^2$ in base 10. In an arbitrary base $c \geq 2$,
we have $(c+1)^2 = (c-k)(c+1) + (k+1)(c+1)$ for $1 \leq k \leq c-1$.
This shows that there is a similar family regardless of the choice of base.
Similarly, $(c+2)^2 = c^2 + 4c + 4 = 1 (c^2+c+1) + 3 (c+1)$ shows that
$111+33 = $ is a square in every base. Also $(2c^2 + c + 1)^2 =
4c^4 + 4c^3 + 5c^2 + 2c + 1 = 4(c^4+c^3+c^2+c+1) + (c-3)(c+1)$, so
$44444+77$ is just the base 10 manifestation of this identiy.
In the more special case that $c= m^2 - 1$, we have the identity
$c^5 + c^4 + 4c^3 + 4c^2 + 4c + 4 = (c+1)(c^2+2)^2 = m^2 (c^2+2)^2$,
which is to say $(111111)_{c} + (3333)_{c}$ is a square.
The most interesting solutions to our equation seem to come in groups.
$111+33=12^2$ and $1111 + 333 = 38^2$ seem related, as do $4+77 = 9^2$,
$44+77=11^2$ and $44444+77 = 221^2$. Also the infinite number of solutions
$9_{2m}+1 = (10^m)^2.$ (Here we're allowing 1-digit solutions.) As
far as we know, this is accidental, but perhaps there's a useful
reason behind this. If so, maybe there's a solution to the problem
that doesn't rely on SAGE computing on elliptic curves(?)
As mentioned earlier, the case of base 10 is easier than some others.
Of particular note is base 7, where there are a number of examples,
the most striking of which is
\[
48060^2 = 2309763600 = (11111111111)_{7} + (3333333)_{7}.
\]
It seems difficult to use modular arithmetic to put a bound on $m$ and $n$
subject to
\[
a \cdot \frac{7^{m} - 1}{6} + b \cdot \frac{7^{n} - 1}{6} = x^{2}.
\]
The solutions to this equation are a subset of those to
\[
A u^{5} + B v^{5} + Cx^{5} = 6w^{2} x^{3}.
\]
This defines a (possibly singular) surface of general type, and it is
conjectured by Bombieri and Lang that the \emph{rational} points
on a surface of general type lie on a finite union of curves. This supports
our suspicion that there should only be finitely many solutions, but a proof
eludes us.
\pagebreak
\section{Appendix: Tables}
{\bf Table A: Quadratic residues modulo $10^n$.} In the following table, generated
with Maple, the entry is O if $-(a+b)$ is a quadratic residue modulo $10^k$ and X otherwise.
\vspace*{.1in}
\begin{tabular}[t]{|c|c|c|c|c|c|}\hline
$-(a+b)$ & $10^2$ & $10^3$ & $10^4$ & $10^5$ & $10^6$ \invs \\ \hline\hline
$-2$ & X & X & X & X & X \invs \\ \hline
$-3$ & X & X & X & X & X \invs \\ \hline
$-4$ & O & O & X & X & X \invs \\ \hline
$-5$ & X & X & X & X & X \invs \\ \hline
$-6$ & X & X & X & X & X \invs \\ \hline
$-7$ & X & X & X & X & X \invs \\ \hline
$-8$ & X & X & X & X & X \invs \\ \hline
$-9$ & X & X & X & X & X \invs \\ \hline
$-10$ & X & X & X & X & X \invs \\ \hline
$-11$ & O & X & X & X & X \invs \\ \hline
$-12$ & X & X & X & X & X \invs \\ \hline
$-13$ & X & X & X & X & X \invs \\ \hline
$-14$ & X & X & X & X & X \invs \\ \hline
$-15$ & X & X & X & X & X \invs \\ \hline
$-16$ & O & O & O & O & X \invs \\ \hline
$-17$ & X & X & X & X & X \invs \\ \hline
$-18$ & X & X & X & X & X \invs \\ \hline
\end{tabular}
\pagebreak
{\bf Table B: Summary of calculations.}
\vspace*{.1in}
\begin{tabular}[t]{|c|c|c|c|c|c|}\hline
Case & $r$ & $x$ & $y$ & $N$ & $x$-coords \invs \\ \hline\hline
$8_m+33$ & 0 & $2\cdot10^l$ & $3k$ & $289$ & $-4, 0, 68 $ \invs \\ \hline
$8_m+33$ & 1 & $2\cdot10^{l+1}$ & $30k$ & $28900$ & $ 0 $ \invs \\ \hline
$8_m+33$ & 2 & $2\cdot10^{l+2}$ & $300k$ & $2890000$ & $ -136, 0, \mathbf{200}, 425 $ \invs \\ \hline
$7_m+99999$ & 0 & $7\cdot10^l$ & $21k$ & $44099216$ & $10577$ \invs \\ \hline
$7_m+99999$ & 1 & $7\cdot10^{l+1} $ & $210k$ & $4409921600$ & $-1064$ \invs \\ \hline
$7_m+99999$ & 2 & $7\cdot10^{l+2} $ & $2100k$ & $440992160000$ & $-5936, -5900, -5516, 2800,$ \invs \\
& & & & & $20825, 21056, 721364000$ \invs \\ \hline
$6_m+55$ & 0 & $6\cdot10^l$ & $18k$ & $17604$ & $-20, -12, 816$ \invs \\ \hline
$6_m+55$ & 1 & $6\cdot10^{l+1} $ & $180k$ & $1760400$ & $-120, 24, 160, 14640$ \invs \\ \hline
$6_m+55$ & 2 & $6\cdot10^{l+2} $ & $1800k$ & $176040000$ & $\mathbf{600}$ \invs \\ \hline
$4_m+77$ & 0 & $4\cdot10^l$ & $12k$ & $11024$ & 1 \invs \\ \hline
$4_m+77$ & 1 & $4\cdot10^{l+1}$ & $120k$ & $1102400$ & $-100, -95, -16, \mathbf{40}, 160,$ \invs \\
& & & & & $584, 1420, 26764 $ \invs \\ \hline
$4_m+77$ & 2 & $4\cdot10^{l+2}$ & $1200k$ & $110240000$ & $-464, -400, 64, \mathbf{400},$ \invs \\
& & & & & $425, 625, 1076, $ \invs \\
& & & & & $\mathbf{4000}, 1154800$ \invs \\ \hline
$3_m+111$ & 0 & $3\cdot10^l$ & $9k$ & $8964$ & $21$\invs \\ \hline
$3_m+111$ & 1 & $3\cdot10^{l+1}$ & $90k$ & $896400$ & $-96, -80, -15, 25, $ \invs \\
& & & & & $40, 49, 120, 256, $ \invs \\
& & & & & $280, 1200, 16576$ \invs \\ \hline
$3_m+111$ & 2 & $3\cdot10^{l+2}$ & $900k$ & $89640000$ & $-375, 124, \mathbf{300},$ \invs \\
& & & & & $700, 5241 $ \invs \\ \hline
$2_m+99$ & 0 & $2\cdot10^l$ & $6k$ & $3556$ & none \invs \\ \hline
$2_m+99$ & 1 & $2\cdot10^{l+1}$ & $60k$ & $355600$ & $ -55, -40, 144, 4320$ \invs \\ \hline
$2_m+99$ & 2 & $2\cdot10^{l+2}$ & $600k$ & $35560000$ & $-216, \mathbf{200}, 704, 800,$ \invs \\
& & & & & $1500, 1800, 7400$ \invs \\
& & & & & $199200$ \invs \\ \hline
$2_m+22$ & 0 & $2\cdot10^l$ & $6k$ & $784$ & $-7, 0, 8, 56$ \invs \\ \hline
$2_m+22$ & 1 & $2\cdot10^{l+1}$ & $60k$ & $78400$ & $-40, 0, 56, 140, 480$ \invs \\ \hline
$2_m+22$ & 2 & $2\cdot10^{l+2}$ & $600k$ & $7840000$ & $-175, 0, 224, 800$ \invs \\ \hline
\end{tabular}
\pagebreak
|
\section{Introduction}
\label{s_intro}
In 1958, Bohr, Mottelson, and Pines~\cite{bmp58} suggested a possible analogy
between the excitation spectra of nuclei and those of the superconducting metallic state.
Since then, a wealth of experimental data have been accumulated,
supporting the important role played by pairing correlations
in defining properties of atomic nuclei,
such as deformation, moments of inertia, alignments, etc.~\cite{Brink05,BCS50}.
Today, the study of pairing correlations
continues to be a subject of active research in nuclear physics,
with an emphasis in exotic nuclei.
Of particular interest is the understanding of the role
played by the isoscalar $(T=0)$ and isovector $(T=1)$ pairing forces~\cite{nprev}
in the structure of $N\approx Z$ nuclei.
Given the charge independence of the nuclear force,
$T=1$ pairing is on an equal footing
between the $T_z=0$ neutron-proton (np)
and $|T_z|=1$ neutron-neutron and proton-proton (nn and pp) components.
In addition, we have the unique possibility
of studying the formation of a condensate of $T=0$ np pairs,
thus implying the possible co-existence of ``Cooper" pairs of isoscalar and isovector type.
Although the nuclear force is stronger in the $T=0$ channel,
it is still not clear how effective the ({\it in-medium}) $T=0$ correlations are
in giving rise to a ground-state isoscalar condensate~\cite{nprev}.
In this paper we consider some interesting properties of the isoscalar condensate in the $jj$ coupling scheme,
in particular with regards to its angular momentum.
Our motivation starts by studying the numerical results of a shell-model calculation,
within the space of single-particle spin-orbit partners,
showing that when the isoscalar component is dominant,
the ground-state is not paired to $J^\pi=0^+$
but, rather, it behaves as a state of aligned $1^+$ {\it quasi-deuterons}.
To gain further insight into the peculiar structure of these condensates,
we develop a boson mapping of the shell model,
leading to an approximate analytic solution.
Group-theoretical solutions of the pairing problem
are known in the isoscalar and isovector limits of $LS$ coupling~\cite{Flowers64,Pang69,Evans81}
and in the isovector limit of $jj$ coupling~\cite{Racah52,Kerman61}
but, to our knowledge, not in the isoscalar limit of $jj$ coupling.
Our results therefore provide, for the first time, approximate analytic formulae
for the energies of the lowest states in that case.
While we are of course aware that this limit is not applicable to real nuclei,
these states might exist close to the ground state
in specific regions of the $N=Z$ line~\cite{Gezerlis11,Bulthuis16}
and, more interestingly perhaps, could be realized in atomic traps.
\section{Single-$j$ shell model}
\label{s_sm}
Single-shell models that capture the main ingredients of the problem
provide a useful framework to understand
the competition of isovector and isoscalar pairing interactions.
Here we start by considering the scattering of $(L=0,S=0,T=1)$ nn, np, and pp pairs
as well as $(L=0,S=1,T=0)$ np pairs,
describing the large spatial overlap of the nucleons' wave function in a relative $L=0$ state.
In the $jj$ coupling scheme the spin-orbit splitting $v_{\ell s}$
increases the energy required to form the ($L=0,S=1,T=0)$ state,
thus favoring a ($J=1, T=0$) {\it quasi-deuteron} configuration.
It then seems of interest to consider a more realistic case,
namely that of a single-$j$ shell that incorporates the $jj$ coupling scheme,
more appropriate in heavier nuclei.
The difference between these simple $LS$ and $jj$ models
has been discussed in terms of the BCS approximation~\cite{Bes00}.
Our approach to study this problem
is to use the shell-model code OXBASH~\cite{Brown88}
with an effective two-body force of the form
\begin{equation}
\hat V(g,x)=-xg\hat V_{J=0,T=1}-(1-x)g\hat V_{J=1,T=0},
\label{e_interaction}
\end{equation}
with
\begin{equation}
\hat V_{J,T}=
{\frac 1 2}(a^\dag_{jt}\times a^\dag_{jt})^{(J,T)}\cdot(\tilde a_{jt}\times\tilde a_{jt})^{(J,T)},
\label{e_defint}
\end{equation}
where $a^\dag_{jm_jtm_t}$ creates a nucleon
with angular momentum $j$ and projection $m_j$,
isospin $t=\frac12$ and projection $m_t$,
and with $\tilde a_{jm_jtm_t}=(-)^{j+m_j+t+m_t}a_{j-m_jt-m_t}$.
The notation $\times$ implies the coupling to angular momentum $J$ and isospin $T$,
and the dot $\cdot$ denotes a scalar product in angular momentum and isospin.
The Hamiltonian~(\ref{e_interaction}) models the mixture of the two types of competing pairing interactions by the parameter $x$,
with $x=0$ corresponding to the isoscalar and $x=1$ to the isovector limits respectively.
The sign convention in Eq.~(\ref{e_interaction}) is such
that $g$ is positive for an attractive interaction $\hat V(g,x)$.
We consider two spin-orbit partners $f_{7/2}$ and $f_{5/2}$
and study the low-lying spectra obtained as a function of the splitting $v_{\ell s}$.
In the limit $v_{\ell s}=0$ we recover the results of the $LS$ coupling scheme.
The results for $v_{\ell s} \gg\langle\hat V\rangle$
agree with those obtained for a single $f_{7/2}$ level only.
The intriguing phenomenon that motivated this study is seen in Fig.~\ref{f_f7n4},
showing the evolution of the two lowest states
in the $N=4$ particle system as a function of $x$.
For an appreciable amount of isoscalar pairing ($x\lesssim0.4$)
the ground state changes from the expected $0^+$ to a $2^+$ state.
Moreover, as seen in Fig.~\ref{f_f7n6},
the ground state for $N=6$ is a $3^+$ and not $1^+$,
and so on for more particles.
Considering that, for two particles interacting with the force~(\ref{e_interaction}),
$x\sim0$ favors deuteron-like pairing with angular momentum $J=1$,
it appears that the ground state of the many-particle system
prefers the aligned configuration of the $n=N/2$ pairs,
{\it i.e.} the configuration with $J=n$.
\begin{figure}
\includegraphics[width=0.4\textwidth, angle=90]{fig1.pdf}
\caption{Energies (in units of the pairing strength $g$) of the lowest two $T=0$ states
for $N=4$ particles in an $f_{7/2}$ shell
as a function of the relative mixture $x$ of isovector and isoscalar pairing.}
\label{f_f7n4}
\end{figure}
\begin{figure}
\includegraphics[width=0.4\textwidth, angle=90]{fig2.pdf}
\caption{Energies (in units of the pairing strength $g$) of the lowest two $T=0$ states
for $N=6$ particles in the $f_{7/2}$ shell
as a function of the relative mixture $x$ of isovector and isoscalar pairing.}
\label{f_f7n6}
\end{figure}
We can trace back the change in the properties of the ground state to the spin-orbit splitting.
In Fig.~\ref{f_fn24} we show the results for the $N=4$ system and a pure isoscalar force.
The energies of the $0^+$ and $2^+$ states
are plotted as a function of the spin-orbit splitting $v_{\ell s}$.
The two states cross, with a $0^+$ ground state in $LS$ coupling,
which becomes a $2^+$ in $jj$ coupling, as we saw above.
To obtain an estimate of the critical value $v_{\ell s}^*$ at which the switch occurs,
we consider the case of the $N=2$ system,
also shown in Fig.~\ref{f_fn24}.
Taking the limit of large $j$, to simplify the $LS$-$jj$ re-coupling coefficients,
we have in $jj$ coupling $E_{jj}(1^+)=-g$
and in $LS$ coupling $E_{LS}(1^+)\approx-6g$.
The $^3S_1$ state can be written in terms of the $jj$-coupled wave functions as~\cite{Talmi93}
\begin{equation}
|^3S_1\rangle\approx
\frac{1}{\sqrt{6}} |j^2_>\rangle+
\frac{2}{\sqrt{6}} |j_>j_<\rangle-
\frac{1}{\sqrt{6}} |j^2_<\rangle,
\end{equation}
from which we can treat perturbatively the effect of the spin-orbit splitting $v_{\ell s}$.
This gives for an intermediate coupling
\begin{equation}
E_{\rm IC}(1^+)\approx E_{LS}(1^+)+\frac{1}{6} 2 v_{\ell s}+\frac{4}{6} v_{\ell s}= E_{LS}(1^+)+ v_{\ell s}.
\end{equation}
The critical value is obtained when the energy above
equals that of the $jj$-coupling limit (dashed lines in Fig.~\ref{f_fn24})
\begin{equation}
E_{jj}(1^+)=E_{LS}(1^+)+ v_{\ell s}^*
\end{equation}
and we find (in the large-$j$ limit)
\begin{equation}
\frac{v_{\ell s}^*}{g}\approx5.
\end{equation}
For the particular case of the $f_{7/2}$-$f_{5/2}$ pair (finite-$j$) we find a value of $\sim3.5$,
in agreement with the estimate shown in Fig.~\ref{f_fn24} (shaded area).
\begin{figure}
\includegraphics[width=0.4\textwidth,angle=90]{fig3.pdf}
\caption{Energies (in units of the pairing strength $g$) of the $T=0$ ground state for $N=2$ particles (top)
and of the lowest two $T=0$ states for $N=4$ particles (bottom)
as a function of the spin-orbit splitting between the $f_{7/2}$ and $f_{5/2}$ orbits
and for pure isoscalar pairing.
The shaded area in the top panel indicates the critical value of the spin-orbit splitting, $v_{\ell s}^*$,
at which the $jj$ coupling takes on. (See text for details).}
\label{f_fn24}
\end{figure}
To shed further light on the properties of the isoscalar condensate discussed above,
we develop in the next section a description based on a mapping
to interacting $p$ bosons of angular momentum $J=1$ and isospin $T=0$.
Based on the results above, and on the fermionic nature of the problem,
we anticipate that the residual interaction between these bosons
favors their aligned coupling.
\section{Isoscalar pairing between fermions in a single $j$ shell}
\label{s_jshell}
Consider a system of $N$ particles, $n$ neutrons and $n$ protons, in a single-$j$ shell,
interacting through an isoscalar pairing interaction
with angular momentum $J=1$ and isospin $T=0$,
corresponding to the $x=0$ limit of Eq.~(\ref{e_interaction}),
$\hat V(g,x=0)=-g\hat V_{10}$.
As discussed in the previous section,
a possible strategy for simplifying the problem
starts from the observation that, by definition of the interaction,
the dominant pair in the two-particle system has $J=1$ and $T=0$.
We attempt to represent a subset of the $2n$-particle eigenstates of this interaction,
including hopefully those at lowest energies,
in terms of a single state $|P\rangle\equiv P^\dag|{\rm o}\rangle$
(with $|{\rm o}\rangle$ the vacuum),
which has $J=1$ and $T=0$,
\begin{equation}
P^\dag_{M_J}\equiv(a^\dag_{jt}\times a^\dag_{jt})^{(J=1,T=0)}_{M_J,M_T=0}.
\label{e_ppair}
\end{equation}
The natural framework to test this idea
is provided by the nucleon-pair shell model (NPSM),
which assumes a basis constructed from nucleon pairs~\cite{Chen93,Chen97,Zhao13un,Fu13}.
In this approximation the full $T=0$ shell-model space
is truncated to one constructed out of $P$ pairs
with basis states $|P^nJ_2\dots J_{n-1}J\rangle$
that are proportional to
\begin{equation}
\left(\cdots\left(\left(P^\dag\times P^\dag\right)^{(J_2)}\times P^\dag\right)^{(J_3)}\times\cdots\times
P^\dag\right)^{(J)}|{\rm o}\rangle.
\label{e_pbasis}
\end{equation}
This $2n$-particle state is characterized
by the set of intermediate angular momenta $\{J_2,\dots,J_{n-1}\}$,
with $J_1=1$ and $J_n=J$, the total angular momentum of the state.
All pairs have $T=0$
and the coupling in isospin need not be considered.
In principle, several intermediate couplings $\{J_2,\dots,J_{n-1}\}$
are possible for a given total angular momentum $J$.
Such is the case for arbitrary pairs but not for $P$ pairs
since the number of independent states
with angular momentum $J$ constructed out of $n$ $P$ pairs
cannot exceed the corresponding number constructed out of $n$ $p$ bosons,
which is 1 if $n-J$ is non-negative and even, and 0 otherwise.
We conclude therefore that, for a given $J$,
at most one state $|P^nJ_2\dots J_{n-1}J\rangle$ exists,
for which the intermediate angular momenta can be chosen as
\begin{equation}
\left\{\begin{array}{lll}
J_i= i\bmod 2,&&1\leq i\leq n-J,\\
J_i=i-n+J,&&n-J\leq i\leq n,\\
\end{array}\right.
\label{e_icoupling}
\end{equation}
where it is implicitly assumed (as will be from now on)
that $n-J$ is non-negative even.
We denote normalized states as $|P^nJ\rangle$,
tacitly assuming the intermediate coupling convention~(\ref{e_icoupling}).
In this convention the paired and spin-aligned states of particular interest here
correspond to the choice
\begin{equation}
\begin{array}{lllll}
\mbox{\rm paired}&:&J_i= i\bmod 2,&&1\leq i\leq n,\\
\mbox{\rm spin-aligned}&:&J_i=i,&&1\leq i\leq n.\\
\end{array}
\label{e_paiali}
\end{equation}
As long as $n\leq(2j+1)/2$ all states~(\ref{e_icoupling}) exist.
This is no longer necessarily true if the shell is more than half filled,
in which case it is advantageous to reconsider the problem in terms of holes.
We then construct basis states $|\tilde P^{2j+1-n}J_2\dots J_{n-1}J\rangle$
that are proportional to
\begin{equation}
\left(\cdots\left(\left(\tilde P\times\tilde P\right)^{(J_2)}\times\tilde P\right)^{(J_3)}\times\cdots\times
\tilde P\right)^{(J)}|\tilde{\rm o}\rangle,
\label{e_hbasis}
\end{equation}
where $|\tilde{\rm o}\rangle$ represents a full shell
and $\tilde P$ annihilates a $P$ pair,
\begin{equation}
\tilde P_{M_J}\equiv(\tilde a_{jt}\times\tilde a_{jt})^{(J=1,T=0)}_{M_J,M_T=0}.
\label{e_hpair}
\end{equation}
The angular momentum and anti-symmetry considerations
concerning the states~(\ref{e_pbasis}) and~(\ref{e_hbasis}) are the same,
and consequently the latter lead to the same allowed basis states~(\ref{e_icoupling})
with $n$ replaced by $\bar n\equiv2j+1-n$.
We denote such states as $|\tilde P^{\bar n}J\rangle$.
In general, $|P^nJ\rangle$ and $|\tilde P^{\bar n}J\rangle$
are {\em not} the same state,
\begin{equation}
|P^nJ\rangle\neq|\tilde P^{\bar n}J\rangle,
\label{e_phuneq}
\end{equation}
and it is possible
that the state on the left-hand side exists
while the one on the right-hand side does not (or {\it vice versa}).
Only if the shell-model state with a given $J$ and $T=0$ is unique,
do the particle and hole representations become equivalent,
as is the case, for example, for the states
\begin{equation}
|P^{2j+1}J=0\rangle=|\tilde {\rm o}\rangle,
\quad
|P^{2j}J=1\rangle=|\tilde PJ=1\rangle.
\label{e_pheq}
\end{equation}
The choice $|P^nJ\rangle$ if $n\leq(2j+1)/2$
and $|\tilde P^{\bar n}J\rangle$ if $n\geq(2j+1)/2$,
apart from being computationally simpler,
gives the best approximation of shell-model states in terms of $P$ pairs.
The summary of the above discussion is that
the truncated shell-model basis constructed out of $P$ pairs
is spanned by the states $|P^nJ\rangle$ if $n\leq(2j+1)/2$
and by the states $|\tilde P^{\bar n}J\rangle$ if $n\geq(2j+1)/2$.
These basis states exist
(provided $n-J$ or $\bar n-J$ is non-negative even)
and are unique for a given $n$ and $J$,
so that no additional labels are needed.
Therefore, in the $P$-pair approximation of the NPSM,
the correlation energy due to isoscalar pairing
in the state with $n$ neutrons and $n$ protons,
coupled to total angular momentum $J$ and isospin $T=0$,
is\footnote{We reserve the notation $E_{\rm f}(n,J)$ and $E_{\rm f}(\bar n,J)$
({\it i.e.}, expressions without tilde)
for the exact correlation energy of the yrast state with angular momentum $J$
calculated in the full shell-model space.}
\begin{equation}
\tilde E_{\rm f}(n,J)\equiv
\langle P^nJ|-g\hat V_{10}|P^nJ\rangle,
\label{e_etfp}
\end{equation}
for $n\leq(2j+1)/2$, and by
\begin{equation}
\tilde E_{\rm f}(\bar n,J)\equiv
\langle\tilde P^{\bar n}J|-g\hat V_{10}|\tilde P^{\bar n}J\rangle,
\label{e_etfh}
\end{equation}
for $n\geq(2j+1)/2$.
The computation of the matrix elements of an arbitrary interaction
between nucleon-pair states
is possible with the recurrence relation devised by Chen~\cite{Chen97}.
In the general formulation of the NPSM
care should be taken of the over-completeness and non-orthogonality of the pair basis.
This is not an issue in the present application
since basis states are unique for a given $n$ and $J$.
It should be stressed that Eqs.~(\ref{e_etfp}) and~(\ref{e_etfh})
yield an approximation to the exact isoscalar-pairing correlation energy.
The energy $E_{\rm f}(n,J)$ of a particle state
is calculated with respect to the vacuum $|{\rm o}\rangle$
while that of a hole state, $E_{\rm f}(\bar n,J)$,
is with respect to the full shell $|\tilde {\rm o}\rangle$.
The particle-hole transformation gives a relation between both quantities,
which is exact in the full shell-model space.
For our particular case of isoscalar pairing this relation is
\begin{equation}
E_{\rm f}(n,J)=
-\frac{3(2j+1-2\bar n)}{2j+1}g+E_{\rm f}(\bar n,J).
\label{e_eph}
\end{equation}
We use the same equation
to relate the approximate energies $\tilde E_{\rm f}(n,J)$ and $\tilde E_{\rm f}(\bar n,J)$.
In the following absolute energies
are quoted with respect to the vacuum $|{\rm o}\rangle$.
For a particle state they are obtained directly
while for a hole state they follow from Eq.~(\ref{e_eph}).
A further approximation is to replace
the $P$ pairs by $p$ bosons,
with single-boson energies and boson-boson interactions
derived from the two-particle and four-particle systems, respectively.
With use of the OAI mapping~\cite{Otsuka78}
a $p$-boson Hamiltonian $\hat H_{\rm b}$ is obtained,
which can be written as
\begin{equation}
\hat H_{\rm b}=
\epsilon_p\,p^\dag\cdot\tilde p+
{\frac 1 2}\sum_{\lambda=0,2}v^{\rm b}_\lambda
(p^\dag\times p^\dag)^{(\lambda)}\cdot(\tilde p\times\tilde p)^{(\lambda)},
\label{e_hamb}
\end{equation}
where $\epsilon_p$ is the $p$-boson energy
and $v^{\rm b}_\lambda$ are the two-body interaction matrix elements
{\em between the $p$ bosons}.
The definition of the adjoint operator $\tilde p_m\equiv(-)^{1-m}p_{-m}$
ensures that $\tilde p_m$ is an annihilation operator
with transformation properties under rotations that are the same
as those for the creation operator $p_m^\dag$~\cite{Iachello06}.
With the above definitions we have
that $p^\dag\cdot\tilde p=\sum_m p_m^\dag p_m$
is the number operator $\hat n_p$.
The single-boson energy is
\begin{equation}
\epsilon_p\equiv
\langle p|\hat H_{\rm b}|p\rangle\doteq
\langle P|-g\hat V_{10}|P\rangle=-g.
\label{e_ebos}
\end{equation}
where the notation $\doteq$ is used to indicate
that the equality holds by virtue of the mapping procedure.
The two-body boson matrix elements with $\lambda=0,2$ are
\begin{align}
v^{\rm b}_\lambda&\equiv
\langle p^2\lambda|\hat H_{\rm b}|p^2\lambda\rangle-2\epsilon_p
\nonumber\\&\doteq
-g\left(\langle j^4[10,10]\lambda0
|\hat V_{10}|
j^4[10,10]\lambda0\rangle-2\right),
\label{e_hmat4}
\end{align}
where the bra and ket represent normalized, anti-symmetric two-pair states,
\begin{equation}
|j^4[J_1T_1,J_2T_2]JT\rangle
\propto{\cal A}|j^2(J_1T_1)j^2(J_2T_2)JT\rangle.
\label{e_ket4}
\end{equation}
The notation in square brackets $[J_1T_1,J_2T_2]$ implies that the state~(\ref{e_ket4})
is constructed from a parent state
with intermediate angular momenta and isospins $J_1T_1$ and $J_2T_2$.
The anti-symmetrized states $|j^4[J_1T_1,J_2T_2]JT\rangle$
can be expanded in terms of the two-pair states $|j^2(J_1T_1)j^2(J_2T_2)JT\rangle$
by means of four-to-two-particle coefficients of fractional parentage (CFPs)~\cite{Talmi93},
\begin{equation}
[j^2(J_aT_a)j^2(J_bT_b)JT|\}j^4[J_1T_1,J_2T_2]JT],
\label{e_cfp}
\end{equation}
which are known in closed form.
From the general expression for the matrix element~(\ref{e_hmat4})
the following results are obtained:
\begin{align}
v^{\rm b}_0/g={}&
-6[j^2(10)j^2(10)00|\}j^4[10,10]00]^2+2,
\nonumber\\
v^{\rm b}_2/g={}&
-6[j^2(10)j^2(10)20|\}j^4[10,10]20]^2
\nonumber\\&-
6[j^2(30)j^2(10)20|\}j^4[10,10]20]^2+2,
\end{align}
which, with the help of
\begin{align}
&[j^2(10)j^2(10)00|\}j^4[10,10]00]^2=
\frac{2j^3-2j+3}{3j(j+1)(2j+1)},
\nonumber\\
&[j^2(10)j^2(10)20|\}j^4[10,10]20]^2=
\frac{10j^3+9j^2-j-3}{15j(j+1)(2j+1)},
\nonumber\\
&[j^2(30)j^2(10)20|\}j^4[10,10]20]^2
\nonumber\\&\qquad=
\frac{9(j-1)(j+2)(2j+3)}{10j(j+1)(2j+1)(5j^2+7j+3)},
\end{align}
lead to the following expressions for the $p$-boson matrix elements:
\begin{align}
v^{\rm b}_0={}&
\frac{6(j^2+j-1)}{j(j+1)(2j+1)}g
\stackrel{j\rightarrow\infty}{\longrightarrow}
\left[\frac{3}{j}+
{\cal O}\left(\frac{1}{j^2}\right)\right]g,
\nonumber\\
v^{\rm b}_2={}&
\frac{3(4j^4+6j^3+j^2+7j+12)}{j(j+1)(2j+1)(5j^2+7j+3)}g
\nonumber\\&
\stackrel{j\rightarrow\infty}{\longrightarrow}
\left[\frac{6}{5j}+
{\cal O}\left(\frac{1}{j^2}\right)\right]g.
\label{e_vbos}
\end{align}
As anticipated, for an attractive isoscalar pairing interaction
the boson-boson matrix elements are repulsive.
This is a finite-space effect, due to the Pauli principle,
since the matrix elements vanish in the large-$j$ limit.
A difference between the $\lambda=0$ and $\lambda=2$ matrix elements
also arises due to Pauli effects,
and it is seen that $v^{\rm b}_2$ is less repulsive.
This favors the spin-aligned ground state,
not only for two but also for more bosons
as a result of the following argument.
Since a system of $n$ interacting identical $p$ bosons is solvable
by virtue of a ${\rm U}(3)\supset{\rm SO}(3)$ dynamical symmetry~\cite{Iachello06},
the eigenvalues of the Hamiltonian~(\ref{e_hamb})
are known in closed form,
\begin{align}
E_{\rm b}(n,J)={}&
n\epsilon_p+
\frac{n(n+1)-J(J+1)}{6}v^{\rm b}_0
\nonumber\\&+
\frac{2n(n-2)+J(J+1)}{6}v^{\rm b}_2,
\label{e_enebos}
\end{align}
where the allowed angular momenta are $J=n,n-2,\dots,1$ or 0.
The only possible ground states of a $p$-boson system
are either paired or spin-aligned~\cite{Law98,Isacker07}.
The paired state has $J=0$ or $J=1$ with energies
\begin{align}
E_{\rm b}(n,J=0)&=
n\epsilon_p+
\frac{n(n+1)}{6}v^{\rm b}_0+
\frac{n(n-2)}{3}v^{\rm b}_2,
\label{e_enepaired}\\
E_{\rm b}(n,J=1)&=
n\epsilon_p+
\frac{(n-1)(n+2)}{6}v^{\rm b}_0+
\frac{(n-1)^2}{3}v^{\rm b}_2,
\nonumber
\end{align}
depending on whether $n$ is even or odd, respectively.
The spin-aligned state has $J=n$ with energy
\begin{equation}
E_{\rm b}(n,J=n)=
n\epsilon_p+
\frac{n(n-1)}{2}v^{\rm b}_2.
\label{e_enealigned}
\end{equation}
The breaking of the rotational invariance in {\it gauge space}~\cite{broglia}
leads to the emergence of isoscalar pairing rotational bands,
as seen in the quadratic dependence of the energies
as a function of the number of pairs $n$, Eqs.~(\ref{e_enepaired}) and (\ref{e_enealigned}).
The difference in energy
between the paired and the spin-aligned states
can be written as
\begin{align}
\Delta_{\rm b}(n)&=
\frac{(n-n_2)(n+1+n_2)}{6}(v^{\rm b}_0-v^{\rm b}_2)
\nonumber\\&\approx
g\frac{3(n-n_2)(n+1+n_2)}{10j},
\end{align}
where $n_2$ is 0 for even $n$ and 1 for odd $n$,
$n_2\equiv n \bmod 2$.
This shows that for all $n$ the difference in energy
between the paired and the spin-aligned states
is positive for an attractive pairing interaction,
that is, the spin-aligned configuration is the ground state.
\begin{table}
\centering
\caption{\label{t_shellj}
Exact energies $E_{\rm f}(n,J)$ of paired ($J=0$ or 1) and aligned ($J=n$) states with $T=0$
of a system of $n$ neutrons and $n$ protons in a single-$j$ shell
interacting through an isoscalar pairing force,
in units of the strength $g$,
and the corresponding energies $\tilde E_{\rm f}(n,J)$ and $E_{\rm b}(n,J)$
obtained in the $P$-pair and $p$-boson approximations.
A dash --- means that a $P$-pair state does not exist
while the absence of an entry indicates
that the numerical result could not be obtained.}
\smallskip
\begin{tabular}{ccccccccccc}
\hline\hline
$E(n,J)$&&$j=7/2$&~&$j=9/2$&~&$j=11/2$&~&$j=13/2$&~&$j=15/2$\\
\hline
$E_{\rm f}(2,0)$&&$-1.298$ &&$-1.424$ &&$-1.514$ &&$-1.580$ &&$-1.631$\\
$\tilde E_{\rm f}(2,0)$&&$-1.298$ &&$-1.424$ &&$-1.514$ &&$-1.580$ &&$-1.631$\\
$E_{\rm b}(2,0)$&&$-1.298$ &&$-1.424$ &&$-1.514$ &&$-1.580$ &&$-1.631$\\
$E_{\rm f}(2,2)$&&$-1.793$ &&$-1.825$ &&$-1.847$ &&$-1.865$ &&$-1.879$\\
$\tilde E_{\rm f}(2,2)$&&$-1.757$ &&$-1.799$ &&$-1.828$ &&$-1.850$ &&$-1.866$\\
$E_{\rm b}(2,2)$&&$-1.757$ &&$-1.799$ &&$-1.828$ &&$-1.850$ &&$-1.866$\\
\hline
$E_{\rm f}(3,1) $&&$-1.793$ &&$-1.953$ &&$-2.086$ &&$-2.192$ &&$-2.277$\\
$\tilde E_{\rm f}(3,1)$&&$-1.636$ &&$-1.848$ &&$-2.010$ &&$-2.135$ &&$-2.233$\\
$E_{\rm b}(3,1)$&&$-1.505$ &&$-1.772$ &&$-1.961$ &&$-2.100$ &&$-2.207$\\
$E_{\rm f}(3,3) $&&$-2.365$ &&$-2.466$ &&$-2.537$ &&$-2.591$ &&$-2.634$\\
$\tilde E_{\rm f}(3,3)$&&$-2.279$ &&$-2.403$ &&$-2.488$ &&$-2.552$ &&$-2.601$\\
$E_{\rm b}(3,3)$&&$-2.271$ &&$-2.397$ &&$-2.484$ &&$-2.549$ &&$-2.599$\\
\hline
$E_{\rm f}(4,0) $&&$-2.080$ &&$-2.251$ &&$-2.424$ &&$ $ &&$ $\\
$\tilde E_{\rm f}(4,0)$&&$-1.628$ &&$-1.887$ &&$-2.141$ &&$-2.353$ &&$-2.526$\\
$E_{\rm b}(4,0)$&&$-1.010$ &&$-1.545$ &&$-1.921$ &&$-2.200$ &&$-2.413$\\
$E_{\rm f}(4,4) $&&$-2.767$ &&$-2.925$ &&$ $ &&$ $ &&$ $\\
$\tilde E_{\rm f}(4,4)$&&$-2.577$ &&$-2.818$ &&$-2.985$ &&$-3.110$ &&$-3.207$\\
$E_{\rm b}(4,4)$&&$-2.541$ &&$-2.794$ &&$-2.968$ &&$-3.098$ &&$-3.198$\\
\hline
$E_{\rm f}(5,1) $&&$-2.543$ &&$ $ &&$ $ &&$ $ &&$ $\\
$\tilde E_{\rm f}(5,1)$&&$-2.386$ &&$-1.975$ &&$-2.284$ &&$-2.566$ &&$-2.806$\\
$E_{\rm b}(5,1)$&&$-2.255$ &&$-1.241$ &&$-1.815$ &&$-2.239$ &&$-2.564$\\
$E_{\rm f}(5,5) $&&--- &&$ $ &&$ $ &&$ $ &&$ $\\
$\tilde E_{\rm f}(5,5)$&&--- &&$-3.052$ &&$-3.324$ &&$-3.528$ &&$-3.687$\\
$E_{\rm b}(5,5)$&&--- &&$-2.990$ &&$-3.281$ &&$-3.496$ &&$-3.663$\\
\hline
$E_{\rm f}(6,0) $&&$-2.798$ &&$-2.851$ &&$ $ &&$ $ &&$ $\\
$\tilde E_{\rm f}(6,0)$&&$-2.798$ &&$-2.487$ &&$-2.206$ &&$-2.539$ &&$-2.847$\\
$E_{\rm b}(6,0)$&&$-2.798$ &&$-2.145$ &&$-1.222$ &&$-1.858$ &&$-2.347$\\
$E_{\rm f}(6,6) $&&--- &&--- &&$ $ &&$ $ &&$ $\\
$\tilde E_{\rm f}(6,6)$&&--- &&--- &&$-3.511$ &&$-3.810$ &&$-4.045$\\
$E_{\rm b}(6,6)$&&--- &&--- &&$-3.421$ &&$-3.744$ &&$-3.994$\\
\hline
$E_{\rm f}(7,1) $&&$-3.250$ &&$-3.153$ &&$ $ &&$ $ &&$ $\\
$\tilde E_{\rm f}(7,1)$&&$-3.250$ &&$-3.048$ &&$-2.784$ &&$-2.534$ &&$-2.892$\\
$E_{\rm b}(7,1)$&&$-3.250$ &&$-2.972$ &&$-2.315$ &&$-1.417$ &&$-2.074$\\
$E_{\rm f}(7,7) $&&--- &&--- &&--- &&$ $ &&$ $\\
$\tilde E_{\rm f}(7,7)$&&--- &&--- &&--- &&$-3.962$ &&$-4.284$\\
$E_{\rm b}(7,7)$&&--- &&--- &&--- &&$-3.842$ &&$-4.192$\\
\hline
$E_{\rm f}(8,0) $&&$-3.000$ &&$-3.224$ &&$-3.424$ &&$ $ &&$ $\\
$\tilde E_{\rm f}(8,0)$&&$-3.000$ &&$-3.224$ &&$-3.141$ &&$-2.968$ &&$-2.776$\\
$E_{\rm b}(8,0)$&&$-3.000$ &&$-3.224$ &&$-2.921$ &&$-2.287$ &&$-1.431$\\
$E_{\rm f}(8,8) $&&--- &&--- &&--- &&--- &&$ $\\
$\tilde E_{\rm f}(8,8)$&&--- &&--- &&--- &&--- &&$-4.408$\\
$E_{\rm b}(8,8)$&&--- &&--- &&--- &&--- &&$-4.256$\\
\hline\hline
\end{tabular}
\end{table}
We recall that the preceding results,
valid for an isoscalar pairing interaction in a single-$j$ shell,
are derived under the following simplifying assumptions:
\begin{enumerate}
\item
The full shell-model space is truncated
to one constructed out of $P$ pairs.
The expectation value of the isoscalar pairing Hamiltonian $-g\hat V_{10}$
in the (unique) $P$-pair state
takes fully account of the Pauli principle
and leads to the approximate correlation energy $\tilde E_{\rm f}(n,J)$.
\item
The fermionic Hilbert space constructed out of $P$ pairs
is mapped onto a corresponding bosonic Hilbert space constructed out of $p$ bosons.
The mapping of the Hamiltonian is carried out in the two- and four-nucleon spaces
and leads to a boson Hamiltonian with up to two-body interactions.
\item
The boson Hamiltonian is used
to calculate the energies $E_{\rm b}(n,J)$ of $n$-boson states.
\end{enumerate}
To gauge the adequacy of the different approximations,
we show in Table~\ref{t_shellj} the exact energies $E_{\rm f}(n,J)$
(wherever they can be calculated)
and the corresponding approximations $\tilde E_{\rm f}(n,J)$ and $E_{\rm b}(n,J)$
for $7/2\leq j\leq15/2$.
Several comments are in order.
First of all, we observe the identity
\begin{equation}
\tilde E_{\rm f}(n=2,J)=E_{\rm b}(n=2,J),
\label{e_n2ident}
\end{equation}
that is, the $P$-pair spectrum of the four-particle system
coincides with that obtained for two $p$ bosons.
This is a generic property of the mapping
and follows from the fact that up to two-body interactions
between the bosons are considered.
In fact, if up to $q$-body interactions are considered,
the identity~(\ref{e_n2ident}) remains valid up to the $n=q$.
Secondly, we observe the identity
\begin{equation}
E_{\rm f}(n=2,J=0)=\tilde E_{\rm f}(n=2,J=0).
\label{e_n2j0dent}
\end{equation}
This is not a generic property
but is valid for the isoscalar pairing interaction,
for which $|P^2J=0\rangle$ decouples from the rest of the shell-model space.
This property of the isoscalar pairing interaction
was already pointed out by Fu {\it et al.}~\cite{Fu14}
on the basis of analytic expressions for four-nucleon overlaps.
Furthermore, we observe from Table~\ref{t_shellj} the following hierarchy:
\begin{equation}
E_{\rm f}(n,J)\leq\tilde E_{\rm f}(n,J)\leq E_{\rm b}(n,J),
\label{e_inequal}
\end{equation}
valid for any $j$, $n$, and $J$.
The first inequality results from the fact that
the lowest eigenvalue of any Hamiltonian in a certain Hilbert space
is lower than the lowest eigenvalue of the same Hamiltonian in a truncated subspace.
We remark that an equality $E_{\rm f}(n=4,J)=\tilde E_{\rm f}(n=4,J)$ can be obtained
by constructing {\em effective} operators in the truncated space,
which is not done in the present application.
The second inequality in Eq.~(\ref{e_inequal})
is a consequence of performing the mapping in the four-particle systems
with an unnormalized ({\it i.e.}, not an effective) Hamiltonian.
For a variety of bosonic systems ($p$, $sd$, $sdp$, etc.)
we have consistently found that the boson Hamiltonian,
as it is derived here from the four-particle system,
gives an upper limit for the fermionic interaction energy of the $n$-particle system.
It is seen from Table~\ref{t_shellj} that the quality of the approximation
varies with $j$, $n$, and $J$.
Two effects are rather obvious:
the approximation becomes (i) better with increasing $j$
and (ii) worse with increasing $n$ [as long as $n\leq(2j+1)/2$].
These effects result from the increasing importance
of Pauli corrections that are neglected
({\it i.e.}, beyond two-body interactions between the bosons).
A more subtle effect is the dependence on $J$.
It is seen that the approximation for the aligned state $J=n$ is adequate,
even close to mid shell, $n\approx(2j+1)/2$, and for low $j$.
On the other hand, it is often rather poor for the paired state with $J=0$ or 1.
It can be conjectured that this is a generic property
of phonon approximations in fermionic systems:
while such descriptions are good for high-angular-momentum states,
they become highly anharmonic at low angular momenta.
Despite the varying quality of the boson approximation,
depending on $j$, $n$, and $J$,
the overall conclusion is that
the predicted feature of the lower energy of the aligned state
as compared to the paired state
is confirmed by the exact fermion calculation.
\section{Conclusion}
\label{s_conc}
We have considered some intriguing properties of a $T=0$ isoscalar condensate in single $j$-shell,
in particular with regards to its angular momentum coupling.
We developed a description based on a mapping of the shell model
to interacting $p$ bosons of angular momentum $J=1$ and isospin $T=0$,
providing for the first time approximate analytic formulae
for the energies of the lowest states.
Our results show that, due to the Pauli principle,
the residual interaction between these bosons favors
({\it a priori } unexpected) the aligned configuration of $n=N/2$ {\sl quasi-deuteron} pairs,
{\it i.e.} that with $J=n$.
While we realize this limit may not be applicable to real nuclei,
these states might exist close to the paired ground states
in specific regions close to the $N=Z$ line.
In fact, it was shown recently in Refs.~\cite{Gezerlis11,Bulthuis16},
using a phenomenological Hamiltonian within the framework of the Hartree-Fock-Bogoliubov theory,
that the spin-triplet phase is favored over the spin-singlet one in the mass region $ A\approx130$ with $Z \approx64$
(for example $^{132}$Gd).
This is found to depend on the occupation of specific low-$j$ orbitals
near the Fermi energy for which the spin-orbit splitting is small.
More interestingly perhaps,
it is envisioned that these condensates might be realized by tunable spin-orbit coupling in ultracold atomic traps~\cite{Lin11},
whereby the control parameter $x$ in Eq.~(\ref{e_interaction})
could be adjusted to drive the system from diamagnetic to magnetic.
A full extension of the present formalism
including the effect of the spin-orbit splitting
will be the subject of a future publication.
\section{Ackowledgements}
This work was supported in part by the FUSTIPEN
(French-U.S. Theory Institute for Physics with Exotic Nuclei)
under U.S. DOE grant No.~ DE-FG02-10ER41700, and by the U.S. DOE contract
No.~DE-AC02-05CH11231 (LBNL).
|
\section{Introduction}
The subject of this paper is the adaptive numerical approximation of second-order
semilinear elliptic problems of the form
\begin{equation}\label{eq:PDE}
\begin{aligned}
-\epsilon\Delta u+u &=f(\cdot,u) \text{ in } \ \Omega,\qquad
u=0 \text{ on } \partial \Omega.
\end{aligned}
\end{equation}
Here, $\Omega\subset\mathbb{R}^2$ is an open and bounded Lipschitz domain, $\epsilon\in(0,1]$ represents a (possibly small singular perturbation) parameter, $f:\,\overline\Omega\times\mathbb{R}\to\mathbb{R}$ is a continuously differentiable function, and~$u:\,\Omega\to\mathbb{R}$ is an unknown solution; in the sequel, we will omit to explicitly express the dependence of~$f$ on the first argument, and simply write~$f(u)$ instead. Problems of this type appear in a wide range of application areas of practical interest, such as, for example, nonlinear reaction-diffusion in ecology and chemical models~\cite{CantrellCosner:03,Edelstein-Keshet:05,BorisyukErmentroutFriedmanTerman:05,Ni:11,OkuboLevin:01}, economy~\cite{BarlesBurdeau:95}, or classical and quantum physics~\cite{BerestyckiLions:83,GibbonJamesMoroz:79,BaroneEspositoMageeScott:71,Strauss:77}.
Partial differential equations (PDEs) of the form~\eqref{eq:PDE} may admit
a unique solution, no solution at all, or more typically a multitude of solutions, or indeed
infinitely many such solutions. Moreover, in the singularly perturbed case, i.e., when
$0< \epsilon \ll 1$, solutions of \eqref{eq:PDE}, when they exist, may contain
sharp layers in the form of interior/boundary layers, or isolated spike--like solutions,
and their numerical approximation represents a challenging computational task.
Indeed, to efficiently and reliably compute discrete approximations to
the analytical solution~$u$ of~\eqref{eq:PDE},
it is essential to exploit {\em a posteriori} bounds which
not only provide information regarding the size of the discretisation error, measured in some
appropriate norm, but also yield local error indicators which may subsequently be employed
to enrich the underlying approximation space in an adaptive manner.
Of course, a key aspect of this general
solution procedure is the design and implementation of a nonlinear solver
which can efficiently compute the approximation $u_h$ to $u$; we shall return
to this issue below.
In general, the traditional approach exploited within the literature for the
design of adaptive finite element methods, for example, is to first discretise
the underlying PDE problem, in our case \eqref{eq:PDE},
and to derive an {\em a posteriori} error bound for the resulting (nonlinear) scheme;
this is typically a very mathematically challenging task. However,
once such a bound has been established, then given a suitable initial
mesh and polynomial approximation order, the underlying nonlinear system of discrete
equations arising from the underlying finite element discretisation may
be solved based on employing, for example, a (damped) Newton iteration.
Denoting this computed numerical approximation by $u_h$, the
size of the error between $u$ and $u_h$ may then
be estimated by exploiting this {\em a posteriori}
error bound. If this bound is below a given user tolerance, then sufficient accuracy has
been attained and the adaptive algorithm may be terminated. Otherwise, the
computational mesh ($h$--refinement) or the polynomial degree ($p$--refinement),
or both ($hp$--refinement) are locally enriched based on identifying regions in
the domain where the elementwise error indicators, which stem from the
{\em a posteriori} error bound, are locally large. On the basis of this new
finite element space, a new approximation $u_h$ to $u$ may be computed, and the whole
process repeated until either the desired accuracy has been attained, or a maximum
number of refinement steps have been completed.
Stimulated by the work undertaken in the recent article \cite{AmreinWihler:15},
we consider an alternative approach based on the
so-called adaptive Newton-Galerkin paradigm for the numerical
approximation of nonlinear problems of the type~\eqref{eq:PDE}.
More precisely, this general technique is based on applying local
Newton-type linearisations on the continuous level that allow for the
approximation of the semilinear PDE~\eqref{eq:PDE} by a sequence of linearised problems.
These resulting \emph{linear} PDEs are then discretised by means of an adaptive finite
element procedure, which, in turn, is based on a suitable \emph{a posteriori}
residual analysis. The adaptive Newton-Galerkin procedure provides an
\emph{interplay} between the (adaptive, or damped) Newton method and
the adaptive finite element approach, whereby we either perform a Newton step
(if the Newton linearisation effect dominates) or enrich the current finite element space based on the above {\em a posteriori} residual indicators (in the case that the finite element discretisation constitutes the main source of error); for related work we refer to~\cite{ChaillouSuri:07,El-AlaouiErnVohralik:11}, or the articles~\cite{BernardiDakroubMansourSayah:15,CongreveWihler:15,GarauMorinZuppa:11} on (derivative-free) fixed-point iteration schemes. Finally, we point to the works~\cite{ChaillouSuri:06,Han:94} dealing with modelling errors in linearised models.
In the current article, we extend the work undertaken in~\cite{AmreinWihler:15} to the framework of $hp$--version adaptive interior penalty discontinuous Galerkin (DG) schemes, thereby giving rise to $hp$--adaptive Newton-discontinuous Galerkin (NDG) methods. Here, the proof of
the resulting {\em a posteriori} residual bound for the interior penalty DG
discretisation of the underlying linearised PDE problem is based on two key steps:
firstly, we introduce a suitable residual operator on a given
enriched space, which, when measured in an appropriate norm, is equivalent
to the error measured in terms of the underlying DG energy norm. Secondly, an upper
bound on the norm of the residual operator is derived based on
exploiting the general techniques developed in the
articles~\cite{HoustonSchotzauWihler:07,HoustonSchotzauWihler:06,Zhu_3D};
we also refer to \cite{SchotzauZhu:11} for the application to
convection--diffusion problems, and to~\cite{HoustonSuliWihler:08,CongreveHoustonSuliWihler:13} for the treatment of strongly monotone
quasilinear PDEs, cf., also,
\cite{CongreveHoustonWihler:13,Congreve_Houston_2014} for $hp$--version
two-grid DG methods.
The proof of this upper bound crucially relies on the approximation of
discontinuous finite element functions by conforming ones,
cf., also, \cite{KarakashianPascal03} for the $h$--version case.
Moreover, in the current setting, following \cite{Verfurth:98}, particular
care is devoted to the derivation of $\epsilon$-robust approximation estimates.
The resulting {\em a posteriori} bound
consists of two key terms: one stemming from the Newton linearisation error,
and the second which measures the approximation error in the underlying DG
scheme. On the basis of this general $hp$--version bound, we devise
a fully automatic $hp$--adaptive NDG scheme for the numerical approximation
of PDEs of the form~\eqref{eq:PDE}. Indeed, the performance of the
resulting adaptive strategy is demonstrated on both the Bratu and
Ginzburg Landau problems; moreover, the superiority of exploiting $hp$--enrichment
of the DG finite element space, in comparison with standard mesh adaptation ($h$--refinement), will be highlighted.
The structure of this article is as follows. In Section~\ref{sec:newton_linearisation}
we briefly outline the adaptive (damped) Newton linearisation
procedure employed within this article. The $hp$--version interior penalty
DG discretisation of the resulting linearised PDE problem is then given in
Section~\ref{sc:dgfem}. Section~\ref{sc:apost} is devoted to the derivation
of a residual-based {\em a posteriori} bound. On the basis of this bound in Section~\ref{sec:adaptivity} we design a suitable adaptive refinement strategy, which controls both the
error arising in the Newton linearisation, as well as the error in the $hp$--DG
finite element scheme; in the latter case, we exploit automatic $hp$--refinement
of the underlying finite element space. The performance of this
proposed algorithm is demonstrated for a series of numerical examples
presented in Section~\ref{sec:numerics}. Finally, in Section~\ref{sec:conlusions} we summarise
the work presented in this article and discuss potential future extensions.
\section{Newton Linearisation} \label{sec:newton_linearisation}
\subsection{An Adaptive Newton Approach}
We will briefly revisit an adaptive `black-box' prediction-type Newton algorithm from~\cite{AmreinWihler:15}, and refer to~\cite{Deuflhard:04} for more sophisticated approaches in more specific situations. Let us consider two Banach spaces $ X, Y $, with norms~$\|\cdot\|_X$ and~$\|\cdot\|_Y$, respectively. Then, given an open subset~$\Xi\subset X$, and a (possibly nonlinear) operator~$\mathsf{F}_{\eps}:\,\Xi\to Y$, we are interested in solving the {\em nonlinear} operator equation
\begin{equation}\label{eq:F=0}
\mathsf{F}_{\eps}(u)=0,
\end{equation}
for some unknown zeros~$u\in\Xi$. Supposing that the Fr\'echet derivative~$\mathsf{F}_{\eps}'$ of~$\mathsf{F}_{\eps}$ exists in~$\Xi$ (or in a suitable subset), the classical Newton method for solving~\eqref{eq:F=0} starts from an initial guess~$u_0\in\Xi$, and generates a sequence~$\{u_n\}_{n\ge 1}\subset X$ that is defined iteratively by the {\em linear} equation
\begin{equation}\label{eq:newton}
\mathsf{F}_{\eps}'(u_n)(u_{n+1}-u_n)=-\mathsf{F}_{\eps}(u_n),\qquad n\ge 0.
\end{equation}
Naturally, for this iteration to be well-defined, we need to assume that~$\mathsf{F}_{\eps}'(u_n)$ is invertible for all~$n\ge 0$, and that~$\{u_n\}_{n\ge 0}\subset\Xi$.
In order to improve the reliability of the Newton method~\eqref{eq:newton} in the case that the initial guess~$u_0$ is relatively far away from a root $u_{\infty}\in\Xi$ of~$\mathsf{F}_{\eps}$, $\mathsf{F}_{\eps}(u_\infty)=0$, introducing some damping in the Newton method is a well-known remedy. In that case~\eqref{eq:newton} is rewritten as
\begin{equation}\label{eq:damped}
u_{n+1}=u_n-\Delta t_n\mathsf{F}_{\eps}'(u_n)^{-1}\mathsf{F}_{\eps}(u_n),\qquad n\ge 0,
\end{equation}
where $\Delta t_n>0$, $n\ge 0$, is a damping parameter that may be adjusted {\em adaptively} in each iteration step. The selection of the Newton parameter~$\Delta t_n$ is based on the following idea from~\cite{AmreinWihler:15}: provided that~$\mathsf{F}_{\eps}'(u)$ is invertible on a suitable subset of~$\Xi\subset X$, we define the {\em Newton-Raphson transform} by
\[
u\mapsto\mathsf{NF}(u):=-\mathsf{F}_{\eps}'(u)^{-1}\mathsf{F}_{\eps}(u);
\]
see, e.g., \cite{SchneebeliWihler:11}. Then, rearranging terms in~\eqref{eq:damped}, we notice that
\begin{equation*}
\frac{u_{n+1}-u_{n}}{\Delta t_n}=\mathsf{NF}(u_n), \qquad n\ge 0,
\end{equation*}
i.e., \eqref{eq:damped} can be seen as the discretisation of the dynamical system
\begin{equation}\label{eq:davy}
\begin{split}
\dot{u}(t)&=\mathsf{NF}(u(t)), \quad t\geq 0,\qquad
u(0)=u_0,
\end{split}
\end{equation}
by the forward Euler scheme, with step size~$\Delta t_n>0$.
For~$t\in[0,\infty)$, the solution~$u(t)$ of~\eqref{eq:davy}, if it exists, defines a trajectory in~$X$ that starts at~$u_0$, and that will potentially converge to a zero of~$\mathsf{F}_{\eps}$ as~$t\to\infty$. Indeed, this can be seen (formally) from the integral form of~\eqref{eq:davy}, that is,
\begin{equation*}
\mathsf{F}_{\eps}(u(t))=\mathsf{F}_{\eps}(u_0)e^{-t},\qquad t\ge 0,
\end{equation*}
which implies that~$\mathsf{F}_{\eps}(u(t))\to 0$ as~$t\to\infty$.
Now taking the view of dynamical systems, our goal is to compute an upper bound for the value of the step sizes~$\Delta t_n>0$ from~\eqref{eq:damped}, $n\ge 0$, so that the discrete forward Euler solution~$\{u_n\}_{n\ge 0}$ from~\eqref{eq:damped} stays reasonably close to the continuous solution of~\eqref{eq:davy}. Specifically, for a prescribed tolerance~$\tau>0$, a Taylor expansion analysis (see~\cite[Section~2]{AmreinWihler:15} for details) reveals that
\[
u(t)=u_0+t\mathsf{NF}(u_0)+\frac{t^2}{2h}\eta_h+\mathcal{O}(t^3)+\mathcal{O}(t^2h\|\mathsf{NF}(u_0)\|_X^2),
\]
where, for any sufficiently small~$h>0$, we let~$\eta_h=\mathsf{NF}(u_0+h\mathsf{NF}(u_0))-\mathsf{NF}(u_0)$. Hence, after the first time step of length~$\Delta t_0>0$ there holds
\begin{equation}\label{eq:O}
u(\Delta t_0)-u_1=\frac{\Delta t_0^2}{2h}\eta_h+\mathcal{O}(\Delta t_0^3)+\mathcal{O}(\Delta t_0^2h\|\mathsf{NF}(u_0)\|_X^2),
\end{equation}
where~$u_1$ is the forward Euler solution from~\eqref{eq:damped}. Therefore, upon setting
\[
\Delta t_0=\sqrt{2\tau h\|\eta_h\|_X^{-1}},
\]
we arrive at
\[
\|u(\Delta t_0)-u_1\|_X\le\tau+\mathcal{O}(\Delta t_0^3)+\mathcal{O}(\Delta t_0^2h\|\mathsf{NF}(u_0)\|_X^2).
\]
In order to balance the~$\mathcal{O}$-terms in~\eqref{eq:O} it is sensible to make the choice
\[
h=\mathcal{O}(\Delta t_0\|\mathsf{NF}(u_0)\|_X^{-2}),
\]
i.e.,
\begin{equation}\label{eq:h}
h=\gamma \Delta t_0\|\mathsf{NF}(u_0)\|_X^{-2},
\end{equation}
for some parameter~$\gamma>0$. This leads to the following \emph{adaptive Newton algorithm}.\\
\begin{algorithm}~\label{al:zs}
Fix a tolerance $\tau>0$ as well as a parameter~$\gamma>0$, and set~$n\gets0$.
\begin{algorithmic}[1]
\State Start the Newton iteration with an initial guess $ u_{0}\in\Xi$.
\If {$n=0$} {choose
\[
\Delta t_0=\min\left\{\sqrt{2\tau\NN{\mathsf{NF}(u_{0})}_{X}^{-1}},1\right\},
\]
based on~\cite[Algorithm~2.1]{AmreinWihler:15} (cf. also~\cite{AmreinWihler:14}),}
\ElsIf {$n\ge 1$} {let $\kappa_{n}=\Delta t_{n-1}$, and $h_n=\gamma \kappa_n\|\mathsf{NF}(u_n)\|_X^{-2}$ based on~\eqref{eq:h}; define the Newton step size
\begin{equation}\label{eq:knew}
\Delta t_n=\min\left\{\sqrt{2\tau h_n\NN{\mathsf{NF}(u_n+h_n\mathsf{NF}(u_n))-\mathsf{NF}(u_n)}^{-1}_{X}},1\right\}.
\end{equation}}
\EndIf
\State Compute~$u_{n+1}$ based on the Newton iteration~\eqref{eq:damped}, and go to~({\footnotesize 3:}) with $n\leftarrow n+1 $.
\end{algorithmic}
\end{algorithm}
We notice that the minimum in~\eqref{eq:knew} ensures that the step size~$\Delta t_n$ is chosen to be~1 whenever possible. Indeed, this is required in order to guarantee quadratic convergence of the Newton iteration close to a root (provided that the root is simple). Furthermore, we remark that the prescribed tolerance~$ \tau $ in the above adaptive strategy will typically be fixed {\em a priori}. Here, for highly nonlinear problems featuring numerous or even infinitely many solutions, it is typically mandatory to select~$\tau\ll 1$ small in order to remain within the attractor of the given initial guess. This is particularly important if the starting value is relatively far away from a solution.
\subsection{Application to Semilinear PDEs}
In this article, we suppose that a (not necessarily unique) solution~$u\in X:=H^1_0(\Omega)$ of~\eqref{eq:PDE} exists; here, we denote by $H^1_0(\Omega)$ the standard Sobolev space of functions in~$H^1(\Omega)=W^{1,2}(\Omega)$ with zero trace on~$\partial\Omega$. Furthermore, signifying by~$X'=H^{-1}(\Omega)$ the dual space of~$X$, and upon defining the map $\mathsf{F}_{\eps}: X\rightarrow X'$ through
\begin{equation}\label{eq:Fweak}
\dprod{\mathsf{F}_{\eps}(u),v}:= \int_{\Omega}\left\{ \epsilon\nabla u\cdot \nabla v+uv-f(u)v\right\}\,\dd\bm x\qquad \forall v\in X,
\end{equation}
where $\dprod{\cdot,\cdot}$ is the dual product in~$X'\times X$, the above problem~\eqref{eq:PDE} can be written as a nonlinear operator equation in~$X'$:
\begin{equation}\label{eq:F0}
u\in X:\qquad \mathsf{F}_{\eps}(u)=0.
\end{equation}
For any subset~$D\subseteq\Omega$, we denote by~$\|\cdot\|_{0,D}$ the $L^2$-norm on~$D$; in
the case when $D=\Omega$, we simply write $\|\cdot\|_{0}$ in lieu of $\|\cdot\|_{0,\Omega}$.
With this notation, we note that the space~$X$ is equipped with the
norm
\[
\|u\|^2_X:=\epsilon\|\nabla u\|^2_{0}+\|u\|_0^2,\qquad u\in X.
\]
The Fr\'echet-derivative of the
operator~$\mathsf{F}_{\eps}$ from~\eqref{eq:F0} at~$u\in X$ is given by
\begin{equation*}
\dprod{\mathsf{F}_{\eps}'(u)w,v}= \int_{\Omega}\left\{\epsilon\nabla w\cdot \nabla v+wv-f'(u)wv\right\}\,\dd\bm x,\qquad v,w\in X=H^1_0(\Omega),
\end{equation*}
where we write~$f'\equiv\partial_u f$. We note that, if there is a constant~$\omega>1$ for which~$f'(u)\in L^{\omega}(\Omega)$, then $\mathsf{F}_{\eps}'(u)$ is a well-defined linear and bounded mapping from~$X$ to~$X'$; see~\cite[Lemma~A.1]{AmreinWihler:15}.
Now given an initial guess~$u_0\in X$, the adaptive Newton method~\eqref{eq:damped} for~\eqref{eq:F0} is defined iteratively to find~$u_{n+1}\in X$ from~$u_n\in X$, $n\ge 0$, such that
\begin{equation*}
\mathsf{F}_{\eps}'(u_n)(u_{n+1}-u_n)=-\Delta t_n\mathsf{F}_{\eps}(u_n),
\end{equation*}
in~$X'$. When applied to~\eqref{eq:Fweak} and~\eqref{eq:F0}, this turns into
\begin{align*}
\int_{\Omega}\{\epsilon\nabla (u_{n+1}-u_n)&\cdot \nabla v+(u_{n+1}-u_n)v-f'(u_n)(u_{n+1}-u_n)v\}\,\dd\bm x\\
&=-\Delta t_n\int_{\Omega}\left\{ \epsilon\nabla u_n\cdot \nabla v+u_nv-f(u_n)v\right\}\,\dd\bm x\qquad\forall v\in X.
\end{align*}
Hence, for~$n\ge 0$, the updated Newton iterate~$u_{n+1}$ is defined through the \emph{linear} weak formulation
\begin{equation}\label{eq:weak}
\begin{split}
\int_\Omega \{\epsilon\nabla \widehat{\mathfrak{u}}_{n+1}\cdot\nabla v&+\widehat{\mathfrak{u}}_{n+1}v -f'(u_n) \widehat{\mathfrak{u}}_{n+1}v \}\,\dd\bm x
=\Delta t_n\int_\Omega \{ f(u_n)-f'(u_n)u_n\} v\,\dd\bm x
\qquad\forall v\in X,
\end{split}
\end{equation}
where~$\widehat{\mathfrak{u}}_{n+1}=u_{n+1}-(1-\Delta t_n)u_n$. Incidentally, if there exists a constant~$\delta$ with~$\epsilon^{-1}(f'(u_n)-1)\le\delta<C_{\PF}^{-2}$ on~$\Omega$, where~$C_{\PF}=C_{\PF}(\Omega)>0$ is the constant in the Poincar\'e-Friedrichs inequality on~$\Omega$,
\[
\|w\|_{0}\le C_{\PF}\|\nabla w\|_{0}\qquad\forall w \in X,
\]
then \eqref{eq:weak} is a linear second-order diffusion-reaction problem that is coercive on~$X$. In particular, \eqref{eq:weak} exhibits a unique solution~$u_{n+1}\in X$ in this case.
\section{$hp$--DG Discretisation}\label{sc:dgfem}
\subsection{Meshes, Spaces, and DG Flux Operators}
We will employ a standard $hp$--DG setting; see, e.g.,~\cite{HoustonSchotzauWihler:07,SchotzauZhu:11}.
\subsubsection{Meshes and DG Spaces}
Let $\mathcal{T}$ be a subdivision of $\Omega$ into disjoint open parallelograms $\kappa$ such that $\overline \Omega =
\bigcup_{\kappa \in \mathcal{T}} {\overline \kappa}$. We assume
that~$\mathcal{T}$ is shape-regular, and that each $\kappa \in \mathcal{T}$ is an affine
image of the unit square~$\widehat\kappa=(0,1)^2$; i.e., for each
$\kappa\in\mathcal{T}$ there exists an affine element mapping $\Psi_\kappa:\,\widehat\kappa\to\kappa$ such that $\kappa = \Psi_\kappa({\widehat \kappa})$. By $h_\kappa$ we denote the
element diameter of $\kappa\in\mathcal{T}$, $h = \max_{\kappa \in {\mathcal
T}_h} h_\kappa$ is the mesh size, and $\bm n_\kappa$ signifies the unit outward
normal vector to $\kappa$ on~$\partial\kappa$. Furthermore, we assume that $\mathcal{T}$ is of bounded
local variation, i.e., there exists a constant $\rho_1\ge 1$,
independent of the element sizes, such that $\rho_1^{-1}\le \nicefrac{h_\kappa}{h_{\kappa'}}\le \rho_1$, for any pair of elements $\kappa, \kappa'\in\mathcal{T}$ which share a common edge $e=(\partial\kappa\cap\partial\kappa')^\circ$. In this context, let us consider the set ${\mathcal E}$ of all one-dimensional open edges of all elements $\kappa \in\mathcal{T}$. Further, we denote by $\mathcal{E}_{\mathcal{I}}$ the set of all edges $e$ in $\mathcal{E}$ that are contained in $\Omega$ (interior edges). Additionally, introduce ${\mathcal E}_{\mathcal{B}}$ to be the set of boundary edges consisting of all $e\in{\mathcal E}$ that are contained in $\partial\Omega$. In our analysis, we allow the meshes to be 1-irregular, i.e., each edge of an element~$\kappa\in\mathcal{T}$ may contain (at most) one hanging node, which we assume to be located at the centre of~$e$. Suppose that $e$ is an edge of an element $\kappa \in \mathcal{T}$; then, by $h_e$, we denote the length of $e$. Due to our assumptions on the subdivision $\mathcal{T}$ we have that, if $e \subset \partial \kappa$, then $h_e$ is commensurate with $h_{\kappa}$, the diameter of
$\kappa$.
For a nonnegative integer $k$, we denote by ${\mathcal Q}_k({\widehat \kappa})$
the set of all tensor-product polynomials on $\widehat \kappa$ of degree $k$ in
each co-ordinate direction. To each $\kappa \in \mathcal{T}$ we
assign a polynomial degree $p_{\kappa}$ (local approximation order). We store the quantities $h_\kappa$ and $p_{\kappa}$ in the
vectors ${\bm h} = \{ h_{\kappa}:\, \kappa \in\mathcal{T}\}$ and ${\bm p} = \{
p_{\kappa}:\, \kappa \in \mathcal{T}\}$, respectively, and
consider the DG finite element space
\begin{equation}\label{eq:VDG}
\mathcal{V}_{\DG} = \{ v \in L^2(\Omega):\;v |_{\kappa} \circ \Psi_{\kappa} \in
{\mathcal Q}_{p_{\kappa}}
(\widehat{\kappa}) \quad \forall \kappa \in \mathcal{T} \} \; .
\end{equation}
We shall suppose that the polynomial degree vector $\mathbf{p}$, with
$p_{\kappa}\geq 1$ for each $\kappa \in \mathcal{T}$, has
bounded local variation, i.e., there exists a constant $\rho_2\geq
1$ independent of $\bm h$ and $\bm p$, such that, for any pair of
neighbouring elements $\kappa, \kappa'\in\mathcal{T}$, we have $\rho_2^{-1} \leq \nicefrac{p_{\kappa}}{p_{\kappa'}} \leq \rho_2$. Moreover, for an edge~$e=(\partial\kappa\cap\partial\kappa')^\circ$ shared by two elements~$\kappa,\kappa'\in\mathcal{T}$, we define~$p_e:=\nicefrac12(p_{\kappa}+p_{\kappa'})$, or~$p_e=p_\kappa$ if~$e=(\partial\kappa\cap\partial\Omega)^\circ$, for some~$\kappa\in\mathcal{T}$, is a boundary edge.
\subsubsection{Jump and Average Operators}
Let $\kappa$ and $\kappa'$ be two adjacent elements of~ $\mathcal{T}$, and
$\bm x$ an arbitrary point on the interior edge $e\in\mathcal{E}_{\mathcal{I}}$ given by
$e=(\partial \kappa\cap\partial \kappa')^\circ$. Furthermore, let $v$
and~$\bm{q}$ be scalar- and vector-valued functions, respectively,
that are sufficiently smooth inside each
element~$\kappa,\kappa'$. Then, the averages of $v$ and
$\bm{q}$ at $\bm{x}\in e$ are given by
\[
\avg{v}=\frac{1}{2}(v|_{\kappa}+v|_{\kappa'}), \qquad \avg{\bm{q}}
=\frac{1}{2}(\bm{q}|_{\kappa}+\bm{q}|_{\kappa'}),
\]
respectively. Similarly, the jumps of $v$ and $\bm{q}$ at $\bm{x}\in
e$ are given by
\[
\jmp{v} =v|_{\kappa}\,\bm{n}_{\kappa}+v|_{\kappa'}\,\bm{n}_{\kappa'},\qquad
\jmp{\bm{q}}=\bm{q}|_{\kappa}\cdot\bm{n}_{\kappa}+\bm{q}|_{\kappa'}\cdot\bm{n}_{\kappa'},
\]
respectively. On a boundary edge $e\in\mathcal{E}_\mathcal{B}$,
we set $\avg{v}=v$, $\avg{\bm{q}}=\bm{q}$ and $\jmp{v}=v\bm{n}$,
with~$\bm{n}$ denoting the unit outward normal vector on the boundary
$\partial\Omega$.
Furthermore, we introduce, for an edge $e\in{\mathcal E}$, the discontinuity penalisation parameter~$\sigma$ by
\begin{equation}\label{sigma}
\sigma|_{e} = \frac{p_e^2}{h_e}.
\end{equation}
We conclude this section by equipping the DG space $\mathcal{V}_{\DG}$ with the DG norm
\begin{equation}\label{eq:DGnorm}
\NN{v}_{\text{\tiny\sf{DG}}}^2 := \epsilon\NN{\nabla_{\T} v}^2_{0}+\NN{v}^2_{0}+\int_{\mathcal{E}}(\epsilon\sigma+\sigma^{-1})|\jmp{v}|^2\,\dd s,
\end{equation}
which is induced by the DG inner product
\begin{equation}\label{eq:ip}
(v,w)_{\text{\tiny\sf{DG}}} = \int_{\Omega}\left\{\epsilon
\nabla_{\T} v\cdot\nabla_{\T} w +vw\right\}\,\dd\bm x + \int_{\mathcal{E}} (\epsilon\sigma+\sigma^{-1})\jmp{w}\cdot\jmp{v}\,\dd s.
\end{equation}
Here, $\nabla_{\T}$ is the element-wise gradient operator. For an element~$\kappa\in\mathcal{T}$ we shall also use the norm
\[
\|v\|_{\epsilon,\kappa}^2:=\epsilon\NN{\nabla v}^2_{0,\kappa}+\NN{v}^2_{0,\kappa},
\]
for~$v\in H^1(\kappa)$.
\subsubsection{Conforming Subspaces}
For a given DG finite element space~$\mathcal{V}_{\DG}$, cf.~\eqref{eq:VDG}, we define the extended space
\[
\mathcal{W}_{\DG}:=H^1_0(\Omega)+\mathcal{V}_{\DG}.
\]
With this notation, the following result holds.
\begin{lemma}\label{lm:A}
There exists a linear operator~$\mathsf{A}_{\DG}:\,\mathcal{W}_{\DG}\to H^1_0(\Omega)$ such that
\begin{equation}\label{eq:equiv}
\begin{split}
\|w-\mathsf{A}_{\DG} w\|^2_{0}&\le C_\eqref{eq:equiv}\sum_{e\in\mathcal{E}}\int_{e}\sigma^{-1}|\jmp{w}|^2\,\dd s,\\
\|\nabla_{\T}(w-\mathsf{A}_{\DG} w)\|^2_{0}&\le C_\eqref{eq:equiv}\sum_{e\in\mathcal{E}}\int_{e}\sigma|\jmp{w}|^2\,\dd s,
\end{split}
\end{equation}
for any~$w\in\mathcal{W}_{\DG}$, where~$C_\eqref{eq:equiv}>0$ is a constant independent of~$\mathcal{T}$ and of~$\bm p$.
\end{lemma}
\begin{proof}
Consider the space~$\mathcal{V}_{\DG}^\|:=\mathcal{V}_{\DG}\cap H^1_0(\Omega)$, and denote by~$\P^\|:\,\mathcal{V}_{\DG}\to\mathcal{V}_{\DG}^\|$
the orthogonal projection with respect to the inner product defined in~\eqref{eq:ip},
i.e.,
\[
w\in\mathcal{V}_{\DG}:\quad (w-\P^\|w,v)_{\text{\tiny\sf{DG}}}=0\qquad\forall v\in \mathcal{V}_{\DG}^\|.
\]
Then, defining the subspace~$\mathcal{V}_{\DG}^\perp:=(\mathsf{id}-\P^\|)\mathcal{V}_{\DG}$, we have the direct
sum~$\mathcal{V}_{\DG}=\mathcal{V}_{\DG}^\|\oplus\mathcal{V}_{\DG}^\perp$, as well as
\begin{equation}\label{eq:spacedecomp}
\mathcal{W}_{\DG}=H^1_0(\Omega)\oplus\mathcal{V}_{\DG}^\perp.
\end{equation}
Based on our assumptions on the mesh~$\mathcal{T}$, and referring to~\cite[Theorem~4.4]{SchotzauZhu:11}, there exists an operator~$\mathsf{I}_{hp}:\,\mathcal{V}_{\DG}\to H^1_0(\Omega)$ that satisfies
\begin{align*}
\sum_{\kappa\in\mathcal{T}}\|v-\mathsf{I}_{hp} v\|^2_{L^2(\kappa)}&\le C\sum_{e\in\mathcal{E}}\int_{e}\sigma^{-1}|\jmp{v}|^2\,\dd s,\\
\sum_{\kappa\in\mathcal{T}}\|\nabla(v-\mathsf{I}_{hp} v)\|^2_{L^2(\kappa)}&\le C\sum_{e\in\mathcal{E}}\int_{e}\sigma|\jmp{v}|^2\,\dd s,
\end{align*}
for any~$v\in\mathcal{V}_{\DG}$. By virtue of~\eqref{eq:spacedecomp}, we can now construct the operator~$\mathsf{A}_{\DG}$ as follows: for any~$w\in\mathcal{W}_{\DG}$, there exist unique representatives~$w_0\in H^1_0(\Omega)$ and~$w_{\text{\tiny\sf{DG}}}^\perp\in\mathcal{V}_{\DG}^\perp$ with~$w=w_0+w_{\text{\tiny\sf{DG}}}^\perp$.
Hence, defining~$\mathsf{A}_{\DG} w:=w_0+\mathsf{I}_{hp} w_{\text{\tiny\sf{DG}}}^\perp\in H^1_0(\Omega)$, and employing the previous estimates, we obtain
\begin{align*}
\|\nabla_{\T}(w-\mathsf{A}_{\DG} w)\|^2_{0}
&=\sum_{\kappa\in\mathcal{T}}\|\nabla(w_{\text{\tiny\sf{DG}}}^\perp-\mathsf{I}_{hp} w_{\text{\tiny\sf{DG}}}^\perp)\|^2_{L^2(\kappa)}
\le C\sum_{e\in\mathcal{E}}\int_{e}\sigma|\jmp{w_{\text{\tiny\sf{DG}}}^\perp}|^2\,\dd s.
\end{align*}
Since~$w_0\in H^1_0(\Omega)$, we notice that~$\jmp{w_0}|_e=\bm 0$ for all~$e\in\mathcal{E}$;
thereby,
\[
\|\nabla_{\T}(w-\mathsf{A}_{\DG} w)\|^2_{0}
\le C\sum_{e\in\mathcal{E}}\int_{e}\sigma|\jmp{w}|^2\,\dd s,
\]
which proves the second bound in~\eqref{eq:equiv}. The first inequality results from an analogous argument.
\end{proof}
\begin{remark}\label{rm:A}
We note that any $v\in H^1_0(\Omega)$ satisfies~$\jmp{v}=\bm 0$ on~$\mathcal{E}$; thereby, in view of~\eqref{eq:equiv}, it follows that~$\mathsf{A}_{\DG} v=v$ for all~$v\in H^1_0(\Omega)$. Furthermore, for~$w\in\mathcal{W}_{\DG}$, upon application of the triangle inequality and Lemma~\ref{lm:A}, we deduce that
\begin{align*}
\|\mathsf{A}_{\DG} w\|_{X}^2
&=\epsilon\|\nabla\mathsf{A}_{\DG} w\|_0^2+\|\mathsf{A}_{\DG} w\|^2_0\\
&\le2\epsilon\|\nabla w\|_0^2+2\|w\|^2_0
+2\epsilon\|\nabla(w-\mathsf{A}_{\DG} w)\|_0^2+2\|w-\mathsf{A}_{\DG} w\|^2_0\\
&\le2\epsilon\|\nabla w\|_0^2+2\|w\|_0^2
+2C_\eqref{eq:equiv}\sum_{e\in\mathcal{E}}\int_{e}(\epsilon\sigma+\sigma^{-1})|\jmp{w}|^2\,\dd s.
\end{align*}
Thus the following stability estimate holds
\begin{equation}\label{eq:stab}
\|\mathsf{A}_{\DG} w\|_{X}
\le C_\eqref{eq:stab}\|w\|_{\text{\tiny\sf{DG}}}\qquad\forall w\in\mathcal{W}_{\DG},
\end{equation}
where~$C_\eqref{eq:stab}=\sqrt{2\max(1,C_\eqref{eq:equiv})}$.
\end{remark}
\subsection{Linear $hp$--DG Approximation}
The $hp$--version interior penalty DG discretisation of~\eqref{eq:weak}
is given by: find~$u^{\DG}_{n+1}\in\mathcal{V}_{\DG}$ from~$u^{\DG}_n$ such that
\begin{equation}\label{eq:dgfem}
\a(u^{\DG}_n;u^{\DG}_{n+1},v)=\l(u^{\DG}_n;v)\qquad\forall v\in\mathcal{V}_{\DG}.
\end{equation}
Here, for a method parameter~$\theta\in[-1,1]$ and a penalty parameter~$C_{\sigma}\ge 0$, we define the forms
\begin{equation}\label{eq:a}
\begin{split}
\a(u^{\DG}_n;u^{\DG}_{n+1},v):=
&\int_\Omega\left\{\epsilon\nabla_{\T}\widehat{\mathfrak{u}}^{\DG}_{n+1}\cdot\nabla_{\T} v+\widehat{\mathfrak{u}}^{\DG}_{n+1} v-f'(u^{\DG}_n)\widehat{\mathfrak{u}}^{\DG}_{n+1}v\right\}\,\dd\bm x\\
&-\int_{\mathcal{E}}\left\{\avg{\epsilon\nabla_{\T}\widehat{\mathfrak{u}}^{\DG}_{n+1}}\cdot\jmp{v}+\theta\jmp{\widehat{\mathfrak{u}}^{\DG}_{n+1}}\cdot\avg{\epsilon\nabla_{\T} v}\right\}\,\dd s\\
&\quad+C_{\sigma}\int_{\mathcal{E}}\epsilon\sigma\jmp{\widehat{\mathfrak{u}}^{\DG}_{n+1}}\cdot\jmp{v}\,\dd s,
\end{split}
\end{equation}
and
\[
\l(u^{\DG}_n;v) = \int_\Omega\f{u^{\DG}_n}v\,\dd\bm x,
\]
for~$v\in\mathcal{V}_{\DG}$, where for $n\ge 0$, we set
\begin{equation} \label{eq:hat}
\begin{split}
\widehat{\mathfrak{u}}^{\DG}_{n+1} & :=u^{\DG}_{n+1}-(1-\Delta t_n)u^{\DG}_n, \\
\f{u^{\DG}_n} & := \Delta t_n (f(u^{\DG}_n)-f'(u^{\DG}_n)u^{\DG}_n).
\end{split}
\end{equation}
The choices~$\theta\in\{-1,0,1\}$ correspond, respectively, to the non-symmetric (NIPG),
incomplete (IIPG), and symmetric (SIPG) interior penalty DG schemes;
cf.~\cite{StammWihler:10}. For the IIPG and SIPG methods,
the penalty parameter $C_{\sigma}$ must be chosen sufficiently large to guarantee
stability of the underlying DG scheme,
cf. \cite{WihlerFrauenfelderSchwab:03}, for example. Furthermore, an additional
constraint on the minimal value of $C_{\sigma}$ will be introduced in
Proposition~\ref{prop:residual_bounds} below.
\section{$hp$--Version \emph{A Posteriori} Analysis}
\label{sc:apost}
\subsection{A DG Residual}
We introduce a residual operator
\[
\mathsf{R}_{\eps}:\,\mathcal{W}_{\DG}\to \mathcal{W}_{\DG}',
\]
where~$\mathcal{W}_{\DG}'$ is the dual space of~$\mathcal{W}_{\DG}$, as follows: given the operator~$\mathsf{A}_{\DG}$ constructed in Lemma~\ref{lm:A}, and~$w\in\mathcal{W}_{\DG}$, let us define
\begin{equation}\label{eq:R}
\begin{split}
\dprod{\mathsf{R}_{\eps}(w),v}:&=\int_\Omega\left\{\epsilon\nabla_{\T} w\cdot\nabla \mathsf{A}_{\DG} v+w\mathsf{A}_{\DG} v-f(w)\mathsf{A}_{\DG} v\right\}\,\dd\bm x\\
&\quad+C_{\sigma}\int_{\mathcal{E}}(\epsilon\sigma+\sigma^{-1})\jmp{w}\cdot\jmp{v}\,\dd s
\qquad\forall v\in \mathcal{W}_{\DG},
\end{split}
\end{equation}
with~$\sigma$ from~\eqref{sigma}, and~$C_{\sigma}$ appearing in~\eqref{eq:a}. Furthermore, for~$w\in\mathcal{W}_{\DG}$, we introduce the norm
\begin{equation}\label{eq:NNN}
\NNN{\mathsf{R}_{\eps}(w)}:=\sup_{\phi\in\mathcal{W}_{\DG}}\frac{\dprod{\mathsf{R}_{\eps}(w),\phi}}{\|\phi\|_{\text{\tiny\sf{DG}}}}.
\end{equation}
For a solution~$u\in H^1_0(\Omega)$ of~\eqref{eq:PDE}, we again note
that~$\jmp{u}=\bm 0$ on~$\mathcal{E}$, and, hence, due to~\eqref{eq:Fweak} and~\eqref{eq:F0}, we conclude that
\begin{equation}\label{eq:R0}
\dprod{\mathsf{R}_{\eps}(u),v}=0\qquad\forall v\in \mathcal{W}_{\DG}.
\end{equation}
Moreover, the following result shows that, under suitable conditions on the nonlinearity~$f$, the norm~$\NNN{\mathsf{R}_{\eps}(\cdot)}$ defined in~\eqref{eq:NNN} is directly related to the DG-norm given in~\eqref{eq:DGnorm}. In this sense, we may employ the norm~$\NNN{\mathsf{R}_{\eps}(\cdot)}$ as a natural measure for the approximation in the Newton-DG formulation~\eqref{eq:dgfem}.
\begin{proposition} \label{prop:residual_bounds}
Suppose that there exist constants~$\varrho_0>-1$ and $L\ge0$ such that~$f$ satisfies
\begin{equation}\label{eq:f}
\varrho_0\le -f',\qquad\text{and}\qquad |f'|\le L,
\end{equation}
on~$\overline\Omega\times\mathbb{R}$. Furthermore, assume that the penalty
parameter~$C_{\sigma}$ is sufficiently large so that
\[
C_{\sigma}\ge \frac{c_0}{2}+\frac{C_\eqref{eq:equiv}(1+L)^2}{2c_0},
\]
where~$C_\eqref{eq:equiv}$ is the constant arising in the bounds~\eqref{eq:equiv}, and~$c_0=1+\min(0,\varrho_0)>0$. Then, for any weak solution~$u\in H^1_0(\Omega)$ of~\eqref{eq:PDE}, the following bounds hold
\begin{equation}\label{eq:sup}
\frac{c_0}{2}\|u-w\|_{\text{\tiny\sf{DG}}}\le\NNN{\mathsf{R}_{\eps}(w)}
\le \sqrt2\max\left(C_{\eqref{eq:stab}}(1+L),C_{\sigma}\right)\|u-w\|_{\text{\tiny\sf{DG}}}
\end{equation}
for all~$w\in\mathcal{W}_{\DG}$, where~$C_\eqref{eq:stab}$ is the constant arising in~\eqref{eq:stab}.
\end{proposition}
\begin{proof}
The two bounds are proved separately. Let~$w\in\mathcal{W}_{\DG}$, then employing~\eqref{eq:R0}, and
noting that~$\mathsf{A}_{\DG} u=u$, cf.~Remark~\ref{rm:A}, we obtain
\begin{align*}
\dprod{\mathsf{R}_{\eps}(w),w-u}
&=\dprod{\mathsf{R}_{\eps}(u)-\mathsf{R}_{\eps}(w),u-w}\\
&=\epsilon\int_\Omega\nabla_{\T}(u-w)\cdot\nabla(u-\mathsf{A}_{\DG} w)\,\dd\bm x+\int_\Omega(u-w)(u-\mathsf{A}_{\DG} w)\,\dd\bm x\\
&\quad-\int_\Omega(f(u)-f(w))(u-\mathsf{A}_{\DG} w)\,\dd\bm x
+C_{\sigma}\int_{\mathcal{E}}(\epsilon\sigma+\sigma^{-1})|\jmp{u-w}|^2\,\dd s\\
&=\|u-w\|_{\text{\tiny\sf{DG}}}^2+\epsilon\int_\Omega\nabla_{\T}(u-w)\cdot\nabla_{\T}(w-\mathsf{A}_{\DG} w)\,\dd\bm x\\
&\quad+\int_\Omega(u-w)(w-\mathsf{A}_{\DG} w)\,\dd\bm x
-\int_\Omega(f(u)-f(w))(u-w)\,\dd\bm x\\
&\quad-\int_\Omega(f(u)-f(w))(w-\mathsf{A}_{\DG} w)\,\dd\bm x
+(C_{\sigma}-1)\int_{\mathcal{E}}(\epsilon\sigma+\sigma^{-1})|\jmp{u-w}|^2\,\dd s.
\end{align*}
Given the assumptions on $f$ stated in~\eqref{eq:f} hold, we conclude that
\[
-(f(u)-f(w))(u-w)\ge\varrho_0|u-w|^2,\qquad |f(u)-f(w)|\le L|u-w|,
\]
on~$\overline\Omega\times\mathbb{R}$. Thus, applying the Cauchy-Schwarz inequality, we arrive at
\begin{align*}
\dprod{\mathsf{R}_{\eps}(w),w-u}
&\ge(1+\min(0,\varrho_0))\|u-w\|_{\text{\tiny\sf{DG}}}^2
-\epsilon\|\nabla_{\T}(u-w)\|_0\|\nabla_{\T}(w-\mathsf{A}_{\DG} w)\|_0\\
&\quad-(1+L)\|u-w\|_0\|w-\mathsf{A}_{\DG} w\|_0\\
&\quad+(C_{\sigma}-1-\min(0,\varrho_0))\int_{\mathcal{E}}(\epsilon\sigma+\sigma^{-1})|\jmp{u-w}|^2\,\dd s.
\end{align*}
Setting~$c_0=1+\min(0,\varrho_0)$, we deduce that
\begin{align*}
\dprod{\mathsf{R}_{\eps}(w),w-u}
&\ge c_0\|u-w\|_{\text{\tiny\sf{DG}}}^2
-\frac{c_0\epsilon}{2}\|\nabla_{\T}(u-w)\|^2_0-\frac{\epsilon}{2c_0}\|\nabla_{\T}(w-\mathsf{A}_{\DG} w)\|^2_0\\
&\quad-\frac{c_0}{2}\|u-w\|_0^2-\frac{(1+L)^2}{2c_0}\|w-\mathsf{A}_{\DG} w\|^2_0\\
&\quad+(C_{\sigma}-c_0)\int_{\mathcal{E}}(\epsilon\sigma+\sigma^{-1})|\jmp{u-w}|^2\,\dd s.
\end{align*}
By virtue of Lemma~\ref{lm:A}, and noting that~$\jmp{u}=\bm 0$ on~$\mathcal{E}$, we get
\begin{align*}
\dprod{\mathsf{R}_{\eps}(w),w-u}
&\ge\frac{c_0}{2}\|u-w\|^2_{\text{\tiny\sf{DG}}}
+\left(C_{\sigma}-\frac{c_0}{2}-\frac{C_\eqref{eq:equiv}(1+L)^2}{2c_0}\right)\int_{\mathcal{E}}(\epsilon\sigma+\sigma^{-1})|\jmp{u-w}|^2\,\dd s\\
&\ge\frac{c_0}{2}\|u-w\|^2_{\text{\tiny\sf{DG}}}.
\end{align*}
This gives the first bound in~\eqref{eq:sup}. In order to show the second estimate, we employ~\eqref{eq:f} and the Cauchy-Schwarz inequality, for any~$v\in \mathcal{W}_{\DG}$, to infer that
\begin{align*}
\dprod{\mathsf{R}_{\eps}(w),v}
&=\dprod{\mathsf{R}_{\eps}(w)-\mathsf{R}_{\eps}(u),v}\\
&=\int_\Omega\left\{\epsilon\nabla_{\T} (w-u)\cdot\nabla \mathsf{A}_{\DG} v+(w-u)\mathsf{A}_{\DG} v-(f(w)-f(u))\mathsf{A}_{\DG} v\right\}\,\dd\bm x\\
&\quad+C_{\sigma}\int_{\mathcal{E}}(\epsilon\sigma+\sigma^{-1})\jmp{w-u}\cdot\jmp{v}\,\dd s\\
&\le\epsilon\|\nabla_{\T}(w-u)\|_0\|\nabla\mathsf{A}_{\DG} v\|_0+(1+L)\|w-u\|_0\|\mathsf{A}_{\DG} v\|_0\\
&\quad+\left(C_{\sigma}^2C_\eqref{eq:stab}^{-2}\int_{\mathcal{E}}(\epsilon\sigma+\sigma^{-1})|\jmp{w-u}|^2\,\dd s\right)^{\nicefrac12}
\left(C_\eqref{eq:stab}^2\int_{\mathcal{E}}(\epsilon\sigma+\sigma^{-1})|\jmp{v}|^2\,\dd s\right)^{\nicefrac12}\\
&\le\max\left(1+L,C_{\sigma} C_\eqref{eq:stab}^{-1}\right)\|u-w\|_{\text{\tiny\sf{DG}}}
\left(\|\mathsf{A}_{\DG} v\|^2_{X}+C_\eqref{eq:stab}^2\int_{\mathcal{E}}(\epsilon\sigma+\sigma^{-1})|\jmp{v}|^2\,\dd s\right)^{\nicefrac12}.
\end{align*}
Recalling the stability of~$\mathsf{A}_{\DG}$ from~\eqref{eq:stab} yields
\[
\dprod{\mathsf{R}_{\eps}(w),v}\le \sqrt2C_{\eqref{eq:stab}}\max\left(1+L,C_{\sigma} C_\eqref{eq:stab}^{-1}\right)\|u-w\|_{\text{\tiny\sf{DG}}}\|v\|_{\text{\tiny\sf{DG}}}.
\]
This implies the second bound in~\eqref{eq:sup}, and, thus, completes the proof.
\end{proof}
\subsection{\emph{A Posteriori} Residual Analysis}
In this section we develop a residual--based \emph{a posteriori} numerical analysis for the $hp$--NDG method~\eqref{eq:dgfem}.
\subsubsection{$hp$--Approximation Estimates}\label{sc:approx}
Let~$v\in\mathcal{W}_{\DG}$ be arbitrary, and consider~$\mathsf{A}_{\DG} v\in H^1_0(\Omega)$ as in Lemma~\ref{lm:A}. Then, we may choose~$\phi^{\text{\tiny\sf{DG}}}\in\mathcal{V}_{\DG}$ such that, for all~$\kappa\in\mathcal{T}$, the stability bound
\[
\|\mathsf{A}_{\DG} v-\phi^{\text{\tiny\sf{DG}}}\|_{0,\kappa}\le \|\mathsf{A}_{\DG} v\|_{0,\kappa},
\]
as well as the approximation estimate
\begin{equation}\label{eq:interp}
\begin{split}
\|\nabla(\mathsf{A}_{\DG} v-\phi^{\text{\tiny\sf{DG}}})\|^2_{0,\kappa}
&+\frac{p_\kappa^2}{h_\kappa^2}\|\mathsf{A}_{\DG} v-\phi^{\text{\tiny\sf{DG}}}\|^2_{0,\kappa}
\le C_\eqref{eq:interp}\left(\|\nabla\mathsf{A}_{\DG} v\|^2_{0,\kappa}+\|\mathsf{A}_{\DG} v\|_{0,\kappa}^2\right)
\end{split}
\end{equation}
hold \emph{simultaneously}, where $C_\eqref{eq:interp}$ is a positive constant, independent of~$\bm h, \bm p$,
and~$\mathsf{A}_{\DG} v$; see~\cite[\S~3.1]{KarkulikMelenk:15}. Since~$\epsilon\in(0,1]$, we infer the bound
\[
\epsilon\|\nabla(\mathsf{A}_{\DG} v-\phi^{\text{\tiny\sf{DG}}})\|^2_{0,\kappa}\le C_\eqref{eq:interp}\|\mathsf{A}_{\DG} v\|_{\epsilon,\kappa}^2,
\]
and
\begin{equation}\label{eq:interp3}
\epsilon^{\nicefrac12}\|\nabla\phi^{\text{\tiny\sf{DG}}}\|_{0,\kappa}
\le\epsilon^{\nicefrac12}\|\nabla(\mathsf{A}_{\DG} v-\phi^{\text{\tiny\sf{DG}}})\|_{0,\kappa}+
\epsilon^{\nicefrac12}\|\nabla\mathsf{A}_{\DG} v\|_{0,\kappa}
\le C_\eqref{eq:interp3}\|\mathsf{A}_{\DG} v\|_{\epsilon,\kappa}.
\end{equation}
Moreover, following the approach outlined in~\cite{Verfurth:98} (see also~\cite{AmreinWihler:15}), we deduce from the above estimates that
\begin{align}\label{eq:L2bound}
\|\mathsf{A}_{\DG} v-\phi^{\text{\tiny\sf{DG}}}\|^2_{0,\kappa}
&\le \min\left(1,C_\eqref{eq:interp}\epsilon^{-1}h_\kappa^2p^{-2}_\kappa\right)\|\mathsf{A}_{\DG} v\|^2_{\epsilon,\kappa}
\le \max\left(1,C_\eqref{eq:interp}\right)\alpha^2_\kappa\|\mathsf{A}_{\DG} v\|^2_{\epsilon,\kappa},
\end{align}
where, for $\kappa\in\mathcal{T}$,
\begin{equation}\label{eq:alpha}
\alpha_\kappa:=\min\left(1,\epsilon^{-\nicefrac12}h_\kappa p^{-1}_\kappa\right).
\end{equation}
Furthermore, applying a multiplicative trace inequality, that is,
\begin{equation}\label{eq:trace}
\|\psi\|^2_{0,\partial\kappa}
\le C_{\eqref{eq:trace}}\left(h_\kappa^{-1}\|\psi\|^2_{0,\kappa}+\|\psi\|_{0,\kappa}\|\nabla\psi\|_{0,\kappa}\right),\qquad \psi\in H^1(\kappa),
\end{equation}
we obtain
\[
\|\mathsf{A}_{\DG} v-\phi^{\text{\tiny\sf{DG}}}\|^2_{0,\partial\kappa}
\le C_\eqref{eq:trace}\max\left(1,C_\eqref{eq:interp}\right)\widetilde\beta^2_\kappa\|\mathsf{A}_{\DG} v\|_{\epsilon,\kappa}^2,
\]
where, for~$\kappa\in\mathcal{T}$, we define
\[
\widetilde\beta_\kappa:=\sqrt{h^{-1}_\kappa\alpha^2_\kappa+\epsilon^{-\nicefrac12}\alpha_\kappa}.
\]
Noting the bound
\[
\widetilde\beta_\kappa^2= \epsilon^{-\nicefrac12}\alpha_\kappa\left(\epsilon^{\nicefrac12}h_\kappa^{-1}\alpha_\kappa+1\right)
\le\epsilon^{-\nicefrac12}\alpha_\kappa(p_\kappa^{-1}+1)\le 2\epsilon^{-\nicefrac12}\alpha_\kappa,
\]
we deduce that
\begin{equation}\label{eq:boundary}
\|\mathsf{A}_{\DG} v-\phi^{\text{\tiny\sf{DG}}}\|_{0,\partial\kappa}
\le C_\eqref{eq:boundary}\beta_\kappa\|\mathsf{A}_{\DG} v\|_{\epsilon,\kappa},
\end{equation}
where
\begin{equation}\label{eq:beta}
\beta_\kappa:=\epsilon^{-\nicefrac14}\alpha_\kappa^{\nicefrac12}.
\end{equation}
\subsubsection{Upper \emph{A Posteriori} Residual Bound}
In order to derive an \emph{a posteriori} residual estimate for the $hp$--NDG discretisation~\eqref{eq:dgfem}, we recall the residual
\begin{align*}
\dprod{\mathsf{R}_{\eps}(u^{\DG}_{n+1}),v}
&=\int_\Omega\left\{\epsilon\nabla_{\T} u^{\DG}_{n+1}\cdot\nabla \mathsf{A}_{\DG} v+u^{\DG}_{n+1}\mathsf{A}_{\DG} v-f(u^{\DG}_{n+1})\mathsf{A}_{\DG} v\right\}\,\dd\bm x\\
&\quad+C_{\sigma}\int_{\mathcal{E}}(\epsilon\sigma+\sigma^{-1})\jmp{u^{\DG}_{n+1}}\cdot\jmp{v}\,\dd s
\equiv T_1+T_2,
\end{align*}
cf.~\eqref{eq:R}, where we define
\begin{align*}
T_1&:=\int_\Omega\left\{\epsilon\nabla_{\T} \widehat{\mathfrak{u}}^{\DG}_{n+1}\cdot\nabla \mathsf{A}_{\DG} v+\widehat{\mathfrak{u}}^{\DG}_{n+1}\mathsf{A}_{\DG} v-(f'(u^{\DG}_{n})\widehat{\mathfrak{u}}^{\DG}_{n+1}+\f{u^{\DG}_n)}\mathsf{A}_{\DG} v\right\}\,\dd\bm x\\
&\quad+ C_{\sigma} \int_{\mathcal{E}}(\epsilon\sigma+\sigma^{-1})\jmp{\widehat{\mathfrak{u}}^{\DG}_{n+1}}\cdot\jmp{v}\,\dd s,\\
T_2&:=(1-\Delta t_n) \dprod{\mathsf{R}_{\eps}(u^{\DG}_{n}),v}
+ \int_\Omega\left\{f(u^{\DG}_n)+f'(u^{\DG}_{n})(u^{\DG}_{n+1}-u^{\DG}_n)-f(u^{\DG}_{n+1})\right\}\mathsf{A}_{\DG} v.
\end{align*}
Here, $\widehat{\mathfrak{u}}^{\DG}_{n+1}$ and~$\f{u^{\DG}_n}$ are given in~\eqref{eq:hat}, and~$v\in\mathcal{W}_{\DG}$ is again arbitrary. Recalling~\eqref{eq:dgfem}, we note that
\begin{align*}
\int_\Omega&\left\{\epsilon\nabla_{\T}\widehat{\mathfrak{u}}^{\DG}_{n+1}\cdot\nabla_{\T} \phi^{\text{\tiny\sf{DG}}}+\widehat{\mathfrak{u}}^{\DG}_{n+1}\phi^{\text{\tiny\sf{DG}}}-(f'(u^{\DG}_n)\widehat{\mathfrak{u}}^{\DG}_{n+1}+\f{u^{\DG}_n})\phi^{\text{\tiny\sf{DG}}}\right\}\,\dd\bm x\\
&=\int_{\mathcal{E}}\left\{\avg{\epsilon\nabla_{\T}\widehat{\mathfrak{u}}^{\DG}_{n+1}}\cdot\jmp{\phi^{\text{\tiny\sf{DG}}}}+\theta\jmp{\widehat{\mathfrak{u}}^{\DG}_{n+1}}\cdot\avg{\epsilon\nabla_{\T} \phi^{\text{\tiny\sf{DG}}}}\right\}\,\dd s
-C_{\sigma}\int_{\mathcal{E}}\epsilon\sigma\jmp{\widehat{\mathfrak{u}}^{\DG}_{n+1}}\cdot\jmp{\phi^{\text{\tiny\sf{DG}}}}\,\dd s,
\end{align*}
with~$\phi^{\text{\tiny\sf{DG}}}\in\mathcal{V}_{\DG}$ as in Section~\ref{sc:approx} above. Therefore,
\begin{align*}
T_1&=\int_\Omega\left\{\epsilon\nabla_{\T}\widehat{\mathfrak{u}}^{\DG}_{n+1}\cdot\nabla_{\T} (\mathsf{A}_{\DG} v-\phi^{\text{\tiny\sf{DG}}})+\widehat{\mathfrak{u}}^{\DG}_{n+1}(\mathsf{A}_{\DG} v-\phi^{\text{\tiny\sf{DG}}})\right\}\,\dd\bm x\\
&\quad-\int_\Omega(f'(u^{\DG}_n)\widehat{\mathfrak{u}}^{\DG}_{n+1} +\f{u^{\DG}_n})(\mathsf{A}_{\DG} v-\phi^{\text{\tiny\sf{DG}}})\,\dd\bm x\\
&\quad+\int_{\mathcal{E}}\left\{\avg{\epsilon\nabla_{\T}\widehat{\mathfrak{u}}^{\DG}_{n+1}}\cdot\jmp{\phi^{\text{\tiny\sf{DG}}}}+\theta\jmp{\widehat{\mathfrak{u}}^{\DG}_{n+1}}\cdot\avg{\epsilon\nabla_{\T} \phi^{\text{\tiny\sf{DG}}}}\right\}\,\dd s\\
&\quad+ C_{\sigma} \int_{\mathcal{E}}(\epsilon\sigma+\sigma^{-1})\jmp{\widehat{\mathfrak{u}}^{\DG}_{n+1}}\cdot\jmp{v}\,\dd s
-C_{\sigma}\int_{\mathcal{E}}\epsilon\sigma\jmp{\widehat{\mathfrak{u}}^{\DG}_{n+1}}\cdot\jmp{\phi^{\text{\tiny\sf{DG}}}}\,\dd s.
\end{align*}
Performing elementwise integration by parts in the first integral, and proceeding as in the proof of~\cite[Theorem~3.2]{HoustonSuliWihler:08}, the following estimate can be established:
\begin{align*}
C|T_1|
&\le\sum_{\kappa\in\mathcal{T}}\|\epsilon\Delta \widehat{\mathfrak{u}}^{\DG}_{n+1}-\widehat{\mathfrak{u}}^{\DG}_{n+1}
+f'(u^{\DG}_n)\widehat{\mathfrak{u}}^{\DG}_{n+1}+\f{u^{\DG}_n}\|_{0,\kappa}\|\mathsf{A}_{\DG} v-\phi^{\text{\tiny\sf{DG}}}\|_{0,\kappa}\\
&\quad+\sum_{\kappa\in\mathcal{T}}\|\epsilon\jmp{\nabla_{\T}\widehat{\mathfrak{u}}^{\DG}_{n+1}}\|_{0,\partial\kappa\setminus\partial\Omega}\|\mathsf{A}_{\DG} v-\phi^{\text{\tiny\sf{DG}}}\|_{0,\partial\kappa}
+\left(\sum_{\kappa\in\mathcal{T}}\frac{\epsilon p_\kappa^2}{h_\kappa}\|\jmp{\widehat{\mathfrak{u}}^{\DG}_{n+1}}\|^2_{0,\partial\kappa}\right)^{\nicefrac12}\epsilon^{\nicefrac12}\|\nabla_{\T}\phi^{\text{\tiny\sf{DG}}}\|_0\\
&\quad +\left(C_{\sigma}^2\int_{\mathcal{E}}(\epsilon\sigma+\sigma^{-1})|\jmp{\widehat{\mathfrak{u}}^{\DG}_{n+1}}|^2\,\dd s\right)^{\nicefrac12}
\left(\int_{\mathcal{E}}(\epsilon\sigma+\sigma^{-1})|\jmp{v}|^2\,\dd s\right)^{\nicefrac12}\\
&\quad +\left(C_{\sigma}^2\sum_{\kappa\in\mathcal{T}}\frac{\epsilon^2 \beta_\kappa^2p_\kappa^4}{h_\kappa^2}\|\jmp{\widehat{\mathfrak{u}}^{\DG}_{n+1}}\|_{0,\partial\kappa}^2\right)^{\nicefrac12}
\left(\sum_{\kappa\in\mathcal{T}}\beta_\kappa^{-2}\|\jmp{\phi^{\text{\tiny\sf{DG}}}}\|_{0,\partial\kappa}^2\right)^{\nicefrac12}.
\end{align*}
Here, $C$ is a positive constant independent of~$\bm h$, $\bm p$, and~$\epsilon$, and~$\beta_\kappa$ is defined in~\eqref{eq:beta}. Observing that~$\jmp{\mathsf{A}_{\DG} v}=\bm 0$ on~$\mathcal{E}$, and recalling~\eqref{eq:boundary}, we infer the bound
\begin{align*}
\sum_{\kappa\in\mathcal{T}}\beta_\kappa^{-2}\|\jmp{\phi^{\text{\tiny\sf{DG}}}}\|^2_{0,\partial\kappa}
&=\sum_{\kappa\in\mathcal{T}}\beta_\kappa^{-2}\|\jmp{\phi^{\text{\tiny\sf{DG}}}-\mathsf{A}_{\DG} v}\|^2_{0,\partial\kappa}
\le C\sum_{\kappa\in\mathcal{T}}\beta_\kappa^{-2}\|\phi^{\text{\tiny\sf{DG}}}-\mathsf{A}_{\DG} v\|^2_{0,\partial\kappa}\le C\|\mathsf{A}_{\DG} v\|_{X}^2.
\end{align*}
Additionally, exploiting~\eqref{eq:interp3},
\eqref{eq:L2bound}, and~\eqref{eq:boundary}, yields
\begin{align*}
C|T_1|
&\le\sum_{\kappa\in\mathcal{T}}\|\epsilon\Delta \widehat{\mathfrak{u}}^{\DG}_{n+1}-\widehat{\mathfrak{u}}^{\DG}_{n+1}
+f'(u^{\DG}_n)\widehat{\mathfrak{u}}^{\DG}_{n+1}+\f{u^{\DG}_n}\|_{0,\kappa}\alpha_{\kappa}\|\mathsf{A}_{\DG} v\|_{\epsilon,\kappa}\\
&\quad+\sum_{\kappa\in\mathcal{T}}\|\epsilon\jmp{\nabla_{\T}\widehat{\mathfrak{u}}^{\DG}_{n+1}}\|_{0,\partial\kappa\setminus\partial\Omega}\beta_\kappa\|\mathsf{A}_{\DG} v\|_{\epsilon,\kappa}
+\left(\sum_{\kappa\in\mathcal{T}}\frac{\epsilon p_\kappa^2}{h_\kappa}\|\jmp{\widehat{\mathfrak{u}}^{\DG}_{n+1}}\|^2_{0,\partial\kappa}\right)^{\nicefrac12}\|\mathsf{A}_{\DG} v\|_{X}\\
&\quad+\left(C_{\sigma}^2\sum_{\kappa\in\mathcal{T}}\left(\frac{\epsilon p_\kappa^2}{h_\kappa}+\frac{h_\kappa}{p_\kappa^2}\right)\|\jmp{\widehat{\mathfrak{u}}^{\DG}_{n+1}}\|_{0,\partial\kappa}^2\,\dd s\right)^{\nicefrac12}\|v\|_{\text{\tiny\sf{DG}}}\\
&\quad +\left(C_{\sigma}^2\sum_{\kappa\in\mathcal{T}}\frac{\epsilon^2 \beta_\kappa^2p_\kappa^4}{h^2_\kappa}\|\jmp{\widehat{\mathfrak{u}}^{\DG}_{n+1}}\|_{0,\partial\kappa}^2\right)^{\nicefrac12}\|\mathsf{A}_{\DG} v\|_X,
\end{align*}
with~$\alpha_\kappa$ defined in~\eqref{eq:alpha}. Observing that~$\alpha_\kappa\le\epsilon^{-\nicefrac12}h_\kappa p_\kappa^{-1}$ yields
\[
\max\left(\frac{\epsilon p_\kappa^2}{h_\kappa}+\frac{h_\kappa}{p_\kappa^2},\frac{\epsilon^2 \beta_\kappa^2p_\kappa^4}{h^2_\kappa}\right)\le \frac{\epsilon p_\kappa^3}{h_\kappa}+\frac{h_\kappa}{p_\kappa^2}.
\]
Hence, applying the Cauchy-Schwarz inequality, and making use of~\eqref{eq:stab}, we arrive at
\begin{align*}
|T_1|\le C\left(\sum_{\kappa\in\mathcal{T}}\eta_{\kappa,n}^2\right)^{\nicefrac12}\|v\|_{\text{\tiny\sf{DG}}},
\end{align*}
where, for any~$\kappa\in\mathcal{T}$, we define the local residual indicators
\begin{equation}\label{eq:eta}
\begin{split}
\eta_{\kappa,n}^2:
&=\alpha_\kappa^2\|\epsilon\Delta \widehat{\mathfrak{u}}^{\DG}_{n+1}-\widehat{\mathfrak{u}}^{\DG}_{n+1}
+f'(u^{\DG}_n)\widehat{\mathfrak{u}}^{\DG}_{n+1}+\f{u^{\DG}_n}\|^2_{0,\kappa}\\
&\quad+\beta^2_\kappa\epsilon^2\|\jmp{\nabla_{\T}\widehat{\mathfrak{u}}^{\DG}_{n+1}}\|^2_{0,\partial\kappa\setminus\partial\Omega}
+\max\left(1,C_{\sigma}^2\right)\left(\frac{\epsilon p_\kappa^3}{h_\kappa}+\frac{h_\kappa}{p_\kappa^2}\right)\|\jmp{\widehat{\mathfrak{u}}^{\DG}_{n+1}}\|^2_{0,\partial\kappa}.
\end{split}
\end{equation}
In order to deal with the term~$T_2$, we apply elementwise integration by parts to obtain
\begin{align*}
\int_\Omega &\{\epsilon\nabla_{\T} u^{\DG}_n\cdot\nabla \mathsf{A}_{\DG} v+u^{\DG}_n\mathsf{A}_{\DG} v-f(u^{\DG}_n)\mathsf{A}_{\DG} v\}\,\dd\bm x\\
&=-\sum_{\kappa\in\mathcal{T}}\int_{\kappa}\{\epsilon\Deltau^{\DG}_n-u^{\DG}_n+f(u^{\DG}_n)\}\mathsf{A}_{\DG} v\,\dd\bm x
+\int_{\mathcal{E}_\mathcal{I}}\jmp{\epsilon\nTu^{\DG}_n}\mathsf{A}_{\DG} v\,\dd s.
\end{align*}
Furthermore, we define the lifting operator
\[
\mathsf{L}:\,\mathcal{V}_{\DG}\to\mathcal{V}_{\DG},\qquad w\mapsto\mathsf{L}(w),
\]
by
\[
\int_\Omega\mathsf{L}(w)\phi^{\text{\tiny\sf{DG}}}\,\dd\bm x=\int_{\mathcal{E}_\mathcal{I}}\jmp{\nabla_{\T} w}\phi^{\text{\tiny\sf{DG}}}\,\dd s
\qquad\forall\phi^{\text{\tiny\sf{DG}}}\in\mathcal{V}_{\DG};
\]
cf., e.g., \cite{ArnoldBrezziCockburnMarini:01,SchotzauSchwabToselli:02}. Thereby, we note that
\begin{align*}
\int_\Omega& \{\epsilon\nabla_{\T} u^{\DG}_n\cdot\nabla \mathsf{A}_{\DG} v+u^{\DG}_n\mathsf{A}_{\DG} v-f(u^{\DG}_n)\mathsf{A}_{\DG} v\}\,\dd\bm x\\
&=-\int_{\Omega}\{\epsilon\Delta_{\mathcal{T}}u^{\DG}_n-u^{\DG}_n+f(u^{\DG}_n)-\epsilon\mathsf{L}(u^{\DG}_n)\}\mathsf{A}_{\DG} v\,\dd\bm x\\
&\quad+\int_{\mathcal{E}_\mathcal{I}}\jmp{\epsilon\nTu^{\DG}_n}(\mathsf{A}_{\DG} v-\phi^{\text{\tiny\sf{DG}}})\,\dd s
-\int_\Omega\epsilon\mathsf{L}(u^{\DG}_n)(\mathsf{A}_{\DG} v-\phi^{\text{\tiny\sf{DG}}})\,\dd\bm x,
\end{align*}
where~$\Delta_{\mathcal{T}}$ is the elementwise Laplacian operator. Applying the Cauchy-Schwarz inequality, and incorporating the bounds from Section~\ref{sc:approx}, we deduce that
\begin{align*}
\bigg|\int_\Omega& \{\epsilon\nabla_{\T} u^{\DG}_n\cdot\nabla \mathsf{A}_{\DG} v+u^{\DG}_n\mathsf{A}_{\DG} v-f(u^{\DG}_n)\mathsf{A}_{\DG} v\}\,\dd\bm x\bigg|\\
&\le\|\epsilon\Delta_{\mathcal{T}}u^{\DG}_n-u^{\DG}_n+f(u^{\DG}_n)-\epsilon\mathsf{L}(u^{\DG}_n)\|_{0}\|\mathsf{A}_{\DG} v\|_{0}\\
&\quad
+\sum_{\kappa\in\mathcal{T}}\epsilon\|\jmp{\nTu^{\DG}_n}\|_{0,\partial\kappa\setminus\partial\Omega}
\|\mathsf{A}_{\DG} v-\phi^{\text{\tiny\sf{DG}}}\|_{0,\partial\kappa}
+\sum_{\kappa\in\mathcal{T}}\epsilon\|\mathsf{L}(u^{\DG}_n)\|_{0,\kappa}\|\mathsf{A}_{\DG} v-\phi^{\text{\tiny\sf{DG}}}\|_{0,\kappa}\\
&\le\|\epsilon\Delta_{\mathcal{T}}u^{\DG}_n-u^{\DG}_n+f(u^{\DG}_n)-\epsilon\mathsf{L}(u^{\DG}_n)\|_{0}\|\mathsf{A}_{\DG} v\|_{X}+C\sum_{\kappa\in\mathcal{T}}\beta_\kappa\epsilon\|\jmp{\nTu^{\DG}_n}\|_{0,\partial\kappa\setminus\partial\Omega}\|\mathsf{A}_{\DG} v\|_{\epsilon,\kappa}\\
&\quad+C\sum_{\kappa\in\mathcal{T}}\alpha_\kappa\epsilon\|\mathsf{L}(u^{\DG}_n)\|_{0,\kappa}\|\mathsf{A}_{\DG} v\|_{\epsilon,\kappa}\\
&\le\|\epsilon\Delta_{\mathcal{T}}u^{\DG}_n-u^{\DG}_n+f(u^{\DG}_n)-\epsilon\mathsf{L}(u^{\DG}_n)\|_{0}\|\mathsf{A}_{\DG} v\|_{X}\\
&\quad
+C\left(\sum_{\kappa\in\mathcal{T}}\left(\beta_\kappa^2\epsilon^2\|\jmp{\nTu^{\DG}_n}\|^2_{0,\partial\kappa\setminus\partial\Omega}
+\alpha_\kappa^2\epsilon^2\|\mathsf{L}(u^{\DG}_n)\|^2_{0,\kappa}\right)\right)^{\nicefrac12}\|\mathsf{A}_{\DG} v\|_{X}.
\end{align*}
Recalling~\eqref{eq:stab}, we get
\begin{align*}
\bigg|\int_\Omega& \{\epsilon\nabla_{\T} u^{\DG}_n\cdot\nabla \mathsf{A}_{\DG} v+u^{\DG}_n\mathsf{A}_{\DG} v-f(u^{\DG}_n)\mathsf{A}_{\DG} v\}\,\dd\bm x\bigg|\\
&\le\|\epsilon\Delta_{\mathcal{T}}u^{\DG}_n-u^{\DG}_n+f(u^{\DG}_n)-\epsilon\mathsf{L}(u^{\DG}_n)\|_{0}\|v\|_{\text{\tiny\sf{DG}}}\\
&\quad
+C\left(\sum_{\kappa\in\mathcal{T}}\left(\beta_\kappa^2\epsilon^2\|\jmp{\nTu^{\DG}_n}\|^2_{0,\partial\kappa\setminus\partial\Omega}
+\alpha_\kappa^2\epsilon^2\|\mathsf{L}(u^{\DG}_n)\|^2_{0,\kappa}\right)\right)^{\nicefrac12}\|v\|_{\text{\tiny\sf{DG}}}.
\end{align*}
Furthermore, we have
\begin{align*}
\bigg|C_{\sigma}\int_{\mathcal{E}}(\epsilon\sigma+\sigma^{-1})\jmp{u^{\DG}_{n}}\cdot\jmp{v}\,\dd s\bigg|
&\le C\left(\sum_{\kappa\in\mathcal{T}}C_{\sigma}^2\left(\frac{\epsilon p_\kappa^2}{h_\kappa}+\frac{h_\kappa}{p_\kappa^2}\right)\|\jmp{u^{\DG}_{n}}\|^2_{0,\partial\kappa}\right)^{\nicefrac12}\|v\|_{\text{\tiny\sf{DG}}}.
\end{align*}
Thus, in summary, we can bound~$T_2$ by
\[
|T_2|
\le C\delta_{n,\Omega}\|v\|_{\text{\tiny\sf{DG}}},
\]
where
\begin{equation}\label{eq:delta1}
\delta_{n,\Omega}:=(1-\Delta t_n)\delta^{(1)}_{n,\Omega}+\delta^{(2)}_{n,\Omega},
\end{equation}
with
\begin{equation}\label{eq:delta2}
\begin{split}
\delta^{(1)}_{n,\Omega}
&:=\|\epsilon\Delta_{\mathcal{T}}u^{\DG}_n-u^{\DG}_n+f(u^{\DG}_n)-\epsilon\mathsf{L}(u^{\DG}_n) \|_{0} \\
&\quad
+ \left(\sum_{\kappa\in\mathcal{T}} \epsilon^2\alpha_\kappa\left(\epsilon^{-\nicefrac12}\|\jmp{\nTu^{\DG}_n}\|^2_{0,\partial\kappa\setminus\partial\Omega} +
\alpha_\kappa\|\mathsf{L}(u^{\DG}_n)\|^2_{0,\kappa}\right) \right)^{\nicefrac12} \\
&\quad
+C_{\sigma} \left(\sum_{\kappa\in\mathcal{T}} \left(\frac{\epsilon p_\kappa^2}{h_\kappa}+\frac{h_\kappa}{p_\kappa^2}\right)\|\jmp{u^{\DG}_{n}}\|^2_{0,\partial\kappa}\right)^{\nicefrac12},
\end{split}
\end{equation}
and
\begin{equation}\label{eq:delta3}
\delta^{(2)}_{n,\Omega}:=\|f(u^{\DG}_n)+f'(u^{\DG}_{n})(u^{\DG}_{n+1}-u^{\DG}_n)-f(u^{\DG}_{n+1})\|_{0}.
\end{equation}
Thus we have proved the following key result.
\begin{theorem}\label{thm:apost}
For the $hp$--NDG method~\eqref{eq:dgfem}, the following upper \emph{a posteriori} residual bound holds
\[
\NNN{\mathsf{R}_{\eps}(u^{\DG}_{n+1})}\le {\mathcal E}(u^{\DG}_n,u^{\DG}_{n+1},{\bm h},{\bm p}) \equiv
C\left(\delta_{n,\Omega}^2+\sum_{\kappa\in\mathcal{T}}\eta_{\kappa,n}^2\right)^{\nicefrac12},
\]
where $C$ is a positive constant, independent of~$\bm h$, $\bm p$, the penalty parameter~$C_{\sigma}$, and~$\epsilon$. Moreover,~$\eta_{\kappa,n}$, $\kappa\in\mathcal{T}$, and~$\delta_{n,\Omega}$ are given in~\eqref{eq:eta} and~\eqref{eq:delta1}--\eqref{eq:delta3}, respectively.
\end{theorem}
\begin{remark}
Following along the lines of~\cite[\S4.4.2]{AmreinWihler:15} and~\cite{HoustonSchotzauWihler:07}, it is possible to prove local lower residual bounds in terms of the error indicators~$\eta_{\kappa}$, $\kappa\in\mathcal{T}$, and some data oscillation terms. In contrast to the $h$--version approach in~\cite{AmreinWihler:15}, however, the local efficiency bounds will be slightly suboptimally scaled with respect to the local polynomial degrees due to the need of applying $p$--dependent norm equivalence results (involving cut-off functions).
\end{remark}
\section{$hp$--Adaptive NDG Scheme}
\label{sec:adaptivity}
In this section, we will discuss how the \emph{a posteriori} bound from Theorem~\ref{thm:apost} can be exploited in the design of an $hp$--adaptive NDG algorithm for the numerical approximation of~\eqref{eq:PDE}.
\subsection{$hp$--Adaptive Refinement Procedure}\label{sc:hpadapt}
In order to enrich the finite element space $\mathcal{V}_{\DG}$, we shall apply an
$hp$--adaptive refinement algorithm which is based on the following two ingredients:
\subsubsection*{(a) Element marking:}
Each element $\kappa$ in the computational mesh $\mathcal{T}$ may be marked for
refinement on the basis of the size of the local residual indicators $\eta_{\kappa,n}$,
cf. \eqref{eq:eta}, $n\geq 0$. To this end, several strategies, such as
equidistribution, fixed fraction, D\"{o}rfler marking, optimized mesh criterion, and so on,
cf. \cite{HoustonSuli02}, for example, have been proposed within the literature.
For the purposes of this article, we employ the {\em maximal strategy}: here,
we refine the set of elements $\kappa\in \mathcal{T}$ which satisfy the
condition
\[
\eta_{\kappa,n} > \Upsilon \max_{\kappa\in\mathcal{T}} \eta_{\kappa,n},
\]
where $0< \Upsilon <1$ is a given parameter. On the basis of \cite{De07,opac-b1101124,Houston2016977},
throughout this article, we set $\Upsilon = \nicefrac13$.
\subsubsection*{(b) $hp$--Refinement criterion:}
Once an element $\kappa\in {\mathcal T}$ has been marked for refinement,
a decision must be made regarding whether to subdivide the element
($h$--refinement) or to increase the local degree of the polynomial
approximation on element $\kappa$ ($p$--refinement). Several strategies have
been proposed within the literature; for a recent review of
$hp$--refinement algorithms, we refer to~\cite{MitchellMcClain:14}.
Here we employ the $hp$--refinement strategy developed in \cite{HoustonSuli:05}
where the local regularity of the analytical solution is estimated on the
basis of truncated local Legendre expansions of the computed numerical solution,
cf., also, \cite{FankhauserWihlerWirz:14,EibnerMelenk:07}.
\subsection{Fully Adaptive Newton-Galerkin Method}
We now propose a procedure that provides an \emph{interplay} of the Newton linearisation and automatic $hp$--finite element mesh refinements based on the {\em a posteriori} residual estimate from Theorem~\ref{thm:apost} (as outlined in the previous Section~\ref{sc:hpadapt}). To this end, we make the assumption that the NDG sequence $\left\{u^{\DG}_{n+1}\right\}_{n\ge 0} $ given by~\eqref{eq:dgfem} is well-defined as long as the iterations are being performed.
\begin{algorithm}\label{al:full}
Given a (coarse) starting mesh $\mathcal{T}$ in~$\Omega$, with an associated (low-order) polynomial degree distribution~$\bm p$, and an initial guess $ u^{\DG}_{0} \in \mathcal{V}_{\DG} $. Set~$n\gets 0$.
\begin{algorithmic}[1]
\State
Determine the Newton step size parameter $\Delta t_n$ based on~$u^{\DG}_n$ by the adaptive procedure from Algorithm~\ref{al:zs}; the Newton-Raphson transform $\mathsf{NF}(u^{\DG}_n)$ required for the computation of the step size parameter~$\Delta t_n$ is approximated using the $hp$--DG method on the current mesh.
\State
Compute the DG solution~$\widehat{\mathfrak{u}}^{\DG}_{n+1}$ from~\eqref{eq:dgfem}, and~$u^{\DG}_{n+1}=\widehat{\mathfrak{u}}^{\DG}_{n+1}+(1-\Delta t_n) u^{\DG}_n$. Furthermore, evaluate the corresponding residual indicators $ \{\eta_{\kappa,n}\}_{\kappa\in\mathcal{T}} $, and $\delta_{n,\Omega} $ from~\eqref{eq:eta} and~\eqref{eq:delta1}--\eqref{eq:delta2}, respectively.
\If {
\begin{equation}\label{eq:test}
\delta_{n,\Omega}^2\le \Lambda \sum_{\kappa\in\mathcal{T}}{\eta_{\kappa,n}^2}
\end{equation}
holds, for some given parameter~$\Lambda >0$,} {$hp$--refine the space $\mathcal{V}_{\DG}$ adaptively based on the marking criterion and the $hp$--strategy outlined in Section~\ref{sc:hpadapt}; go back to step~({\footnotesize\sc 1:}) with the new mesh~$\mathcal{T}$ (and based on the previously computed solution~$u^{\DG}_{n+1}$ interpolated on the refined mesh).}
\Else{, i.e., if~\eqref{eq:test} is not fulfilled, then set~$n\leftarrow n+1$, and perform another Newton step by going back to~({\footnotesize\sc 1:}).}
\EndIf
\end{algorithmic}
\end{algorithm}
\begin{remark}
We note that our computational experience suggests that the choice of the element marking strategy can directly affect the robustness of the NDG scheme, particularly, when the numerical solution is far away from a given solution. Indeed, it is essential to employ a marking scheme which adaptively adjusts the number of elements marked for refinement at each step of the adaptive process; algorithms such as the fixed fraction method which only mark a fixed percentage of elements at each refinement level can lead
to slow convergence of the combined adaptive Newton-Galerkin approach.
\end{remark}
\section{Numerical Experiments} \label{sec:numerics}
In this section we present a series of numerical experiments to demonstrate the
practical performance of the proposed $hp$--adaptive refinement strategy
outlined in Algorithm~\ref{al:full}. To this end, throughout this section
we select $\tau=0.1$ and $\gamma=0.5$ in Algorithm~\ref{al:zs}, the penalty
parameter $C_{\sigma}=10$ and $\theta=1$ (SIPG) in the interior penalty
DG scheme \eqref{eq:dgfem}, cf. \eqref{eq:a}, and
$\Lambda = 0.5$ in Algorithm~\ref{al:full}, cf.~\cite{AmreinWihler:15}.
Throughout this section we shall compare the performance of the
proposed $hp$--adaptive refinement strategy with the corresponding
algorithm based on exploiting only local mesh subdivision, i.e.,
$h$--refinement.
Furthermore, within each inner linear iteration, we employ the \emph{direct} MUltifrontal Massively Parallel Solver (MUMPS) \cite{MUMPS:2,MUMPS:1,MUMPS:3}; in particular, in Theorem~\ref{thm:apost}, we do not take into account any linear algebra errors resulting from iterative solvers (cf., e.g., \cite{El-AlaouiErnVohralik:11}).
\begin{figure}[t!]
\begin{center}
\includegraphics[scale=0.5]{bratu_soln_slices}
\end{center}
\caption{Bratu Problem. Slice at $y=0.5$, $0\leq x\leq 1$, of the upper
and lower solutions computed
with $\epsilon=1$ and $\epsilon = 0.5$, together with the critical solution
($\epsilon=\epsilon_c$).}
\label{fig:bratu_slices}
\end{figure}
\begin{figure}[t!]
\begin{center}
\begin{tabular}{cc}
\includegraphics[scale=0.19]{bratu_upper_eps1} &
\includegraphics[scale=0.14]{bratu_upper_eps_pt_5} \\
(a) & (b)
\end{tabular}
\end{center}
\caption{Bratu Problem. Upper solution computed with:
(a) $\epsilon=1$; (b) $\epsilon = 0.5$.}
\label{fig:bratu_solutions}
\end{figure}
\begin{example}
In this first example, we consider the Bratu problem
\[
\epsilon\Delta u + {\rm e}^u = 0 \quad \mbox{ in } (0,1)^2,
\]
i.e., $f(u)={\rm e}^u+u$, subject to homogeneous Dirichlet boundary conditions on
$\partial\Omega$. Writing $\lambda = \nicefrac{1}{\epsilon}$, we recall that there
exists a critical parameter value $\lambda_c$ $(=1/\epsilon_c)$, such that for
$\lambda > \lambda_c$ $(\epsilon < \epsilon_c)$ the problem has no solution, for
$\lambda = \lambda_c$ $(\epsilon = \epsilon_c)$ there exists exactly one
solution, and for $\lambda < \lambda_c$ $(\epsilon > \epsilon_c)$ there are two solutions.
In the one--dimensional
setting, an analytical expression for $\lambda_c$ is available, cf.
\cite{aschermattheijhrussell,CHHPSbratu,calvetti_2000}; for the two--dimensional case,
calculations have revealed that $\lambda_c = 6.808124423$
$(\epsilon_c=0.146883332)$ to 9 decimal places,
see \cite{CHHPSbratu,mohsen08,mohsen_2014}, and the references cited therein.
\begin{figure}[t!]
\begin{center}
\begin{tabular}{cc}
\includegraphics[scale=0.38]{bratu_one_lower_convergence} &
\includegraphics[scale=0.38]{bratu_one_upper_convergence} \\
(a) & (b) \\
~ & ~ \\
\includegraphics[scale=0.38]{bratu_eps_pt_5_lower_convergence} &
\includegraphics[scale=0.38]{bratu_eps_pt_5_upper_convergence} \\
(c) & (d) \\
~ & ~ \\
\multicolumn{2}{c}{\includegraphics[scale=0.38]{bratu_critical_convergence}} \\
\multicolumn{2}{c}{(e)}
\end{tabular}
\end{center}
\caption{Bratu Problem. Comparison between $h$-- and $hp$--refinement.
(a)~$\epsilon=1$ (lower solution);
(b) $\epsilon=1$ (upper solution);
(c) $\epsilon=\nicefrac{1}{2}$ (lower solution);
(d) $\epsilon=\nicefrac{1}{2}$ (upper solution);
(e) $\epsilon=\epsilon_c$ (critical solution); }
\label{fig:bratu_convergence}
\end{figure}
\begin{figure}[t!]
\begin{center}
\begin{tabular}{cc}
\includegraphics[scale=0.38]{bratu_one_upper_h_delta_t} &
\includegraphics[scale=0.38]{bratu_one_upper_hp_delta_t} \\
(a) & (b) \\
~ & ~\\
\includegraphics[scale=0.38]{bratu_eps_pt_5_upper_h_delta_t} &
\includegraphics[scale=0.38]{bratu_eps_pt_5_upper_hp_delta_t} \\
(c) & (d) \\
~ & ~\\
\includegraphics[scale=0.38]{bratu_eps_pt_5_lower_h_delta_t} &
\includegraphics[scale=0.38]{bratu_eps_pt_5_lower_hp_delta_t} \\
(e) & (f)
\end{tabular}
\end{center}
\caption{Bratu Problem. Damping parameter $\Delta t_n$.
Left: $h$--refinement; right: $hp$--refinement.
(a) \& (b) $\epsilon=1$ (upper solution);
(c) \& (d) $\epsilon=\nicefrac{1}{2}$ (upper solution);
(e) \& (f) $\epsilon=\nicefrac{1}{2}$ (lower solution).}
\label{fig:bratu_deltat}
\end{figure}
Following \cite{mohsen_2014}, we select the initial guess $ u^{\DG}_{0} \in \mathcal{V}_{\DG} $
to be the $L^2$--projection of the function $u_0$ onto $\mathcal{V}_{\DG}$, where
$$
u_0 = a \sin(\pi x)\sin(\pi y)
$$
and $a$ is a given amplitude. Noting that the maximum amplitude of the critical solution
computed with $\epsilon = \epsilon_c$ is approximately $1.39$, selecting $a$ to be
smaller/larger than this value leads to convergence to the so--called
lower/upper solution, respectively. With this in mind we select
$a=2$ when $\epsilon = \epsilon_c$, $a\in\{\nicefrac{1}{10},6\}$ for
$\epsilon = 1$, and $a\in\{1,4\}$ for $\epsilon=\nicefrac{1}{2}$; in the latter
two cases the smaller value of $a$ is employed for the computation
of the lower solution, while the larger value ensures convergence to
the upper solution.
In Figure~\ref{fig:bratu_slices} we plot a slice of each of the computed
numerical solutions at $y=0.5$, $0\leq x\leq 1$. Here, we observe that the
lower solutions tend to be rather flat in profile, while the upper solutions
have a stronger peak in the middle of the computational domain, cf.,
also, Figure~\ref{fig:bratu_solutions}.
\begin{figure}[t!]
\begin{center}
\begin{tabular}{cc}
\includegraphics[scale=0.4]{bratu_hmesh_upper_eps1} &
\includegraphics[scale=0.39]{bratu_hpmesh_upper_eps1} \\
(a) & (b) \\
\includegraphics[scale=0.4]{bratu_hmesh_upper_eps_pt5} &
\includegraphics[scale=0.39]{bratu_hpmesh_upper_eps_pt5} \\
(c) & (d) \\
\includegraphics[scale=0.4]{bratu_hmesh_critical} &
\includegraphics[scale=0.39]{bratu_hpmesh_critical} \\
(e) & (f)
\end{tabular}
\end{center}
\caption{Bratu Problem. Computational meshes.
Left: $h$--refinement; right: $hp$--refinement.
(a) \& (b) Upper solution computed with $\epsilon=1$;
(c) \& (d) Upper solution computed with $\epsilon = 0.5$.
(e) \& (f) Critical solution.}
\label{fig:bratu_meshes}
\end{figure}
In Figure~\ref{fig:bratu_convergence} we demonstrate the performance of the
proposed $hp$--adaptive NDG algorithm, cf. Algorithm~\ref{al:full},
for the computation of the lower and upper solutions when
$\epsilon=1$ and $\epsilon=\nicefrac{1}{2}$, as well as for the numerical
approximation of the critical solution when $\epsilon = \epsilon_c$.
In each case we plot the residual estimator
${\mathcal E}={\mathcal E}(u^{\DG}_n,u^{\DG}_{n+1},{\bm h},{\bm p})$ versus
the square root of the number of degrees of freedom in the finite element
space $\mathcal{V}_{\DG}$, based on employing both $h$-- and $hp$--refinement.
For each parameter value we observe that the $hp$--refinement algorithm
leads to an exponential decay of the residual estimator ${\mathcal E}$ as the
finite element space $\mathcal{V}_{\DG}$ is adaptively enriched: on a linear-log
plot, the convergence lines are roughly straight. Moreover, we observe the
superiority of $hp$--refinement in comparison with a standard
$h$--refinement algorithm, in the sense that the former refinement strategy
leads to several orders of magnitude reduction in ${\mathcal E}$,
for a given number of degrees of freedom, than the corresponding
quantity computed exploiting mesh subdivision only.
In Figure~\ref{fig:bratu_deltat} we plot the size of the Newton damping
$\Delta t_n$ versus the global iteration number. In many of the cases
considered here $\Delta t_n=1$ at all steps; for brevity, these results
have been omitted. For the cases presented in Figure~\ref{fig:bratu_deltat},
we observe that initially the damping parameter slowly increases when we are
far away from the solution; once the damping parameter is close to unity,
the condition
$$
\delta_{n,\Omega}^2\le \Lambda \sum_{\kappa\in\mathcal{T}}{\eta_{\kappa,n}^2}
$$
in Algorithm~\ref{al:full} becomes fulfilled in which case the
finite element space $\mathcal{V}_{\DG}$ is adaptively enriched. In some cases,
particularly at the early stages of the algorithm, refinement
of $\mathcal{V}_{\DG}$ may then lead to a reduction in $\Delta t_n$, in which
case further Newton steps are required before the next refinement can be undertaken. As
the iterates approach the solution more closely, the size of the damping parameter typically remains approximately~1.
Finally, in Figure~\ref{fig:bratu_meshes} we show the $h$-- and $hp$--refined
meshes generated for the numerical approximation of the upper solutions
when $\epsilon=1$ and $\epsilon=\nicefrac{1}{2}$, as well as for the
critical solution. Here we observe that when $h$--refinement is employed,
the mesh is concentrated in the vicinity of the peak in the solution
located at the centre of the computational domain, cf.
Figures~\ref{fig:bratu_slices} \&~\ref{fig:bratu_solutions}.
In the $hp$--setting, we observe that while some mesh refinement
has been undertaken in the centre of the domain $\Omega$, the corners
of $\Omega$ have been significantly refined in order to resolve corner singularities typical for elliptic problems. Moreover, $p$--enrichement
has been employed both in these corner regions, as well as in the vicinity
of the peak in the computed solution. The corresponding meshes for the lower
solutions are largely uniformly refined, due to the flat nature of the
solution; for brevity, these have been omitted.
\end{example}
\begin{figure}[t!]
\begin{center}
\begin{tabular}{cc}
\includegraphics[scale=0.12]{ginzburg_soln_eps_m3} &
\includegraphics[scale=0.12]{ginzburg_soln_eps_m6} \\
(a) & (b)
\end{tabular}
\end{center}
\caption{Ginzburg-Landau equation. Solution computed with:
(a) $\epsilon=10^{-3}$; (b) $\epsilon = 10^{-6}$.}
\label{fig:ginzburg_solutions}
\end{figure}
\begin{example}
In this example, we consider the
Ginzburg-Landau equation given by
\[
-\epsilon\Delta u+u=u(2-u^2) \quad \mbox{ in } (-1,1)^2,
\]
subject to homogeneous Dirichlet boundary conditions on $\partial\Omega$.
Following \cite{AmreinWihler:15}, we first note that $u\equiv 0$ is a
solution; moreover, any solution $u$ appears in a pairwise fashion
as $-u$. In the absence of boundary conditions, it is clear that $u=\pm 1$
are solutions of the Ginzburg-Landau equation. Thereby, in the presence
of homogeneous Dirichlet boundary conditions, boundary layers will arise
in the vicinity of $\partial\Omega$, whose width will be governed by
the size of the diffusion coefficient $\epsilon$. Here,
we select the initial guess $ u^{\DG}_{0} \in \mathcal{V}_{\DG} $
to be the $L^2$--projection of the function $u_0(x,y) = -\mbox{sgn}(x)$ onto $\mathcal{V}_{\DG}$,
subject to the enforcement of the boundary conditions. In this case
the solution to the Ginzburg-Landau equation will possess not only
boundary layers, but also an internal layer along $x=0$;
in Figure~\ref{fig:ginzburg_solutions} we plot the solution computed
with both $\epsilon=10^{-3}$ and $\epsilon=10^{-6}$.
\begin{figure}[t!]
\begin{center}
\begin{tabular}{cc}
\includegraphics[scale=0.38]{ginzburg_landau_semilog_conv_eps_m3} &
\includegraphics[scale=0.38]{ginzburg_landau_semilog_conv_eps_m4} \\
(a) & (b) \\
~ & ~ \\
\includegraphics[scale=0.38]{ginzburg_landau_semilog_conv_eps_m5} &
\includegraphics[scale=0.38]{ginzburg_landau_semilog_conv_eps_m6} \\
(c) & (d)
\end{tabular}
\end{center}
\caption{Ginzburg-Landau equation. Comparison between $h$-- and $hp$--refinement.
(a)~$\epsilon=10^{-3}$;
(b) $\epsilon=10^{-4}$;
(c) $\epsilon=10^{-5}$;
(d) $\epsilon=10^{-6}$.}
\label{fig:ginzburg_convergence}
\end{figure}
In Figure~\ref{fig:ginzburg_convergence} we demonstrate the performance of the
proposed $hp$--adaptive NDG algorithm, cf. Algorithm~\ref{al:full},
for the computation of the solution to the Ginzburg-Landau equation when
$\epsilon=10^{-3},~10^{-4},~10^{-5},~10^{-6}$.
In each case we plot the residual estimator
${\mathcal E}$ versus
the square root of the number of degrees of freedom in the finite element
space $\mathcal{V}_{\DG}$, based on employing both $h$-- and $hp$--refinement.
For each value of $\epsilon$ we again observe that the $hp$--refinement algorithm
leads to an exponential decay of the residual estimator ${\mathcal E}$ as the
finite element space $\mathcal{V}_{\DG}$ is adaptively enriched. Moreover, we again
observe the superiority of exploiting $hp$--refinement in comparison with a standard
$h$--refinement algorithm, in the sense that the former refinement strategy
leads to several orders of magnitude reduction in ${\mathcal E}$,
for a given number of degrees of freedom, than the corresponding
quantity computed using $h$--refinement only. Furthermore, we note
that as $\epsilon$ is reduced, additional $h$--enrichment of the computational
mesh is required before $p$--refinement is employed. Indeed, for $\epsilon=10^{-6}$
we observe that there is an initial transient, before the $hp$--version
convergence line becomes straight and exponential convergence is observed.
In Figure~\ref{fig:ginzburg_deltat_eps_m3} we plot
$\Delta t_n$ versus the global iteration number for $\epsilon=10^{-3}$;
for the other values of $\epsilon$ considered here, the damping parameter
was close to one on all of the meshes considered. As in the previous
example, we again see an initial increase in $\Delta t_n$ as the
adaptive Newton algorithm proceeds, before the underlying mesh
is adaptively refined. Again, in the early stages of the algorithm,
enrichment of $\mathcal{V}_{\DG}$ may lead to some additional damping, before
$\Delta t_n$ tends to one.
Finally, in Figure~\ref{fig:ginzburg_meshes} we plot the
corresponding $h$-- and $hp$--meshes generated for
$\epsilon=10^{-3}$ and $\epsilon=10^{-6}$. Here, we clearly observe
that the boundary and internal layers present in the analytical
solution are refined by our adaptive mesh adaptation strategy; in particular, we emphasise that the NDG iterates converge to a solution which features the same topology as the initial guess, and, hence, does not switch between various attractors (corresponding to different solutions; see, e.g, \cite{AmreinWihler:15}). In the $hp$--setting, we see that once the $h$--mesh has been sufficiently refined, then $p$--enrichment is employed.
\begin{figure}[t!]
\begin{center}
\begin{tabular}{cc}
\includegraphics[scale=0.38]{ginzburg_landau_h_delta_t_eps_m3} &
\includegraphics[scale=0.38]{ginzburg_landau_hp_delta_t_eps_m3} \\
(a) & (b)
\end{tabular}
\end{center}
\caption{Ginzburg-Landau equation. Damping parameter $\Delta t_n$ for $\epsilon=10^{-3}$.
(a) $h$--refinement;
(b) $hp$--refinement.}
\label{fig:ginzburg_deltat_eps_m3}
\end{figure}
\begin{figure}[t!]
\begin{center}
\begin{tabular}{cc}
\includegraphics[scale=0.4]{ginzburg_hmesh_eps_m3} &
\includegraphics[scale=0.39]{ginzburg_hpmesh_eps_m3} \\
(a) & (b) \\
\includegraphics[scale=0.4]{ginzburg_hmesh_eps_m6} &
\includegraphics[scale=0.39]{ginzburg_hpmesh_eps_m6} \\
(c) & (d)
\end{tabular}
\end{center}
\caption{Ginzburg-Landau equation. Computational meshes.
Left: $h$--refinement; right: $hp$--refinement.
(a) \& (b) $\epsilon=10^{-3}$;
(c) \& (d) $\epsilon = 10^{-6}$.}
\label{fig:ginzburg_meshes}
\end{figure}
\end{example}
\section{Concluding remarks}
\label{sec:conlusions}
In this article we have introduced the $hp$--version of the NDG scheme for the
numerical approximation of second-order, singularly perturbed, semilinear
elliptic boundary value problems. Here, the general approach is based on first
linearising the underlying PDE problem on a continuous level, followed by
subsequent discretisation of the resulting sequence of linear PDEs. For this latter
task, in the current article we have exploited the $hp$--version of the
interior penalty DG method. Furthermore, we have derived an $\epsilon$-robust
{\em a posteriori} bound which takes into account both the linearisation
and discretisation errors. On the basis of this residual estimate, we have
designed and implemented an $hp$--adaptive refinement algorithm
which automatically controls both of these sources of error; the
practical performance of this strategy has been studied for a series
of numerical test problems. Future work will be devoted to the
extension of this technique to more general nonlinear PDE problems,
as well as to problems in three dimensions.
\bibliographystyle{amsplain}
|
\section*{Acknowledgement}
We would like to thank Sayantani Bhattacharyya, Suvankar Dutta and Urjit Yajnik for various helpful discussions. We thank Utkarsh Sharma for the collaborations in the initial stages of this project. LKJ would like to thank organizers of \textit{Indian Strings Meet-14, Puri, India} and \textit{Applications of AdS/CFT to QCD and CMT, CRM Montreal, Canada}, where parts of this work were discussed.
\section*{References}
\bibliographystyle{unsrt}
|
\section{Introduction}
\label{sec-intro}
Let $X\subset S^3$ be a smoothly embedded closed surface of genus
$g>1$. As we will explain, $X$ automatically has a very rich and
rigid geometric structure. Indeed, $X$ inherits a Riemannian metric
from $S^3$, and after specifying some conventions we also obtain a
well-defined orientation. Now consider a point $x\in X$, and let
$T_xX$ denote the corresponding tangent space. Let $J_x\:T_xX\to
T_xX$ be the anticlockwise rotation through $\pi/2$ (which is
meaningful given the metric and orientation). This depends smoothly
on $x$ and satisfies $J_x^2=-1$, so it gives an almost complex
structure on $X$. It has been known since the early twentieth century
that any almost complex structure on a manifold of real dimension two
integrates to give a genuine complex structure. Thus, $X$ can be
regarded as a compact Riemann surface. It is known that any compact
Riemann surface can be regarded as a projective algebraic variety over
$\C$, and also as a branched cover of the Riemann sphere.
Alternatively, as we have assumed that the genus is larger than one,
the universal cover of $X$ is conformally equivalent to the open unit
disc $\Dl$. This means that $X$ is conformally equivalent to the
quotient $\Dl/\Pi$ for some Fuchsian group $\Pi$.
To the best of our knowledge, the literature contains no examples
where a significant fraction of this structure can be made explicit.
This monograph is a partially successful attempt to provide such
an example, involving the surface
\[ EX^* = \{x\in S^3\st
(3x_3^2-2)x_4+\rt(x_1^2-x_2^2)x_3=0
\},
\]
with weaker results for a one-parameter family of surfaces in which
$EX^*$ appears. To display $EX^*$ visually, we apply the
stereographic projection map $s\:S^3\to\R^3\cup\{\infty\}$ defined as
follows:
\begin{align*}
s(x) &=
\left(\frac{x_1}{1-x_4},\frac{x_2}{1-x_4},\frac{x_3}{1-x_4}\right)
\\
s^{-1}(u) &=
\left(\frac{2u_1}{\|u\|^2+1},
\frac{2u_2}{\|u\|^2+1},
\frac{2u_3}{\|u\|^2+1},
\frac{\|u\|^2-1}{\|u\|^2+1}\right).
\end{align*}
The image $s(EX^*)$ looks like this:
\[ \includegraphics[scale=0.4,angle=90,clip=true,
trim=10cm 8cm 10cm 8cm]{images/EX.jpg}
\]
Our work is organised around the following
definitions:
\begin{definition}\label{defn-G}
Let $G$ be the group of order $16$ generated by $\lm$, $\mu$ and
$\nu$ subject to relations
\[ \lm^4=\mu^2=\nu^2=(\mu\nu)^2=(\lm\mu)^2=(\lm\nu)^2=1, \]
so
\[ G=\{\lm^i\mu^j\nu^k\st 0\leq i<4,\;0\leq j,k<2\}. \]
We use the following notation for subgroups:
\begin{align*}
D_8 &= \ip{\lm,\mu} &
C_4 &= \ip{\lm} \\
C_2 &= \ip{\lm^2} &
C'_2 &= \ip{\mu\nu}.
\end{align*}
\end{definition}
\begin{definition}\label{defn-V-star}
We write $V^*$ for the set $\{0,\dotsc,13\}$ equipped with the action
of $G$ by the following permutations:
\begin{align*}
\lm &\mapsto (2\;3\;4\;5)\;(6\;7\;8\;9)\;(10\;11)\;(12\;13) \\
\mu &\mapsto (0\;1)\;(3\;5)\;(6\;9)\;(7\;8)\;(10\;12)\;(11\;13) \\
\nu &\mapsto (3\;5)\;(6\;9)\;(7\;8).
\end{align*}
\end{definition}
\begin{remark}\label{rem-V-star}
The orbits in $V^*$ are
\begin{align*}
\{0,1\} & \simeq G/\ip{\lm,\nu} \\
\{2,3,4,5\} & \simeq G/\ip{\mu,\nu} \\
\{6,7,8,9\} & \simeq G/\ip{\lm\mu,\lm\nu} \\
\{10,11,12,13\} & \simeq G/\ip{\lm^2,\nu}.
\end{align*}
The action can be displayed as follows:
\begin{center}
\begin{tikzpicture}[scale=1.7,auto,shorten >= 1pt]
\node (v0) at ( 0, 1) [circle,draw] {$\ss 0$};
\node (v1) at ( 0, 0) [circle,draw] {$\ss 1$};
\node (v2) at ( 1, 1) [circle,draw] {$\ss 2$};
\node (v3) at ( 2, 1) [circle,draw] {$\ss 3$};
\node (v4) at ( 2, 0) [circle,draw] {$\ss 4$};
\node (v5) at ( 1, 0) [circle,draw] {$\ss 5$};
\node (v6) at ( 3, 1) [circle,draw] {$\ss 6$};
\node (v7) at ( 4, 1) [circle,draw] {$\ss 7$};
\node (v8) at ( 4, 0) [circle,draw] {$\ss 8$};
\node (v9) at ( 3, 0) [circle,draw] {$\ss 9$};
\node (va) at ( 5, 1) [circle,draw] {$\ss {10}$};
\node (vb) at ( 5, 0) [circle,draw] {$\ss {11}$};
\node (vc) at ( 6, 1) [circle,draw] {$\ss {12}$};
\node (vd) at ( 6, 0) [circle,draw] {$\ss {13}$};
\draw[->,red,thick] (v2) to (v3);
\draw[->,red,thick] (v3) to (v4);
\draw[->,red,thick] (v4) to (v5);
\draw[->,red,thick] (v5) to (v2);
\draw[->,red,thick] (v6) to (v7);
\draw[->,red,thick] (v7) to (v8);
\draw[->,red,thick] (v8) to (v9);
\draw[->,red,thick] (v9) to (v6);
\draw[<->,red,thick] (va) to (vb);
\draw[<->,red,thick] (vc) to (vd);
\draw[<->,olivegreen,thick,dotted] (v3) to (v5);
\draw[<->,olivegreen,thick,bend left,dotted] (v6) to (v9);
\draw[<->,olivegreen,thick,bend right,dotted] (v7) to (v8);
\draw[<->,blue,thick,dashed] (v0) to (v1);
\draw[<->,blue,thick,dashed] (va) to (vc);
\draw[<->,blue,thick,dashed] (vb) to (vd);
\end{tikzpicture}
\end{center}
The solid red arrows show the action of $\lm$, the dotted green
arrows show the action of $\nu$, and the dashed blue arrows show the
action of $\mu\nu$.
\end{remark}
\begin{definition}\label{defn-precromulent}
A \emph{precromulent surface} is a compact Riemann surface $X$ of genus
$2$ with an action of $G$ such that
\begin{itemize}
\item[(a)] The elements of $D_8$ act conformally, and the elements
of $G\sm D_8$ act anticonformally.
\item[(b)] The set $V=\{v\in X\st \stab_{D_8}(v)\neq 1\}$ is
isomorphic to $V^*$ as a $G$-set.
\end{itemize}
A \emph{precromulent labelling} of $X$ is a specific choice of
isomorphism $V^*\simeq V$, or equivalently, a listing of the points
in $V$ as $v_0,\dotsc,v_{13}$ such that $G$ permutes these points in
accordance with the permutations listed in
Definition~\ref{defn-V-star}. A \emph{cromulent labelling} is a
precromulent labelling such that
\begin{itemize}
\item[(c)] $\lm$ acts on the tangent space $T_{v_0}X$ as
multiplication by $i$.
\item[(d)] In the set $X'=\{x\in X\st\stab_G(x)=1\}$, there is a
connected component $F'$ whose closure contains
$\{v_0,v_3,v_6,v_{11}\}$.
\end{itemize}
We will show in Proposition~\ref{prop-labellings} that every
precromulent surface has precisely two cromulent labellings, which
are exchanged by the action of $\lm^2$. A \emph{cromulent surface}
is a precromulent surface with a choice of cromulent labelling.
A \emph{(precromulent) isomorphism} between precromulent surfaces
will mean a $G$-equivariant conformal isomorphism. A
\emph{(cromulent) isomorphism} between cromulent surfaces
will mean a $G$-equivariant conformal isomorphism that is compatible
with the specified labellings.
\end{definition}
Now fix a parameter $a\in (0,1)$. We put
\[ EX(a) = \{x\in\R^4\st \|x\|=1,\;
((a^{-2}+1)x_3^2-2)x_4+a^{-1}(x_1^2-x_2^2)x_3=0\},
\]
and observe that $EX^*=EX(1/\rt)$. In Section~\ref{sec-E} we
will give $EX(a)$ a $G$-action and labelling making it a cromulent
surface. We call these surfaces the \emph{embedded family}. Although
our central problem is to study uniformizations of the surfaces
$EX(a)$, we will also discuss many other features of their geometry
and topology. In particular, we will give an alternative definition
which is much more geometric but takes longer to state. Two special
features of the case $a=1/\rt$ are as follows:
\begin{itemize}
\item[(a)] By a \emph{great circle} we mean the intersection of $S^3$
with a two-dimensional vector subspace of $\R^4$. For all $a$, the
fixed set of the element $\nu\in G$ is the disjoint union of three
curves, each of which is diffeomorphic to $S^1$. If
$a=1/\rt$ (but for no other value) then one of those components
is a great circle.
\item[(b)] One can show that the complexified variety
\[ CEX(a) = \{x\in\C^4\st \sum_ix_i^2=1,\;
((a^{-2}+1)x_3^2-2)x_4+a^{-1}(x_1^2-x_2^2)x_3=0\}
\]
is smooth for all $a\neq 1/\rt$, but $CEX(1/\rt)$ is
singular at the eight points in the $G$-orbit of
$(i\rt,0,\rt,1)$.
\end{itemize}
Next, put
\[ PX_0(a) = \{(w,z)\in\C^2\st w^2=z^5-(a^2+a^{-2})z^3+z\}. \]
This is an affine hyperelliptic curve. By well-known methods we can
construct a compact Riemann surface $PX(a)$ which is the union of
$PX_0(a)$ with a single extra point. In
Section~\ref{sec-P} we will give this a $G$-action and
labelling making it a cromulent surface. We call these surfaces the
\emph{projective family}.
Finally, let $\Pi$ be the abstract group generated by symbols $\bt_k$
(for $k\in\Z/8$) subject to the following relations:
\begin{align*}
\bt_{k+4} &= \bt_k^{-1} \\
\bt_0\bt_1\bt_2\bt_3\bt_4\bt_5\bt_6\bt_7 &= 1.
\end{align*}
In Section~\ref{sec-H} we will give a free action of $\Pi$ on
the unit disc $\Dl=\{z\in\C\st |z|<1\}$, depending on a parameter
$a\in(0,1)$. We will show that the orbit space $HX(a)$ for this
action is a compact Riemann surface of genus two, and give it a
cromulent structure. We call these surfaces the \emph{hyperbolic family}.
In Theorem~\ref{thm-classify-cromulent},
Corollary~\ref{cor-cromulent-iso} and Theorem~\ref{thm-H-universal}
we will show that
\begin{itemize}
\item For any two cromulent surfaces, there is at most one
isomorphism between them.
\item For any cromulent surface $X$ there is a unique $a\in (0,1)$
such that $X\simeq PX(a)$, and there is a unique $b\in (0,1)$ such
that $X\simeq HX(b)$.
\end{itemize}
In other words, the projective family and the hyperbolic family are
both universal. We conjecture that the embedded family is also
universal, but we have not proved this.
As a consequence of universality, for every $a\in (0,1)$ there is a
unique $b\in (0,1)$ such that $PX(a)\simeq HX(b)$. In
Section~\ref{sec-P-H} we will develop two different methods for
computing $a$ as a function of $b$ or \emph{vice versa}, and for
computing the corresponding cromulent isomorphism. One method
involves a rich theory based on Fuchsian differential equations and
the Schwarzian derivative; the other is less illuminating, but in some
respects more efficient and direct. The graph of $a$ against $b$ is
as follows.
\begin{center}
\begin{tikzpicture}[scale=4]
\draw[black,->] (-0.05,0) -- (1.05,0);
\draw[black,->] (0,-0.05) -- (0,1.05);
\draw[black] (1,-0.05) -- (1,0);
\draw[black] (-0.05,1) -- (0,1);
\draw ( 0.00,-0.05) node[anchor=north] {$0$};
\draw ( 1.00,-0.05) node[anchor=north] {$1$};
\draw (-0.05, 0.00) node[anchor=east ] {$0$};
\draw (-0.05, 1.00) node[anchor=east ] {$1$};
\draw ( 1.05, 0.00) node[anchor=west ] {$b$};
\draw ( 0.00, 1.05) node[anchor=south] {$a$};
\draw[red] plot[smooth] coordinates{ (0.000,1.000) (0.060,1.000) (0.080,1.000) (0.100,1.000) (0.120,1.000) (0.140,1.000) (0.160,1.000) (0.180,1.000) (0.200,1.000) (0.220,1.000) (0.240,0.999) (0.260,0.997) (0.280,0.994) (0.300,0.990) (0.320,0.983) (0.340,0.974) (0.360,0.961) (0.380,0.944) (0.400,0.923) (0.420,0.898) (0.440,0.869) (0.460,0.835) (0.480,0.798) (0.500,0.757) (0.520,0.713) (0.540,0.667) (0.560,0.620) (0.580,0.571) (0.600,0.521) (0.620,0.472) (0.640,0.423) (0.660,0.375) (0.680,0.328) (0.700,0.283) (0.720,0.241) (0.740,0.201) (0.760,0.163) (0.780,0.129) (0.800,0.099) (0.820,0.073) (0.840,0.050) (0.860,0.032) (0.880,0.019) (0.900,0.009) (0.920,0.004) (0.940,0.001) (1.000,0.000) };
\fill[black] (0.801,0.098) circle(0.015);
\end{tikzpicture}
\end{center}
We have conducted a fairly extensive heuristic
search for closed-form relationships between the above graph and
various other functions that we know to be relevant, but without
success.
Finally, we want to find $a$ and $b$ such that
$EX^*\simeq PX(a)\simeq HX(b)$. Our best estimates are
$a\simeq 0.0983562$ and $b\simeq 0.8005319$, corresponding to the
marked point on the above graph. We have some reason to hope that all
the quoted digits are accurate, but we have not performed a rigorous
error analysis. In Section~\ref{sec-classify-roothalf} we will
explain the methods used to calculate $b$ (and then $a$ is calculated
from $b$ as described previously). The first step is to find the
unique smooth function $f$ on $EX^*$ such that $e^{2f}$ times the
standard metric has curvature equal to $-1$. We can then find the
lengths of certain curves with respect to this rescaled metric, and
the value of $b$ can be determined from these lengths.
\subsection{Maple code}
To carry out the work described above, we need to check a very large
number of reasonably complex formulae and combinatorial facts, and we
also need to perform extensive numerical calculations. Most of the
formulae could individually be checked by hand with sufficient effort.
However, the number and size of the formulae are so large that
computer assistance is required for the project as a whole. We have
used Maple for this. The code and documentation are distributed
alongside this monograph, and there is an overview of the structure in
Section~\ref{sec-maple}. This monograph contains many lines like
this:
\begin{checks}
group_check.mpl: check_group_properties(), check_character_table()
\end{checks}
This indicates that some set of claims that have recently been made in
the text can be checked by executing the functions
\mcode+check_group_properties()+ and \mcode+check_character_table()+,
which are defined in the file \fname+group_check.mpl+. These functions
are set up so that they will print their own names, then they will run
silently unless they detect any errors. One can set the global
variable \mcode+assert_verbosely+ to \mcode+true+, and then the checking
functions will print additional information about the individual
claims being checked. One can check the complete set of claims for
the whole monograph by reading the file \fname+check_all.mpl+. While
this does not quite reach the level of rigour provided by formal proof
assistants such as Isabelle, it is a major step in that direction.
The worksheet \fname+text_check.mw+ also provides another means to check
the consistency of the text with the Maple code. (Some fragments of
\LaTeX code were generated automatically by Maple to ensure
correctness, but technical problems with precise control of formatting
dissuaded us from using this approach more extensively.)
One can also repeat all the numerical calculations by following the
instructions in Section~\ref{sec-build}.
The most convenient place to view and download the code, documentation
and other associated files is the page
\url{https://neilstrickland.github.io/genus2/}. A snapshot will also
be placed on the arxiv, as a set of ancilliary files.
\section{General theory of precromulent surfaces}
\label{sec-general}
\subsection{Representations of \texorpdfstring{$G$}{G}}
\label{sec-representations}
We first discuss the representation theory of $G$, which will be useful
for organising various algebraic calculations later. We assume that
the reader is familiar with the basic ideas of representation theory,
which are discussed in~\cite{se:lrf}, for example.
\begin{proposition}\label{prop-characters}
The centre of $G$ is $\{1,\lm^2,\mu\nu,\lm^2\mu\nu\}$, and the
commutator subgroup is $\{1,\lm^2\}$. The character table is as
follows:
\[ \renewcommand{\arraystretch}{1.5}
\begin{array}{|c|c|c|c|c|c|c|c|c|c|c|} \hline
& \chi_0 & \chi_1 & \chi_2 & \chi_3 & \chi_4 & \chi_5 & \chi_6 & \chi_7 & \chi_8 & \chi_9 \\ \hline
1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 2 & 2 \\ \hline
\lm^2 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & -2 & -2 \\ \hline
\mu\nu & 1 & 1 & -1 & -1 & 1 & 1 & -1 & -1 & 2 & -2 \\ \hline
\lm^2\mu\nu & 1 & 1 & -1 & -1 & 1 & 1 & -1 & -1 & -2 & 2 \\ \hline
\lm^{\pm 1} & 1 & -1 & 1 & -1 & -1 & 1 & -1 & 1 & 0 & 0 \\ \hline
\mu,\lm^2\mu & 1 & 1 & -1 & -1 & -1 & -1 & 1 & 1 & 0 & 0 \\ \hline
\lm^{\pm 1}\mu & 1 & -1 & -1 & 1 & 1 & -1 & -1 & 1 & 0 & 0 \\ \hline
\nu,\lm^2\nu & 1 & 1 & 1 & 1 & -1 & -1 & -1 & -1 & 0 & 0 \\ \hline
\lm^{\pm 1}\nu & 1 & -1 & 1 & -1 & 1 & -1 & 1 & -1 & 0 & 0 \\ \hline
\lm^{\pm 1}\mu\nu & 1 & -1 & -1 & 1 & -1 & 1 & 1 & -1 & 0 & 0 \\ \hline
\end{array}
\]
\end{proposition}
\begin{proof}
The commutator of $\lm$ and $\mu$ is $\lm^2$, and it is clear from
the form of the defining relations that $\{1,\lm^2\}$ is normal and
that $G/\{1,\lm^2\}$ is elementary abelian. It follows that the
commutator subgroup is precisely $\{1,\lm^2\}$. It is a
straightforward calculation that the elements $1$, $\lm^2$, $\mu\nu$
and $\lm^2\mu\nu$ are central, but that no other element commutes
with $\lm$. It follows that the centre is as claimed. We now see
that if $\al$ is a non-central element then the corresponding
conjugacy class is just $\{\al,\lm^2\al\}$. This means that there
are ten conjugacy classes, as listed in the left hand column. The
characters of degree one are the same as the homomorphisms from the
abelianization $G/\{1,\lm^2\}$ to $S^1$. As $G/\{1,\lm^2\}$ is
elementary abelian of order $8$, it is easy to check that
$\chi_0,\dotsc,\chi_7$ is a complete list of such characters. There
are two different retractions of $G$ onto $D_8$, one sending $\mu\nu$
to the identity, and the other sending $\mu\nu$ to $\lm^2$. There is
a standard action of $D_8$ as the isometries of a square in $\R^2$,
and by pulling this back along the two projections we get two
two-dimensional representations of $G$, with characters $\chi_8$ and
$\chi_9$. These are irreducible, because in each case the sum of the
squares of the character values is equal to the group order. We now
have ten different irreducible representations, which matches the
number of conjugacy classes, so the list is complete.
\begin{checks}
group_check.mpl: check_group_properties(), check_character_table()
\end{checks}
\end{proof}
\begin{remark}
Maple notation for the elements of $G$ is as follows:
\begin{align*}
1 &= 1 & \lm &= L & \lm^2 &= LL & \lm^3 &= LLL \\
\mu &= M & \lm\mu &= LM & \lm^2\mu &= LLM & \lm^3\mu &= LLLM \\
\nu &= N & \lm\nu &= LN & \lm^2\nu &= LLN & \lm^3\nu &= LLLN \\
\mu\nu &= MN & \lm\mu\nu &= LMN & \lm^2\mu\nu &= LLMN & \lm^3\mu\nu &= LLLMN
\end{align*}
To make this work reliably, the code in the file \fname+group.mpl+
protects the symbols \mcode+L+, \mcode+N+, \mcode+LLMN+ and so on, so
they cannot be assigned values. The function \mcode+G_mult+ computes
the group operation, so \mcode+G_mult(M,L)+ returns \mcode+LLLM+, for
example. The functions \mcode+G_inv+ and \mcode+G_conj+ compute
inverses and conjugates. To retrieve $\chi_8(\lm^2)$ (for example),
one can enter \mcode+character[8][LL]+. The variable \mcode+G16+
contains the list of all elements of $G$. All of this is set up by
the file \fname+group.mpl+.
Note that this discussion of the contents of \fname+group.mpl+ is
incomplete, as will be the case with similar comments throughout this
monograph. For full information, the reader should consult the code
itself, and the comments therein. The full set of files for this
project contains a \fname+doc+ directory. The file \fname+defs.html+
in that directory is an index of all defined symbols, with links to
the relevant lines in in the files of Maple code.
\end{remark}
\subsection{Automorphisms of \texorpdfstring{$V^*$}{V*}}
\label{sec-aut-V}
As we stated in the introduction, every precromulent surface has
precisely two cromulent labellings. In order to prove this, we will
need to understand the automorphisms of the $G$-set $V^*$, and it is
convenient to treat that question now.
\begin{proposition}\label{prop-aut-V}
$\Aut(V^*)$ is isomorphic to $C_2^5$, with the following generators:
\begin{align*}
\phi_0 &= (0\;1) \\
\phi_1 &= (2\;4)(3\;5) \\
\phi_2 &= (6\;8)(7\;9) \\
\phi_3 &= (10\;11)(12\;13) \\
\phi_4 &= (10\;12)(11\;13).
\end{align*}
\end{proposition}
Readers may find it helpful to consider the picture in
Remark~\ref{rem-V-star} when reading the argument below. The
permutation $\phi_i$ is represented in Maple as \mcode+aut_V_phi[i]+.
\begin{proof}
First, it is straightforward to check directly that the above
permutations commute with $\lm$, $\mu$ and $\nu$ (so they define
automorphisms of $V^*$). It is also easy to see that they are
commuting involutions and that they generate a group $A$ isomorphic
to $C_2^5$.
Now consider an arbitrary permutation $\phi$ that commutes with
$\lm$, $\mu$ and $\nu$; we must show that $\phi\in A$. As $\phi$
commutes with $G$, we must have $\stab_G(\phi(i))=\stab_G(i)$ for all
$i$. The stabilisers are as follows:
\begin{align*}
\stab_G(0) = \stab_G(1) &= \ip{\lm,\nu} \\
\stab_G(2) = \stab_G(4) &= \ip{\mu,\nu} \\
\stab_G(3) = \stab_G(5) &= \ip{\lm^2\mu,\lm^2\nu} \\
\stab_G(6) = \stab_G(8) &= \ip{\lm\mu,\lm\nu} \\
\stab_G(7) = \stab_G(9) &= \ip{\lm^{-1}\mu,\lm^{-1}\nu} \\
\stab_G(10) = \dotsb = \stab_G(v_{13}) &= \ip{\lm^2,\nu}.
\end{align*}
It follows that $\phi$ must preserve each of the following sets:
\[ \{0,1\},\;\{2,4\},\;\{3,5\},\;\{6,8\},\;\{7,9\},\;
\{10,11,12,13\}.
\]
The restriction of $\phi$ to $\{2,3,4,5\}$ must commute with the
restrictions of $\lm$ and $\mu$, which are $(2\;3\;4\;5)$ and
$(3\;5)$. It follows easily that the restriction of $\phi$ is
$\phi_1=(2\;4)(3\;5)$ or the identity. A similar argument shows that
the restriction of $\phi$ to $\{6,7,8,9\}$ is $\phi_2=(6\;8)(7\;9)$
or the identity, and the restriction to $\{10,11,12,13\}$ must be a
transposition pair or the identity. Here the possible transposition
pairs are $\phi_3=(10\;11)(12\;13)$ and $\phi_4=(10\;12)(11\;13)$ and
$\phi_3\phi_4=(10\;13)(11\;12)$. The claim follows easily.
\begin{checks}
group_check.mpl: check_aut_V()
\end{checks}
\end{proof}
\subsection{Quotients}
\label{sec-quotients}
Let $X$ be a precromulent surface. It is standard that for any finite
group $H$ of conformal automorphisms, the quotient $X/H$ always has a
canonical structure as a compact connected Riemann surface such that
the projection $X\to X/H$ is a branched cover. We will need to
understand the genus of $X/H$, which is determined by its Euler
characteristic, which is given by the following result:
\begin{lemma}\label{lem-chi-quotient}
For any subgroup $H\leq D_8$ we have $\chi(X/H)=|V/H|-16/|H|$.
\end{lemma}
\begin{proof}
We can write $X=A\cup B$, where $A$ is a union of small discs
around the points of $V$, and $B$ is the closure of the complement of
$A$. This means that the set $C=A\cap B$ is a disjoint union of
circles, so $\chi(C)=0$, so $\chi(A)+\chi(B)=\chi(X)$. Similarly,
$C/H$ is again a union of circles, so $\chi(C/H)=0$, so
$\chi(A/H)+\chi(B/H)=\chi(X/H)$. Now $A$ and $A/H$ are homotopy
equivalent to $V$ and $V/H$, so $\chi(A)=14$ and $\chi(A/H)=|V/H|$.
As $X$ has genus $g=2$ we have $\chi(X)=2-2g=-2$, so
$\chi(B)=-2-14=-16$. Next, note that the action of $H$ on $B$ is
free. Thus, if we choose a finite regular cell
structure on $B/H$, then the preimage in $B$ of each cell in $B/H$
will be a disjoint union of $|H|$ cells. Using this we see that
$\chi(B/H)=\chi(B)/|H|=-16/|H|$, so $\chi(X/H)=|V/H|-16/|H|$.
\end{proof}
Recall that we use the following notation for subgroups of $D_8$
\[
D_8 = \ip{\lm,\mu} \hspace{5em}
C_4 = \ip{\lm} \hspace{5em}
C_2 = \ip{\lm^2}.
\]
\begin{corollary}\label{cor-quotient-types}
The surfaces $X/C_2$, $X/C_4$ and $X/D_8$ are all conformally
equivalent to $\C_\infty$. However, $X/\ip{\mu}$ and $X/\ip{\lm\mu}$
are elliptic curves.
\end{corollary}
\begin{proof}
By the classification of compact connected Riemann surfaces, it will
suffice to show that $\chi(X/C_2)=\chi(X/C_4)=\chi(X/D_8)=2$ and
$\chi(X/\ip{\mu})=\chi(X/\ip{\lm\mu})=0$. If $H\leq D_8$ is
generated by a single element $\sg$, then $|V/H|$ is just the number
of cycles (including $1$-cycles) in the permutation corresponding to
$\sg$. This gives everything in the following table except for the
case $H=D_8$, which is easily handled in an \emph{ad-hoc} way.
\[ \begin{array}{|c|c|c|c|c|}
\hline
H & |H| & \sg & |V/H| & \chi(X/H) \\ \hline
C_2 & 2 & \lm^2 & 10 & 2 \\ \hline
C_4 & 4 & \lm & 6 & 2 \\ \hline
\ip{\mu} & 2 & \mu & 8 & 0 \\ \hline
\ip{\lm\mu} & 2 & \lm\mu & 8 & 0 \\ \hline
D_8 & 8 & & 4 & 2 \\ \hline
\end{array}
\]
\end{proof}
\begin{remark}\label{rem-conj-quot}
Recall that the action of an element $g\in G$ gives an isomorphism
$X/H\to X/gHg^{-1}$. In particular, the action of $\lm$ gives
isomorphisms $X/\ip{\mu}\to X/\ip{\lm^2\mu}$ and
$X/\ip{\lm\mu}\to X/\ip{\lm^3\mu}$. Because of this, we will mostly
restrict attention to $X/\ip{\mu}$ and $X/\ip{\lm\mu}$, and ignore
$X/\ip{\lm^2\mu}$ and $X/\ip{\lm^3\mu}$.
\end{remark}
\begin{remark}\label{rem-smooth-branch}
All of the above relies on the standard fact that if $Z$ is a Riemann
surface and $H$ is a finite group of holomorphic automorphisms, then
$Z/H$ has a natural structure as a Riemann surface, and in particular
has a smooth structure. We offer some remarks about this, some of
which will be needed later.
More precisely, the claim is that this structure makes $Z/H$ into a
coequaliser for the action in the analytic category: if $U$ is an
$H$-invariant open subset of $Z$, and $f\:U\to W$ is an $H$-invariant
analytic function to another Riemann surface $W$, then $U/H$ is open
in $Z/H$, and there is a unique analytic function $g\:U/H\to W$ such
that the composite $U\to U/H\xra{g}W$ is $f$. Note here that
coequalisers are automatically unique up to unique isomorphism.
Thus, it does not matter if we make some arbitrary choices in the
process of constructing a coequaliser; the result will be independent
of those choices.
The proof of the claim is local on $Z$. Given $z\in Z$, put
\begin{align*}
C_0 &= \{\al\in H\st \al(z)=z\} \\
C_1 &= \{\al\in H\st \al=1 \text{ on some neighbourhood of } z\} \\
C &= C_0/C_1.
\end{align*}
Then each element $\al\in C$ must act
on $T^*_zZ$ as multiplication by some scalar $\chi(\al)\in\C^\tm$;
this defines a homomorphism $\chi\:C\to\C^\tm$. By power series
methods, one can check that $\chi$ must be injective, and thus that
$C$ must be cyclic, of order $n$ say. Now choose a local parameter
$f_0$ with $f_0(z)=0$, and put
\[ f(w) = |C|^{-1}\sum_{\al\in C}\chi(\al)^{-1}f_0(\al(w)). \]
This is the same as $f_0$ to first order, so it is again a local
parameter, and it satisfies $f(\al(w))=\chi(\al)\,f(w)$. Using this,
we reduce to the case where the group $\mu_n$ of $n$'th roots of
unity acts on $\C$ by multiplication. Here, the map
$\sg_n\:z\mapsto z^n$ is easily seen to be a coequaliser.
Note, however, that the map $\sg_n\:\C\to\C$ is not a coequaliser in
the smooth category (provided that $n>1$). Indeed, the function
$f(z)=|z|^2$ is smooth and $\mu_n$-invariant. There is a unique map
$g\:\C\to\R$ with $f=g\circ\sg_n$, namely $g(w)=|w|^{2/n}$. However,
$g$ is not smooth. Because of this, if we start with a smooth
surface $Z$ and an orientation-preserving action of a finite group
$H$, there is no obvious way to obtain a smooth structure on $Z/H$.
Given $z$ and $C$ as above, we can choose a chart $\phi$ at $z$ on
which $C$ acts by rotation, and using this we obtain a chart
$\ov{\phi}$ on the quotient. However, if $\psi$ is another local
chart at $z$ on which $C$ acts by rotation, then
$\ov{\psi}^{-1}\circ\ov{\phi}$ need not be smooth.
We can always obtain a smooth structure on $Z/H$ by choosing a smooth
invariant Riemannian metric, using this to give $Z$ a conformal
structure, and then taking a major detour through the analytic
category as above. However, the result will depend on the choice of
metric, and we do not know any way to shortcut the detour.
\end{remark}
\subsection{Curve systems}
\label{sec-curve-systems}
In this section, we define what we mean by a \emph{curve system} on a
precromulent surface. Later we will exhibit curve systems for the
projective family, the hyperbolic family and the embedded family. We
will also show that the projective family is universal, so in fact
every precromulent surface has a curve system.
\begin{definition}\label{defn-precromulent-C}
Let $X$ be a labelled precromulent surface. For any $\gm\in G$ we
put $X^\gm=\{x\in X\st \gm(x)=x\}$. We then put
\begin{align*}
C_0 &= \text{ the component of $v_2$ in } X^{\mu\nu} \\
C_1 &= \text{ the component of $v_0$ in } X^{\lm\nu} \\
C_2 &= \text{ the component of $v_0$ in } X^{\lm^3\nu} \\
C_3 &= \text{ the component of $v_{11}$ in } X^{\lm^2\nu} \\
C_4 &= \text{ the component of $v_{10}$ in } X^\nu \\
C_5 &= \text{ the component of $v_0$ in } X^\nu \\
C_6 &= \text{ the component of $v_0$ in } X^{\lm^2\nu} \\
C_7 &= \text{ the component of $v_1$ in } X^\nu \\
C_8 &= \text{ the component of $v_1$ in } X^{\lm^2\nu}.
\end{align*}
\end{definition}
\begin{remark}\label{rem-Ck-circle}
The elements $\mu\nu$ and $\lm^k\nu$ act on $X$ as antiholomorphic
involutions. A standard result, which we will recall as
Corollary~\ref{cor-fixed-circles}, shows that the fixed set of an
antiholomorphic involution on a compact Riemann surface is always
diffeomorphic to a finite disjoint union of circles. Thus, each of
the sets $C_k$ above is a circle. If $\al$ and $\bt$ are distinct
antiholomorphic involutions in $G$ then $X^\al\cap X^\bt$ is fixed by
the holomorphic element $\al\bt$ and so is contained in the finite
set $V$. Thus, for example, we have $C_0\cap C_1\sse V$. On the
other hand, $C_4$ and $C_5$ are two components in $X^\nu$, so they
are either equal or disjoint. In fact, we will see later that they
are always disjoint, but this will require some further theory.
More generally, $C_4$, $C_5$ and $C_7$ are disjoint, and $C_3$, $C_6$
and $C_8$ are disjoint.
\end{remark}
\begin{remark}
The antiholomorphic involution that fixes $C_k$ is represented in
Maple as \mcode+c_involution[k]+. For example, \mcode+c_involution[6]+
evaluates to \mcode+LLN+, which is our Maple notation for $\lm^2\nu$.
\end{remark}
\begin{definition}\label{defn-curve-system}
Let $X$ be a labelled precromulent surface. A \emph{curve system} on
$X$ is a system of maps $c_k\:\R\to X$ (for $0\leq k\leq 8$) such
that:
\begin{itemize}
\item[(a)] Each $c_k$ is real-analytic and $2\pi$-periodic and induces an
embedding $\R/2\pi\Z\to X$.
\item[(b)] The vertices $v_0,\dotsc,v_{13}$ occur as values of the maps
$c_0,\dotsc,c_8$, as follows:
\[ \begin{array}{|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
&0&1&2&3&4&5&6&7&8&9&10&11&12&13\\ \hline
0&&&0&\ppi&\pi&-\ppi&\tfrac{\pi}{4}&\tfrac{3\pi}{4}&-\tfrac{3\pi}{4}&-\tfrac{\pi}{4}&&&&\\ \hline
1&0&\pi&&&&&\ppi&&-\ppi&&&&&\\ \hline
2&0&\pi&&&&&&\ppi&&-\ppi&&&&\\ \hline
3&&&&\ppi&&-\ppi&&&&&&0&&\pi\\ \hline
4&&&-\ppi&&\ppi&&&&&&0&&\pi&\\ \hline
5&0&&&&&&&&&&&\pi&&\\ \hline
6&0&&&&&&&&&&\pi&&&\\ \hline
7&&0&&&&&&&&&&&&\pi\\ \hline
8&&0&&&&&&&&&&&\pi&\\ \hline
\end{array} \]
In more detail, if the above table has an angle $\tht$ in column
$j$ of row $i$, then $c_i(\tht)=v_j$, but if column $j$ of row $i$
is empty, then $v_j\not\in c_i(\R)$.
\item[(c)] The group $G$ acts on the curves $c_k$ as follows:
\begin{align*}
\lm(c_{ 0}(t)) &= c_{ 0}( t+\pi/2) &
\mu(c_{ 0}(t)) &= c_{ 0}(-t) &
\nu(c_{ 0}(t)) &= c_{ 0}(-t) \\
\lm(c_{ 1}(t)) &= c_{ 2}( t) &
\mu(c_{ 1}(t)) &= c_{ 2}( t + \pi) &
\nu(c_{ 1}(t)) &= c_{ 2}(-t) \\
\lm(c_{ 2}(t)) &= c_{ 1}(-t) &
\mu(c_{ 2}(t)) &= c_{ 1}( t + \pi) &
\nu(c_{ 2}(t)) &= c_{ 1}(-t) \\
\lm(c_{ 3}(t)) &= c_{ 4}( t) &
\mu(c_{ 3}(t)) &= c_{ 3}( t + \pi) &
\nu(c_{ 3}(t)) &= c_{ 3}(-t) \\
\lm(c_{ 4}(t)) &= c_{ 3}(-t) &
\mu(c_{ 4}(t)) &= c_{ 4}(-t - \pi) &
\nu(c_{ 4}(t)) &= c_{ 4}( t) \\
\lm(c_{ 5}(t)) &= c_{ 6}( t) &
\mu(c_{ 5}(t)) &= c_{ 7}( t) &
\nu(c_{ 5}(t)) &= c_{ 5}( t) \\
\lm(c_{ 6}(t)) &= c_{ 5}(-t) &
\mu(c_{ 6}(t)) &= c_{ 8}(-t) &
\nu(c_{ 6}(t)) &= c_{ 6}(-t) \\
\lm(c_{ 7}(t)) &= c_{ 8}( t) &
\mu(c_{ 7}(t)) &= c_{ 5}( t) &
\nu(c_{ 7}(t)) &= c_{ 7}( t) \\
\lm(c_{ 8}(t)) &= c_{ 7}(-t) &
\mu(c_{ 8}(t)) &= c_{ 6}(-t) &
\nu(c_{ 8}(t)) &= c_{ 8}(-t) \\
\end{align*}
\end{itemize}
\end{definition}
\begin{remark}
The details of axiom~(b) are represented in Maple in several
different ways, which are useful for different purposes. Consider,
for example, the fact that $c_2(\pi/2)=v_7$ but $v_7$ does not lie on
$C_3$.
\begin{itemize}
\item \mcode+v_on_c[7,2]+ is \mcode+Pi/2+, but \mcode+v_on_c[7,3]+ is
\mcode+NULL+.
\item \mcode+c_gen[2](Pi/2)+ evaluates to \mcode+v_gen[7]+. Here
\mcode+v_gen[7]+ is just a symbol, with no assigned value. On the
other hand, \mcode+c_gen[2](Pi/4)+ just evaluates to itself,
corresponding to the fact that we have no axiom about the value of
$c_2(\pi/4)$.
\item \mcode+v_track[7]+ is a list of equations, one of which
is the equation \mcode+2=Pi/2+. There is no equation in the list
with $3$ on the left hand side.
\item \mcode+c_track[2]+ is a list of equations, one of which is the
equation \mcode+7=Pi/2+. On the other hand, \mcode+c_track[3]+ has no
equation with $7$ on the left hand side.
\end{itemize}
The details of axiom~(c) are encoded in the table \mcode+act_c_data+,
which is indexed by pairs \mcode+[g,i]+ with \mcode+g+ in $G$ and
\mcode+i+ in $\{0,\dotsc,8\}$. If \mcode+act_c_data[g,i]+ evaluates to
\mcode+[j,m,a]+ then the corresponding axiom is $g.c_i(t)=c_j(mt+a)$.
\end{remark}
\begin{remark}
Suppose we have a curve system $(c_k)_{k=0}^8$ and a strictly
increasing analytic diffeomorphism $u\:\R\to\R$ with $u(-t)=-u(t)$
and $u(t+\pi/4)=u(t)+\pi/4$; then the maps $c_k\circ u$ give another
curve system. Thus, curve systems are not unique. However, they are
unique up to a kind of reparametrisation slightly more general than
that described above; we will not spell out the details.
\end{remark}
\begin{proposition}\label{prop-curve-system}
Let $(c_k)_{k=0}^8$ be a curve system on a labelled precromulent
surface $X$. Then:
\begin{itemize}
\item[(1)] For each $k$ the map $c_k$ gives a diffeomorphism
$\R/2\pi\Z\to C_k$.
\item[(2)] The sets $C_4$, $C_5$ and $C_7$ are disjoint.
\item[(3)] The sets $C_3$, $C_6$ and $C_8$ are disjoint.
\item[(4)] For all $i\neq j$ we have $C_i\cap C_j\sse V$ (so a
precise list of elements of $C_i\cap C_j$ can be read off from
axiom~(b)).
\end{itemize}
\end{proposition}
\begin{proof}
Axiom~(c) gives $\mu\nu(c_0(t))=\mu(c_0(-t))=c_0(t)$, so
$c_0(\R)\sse X^{\mu\nu}$. Moreover, $c_0(\R)$ is connected and
contains $c_0(0)$, which is $v_2$ by axiom~(b). This proves that
$c_0(\R)\sse C_0$. Axiom~(a) tells us that $c_0$ gives a smooth
embedding $\R/2\pi\Z\to C_0$, but $C_0$ is diffeomorphic to a circle
by Remark~\ref{rem-Ck-circle}, and any smooth embedding of a circle
in a circle is necessarily a diffeomorphism. The same line of
argument shows that $c_k$ induces a diffeomorphism $\R/2\pi\Z\to C_k$
for all $k$. Next, axiom~(b) tells us that $v_0\not\in c_4(\R)=C_4$,
so $C_5$ is a component of $X^\nu$ which is different from $C_4$ and
therefore disjoint from $C_4$. The same line of argument shows that
$C_4$, $C_5$ and $C_7$ are disjoint, and also that $C_3$, $C_6$ and
$C_8$ are disjoint. Now consider an intersection $C_i\cap C_j$ that
is not covered by~(b) or~(c). We then find that $C_i\sse X^\gm$ and
$C_j\sse X^\dl$ for some antiholomorphic involutions $\gm,\dl\in G$
with $\gm\neq\dl$, so $\gm\dl$ is a nontrivial element of $D_8$. Any
element of $C_i\cap C_j$ is fixed by $\gm\dl$, and so lies in $V$ by
the definition of $V$.
\end{proof}
\begin{proposition}\label{prop-empty-boxes}
Suppose we have a system of maps $c_k\:\R\to X$ such that axioms~(a)
and~(c) are satisfied. Suppose also that
\begin{itemize}
\item[(p)] The part of axiom~(b) corresponding to the nonempty boxes
in the table is satisfied.
\item[(q)] The sets $c_3(\R)$, $c_6(\R)$ and $c_8(\R)$ are disjoint.
\end{itemize}
Then the maps $c_k$ give a curve system.
\end{proposition}
\begin{proof}
First note that part~(a) of Proposition~\ref{prop-curve-system} used
only axioms that we are still assuming here, so we again have
$C_k=c_k(\R)$ for all $k$. We also see from axiom~(c) that
$\lm(C_3)=C_4$ and $\lm(C_6)=C_5$ and $\lm(C_8)=C_7$, so $C_4$, $C_5$
and $C_7$ are also disjoint.
Next, we can redraw the table in axiom~(b) as follows:
\[ \begin{array}{|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
&0&1 &2 &3 &4 &5 &6&7&8&9&10&11&12&13\\ \hline
0&A&A &0 &\ppi&\pi &-\ppi&\tfrac{\pi}{4}&\tfrac{3\pi}{4}&-\tfrac{3\pi}{4}&-\tfrac{\pi}{4}&A&A&A&A\\ \hline
1&0&\pi&A &A &A &A &\ppi&A&-\ppi&A&A&A&A&A\\ \hline
2&0&\pi&A &A &A &A &A&\ppi&A&-\ppi&A&A&A&A\\ \hline
3&B&B &A &\ppi&A &-\ppi&A&A&A&A&B&0&B&\pi\\ \hline
4&B&B &\ppi&A &\ppi&A &A&A&A&A&0&B&\pi&B\\ \hline
5&0&B &B &A &B &A &A&A&A&A&B&\pi&B&B\\ \hline
6&0&B &A &B &A &B &A&A&A&A&\pi&B&B&B\\ \hline
7&B&0 &B &A &B &A &A&A&A&A&B&B&B&\pi\\ \hline
8&B&0 &A &B &A &B &A&A&A&A&B&B&\pi&B\\ \hline
\end{array} \]
All the boxes that were blank in the original table have been marked
$A$ or $B$. Consider for example column $5$, corresponding to
$v_5$. It follows from the definition of a precromulent labelling
that the stabiliser group of $v_5$ is
$\{1,\mu\nu,\lm^2\mu,\lm^2\nu\}$, so in particular $v_5$ is not fixed
by $\nu$, $\lm\nu$ or $\lm^3\nu$, so it cannot lie in $c_i(\R)$ for
$i\in\{1,2,4,5,7\}$. This accounts for all the boxes in column $5$
marked $A$. We also have $-\pi/2$ in row $3$, indicating that
$v_5=c_3(-\pi/2)\in C_3$. As $C_3$, $C_6$ and $C_8$ are disjoint, we
see that $v_5\not\in c_6(\R)$ and $v_5\not\in c_8(\R)$, which
accounts for the remaining two boxes in column $5$ marked $B$. The
same line of argument works for all the other columns.
\end{proof}
\begin{definition}\label{defn-std-isotropy}
We say that $X$ has \emph{standard isotropy} if
\begin{align*}
X^{\mu\nu} &= C_0 \\
X^{\lm\nu} &= C_1 \\
X^{\lm^3\nu} &= C_2 \\
X^{\nu} &= C_4\amalg C_5 \amalg C_7 \\
X^{\lm^2\nu} &= C_3\amalg C_6 \amalg C_8 \\
X^{\lm^2\mu\nu} &= \emptyset.
\end{align*}
\end{definition}
We will show later that every cromulent surface has standard
isotropy.
\subsection{Holomorphic curve systems}
\label{sec-holomorphic-curves}
For any curve system, it turns out that each map $c_k\:\R\to X$ can be
extended to give a holomorphic map defined on a suitable neighbourhood
of $\R$ in $\C$. We will start by developing the relevant theory in a
slightly more abstract setting.
\begin{proposition}\label{prop-chart}
Let $X$ be a Riemann surface, and let $c\:\R\to X$ be a real analytic
map with $c'(0)\neq 0$. Then there is a unique germ of an analytic
map $\phi\:\C\to X$ with $\phi(t)=c(t)$ for small real values of
$t$. Similarly, there is a unique local conformal
parameter $z$ at $c(0)$ such that $z(c(t))=t$ for small $t\in\R$.
Moreover, if $\tau$ is an antiholomorphic involution on $X$ with
$\tau(c(t))=c(t)$ for all $t$, then $z(\tau(u))=\ov{z(u)}$.
\end{proposition}
\begin{proof}
The claim is local on $X$, so we may assume that $X=\C$ and
$c(0)=0$. As $c$ is real analytic, there are coefficients $a_k\in\C$
such that $c(t)=\sum_ka_kt^k$, with the sum being absolutely
convergent for small real values of $t$. It follows in a standard
way that the sum is still absolutely convergent for small complex
values of $t$, and this gives us a germ of a complex analytic map
$\phi\:\C\to X$ extending $c$. This is unique, by the Identity
Principle. We also have $\phi'(0)=c'(0)\neq 0$, so $\phi$ is locally
invertible near $0$, and the inverse is the unique local parameter
$z$ such that $z(c(t))=t$.
Now suppose that $\tau$ is an antiholomorphic involution on $X$ with
$\tau(c(t))=c(t)$ for all $t$. Then the map
$u\mapsto\ov{z(\tau(u))}$ has the defining property of $z$, and so is
the same as $z$, as claimed.
\end{proof}
Given $c\:\R\to X$, we can apply the above proposition to $c(t_0+t)$
for various different values of $t_0$, and then patch the results
together. To organise this construction, we introduce the
following definitions:
\begin{definition}\label{defn-band}
Consider a point $z=x+iy\in\C$. We let $\CQ(z)$ denote the set of
pairs $(U_0,\tc_0)$, where $U_0$ is a convex open subset of $\C$
containing $x$ and $z$, and $\tc_0\:U_0\to X$ is a holomorphic map with
$\tc_0|_{U_0\cap\R}=c|_{U_0\cap\R}$. We then put
$V=\{z\st\CQ(z)\neq\emptyset\}$. If $z\in V$ then we choose any
$(U_0,\tc_0)\in\CQ(z)$ and put $\tc(z)=\tc_0(z)$; a straightforward
argument with the identity principle shows that this is independent
of the choice of $(U_0,\tc_0)$.
\end{definition}
\begin{proposition}\label{prop-band-chart}
The set $V$ is open in $\C$ and contains $\R$, and it is closed under
conjugation. The map $\tc\:V\to X$ is holomorphic, and satisfies
$\tc|_{V\cap\R}=c$ and $\tc(\ov{z})=\tau(\tc(z))$. Moreover, if
$c(t+2\pi)=c(t)$ for all $t$, then we also have $V+2\pi=V$ and
$\tc(t+2\pi)=\tc(t)$.
\end{proposition}
\begin{proof}
Straightforward.
\end{proof}
\begin{corollary}\label{cor-band-charts}
Suppose that $X$ is a cromulent surface, and $(c_k)_{k=0}^8$ is a
curve system. Then each map $c_k\:\R\to X$ has a canonical
holomorphic extension $\tc_k\:V_k\to X$, where $V_k$ is a
$2\pi$-periodic open neighbourhood of $\R$ in $\C$. \qed
\end{corollary}
\subsection{Fundamental domains}
\label{sec-fundamental}
\begin{definition}\label{defn-fundamental}
Let $X$ be a compact Hausdorff space, and let $H$ be a finite group acting
continuously on $X$. Let $F$ be a closed subset of $X$.
\begin{itemize}
\item[(a)] We say that $F$ is a \emph{fundamental domain} for $H$ if
$X=\bigcup_{\gm\in H}\gm(F)$, and $\text{int}(F)\cap\gm(F)=\emptyset$
for all $\gm\in H\sm\{1\}$.
\item[(b)] We say that $F$ is a \emph{retractive fundamental domain} if,
in addition, there is a continuous map $r\:X\to F$ such that
\begin{itemize}
\item[(i)] $r(x)=x$ for all $x\in F$ (so $r$ is a retraction).
\item[(ii)] $r(\gm(x))=r(x)$ for all $x\in X$ and $\gm\in H$.
\end{itemize}
\end{itemize}
\end{definition}
\begin{proposition}\label{prop-fundamental}
Let $F$ be a retractive fundamental domain for $H$, with retraction $r$.
Then
\begin{itemize}
\item[(a)] For all $x\in X$, the point $r(x)$ lies in the same
$H$-orbit as $x$.
\item[(b)] For all $\gm\in H$ we have
$F\cap\gm(F)=\{x\in F\st \gm(x)=x\}$.
\item[(c)] There is a canonical homeomorphism $X/H\simeq F$.
\item[(d)] There is a canonical homeomorphism
$X\simeq(G\tm F)/\sim$, where $(\gm_0,x_0)\sim(\gm_1,x_1)$ iff
$x_0=x_1$ and $\gm_1^{-1}\gm_0\in\stab_G(x_0)$.
\end{itemize}
\end{proposition}
\begin{proof}
\begin{itemize}
\item[(a)] As $F$ is a fundamental domain, we have $x=\gm(y)$ for
some $y\in F$ and $\gm\in H$. This gives $r(x)=r(\gm(y))$, but we
can use axioms~(ii) and~(i) to see that $r(\gm(y))=r(y)=y$, so
$r(x)=y$. Thus, $x$ and $r(x)$ lie in the same $H$-orbit.
\item[(b)] Now suppose that $x\in F\cap\gm(F)$, so $x=\gm(y)$ for
some $y\in F$. We now have $x=r(x)=r(\gm(y))=r(y)=y$, so
$x=\gm(x)$ as required.
\item[(c)] We have an inclusion $j\:F\to X$ and a projection
$p\:X\to X/H$. We will show that $pj$ is a homeomorphism.
As $r$ is continuous with $r(\gm(x))=r(x)$ for all $x$ and $\gm$,
we see that there is a unique map $\ov{r}\:X/H\to F$ with
$\ov{r}p=r$, and that this is continuous. As $r$ is a retraction
we have $\ov{r}pj=rj=1$. Next, as $x$ is in the same orbit as
$r(x)$, we have $p(x)=pjr(x)=pj\ov{r}p(x)$. As $p$ is surjective
it follows that $pj\ov{r}=1$. This proves that $\ov{r}$ is an
inverse for $pj$, as required.
\item[(d)] We have a continuous map $m\:G\tm F\to X$ given by
$m(\gm,x)=\gm(x)$. As $X=\bigcup_\gm\gm(F)$ we see that $m$ is
surjective. The source and target are compact Hausdorff spaces, so
$m$ is automatically a quotient map. If
$m(\gm_0,x_0)=m(\gm_1,x_1)$ then the element $\gm=\gm_1^{-1}\gm_0$
has $\gm(x_0)=x_1$. Applying $r$ to this gives $x_0=x_1$, and it
follows that $\gm\in\stab_G(x_0)$. It follows that $m$ induces a
homeomorphism $(G\tm F)/\sim\to X$, as claimed.
\end{itemize}
\end{proof}
Now let $X$ be a labelled precromulent surface with a curve system.
By the axioms for a curve system, we have
\begin{align*}
v_0 &= c_1(0) = c_5(0) &
v_3 &= c_0(\pi/2) = c_3(\pi/2) \\
v_6 &= c_0(\pi/4) = c_1(\pi/2) &
v_{11} &= c_3(0) = c_5(\pi).
\end{align*}
This means that the set
\[ DF_{16} =
c_0([\tfrac{\pi}{4},\ppi]) \cup
c_1([0,\ppi]) \cup
c_3([0,\ppi]) \cup
c_5([0,\pi])
\]
fits together as follows:
\begin{center}
\begin{tikzpicture}[scale=5]
\begin{scope}
\draw[blue] (1.00,0.00) -- (1.00,1.00);
\draw[blue,->] (1.00,0.00) -- (1.00,0.50);
\draw[green] (1.00,0.00) -- (0.00,0.00);
\draw[green,->] (1.00,0.00) -- (0.50,0.00);
\draw[magenta] (1.00,1.00) -- (0.00,1.00);
\draw[magenta,->] (1.00,1.00) -- (0.50,1.00);
\draw[cyan] (0.00,0.00) -- (0.00,1.00);
\draw[cyan,->] (0.00,0.00) -- (0.00,0.50);
\fill (1.00,0.00) circle(0.005);
\fill (1.00,1.00) circle(0.005);
\fill (0.00,0.00) circle(0.005);
\fill (0.00,1.00) circle(0.005);
\draw (1.00,0.00) node[anchor=north] {$\ss v_0$};
\draw (1.00,1.00) node[anchor=south] {$\ss v_{11}$};
\draw (0.00,0.00) node[anchor=north] {$\ss v_{6}$};
\draw (0.00,1.00) node[anchor=south] {$\ss v_3$};
\draw (1.00,0.50) node[anchor=west] {$\ss c_5([0,\pi])$};
\draw (0.00,0.50) node[anchor=east] {$\ss c_0([\pi/4,\pi/2])$};
\draw (0.50,0.00) node[anchor=north] {$\ss c_1([0,\pi/2])$};
\draw (0.50,1.00) node[anchor=south] {$\ss c_3([0,\pi/2])$};
\end{scope}
\end{tikzpicture}
\end{center}
(Using Proposition~\ref{prop-curve-system}, we see that the four
boundary arcs cannot have any additional intersection points.)
\begin{remark}
Information about the above picture is stored as a table in Maple in
the global variable \mcode+F16_curve_limits+, which is defined in
\fname+cromulent.mpl+. For example, \mcode+F16_curve_limits[1]+ is
the range \mcode+0..Pi/2+, whereas \mcode+F16_curve_limits[2]+ is
\mcode+NULL+ (because none of the sides of $DF_{16}$ lies along
$C_2$). Note that Maple does not display the full structure of
tables by default; to see all entries in \mcode+F16_curve_limits+, one
needs to enter \mcode+eval(F16_curve_limits)+, not just
\mcode+F16_curve_limits+.
\end{remark}
\begin{remark}
The colours in the above diagram will be used systematically
throughout this monograph: the curve $c_0$ is cyan, the curves $c_1$
and $c_2$ are green, the curves $c_3$ and $c_4$ are magenta, and the
curves $c_5$ to $c_8$ are blue. The colour of $c_k$ is represented
in Maple as \mcode+c_colour[k]+. Readers who have trouble
distinguishing these colours can try changing the definitions of
\mcode+c_colour[k]+ in the file \fname+cromulent.mpl+ and
regenerating the diagrams using the functions in various files called
\fname+plots.mpl+ appearing in several different directories.
However, colour should not be strictly necessary for any of the
diagrams.
\end{remark}
\begin{lemma}\label{lem-F-stabilisers}
The stabilisers of points in $DF_{16}$ are as follows:
\begin{align*}
\stab_G(v_0) &= \ip{\lm,\nu} &
\stab_G(v_3) &= \ip{\lm^2\mu,\lm^2\nu} \\
\stab_G(v_6) &= \ip{\lm\mu,\lm\nu} &
\stab_G(v_{11}) &= \ip{\lm^2,\nu}
\end{align*}
\begin{align*}
\stab_G(c_0(t)) &= \{1,\mu\nu\} & \text{ for } & \pi/4<t<\pi/2 \\
\stab_G(c_1(t)) &= \{1,\lm\nu\} & \text{ for } & 0<t<\pi/2 \\
\stab_G(c_3(t)) &= \{1,\lm^2\nu\} & \text{ for } & 0<t<\pi/2 \\
\stab_G(c_5(t)) &= \{1,\nu\} & \text{ for } & 0<t<\pi.
\end{align*}
\end{lemma}
\begin{proof}
The stabilisers for the points $v_i$ are determined by
Definition~\ref{defn-V-star}. Next, axiom~(c) in
Definition~\ref{defn-curve-system} tells us that
$\stab_G(c_0(t))\supseteq\{1,\mu\nu\}$ for all $t$. Moreover, if
$\pi/4<t<\pi/2$ then axiom~(b) tells us that $c_0(t)\neq v_i$ for all
$i$, so $\stab_G(c_0(t))\cap D_8=\{1\}$. It is easy to see that any
subgroup strictly larger than $\{1,\mu\nu\}$ has nontrivial
intersection with $D_8$, so we must have
$\stab_G(c_0(t))=\{1,\mu\nu\}$ as claimed. The same line of argument
works for $c_1$, $c_3$ and $c_5$.
\end{proof}
\begin{definition}\label{defn-standard-F}
A \emph{standard fundamental domain} is a subset $F_{16}\sse X$ that
is a retractive fundamental domain for $G$ and is homeomorphic to a
square and has boundary $DF_{16}$.
\end{definition}
\begin{remark}\label{rem-standard-F}
We will see later that every cromulent surface has a unique standard
fundamental domain. Conversely, suppose that $X$ is a labelled
precromulent surface with a given curve system and a standard
fundamental domain, and that $\lm_*=i\:T_{v_0}X\to T_{v_0}X$. Then
the interior of the standard fundamental domain has the property
specified in Definition~\ref{defn-precromulent}(d), which proves that
$X$ is actually cromulent.
\end{remark}
\begin{remark}
If $F_{16}$ is a standard fundamental domain, the
Proposition~\ref{prop-fundamental}(d) allows us to identify $X$ with
$(G\tm F_{16})/\sim$ for a certain equivalence relation $\sim$. This
relation depends only on the stabilisers of points in $DF_{16}$,
which are given by Lemma~\ref{lem-F-stabilisers}. We can also
identify $F_{16}$ with $[0,1]^2$ and thus identify $X$ with a
quotient of $G\tm [0,1]^2$.
\end{remark}
If we perform only some of the identifications given by the above
equivalence relation, we obtain the following space, which we call
$\Net_0$. It is clearly homeomorphic to a square.
\begin{center}
\begin{tikzpicture}[scale=.5]
\draw[ green] ( 0, 0) -- ( 6, 6);
\draw[ green] ( 0, 0) -- ( -6, 6);
\draw[ green] ( 0, 0) -- ( -6, -6);
\draw[ green] ( 0, 0) -- ( 6, -6);
\draw[ blue] ( 0, 0) -- ( 0, 8);
\draw[ blue] ( 0, 0) -- ( 0, -8);
\draw[ blue] ( 0, 0) -- ( 8, 0);
\draw[ blue] ( 0, 0) -- ( -8, 0);
\draw[ blue] ( 12, 12) -- ( 4, 12);
\draw[ blue] ( 12, 12) -- ( 12, 4);
\draw[ green] ( 12, 12) -- ( 6, 6);
\draw[ green] (-12, 12) -- ( -6, 6);
\draw[ green] (-12,-12) -- ( -6, -6);
\draw[ green] ( 12,-12) -- ( 6, -6);
\draw[ cyan] ( 4, 8) -- ( 6, 6);
\draw[magenta] ( 4, 8) -- ( 0, 8);
\draw[magenta] ( 4, 8) -- ( 4, 12);
\draw[ cyan] ( 8, 4) -- ( 6, 6);
\draw[magenta] ( 8, 4) -- ( 8, 0);
\draw[magenta] ( 8, 4) -- ( 12, 4);
\draw[ cyan] ( -4, 8) -- ( -6, 6);
\draw[magenta] ( -4, 8) -- ( 0, 8);
\draw[magenta] ( -4, 8) -- ( -4, 12);
\draw[ cyan] ( 8, -4) -- ( 6, -6);
\draw[magenta] ( 8, -4) -- ( 8, 0);
\draw[magenta] ( 8, -4) -- ( 12, -4);
\draw[ blue] (-12, 12) -- ( -4, 12);
\draw[ blue] (-12, 12) -- (-12, 4);
\draw[ blue] (-12,-12) -- ( -4,-12);
\draw[ blue] (-12,-12) -- (-12, -4);
\draw[ blue] ( 12,-12) -- ( 4,-12);
\draw[ blue] ( 12,-12) -- ( 12, -4);
\draw[ cyan] ( 4, -8) -- ( 6, -6);
\draw[magenta] ( 4, -8) -- ( 0, -8);
\draw[magenta] ( 4, -8) -- ( 4,-12);
\draw[ cyan] ( -8, 4) -- ( -6, 6);
\draw[magenta] ( -8, 4) -- ( -8, 0);
\draw[magenta] ( -8, 4) -- (-12, 4);
\draw[ cyan] ( -4, -8) -- ( -6, -6);
\draw[magenta] ( -4, -8) -- ( 0, -8);
\draw[magenta] ( -4, -8) -- ( -4,-12);
\draw[ cyan] ( -8, -4) -- ( -6, -6);
\draw[magenta] ( -8, -4) -- ( -8, 0);
\draw[magenta] ( -8, -4) -- (-12, -4);
\draw[-angle 90, cyan] ( 4, 8) -- ( 5, 7);
\draw[-angle 90, cyan] ( 6, 6) -- ( 7, 5);
\draw[-angle 90, cyan] ( 8, -4) -- ( 7, -5);
\draw[-angle 90, cyan] ( 6, -6) -- ( 5, -7);
\draw[-angle 90, cyan] ( -4, -8) -- ( -5, -7);
\draw[-angle 90, cyan] ( -6, -6) -- ( -7, -5);
\draw[-angle 90, cyan] ( -8, 4) -- ( -7, 5);
\draw[-angle 90, cyan] ( -6, 6) -- ( -5, 7);
\draw ( -5, 7) node[anchor=north west] {$\ss c_0$};
\draw ( 5, 7) node[anchor=north east] {$\ss c_0$};
\draw ( -5, -7) node[anchor=south west] {$\ss c_0$};
\draw ( 5, -7) node[anchor=south east] {$\ss c_0$};
\draw[-angle 90,green] ( 3, 3) -- ( 4, 4);
\draw[-angle 90,green] ( 7, 7) -- ( 8, 8);
\draw[-angle 90,green] ( -5, -5) -- ( -4, -4);
\draw[-angle 90,green] ( -9, -9) -- ( -8, -8);
\draw ( 4, 4) node[anchor=north west] {$\ss c_1$};
\draw[-angle 90,green] ( -3, 3) -- ( -4, 4);
\draw[-angle 90,green] ( -7, 7) -- ( -8, 8);
\draw[-angle 90,green] ( 5, -5) -- ( 4, -4);
\draw[-angle 90,green] ( 9, -9) -- ( 8, -8);
\draw ( -4, 4) node[anchor=north east] {$\ss c_2$};
\draw[-angle 90,magenta] (-11, -4) -- (-10, -4);
\draw[-angle 90,magenta] ( -8, -4) -- ( -8, -2);
\draw[-angle 90,magenta] ( -8, 0) -- ( -8, 2);
\draw[-angle 90,magenta] ( -9, 4) -- (-10, 4);
\draw ( -8, -2) node[anchor=east] {$\ss c_3$};
\draw[-angle 90,magenta] ( 11, -4) -- ( 10, -4);
\draw[-angle 90,magenta] ( 8, -4) -- ( 8, -2);
\draw[-angle 90,magenta] ( 8, 0) -- ( 8, 2);
\draw[-angle 90,magenta] ( 9, 4) -- ( 10, 4);
\draw ( 8, -2) node[anchor=west] {$\ss c_3$};
\draw[-angle 90,magenta] ( 4, 11) -- ( 4, 10);
\draw[-angle 90,magenta] ( 4, 8) -- ( 2, 8);
\draw[-angle 90,magenta] ( 0, 8) -- ( -2, 8);
\draw[-angle 90,magenta] ( -4, 9) -- ( -4, 10);
\draw ( -2, 8) node[anchor=south] {$\ss c_4$};
\draw[-angle 90,magenta] ( 4,-11) -- ( 4,-10);
\draw[-angle 90,magenta] ( 4, -8) -- ( 2, -8);
\draw[-angle 90,magenta] ( 0, -8) -- ( -2, -8);
\draw[-angle 90,magenta] ( -4, -9) -- ( -4,-10);
\draw ( -2, -8) node[anchor=north] {$\ss c_4$};
\draw[-angle 90,blue] ( 3, 0) -- ( 4, 0);
\draw[-angle 90,blue] ( -5, 0) -- ( -4, 0);
\draw ( 4, 0) node[anchor=north] {$\ss c_5$};
\draw[-angle 90,blue] ( 0, 3) -- ( 0, 4);
\draw[-angle 90,blue] ( 0, -5) -- ( 0, -4);
\draw ( 0, 4) node[anchor=east] {$\ss c_6$};
\draw[-angle 90,blue] (-12, 7) -- (-12, 8);
\draw[-angle 90,blue] (-12, -7) -- (-12, -8);
\draw[-angle 90,blue] ( 12, 9) -- ( 12, 8);
\draw[-angle 90,blue] ( 12, -9) -- ( 12, -8);
\draw (-12, 8) node[anchor=east] {$\ss c_7$};
\draw (-12, -8) node[anchor=east] {$\ss c_7$};
\draw ( 12, 8) node[anchor=west] {$\ss c_7$};
\draw ( 12, -8) node[anchor=west] {$\ss c_7$};
\draw[-angle 90,blue] ( -9, 12) -- ( -8, 12);
\draw[-angle 90,blue] ( 9, 12) -- ( 8, 12);
\draw[-angle 90,blue] ( -7,-12) -- ( -8,-12);
\draw[-angle 90,blue] ( 7,-12) -- ( 8,-12);
\draw ( 8,-12) node[anchor=north] {$\ss c_8$};
\draw ( -8,-12) node[anchor=north] {$\ss c_8$};
\draw ( 8, 12) node[anchor=south] {$\ss c_8$};
\draw ( -8, 12) node[anchor=south] {$\ss c_8$};
\fill[black] ( 0, 0) circle(0.10);
\fill[black] ( -4, -8) circle(0.10);
\fill[black] ( 12, 12) circle(0.10);
\fill[black] ( 4, 8) circle(0.10);
\fill[black] ( 8, 4) circle(0.10);
\fill[black] ( -4, 8) circle(0.10);
\fill[black] ( 8, -4) circle(0.10);
\fill[black] ( 6, 6) circle(0.10);
\fill[black] ( -6, 6) circle(0.10);
\fill[black] ( 6, -6) circle(0.10);
\fill[black] ( -6, -6) circle(0.10);
\fill[black] ( 8, 0) circle(0.10);
\fill[black] ( 0, 8) circle(0.10);
\fill[black] ( 12, 4) circle(0.10);
\fill[black] ( 4, 12) circle(0.10);
\fill[black] ( -8, 4) circle(0.10);
\fill[black] ( 0, -8) circle(0.10);
\fill[black] ( -8, -4) circle(0.10);
\fill[black] (-12, 4) circle(0.10);
\fill[black] ( -4,-12) circle(0.10);
\fill[black] (-12, 12) circle(0.10);
\fill[black] (-12,-12) circle(0.10);
\fill[black] ( -8, 0) circle(0.10);
\fill[black] (-12, -4) circle(0.10);
\fill[black] ( 4,-12) circle(0.10);
\fill[black] ( 12,-12) circle(0.10);
\fill[black] ( 12, -4) circle(0.10);
\fill[black] ( 4, -8) circle(0.10);
\fill[black] ( -4, 12) circle(0.10);
\draw ( 0, 0) node[anchor=north]{$\ss v_{0}$};
\draw ( -4, -8) node[anchor=north west]{$\ss v_{4.1}$};
\draw ( 12, 12) node[anchor=south west]{$\ss v_{1}$};
\draw ( 4, 8) node[anchor=south east]{$\ss v_{2}$};
\draw ( 8, 4) node[anchor=north west]{$\ss v_{3}$};
\draw ( -4, 8) node[anchor=south west]{$\ss v_{4}$};
\draw ( 8, -4) node[anchor=south west]{$\ss v_{5}$};
\draw ( 6, 6) node[anchor=north]{$\ss v_{6}$};
\draw ( -6, 6) node[anchor=north]{$\ss v_{7}$};
\draw ( 6, -6) node[anchor=north]{$\ss v_{9}$};
\draw ( -6, -6) node[anchor=north]{$\ss v_{8}$};
\draw ( 8, 0) node[anchor=west]{$\ss v_{11}$};
\draw ( 0, 8) node[anchor=south]{$\ss v_{10}$};
\draw ( 12, 4) node[anchor=north west]{$\ss v_{13}$};
\draw ( 4, 12) node[anchor=south east]{$\ss v_{12}$};
\draw ( -8, 4) node[anchor=north east]{$\ss v_{3.1}$};
\draw ( 0, -8) node[anchor=north]{$\ss v_{10.1}$};
\draw ( -8, -4) node[anchor=south east]{$\ss v_{5.1}$};
\draw (-12, 4) node[anchor=north east]{$\ss v_{13.3}$};
\draw ( -4,-12) node[anchor=north west]{$\ss v_{12.1}$};
\draw (-12, 12) node[anchor=south east]{$\ss v_{1.1}$};
\draw (-12,-12) node[anchor=north east]{$\ss v_{1.2}$};
\draw ( -8, 0) node[anchor=east]{$\ss v_{11.1}$};
\draw (-12, -4) node[anchor=south east]{$\ss v_{13.1}$};
\draw ( 4,-12) node[anchor=north east]{$\ss v_{12.3}$};
\draw ( 12,-12) node[anchor=north west]{$\ss v_{1.3}$};
\draw ( 12, -4) node[anchor=south west]{$\ss v_{13.2}$};
\draw ( 4, -8) node[anchor=north east]{$\ss v_{2.1}$};
\draw ( -4, 12) node[anchor=south west]{$\ss v_{12.2}$};
\draw (6.00,2.00) node{$1$};
\draw (-2.00,6.00) node{$\lm$};
\draw (-6.00,-2.00) node{$\lm^2$};
\draw (2.00,-6.00) node{$\lm^3$};
\draw (10.00,-6.00) node{$\mu$};
\draw (6.00,10.00) node{$\lm\mu$};
\draw (-10.00,6.00) node{$\lm^2\mu$};
\draw (-6.00,-10.00) node{$\lm^3\mu$};
\draw (6.00,-2.00) node{$\nu$};
\draw (2.00,6.00) node{$\lm\nu$};
\draw (-6.00,2.00) node{$\lm^2\nu$};
\draw (-2.00,-6.00) node{$\lm^3\nu$};
\draw (10.00,6.00) node{$\mu\nu$};
\draw (-6.00,10.00) node{$\lm\mu\nu$};
\draw (-10.00,-6.00) node{$\lm^2\mu\nu$};
\draw (6.00,-10.00) node{$\lm^3\mu\nu$};
\end{tikzpicture}
\end{center}
\begin{checks}
nets_check.mpl: check_nets()
\end{checks}
\begin{remark}
There is code for dealing with nets in the file \fname+nets.mpl+.
This uses the object oriented programming framework described in
Section~\ref{sec-oo-maple}. Information about $\Net_0$ is stored in
the variable \mcode+net_0+, as an instance of the class \mcode+net+.
This means that
\begin{itemize}
\item One can enter \mcode+net_0["v"][12.1]+ to retrieve the
coordinates of the point $v_{12.1}$ in the above picture.
\item One can enter \mcode+net_0["squares"][M]+ to retrieve the list
\mcode+[9,1,13.2,5]+ corresponding to the vertices of the region
marked $\mu$ (recall that $\mu$ is represented as \mcode+M+ in Maple).
\item One can enter \mcode+net_0["plot"]+ to generate a picture of
the net as a Maple plot structure. Note that this is an example of
a method rather than a property: it performs an operation rather
than simply returning information that was previously stored.
\item One can enter \mcode+net_0["check"]+ to perform various
consistency checks on the combinatorial structure of the net.
\end{itemize}
There are also various other properties and methods.
\end{remark}
\begin{remark}\label{rem-match-net}
Elsewhere we will consider a number of other constructions which give
partial or global maps between cromulent surfaces and $\R^2$ or $\C$
or $\C_\infty\simeq S^2$ or $\R^2/\Z^2$. We will usually arrange the
details of such maps so that they match up with $\Net_0$ as far as
possible: $v_0$ will go to the origin, $c_5(t)$ will go to the
positive $x$-axis for small $t>0$, $c_6(t)$ will go to the positive
$y$-axis for small $t>0$, and so on.
\end{remark}
We can obtain the space $X$ by performing some additional
identifications on the boundary. The points marked $v_{12}$,
$v_{12.1}$, $v_{12.2}$ and $v_{12.3}$ all map to $v_{12}$, and
similarly for the other points with fractional labels.
The above net inherits an orientation from $\R^2$, but it also
inherits an orientation from $X$, so we can ask whether these
orientations are the same. To see that they are, recall that $\lm$
acts on the tangent space $T_{v_0}X$ as multiplication by $i$. (This
was part of the definition of a cromulent labelling.) On the other
hand, we have $\lm(c_1(t))=c_2(t)$, and from this we see that $\lm$
acts on the net near $v_0$ as an anticlockwise turn through $\pi/2$.
This implies that the orientations are compatible as claimed.
To explain the gluing conditions on the boundary in more detail, we
use the following, less cluttered version of the above diagram:
\begin{center}
\begin{tikzpicture}[scale=0.3]
\draw[blue] (-12, 12) -- ( -4, 12);
\draw[blue] ( 4, 12) -- ( 12, 12);
\draw[blue] (-12,-12) -- ( -4,-12);
\draw[blue] ( 4,-12) -- ( 12,-12);
\draw[blue] ( 12,-12) -- ( 12, -4);
\draw[blue] ( 12, 4) -- ( 12, 12);
\draw[blue] (-12,-12) -- (-12, -4);
\draw[blue] (-12, 4) -- (-12, 12);
\draw[magenta] ( -4, 12) -- ( -4, 8) -- ( 4, 8) -- ( 4, 12);
\draw[magenta] ( -4,-12) -- ( -4, -8) -- ( 4, -8) -- ( 4,-12);
\draw[magenta] ( 12, -4) -- ( 8, -4) -- ( 8, 4) -- ( 12, 4);
\draw[magenta] (-12, -4) -- ( -8, -4) -- ( -8, 4) -- (-12, 4);
\draw[-angle 90,blue] ( 12, 12) -- ( 8, 12);
\draw[-angle 90,blue] (-12, 12) -- ( -8, 12);
\draw[-angle 90,blue] ( 4,-12) -- ( 8,-12);
\draw[-angle 90,blue] ( -4,-12) -- ( -8,-12);
\draw[-angle 90,blue] ( 12, 12) -- ( 12, 8);
\draw[-angle 90,blue] ( 12,-12) -- ( 12, -8);
\draw[-angle 90,blue] (-12, 4) -- (-12, 8);
\draw[-angle 90,blue] (-12, -4) -- (-12, -8);
\draw[-angle 90,magenta] ( 4, 8) -- ( 0, 8);
\draw[-angle 90,magenta] ( 4, -8) -- ( 0, -8);
\draw[-angle 90,magenta] ( -8, -4) -- ( -8, 0);
\draw[-angle 90,magenta] ( 8, -4) -- ( 8, 0);
\draw ( 8, 0) node[anchor=east] {$c_3$};
\draw ( -8, 0) node[anchor=west] {$c_3$};
\draw ( 0, 8) node[anchor=north] {$c_4$};
\draw ( 0, -8) node[anchor=south] {$c_4$};
\draw ( 12, -8) node[anchor=west] {$c_7^+$};
\draw ( 12, 8) node[anchor=west] {$c_7^+$};
\draw (-12, -8) node[anchor=east] {$c_7^-$};
\draw (-12, 8) node[anchor=east] {$c_7^-$};
\draw ( -8, 12) node[anchor=south] {$c_8^+$};
\draw ( 8, 12) node[anchor=south] {$c_8^+$};
\draw ( -8,-12) node[anchor=north] {$c_8^-$};
\draw ( 8,-12) node[anchor=north] {$c_8^-$};
\end{tikzpicture}
\end{center}
The gluing rules are as follows:
\begin{itemize}
\item The two edges marked $c_7^+$ are identified together.
\item The two edges marked $c_7^-$ are identified together.
\item The two edges marked $c_8^+$ are identified together.
\item The two edges marked $c_8^-$ are identified together.
\item The curve marked $c_3$ consisting of three edges at the left of
the diagram is identified with the corresponding curve at the right
of the diagram.
\item The curve marked $c_4$ consisting of three edges at the top of
the diagram is identified with the corresponding curve at the bottom
of the diagram.
\end{itemize}
The edges $c_k^+$ (for $k\in\{7,8\}$) become the arcs
$c_k([0,\pi])\sse X$, and the edges $c_k^-$ become the arcs
$c_k([-\pi,0])$.
Here are three more ways we can perform partial gluing to get a net
for $X$; we will call them $\Net_1$, $\Net_2$ and $\Net_3$.
\begin{center}
\begin{tikzpicture}[scale=1]
\draw[ green] ( 4, 4) -- ( 3, 3);
\draw[ blue] ( 4, 4) -- ( 0, 4);
\draw[ blue] ( 4, 4) -- ( 4, 0);
\draw[ green] ( 2, 2) -- ( 3, 3);
\draw[ blue] ( 2, 2) -- ( 0, 2);
\draw[ blue] ( 2, 2) -- ( 2, 0);
\draw[ cyan] ( 0, 3) -- ( -3, 3);
\draw[ cyan] ( 0, 3) -- ( 3, 3);
\draw[magenta] ( 0, 3) -- ( 0, 4);
\draw[magenta] ( 0, 3) -- ( 0, 2);
\draw[ cyan] ( -3, 0) -- ( -3, 3);
\draw[ cyan] ( -3, 0) -- ( -3, -3);
\draw[magenta] ( -3, 0) -- ( -4, 0);
\draw[magenta] ( -3, 0) -- ( -2, 0);
\draw[ cyan] ( 0, -3) -- ( -3, -3);
\draw[ cyan] ( 0, -3) -- ( 3, -3);
\draw[magenta] ( 0, -3) -- ( 0, -4);
\draw[magenta] ( 0, -3) -- ( 0, -2);
\draw[ cyan] ( 3, 0) -- ( 3, -3);
\draw[ cyan] ( 3, 0) -- ( 3, 3);
\draw[magenta] ( 3, 0) -- ( 4, 0);
\draw[magenta] ( 3, 0) -- ( 2, 0);
\draw[ green] ( -4, 4) -- ( -3, 3);
\draw[ blue] ( -4, 4) -- ( 0, 4);
\draw[ blue] ( -4, 4) -- ( -4, 0);
\draw[ green] ( -4, -4) -- ( -3, -3);
\draw[ blue] ( -4, -4) -- ( 0, -4);
\draw[ blue] ( -4, -4) -- ( -4, 0);
\draw[ green] ( 4, -4) -- ( 3, -3);
\draw[ blue] ( 4, -4) -- ( 4, 0);
\draw[ blue] ( 4, -4) -- ( 0, -4);
\draw[ green] ( -2, 2) -- ( -3, 3);
\draw[ blue] ( -2, 2) -- ( 0, 2);
\draw[ blue] ( -2, 2) -- ( -2, 0);
\draw[ green] ( -2, -2) -- ( -3, -3);
\draw[ blue] ( -2, -2) -- ( 0, -2);
\draw[ blue] ( -2, -2) -- ( -2, 0);
\draw[ green] ( 2, -2) -- ( 3, -3);
\draw[ blue] ( 2, -2) -- ( 2, 0);
\draw[ blue] ( 2, -2) -- ( 0, -2);
\draw ( 4, 4) node[anchor=south west]{$\ss 0$};
\draw ( 2, 2) node[anchor=north east]{$\ss 1$};
\draw ( 0, 3) node{$\ss 2$};
\draw ( -3, 0) node{$\ss 3$};
\draw ( 0, -3) node{$\ss 4$};
\draw ( 3, 0) node{$\ss 5$};
\draw ( -3, 3) node{$\ss 6$};
\draw ( -3, -3) node{$\ss 7$};
\draw ( 3, 3) node{$\ss 9$};
\draw ( 3, -3) node{$\ss 8$};
\draw ( 4, 0) node[anchor=west]{$\ss 11$};
\draw ( 0, 4) node[anchor=south]{$\ss 10$};
\draw ( 2, 0) node[anchor=east]{$\ss 13$};
\draw ( 0, 2) node[anchor=north]{$\ss 12$};
\draw ( 0, -4) node[anchor=north]{$\ss 10.1$};
\draw ( 0, -2) node[anchor=south]{$\ss 12.1$};
\draw ( -2, 2) node[anchor=north west]{$\ss 1.1$};
\draw ( 4, -4) node[anchor=north west]{$\ss 0.3$};
\draw ( -2, -2) node[anchor=south west]{$\ss 1.2$};
\draw ( -4, 0) node[anchor=east]{$\ss 11.1$};
\draw ( -4, 4) node[anchor=south east]{$\ss 0.1$};
\draw ( -2, 0) node[anchor=west]{$\ss 13.1$};
\draw ( 2, -2) node[anchor=south east]{$\ss 1.3$};
\draw ( -4, -4) node[anchor=north east]{$\ss 0.2$};
\draw (-3.50,1.75) node{$1$};
\draw (-1.75,-3.50) node{$\lm$};
\draw (3.50,-1.75) node{$\lm^2$};
\draw (1.75,3.50) node{$\lm^3$};
\draw (2.50,1.25) node{$\mu$};
\draw (-1.25,2.50) node{$\lm\mu$};
\draw (-2.50,-1.25) node{$\lm^2\mu$};
\draw (1.25,-2.50) node{$\lm^3\mu$};
\draw (3.50,1.75) node{$\nu$};
\draw (-1.75,3.50) node{$\lm\nu$};
\draw (-3.50,-1.75) node{$\lm^2\nu$};
\draw (1.75,-3.50) node{$\lm^3\nu$};
\draw (-2.50,1.25) node{$\mu\nu$};
\draw (-1.25,-2.50) node{$\lm\mu\nu$};
\draw (2.50,-1.25) node{$\lm^2\mu\nu$};
\draw (1.25,2.50) node{$\lm^3\mu\nu$};
\end{tikzpicture}
\end{center}
\vspace{2ex}
\begin{center}
\begin{tikzpicture}[scale=.4]
\draw[ green] ( -4, 0) -- ( 0, 0);
\draw[ blue] ( -4, 0) -- ( -4, 4);
\draw[ blue] ( -4, 0) -- ( -4, -4);
\draw[ green] ( -4, 0) -- ( -8, 1);
\draw[ green] ( -4, 0) -- ( -8, -1);
\draw[ green] ( 4, 0) -- ( 0, 0);
\draw[ green] ( 4, 0) -- ( 8, 1);
\draw[ green] ( 4, 0) -- ( 8, -1);
\draw[ blue] ( 4, 0) -- ( 4, 4);
\draw[ blue] ( 4, 0) -- ( 4, -4);
\draw[ cyan] ( 0, 4) -- ( 0, 0);
\draw[magenta] ( 0, 4) -- ( -4, 4);
\draw[magenta] ( 0, 4) -- ( 4, 4);
\draw[ cyan] ( 0, 4) -- ( 0, 8);
\draw[ cyan] ( 0, -4) -- ( 0, 0);
\draw[magenta] ( 0, -4) -- ( -4, -4);
\draw[magenta] ( 0, -4) -- ( 4, -4);
\draw[ cyan] ( 0, -4) -- ( 0, -8);
\draw[ cyan] ( 8, 4) -- ( 8, 1);
\draw[ cyan] ( 8, 4) -- ( 8, 8);
\draw[magenta] ( 8, 4) -- ( 4, 4);
\draw[ cyan] ( 8, -4) -- ( 8, -1);
\draw[magenta] ( 8, -4) -- ( 4, -4);
\draw[ cyan] ( 8, -4) -- ( 8, -8);
\draw[ blue] ( -4, 8) -- ( -4, 4);
\draw[ green] ( -4, 8) -- ( -8, 8);
\draw[ green] ( -4, 8) -- ( 0, 8);
\draw[ blue] ( -4, -8) -- ( -4, -4);
\draw[ green] ( -4, -8) -- ( 0, -8);
\draw[ green] ( -4, -8) -- ( -8, -8);
\draw[ green] ( 4, 8) -- ( 8, 8);
\draw[ blue] ( 4, 8) -- ( 4, 4);
\draw[ green] ( 4, 8) -- ( 0, 8);
\draw[ blue] ( 4, -8) -- ( 4, -4);
\draw[ green] ( 4, -8) -- ( 0, -8);
\draw[ green] ( 4, -8) -- ( 8, -8);
\draw[magenta] ( -8, 4) -- ( -4, 4);
\draw[ cyan] ( -8, 4) -- ( -8, 1);
\draw[ cyan] ( -8, 4) -- ( -8, 8);
\draw[magenta] ( -8, -4) -- ( -4, -4);
\draw[ cyan] ( -8, -4) -- ( -8, -8);
\draw[ cyan] ( -8, -4) -- ( -8, -1);
\draw ( -4, 0) node[anchor=east]{$\ss 0$};
\draw ( -8, 4) node[anchor=east]{$\ss 4.1$};
\draw ( 4, 0) node[anchor=west]{$\ss 1$};
\draw ( 0, 4) node{$\ss 2$};
\draw ( 0, -4) node{$\ss 3$};
\draw ( 8, 4) node[anchor=west]{$\ss 4$};
\draw ( 8, -4) node[anchor=west]{$\ss 5$};
\draw ( 0, 0) node{$\ss 6$};
\draw ( 8, 1) node[anchor=west]{$\ss 7$};
\draw ( 8, -1) node[anchor=west]{$\ss 9$};
\draw ( 8, 8) node[anchor=south west]{$\ss 8$};
\draw ( -8, 1) node[anchor=east]{$\ss 7.1$};
\draw ( -4, -4) node{$\ss 11$};
\draw ( -4, 4) node{$\ss 10$};
\draw ( 4, -4) node{$\ss 13$};
\draw ( 4, 4) node{$\ss 12$};
\draw ( -8, 8) node[anchor=south east]{$\ss 8.1$};
\draw ( -8, -4) node[anchor=east]{$\ss 5.1$};
\draw ( 0, -8) node[anchor=north]{$\ss 7.2$};
\draw ( -8, -8) node[anchor=north east]{$\ss 8.2$};
\draw ( 4, 8) node[anchor=south]{$\ss 1.1$};
\draw ( 8, -8) node[anchor=north west]{$\ss 8.3$};
\draw ( 4, -8) node[anchor=north]{$\ss 1.2$};
\draw ( -4, 8) node[anchor=south]{$\ss 0.1$};
\draw ( -8, -1) node[anchor=east]{$\ss 9.1$};
\draw ( 0, 8) node[anchor=south]{$\ss 9.2$};
\draw ( -4, -8) node[anchor=north]{$\ss 0.2$};
\draw (-2.00,-2.00) node{$1$};
\draw (-6.00,2.25) node{$\lm$};
\draw (-6.00,-6.00) node{$\lm^2$};
\draw (-2.00,6.00) node{$\lm^3$};
\draw (6.00,-2.25) node{$\mu$};
\draw (2.00,2.00) node{$\lm\mu$};
\draw (2.00,-6.00) node{$\lm^2\mu$};
\draw (6.00,6.00) node{$\lm^3\mu$};
\draw (-6.00,-2.25) node{$\nu$};
\draw (-2.00,2.00) node{$\lm\nu$};
\draw (-2.00,-6.00) node{$\lm^2\nu$};
\draw (-6.00,6.00) node{$\lm^3\nu$};
\draw (2.00,-2.00) node{$\mu\nu$};
\draw (6.00,2.25) node{$\lm\mu\nu$};
\draw (6.00,-6.00) node{$\lm^2\mu\nu$};
\draw (2.00,6.00) node{$\lm^3\mu\nu$};
\end{tikzpicture}
\hspace{1em}
\begin{tikzpicture}[scale=.4]
\draw[ green] ( 8, 4) -- ( 4, 4);
\draw[ blue] ( 8, 4) -- ( 8, 8);
\draw[ blue] ( 8, 4) -- ( 8, 0);
\draw[ green] ( 0, 4) -- ( -4, 4);
\draw[ green] ( 0, 4) -- ( 4, 4);
\draw[ blue] ( 0, 4) -- ( 1, 8);
\draw[ blue] ( 0, 4) -- ( 0, 0);
\draw[ blue] ( 0, 4) -- ( -1, 8);
\draw[ cyan] ( 4, 8) -- ( 4, 4);
\draw[magenta] ( 4, 8) -- ( 8, 8);
\draw[magenta] ( 4, 8) -- ( 1, 8);
\draw[ cyan] ( -4, 0) -- ( -4, 4);
\draw[ cyan] ( -4, 0) -- ( -4, -4);
\draw[magenta] ( -4, 0) -- ( 0, 0);
\draw[magenta] ( -4, 0) -- ( -8, 0);
\draw[ cyan] ( 4, -8) -- ( 4, -4);
\draw[magenta] ( 4, -8) -- ( 8, -8);
\draw[magenta] ( 4, -8) -- ( 1, -8);
\draw[ cyan] ( 4, 0) -- ( 4, -4);
\draw[ cyan] ( 4, 0) -- ( 4, 4);
\draw[magenta] ( 4, 0) -- ( 8, 0);
\draw[magenta] ( 4, 0) -- ( 0, 0);
\draw[ green] ( -8, 4) -- ( -4, 4);
\draw[ blue] ( -8, 4) -- ( -8, 8);
\draw[ blue] ( -8, 4) -- ( -8, 0);
\draw[ green] ( -8, -4) -- ( -4, -4);
\draw[ blue] ( -8, -4) -- ( -8, -8);
\draw[ blue] ( -8, -4) -- ( -8, 0);
\draw[ green] ( 8, -4) -- ( 4, -4);
\draw[ blue] ( 8, -4) -- ( 8, 0);
\draw[ blue] ( 8, -4) -- ( 8, -8);
\draw[ green] ( 0, -4) -- ( -4, -4);
\draw[ green] ( 0, -4) -- ( 4, -4);
\draw[ blue] ( 0, -4) -- ( 0, 0);
\draw[ blue] ( 0, -4) -- ( -1, -8);
\draw[ blue] ( 0, -4) -- ( 1, -8);
\draw[ cyan] ( -4, 8) -- ( -4, 4);
\draw[magenta] ( -4, 8) -- ( -8, 8);
\draw[magenta] ( -4, 8) -- ( -1, 8);
\draw[ cyan] ( -4, -8) -- ( -4, -4);
\draw[magenta] ( -4, -8) -- ( -8, -8);
\draw[magenta] ( -4, -8) -- ( -1, -8);
\draw ( 8, 4) node[anchor=west]{$\ss 0$};
\draw ( -4, -8) node[anchor=north]{$\ss 4.1$};
\draw ( 0, 4) node[anchor=south]{$\ss 1$};
\draw ( 4, 8) node[anchor=south]{$\ss 2$};
\draw ( -4, 0) node{$\ss 3$};
\draw ( 4, -8) node[anchor=north]{$\ss 4$};
\draw ( 4, 0) node{$\ss 5$};
\draw ( -4, 4) node{$\ss 6$};
\draw ( -4, -4) node{$\ss 7$};
\draw ( 4, 4) node{$\ss 9$};
\draw ( 4, -4) node{$\ss 8$};
\draw ( 8, 0) node[anchor=west]{$\ss 11$};
\draw ( 8, 8) node[anchor=south west]{$\ss 10$};
\draw ( 0, 0) node{$\ss 13$};
\draw ( 1, 8) node[anchor=south]{$\ss 12$};
\draw ( -8, 8) node[anchor=south east]{$\ss 10.1$};
\draw ( 8, -8) node[anchor=north west]{$\ss 10.3$};
\draw ( -1, 8) node[anchor=south]{$\ss 12.1$};
\draw ( -8, -8) node[anchor=north east]{$\ss 10.2$};
\draw ( 0, -4) node[anchor=north]{$\ss 1.1$};
\draw ( 8, -4) node[anchor=west]{$\ss 0.3$};
\draw ( -8, 0) node[anchor=east]{$\ss 11.1$};
\draw ( -8, 4) node[anchor=east]{$\ss 0.1$};
\draw ( 1, -8) node[anchor=north]{$\ss 12.3$};
\draw ( -4, 8) node[anchor=south]{$\ss 2.1$};
\draw ( -8, -4) node[anchor=east]{$\ss 0.2$};
\draw ( -1, -8) node[anchor=north]{$\ss 12.2$};
\draw (-6.00,2.00) node{$1$};
\draw (-6.00,-6.00) node{$\lm$};
\draw (6.00,-2.00) node{$\lm^2$};
\draw (6.00,6.00) node{$\lm^3$};
\draw (2.00,2.00) node{$\mu$};
\draw (-2.25,6.00) node{$\lm\mu$};
\draw (-2.00,-2.00) node{$\lm^2\mu$};
\draw (2.25,-6.00) node{$\lm^3\mu$};
\draw (6.00,2.00) node{$\nu$};
\draw (-6.00,6.00) node{$\lm\nu$};
\draw (-6.00,-2.00) node{$\lm^2\nu$};
\draw (6.00,-6.00) node{$\lm^3\nu$};
\draw (-2.00,2.00) node{$\mu\nu$};
\draw (-2.25,-6.00) node{$\lm\mu\nu$};
\draw (2.00,-2.00) node{$\lm^2\mu\nu$};
\draw (2.25,6.00) node{$\lm^3\mu\nu$};
\end{tikzpicture}
\end{center}
\begin{checks}
nets_check.mpl: check_nets()
\end{checks}
In all of $\Net_0,\dotsc,\Net_3$, the labels $0$, $11$, $3$ and $6$ occur
in anticlockwise order around the region marked $1$. This shows that
all the nets have orientation compatible with each other and thus also
compatible with the orientation of $X$.
Here is another way to assemble the pieces. The left hand picture
(which we call $\Net_4^+$) consists of eight distorted copies of
$F_{16}$, and is homeomorphic to a disc with two holes, or a ``pair of
pants''. The right hand picture ($\Net_4^-$) consists of the other eight
translates of $F_{16}$. The surface can be obtained by gluing the two
pictures together along $C_3\amalg C_6\amalg C_8$: this is a ``pair of
pants decomposition''.
\begin{center}
\begin{tikzpicture}[scale=0.8]
\draw[magenta] (-4,-3) -- ( 4,-3) -- ( 4, 3) -- (-4, 3) -- (-4,-3);
\draw[green] (-3,-2) -- ( 3,-2) -- ( 3, 2) -- (-3, 2) -- (-3,-2);
\draw[cyan] ( 0,-3) -- ( 0, 3);
\draw[blue] (-4, 0) -- (-3, 0);
\draw[blue] ( 3, 0) -- ( 4, 0);
\draw[magenta] (-1, 0) -- ( 1, 0);
\filldraw[draw=blue,fill=gray!20]
(-3, 0) -- (-2, 1) -- (-1, 0) -- (-2,-1) -- (-3, 0);
\filldraw[draw=blue,fill=gray!20]
( 3, 0) -- ( 2, 1) -- ( 1, 0) -- ( 2,-1) -- ( 3, 0);
\fill[black] (-4, 0) circle(0.04);
\fill[black] (-3, 0) circle(0.04);
\fill[black] (-1, 0) circle(0.04);
\fill[black] ( 0, 0) circle(0.04);
\fill[black] ( 1, 0) circle(0.04);
\fill[black] ( 3, 0) circle(0.04);
\fill[black] ( 4, 0) circle(0.04);
\fill[black] ( 0,-3) circle(0.04);
\fill[black] ( 0,-2) circle(0.04);
\fill[black] ( 0, 2) circle(0.04);
\fill[black] ( 0, 3) circle(0.04);
\draw (-4, 0) node[anchor=east] {$\ss v_{13}$};
\draw (-3, 0) node[anchor=west] {$\ss v_1$};
\draw (-1, 0) node[anchor=east] {$\ss v_{12}$};
\draw ( 0, 0) node[anchor=north east] {$\ss v_2$};
\draw ( 1, 0) node[anchor=west] {$\ss v_{10}$};
\draw ( 3, 0) node[anchor=east] {$\ss v_0$};
\draw ( 4, 0) node[anchor=west] {$\ss v_{11}$};
\draw ( 0,-3) node[anchor=north] {$\ss v_5$};
\draw ( 0,-2) node[anchor=north east] {$\ss v_9$};
\draw ( 0, 2) node[anchor=south east] {$\ss v_6$};
\draw ( 0, 3) node[anchor=south] {$\ss v_3$};
\draw ( 3.5, 2.5) node {$\ss 1$};
\draw ( 3.5,-2.5) node {$\ss \nu$};
\draw (-3.5, 2.5) node {$\ss \mu\nu$};
\draw (-3.5,-2.5) node {$\ss \mu$};
\draw ( 0.9, 1.2) node {$\ss \lm\nu$};
\draw ( 0.9,-1.2) node {$\ss \lm^3$};
\draw (-0.9, 1.2) node {$\ss \lm\mu$};
\draw (-0.9,-1.2) node {$\ss \lm^3\mu\nu$};
\draw[->,cyan] ( 0.0, 2.4) -- ( 0.0, 2.5);
\draw[->,cyan] ( 0.0, 0.9) -- ( 0.0, 1.0);
\draw[->,cyan] ( 0.0,-1.1) -- ( 0.0,-1.0);
\draw[->,cyan] ( 0.0,-2.6) -- ( 0.0,-2.5);
\draw ( 0.0, 0.9) node[anchor=west] {$\ss c_0$};
\draw[->,green] ( 1.6, 2.0) -- ( 1.5, 2.0);
\draw[->,green] (-1.4, 2.0) -- (-1.5, 2.0);
\draw ( 1.5, 2.0) node[anchor=south] {$\ss c_1$};
\draw (-1.5, 2.0) node[anchor=south] {$\ss c_1$};
\draw[->,green] ( 1.4,-2.0) -- ( 1.5,-2.0);
\draw[->,green] (-1.6,-2.0) -- (-1.5,-2.0);
\draw ( 1.5,-2.0) node[anchor=north] {$\ss c_2$};
\draw (-1.5,-2.0) node[anchor=north] {$\ss c_2$};
\draw[->,magenta] ( 4.0, 1.4) -- ( 4.0, 1.5);
\draw[->,magenta] ( 4.0,-1.6) -- ( 4.0,-1.5);
\draw[->,magenta] (-4.0, 1.6) -- (-4.0, 1.5);
\draw[->,magenta] (-4.0,-1.4) -- (-4.0,-1.5);
\draw ( 4.0, 1.5) node[anchor=west] {$\ss c_3$};
\draw ( 4.0,-1.5) node[anchor=west] {$\ss c_3$};
\draw (-4.0, 1.5) node[anchor=east] {$\ss c_3$};
\draw (-4.0,-1.5) node[anchor=east] {$\ss c_3$};
\draw[->,blue] ( 3.4, 0.0) -- ( 3.5, 0.0);
\draw ( 3.5, 0.0) node[anchor=north] {$\ss c_5$};
\draw[->,blue] (-3.4, 0.0) -- (-3.5, 0.0);
\draw (-3.5, 0.0) node[anchor=north] {$\ss c_7$};
\draw[->,blue] ( 2.6, 0.4) -- ( 2.5, 0.5);
\draw[->,blue] ( 1.6, 0.6) -- ( 1.5, 0.5);
\draw[->,blue] ( 1.4,-0.4) -- ( 1.5,-0.5);
\draw[->,blue] ( 2.4,-0.6) -- ( 2.5,-0.5);
\draw ( 2.5, 0.5) node[anchor=south west] {$\ss c_6$};
\draw[->,blue] (-2.6, 0.4) -- (-2.5, 0.5);
\draw[->,blue] (-1.6, 0.6) -- (-1.5, 0.5);
\draw[->,blue] (-1.4,-0.4) -- (-1.5,-0.5);
\draw[->,blue] (-2.4,-0.6) -- (-2.5,-0.5);
\draw (-2.5, 0.5) node[anchor=south east] {$\ss c_8$};
\draw[->,magenta] (-0.6, 0.0) -- (-0.5, 0.0);
\draw[->,magenta] ( 0.4, 0.0) -- ( 0.5, 0.0);
\draw ( 0.5, 0.0) node[anchor=north] {$\ss c_4$};
\end{tikzpicture}
\qquad
\begin{tikzpicture}[scale=0.8]
\draw[magenta] (-4,-3) -- ( 4,-3) -- ( 4, 3) -- (-4, 3) -- (-4,-3);
\draw[green] (-3,-2) -- ( 3,-2) -- ( 3, 2) -- (-3, 2) -- (-3,-2);
\draw[cyan] ( 0,-3) -- ( 0, 3);
\draw[blue] (-4, 0) -- (-3, 0);
\draw[blue] ( 3, 0) -- ( 4, 0);
\draw[magenta] (-1, 0) -- ( 1, 0);
\filldraw[draw=blue,fill=gray!20]
(-3, 0) -- (-2, 1) -- (-1, 0) -- (-2,-1) -- (-3, 0);
\filldraw[draw=blue,fill=gray!20]
( 3, 0) -- ( 2, 1) -- ( 1, 0) -- ( 2,-1) -- ( 3, 0);
\fill[black] (-4, 0) circle(0.04);
\fill[black] (-3, 0) circle(0.04);
\fill[black] (-1, 0) circle(0.04);
\fill[black] ( 0, 0) circle(0.04);
\fill[black] ( 1, 0) circle(0.04);
\fill[black] ( 3, 0) circle(0.04);
\fill[black] ( 4, 0) circle(0.04);
\fill[black] ( 0,-3) circle(0.04);
\fill[black] ( 0,-2) circle(0.04);
\fill[black] ( 0, 2) circle(0.04);
\fill[black] ( 0, 3) circle(0.04);
\draw (-4, 0) node[anchor=east] {$\ss v_{13}$};
\draw (-3, 0) node[anchor=west] {$\ss v_1$};
\draw (-1, 0) node[anchor=east] {$\ss v_{12}$};
\draw ( 0, 0) node[anchor=north east] {$\ss v_4$};
\draw ( 1, 0) node[anchor=west] {$\ss v_{10}$};
\draw ( 3, 0) node[anchor=east] {$\ss v_0$};
\draw ( 4, 0) node[anchor=west] {$\ss v_{11}$};
\draw ( 0,-3) node[anchor=north] {$\ss v_5$};
\draw ( 0,-2) node[anchor=north east] {$\ss v_8$};
\draw ( 0, 2) node[anchor=south east] {$\ss v_7$};
\draw ( 0, 3) node[anchor=south] {$\ss v_3$};
\draw ( 3.5, 2.5) node {$\ss \lm^2\nu$};
\draw ( 3.5,-2.5) node {$\ss \lm^2$};
\draw (-3.5, 2.5) node {$\ss \lm^2\mu$};
\draw (-3.5,-2.5) node {$\ss \lm^2\mu\nu$};
\draw ( 0.9, 1.2) node {$\ss \lm$};
\draw ( 0.9,-1.2) node {$\ss \lm^3\nu$};
\draw (-0.9, 1.2) node {$\ss \lm\mu\nu$};
\draw (-0.9,-1.2) node {$\ss \lm^3\mu$};
\draw[->,cyan] ( 0.0, 2.6) -- ( 0.0, 2.5);
\draw[->,cyan] ( 0.0, 1.1) -- ( 0.0, 1.0);
\draw[->,cyan] ( 0.0,-0.9) -- ( 0.0,-1.0);
\draw[->,cyan] ( 0.0,-2.4) -- ( 0.0,-2.5);
\draw ( 0.0, 0.9) node[anchor=west] {$\ss c_0$};
\draw[->,green] ( 1.6, 2.0) -- ( 1.5, 2.0);
\draw[->,green] (-1.4, 2.0) -- (-1.5, 2.0);
\draw ( 1.5, 2.0) node[anchor=south] {$\ss c_2$};
\draw (-1.5, 2.0) node[anchor=south] {$\ss c_2$};
\draw[->,green] ( 1.4,-2.0) -- ( 1.5,-2.0);
\draw[->,green] (-1.6,-2.0) -- (-1.5,-2.0);
\draw ( 1.5,-2.0) node[anchor=north] {$\ss c_1$};
\draw (-1.5,-2.0) node[anchor=north] {$\ss c_1$};
\draw[->,magenta] ( 4.0, 1.4) -- ( 4.0, 1.5);
\draw[->,magenta] ( 4.0,-1.6) -- ( 4.0,-1.5);
\draw[->,magenta] (-4.0, 1.6) -- (-4.0, 1.5);
\draw[->,magenta] (-4.0,-1.4) -- (-4.0,-1.5);
\draw ( 4.0, 1.5) node[anchor=west] {$\ss c_3$};
\draw ( 4.0,-1.5) node[anchor=west] {$\ss c_3$};
\draw (-4.0, 1.5) node[anchor=east] {$\ss c_3$};
\draw (-4.0,-1.5) node[anchor=east] {$\ss c_3$};
\draw[->,blue] ( 3.6, 0.0) -- ( 3.5, 0.0);
\draw ( 3.5, 0.0) node[anchor=north] {$\ss c_5$};
\draw[->,blue] (-3.6, 0.0) -- (-3.5, 0.0);
\draw (-3.5, 0.0) node[anchor=north] {$\ss c_7$};
\draw[->,blue] ( 2.6, 0.4) -- ( 2.5, 0.5);
\draw[->,blue] ( 1.6, 0.6) -- ( 1.5, 0.5);
\draw[->,blue] ( 1.4,-0.4) -- ( 1.5,-0.5);
\draw[->,blue] ( 2.4,-0.6) -- ( 2.5,-0.5);
\draw ( 2.5, 0.5) node[anchor=south west] {$\ss c_6$};
\draw[->,blue] (-2.6, 0.4) -- (-2.5, 0.5);
\draw[->,blue] (-1.6, 0.6) -- (-1.5, 0.5);
\draw[->,blue] (-1.4,-0.4) -- (-1.5,-0.5);
\draw[->,blue] (-2.4,-0.6) -- (-2.5,-0.5);
\draw (-2.5, 0.5) node[anchor=south east] {$\ss c_8$};
\draw[->,magenta] (-0.4, 0.0) -- (-0.5, 0.0);
\draw[->,magenta] ( 0.6, 0.0) -- ( 0.5, 0.0);
\draw ( 0.5, 0.0) node[anchor=north] {$\ss c_4$};
\end{tikzpicture}
\end{center}
In this case the orientation of $\Net_4^+$ is compatible with the
orientation of $X$, but the orientation of $\Net_4^-$ is reversed.
We now give another net which we call $\Net_5$. Note that the central
octagon is the same as for $\Net_0$, but the outer pieces have been
rearranged.
\begin{center}
\begin{tikzpicture}[scale=1]
\draw[ green] ( 3, 2) -- ( 2, 2);
\draw[ blue] ( 3, 2) -- ( 3, 5);
\draw[ green] ( 0, 0) -- ( 2, 2);
\draw[ green] ( 0, 0) -- ( 2, -2);
\draw[ green] ( 0, 0) -- ( -2, -2);
\draw[ green] ( 0, 0) -- ( -2, 2);
\draw[ blue] ( 0, 0) -- ( 3, 0);
\draw[ blue] ( 0, 0) -- ( 0, 3);
\draw[ blue] ( 0, 0) -- ( -3, 0);
\draw[ blue] ( 0, 0) -- ( 0, -3);
\draw[ cyan] ( 3, 1) -- ( 2, 2);
\draw[magenta] ( 3, 1) -- ( 3, 0);
\draw[ cyan] ( 1, -3) -- ( 2, -2);
\draw[magenta] ( 1, -3) -- ( 0, -3);
\draw[ cyan] ( -3, -1) -- ( -2, -2);
\draw[magenta] ( -3, -1) -- ( -3, 0);
\draw[ cyan] ( -1, 3) -- ( -2, 2);
\draw[magenta] ( -1, 3) -- ( 0, 3);
\draw[ blue] ( 2, 6) -- ( 3, 5);
\draw[ green] ( 2, 6) -- ( 1, 6);
\draw[ green] ( -2, 3) -- ( -2, 2);
\draw[ blue] ( -2, 3) -- ( -5, 3);
\draw[ green] ( -6, 2) -- ( -6, 1);
\draw[ blue] ( -6, 2) -- ( -5, 3);
\draw[ green] ( -3, -2) -- ( -2, -2);
\draw[ blue] ( -3, -2) -- ( -3, -5);
\draw[ green] ( -2, -6) -- ( -1, -6);
\draw[ blue] ( -2, -6) -- ( -3, -5);
\draw[ green] ( 2, -3) -- ( 2, -2);
\draw[ blue] ( 2, -3) -- ( 5, -3);
\draw[ blue] ( 6, -2) -- ( 5, -3);
\draw[ green] ( 6, -2) -- ( 6, -1);
\draw[ cyan] ( -3, 1) -- ( -2, 2);
\draw[ cyan] ( -3, 1) -- ( -6, 1);
\draw[magenta] ( -3, 1) -- ( -5, 3);
\draw[magenta] ( -3, 1) -- ( -3, 0);
\draw[ cyan] ( 1, 3) -- ( 2, 2);
\draw[magenta] ( 1, 3) -- ( 3, 5);
\draw[magenta] ( 1, 3) -- ( 0, 3);
\draw[ cyan] ( 1, 3) -- ( 1, 6);
\draw[ cyan] ( 3, -1) -- ( 2, -2);
\draw[magenta] ( 3, -1) -- ( 5, -3);
\draw[magenta] ( 3, -1) -- ( 3, 0);
\draw[ cyan] ( 3, -1) -- ( 6, -1);
\draw[ cyan] ( -1, -3) -- ( -2, -2);
\draw[ cyan] ( -1, -3) -- ( -1, -6);
\draw[magenta] ( -1, -3) -- ( -3, -5);
\draw[magenta] ( -1, -3) -- ( 0, -3);
\draw ( 3, 2) node[anchor=west]{$\ss 0$};
\draw ( 0, 0) node[anchor=east]{$\ss 1$};
\draw ( 3, 1) node[anchor=west]{$\ss 2$};
\draw ( 1, -3) node[anchor=north]{$\ss 3$};
\draw ( -3, -1) node[anchor=east]{$\ss 4$};
\draw ( -5, 3) node[anchor=south]{$\ss 10.1$};
\draw ( -1, 3) node[anchor=south]{$\ss 5$};
\draw ( -6, 2) node[anchor=south east]{$\ss 0.3$};
\draw ( 2, 2) node[anchor=south]{$\ss 6$};
\draw ( 2, -2) node[anchor=west]{$\ss 7$};
\draw ( -2, 2) node[anchor=east]{$\ss 9$};
\draw ( -2, -2) node[anchor=north]{$\ss 8$};
\draw ( 3, 5) node[anchor=west]{$\ss 11$};
\draw ( 5, -3) node[anchor=north]{$\ss 10$};
\draw ( 0, 3) node[anchor=south]{$\ss 13$};
\draw ( 3, 0) node[anchor=west]{$\ss 12$};
\draw ( -6, 1) node[anchor=north]{$\ss 6.1$};
\draw ( -1, -3) node[anchor=north west]{$\ss 5.1$};
\draw ( 2, -3) node[anchor=north]{$\ss 0.6$};
\draw ( -1, -6) node[anchor=west]{$\ss 9.1$};
\draw ( 3, -1) node[anchor=south west]{$\ss 4.1$};
\draw ( -3, -2) node[anchor=east]{$\ss 0.4$};
\draw ( 1, 3) node[anchor=south east]{$\ss 3.1$};
\draw ( 6, -2) node[anchor=north west]{$\ss 0.7$};
\draw ( 0, -3) node[anchor=north]{$\ss 13.1$};
\draw ( 2, 6) node[anchor=south west]{$\ss 0.1$};
\draw ( 1, 6) node[anchor=east]{$\ss 7.1$};
\draw ( -2, 3) node[anchor=south]{$\ss 0.2$};
\draw ( -3, 1) node[anchor=north east]{$\ss 2.1$};
\draw ( -2, -6) node[anchor=north east]{$\ss 0.5$};
\draw ( -3, 0) node[anchor=east]{$\ss 12.1$};
\draw ( 6, -1) node[anchor=south]{$\ss 8.1$};
\draw ( -3, -5) node[anchor=east]{$\ss 11.1$};
\draw (2.25,3.00) node{$1$};
\draw (3.00,-2.25) node{$\lm$};
\draw (-2.25,-3.00) node{$\lm^2$};
\draw (-3.00,2.25) node{$\lm^3$};
\draw (-0.75,2.00) node{$\mu$};
\draw (2.00,0.75) node{$\lm\mu$};
\draw (0.75,-2.00) node{$\lm^2\mu$};
\draw (-2.00,-0.75) node{$\lm^3\mu$};
\draw (-1.75,-5.00) node{$\nu$};
\draw (-5.00,1.75) node{$\lm\nu$};
\draw (1.75,5.00) node{$\lm^2\nu$};
\draw (5.00,-1.75) node{$\lm^3\nu$};
\draw (0.75,2.00) node{$\mu\nu$};
\draw (2.00,-0.75) node{$\lm\mu\nu$};
\draw (-0.75,-2.00) node{$\lm^2\mu\nu$};
\draw (-2.00,0.75) node{$\lm^3\mu\nu$};
\end{tikzpicture}
\end{center}
The point about $\Net_5$ is that it allows us to read off a convenient
presentation of the fundamental group $\pi_1(X,v_0)$.
\begin{definition}
We define $\bt_i\:[0,1]\to X$ for $i\in\Z/8$ by
\begin{align*}
\bt_0(t) &= c_5(2\pi t) \\
\bt_1(t) &= \begin{cases}
c_1( -3t \pi ) & 0 \leq t \leq 1/6 \\
c_0((-1-3t)\pi/2) & 1/6 \leq t \leq 2/6 \\
c_4(( 6t-1)\pi/2) & 2/6 \leq t \leq 4/6 \\
c_0(( 3t-2)\pi/2) & 4/6 \leq t \leq 5/6 \\
c_1(( 3-3t)\pi ) & 5/6 \leq t \leq 1,
\end{cases}
\end{align*}
then $\bt_{i+2j}(t)=\lm^j\bt_i(t)$ for $i\in\{0,1\}$ and
$j\in\{1,2,3\}$.
\end{definition}
It is straightforward to check that $\bt_k(0)=\bt_k(1)=v_0$ for all
$k$, so $\bt_k$ represents an element of $\pi_1(X,v_0)$.
\begin{proposition}\label{prop-pi-one}
$\pi_1(X,v_0)$ is generated by the elements $\bt_i$, subject only to
the relations $\bt_i\bt_{i+4}=1$ and
\[ \bt_0\bt_1\bt_2\bt_3\bt_4\bt_5\bt_6\bt_7 = 1. \]
\end{proposition}
\begin{proof}
Inspection of the definitions, together with part~(c) of
Definition~\ref{defn-curve-system}, shows that
$\bt_{i+4}(t)=\bt_i(1-t)$ for all $i$, so $\bt_{i+4}$ is inverse to
$\bt_i$ in $\pi_1(X,v_0)$. The paths $\bt_i$ appear in $\Net_5$ as
follows:
\begin{center}
\begin{tikzpicture}[scale=0.6]
\draw[ blue,-<] ( 3, 2) -- ( 3, 4);
\draw[ blue,-<] ( -2, 3) -- ( -4, 3);
\draw[ blue,-<] ( -3, -2) -- ( -3, -4);
\draw[ blue,-<] ( 2, -3) -- ( 4, -3);
\draw[orange,-<] ( 3, -1) -- ( 3, 0);
\draw[orange,-<] ( 1, 3) -- ( 0, 3);
\draw[orange,-<] ( -3, 1) -- ( -3, 0);
\draw[orange,-<] ( -1, -3) -- ( 0, -3);
%
\draw[ orange] ( 3, 2) -- ( 2, 2);
\draw[ blue] ( 3, 2) -- ( 3, 5);
\draw[ orange] ( 3, 1) -- ( 2, 2);
\draw[ orange] ( 3, 1) -- ( 3, 0);
\draw[ orange] ( 1, -3) -- ( 2, -2);
\draw[ orange] ( 1, -3) -- ( 0, -3);
\draw[ orange] ( -3, -1) -- ( -2, -2);
\draw[ orange] ( -3, -1) -- ( -3, 0);
\draw[ orange] ( -1, 3) -- ( -2, 2);
\draw[ orange] ( -1, 3) -- ( 0, 3);
\draw[ blue] ( 2, 6) -- ( 3, 5);
\draw[ orange] ( 2, 6) -- ( 1, 6);
\draw[ orange] ( -2, 3) -- ( -2, 2);
\draw[ blue] ( -2, 3) -- ( -5, 3);
\draw[ orange] ( -6, 2) -- ( -6, 1);
\draw[ blue] ( -6, 2) -- ( -5, 3);
\draw[ orange] ( -3, -2) -- ( -2, -2);
\draw[ blue] ( -3, -2) -- ( -3, -5);
\draw[ orange] ( -2, -6) -- ( -1, -6);
\draw[ blue] ( -2, -6) -- ( -3, -5);
\draw[ orange] ( 2, -3) -- ( 2, -2);
\draw[ blue] ( 2, -3) -- ( 5, -3);
\draw[ blue] ( 6, -2) -- ( 5, -3);
\draw[ orange] ( 6, -2) -- ( 6, -1);
\draw[ orange] ( -3, 1) -- ( -6, 1);
\draw[ orange] ( -3, 1) -- ( -3, 0);
\draw[ orange] ( 1, 3) -- ( 0, 3);
\draw[ orange] ( 1, 3) -- ( 1, 6);
\draw[ orange] ( 3, -1) -- ( 3, 0);
\draw[ orange] ( 3, -1) -- ( 6, -1);
\draw[ orange] ( -1, -3) -- ( -1, -6);
\draw[ orange] ( -1, -3) -- ( 0, -3);
\draw ( -3, -5) node[anchor=south west]{$\bt_0$};
\draw ( -3, -1) node[anchor=south west]{$\bt_1$};
\draw ( -5, 3) node[anchor=north west]{$\bt_2$};
\draw ( -1, 3) node[anchor=north west]{$\bt_3$};
\draw ( 3, 5) node[anchor=north east]{$\bt_4$};
\draw ( 3, 1) node[anchor=north east]{$\bt_5$};
\draw ( 5, -3) node[anchor=south east]{$\bt_6$};
\draw ( 1, -3) node[anchor=south east]{$\bt_7$};
%
\fill ( 3, 2) circle(0.05);
\fill ( -6, 2) circle(0.05);
\fill ( 2, -3) circle(0.05);
\fill ( -3, -2) circle(0.05);
\fill ( 6, -2) circle(0.05);
\fill ( 2, 6) circle(0.05);
\fill ( -2, 3) circle(0.05);
\fill ( -2, -6) circle(0.05);
\end{tikzpicture}
\end{center}
The surface $X$ can be recovered by gluing $\bt_i$ to the reverse of
$\bt_{i+4}$ for all $i$, so we have a presentation of $X$ of the type
used in the standard approach to the classification of surfaces,
which gives the claimed presentation of the fundamental group.
\end{proof}
\begin{remark}\label{rem-groupoid}
For some purposes it is more convenient to work with the fundamental
groupoid $\Pi_1(X)$, or the full subgroupoid $\Gm\subset\Pi_1(X)$
with objects $\{v_i\st 0\leq i<14\}$. This has an action of $G$ by
groupoid automorphisms. Each side of $F_{16}$ gives a generator, and
the interior of $F_{16}$ gives a relation. One can check that the
$G$-orbits of these generators and relations give an equivariant
presentation for $\Gm$, fom which one can recover our earlier
presentation of the group $\pi_1(X,v_0)=\Gm(v_0,v_0)$. Details are
in the file \fname+groupoid.mpl+.
\end{remark}
Given a subgroup $H\leq G$, it is usually straightforward to find a
subset of $\Net_0$ that gives a fundamental domain for the action of
$H$, and thus to understand the topology of $X/H$. In the case
$H\leq D_8$, this will be consistent with
Corollary~\ref{cor-quotient-types}. We will do this explicitly for
the cases $H=\ip{\lm}$ and $H=\ip{\lm\mu}$, where $X/H$ is an elliptic
curve. First, the inner octagon in $\Net_0$ is a fundamental domain
for $\ip{\mu}$. We can redraw this octagon in a slightly distorted
form as follows:
\begin{center}
\begin{tikzpicture}[scale=1]
\draw[magenta] (-3,-3) -- (-3, 3);
\draw[magenta] ( 3,-3) -- ( 3, 3);
\draw[magenta] (-1, 3) -- ( 1, 3);
\draw[magenta] (-1,-3) -- ( 1,-3);
\draw[blue ] (-3, 0) -- ( 3, 0);
\draw[blue ] ( 0,-3) -- ( 0, 3);
\draw[green ] (-2,-3) -- ( 2, 3);
\draw[green ] (-2, 3) -- ( 2,-3);
\draw[cyan ] (-3, 3) -- (-1, 3);
\draw[cyan ] ( 3, 3) -- ( 1, 3);
\draw[cyan ] (-3,-3) -- (-1,-3);
\draw[cyan ] ( 3,-3) -- ( 1,-3);
\fill[black] (-3,-3) circle(0.03);
\fill[black] (-2,-3) circle(0.03);
\fill[black] (-1,-3) circle(0.03);
\fill[black] ( 0,-3) circle(0.03);
\fill[black] ( 1,-3) circle(0.03);
\fill[black] ( 2,-3) circle(0.03);
\fill[black] ( 3,-3) circle(0.03);
\fill[black] (-3, 0) circle(0.03);
\fill[black] ( 0, 0) circle(0.03);
\fill[black] ( 3, 0) circle(0.03);
\fill[black] (-3, 3) circle(0.03);
\fill[black] (-2, 3) circle(0.03);
\fill[black] (-1, 3) circle(0.03);
\fill[black] ( 0, 3) circle(0.03);
\fill[black] ( 1, 3) circle(0.03);
\fill[black] ( 2, 3) circle(0.03);
\fill[black] ( 3, 3) circle(0.03);
\draw (-2.0, 1) node{$\ss\lm^2\nu$};
\draw (-0.7, 2) node{$\ss\lm$};
\draw ( 0.7, 2) node{$\ss\lm\nu$};
\draw ( 2.0, 1) node{$\ss 1$};
\draw (-2.0,-1) node{$\ss\lm^2$};
\draw (-0.7,-2) node{$\ss\lm^3\nu$};
\draw ( 0.7,-2) node{$\ss\lm^3$};
\draw ( 2.0,-1) node{$\ss\nu$};
\end{tikzpicture}
\end{center}
One can now check that $X/\ip{\mu}$ is obtained by gluing the left
edge of the square to the right edge, and the top to the bottom, which
gives a torus as expected.
For $X/\ip{\lm\mu}$, it is best to cut some corners off the inner
octagon as shown on the left below, and then rearrange the pieces as
shown on the right.
\begin{center}
\begin{tikzpicture}[scale=0.6]
\draw[blue ] (-4, 0) -- ( 4, 0);
\draw[blue ] ( 0,-4) -- ( 0, 4);
\draw[magenta] (-4,-2) -- (-4, 2);
\draw[magenta] ( 4,-2) -- ( 4, 2);
\draw[magenta] (-2,-4) -- ( 2,-4);
\draw[magenta] (-2, 4) -- ( 2, 4);
\draw[cyan ] (-4, 2) -- (-2, 4);
\draw[cyan ] (-4,-2) -- (-2,-4);
\draw[cyan ] ( 4, 2) -- ( 2, 4);
\draw[cyan ] ( 4,-2) -- ( 2,-4);
\draw[green ] (-3,-3) -- ( 3, 3);
\draw[green ] (-3, 3) -- ( 3,-3);
\draw[orange ] ( 0, 4) -- (-3, 3) -- (-4, 0);
\draw[orange ] ( 4, 0) -- ( 3,-3) -- ( 0,-4);
\fill[black ] ( 4, 0) circle(0.05);
\fill[black ] ( 4, 2) circle(0.05);
\fill[black ] ( 3, 3) circle(0.05);
\fill[black ] ( 2, 4) circle(0.05);
\fill[black ] ( 0, 4) circle(0.05);
\fill[black ] (-2, 4) circle(0.05);
\fill[black ] (-3, 3) circle(0.05);
\fill[black ] (-4, 2) circle(0.05);
\fill[black ] (-4, 0) circle(0.05);
\fill[black ] (-4,-2) circle(0.05);
\fill[black ] (-3,-3) circle(0.05);
\fill[black ] (-2,-4) circle(0.05);
\fill[black ] ( 0,-4) circle(0.05);
\fill[black ] ( 2,-4) circle(0.05);
\fill[black ] ( 3,-3) circle(0.05);
\fill[black ] ( 4,-2) circle(0.05);
\fill[black ] ( 0, 0) circle(0.05);
\draw ( 2, 1) node{$\ss 1$};
\draw ( 1, 2) node{$\ss \lm\nu$};
\draw (-1, 2) node{$\ss \lm$};
\draw (-2, 1) node{$\ss \lm^2\nu$};
\draw (-2,-1) node{$\ss \lm^2$};
\draw (-1,-2) node{$\ss \lm^3\nu$};
\draw ( 1,-2) node{$\ss \lm^3$};
\draw ( 2,-1) node{$\ss \nu$};
\end{tikzpicture}
\hspace{8em}
\begin{tikzpicture}[scale=0.8]
\draw[orange ] (-3,-3) -- ( 3,-3) -- ( 3, 3) -- (-3, 3) -- (-3,-3);
\draw[blue ] ( 0,-3) -- ( 0, 3);
\draw[blue ] (-3, 0) -- ( 3, 0);
\draw[green ] (-1,-1) -- ( 1, 1);
\draw[green ] (-3, 3) -- ( 3,-3);
\draw[cyan ] (-3,-3) -- (-1,-1);
\draw[cyan ] ( 1, 1) -- ( 3, 3);
\draw[magenta] ( 3, 0) -- ( 2, 2) -- ( 0, 3);
\draw[magenta] (-3, 0) -- (-2,-2) -- ( 0,-3);
\fill[black ] ( 3,-3) circle(0.04);
\fill[black ] ( 3, 0) circle(0.04);
\fill[black ] ( 3, 3) circle(0.04);
\fill[black ] ( 0, 3) circle(0.04);
\fill[black ] (-3, 3) circle(0.04);
\fill[black ] (-3, 0) circle(0.04);
\fill[black ] (-3,-3) circle(0.04);
\fill[black ] ( 0,-3) circle(0.04);
\fill[black ] ( 2, 2) circle(0.04);
\fill[black ] (-2,-2) circle(0.04);
\fill[black ] ( 1, 1) circle(0.04);
\fill[black ] (-1,-1) circle(0.04);
\fill[black ] ( 0, 0) circle(0.04);
\draw ( 1.6, 0.8) node{$\ss 1$};
\draw ( 0.8, 1.6) node{$\ss \lm\nu$};
\draw (-0.8, 1.6) node{$\ss \lm$};
\draw (-1.6, 0.8) node{$\ss \lm^2\nu$};
\draw (-1.6,-0.8) node{$\ss \lm^2$};
\draw (-0.8,-1.6) node{$\ss \lm^3\nu$};
\draw ( 0.8,-1.6) node{$\ss \lm^3$};
\draw ( 1.6,-0.8) node{$\ss \nu$};
\draw ( 2.6, 2.0) node{$\ss \lm^2\nu$};
\draw ( 2.0, 2.6) node{$\ss \lm^3$};
\draw (-2.6,-2.0) node{$\ss \nu$};
\draw (-2.0,-2.6) node{$\ss \lm$};
\end{tikzpicture}
\end{center}
One can again check that $X/\ip{\lm\mu}$ is obtained by gluing the left
edge of the square to the right edge, and the top to the bottom, which
gives a torus as expected.
\subsection{Homology}
\label{sec-homology}
We next consider the homology groups of a cromulent surface. For any
compact Riemann surface of genus $2$, it is standard that
$H_0(X)\simeq H_2(X)\simeq\Z$ and $H_1(X)\simeq\Z^4$, and that all
other homology groups are zero. Our main task is to give specific
generators for $H_2(X)$, and understand the action of $G$ in terms of
those generators. One approach is to recall that $H_1(X)$ is just the
abelianization of $\pi_1(X,v_0)$; we see from
Proposition~\ref{prop-pi-one} that this is the free abelian group
generated by $\{\bt_0,\bt_1,\bt_2,\bt_3\}$. However, we will use
different generators that interact with the group action in a more
convenient way.
\begin{proposition}\label{prop-homology}
Let $X$ be a cromulent surface with a curve system. Then there is an
isomorphism $\psi\:H_1(X)\to\Z^4$, with the following effect on the
homology classes of the curves $c_k$:
\begin{align*}
\psi(c_{ 0}) &= (\pp 0,\pp 0,\pp 0,\pp 0) \\
\psi(c_{ 1}) &= (\pp 1,\pp 1, -1, -1) &
\psi(c_{ 2}) &= ( -1,\pp 1,\pp 1, -1) \\
\psi(c_{ 3}) &= (\pp 0,\pp 1,\pp 0, -1) &
\psi(c_{ 4}) &= ( -1,\pp 0,\pp 1,\pp 0) \\
\psi(c_{ 5}) &= (\pp 1,\pp 0,\pp 0,\pp 0) &
\psi(c_{ 6}) &= (\pp 0,\pp 1,\pp 0,\pp 0) \\
\psi(c_{ 7}) &= (\pp 0,\pp 0,\pp 1,\pp 0) &
\psi(c_{ 8}) &= (\pp 0,\pp 0,\pp 0,\pp 1).
\end{align*}
This is equivariant with respect to the following action of $G$ on
$\Z^4$:
\begin{align*}
\lm(n) &= ( -n_2,\pp n_1, -n_4,\pp n_3) \\
\mu(n) &= (\pp n_3, -n_4,\pp n_1, -n_2) \\
\nu(n) &= (\pp n_1, -n_2,\pp n_3, -n_4).
\end{align*}
Moreover, the intersection product on $H_1(X)$ corresponds to the
following bilinear form on $\Z^4$:
\[ (n,m) = n_1m_2 - n_2m_1 - n_3m_4 + n_4m_3. \]
\end{proposition}
\begin{proof}
We use the net $\Net_0$ for $X$ discussed in
Section~\ref{sec-fundamental}. This gives a CW structure on $X$ with
a single $0$-cell, a single $2$-cell, and four $1$-cells
corresponding to $c_3$, $c_4$, $c_7$ and $c_8$. The attaching map
for the $2$-cell is a product of commutators and so is homologically
trivial. It follows that the homology classes $[c_k]$ for
$k\in\{3,4,7,8\}$ give a basis for $H_1(X)$.
We next derive some relations. Consider the following subsets of $\Net_0$:
\begin{center}
\begin{tikzpicture}[scale=.18]
\begin{scope}
\fill[gray!20] ( -4, -8) -- ( 4, -8) -- ( 8, -4) -- ( 8, 4) --
( 4, 8) -- ( -4, 8) -- ( -8, 4) -- ( -8, -4) -- cycle;
\draw ( 0,-14) node {$r_0$};
\draw[ green] ( 0, 0) -- ( 6, 6);
\draw[ green] ( 0, 0) -- ( -6, 6);
\draw[ green] ( 0, 0) -- ( -6, -6);
\draw[ green] ( 0, 0) -- ( 6, -6);
\draw[ blue] ( 0, 0) -- ( 0, 8);
\draw[ blue] ( 0, 0) -- ( 0, -8);
\draw[ blue] ( 0, 0) -- ( 8, 0);
\draw[ blue] ( 0, 0) -- ( -8, 0);
\draw[ blue] ( 12, 12) -- ( 4, 12);
\draw[ blue] ( 12, 12) -- ( 12, 4);
\draw[ green] ( 12, 12) -- ( 6, 6);
\draw[ green] (-12, 12) -- ( -6, 6);
\draw[ green] (-12,-12) -- ( -6, -6);
\draw[ green] ( 12,-12) -- ( 6, -6);
\draw[ cyan] ( 4, 8) -- ( 6, 6);
\draw[magenta] ( 4, 8) -- ( 0, 8);
\draw[magenta] ( 4, 8) -- ( 4, 12);
\draw[ cyan] ( 8, 4) -- ( 6, 6);
\draw[magenta] ( 8, 4) -- ( 8, 0);
\draw[magenta] ( 8, 4) -- ( 12, 4);
\draw[ cyan] ( -4, 8) -- ( -6, 6);
\draw[magenta] ( -4, 8) -- ( 0, 8);
\draw[magenta] ( -4, 8) -- ( -4, 12);
\draw[ cyan] ( 8, -4) -- ( 6, -6);
\draw[magenta] ( 8, -4) -- ( 8, 0);
\draw[magenta] ( 8, -4) -- ( 12, -4);
\draw[ blue] (-12, 12) -- ( -4, 12);
\draw[ blue] (-12, 12) -- (-12, 4);
\draw[ blue] (-12,-12) -- ( -4,-12);
\draw[ blue] (-12,-12) -- (-12, -4);
\draw[ blue] ( 12,-12) -- ( 4,-12);
\draw[ blue] ( 12,-12) -- ( 12, -4);
\draw[ cyan] ( 4, -8) -- ( 6, -6);
\draw[magenta] ( 4, -8) -- ( 0, -8);
\draw[magenta] ( 4, -8) -- ( 4,-12);
\draw[ cyan] ( -8, 4) -- ( -6, 6);
\draw[magenta] ( -8, 4) -- ( -8, 0);
\draw[magenta] ( -8, 4) -- (-12, 4);
\draw[ cyan] ( -4, -8) -- ( -6, -6);
\draw[magenta] ( -4, -8) -- ( 0, -8);
\draw[magenta] ( -4, -8) -- ( -4,-12);
\draw[ cyan] ( -8, -4) -- ( -6, -6);
\draw[magenta] ( -8, -4) -- ( -8, 0);
\draw[magenta] ( -8, -4) -- (-12, -4);
\end{scope}
\begin{scope}[xshift=30cm]
\fill[gray!20] (-12,-12) -- ( -4,-12) -- ( -4, -8) -- ( 4, -8) --
( 4,-12) -- ( 12,-12) -- ( 12, -4) -- ( 8, -4) --
( 8, 4) -- ( 12, 4) -- ( 12, 12) -- cycle;
\draw ( 0,-14) node {$r_1$};
\draw[ green] ( 0, 0) -- ( 6, 6);
\draw[ green] ( 0, 0) -- ( -6, 6);
\draw[ green] ( 0, 0) -- ( -6, -6);
\draw[ green] ( 0, 0) -- ( 6, -6);
\draw[ blue] ( 0, 0) -- ( 0, 8);
\draw[ blue] ( 0, 0) -- ( 0, -8);
\draw[ blue] ( 0, 0) -- ( 8, 0);
\draw[ blue] ( 0, 0) -- ( -8, 0);
\draw[ blue] ( 12, 12) -- ( 4, 12);
\draw[ blue] ( 12, 12) -- ( 12, 4);
\draw[ green] ( 12, 12) -- ( 6, 6);
\draw[ green] (-12, 12) -- ( -6, 6);
\draw[ green] (-12,-12) -- ( -6, -6);
\draw[ green] ( 12,-12) -- ( 6, -6);
\draw[ cyan] ( 4, 8) -- ( 6, 6);
\draw[magenta] ( 4, 8) -- ( 0, 8);
\draw[magenta] ( 4, 8) -- ( 4, 12);
\draw[ cyan] ( 8, 4) -- ( 6, 6);
\draw[magenta] ( 8, 4) -- ( 8, 0);
\draw[magenta] ( 8, 4) -- ( 12, 4);
\draw[ cyan] ( -4, 8) -- ( -6, 6);
\draw[magenta] ( -4, 8) -- ( 0, 8);
\draw[magenta] ( -4, 8) -- ( -4, 12);
\draw[ cyan] ( 8, -4) -- ( 6, -6);
\draw[magenta] ( 8, -4) -- ( 8, 0);
\draw[magenta] ( 8, -4) -- ( 12, -4);
\draw[ blue] (-12, 12) -- ( -4, 12);
\draw[ blue] (-12, 12) -- (-12, 4);
\draw[ blue] (-12,-12) -- ( -4,-12);
\draw[ blue] (-12,-12) -- (-12, -4);
\draw[ blue] ( 12,-12) -- ( 4,-12);
\draw[ blue] ( 12,-12) -- ( 12, -4);
\draw[ cyan] ( 4, -8) -- ( 6, -6);
\draw[magenta] ( 4, -8) -- ( 0, -8);
\draw[magenta] ( 4, -8) -- ( 4,-12);
\draw[ cyan] ( -8, 4) -- ( -6, 6);
\draw[magenta] ( -8, 4) -- ( -8, 0);
\draw[magenta] ( -8, 4) -- (-12, 4);
\draw[ cyan] ( -4, -8) -- ( -6, -6);
\draw[magenta] ( -4, -8) -- ( 0, -8);
\draw[magenta] ( -4, -8) -- ( -4,-12);
\draw[ cyan] ( -8, -4) -- ( -6, -6);
\draw[magenta] ( -8, -4) -- ( -8, 0);
\draw[magenta] ( -8, -4) -- (-12, -4);
\end{scope}
\begin{scope}[xshift=60cm]
\fill[gray!20] ( -8, 0) -- ( 8, 0) -- ( 8, -4) -- ( 12, -4) --
( 12,-12) -- ( 4,-12) -- ( 4, -8) -- ( -4, -8) --
( -4,-12) -- (-12,-12) -- (-12, -4) -- ( -8, -4) -- cycle;
\draw ( 0,-14) node {$r_2$};
\draw[ green] ( 0, 0) -- ( 6, 6);
\draw[ green] ( 0, 0) -- ( -6, 6);
\draw[ green] ( 0, 0) -- ( -6, -6);
\draw[ green] ( 0, 0) -- ( 6, -6);
\draw[ blue] ( 0, 0) -- ( 0, 8);
\draw[ blue] ( 0, 0) -- ( 0, -8);
\draw[ blue] ( 0, 0) -- ( 8, 0);
\draw[ blue] ( 0, 0) -- ( -8, 0);
\draw[ blue] ( 12, 12) -- ( 4, 12);
\draw[ blue] ( 12, 12) -- ( 12, 4);
\draw[ green] ( 12, 12) -- ( 6, 6);
\draw[ green] (-12, 12) -- ( -6, 6);
\draw[ green] (-12,-12) -- ( -6, -6);
\draw[ green] ( 12,-12) -- ( 6, -6);
\draw[ cyan] ( 4, 8) -- ( 6, 6);
\draw[magenta] ( 4, 8) -- ( 0, 8);
\draw[magenta] ( 4, 8) -- ( 4, 12);
\draw[ cyan] ( 8, 4) -- ( 6, 6);
\draw[magenta] ( 8, 4) -- ( 8, 0);
\draw[magenta] ( 8, 4) -- ( 12, 4);
\draw[ cyan] ( -4, 8) -- ( -6, 6);
\draw[magenta] ( -4, 8) -- ( 0, 8);
\draw[magenta] ( -4, 8) -- ( -4, 12);
\draw[ cyan] ( 8, -4) -- ( 6, -6);
\draw[magenta] ( 8, -4) -- ( 8, 0);
\draw[magenta] ( 8, -4) -- ( 12, -4);
\draw[ blue] (-12, 12) -- ( -4, 12);
\draw[ blue] (-12, 12) -- (-12, 4);
\draw[ blue] (-12,-12) -- ( -4,-12);
\draw[ blue] (-12,-12) -- (-12, -4);
\draw[ blue] ( 12,-12) -- ( 4,-12);
\draw[ blue] ( 12,-12) -- ( 12, -4);
\draw[ cyan] ( 4, -8) -- ( 6, -6);
\draw[magenta] ( 4, -8) -- ( 0, -8);
\draw[magenta] ( 4, -8) -- ( 4,-12);
\draw[ cyan] ( -8, 4) -- ( -6, 6);
\draw[magenta] ( -8, 4) -- ( -8, 0);
\draw[magenta] ( -8, 4) -- (-12, 4);
\draw[ cyan] ( -4, -8) -- ( -6, -6);
\draw[magenta] ( -4, -8) -- ( 0, -8);
\draw[magenta] ( -4, -8) -- ( -4,-12);
\draw[ cyan] ( -8, -4) -- ( -6, -6);
\draw[magenta] ( -8, -4) -- ( -8, 0);
\draw[magenta] ( -8, -4) -- (-12, -4);
\end{scope}
\end{tikzpicture}
\end{center}
The boundary of $r_0$ consists of four fragments of $c_0$ (which
together make up the whole of $c_0$) together with a fragment of
$c_3$ repeated twice with opposite orientations, and a fragment of
$c_4$ repeated twice with opposite orientations. From this we
conclude that $[c_0]=0$ in $H_1(X)$. Similarly, the boundary of
$r_1$ consists of $c_1$, $c_3$ and $c_4$ together with mutually
cancelling fragments of $c_7$ and $c_8$. Here $c_1$ and $c_4$ run
clockwise but $c_3$ runs anticlockwise. We therefore have
$[c_1]-[c_3]+[c_4]=0$. Applying the same method to $r_2$ gives
$[c_4]+[c_5]-[c_7]=0$. Next, part of the definition of a curve
system is that $\lm(c_1(t))=c_2(t)$ and $\lm(c_3(t))=c_4(t)$ and
$\lm(c_4(t))=c_3(-t)$, which gives $\lm_*[c_1]=[c_2]$ and
$\lm_*[c_3]=[c_4]$ and $\lm_*[c_4]=-[c_3]$. We can therefore apply
$\lm_*$ to the relation $[c_1]-[c_3]+[c_4]=0$ to get
$[c_2]-[c_3]-[c_4]=0$. Similarly, we can apply $\lm_*$ to the
relation $[c_4]+[c_5]-[c_7]=0$ to get $-[c_3]+[c_6]-[c_8]=0$. This
is enough to show that $[c_5]$, $[c_6]$, $[c_7]$ and $[c_8]$ form an
alternative basis for $H_1(X)$. By writing everything in terms of
this basis we get an isomorphism $\psi\:H_1(X)\to\Z^4$, and by
inspecting the above relations we see that this is given by the
claimed formulae.
Next, it is part of the definition of cromulence that
\begin{align*}
\lm(c_{ 5}(t)) &= c_{ 6}( t) &
\mu(c_{ 5}(t)) &= c_{ 7}( t) &
\nu(c_{ 5}(t)) &= c_{ 5}( t) \\
\lm(c_{ 6}(t)) &= c_{ 5}(-t) &
\mu(c_{ 6}(t)) &= c_{ 8}(-t) &
\nu(c_{ 6}(t)) &= c_{ 6}(-t) \\
\lm(c_{ 7}(t)) &= c_{ 8}( t) &
\mu(c_{ 7}(t)) &= c_{ 5}( t) &
\nu(c_{ 7}(t)) &= c_{ 7}( t) \\
\lm(c_{ 8}(t)) &= c_{ 7}(-t) &
\mu(c_{ 8}(t)) &= c_{ 6}(-t) &
\nu(c_{ 8}(t)) &= c_{ 8}(-t).
\end{align*}
The action in homology can be read off from this in an obvious way,
and we find that it works as described in the statement of the
Proposition.
Finally, we need to analyse the intersection pairing. From the
definition of a curve system and associated discussion, we see that
$C_5\cap C_6=\{v_0\}$ and $C_7\cap C_8=\{v_1\}$ and
\[ C_5\cap C_7 = C_5\cap C_8 = C_6\cap C_7 = C_6\cap C_8 = \emptyset.
\]
This means that the corresponding products in homology are
$[c_5]\cdot[c_6]=\pm 1$ and $[c_7]\cdot[c_8]=\pm 1$ and
\[ [c_5]\cdot [c_7] = [c_5]\cdot [c_8] =
[c_6]\cdot [c_7] = [c_6]\cdot [c_8] = 0.
\]
In the net, $v_0$ is the origin, $c_5$ runs to the right along the
$x$-axis and $c_6$ runs upwards along the $y$-axis, so
$[c_5]\cdot [c_6]=+1$. Moreover, the map $\mu$ preserves orientation
and has $\mu_*[c_5]=[c_7]$ and $\mu_*[c_6]=-[c_8]$; this means that
$[c_7]\cdot[c_8]=-1$. The claimed description of the intersection
product follows easily.
\begin{checks}
homology_check.mpl: check_homology()
\end{checks}
\end{proof}
\begin{remark}
The vector $\psi(c_4)=(-1,0,1,0)$ (for example) is represented in
Maple as \mcode+c_homology[4]+. The action of $G$ on $\Z^4$ is
represented by \mcode+act_Z4+, which is a table indexed by the
elements of $G$, whose entries are functions. For example, the
expression \mcode+act_Z4[LMN]([1,2,3,4])+ evaluates to
\mcode+[-4,3,-2,1]+, corresponding to the fact that
$\lm\mu\nu(1,2,3,4)=(-4,3,-2,1)$. Most other actions of $G$ on other
sets are also represented in this way.
\end{remark}
We now relate the above description to the fundamental group.
\begin{lemma}
The homology classes of the generators $\bt_i\in\pi_1(X,v_0)$ are as
follows:
\begin{align*}
[\bt_0] &= (\pp 1,\pp 0,\pp 0,\pp 0) &
[\bt_1] &= ( -1, -1,\pp 1,\pp 0) \\
[\bt_2] &= (\pp 0,\pp 1,\pp 0,\pp 0) &
[\bt_3] &= (\pp 1, -1,\pp 0,\pp 1) \\
[\bt_4] &= ( -1,\pp 0,\pp 0,\pp 0) &
[\bt_5] &= (\pp 1,\pp 1, -1,\pp 0) \\
[\bt_6] &= (\pp 0, -1,\pp 0,\pp 0) &
[\bt_7] &= ( -1,\pp 1,\pp 0, -1)
\end{align*}
\end{lemma}
\begin{proof}
The claim for $\bt_0$ is immediate from the definitions, and the
claims for $\bt_2$, $\bt_4$ and $\bt_6$ follow using the group
action. Now consider the left hand half of $\Net_5$:
\begin{center}
\begin{tikzpicture}[scale=0.6]
\fill[gray!20]
( 0, -3) -- ( 0, 3) -- ( -1, 3) -- ( -2, 2) --
( -2, 3) -- ( -5, 3) -- ( -6, 2) -- ( -6, 1) --
( -3, 1) -- ( -3, -1) -- ( -2, -2) -- ( -3, -2) --
( -3, -5) -- ( -2, -6) -- ( -1, -6) -- ( -1, -3) -- cycle;
\draw[ blue,-<] ( 3, 2) -- ( 3, 4);
\draw[ blue,-<] ( -2, 3) -- ( -4, 3);
\draw[ blue,-<] ( -3, -2) -- ( -3, -4);
\draw[ blue,-<] ( 2, -3) -- ( 4, -3);
\draw[orange,-<] ( 3, -1) -- ( 3, 0);
\draw[orange,-<] ( 1, 3) -- ( 0, 3);
\draw[orange,-<] ( -3, 1) -- ( -3, 0);
\draw[orange,-<] ( -1, -3) -- ( 0, -3);
%
\draw[blue ] ( 0, -3) -- ( 0, 3);
\draw[blue,->] ( 0, -3) -- ( 0, 0);
\draw (0,0) node[anchor=west] {$\ss c_7$};
%
\draw[ orange] ( 3, 2) -- ( 2, 2);
\draw[ blue] ( 3, 2) -- ( 3, 5);
\draw[ orange] ( 3, 1) -- ( 2, 2);
\draw[ orange] ( 3, 1) -- ( 3, 0);
\draw[ orange] ( 1, -3) -- ( 2, -2);
\draw[ orange] ( 1, -3) -- ( 0, -3);
\draw[ orange] ( -3, -1) -- ( -2, -2);
\draw[ orange] ( -3, -1) -- ( -3, 0);
\draw[ orange] ( -1, 3) -- ( -2, 2);
\draw[ orange] ( -1, 3) -- ( 0, 3);
\draw[ blue] ( 2, 6) -- ( 3, 5);
\draw[ orange] ( 2, 6) -- ( 1, 6);
\draw[ orange] ( -2, 3) -- ( -2, 2);
\draw[ blue] ( -2, 3) -- ( -5, 3);
\draw[ orange] ( -6, 2) -- ( -6, 1);
\draw[ blue] ( -6, 2) -- ( -5, 3);
\draw[ orange] ( -3, -2) -- ( -2, -2);
\draw[ blue] ( -3, -2) -- ( -3, -5);
\draw[ orange] ( -2, -6) -- ( -1, -6);
\draw[ blue] ( -2, -6) -- ( -3, -5);
\draw[ orange] ( 2, -3) -- ( 2, -2);
\draw[ blue] ( 2, -3) -- ( 5, -3);
\draw[ blue] ( 6, -2) -- ( 5, -3);
\draw[ orange] ( 6, -2) -- ( 6, -1);
\draw[ orange] ( -3, 1) -- ( -6, 1);
\draw[ orange] ( -3, 1) -- ( -3, 0);
\draw[ orange] ( 1, 3) -- ( 0, 3);
\draw[ orange] ( 1, 3) -- ( 1, 6);
\draw[ orange] ( 3, -1) -- ( 3, 0);
\draw[ orange] ( 3, -1) -- ( 6, -1);
\draw[ orange] ( -1, -3) -- ( -1, -6);
\draw[ orange] ( -1, -3) -- ( 0, -3);
\draw ( -3, -5) node[anchor=east ]{$\bt_0$};
\draw ( -3, -1) node[anchor=east ]{$\bt_1$};
\draw ( -5, 3) node[anchor=south]{$\bt_2$};
\draw ( -1, 3) node[anchor=south]{$\bt_3$};
\draw ( 1, -3) node[anchor=north]{$\bt_7$};
%
\fill ( 3, 2) circle(0.05);
\fill ( -6, 2) circle(0.05);
\fill ( 2, -3) circle(0.05);
\fill ( -3, -2) circle(0.05);
\fill ( 6, -2) circle(0.05);
\fill ( 2, 6) circle(0.05);
\fill ( -2, 3) circle(0.05);
\fill ( -2, -6) circle(0.05);
\end{tikzpicture}
\end{center}
Define $u(t)=\bt_3(t/2)=\bt_7(1-t/2)$ for $0\leq t\leq 1$. The
boundary of the above region gives a relation
\[ [\bt_0] + [\bt_1] + [\bt_2] + [u] - [c_7] - [u] = 0, \]
so
\[ [\bt_1] = [c_7] - [\bt_0] - [\bt_2] = (-1,-1,1,0) \]
as claimed. The remaining claims for $\bt_3$, $\bt_5$ and $\bt_7$
now follow using the group action.
\end{proof}
For some purposes it is convenient to use a different basis. We put
\begin{align*}
u_{11} &= ( 1,0,\pp 1,0) & u_{12} &= (0, 1,0,\pp 1) \\
u_{21} &= ( 1,0,-1,0) & u_{22} &= (0, 1,0,-1).
\end{align*}
These elements do not generate all of $\Z^4$, but only a subgroup of
index $4$. However, they give a basis for $\Q^4$. One can check that
\begin{align*}
\lm(u_{11}) &= \pp u_{12} & \mu(u_{11}) &= \pp u_{11} & \nu(u_{11}) &= \pp u_{11} \\
\lm(u_{12}) &= -u_{11} & \mu(u_{12}) &= -u_{12} & \nu(u_{12}) &= -u_{12} \\
\lm(u_{21}) &= \pp u_{22} & \mu(u_{21}) &= -u_{21} & \nu(u_{21}) &= \pp u_{21} \\
\lm(u_{22}) &= -u_{21} & \mu(u_{22}) &= \pp u_{22} & \nu(u_{22}) &= -u_{22}.
\end{align*}
This shows that $\{u_{11},u_{12}\}$ is a basis for a subrepresentation
$U_1$ with character $\chi_8$, whereas $\{u_{21},u_{22}\}$ is a basis
for a subrepresentation $U_2$ with character $\chi_9$.
We now consider the homology of certain quotient spaces $X/H$. For
most subgroups $H\leq G$, the quotient $X/H$ is either a disc or a
sphere, and so the first homology group is trivial. The interesting
cases are the elliptic curves $X/\ip{\mu}$ and $X/\ip{\lm\mu}$. We
write $q_+\:X\to X/\ip{\mu}$ and $q_-\:X\to X/\ip{\lm\mu}$ for the
quotient maps.
\begin{proposition}\label{prop-elliptic-homology}
There is a commutative diagram
\[ \xymatrix{
\Z^2 \ar[d]^\simeq_{\psi_+} &
\Z^4 \ar@{->>}[l]_{\tht_+} \ar@{->>}[r]^{\tht_-} \ar[d]_\psi^\simeq &
\Z^2 \ar[d]_\simeq^{\psi_-} \\
H_1(X/\ip{\mu}) &
H_1(X) \ar@{->>}[l]^{(q_+)_*} \ar@{->>}[r]_{(q_-)_*} &
H_1(X/\ip{\lm\mu})
} \]
where
\begin{align*}
\tht_+(n) &= (n_1+n_3,\;n_2-n_4) \\
\tht_-(n) &= (n_2+n_3,\;n_1+n_4).
\end{align*}
\end{proposition}
\begin{proof}
First, we can regard $H_1(X)$ as the abelianization of
$\pi_1(X,v_2)$, and $H_1(X/\ip{\mu})$ as the abelianization of
$\pi_1(X/\ip{\mu},q_+(v_2))$. Because $q_+$ is a branched covering,
any loop $u$ based at $q_+(v_2)$ can be lifted to give a path $\tu$
in $X$ starting at $v_2$. The endpoint $\tu(1)$ will lie over
$q_+(v_2)$, but $\mu(v_2)=v_2$ so this ensures that $\tu(1)=v_2$.
This means that $\tu$ is again a loop, and we deduce that the map
\[ (q_+)_*\:\pi_1(X,v_2)\to \pi_1(X/\ip{\mu},q_+(v_2)) \]
is surjective. It follows that the map
\[ (q_+)_* \: H_1(X) \to H_1(X/\ip{\mu}) \]
is also surjective. A similar argument (using the basepoint
$v_6=\lm\mu(v_6)$) shows that $(q_-)_*$ is also surjective on $H_1$.
Now recall that $\psi$ is equivariant for the following action:
\begin{align*}
\mu(n) &= (n_3,-n_4,n_1,-n_2) \\
\lm\mu(n) &= (n_4,n_3,n_2,n_1).
\end{align*}
It follows easily that $\tht_+$ is a coequaliser for the action of
$\mu$, and $\tht_-$ is a coequaliser for the action of $\lm\mu$.
This implies that there are unique maps $\psi_+$ and $\psi_-$ making
the diagram commute. As $\psi$ is an isomorphism and $(q_\pm)_*$ is
surjective, we see that $\psi_\pm$ is also surjective. However,
$X/\ip{\mu}$ and $X/\ip{\lm\mu}$ are both homeomorphic to the torus,
so in each case $H_1\simeq\Z^2$, and this implies that any surjective
homomorphism $\Z^2\to H_1$ is automatically an isomorphism.
\end{proof}
We next discuss the Jacobian variety $JX$. This is usually
constructed as an abelian variety using methods of algebraic geometry,
as we will recall in Section~\ref{sec-ellquot}. However, in our
cromulent setting it is possible to construct $JX$ as a space using
only topological methods.
We first give some definitions related to coverings and subgroups of
the fundamental group. These are standard, but we just want to pin
down some issues of naturality.
\begin{definition}
Let $Y$ be a (locally tame) path-connected space with a basepoint
$y_0$. We define $\tY$ to be the space of paths in $Y$ starting at
$y_0$, modulo the equivalence relation of homotopy relative to
endpoints. This has a natural basepoint $\ty_0$, which is the
equivalence class of the constant path at $y_0$. All this is clearly
functorial for based maps. Moreover, path join gives a free left
action of $\pi_1(Y,y_0)$ on $\tY$. Evaluation at $1$ gives a
projection $\pi\:\tY\to Y$, which induces a homeomorphism
$\tY/\pi_1(Y,y_0)\to Y$. Given a subgroup $H\leq\pi_1(Y,y_0)$, we
have a based covering map $\tY/H\to Y$, whose effect in $\pi_1$ is
the inclusion $H\to\pi_1(Y,y_0)$. Given a based homeomorphism
$f\:Y_0\to Y_1$ with $f_*(H_0)=H_1$, we get an induced based
homeomorphism $\tY_0/H_0\to\tY_1/H_1$.
\end{definition}
\begin{definition}\label{defn-JX}
We put $J'X=(X/\ip{\mu})\tm(X/\ip{\lm\mu})$, and define
$q'\:X\to J'X$ by $q'(x)=(q_+(x),q_-(x))$. The resulting map
\[ \tht' = (\Z^4 \xra{\psi} H_1(X)
\xra{q'_+} H_1(X/\ip{\mu})\oplus H_1(X/\ip{\lm\mu})
\xra{(\psi_+,\psi_-)^{-1}} \Z^4)
\]
is then given by
\[ \tht'(n) = (\tht_+(n),\tht_-(n))
= (n_1+n_3,n_2-n_4,n_1+n_4,n_2+n_3).
\]
One can check that the image of $\tht'$ is the subgroup
\[ \Tht' = \{m\in\Z^4\st\sum_im_i = 0\pmod{2}\}, \]
which has index two. Note also that $J'X\simeq(S^1)^4$, so the natural
map $\pi_1(J'X,q'(v_0))\to H_1(J'X)$ is an isomorphism. Thus, $\Tht'$
corresponds to a subgroup of index two in $\pi_1(J'X,q'(v_0))$, and
so gives rise to a double cover of $J'X$, which we call $JX$.
Because $\pi_1(JX)=\Tht'$, we see that $q'$ lifts to give a map
$q\:X\to JX$. By construction, this induces an isomorphism
$H_1(X)\to H_1(JX)$.
\end{definition}
\begin{remark}
If $Y$ and $Z$ are homotopy equivalent to $(S^1)^n$ and $(S^1)^m$,
then standard methods of homotopy theory show that the natural map
\[ H_1 \: [Y,Z] \to \Hom(H_1(Y),H_1(Z)) \]
is bijective. (It makes no difference here whether we work in a
based context or an unbased context.) Using this, we can produce a
map $JX\tm JX\to JX$ which makes $JX$ into an abelian group up to
homotopy, and we can produce a map $G\to[JX,JX]$ giving an action of
$G$ on $JX$ up to homotopy. Using analytic methods, we can improve
this: $JX$ becomes an abelian variety, with an action of $G$ by
automorphisms of abelian varieties. However, we cannot
build a topological group structure or a $G$-action by homeomorphisms
without using analysis. The element $\lm^2\in G$ normalises the
subgroups $\ip{\mu}$ and $\ip{\lm\mu}$, and preserves the basepoint
$v_0$, so this induces an involution on $J'X$ and on $JX$. However,
no other nontrivial element of $G$ shares these properties.
\end{remark}
\section{The projective family}
\label{sec-P}
In this section we construct a family of precromulent surfaces $PX(a)$
(for $a\in (0,1)$) as branched covers of the Riemann sphere
$\C_\infty$. Later we will consider an arbitrary precromulent surface
$X$, and attempt to find an isomorphism $X\simeq PX(a)$ for some $a$.
The notion of a cromulent labelling will emerge naturally from this
analysis.
\subsection{Definitions}
\label{sec-P-defs}
\begin{definition}\label{defn-P}
For any $a\in(0,1)$ we put $A=a^2+1/a^2\in(2,\infty)$ and define
$r_a\:\C\to\C$ by
\[ r_a(z) = z(z-a)(z+a)(z-1/a)(z+1/a)
= z^5 - A z^3 + z.
\]
Next, we put
\begin{align*}
R_0(a) &= \C[z,w]/(w^2-r_a(z)) \\
PX_0(a) &= \spec(R_0(a)) = \{(w,z)\in \C^2 \st w^2=r_a(z) \},
\end{align*}
so $PX_0(a)$ is a smooth affine hyperelliptic curve. Unfortunately,
if we just take the closure in $\C P^2$, the resulting curve is
singular at infinity. To get a nonsingular completion, we put
\[ PX(a) = \{[z]\in\C P^4\st
z_1^2-z_2z_3-z_4z_5+Az_3z_4=0,\;
z_2z_4=z_3^2,\;z_2z_5=z_3z_4,\;z_3z_5=z_4^2
\}.
\]
There is a map $j\:PX_0(a)\to PX(a)$ given by
\[ j(w,z) = [w:1:z:z^2:z^3]. \]
We define points $v_0,\dotsc,v_{13}\in PX(a)$ by
\begin{align*}
v_0 &= [0:1:0:0:0] = j(0,0) \\
v_1 &= [0:0:0:0:1] \\
v_2 &= j( -(a^{-1}-a),-1) &
v_6 &= j\left(\pp \frac{1+i}{\rt}(a^{-1}+a),\pp i\right) &
v_{10} &= j(0,-a) \\
v_3 &= j(-i(a^{-1}-a),\pp 1) &
v_7 &= j\left(-\frac{1-i}{\rt}(a^{-1}+a),-i\right) &
v_{11} &= j(0,\pp a) \\
v_4 &= j(\pp (a^{-1}-a),-1) &
v_8 &= j\left(-\frac{1+i}{\rt}(a^{-1}+a),\pp i\right) &
v_{12} &= j(0,-a^{-1}) \\
v_5 &= j(\pp i(a^{-1}-a),\pp 1) &
v_9 &= j\left(\pp \frac{1-i}{\rt}(a^{-1}+a),-i\right) &
v_{13} &= j(0,\pp a^{-1}).
\end{align*}
We let $G$ act on $\C P^4$ and $PX(a)$ by
\begin{align*}
\lm[z] &= [iz_1:z_2:-z_3:z_4:-z_5] \\
\mu[z] &= [z_1:z_5:z_4:z_3:z_2] \\
\nu[z] &= [\ov{z}_1:\ov{z}_2:\ov{z}_3:\ov{z}_4:\ov{z}_5].
\end{align*}
\end{definition}
\begin{checks}
cromulent.mpl: check_precromulent("P")
projective/PX_check.mpl: check_P_action()
projective/PX_check.mpl: check_j_P()
\end{checks}
We will prove as Proposition~\ref{prop-P-precromulent} that this
gives a precromulent surface; then we will see in
Proposition~\ref{prop-P-fundamental} that it is actually cromulent.
\begin{remark}
The parameters $a$ and $A$ are \mcode+a_P+ and \mcode+A_P+ in Maple.
The polynomial $r_a(z)$ is \mcode+r_P(z)+. Points in $\C^2$ are
represented as lists of length two. (Maple distinguishes between
lists and vectors, and we have generally preferred to use lists, for
technical reasons that we will not explore here.) One can check
whether a list lies in $PX_0(a)$ using the function
\mcode+is_member_P_0(z)+. Points in $\C P^4$ are represented by lists
of length $5$, and one can check whether two lists are projectively
equivalent using the function \mcode+is_equal_P(w,z)+. One can also
check whether a point lies in $PX(a)$ using the function
\mcode+is_member_P(z)+. The map $j\:PX_0(a)\to PX(a)$ is
\mcode+j_P(z)+. The points $v_i\in PX(a)$ are \mcode+v_P[i]+. The action
of $g\in G$ on $PX(a)$ is given by \mcode+act_P[g](z)+, and the
corresponding action on $PX_0(a)$ is \mcode+act_P_0[g](z)+. All this
comes from the file \fname+projective/PX.mpl+.
\end{remark}
\begin{remark}\label{rem-a-Po}
All the above functions treat \mcode+a_P+ as a symbol. There is also
a global variable \mcode+a_P0+ which holds a numerical value for
\mcode+a_P+. In fact, there are two such variables, called
\mcode+a_P0+ and \mcode+a_P1+. This is intended to cover the case
where \mcode+a_P0+ is an exact expression (such as a rational number)
and \mcode+a_P1+ is a floating point approximation to the same
number. However, we have usually taken \mcode+a_P0+ to be a floating
point number, so that there is no distinction between \mcode+a_P0+ and
\mcode+a_P1+. These variables should be set using the function
\mcode+set_a_P0+, defined in the file \fname+projective/PX0.mpl+.
This will then set a large number of other variables by substituting
\mcode+a_P0+ or \mcode+a_P1+ for \mcode+a_P+. For example,
\mcode+v_P0[i]+ and \mcode+v_P1[i]+ are obtained by applying these
substitutions to \mcode+v_P[i]+. When the file
\fname+projective/PX0.mpl+ is loaded, it calls the function
\mcode+set_a_P0+ to set \mcode+a_P0+ to a particular value
(approximately $0.0984$), which is close to the value for which
$EX^*$ is isomorphic to $PX(a)$. However, one can call
\mcode+set_a_P0+ again to change the value if desired.
\end{remark}
\begin{remark}\label{rem-simplify-P}
When working with an expression $m$ involving \mcode+a_P+, it is
sometimes convenient to use the function \mcode+simplify_P(m)+
(defined in \fname+projective/PX.mpl+). This will try some
substitutions like $\sqrt{1-a^2}=\sqrt{1-a}\sqrt{1+a}$ that are not
always used by the default simplification functions.
\end{remark}
\begin{remark}\label{rem-r-coprime}
As $a\in(0,1)$ we see that $r_a(z)$ has no repeated roots, so
$r_a(z)$ and $r'_a(z)$ are coprime in $\R[z]$ or $\C[z]$. More
specifically, one can check by direct expansion that
$m_0(z)r_a(z)+m_1(z)r'_a(z)=1$, where
\begin{align*}
m_0(z) &= \frac{(100-30A^2)z^3+(18A^3-70A)z}{4(A^2-4)} \\
m_1(z) &= \frac{(6A^2-20)z^4-(6A^3-22A)z^2+(4A^2-16)}{4(A^2-4)}.
\end{align*}
(These are \mcode+r_P_cofactor0(z)+ and \mcode+r_P_cofactor1(z)+ in the
Maple code.)
\begin{checks}
projective/PX_check.mpl: check_r_P_cofactors()
\end{checks}
\end{remark}
\begin{lemma}\label{lem-R-domain}
The ring $R_0(a)$ is an integral domain, and the set
\[ B = \{z^iw^j\st i\in\N,\;j\in\{0,1\}\} \]
is a basis for $R_0(a)$ over $\C$.
\end{lemma}
\begin{proof}
From the description $R_0(a)=\C[z,w]/(w^2-r_a(z))$ it is clear that
$\{1,w\}$ is a basis for $R_0(a)$ as a module over $\C[z]$. It
follows that $B$ is a basis for $R_0(a)$ over $\C$. Next, for any
element $f=p(z)+q(z)w\in R_0(a)$, we put
\[ N(f) = (p(z)+q(z)w)(p(z)-q(z)w)=p(z)^2-q(z)^2r_a(z) \in\C[z]. \]
It is easy to check that $N(fg)=N(f)N(g)$. Moreover, by considering
the highest power of $z$ that divides the various terms, we see that
$N(f)\neq 0$ whenever $f\neq 0$. Thus, if $g$ is also nonzero we
have $N(fg)=N(f)N(g)$, which is nonzero because $\C[z]$ is a domain,
so $fg\neq 0$ as required.
\end{proof}
\begin{lemma}\label{lem-P-differentials}
The module $\Om^1(PX_0(a))$ of K\"ahler differentials is freely
generated over $R_0(a)$ by the element
\[ \om_0 = m_0(z)w\,dz + 2m_1(z)\,dw \]
(where $m_0$ and $m_1$ are as in Remark~\ref{rem-r-coprime}). In
particular, we have
\begin{align*}
dz &= w\,\om_0 \\
dw &= \half r'_a(z)\om_0.
\end{align*}
\end{lemma}
\begin{proof}
We will put $\Om^1=\Om^1(PX_0(a))$ for brevity. Differentiating the
equation $w^2=r_a(z)$ gives $2w\,dw=r'_a(z)\,dz$, and (essentially by
definition) the module $\Om^1$ is generated by $dw$ and $dz$ subject
only to this relation. Now
\begin{align*}
w\,\om_0 &= m_0(z)w^2\,dz + 2m_1(z)w\,dw \\
&= m_0(z)r_a(z)\,dz + m_1(z)r'_a(z)\,dz \\
&= (m_0(z)r_a(z)+m_1(z)r'_a(z))\,dz = dz.
\end{align*}
A similar argument shows that $\half r'_a(z)\om_0=dw$, so both $dz$
and $dw$ lie in the submodule of $\Om^1$ generated by $\om_0$. This
implies that $\om_0$ generates all of $\Om^1$. Now note that our
original presentation of $\Om^1$ implies that $\Om^1[w^{-1}]$ is
freely generated over $R_0(a)[w^{-1}]$ by $dw$. Thus, if
$f\in R_0(a)$ satisfies $f\om_0=0$ then $f\Om^1=0$ so
$f\Om^1[w^{-1}]=0$ so $f$ must map to zero in $R_0(a)[w^{-1}]$.
However, as $R_0(a)$ is an integral domain we see that the map
$R_0(a)\to R_0(a)[w^{-1}]$ is injective, so $f$ must be zero. It
follows that $\Om^1$ is \emph{freely} generated by $\om_0$, as
claimed.
\end{proof}
\begin{lemma}\label{lem-nonzero}
Consider a point $[z]\in PX(a)$.
\begin{itemize}
\item[(a)] If any of the coordinates $z_2,\dotsc,z_5$ is zero, then
$[z]\in\{v_0,v_1\}$.
\item[(b)] Either $z_2\neq 0$ or $z_5\neq 0$.
\end{itemize}
\end{lemma}
\begin{proof}
We put
\begin{align*}
r_0(z) &= z_1^2-z_2z_3-z_4z_5+Az_3z_4 \\
r_1(z) &= z_2 z_4 - z_3^2 \\
r_2(z) &= z_2 z_5 - z_3 z_4 \\
r_3(z) &= z_3 z_5 - z_4^2,
\end{align*}
so that $PX(a)=\{[z]\st r_0(z)=\dotsb=r_3(z)=0\}$.
\begin{itemize}
\item[(a)] Suppose that $[z]\in PX(a)$ and $z_2z_3z_4z_5=0$. Using
$r_2$ we see that $(z_2z_5)^2$ and $(z_3z_4)^2$ are both equal to
$z_2z_3z_4z_5$ and thus to zero, so $z_2z_5=z_3z_4=0$. Thus, either
$z_2=0$ or $z_5=0$; and either $z_3=0$ or $z_4=0$. If $z_3=0$ then
$r_3$ gives $z_4=0$; conversely, if $z_4=0$ then $r_1$ gives
$z_3=0$. We must therefore have $z_3=z_4=0$. Substituting this in
$r_0$ gives $z_1=0$. In summary, at most one of the coordinates
$z_i$ can be nonzero, and the nonzero coordinate must be $z_2$ or
$z_5$. It follows that $[z]\in\{v_0,v_1\}$ as required.
\item[(b)] The claim is clear if $z_2z_3z_4z_5\neq 0$, and follows
from~(a) if $z_2z_3z_4z_5=0$.
\end{itemize}
\end{proof}
\begin{proposition}\label{prop-P-j}
The map $j$ gives an isomorphism $PX_0(a)\simeq PX(a)\sm\{v_1\}$.
\end{proposition}
\begin{proof}
The lemma shows that for $[z]\in PX(a)\sm\{v_1\}$ we have
$z_2\neq 0$. We can thus define $k\:PX(a)\sm\{v_1\}\to\C^2$ by
$k[z]=(z_1/z_2,z_3/z_2)$. It will be harmless to rescale $z$ so
that $z_2=1$, and then the last three defining relations for $PX(a)$
become $z_4=z_3^2$ and $z_5=z_4z_4$ and $z_3z_5=z_4^2$, so
$(z_2,z_3,z_4,z_5)=(1,z_3,z_3^2,z_3^3)$. If we use these to rewrite
the first relation we get
\[ z_1^2-z_3-z_3^5+(a^2+a^{-2})z_3^3 = 0, \]
so $k[z]\in PX_0(a)$. The equations $jk=1$ and $kj=1$ are now
clear.
\begin{checks}
projective/PX_check.mpl: check_j_P()
\end{checks}
\end{proof}
\begin{definition}\label{defn-P-cover}
We put
\begin{align*}
U_0 &= PX(a)\sm\{v_1\} \\
U_1 &= PX(a)\sm\{v_0\}=\mu(U_0)\simeq U_0 \\
U_{01} &= U_0\cap U_1.
\end{align*}
\end{definition}
\begin{remark}\label{rem-P-cover}
We use $j$ to silently identify $U_0$ with $PX_0(a)$, which is the
spectrum of the ring $R_0(a)=\C[z,w]/(w^2-r_a(z))$. This in turn
identifies $U_{01}$ with the spectrum of the ring
\[ R'(a) = R_0(a)[z^{-1}] = \C[z^{\pm 1},w]/(w^2-r_a(z)). \]
The set
\[ B_0=\{z^i\st i\geq 0\} \amalg\{z^iw\st i\geq 0\} \]
is a basis for the subring $R_0(a)\subset R'(a)$ over $\C$. The
set $R_1(a)=\mu(R_0(a))$ is also a subring of $R'(a)$, with basis
\[ B_1=\{z^i\st i\leq 0\} \amalg\{z^iw\st i\leq -3\}. \]
\end{remark}
\begin{proposition}\label{prop-P-precromulent}
The above definitions make $PX(a)$ into a precromulent surface.
\end{proposition}
\begin{proof}
First, the standard Jacobian condition shows that $PX_0(a)$ is
smooth, so Proposition~\ref{prop-P-j} shows that $PX(a)$ is smooth
except possibly at $v_1$. Next, straightforward calculation shows
that the action of $G$ on $\C P^4$ preserves $PX(a)$, with $D_8$
acting conformally and $G\sm D_8$ acting anticonformally, and
\begin{align*}
\lm(j(w,z)) &= j(iw,-z) \\
\mu(j(w,z)) &= j(-w/z^3,1/z) \\
\nu(j(w,z)) &= j(\ov{w},\ov{z}).
\end{align*}
In particular, we see that $\mu$ gives an isomorphism between
$PX(a)\sm\{v_1\}$ and $PX(a)\sm\{v_0\}$, showing that $PX(a)$ is
smooth everywhere. It is clearly closed in $\C P^4$ and therefore
compact. It is standard that for any polynomial $r(z)$ of degree
$2g+1$, the hyperelliptic curve $w^2=r(z)$ has genus $g$; in
particular, $PX(a)$ has genus $2$. A straightforward but lengthy
check shows that $G$ permutes the points $v_i$ in accordance with
Definition~\ref{defn-precromulent}(b). The tangent space to
$PX_0(a)$ at $(0,0)$ is $\C\oplus 0$, and $\lm$ acts on this is
multiplication by $i$. All that is left is to check that $D_8$ acts
freely on $PX(a)\sm V$, where $V=\{v_0,\dotsc,v_{13}\}$. The fixed
points of $\lm^2$ on $PX_0(a)$ are pairs $(w,z)$ with
$(-w,z)=(w,z)$ and $w^2=z(z^2-a^2)(z^2-a^{-2})$, which means that
$w=0$ and $z\in\{0,a,-a,a^{-1},-a^{-1}\}$. It follows that
\[ PX(a)^{\langle\lm^2\rangle} =
\{v_0,v_1,v_{10},v_{11},v_{12},v_{13}\} \sse V.
\]
All fixed points of $\lm$ or $\lm^3$ are also fixed by $\lm^2$ and so
lie in $V$. A similar analysis shows that the fixed points of $\mu$,
$\lm\mu$, $\lm^2\mu$ and $\lm^3\mu$ also lie in $V$, as required.
\end{proof}
\begin{remark}\label{rem-P-quotient}
We can define $p\:PX(a)\to\C_\infty$ by $p([z])=z_3/z_2$, so
$pj(w,z)=z$. Using $\lm^2j(w,z)=(-w,z)$, it is not hard to check
that $p$ gives an isomorphism $PX(a)/\ip{\lm^2}\to\C_\infty$.
Moreover, $p$ is equivariant if we use the following action on
$\C_\infty$:
\[ \lm(z) = -z \hspace{4em}
\mu(z) = 1/z \hspace{4em}
\nu(z) = \ov{z}.
\]
This action is represented by \mcode+act_C+ in Maple. For example,
\mcode+act_C[L](3)+ evaluates to \mcode+-3+.
\end{remark}
\subsection{The curve system}
\label{sec-P-curves}
\begin{definition}\label{defn-P-curves}
We define maps $j'\:\C^3\sm\{0\}\to\C P^4$ by
\[ j'(w,x,y) = [w:x^3:x^2y:xy^2:y^3]. \]
Note that $j(w,z)=j'(w,1,z)$, and when $x\neq 0$ we have
$j'(w,x,y)=j(w/x^3,y/x)$. We then define $c_k\:\R\to PX(a)$ as follows:
\begin{align*}
c_0(t) &=
j'(-\sqrt{(a^{-1}-a)^2+4\sin^2(2t)},\;e^{it},\;-e^{-it}) \\
c_1(t) &=
j'\left(\frac{1+i}{8\rt}\sin(t)
\sqrt{16\cos(t)^2+(a+a^{-1})^2\sin(t)^4},\;
\frac{1+\cos(t)}{2},\;\frac{1-\cos(t)}{2}i\right) \\
c_2(t) &= \lm(c_1(t)) \\
c_3(t) &=
j'\left(-i\frac{a^{-1}-a}{8}
\sin(t)\sqrt{(1+a)^4-(1-a)^4\cos(t)^2}
\sqrt{(1+a)^2-(1-a)^2\cos(t)^2},\right. \\
& \left.\qquad\qquad\vphantom{\frac{a^{-1}-a}{8}}
\frac{(1+a)+(1-a)\cos(t)}{2},\;
\frac{(1+a)-(1-a)\cos(t)}{2}\right) \\
c_4(t) &= \lm(c_3(t)) \\
c_5(t) &= j\left(\frac{\sin(t)}{8}\sqrt{2a(3-\cos(t))(4-a^4(1-\cos(t))^2)},\;
a \frac{1 - \cos(t)}{2}\right) \\
c_6(t) &= \lm(c_5(t)) \\
c_7(t) &= \mu(c_5(t)) \\
c_8(t) &= \lm\mu(c_5(t)).
\end{align*}
\end{definition}
Maple notation for $c_k(t)$ is \mcode+c_P[k](t)+. The versions with a
numerical values for $a$ are \mcode+c_P0[k](t)+ and
\mcode+c_P1[k](t)+. The map $j'$ is \mcode+jj_P+.
\begin{proposition}\label{prop-P-curves}
The above maps give a curve system on $PX(a)$ (in the sense of
Definition~\ref{defn-curve-system}).
\end{proposition}
\begin{proof}
Combine Lemmas~\ref{lem-P-curves-bc} and~\ref{lem-P-curves-a}.
\begin{checks}
cromulent.mpl: check_precromulent("P")
\end{checks}
\end{proof}
\begin{lemma}\label{lem-P-curves-bc}
For $0\leq k\leq 8$, the map $c_k\:\R\to\C P^4$ is smooth, with
$c_k(t+2\pi)=c_k(t)$, and the image is contained in $PX(a)$.
Moreover, parts~(b) and~(c) of Definition~\ref{defn-curve-system} are
satisfied.
\end{lemma}
\begin{proof}
Direct calculation.
\end{proof}
\begin{lemma}\label{lem-P-pc}
The composites $\R\xra{c_k}PX(a)\xra{p}\C_\infty$ and their images
are as follows:
\begin{align*}
pc_0(t) &= - e^{-2it} &
pc_0(\R) &= S^1 \\
pc_1(t) &= \pp\frac{1-\cos(t)}{1+\cos(t)}i &
pc_1(\R) &= [0,\infty] i \\
pc_2(t) &= -\frac{1-\cos(t)}{1+\cos(t)}i &
pc_2(\R) &= [-\infty,0] i \\
pc_3(t) &= \pp\frac{(1+a)-(1-a)\cos(t)}{(1+a)+(1-a)\cos(t)} &
pc_3(\R) &= [a,a^{-1}] \\
pc_4(t) &= -\frac{(1+a)-(1-a)\cos(t)}{(1+a)+(1-a)\cos(t)} &
pc_4(\R) &= [-a^{-1},-a] \\
pc_5(t) &= \pp\frac{1-\cos(t)}{2}a &
pc_5(\R) &= [0,a] \\
pc_6(t) &= -\frac{1-\cos(t)}{2}a &
pc_6(\R) &= [-a,0] \\
pc_7(t) &= \pp\frac{2}{1-\cos(t)}a^{-1} &
pc_7(\R) &= [a^{-1},\infty] \\
pc_8(t) &= -\frac{2}{1-\cos(t)}a^{-1} &
pc_8(\R) &= [-\infty,-a^{-1}] \\
\end{align*}
\end{lemma}
\begin{proof}
Direct calculation.
\begin{checks}
projective/PX_check.mpl: check_pc_P()
\end{checks}
\end{proof}
\begin{lemma}\label{lem-jj-equal}
Suppose that $j'(w_0,x,y)=j'(w_1,x,y)$ and that this point lies in
$PX(a)$. Then $w_0=w_1$.
\end{lemma}
\begin{proof}
By assumption we have
\[ [w_0:x^3:x^2y:xy^2:y^3]=[w_1:x^3:x^2y:xy^2:y^3]. \]
It follows easily that $w_0=w_1$ unless $x=y=0$. However, as this
point lies in $PX(a)$, Lemma~\ref{lem-nonzero} tells us that $x$ and
$y$ cannot both vanish.
\end{proof}
\begin{lemma}\label{lem-P-curves-a}
For $0\leq k\leq 8$, the induced map $c_k\:\R/2\pi\Z\to PX(a)$ is a
smooth embedding.
\end{lemma}
\begin{proof}
Because of the group action, it will suffice to treat the cases
$k=0,1,3,5$.
Using Lemma~\ref{lem-P-pc} we see that $(pc_k)'(t)\neq 0$ except
when $k>0$ and $t\in\pi\Z$. It follows that $c_k'(t)\neq 0$ except
possibly when $k>0$ and $t\in\pi\Z$. Moreover, one can check that to
first order in $\ep$ we have
\begin{align*}
c_1(\ep) &\simeq \left[\frac{1+i}{2\rt}\ep:1:0:0:0\right] \\
c_1(\ep+\pi) &\simeq \left[\frac{1-i}{2\rt}\ep:0:0:0:1\right].
\end{align*}
It follows easily that $c'_1(t)\neq 0$ for all $t$, so $c_1$ is at
least an immersion. Similar calculations show that $c_0,\dotsc,c_8$
are all immersions.
We now just need to show that $c_k\:\R/2\pi\Z\to PX(a)$ is
injective.
\begin{itemize}
\item[(a)] Consider the case $k=0$. If $u=c_0(t)$ we have
$p(u)=-e^{-2it}$, and it follows that the quantity
$a^{-2}+a^2-p(u)^2-p(u)^{-2}$ is equal to $a^{-2}+a^2-2\cos(4t)$
and so is strictly positive. It follows in turn that
\[ -q(u)\,p(u)^{-2}\,(a^{-2}+a^2-p(u)^2-p(u)^{-2})^{-1/2} = e^{it}.
\]
From this it is clear that when $c_0(s)=c_0(t)$ we have
$e^{is}=e^{it}$ and so $s-t\in 2\pi\Z$ as required.
\item[(b)] Now suppose instead that $k\in\{1,3,5\}$ and
$c_k(s)=c_k(t)$. We then have $pc_k(s)=pc_k(t)$, and using
Lemma~\ref{lem-P-pc} we can deduce that $\cos(s)=\cos(t)$. Now, we
can use the identities $\sin(2t)=\sin(t)\cos(t)$ and
$\sin^2(t)=1-\cos^2(t)$ to rewrite $c_k(t)$ in the form
\[ c_k(t)=j'(u(\cos(t))\sin(t),\;v(\cos(t)),\;w(\cos(t))), \]
for some functions $u$, $v$ and $w$. As
$\cos(s)=\cos(t)$ and $c_k(s)=c_k(t)$ we can use
Lemma~\ref{lem-jj-equal} to see that
\[ u(\cos(t))\sin(s) = u(\cos(t))\sin(t). \]
Moreover, in each case one can check that $u(\cos(t))$ is never
zero, so $\sin(s)=\sin(t)$. We thus have $s-t\in 2\pi\Z$ again.
\end{itemize}
\end{proof}
\begin{proposition}\label{prop-P-std-isotropy}
$PX(a)$ has standard isotropy (as in
Definition~\ref{defn-std-isotropy}).
\end{proposition}
\begin{proof}
First, we put $D_k=p(C_k)\sse\C_\infty$; these sets are described by
Lemma~\ref{lem-P-pc}. Note that we have
$\lm^2(c_0(t))=c_0(t+\pi)$, and $\lm^2(c_k(t))=c_k(-t)$ for $1\leq
k\leq 8$. It follows that for all $k$ we have $\lm^2(C_k)=C_k$. In
view of Remark~\ref{rem-P-quotient}, it follows that
$C_k=p^{-1}(D_k)$.
It is clear that the sets $D_4=[-a^{-1},-a]$, $D_5=[0,a]$ and
$D_7=[a^{-1},\infty]$ are disjoint, and it follows that $C_4$, $C_5$
and $C_7$ are disjoint. Similarly, $C_3$, $C_6$ and $C_8$ are
disjoint.
Next, recall that the map $p\:PX(a)\to\C_\infty$ is equivariant with
respect to the action described in Remark~\ref{rem-P-quotient}
(given by $\lm(z)=-z$ and $\mu(z)=z^{-1}$ and $\nu(z)=\ov{z}$). In
particular, if $x\in PX(a)$ is fixed by an element $\al\in G$, then
$p(x)$ is also fixed by $\al$.
\begin{itemize}
\item[(a)] Consider a point $x\in PX(a)$ with $\mu\nu(x)=x$. Then
$p(x)$ is also fixed by $\mu\nu$, which means that
$p(x)=1/\ov{p(x)}$, so $p(x)\in S^1=D_0$, so $x\in
p^{-1}(D_0)=C_0$. We thus have $PX(a)^{\ip{\mu\nu}}=C_0$ as claimed.
\item[(b)] Consider a point $x\in PX(a)$ with $\lm\nu(x)=x$. If
$x=v_0=c_1(0)$ or $x=v_1=c_1(\pi)$ then it is clear that
$x\in C_1$. Suppose instead that $x\not\in\{v_0,v_1\}$, so can be
written as $j(w,z)$ with $z\neq 0$. In general we have
$\lm\nu(j(w,z))=j(i\ov{w},-\ov{z})$. As $\lm\nu(x)=x$ we see that
$w=i\ov{w}$ and $z=-\ov{z}$. Put
$\om=e^{i\pi/4}=\frac{1+i}{\rt}$, so $\om^2=i$ and
$\om=i\ov{\om}$. We find that $w=\om w_1$ and $z=iz_1$ for some
$w_1,z_1\in\R$. The equation $w^2=r_a(z)$ becomes
$w_1^2=z_1(z_1^2+a^2)(z_1^2+a^{-2})$. From this it is clear that
$z_1>0$. By elementary calculus, there is a unique $t\in(0,\pi)$
with $z_1=(1-\cos(t))/(1+\cos(t))$. For this $t$ we find that
$p(c_1(t))=z$, and thus that $x$ is either $c_1(t)$ or
$\lm^2(c_1(t))=c_1(-t)$. Either way we have $x\in c_1(\R)=C_1$,
so $PX(a)^{\ip{\lm\nu}}=C_1$ as claimed.
\item[(c)] As $\lm^3\nu$ is conjugate to $\lm\nu$, we can use the
group action to deduce that $PX(a)^{\ip{\lm^3\nu}}=C_2$.
\item[(d)] Now consider a point $x\in PX(a)$ with $\nu(x)=x$. If
$x=v_1$ then $x\in C_7$. Otherwise, we have $x=j(w,z)$ for some
$(w,z)\in PX_0(a)$. As $\nu(x)=x$ we see that $w$ and $z$ are
real. As $r_a(z)=w^2$ we have $r_a(z)\geq 0$. Recall
that the roots of $r_a(z)$, listed in increasing order, are
$-a^{-1},-a,0,a$ and $a^{-1}$. It follows that
\[ p(x) = z \in [-a^{-1},-a] \amalg [0,a] \amalg [a^{-1},\infty] \\
= D_4\amalg D_5 \amalg D_7,
\]
and thus that $x\in C_4\amalg C_5\amalg C_7$.
\item[(e)] Consider instead a point $x\in PX(a)$ with
$\lm^2\nu(x)=x$. Then the point $y=\lm(x)$ satisfies $\nu(x)=x$
and so lies in $C_4\amalg C_5\amalg C_7$. However, one can check
from Definition~\ref{defn-precromulent-C} that
\begin{align*}
\lm(C_3) &= C_4 & \lm(C_4) &= C_3 \\
\lm(C_5) &= C_6 & \lm(C_6) &= C_5 \\
\lm(C_7) &= C_8 & \lm(C_8) &= C_7,
\end{align*}
so $x\in C_3\amalg C_6\amalg C_8$.
\item[(f)] Finally, consider a point $x\in PX(a)$ with
$x=\lm^2\mu\nu(x)$. Then $x$ cannot be equal to $v_1$, so
$x=j(w,z)$ for some $w$ and $z$. The equation $x=\lm^2\mu\nu(x)$
gives $z=-1/\ov{z}$ and so $|z|^2=-1$, which is impossible. Thus,
there are no such points $x$.
\end{itemize}
\end{proof}
This is a convenient place to record the following result, which will
be needed later.
\begin{lemma}\label{lem-right-angle}
\begin{itemize}
\item[(a)] The curves $C_0$ and $C_3$ cross at right angles at $v_3$
\item[(b)] The curves $C_0$ and $C_1$ cross at right angles at $v_6$
\item[(c)] The curves $C_3$ and $C_5$ cross at right angles at $v_{11}$.
\end{itemize}
\end{lemma}
If we were willing to wait until we had proved
Lemma~\ref{lem-bt-al-fixed}, we could give a non-computational
argument based on that. However, we will just calculate the relevant
derivatives instead.
\begin{proof}
For~(a), recall that $v_3=c_0(\pi/2)=c_3(\pi/2)$, so we need to
compare $c'_0(\pi/2)$ with $c'_2(\pi/2)$. Because the map
$p\:PX(a)\to\C_\infty$ is conformal, it will suffice to show that
$(pc_0)'(\pi/2)$ and $(pc_3)'(\pi/2)$ are nonzero and that the ratio
between them is purely imaginary. This is easy to do using the
formulae in Lemma~\ref{lem-P-pc}. Specifically, we have
$(pc_0)'(\pi/2)=-2i$ and $(pc_3)'(\pi/2)=2(1-a)/(1+a)$. We can
prove~(b) in the same way using $(pc_0)'(\pi/4)=2$ and
$(pc_1)'(\pi/2)=2i$. We need a slightly different method for
$v_{11}$ because $p$ has derivative zero there. We instead define a
rational map $q\:PX(a)\to\C_\infty$ by $q(z)=z_1/z_2$, so
$qj(w,z)=w$. This is conformal and satisfies $q(v_{11})=0$, so it
will suffice to show that $(qc_3)'(0)$ and $(qc_5)'(\pi)$ are
nonzero, and that the ratio between them is purely imaginary. A
standard calculation from the definitions gives
\begin{align*}
(qc_3)'(0) &= -i(1-a^2)(1+a^2)^{1/2}/\rt \\
(qc_5)'(\pi) &= -(1-a^2)^{1/2}(1+a^2)^{1/2}a^{1/2}/\rt
\end{align*}
as required.
\end{proof}
\subsection{Fundamental domains}
\label{sec-P-fundamental}
\begin{proposition}\label{prop-P-fundamental}
If we put
\begin{align*}
PF'_{16}(a) &= \{(w,z)\in PX_0(a)\st
\text{Re}(z),\text{Im}(z),\text{Re}(w)\geq 0,\;
\text{Re}(w)\geq\text{Im}(w),\;|z|\leq 1\} \\
PF_{16}(a) &= j(PF'_{16}(a))\subset PX(a),
\end{align*}
then $PF_{16}(a)$ is a standard fundamental domain for $PX(a)$ (as in
Definition~\ref{defn-standard-F}). Thus, $PX(a)$ is cromulent (by
Remark~\ref{rem-standard-F}).
\end{proposition}
\begin{proof}
For brevity, we will write $F'$ and $F$ for $PF'_{16}(a)$ and
$PF_{16}(a)$. Put
\begin{align*}
Z &= \{x+iy\in\C\st x,y\geq 0,\;x^2+y^2\leq 1\} &
&= \{r\,e^{i\tht}\st 0\leq r\leq 1,\;0\leq\tht\leq\pi/2\} \\
W &= \{x+iy\in\C\st x\geq 0,\; x\geq y\} &
&= \{r\,e^{i\tht}\st 0\leq r,\; -\pi/2\leq\tht\leq\pi/4\} \\
W^2 &= \{x+iy\in\C\st x\geq 0 \text{ or } y\leq 0\} &
&= \{r\,e^{i\tht}\st 0\leq r,\; -\pi\leq\tht\leq\pi/2\}.
\end{align*}
We then have $F'=(W\tm Z)\cap PX_0(a)$.
We now claim that $r_a(Z)\sse W^2$. Indeed, it is clear that
$\partial Z$ is a simple closed curve. The image $r_a(\partial Z)$
consists of the points $r_a(t)\in\R$ (for $0\leq t\leq 1$) and
$r_a(it)\in i\R$ (for $0\leq t\leq 1$) and $r_a(e^{it})$ (for
$0\leq t\leq\pi/2$). Here
\[ r_a(e^{it}) = - (4\sin(t)^2+(a^{-1}-a)^2) e^{3it}, \]
so $\text{arg}(r_a(e^{it}))=3t-\pi\in[-\pi,\pi/2]$, so
$r_a(e^{it})\in W^2$. We now see that $r_a(\partial Z)$ is a simple
closed curve in $W^2$. The argument principle shows that $r_a(Z)$ is
the interior of $r_a(\partial Z)$, and this is contained in $W^2$ as
claimed.
Now consider a point $v\in PX(a)$. If $p(v)=x+iy\in\C$ then we put
\begin{align*}
s_0(v) &= \frac{|x| + i|y|}{\max(1,x^2+y^2)} \in Z \\
s(v) &= j(\sqrt{r_a(s_0(v))},s_0(v)) \in F'.
\end{align*}
(Here $r_a(s_0(v))$ lies in $W^2$, and $\sqrt{r_a(s_0(v))}$ refers to
the unique choice of square root that lies in $W$.) For the
exceptional case $v=v_1$, we put $s(v_1)=(0,0)$. It is easy to see
that $s$ is a retraction. Using Remark~\ref{rem-P-quotient}, we
see that $s_0(v)=s_0(v')$ iff $Gv=Gv'$, and also that
$s_0(v)\in G.p(v)$. After recalling that
$\lm^2j(w,z)=j(-w,z)$, we deduce that $s(v)=s(v')$ iff $Gv=Gv'$, and
also that $s(v)\in G.v$. It follows that
$PX(a)=\bigcup_{\gm\in G}\gm.F'$, with
\[ F'\cap\gm F' = \{v\in F'\st \gm(v)=v\}. \]
If $v$ lies in the interior of $F'$ then it is easy to see that
$p(v)$ lies in the interior of $Z$, and thus that
$\stab_G(v)\sse\stab_G(p(v))=\{1,\lm^2\}$. On the other hand, for
$v$ in the interior of $F'$ we also have $r_a(p(v))\neq 0$, so $v$ is
not fixed by $\lm^2$, so $\stab_G(v)=1$. We now see that
$\text{int}(F')\cap\gm(F')=\emptyset$ for $\gm\neq 1$, so $F'$ is a
retractive fundamental domain for $PX(a)$.
Next, the formulae in Lemma~\ref{lem-P-pc} show that
\begin{align*}
\partial Z &= [0,a] \cup [a,1] \cup e^{[0,\pi/2]i} \cup [i,0] \\
&= pc_5([0,\pi])\cup pc_3([0,\pi/2])\cup
pc_0([\pi/4,\pi/2]) \cup pc_1([0,\pi/2]).
\end{align*}
From this we deduce that
\[ \partial F' =
c_5([0,\pi])\cup c_3([0,\pi/2])\cup
c_0([\pi/4,\pi/2]) \cup c_1([0,\pi/2]) = DF_{16},
\]
so $F'$ is a standard fundamental domain.
\end{proof}
We can illustrate the surface $PX(a)$ as follows. The picture on the
left shows the image under $p\:PX(a)\to\C_\infty$ of the fundamental
domain $F$, and the picture on the right shows the image under $q$.
(In both cases the origin is at $v_0$.)
\[
\begin{tikzpicture}[scale=4]
\draw[blue] (0,0) -- (0.600,0);
\draw[magenta] (0.600,0) -- (1,0);
\draw[cyan] (0,0) (1,0) arc(0:90:1);
\draw[green] (0,1) -- (0,0);
\fill[black] (0.600,0) circle(0.015);
\fill[black] (0,0) circle(0.015);
\fill[black] (1,0) circle(0.015);
\fill[black] (0,1) circle(0.015);
\draw (0.000,0.000) node[anchor=north] {$\ss v_0$};
\draw (0.600,0.000) node[anchor=north] {$\ss v_{11}$};
\draw (0.000,1.000) node[anchor=south] {$\ss v_6$};
\draw (1.000,0.000) node[anchor=north] {$\ss v_3$};
\end{tikzpicture}
\hspace{10em}
\begin{tikzpicture}[scale=2]
\draw[magenta] (0,0) -- (0,-1.067);
\draw[cyan] plot[smooth]
coordinates{ (0.000,-1.067) (0.259,-1.081) (0.560,-1.098)
(0.910,-1.065) (1.284,-0.933) (1.637,-0.678)
(1.914,-0.303) (2.070,0.163) (2.074,0.674)
(1.914,1.173) (1.603,1.603) };
\draw[green] (1.603,1.603) -- (0,0);
\draw[blue] (0,0) -- (0.470,0);
\fill[black] (0.000,0.000) circle(0.03);
\fill[black] (0.000,-1.067) circle(0.03);
\fill[black] (1.603,1.603) circle(0.03);
\draw (0.000,0.000) node[anchor=east] {$\ss v_0,v_{11}$};
\draw (0.000,-1.067) node[anchor=east] {$\ss v_3$};
\draw (1.603,1.603) node[anchor=south] {$\ss v_6$};
\end{tikzpicture}
\]
We next consider differential forms on $PX_0(a)$ and $PX(a)$.
\begin{remark}\label{rem-anticonformal-forms}
Holomorphic differential forms are clearly functorial for conformal
isomorphisms. In fact, they are also functorial for anticonformal
isomorphisms. Indeed, given an anticonformal map $\phi\:Z_0\to Z_1$
of Riemann surfaces and a holomorphic function $f\in\CO(Z_1)$, we can
define $\phi^\#(f)\in\CO(Z_0)$ by $\phi^\#(f)(z)=\ov{f(\phi(z))}$.
It is not hard to see that there is a unique locally determined map
$\phi^\#\:\Om^1(Z_1)\to\Om^1(Z_0)$ satisfying
$\phi^\#(f\,dg)=\phi^\#(f)\,d\phi^\#(g)$ for all $f,g\in\CO(Z_1)$.
We therefore have an action of $G$ on $\Om^1(PX(a))$.
\end{remark}
\begin{proposition}\label{prop-holomorphic-forms}\leavevmode
\begin{itemize}
\item[(a)] The differential form $\om_0\in\Om^1(U_0)$ (from
Lemma~\ref{lem-P-differentials}) extends to give a holomorphic
differential form on all of $PX(a)$ (which we also call $\om_0$).
\item[(b)] The form $\om_1=\mu^*(\om_0)$ satisfies $\om_1=z\,\om_0$
when restricted to $U_0$.
\item[(c)] The set $\{\om_0,\om_1\}$ is a basis for $\Om^1(PX(a))$
over $\C$.
\item[(d)] The group $G$ acts on this space by
\begin{align*}
\lm^*(\om_0) &= \pp i\om_0 &
\mu^*(\om_0) &= \om_1 &
\nu^\#(\om_0) &= \om_0 \\
\lm^*(\om_1) &= -i\om_1 &
\mu^*(\om_1) &= \om_0 &
\nu^\#(\om_1) &= \om_1.
\end{align*}
\end{itemize}
\end{proposition}
\begin{proof}
As $z\om_0\in\Om^1(U_0)$ and $\mu(U_1)=U_0$ we have a holomorphic
form $\om'_0=\mu^*(z\,\om_0)\in\Om^1(U_1)$. Recall that
$w\om_0=dz$. After restricting to $U_{01}$ we can apply $\mu^*$ to
this equation, giving
\[ -wz^{-3}\mu^*(\om_0) = d(z^{-1})
= -z^{-2}\,dz
= -z^{-2}w\om_0,
\]
which implies that $\mu^*(\om_0)=z\om_0$, and thus that
$\mu^*(z\om_0)=\om_0$. This implies that $\om_0$ and $\om'_0$ have
the same restriction to $U_{01}$, so we can patch them together to
give a holomorphic form on all of $U_0\cup U_1=PX(a)$. Claims~(a)
and~(b) are now clear, and~(d) is a straightforward calculation.
This just leaves~(c). Consider a holomorphic form
$\al\in\Om^1(PX(a))$. Lemma~\ref{lem-P-differentials} tells us
that there is a unique function $f_0\in R_0(a)$ such that
$\al=f_0\om_0$ on $U_0$. By applying the same logic to $\mu^*(\al)$,
and applying $\mu^*$ again, we see that there is also a unique
function $f_1\in R_1(a)$ such that $\al=f_1\om_1$ on $U_1$. On
$U_{01}$ we now see that $f_0\om_0=\al=f_1\om_1=f_1z\om_0$, so
$f_0=f_1z$. Using the bases described in Remark~\ref{rem-P-cover}
we see that $f_0\in R_0(a)\cap R_1(a)z=\C\{1,z\}$, and it follows
that $\al\in\C\{\om_0,\om_1\}$. Moreover, it is easy to see that
$\om_1$ vanishes at $v_0$ but $\om_0$ does not, and the other way
around at $v_1$. This means that $\om_0$ and $\om_1$ are linearly
independent, so they form a basis for $\Om^1(PX(a))$.
\end{proof}
\begin{remark}\label{rem-P-parameter}
The coordinate $w$ is a local parameter on $PX_0(a)$ at the point
$v_0=(0,0)$. In terms of this parameter we have $\om_0=2\,dw+O(w^4)$
and $\om_1=2w^2\,dw+O(w^6)$.
\end{remark}
\begin{definition}\label{defn-period-p}
The \emph{periods} for $PX(a)$ are the numbers
$p_{jk}(a)=\int_{c_j}\om_k\in\C$ (for $0\leq j\leq 8$ and
$k\in\{0,1\}$).
\end{definition}
\subsection{Galois theory}
\label{sec-galois}
Let $PK(a)$ denote the field of rational functions on $PX_0(a)$ (which
is the same as the field of rational functions on $PX(a)$). This can
be described as
\[ PK(a) = \C(z)[w]/(w^2-r_a(z)). \]
This field has an action of the group $D_8$, and for any subgroup
$H\leq D_8$, we can identify the fixed field $PK(a)^H$ with the field
of rational functions on the quotient $PX(a)/H$, with its standard
structure as a Riemann surface. The subgroups of $D_8$ can be
enumerated as follows:
\begin{center}
\begin{tikzpicture}[scale=3.1]
\def0.7{0.7}
\def1.4{1.4}
\def2.1{2.1}
\def0.8{0.8}
\def1.6{1.6}
\def( 0.0, 0.0){( 0.0, 0.0)}
\def(-\xa, \ya){(-0.8, 0.7)}
\def( \xa, \ya){( 0.8, 0.7)}
\def( 0, \ya){( 0, 0.7)}
\def(-\xb, \yb){(-1.6, 1.4)}
\def( \xa, \yb){( 0.8, 1.4)}
\def(-\xa, \yb){(-0.8, 1.4)}
\def( \xb, \yb){( 1.6, 1.4)}
\def( 0, \yb){( 0, 1.4)}
\def( 0, \yc){( 0, 2.1)}
\begin{scope}[xshift=1cm]
\draw ( 0.0, 0.0) node{$H_{10}=D_8$};
\draw ( 0, \ya) node{$H_3=\ip{\lm}$};
\draw (-\xa, \ya) node{$H_8=\ip{\lm^2,\mu}$};
\draw ( \xa, \ya) node{$H_9=\ip{\lm^2,\lm\mu}$};
\draw (-\xb, \yb) node{$H_6=\ip{\lm^2\mu}$};
\draw ( \xa, \yb) node{$H_5=\ip{\lm\mu}$};
\draw (-\xa, \yb) node{$H_4=\ip{\mu}$};
\draw ( \xb, \yb) node{$H_7=\ip{\lm^3\mu}$};
\draw ( 0, \yb) node{$H_2=\ip{\lm^2}$};
\draw ( 0, \yc) node{$H_1=1$};
\draw[<-,shorten <=15pt,shorten >=15pt] ( 0.0, 0.0) -- (-\xa, \ya);
\draw[<-,shorten <=15pt,shorten >=15pt] ( 0.0, 0.0) -- ( \xa, \ya);
\draw[<-,shorten <=15pt,shorten >=15pt] ( 0.0, 0.0) -- ( 0, \ya);
\draw[<-,shorten <=15pt,shorten >=15pt] ( 0, \ya) -- ( 0, \yb);
\draw[<-,shorten <=15pt,shorten >=15pt] (-\xa, \ya) -- (-\xb, \yb);
\draw[<-,shorten <=15pt,shorten >=15pt] (-\xa, \ya) -- (-\xa, \yb);
\draw[<-,shorten <=15pt,shorten >=15pt] (-\xa, \ya) -- ( 0, \yb);
\draw[<-,shorten <=15pt,shorten >=15pt] ( \xa, \ya) -- ( \xa, \yb);
\draw[<-,shorten <=15pt,shorten >=15pt] ( \xa, \ya) -- ( \xb, \yb);
\draw[<-,shorten <=15pt,shorten >=15pt] ( \xa, \ya) -- ( 0, \yb);
\draw[<-,shorten <=19pt,shorten >=15pt] (-\xb, \yb) -- ( 0, \yc);
\draw[<-,shorten <=15pt,shorten >=15pt] ( \xa, \yb) -- ( 0, \yc);
\draw[<-,shorten <=15pt,shorten >=15pt] (-\xa, \yb) -- ( 0, \yc);
\draw[<-,shorten <=19pt,shorten >=15pt] ( \xb, \yb) -- ( 0, \yc);
\draw[<-,shorten <=15pt,shorten >=15pt] ( 0, \yb) -- ( 0, \yc);
\end{scope}
\end{tikzpicture}
\end{center}
We will describe the fixed fields $L_i=PK(a)^{H_i}$ in terms of the
following elements:
\begin{align*}
t_0 &= z & u_0 &= w \\
t_1 &= z^2 &
u_1 &= \frac{2w(1-z)}{(1+z^2)^2} &
v_1 &= w(1-1/z^3)/2 \\
t_2 &= \frac{2z}{1+z^2} &
u_2 &= -\frac{1+i}{\rt}\,\frac{2w(i+z)}{(1-z^2)^2} &
v_2 &= w(1-i/z^3)/2 \\
t_3 &= \frac{2iz}{1-z^2} &
u_3 &= \frac{-2iw(1+z)}{(1+z^2)^2} &
v_3 &= w(1+1/z^3)/2 \\
t_4 &= \frac{2z^2}{1+z^4} &
u_4 &= -\frac{1-i}{\rt}\,\frac{2w(i-z)}{(1-z^2)^2} &
v_4 &= w(1+i/z^3)/2.
\end{align*}
One can check that
\begin{align*}
L_1 &= \C(t_0)\{1,u_0\} \\
L_2 &= \C(t_0) \\
L_3 &= \C(t_1) \\
L_4 &= \C(t_2)\{1,u_1\} = \C(t_2)\{1,v_1\} \\
L_5 &= \C(t_3)\{1,u_2\} = \C(t_3)\{1,v_2\} \\
L_6 &= \C(t_2)\{1,u_3\} = \C(t_2)\{1,v_3\} \\
L_7 &= \C(t_3)\{1,u_4\} = \C(t_3)\{1,v_4\} \\
L_8 &= \C(t_2) \\
L_9 &= \C(t_3) \\
L_{10} &= \C(t_4).
\end{align*}
Details for $L_4$ and $L_5$ will be given in
Section~\ref{sec-ellquot}. The cases $L_6$ and $L_7$ can be recovered
from this, because $H_6$ and $H_7$ are conjugate to $H_4$ and $H_5$
respectively. The other cases are relatively easy (and are easily
seen to be consistent with Corollary~\ref{cor-quotient-types}, which
gives the genera of the quotients $PX(a)/H_i$). One can find further
information in the files \fname+parabolic/galois.mpl+ and
\fname+parabolic/PK_subfields.mpl+.
\begin{checks}
parabolic/galois_check.mpl: check_PK(), check_PK_subfields()
\end{checks}
\subsection{Elliptic quotients}
\label{sec-ellquot}
We next study the quotients $PX(a)/\ip{\mu}$ and $PX(a)/\ip{\lm\mu}$.
First note that there is no natural action of the full group $G$ on
$PX(a)/\ip{\mu}$. Instead, there is an action of the centraliser of
$\mu$, which is $\ip{\lm^2,\mu,\nu}\simeq C_2^3$. This action factors
through the quotient group $\ip{\lm^2,\mu,\nu}/\ip{\mu}\simeq C_2^2$.
Similarly, we have a natural action of the group
$\ip{\lm^2,\lm\mu,\mu\nu}/\ip{\lm\mu}\simeq C_2^2$ on $PX(a)/\ip{\lm\mu}$.
\begin{definition}\label{defn-ellquot}
We put $b_{\pm}=(a^{-1}\pm a)/2$, and define affine curves
$E_0^{\pm}(a)$ as follows.
\begin{align*}
q^+_a(x) &= 2x(x-1)\left(b_+^2x^2-1\right) \\
q^-_a(x) &= 2x(x-1)\left(b_-^2x^2+1\right) \\
E^+_0(a) &= \{(y,x)\in\C^2\st y^2=q_+(x)\} \\
E^-_0(a) &= \{(y,x)\in\C^2\st y^2=q_-(x)\}.
\end{align*}
We can obtain smooth projective completions of these curves by taking
the closures of their images under the map $j\:\C^2\to\C P^3$ given
by
\[ j(y,x) = [y:1:x:x^2]. \]
The results are
\begin{align*}
E^+(a) &=
\{[z]\st z_2z_4-z_3^2 = z_1^2 - 2(z_4-z_3)(b_+^2z_4-z_2)=0\} \\
E^-(a) &=
\{[z]\st z_2z_4-z_3^2 = z_1^2 - 2(z_4-z_3)(b_-^2z_4+z_2)=0\}.
\end{align*}
We define an action of the group $\ip{\lm^2,\mu,\nu}$ on $E^+(a)$ as
follows:
\begin{align*}
\lm^2[z] &= [ -z_1:z_2:z_3:z_4] & \lm^2j(y,x) &= j( -y,x) \\
\mu[z] &= [\pp z_1:z_2:z_3:z_4] & \mu j(y,x) &= j(\pp y,x) \\
\nu[z] &= [\pp \ov{z_1}:\ov{z_2}:\ov{z_3}:\ov{z_4}] &
\nu j(y,x) &= j(\pp\ov{y},\ov{x}).
\end{align*}
Similarly, we define an action of the group
$\ip{\lm^2,\lm\mu,\mu\nu}$ on $E^-(a)$ as follows:
\begin{align*}
\lm^2[z] &= [ -z_1:z_2:z_3:z_4] & \lm^2j(y,x) &= j( -y,x) \\
\lm\mu[z] &= [\pp z_1:z_2:z_3:z_4] & \lm\mu j(y,x) &= j(\pp y,x) \\
\mu\nu[z] &= [\pp \ov{z_1}:\ov{z_2}:\ov{z_3}:\ov{z_4}] &
\mu\nu j(y,x) &= j(\pp\ov{y},\ov{x}).
\end{align*}
\end{definition}
\begin{checks}
projective/ellquot_check.mpl: check_ellquot()
\end{checks}
\begin{remark}
Code for all this is in \fname+ellquot.mpl+. The polynomials $q^+_a$
and $q^-_a$ are \mcode+q_Ep+ and \mcode+q_Em+. Elements of
$E^+_0(a)$ and $E^-_0(a)$ are represented as lists of length two,
whereas elements of $E^+(a)$ and $E^-(a)$ are represented as lists of
length four. The functions \mcode+is_equal_Ep+ and
\mcode+is_equal_Em+ (which are actually the same) can be used to test
projective equality. The function \mcode+is_member_Ep_0+ can be used
to test whether a point lies in $E^+_0(a)$, and similarly for
\mcode+is_member_Em_0+, \mcode+is_member_Ep+ and
\mcode+is_member_Em+. The inclusion $E^+_0(a)\to E^+(a)$ and its
inverse are \mcode+j_Ep+ and \mcode+j_inv_Ep+, and similarly for
\mcode+j_Em+ and \mcode+j_inv_Em+. The function \mcode+NF_Ep+ can be
used to reduce a polynomial in $z_1,\dotsc,z_5$ to normal form modulo
the Gr\"obner basis for the ideal that defines $E^+(a)$. There is a
similar function \mcode+NF_Em+ for $E^-(a)$. Actions of $G$ are
given by \mcode+act_Ep_0+, \mcode+act_Em_0+, \mcode+act_Ep+ and
\mcode+act_Em+.
\end{remark}
\begin{proposition}\label{prop-P-Ep}
There is a unique morphism $\phi^+\:PX(a)\to E^+(a)$ satisfying
\[ \phi^+(j(w,z)) = j\left(\frac{2w(1-z)}{(1+z^2)^2},
\frac{2z}{1+z^2}\right)
\]
for all $(w,z)\in PX_0(a)$ with $z\neq\pm i$. Moreover, this is
equivariant with respect to $\ip{\lm^2,\mu,\nu}$, and it induces an
isomorphism $PX(a)/\ip{\mu}\to E^+(a)$.
\end{proposition}
Two variants of this map are represented by \mcode+P_to_Ep+ and
\mcode+P_to_Ep_0+.
\begin{proof}
First define $\psi\:\C^5\to\C^4$ by
\[ \psi(z) = (2(z_2-z_3)z_1,\;z_2^2+2z_2z_4+z_3z_5,\;
2z_2(z_3+z_5),\;4z_2z_4).
\]
This is homogeneous of degree two, so it induces a map
$\ov{\psi}\:U\to\C P^3$, where $U=\{[z]\in\C P^5\st\psi(z)\neq 0\}$.
Now put
\[ V = \{j(w,z)\in j(PX_0(a))\st z\not\in\{0,i,-i\}\}, \]
and note that this is open and dense in $PX(a)$ and is preserved by
$G$. It is straightforward to check that $V\sse j(PX_0(a))\sse U$,
so we can define $\phi^+_0$ to be the restriction of $\ov{\psi}$ to
$PX_0(a)$. It follows easily from the definitions that for
$j(w,z)\in V$ we have
\[ \phi^+_0j(w,z) = j\left(\frac{2w(1-z)}{(1+z^2)^2},
\frac{2z}{1+z^2}\right),
\]
and that this lies in $E^+(a)$. From this we also see that the
restriction of $\phi^+_0$ to $V$ is equivariant, and in particular
that $\phi^+_0=\phi^+_0\mu$ on $V$. By continuity, we must have
$\phi^+_0=\phi^+_0\mu$ on all of $PX_0(a)$. We can thus patch
together $\phi^+_0$ and $\phi^+_0\mu$ to get a morphism from
$PX(a)=j(PX_0(a))\cup\mu j(PX_0(a))$ to $E^+(a)$. The equivariance
conditions are satisfied on the open dense subset $V$, so they are
satisfied everywhere.
Now consider a point $(y,x)\in E^+_0(a)$ with $x\not\in\{0,1\}$. Let
$u$ be a square root of $1-x^2$, and put
\[ v_{\pm} = \left(
\frac{\pm(2-x)u-(x+2)(x-1)}{x^3(x-1)}y,\;
\frac{1\pm u}{x}
\right).
\]
Straightforward algebra shows that $v_+,v_-\in PX_0(a)$ with
$\mu(v_+)=v_-$, and that $(\phi^+)^{-1}\{(y,x)\}=\{v_+,v_-\}$. It
follows that the induced map $PX(a)/\ip{\mu}\to E^+(a)$ is
generically bijective. As the source and target are both smooth and
complete algebraic curves, it follows that the map is an isomorphism,
as claimed.
\begin{checks}
projective/ellquot_check.mpl: check_ellquot()
\end{checks}
\end{proof}
\begin{proposition}\label{prop-P-Em}
There is a unique map $\phi^-\:PX(a)\to E^-(a)$ satisfying
\[ \phi^-(j(w,z)) = \left(-\frac{\rt(1+i)w(i+z)}{(1-z^2)^2},
\frac{2iz}{1-z^2}\right) \in E^0_+(a)
\]
for all $(w,z)\in PX_0(a)$ with $z\neq\pm 1$. Moreover, this is
equivariant with respect to $\ip{\lm^2,\lm\mu,\mu\nu}$, and it
induces an isomorphism $PX(a)/\ip{\lm\mu}\to E^-(a)$.
\end{proposition}
Two variants of this map are represented by \mcode+P_to_Em+ and
\mcode+P_to_Em_0+.
\begin{proof}
Similar to the previous proposition, using the formulae
\[ \psi(z) = (\rt (1-i)z_1(z_2-iz_3),\;
z_2^2-2z_2z_4+z_3z_5,\;
2iz_2(z_3-z_5),\;
-4z_2z_4)
\]
and
\[ v_{\pm} = \left(
\frac{1+i}{\rt}\;\frac{\pm(x-2)u-(x+2)(x-1)}{x^3(x-1)}y,\;
\frac{\pm u-1}{x}i
\right).
\]
\begin{checks}
projective/ellquot_check.mpl: check_ellquot()
\end{checks}
\end{proof}
\begin{definition}
We will write $v_i^+=\phi^+(v_i)\in E^+(a)$, and similarly for
$v_i^-$.
\end{definition}
\begin{remark}\label{rem-Ep-infinity}
One can check that
\begin{align*}
v^+_6 = v^+_9 &= [\pp a+a^{-1}:0:0:-\rt] \\
v^+_7 = v^+_8 &= [ -a-a^{-1}:0:0:-\rt],
\end{align*}
and these are the only points in $E^+(a)\sm j(E^+_0(a))$. Similarly,
we have
\begin{align*}
v^-_2 = v^-_3 &= [\pp a-a^{-1}:0:0:-\rt] \\
v^-_4 = v^-_5 &= [ -a+a^{-1}:0:0:-\rt],
\end{align*}
and these are the only points in $E^-(a)\sm j(E^-_0(a))$.
\end{remark}
\begin{remark}\label{rem-elliptic-group}
One can also check that
\begin{align*}
v^+_0 = v^+_1 &= j(0,0) = [0:1:0:0] \in E^+(a) \\
v^-_0 = v^-_1 &= j(0,0) = [0:1:0:0] \in E^-(a).
\end{align*}
We use these points as the basepoints in $E^+(a)$ and $E^-(a)$. As
these are elliptic curves, each of them has a unique group structure
for which the specified basepoint is the zero element. We also find
that
\begin{align*}
j^{-1}\phi^+c_5(t) &= (\sqrt{a}t,0) + O(t^2) \\
j^{-1}\phi^+c_1(t) &= (e^{i\pi/4}t,0) + O(t^2) \\
j^{-1}\phi^-c_5(t) &= (e^{-i\pi/4}\sqrt{a}t,0) + O(t^2) \\
j^{-1}\phi^-c_1(t) &= (t,0) + O(t^2).
\end{align*}
Thus, if we use the map $z\mapsto j^{-1}(z)_1$ as a local coordinate
at the basepoint $v^\pm_0$, then $E^+(a)$ looks like our
standard picture $\Net_0$, but $E^-(a)$ is rotated clockwise by
$\pi/4$.
\begin{checks}
projective/ellquot_check.mpl: check_ellquot_origin()
\end{checks}
\end{remark}
\begin{definition}\label{defn-T-matrices}
We define matrices $T_i^\pm$ as follows:
{\tiny\begin{align*}
T_1^+ &= \bbm
b_+^2-1 & 0 & 0 & 0 \\
0 & 1 & -2b_+^2 & b_+^4 \\
0 & 1 & -b_+^2-1 & b_+^2 \\
0 & 1 & -2 & 1
\ebm &
T_1^- &= \bbm
b_-^2+1 & 0 & 0 & 0 \\
0 & -1 & -2b_-^2 & -b_-^4 \\
0 & -1 & -b_-^2+1 & b_-^2 \\
0 & -1 & 2 & -1
\ebm \\
T_2^+ &= \bbm
2(1-b_+) & 0 & 0 & 0 \\
0 & b_+ & 2b_+(b_+-2) & b_+(b_+-2)^2 \\
0 & 1 & -2 & -b_+(b_+-2) \\
0 & b_+^{-1} & -2 & b_+
\ebm &
T_2^- &= \bbm
2(1-b_-) & 0 & 0 & 0 \\
0 & b_- & 2ib_-(2i-b_-) & -b_-(2i-b_-)^2 \\
0 & i & -2i & -ib_-(2i-b_-) \\
0 & -b_-^{-1} & -2i & b_-
\ebm \\
T_3^+ &= \bbm
-2(1+b_+) & 0 & 0 & 0 \\
0 & b_+ & -2b_+(b_++2) & b_+(b_++2)^2 \\
0 & -1 & 2 & b_+(b_++2) \\
0 & b_+^{-1} & 2 & b_+
\ebm &
T_3^- &= \bbm
2(i+b_-) & 0 & 0 & 0 \\
0 & -b_- & -2ib_-(b_-+2i) & b_-(b_-+2i)^2 \\
0 & i & -2i & ib_-(b_-+2i) \\
0 & b_-^{-1} & -2i & -b_-
\ebm
\end{align*}}
One can check that in $PGL_4(\C)$ we have
\[ (T_1^+)^2 = (T_2^+)^2 = (T_3^+)^2 = T_1^+T_2^+T_3^+ =
(T_1^-)^2 = (T_2^-)^2 = (T_3^-)^2 = T_1^-T_2^-T_3^- = 1.
\]
One can also check that these matrices preserve the defining
equations for $E^+(a)$ or $E^-(a)$ as appropriate, so we have
holomorphic involutions $\tau_i^+\:E^+(a)\to E^+(a)$ and
$\tau_i^-\:E^-(a)\to E^-(a)$ for $i=1,2,3$. We also define
$\tau_0^\pm$ to be the identity.
\begin{checks}
projective/ellquot_check.mpl: check_translations()
\end{checks}
\end{definition}
The maps $\tau_i^+$ are \mcode+Ep_trans[i]+ (on $E^+_0(a)$) or
\mcode+Ep_0_trans[i]+ (on $E^+(a)$), and the maps $\tau_i^-$ are
\mcode+Em_trans[i]+ or \mcode+Em_0_trans[i]+.
Because $E^\pm(a)$ is an elliptic curve, it is standard that the line
bundle $\Om^1$ is trivial. For an elliptic curve in Weierstrass form
$y^2=x^3+ax+b$, it is also standard that $dx/y$ is a generator for
$\Om^1$. As our conventions are slightly different, it is not quite
so standard that the same formula remains valid, but we will now prove
that it is.
\begin{proposition}\label{prop-elliptic-forms}
There is a unique differential form $\om^\pm$ on $E^\pm(a)$ such that
$j^*(\om^\pm)=dx/y$ on $E_0^\pm(a)$. Moreover, this is everywhere
finite and nonzero, so it generates the module $\Om^1_{E^{\pm}(a)}$.
Near the origin we have $j^*(\om^\pm)=(1+O(y^2))dy$.
\end{proposition}
\begin{proof}
Differentiating the relation $y^2=q^+_a(x)$ gives $2y\,dy=(q^+_a)'(x)\,dx$
and thus $dx/y=y\,dx/(q^+_a)(x)=2dy/(q^+_a)'(x)$. Here $q_a^+(x)$ has no
repeated roots, so there are no points where $q_a^+(x)$ and $(q_a^+)'(x)$
both vanish. It follows that $dx/y$ is finite and nonzero everywhere
in $E^+_0(a)$. Next, recall that
$E^+(a)=j(E^+_0(a))\cup\{v_6^+,v_7^+\}$. Calculation shows that
$\tau^+_1(v^+_6)$ and $\tau^+_1(v^+_7)$ lie in $j(E^+_0(a))$, so
$E^+(a)=j(E^+_0(a))\cup\tau^+_1j(E^+_0(a))$. One can check from the
definitions that
\[ \tau^+_1j(y,x) = j\left(
\frac{b_-^2y}{(b_+^2x-1)^2},
\frac{x-1}{b_+^2x-1}
\right),
\]
and thus that $(\tau^+_1)^*(x/dy)$ agrees with $x/dy$ on their common
domain. We can thus patch together $dx/y$ with $(\tau^+_1)^*(x/dy)$
to get a form $\om^+$ which is finite and nonzero everywhere on
$E^+(a)$, as required. One can check that $(q_a^+)'(x)=2+O(x)$, and
the relation $y^2=q^+_a(x)$ gives $x=O(y^2)$. We have seen that
$j^*(\om^+)=2dy/(q^+_a)'(x)$, so $j^*(\om^+)=(1+O(y^2))dy$ as claimed.
The same method works for $E^-(a)$.
\begin{checks}
projective/ellquot_check.mpl: check_translations()
\end{checks}
\end{proof}
\begin{remark}\label{rem-translate}
For any $i$, the form $(\tau^+_i)^*(\om^+)$ must have the form
$u\,\om^+$ for some function $u$ which is holomorphic everywhere on
$E^+(a)$, and so is constant. As $(\tau^+_i)^2=1$ we see that
$u^2=1$, so $u=\pm 1$. In the case $i=1$, we saw in the proof of the
above proposition that $u=1$. By the same method one can check that
$u=1$ for $i=2,3$ as well. This implies that all the maps $\tau^+_i$
are actually translations with respect to the standard group
structure on $E^+(a)$. More specifically, the zero element is
$o=v^+_0=v^+_1$, and one can check that
\begin{align*}
\tau^+_1(o) &= v^+_3 = v^+_5 \\
\tau^+_2(o) &= v^+_{11} = v^+_{13} \\
\tau^+_3(o) &= v^+_{10} = v^+_{12}.
\end{align*}
Thus, we have $\tau^+_1(p)=p+v^+_3$ and so on.
The situation for $E^-(a)$ is similar, but with
\begin{align*}
\tau^-_1(o) &= v^-_7 = v^-_9 \\
\tau^-_2(o) &= v^-_{11} = v^-_{13} \\
\tau^-_3(o) &= v^-_{10} = v^-_{12}.
\end{align*}
\begin{checks}
projective/ellquot_check.mpl: check_translations()
\end{checks}
\end{remark}
\begin{proposition}\label{prop-Ep-Em}
There are (unbranched) double covering maps
\[ E^+(a)\xra{\pi^+} E^-(a) \xra{\pi^-} E^+(a) \]
given generically by
\begin{align*}
\pi^+(j(y,x)) &=
j\left(
\frac{\rt y((1-x)^2+b_-^2x^2)}{((1-x)^2-b_-^2x^2)^2},\;
\frac{2x(x-1)}{((1-x)^2-b_-^2x^2)}
\right) \\
\pi^-(j(y,x)) &=
j\left(
\frac{\rt y((1-x)^2-b_+^2x^2)}{((1-x)^2+b_+^2x^2)^2},\;
\frac{2x(x-1)}{((1-x)^2+b_+^2x^2)}
\right).
\end{align*}
(More precisely, the above formulae are valid for all points $(y,x)$
where the denominators are nonzero.) These are in fact surjective
group homomorphisms, with
\begin{align*}
\ker(\pi^+) &= \{j(0,0),j(0,1)\} = \{v_0^+,v_3^+\} \\
\ker(\pi^-) &= \{j(0,0),j(0,1)\} = \{v_0^-,v_7^-\}.
\end{align*}
\end{proposition}
These maps are \mcode+Ep_to_Em+ and \mcode+Em_to_Ep+ (or
\mcode+Ep_0_to_Em_0+ and \mcode+Em_0_to_Ep_0+).
\begin{proof}
We define $\tpi^+,\tpi^-\:\C^4\to\C^4$ by
\begin{align*}
\tpi^+(z)_1 &= \rt z_1(z_2-2z_3+b_+^2z_4) \\
\tpi^+(z)_2 &= (2-b_+^2)^2z_4^2 +(z_2-2z_3+2(2-b_+^2)z_4)(z_2-2z_3) \\
\tpi^+(z)_3 &= 2(z_2-b_+^2z_4)(z_4-z_3)+4(z_4^2-2z_3z_4+z_2z_4) \\
\tpi^+(z)_4 &= 4(z_4-2z_3+z_2)z_4 \\
\tpi^-(z)_1 &= \rt z_1(z_2-2z_3-b_-^2z_4) \\
\tpi^-(z)_2 &= (2+b_-^2)^2z_4^2 +(z_2-2z_3+2(2+b_-^2)z_4)(z_2-2z_3) \\
\tpi^-(z)_3 &= 2(z_2+b_-^2z_4)(z_4-z_3)+4(z_4^2-2z_3z_4+z_2z_4) \\
\tpi^-(z)_4 &= 4(z_4-2z_3+z_2)z_4.
\end{align*}
Recall that
\begin{align*}
E^+(a) &= \{[z]\st \rho_0(z)=\rho_1(z)=0\} \\
E^-(a) &= \{[z]\st \rho_0(z)=\rho_2(z)=0\}
\end{align*}
where
\begin{align*}
\rho_0(z) &= z_2z_4-z_3^2 \\
\rho_1(z) &= z_1^2 - 2(z_4-z_3)(b_+^2z_4-z_2) \\
\rho_2(z) &= z_1^2 - 2(z_4-z_3)(b_-^2z_4+z_2).
\end{align*}
We claim that there are no nonzero points in $\C^4$ where
$\rho_0(z)=\rho_1(z)=0$ and $\tpi^+(z)=0$. This can be proved by
using Gr\"obner basis methods to prove that the ideals
\[ I_k = (z_k-1,\rho_0(z),\rho_1(z),\tpi^+(z)_1,\dotsc,\tpi^+(z)_4) \]
all contain $1$, or just by solving the equations in a more
elementary way. One can also use Gr\"obner bases to check that
$\rho_i(\tpi^+(z))\in(\rho_0(z),\rho_1(z))$ for $i\in\{0,2\}$. It
follows that the rule $\pi^+([z])=[\tpi^+(z)]$ gives a well-defined
morphism $\pi^+\:E^+(a)\to E^-(a)$. Straightforward algebra shows
that $\pi^+(j(y,x))$ is given by the stated formula whenever
$(1-x)^2-b_-^2x^2\neq 0$. In particular, we have
$\pi^+(j(0,0))=j(0,0)$, so $\pi^+$ preserves basepoints. It is a
standard fact that any basepoint preserving morphism of elliptic
curves is a group homomorphism, and in this context, any non-constant
group homomorphism is a covering map, so we just need to identify
$\ker(\pi^+)$. If $(1-x)^2-b_-^2x^2\neq 0$ then the stated formula
for $\pi^+(j(x,y))$ is valid, and we see that $\pi^+(j(x,y))=j(0,0)$
iff $(x,y)\in\{(0,0),(0,1)\}$. The exceptional points where
$(1-x)^2-b_-^2x^2=0$ are as follows:
\begin{align*}
w_1 &= j(\pp 2b_-/(1+b_-)^2,\;1/(1+b_-)) &
w_2 &= j(\pp 2b_-/(1-b_-)^2,\;1/(1-b_-)) \\
w_3 &= j( - 2b_-/(1+b_-)^2,\;1/(1+b_-)) &
w_4 &= j( - 2b_-/(1-b_-)^2,\;1/(1-b_-)).
\end{align*}
These satisfy $\pi^+(w_1)=\pi^+(w_2)=v_2^+\neq v_0^+$ and
$\pi^+(w_3)=\pi^+(w_4)=v_4^+\neq v_0^+$, so they do not contribute to
the kernel. This just leaves the points in $E^+(a)$ that do not lie
in the image of $j$, which are $v_6^+$ and $v_7^+$; direct
calculation shows again that these are not in the kernel of $\pi^+$.
This completes the proof that $\ker(\pi^+)=\{v_0^+,v_3^+\}$. (As an
alternative, we could reach the same conclusion by calculating
Gr\"obner bases for each of the ideals
\[ (z_i-1,\rho_0(z),\rho_1(z),\tpi^+(z)_1,\tpi^+(z)_3,\tpi^+(z)_4),
\]
or by showing that the degree of the relevant field extension is two
and appealing to some more abstract arguments.)
The proof for $\pi^-$ is essentially the same.
\begin{checks}
projective/ellquot_check.mpl: check_isogenies()
\end{checks}
\end{proof}
\begin{definition}\label{defn-PJ}
We note that Proposition~\ref{prop-Ep-Em} implies that the elements
$v_3^+$ and $v_7^-$ have order two, so $(v_7^-,v_3^+)$ generates
a subgroup $Z$ of order two in $(E^-(a)\tm E^+(a))$. We define
\[ PJ(a) = (E^-(a)\tm E^+(a))/Z. \]
We also define $\tht^+\:PJ(a)\to E^+(a)$ and
$\tht^-\:PJ(a)\to E^-(a)$ by
\begin{align*}
\tht^+((w^-,w^+)+Z) &= \pi^-(w^-) \\
\tht^-((w^-,w^+)+Z) &= \pi^+(w^+).
\end{align*}
\end{definition}
\begin{proposition}\label{prop-P-to-PJ}
There is a unique morphism $\phi\:PX(a)\to PJ(a)$ such that
$\phi(v_0)=o$ and $\tht^+\phi=\phi^+$ and $\tht^-\phi=\phi^-$.
\end{proposition}
\begin{proof}
Suppose we have $(w,z)\in PX_0(a)$ and $u\in\C$ with $u^2=z$. When the
relevant denominators are nonzero, we then put
\begin{align*}
x_- &= \;\frac{w/u - (1-z)^2}{2(b_-^2z-\;(1-z)^2)} \\
x_+ &= i\frac{w/u + (i+z)^2}{2(b_+^2z+ i(i+z)^2)} \\
y_- &= \frac{1}{\rt}\; \frac{u(1-z)(2b_+^2(1-z)^2-b_-^2(w/u+1+z^2))}{(b_-^2z-\;(1-z)^2)^2} \\
y_+ &= \frac{1+i}{2}\; \frac{u(i+z)(2b_-^2(i+z)^2+b_+^2(w/u+1-z^2))}{(b_+^2z+i (i+z)^2)^2},
\end{align*}
and $\phi_0(w,z,u)=(y_-,x_-,y_+,x_+)$. One can check that
\begin{itemize}
\item[(a)] $y_-^2=q^-_a(x_-)$ and $y_+^2=q^+_a(x_+)$, so
$\phi_0(w,z,u)\in E^-_0(a)\tm E^+_0(a)$.
\item[(b)] $\phi_0(w,z,-u)=(\tau_1^-\tm\tau_1^+)\phi_0(w,z,u)$.
\item[(c)] $\pi^-(j(y^-,x^-))=\phi^+_0(w,z)$ and
$\pi^+(j(y^+,x^+))=\phi^-_0(w,z)$.
\end{itemize}
It follows that the image of $\phi_0(w,z,u)$ in $PJ(a)$ is
independent of the choice of $u$, so we can call it $\phi_0(w,z)$.
This defines a rational map from $EX_0(a)$ to $PJ(a)$, but $EX_0(a)$
is dense subset of the smooth curve $EX(a)$, and $PJ(a)$ is complete,
so this extends uniquely to give a morphism $\phi\:PX(a)\to PJ(a)$.
Point~(c) above shows that $\tht^\pm\phi=\phi^\pm$.
\begin{checks}
projective/ellquot_check.mpl: check_isogenies()
\end{checks}
\end{proof}
The map $\phi$ is represented in Maple as \mcode+P_0_to_J_0+.
\begin{corollary}\label{cor-jacobian}
$PJ(a)$ can be regarded as the Jacobian variety of $PX(a)$.
\end{corollary}
\begin{proof}
The Jacobian variety $J$ can be characterised by the fact that it is
an abelian variety equipped with a map $\dl\:PX(a)\to J$ of varieties
such that the induced map $H_1PX(a)\to H_1J$ is an isomorphism. We
saw in Definition~\ref{defn-JX} that the map
\[ (\phi^+,\phi^-)_*\: H_1PX(a) \to H_1(E^+(a)\tm E^-(a)) \]
is injective, and has image of index two. As the map
\[ \tht=(\tht^+,\tht^-)\:PJ(a)\to E^+(a)\tm E^-(a) \]
is a connected double covering, it also gives an index two subgroup
of $\pi_1=H_1$. As $(\phi^+,\phi^-)=\tht\phi$, we deduce that
$\phi_*\:H_1PX(a)\to H_1PJ(a)$ is an isomorphism, as required.
\end{proof}
It is standard that any elliptic curve has an analytic parametrisation
via the Weierstrass $\wp$-function. Details for the present case are
as follows.
\begin{definition}\label{defn-weierstrass-xi}
We put
\begin{align*}
g_2^+ &= 4(\tfrac{1}{3}+b_+^2) &
g_2^- &= 4(\tfrac{1}{3}-b_-^2) \\
g_3^+ &= \tfrac{8}{3}(\tfrac{1}{9}-b_+^2) &
g_3^- &= \tfrac{8}{3}(\tfrac{1}{9}+b_-^2) \\
p_0^+(z) &= \wp(z/\rt;g_2^+,g_3^+) &
p_0^-(z) &= \wp(iz/\rt;g_2^-,g_3^-) \\
p_1^+(z) &= \wp'(z/\rt;g_2^+,g_3^+) &
p_1^-(z) &= \wp'(iz/\rt;g_2^-,g_3^-)
\end{align*}
and then
\begin{align*}
\xi^+(z) &= j\left(-\frac{p_1^+(z)}{\rt(p_0^+(z)+1/3)^2},
\frac{1}{(p_0^+(z)+1/3)}\right)\in\C P^3 \\
\xi^-(z) &= j\left(i\frac{p_1^-(z)}{\rt(p_0^-(z)+1/3)^2},
\frac{1}{(p_0^-(z)+1/3)}\right)\in\C P^3.
\end{align*}
\end{definition}
In Maple, the parameters $g_i^\pm$ are
\mcode+Wg2p+, \mcode+Wg3p+, \mcode+Wg2m+ and \mcode+Wg3m+. The maps
$\xi^\pm$ are \mcode+C_to_Ep_0+ and \mcode+C_to_Em_0+.
\begin{remark}\label{rem-xi-differential}
It is a standard fact that $\wp(z)=z^{-2}+O(z^2)$, so
$\wp'(z)=-2z^{-3}+O(z)$. Using this, we find that
$j^{-1}\xi^+(z)=(z,0)+O(z^2)$ and $j^{-1}\xi^-(z)=(z,0)+O(z^2)$.
\end{remark}
\begin{proposition}\label{prop-xi-functions}
There are lattices $\Lm^+,\Lm^-\subset\C$ such that $\xi^{\pm}$
induces an isomorphism $\C/\Lm^{\pm}\to E^{\pm}(a)$. Moreover, the
forms $\om^\pm$ on $E^{\pm}(a)$ (from
Proposition~\ref{prop-elliptic-forms}) satisfy
$(\xi^\pm)^*(\om^\pm)=dz$.
\end{proposition}
\begin{proof}
Given any $g_2,g_3$ we can define $f(x)=4x^3-g_2x-g_3$ and
$F_0=\{(y,x)\in\C^2\st y^2=f(x)\}$, and we can then define $F$ to be
the normalisation of $F_0$. It is standard that
\[ \wp'(z;g_2,g_3)^2 = f(\wp(z;g_2,g_3)), \]
so we have a meromorphic function $z\mapsto(\wp'(z),\wp(z))$ from
$\C$ to $F$. It is also standard that this induces an isomorphism
$\C/\Lm\to F$, for a suitable lattice $\Lm\subset\C$. The first
claim follows from this by a change of coordinates.
Next, we have $(\xi^\pm)^*(\om^\pm)=u(z)\,dz$ for some function
$u(z)$ which is holomorphic, nowhere zero, and periodic with respect
to $\Lm^\pm$. This forces $u(z)$ to be constant.
Remark~\ref{rem-xi-differential}, together with the last part of
Proposition~\ref{prop-elliptic-forms}, shows that $u=1$.
\begin{checks}
projective/ellquot_check.mpl: check_weierstrass()
\end{checks}
\end{proof}
We now want to understand the lattices $\Lm^+$ and $\Lm^-$ in more
detail, which is essentially the same as calculating the periods
$p_{jk}(a)=\int_{c_j}\om_k$ as in Definition~\ref{defn-period-p}.
\begin{definition}
We define the complete elliptic integral $K(k)$ (for $0<k<1$) by
\[ K(k) = \int_0^1 \frac{dt}{\sqrt{1-t^2}\sqrt{1-k^2t^2}}. \]
In this range the square roots are real and positive and there is no
need for branch cuts.
\end{definition}
Our definition is the same as Maple's \mcode+EllipticK(k)+, but
slightly different conventions are used in some other sources.
\begin{definition}\label{defn-period-rs}
We put
\begin{align*}
m_+ &= \frac{1 + a}{\sqrt{2(1+a^2)}} &
m_- &= \frac{1 - a}{\sqrt{2(1+a^2)}} \\
\al_+ &= 2b_+^{-1/2} K(m_+) &
\al_- &= 2b_+^{-1/2} K(m_-).
\end{align*}
(Note here that the first factor in $\al_-$ involves $b_+$, not $b_-$.)
\end{definition}
In Maple, these are \mcode+mp_period+, \mcode+mm_period+,
\mcode+ap_period+ and \mcode+am_period+. These are defined in the file
\fname+projective/picard_fuchs.mpl+.
\begin{remark}\label{rem-legendre}
A mixture of theoretical arguments and numerical calculations makes
it clear that we also have
\begin{align*}
\al_+ &= \pi P_{-1/4}(A/2) + 2Q_{-1/4}(A/2) \\
\al_- &= \pi P_{-1/4}(A/2)
\end{align*}
where $A=a^{-2}+a^2$ as before, and $P$ and $Q$ are Legendre
functions. We do not have a complete proof, but we will mention some
ingredients. Consider the differential operator
\[ \CL =
198 \frac{\partial^2}{\partial A^2} +
192 A \frac{\partial^3}{\partial A^3} +
(32A^2 - 128) \frac{\partial^4}{\partial A^4}.
\]
One can check by direct calculation that
\[ \CL((z^5-Az^3+z)^{-1/2}) \,dz =
\frac{d}{dz}\left(
\frac{33z^8-3Az^{10}-27z^{12}}{(z^5-Az^3+z)^{7/2}}
\right) \,dz.
\]
The terms on the left can be interpreted as meromorphic differential
forms on $PX(a)$, all of whose residues are zero. This means that
their integrals round a loop depend only on the homology class of the
loop. On the other hand, the integral of the right hand side around
any loop is zero. Using this, we see that the periods (when
expressed as a function of $A$) are annihilated by $\CL$. This is
the Picard-Fuchs equation for the family $\{PX(a)\st a\in(0,1)\}$.
Maple asserts that the annihilator of $\CL$ is spanned by $1$, $A$,
$P_{-1/4}(A/2)$ and $Q_{-1/4}(A/2)$. This could presumably be
checked using hypergeometric series, but we have not attempted that.
The specific coefficients in the stated expressions for $\al_+$ and
$\al_-$ were obtained by graphical and numerical experimentation.
\begin{checks}
projective/picard_fuchs_check.mpl: check_picard-fuchs()
\end{checks}
\end{remark}
\begin{proposition}\label{prop-periods}
The periods are
\begin{align*}
p_{0,0} &= 0 & p_{0,1} &= 0 \\
p_{1,0} &= (i+1)\al_- & p_{1,1} &= (i-1)\al_- \\
p_{2,0} &= (i-1)\al_- & p_{2,1} &= (i+1)\al_- \\
p_{3,0} &= i\al_- & p_{3,1} &= i\al_- \\
p_{4,0} &= -\al_- & p_{4,1} &= \al_- \\
p_{5,0} &= (\al_++\al_-)/2 & p_{5,1} &= (\al_+-\al_-)/2 \\
p_{6,0} &= i(\al_++\al_-)/2 & p_{6,1} &= -i(\al_+-\al_-)/2 \\
p_{7,0} &= (\al_+-\al_-)/2 & p_{7,1} &= (\al_++\al_-)/2 \\
p_{8,0} &= i(\al_+-\al_-)/2 & p_{8,1} &= -i(\al_++\al_-)/2.
\end{align*}
\end{proposition}
In Maple, $p_{ij}$ is \mcode+p_period[i,j]+.
The proof will be given after some preparatory lemmas.
\begin{lemma}\label{lem-int-q}
\begin{align*}
\int_0^{1/b_+} \frac{dx}{\sqrt{ q_a^+(x)}} &= \al_+/2 \\
\int_{-1/b_+}^0 \frac{dx}{\sqrt{-q_a^+(x)}} &= \al_-/2.
\end{align*}
\end{lemma}
\begin{proof}
For the first integral we use the substitution
$x=(1-t^2)/(b_+-t^2)$, and for the second we use the
substitution $x=b_+^{-1}(1-2/t^2)^{-1}$. In both cases we get an
extra factor of $-1$ in the integrand, which is cancelled by the fact
that the limits are reversed, because $x$ is a decreasing function of
$t$.
\begin{checks}
projective/picard_fuchs_check.mpl: check_period_integrals()
\end{checks}
\end{proof}
\begin{lemma}\label{lem-elliptic-periods}
\begin{align*}
\int_{\phi^+\circ c_5}\frac{dx}{y} &= \al_+ \\
\int_{\phi^+\circ c_6}\frac{dx}{y} &= i\al_-.
\end{align*}
\end{lemma}
\begin{proof}
From the definitions we have $j^{-1}(c_5(t))=(w(t),z(t))$, where
$z(t)=a\sin(t/2)^2\in [0,a]$ and $w(t)$ is a positive multiple of
$\sin(t)$. It follows that $j^{-1}\phi^+(c_5(t))=(y(t),x(t))$, where
$x(t)=2/(z(t)+z(t)^{-1})$, and $y(t)$ is again a positive multiple of
$\sin(t)$. On the other hand, we have $(y(t),x(t))\in E^+_0(a)$ so
$y(t)^2=q_a^+(x(t))$, so for $0\leq t\leq \pi$ we must have
$y(t)=\sqrt{q^+_a(x(t))}$, and for $\pi\leq t\leq 2\pi$ we must have
$y(t)=-\sqrt{q^+_a(x(t))}$. Moreover, as $t$ runs from $0$ to $\pi$
we see that $z(t)$ increases from $0$ to $a$, and so $x(t)$ increases
from $0$ to $1/b_+$. On the other hand, as $t$ increases from $\pi$
to $2\pi$ we see that $x(t)$ decreases from $1/b_+$ to $0$. It
follows that
\[ \int_{\phi^+\circ c_5}\frac{dx}{y} =
\int_{x=0}^{1/b_+}\frac{dx}{\sqrt{q_a^+(x)}} +
\int_{x=1/b_+}^{0}\frac{dx}{-\sqrt{q_a^+(x)}} =
2\int_{0}^{1/b_+}\frac{dx}{\sqrt{q_a^+(x)}} =
\al_+.
\]
The second integral is similar. We have
$j^{-1}(c_6(t))=(i\,w(t),-z(t))$, where $z(t)=a\sin(t/2)^2\in [0,a]$ and
$w(t)$ is a positive multiple of $\sin(t)$. It follows that
$j^{-1}\phi^+(c_5(t))=(y(t),x(t))$, where $x(t)=-2/(z(t)+z(t)^{-1})$,
and $y(t)$ is again a positive multiple of $i\sin(t)$. In this range
$q_a^+(x(t))\leq 0$ so it is natural to consider
$\sqrt{-q_a^+(x(t))}$. As $(y(t),x(t))\in E^+_0(a)$, we must have
$(y(t)/i)^2=-q_a^+(x(t))$, so for $0\leq t\leq \pi$ we must have
$y(t)=i\sqrt{-q^+_a(x(t))}$, and for $\pi\leq t\leq 2\pi$ we must
have $y(t)=-i\sqrt{-q^+_a(x(t))}$. Moreover, as $t$ runs from $0$ to
$\pi$ and then to $2\pi$, we see that $x(t)$ decreases from $0$ to
$-1/b_+$, and then increases back to $0$ again. It follows that
\[ \int_{\phi^+\circ c_6}\frac{dx}{y} =
\int_{0}^{-1/b_+}i\frac{dx}{\sqrt{-q_a^+(x)}} +
\int_{-1/b_+}^{0}-i\frac{dx}{\sqrt{-q_a^+(x)}} =
2i\int_{0}^{1/b_+}\frac{dx}{\sqrt{-q_a^+(x)}} =
i\al_-.
\]
\begin{checks}
projective/picard_fuchs_check.mpl: check_period_integrals()
\end{checks}
\end{proof}
\begin{lemma}\label{lem-form-pullback}
For the forms $\om^{\pm}$ on $E^{\pm}(a)$ and $\om_i$ on $PX(a)$ (as in
Propositions~\ref{prop-holomorphic-forms}
and~\ref{prop-elliptic-forms}), we have
\begin{align*}
(\phi^+)^*(\om^+) &= \om_0+\om_1 \\
(\phi^-)^*(\om^+) &= \frac{1+i}{\rt}(i\om_0-\om_1).
\end{align*}
\end{lemma}
\begin{proof}
On the affine pieces $PX_0(a)$ and $E^+_0(a)$ we have $\om^+=dx/y$
and
\[ \phi^+(w,z) = \left(\frac{2w(1-z)}{(1+z^2)^2},
\frac{2z}{1+z^2}\right),
\]
so
\begin{align*}
(\phi^+)^*(\om^+) &=
\frac{(1+z^2)^2}{2w(1-z)}d\left(\frac{2z}{1+z^2}\right) \\
&= \frac{(1+z^2)^2}{2w(1-z)} \; \frac{2-2z^2}{(1+z^2)^2}dz
= \frac{1+z}{w}dz.
\end{align*}
On the other hand, we have $\om_0=dz/w$ and $\om_1=z\,dz/w$, so
$(\phi^+)^*(\om^+)=\om_0+\om_1$. The argument for
$(\phi^-)^*(\om^-)$ is similar.
\begin{checks}
projective/picard_fuchs_check.mpl: check_period_integrals()
\end{checks}
\end{proof}
\begin{corollary}\label{cor-basic-periods}
We have $p_{5,0}=(\al_++\al_-)/2$ and $p_{5,1}=(\al_+-\al_-)/2$.
\end{corollary}
\begin{proof}
Lemma~\ref{lem-elliptic-periods} says that
\[ \int_{c_5}(\phi^+)^*\left(\frac{dx}{y}\right) = \al_+. \]
On the left hand side, Lemma~\ref{lem-form-pullback} says that the
integrand is $\om_0+\om_1$, so the integral is $p_{5,0}+p_{5,1}$.
Similarly, the second equation in Lemma~\ref{lem-elliptic-periods}
becomes $p_{6,0}+p_{6,1}=i\al_-$. On the other hand, we have seen that
$\lm\circ c_5=c_6$ and $\lm^*\om_0=i\om_0$ and $\lm^*\om_1=-i\om_1$;
it follows that $p_{6,0}+p_{6,1}=ip_{5,0}-ip_{5,1}$. Linear algebra
now gives $p_{5,0}=(\al_++\al_-)/2$ and $p_{5,1}=(\al_+-\al_-)/2$ as
claimed.
\end{proof}
\begin{proof}[Proof of Proposition~\ref{prop-periods}]
It is standard that there is a well-defined pairing
$H_1(PX(a))\ot\Om^1(PX(a))\to\C$ such that $([\gm],\al)=\int_\gm\al$
for all closed curves $\gm$ in $PX(a)$ and all $\om\in\Om^1(PX(a))$.
In particular, we have $p_{i,j}=([c_i],\om_j)$. Thus, relations in
$H_1(PX(a))$ will give relations between periods. Note also that for
$g\in D_8$ we have
$(g_*[c_i],\om_j)=([g\circ c_i],\om_j)=([c_i],g^*\om_j)$.
Proposition~\ref{prop-holomorphic-forms} gives the action of $D_8$ on
$\om_0$ and $\om_1$, whereas Definition~\ref{defn-curve-system}(c)
and Proposition~\ref{prop-P-curves} give the action on the curves
$c_i$. In particular, we have
\[ [c_6] = \lm_*[c_5] \hspace{4em}
[c_7] = \mu_*[c_5] \hspace{4em}
[c_8] = (\mu\lm)_*[c_5],
\]
whereas
\[ \lm^*\om_0 = i\om_0 \hspace{4em}
\lm^*\om_1 = -i\om_1 \hspace{4em}
\mu^*\om_0 = \om_1 \hspace{4em}
\mu^*\om_1 = \om_0.
\]
It therefore follows from Corollary~\ref{cor-basic-periods} that
\begin{align*}
p_{5,0} &= (\al_++\al_-)/2 & p_{5,1} &= (\al_+-\al_-)/2 \\
p_{6,0} &= i(\al_++\al_-)/2 & p_{6,1} &= -i(\al_+-\al_-)/2 \\
p_{7,0} &= (\al_+-\al_-)/2 & p_{7,1} &= (\al_++\al_-)/2 \\
p_{8,0} &= i(\al_+-\al_-)/2 & p_{8,1} &= -i(\al_++\al_-)/2.
\end{align*}
Next, we see from Proposition~\ref{prop-homology} that
\begin{align*}
[c_0] &= 0 \\
[c_1] &= [c_5]+[c_6]-[c_7]-[c_8] \\
[c_2] &= -[c_5]+[c_6]+[c_7]-[c_8] \\
[c_3] &= [c_6] -[c_8] \\
[c_4] &= -[c_5] +[c_7].
\end{align*}
We can now apply the maps $(-,\om_0)$ and $(-,\om_1)$ to deduce that
\begin{align*}
p_{0,0} &= 0 & p_{0,1} &= 0 \\
p_{1,0} &= (i+1)\al_- & p_{1,1} &= (i-1)\al_- \\
p_{2,0} &= (i-1)\al_- & p_{2,1} &= (i+1)\al_- \\
p_{3,0} &= i\al_- & p_{3,1} &= i\al_- \\
p_{4,0} &= -\al_- & p_{4,1} &= \al_-.
\end{align*}
\begin{checks}
projective/picard_fuchs_check.mpl: check_periods()
\end{checks}
\end{proof}
\begin{proposition}
The lattices $\Lm^+$ and $\Lm^-$ (from
Proposition~\ref{prop-xi-functions}) are
\begin{align*}
\Lm^+ &= \{n\al_++m\al_-i \st n,m\in\Z\} \\
\Lm^- &= \{(n\al_++m\al_-i)/\sqrt{2}\st n,m\in\Z,\; n=m\pmod{2}\}.
\end{align*}
\end{proposition}
\begin{proof}
We can identify $\Lm^+$ with
$\{\int_\gm dz\st\gm\in\pi_1(\C/\Lm^+)\}$. Now $\xi^+$ induces an
isomorphism $\C/\Lm^+\to E^+(a)$, under which $dz$ corresponds to
$\om^+$ (by the last part of Proposition~\ref{prop-xi-functions}).
This means that $\Lm^+=\{\int_\gm\om^+\st\gm\in\pi_1(E^+(a))\}$. On
the other hand, we know from Proposition~\ref{prop-elliptic-homology}
that the group $\pi_1(E^+(a))=H_1(E^+(a))$ is a quotient of
$H_1(PX(a))$, so
\[ \Lm^+=\{\int_\gm(\phi^+)^*(\om^+)\st\gm\in\pi_1(PX(a))\}. \]
Using Lemma~\ref{lem-form-pullback} we now see that $\Lm^+$ is
spanned by the numbers $p_{k0}+p_{k1}$, and by inspecting
Proposition~\ref{prop-periods} we conclude that
$\Lm^+=\{n\al_++m\al_-i\st n,m\in\Z\}$ as claimed.
In the same way, using the relation
$(\phi^-)^*(\om^-)=\frac{1+i}{\rt}(i\om_0-\om_1)$ we see that $\Lm^-$
is generated by the numbers $\frac{1+i}{\rt}(ip_{k0}-p_{k1})$. These
numbers can again be read off from Proposition~\ref{prop-periods},
giving
\[ \Lm^- = \{(n\al_++m\al_-i)/\sqrt{2}\st n,m\in\Z,\; n=m\pmod{2}\}
\]
as claimed.
\begin{checks}
projective/ellquot_check.mpl: check_weierstrass()
\end{checks}
\end{proof}
The closed curves $c_k(t)\in PX(a)$ can be mapped to $E^+(a)$ using
$\phi^+$ and then lifted via $\xi^+$ to give curves in $\C$ which are
usually not closed. There are no closed formulae for these curves,
but the functions \mcode+c_TEp_approx[k]+ and \mcode+c_TEm_approx[k]+
give good approximations. The set
\[ \{x+iy\st |x|\leq\al_+/2,|y|\leq\al_-/2\} \]
is a fundamental domain for the action of $\Lm^+$ on $\C$, and the
parts of the lifted curves lying in that domain can be illustrated as
follows:
\begin{center}
\begin{tikzpicture}[scale=4]
\draw[magenta] (0.859,-0.785) -- (0.859,0.785);
\draw[magenta] (-0.859,-0.785) -- (-0.859,0.785);
\draw[blue] (-0.859,0.000) -- (0.859,0.000);
\draw[blue] (0.000,-0.785) -- (0.000,0.785);
\draw[magenta] (-0.466,0.785) -- (0.466,0.785);
\draw[magenta] (-0.466,-0.785) -- (0.466,-0.785);
\draw[cyan] (-0.859,0.785) -- (-0.466,0.785);
\draw[cyan] (0.859,0.785) -- (0.466,0.785);
\draw[cyan] (-0.859,-0.785) -- (-0.466,-0.785);
\draw[cyan] (0.859,-0.785) -- (0.466,-0.785);
\draw[smooth,green] (-0.581,-0.785) -- (-0.576,-0.736) -- (-0.560,-0.679) -- (-0.528,-0.608) -- (-0.473,-0.520) -- (-0.390,-0.412) -- (-0.280,-0.286) -- (-0.146,-0.147) -- (0.000,0.000) -- (0.146,0.147) -- (0.280,0.286) -- (0.390,0.412) -- (0.473,0.520) -- (0.528,0.608) -- (0.560,0.679) -- (0.576,0.736) -- (0.581,0.785) ;
\draw[smooth,green] (0.581,-0.785) -- (0.576,-0.736) -- (0.560,-0.679) -- (0.528,-0.608) -- (0.473,-0.520) -- (0.390,-0.412) -- (0.280,-0.286) -- (0.146,-0.147) -- (0.000,0.000) -- (-0.146,0.147) -- (-0.280,0.286) -- (-0.390,0.412) -- (-0.473,0.520) -- (-0.528,0.608) -- (-0.560,0.679) -- (-0.576,0.736) -- (-0.581,0.785) ;
\fill (-0.859, 0.785) circle(0.011);
\fill (-0.581, 0.785) circle(0.011);
\fill (-0.466, 0.785) circle(0.011);
\fill ( 0.000, 0.785) circle(0.011);
\fill ( 0.466, 0.785) circle(0.011);
\fill ( 0.581, 0.785) circle(0.011);
\fill ( 0.859, 0.785) circle(0.011);
\fill (-0.859,-0.785) circle(0.011);
\fill (-0.581,-0.785) circle(0.011);
\fill (-0.466,-0.785) circle(0.011);
\fill ( 0.000,-0.785) circle(0.011);
\fill ( 0.466,-0.785) circle(0.011);
\fill ( 0.581,-0.785) circle(0.011);
\fill ( 0.859,-0.785) circle(0.011);
\fill (-0.859, 0.000) circle(0.011);
\fill ( 0.000, 0.000) circle(0.011);
\fill ( 0.859, 0.000) circle(0.011);
\end{tikzpicture}
\end{center}
This is combinatorially equivalent to the net for $X/\ip{\mu}$ which
we exhibited in Section~\ref{sec-fundamental}. We have used the value
$a=0.1$, which is roughly right for the embedded surface $EX^*$. The
above domain is not square; the ratio
$\text{width}/\text{height}=\al_+/\al_-$ is approximately $1.1$.
The situation for $E^-(a)$ is a little more complicated. The picture
below shows the domain
\[ \{x+iy\st |x|\leq\al_+/\rt,|y|\leq\al_-/\rt\}, \]
which covers $\C/\Lm^-$ twice. The diagonal blue curves, which
represent $c_5$ and $c_6$, are close to being straight, but they are
not exactly straight. The dashed orange lines enclose a fundamental
domain for $\Lm^-$. This is a rhombus, but the angles are not
$\pi/2$. It is combinatorially equivalent to the net for
$X/\ip{\lm\mu}$ which we exhibited in Section~\ref{sec-fundamental},
but is rotated through $\pi/4$ as well as being slightly
distorted.
\begin{center}
\begin{tikzpicture}[scale=3]
\draw[cyan] (-0.394, 1.111) -- ( 0.394, 1.111);
\draw[cyan] (-0.394,-1.111) -- ( 0.394,-1.111);
\draw[cyan] (-1.215, 0.000) -- (-0.822, 0.000);
\draw[cyan] ( 1.215, 0.000) -- ( 0.822, 0.000);
\draw[green] (-0.822, 0.000) -- ( 0.822, 0.000);
\draw[green] ( 0.394, 1.111) -- ( 1.215, 1.111);
\draw[green] (-0.394, 1.111) -- (-1.215, 1.111);
\draw[green] ( 0.394,-1.111) -- ( 1.215,-1.111);
\draw[green] (-0.394,-1.111) -- (-1.215,-1.111);
\draw[green] (-1.215,-1.111) -- (-1.215, 1.111);
\draw[green] ( 0.000,-1.111) -- ( 0.000, 1.111);
\draw[green] ( 1.215,-1.111) -- ( 1.215, 1.111);
\draw[smooth,magenta] (0.278,-1.111) -- (0.292,-1.024) -- (0.344,-0.907) -- (0.453,-0.746) -- (0.608,-0.555) -- (0.762,-0.364) -- (0.871,-0.204) -- (0.924,-0.087) -- (0.937,0.000) -- (0.924,0.087) -- (0.871,0.204) -- (0.762,0.364) -- (0.608,0.555) -- (0.453,0.746) -- (0.344,0.907) -- (0.292,1.024) -- (0.278,1.111) ;
\draw[smooth,magenta] (-0.278,-1.111) -- (-0.292,-1.024) -- (-0.344,-0.907) -- (-0.453,-0.746) -- (-0.608,-0.555) -- (-0.762,-0.364) -- (-0.871,-0.204) -- (-0.924,-0.087) -- (-0.937,0.000) -- (-0.924,0.087) -- (-0.871,0.204) -- (-0.762,0.364) -- (-0.608,0.555) -- (-0.453,0.746) -- (-0.344,0.907) -- (-0.292,1.024) -- (-0.278,1.111) ;
\draw[smooth,blue] (-1.215,1.111) -- (-1.129,1.024) -- (-1.044,0.941) -- (-0.964,0.864) -- (-0.887,0.794) -- (-0.814,0.729) -- (-0.744,0.668) -- (-0.675,0.611) -- (-0.608,0.555) -- (-0.540,0.500) -- (-0.471,0.442) -- (-0.401,0.382) -- (-0.328,0.317) -- (-0.251,0.246) -- (-0.171,0.169) -- (-0.087,0.086) -- (0.000,0.000) -- (0.087,-0.086) -- (0.171,-0.169) -- (0.251,-0.246) -- (0.328,-0.317) -- (0.401,-0.382) -- (0.471,-0.442) -- (0.540,-0.500) -- (0.608,-0.555) -- (0.675,-0.611) -- (0.744,-0.668) -- (0.814,-0.729) -- (0.887,-0.794) -- (0.964,-0.864) -- (1.044,-0.941) -- (1.129,-1.024) -- (1.215,-1.111) ;
\draw[smooth,blue] (-1.215,-1.111) -- (-1.129,-1.024) -- (-1.044,-0.941) -- (-0.964,-0.864) -- (-0.887,-0.794) -- (-0.814,-0.729) -- (-0.744,-0.668) -- (-0.675,-0.611) -- (-0.608,-0.555) -- (-0.540,-0.500) -- (-0.471,-0.442) -- (-0.401,-0.382) -- (-0.328,-0.317) -- (-0.251,-0.246) -- (-0.171,-0.169) -- (-0.087,-0.086) -- (0.000,0.000) -- (0.087,0.086) -- (0.171,0.169) -- (0.251,0.246) -- (0.328,0.317) -- (0.401,0.382) -- (0.471,0.442) -- (0.540,0.500) -- (0.608,0.555) -- (0.675,0.611) -- (0.744,0.668) -- (0.814,0.729) -- (0.887,0.794) -- (0.964,0.864) -- (1.044,0.941) -- (1.129,1.024) -- (1.215,1.111) ;
\fill (-1.215,-1.111) circle(0.014);
\fill (-0.394,-1.111) circle(0.014);
\fill (-0.278,-1.111) circle(0.014);
\fill ( 0.000,-1.111) circle(0.014);
\fill ( 0.278,-1.111) circle(0.014);
\fill ( 0.394,-1.111) circle(0.014);
\fill ( 1.215,-1.111) circle(0.014);
\fill (-1.215, 1.111) circle(0.014);
\fill (-0.394, 1.111) circle(0.014);
\fill (-0.278, 1.111) circle(0.014);
\fill ( 0.000, 1.111) circle(0.014);
\fill ( 0.278, 1.111) circle(0.014);
\fill ( 0.394, 1.111) circle(0.014);
\fill ( 1.215, 1.111) circle(0.014);
\fill (-0.608,-0.555) circle(0.014);
\fill ( 0.608,-0.555) circle(0.014);
\fill (-0.608, 0.555) circle(0.014);
\fill ( 0.608, 0.555) circle(0.014);
\fill (-1.215, 0.000) circle(0.014);
\fill (-0.822, 0.000) circle(0.014);
\fill (-0.937, 0.000) circle(0.014);
\fill ( 0.000, 0.000) circle(0.014);
\fill ( 0.937, 0.000) circle(0.014);
\fill ( 0.822, 0.000) circle(0.014);
\fill ( 1.215, 0.000) circle(0.014);
\draw[orange,dashed] (1.215,0.000) -- (0.000,1.111) -- (-1.215,0.000) -- (0.000,-1.111) -- cycle;
\end{tikzpicture}
\end{center}
\subsection{Some general theory of Riemann surfaces}
\label{sec-riemann}
In the next section, we will give a classification of (pre)cromulent
surfaces. In the present section, we develop some more general theory
of Riemann surfaces, which will feed into that classification. All of
it is essentially standard; we discuss it here in order to have a
convenient reference with a uniform approach.
\subsubsection{Involutions}
Throughout this section, $Z$ will be a compact connected Riemann
surface, with a conformal involution $\al\:Z\to Z$, and an
anticonformal involution $\bt\:Z\to Z$ that commutes with $\al$. We
also assume that $\al$ has only isolated fixed points (which means
that the total number of fixed points is finite). We will write $\Dl$
for the open unit disc in $\C$.
\begin{lemma}\label{lem-invariant-metric}
$Z$ admits a smooth Riemannian metric that is compatible with the
conformal structure and is invariant under the action of $\al$ and
$\bt$.
\end{lemma}
\begin{proof}
Any coordinate patch clearly admits a smooth conformal Riemannian
metric, and one can use a partition of unity to combine such local
metrics to give a local metric, say $\mu_0$. Any smooth automorphism
of $Z$ acts in an evident way on the set of metrics, and that set is
convex, so we can define
\[ \mu = (\mu_0 + \al^*\mu_0 + \bt^*\mu_0 + \al^*\bt^*\mu_0)/4. \]
This is the required invariant metric.
\end{proof}
For the rest of this section we will assume that an invariant metric
has been chosen.
\begin{remark}\label{rem-std-param}
Let $C$ be a closed connected one-dimensional smooth submanifold of
$Z$. Then $C$ is necessarily diffeomorphic to the circle. Now fix a
point $a\in C$, and a unit vector $v\in T_aC$. It is then standard
that there is a unique smooth map $c_1\:\R\to C$ with $c_1(0)=a$ and
$c_1'(0)=v$ and $\|c'_1(t)\|=1$ for all $t$. One can check that
$c_1$ induces a diffeomorphism $\R/\Z d\to C$ for some $d>0$. We put
$c(t)=c_1(td/2\pi)$, so $c$ induces a diffeomorphism
$\R/2\pi\Z\to C$, which we call a \emph{standard parametrisation}
of $C$. It depends on the choice of $a$ and $v$, but if we make
different choices then the new standard parametrisation will be of
the form $t\mapsto c(p+t)$ or $t\mapsto c(p-t)$ for some constant $p$.
\end{remark}
\begin{remark}\label{rem-circle-involution}
Let $c\:\R/2\pi\Z\to C$ be as in the previous remark, let
$\gm\:Z\to Z$ be an involution that preserves the metric, and suppose
that $\gm(C)=C$. Then $\gm(c(t))$ must have the form $c(p+t)$ or
$c(p-t)$ for some constant $p$ (which is well-defined modulo
$2\pi$).
\begin{itemize}
\item[(a)] If $\gm(c(t))=c(p+t)$ then the equation $\gm^2=1$ gives
$2p=0\pmod{2\pi}$, so we can take $p=0$ or $p=\pi$. If $p=0$ then
of course $\gm|_C=1$. If $p=\pi$ then $\gm$ acts freely on $C$, so
$C/\ip{\gm}$ is again a circle.
\item[(b)] If $\gm(c(t))=c(p-t)$ then we can define a new standard
parametrisation by $c^*(t)=c(p/2+t)$, and this satisfies
$\gm(c^*(t))=c^*(-t)$. It follows that the points $a=c^*(0)$ and
$b=c^*(\pi)$ are fixed by $\gm$, but that $\gm$ acts freely on
$C\sm\{a,b\}$. If we put $P=c^*([0,\pi])$ and $Q=c^*([-\pi,0])$
then the evident maps
\[ [0,\pi] \xra{c^*} P \xra{} C/\ip{\gm}
\xla{} Q \xla{c^*} [-\pi,0]
\]
are homeomorphisms. Moreover, we have $P\cup Q=C$ and
$P\cap Q=\{a,b\}$.
\end{itemize}
\end{remark}
\begin{definition}\label{defn-local-parameter}
Suppose that $a\in Z$, that $U_0$ is an open neighbourhood of $a$,
and that $f_0\:U_0\to\C$ is a holomorphic map. We say that $f_0$ is
a \emph{centred local parameter} at $a$ if $f_0(a)=0$, and that
$df_0$ generates the cotangent space to $Z$ at $a$. We say that the
pair $(U_0,f_0)$ is \emph{normalised} if $f_0$ gives a conformal
isomorphism from $U_0$ to the open unit disc $\Dl=\{z\in\C\st |z|<1\}$.
\end{definition}
\begin{remark}\label{rem-shrinking}
Let $f_0\:U_0\to\C$ be a centred local parameter that need not be
normalised. The holomorphic inverse function theorem then guarantees
that there is a smaller open neighbourhood $U$ with $a\in U\sse U_0$
and a number $\ep>0$ such that $f_0$ restricts to give a conformal
isomorphism $U\to\{z\in\C\st |z|<\ep\}$. This means that the map
$f=\ep^{-1}f_0|_U\:U\to\Dl$ is a conformal isomorphism, so $(U,f)$ is
normalised. The operation that converts $(U_0,f_0)$ to $(U,f)$ will
be called \emph{shrinking}.
\end{remark}
\begin{lemma}\label{lem-al-fixed}
Suppose that $a\in Z$ with $\al(a)=a$. Then there is a normalised
local parameter $f\:U\to\Dl$ at $a$ such that $\al(U)=U$ and
$f(\al(u))=-f(u)$ for all $u\in U$.
\end{lemma}
\begin{proof}
Choose any centred local parameter $f_0\:U_0\to\C$. Put
$U_1=U_0\cap\al(U_0)$ and $f_1=f_0|_{U_1}$; this still gives a
centred local parameter. We can expand $f_1(\al(u))$ as a power
series $\sum_{k=1}^\infty a_kf_1(u)^k$. Because $\al^2=1$, we have
$a_1=\pm 1$. We claim that $a_1$ cannot be equal to $1$. Indeed, as
$\al$ has isolated fixed points, we cannot have $f_1(\al(u))=f_1(u)$
as a power series. Thus, if $a_1=1$ then there must exist $k>1$ such
that $a_i=0$ for $1<i<k$ and $a_k\neq 0$, so
\[ f_1(\al(u)) = f_1(u) + a_kf_1(u)^k + O(f_1(u)^{k+1}). \]
If we substitute this into itself and use $\al^2=1$ we get $2a_k=0$,
which is a contradiction. We must therefore have $a_1=-1$.
Now put $f_2(u)=(f_1(u)-f_1(\al(u)))/2$, so $f_2\:U_1\to\C$ is
holomorphic with $f_2(a)=0$ and $f_2(\al(u))=-f_2(u)$. We also have
$f_2(u)=f_1(u)+O(f_1(u)^2)$, and thus that $f_2(u)$ is again a
centred local parameter. We can therefore produce the required pair
$(U,f)$ by shrinking.
\end{proof}
\begin{lemma}\label{lem-al-not-fixed}
Suppose that $a\in Z$ with $\al(a)\neq a$. Then there is a
normalised local parameter $f\:U\to \Dl$ at $a$ such that
$\al(\ov{U})\cap\ov{U}=\emptyset$.
\end{lemma}
\begin{proof}
By standard arguments with compact Hausdorff spaces, we can choose
open neighbourhoods $V$ of $a$ and $W$ of $\al(a)$ such that
$\ov{V}\cap\ov{W}=\emptyset$. Put $U_0=V\cap\al(W)$, so $U_0$ is an
open neighbourhood of $a$ with
$\ov{U_0}\cap\al(\ov{U_0})=\emptyset$. Now let $f\:U\to \Dl$ be any
normalised local parameter at $a$ with $U\sse U_0$.
\end{proof}
\begin{lemma}\label{lem-bt-fixed}
Suppose that $a\in Z$ with $\bt(a)=a$. Then there is a normalised
local parameter $f\:U\to \Dl$ at $a$ with $\bt(U)=U$ and
$f(\bt(u))=\ov{f(u)}$ for all $u\in U$.
\end{lemma}
\begin{proof}
Choose any centred local parameter $f_0\:U_0\to\C$. Put
$U_1=U_0\cap\bt(U_0)$ and $f_1=f_0|_{U_1}$; this still gives a
centred local parameter. The map $u\mapsto\ov{f_1(\bt(u))}$ is
another centred local parameter on $U_1$, so we have
$\ov{f_1(\bt(u))}=c\,f_1(u)+O(f_1(u)^2)$ for some $c\neq 0$. Using
$\bt^2=1$ we find that $\ov{c}\,c=1$, so $c=e^{2i\tht}$ for some
$\tht\in\R$. Now put
\[ f_2(u) = (e^{i\tht} f_1(u) + e^{-i\tht}\ov{f_1(\bt(u))})/2
= e^{i\tht} f_1(u) + O(f_1(u)^2).
\]
This is again a centred local parameter at $a$, and it satisfies
$f_2(\bt(u))=\ov{f_2(u)}$. Shrinking now gives the required pair
$(U,f)$.
\end{proof}
\begin{corollary}\label{cor-fixed-circles}
The fixed set $Z^{\ip{\bt}}$ is a closed submanifold of $Z$, and so
is diffeomorphic to a finite disjoint union of circles. The same
applies to $Z^{\ip{\al\bt}}$.
\end{corollary}
\begin{proof}
Any normalised local parameter $f\:U\to \Dl$ as in the lemma gives a
diffeomorphism
\[ U\cap Z^{\ip{\bt}}\to \Dl\cap\R = (-1,1), \]
and it follows easily from this that $Z^{\ip{\bt}}$ is a closed
submanifold of $Z$. As $\al\bt$ is an equally good example of an
anticonformal involution, we see that $Z^{\ip{\al\bt}}$ is also a
closed submanifold.
\end{proof}
\begin{remark}\label{rem-preserved-circle}
As $\al$ commutes with $\bt$, it preserves the set $Z^{\ip{\bt}}$.
However, if $Z^{\ip{\bt}}$ has several components, then they need not
be preserved individually. If a certain component is preserved, then
Remark~\ref{rem-circle-involution} will apply.
\end{remark}
\begin{lemma}\label{lem-bt-al-fixed}
Suppose that $a\in Z$ satisfies $\al(a)=\bt(a)=a$. Then there is a
normalised local parameter $f\:U\to \Dl$ at $a$ such that
$f(\al(u))=-f(u)$ and $f(\bt(u))=\ov{f(u)}$ for all $u\in U$.
Moreover:
\begin{itemize}
\item[(a)] $f$ induces a conformal isomorphism
$g\:U/\ip{\al}\to \Dl$ with $g([u])=f(u)^2$.
\item[(b)] $f$ restricts to give a diffeomorphism from
$U^{\ip{\bt}}$ to the real axis in $\Dl$.
\item[(c)] Similarly, $f$ restricts to give a diffeomorphism from
$U^{\ip{\al\bt}}$ to the imaginary axis in $\Dl$.
\item[(d)] $g$ restricts to give homeomorphisms
$U^{\ip{\bt}}/\ip{\al}\to [0,1)$ and
$U^{\ip{\al\bt}}/\ip{\al}\to(-1,0]$.
\end{itemize}
\end{lemma}
\begin{proof}
We choose a local parameter $f_0$ with $f_0(\al(u))=-f(u)$ as in
Lemma~\ref{lem-al-fixed}. We then take this as the initial choice in
the proof of Lemma~\ref{lem-bt-fixed}. This gives a normalised local
parameter $f$ with $f(\bt(u))=\ov{f(u)}$, and by inspecting the
construction we see that the property $f(\al(u))=-f(u)$ is retained
as well. The additional properties~(a) to~(d) follow easily.
\end{proof}
\subsubsection{Branched coverings}
First, we give a formal definition:
\begin{definition}
We put $\Dl=\{z\in\C\st |z|<1\}$, and $\Dl'=\Dl\sm\{0\}$. We let
$\pi\:\Dl\tm\{0,1\}\to \Dl$ denote the projection, and we let
$\sg\:\Dl\to \Dl$ denote the squaring map.
Let $f\:X\to Y$ be a holomorphic map between Riemann surfaces. We
say that $f$ is a \emph{branched double covering} if for each
$y\in Y$ there is a diagram of one of the following types:
\[ \xymatrix{
\Dl\tm\{0,1\} \ar[r]^q \ar[d]_{\pi} &
X \ar[d]^f &&
\Dl \ar[r]^q \ar[d]_{\sg} &
X \ar[d]^f \\
\Dl \ar[r]_p &
Y &&
\Dl \ar[r]_p &
Y
} \]
where
\begin{itemize}
\item[(a)] $p$ is a holomorphic chart with $p(0)=y$.
\item[(b)] The square is a pullback, so $q$ gives a holomorphic
isomorphism from $\Dl\tm\{0,1\}$ or $\Dl$ to $f^{-1}(p(\Dl))$.
\end{itemize}
\end{definition}
\begin{remark}
We note that in the left hand case we have $|p^{-1}\{y\}|=2$, whereas
in the right hand case we have $|p^{-1}\{y\}|=1$, so the two cases are
disjoint. In the right hand case we say that $y$ is a \emph{branch
point}, and we write $B(f)$ for the set of branch points. We note
that all points in $p(\Dl)\sm\{y\}$ have two preimages and so are not
branch points; thus, the set $B(f)$ is discrete.
\end{remark}
\begin{lemma}\label{lem-removable}
Let $X$ and $Y$ be Riemann surfaces such that $Y$ is isomorphic to
$\Dl$ or $\Dl\amalg \Dl$, and let $a$ be a point in $X$. Then any
holomorphic map $f\:X\sm\{a\}\to Y$ has a unique holomorphic
extension $f\:X\to Y$.
\end{lemma}
\begin{proof}
By choosing a chart around $a$ we can reduce to the case where $X=\Dl$
and $a=0$. Now, even if $Y\simeq \Dl\amalg \Dl$ the image $f(\Dl')$ will
be connected and therefore contained in one of the two copies of
$\Dl$. We can thus assume that $Y=\Dl$. This case is just the standard
theorem on removable singularities in complex analysis.
\end{proof}
\begin{proposition}\label{prop-un-branched}
Let $Y$ be a Riemann surface, let $V$ be a discrete subset of $Y$,
and put $Y'=Y\sm V$. Let $\CX$ be the category of branched double
coverings $f\:X\to Y$ with $B(f)\sse V$, and let $\CX'$ be the
category of unbranched double coverings of $Y'$. (In both cases, the
morphisms are holomorphic isomorphisms covering the identity on $Y$
or $Y'$.) Let $R\:\CX\to\CX'$ be the evident restriction functor,
given by $R(X\xra{f}Y)=(X'\xra{f'}Y')$ where $X'=f^{-1}(Y')$ and
$f'=f|_{X'}$. Then $R$ is an equivalence of categories.
\end{proposition}
\begin{proof}
Suppose we have objects $(X_i\xra{f_i}Y)\in\CX$ for $i=0,1$ and an
isomorphism $g'\:X'_0\to X'_1$ with $f'_1g_1=f'_0$. We claim that
there is a unique holomorphic extension $g\:X_0\to X_1$, and that
this satisfies $f_1g=f_0$. This can be checked locally on $Y$, so we
can restrict attention to a small neighbourhood of a point in $V$ and
therefore assume that $(X_1\xra{f_1}Y)$ is either
$(\Dl\tm\{0,1\}\xra{\pi}\Dl)$ or $(\Dl\xra{\sg}\Dl)$. In either case it is
clear from Lemma~\ref{lem-removable} that $g'$ has a unique
holomorphic extension $g\:X_0\to X_1$. We can also apply the
uniqueness clause in the same lemma to the map $f'_1g'=f'_0$; this
gives $f_1g=f_0$. We now see that $R$ is full and faithful.
We now need to show that $R$ is essentially surjective. Consider an
unbranched covering $f'\:X'\to Y'$. For each $v\in V$ we can choose
a chart $p_v\:\Dl\to Y$ with $p_v(0)=v$. As $V$ is discrete we may
assume, after shrinking the charts if necessary, that the sets
$p_v(\Dl)$ are disjoint. In particular, this means that
$p_v(\Dl)\cap V=\{v\}$, so $p_v(\Dl')\sse Y'$. The pullback $p_v^*(X')$
is an unbranched double cover of $\Dl'$, so it is isomorphic to
$(\Dl'\tm\{0,1\}\xra{\pi'}\Dl')$ or to $(\Dl'\xra{\sg'}\Dl')$. In either
case there is an evident way to extend $p_v^*(X')$ to give a branched
cover of all of $\Dl$, and this in turn extends $X'$ to give a branched
cover of $p_v(\Dl)$. These extensions can be patched together to give
a branched cover $(X\xra{f}Y)$ extending the original unbranched
cover, as required.
\end{proof}
\begin{definition}\label{defn-monodromy}
Let $(X\xra{f}Y)$ be a branched double cover with $B(f)\sse V$,
giving an unbranched cover $(X'\xra{f'}Y')$. For any closed loop
$u\:[0,1]\to Y'$, we define the \emph{monodromy} $\mu_X(u)\in\Z/2$ as
follows. The fibre $F=(f')^{-1}\{u(0)\}$ will have precisely two
elements. For any element $a$, there is a unique continuous lift
$\tu\:[0,1]\to X'$ with $f\tu=u$ and $u(0)=a$. We put
$\sg(a)=\tu(1)\in F$. This defines a permutation $\sg\:F\to F$; we
put $\mu_X(u)=0$ if $\sg$ is the identity, and $\mu_X(u)=1$ if $\sg$
is the transposition. This depends only on the homotopy class of the
loop $u$.
Next, for $v\in V$ we let $\om_v$ denote a small loop in $Y'$ that
winds once around $v$ and does not wind around any of the other
points in $V$. It is clear that $v$ is a branch point for $X$ iff
$\mu_X(\om_v)=1$.
\end{definition}
\begin{lemma}\label{lem-homology-monodromy}
There is a unique homomorphism $\ov{\mu}_X\:H_1(Y')\to\Z/2$ such that
$\mu_X(u)=\ov{\mu}_X([u])$ for all loops $u$, where $[u]$
denotes the homology class represented by $u$.
\end{lemma}
\begin{proof}
We can reduce to the case where $Y'$ is connected, and choose a
basepoint $b\in Y'$. The Hurewicz map then
gives an isomorphism $h_b\:\pi_1(Y',b)_{\text{ab}}\to H_1(Y')$, and
it follows that there is a unique homomorphism
$\ov{\mu}\:H_1(Y')\to\Z/2$ with $\mu(u)=\ov{\mu}(\ip{u})$ for all
loops $u$ based at $b$. If $u$ is a loop that is not based at $b$,
we can choose a path $w$ from $b$ to $u(0)$, and let $u'$ be the loop
given by $w$ followed by $u$ followed by the reverse of $w$, so $u'$
is based at $b$. We then have $[u']=[u]$ and $\mu(u)=\mu(u')$,
so we still have $\mu(u)=\ov{\mu}([u])$.
\end{proof}
\begin{lemma}\label{lem-covering-classification}
The map
\[ [X\xra{f}Y]\mapsto\ov{\mu}_X \]
gives a bijection from the set of isomorphism classes in $\CX$ (or
$\CX'$) to $\Hom(H_1(Y'),\Z/2)$.
\end{lemma}
\begin{proof}
We can again reduce to the case where $Y'$ is connected, choose a
basepoint $b\in Y'$, and use the Hurewicz isomorphism
$H_1(Y')=\pi_1(Y',b)_{\text{ab}}$. The claim is then that unbranched
double covers of $Y'$ are classified by homomorphisms
$\pi_1(Y',b)\to\Z/2$, which is standard covering theory.
\end{proof}
\begin{definition}
Let $f\:X\to Y$ be a branched double covering. We define
$\chi\:X\to X$ by
\[ \chi(x) = \begin{cases}
x' & \text{ if } f^{-1}\{f(x)\} = \{x,x'\}
\text{ with } x'\neq x \\
x & \text{ if } f^{-1}\{f(x)\} = \{x\}.
\end{cases}
\]
This is easily seen to be holomorphic.
\end{definition}
\begin{proposition}\label{prop-branched-unique}
Suppose that $Y$ is isomorphic to $\C_\infty$, and that $V\subset Y$
is a finite subset of even size. Then:
\begin{itemize}
\item[(a)] There is a branched covering $f\:X\to Y$ for which
$B(f)=V$.
\item[(b)] If $f_0\:X_0\to Y$ and $f_1\:X_1\to Y$ are as in~(a),
then there is an isomorphism $g\:X_0\to X_1$ with $f_1g=f_0$, and
we have $\chi_1g=g\chi_0$.
\item[(c)] If $g,h\:X_0\to X_1$ are as in~(b), then either $h=g$
or $h=\chi_1g$.
\end{itemize}
\end{proposition}
\begin{proof}
Recall that for each $v\in V$ we have a loop $\om_v$ and a homology
class $[\om_v]\in H_1(Y')$. A standard calculation using the
Mayer-Vietoris sequence shows that $H_1(Y')$ is generated by these
classes subject only to the relation $\sum_{v\in V}[\om_v]=0$. As
$|V|$ is even it follows that there is a unique homomorphism
$\nu\:H_1(Y')\to\Z/2$ with $\nu([\om_v])=1$ for all $v\in V$. By
Lemma~\ref{lem-covering-classification}, there exist unbranched
coverings of $Y'$ with monodromy $\nu$, and any two such are
isomorphic. By Proposition~\ref{prop-un-branched}, it follows that
there exist branched coverings of $Y$ whose branch set is precisely
$V$, and any two such are isomorphic. This proves~(a) and~(b) except
for the fact that $\chi_1g=g\chi_0$. This fact and claim~(c) are
standard covering theory and are left to the reader.
\end{proof}
\subsection{The projective family is universal}
\label{sec-P-universal}
We will prove the following two theorems.
\begin{theorem}\label{thm-classify-precromulent}
Let $X$ be a precromulent surface. Then there is a unique number
$a\in(0,1)$ such that $X\simeq PX(a)$ as $G$-equivariant Riemann
surfaces. Moreover, there are precisely two isomorphisms
$X\to PX(a)$, which are related by the action of $\lm^2$.
\end{theorem}
\begin{theorem}\label{thm-classify-cromulent}
Let $X$ be a cromulent surface. Then there is a unique number
$a\in(0,1)$ such that $X\simeq PX(a)$ as $G$-equivariant Riemann
surfaces. Moreover, there is precisely one cromulent isomorphism
$X\to PX(a)$.
\end{theorem}
The proofs will be given after some preliminary results. First,
however, we record a consequence of
Theorem~\ref{thm-classify-cromulent}:
\begin{corollary}
Let $X$ be a cromulent surface. Then $X$ admits a curve system (as
in Definition~\ref{defn-curve-system}) and has standard isotropy (as
in Definition~\ref{defn-std-isotropy}).
\end{corollary}
\begin{proof}
Definition~\ref{defn-P-curves} and
Proposition~\ref{prop-P-std-isotropy} show that this holds for
$PX(a)$, which is sufficient by
Theorem~\ref{thm-classify-cromulent}.
\end{proof}
\begin{proposition}\label{prop-v-zero}
There is a unique point $v_0\in X$ such that $\lm(v_0)=v_0$ and
$\lm_*=i\:T_{v_0}X\to T_{v_0}X$. Similarly, there is a unique point
$v_1=\mu(v_0)\in X$ such that $\lm(v_1)=v_1$ and
$\lm_*=-i\:T_{v_1}X\to T_{v_1}X$.
\end{proposition}
\begin{proof}
In $V^*$ there are precisely two points that are fixed by $\lm$, and
they are exchanged by $\mu$. The same must therefore be true in $X$.
Let $a$ be one of these points, so the other one is $b=\mu(a)$. Note
that the holomorphic involution $\lm^2$ fixes $a$ and $b$, so it acts
as $-1$ on $T_aX$ and $T_bX$ by Lemma~\ref{lem-al-fixed}, so $\lm$
acts as $\pm i$. Now consider the commutative diagram on the left
below, and the resulting commutative diagram on the right:
\[ \xymatrix{
X \ar[r]^\mu \ar[d]_\lm &
X \ar[d]^{\lm^{-1}} & &
T_{a}X \ar[r]^{\mu_*} \ar[d]^{\lm_*} &
T_{b}X \ar[d]^{\lm^{-1}_*} \\
X \ar[r]_\mu & X && T_{a}X \ar[r]_{\mu_*} & T_{b}X.
}
\]
From this we see that the eigenvalue of $\lm$ on $T_aX$ is the same
as the eigenvalue of $\lm^{-1}$ on $T_bX$. The claim follows easily
from this.
\end{proof}
\begin{definition}\label{defn-e-points}
We define points $e_0,e_1,e_\infty\in X/\ip{\lm^2}$ as follows.
First, we let $v_0$ and $v_1$ be as in Proposition~\ref{prop-v-zero},
and we put $e_0=[v_0]$ and $e_\infty=[v_1]$. Next, we note that in
$V^*$ there are precisely two points with stabiliser
$\ip{\lm^2\mu,\lm^2\nu}$ (namely $3$ and $5$), and that these are
exchanged by $\lm^2$. It follows that the same is true in $V$.
These points therefore form an equivalence class in $X/\ip{\lm^2}$,
which we call $e_1$.
\end{definition}
We saw in Corollary~\ref{cor-quotient-types} that $X/\ip{\lm^2}$ is
isomorphic to $\C_\infty$. It is well-known that the conformal
automorphisms of $\C_\infty$ are the M\"obius transformations, and
that these act freely and transitively on the triples of distinct
points in $\C_\infty$. This validates the following definition:
\begin{definition}\label{defn-p}
We let $p\:X/\ip{\lm^2}\to\C_\infty$ denote the unique conformal
isomorphism such that $p(e_i)=i$ for $i\in\{0,1,\infty\}$. We will
also use the symbol $p$ for the composite
$X\to X/\ip{\lm^2}\xra{p}\C_\infty$.
\end{definition}
\begin{lemma}\label{lem-p-equivariance}
If we let $G$ act on $\C_\infty$ by
\[ \lm(z)=-z \hspace{5em}
\mu(z)=1/z \hspace{5em}
\nu(z)=\ov{z}
\]
then the map $p$ is equivariant.
\end{lemma}
\begin{proof}
The maps $x\mapsto -p(\lm(x))$ and $x\mapsto 1/p(\mu(x))$ and
$x\mapsto \ov{p(\nu(x))}$ all have the defining property of $p$.
\end{proof}
\begin{lemma}\label{lem-v-eleven}
There is a unique point $v_{11}\in V$ such that the number
$a=p(v_{11})$ lies in $(0,1)$. Moreover, if we define
$v_{10}=\lm(v_{11})$ and $v_{12}=\lm\mu(v_{11})$ and
$v_{13}=\mu(v_{11})$, then we have
\[ p(v_{10}) = -a \hspace{4em}
p(v_{11}) = a \hspace{4em}
p(v_{12}) = -1/a \hspace{4em}
p(v_{13}) = 1/a.
\]
\end{lemma}
\begin{proof}
Put $W=\{x\in X\st\stab_G(x)=\ip{\lm^2,\nu}\}$. Because the group
$\ip{\lm^2,\nu}$ is normal in $G$, this is a $G$-set. As $X$ is
precromulent, it is equivariantly isomorphic to the $G$-set
\[ W^* = \{i\in V^*\st \stab_G(i)=\ip{\lm^2,\nu}\} =
\{10,11,12,13\} \simeq G/\ip{\lm^2,\nu}.
\]
As $p$ gives an equivariant isomorphism $X/\ip{\lm^2}\to\C_\infty$,
it must restrict to give an equivariant injection
\[ W/\ip{\lm^2} \to
\{z\in\C_\infty\st\stab_G(z)=\ip{\lm^2,\nu}\}.
\]
The domain here is just $W$, and the codomain is the set
\[ U = \R_\infty\sm\{0,1,-1,\infty\} =
(-\infty,-1) \amalg
(-1, 0) \amalg
( 0, 1) \amalg
(1,\infty).
\]
The action of $G$ on $\C_\infty$ permutes the four components of $U$
transitively, so the preimage under $p$ of each component must
contain precisely one point of $W$. The claim is clear from this.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm-classify-precromulent}]
The map $p\:X\to\C_\infty$ is a branched covering, with branch set
$p(U)$, where $U=\{x\in X\st\lm^2(x)=x\}$. This is equivariantly
isomorphic to the $G$-set
\[ U^*=\{i\in V^*\st\lm^2(i)=i\}=\{0,1,10,11,12,13\}, \]
and using this we see that $p(U)=\{0,\infty,\pm a,\pm 1/a\}$. This
is the same as the branch set for the map $p\:PX(a)\to\C_\infty$
defined in Remark~\ref{rem-P-quotient}. Our claim now follows from
Proposition~\ref{prop-branched-unique}.
\end{proof}
\begin{corollary}\label{cor-precromulent-aut}\leavevmode
\begin{itemize}
\item[(a)] If $X$ is a precromulent surface, then the group of
precromulent automorphisms of $X$ is $C_2=\{1,\lm^2\}$.
\item[(b)] If $X$ and $Y$ are isomorphic precromulent surfaces, then
there are precisely two precromulent isomorphisms between them. If
one of them is $\phi$, then the other is $\phi\lm_X^2=\lm_Y^2\phi$.
\end{itemize}
\end{corollary}
\begin{proof}
Clear from the theorem.
\end{proof}
\begin{corollary}\label{cor-cromulent-aut}
If $X$ is a cromulent surface, then the only cromulent automorphism
is the identity.
\end{corollary}
\begin{proof}
Any cromulent automorphism is also a precromulent automorphism, and
so is $1$ or $\lm^2$; but $\lm^2$ does not preserve the labelling.
\end{proof}
\begin{corollary}\label{cor-cromulent-iso}
Let $X$ and $Y$ be isomorphic cromulent surfaces; then there is a
unique cromulent isomorphism between them. \qed
\end{corollary}
\begin{proposition}\label{prop-labellings}
Let $X$ be a precromulent surface. Then $X$ has precisely two
cromulent labellings, which are related by the action of $\lm^2$.
\end{proposition}
\begin{proof}
By Theorem~\ref{thm-classify-precromulent} we can reduce to the
case $X=PX(a)$. In particular, this means that we have a curve
system, and nets as in Section~\ref{sec-fundamental}.
The labelling given in Definition~\ref{defn-P} is cromulent, by
Proposition~\ref{prop-P-fundamental}. It follows easily that if we
change the labelling by $\lm^2$, then it remains cromulent.
Now suppose we have another cromulent labelling, say
$(v'_i)_{i=0}^{13}$. This must have the form $v'_i=v_{\phi(i)}$ for
some $\phi$ in the group $\Aut(V^*)$, which is described by
Proposition~\ref{prop-aut-V}. Proposition~\ref{prop-v-zero} shows
that we must have $\phi(0)=0$ and $\phi(1)=1$.
Proposition~\ref{prop-aut-V} shows that $\phi(2)\in\{2,4\}$, and
after replacing $\phi$ by $\phi\lm^2$ if necessary, we may assume
that $\phi(2)=2$. Assuming this, we see from
Proposition~\ref{prop-aut-V} that $\phi(i)=i$ for $i\in\{2,3,4,5\}$,
and that $\phi(6)\in\{6,8\}$.
Next, as the new labelling is cromulent, there must be a connected
component $F'\sse\{x\in X\st\stab_G(x)=1\}$ whose closure contains
the set $U=\{v_{\phi(0)},v_{\phi(3)},v_{\phi(6)},v_{\phi(11)}\}$.
From the discussion in Section~\ref{sec-fundamental} we see that $F'$
must be $\gm(PF'_{16}(a))$ for some $\gm\in G$. We have seen that
$U$ contains $v_0$, $v_3$ and either $v_6$ or $v_8$. By inspecting
the nets in Section~\ref{sec-fundamental}, we see that this can only
be consistent if $\gm=1$ and $\phi(6)=6$ and $\phi(11)=11$. By
consulting Proposition~\ref{prop-aut-V} again, we conclude that
$\phi=1$, as required.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm-classify-cromulent}]
Let $X$ be a cromulent surface. By
Theorem~\ref{thm-classify-precromulent}, there is a unique $a$ such
that $X\simeq PX(a)$ as precromulent surfaces. Choose a precromulent
isomorphism $\phi\:X\to PX(a)$. The points $\phi^{-1}(v_i)$ give a
cromulent labelling of $X$. By Proposition~\ref{prop-labellings},
this must either be the given labelling of $X$, or its twist by
$\lm^2$. Thus, after replacing $\phi$ by $\phi\lm^2$ if necessary,
we may assume that $\phi$ is a cromulent isomorphism. It is unique
by Corollary~\ref{cor-cromulent-iso}.
\end{proof}
\begin{remark}\label{rem-p-extra}
We now see that the map $p\:X/\ip{\lm^2}\to\C_\infty$ from
Definition~\ref{defn-p} must factor as a cromulent isomorphism
$X\to PX(a)$, followed by the map $p\:PX(a)\to\C_\infty$ from
Remark~\ref{rem-P-quotient}. This implies that we have the
following additional properties:
\begin{align*}
p(v_0) &= 0 \\
p(v_1) &= \infty \\
p(v_2) = p(v_4) &= -1 \\
p(v_3) = p(v_5) &= \pp 1 \\
p(v_6) = p(v_8) &= \pp i \\
p(v_7) = p(v_9) &= -i \\
p(v_{10}) &= -a \\
p(v_{11}) &= \pp a \\
p(v_{12}) &= -1/a \\
p(v_{13}) &= \pp 1/a.
\end{align*}
\end{remark}
The number $p(v_i)$ is recorded in Maples as \mcode+v_C[i]+.
\begin{remark}\label{rem-p-hat}
For some purposes it is more convenient to work with the round sphere
$S^2\subset\R^3$ rather than the Riemann sphere $\C_\infty$. We
identify them using the stereographic projection map
\[ \xi(x+iy) =
\left(
\frac{2x}{x^2+y^2+1},
\frac{2y}{x^2+y^2+1},
\frac{x^2+y^2-1}{x^2+y^2+1}
\right).
\]
This has been normalised so that the unit circle in $\C$ is sent to
the equator. It is a standard fact that $\xi$ is conformal. The
resulting complex structure on the tangent spaces
$T_yS^2=\{t\in\R^3\st t.y=0\}$ can be described in terms of the cross
product of vectors: we have $(a+ib)t=at+bt\tm y$. This means that an
ordered basis $(u,v)$ for $T_yS^2$ is oriented iff $\det(y,v,u)>0$.
Note also that if $\xi(z)=y$ then
\begin{align*}
\xi(-z) &= ( - y_1,\; - y_2,\; \pp y_3) \\
\xi(1/z) &= (\pp y_1,\; - y_2,\; - y_3) \\
\xi(\ov{z}) &= (\pp y_1,\; - y_2,\; \pp y_3).
\end{align*}
It follows that the composite $\hp=\xi p\:X\to S^2$ has properties as
follows:
\begin{align*}
\hp(v_0) &= (\pp 0,\pp 0, -1) \\
\hp(v_1) &= (\pp 0,\pp 0,\pp 1) \\
\hp(v_2) &= \hp(v_4) = ( - 1,\pp 0,\pp 0) \\
\hp(v_3) &= \hp(v_5) = (\pp 1,\pp 0,\pp 0) \\
\hp(v_6) &= \hp(v_8) = (\pp 0,\pp 1,\pp 0) \\
\hp(v_7) &= \hp(v_9) = (\pp 0, - 1,\pp 0) \\
\hp(v_{10}) &= ( - 2a,0,a^2-1)/(a^2+1) \\
\hp(v_{11}) &= (\pp 2a,0,a^2-1)/(a^2+1) \\
\hp(v_{12}) &= ( - 2a,0,1-a^2)/(a^2+1) \\
\hp(v_{13}) &= (\pp 2a,0,1-a^2)/(a^2+1)
\end{align*}
\begin{align*}
\hp_1(\lm(x)) &= -\hp_1(x) & \hp_2(\lm(x)) &= -\hp_2(x) & \hp_3(\lm(x)) &= \pp\hp_3(x) \\
\hp_1(\mu(x)) &= \pp\hp_1(x) & \hp_2(\mu(x)) &= -\hp_2(x) & \hp_3(\mu(x)) &= -\hp_3(x) \\
\hp_1(\nu(x)) &= \pp\hp_1(x) & \hp_2(\nu(x)) &= -\hp_2(x) & \hp_3(\nu(x)) &= \pp\hp_3(x).
\end{align*}
The points $\hp(v_i)$ are recorded in Maple as \mcode+v_S2[i]+, and
the induced action of $g\in G$ on $u\in S^2$ is \mcode+act_S2[g](u)+.
\end{remark}
\section{The hyperbolic family}
\label{sec-H}
\subsection{The groups \texorpdfstring{$\Pi$ and $\tPi$}{Pi and Pi tilde}}
\label{sec-Pi}
Later we will construct a family of cromulent surfaces as quotients of
the unit disc by different actions of a certain group $\Pi$. In this
section we define and study $\Pi$, together with a larger group $\tPi$
such that $\tPi/\Pi=G$. Actions of $\tPi$ and $\Pi$ will be given in
the following section.
\begin{definition}\label{defn-Pi}
Let $\Pi$ be the abstract group generated by symbols $\bt_k$ (for
$k\in\Z/8$) subject to the following relations:
\begin{align*}
\bt_{k+4} &= \bt_k^{-1} \\
\bt_0\bt_1\bt_2\bt_3\bt_4\bt_5\bt_6\bt_7 &= 1.
\end{align*}
\end{definition}
We saw in Proposition~\ref{prop-pi-one} that any cromulent surface has
fundamental group isomorphic to $\Pi$. Note also that the
abelianization $\Pi_{\text{ab}}$ is freely generated (as an abelian
group) by $\bt_0,\dotsc,\bt_3$, and so is isomorphic to $\Z^4$.
We can introduce an alternative set of generators as follows:
\begin{align*}
\al_0 &= \bt_3^{-1}\bt_2^{-1}\bt_0 &
\al_1 &= \bt_1\bt_2\bt_3 &
\al_2 &= \bt_2^{-1} &
\al_3 &= \bt_3^{-1} \\
\bt_0 &= \al_2^{-1}\al_3^{-1}\al_0 &
\bt_1 &= \al_1\al_3\al_2 &
\bt_2 &= \al_2^{-1} &
\bt_3 &= \al_3^{-1}.
\end{align*}
In terms of these, the relation $\bt_0\dotsb\bt_7=1$ becomes the
standard relation
\[ [\al_0,\al_1][\al_2,\al_3] =
\al_0\al_1\al_0^{-1}\al_1^{-1}
\al_2\al_3\al_2^{-1}\al_3^{-1} = 1
\]
for the fundamental group of a surface of genus $2$. (However, we
prefer to use the generators $\bt_k$, for reasons of symmetry.)
\begin{checks}
hyperbolic/Pi_check.mpl: check_Pi_alpha()
\end{checks}
\begin{remark}\label{rem-runs}
Note that the relation $\bt_0\dotsb\bt_7=1$ can be conjugated to give
\[ \bt_k\bt_{k+1}\dotsb\bt_{k+7} = 1 \]
for any $k\in\Z/8$. This in turn gives
\begin{align*}
\bt_k\bt_{k+1}\bt_{k+2}\bt_{k+3}\bt_{k+4}\bt_{k+5}\bt_{k+6} &=
\bt_{k+7}^{-1} &&\hspace{-1em}=
\bt_{k+3} \\
\bt_k\bt_{k+1}\bt_{k+2}\bt_{k+3}\bt_{k+4}\bt_{k+5} &=
(\bt_{k+6}\bt_{k+7})^{-1} &&\hspace{-1em}=
\bt_{k+3}\bt_{k+2} \\
\bt_k\bt_{k+1}\bt_{k+2}\bt_{k+3}\bt_{k+4} &=
(\bt_{k+5}\bt_{k+6}\bt_{k+7})^{-1} &&\hspace{-1em}=
\bt_{k+3}\bt_{k+2}\bt_{k+1} \\
\bt_k\bt_{k+1}\bt_{k+2}\bt_{k+3} &=
(\bt_{k+4}\bt_{k+5}\bt_{k+6}\bt_{k+7})^{-1} &&\hspace{-1em}=
\bt_{k+3}\bt_{k+2}\bt_{k+1}\bt_k \\
\bt_k\bt_{k+1}\bt_{k+2} &=
(\bt_{k+3}\bt_{k+4}\bt_{k+5}\bt_{k+6}\bt_{k+7})^{-1} &&\hspace{-1em}=
\bt_{k+3}\bt_{k+2}\bt_{k+1}\bt_k\bt_{k-1} \\
\bt_k\bt_{k+1} &=
(\bt_{k+2}\bt_{k+3}\bt_{k+4}\bt_{k+5}\bt_{k+6}\bt_{k+7})^{-1} &&\hspace{-1em}=
\bt_{k+3}\bt_{k+2}\bt_{k+1}\bt_k\bt_{k-1}\bt_{k-2} \\
\bt_k &=
(\bt_{k+1}\bt_{k+2}\bt_{k+3}\bt_{k+4}\bt_{k+5}\bt_{k+6}\bt_{k+7})^{-1} &&\hspace{-1em}=
\bt_{k+3}\bt_{k+2}\bt_{k+1}\bt_k\bt_{k-1}\bt_{k-2}\bt_{k-3}.
\end{align*}
Thus:
\begin{itemize}
\item Any increasing or decreasing run of length at least $5$ can be
replaced by a strictly shorter run in the opposite direction.
\item Any run of length $4$ can be reversed.
\end{itemize}
\end{remark}
\begin{definition}\label{defn-beta-reduced}
A word in the letters $\{\bt_i\st i\in\Z/8\}$ is \emph{reduced} if
\begin{itemize}
\item[(a)] There are no subwords of the form $\bt_i\bt_{i+4}$
\item[(b)] There are no subwords of the form
$\bt_i\bt_{i+1}\bt_{i+2}\bt_{i+3}\bt_{i+4}$.
\item[(c)] There are no subwords of the form
$\bt_i\bt_{i-1}\bt_{i-2}\bt_{i-3}$.
\end{itemize}
\end{definition}
\begin{proposition}\label{prop-beta-reduced}
Every element $\pi\in\Pi$ can be expressed in a unique way as a reduced
word in the letters $\bt_i$. Moreover, if $\pi$ is represented by a
word $w$, then the corresponding reduced word $w'$ can be obtained
from $w$ by repeatedly cancelling pairs of the form $\bt_i\bt_{i+4}$,
and shortening or reversing runs as in Remark~\ref{rem-runs}.
\end{proposition}
\begin{proof}
A straightforward argument by induction on the length shows that for
every word there is a reduced word (of the same length or less) that
represents the same element of $\Pi$. Uniqueness is less obvious,
but follows from Dehn's algorithm and small cancellation theory. In
more detail, put $X=\{\bt_0,\bt_1,\bt_2,\bt_3\}$, then put
\begin{align*}
\sg_i^+ &= \bt_i\bt_{i+1}\dotsb\bt_{i+7} \\
\sg_i^- &= \bt_i\bt_{i-1}\dotsb\bt_{i-7},
\end{align*}
where each $\bt_j$ is replaced by an element of $X\amalg X^{-1}$ in
the obvious way. Then the set
\[ S=\{\sg_i^+\st i\in\Z/8\} \amalg \{\sg_i^-\st i\in\Z/8\} \]
consists of reduced words in the free group $FX$, and is closed under
taking inverses and cyclic permutations. If $\al$ and $\bt$ are
distinct elements of $S$, then they have length $8$, but they share at
most one initial letter. The shared fraction of $1/8$ is less than
$1/6$, so the main theorem of~\cite{gr:daw} is applicable, and the
conclusion follows easily. (For a textbook treatment, see Chapter~V
of~\cite{lysc:cgt}.)
\end{proof}
\begin{remark}
Elements of $\Pi$ are represented in Maple as lists of integers. For
example, the list \mcode+[5,3,6]+ represents $\bt_5\bt_3\bt_6$. The
function \mcode+is_Pi_reduced(L)+ decides whether a list \mcode+L+
corresponds to a reduced word. The function \mcode+Pi_reduce(L)+
finds the unique reduced word that represents the same element as
\mcode+L+. Here \mcode+L+ is allowed to have arbitrary integer
entries, but the first step in the reduction process is to reduce
them modulo $8$ so that they lie in $\{0,1,\dotsc,7\}$.
Multiplication and inversion are implemented by the functions
\mcode+Pi_mult+ and \mcode+Pi_inv+. All of this (together with
various other things) is in the file \fname+hyperbolic/Pi.mpl+.
\end{remark}
\begin{corollary}\label{cor-trivial-centre}
The centre of $\Pi$ is trivial, so the conjugation map
$\Pi\to\Aut(\Pi)$ is injective.
\end{corollary}
\begin{proof}
Any nontrivial element $\pi\in\Pi\sm\{1\}$ can be represented by a
nonempty reduced word $w$. Let $i$ and $j$ be the
indices of the first and last letters in $w$, and put
\[ L = \{i-1,i,i+1,i+4,j-1,j+1,j+4\} \]
(so $|L|\leq 7$). Choose $k\in(\Z/8)\sm L$, so that $\bt_kw$ and
$w\bt_k$ are reduced words. Their first letters are $\bt_k$ and
$\bt_i$, so they are different. It follows that $\pi$ is
not central.
\end{proof}
\begin{proposition}\label{prop-Aut-Pi}
The group $\Pi$ has automorphisms $\lm_*$, $\mu_*$ and $\nu_*$, which
act on the generators $\bt_i$ as follows:
\[ \begin{array}{|c|c|c|c|c|c|c|c|c|} \hline
& \bt_0 & \bt_1 & \bt_2 & \bt_3 & \bt_4 & \bt_5 & \bt_6 & \bt_7
\\ \hline
\lm_* &
\bt_2 & \bt_3 & \bt_4 & \bt_5 & \bt_6 & \bt_7 & \bt_0 & \bt_1
\\ \hline
\mu_* &
\bt_2\bt_0\bt_1 &
\bt_5\bt_4\bt_3 &
\bt_0\bt_7\bt_6 &
\bt_2\bt_3\bt_1 &
\bt_5\bt_4\bt_6 &
\bt_7\bt_0\bt_1 &
\bt_2\bt_3\bt_4 &
\bt_5\bt_7\bt_6 \\ \hline
\nu_* &
\bt_0 &
\bt_2\bt_1\bt_2 &
\bt_6 &
\bt_0\bt_7\bt_0 &
\bt_4 &
\bt_6\bt_5\bt_6 &
\bt_2 &
\bt_4\bt_3\bt_4 \\ \hline
\end{array}
\]
Moreover, we have $\lm_*^4=\mu_*^2=\nu_*^2=(\lm_*\nu_*)^2=1$, and the
automorphisms $(\lm_*\mu_*)^2$ and $(\nu_*\mu_*)^2$ are inner. More precisely, for
any $\pi\in\Pi$ we have
\begin{align*}
(\lm_*\mu_*)^2(\pi) &=
(\bt_7\bt_6)\;\pi\;(\bt_7\bt_6)^{-1} \\
(\nu_*\mu_*)^2(\pi) &=
(\bt_6\bt_0\bt_7\bt_6)\;\pi\;(\bt_6\bt_0\bt_7\bt_6)^{-1}.
\end{align*}
\end{proposition}
\begin{proof}
We will only discuss $\mu_*$; similar arguments cover the other two
cases. Put $B=\{\bt_k\:k\in\Z/8\}\subset\Pi$ and define
$\mu_*\:B\to\Pi$ by the above table. To show that this extends to an
endomorphism of $\Pi$, we must check that it respects the relations,
or in other words that
\begin{align*}
\mu_*(\bt_k)\mu_*(\bt_{k+4}) &= 1 \\
\mu_*(\bt_0)\mu_*(\bt_1)\mu_*(\bt_2)\mu_*(\bt_3)
\mu_*(\bt_4)\mu_*(\bt_5)\mu_*(\bt_6)\mu_*(\bt_7) &= 1.
\end{align*}
The first of these follows easily by inspecting the definition of
$\mu_*(\bt_k)$ and using the relation $\bt_j\bt_{j+4}=1$ three
times. For the second, the left hand side can be grouped as
\[
\bt_2(\bt_0\bt_1
\bt_5\bt_4)\bt_3
(\bt_0(\bt_7\bt_6
\bt_2\bt_3)(\bt_1
\bt_5)\bt_4)(\bt_6
\bt_7\bt_0\bt_1
\bt_2\bt_3\bt_4
\bt_5)\bt_7\bt_6.
\]
Working from the inside out, we see that the content of each matched
pair of parentheses cancels down to $1$. This leaves
$\bt_2\bt_3\bt_7\bt_6$, which cancels further down to $1$. We thus
have an endomorphism $\mu_*$ as claimed. It satisfies
\[ \mu_*^2(\bt_0) = \mu_*(\bt_2\bt_0\bt_1) =
\mu_*(\bt_2)\mu_*(\bt_0)\mu_*(\bt_1) =
\bt_0(\bt_7(\bt_6\;\bt_2)(\bt_0(\bt_1\;\bt_5)\bt_4)\bt_3) =
\bt_0.
\]
By similar arguments, it also satisfies $\mu_*^2(\bt_i)=\bt_i$ for
all other indices $i$, so $\mu_*^2=1$. In particular, this means
that $\mu_*$ is an automorphism. The identities
$\lm_*^4=\nu_*^2=(\lm_*\nu_*)^2=1$ can be verified in a similar way.
Moreover, it will also suffice to check the identities for
$(\lm_*\mu_*)^2(\pi)$ and $(\nu_*\mu_*)^2(\pi)$ when $\pi=\bt_k$ for
some $k$, and this is again straightforward (but long).
\begin{checks}
hyperbolic/Pi_check.mpl: check_Pi_relations()
\end{checks}
\end{proof}
\begin{definition}\label{defn-tPi}
We let $\tPi$ be the abstract group generated by symbols
$\lm,\mu,\nu,\bt_k$ (for $k\in\Z/8$) subject to the following relations:
\begin{align*}
\bt_{k+4} &= \bt_k^{-1} \\
\bt_0\bt_1\bt_2\bt_3\bt_4\bt_5\bt_6\bt_7 &= 1 \\
\lm^4 &= \mu^2 = \nu^2 = (\lm\nu)^2 = 1 \\
(\lm\mu)^2 &= \bt_7\bt_6 \\
(\nu\mu)^2 &= \bt_6\bt_0\bt_7\bt_6 \\
\lm\bt_k\lm^{-1} &= \lm_*(\bt_k) \\
\mu\bt_k\mu &= \mu_*(\bt_k) \\
\nu\bt_k\nu &= \nu_*(\bt_k).
\end{align*}
(Here $\lm_*(\bt_k)$, $\mu_*(\bt_k)$ and $\nu_*(\bt_k)$ refer to the
words given in the table in Proposition~\ref{prop-Aut-Pi}.)
\end{definition}
\begin{proposition}\label{prop-tPi}
$\Pi$ can be identified with the subgroup of $\tPi$ generated by
$\bt_0,\dotsc,\bt_7$. Moreover, this subgroup is normal, and there
is a canonical isomorphism $\tPi/\Pi\simeq G$.
\end{proposition}
\begin{proof}
Let $\Pi'$ be the subgroup of $\tPi$ generated by
$\bt_0,\dotsc,\bt_7$. Using the defining relations for
$\lm\bt_k\lm^{-1}$, $\mu\bt_k\mu$ and $\nu\bt_k\nu$ we see that
$\Pi'$ is normal. If we just set all the elements $\bt_k$ to the
identity in our presentation for $\tPi$, we obtain a presentation for
$\tPi/\Pi'$; from this it is clear that $\tPi/\Pi'=G$. There is an
evident surjective homomorphism $\phi\:\Pi\to\Pi'$, sending $\bt_k$
to $\bt_k$ for all $k$. All that is left is to show that this is
injective. For this, we let $\gm\:\Pi\to\Inn(\Pi)$ be the usual
conjugation map, given by $\gm(\al)(\pi)=\al\pi\al^{-1}$. This is
surjective by the definition of $\Inn(\Pi)$, and injective by
Corollary~\ref{cor-trivial-centre}, so it is an isomorphism. Next,
we define
\begin{align*}
\dl(\lm) &= \lm_* \in \Aut(\Pi) \\
\dl(\mu) &= \mu_* \in \Aut(\Pi) \\
\dl(\nu) &= \nu_* \in \Aut(\Pi) \\
\dl(\bt_k) &= \gm(\bt_k) \in \Inn(\Pi) \lhd \Aut(\Pi).
\end{align*}
Using Proposition~\ref{prop-Aut-Pi} we see that this is compatible
with the defining relations for $\tPi$, so it extends to give a
homomorphism $\tPi\to\Aut(\Pi)$. We can now chase the diagram
\[ \xymatrix{
\Pi \ar[d]_\gm^\simeq \ar[r]^\phi &
\tPi \ar[d]_{\dl} \ar@{->>}[r] &
G \ar[d] \\
\Inn(\Pi) \ar@{ >->}[r]_{\text{inc}} &
\Aut(\Pi) \ar@{->>}[r] & \Out(\Pi)
} \]
to see that $\phi$ is injective as required.
\end{proof}
\begin{remark}
Elements of $\tPi$ are represented as pairs \mcode+[T,L]+, with
\mcode+T+ in $G$ and \mcode+L+ in $\Pi$. The multiplication rule is
not obvious; it is implemented by the function \mcode+Pi_tilde_mult+,
which uses data stored in the table \mcode+G_Pi_cocycle+. Inversion
is implemented by the function \mcode+Pi_tilde_inv+. All this is in
\fname+hyperbolic/Pi.mpl+.
\end{remark}
We next give an alternate presentation of $\Pi$, which will be helpful
when we want to analyse actions on the unit disc.
\begin{proposition}\label{prop-Pi-Sg}
$\Pi$ is generated by the elements
\begin{align*}
\sg_a &= \bt_5\bt_6 &
\sg_b &= \bt_2 &
\sg_c &= \bt_7\bt_0 \\
\sg_d &= \bt_4 &
\sg_e &= \bt_1\bt_2 &
\sg_f &= \bt_3\bt_4
\end{align*}
subject only to the relations
\[ \sg_a\sg_c\sg_e\sg_f =
\sg_b\sg_e^{-1}\sg_b^{-1}\sg_a^{-1} =
\sg_d\sg_f^{-1}\sg_d^{-1}\sg_c^{-1} = 1.
\]
\end{proposition}
\begin{proof}
First, it is straightforward to check that the stated relations hold
in $\Pi$. Thus, if we let $\Sg$ denote the abstract group with the
indicated generators and relations, we have a homomorphism
$\phi\:\Sg\to\Pi$ given by $\phi(\sg_t)=\sg_t$ for all
$t\in\{a,b,c,d,e,f\}$. Next, we define elements
$\bt_0,\dotsc,\bt_7\in\Sg$ by
\[ \bt_0 = \sg_d^{-1} \hspace{3em}
\bt_1 = \sg_e\sg_b^{-1} \hspace{3em}
\bt_2 = \sg_b \hspace{3em}
\bt_3 = \sg_f\sg_d^{-1} \hspace{3em}
\bt_{i+4}=\bt_i^{-1}.
\]
We claim that $\bt_0\bt_1\dotsb\bt_7=1$ in $\Sg$. It will suffice to
prove the conjugate relation
$\bt_5\bt_6\bt_7\bt_0\bt_1\bt_2\bt_3\bt_4=1$. We can write out the
left hand side and group the terms as
\[ (\sg_b\sg_e^{-1}\sg_b^{-1})(\sg_d\sg_f^{-1}\sg_d^{-1})
\sg_e(\sg_b^{-1}\sg_b)\sg_f(\sg_d^{-1}\sg_d).
\]
The relation $\sg_b\sg_e^{-1}\sg_b^{-1}\sg_a^{-1}=1$ converts the
first parenthesised term to $\sg_a$, and similarly the second becomes
$\sg_c$. The third parenthesised term is clearly the identity, as is
the fourth. This leaves $\sg_a\sg_c\sg_e\sg_f$, which is the
identity by the first defining relation for $\Sg$. This means that
we have a homomorphism $\psi\:\Pi\to\Sg$ given by $\psi(\bt_i)=\bt_i$
for all $i$. Straightforward calculations now show that
$\phi\psi(\bt_i)=\bt_i$ for all $i$, and $\psi\phi(\sg_t)=\sg_t$ for
all $t$, so $\phi\psi=1_\Pi$ and $\psi\phi=1_\Sg$.
\begin{checks}
hyperbolic/Pi_check.mpl: check_Pi_sigma
\end{checks}
\end{proof}
\subsection{Cromulent actions}
\label{sec-H-action}
We next describe a family of actions of $\tPi$ on the unit disc
\[ \Dl = \{z\in\C\st |z|<1\}. \]
This has a standard Riemannian metric as follows:
\[ ds^2 = 4|dz|^2 / (1-|z|^2)^2. \]
We recall some standard facts about this, most of which can be found
in~\cite[Section 4.1]{an:hg}, for example. The Gaussian curvature of
the metric is equal to $-1$. The conformal automorphisms of $\Dl$
have the form
\[ z \mapsto \lm \frac{z-\al}{1-\ov{\al}z}, \]
with $|\al|<1$ and $|\lm|=1$, and that these all preserve the
metric. Similarly, the anticonformal automorphisms have the form
\[ z \mapsto \lm \frac{\ov{z}-\al}{1-\ov{\al}\,\ov{z}}. \]
The geodesic distance function is
\[ \dhyp(z,w) =
2\arctanh\left|\frac{z-w}{1-\ov{z}w}\right|.
\]
\begin{checks}
hyperbolic/HX_check.mpl: check_hyperbolic_metric()
\end{checks}
Now fix a number $b$ with $0<b<1$, and put $b_+=\sqrt{1+b^2}$ and
$b_-=\sqrt{1-b^2}$. We define an action of $\tPi$ on the
unit disc $\Dl=\{z\in\C\st |z|<1\}$ as follows:
\begin{align*}
\lm(z) &= iz &
\bt_0(z) &= \frac{b_+z+1}{z+b_+} \\
\mu(z) &= \frac{b_+z-b^2-i}{(b^2-i)z-b_+} &
\bt_1(z) &= \frac{b_+^3z-(2+i)b^2-i}{((i-2)b^2+i)z+b_+^3} \\
\nu(z) &= \ov{z} &
\bt_{2n}(z) &= i^n\bt_0(z/i^n) \\
&&
\bt_{2n+1}(z) &= i^n\bt_1(z/i^n).
\end{align*}
One can check by direct calculation that the defining relations for
$\tPi$ are satisfied.
\begin{checks}
hyperbolic/HX_check.mpl: check_Pi_action()
\end{checks}
\begin{remark}
The parameters $b$, $b_+$ and $b_-$ are \mcode+a_H+, \mcode+ap_H+ and
\mcode+am_H+ in the Maple code. The action of $g\in\tPi$ on $z\in\Dl$
is given by \mcode+act_Pi_tilde(g,z)+. The group $G$ can be
considered as a subset of $\tPi$ which is not a subgroup. For
$g\in G$, the alternative notation \mcode+act_H[g](z)+ also works for
\mcode+act_Pi_tilde(g,z)+. This notation is potentially misleading,
because we do not actually have an action of $G$.
\end{remark}
\begin{remark}\label{rem-a-Ho}
The evident analogue of Remark~\ref{rem-a-Po} applies to \mcode+a_H+.
Most parts of the code treat \mcode+a_H+ as a symbol, but there are
also global variables \mcode+a_H0+ and \mcode+a_H1+ which holds
numerical values for \mcode+a_H+. These should be set using the
function \mcode+set_a_H0+, which is defined in
\fname+hyperbolic/HX0.mpl+. There is also a function
\mcode+simplify_H+ analogous to the function \mcode+simplify_P+
mentioned in Remark~\ref{rem-simplify-P}: it applies some
simplification rules like $\sqrt{1-b^4}=b_+b_-$ which are not always
used by Maple.
\end{remark}
We now put $HX(b)=\Dl/\Pi$. Later we will give this the structure of a
cromulent surface. As a first step, we note that $\Pi$ is normal in
$\tPi$, so there is an induced action of $G=\tPi/\Pi$ on $HX(b)$.
\begin{definition}\label{defn-v-H}
We define points $v_0,\dotsc,v_{13}\in \Dl$ as follows:
\begin{align*}
v_0 &= 0 &
v_6 &= \frac{1+i}{\rt}\;\frac{\rt-b_-}{b_+} &
v_{10} &= i(b_+-b) \\
v_1 &= \frac{1+i}{2}b_+ &
v_7 &= i \, v_6 &
v_{11} &= b_+-b \\
v_2 &= \frac{b\,b_--b_+}{i-b^2} &
v_8 &= -v_6 &
v_{12} &= (b+b_+)\frac{i+(i+2)b^2}{(b+b_+)^2+b^2} \\
v_3 &= \frac{b\,b_--b_+}{ib^2-1} &
v_9 &= -i\, v_6 &
v_{13} &= i\,\ov{v_{12}} \\
v_4 &= i \, v_3 \\
v_5 &= -i \, v_2.
\end{align*}
\end{definition}
\begin{definition}\label{defn-v-H-extra}
We also consider additional points in the same $\Pi$-orbits:
\[ \begin{array}{rll}
v_{ 1.1} &= \pp i\,v_1 &= \bt_2\bt_1(v_1) \\
v_{ 1.2} &= - v_1 &= \bt_1\bt_2\bt_3\bt_4(v_1) \\
v_{ 1.3} &= -i\, v_1 &= \bt_3\bt_4(v_1) \\
v_{ 2.1} &= \pp\ov{v_2 } &= \bt_6(v_2) \\
v_{ 3.1} &= - \ov{v_3 } &= \bt_4(v_3) \\
v_{ 4.1} &= \pp\ov{v_4 } &= \bt_6(v_4) \\
v_{ 5.1} &= - \ov{v_5 } &= \bt_4(v_5) \\
v_{10.1} &= \pp\ov{v_{10}} &= \bt_6(v_{10}) \\
v_{11.1} &= - \ov{v_{11}} &= \bt_4(v_{11}) \\
v_{12.1} &= - v_{12} &= \bt_1(v_{12}) \\
v_{12.2} &= - \ov{v_{12}} &= \bt_2\bt_1(v_{12}) \\
v_{12.3} &= \pp\ov{v_{12}} &= \bt_6(v_{12}) \\
v_{13.1} &= - v_{13} &= \bt_4\bt_3\bt_4(v_{13}) \\
v_{13.2} &= \pp\ov{v_{13}} &= \bt_3\bt_4(v_{13}) \\
v_{13.3} &= - \ov{v_{13}} &= \bt_4(v_{13}).
\end{array} \]
\end{definition}
\begin{remark}
The points $v_i$ are represented as \mcode+v_H[i]+, and there are also
points \mcode+v_H0[i]+ and \mcode+v_H1[i]+ that are obtained by
substituting the numerical values \mcode+a_H0+ or \mcode+a_H1+ for
the symbol \mcode+a_H+. The entries in the above table have the form
$v_i=\gm_i(v_j)$ where $i$ is not an integer, $j$ is the integer part
of $i$, and $\gm_i$ is an element of $\Pi$. The element $\gm_i$
is represented in maple as \mcode+v_H_fraction_offset[i]+.
\end{remark}
One can check that $\lm$, $\mu$ and $\nu$ act as follows:
\begin{align*}
\lm(v_{ 0}) &= v_{ 0} &
\mu(v_{ 0}) &= \bt_2(\bt_3(\bt_4(v_{ 1}))) &
\nu(v_{ 0}) &= v_{ 0} \\
\lm(v_{ 1}) &= \bt_2(\bt_1(v_{ 1})) &
\mu(v_{ 1}) &= \bt_2(v_{ 0}) &
\nu(v_{ 1}) &= \bt_3(\bt_4(v_{ 1})) \\
\lm(v_{ 2}) &= \bt_4(v_{ 3}) &
\mu(v_{ 2}) &= v_{ 2} &
\nu(v_{ 2}) &= \bt_6(v_{ 2}) \\
\lm(v_{ 3}) &= v_{ 4} &
\mu(v_{ 3}) &= \bt_2(v_{ 5}) &
\nu(v_{ 3}) &= v_{ 5} \\
\lm(v_{ 4}) &= \bt_4(v_{ 5}) &
\mu(v_{ 4}) &= \bt_5(\bt_6(v_{ 4})) &
\nu(v_{ 4}) &= \bt_6(v_{ 4}) \\
\lm(v_{ 5}) &= v_{ 2} &
\mu(v_{ 5}) &= \bt_2(\bt_3(\bt_4(v_{ 3}))) &
\nu(v_{ 5}) &= v_{ 3} \\
\lm(v_{ 6}) &= v_{ 7} &
\mu(v_{ 6}) &= \bt_2(v_{ 9}) &
\nu(v_{ 6}) &= v_{ 9} \\
\lm(v_{ 7}) &= v_{ 8} &
\mu(v_{ 7}) &= \bt_5(v_{ 8}) &
\nu(v_{ 7}) &= v_{ 8} \\
\lm(v_{ 8}) &= v_{ 9} &
\mu(v_{ 8}) &= \bt_5(\bt_4(\bt_3(v_{ 7}))) &
\nu(v_{ 8}) &= v_{ 7} \\
\lm(v_{ 9}) &= v_{ 6} &
\mu(v_{ 9}) &= \bt_2(\bt_3(\bt_4(v_{ 6}))) &
\nu(v_{ 9}) &= v_{ 6} \\
\lm(v_{10}) &= \bt_4(v_{11}) &
\mu(v_{10}) &= v_{12} &
\nu(v_{10}) &= \bt_6(v_{10}) \\
\lm(v_{11}) &= v_{10} &
\mu(v_{11}) &= \bt_2(\bt_3(\bt_4(v_{13}))) &
\nu(v_{11}) &= v_{11} \\
\lm(v_{12}) &= \bt_4(v_{13}) &
\mu(v_{12}) &= v_{10} &
\nu(v_{12}) &= \bt_6(v_{12}) \\
\lm(v_{13}) &= \bt_2(\bt_1(v_{12})) &
\mu(v_{13}) &= \bt_2(v_{11}) &
\nu(v_{13}) &= \bt_3(\bt_4(v_{13})) \\
\end{align*}
This shows that $G$ permutes the corresponding points in $HX(b)$ in
accordance with Definition~\ref{defn-precromulent}. The elements of
$\Pi$ appearing here are recorded in the table
\mcode+v_action_witness_H+, which is defined in the file
\fname+hyperbolic/HX.mpl+.
\begin{checks}
cromulent.mpl: check_precromulent("H")
hyperbolic/HX_check.mpl: check_v_H()
\end{checks}
\subsection{The curve system}
\label{sec-H-curves}
We would now like to construct curves $C_0,\dotsc,C_8$ in $\Dl$ or
$HX(b)$. These will be the fixed sets of certain anticonformal
involutions of $\Dl$, or the images in $HX(b)$ of those fixed sets.
Such involutions can be classified as follows:
\begin{itemize}
\item[(a)] Suppose that $m\in\C$ with $|m|>1$. Put
\[ \xi_m(z) = \frac{m\,\ov{z}-1}{\ov{z}-\ov{m}}. \]
This is an anticonformal involution on $\C_\infty$ that preserves
$\Dl$. Maple notation (defined in the file \fname+hyperbolic/HX.mpl+) is
\mcode+xi(m,z)+. We put
\[ \Xi_m = \{z\in \Dl \st \xi_m(z)=z\}. \]
If we put $d=\sqrt{|m|^2-1}$, then the fixed set of $\xi_m$ in $\C$
is the circle of radius $d$ centred at $m$, and $\Xi_m$ is the
intersection of this circle with $\Dl$. This is a geodesic in $\Dl$,
and the following formula gives an isometric parametrisation
$\om_m\:\R\to\Xi_m$:
\[ \om_m(s) =
\left(\frac{id-1}{\ov{m}}\right)
\frac{(id+1) - i |m| e^s}{i |m| e^s + (id-1)}.
\]
Maple notation is \mcode+xi_curve(m,s)+.
The endpoints of $\Xi_m$ on the unit circle are
$(1\pm id)/\ov{m}$. If we have another involution of the same type
with parameter $m'$, then $\xi_m\xi_{m'}=\xi_{m'}\xi_m$ iff $\Xi_m$
and $\Xi_{m'}$ cross at right angles, iff $\text{Re}(\ov{m}m')=1$.
\item[(b)] Suppose instead that $u\in\C$ with $|u|=1$. We then have
an anticonformal involution $z\mapsto u\ov{z}$. The fixed set in
$\Dl$ is the straight line joining the two square roots of $u$. This
is isometrically parameterised by the map
\[ s\mapsto \sqrt{u}\frac{e^s - 1}{e^s+1} =
\sqrt{u}\tanh(s/2).
\]
\item[(c)] Every anticonformal involution on $\Dl$ arises in one of the
above two ways.
\end{itemize}
\begin{checks}
hyperbolic/HX_check.mpl: check_xi()
\end{checks}
\begin{definition}\label{defn-H-curves}
We define constants $s_0,\dotsc,s_4$ as follows:
\begin{align*}
s_0 &= 2\log\left(\frac{\rt b}{b_+-b_-}\right) &
s_1 &= \frac{1}{2}\log\left(\frac{\rt+b_+}{\rt-b_+}\right) &
s_2 &= \log\left(\frac{1+b}{b_-}\right) \\
s_3 &= \frac{1}{2}\log\left(\frac{b+b_++1}{b+b_+-1}\right) &
s_4 &= \frac{1}{4}\log\left(\frac{b_+^2+2b_++2}{b_+^2-2b_++2}\right).
\end{align*}
We then define maps $\tc_k\:\R\to \Dl$ for $0\leq k\leq 8$ as follows:
\begin{align*}
\tc_0(t) &= \om_{(1+i)/b_+}((t/\pi-1/4)s_0) \\
\tc_1(t) &= e^{i\pi/4}\tanh(t\,s_1/\pi) &
\tc_2(t) &= e^{3i\pi/4}\tanh(t\,s_1/\pi) \\
\tc_3(t) &= \om_{b_+}(-t\,s_2/\pi) &
\tc_4(t) &= \om_{ib_+}(-t\,s_2/\pi) \\
\tc_5(t) &= \tanh(t\,s_3/\pi) &
\tc_6(t) &= i\,\tanh(t\,s_3/\pi) \\
\tc_7(t) &= \om_{ib_+/2+1/b_+}( t\,s_3/\pi - s_4) &
\tc_8(t) &= \om_{ b_+/2+i/b_+}(-t\,s_3/\pi + s_4).
\end{align*}
We write $c_k$ for the composite
\[ \R \xra{\tc_k} \Dl \xra{} \Dl/\Pi = HX(b). \]
We also put $\tC_k=\tc_k(\R)\subset \Dl$ and $C_k=c_k(\R)\subset \Dl/\Pi$.
\end{definition}
\begin{remark}
The function $\tc_k(t)$ is \mcode+c_H[k](t)+, and the constant $s_j$
is \mcode+s_H[j]+. For $k\in\{0,3,4,7,8\}$ we see that $\tc_k(\R)$ is
a circular arc. The centre is recorded as \mcode+c_H_p[k]+, and the
radius as \mcode+c_H_r[k]+.
\end{remark}
The factors $s_k$ are chosen to make the maps $c_k$ periodic with
period $2\pi$. In more detail:
\begin{proposition}\label{prop-c-H-cycle}
For all $t\in\R$ we have
\begin{align*}
\tc_0(t+2\pi) &= \bt_0\bt_2\bt_4\bt_6(\tc_0(t)) \\
\tc_1(t+2\pi) &= \bt_0\bt_7\bt_6\bt_5(\tc_1(t)) &
\tc_2(t+2\pi) &= \bt_2\bt_1\bt_0\bt_7(\tc_2(t)) \\
\tc_3(t+2\pi) &= \bt_0\bt_7(\tc_3(t)) &
\tc_4(t+2\pi) &= \bt_2\bt_1(\tc_4(t)) \\
\tc_5(t+2\pi) &= \bt_0(\tc_5(t)) &
\tc_6(t+2\pi) &= \bt_2(\tc_6(t)) \\
\tc_7(t+2\pi) &= \bt_0\bt_2\bt_1(\tc_7(t)) &
\tc_8(t+2\pi) &= \bt_2\bt_3\bt_4(\tc_8(t)).
\end{align*}
Thus, for $0\leq k\leq 8$ we have $c_k(t+2\pi)=c_k(t)$.
\end{proposition}
(The group element for $\tc_k(t)$ is recorded as \mcode+c_H_cycle[k]+.)
\begin{proof}
Computer calculation.
\begin{checks}
hyperbolic/HX_check.mpl: check_c_H_monodromy()
\end{checks}
\end{proof}
For the equivariance properties of a curve system, it will suffice to check that
the equations below hold in $\Dl$, and this can be done by direct calculation.
\begin{align*}
\lm(\tc_0(t)) &= \bt_4(\tc_0(t+\pi/2)) &
\mu(\tc_0(t)) &= \tc_0(-t) &
\nu(\tc_0(t)) &= \bt_6(\tc_0(-t)) \\
\lm(\tc_1(t)) &= \tc_2(t) &
\mu(\tc_1(t)) &= \bt_5\bt_4\bt_3(\tc_2(\pi+t)) &
\nu(\tc_1(t)) &= \tc_2(-t) \\
\lm(\tc_2(t)) &= \tc_1(-t) &
\mu(\tc_2(t)) &= \bt_2\bt_3\bt_4(\tc_1(\pi+t)) &
\nu(\tc_2(t)) &= \tc_1(-t) \\
\lm(\tc_3(t)) &= \tc_4(t) &
\mu(\tc_3(t)) &= \bt_2\bt_3\bt_4(\tc_3(\pi+t)) &
\nu(\tc_3(t)) &= \tc_3(-t) \\
\lm(\tc_4(t)) &= \bt_4(\tc_3(-t)) &
\mu(\tc_4(t)) &= \tc_4(-t-\pi) &
\nu(\tc_4(t)) &= \bt_6(\tc_4(t)) \\
\lm(\tc_5(t)) &= \tc_6(t) &
\mu(\tc_5(t)) &= \bt_2\bt_3\bt_4(\tc_7(t)) &
\nu(\tc_5(t)) &= \tc_5(t) \\
\lm(\tc_6(t)) &= \tc_5(-t) &
\mu(\tc_6(t)) &= \bt_2\bt_3\bt_4(\tc_8(-t)) &
\nu(\tc_6(t)) &= \tc_6(-t) \\
\lm(\tc_7(t)) &= \bt_2\bt_1(\tc_8(t)) &
\mu(\tc_7(t)) &= \bt_2(\tc_5(t)) &
\nu(\tc_7(t)) &= \bt_3\bt_4(\tc_7(t)) \\
\lm(\tc_8(t)) &= \bt_2\bt_1(\tc_7(-t)) &
\mu(\tc_8(t)) &= \bt_2(\tc_6(-t)) &
\nu(\tc_8(t)) &= \bt_3\bt_4(\tc_8(-t))
\end{align*}
(The elements of $\Pi$ appearing here are recorded in the table
\mcode+c_action_witness_H+.)
\begin{checks}
cromulent.mpl: check_precromulent("H")
\end{checks}
In the case $b=0.75$, these curves and vertices can be illustrated as follows:
\begin{center}
\begin{tikzpicture}[scale=5]
\draw (0,0) circle(1);
\draw[blue] (-1,0) -- (1,0);
\draw[blue] (0,-1) -- (0,1);
\draw[green] (-0.707,-0.707) -- ( 0.707, 0.707);
\draw[green] (-0.707, 0.707) -- ( 0.707,-0.707);
\draw[cyan] (0.800,0.800) (0.294,0.956) arc(-197:-73:0.529);
\draw[cyan,dotted] (-0.800,0.800) (-0.956,0.294) arc(-107:17:0.529);
\draw[cyan,dotted] (-0.800,-0.800) (-0.294,-0.956) arc(-17:107:0.529);
\draw[cyan,dotted] (0.800,-0.800) (0.956,-0.294) arc(73:197:0.529);
\draw[magenta] (1.250,0.000) (0.800,0.600) arc(127:233:0.750);
\draw[magenta,dotted] (0.000,1.250) (-0.600,0.800) arc(-143:-37:0.750);
\draw[magenta,dotted] (-1.250,0.000) (-0.800,-0.600) arc(-53:53:0.750);
\draw[magenta,dotted] (0.000,-1.250) (0.600,-0.800) arc(37:143:0.750);
\draw[magenta] (0.000,1.250) (-0.600,0.800) arc(-143:-37:0.750);
\draw[magenta,dotted] (-1.250,0.000) (-0.800,-0.600) arc(-53:53:0.750);
\draw[magenta,dotted] (0.000,-1.250) (0.600,-0.800) arc(37:143:0.750);
\draw[magenta,dotted] (1.250,0.000) (0.800,0.600) arc(127:233:0.750);
\draw[blue] (0.800,0.625) (0.670,0.742) arc(-222:-62:0.175);
\draw[blue,dotted] (-0.625,0.800) (-0.742,0.670) arc(-132:28:0.175);
\draw[blue,dotted] (-0.800,-0.625) (-0.670,-0.742) arc(-42:118:0.175);
\draw[blue,dotted] (0.625,-0.800) (0.742,-0.670) arc(48:208:0.175);
\draw[blue] (0.625,0.800) (0.471,0.882) arc(-208:-48:0.175);
\draw[blue,dotted] (-0.800,0.625) (-0.882,0.471) arc(-118:42:0.175);
\draw[blue,dotted] (-0.625,-0.800) (-0.471,-0.882) arc(-28:132:0.175);
\draw[blue,dotted] (0.800,-0.625) (0.882,-0.471) arc(62:222:0.175);
\fill[black](0.000,0.000) circle(0.007);
\fill[black](0.625,0.625) circle(0.007);
\fill[black](0.322,0.573) circle(0.007);
\fill[black](-0.573,0.322) circle(0.007);
\fill[black](-0.322,-0.573) circle(0.007);
\fill[black](0.573,-0.322) circle(0.007);
\fill[black](0.426,0.426) circle(0.007);
\fill[black](-0.426,0.426) circle(0.007);
\fill[black](-0.426,-0.426) circle(0.007);
\fill[black](0.426,-0.426) circle(0.007);
\fill[black](0.000,0.500) circle(0.007);
\fill[black](0.500,0.000) circle(0.007);
\fill[black](-0.500,0.000) circle(0.007);
\fill[black](-0.493,0.685) circle(0.007);
\fill[black](0.685,0.493) circle(0.007);
\fill[black](-0.625,-0.625) circle(0.007);
\fill[black](-0.322,0.573) circle(0.007);
\fill[black](-0.625,0.625) circle(0.007);
\fill[black](0.000,-0.500) circle(0.007);
\fill[black](0.493,0.685) circle(0.007);
\fill[black](0.625,-0.625) circle(0.007);
\fill[black](-0.573,-0.322) circle(0.007);
\fill[black](0.685,-0.493) circle(0.007);
\fill[black](0.493,-0.685) circle(0.007);
\fill[black](-0.685,-0.493) circle(0.007);
\fill[black](0.322,-0.573) circle(0.007);
\fill[black](-0.493,-0.685) circle(0.007);
\fill[black](-0.685,0.493) circle(0.007);
\fill[black](0.573,0.322) circle(0.007);
\draw( 0.30, 0.25) node{$c_1$};
\draw(-0.30, 0.25) node{$c_2$};
\draw( 0.80,-0.03) node{$c_5$};
\draw(-0.05, 0.80) node{$c_6$};
\draw( 0.773, 0.246) node{$c_{0}$};
\draw( 0.486, 0.179) node{$c_{3}$};
\draw(-0.229, 0.566) node{$c_{4}$};
\draw( 0.771, 0.422) node{$c_{7}$};
\draw( 0.422, 0.771) node{$c_{8}$};
\draw( 0.060, 0.020) node{$v_{0}$};
\draw( 0.655, 0.595) node{$v_{1}$};
\draw( 0.272, 0.583) node{$v_{2}$};
\draw(-0.623, 0.322) node{$v_{3.1}$};
\draw(-0.372,-0.573) node{$v_{4.1}$};
\draw( 0.623,-0.322) node{$v_{5}$};
\draw( 0.386, 0.426) node{$v_{6}$};
\draw(-0.386, 0.426) node{$v_{7}$};
\draw(-0.396,-0.456) node{$v_{8}$};
\draw( 0.396,-0.456) node{$v_{9}$};
\draw(-0.040, 0.470) node{$v_{10}$};
\draw( 0.470,-0.030) node{$v_{11}$};
\draw(-0.440,-0.030) node{$v_{11.1}$};
\draw( 0.443, 0.685) node{$v_{12}$};
\draw( 0.635, 0.493) node{$v_{13}$};
\draw(-0.655,-0.595) node{$v_{1.2}$};
\draw(-0.382, 0.573) node{$v_{4}$};
\draw(-0.675, 0.595) node{$v_{1.1}$};
\draw(-0.060,-0.530) node{$v_{10.1}$};
\draw(-0.433, 0.685) node{$v_{12.2}$};
\draw( 0.655,-0.595) node{$v_{1.3}$};
\draw(-0.623,-0.322) node{$v_{5.1}$};
\draw( 0.745,-0.493) node{$v_{13.2}$};
\draw(-0.433,-0.685) node{$v_{12.1}$};
\draw(-0.745,-0.493) node{$v_{13.1}$};
\draw( 0.382,-0.563) node{$v_{2.1}$};
\draw( 0.433,-0.685) node{$v_{12.3}$};
\draw(-0.745, 0.493) node{$v_{13.3}$};
\draw( 0.523, 0.312) node{$v_{3}$};
\end{tikzpicture}
\end{center}
The dotted curves are images of the undotted curves under the action
of various elements of $\Pi$. Combinatorially, the whole picture is
essentially the same as the first net described in
Section~\ref{sec-fundamental}.
The following diagrams show how the picture varies as $b$ varies from
$0$ to $1$.
\begin{center}
\begin{tikzpicture}[scale=3]
\begin{scope}
\draw (0,0) (1,0) arc(0:90:1);
\draw[blue] (0,1) -- (0,0) -- (1,0);
\draw[green] (0,0) -- ( 0.707, 0.707);
\draw[cyan] (0.980,0.980) (0.021,1.000) arc(-181:-89:0.960);
\draw[blue] (0.980,0.510) (0.606,0.795) arc(-217:-88:0.470);
\draw[blue] (0.510,0.980) (0.041,0.999) arc(-182:-53:0.470);
\draw[magenta](1.020,0) (0.817,0) arc(180:101:0.203);
\draw[magenta](0,1.020) (0,0.817) arc(270:349:0.203);
\draw (0.5,-0.1) node {$b=0.20$};
\end{scope}
\begin{scope}[xshift=1.3cm]
\draw (0,0) (1,0) arc(0:90:1);
\draw[blue] (0,1) -- (0,0) -- (1,0);
\draw[green] (0,0) -- ( 0.707, 0.707);
\draw[cyan] (0.900,0.900) (0.118,0.993) arc(-187:-83:0.787);
\draw[blue] (0.900,0.556) (0.633,0.774) arc(-219:-77:0.344);
\draw[blue] (0.556,0.900) (0.220,0.976) arc(-193:-51:0.344);
\draw[magenta](1.111,0) (0.627,0) arc(180:116:0.484);
\draw[magenta](0,1.111) (0,0.627) arc(270:334:0.484);
\draw (0.5,-0.1) node {$b=0.50$};
\end{scope}
\begin{scope}[xshift=2.6cm]
\draw (0,0) (1,0) arc(0:90:1);
\draw[blue] (0,1) -- (0,0) -- (1,0);
\draw[green] (0,0) -- ( 0.707, 0.707);
\draw[cyan] (0.800,0.800) (0.294,0.956) arc(-197:-73:0.529);
\draw[blue] (0.800,0.625) (0.670,0.742) arc(-222:-62:0.175);
\draw[blue] (0.625,0.800) (0.471,0.882) arc(-208:-48:0.175);
\draw[magenta](1.250,0) (0.500,0) arc(180:127:0.750);
\draw[magenta](0,1.250) (0,0.500) arc(270:323:0.750);
\draw (0.5,-0.1) node {$b=0.75$};
\end{scope}
\begin{scope}[xshift=3.9cm]
\draw (0,0) (1,0) arc(0:90:1);
\draw[blue] (0,1) -- (0,0) -- (1,0);
\draw[green] (0,0) -- ( 0.707, 0.707);
\draw[cyan] (0.720,0.720) (0.561,0.828) arc(-214:-56:0.192);
\draw[blue] (0.720,0.694) (0.702,0.712) arc(-225:-47:0.026);
\draw[blue] (0.694,0.720) (0.676,0.737) arc(-223:-45:0.026);
\draw[magenta](1.389,0) (0.425,0) arc(180:134:0.964);
\draw[magenta](0,1.389) (0,0.425) arc(270:316:0.964);
\draw (0.5,-0.1) node {$b=0.95$};
\end{scope}
\end{tikzpicture}
\end{center}
In particular, we see that there is no change in the combinatorial
structure.
We write $HF_1(b)$ for the following region.
\begin{center}
\begin{tikzpicture}[scale=5]
\draw[magenta]( 1.250, 0.000) +(139:0.750) arc(139:221:0.750);
\draw[blue] ( 0.800, 0.625) +(180:0.175) arc(180:229:0.175);
\draw[blue] ( 0.625, 0.800) +(221:0.175) arc(221:270:0.175);
\draw[magenta]( 0.000, 1.250) +(229:0.750) arc(229:311:0.750);
\draw[blue] (-0.625, 0.800) +(270:0.175) arc(270:319:0.175);
\draw[blue] (-0.800, 0.625) +(311:0.175) arc(311:360:0.175);
\draw[magenta](-1.250, 0.000) +(-41:0.750) arc(-41: 41:0.750);
\draw[blue] (-0.800,-0.625) +( 0:0.175) arc( 0: 49:0.175);
\draw[blue] (-0.625,-0.800) +( 41:0.175) arc( 41: 90:0.175);
\draw[magenta]( 0.000,-1.250) +( 49:0.750) arc( 49:131:0.750);
\draw[blue] ( 0.625,-0.800) +( 90:0.175) arc( 90:139:0.175);
\draw[blue] ( 0.800,-0.625) +(131:0.175) arc(131:180:0.175);
\draw[magenta,->] ( 0.500, 0.000) -- ( 0.500,-0.001);
\draw[magenta,->] (-0.500, 0.000) -- (-0.500,-0.001);
\draw[magenta,->] ( 0.000, 0.500) -- (-0.001, 0.500);
\draw[magenta,->] ( 0.000,-0.500) -- (-0.001,-0.500);
\draw[blue,->] ( 0.800, 0.625) +(205:0.175) -- +(206:0.175);
\draw[blue,->] ( 0.625, 0.800) +(246:0.175) -- +(245:0.175);
\draw[blue,->] (-0.625, 0.800) +(295:0.175) -- +(296:0.175);
\draw[blue,->] (-0.800, 0.625) +(335:0.175) -- +(334:0.175);
\draw[blue,->] (-0.800,-0.625) +( 25:0.175) -- +( 26:0.175);
\draw[blue,->] (-0.625,-0.800) +( 65:0.175) -- +( 64:0.175);
\draw[blue,->] ( 0.625,-0.800) +(115:0.175) -- +(116:0.175);
\draw[blue,->] ( 0.800,-0.625) +(155:0.175) -- +(154:0.175);
\fill( 0.685, 0.493) circle(0.01);
\fill( 0.625, 0.625) circle(0.01);
\fill( 0.493, 0.685) circle(0.01);
\fill(-0.493, 0.685) circle(0.01);
\fill(-0.625, 0.625) circle(0.01);
\fill(-0.685, 0.493) circle(0.01);
\fill(-0.685,-0.493) circle(0.01);
\fill(-0.625,-0.625) circle(0.01);
\fill(-0.493,-0.685) circle(0.01);
\fill( 0.493,-0.685) circle(0.01);
\fill( 0.625,-0.625) circle(0.01);
\fill( 0.685,-0.493) circle(0.01);
\draw( 0.685, 0.493) node[anchor=west]{$v_{13}$};
\draw( 0.625, 0.625) node[anchor=south west]{$v_{1}$};
\draw( 0.493, 0.685) node[anchor=south]{$v_{12}$};
\draw(-0.493, 0.685) node[anchor=south]{$v_{12.2}$};
\draw(-0.625, 0.625) node[anchor=south east]{$v_{1.1}$};
\draw(-0.685, 0.493) node[anchor=east]{$v_{13.3}$};
\draw(-0.685,-0.493) node[anchor=east]{$v_{13.1}$};
\draw(-0.625,-0.625) node[anchor=north east]{$v_{1.2}$};
\draw(-0.493,-0.685) node[anchor=north]{$v_{12.1}$};
\draw( 0.493,-0.685) node[anchor=north]{$v_{12.3}$};
\draw( 0.625,-0.625) node[anchor=north west]{$v_{1.3}$};
\draw( 0.685,-0.493) node[anchor=west]{$v_{13.2}$};
\draw( 0.535, 0.605) node{$a$};
\draw(-0.530, 0.605) node{$*a$};
\draw( 0.000, 0.450) node{$b$};
\draw( 0.000,-0.450) node{$*b$};
\draw(-0.595, 0.532) node{$c$};
\draw(-0.590,-0.532) node{$*c$};
\draw(-0.450, 0.000) node{$d$};
\draw( 0.445, 0.000) node{$*d$};
\draw(-0.535,-0.605) node{$e$};
\draw( 0.539,-0.605) node{$*e$};
\draw( 0.595,-0.532) node{$f$};
\draw( 0.590, 0.532) node{$*f$};
\end{tikzpicture}
\end{center}
Note that we have marked each edge with a direction and a label.
These edges are portions of the curves $\tC_k$, or images of those
curves under the action of elements of $\Pi$, so in particular they
are geodesics. One can check that when
$b=\sqrt{2/\sqrt{3}-1}\simeq 0.3933$, the fundamental domain is
actually a right angled regular dodecagon.
\begin{checks}
hyperbolic/HX_check.mpl: check_H_F1()
\end{checks}
\begin{proposition}\label{prop-Pi-free}
The group $\Pi$ acts freely on $\Dl$, and $HF_1(b)$ is a fundamental
domain for this action.
\end{proposition}
\begin{proof}
We will use a standard theorem of Poincar\'e (which is presented as
Theorem VII.1.7 in the textbook~\cite{iv:hg}, for example).
Recall that in Proposition~\ref{prop-Pi-Sg} we introduced elements
$\sg_t\in\Pi$ for $t\in\{a,b,c,d,e,f\}$. We claim that $\sg_t$
carries the edge labelled $*t$ to the edge labelled $t$, and carries
the interior of $HF_1(b)$ to the exterior. As the edges are geodesic,
this claim can be checked by calculating the effect of $\sg_t$ on the
ends of the relevant edge, which is straightforward. Thus, we have a
system of edge pairings as in the Poincar\'e Theorem.
For the next ingredient, we need to know the internal angles of the
hyperbolic polygon $HF_1(b)$. We claim that they are all $\pi/2$. For
example, side $a$ is part of the curve $\tC_8$, and by inspecting the
formula for $\tc_8(t)$ we see that $\tC_8=\Xi_{m_8}$, where
$m_8=b_+/2+i/b_+$. Similarly, side $b$ is part of $\tC_4=\Xi_{m_4}$,
where $m_4=ib_+$. As we mentioned previously, geodesics $\Xi_m$ and
$\Xi_{m'}$ are orthogonal if and only if $\text{Re}(\ov{m}m')=1$. It
is clear that $\text{Re}(\ov{m_4}m_8)=1$, so edges $a$ and $b$ meet
at right angles. Similarly, $*f$ is part of $\tC_7=\Xi_{m_7}$, where
$m_7=ib_+/2+1/b_+$, and this also meets $a$ at right angles, by the same
test. It follows by symmetry that the remaining internal angles are
$\pi/2$.
We next need to understand the edge cycle map. We write $\ov{a}$ for
the edge $a$ considered in the reverse direction, and similarly for
the other edges. We write $E$ for the set of directed edges, so
$|E|=24$. We define $\al\:E\to E$ by
\[ \al(t) = *t \qquad
\al(*t) = t \qquad
\al(\ov{t}) = \ov{*t} \qquad
\al(\ov{*t}) = \ov{t}.
\]
We also define $\bt(u)$ to be the directed edge different from $u$
that has the same initial point as $u$. For example, we have
$\bt(a)=*f$, $\bt(b)=\ov{a}$ and so on. The edge cycle map $\gm$ is
the composite $\bt\al\:E\to E$. It can be written in disjoint cycle
notation as
\[ \gm = (a\;c\;e\;f)
(*f\;*e\;*c\;*a)
(b\;\ov{*e}\;\ov{*b}\;\ov{*a})
(\ov{a}\;\ov{b}\;\ov{e}\;*b)
(d\;\ov{*f}\;\ov{*d}\;\ov{*c})
(\ov{c}\;\ov{d}\;\ov{f}\;*d).
\]
For each cycle in $\gm$, we can consider the sum of the internal
angles at the initial points of the corresponding edges. The key
hypothesis in the Poincar\'e Theorem is that this sum must be a
multiple of $2\pi$. This is clearly satisfied here, as each cycle
has length $4$ and all internal angles are $\pi/2$.
The theorem now tells us that the side pairing maps $\sg_t$ generate
a group $\Sg$ that acts freely on $\Dl$, with $HF_1(b)$ as a fundamental
domain. Moreover, $\Sg$ is generated freely by the $\sg_t$ subject
only to a small family of relations, one for each cycle in $\gm$.
The relations are constructed in an obvious way from the cycles, with
a factor $\sg_t$ for each entry $t$ or $*t$, and a factor
$\sg_t^{-1}$ for each entry $\ov{t}$ or $\ov{*t}$. Specifically, the
relations are as follows:
\begin{align*}
\sg_a\sg_c\sg_e\sg_f &= 1 &
\sg_f^{-1}\sg_e^{-1}\sg_c^{-1}\sg_a^{-1} &= 1 \\
\sg_b\sg_e^{-1}\sg_b^{-1}\sg_a^{-1} &= 1 &
\sg_a\sg_b\sg_e\sg_b^{-1} &= 1 \\
\sg_d\sg_f^{-1}\sg_d^{-1}\sg_c^{-1} &= 1 &
\sg_c\sg_d\sg_f\sg_d^{-1} &= 1.
\end{align*}
The relations on the right are equivalent to those on the left and so
can be ignored. Proposition~\ref{prop-Pi-Sg} now allows us to
identify $\Sg$ with $\Pi$.
\begin{checks}
hyperbolic/HX_check.mpl: check_H_F1()
\end{checks}
\end{proof}
\begin{remark}\label{rem-move-inwards}
We can make a more constructive statement as follows. Any conformal
automorphism of $\Dl$ has the form $\gm(z)=\lm(z-\al)/(1-\ov{\al}z)$
for some $\lm,\al$ with $|\lm|=1$ and $|\al|<1$. Suppose that
$\al\neq 0$. One can check that
\[ |\gm(z)|^2 - |z|^2 =
\left|\frac{\al}{1-\ov{\al}z}\right|^2
(1-|z|^2) \left(|z-\ov{\al}^{-1}|^2-(|\al|^{-2}-1)\right).
\]
\begin{checks}
hyperbolic/HX_check.mpl: check_move_inwards()
\end{checks}
This means that $|\gm(z)|=|z|$ if and only if
$|z-\ov{\al}^{-1}|=\sqrt{|\al|^{-2}-1}$, and this locus describes a circle
that cuts $\partial\Dl$ at right angles, or in other words, a
hyperbolic geodesic. Now put
\[ B = \{\bt_0,\bt_2,\bt_4,\bt_6,
\bt_0\bt_7,\bt_1\bt_2,\bt_2\bt_1,\bt_3\bt_4,
\bt_4\bt_3,\bt_5\bt_6,\bt_6\bt_5,\bt_7\bt_0\}.
\]
Using the above analysis one can check that
\[ HF_1(b) = \{z\in\Dl\st |z|\leq |\gm(z)| \text{ for all } \gm\in B\}.
\]
(Note that there is one element of $B$ for each of the twelve sides
of $HF_1(b)$.) Now suppose we have a point $z_0\in\Dl$ that lies
outside $HF_1(b)$. We can choose $\gm_0\in B$ such that the point
$z_1=\gm_0(z_0)$ has $|z_1|<|z_0|$. If $z_1$ still does not lie in
$HF_1(b)$, we can choose $\gm_1\in B$ such that the point
$z_2=\gm_1(z_1)$ has $|z_2|<|z_1|$, and so on. As the orbit
$\Pi z_0$ is discrete, it can only contain finitely many points of
absolute value less than or equal to $|z_0|$, so this process will
terminate after a finite number of steps. This gives us a point
$z_n\in HF_1(b)\cap\Pi z_0$, which will be unique unless it lies on
the boundary of $HF_1(b)$. This algorithm is implemented by the
function \mcode+retract_F1_H0_aux+, defined in
\fname+hyperbolic/HX0.mpl+.
\end{remark}
\begin{corollary}\label{cor-bound-i}
For $\gm\in\Pi\sm\{1\}$ we have $1/\sqrt{2}\leq|\gm(0)|<1$.
\end{corollary}
\begin{proof}
Because $0$ lies in the interior of $F$ we see that the orbit $\Pi.0$
is discrete. We can thus choose $\gm\in\Pi\sm\{1\}$ such that
$|\gm(0)|$ is minimal. Because $F$ is a fundamental domain, it is
clear that $\gm(0)\not\in F$. Thus, by
Remark~\ref{rem-move-inwards}, there is an element $\dl\in B$
such that $|\dl\gm(0)|<|\gm(0)|$. By our choice of $\gm$, we must
have $\dl\gm=1$, so $\gm=\dl^{-1}$. As $B$ is closed under
inversion, we must have $\gm\in B$. From the definitions we see
that when $\gm\in B$ we have
\[ |\gm(0)|^2 \in \left\{
\frac{1}{2}\left(1+\frac{1-b^2}{1+b^2}\right),
\frac{4}{5}\left(1+\frac{b^2(3-b^2)}{5+2b^2+b^4}\right)
\right\}.
\]
From these expressions it is clear that $|\gm(0)|^2\geq 1/2$, as
required.
\begin{checks}
hyperbolic/HX_check.mpl: check_Pi_bound()
\end{checks}
\end{proof}
\begin{corollary}\label{cor-bound-ii}
Consider an element $\gm\in\Pi\sm\{1\}$, given by
$\gm(z)=(az+b)/(cz+d)$ with $ad-bc=1$. Then
\begin{align*}
|c| &\geq 1 \\
1 &< |d/c|\leq \sqrt{2} \\
|cz+d| &\leq (1+\sqrt{2})|c|.
\end{align*}
\end{corollary}
\begin{proof}
As $\gm$ preserves $\Dl$, it can be written in the form
$\gm(z)=\mu^2(z+\al)/(\ov{\al}z+1)$ with $|\al|<1$ and $|\mu|=1$.
This can be rewritten as $\gm(z)=(az+b)/(cz+d)$, where
\begin{align*}
a &= \mu/\sqrt{1-|\al|^2} &
b &= \mu\al/\sqrt{1-|\al|^2} \\
c &= \ov{\mu}\ov{\al}/\sqrt{1-|\al|^2} &
d &= \ov{\mu}/\sqrt{1-|\al|^2},
\end{align*}
and then $ad-bc=1$.
As $\gm(0)=\mu^2\al$, Corollary~\ref{cor-bound-i} tells us that
$1/\sqrt{2}\leq|\al|<1$. Now $|c|=|\al|/\sqrt{1-|\al|^2}$, which is
a strictly increasing function of $|\al|$, and is equal to $1$ when
$|\al|=1/\sqrt{2}$. We deduce that $|c|\geq 1$ as claimed.
Similarly, we have $|d/c|=1/|\al|\in(1,\sqrt{3}]$. Finally, we have
\[ |cz+d|/|c|=|z+d/c| \leq |z|+|d/c| < 1 + \sqrt{2}, \]
so $|cz+d|\leq (1+\sqrt{2})c$ as claimed.
\end{proof}
\subsection{Fundamental domains}
\label{sec-H-fundamental}
\begin{definition}\label{defn-HF-sixteen}
We define $HF_{16}(b)$ to be the region indicated below:
\begin{center}
\begin{tikzpicture}[scale=8]
\draw[blue] (0,0) -- (0.5,0);
\draw[green] (0,0) -- ( 0.426, 0.426);
\draw[cyan] (0.800,0.800) +(-135:0.529) arc(-135:-115:0.529);
\draw[magenta] (1.250,0.000) +(155:0.750) arc(155:180:0.750);
\fill[black](0.000,0.000) circle(0.004);
\fill[black](0.426,0.426) circle(0.004);
\fill[black](0.500,0.000) circle(0.004);
\fill[black](0.573,0.322) circle(0.004);
\draw( 0.000,-0.030) node{$v_{0}$};
\draw( 0.603, 0.312) node{$v_{3}$};
\draw( 0.386, 0.426) node{$v_{6}$};
\draw( 0.500,-0.030) node{$v_{11}$};
\draw( 0.505, 0.385) node{$c_{0}$};
\draw( 0.200, 0.250) node{$c_{1}$};
\draw( 0.556, 0.160) node{$c_{3}$};
\draw( 0.250,-0.030) node{$c_{5}$};
\end{tikzpicture}
\end{center}
\end{definition}
\begin{proposition}\label{prop-HF-sixteen}
$HF_{16}(b)$ is a fundamental domain for the action of $\tPi$ on
$\Dl$. Similarly, the image in $HX(b)$ is a fundamental domain for
the action of $G$ on $HX(b)$.
\end{proposition}
\begin{proof}
First put
\[ HF_8(b) = \{z\in HF_1(b)\st 0\leq\arg(z)\leq\pi/4\}. \]
By inspecting the following picture, we see that
\begin{align*}
HF_{16}(b)\cup \lm\nu\mu(HF_{16}(b)) &= HF_8(b) \\
HF_{16}(b)\cap \lm\nu\mu(HF_{16}(b)) &= c_0([\pi/4,\pi/2]).
\end{align*}
\begin{center}
\begin{tikzpicture}[scale=8]
\draw[blue] (0,0) -- (0.5,0);
\draw[green] (0,0) -- ( 0.625, 0.625);
\draw[cyan] (0.800,0.800) +(-135:0.529) arc(-135:-115:0.529);
\draw[magenta] (1.250,0.000) +(139:0.750) arc(139:180:0.750);
\draw[blue] (0.800,0.625) +(-180:0.175) arc(-180:-130:0.175);
\fill[black](0.000,0.000) circle(0.004);
\fill[black](0.625,0.625) circle(0.004);
\fill[black](0.426,0.426) circle(0.004);
\fill[black](0.500,0.000) circle(0.004);
\fill[black](0.685,0.493) circle(0.004);
\fill[black](0.573,0.322) circle(0.004);
\draw( 0.000,-0.030) node{$v_{0}$};
\draw( 0.625, 0.650) node{$v_{1}$};
\draw( 0.603, 0.312) node{$v_{3}$};
\draw( 0.386, 0.426) node{$v_{6}$};
\draw( 0.500,-0.030) node{$v_{11}$};
\draw( 0.715, 0.493) node{$v_{13}$};
\draw( 0.490, 0.340) node{$c_{0}$};
\draw( 0.200, 0.250) node{$c_{1}$};
\draw( 0.556, 0.160) node{$c_{3}$};
\draw( 0.250,-0.030) node{$c_{5}$};
\draw( 0.670, 0.550) node{$c_{7}$};
\end{tikzpicture}
\end{center}
Now put
\begin{align*}
T_8 &= \{ 1,\lm,\lm^2,\lm^3,\nu,\lm\nu,\lm^2\nu,\lm^3\nu \} \\
T_{16} &= T_8 \amalg \{\tau\lm\mu\nu\st \tau\in T_8\}.
\end{align*}
It is easy to check that
\[ \bigcup_{\tau\in T_{16}}HF_{16}(b) =
\bigcup_{\tau\in T_8}HF_8(b) = HF_1(b).
\]
It is also easy to check that the homomorphism $\pi\:\tPi\to G$
restricts to give a bijection $T_{16}\to G$, so that
\[ \tPi=\coprod_{\tau\in T_{16}} \Pi\tau. \]
We also know that $HF_1(b)$ is a fundamental domain for $\Pi$, so
$\Dl=\bigcup_{\phi\in\Pi}\phi(HF_1(b))$. Putting this together, we
deduce that $\Dl=\bigcup_{\phi\in\tPi}\phi(HF_{16}(b))$.
Now consider an element $\phi\in\tPi\sm\{1\}$ and a point
$z\in HF_{16}(b)$ such that $\phi(z)$ also lies in $HF_{16}(b)$. We
can write $\phi=\psi\tau$ with $\psi\in\Pi$ and $\tau\in T_{16}$.
Now the point $z'=\tau(z)$ lies in $HF_1(b)$, so the point
$z''=\phi(z)=\psi(z')$ lies in $\psi(HF_1(b))$, but it also lies in
$HF_{16}(b)$ by assumption, and one can see directly that
$HF_{16}(b)$ is in the interior of $HF_1(b)$. As $HF_1(b)$ is a
fundamental domain for $\Pi$, this can only be consistent if
$\psi=1$. This means that
$\tau(z)\in HF_{16}(b)\cap\tau(HF_{16}(b))$, and the action of
$T_{16}$ is simple enough that we can just do a check of cases to
show that $\tau(z)\in\partial(HF_{16}(b))$. This proves that
$HF_{16}(b)$ is a fundamental domain as claimed.
\end{proof}
\begin{proposition}\label{prop-square-diffeo-H}
$HF_{16}(b)$ is homeomorphic to the unit square.
\end{proposition}
\begin{proof}
This is visually obvious, but we will outline a construction of an
explicit homeomorphism. Suppose we have two disjoint geodesics in
$\Dl$, the first with endpoints $a_1,a_2\in S^1$, and the second with
endpoints $a_3,a_4\in S^1$. It is easy to produce a M\"obius
transformation $m$ that sends $\Dl$ to the upper half-plane and
$\{a_1,a_2\}$ to $\{1,-1\}$. This means that $m$ sends our first
geodesic to the upper half of the unit circle, and our second
geodesic to another semicircle that crosses the real line
orthogonally at $m(a_3)$ and $m(a_4)$. As the two geodesics are
disjoint, the second circle must either be wholly inside or wholly
outside the unit circle. By composing $m$ with $z\mapsto -1/z$ if
necessary, we may assume that it lies outside. There is in fact a
one-parameter family of choices of $m$ that have the properties
mentioned so far, and one can check that there is precisely one
choice for which $m(a_3)+m(a_4)=0$. With this condition, we see that
$m$ sends the second geodesic to a semicircle centred at the origin,
of radius $r>1$ say. Now the function $p(z)=\log|m(z)|/\log(r)$
takes the values $0$ and $1$ on our two geodesics. In the cases of
interest it is convenient to adjust this procedure slightly. Rather
than explaining the intermediate steps, we just describe the
outcome. We put
\begin{align*}
\zt_1 &= \sqrt{\frac{i-b}{i+b}} = \frac{1+ib}{b_+} \in S^1 &
\zt_2 &= \frac{b_++ib_-}{1+i} \in S^1 \\
r_1 &= \sqrt{\frac{1-b}{1+b}} = \frac{b_-}{1+b} \in\R^+ &
r_2 &= \frac{b_++b_-}{\rt b} \in\R^+ \\
m_1(z) &= \frac{\zt_1 - z}{1 - \zt_1 z} &
m_2(z) &= \frac{i\zt_2 - z}{1 - \zt_2 z} \\
p_1(z) &= \log|m_1(z)|/\log(r_1) &
p_2(z) &= \log|m_2(z)|/\log(r_2),
\end{align*}
then $p(z)=(1-p_1(z),p_2(z))$. This defines a map
$p\:HF_{16}(b)\to [0,1]^2$, with boundary behaviour as discussed in
Section~\ref{sec-fundamental}. Recall that M\"obius transformations
send circles to circles (provided that we interpret straight lines as
circles of infinite radius). It follows that for any point
$t\in[0,1]^2$, the fibres $p_1^{-1}\{1-t_1\}$ and $p_2^{-1}\{t_2\}$
are circles, with centres $c_1$ and $c_2$ say. It is not too hard to
obtain formulae for $c_1$ and $c_2$; the observation that
$m_1=m_1^{-1}$ and $m_2=m_2^{-1}$ is helpful for this. The relevant
circles must cross the unit circle orthogonally, so the radii are
$\sqrt{|c_1|^2-1}$ and $\sqrt{|c_2|^2-1}$. The point $p^{-1}(t)$
lies on the intersection of the two circles, and so can be found by
an exercise in coordinate geometry. After some simplification we
arrive at the following formulae:
\begin{align*}
s_1 &= r_1^{1-t_1} &
s_2 &= r_2^{t_2} \\
c_1 &= \frac{s_1^2/\zt_1-\zt_1}{s_1^2-1} &
c_2 &= \frac{s_2^2/\zt_2-i\zt_2}{s_2^2-1} \\
\al &= \text{Re}\left(\frac{c_1(\ov{c_1}-\ov{c_2})}{|c_1-c_2|^2}\right) &
\bt &= \sqrt{\frac{|c_1|^2-1}{|c_1-c_2|^2}-\al^2},
\end{align*}
then
\[ p^{-1}(t) = (1-\al+i\bt)c_1 + (\al-i\bt)c_2. \]
Unfortunately this formula is not meaningful when $t_1=1$ or $t_2=0$,
but those cases can be handled in an \emph{ad hoc} way.
\begin{checks}
hyperbolic/HX_check.mpl: check_square_diffeo_H()
\end{checks}
\end{proof}
\subsection{The hyperbolic family is universal}
\label{sec-H-universal}
\begin{theorem}\label{thm-H-universal}
Let $X$ be a cromulent surface. Then there is a unique number
$b\in(0,1)$ such that $X$ is isomorphic to $HX(b)$, and the
isomorphism is also unique.
\end{theorem}
The rest of this section will constitute the proof. The threads will
be gathered together in Proposition~\ref{prop-H-universal}.
Theorem~\ref{thm-classify-cromulent} says that every cromulent surface
is isomorphic to $PX(a)$ for some $a$, so we may assume that
$X=PX(a)$. We therefore have a curve system as in
Definition~\ref{defn-curve-system}, and a fundamental domain
$PF_{16}(a)$ as in Proposition~\ref{prop-P-fundamental}.
\begin{definition}\label{defn-lm-nu-Dl}
As before we define $\lm,\nu\:\Dl\to\Dl$ by $\lm(z)=iz$ and
$\nu(z)=\ov{z}$. These maps clearly satisfy
$\lm^4=\nu^2=(\lm\nu)^2=1$.
\end{definition}
\begin{lemma}\label{lem-p-Dl-X}
There is a unique covering map $p\:\Dl\to X$ such that $p(0)=v_0$ and
$p'(0)$ is a positive multiple of $c'_5(0)$. Moreover, this
satisfies $\lm p=p\lm$ and $\nu p=p\nu$.
\end{lemma}
\begin{proof}
The uniformization theorem for Riemann surfaces shows that the
universal cover of $X$ is conformally equivalent to $\Dl$, so we can
choose a covering map $p_0\:\Dl\to X$. We can then choose a point
$a\in\Dl$ with $p_0(a)=v_0$, and put $m(z)=(z+a)/(1+\ov{a}z)$. Then
$m$ is an automorphism of $\Dl$ with $m(0)=a$, so the composite
$p_1=p_0\circ m$ is a covering with $p_1(0)=v_0$. Now
$p'_1(0)=re^{i\tht}\,c'_5(0)$ for some $r>0$ and $\tht\in\R$, and the
map $p(z)=p_1(z/e^{i\tht})$ is a covering with the required
properties. If $q$ is another such map, then standard theory of
coverings gives an automorphism $f\:\Dl\to\Dl$ such that $q=pf$. Our
condition on $p$ and $q$ implies that $f(0)=0$ and $f'(0)>0$.
However, any conformal automorphism of $\Dl$ has the form
$f(z)=\al(z-\bt)/(1-\ov{\bt}z)$ for some $\al,\bt$ with $|\al|=1$ and
$|\bt|<1$. The condition $f(0)=0$ gives $\bt=0$, and then the
condition $f'(0)>0$ gives $\al=1$, so $f$ is the identity and $p=q$
as claimed.
The maps $\lm^{-1}p\lm$ and $\nu^{-1}p\nu$ are easily seen to satisfy
the defining conditions for $p$, so $\lm p=p\lm$ and $\nu p=p\nu$.
\end{proof}
\begin{proposition}\label{prop-classify-a}
There is a unique system of points $v_i^*\in\Dl$ (for $0\leq i\leq
13$) and continuous maps $c_j^*\:\R\to\Dl$ (for $0\leq j\leq 8$) such
that the following hold:
\begin{itemize}
\item[(a)] $v_0^*=0$
\item[(b)] For all $i$ we have $p(v_i^*)=v_i$, and for all $j$ and
$t$ we have $p(c_j^*(t))=c_j(t)$.
\item[(c)] Whenever a number $t$ appears in the $i$'th column of the
$j$'th row of the table below, we have $c^*_j(t)=v_i^*$.
\[ \begin{array}{|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
&0&1&2&3&4&5&6&7&8&9&10&11&12&13\\ \hline
0&&&0&\ppi&&&\tfrac{\pi}{4}&&&&&&&\\ \hline
1&0&\pi&&&&&\ppi&&-\ppi&&&&&\\ \hline
2&0&&&&&&&\ppi&&-\ppi&&&&\\ \hline
3&&&&\ppi&&-\ppi&&&&&&0&&\pi\\ \hline
4&&&-\ppi&&\ppi&&&&&&0&&-\pi&\\ \hline
5&0&&&&&&&&&&&\pi&&\\ \hline
6&0&&&&&&&&&&\pi&&&\\ \hline
7&&0&&&&&&&&&&&&\\ \hline
8&&0&&&&&&&&&&&&\\ \hline
\end{array} \]
\item[(d)] The action of $\lm$ and $\nu$ on the points $v_i^*$ is
partially given by the following table:
\[ \begin{array}{|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
&v_0^*&v^*_1&v^*_2&v^*_3&v^*_4&v^*_5&v^*_6&v^*_7&v^*_8&v^*_9&
v^*_{10}&v^*_{11}&v^*_{12}&v^*_{13}\\ \hline
\lm & v^*_0 & & & v^*_4 & & v^*_3 &
v^*_7 & v^*_8 & v^*_9 & v^*_6 & & v^*_{10} & & \\ \hline
\nu & v^*_0 & & & v^*_5 & & v^*_3 &
v^*_9 & v^*_8 & v^*_7 & v^*_6 & & v^*_{11} & & \\ \hline
\lm\nu & v^*_0 & v^*_1 & v^*_3 & v^*_2 & v^*_5 & v^*_4 &
v^*_6 & v^*_9 & v^*_8 & v^*_7 &
v^*_{11} & v^*_{10} & v^*_{13} & v^*_{12} \\ \hline
\end{array} \]
\item[(e)] The action of $\lm$ and $\nu$ on the curves $c_j^*$ is
partially given by the following table:
\[ \begin{array}{|c|c|c|c|c|c|c|c|c|c|}
\hline
&c_0^*(t)&c^*_1(t)&c^*_2(t)&c^*_3(t)&c^*_4(t)&c^*_5(t)&c^*_6(t)&c^*_7(t)&c^*_8(t)\\ \hline
\lm & & c_2^*(t) & c_1^*(-t) & c_4^*(t) & &
c_6^*(t) & c_5^*(-t) & & \\ \hline
\nu & & c_2^*(-t) & c_1^*(-t) & c_3^*(-t) & &
c_5^*(t) & c_6^*(-t) & & \\ \hline
\lm\nu & c_0^*(\pi/2-t) & c_1^*(t) & c_2^*(-t) &
c_4^*(-t) & c_3^*(-t) & c_6^*(t) & c_5^*(t) &
c_8^*(t) & c_7^*(t) \\ \hline
\end{array} \]
\end{itemize}
\end{proposition}
\begin{proof}\leavevmode
\begin{itemize}
\item[(0)] We define $v^*_0=0$ and note that this is fixed by $\lm$
and $\nu$, and that $p(v_0^*)=v_0$ by the definition of $p$.
\item[(1)] For $j\in\{1,2,5,6\}$ we define $c_j^*$ to be the unique
continuous lift of $c_j$ that satisfies $c^*_j(0)=v^*_0$. The
claimed formulae for $\lm(c^*_j(t))$ and $\nu(c^*_j(t))$ then hold
by an evident uniqueness argument. For example, we have seen
previously that $\lm(c_2(t))=c_1(-t)$ in $PX(a)$, so the maps
$t\mapsto\lm(c_2^*(t))$ and $t\mapsto c^*_1(-t)$ are both
continuous lifts of the map $t\mapsto c_1(-t)$. Both give $v^*_0$
when $t=0$, so they must agree for all $t$.
\item[(2)] We define
\begin{align*}
v^*_1 &= c_1^*(\pi) &
v^*_6 &= c_1^*(\pi/2) &
v^*_8 &= c_1^*(-\pi/2) \\
& &
v^*_7 &= c_2^*(\pi/2) &
v^*_9 &= c_2^*(-\pi/2) \\
v^*_{10} &= c^*_6(\pi) \\
v^*_{11} &= c^*_5(\pi).
\end{align*}
By taking $t=\pi$ or $t=\pm\pi/2$ in~(1) we see that $\lm$ and
$\nu$ act on $\{v^*_1,v^*_6,v^*_7,v^*_8,v^*_9,v^*_{10},v^*_{11}\}$
as indicated in~(d). Because the maps $c_j\:\R\to PX(a)$ form a
curve system, we see that $p(v^*_i)=v_i$ in all these cases.
\item[(3)] For $j\in\{3,4,7,8\}$ we let $c^*_j$ denote the unique
lift of $c_j(t)$ satisfying the following initial condition:
\[ c_3^*(0) = v_{11}^* \hspace{4em}
c_4^*(0) = v_{10}^* \hspace{4em}
c_7^*(0) = c_8^*(0) = v_1^*.
\]
The claimed formulae for $\lm(c^*_j(t))$ and $\nu(c^*_j(t))$ then
hold by an evident uniqueness argument.
\item[(4)] We define
\begin{align*}
v^*_{13} &= c_3^*(\pi) &
v^*_3 &= c_3^*(\pi/2) &
v^*_5 &= c_3^*(-\pi/2) \\
v^*_{12} &= c_4^*(-\pi) &
v^*_4 &= c_4^*(\pi/2) &
v^*_2 &= c_4^*(-\pi/2).
\end{align*}
By taking $t=\pi$ or $t=\pm\pi/2$ in~(3) we see that $\lm$ and
$\nu$ act on $\{v^*_2,v^*_3,v^*_4,v^*_5,v^*_{12},v^*_{13}\}$
as indicated in~(d). Because the maps $c_j\:\R\to PX(a)$ form a
curve system, we see that $p(v^*_i)=v_i$ in all these cases.
\item[(5)] We let $c^*_0$ denote the unique lift of $c_0$ satisfying
$c^*_0(\pi/4)=v^*_6$. By the usual uniqueness argument, so have
$\lm\nu(c^*_0(t))=c^*_0(\pi/2-t)$.
\item[(6)] Now all of~(a), (b) and~(c) is true by construction
except for the identities $c^*_0(0)=v^*_2$ and
$c^*_0(\pi/2)=v^*_3$. All parts of~(d) and~(e) have also been
established. For the remaining facts, we consider the fundamental
domain $PF_{16}(a)$ in $PX(a)$. We have two paths in $PF_{16}(a)$
from $v_0$ to $v_3$: one given by $c_5([0,\pi])$ followed by
$c_3([0,\pi/2])$, and the other by $c_1([0,\pi/2])$ followed by
$c_0([\pi/4,\pi/2])$. If we lift the first path starting with
$v^*_0$ then the endpoint is $c^*_3(\pi/2)=v^*_3$, and if we lift
the second then the endpoint is $c_0^*(\pi/2)$. Now $PF_{16}(a)$
is homeomorphic to a square and so is contractible. In particular,
our two paths are homotopic relative to the endpoints, and it
follows that the two lifts have the same endpoints, so
$c_0^*(\pi/2)=v_3$ as claimed. By applying the map $\lm\nu$ we
deduce that $c_0^*(0)=v_2$.
\end{itemize}
\end{proof}
\begin{proposition}
For $0\leq j\leq 8$ there is an antiholomorphic involution
$\tht_j\:\Dl\to\Dl$ such that $c_j^*$ gives a diffeomorphism from
$\R$ to the geodesic $C^*_j=\{z\in\Dl\st\tht_j(z)=z\}$.
Specifically, for $j\in\{1,2,5,6\}$ we have
\begin{align*}
\tht_1 &= \lm\nu & C^*_1 &= (-1,1).e^{ i\pi/4} \\
\tht_2 &= \lm^3\nu & C^*_2 &= (-1,1).e^{-i\pi/4} \\
\tht_5 &= \nu & C^*_5 &= (-1,1) \\
\tht_6 &= \lm^2\nu & C^*_6 &= (-1,1).i.
\end{align*}
\end{proposition}
\begin{proof}
Put $C_j=c_j(\R)\subset PX(a)$. We have seen previously that in each
case there is an antiholomorphic involution $\rho_j\in G$ such that
$\rho_j(c_j(t))=c_j(t)$ for all $t\in\R$, and in fact $C_j$ is a
connected component of the set
$PX(a)^{\rho_j}=\{z\in PX(a)\st\rho_j(z)=z\}$. Moreover, $C_j$ is
diffeomorphic to $S^1$ and the map $c_j\:\R\to C_j$ is a universal
covering.
Next, standard covering theory shows that there is a
unique continuous map $\tht_j\:\Dl\to\Dl$ with $p\tht_j=\rho_jp$ and
$\tht_j(c^*_j(0))=c_j^*(0)$. As $p$ is a holomorphic covering, the
equation $p\tht_j=\rho_jp$ implies that $\tht_j$ is antiholomorphic.
The map $\tht_j^2$ covers $\rho_j^2=1$ and fixes $c_j^*(0)$; it
follows that $\tht_j^2=1$. Now $c^*_j$ and $\rho_j\circ c^*_j$ are
both lifts of $c_j$ with the same value at $t=0$, so they must be the
same, so $c_j^*(\R)\sse C_j^*$. We previously classified the
antiholomorphic involutions on $\Dl$, and using that classification
we see that $C_j^*$ is a geodesic in $\Dl$ and is diffeomorphic to
$\R$. It follows that $p(C_j^*)$ is a connected subset of
$PX(a)^{\rho_j}$ containing $p(c^*_j(0))=c_j(0)$, so
$p(C_j^*)\sse C_j$. Now $p$ is a proper map with nonzero complex
derivative everywhere in $\Dl$. It follows that $p\:C_j^*\to C_j$ is
also proper with nonzero real derivative everywhere. This means that
$p\:C_j^*\to C_j$ is another universal covering, and by the
uniqueness of universal coverings, we see that $c^*_j\:\R\to C_j^*$
must be a diffeomorphism.
For the case $j=5$, we have seen that $\nu(c_5^*(t))=c_5^*(t)$ for
all $t$ and it follows that $\tht_5=\nu$. We also have
$\nu(z)=\ov{z}$ so $C_5^*=(-1,1)$. The cases $j\in\{1,2,6\}$ are
similar.
\end{proof}
\begin{lemma}\label{lem-Dl-involutions}
For any point $v\in\Dl$, there is a unique holomorphic involution on
$\Dl$ that fixes $v$.
\end{lemma}
\begin{proof}
As $\Aut(\Dl)$ acts transitively on $\Dl$, we may assume that $v=0$.
In this case the map $z\mapsto -z$ is a holomorphic involution that
fixes $v$. Let $\phi$ be any other holomorphic involution that
fixes $v$. As $\phi$ is an automorphism of $\Dl$ we have
$\phi(z)=\lm(z-\al)/(1-\ov{\al}z)$ for some $\lm,\al$ with $|\lm|=1$
and $|\al|<1$. As $\phi$ fixes $v=0$ we have $\al=0$, so
$\phi(z)=\lm z$. As $\phi$ is an involution we must have $\lm=-1$.
\end{proof}
\begin{proposition}\label{prop-Dl-kp}
Let $\kp$ be the unique holomorphic involution on $\Dl$ that fixes
$v^*_6$. Then $p\kp=\lm\mu p$, and the action on the points $v^*_i$
is partially given by the following table:
\[ \begin{array}{|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
&v_0^*&v^*_1&v^*_2&v^*_3&v^*_4&v^*_5&v^*_6&v^*_7&v^*_8&v^*_9&
v^*_{10}&v^*_{11}&v^*_{12}&v^*_{13}\\ \hline
\kp & v^*_1 & v^*_0 & v^*_3 & v^*_2 & & &
v^*_6 & & & & v^*_{13} & v^*_{12} & v^*_{11}& v^*_{10}\\ \hline
\end{array} \]
Also, the action on the curves $c^*_j$ is partially given by the
following table:
\[ \begin{array}{|c|c|c|c|c|c|c|c|c|c|}
\hline
&c_0^*(t)&c^*_1(t)&c^*_2(t)&c^*_3(t)&c^*_4(t)&
c^*_5(t)&c^*_6(t)&c^*_7(t)&c^*_8(t)\\ \hline
\kp & c_0^*(\pi/2-t) & c_1^*(\pi-t) & &
c_4^*(t-\pi) & c_3^*(t+\pi) &
c_8^*(t) & c_7^*(t) & c_6^*(t) & c_5^*(t) \\ \hline
\end{array} \]
\end{proposition}
\begin{proof}
Given a point $z\in\Dl$, we let $u$ be any path in $\Dl$ from $v_6^*$
to $z$ in $\Dl$. We recall that $\lm\mu(v_6)=v_6$, so the path
$\lm\mu p u$ starts at $p(v^*_6)$, so there is a unique lifting $u'$
of $\lm\mu p u$ with $u'(0)=v^*_6$. We define $\kp(z)=u'(1)$.
Standard covering theory shows that this is independent of the choice
of $u$ and gives the unique continuous map $\kp\:\Dl\to\Dl$ with
$\kp(v^*_6)=v^*_6$ and $p\kp=\lm\mu p$. As $p$ is a holomorphic
covering and holomorphy can be checked locally it follows that $\kp$
is holomorphic. The map $\kp^2$ covers $(\lm\mu)^2=1$ and fixes
$v^*_6$; it follows that $\kp^2=1$. Thus, $\kp$ is the unique
holomorphic involution on $\Dl$ that fixes $v^*_6$.
The curves $\kp(c_0^*(t))$ and $c_0^*(\pi/2-t)$ both lift
$c_0(\pi/2-t)$ and pass through $v^*_6$ at $t=\pi/4$, so they are the
same. Taking $t\in\{0,\pi/2\}$ we deduce that $\kp$ exchanges
$v^*_2$ and $v^*_3$. Essentially the same argument gives
$\kp(c^*_1(t))=c^*_1(\pi-t)$, and shows that $\kp$ exchanges $v^*_0$
and $v^*_1$. Now that we know the action on the point
$v^*_3=c^*_3(\pi/2)$ we can see that the curves $\kp(c^*_3(t))$ and
$c^*_4(t-\pi)$ both lift $c_4(t-\pi)$ and pass through $v^*_2$ at
$t=\pi/2$ so they are the same. As $\kp^2=1$ we can also deduce that
$\kp(c^*_4(t))=c^*_3(t+\pi)$. By taking $t=0$ or $t=\pm\pi$ we
deduce that $\kp$ exchanges $v^*_{10}$ and $v^*_{13}$, and also
exchanges $v^*_{11}$ and $v^*_{12}$. This just leaves the action on
$c^*_j(t)$ for $j\in\{5,6,7,8\}$, which can be checked in the same
way using $v^*_0$ and $v^*_1$ as basepoints.
\end{proof}
\begin{lemma}\label{lem-v-eleven-sign}
$v^*_{11}$ is a positive real number, and $v^*_6$ is a positive
multiple of $e^{i\pi/4}=(1+i)/\rt$.
\end{lemma}
\begin{proof}
We have seen that $c^*_5$ gives a diffeomorphism from $\R$ to
$(-1,1)$. We also have $p(c^*_5(t))=c_5(t)$ so
$(c_5^*)'(0)=c'_5(0)/p'(0)$, and this is a positive real number by
the definition of $p$. As $c^*_5$ is a diffeomorphism the derivative
cannot change sign, so it is a strictly increasing map. It follows
that the point $v^*_{11}=c^*_5(\pi)$ lies on the positive real axis
as claimed.
Next, we also know that $c^*_1$ gives a diffeomorphism from $\R$
to $(-1,1).e^{i\pi/4}$. By examining the formula in
Definition~\ref{defn-P-curves} we see that to first order in $t$ we
have
\begin{align*}
c_1(t) &= [t e^{i\pi/4}/2:1:0:0] \\
c_5(t) &= [t \sqrt{a}/2:1:0:0],
\end{align*}
so $c'_1(0)$ is a positive multiple of $e^{i\pi/4}\,c'_5(0)$. Using
this we see that $c^*_5$ must carry $(0,\infty)$ to
$(0,1).e^{i\pi/4}$. In particular, the point $v^*_6=c^*_1(\pi/2)$ is
a positive multiple of $e^{i\pi/4}$ as claimed.
\end{proof}
\begin{lemma}\label{lem-right-circles}
Suppose that two circles in $\R^2$ meet at right angles. Let $r_1$
and $r_2$ be the radii, and let $d$ be the distance between the
centres; then $d^2=r_1^2+r_2^2$.
\end{lemma}
\begin{proof}
Elementary.
\end{proof}
\begin{proposition}\label{prop-b-H}
There is a unique number $b\in(0,1)$ such that
\begin{align*}
v^*_0 &= 0 &
v^*_1 &= \frac{1+i}{2}b_+ \\
v^*_2 &= \frac{b\,b_--b_+}{i-b^2} &
v^*_5 &= -i \, v^*_2 \\
v^*_3 &= \frac{b\,b_--b_+}{ib^2-1} &
v^*_4 &= i \, v^*_3 \\
v^*_6 &= \frac{1+i}{\rt}\;\frac{\rt-b_-}{b_+} &
v^*_7 &= i \, v^*_6 \\
v^*_8 &= -v^*_6 &
v^*_9 &= -i\, v^*_6 \\
v^*_{10} &= i(b_+-b) &
v^*_{11} &= b_+-b \\
v^*_{12} &= (b+b_+)\frac{i+(i+2)b^2}{(b+b_+)^2+b^2} &
v^*_{13} &= i\,\ov{v^*_{12}}
\end{align*}
(where $b_{\pm}=\sqrt{1\pm b^2}$ as before). Moreover, the map $\kp$
is given by
\[ \kp(z) = \frac{b_+i - (1+i)z}{1+i - b_+z}. \]
\end{proposition}
\begin{proof}
We will use the curves $C_j^*\subset\Dl$ for $j\in\{0,1,3,5\}$. We
have already seen that $C_5^*=(-1,1)$ and $C_1^*=(-1,1).e^{i\pi/4}$.
The set $C_3^*$ is a geodesic in $\Dl$ that does not pass through the
origin, so it is the intersection of $\Dl$ with a circle centred
outside $\Dl$ that crosses $\partial\Dl$ at right angles. (This is a
standard fact of hyperbolic geometry.) We let $b$ denote the radius
of $C_3^*$ (so $b>0$).
We next claim that
\begin{itemize}
\item[(a)] The curves $C_0^*$ and $C_3^*$ cross at right angles at $v_3^*$
\item[(b)] The curves $C_0^*$ and $C_1^*$ cross at right angles at $v_6^*$
\item[(c)] The curves $C_3^*$ and $C_5^*$ cross at right angles at $v_{11}^*$.
\end{itemize}
Indeed, we see from Proposition~\ref{prop-classify-a} that the
indicated curves cross at the indicated points, and that in all
relevant cases we have $p(c_i^*(t))=c_i(t)$ and $p(v_j^*)=v_j$ in
$PX(a)$. As $p$ is a holomorphic covering it preserves angles, so
the claim follows from Lemma~\ref{lem-right-angle}.
As $C^*_3$ meets the curve $C^*_5=(-1,1)$ at right angles at the
point $v^*_{11}>0$, we see that the centre of $C^*_3$ must be on the
positive real axis. As $C^*_3$ also meets $\partial\Dl$
orthogonally, Lemma~\ref{lem-right-circles} shows that the centre is
$\sqrt{1+b^2}=b_+$. It then follows that $v^*_{11}=b_+-b$.
Now put $\om=e^{i\pi/4}=(1+i)/\rt$. By a similar argument,
there is a number $c>0$ such that $\tC_0$ is a circular arc with
centre $\sqrt{1+c^2}\om$ and radius $c$, and we have
$v^*_6=(\sqrt{1+c^2}-c)\om$.
As $\tC_0$ and $\tC_3$ meet at right angles, we must have
\[ b^2+c^2 = |\sqrt{1+c^2}\om - \sqrt{1+b^2}|^2 =
\left(\frac{\sqrt{1+c^2}}{\rt} - \sqrt{1+b^2}\right)^2 +
\left(\frac{\sqrt{1+c^2}}{\rt}\right)^2.
\]
After some manipulation this gives $c^2=(1-b^2)/(1+b^2)$. This
ensures that $b<1$, and we can take square roots to get $c=b_-/b_+$.
We also get $\sqrt{1+c^2}=\rt/b_+$ and so
\[ v^*_6 = (\sqrt{1+c^2}-c)\om =
\frac{1+i}{\rt}\;\frac{\rt-b_-}{b_+}
\]
as claimed.
Now put $w=(b_+-b\,b_-)/(1-ib^2)$. One can check that
\[ 1 - |w|^2 = 2b\,b_-\,(b_+-b\,b_-)/(1+b^4) > 0, \]
so $w\in\Dl$. Long but fairly straightforward calculations also show
that $|w-b_+|^2=b^2$ and $|w-\sqrt{1+c^2}\om|^2=c^2$, so
$w\in\tC_3\cap\tC_0$. It is standard that distinct geodesics in
$\Dl$ meet in only one place, so we must have $v^*_3=w$.
Next, if we define
\[ \kp^*(z) = \frac{b_+i - (1+i)z}{1+i - b_+z}, \]
a straightforward calculation shows that this is a holomorphic
involution on $\Dl$ that fixes $v^*_6$. However,
Lemma~\ref{lem-Dl-involutions} shows that there is only one such
involution, so $\kp$ must be the same as $\kp^*$
Above we have established formulae for $v^*_0$, $v^*_3$, $v^*_6$ and
$v^*_{11}$ in terms of $b$. Propositions~\ref{prop-classify-a}
and~\ref{prop-Dl-kp} also give
\begin{align*}
v^*_1 &= \kp(v^*_0) &
v^*_2 &= \kp(v^*_3) &
v^*_4 &= \lm(v^*_3) &
v^*_5 &= \nu(v^*_3) \\
v^*_7 &= \lm(v^*_6) &
v^*_8 &= \lm^2(v^*_6) &
v^*_9 &= \lm^3(v^*_6) \\
v^*_{10} &= \lm(v_{11}) &
v^*_{12} &= \kp(v_{11}) &
v^*_{13} &= \lm\nu(v_{12}).
\end{align*}
Using these we can deduce the stated formulae for all points
$v^*_i$.
\end{proof}
From now on we use the value of $b$ coming from the previous
Proposition. This gives maps $\lm,\mu,\nu,\bt_0,\dotsc,\bt_7$
generating a group $\Pi$ as in Section~\ref{sec-H}. The maps
$\lm$ and $\nu$ from that section are of course the same as the ones
we have been using already in this section.
\begin{definition}\label{defn-tPhi}
We say that a conformal or anticonformal automorphism
$\phi\:\Dl\to\Dl$ is \emph{$G$-compatible} if there is an element
$\phi_1\in G$ such that the following diagram commutes:
\[ \xymatrix{
\Dl \ar[d]_p \ar[r]^\phi & \Dl \ar[d]^p \\
X \ar[r]_{\phi_1} & X.
} \]
We write $\tPhi$ for the group of all $G$-compatible automorphisms,
and note that the construction $\phi\mapsto\phi_1$ gives a
homomorphism $\tPhi\to G$. We write $\Phi$ for the kernel, which is
just the group of automorphisms $\phi$ satisfying $p\phi=p$, or in
other words deck transformations. Standard covering theory shows
that $p$ induces a conformal isomorphism $\Dl/\Phi\to PX(a)$.
\end{definition}
\begin{proposition}\label{prop-tPhi-gens}
The maps $\lm$, $\mu$, $\nu$ and $\kp$ are elements of $\tPhi$, with
$\pi(\lm)=\lm$ and $\pi(\mu)=\mu$ and $\pi(\nu)=\nu$ and
$\pi(\kp)=\lm\mu$.
\end{proposition}
\begin{proof}
The claims for $\lm$, $\nu$ and $\kp$ are clear by construction. By
the same argument that we used in Proposition~\ref{prop-Dl-kp}, if we
let $\mu'$ denote the unique holomorphic involution on $\Dl$ that
fixes $v_2$, then $p\mu'=\mu p$, so $\mu'\in\tPhi$ with
$\pi(\mu')=\mu$. However, if we define
\[ \mu(z) = \frac{b_+z-b^2-i}{(b^2-i)z-b_+} \]
as in Section~\ref{sec-H}, then straightforward calculation
shows that $\mu$ is a holomorphic involution with $\mu(v^*_2)=v^*_2$,
so $\mu$ is the same as $\mu'$.
\end{proof}
\begin{proposition}\label{prop-Pi-in-Phi}
We have $\Pi\leq\Phi$ and $\tPi\leq\tPhi$.
\end{proposition}
\begin{proof}
We first claim that $\bt_0\in\Phi$. Note that $\lm^2(v_{11})=v_{11}$ in
$PX(a)$, so the points $v^*_{11}=b_+-b$ and $\lm^2(v^*_{11})=b-b_+$
have the same image under $p$. Recall that
$\bt_0(z)=(b_+z+1)/(z+b_+)$; this implies that
$\bt_0(\lm^2(v^*_{11}))=v^*_{11}$, and that $\bt_0$ restricts to give
a strictly increasing automorphism of $(-1,1)$.
We originally introduced $p$ as the unique holomorphic covering map
$\Dl\to PX(a)$ such that $p(0)=v_0$ and $p'(0)$ is a positive
multiple of $c'_5(0)$. However, the same line of argument shows that
$p$ is also the unique holomorphic covering map $\Dl\to PX(a)$ such
that $p(\lm^2(v^*_{11}))=v_{11}$ and $p'(\lm^2(v^*_{11}))$ is a
positive multiple of $c'_5(-\pi)$. The composite $p\bt_0$ has these
properties, so $p\bt_0=p$, so $\bt_0\in\Phi$ as claimed.
It is also clear that $\Phi$ is normal in $\tPhi$ and $\lm\in\tPhi$
so the conjugates $\bt_{2k}=\lm^k\bt_0\lm^{-k}$ also lie in $\Phi$.
We now recall that the elements $\lm,\mu\in\tPi$ lie in $\tPhi$ and
satisfy $(\lm\mu)^2=\bt_7\bt_6$. As $(\lm\mu)^2=1$ in $G$ we can
deduce that $\bt_7\bt_6\in\Phi$. We saw above that $\bt_6\in\Phi$,
so $\bt_7\in\Phi$. Using $\lm^k\bt_7\lm^{-k}=\bt_{7+2k}$ we deduce
that $\bt_j\in\Phi$ for all $j$, so $\Pi\leq\Phi$. As
$\lm,\mu,\nu\in\tPhi$ it also follows that $\tPi\leq\tPhi$.
\end{proof}
\begin{proposition}\label{prop-H-universal}
We have $\Pi=\Phi$ and $\tPi=\tPhi$, and the map $p\:\Dl\to PX(a)$
induces an isomorphism $\ov{p}\:HX(b)=\Dl/\Pi\to PX(a)$ of cromulent
surfaces.
\end{proposition}
\begin{proof}
As $\Pi\leq\Phi$, we can factor the map $p$ as
\[ \Dl \xra{q} HX(b) = \Dl/\Pi \xra{\ov{p}} \Dl/\Phi \simeq PX(a).
\]
As $p$ and $q$ are holomorphic coverings, we see that $\ov{p}$ is
also a holomorphic covering. We have also seen that both $HX(b)$ and
$PX(a)$ are compact, so $\ov{p}$ has degree $d<\infty$ say. It
follows (by choosing compatible triangulations, say) that
$\chi(HX(b))=d\;\chi(PX(a))$ (where $\chi$ denotes the Euler
characteristic). However, both $HX(b)$ and $PX(a)$ have genus $g=2$
and therefore Euler characteristic $2-2g=-2$, so we must have $d=1$,
so $\ov{p}$ is an isomorphism of Riemann surfaces. By construction
it is $G$-equivariant and sends $v_i$ to $v_i$ so it is an
isomorphism of cromulent surfaces.
\end{proof}
\section{Relating the projective and hyperbolic families}
\label{sec-P-H}
Recall that the projective and algebraic families are both universal,
so for each $a\in(0,1)$ there is a unique $b\in (0,1)$ such that
$PX(a)$ is isomorphic (in a unique way) to $HX(b)$, and
\emph{vice-versa}. In this section we will give two different methods
for calculating this correspondence. The first method starts with $b$
and calculates $a$, and the second works in the opposite direction.
\subsection{Preliminaries}
\label{sec-P-H-prelim}
The space $HX(b)$ is by definition a quotient of $\Dl$. We have an
isomorphism $HX(b)\to PX(a)$, and an isomorphism
$PX(a)/\ip{\lm^2}\to\C_\infty$ sending $j(w,z)$ to $z$. Composing
these maps gives a map $p\:\Dl\to\C_\infty$. Our main task is to
calculate this map.
In Definition~\ref{defn-v-H} we defined certain points $v_i\in\Dl$
(depending on $b$), which become the labelled points in $HX(b)$. In
this section, we will write $v_{Hi}$ for these points. Similarly, we
will write $c_{Hj}$ for the curves $\tc_j\:\R\to\Dl$ defined in
Definition~\ref{defn-H-curves}. Moreover, the images in
$\C_\infty$ of the points $v_i\in PX(a)$ and the curves
$c_j\:\R\to P(a)$ will be denoted by $v_{Ci}$ and $c_{Cj}$. We are
primarily interested in the corners of the fundamental domain, which
we can tabulate as follows:
\begin{align*}
v_{H0} &= 0 &
v_{C0} &= 0 \\
v_{H3} &= \frac{b\,b_--b_+}{ib^2-1} &
v_{C3} &= 1 \\
v_{H6} &= \frac{1+i}{\rt}\;\frac{\rt-b_-}{b_+} &
v_{C6} &= i \\
v_{H11} &= b_+-b &
v_{C11} &= a.
\end{align*}
Because the map $HX(b)\to PX(a)$ is cromulent, we have
$p(v_{Hi})=v_{Ci}$ and $p(c_{Hj}(\R))=c_{Cj}(\R)$. In particular, we
have $v_{H0}=0$ and $v_{C0}=0$, so $p(0)=0$. Using equivariance with
respect to $\lm$, we also see that $p(iz)=-p(z)$, so $p(-z)=p(z)$, and
it follows that $p'(0)=0$. This makes it inconvenient to work with
$p$ itself. Instead, we will work with a certain map of the form
$p_1=\phi p\psi$, where $\phi\in\Aut(\C_\infty)$ and
$\psi\in\Aut(\Dl)$. This will be arranged so that $p_1(0)=0$ and
$p'_1(0)>0$. Details are as follows:
\begin{definition}\label{defn-schwarz-phi}
We define $\phi\in\Aut(\C_\infty)$ and $\psi\in\Aut(\Dl)$ by
\begin{align*}
\phi(z) &= \frac{i-z}{i+z} &
\phi^{-1}(z) &= i \frac{1-z}{1+z} \\
\psi(z) &= \frac{1+i}{\rt}
\frac{\rt-b_- - b_+z}{b_+ (b_--\rt)z} &
\psi^{-1}(z) &= -\frac{\rt-b_--(1-i)b_+z/\rt}{
(1-i)(1-b_-/\rt)z-b_+}.
\end{align*}
We then define $p_1=\phi\circ p\circ\psi\:\Dl\to\C_\infty$. We also
put $v_{HSi}=\psi^{-1}(v_{Hi})\in\Dl$ and
$v_{PSi}=\phi(v_{Ci})\in\C_\infty$, so that $p_1(v_{HSi})=v_{PSi}$.
We define curves $c_{HSj}$ and $c_{PSj}$ in the same way. Finally,
we put
\[ t = b^2\frac{\rt b_+ - 2bb_-}{1-b^2+2b^4}
\hspace{6em}
s = \frac{(b_+b_--\rt)(b-b^3)}{1-b^2+2b^4}.
\]
\end{definition}
\begin{remark}
In Maple, the maps $\phi$ and $\psi$ are \mcode+schwarz_phi+ and
\mcode+schwarz_psi+, and the inverse maps are \mcode+schwarz_phi_inv+ and
\mcode+schwarz_psi_inv+. Maple notation for $v_{HSi}$ and
$c_{HSj}(t)$ is \mcode+v_HS[i]+ and \mcode+c_HS[j](t)+, and similarly
for $v_{PSi}$ and $c_{PSj}(t)$. Maple notation for $t$ and $s$ is
\mcode+t_schwarz+ and \mcode+s_schwarz+. All of this is in
\fname+hyperbolic/schwarz.mpl+.
\end{remark}
By direct calculation, we have
\begin{align*}
v_{HS0} &= \frac{\rt-b_-}{b_+} &
v_{PS0} &= 1 \\
v_{HS3} &= i \frac{b_-}{b_++\rt b} &
v_{PS3} &= i \\
v_{HS6} &= 0 &
v_{PS6} &= 0 \\
v_{HS11} &= t+is &
v_{PS11} &= \frac{i-a}{i+a}
\end{align*}
(where $t$ and $s$ are as in Definition~\ref{defn-schwarz-phi}).
\begin{checks}
hyperbolic/schwarz_check.mpl: check_schwarz()
\end{checks}
\begin{lemma}\label{lem-psi-edges}
$\psi^{-1}$ acts as follows on the edges of $HF_{16}$:
\begin{itemize}
\item $\psi^{-1}(C_0)=(-1,1).i$
\item $\psi^{-1}(C_1)=(-1,1)$
\item $\psi^{-1}(C_3)$ is the intersection of $\Dl$ with the circle
of radius $\rt b/b_-$ centred at $ib_+/b_-$
\item $\psi^{-1}(C_5)$ is the intersection of $\Dl$ with the circle
of radius $\rt b_-/b_+$ centred at $(\rt+b_-)/b_+$.
\end{itemize}
\end{lemma}
\begin{proof}
Let $A_k$ denote the set that is claimed to be equal to
$\psi^{-1}(C_k)$. In each case, it is easy to check that $A_k$ is a
geodesic. We also know that $C_k$ is a geodesic and $\psi$ is an
isometry, so $\psi^{-1}(C_k)$ is also a geodesic. Because of this,
it will suffice to check that $|A_k\cap\psi^{-1}(C_k)|\geq 2$, which
we can do by considering the points
$\{\psi^{-1}(v_i)\st i\in\{0,3,6,11\}\}$.
\begin{checks}
hyperbolic/schwarz_check.mpl: check_schwarz()
\end{checks}
\end{proof}
We can illustrate the maps $p$, $p_1$, $\phi$ and $\psi$ as follows.
\begin{center}
\begin{tikzpicture}
\path[use as bounding box] (-0.5,-9) rectangle(16,6);
\draw[->] ( 2.0,-2.5) -- ( 2.0,-1.0);
\draw[->] (11.6,-1.0) -- (11.6,-2.5);
\draw[->] ( 5.0, 1.5) -- ( 7.0, 1.5);
\draw[->] ( 5.5,-5.5) -- ( 7.5,-5.5);
\draw ( 2.0,-1.75) node[anchor=east] {$\psi$};
\draw (11.6,-1.75) node[anchor=west] {$\phi$};
\draw ( 6.0, 1.50) node[anchor=south] {$p$};
\draw ( 6.5,-5.55) node[anchor=north] {$p_1$};
\begin{scope}[scale=6]
\draw[blue] (0,0) -- (0.500,0);
\draw[blue] (0,0) -- (0,0.500);
\draw[green] (0,0) -- ( 0.625, 0.625);
\draw[cyan] (0.800,0.800) (0.426,0.426) arc(225:245:0.529);
\draw[cyan] (0.800,0.800) (0.426,0.426) arc(225:205:0.529);
\draw[magenta] (1.250,0.000) (0.500,0.000) arc(180:139:0.750);
\draw[magenta] (0.000,1.250) (0.000,0.500) arc(270:311:0.750);
\draw[blue] (0.800,0.625) (0.625,0.625) arc( 180:230:0.175);
\draw[blue] (0.625,0.800) (0.625,0.625) arc( 270:220:0.175);
\fill[black](0.000,0.000) circle(0.006);
\fill[black](0.625,0.625) circle(0.006);
\fill[black](0.322,0.573) circle(0.006);
\fill[black](0.426,0.426) circle(0.006);
\fill[black](0.000,0.500) circle(0.006);
\fill[black](0.500,0.000) circle(0.006);
\fill[black](0.685,0.493) circle(0.006);
\fill[black](0.493,0.685) circle(0.006);
\fill[black](0.573,0.322) circle(0.006);
\draw( 0.000,-0.030) node{$\ss 0$};
\draw( 0.650, 0.620) node{$\ss 1$};
\draw( 0.292, 0.585) node{$\ss 2$};
\draw( 0.613, 0.312) node{$\ss 3$};
\draw( 0.386, 0.426) node{$\ss 6$};
\draw(-0.040, 0.480) node{$\ss 10$};
\draw( 0.493,-0.030) node{$\ss 11$};
\draw( 0.500, 0.715) node{$\ss 12$};
\draw( 0.715, 0.493) node{$\ss 13$};
\end{scope}
\begin{scope}[scale=4,xshift=0.5cm,yshift=-1.4cm]
\draw[magenta] (0.000, 1.890) +(-105:1.604) arc(-105: -75:1.604);
\draw[magenta] (0.000,-1.890) +( 105:1.604) arc( 105: 75:1.604);
\draw[blue] ( 1.131, 0.529) +(225:0.748) arc(225:195:0.748);
\draw[blue] ( 1.131,-0.529) +(135:0.748) arc(135:165:0.748);
\draw[blue] (-1.131, 0.529) +(315:0.748) arc(315:345:0.748);
\draw[blue] (-1.131,-0.529) +( 45:0.748) arc( 45: 15:0.748);
\draw[green] (-0.602,0.000) -- ( 0.602,0.000);
\draw[cyan] (0.000,-0.286) -- (0.000, 0.286);
\fill[black]( 0.000, 0.000) circle(0.01);
\fill[black]( 0.000, 0.286) circle(0.01);
\fill[black]( 0.000,-0.286) circle(0.01);
\fill[black]( 0.602, 0.000) circle(0.01);
\fill[black](-0.602, 0.000) circle(0.01);
\fill[black]( 0.408, 0.339) circle(0.01);
\fill[black]( 0.408,-0.339) circle(0.01);
\fill[black](-0.408, 0.339) circle(0.01);
\fill[black](-0.408,-0.339) circle(0.01);
\draw( 0.650, 0.000) node{$\ss 0$};
\draw(-0.650, 0.000) node{$\ss 1$};
\draw( 0.000,-0.340) node{$\ss 2$};
\draw( 0.000, 0.340) node{$\ss 3$};
\draw( 0.030, 0.030) node{$\ss 6$};
\draw( 0.420,-0.390) node{$\ss 10$};
\draw( 0.420, 0.390) node{$\ss 11$};
\draw(-0.420,-0.390) node{$\ss 12$};
\draw(-0.420, 0.390) node{$\ss 13$};
\end{scope}
\begin{scope}[scale=2.5,xshift=4.5cm]
\draw[magenta] (-1.333,0.000) -- (-0.750,0.000);
\draw[magenta] ( 1.333,0.000) -- ( 0.750,0.000);
\draw[blue] (-1.667,0.000) -- (-1.333,0.000);
\draw[blue] (-0.750,0.000) -- ( 0.750,0.000);
\draw[blue] ( 1.667,0.000) -- ( 1.333,0.000);
\draw[green] (0.000,0.000) -- ( 0.000,1.667);
\draw[cyan] (0.000,0.000) (1.000,0.000) arc(0:180:1.000);
\fill[black](-1.333,0.000) circle(0.015);
\fill[black](-1.000,0.000) circle(0.015);
\fill[black](-0.750,0.000) circle(0.015);
\fill[black]( 0.000,0.000) circle(0.015);
\fill[black]( 0.750,0.000) circle(0.015);
\fill[black]( 1.000,0.000) circle(0.015);
\fill[black]( 1.333,0.000) circle(0.015);
\fill[black]( 0.000,1.000) circle(0.015);
\draw (-1.333,-0.100) node{$\ss 12$};
\draw (-1.000,-0.100) node{$\ss 2$};
\draw (-0.750,-0.100) node{$\ss 10$};
\draw ( 0.000,-0.100) node{$\ss 0$};
\draw ( 0.750,-0.100) node{$\ss 11$};
\draw ( 1.000,-0.100) node{$\ss 3$};
\draw ( 1.333,-0.100) node{$\ss 13$};
\draw ( 0.100, 1.100) node{$\ss 6$};
\end{scope}
\begin{scope}[scale=2,xshift=5.75cm,yshift=-2.7cm]
\draw[magenta] ( 0.000, 0.000) ( 23:1.000) arc( 23:157:1.000);
\draw[magenta] ( 0.000, 0.000) (203:1.000) arc(203:337:1.000);
\draw[blue] ( 0.000, 0.000) (157:1.000) arc(157:203:1.000);
\draw[blue] ( 0.000, 0.000) (337:1.000) arc(337:383:1.000);
\draw[green] (-1.000, 0.000) -- ( 1.000, 0.000);
\draw[cyan] ( 0.000,-1.000) -- ( 0.000, 1.000);
\fill[black] ( 0.000, 0.000) circle(0.02);
\fill[black] ( 0.000, 1.000) circle(0.02);
\fill[black] ( 0.000,-1.000) circle(0.02);
\fill[black] (-1.000, 0.000) circle(0.02);
\fill[black] ( 1.000, 0.000) circle(0.02);
\fill[black] ( 157:1.000) circle(0.02);
\fill[black] ( 203:1.000) circle(0.02);
\fill[black] ( 337:1.000) circle(0.02);
\fill[black] ( 383:1.000) circle(0.02);
\draw( 1.090, 0.000) node{$\ss 0$};
\draw(-1.090, 0.000) node{$\ss 1$};
\draw( 0.000, 1.090) node{$\ss 3$};
\draw( 0.000,-1.090) node{$\ss 2$};
\draw( 0.050, 0.050) node{$\ss 6$};
\draw( 337:1.120) node{$\ss 10$};
\draw( 383:1.120) node{$\ss 11$};
\draw( 203:1.120) node{$\ss 12$};
\draw( 157:1.120) node{$\ss 13$};
\end{scope}
\end{tikzpicture}
\end{center}
\begin{lemma}\label{lem-p-one-props}
$p_1(-z)=-p_1(z)$ and $p_1(\ov{z})=\ov{p_1(z)}$. Thus, $p_1(z)$ has
a Taylor series $\sum_ia_iz^{2i+1}$ with $a_i\in\R$.
\end{lemma}
\begin{proof}
We know that $p$ is $\tPi$-equivariant, so $p\bt_0\lm\mu=\lm\mu p$.
Thus, if we put $\pi=\psi^{-1}\bt_0\lm\mu\psi\in\Aut(\Dl)$ and
$\pi'=\phi\lm\mu\phi^{-1}\in\Aut(\C_\infty)$, we have
$p_1\pi=\pi'p_1$. Direct calculation
shows that $\bt_0\lm\mu$ is the holomorphic involution fixing $v_{H6}$, so
$\pi$ is the holomorphic involution fixing $v_{HS6}=0$, or in other
words $\pi(z)=-z$. Direct calculation also gives $\pi'(z)=-z$, so
$p_1(-z)=-p_1(z)$ as claimed.
Similarly, we have $p\lm\nu=\lm\nu p$. Thus, if we put
$\xi=\psi^{-1}\lm\nu\psi\in\Aut(\Dl)$ and
$\xi'=\phi\lm\nu\phi^{-1}\in\Aut(\C_\infty)$, we have
$p_1\xi=\xi'p_1$. Here $\lm\nu$ is the antiholomorphic involution
of $\Dl$ that fixes $v_{H0}$ and $v_{H6}$, so $\xi$ is the antiholomorphic
involution of $\Dl$ that fixes $v_{HS0}$ and $v_{HS6}$, which gives
$\xi(z)=\ov{z}$. Direct calculation also gives $\xi'(z)=\ov{z}$, so
$p_1(\ov{z})=\ov{p_1(z)}$ as claimed.
\end{proof}
\begin{lemma}\label{lem-p-one-poles}
The set of poles of $p_1$ is $\psi^{-1}(\Pi.\{v_{H1},v_{H9}\})$, and
all these poles are simple. Moreover, the points $\pm ib_-/b_+$ are
poles, and the corresponding residues are equal and are real.
\end{lemma}
\begin{proof}
We are interested in the preimage of $\infty$ under the composite
\[ \xymatrix{
\Dl \ar[r]^{m_1=\psi}_{\simeq} &
\Dl \ar[r]^(0.3){m_2} &
\Dl/\Pi = HX(b) \ar[r]^(0.6){m_3}_(0.6){\simeq} &
PX(a) \ar[r]^{m_4} &
\C_\infty \ar[r]^{m_5=\phi}_{\simeq} &
\C_\infty.
} \]
First, we have $m_5^{-1}\{\infty\}=\{\phi^{-1}(\infty)\}=\{-i\}$. The map
$m_4\:PX(a)\to\C_\infty$ induces a bijection
\[ PX(a)/\ip{\lm^2}\to\C_\infty, \]
and it sends both $v_{P7}$ and $v_{P9}$ to $-i$, so
$m_4^{-1}\{-i\}=\{v_{P7},v_{P9}\}$. As $m_3$ is a cromulent
isomorphism, it follows that
\[ (m_3m_2)^{-1}\{v_{P7},v_{P9}\}=\Pi\{v_{H7},v_{H9}\}. \]
It follows easily that the set of poles is as claimed. The points
$v_7$ and $v_9$ in $PX(a)$ are not fixed by $\lm^2$, so they are not
branch points for the map $PX(a)\to\C_\infty$; it follows that all
the poles are simple. Now put $\al=b_-/b_+\in\R$. After unwinding
the definitions and performing some algebraic simplification we find
that $\psi^{-1}\bt_0(v_{H7})=i\al$ and
$\psi^{-1}\bt_2(v_{H9})=-i\al$, which shows that the points
$\pm i\al$ are poles. This means that there are constants
$r_1,r_2\in\C$ and a meromorphic function $q(z)$ on $\Dl$ such that
$q$ is holomorphic at $\pm i\al$ and
\[ p_1(z) = \frac{r_1}{z-i\al} + \frac{r_2}{z+i\al} + q(z). \]
Moreover, the triple $(r_1,r_2,q(z))$ is characterised uniquely by
these properties. Next, recall that $p_1(z)=-p_1(-z)$. This shows that
$(r_1,r_2,q(z))=(r_2,r_1,-q(-z))$. Similarly, the fact that
$p_1(z)=\ov{p_1(\ov{z})}$ shows that
$(r_1,r_2,q(z))=(\ov{r_2},\ov{r_1},\ov{q(\ov{z})})$. This means that
$r_1=r_2\in\R$ as claimed.
\end{proof}
In the case $b=0.75$, the poles can be illustrated as shown below.
The inner dotted circle (with radius $0.6$) is the smallest circle
centred at the origin that contains $\psi^{-1}(HF_4)$. The only two
poles inside this circle are $\pm ib_-/b_+$; these are shown as solid
red dots. A further $22$ poles are also shown; they are on or outside
the outer dashed circle, which has radius $0.8$. All remaining poles
are even closer to the unit circle.
\begin{center}
\begin{tikzpicture}
\begin{scope}[scale=3]
\draw[black] circle(1);
\draw[black,dotted] circle(0.6);
\draw[black,dashed] circle(0.8);
\draw[magenta] (0.000, 1.890) +(-105:1.604) arc(-105: -75:1.604);
\draw[magenta] (0.000,-1.890) +( 105:1.604) arc( 105: 75:1.604);
\draw[blue] ( 1.131, 0.529) +(225:0.748) arc(225:195:0.748);
\draw[blue] ( 1.131,-0.529) +(135:0.748) arc(135:165:0.748);
\draw[blue] (-1.131, 0.529) +(315:0.748) arc(315:345:0.748);
\draw[blue] (-1.131,-0.529) +( 45:0.748) arc( 45: 15:0.748);
\draw[green] (-0.602,0.000) -- ( 0.602,0.000);
\draw[cyan] (0.000,-0.286) -- (0.000, 0.286);
\fill[red] ( 0.0000, -0.5292) circle(0.015);
\fill[red] ( 0.0000, 0.5292) circle(0.015);
\draw[red] ( 0.0000, -0.9433) circle(0.015);
\draw[red] ( 0.0000, 0.9433) circle(0.015);
\draw[red] ( 0.7252, -0.3392) circle(0.015);
\draw[red] ( 0.5222, -0.7734) circle(0.015);
\draw[red] ( 0.9283, 0.0950) circle(0.015);
\draw[red] ( 0.6684, 0.6509) circle(0.015);
\draw[red] ( 0.2616, -0.8956) circle(0.015);
\draw[red] (-0.2616, -0.8956) circle(0.015);
\draw[red] (-0.7252, -0.3392) circle(0.015);
\draw[red] (-0.5222, -0.7734) circle(0.015);
\draw[red] (-0.6684, 0.6509) circle(0.015);
\draw[red] (-0.9283, 0.0950) circle(0.015);
\draw[red] ( 0.7252, 0.3392) circle(0.015);
\draw[red] ( 0.5222, 0.7734) circle(0.015);
\draw[red] ( 0.9283, -0.0950) circle(0.015);
\draw[red] ( 0.2616, 0.8956) circle(0.015);
\draw[red] (-0.7252, 0.3392) circle(0.015);
\draw[red] ( 0.6684, -0.6509) circle(0.015);
\draw[red] (-0.6684, -0.6509) circle(0.015);
\draw[red] (-0.9283, -0.0945) circle(0.015);
\draw[red] (-0.2616, 0.8956) circle(0.015);
\draw[red] (-0.5222, 0.7734) circle(0.015);
\end{scope}
\end{tikzpicture}
\end{center}
\begin{lemma}\label{lem-p-one-branches}
$p'_1(v_{HS0})=p'_1(v_{HS11})=0$.
\end{lemma}
\begin{proof}
This follows from the fact that the map $HX(b)\to PX(a)\to\C_\infty$
is $\lm^2$-invariant, and $v_0$ and $v_{11}$ are fixed by $\lm^2$ in
$HX(b)$.
\end{proof}
\subsection{Finding \texorpdfstring{$a$}{a} from \texorpdfstring{$b$}{b}}
\label{sec-a-from-b}
In this section, we describe an algorithm to calculate $a$ from $b$.
This algorithm is implemented by the methods of the class
\mcode+H_to_P_map+, which is defined in the file
\fname+hyperbolic/H_to_P.mpl+. In more detail, if we want to
take $b=0.75$, we can enter
\begin{mcodeblock}
HP := `new/H_to_P_map`():
HP["set_a_H",0.75]:
HP["make_samples"]:
HP["find_p1"]:
HP["a_P"];
HP["p1"](z);
HP["err"];
\end{mcodeblock}
The \mcode+find_p1+ method takes about 22 seconds on a fairly capable
PC. The line \mcode+HP["a_P"]+ returns the value of $a$, which is
about $0.1816$. The line \mcode+HP["p1"](z)+ returns a rational
function of $z$, with poles only at the points $\pm ib_-/b_+$
mentioned in Lemma~\ref{lem-p-one-poles}. When restricted to
$\psi^{-1}(HF_4)$, this is a good approximation to $p_1(z)$. The line
\mcode+HP["err"]+ returns a measure of the quality of approximation,
which is about $7\tm 10^{-11}$ in this case. It could be improved by
increasing the degree of polynomials and the number of sample points
used in the algorithm; the code has options for this. The code also
has methods to generate various different visualisations of the
behaviour of $p_1$, and to analyse the errors in more detailed ways.
If one wants to perform the above calculation for several different
values of $b$, and to compare the results with those obtained by the
method of Section~\ref{sec-b-from-a}, then we can instead use the
class \mcode+HP_table+, defined in \fname+hyperbolic/HP_table.mpl+.
For example, we can enter the following:
\begin{mcodeblock}
HPT := `new/HP_table`():
HPT["add_a_H",0.75];
HPT["add_a_H",0.76];
HPT["add_a_H",0.77];
\end{mcodeblock}
This will perform the above calculation for the valuse $b=0.75$,
$b=0.76$ and $b=0.77$. The object of class \mcode+H_to_P_map+ for
$b=0.75$ can then be retrieved as \mcode+HPT["H_to_P_maps"][0.75]+.
After calcuating a sufficient range of values of $b$, one can enter
\mcode+HPT["set_spline"]+ and then \mcode+HPT["full_plot"]+ to
generate a plot of $a$ against $b$.
Alternatively (as discussed in Section~\ref{sec-build}), one can read
the file \fname+build_data.mpl+ and execute
\begin{mcodeblock}
build_data["HP_table"]();
\end{mcodeblock}
to perform all calculations for
$b=0.06$ to $b=0.94$ in steps of $0.02$, and to do various other
related work. Here one may wish to enter
\mcode+infolevel[genus2] := 7;+ before starting the calculation; this
will instruct Maple to print various progress reports as it proceeds.
\begin{lemma}
There is a unique sequence of polynomials
$p_{10}(z),p_{11}(z),p_{12}(z)$ and $p_{14}(z)$, such that:
\begin{itemize}
\item[(a)] All the polynomials are odd, with real coefficients.
\item[(b)] The polynomials $p_{10}(z)$ to $p_{12}(z)$ have degree
$13$.
\item[(c)] The polynomial $p_{14}(z)$ has degree $15$, and has the
form $p_{14}(z)=z+O(z^3)$.
\item[(d)] For all $k$ we have
$p'_{1k}(v_{HS0})=p'_{1k}(v_{HS11})=0$.
\item[(e)] Values at $v_{HS0}$, $v_{HS3}$ and $v_{HS11}$ are as
follows:
\[ \renewcommand{\arraystretch}{1.5}
\begin{array}{|c|c|c|c|} \hline
& v_0 & v_3 & v_{11} \\ \hline
p_{10} & 1 & i & 0 \\ \hline
p_{11} & 0 & 0 & 1 \\ \hline
p_{12} & 0 & 0 & i \\ \hline
p_{14} & 0 & 0 & 0 \\ \hline
\end{array}
\]
\end{itemize}
\end{lemma}
\begin{proof}
Put $\al=v_{HS0}\in\R$ and $\bt=v_{HS3}/i\in\R$ and $\gm=v_{HS11}$.
Let $F$ denote the space of all odd polynomials of degree at most
$13$, and note that this has dimension $7$. Note also that for
$f\in F$ we automatically have $f(\R)\sse\R$ and $f(i\R)\sse i\R$ and
$f'(i\R)\sse\R$, and the roots and their multiplicities are invariant
under the maps $z\mapsto -z$ and $z\mapsto\ov{z}$. Using this, we
see that the only possibility for $p_{14}(z)$ is the polynomial
\[ p_{14}(z) =
z(1-z^2/\al^2)^2(1+z^2/\bt^2)(1-z^2/\gm^2)^2(1-z^2/\ov{\gm}^2)^2.
\]
Next, we can define $\ep\:F\to\R^7$ by
\[ \ep(f) = (f(\al),\;
f(i\bt)/i,\;
\text{Re}(f(\gm)),\;
\text{Im}(f(\gm)),\;
f'(\al),
\text{Re}(f'(\gm)),\;
\text{Im}(f'(\gm))).
\]
If $\ep(f)=0$ then $f$ must be divisible by $p_{14}(z)$, but we can
then compare degrees to see that $f=0$. This means that $\ep$ is an
injective linear map between spaces of dimension $7$, so it is an
isomorphism. The polynomials $p_{10}(z)$, $p_{11}(z)$ and
$p_{12}(z)$ can be obtained by applying $\ep^{-1}$ to suitable
vectors in $\R^7$.
\end{proof}
\begin{definition}
We put
\[ p_{15}(z) =
(z-ib_-/b_+)(z+ib_-/b_+) = z^2 + \frac{1-b^2}{1+b^2},
\]
and then $p_{13}(z)=p_{14}(z)/p_{15}(z)$. Then, given $a\in\R^d$, we
put
\[ P(a)(z) = p_{10}(z) + \sum_{i=1}^3 a_ip_{1i}(z) +
p_{14}(z)\sum_{i=4}^da_iz^{2(i-4)}.
\]
\end{definition}
We now choose $\al\in(0,1)$ which we believe is a reasonable
approximation to the required value of $a$. It would not be too
harmful to just take $\al=1/2$. Alternatively, the code defines a
polynomial \mcode+f(t)=schwarz_b_approx(t)+ of degree ten, which is a
good approximation to the function $a\mapsto b$; we can thus find
$\al$ by solving $f(\al)=b$ numerically. We then put
$z_0=(i-\al)/(i+\al)$, which is the value of $v_{PS11}$ corresponding
to $a=\al$.
Now consider the map $p_1(z)$. Let $a_1$ and $a_2$ be the real and
imaginary parts of $p_1(v_{HS11})$, and let $a_3$ be the unique
real constant such $p_1(z)-a_3p_{13}(z)$ has residue zero at
$ib_-/b_+$. We then find that the function
$p_1(z)-p_{10}(z)-\sum_{i=1}^3a_ip_{1i}(z)$ is holomorphic on a disc
$\Dl'$ centred at $0$ that includes all of $HF_4$. Moreover, it is
odd, with real Taylor coefficients. By considering it order of
vanishing at the various points $v_{HSj}$, we see that is the product
of $p_{14}(z)$ with an even function that is also holomorphic on $\Dl'$.
This means that when $d$ is sufficiently large, $p_1(z)$ can be well
approximated by $P(a)(z)$ for some $a\in\R^d$. To find $a$, we note
that
\[ p_1(C_{HS3}\cup C_{HS5})=C_{PS3}\cup C_{PS5} = S^1. \]
We therefore choose a reasonably large number $n$ (say $n=200$) and a
list of closely spaced points $s=(s_1,\dotsc,s_n)$ lying in
$\psi^{-1}(HF_4\cap(C_{H3}\cup C_{H5}))$. We then define
$\eta\:\R^d\to\R^n$ by
\[ \eta(a)_j = |P(a)(s_j)|^2 - 1. \]
It is not hard to see that this has the form
\[ \eta(a)_j = |(Ma+c)_j|^2 - 1 \]
for some matrix $M\in M_{nd}(\C)$ and some vector $c\in\C^n$ which can
be precomputed. Using this, we get
\[ \frac{\partial}{\partial a_k}\eta(a)_j =
2\text{Re}(\ov{M_{jk}}(Ma+c)_j).
\]
This makes it easy to minimise $\|\eta(a)\|^2$ by an iterative process.
In the case $b\simeq 0.80053190489236$ which is relevant for
uniformising $EX^*$, we have taken $d=50$ and $n=300$, and have ended
up with errors $|P(a)(s_j)|^2-1$ that are less than $10^{-23}$.
Calculations using the above method give the following graph of $a$ as
a function of $b$ (with the marked point representing $EX^*$):
\begin{center}
\begin{tikzpicture}[scale=4]
\draw[black,->] (-0.05,0) -- (1.05,0);
\draw[black,->] (0,-0.05) -- (0,1.05);
\draw[black] (1,-0.05) -- (1,0);
\draw[black] (-0.05,1) -- (0,1);
\draw ( 0.00,-0.05) node[anchor=north] {$0$};
\draw ( 1.00,-0.05) node[anchor=north] {$1$};
\draw (-0.05, 0.00) node[anchor=east ] {$0$};
\draw (-0.05, 1.00) node[anchor=east ] {$1$};
\draw ( 1.05, 0.00) node[anchor=west ] {$b$};
\draw ( 0.00, 1.05) node[anchor=south] {$a$};
\draw[red] plot[smooth] coordinates{ (0.000,1.000) (0.060,1.000) (0.080,1.000) (0.100,1.000) (0.120,1.000) (0.140,1.000) (0.160,1.000) (0.180,1.000) (0.200,1.000) (0.220,1.000) (0.240,0.999) (0.260,0.997) (0.280,0.994) (0.300,0.990) (0.320,0.983) (0.340,0.974) (0.360,0.961) (0.380,0.944) (0.400,0.923) (0.420,0.898) (0.440,0.869) (0.460,0.835) (0.480,0.798) (0.500,0.757) (0.520,0.713) (0.540,0.667) (0.560,0.620) (0.580,0.571) (0.600,0.521) (0.620,0.472) (0.640,0.423) (0.660,0.375) (0.680,0.328) (0.700,0.283) (0.720,0.241) (0.740,0.201) (0.760,0.163) (0.780,0.129) (0.800,0.099) (0.820,0.073) (0.840,0.050) (0.860,0.032) (0.880,0.019) (0.900,0.009) (0.920,0.004) (0.940,0.001) (1.000,0.000) };
\fill[black] (0.801,0.098) circle(0.015);
\end{tikzpicture}
\end{center}
The middle section of this graph can be obtained by the methods of
this section, or those of Section~\ref{sec-b-from-a}. However, the
methods of Section~\ref{sec-b-from-a} behave poorly when $(a,b)$
approaches $(1,0)$.
An optimist might hope for an explicit formula for the above graph,
perhaps involving the elliptic integrals from
Section~\ref{sec-P}, the constants $s_k$ in
Definition~\ref{defn-H-curves}, or other standard special
functions or quantities found elsewhere in this document. We have
performed a fairly extensive experimental search for such formulae,
but without success. Of course one can find polynomials of high
degree fitting the graphs to any desired accuracy, but the answers are
not illuminating. It might be helpful if we had a meaningful
interpretation of the endpoints $(a,b)=(1,0)$ and $(a,b)=(0,1)$,
perhaps in terms of a Deligne-Mumford compactification of an
appropriate moduli space of stable curves, but we have not
investigated this seriously.
\subsection{Recollections on the Schwarzian derivative}
\label{sec-schwarz}
We now start to discuss our second method, where we are given
$a\in(0,1)$ and we try to find $b\in (0,1)$ such that
$PX(a)\simeq HX(b)$. This method is based on the Schwarzian
derivative, whose definition and properties we now recall.
\begin{definition}
Let $f$ be a nonconstant meromorphic function on a connected domain
$U\sse\C$. The Schwarzian derivative $S(f)$ is defined by
\[ S(f) = (f''/f')' - \tfrac{1}{2} (f''/f')^2
= f'''/f' - \tfrac{3}{2} (f''/f')^2.
\]
\end{definition}
\begin{proposition}\label{prop-vanishing-schwarzian}
We have $S(f)=0$ if and only if there are constants $a,b,c,d\in\C$
with $ad-bc\neq 0$ and $f(z)=(az+b)/(cz+d)$ on $U$ (or in other
words, $f$ is a M\"obius function).
\end{proposition}
\begin{proof}
This is standard. One can check by direct calculation that
functions $f(z)=(az+b)/(cz+d)$ have $S(f)=0$. Conversely, suppose
that $f$ is meromorphic on $U$ with $S(f)=0$. Then the function
$g=f''/f'$ satisfies $g'=g^2/2$. If $g=0$ then we deduce that
$f''=0$ so $f$ has the form $f(z)=az+b$ for some $a$ and $b$, as
required. Otherwise, we can consider the meromorphic function $1/g$
and we deduce that $(1/g)'=-g'/g^2=-1/2$, which gives
$g(z)=-2/(z+\dl)$ for some constant $\dl$, or in other words
$(z+\dl)f''(z)=-2f'(z)$. From this it follows that the function
$(z+\dl)^2f'(z)$ has zero derivative and so is constant. This gives
that $f'(z)=\gm/(z+\dl)^2$ and then $f(z)=\bt-\gm/(z+\dl)$ for some
constants $\bt$ and $\dl$, and this can evidently be rewritten in the
form $(az+b)/(cz+d)$.
\end{proof}
\begin{proposition}
Given meromorphic functions $V\xra{g}U\xra{f}\C$ we have
\[ S(f\circ g) = (S(f)\circ g) \cdot (g')^2 + S(g). \]
(We call this the \emph{Schwarzian chain rule}.) In particular, if
$f$ is a M\"obius function then $S(f\circ g)=S(g)$.
\end{proposition}
\begin{proof}
This is also standard. We can use the ordinary chain rule repeatedly to
express the first three derivatives of $f\circ g$ in terms of those
of $f$ and $g$, and the rest is pure algebra.
\end{proof}
We next recall the relationship between the Schwarzian derivative and
certain types of second order linear differential equations. Again,
almost all of this is well-known, but we give a self-contained account
to serve as a convenient basis for discussing some additional points
that are less standard.
\begin{proposition}\label{prop-schwarzian-solutions}
Let $U$ be a simply connected open subset of $\C$, and let $s$ be a
holomorphic function on $U$. Put
\begin{align*}
F &= \{ \text{ meromorphic functions $f$ on $U$
such that $f''+\half sf=0$} \} \\
G &= \{ \text{ nonconstant meromorphic functions $g$ on $U$
such that $S(g)=s$} \}.
\end{align*}
Then
\begin{itemize}
\item[(a)] Every function in $F$ is actually holomorphic.
\item[(b)] For any $z_0\in U$ and any $u_0,u_1\in\C$ there is a
unique function $f\in F$ such that
$f(z)=u_0+u_1(z-z_0)+O((z-z_0)^2)$. In particular, $F$ has
dimension two over $\C$.
\item[(c)] $G$ is precisely the set of functions of the form
$f_1/f_0$, where $f_0$ and $f_1$ are linearly independent elements
of $F$.
\item[(d)] For any $z_0\in U$ and any $v_0,v_1,v_2\in\C$ with
$v_1\neq 0$ there is a unique function $g\in G$ such that
$g(z)=v_0+v_1(z-z_0)+v_2(z-z_0)^2+O((z-z_0)^3)$.
\end{itemize}
\end{proposition}
\begin{proof}
\begin{itemize}
\item[(a)] This is Lemma~\ref{lem-F-holomorphic}.
\item[(b)] This is Corollary~\ref{cor-F-disc}.
\item[(c)] Combine Lemma~\ref{lem-quotient-schwarzian} and
Corollary~\ref{cor-quotient-schwarzian}.
\end{itemize}
\end{proof}
\begin{lemma}\label{lem-F-holomorphic}
Every function in $F$ is holomorphic.
\end{lemma}
\begin{proof}
Consider an element $f\in F$ and a point $z_0\in U$. If $f$ is not
holomorphic at $z_0$, then it must have a pole of order $d>0$ at
$z_0$, so $f''$ has a pole of order $d+2$, whereas $\half sf$ has a
pole of order at most $d$ (or is holomorphic). This contradicts the
equation $f''+\half sf=0$.
\end{proof}
\begin{lemma}\label{lem-F-disc}
Let $U$ be the open disc centred at $z_0$ with radius $r>0$, and let
$s$ be a holomorphic function on $U$. Define $F$ and $G$ as above.
Then for any $u_0,u_1\in\C$ there is a unique function $f\in F$ such
that $f(z)=u_0+u_1(z-z_0)+O((z-z_0)^2)$. In particular, $F$ has
dimension two over $\C$.
\end{lemma}
\begin{proof}
We can expand $s$ as a power series, say $s(z)=\sum_ks_k(z-z_0)^k$.
Put $a_0=u_0$ and $a_1=u_1$, then define $a_k$ recursively for $k>1$
by
\[ a_k = \frac{-1}{2k(k-1)}\sum_{j=0}^{k-2}a_js_{k-2-j}. \] It is
then easy to see that the formal power series
$f_0(z)=\sum_ka_k(z-z_0)^k$ is the unique one with $f''_0+\half
sf_0=0$ and $f_0(z)=u_0+u_1(z-z_0)+O((z-z_0)^2)$. Moreover, the
coefficients $a_k$ grow at a rate comparable to that of the
coefficients $s_k$, so they have the same radius of convergence.
Thus, the above expression defines a holomorphic function on $U$.
\end{proof}
\begin{corollary}\label{cor-F-disc}
Claim~(b) in Proposition~\ref{prop-schwarzian-solutions} holds for
an arbitrary simply connected domain $U$.
\end{corollary}
\begin{proof}
This follows by analytic continuation. In more detail, for any
$z\in U$ we can choose a path $\gm$ from $z_0$ to $z$ in $U$. We can
then choose closely spaced points $\gm(t_i)$ and radii $r_i>0$ such
that $t_0=1$ and $t_N=1$ and the disc $U_i$ of radius $r_i$ centred
at $\gm(t_i)$ is contained in $U$ and contains $\gm(t_{i+1})$. We
then let $f_0$ be the unique holomorphic function on $U_0$ with
$f''_0+\half sf_0=0$ and $f_0(z)=u_0+u_1(z-z_0)+O((z-z_0)^2)$. Using
this as a starting point, we let $f_k$ denote the unique holomorphic
function on $U_k$ that satisfies $f''_k+\half sf_k=0$ and agrees with
$f_{k-1}$ to second order at $\gm(t_k)$. We then define
$f(z)=f_N(z)$. It is easy to check that this does not depend on the
precise choice of points $t_i$, nor does it change if we move $\gm$
by a small homotopy fixing the endpoints. As $U$ is simply connected
any two paths from $z_0$ to $z$ are homotopic relative to endpoints,
and any homotopy can be broken down into small homotopies. It
follows that $f(z)$ is independent of all choices, and it defines an
element of $F$ with the required behaviour near $z_0$. Any other
such element will agree with $f$ at least on $U_0$, but then it must
agree everywhere on $U$ by analytic continuation.
\end{proof}
\begin{lemma}\label{lem-quotient-schwarzian}
Suppose that $f_0$ and $f_1$ are linearly independent elements of
$F$, so the quotient $g=f_1/f_0$ is a nonconstant meromorphic
function. Then $S(g)=s$.
\end{lemma}
\begin{proof}
Put $W=f_0f'_1-f'_0f_1$ (the Wronskian of $f_0$ and $f_1$). This is
easily seen to satisfy $W'=0$, so it is constant. The quotient
$f=f_1/f_0$ has $f'=W/f_0^2$, so $f''=-2Wf'_0/f_0^3$, so
$f''/f'=-2f'_0/f_0$. From this we find that $S(f)=s$ as claimed.
\end{proof}
\begin{lemma}\label{lem-G-transitive}
If $g,h\in G$ then $g=m\circ h$ for some M\"obius function $m$.
\end{lemma}
\begin{proof}
As $h$ is nonconstant and meromorphic, we can choose a small disc
$V\sse U$ such that $h$ is holomorphic on $V$ and $h'$ is nonzero
everywhere in $V$. After shrinking $V$ if necessary, we can then
assume that the map $h\:V\to h(V)$ is a conformal isomorphism. Put
$m=g\circ h^{-1}$, which is meromorphic on $h(V)$ and satisfies
$g=m\circ h$. The Schwarzian chain rule gives
$S(g)=(S(m)\circ h)\cdot(h')^2+S(h)$ on $V$. However, $S(g)=S(h)=s$
and $h'$ is nowhere zero, so $S(m)\circ h=0$ on $V$, so $S(m)=0$ on
$h(V)$. It follows that $m$ is a M\"obius function. The functions
$g$ and $m\circ h$ are both meromorphic and they agree on a disc so
they must agree everywhere in $U$.
\end{proof}
\begin{corollary}\label{cor-quotient-schwarzian}
Every element $g\in G$ can be written as $g=f_1/f_0$ for some
linearly independent pair of elements $f_0,f_1\in F$.
\end{corollary}
\begin{proof}
Let $e_0$ and $e_1$ be any basis for $F$, and put $h=e_1/e_0$.
Lemma~\ref{lem-quotient-schwarzian} tells us that $h\in G$, so
Lemma~\ref{lem-G-transitive} tells us that $g=(ah+b)/(ch+d)$ for some
$a,b,c,d$ with $ad-bc\neq 0$. This means that the functions
$f_1=ae_1+be_0$ and $f_0=ce_1+de_0$ are linearly independent elements
of $F$ with $g=f_1/f_0$.
\end{proof}
\begin{lemma}\label{lem-mobius-straighten}
Suppose that $f$ is holomorphic at $z_0$, with $f'(z_0)\neq 0$. Then
there is a unique M\"obius function $m$ such that
$m(f(z))=z-z_0+O((z-z_0)^3)$. Specifically, if
\[ f(z) = u_0 + u_1(z-z_0) + u_2(z-z_0)^2 + O((z-z_0)^3) \]
then
\[ m(z) = \frac{u_1(z-u_0)}{u_2(z-u_0)+u_1^2}. \]
\end{lemma}
\begin{proof}
If we define $m(z)$ as above, then it is straightforward to check
that $m(f(z))=z-z_0+O((z-z_0)^3)$. If $n$ is another M\"obius
function with $n(f(z))=z-z_0+O((z-z_0)^3)$ then the function
$k=n\circ m^{-1}$ must have the form $k(z)=(az+b)/(cz+d)$ for some
$a,b,c,d$, but also $k(z)=z+O(z^3)$. In particular, we have
$k(0)=0$, which gives $b=0$. We then have $k'(0)=1$, which gives
$a=d$. After cancelling we may assume that $a=d=1$. Finally, we
have $k''(0)=0$, so $c=0$, so $k$ is the identity as required.
\end{proof}
\begin{corollary}\label{cor-mobius-straighten}
For any $z_0\in U$ and any $v_0,v_1,v_2\in\C$ with $v_1\neq 0$ there
is a unique function $g\in G$ such that
$g(z)=v_0+v_1(z-z_0)+v_2(z-z_0)^2+O((z-z_0)^3)$.
\end{corollary}
\begin{proof}
It will be harmless, and notationally convenient, to assume that
$z_0=0$. By Corollary~\ref{cor-F-disc} we can choose
$f_0,f_1\in F$ with $f_0(z)=1+O(z^2)$ and $f_1(z)=z+O(z^2)$. The
function $g_0=f_1/f_0$ now lies in $G$ and has $g'_0(0)\neq 0$, so we
can find a M\"obius function $m_1$ such that the function
$g_1(z)=m_1(g_0(z))$ satisfies $g_1(z)=z+O(z^3)$. We then put
\[ m_2(z)=(v_0v_1-(v_1^2-v_0v_2)z)/(v_1-v_2z) \]
and $g(z)=m_2(g_1(z))$.
\end{proof}
The next result refers to circles in $\C_\infty$. Here we regard
straight lines as circles of infinite radius. With this convention,
it is well-known that M\"obius functions send circles to circles.
\begin{proposition}\label{prop-image-circle}
Suppose that a real interval $(a,b)$ is contained in $U$, and that
$f'(t)\neq 0$ for all $t\in(a,b)$, so $S(f)$ is holomorphic on
$(a,b)$. Then the following are equivalent:
\begin{itemize}
\item[(a)] $f((a,b))$ is contained in some circle
$C\subset\C_\infty$.
\item[(b)] $S(f)$ is real on $(a,b)$
\end{itemize}
\end{proposition}
\begin{proof}
Suppose that~(a) holds. We can choose a M\"obius function $m$ such
that $m(C)=\R$, and put $g=m\circ f$, so $g((a,b))\sse\R$. From the
definition of $S(g)$ it is clear that $S(g)$ is real on $(a,b)$, but
the Schwarzian chain rule shows that $S(f)=S(g)$, so~(b) holds.
Conversely, suppose that the function $s=S(f)$ is real on $(a,b)$.
Choose a point $t_0\in(a,b)$. Lemma~\ref{lem-mobius-straighten}
gives us a M\"obius function $m$ such that the function $g=m\circ f$
has $g(t)=(t-t_0)+O((t-t_0)^3)$. Note also that $S(g)$ is again
equal to $s$. Now put $h(z)=\ov{g(\ov{z})}$, and note that this is
again meromorphic. Using power series representations we can see
that $S(h)=r$, where $r(z)=\ov{s(\ov{z})}$. Now $r$ is also
holomorphic, and agrees with $s$ on $(a,b)$, so it must agree with
$s$ everywhere in $U$. This means that $g$ and $h$ are both elements
of $G$, so Lemma~\ref{lem-G-transitive} gives us a M\"obius
function $n$ with $g=n\circ h$. However, both $g(t)$ and $h(t)$ are
of the form $(t-t_0)+O((t-t_0)^3)$. It follows that $n(z)=z+O(z^3)$,
and thus that $n$ is the identity, so $g=h$. This means that
$g((a,b))\sse\R$, so $f((a,b))\sse m^{-1}(\R)$. Moreover, as $m$ is
a M\"obius function, the set $m^{-1}(\R)$ is a circle as required.
\end{proof}
\begin{proposition}
Let $U$ be the disc of radius $r$ centred at $z_0$, and let $s(z)$ be
a function that is holomorphic on $U\sm\{z_0\}$ and has Laurent
expansion $\sum_{k\geq -2}s_k(z-z_0)^k$ with $s_{-2}=3/8$. Let $U'$
be obtained from $U$ by removing a line segment from $z_0$ to the
edge, and let $\xi(z)$ be a holomorphic branch of $(z-z_0)^{1/4}$ on
$U'$. Let $F'$ be the space of holomorphic solutions of
$f''+\half sf=0$ on $U'$. Then
\begin{itemize}
\item There is a unique holomorphic function $e_0$ on $U$ such that
$e_0(0)=1$ and the function $f_0=e_0\xi$ lies in $F'$.
\item There is a unique holomorphic function $e_1$ on $U$ such that
$e_1(0)=1$ and the function $f_1=e_1\xi^3$ lies in $F'$.
\end{itemize}
\end{proposition}
\begin{proof}
It will be harmless, and notationally convenient, to assume that
$z_0=0$.
By assumption we have $\xi(z)^4=z$, so $4\xi^3\xi'=1$, so
$\xi'(z)=1/(4\xi(z)^3)=\xi(z)/(4z)$.
Consider a function of the form $f=e\xi$, where
$e(z)=\sum_{k\geq 0}a_kz^k$ (and we take $a_k=0$ for $k<0$). We then have
\[ f'(z)=e'(z)\xi(z)+e(z)\xi'(z)=(e'(z)+\tfrac{1}{4}e(z)z^{-1})\xi(z). \]
By a similar argument, we have
\[ f''(z) =
(e''(z) + \tfrac{1}{2}e'(z)z^{-1} -
\tfrac{3}{16}e(z)z^{-2})\xi(z).
\]
Thus, the equation $f''+\half sf=0$ is equivalent to
\[ e''(z) + \half z^{-1} e'(z) + \half t(z) e(z) = 0, \]
where
\[ t(z) = s(z) - \tfrac{3}{8} z^{-2} = \sum_{k\geq -1}s_kz^k. \]
Looking at the coefficient of $z^k$, we get
\[ (k+\tfrac{3}{2})(k+2) a_{k+2} +
\half\sum_{j=0}^{k+1}s_{k-j}a_j = 0.
\]
For $k\leq -2$ this is trivially satisfied, and for $k\geq -1$ it can be
rearranged to express $a_{k+2}$ in terms of $a_0,\dotsc,a_{k+1}$. It
follows that there is a unique power series solution with $a_0=1$.
The rate of growth of the coefficients $a_k$ is comparable with that
of the coefficients $s_k$, so the series $\sum_ka_kz^k$ converges to
give the claimed function $e_0(z)$.
Now consider instead a function of the form $f=e\xi^3$. We find in
the same way that
\[ f''(z) =
(e''(z) + \tfrac{3}{2}e'(z)z^{-1} -
\tfrac{3}{16}e(z)z^{-2})\xi(z)^3,
\]
and thus that the equation $f''+\half sf=0$ is equivalent to
\[ e''(z) + \tfrac{3}{2} z^{-1} e'(z) + \half t(z) e(z) = 0, \]
or
\[ (k+\tfrac{5}{2})(k+2) a_{k+2} +
\half\sum_{j=0}^{k+1}s_{k-j}a_j = 0.
\]
Just as in the previous case, there is a unique solution, as
claimed. The functions $f_0$ and $f_1$ are linearly independent (as
we can see by considering their rate of growth near $z=0$), and we
have seen that $F'$ has dimension two, so they must form a basis.
\end{proof}
\subsection{Application to cromulent surfaces}
\label{sec-b-from-a}
For each $a\in(0,1)$ we have shown that there is a unique $b\in(0,1)$
such that $PX(a)\simeq HX(b)$ as cromulent surfaces, and in fact there
is a unique cromulent isomorphism $HX(b)\to PX(a)$. As in
Section~\ref{sec-P-H-prelim}, we write $p$ for the canonical map
\[ \Dl \to \Dl/\Pi = HX(b) \to PX(a) \to
PX(a)/\ip{\lm}^2 \to \C_\infty,
\]
and we also consider the map $p_1=\phi p\psi\:\Dl\to\C_\infty$.
In this section, we describe an algorithm to calculate $b$ from $a$.
This algorithm is implemented by the methods of the class
\mcode+P_to_H_map+, which is defined in the file
\fname+hyperbolic/P_to_H.mpl+. In more detail, if we want to
take $a=0.1$, we can enter
\begin{mcodeblock}
PH := `new/P_to_H_map`():
PH["set_a_P",0.1]:
PH["add_charts"]:
PH["find_p1_inv"];
PH["a_H"];
PH["p1_inv"](z);
PH["err"];
\end{mcodeblock}
The line \mcode+PH["a_H"]+ returns the value of $b$, which is
about $0.7994$. The line \mcode+PH["p1_inv"](z)+ returns a polynomial
in $z$, which is a good approximation to $p_1^{-1}(z)$. The line
\mcode+HP["err"]+ returns a measure of the quality of approximation,
which is about $10^{-21}$ in this case.
If one wants to perform the above calculation for several different
values of $a$, and to compare the results with those obtained by the
method of Section~\ref{sec-a-from-b}, then we can instead use the
class \mcode+HP_table+, defined in \fname+hyperbolic/HP_table.mpl+.
For example, we can enter the following:
\begin{mcodeblock}
HPT := `new/HP_table`():
HPT["add_a_P",0.1];
HPT["add_a_P",0.2];
HPT["add_a_P",0.3];
\end{mcodeblock}
This will perform the above calculation for the valuse $a=0.1$,
$a=0.2$ and $a=0.3$. The object of class \mcode+P_to_H_map+ for
$a=0.1$ can then be retrieved as \mcode+HPT["P_to_H_maps"][0.1]+.
The function \mcode+build_data["HP_table"]()+ (defined in
\fname+build_data.mpl+) does the calculation for $a$ from $0.06$ to
$0.94$ in steps of $0.02$ (as well as implementing the method of
Section~\ref{sec-a-from-b} and performing various other work).
We can think of the inverse of $p$ as giving a multivalued function
$p^{-1}\:\C_\infty\to\Dl$. It is a key point that the Schwarzian
derivative $S(p^{-1})$ is single-valued, and in fact is a rational
function whose properties we can understand quite explicitly. To
explain this, we will use the following definitions
\begin{definition}\label{defn-schwarz-s}
We put
\begin{align*}
B &= \{v_{Ci}\st i\in\{0,1,10,11,12,13\}\} \\
&= \{0,\infty,a,-a,1/a,-1/a\} \subset\C_\infty \\
B_0 &= B\sm\{\infty\} = \{0,a,-a,1/a,-1/a\} \\
r_a(z) &= \prod_{u\in B_0}(z-u) = z^5-(a^2+a^{-2})z^3+z \\
s_0(z) &= \frac{-3z^3/2}{r_a(z)} +
\frac{3}{8}\sum_{u\in B_0}\frac{1}{(z-u)^2} \\
s_1(z) &= \frac{z}{r_a(z)}.
\end{align*}
\end{definition}
We have seen $r_a(z)$ before; the definition is repeated for ease of
reference and to display the connection with $B$ and $B_0$. Recall
also that $B$ is the set of critical values of $p$, so $p$
restricts to give a covering map $\Dl\sm p^{-1}(B)\to\C_\infty\sm B$.
For any sufficiently small connected open set
$U\subset\C_\infty\sm B$, we can choose a holomorphic map $f\:U\to\Dl$
with $pf=1_U$ (so $f$ is a local branch of $p^{-1}$); we then define
$S(p^{-1})_U=S(f)$. This is independent of the choice of $f$, because
any other choice has the form $m\circ f$ for some M\"obius map
$m\in\ip{\Pi,\lm^2}$, and the Schwarzian chain rule gives
$S(m\circ f)=S(f)$. Given this invariance, it is clear that
$S(p^{-1})_U=S(p^{-1})_V$ on $U\cap V$. We can therefore patch these
local functions together to get a meromorphic function $S(p^{-1})$ on
$\C_\infty\sm B$.
\begin{proposition}\label{prop-S-p-inv}
There is a constant $d\in\R$ such that $S(p^{-1})=s_0+ds_1$ (where
$s_0$ and $s_1$ are as in Definition~\ref{defn-schwarz-s}).
\end{proposition}
\begin{proof}
For convenience, we write $s(z)=S(p^{-1})(z)$. Put
\[ d(z)=(s(z)-s_0(z))/s_1(z)=(s(z)-s_0(z))r_a(z)/z. \]
The claim is that this is a real constant.
First let $U$ and $f$ be as in our definition of $s(z)$. The
equation $pf=1_U$ implies that $f'$ is nonzero everywhere in $U$, so
the Schwarzian derivative $s=f'''/f'-\tfrac{3}{2}(f''/f')^2$ is
holomorphic in $U$. This shows that all singularities of $s$ must
lie in $B$.
Now consider a point $u\in B_0$, and choose $\tu\in\Dl$ with
$p(\tu)=u$. By a standard argument with branched double covers, we
have $p(z)=u+c(z-\tu)^2+O((z-\tu)^3)$ for some $c\neq 0$. It follows
that on any small open set close to $u$, we have
$p^{-1}(z)=((z-u)/c)^{1/2}+O((z-u)^{3/2})$. Computing the
Schwarzian derivative from this approximation gives
$s(z)=\tfrac{3}{8}(z-u)^{-2}+O((z-u)^{-1})$, which matches the
behaviour of $s_0(z)$. We therefore see that $s(z)-s_0(z)$ has at
worst a simple pole at $u$. As this holds for all $u\in B_0$, we see
that the product $e(z)=(s(z)-s_0(z))r_a(z)=z\,d(z)$ is holomorphic
everywhere in $\C$, so $d(z)$ has at worst a simple pole at $z=0$.
Next, recall that there is an element $\mu\in\tPi$ that satisfies
$p(\mu(z))=1/p(z)$ for all $z\in\Dl$. In other words, if we define
$\mu_0\:\C_\infty\to\C_\infty$ by $\mu_0(z)=1/z$, then we have
$\mu_0\circ p=p\circ\mu$, and so $p^{-1}\circ\mu_0=\mu\circ p^{-1}$.
Here $\mu$ and $\mu_0$ are M\"obius maps so the Schwarzian chain
rule gives $(S(p^{-1})\circ\mu_0)\;(\mu_0')^2=S(p^{-1})$, and thus
$s(z^{-1})=z^4s(z)$. A similar argument with $\lm$ gives
$s(-z)=s(z)$. By direct calculation we also see that
\begin{align*}
s_0(z^{-1}) &= z^4s_0(z^{-1}) & s_0(-z) &= s_0(z) \\
s_1(z^{-1}) &= z^4s_1(z^{-1}) & s_1(-z) &= s_1(z).
\end{align*}
It follows that $d(z^{-1})=d(-z)=d(z)$. As $d$ is even and has at
worst a simple pole at $0$, it must actually be holomorphic at $0$.
It is also holomorphic elsewhere in $\C$ and it satisfies
$d(z^{-1})=d(z)$ so it must be bounded on $\C$ and thus constant.
Finally, recall that $p$ is equivariant with respect to the map
$\nu$, or in other words $p(\ov{z})=\ov{p(z)}$. This means that $p$
is real on $\R\cap\Dl$, so $p^{-1}$ can be chosen to be real on $\R$,
so $S(p^{-1})$ is real on $\R$. From this it is clear that $d\in\R$.
\end{proof}
\begin{remark}
Maple notation for $d$ and $s(z)$ is \mcode+d_p_inv+ and
\mcode+S_p_inv+.
\end{remark}
\begin{remark}
We can regard the parameter $d$ in the proposition as a function of
$a$ or of $b$. Numerical calculations (by a method to be described
below) suggest that $d$ grows like $a^{-2}$ as $a\to 0$, and give the
following graph of $a^2d$ against $a$:
\begin{center}
\begin{tikzpicture}[xscale=10,yscale=20]
\draw[->] (-0.02,0.45) -- (1.05,0.45);
\draw[->] (0,0.4) -- (0,0.63);
\draw (0,0.63) node[anchor=south] {$\ss a^2d$};
\draw (1.07,0.45) node[anchor=west] {$\ss a$};
\draw (-0.02,0.45) node[anchor=east] {$\ss 0.45$};
\draw (-0.02,0.50) node[anchor=east] {$\ss 0.50$};
\draw (-0.02,0.55) node[anchor=east] {$\ss 0.55$};
\draw (-0.02,0.60) node[anchor=east] {$\ss 0.60$};
\draw (0,0.45) -- (-0.02,0.45);
\draw (0,0.50) -- (-0.02,0.50);
\draw (0,0.55) -- (-0.02,0.55);
\draw (0,0.60) -- (-0.02,0.60);
\draw (0.1,0.45) -- (0.1,0.44);
\draw (0.2,0.45) -- (0.2,0.44);
\draw (0.3,0.45) -- (0.3,0.44);
\draw (0.4,0.45) -- (0.4,0.44);
\draw (0.5,0.45) -- (0.5,0.44);
\draw (0.6,0.45) -- (0.6,0.44);
\draw (0.7,0.45) -- (0.7,0.44);
\draw (0.8,0.45) -- (0.8,0.44);
\draw (0.9,0.45) -- (0.9,0.44);
\draw (1.0,0.45) -- (1.0,0.44);
\draw (0.1,0.44) node[anchor=north] {$\ss 0.1$};
\draw (0.2,0.44) node[anchor=north] {$\ss 0.2$};
\draw (0.3,0.44) node[anchor=north] {$\ss 0.3$};
\draw (0.4,0.44) node[anchor=north] {$\ss 0.4$};
\draw (0.5,0.44) node[anchor=north] {$\ss 0.5$};
\draw (0.6,0.44) node[anchor=north] {$\ss 0.6$};
\draw (0.7,0.44) node[anchor=north] {$\ss 0.7$};
\draw (0.8,0.44) node[anchor=north] {$\ss 0.8$};
\draw (0.9,0.44) node[anchor=north] {$\ss 0.9$};
\draw[red] (0.030,0.592) -- (0.040,0.589) -- (0.050,0.585) --
(0.060,0.583) -- (0.070,0.580) -- (0.080,0.577) -- (0.090,0.575) --
(0.100,0.572) -- (0.110,0.570) -- (0.120,0.568) -- (0.130,0.566) --
(0.140,0.564) -- (0.150,0.561) -- (0.160,0.559) -- (0.170,0.557) --
(0.180,0.555) -- (0.190,0.553) -- (0.200,0.551) -- (0.210,0.549) --
(0.220,0.547) -- (0.230,0.546) -- (0.240,0.544) -- (0.250,0.542) --
(0.260,0.540) -- (0.270,0.538) -- (0.280,0.536) -- (0.290,0.535) --
(0.300,0.533) -- (0.310,0.531) -- (0.320,0.529) -- (0.330,0.528) --
(0.340,0.526) -- (0.350,0.524) -- (0.360,0.523) -- (0.370,0.521) --
(0.380,0.520) -- (0.390,0.518) -- (0.400,0.517) -- (0.410,0.515) --
(0.420,0.514) -- (0.430,0.513) -- (0.440,0.511) -- (0.450,0.510) --
(0.460,0.509) -- (0.470,0.508) -- (0.480,0.507) -- (0.490,0.506) --
(0.500,0.505) -- (0.510,0.504) -- (0.520,0.503) -- (0.530,0.502) --
(0.540,0.502) -- (0.550,0.501) -- (0.560,0.501) -- (0.570,0.500) --
(0.580,0.500) -- (0.590,0.500) -- (0.600,0.500) -- (0.610,0.500) --
(0.620,0.500) -- (0.630,0.500) -- (0.640,0.500) -- (0.650,0.500) --
(0.660,0.501) -- (0.670,0.502) -- (0.680,0.502) -- (0.690,0.503) --
(0.700,0.504) -- (0.710,0.505) -- (0.720,0.506) -- (0.730,0.508) --
(0.740,0.509) -- (0.750,0.511) -- (0.760,0.513) -- (0.770,0.515) --
(0.780,0.517) -- (0.790,0.519) -- (0.800,0.521) -- (0.810,0.524) --
(0.820,0.527) -- (0.830,0.529) -- (0.840,0.532) -- (0.850,0.535) --
(0.860,0.539) -- (0.870,0.542) -- (0.880,0.545) -- (0.890,0.549) --
(0.900,0.552) -- (0.910,0.556) -- (0.920,0.557);
\end{tikzpicture}
\end{center}
\end{remark}
We can now try to find $p^{-1}$ and $p$ by power series methods. As
before, it turns out to be convenient to focus on $p_1^{-1}$ rather
than $p^{-1}$, because $p_1^{-1}$ is unbranched at the origin, and the
associated power series has a reasonably large radius of convergence.
\begin{proposition}\label{prop-p-one}
We have $S(p_1^{-1})=s^*_0+ds^*_1$, where $d$ is the same real
constant as in Proposition~\ref{prop-S-p-inv}, and
\begin{align*}
s^*_0(z) &= \frac{192a^4z^2(1+z^2)^2-9(1-a^4)^2(1-z^2)^4}{
2(1-z^2)^2((1+a^2)^2(1-z^2)^2+16a^2z^2)^2} &
s^*_1(z) &= \frac{4a^2}{(1+a^2)^2(1-z^2)^2+16a^2z^2}.
\end{align*}
The set of poles of $S(p_1^{-1})$ is
\[ \psi^{-1}(B) = \{v_{PSi}\st i\in\{0,1,10,11,12,13\}\} =
\left\{\pm 1,
\frac{i-a}{i+a},\frac{i+a}{i-a},
\frac{a-i}{a+i},\frac{a+i}{a-i}\right\}.
\]
\end{proposition}
\begin{proof}
Recall that $p_1=\phi p\psi$, so $p_1^{-1}=\psi^{-1}p^{-1}\phi^{-1}$.
Here $\phi^{-1}$ and $\psi^{-1}$ are both M\"obius maps, so
$S(\phi^{-1})=0$ and $S(\psi^{-1})=0$. The Schwarzian chain rule
therefore gives
\[ S(p_1^{-1}) =
S(\psi^{-1}\circ p^{-1}\circ\phi^{-1}) =
(S(p^{-1})\circ\phi^{-1}) \cdot ((\phi^{-1})')^2.
\]
Here $S(p^{-1})$ is given by Proposition~\ref{prop-S-p-inv}, and
\[ (\phi^{-1})'(z) = -\frac{2i}{(1+z)^2}. \]
It is now a lengthy but straightforward calculation to show that this
is equivalent to the claimed formula. We also see that the poles of
$S(p_1^{-1})$ are the images under $\phi$ of the poles of
$S(p^{-1})$, together with the poles of $(\phi^{-1})'$. We saw
previously that $S(p^{-1})$ has poles only in the set
\[ \{0,\pm a,\pm 1/a\}=\{v_{C0},v_{C10},v_{C11},v_{C12},v_{C13}\}.
\]
These points $v_{Ci}$ are mapped by $\phi$ to the corresponding
points $v_{PSi}$, and the only pole of $(\phi^{-1})'$ is at
$-1=v_{PS1}$, so the poles of $S(p_1^{-1})$ are as claimed.
\end{proof}
\begin{remark}
Maple notation for $s_0^*(z)$ and $s_1^*(z)$ is \mcode+S0_p1_inv(z)+
and \mcode+S1_p1_inv(z)+.
\end{remark}
\begin{remark}\label{rem-heun}
We can also consider the map $p_2=\xi\circ p_1$, where
$\xi(z)=2/(z^2+z^{-2})$. If we choose domains so that the maps
$U\xra{p_1}V\xra{\xi}W$ are invertible, then we can use the
schwarzian chain rule to obtain a formula for $S(p_2^{-1})$ on $W$.
It turns out that this is a rational function with poles (of various
orders) at $0$, $1$, $-1$ and $-(1+a^2)/(1-6a^2+a^4)$. Because there
are only four poles, the basic solutions to the differential equation
$f''+S(p_2^{-1})f/2$ can be expressed as products of certain factors
$(z-\al)^{n/8}$ with suitable instances of the Heun $G$-function.
Further details are given in the Maple code, but we will not discuss
them here, because we did not find this representation to be useful.
However, we believe that this is the closest possible connection with
special functions that have previously been named and studied.
\begin{checks}
hyperbolic/schwarz_check.mpl: check_heun()
\end{checks}
\end{remark}
\begin{definition}\label{defn-star-U}
We put
\[ U = \C \sm\{t\,\psi^{-1}(u)\st t\geq 1,\;u\in B\}. \]
\end{definition}
In other words, $U$ is the domain obtained from $\C$ by deleting
rays from the points of $\psi^{-1}(B)$ to $\infty$. This is
simply connected and contains $i\R\cup\Dl$, and the maps $s^*_0$ and
$s^*_1$ are holomorphic on $U$.
\begin{proposition}\label{prop-p-one-section}
There is a unique holomorphic map $g\:U\to\Dl$ satisfying $g(0)=0$
and $p_1g=1$. This satisfies $g(\ov{z})=\ov{g(z)}$ and
$g(-z)=-g(z)$. Moreover, $g(z)$ can be written in the form
$c\,f_1(z)/f_0(z)$, where
\begin{itemize}
\item The maps $f_k$ are holomorphic on $U$ and satisfy
$f''_k+\half(s^*_0+ds^*_1)f_k=0$.
\item $f_0$ is an even function with $f_0(0)=1$ and
$f_0(\ov{z})=\ov{f_0(z)}$.
\item $f_1$ is an odd function with $f'_1(0)=1$ and
$f_1(\ov{z})=\ov{f_1(z)}$.
\item $c$ is a positive real number.
\end{itemize}
\end{proposition}
\begin{proof}
As $U$ contains none of the critical values of $p_1$, we see that
the map $p_1\:p_1^{-1}(U)\to U$ is a covering. We have also seen
that $p_1(0)=0$. As $U$ is simply connected, standard covering
theory tells us that there is a unique section $g\:U\to\Dl$ with
$p_1g=1$ and $g(0)=0$. As $p_1$ is holomorphic, it is easy to check
that the same holds for $g$. As $p_1(-z)=-p_1(z)$ and
$p_1((-1,1))\sse\R$ we see that $g(-z)=-g(z)$ and
$g(\R)\sse(-1,1)$.
Now put $s^*=S(p_1^{-1})$, which has the form $s^*_0+ds_1^*$ as in
Proposition~\ref{prop-p-one}. Let $F$ denote the space of
holomorphic functions on $U$ that satisfy $e''+\half s^*e=0$.
Proposition~\ref{prop-schwarzian-solutions} tells us that there are
unique functions $f_0,f_1\in F$ with $f_0(z)=1+O(z^2)$ and
$f_1(z)=z+O(z^2)$, and these functions form a basis for $F$. From
the definition of $s^*_0$ and $s^*_1$ we see that $s^*(z)=-z$. It
follows that the maps $z\mapsto f_0(-z)$ and $z\mapsto -f_1(-z)$
satisfy the defining conditions of $f_0$ and $f_1$ respectively; so
$f_0$ is even and $f_1$ is odd. Similarly, we see that
$s^*(z)=\ov{s^*(\ov{z})}$, and it follows that
$f_k(z)=\ov{f_k(\ov{z})}$. Proposition~\ref{prop-schwarzian-solutions}
also tells us that $g$ is the quotient of two linearly independent
elements of $F$, so there are constants $A,B,C$ and $D$ such that
$g=(Af_0+Bf_1)/(Cf_0+Df_1)$ and $AD-BC\neq 0$. As $g(0)=0$ we must
have $A=0$, so $Cf_0+Df_1=Bf_1/g$. As both $g$ and $f_1$ are odd we
see that $Cf_0+Df_1$ is even, so $D=0$. Thus, if we put $c=B/C$ we
have $g=c\,f_1/f_0$. By considering derivatives at the origin, we
see that $c$ is real and positive.
\end{proof}
\subsection{Methods for explicit calculation}
\label{sec-schwarz-methods}
The parameters $b$ and $d$ depend on $a$; in this section we will call
them $\bt(a)$ and $\dl(a)$. While
Proposition~\ref{prop-p-one-section} is very satisfactory, we cannot
immediately use it for computation, because we do not know the
values of $\bt(a)$ and $\dl(a)$. We will describe some calculations
that we can do with an arbitrary pair $(b,d)$, which will enable us to
test whether $(b,d)\simeq(\bt(a),\dl(a))$, and to improve the degree
of approximation if necessary. First, we define
$s^*(z)=s_0^*(z)+ds_1^*(z)$. Next, for any open set $V\subset\C$, we
put
\begin{align*}
F^*(V) &= \{f\in\text{Hol}(V)\st f''+s^*f/2=0\} \\
G^*(V) &= \{g\in\text{Mer}(V)\st S(g)=s^*\},
\end{align*}
and note that these are described by
Proposition~\ref{prop-schwarzian-solutions}. In particular, we can
consider $F^*(U)$ and $G^*(U)$, where $U$ is as in
Definition~\ref{defn-star-U}. Just as in
Proposition~\ref{prop-p-one-section}, we see that there is a unique
basis $\{f_0,f_1\}$ for $F^*(U)$ such that $f_k(z)=z^k+O(z^2)$. Both
the power series solution method in Lemma~\ref{lem-F-disc} and the
analytic continuation method in Corollary~\ref{cor-F-disc} are
straightforwardly computable, so we can calculate $f_k(z)$ for any
$z\in U$. (To calculate $f_k(u)$ when $|u|=1$, we have typically
computed power series solutions $f_{kj}(z)$ centred at $ju/10$ for
$0\leq j\leq 10$, and compared then by evaluating $f_{kj}((j+1)u/10)$
and $f'_{kj}((j+1)u/10)$. This works provided that $u$ is not too
close to any of the branch points. If it is very close to a branch
point, then we need to take smaller steps as we approach it.) We also
put $g_0(z)=f_1(z)/f_0(z)$. It is not hard to see that this is odd
with real coefficients, so $g_0(i\R)\sse i\R$.
\begin{definition}\label{defn-circle-fit}
Suppose that
\[ g_0(ie^{it})/i = u_0+iu_1t+u_2t^2+O(t^3). \]
Using the fact that $g_0(-\ov{z})=-\ov{g(z)}$, we see that the
coefficients $u_j$ are real. We put
\begin{align*}
b &= \frac{u_1^2}{\sqrt{8u_0u_1(u_0u_2+u_1^2)+u_1^4}} \\
c &= \sqrt{\frac{u_2}{u_0(u_0u_2+u_1^2)}} \\
g(z) &= c\,g_0(z).
\end{align*}
\end{definition}
These formulae are embedded in the Maple function
\mcode+series_circle_fit(u0,u1,u2)+.
\begin{proposition}
If $d=\dl(a)$ then $b=\bt(a)$ and $g(z)$ is the function described in
Proposition~\ref{prop-p-one-section}. Moreover, for any $z_0$ on the
arc of $S^1$ between $1$ and $(i-a)/(i+a)$, we have
$\text{Im}(\psi(g(z_0)))=0$.
\end{proposition}
\begin{proof}
Because $d=\dl(a)$, we see that $f_0$ and $f_1$ are as described in
Proposition~\ref{prop-p-one-section}. Put $b^*=\bt(a)$, and let
$c^*$ and $g^*$ be the number and the function denoted by $c$ and $g$
in Proposition~\ref{prop-p-one-section}, so $g^*=c^*g_0$; our
task is to prove that $b^*=b$ and $c^*=c$ and $g^*=g$.
We saw in Lemma~\ref{lem-psi-edges} that $C_{HS3}$ is part
of the circle of radius $R=\rt b^*/b^*_-$ centred at the point
$iA=ib^*_+/b^*_-$. We also have
$p_1(C_{HS3})\sse C_{PS3}\sse S^1$, and it follows that
$g^*(ie^{it})\sse C_{HS3}$ for small $t$, or in other words
$|A-c^*g_0(ie^{it})/i|^2=R^2$. This gives
\[ R^2 = (A-c^*u_0-c^*u_2t^2)^2 + u_1^2t^2 + O(t^3), \]
and we can expand out and compare coefficients to get
\begin{align*}
R^2 &= (A-c^*u_0)^2 \\
u_1^2 &= 2(A-c^*u_0)c^*u_2.
\end{align*}
Note also that the definitions $R=\rt b^*/b^*_-$ and $A=b^*/b^*_-$
imply $A^2-R^2=1$ (corresponding to the fact that $C_{HS3}$ crosses
the unit circle orthogonally). It is now an exercise in algebra to
solve these equations giving
\begin{align*}
c^* &= \sqrt{\frac{u_2}{u_0(u_0u_2+u_1^2)}}
= c \\
A &= \frac{u_0u_2 + \half u_1^2}{\sqrt{u_0u_2(u_0u_2+u_1^2)}} \\
R &= \frac{\half u_1^2}{\sqrt{u_0u_2(u_0u_2+u_1^2)}}.
\end{align*}
From the $c^*=c$ we deduce that $g^*=g$. We also have
$A=\rt b^*/\sqrt{1-(b^*)^2}$, which gives $b^*=A/\sqrt{2+A^2}$; after
some further manipulation this gives $b^*=b$.
Finally, let $L$ denote the arc between $1$ and $(i-a)/(i+a)$. This
is part of $C_{PS5}=p_1(C_{HS5})=p\psi^{-1}(C_{H5})$, but
$C_{H5}=(-1,1)$; it follows that $\psi(g(L))\sse (-1,1)$, so
$\text{Im}(\psi(g(L)))=0$.
\begin{checks}
hyperbolic/schwarz_check.mpl: check_schwarz()
\end{checks}
\end{proof}
We thus arrive at the following method. Given $a$ we fix a point
$z_0$ on $S^1$ between $1$ and $(i-a)/(i+a)$. We then choose $d$ and
compute $g_0(z)$. The power series for $g_0(ie^{it})/i$ gives the
coefficients $u_0$, $u_1$ and $u_2$, from which we compute $b$, $c$
and $g(z)$. We then put $\ep(d)=\text{Im}(\psi(z_0))$. If $d=\dl(a)$
then we will have $\ep(d)=0$. If $\ep(d)\neq 0$ then we can adjust
our value of $d$ and try again. As evaluation of $\ep(d)$ is quite
expensive, it is important to use an efficient search algorithm. It
is also useful to retain a lot of information generated in the course
of the calculation which we have found to be awkward when using
Maple's built in \mcode+fsolve+ function. We have therefore used our
own implementation of Brent's method~\cite{br:amw} (closely following
the Matlab implementation by John Burkardt~\cite{bu:br}). This gives
us a value of $d$ such that $\ep(d)=0$ to high accuracy. We have not
given a general proof that $\ep(d)=0$ implies $d=\dl(a)$, but in any
given case it is easy to perform additional checks to verify that this
is the case; for example, we can feed our new value of $b$ into the
method of Section~\ref{sec-a-from-b} and check that everything is
consistent.
Maple commands for the above algorithm were given at the beginning of
Section~\ref{sec-b-from-a}. For more detail, see the comments in the
code.
\subsection{Holomorphic forms}
\label{sec-hol-forms}
Suppose we have constructed an isomorphism $f\:HX(b)\to PX(a)$, given
generically by
\[ f(z)=j(q(z),p(z)). \]
Recall that Proposition~\ref{prop-holomorphic-forms} gives a basis
$\{\om_0,\om_1\}$ for $\Om^1(PX(a))$. Let $m(z)$ be the function on
$\Dl$ given by $f^*(\om_0)=m(z)\,dz$. Note that this is holomorphic
on $\Dl$ and so is given by a power series that converges everywhere
on $\Dl$, unlike the functions $p(z)$ and $p_1(z)$ which have
infinitely many poles. In this section we will investigate the
properties of $m(z)$.
\begin{proposition}\label{prop-m-props}
The function $m(z)$ has the following automorphy properties:
\begin{itemize}
\item[(a)] For $\gm\in\Pi$ we have $m(z)=m(\gm(z))\,\gm'(z)$.
\item[(b)] $m(z)=m(iz)$ (so $m(z)$ is a power series in $z^4$).
\item[(c)] $m(\ov{z})=\ov{m(z)}$ (so the power series for $m(z)$ has
real coefficients).
\item[(d)]
\[ m(\mu(z)) = -\frac{b^2(1-b^2)\,m(z)}{((i-b^2)z+b_+)^2\,p(z)}.
\]
\end{itemize}
Moreover, for a suitable branch of the square root we have
\[ m(z) = p'(z)\, r_a(p(z))^{-1/2} \]
(where $r_a(z)=z(z-a)(z+a)(z-1/a)(z+1/a)$ as in
Definition~\ref{defn-P}).
\end{proposition}
Note that property~(d) means that $p$ can be computed from $m$.
\begin{proof}
\begin{itemize}
\item[(a)] For $\gm\in\Pi$ we have $f\gm=f\:\Dl\to PX(a)$ and so
\[ m(\gm(z))\gm'(z)\,dz = \gm^*(m(z)\,dz) = (f\gm)^*(\om_0) =
f^*(\om_0) =m(z)\,dz.
\]
\item[(b)] We have $f\lm=\lm f\:\Dl\to PX(a)$, and $\lm^*dz=i\,dz$
on $\Dl$ whereas $\lm^*\om_0=i\,\om_0$ on $PX(a)$ (by
Proposition~\ref{prop-holomorphic-forms}). It follows easily from
this that $m(iz)=m(\lm(z))=m(z)$.
\item[(c)] This follows in the same way, using the fact that
$\nu^\#(dz)=dz$ and $\nu^\#(\om_0)=\om_0$.
\item[(d)] Recall that $\mu^*(\om_0)=\om_1$, and on $PX_0(a)$ we
have $\om_1=z\,\om_0$, which means that
\[ f^*(\om_1)=p(z)\,f^*(\om_0)=p(z)m(z)\,dz. \]
On the other hand, as
$f\mu=\mu f$ we have
\[ f^*(\om_1)=\mu^*(\om_0)=m(\mu(z))\mu'(z)\,dz. \]
A calculation
gives $\mu'(z)=-b^2(1-b^2)/((i-b^2)z+b_+)^2$. Putting this
together gives the claimed equation.
\end{itemize}
Finally, $\om_0$ is given on $PX_0(a)$ by $dz/w$, and this gives
$f^*(\om_0)=p'(z)\,dz/q(z)$. Also, as $(p(z),q(z))\in PX_0(a)$ we
have $q(z)=\pm\sqrt{r_a(p(z))}$.
\end{proof}
Using the methods of Sections~\ref{sec-a-from-b}
and~\ref{sec-b-from-a}, we can calculate $p_1(z)$ quite accurately on
a domain including $\psi^{-1}(HF_{16}(b))$, and this lets us calculate
$p(z)$ on $HF_{16}(b)$. Given an arbitrary point $z\in\Dl$ we can use
the method in Remark~\ref{rem-move-inwards} to find $\gm\in\Pi$ such
that $\gm(z)\in HF_1(b)$, then we can find $\bt\in G$ such that
$\bt\gm(z)\in HF_{16}(b)$. Using the automorphy properties of $m$ we
can then find $m(z)$ from $m(\bt\gm(z))$. Given an object \mcode+HP+
of the class \mcode+H_to_P_map+, the method \mcode+HP["m_piecewise",z]+
will calculate $m(z)$ by the above algorithm.
To obtain the power series for $m(z)$, it is best to calculate $m(s)$
for all $s$ in some finite subset $S\subset\C$, and then find a
polynomial $m_0(z)=\sum_{i=0}^da_iz^{4i}$ which minimises
$\sum_s\|m_0(s)-m(s)\|^2$. As this objective function depends
quadratically on the coefficients $a_i\in\R$, the minimisation problem
reduces easily to a matrix calculation. We have generally taken
\[ S = \{re^{k\pi i/(2n)}\st 0\leq k<n\} \]
for some radius $r\in(0,1)$ (say $r=0.95$) and some integer $n>0$ (say
$n=400$). Note that the relation $m(iz)=m(z)$ makes it natural to
consider only sample points in the first quadrant, and the maximum
principle of complex analysis makes it reasonable to consider only
sample points on the boundary of the region where we want our
approximation to be accurate. This algorithm is implemented by the
method \mcode+HP["find_m_series",r,n,d]+.
The following plot shows the curve $m(0.93e^{it})$ with the parameters
$a$ and $b$ that are relevant for $EX^*$. (The real axis is drawn
vertically.)
\begin{center}
\begin{tikzpicture}[scale=.4]
\draw[gray] (-13, -7) rectangle (13,13);
\draw[gray] (-13, 0) -- (13,0);
\draw[gray] ( 0, -7) -- (0,13);
\fill[black] ( 13, 0) circle(0.15);
\fill[black] (-13, 0) circle(0.15);
\fill[black] ( 0, -7) circle(0.15);
\fill[black] ( 0, 13) circle(0.15);
\draw ( 13, 0) node[anchor=west ] {$-13i$};
\draw (-13, 0) node[anchor=east ] {$ 13i$};
\draw ( 0, -8) node[anchor=north] {$ 8$};
\draw ( 0, 13) node[anchor=south] {$ 13$};
\draw[smooth,red] (-0.4839,11.0507) -- (-0.9605,10.9726) -- (-1.4223,10.8445) -- (-1.8627,10.6692) -- (-2.2754,10.4508) -- (-2.6550,10.1940) -- (-2.9970,9.9047) -- (-3.2979,9.5889) -- (-3.5556,9.2533) -- (-3.7687,8.9046) -- (-3.9375,8.5493) -- (-4.0629,8.1938) -- (-4.1471,7.8438) -- (-4.1929,7.5043) -- (-4.2037,7.1796) -- (-4.1836,6.8732) -- (-4.1365,6.5876) -- (-4.0667,6.3245) -- (-3.9780,6.0850) -- (-3.8742,5.8694) -- (-3.7588,5.6775) -- (-3.6348,5.5088) -- (-3.5047,5.3624) -- (-3.3708,5.2373) -- (-3.2349,5.1324) -- (-3.0986,5.0466) -- (-2.9630,4.9788) -- (-2.8292,4.9280) -- (-2.6979,4.8931) -- (-2.5700,4.8735) -- (-2.4459,4.8684) -- (-2.3261,4.8771) -- (-2.2111,4.8991) -- (-2.1013,4.9340) -- (-1.9969,4.9814) -- (-1.8985,5.0411) -- (-1.8063,5.1130) -- (-1.7208,5.1971) -- (-1.6426,5.2935) -- (-1.5724,5.4024) -- (-1.5109,5.5240) -- (-1.4593,5.6585) -- (-1.4187,5.8063) -- (-1.3908,5.9673) -- (-1.3773,6.1417) -- (-1.3803,6.3291) -- (-1.4019,6.5291) -- (-1.4445,6.7406) -- (-1.5108,6.9624) -- (-1.6032,7.1926) -- (-1.7244,7.4287) -- (-1.8764,7.6677) -- (-2.0614,7.9060) -- (-2.2807,8.1394) -- (-2.5351,8.3635) -- (-2.8247,8.5732) -- (-3.1487,8.7635) -- (-3.5052,8.9294) -- (-3.8917,9.0661) -- (-4.3048,9.1693) -- (-4.7403,9.2355) -- (-5.1934,9.2615) -- (-5.6592,9.2455) -- (-6.1325,9.1862) -- (-6.6082,9.0832) -- (-7.0814,8.9369) -- (-7.5475,8.7482) -- (-8.0024,8.5185) -- (-8.4422,8.2497) -- (-8.8635,7.9438) -- (-9.2632,7.6030) -- (-9.6386,7.2297) -- (-9.9870,6.8264) -- (-10.3063,6.3958) -- (-10.5944,5.9408) -- (-10.8495,5.4642) -- (-11.0703,4.9693) -- (-11.2557,4.4593) -- (-11.4051,3.9373) -- (-11.5182,3.4065) -- (-11.5952,2.8698) -- (-11.6365,2.3300) -- (-11.6427,1.7895) -- (-11.6145,1.2503) -- (-11.5524,0.7142) -- (-11.4571,0.1829) -- (-11.3286,-0.3423) -- (-11.1671,-0.8598) -- (-10.9722,-1.3680) -- (-10.7435,-1.8649) -- (-10.4806,-2.3479) -- (-10.1829,-2.8143) -- (-9.8506,-3.2605) -- (-9.4839,-3.6826) -- (-9.0841,-4.0762) -- (-8.6531,-4.4369) -- (-8.1938,-4.7602) -- (-7.7101,-5.0420) -- (-7.2068,-5.2784) -- (-6.6896,-5.4667) -- (-6.1648,-5.6050) -- (-5.6391,-5.6924) -- (-5.1193,-5.7295) -- (-4.6121,-5.7181) -- (-4.1237,-5.6610) -- (-3.6598,-5.5623) -- (-3.2248,-5.4268) -- (-2.8225,-5.2601) -- (-2.4552,-5.0679) -- (-2.1243,-4.8562) -- (-1.8298,-4.6308) -- (-1.5710,-4.3971) -- (-1.3462,-4.1600) -- (-1.1532,-3.9236) -- (-0.9892,-3.6914) -- (-0.8511,-3.4662) -- (-0.7359,-3.2498) -- (-0.6405,-3.0436) -- (-0.5619,-2.8483) -- (-0.4974,-2.6642) -- (-0.4447,-2.4909) -- (-0.4016,-2.3279) -- (-0.3667,-2.1747) -- (-0.3385,-2.0302) -- (-0.3160,-1.8938) -- (-0.2988,-1.7645) -- (-0.2863,-1.6416) -- (-0.2784,-1.5247) -- (-0.2749,-1.4133) -- (-0.2759,-1.3073) -- (-0.2813,-1.2066) -- (-0.2909,-1.1112) -- (-0.3046,-1.0214) -- (-0.3221,-0.9372) -- (-0.3432,-0.8588) -- (-0.3673,-0.7863) -- (-0.3942,-0.7196) -- (-0.4236,-0.6588) -- (-0.4551,-0.6039) -- (-0.4885,-0.5548) -- (-0.5238,-0.5116) -- (-0.5608,-0.4745) -- (-0.5993,-0.4438) -- (-0.6391,-0.4200) -- (-0.6797,-0.4036) -- (-0.7206,-0.3952) -- (-0.7608,-0.3955) -- (-0.7990,-0.4050) -- (-0.8337,-0.4239) -- (-0.8632,-0.4524) -- (-0.8855,-0.4898) -- (-0.8985,-0.5353) -- (-0.9004,-0.5872) -- (-0.8895,-0.6436) -- (-0.8647,-0.7018) -- (-0.8252,-0.7587) -- (-0.7714,-0.8113) -- (-0.7041,-0.8562) -- (-0.6252,-0.8906) -- (-0.5372,-0.9120) -- (-0.4431,-0.9185) -- (-0.3466,-0.9093) -- (-0.2512,-0.8842) -- (-0.1603,-0.8441) -- (-0.0770,-0.7906) -- (-0.0039,-0.7260) -- (0.0573,-0.6530) -- (0.1054,-0.5747) -- (0.1402,-0.4939) -- (0.1620,-0.4136) -- (0.1717,-0.3362) -- (0.1707,-0.2641) -- (0.1607,-0.1988) -- (0.1437,-0.1418) -- (0.1216,-0.0940) -- (0.0966,-0.0557) -- (0.0709,-0.0270) -- (0.0464,-0.0077) -- (0.0251,0.0032) -- (0.0088,0.0066) -- (-0.0012,0.0039) -- (-0.0036,-0.0032) -- (0.0025,-0.0128) -- (0.0173,-0.0229) -- (0.0409,-0.0316) -- (0.0730,-0.0368) -- (0.1127,-0.0367) -- (0.1591,-0.0297) -- (0.2108,-0.0143) -- (0.2663,0.0105) -- (0.3241,0.0458) -- (0.3823,0.0923) -- (0.4392,0.1504) -- (0.4929,0.2206) -- (0.5413,0.3029) -- (0.5823,0.3972) -- (0.6136,0.5032) -- (0.6325,0.6202) -- (0.6366,0.7469) -- (0.6230,0.8816) -- (0.5894,1.0217) -- (0.5333,1.1642) -- (0.4530,1.3050) -- (0.3473,1.4398) -- (0.2163,1.5633) -- (0.0607,1.6704) -- (-0.1172,1.7555) -- (-0.3138,1.8138) -- (-0.5244,1.8408) -- (-0.7433,1.8332) -- (-0.9638,1.7888) -- (-1.1787,1.7071) -- (-1.3806,1.5890) -- (-1.5627,1.4372) -- (-1.7184,1.2559) -- (-1.8425,1.0505) -- (-1.9310,0.8277) -- (-1.9814,0.5947) -- (-1.9931,0.3591) -- (-1.9668,0.1281) -- (-1.9051,-0.0912) -- (-1.8118,-0.2929) -- (-1.6917,-0.4722) -- (-1.5503,-0.6255) -- (-1.3937,-0.7508) -- (-1.2275,-0.8473) -- (-1.0574,-0.9155) -- (-0.8882,-0.9568) -- (-0.7242,-0.9736) -- (-0.5687,-0.9686) -- (-0.4240,-0.9451) -- (-0.2918,-0.9061) -- (-0.1729,-0.8547) -- (-0.0678,-0.7934) -- (0.0236,-0.7247) -- (0.1016,-0.6506) -- (0.1666,-0.5727) -- (0.2190,-0.4923) -- (0.2592,-0.4108) -- (0.2875,-0.3292) -- (0.3039,-0.2486) -- (0.3086,-0.1702) -- (0.3019,-0.0955) -- (0.2840,-0.0259) -- (0.2555,0.0369) -- (0.2175,0.0913) -- (0.1710,0.1356) -- (0.1179,0.1684) -- (0.0602,0.1886) -- (0.0000,0.1954) -- (-0.0602,0.1886) -- (-0.1179,0.1684) -- (-0.1710,0.1356) -- (-0.2175,0.0913) -- (-0.2555,0.0369) -- (-0.2840,-0.0259) -- (-0.3019,-0.0955) -- (-0.3086,-0.1702) -- (-0.3039,-0.2486) -- (-0.2875,-0.3292) -- (-0.2592,-0.4108) -- (-0.2190,-0.4923) -- (-0.1666,-0.5727) -- (-0.1016,-0.6506) -- (-0.0236,-0.7247) -- (0.0678,-0.7934) -- (0.1729,-0.8547) -- (0.2918,-0.9061) -- (0.4240,-0.9451) -- (0.5687,-0.9686) -- (0.7242,-0.9736) -- (0.8882,-0.9568) -- (1.0574,-0.9155) -- (1.2275,-0.8473) -- (1.3937,-0.7508) -- (1.5503,-0.6255) -- (1.6917,-0.4722) -- (1.8118,-0.2929) -- (1.9051,-0.0912) -- (1.9668,0.1281) -- (1.9931,0.3591) -- (1.9814,0.5947) -- (1.9310,0.8277) -- (1.8425,1.0505) -- (1.7184,1.2559) -- (1.5627,1.4372) -- (1.3806,1.5890) -- (1.1787,1.7071) -- (0.9638,1.7888) -- (0.7433,1.8332) -- (0.5244,1.8408) -- (0.3138,1.8138) -- (0.1172,1.7555) -- (-0.0607,1.6704) -- (-0.2163,1.5633) -- (-0.3473,1.4398) -- (-0.4530,1.3050) -- (-0.5333,1.1642) -- (-0.5894,1.0217) -- (-0.6230,0.8816) -- (-0.6366,0.7469) -- (-0.6325,0.6202) -- (-0.6136,0.5032) -- (-0.5823,0.3972) -- (-0.5413,0.3029) -- (-0.4929,0.2206) -- (-0.4392,0.1504) -- (-0.3823,0.0923) -- (-0.3241,0.0458) -- (-0.2663,0.0105) -- (-0.2108,-0.0143) -- (-0.1591,-0.0297) -- (-0.1127,-0.0367) -- (-0.0730,-0.0368) -- (-0.0409,-0.0316) -- (-0.0173,-0.0229) -- (-0.0025,-0.0128) -- (0.0036,-0.0032) -- (0.0012,0.0039) -- (-0.0088,0.0066) -- (-0.0251,0.0032) -- (-0.0464,-0.0077) -- (-0.0709,-0.0270) -- (-0.0966,-0.0557) -- (-0.1216,-0.0940) -- (-0.1437,-0.1418) -- (-0.1607,-0.1988) -- (-0.1707,-0.2641) -- (-0.1717,-0.3362) -- (-0.1620,-0.4136) -- (-0.1402,-0.4939) -- (-0.1054,-0.5747) -- (-0.0573,-0.6530) -- (0.0039,-0.7260) -- (0.0770,-0.7906) -- (0.1603,-0.8441) -- (0.2512,-0.8842) -- (0.3466,-0.9093) -- (0.4431,-0.9185) -- (0.5372,-0.9120) -- (0.6252,-0.8906) -- (0.7041,-0.8562) -- (0.7714,-0.8113) -- (0.8252,-0.7587) -- (0.8647,-0.7018) -- (0.8895,-0.6436) -- (0.9004,-0.5872) -- (0.8985,-0.5353) -- (0.8855,-0.4898) -- (0.8632,-0.4524) -- (0.8337,-0.4239) -- (0.7990,-0.4050) -- (0.7608,-0.3955) -- (0.7206,-0.3952) -- (0.6797,-0.4036) -- (0.6391,-0.4200) -- (0.5993,-0.4438) -- (0.5608,-0.4745) -- (0.5238,-0.5116) -- (0.4885,-0.5548) -- (0.4551,-0.6039) -- (0.4236,-0.6588) -- (0.3942,-0.7196) -- (0.3673,-0.7863) -- (0.3432,-0.8588) -- (0.3221,-0.9372) -- (0.3046,-1.0214) -- (0.2909,-1.1112) -- (0.2813,-1.2066) -- (0.2759,-1.3073) -- (0.2749,-1.4133) -- (0.2784,-1.5247) -- (0.2863,-1.6416) -- (0.2988,-1.7645) -- (0.3160,-1.8938) -- (0.3385,-2.0302) -- (0.3667,-2.1747) -- (0.4016,-2.3279) -- (0.4447,-2.4909) -- (0.4974,-2.6642) -- (0.5619,-2.8483) -- (0.6405,-3.0436) -- (0.7359,-3.2498) -- (0.8511,-3.4662) -- (0.9892,-3.6914) -- (1.1532,-3.9236) -- (1.3462,-4.1600) -- (1.5710,-4.3971) -- (1.8298,-4.6308) -- (2.1243,-4.8562) -- (2.4552,-5.0679) -- (2.8225,-5.2601) -- (3.2248,-5.4268) -- (3.6598,-5.5623) -- (4.1237,-5.6610) -- (4.6121,-5.7181) -- (5.1193,-5.7295) -- (5.6391,-5.6924) -- (6.1648,-5.6050) -- (6.6896,-5.4667) -- (7.2068,-5.2784) -- (7.7101,-5.0420) -- (8.1938,-4.7602) -- (8.6531,-4.4369) -- (9.0841,-4.0762) -- (9.4839,-3.6826) -- (9.8506,-3.2605) -- (10.1829,-2.8143) -- (10.4806,-2.3479) -- (10.7435,-1.8649) -- (10.9722,-1.3680) -- (11.1671,-0.8598) -- (11.3286,-0.3423) -- (11.4571,0.1829) -- (11.5524,0.7142) -- (11.6145,1.2503) -- (11.6427,1.7895) -- (11.6365,2.3300) -- (11.5952,2.8698) -- (11.5182,3.4065) -- (11.4051,3.9373) -- (11.2557,4.4593) -- (11.0703,4.9693) -- (10.8495,5.4642) -- (10.5944,5.9408) -- (10.3063,6.3958) -- (9.9870,6.8264) -- (9.6386,7.2297) -- (9.2632,7.6030) -- (8.8635,7.9438) -- (8.4422,8.2497) -- (8.0024,8.5185) -- (7.5475,8.7482) -- (7.0814,8.9369) -- (6.6082,9.0832) -- (6.1325,9.1862) -- (5.6592,9.2455) -- (5.1934,9.2615) -- (4.7403,9.2355) -- (4.3048,9.1693) -- (3.8917,9.0661) -- (3.5052,8.9294) -- (3.1487,8.7635) -- (2.8247,8.5732) -- (2.5351,8.3635) -- (2.2807,8.1394) -- (2.0614,7.9060) -- (1.8764,7.6677) -- (1.7244,7.4287) -- (1.6032,7.1926) -- (1.5108,6.9624) -- (1.4445,6.7406) -- (1.4019,6.5291) -- (1.3803,6.3291) -- (1.3773,6.1417) -- (1.3908,5.9673) -- (1.4187,5.8063) -- (1.4593,5.6585) -- (1.5109,5.5240) -- (1.5724,5.4024) -- (1.6426,5.2935) -- (1.7208,5.1971) -- (1.8063,5.1130) -- (1.8985,5.0411) -- (1.9969,4.9814) -- (2.1013,4.9340) -- (2.2111,4.8991) -- (2.3261,4.8771) -- (2.4459,4.8684) -- (2.5700,4.8735) -- (2.6979,4.8931) -- (2.8292,4.9280) -- (2.9630,4.9788) -- (3.0986,5.0466) -- (3.2349,5.1324) -- (3.3708,5.2373) -- (3.5047,5.3624) -- (3.6348,5.5088) -- (3.7588,5.6775) -- (3.8742,5.8694) -- (3.9780,6.0850) -- (4.0667,6.3245) -- (4.1365,6.5876) -- (4.1836,6.8732) -- (4.2037,7.1796) -- (4.1929,7.5043) -- (4.1471,7.8438) -- (4.0629,8.1938) -- (3.9375,8.5493) -- (3.7687,8.9046) -- (3.5556,9.2533) -- (3.2979,9.5889) -- (2.9970,9.9047) -- (2.6550,10.1940) -- (2.2754,10.4508) -- (1.8627,10.6692) -- (1.4223,10.8445) -- (0.9605,10.9726) -- (0.4839,11.0507) -- (0.0000,11.0769) -- cycle;
\end{tikzpicture}
\end{center}
We will mention one other approach to the calculation of $m(z)$, but
we will not go into great detail because we have not found it to be
computationally efficient.
\begin{definition}\label{defn-tensor-sections}
Put
\[ A_k = \{f\in\Hol(\Dl)\st f(z) = f(\gm(z)) \gm'(z)^k
\text{ for all } \gm\in\Pi \text{ and } z\in\Dl
\}.
\]
Multiplication by $dz^{\ot k}$ identifies $A_k$ with the space of
holomorphic sections of the $k$'th tensor power of the cotangent
bundle of $HX(b)$. In particular, we have $A_1=\Om^1(HX(b))$.
\end{definition}
\begin{lemma}\label{lem-riemann-roch}
\[ \dim_\C(A_k) = \begin{cases}
0 & \text{ if } k < 0 \\
1 & \text{ if } k = 0 \\
2 & \text{ if } k = 1 \\
2k-1 & \text{ if } k > 1.
\end{cases}
\]
\end{lemma}
\begin{proof}
This is a standard consequence of the Riemann-Roch theorem
(see~\cite[Section IV.1]{ha:ag}, for example). In more
detail, that theorem tells us that for any line bundle $\CL$ over a
compact Riemann surface $Z$ of genus $g$, we have
\[ \dim(H^0(Z;\CL)) - \dim(H^0(Z;\Om^1\ot\CL^*)) =
\deg(\CL) + 1 - g.
\]
Now take $Z=HX(b)$ (so $g=2$) and put
$f(n)=\dim(H^0(HX(b);(\Om^1)^{\ot n}))$ for $n\in\Z$. We have seen
that $\{\om_0,\om_1\}$ is a basis for $H^0(HX(b);\Om^1)$ over $\C$,
so $f(1)=2$. On any compact Riemann surface, the only holomorphic
$\C$-valued functions are constant, so $f(0)=1$. The Riemann-Roch
theorem gives $f(n)-f(1-n)=n\,\deg(\Om^1)-1$. Taking $n=1$ gives
$\deg(\Om^1)=2$, so we get $f(n)-f(1-n)=2n-1$. Moreover, as
$\deg(\Om^1)>0$ we have $f(n)=0$ for $n<0$. Thus, when $n>1$ we have
$f(1-n)=0$ and so $f(n)=2n-1$ as required.
\end{proof}
\begin{definition}\label{defn-automorphic-series}
For $j\geq 0$ and $k\geq 2$ we put
\[ p_{jk}(z) = \sum_{\gm\in\Pi} \gm(z)^j\gm'(z)^k. \]
\end{definition}
It is a standard theorem that the above series is absolutely uniformly
convergent on compact subsets of $\Dl$. We will give a proof that
includes explicit bounds in the case of interest.
\begin{definition}\label{defn-automorphic-area}
For any subset $A\sse\Pi$ we put
\[ a(A) =
\text{area}\left(\bigcup_{\gm\in A}\gm(HF_1(b))\right) =
\sum_{\gm\in A}\text{area}(\gm(HF_1(b))).
\]
(Here areas are defined in terms of the Euclidean metric on $\Dl$,
not the hyperbolic metric.) We also write
$a(\gm)=a(\{\gm\})=\text{area}(\gm(HF_1(b)))$. In particular,
$a(1)$ is just the area of $HF_1(b)$.
\end{definition}
\begin{remark}\label{rem-automorphic-area}
We can find $a(1)$ as follows. It is easy to see that $a(1)$ is $8$
times the area of $HF_8(b)$. The boundary of this region consists
of straight lines joining $v_0=0$ to $v_1$ and $v_{11}$, together
with an arc of the circle $C_3$ joining $v_{11}$ to $v_{13}$, and an
arc of the circle $C_7$ joining $v_1$ to $v_{13}$. Recall that $C_3$
has centre $a_3=b_+$ and radius $r_3=\sqrt{|a_3|^2-1}=b$, whereas
$C_7$ has centre $a_7=b_+^{-1}+\half ib_+$ and radius
$r_7=\sqrt{|a_7|^2-1}=\half b_-^2b_+^{-1}$. One can check that
$v_{11}=a_3-r_3$ and $v_1=a_7-r_7$ and
\[ v_{13} = a_3 - r_3 e^{-i\tht} = a_7 - i\,r_7 e^{-i\tht} \]
where $\tht=2\arctan(b(b_+-b))$. Using this, it is easy to
parameterise the boundary of $HF_8(b)$ and use Green's theorem (in the
form $\text{area}(D)=-\oint_{\partial D}y\,dx$) to calculate the
area. We eventually arrive at the following formula:
\[ a(1) =
3+b^2
- \frac{2bb_-^2}{b_+} - \frac{b_-^4}{b_+^2}\frac{\pi}{2}
- \frac{3b^4+6b^2-1}{1+b^2}\tht.
\]
The graph is as follows:
\begin{center}
\begin{tikzpicture}[scale=6]
\draw[->] (-0.03,0) -- (1.05,0);
\draw[->] (0,-0.03) -- (0,0.55);
\draw[smooth,red] (0.000,0.455) -- (0.100,0.464) -- (0.200,0.480) -- (0.300,0.493) -- (0.400,0.498) -- (0.500,0.491) -- (0.600,0.472) -- (0.700,0.440) -- (0.800,0.396) -- (0.900,0.340) -- (1.000,0.273);
\draw (0.2,0) -- (0.2,-0.02);
\draw (0.4,0) -- (0.4,-0.02);
\draw (0.6,0) -- (0.6,-0.02);
\draw (0.8,0) -- (0.8,-0.02);
\draw (1.0,0) -- (1.0,-0.03);
\draw (0,0.1) -- (-0.02,0.1);
\draw (0,0.2) -- (-0.02,0.2);
\draw (0,0.3) -- (-0.02,0.3);
\draw (0,0.4) -- (-0.02,0.4);
\draw (0,0.5) -- (-0.03,0.5);
\draw (1.05,0) node[anchor=west] {$b$};
\draw (0.8,0.47) node {$a(1)$};
\draw (0,-0.03) node[anchor=north] {$0$};
\draw (1,-0.03) node[anchor=north] {$1$};
\draw (-0.03,0.0) node[anchor=east] {$0$};
\draw (-0.03,0.5) node[anchor=east] {$\pi/2$};
\end{tikzpicture}
\end{center}
\begin{checks}
hyperbolic/HX_check.mpl: check_F1_area()
\end{checks}
\end{remark}
\begin{proposition}\label{prop-automorphic-conv}
Fix $k\geq 2$ and $\dl>0$, and let $K$ be the closed disc of radius
$1-\dl$ centred at the origin. Then the series defining $p_{jk}(z)$
converges absolutely and uniformly on $K$. More precisely, for any
subset $A\sse\Pi$ and any $z\in K$ we have
\[ \left|p_{jk}(z)-\sum_{\gm\in A}\gm(z)^j\gm'(z)^k\right|
\leq \sum_{\gm\in\Pi\sm A} |\gm'(z)|^k
\leq\dl^{-2k}a(\Pi\sm A)/a(1) = \dl^{-2k}(\pi - a(A))/a(1).
\]
\end{proposition}
\begin{proof}
Put $F=HF_1(b)$. Note that the images $\{\gm(F)\st\gm\in\Pi\}$
cover $\Dl$, and any two of these images have intersection of measure
zero, so
\[ \sum_{\gm\in\Pi}a(\gm)=\text{area}(\Dl)=\pi. \]
Next, if we write $\gm(z)$ in the form $(az+b)/(cz+d)$ with
$ad-bc=1$, then we find that $\gm'(z)=(cz+d)^{-2}$, so the Jacobian
determinant is $|cz+d|^{-4}$. It follows that
$a(\gm)=\iint_{F}|cz+d|^{-4}$. Corollary~\ref{cor-bound-ii} tells
us that $|cz+d|\leq(\sqrt{2}+1)|c|$, so
$|c|^{-4}\leq(\sqrt{2}+1)^4|cz+d|^{-4}$, so $|c|^{-4}a(1)\leq a(\gm)$.
Now let $z$ be any point in $K$. Clearly $|\gm(z)^j|\leq 1$. We have
$|cz+d|^{-2k}=|c|^{-2k}|z+d/c|^{-2k}$. Corollary~\ref{cor-bound-ii}
also gives $|d/c|\geq 1$ so for $z\in K$ we have
$|z+d/c|^{-2k}\leq\dl^{-2k}$. The same result also gives $|c|\geq 1$
and $k\geq 2$ by assumption so
$|c|^{-2k}\leq|c|^{-4}\leq a(\gm)/a(1)$. Putting this together gives
$|\gm(z)^j\gm'(z)^k|\leq\dl^{-2k}a(\gm)/a(1)$. Taking the sum over $\gm$
gives
\[ \sum_{\gm\not\in A}|\gm'(z)|^{-k}\leq\dl^{-2k}a(\Pi\sm A)/a(1). \]
\end{proof}
\begin{corollary}\label{cor-automorphic-series}
$p_{jk}\in A_k$ for all $j$ and $k$.
\end{corollary}
\begin{proof}
The proposition shows that the series for $p_{jk}(z)$ is absolutely
uniformly convergent on compact subsets of $\Dl$. This validates the
following manipulation:
\begin{align*}
p_{jk}(\dl(z))\dl'(z)^k &=
\sum_{\gm\in\Pi} (\gm\dl)(z)^j \gm'(\dl(z))^k\dl'(z)^k \\
&= \sum_{\gm\in\Pi} (\gm\dl)(z)^j (\gm\dl)'(z)^k \\
&= \sum_{\ep\in\Pi} \ep(z)^j\ep'(z)^k = p_{jk}(z).
\end{align*}
\end{proof}
Next, the isomorphism $A_2=\Gamma(HX(b);\Om^{\ot 2})$ gives rise to an
action of $G$ on $A_2$; we will write $\gm^\bullet f$ for the action
of $\gm$ on $f$. Note that the functions $m(z)$ and
\[ n(z) = (\mu^\bullet m)(z) = m(\mu(z))\mu'(z) \]
give a basis for $A_1$. It follows that the functions $m^2$, $n^2$
and $mn$ are linearly independent in $A_2$, and $A_2$ has dimension
$3$ by Lemma~\ref{lem-riemann-roch}, so the indicated elements must
give a basis. In particular, we have $p_{02}=a_0m^2+a_1n^2+a_2mn$ for
some constants $a_i$. One can check that $p_{02}$, $m^2$ and $n^2$
are fixed by $\lm$ whereas $mn$ is negated, so $a_2=0$. Using the
automorphy properties of $m$ together with the relation
$\bt_6\mu(v_1)=v_0$ one can check that $m(v_1)=n(v_0)=0$ and
\[ n(v_1)/m(v_0) = (\bt_6\mu)'(v_1) = 2i/(1-b^2). \]
From this we obtain $a_1=a^*a_0$, where
\[ a^* = -\tfrac{1}{4}(1-b^2) p_{02}(v_1)/p_{02}(v_0). \]
From this it follows that $p_{02}-a^*\mu^\bullet p_{02}$ is a constant
multiple of $m^2$. Thus, if we can calculate $p_{02}$ effectively,
then we can recover the function $m$ up to a constant multiplier,
without using the methods in Sections~\ref{sec-a-from-b}
and~\ref{sec-b-from-a}. We can then find the map $p$ from the
relation $p=(\mu^\bullet m)/m$, noting that the unknown constant
cancels out.
However, in practice we need an extremely large number of terms to
calculate $p_{02}$ accurately, so this is not an efficient approach.
Various tactics are available to streamline the calculation, but they
are not sufficient to change the conclusion.
The above algorithm is implemented by the class
\mcode+automorphy_system+, which is declared in the file
\fname+hyperbolic/automorphic.mpl+. In more detail, we can enter the
following:
\begin{mcodeblock}
AS := `new/automorphy_system`();
AS["a_H"] := a_H0;
AS["poly_deg"] := 100;
AS["band_number"] := 4;
AS["set_p0_series",2]:
AS["set_m_series"]:
AS["m_series"](z);
\end{mcodeblock}
The last line will give a polynomial $m^*(z)$ of degree $100$ which
approximates $m(z)/m(0)$. It is based on a calculation of $p_{02}(z)$
obtained by summing over a certain subset of $\Pi$. More
specifically, we can put $B_0=\{1\}$ and
\[ B_1 = \{\gm\in\Pi \st \gm(HF_1(b))\cap HF_1(b)\neq\emptyset\}, \]
then we can define $B_n=B_1.B_{n-1}$ recursively. There are $25$
elements in $B_1$, and it is not hard to list them explicitly, and
then to give an algorithm which enumerates $B_n$ for all $n$. The
line
\begin{mcodeblock}
AS["band_number"] := 4;
\end{mcodeblock}
specifies that sums should be taken over the set $B_4$, which has
$156772$ elements. We find that $|m(0)m^*(z)-m(z)|\leq 10^{-5}$ for
$|z|\leq 0.5$, but the error grows to about $10^{-3}$ when $|z|=0.65$,
and becomes very large when $|z|>0.8$.
\section{The embedded family}
\label{sec-E}
\subsection{Geometry behind the definition}
\label{sec-E-geometry}
\begin{definition}\label{defn-X}
Fix $a\in (0,1)$. For $x=(x_1,x_2,x_3,x_4)\in\R^4$ we put
\begin{align*}
\rho(x) &= x_1^2+x_2^2+x_3^2+x_4^2 = \sum_i x_i^2 \\
f_1(x) &= 2 x_2^2 + (x_4-1-x_3/a)^2 \\
f_2(x) &= 2 x_1^2 + (x_4-1+x_3/a)^2 \\
f(x) &= f_1(x)f_2(x) \\
EX(a) &= \{x\in\R^4\st \rho(x)=1 \text{ and } f(x)=f(-x) \}.
\end{align*}
\end{definition}
Straightforward algebra shows that this is the same as the definition
in the Introduction.
Now put
\[ \Om^+_1 = \{x\in S^3 \st f_1(x)=0\}. \]
Recall that $f_1(x)=2 x_2^2 + (x_4-1-x_3/a)^2$, and a sum of two real
squares can only be zero if the individual terms are zero. It follows that
\begin{align*}
\Om^+_1 &= \{x\in S^3 \st x_2=0 \text{ and } x_4=1+x_3/a\} \\
&= \{(x_1,0,x_3,1+x_3/a)\in\R^4 \st x_1^2+x_3^2+(1+x_3/a)^2=1\}.
\end{align*}
This is the intersection of $S^3$ with a two-dimensional affine
subspace of $\R^4$, so it is a circle. This circle passes through
the point $(0,0,0,1)$, which corresponds to infinity under the
stereographic projection map $s\:S^3\to\R^3\cup\{\infty\}$ that we
defined in the Introduction. This means that the image $s(\Om^+_1)$
is a ``circle through $\infty$'', or in other
words a straight line. In fact, one can check that
\[ s(\Om^+_1) = \{(x,y,z)\in\R^3\st y=0,z=-a\}. \]
Similarly, the set
\[ \Om^+_2 = \{x\in S^3 \st f_2(x)=0\} \]
is another circle in $S^3$, with stereographic projection
\[ s(\Om^+_2) = \{(x,y,z)\in\R^3\st x=0,z=+a\}. \]
\begin{checks}
embedded/geometry_check.mpl: check_omega()
\end{checks}
Note that the lines $s(\Om^+_1)$ and $s(\Om^+_2)$ are at right angles
to each other, but they do not touch except at $\infty$. We can also
put
\[ \Om^+ = \{x\in S^3\st f(x)=0\} = \{x\in S^3\st f_1(x)f_2(x)=0\}
= \Om^+_1\cup\Om^+_2.
\]
Now recall that the definition of $X$ also involves $f(-x)$, so we
should study the sets $\Om^-_i=\{x\in S^3\st f_i(-x)=0\}$ and
$\Om^-=\{x\in S^3\st f(-x)=0\}=\Om^-_1\cup\Om^-_2$. The sets
$\Om_i^-$ are again circles in $S^3$, but they do not pass through
$(0,0,0,1)$ so their stereographic projections are circles rather than
straight lines. In fact one can check that
\begin{align*}
s(\Om^-_1) &= \{(x,0,z)\in\R^3\st x^2+(z-1/(2a))^2=1/(2a)^2\} \\
&= \text{ the circle of radius $1/(2a)$ in the $(x,z)$-plane centred at $(0,0,1/(2a))$} \\
s(\Om^-_2) &= \{(0,y,z)\in\R^3\st y^2+(z+1/(2a))^2=1/(2a)^2\} \\
&= \text{ the circle of radius $1/(2a)$ in the $(y,z)$-plane centred at $(0,0,-1/(2a))$}.
\end{align*}
\begin{checks}
embedded/geometry_check.mpl: check_omega()
\end{checks}
The picture for $a=1/\rt$ is as follows:
\[ \includegraphics[scale=0.25,clip=true,
trim=6cm 6cm 6cm 6cm]{images/Omega.jpg} \]
(The set $s(\Omega^+)$ is shown in red, and the set $s(\Omega^-)$ is
shown in blue.)
On $\Om^+$ we have $f(x)=0$ and $f(-x)>0$, whereas on $\Om^-$ we have
$f(x)>0$ and $f(-x)=0$. We defined $EX(a)=\{x\in S^3\st f(x)=f(-x)\}$, and
we now see that this fits between $\Om^+$ and $\Om^-$. This can be
displayed as follows:
\[ \includegraphics[scale=0.25,clip=true,
trim=6cm 6cm 6cm 6cm]{images/XOmega.jpg}
\]
It will be convenient to describe $EX(a)$ using the functions $g$ and
$g_0$ given below.
\begin{definition}\label{defn-g}
We put
\begin{align*}
g_0(x) &= (f(x)-f(-x))/8 + x_4(\rho(x)-1) \\
&= ((1+a^{-2})x_3^2-2)x_4+a^{-1}(x_1^2-x_2^2)x_3 \\
g(x) &= (f(x)-f(-x))/8 - x_4(\rho(x)-1) \\
&= (a^{-2}-1)x_3^2x_4-2(x_1^2+x_2^2)x_4-2x_4^3+a^{-1}(x_1^2-x_2^2)x_3.
\end{align*}
(The advantage of $g_0(x)$ is that it has few terms, and the advantage
of $g(x)$ is that it is a homogeneous cubic.)
\begin{checks}
embedded/geometry_check.mpl: check_g()
\end{checks}
It is now straightforward to check that
\[ EX(a) = \{x\in S^3 \st g_0(x)=0\} = \{x\in S^3\st g(x)=0\}. \]
We put
\begin{align*}
\tA &= \R[x_1,x_2,x_3,x_4] \\
A &= \CO_{EX(a)} = \tA/(\rho(x)-1,g(x)),
\end{align*}
so $A$ is the ring of polynomial functions on $EX(a)$.
\end{definition}
\begin{remark}
Maple notation for the parameter $a$ is $a_E$. The global variable
\mcode+a_E0+ is set to $1/\rt$, and \mcode+a_E1+ is a 100 digit
approximation to that. Elements of $\R^4$ are represented in Maple
as lists of length $4$. The functions $\rho(x)$, $f_1(x)$, $f_2(x)$,
$f(x)$, $g_0(x)$ and $g(x)$ are \mcode+rho(x)+, \mcode+f_1(x)+,
\mcode+f_2(x)+, \mcode+f(x)+, \mcode+g_0(x)+ and \mcode+g(x)+. The
functions obtained from these by setting $a=1/\rt$ are \mcode+f_10(x)+,
\mcode+f_20(x)+, \mcode+f0(x)+, \mcode+g_00(x)+ and \mcode+g0(x)+.
Note in particular the difference between \mcode+g_0(x)+ and
\mcode+g0(x)+.
\end{remark}
\begin{proposition}\label{prop-X-smooth}
The space $EX(a)$ is a compact smooth oriented embedded submanifold in
$S^3$.
\end{proposition}
We will use the following notation:
\begin{definition}\label{defn-nx}
We put
\begin{align*}
n(x) &= (\nabla g)_x =
(\partial g/\partial x_1,\dotsc,\partial g/\partial x_4) \\
&= ( 2x_1(x_3/a-2x_4),\;
-2x_2(x_3/a+2x_4), \\
&\qquad
2(a^{-2}-1)x_3x_4+a^{-1}(x_1^2-x_2^2),\;
(a^{-2}-1)x_3^2-6x_4^2-2(x_1^2+x_2^2)).
\end{align*}
(This is \mcode+dg(x)+ in Maple.)
\end{definition}
\begin{proof}
Define $p\:\R^4\to\R^2$ by $p(x)=(\rho(x)-1,g(x))$, so
$EX(a)=p^{-1}\{0\}$. We must first show that $0$ is a regular value
of $p$, or equivalently that the gradients of $\rho-1$ and $g$ are
linearly independent at every point in $EX(a)$. Note here that the
gradient of $\rho-1$ at $x$ is just $2x$, which is certainly nonzero
at all points in $EX(a)$. Moreover, as $g$ is a homogeneous cubic
function we have $(2x).n(x)=6g(x)$, which is zero on $EX(a)$, so the
two gradients are orthogonal. It will thus be enough to show that
$n(x)$ is nonzero everywhere on $EX(a)$. By direct expansion one can
check that
\[ \frac{1-a^2}{16}(n(x)_1^2+n(x)_2^2) +
\frac{a}{2}(x_1^2-x_2^2)n(x)_3 -
\frac{1}{4}(x_1^2+x_2^2)n(x)_4 =
x_1^4+x_2^4+(\tfrac{5}{2}-a^2)(x_1^2+x_2^2)x_4^2.
\]
\begin{checks}
embedded/geometry_check.mpl: check_g()
embedded/EX_check.mpl: check_smoothness()
\end{checks}
If $n(x)=0$ then the left hand side vanishes so the right hand side
must also vanish. After noting that $a\in (0,1)$ so
$\frac{5}{2}-a^2>0$, it follows that $x_1=x_2=0$. After substituting
this back into the equations for $n(x)$, we see that $x_3x_4=0$ and
$(a^{-2}-1)x_3^2-6x_4^2=0$, which easily implies that $x_3=x_4=0$.
Thus, the only place in $\R^4$ where $n(x)=0$ is the origin, so in
particular there are no points in $EX(a)$ with this property.
This completes the proof that $0$ is a regular value of $p$, which
implies in a standard way that $EX(a)$ is a smooth closed submanifold
of $\R^4$. It is compact because it is closed in $S^3$. Now let
$\om_k$ denote the standard volume form in $\Lm^k(\R^k)$. Standard
exterior algebra now tells us that for each $x\in EX(a)$ there is a
unique element $\al_x\in\Lm^2(T_xEX(a))<\Lm^2(\R^4)$ such that
$\al_x\wedge p^*(\om_2)=\om_4$. These forms give a smooth, nowhere
vanishing section of $\Lm^2(T)$ over $EX(a)$, and thus an orientation
of $EX(a)$.
\end{proof}
\begin{remark}\label{rem-move-to-X}
The above considerations also give an efficient practical method to
compute points on $X$ numerically. For any $y\in\R^4$, we define
\[ \sg(y) =
\frac{1}{\|y\|}\left(y - \frac{g(y)}{\|n(y)\|^2}n(y)\right).
\]
One can show that if the distance from $y$ to $X$ is of order
$\ep\ll 1$, then the distance from $\sg(y)$ to $X$ is of order
$\ep^2$. This implies that the sequence $(\sg^k(x))_{k\geq 0}$
converges rapidly to a point $\sg^\infty(x)\in X$. This is
implemented by the Maple function \mcode+move_to_X(x)+.
\end{remark}
Next, we use the metric and the orientation to give an almost complex
structure on $X$. Explicitly, for each $x\in X$ there is a unique
isometric linear map $J_x\:T_xX\to T_xX$ such that $\ip{u,J_xu}=0$ for
all $u$, and $u\wedge J_xu$ is a positive multiple of $\al_x$ for all
$u\neq 0$. This satisfies $J_x^2=-1$ and so gives a complex structure
on $T_xX$. All this can be verified easily after choosing an oriented
orthonormal basis. As mentioned in the introduction, this can be
integrated to give a complex structure on $X$. In more detail, if
$U\sse X$ is open and $q\:U\to\C$ is a smooth map, we say that $q$ is
\emph{holomorphic} (or \emph{conformal}) if for each $x\in U$, the
derivative $Dq_x\:T_xX\to\C$ is $\C$-linear. It is a nontrivial fact
that for each $x\in X$ there is an open neighbourhood $U$ of $X$ and
an injective holomorphic function $q\:U\to\C$ such that
$Dq_y\:T_yX\to\C$ is an isomorphism for all $y\in U$. This has been
known at least since the work of Korn and Lichtenstein in~1916. A
modern proof is given in~\cite{deka:srt}. As we mentioned in the
introduction, it can also be seen as a special case of the
Newlander-Nirenberg theorem. It is clear that the transition map
between any pair of holomorphic charts as above, is a function on an
open domain in $\C$ that is holomorphic in the traditional sense. We
can thus use these charts to regard $X$ as a Riemann surface.
\subsection{The group action}
\label{sec-E-G}
We define linear isometric maps $\lm,\mu,\nu:\R^4\to\R^4$ as follows:
\begin{align*}
\lm(x_1,x_2,x_3,x_4) &= ( -x_2,\pp x_1,\pp x_3, -x_4) \\
\mu(x_1,x_2,x_3,x_4) &= (\pp x_1, -x_2, -x_3, -x_4) \\
\nu(x_1,x_2,x_3,x_4) &= (\pp x_1, -x_2,\pp x_3,\pp x_4).
\end{align*}
Maple notation for $g.x$ is \mcode+act_R4[g](x)+.
It is straightforward to check that
\[ \lm^4=\mu^2=\nu^2=(\mu\nu)^2=(\mu\lm)^2=(\nu\lm)^2=1, \]
so we have an isometric action of our group $G$ on $\R^4$.
\begin{checks}
embedded/EX_check.mpl: check_R4_action()
\end{checks}
All three of the generators have determinant $-1$ and so reverse the
orientation of $\R^4$. Note also that $\lm^2\mu\nu(x)=-x$.
Directly from the definitions we see that
\begin{align*}
\rho(\lm(x)) &= \rho(\mu(x))=\rho(\nu(x))=\rho(x) \\
g(\lm(x)) &= g(\mu(x))=-g(x) \\
g(\nu(x)) &= g(x)
\end{align*}
This means that $\lm$, $\mu$ and $\nu$ send $X$ to itself.
\begin{checks}
embedded/EX_check.mpl: check_symmetry()
\end{checks}
\begin{remark}\label{rem-heegard}
If we put $D_+=\{x\in S^3\st g(x)\geq 0\}$ and
$D_-=\{x\in S^3\st g(x)\leq 0\}$ then we have $D_+\cup D_-=S^3$ and
$D_+\cap D_-=EX(a)$, and the map $\mu$ gives a homeomorphism between
$D_+$ and $D_-$. Moreover, $D_+$ and $D_-$ are handlebodies, with
the figure eight curve $\Om^-$ as a deformation retract of $D_+$, and
$\Om^+$ as a deformation retract of $D_-$. This gives an unusually
explicit Heegaard splitting of $S^3$ along $EX(a)$.
\end{remark}
Recall that the orientation form $\al_x\in\Lm^2(T_xX)$ is
characterised by the equation $\al_x\wedge (\rho-1,g)^*(\om_2)=\om_4$.
As $\lm$ preserves $\rho-1$ and changes the sign of $g$ and $\om_4$,
we deduce that $\lm^*\al=\al$, so $\lm$ preserves orientation. As the
complex structure on $EX(a)$ is determined by the metric and the
orientation, it follows that $\lm$ is a conformal automorphism of
$EX(a)$. By the same kind of logic, the map $\mu$ preserves
orientation and is conformal, whereas $\nu$ reverses orientation and
is anticonformal.
\subsection{Isotropy}
\label{sec-E-isotropy}
We now define points $v_i\in\R^4$ as follows:
\begin{align*}
v_{ 0} &= (\pp 0,\pp 0,\pp 1,\pp 0) &
v_{ 6} &= (\pp 1,\pp 1,\pp 0,\pp 0)/\rt \\
v_{ 1} &= (\pp 0,\pp 0, -1,\pp 0) &
v_{ 7} &= ( -1,\pp 1,\pp 0,\pp 0)/\rt \\
v_{ 2} &= (\pp 1,\pp 0,\pp 0,\pp 0) &
v_{ 8} &= ( -1, -1,\pp 0,\pp 0)/\rt \\
v_{ 3} &= (\pp 0,\pp 1,\pp 0,\pp 0) &
v_{ 9} &= (\pp 1, -1,\pp 0,\pp 0)/\rt \\
v_{ 4} &= ( -1,\pp 0,\pp 0,\pp 0) &
v_{10} &= (0,0,\pp\sqrt{\frac{2a^2}{1+a^2}},\pp\sqrt{\frac{1-a^2}{1+a^2}}) \\
v_{ 5} &= (\pp 0,-1,\pp 0,\pp 0) &
v_{11} &= (0,0,\pp\sqrt{\frac{2a^2}{1+a^2}}, -\sqrt{\frac{1-a^2}{1+a^2}}) \\
&&
v_{12} &= (0,0, -\sqrt{\frac{2a^2}{1+a^2}}, -\sqrt{\frac{1-a^2}{1+a^2}}) \\
&&
v_{13} &= (0,0, -\sqrt{\frac{2a^2}{1+a^2}},\pp\sqrt{\frac{1-a^2}{1+a^2}})
\end{align*}
These are \mcode+v_E[i]+ in Maple.
Routine calculation shows that these points all lie in $X$, and that
the action of $G$ permutes them, according to the following rules:
\begin{align*}
\lm &\mapsto (2\;3\;4\;5)\;(6\;7\;8\;9)\;(10\;11)\;(12\;13) \\
\mu &\mapsto (0\;1)\;(3\;5)\;(6\;9)\;(7\;8)\;(10\;12)\;(11\;13) \\
\nu &\mapsto (3\;5)\;(6\;9)\;(7\;8).
\end{align*}
(For example, the cycle $(2\;3\;4\;5)$ in $\lm$ means that
$\lm(v_2)=v_3$, $\lm(v_3)=v_4$, $\lm(v_4)=v_5$ and $\lm(v_5)=v_2$.)
This almost shows that we have a precromulent surface, except that we
need to check that there are no further points with nontrivial
stabiliser in $D_8$.
\begin{proposition}\label{prop-no-more-isotropy}
If $x\in EX(a)\sm\{v_0,\dotsc,v_{13}\}$ then $\stab_{D_8}(x)=1$.
\end{proposition}
\begin{proof}
Consider for example a point $x\in EX(a)$ with $\lm^2(x)=x$, or in other
words $(-x_1,-x_2,x_3,x_4)=(x_1,x_2,x_3,x_4)$ or $x_1=x_2=0$. The
equations for $EX(a)$ become $x_3^2+x_4^2=1$ and
$x_4((a^{-2}-1)x_3^2-2x_4^3)=0$. If $x_4=0$ we must have
$x=(0,0,\pm 1,0)$, so $x=v_0$ or $x=v_1$. Otherwise we must have
$(a^{-2}-1)x_3^2=2x_4^2$. In conjunction with $x_3^2+x_4^2=1$ this gives
$x_3^2=2a^2/(1+a^2)$ and $x_4^2=(1-a^2)/(1+a^2)$ so
$x\in\{v_{10},v_{11},v_{12},v_{13}\}$. Note that if $x$ is fixed by
$\lm$ or $\lm^3=\lm^{-1}$ then it is certainly fixed by $\lm^2$, so
again it is one of the $v_i$.
Similarly:
\begin{itemize}
\item We have $\mu(x)=x$ iff $x_2=x_3=x_4=0$, and $\lm^2\mu(x)=x$
iff $x_1=x_3=x_4=0$. In these cases it is clear that
$x\in\{v_2,v_3,v_4,v_5\}$.
\item We have $\lm\mu(x)=x$ iff $x_1=x_2$ and $x_3=0$. In this
context we have $g_0(x)=-2x_4$, so we must also have $x_4=0$, which
makes it clear that $x\in\{v_6,v_8\}$.
\item A similar argument shows that if $\lm^3\mu(x)=x$ then
$x\in\{v_7,v_9\}$.
\end{itemize}
\begin{checks}
embedded/EX_check.mpl: check_fixed_points()
\end{checks}
\end{proof}
\begin{remark}\label{rem-tangent}
It will be useful to understand the tangent space $T_{v_0}X$ more
explicitly. The formula in Definition~\ref{defn-nx} shows that the
gradient of $g$ at $v_0$ is
\[ n(v_0) = n(e_3) = (0,0,0,a^{-2}-1) = (a^{-2}-1)e_4. \]
On the other hand, we have $\nabla(\rho-1)_{v_0}=2v_0=2e_3$, so
\[ (\rho-1,g)^*(\om_2) =
\nabla(\rho-1)_{v_0}\wedge\nabla(g)_{v_0} =
2(a^{-2}-1)e_3\wedge e_4.
\]
The tangent space is the orthogonal complement to $v_0$ and $n(v_0)$,
so it is spanned by the basis vectors $e_1$ and $e_2$ (which are
orthonormal). The orientation form $\al_{v_0}$ must therefore be
$(e_1\wedge e_2)/(2(a^{-2}-1))$, which is a positive multiple of
$e_1\wedge e_2$, so $J_{v_0}(e_1)=e_2$. In other words, the map
$(x+iy)\mapsto(x,y,0,0)$ gives an isometric $\C$-linear isomorphism
$\C\to T_{v_0}X$. It is also clear from this that $\lm$ acts on
$T_{v_0}EX(a)$ as multiplication by $i$.
\end{remark}
\subsection{Associated complex varieties}
\label{sec-E-complex}
For $x\in\C^4$ we again put $\rho(x)=\sum_jx_j^2$ (not
$\sum_j|x_j|^2$). We then put
\begin{align*}
CEX(a) &= \{x\in\C^4 \st g(x)=0,\;\rho(x)=1\} \\
PEX(a) &= \{[x]\in\C P^3\st g(x)=0\} \\
PEX'(a) &= \{[x]\in\C P^3\st g(x)=0,\;\rho(x)\neq 0\}.
\end{align*}
It is clear that $CEX(a)$ is an affine variety, and $PEX(a)$ is a
projective variety, and $PEX'(a)$ is a quasiprojective open subvariety
of $PEX(a)$. The map $x\mapsto [x]$ gives a double covering
$CEX(a)\to PEX'(a)$. We can identify $EX(a)$ with the set of real
points in $CEX(a)$.
\begin{proposition}
If $a\neq 1/\rt$, then $CEX(a)$, $PEX(a)$ and $PEX'(a)$ are all
smooth.
\end{proposition}
\begin{proof}
As $PEX'(a)$ is open in $PEX(a)$ and $CEX(a)$ is a double cover of
$PEX'(a)$, it will suffice to treat the case of $PEX(a)$. The
partial derivatives of $g$ are
\begin{align*}
n_1 = \partial g/\partial x_1 &= \pp 2x_1(x_3/a - 2x_4) \\
n_2 = \partial g/\partial x_2 &= -2x_2(x_3/a + 2x_4) \\
n_3 = \partial g/\partial x_3 &= 2(a^{-2}-1)x_3x_4 + (x_1^2-x_2^2)/a \\
n_4 = \partial g/\partial x_4 &= (a^{-2}-1)x_3^2 - 2x_1^2 - 2x_2^2 - 6x_4^2.
\end{align*}
Put
\[ U_i = \{x\in\C^4\st x_i=1,\; n_1=n_2=n_3=n_4=0\}. \]
By well-known arguments, it will suffice to check that the sets $U_i$
are all empty. Consider a point $x\in U_1$. The equation $n_1=0$
gives $x_3=2ax_4$, and we can substitute this in $n_2=0$ to get
$x_2x_4=0$. If $x_2\neq 0$ then this gives $x_4=0$ so $x_3=0$, and
the relation $n_3=0$ gives a contradiction. We must therefore have
$x_2=0$. Putting $x_1=1$ and $x_2=0$ and $x_3=4ax_4$ in $n_3=n_4=0$
we get $2a^2=1$ and $2x_4^2=-1$. We are assuming explicitly that
$a\neq 1/\rt$ and implicitly that $a\in(0,1)$, so this is
impossible. We conclude that $U_1=\emptyset$, and a similar argument
gives $U_2=\emptyset$. Now consider $x\in U_3$. If we had
$x_1\neq 0$ then $x/x_1$ would be in $U_1$, which is impossible.
Thus $x_1=0$, and similarly $x_2=0$, and $x_3=1$ by assumption.
Substituting this into $n_3=n_4=0$ gives a contradiction. A similar
argument works for $U_4$.
\begin{checks}
embedded/EX_check.mpl: check_PEX_smoothness()
\end{checks}
\end{proof}
\begin{proposition}
None of the surfaces $CEX(1/\rt)$, $PEX(1/\rt)$ or $PEX'(1/\rt)$ is
smooth. Moreover, $PEX(1/\rt)$ is isomorphic to the singular Cayley
cubic with equation
\[ X_1X_2X_3 + X_1X_2X_4 + X_1X_3X_4 + X_2X_3X_4 = 0. \]
\end{proposition}
\begin{remark}
Although the isomorphism with the Cayley cubic is interesting, it
interacts poorly with the underlying real structure, and so is not
too helpful for studying the cromulent surface $EX(1/\rt)$.
\end{remark}
\begin{proof}
At the point $w=(\sqrt{-2},0,\rt,1)$ we find that $\rho(w)=1$ and
$g(w)=0$ and all partial derivatives of $g$ also vanish. It follows
that $w$ is a singular point in $CEX(1/\rt)$ and that $[w]$ is a
singular point in $PEX'(1/\rt)\subset PEX(1/\rt)$. The same holds
for all points in the $G$-orbit of $w$. (One can check that
$\nu(w)=w$ and $\lm^2\mu\nu(w)=-w$ and $|G.w|=8$ and $|G.[w]|=4$.)
Now put
\begin{align*}
X_1 &= x_4 + x_3/\rt + x_1\sqrt{-2} &
X_2 &= x_4 + x_3/\rt - x_1\sqrt{-2} \\
X_3 &= x_4 - x_3/\rt + x_2\sqrt{-2} &
X_4 &= x_4 - x_3/\rt - x_2\sqrt{-2}.
\end{align*}
We find that
\[ X_1X_2X_3 + X_1X_2X_4 + X_1X_3X_4 + X_2X_3X_4 = -2g(x), \]
so this gives an isomorphism with the Cayley cubic.
\begin{checks}
embedded/roothalf/cayley_surface_check.mpl: check_cayley_surface();
\end{checks}
\end{proof}
\subsection{The ring of functions}
\label{sec-E-functions}
\begin{definition}\label{defn-yzu}
We put
\begin{align*}
y_1 &= x_3 & y_2 &= (x_2^2 - x_1^2 - (a^{-1}+a)x_3x_4)/(2a) \\
z_1 &= y_1^2 & z_2 &= y_2^2
\end{align*}
and
\begin{align*}
u_1 &= (1-2ay_2)/2 - \half(y_2-a)(y_2-a^{-1})y_1^2 \\
u_2 &= (1+2ay_2)/2 - \half(y_2+a)(y_2+a^{-1})y_1^2 \\
u_3 &= 4u_1u_2
= (1-z_1-z_1z_2)^2 - z_2((a+a^{-1})z_1-2a)^2 \\
u_4 &= u_1+u_2 = 1 - z_1 - z_1z_2
\end{align*}
\end{definition}
\begin{checks}
embedded/invariants_check.mpl: check_invariants()
\end{checks}
\begin{proposition}\label{prop-OX-basis}
The ring $A$ of polynomial functions on $EX(a)$ can be described as
\[ A = \R[y_1,y_2][x_1,x_2]/(x_1^2-u_1,x_2^2-u_2), \]
with $x_3=y_1$ and $x_4=-y_1y_2$. The set
\[ M = \{x_1^ix_2^jy_1^ky_2^l\st i,j,k,l\in\{0,1\}\} \]
is a basis for $A$ over the subring $\R[z_1,z_2]$.
\end{proposition}
\begin{proof}
We have $x_3=y_1$ by definition, and the relation $g_0(x)=0$ can
easily be rearranged as $x_4=-y_1y_2$. We also have
\begin{align*}
x_1^2+x_2^2 &= 1-x_3^2-x_4^2 = 1-y_1^2-y_1^2y_2^2 \\
x_1^2-x_2^2 &= -2ay_2-(a+a^{-1})x_3x_4
= -2ay_2+(a+a^{-1})y_1^2y_2.
\end{align*}
Adding these equations gives $x_1^2=u_1$, and subtracting them gives
$x_2^2=u_2$. We now see that $A=\R[y_1,y_2]/(x_i^2-u_i)$. It is
clear from this that $\{1,x_1,x_2,x_1x_2\}$ is a basis for $A$ over
$\R[y_1,y_2]$, and $\{1,y_1,y_2,y_1y_2\}$ is a basis for
$\R[y_1,y_2]$ over $\R[z_1,z_2]$, so $M$ is a basis for $A$ over
$\R[z_1,z_2]$.
\end{proof}
\begin{remark}
The global symbols \mcode+x+, \mcode+y+, \mcode+z+ (and various
others) are protected (by a call to the \mcode+protect+ command in
the file \fname+Rn.mpl+). This prevents the user from assigning
values to these symbols, which is necessary in for our use of
Gr\"obner bases to work properly. Variables such as \mcode+x0+ or
\mcode+X+ can be used instead of \mcode+x+ for storing points in
$\R^4$. The variable \mcode+xx+ is set equal to
\mcode+[x[1],x[2],x[3],x[4]]+ (a list of length $4$, whose entries
are unassigned symbols). Similarly, \mcode+yy+ is
\mcode+[y[1],y[2]]+, and \mcode+zz+ is \mcode+[z[1],z[2]]+. Note
although $z_1=x_3^2$, the symbol \mcode+z[1]+ does not have the value
\mcode+x[3]^2+. Instead, there is another variable \mcode+zx[1]+
with the value \mcode+x[3]^2+. If we have an expression \mcode+m+
involving \mcode+z[1]+ and \mcode+z[2]+, we can convert it to an
expression in \mcode+x[1]+ to \mcode+x[4]+ using the syntax
\mcode+subs({z[1]=zx[1],z[2]=zx[2]},m)+ or
\mcode+eval(subs({z=zx},m))+. Some esoteric features of Maple mean
that the second form will not work correctly without \mcode+eval()+.
Similarly, there are variables \mcode+yx[i]+ which contain
expressions for $y_i$ in terms of $x_j$, and variables \mcode+uy[1]+,
\mcode+uy[2]+, \mcode+uz[3]+ and \mcode+uz[4]+ which contain
expressions for $u_i$ in terms of $y_j$ or $z_j$. We can also regard
the rule $x\mapsto y$ as giving a function $\R^4\to\R^2$. Maple
notation for this function is \mcode+y_proj(x)+, and
\mcode+z_proj(x)+ is similar.
\end{remark}
\begin{remark}
The functions \mcode+NF_x+, \mcode+NF_y+ and \mcode+NF_z+ can be used
to simplify elements of the ring $A$, by reducing them modulo a
suitable Gr\"obner basis. The function \mcode+NF_x+ will convert any
expression to one that involves only the variables $x_i$, whereas
\mcode+NF_y+ converts $x$'s to $y$'s as far as possible, and
similarly for \mcode+NF_z+.
\end{remark}
It is straightforward to check that $G$ acts on $A$ as follows:
\begin{align*}
\lm^*(x_1) &= -x_2 & \mu^*(x_1) &= \pp x_1 & \nu^*(x_1) &= \pp x_1 \\
\lm^*(x_2) &= \pp x_1 & \mu^*(x_2) &= -x_2 & \nu^*(x_2) &= -x_2 \\
\lm^*(x_3) &= \pp x_3 & \mu^*(x_3) &= -x_3 & \nu^*(x_3) &= \pp x_3 \\
\lm^*(x_4) &= -x_4 & \mu^*(x_4) &= -x_4 & \nu^*(x_4) &= \pp x_4 \\
\lm^*(y_1) &= \pp y_1 & \mu^*(y_1) &= -y_1 & \nu^*(y_1) &= \pp y_1 \\
\lm^*(y_2) &= -y_2 & \mu^*(y_2) &= \pp y_2 & \nu^*(y_2) &= \pp y_2 \\
\lm^*(z_1) &= \pp z_1 & \mu^*(z_1) &= \pp z_1 & \nu^*(z_1) &= \pp z_1 \\
\lm^*(z_2) &= \pp z_2 & \mu^*(z_2) &= \pp z_2 & \nu^*(z_2) &= \pp z_2.
\end{align*}
In particular, the group acts as the identity on $\R[z_1,z_2]$, and
permutes the set $M\cup(-M)$. Some of this can also be expressed in
terms of the characters listed in Proposition~\ref{prop-characters}:
\begin{align*}
\gm^*(x_3) &= \chi_2(\gm) x_3 &
\gm^*(x_4) &= \chi_3(\gm) x_4 \\
\gm^*(y_1) &= \chi_2(\gm) y_1 &
\gm^*(y_2) &= \chi_1(\gm) y_2 \\
\gm^*(x_1x_2) &= \chi_4(\gm)x_1x_2.
\end{align*}
This makes it easy to analyse the invariants for various subgroups of
$G$. The most important cases are as follows:
\begin{proposition}\label{prop-invariants}
$A^G=\R[z_1,z_2]$ and $A^{\ip{\lm^2,\nu}}=\R[y_1,y_2]$.
\end{proposition}
\begin{proof}
Any element $a\in A$ can be written uniquely as
$a=a_0+a_1x_1+a_2x_2+a_3x_1x_2$ with $a_0,\dotsc,a_3\in\R[y_1,y_2]$.
We then find that
\begin{align*}
(\lm^2)^*(a) &= a_0 - a_1x_1 - a_2x_2 + a_3x_1x_2 \\
\nu^*(a) &= a_0 + a_1x_1 - a_2x_2 - a_3x_1x_2,
\end{align*}
so $a$ is invariant under $\ip{\lm^2,\nu}$ if and only if
$a_1=a_2=a_3=0$ and $a=a_0\in\R[y_1,y_2]$. If this holds, we can
write $a$ uniquely as $b_0+b_1y_1+b_2y_2+b_3y_1y_2$ with
$b_0,\dotsc,b_3\in\R[z_1,z_2]$. We then find that
\begin{align*}
\lm^*(a) &= b_0 + b_1y_1 - b_2y_2 - b_3y_1y_2 \\
\mu^*(a) &= b_0 - b_1y_1 + b_2y_2 - b_3y_1y_2,
\end{align*}
so $a$ is invariant under all of $G$ if and only if $b_1=b_2=b_3=0$
and $a=b_0\in\R[z_1,z_2]$.
\end{proof}
\subsection{The curve system}
\label{sec-E-curves}
In this section we will construct a curve system for $EX(a)$.
\begin{definition}\label{defn-slices}
We put $X_k=\{x\in EX(a)\st x_k=0\}$.
\end{definition}
\begin{proposition}\label{prop-slices-a}
The fixed points of antiholomorphic elements of $G$ are as follows:
\begin{align*}
EX(a)^{\mu\nu} &= \{x\in EX(a)\st x_3=x_4=0\} = X_3\cap X_4 \\
EX(a)^{\lm\nu} &= \{x\in EX(a)\st x_1=x_2,\;x_4=0\} \sse X_4 \\
EX(a)^{\lm^3\nu} &= \{x\in EX(a)\st x_1=-x_2,\;x_4=0\} \sse X_4 \\
EX(a)^{\lm^2\nu} &= \{x\in EX(a)\st x_1=0\} = X_1 \\
EX(a)^\nu &= \{x\in EX(a)\st x_2=0\} = X_2.
\end{align*}
\end{proposition}
\begin{proof}
Immediate from formulae for the action of the relevant group elements
on $\R^4$.
\end{proof}
\begin{lemma}\label{lem-a-order}
Put $a^*=\sqrt{(a^{-2}-1)/2}$, so $2at^2+a-a^{-1}=0$ iff $t=\pm a^*$.
Then:
\begin{itemize}
\item If $0<a<1/\rt$ we have
\[ -a^* < 0 < a < \frac{1}{2a} < a^* < \frac{1}{a}. \]
\item If $a=1/\rt$ we have
\[ -a^* < 0 < a = \frac{1}{2a} = a^* < \frac{1}{a}. \]
\item If $1/\rt<a<1$ we have
\[ -a^* < 0 < a^* < \frac{1}{2a} < a < \frac{1}{a}. \]
\end{itemize}
\end{lemma}
\begin{proof}
Straightforward.
\end{proof}
\begin{definition}\label{defn-T-alg}
For all $a\in(0,1)$ we put
\begin{align*}
T_{\alg}^- &= [-a^*,0] \\
T_{\alg}^+ &= [\min(1/(2a),a^*),\max(1/(2a),a^*)]
= \begin{cases}
[1/(2a),a^*] & \text{ if } 0 < a \leq 1/\rt \\
[a^*,1/(2a)] & \text{ if } 1/\rt \leq a < 1.
\end{cases} \\
T_{\alg} &= T_{\alg}^- \cup T_{\alg}^+
\end{align*}
Note that $1/a$ is never in $T_{\alg}$, and $a$ is in $T_{\alg}$ if
and only if $a=1/\rt$. Note also that $T_{\alg}^+$ and
$T_{\alg}^-$ are nonempty disjoint closed sets.
\end{definition}
\begin{proposition}\label{prop-T-alg}
Consider a point $x\in X_1$ (so $x_1=0$), and define
$y_1=x_3$ and $y_2=(x_2^2-(a+a^{-1})x_3x_4)/(2a)$ as in
Definition~\ref{defn-yzu}. Then
\begin{align*}
y_1^2(y_2-a)(y_2-a^{-1}) &= -2a(y_2-1/(2a)) \tag{A} \\
x_2^2(y_2-a)(y_2-a^{-1}) &= 2ay_2(y_2^2-(a^*)^2) \tag{B} \\
\left(y_1^2-\frac{2}{a^{-2}+1}\right)(y_2-a)(y_2-a^{-1})
&= \frac{2}{a^{-2}+1}((a^*)^2-y_2^2) \tag{C}
\end{align*}
Thus $y_2\in T_{\alg}^+\amalg T_{\alg}^-$, and if
$y_2\in T_{\alg}^-$ then
\[ y_1 \in \left[-1,-\sqrt{\frac{2}{a^{-2}+1}}\right] \amalg
\left[\sqrt{\frac{2}{a^{-2}+1}},1\right].
\]
\end{proposition}
\begin{proof}
Recall from Proposition~\ref{prop-OX-basis} that $x_1^2=u_1$, where
\[ u_1 = (1-2ay_2)/2 - \half(y_2-a)(y_2-a^{-1})y_1^2. \]
Here $x_1=0$ so $u_1=0$, and this can be rearranged to give
equation~(A). Next, as $x_1=0$ and $x_3=y_1$ and $x_4=-y_1y_2$
we have $x_2^2=1-x_3^2-x_4^2=1-(1+y_2^2)y_1^2$. We can thus take the
equation
\[ (y_2-a)(y_2-a^{-1})=y_2^2-(a+a^{-1})y_2+1 \tag{D} \]
and subtract $(1+y_2^2)$ times equation~(A) to get equation~(B).
Alternatively, we can subtract $2/(a^{-2}+1)$ times~(D) from~(A) and
rearrange slightly to get~(C). Now consider the signs of the left
and right hand sides of~(A), bearing in mind Lemma~\ref{lem-a-order};
it follows that we must have $y_2\in T_{\alg}$. Suppose in fact that
$y_2\in T_{\alg}^-$, so $-a^*\leq y_2\leq 0$. On the left hand side
of~(C) the factors $y_2-a^{\pm 1}$ are both strictly negative, and
on the right hand side $(a^*)^2-y_2^2\geq 0$. It therefore follows
from~(D) that $y_1^2\geq 2/(a^{-2}-1)$. On the other hand, we also
have $y_1^2=x_3^2=1-x_2^2-x_4^2\leq 1$. The relation
\[ y_1 \in \left[-1,-\sqrt{\frac{2}{a^{-2}+1}}\right] \amalg
\left[\sqrt{\frac{2}{a^{-2}+1}},1\right]
\]
is now clear.
\end{proof}
\begin{remark}\label{rem-T-alg}
We have $\lm(X_1)=\lm^{-1}(X_1)=X_2$ and $\lm^*(y_1)=y_1$ and
$\lm^*(y_2)=-y_2$. Using this we deduce that when $x\in X_2$ we have
$-y_2\in T_{\alg}$, and if $-y_2\in T_{\alg}^-$ we again have
\[ y_1 \in \left[-1,-\sqrt{\frac{2}{a^{-2}+1}}\right] \amalg
\left[\sqrt{\frac{2}{a^{-2}+1}},1\right].
\]
\end{remark}
\begin{definition}\label{defn-C-star}
We put
\begin{align*}
C_0^* &= \{(\cos(t),\;\sin(t),\;0,\;0)\st t\in\R\} \\
C_1^* &= \{(\pp\sin(t)/\rt,\;\sin(t)/\rt,\;\cos(t),\;0)\st t\in\R\} \\
C_2^* &= \{( -\sin(t)/\rt,\;\sin(t)/\rt,\;\cos(t),\;0)\st t\in\R\} \\
C_3^* &= \{x\in X_1\;|\hspace{0.75em} \pp y_2\in T_{\alg}^+\} \\
C_4^* &= \{x\in X_2\;|\hspace{0.30em} -y_2\in T_{\alg}^+\} \\
C_5^* &= \{x\in X_2\;|\hspace{0.30em} -y_2\in T_{\alg}^-,\; y_1\geq 0\} \\
C_6^* &= \{x\in X_1\;|\hspace{0.75em} \pp y_2\in T_{\alg}^-,\; y_1\geq 0\} \\
C_7^* &= \{x\in X_2\;|\hspace{0.30em} -y_2\in T_{\alg}^-,\; y_1\leq 0\} \\
C_8^* &= \{x\in X_1\;|\hspace{0.75em} \pp y_2\in T_{\alg}^-,\; y_1\leq 0\}.
\end{align*}
\end{definition}
It is straightforward to check that $C_k=C_k^*$ for $k\in\{0,1,2\}$.
Using Proposition~\ref{prop-T-alg} and Remark~\ref{rem-T-alg} we see
that
\begin{align*}
X^{\lm^2\nu} &= X_1 = C_3^* \amalg C_6^* \amalg C_8^* \\
X^\nu &= X_2 = C_4^* \amalg C_5^* \amalg C_7^*.
\end{align*}
Next, recall that $C_3$ is the component of $X^{\lm^2\nu}=X_1$
containing $v_{11}$. One can check that $y_2=a^*$ at $v_{11}$, so
$v_{11}\in C_3^*$. By connectivity, it follows that
$C_3\sse C_3^*$. The same line of argument shows that
$C_k\sse C_k^*$ for all $k\in\{3,\dotsc,8\}$, and $C_k$ will be the
same as $C_k^*$ if $C_k^*$ is connected. To prove that this holds
we need to parameterise $C_k^*$. We will do this in two different
ways.
\begin{definition}\label{defn-c-alg}
We put
\[ c_{\alg}(t) =
\left(0,
\sqrt{\frac{(2at^2+a-a^{-1})t}{(t-a)(t-a^{-1})}},
\sqrt{\frac{1-2at}{(t-a)(t-a^{-1})}},
-t\sqrt{\frac{1-2at}{(t-a)(t-a^{-1})}}
\right).
\]
One can check using Lemma~\ref{lem-a-order} that this defines a
continuous map $c_{\alg}\:T_{\alg}\to\R^4$ except in
the case $a=1/\rt$, when the domain of $c_{\alg}$ is
$T_{\alg}\sm\{a\}=T_{\alg}^-$.
\end{definition}
The map $c_{\alg}(t)$ is defined in \fname+embedded/extra_curves.mpl+
as \mcode+c_algebraic(t)+.
We will show that the map $c_{\alg}$ is essentially inverse to
$y_2\:X_1\to T_{\alg}$. The formula is forced by the identities in
Proposition~\ref{prop-T-alg}. A precise statement of the key property
is as follows.
\begin{proposition}\label{prop-c-alg}
The image of $c_{\alg}$ is contained in $X_1$, and we have
$y_2(c_{\alg}(t))=t$. Conversely, consider a point $x\in X_1$, so
$y_2\in T_{\alg}$. Except in the case where $y_2=a=1/\rt$,
there is an element $\gm\in\{1,\mu,\nu,\mu\nu\}$ such that
$x=\gm(c_{\alg}(y_2))$. If $x_2,x_3\geq 0$ then we can take $\gm=1$.
\end{proposition}
\begin{proof}
First, it is straightforward to check that $\rho(c_{\alg}(t))=1$ and
$g_0(c_{\alg}(t))=0$ so the image of $c_{\alg}$ is contained in
$X_1$. When $x_3\neq 0$ we have $y_2=-x_4/x_3$, and using this we
see that $y_2(c_{\alg}(t))=t$ except possibly when $t=1/(2a)$, but
that case can be recovered by continuity.
For the converse, consider a point $x\in X_1$, and suppose we are not
in the exceptional case where $y_2=a=1/\rt$, so
$(y_2-a)(y_2-a^{-1})\neq 0$ and the identities
in Proposition~\ref{prop-T-alg} can be rearranged to give formulae
for $x_2^2$ and $x_3^2$ in terms of $y_2$. Recall also
that $x_4=-x_3y_2$. Using this, we see that the point
$x'=c_{\alg}(y_2)$ satisfies $x_2=\pm x'_2$ and $x_3=\pm x'_3$.
Recall also that
\begin{align*}
\mu(0,x_2,x_3,x_4) &= (0, -x_2, -x_3, -x_4) \\
\nu(0,x_2,x_3,x_4) &= (0, -x_2,\pp x_3,\pp x_4),
\end{align*}
and that $y_2$ is invariant under $\mu$ and $\nu$. The claim now
follows easily.
\end{proof}
It would be possible to prove $C_k=C_k^*$ using only $c_{\alg}$, with
a slight digression to cover the case $a=1/\rt$, but we prefer to
use a full curve system instead.
\begin{definition}\label{defn-E-curves}
We define maps $c_k\:\R\to\R^4$ as follows.
\begin{align*}
c_0(t) &= (\cos(t),\sin(t),0,0) \\
c_1(t) &= (\sin(t)/\rt,\sin(t)/\rt,\cos(t),0) \\
c_2(t) &= \lm(c_1(t)) \\
p_3(t) &= (1+a^2)\sin(t)^2 +
\sqrt{(1+a^2)(1-a^2+2a^2\sin(t)^2)+(1-a^2)^2\cos(t)^4} \\
c_3(t) &= \left(0,
\sqrt{\frac{2(1-a^2)+4a^2\sin(t)^2}{p_3(t)}}\sin(t),\right.\\
&\qquad \left.\sqrt{\frac{2}{1+a^{-2}}}\cos(t),
\sqrt{\frac{2}{1+a^{-2}}}\frac{a^{-1}-a+2a\sin(t)^2}{p_3(t)}\cos(t)\right)\\
c_4(t) &= \lm(c_3(t)) \\
\tau_5(t) &= -\sqrt{(a^{-2}-1)/2}\sin(t/2)^2 \\
p_5(t) &= (\tau_5(t)-a)(\tau_5(t)-a^{-1}) \\
c_5(t) &= \left(
\frac{(a^{-2}-1)^{3/4}}{2^{5/4}}\sqrt{\frac{a(1+\sin(t/2)^2)}{p_5(t)}}\sin(t),
0, \sqrt{\frac{1-2a\tau_5(t)}{p_5(t)}},
-\tau_5(t)\sqrt{\frac{1-2a\tau_5(t)}{p_5(t)}}
\right) \\
c_6(t) &= \lm(c_5(t)); \quad
c_7(t) = \mu(c_5(t)); \quad
c_8(t) = \lm(\mu(c_5(t))).
\end{align*}
Maple notation for these is \mcode+c_E[k](t)+.
\end{definition}
\begin{remark}
We have written $p_3(t)$ in a form which makes it clear that it is
always strictly positive, and using this it is not hard to see
that $c_3(t)$ is well-defined. It is also clear that
$\tau_5(t)\leq 0$ and so $p_5(t)\geq 1$, which in turn implies that
$c_5(t)$ is well-defined. The map $c_5$ is essentially
$c_{\alg}\circ\tau_5$ except that $c_{\alg}$ involves nonnegative
square roots of certain quantities, whereas $c_5$ uses different
branches of these roots that are sometimes negative. A similar
approach would be possible for $c_3$ but would run into problems as
$a$ passes through $1/\rt$.
\end{remark}
\begin{proposition}\label{prop-E-curves}
The images of the above maps lie in $EX(a)$, and they give a curve
system for $EX(a)$. Moreover, we have $c_k(\R)=C_k=C_k^*$ for all
$k$.
\end{proposition}
\begin{proof}
We first need to check that $c_k(\R)\sse EX(a)$, or equivalently
$\rho(c_k(t))=1$ and $g_0(c_k(t))=0$. This is easy for
$k\in\{0,1,2\}$. The algebra is harder for $k\in\{3,5\}$ but there
is no conceptual difficulty, it is just a lengthy exercise in
trigonometric simplification. The remaining cases $k\in\{4,6,7,8\}$
follow from the cases $k\in\{3,5\}$ using the group action.
Next, it is elementary to check all the group transformation
equations in axiom~(c) of Definition~\ref{defn-curve-system}, and to
check all the identities $c_i(\tht)=v_j$ corresponding to the
nonempty boxes in axiom~(b).
We now address axiom~(a). The cases $k\in\{0,1,2\}$ are again easy,
and the cases $k\in\{4,6,7,8\}$ will follow from $k\in\{3,5\}$ by
symmetry. Suppose that $c_3(t)=c_3(u)$. By looking at the third
component, we see that $\cos(t)=\cos(u)$, and thus that
$\sin(t)=\pm\sin(u)$, and thus that $p_3(t)=p_3(u)$. After recalling
that $p_3>0$ and inspecting the second component we deduce that
$\sin(t)=\sin(u)$, so $t-u\in 2\pi\Z$. Similarly, if $c'_3(t)=0$
then by looking at the third component we see that $\sin(t)=0$, so
$t=n\pi$ for some $n\in\Z$. This means that the functions $\cos$ and
$\sin^2$ are both constant to first order near $t$, so the same is
true of $p_3$. By inspecting the second component of $c_3$ we deduce
that $\sin$ must also be constant to first order, so $\cos(t)=0$,
which is impossible. This proves all claims for $c_3$.
For $c_5$, we first observe that the map $y_2\:EX(a)\to\R$ is
invariant under $\mu$ and $\nu$ and satisfies
$y_2(c_5(t))=-\tau_5(t)$. Thus, if $c_5(t)=c_5(u)$ then
$\tau_5(t)=\tau_5(u)$, which easily gives $\cos(t)=\cos(u)$. Given
this, the rest of the argument is essentially the same as for $c_3$.
This completes the proof of axiom~(a).
Now note that $c_3(\R)$ is contained in $X^{\lm^2\nu}=X_1$ and is
connected and contains $c_3(0)=v_{11}$, but $C_3$ is defined to be
the component of $v_{11}$ in $X^{\lm^2\nu}$, so
$c_3(\R)\sse C_3\sse C_3^*$. Moreover, $y_2(c_3(\R))$ is a connected
subset of $T_{\alg}^+$ containing both of the endpoints
$a^*=y_2(c_3(0))$ and $1/(2a)=y_2(c_3(\pi/2))$, so
$y_2(c_3(\R))=T_{\alg}^+$. Thus, if $x\in C_3^*$ then there exists
$t\in\R$ with $y_2(c_3(t))=y_2(x)$, and it follows that
$x=\gm(c_3(t))$ for some $\gm\in\{1,\mu,\nu,\mu\nu\}$. As
$\mu(c_3(t))=c_3(t+\pi)$ and $\nu(c_3(t))=c_3(-t)$ we deduce that
$x\in c_3(\R)$. In conclusion, we have $c_3(\R)=C_3=C_3^*$. Similar
arguments give $c_k(\R)=C_k=C_k^*$ for all $k\in\{3,\dotsc,8\}$.
Recall also that the sets $C_3^*$, $C_6^*$ and $C_8^*$ are disjoint
(immediately from the definitions).
Proposition~\ref{prop-empty-boxes} therefore guarantees that we have
a curve system.
\end{proof}
\begin{proposition}\label{prop-slices}
$X_3=C_0$ and $X_4=C_0\cup C_1\cup C_2$.
\end{proposition}
\begin{proof}
First, it is straightforward to check that $X_3\supseteq C_0$ and
$X_4\supseteq C_0\cup C_1\cup C_2$. For the converse, suppose that
$x\in X_3$. Then the relation $x_4=-y_1y_2=-x_3y_2$ shows that
$x_4$ vanishes as well as $x_3$, and we also have $\|x\|=1$ so
$x\in C_0$. Suppose instead that $x\in X_4\sm X_3$. As
$x_4=0\neq x_3$ the relation $g_0(x)=0$ becomes
$2(x_1-x_2)(x_1+x_2)x_3=0$ and we can divide by $x_3$ to get
$x_1=\pm x_2$. We also have $\|x\|=1$ and it follows easily that
$x\in C_1\cup C_2$.
\end{proof}
\begin{remark}\label{rem-curves-roothalf}
The formulae for $c_3$ and $c_5$ simplify significantly in the case $a=1/\rt$:
\begin{align*}
c_3(t) &= \left(0,\sin(t),\sqrt{2/3}\cos(t),-\sqrt{1/3}\cos(t)\right) \\
c_5(t) &= \left(-\sin(t),0,2^{3/2},\cos(t)-1\right)/\sqrt{10-2\cos(t)}.
\end{align*}
The following picture shows all the curves $c_k(t)$ in that case.
\[ \includegraphics[scale=0.25,clip=true,
trim=6cm 6cm 6cm 6cm]{images/curves_E.jpg}
\]
\end{remark}
\subsection{Fundamental domains}
\label{sec-E-fundamental}
In this section we define and study retractive fundamental domains in
$EX(a)$ for certain subgroups of $G$. It is convenient to start with
an easy case:
\begin{definition}\label{defn-F-two}
We put $H_2=\{1,\lm^2\nu\}$ and recall that $C_2=\{1,\lm^2\}$, noting
also that
\begin{align*}
\lm^2(x) &= (-x_1,\;-x_2,\;x_3,\;x_4) \\
\lm^2\nu(x) &= (-x_1,\;\pp x_2,\;x_3,\;x_4).
\end{align*}
We then put $F_2=\{x\in EX(a)\st x_1\geq 0\}$, and define
$r_2\:EX(a)\to F_2$ by
\[ r_2(x) = (|x_1|,\;x_2,\;x_3,\;x_4). \]
\end{definition}
\begin{proposition}\label{prop-F-two}
$F_2$ is a retractive fundamental domain for $H_2$, with retraction
$r_2$. Moreover, $F_2$ is also a non-retractive fundamental domain
for $C_2$.
\end{proposition}
\begin{proof}
Clear.
\end{proof}
We will see later that $F_2$ is homeomorphic to a disc with two holes.
\begin{definition}\label{defn-F-four}
We put $H_4=\{1,\lm^2,\nu,\lm^2\nu\}$, recalling that
\begin{align*}
\lm^2(x) &= (-x_1,\;-x_2,\;x_3,\;x_4) \\
\nu(x) &= (\pp x_1,\;-x_2,\;x_3,\;x_4) \\
\lm^2\nu(x) &= (-x_1,\;\pp x_2,\;x_3,\;x_4).
\end{align*}
We then put $F_4=\{x\in EX(a)\st x_1,x_2\geq 0\}$, and define
$r_4\:EX(a)\to F_4$ by
\[ r_4(x) = (|x_1|,\;|x_2|,\;x_3,x_4). \]
We also put
\[ F_4^* = \{y\in\R^2\st u_1,u_2\geq 0\}, \]
where $u_1$ and $u_2$ are defined in terms of $y_1$ and $y_2$ as in
Definition~\ref{defn-yzu}:
\begin{align*}
u_1 &= (1-2ay_2)/2 - \half(y_2-a)(y_2-a^{-1})y_1^2 \\
u_2 &= (1+2ay_2)/2 - \half(y_2+a)(y_2+a^{-1})y_1^2.
\end{align*}
\end{definition}
\begin{proposition}\label{prop-F-four}
$F_4$ is a retractive fundamental domain for $H_4$, with retraction
$r_4$. Moreover, there is a map $p_4\:EX(a)\to F_4^*$ given by
\[ p_4(x) = (x_3,(x_2^2 - x_1^2 - (a^{-1}+a)x_3x_4)/(2a))=(y_1,y_2), \]
and a map $s_4\:F_4^*\to F_4$ given by
\[ s_4(y) = (\sqrt{u_1},\sqrt{u_2},y_1,-y_1y_2), \]
and these satisfy $p_4s_4=1$ and $s_4p_4=r_4$. Thus, $p_4$ restricts
to give a homeomorphism $F_4\to F_4^*$ with inverse $s_4$.
\end{proposition}
Maple notation for $p_4(x)$ and $s_4(t)$ is \mcode+y_proj(x)+ and
\mcode+y_lift(t)+.
\begin{proof}
It is clear that $F_4$ is a retractive fundamental domain for $H_4$, with
retraction $r_4$. Recall that the ring of functions on $EX(a)$ is
generated by $y_1$, $y_2$, $x_1$ and $x_2$, with $x_i^2=u_i$ for
$i=1,2$ and $x_3=y_1$ and $x_4=-y_1y_2$. It follows that $u_1$ and
$u_2$ are nonnegative as functions on $EX(a)$, or equivalently that
$p_4(EX(a))\sse F_4^*$. It also follows that the stated formula gives a
well-defined map $s_4\:F_4^*\to F_4$, and it is straightforward to
check that $p_4s_4=1$ and $s_4p_4=r_4$.
\begin{checks}
embedded/EX_check.mpl: check_E_F4()
\end{checks}
\end{proof}
The following picture shows $F_4$ together with the curves
$c_3,\dotsc,c_8$.
\[ \includegraphics[scale=0.35,clip=true,
trim=6cm 6cm 6cm 7cm]{images/F4.jpg}
\]
Formulae for the action of $p_4$ on some of the curves $c_i$ are as follows:
\begin{align*}
p_4(c_{\alg}(t)) &=
\left(\sqrt{\frac{1-2at}{(t-a)(t-a^{-1})}},t\right) \\
p_4(c_0(t)) &= (0,-\cos(2t)/(2a)) \\
p_4(c_1(t)) = p_4(c_2(t)) &= (\cos(t),0).
\end{align*}
Formulae for the remaining curves are not illuminating.
The following picture shows the set $F_4^*$ for three different values
of the parameter $a$, together with the images of the points $v_i$
under the map $p_4$.
\begin{center}
\begin{tikzpicture}[scale=1.5]
\begin{scope}
\draw[cyan] plot[smooth] coordinates{ (0.000,-1.000) (0.000,-0.309) (0.000,0.809) (0.000,0.809) (0.000,-0.309) (0.000,-1.000) (0.000,-0.309) (0.000,0.809) (0.000,0.809) (0.000,-0.309) (0.000,-1.000) };
\draw[green] plot[smooth] coordinates{ (1.000,0.000) (0.809,0.000) (0.309,0.000) (-0.309,0.000) (-0.809,0.000) (-1.000,0.000) (-0.809,0.000) (-0.309,0.000) (0.309,0.000) (0.809,0.000) (1.000,0.000) };
\draw[green] plot[smooth] coordinates{ (1.000,0.000) (0.809,0.000) (0.309,0.000) (-0.309,0.000) (-0.809,0.000) (-1.000,0.000) (-0.809,0.000) (-0.309,0.000) (0.309,0.000) (0.809,0.000) (1.000,0.000) };
\draw[magenta] plot[smooth] coordinates{ (0.632,1.225) (0.512,1.144) (0.195,1.019) (-0.195,1.019) (-0.512,1.144) (-0.632,1.225) (-0.512,1.144) (-0.195,1.019) (0.195,1.019) (0.512,1.144) (0.632,1.225) };
\draw[magenta] plot[smooth] coordinates{ (0.632,-1.225) (0.512,-1.144) (0.195,-1.019) (-0.195,-1.019) (-0.512,-1.144) (-0.632,-1.225) (-0.512,-1.144) (-0.195,-1.019) (0.195,-1.019) (0.512,-1.144) (0.632,-1.225) };
\draw[blue] plot[smooth] coordinates{ (1.000,0.000) (0.925,0.117) (0.798,0.423) (0.703,0.802) (0.649,1.108) (0.632,1.225) (0.649,1.108) (0.703,0.802) (0.798,0.423) (0.925,0.117) (1.000,0.000) };
\draw[blue] plot[smooth] coordinates{ (1.000,0.000) (0.925,-0.117) (0.798,-0.423) (0.703,-0.802) (0.649,-1.108) (0.632,-1.225) (0.649,-1.108) (0.703,-0.802) (0.798,-0.423) (0.925,-0.117) (1.000,-0.000) };
\draw[blue] plot[smooth] coordinates{ (-1.000,0.000) (-0.925,0.117) (-0.798,0.423) (-0.703,0.802) (-0.649,1.108) (-0.632,1.225) (-0.649,1.108) (-0.703,0.802) (-0.798,0.423) (-0.925,0.117) (-1.000,0.000) };
\draw[blue] plot[smooth] coordinates{ (-1.000,0.000) (-0.925,-0.117) (-0.798,-0.423) (-0.703,-0.802) (-0.649,-1.108) (-0.632,-1.225) (-0.649,-1.108) (-0.703,-0.802) (-0.798,-0.423) (-0.925,-0.117) (-1.000,-0.000) };
\fill (1.000,0.000) circle(0.015);
\fill (-1.000,0.000) circle(0.015);
\fill (0.000,-1.000) circle(0.015);
\fill (0.000,1.000) circle(0.015);
\fill (0.000,-1.000) circle(0.015);
\fill (0.000,1.000) circle(0.015);
\fill (0.000,0.000) circle(0.015);
\fill (0.000,0.000) circle(0.015);
\fill (0.000,0.000) circle(0.015);
\fill (0.000,0.000) circle(0.015);
\fill (0.632,-1.225) circle(0.015);
\fill (0.632,1.225) circle(0.015);
\fill (-0.632,-1.225) circle(0.015);
\fill (-0.632,1.225) circle(0.015);
\draw (0.000,0.000) node[anchor=north] {$v_{6},v_{7},v_{8},v_{9}$};
\draw (-1.000,0.000) node[anchor=east] {$v_{1}$};
\draw (0.000,-1.000) node[anchor=north] {$v_{2},v_{4}$};
\draw (1.000,0.000) node[anchor=west] {$v_{0}$};
\draw (-0.632,1.225) node[anchor=south] {$v_{13}$};
\draw (0.632,1.225) node[anchor=south] {$v_{11}$};
\draw (0.000,1.000) node[anchor=south] {$v_{3},v_{5}$};
\draw (-0.632,-1.225) node[anchor=north] {$v_{12}$};
\draw (0.632,-1.225) node[anchor=north] {$v_{10}$};
\draw (0,-1.7) node {$a=1/2$};
\end{scope}
\begin{scope}[xshift=3.5cm]
\draw[cyan] plot[smooth] coordinates{ (0.000,-0.707) (0.000,-0.219) (0.000,0.572) (0.000,0.572) (0.000,-0.219) (0.000,-0.707) (0.000,-0.219) (0.000,0.572) (0.000,0.572) (0.000,-0.219) (0.000,-0.707) };
\draw[green] plot[smooth] coordinates{ (1.000,0.000) (0.809,0.000) (0.309,0.000) (-0.309,0.000) (-0.809,0.000) (-1.000,0.000) (-0.809,0.000) (-0.309,0.000) (0.309,0.000) (0.809,0.000) (1.000,0.000) };
\draw[green] plot[smooth] coordinates{ (1.000,0.000) (0.809,0.000) (0.309,0.000) (-0.309,0.000) (-0.809,0.000) (-1.000,0.000) (-0.809,0.000) (-0.309,0.000) (0.309,0.000) (0.809,0.000) (1.000,0.000) };
\draw[magenta] plot[smooth] coordinates{ (0.816,0.707) (0.661,0.707) (0.252,0.707) (-0.252,0.707) (-0.661,0.707) (-0.816,0.707) (-0.661,0.707) (-0.252,0.707) (0.252,0.707) (0.661,0.707) (0.816,0.707) };
\draw[magenta] plot[smooth] coordinates{ (0.816,-0.707) (0.661,-0.707) (0.252,-0.707) (-0.252,-0.707) (-0.661,-0.707) (-0.816,-0.707) (-0.661,-0.707) (-0.252,-0.707) (0.252,-0.707) (0.661,-0.707) (0.816,-0.707) };
\draw[blue] plot[smooth] coordinates{ (1.000,0.000) (0.977,0.068) (0.923,0.244) (0.868,0.463) (0.830,0.640) (0.816,0.707) (0.830,0.640) (0.868,0.463) (0.923,0.244) (0.977,0.068) (1.000,0.000) };
\draw[blue] plot[smooth] coordinates{ (1.000,0.000) (0.977,-0.068) (0.923,-0.244) (0.868,-0.463) (0.830,-0.640) (0.816,-0.707) (0.830,-0.640) (0.868,-0.463) (0.923,-0.244) (0.977,-0.068) (1.000,-0.000) };
\draw[blue] plot[smooth] coordinates{ (-1.000,0.000) (-0.977,0.068) (-0.923,0.244) (-0.868,0.463) (-0.830,0.640) (-0.816,0.707) (-0.830,0.640) (-0.868,0.463) (-0.923,0.244) (-0.977,0.068) (-1.000,0.000) };
\draw[blue] plot[smooth] coordinates{ (-1.000,0.000) (-0.977,-0.068) (-0.923,-0.244) (-0.868,-0.463) (-0.830,-0.640) (-0.816,-0.707) (-0.830,-0.640) (-0.868,-0.463) (-0.923,-0.244) (-0.977,-0.068) (-1.000,-0.000) };
\fill (1.000,0.000) circle(0.015);
\fill (-1.000,0.000) circle(0.015);
\fill (0.000,-0.707) circle(0.015);
\fill (0.000,0.707) circle(0.015);
\fill (0.000,-0.707) circle(0.015);
\fill (0.000,0.707) circle(0.015);
\fill (0.000,0.000) circle(0.015);
\fill (0.000,0.000) circle(0.015);
\fill (0.000,0.000) circle(0.015);
\fill (0.000,0.000) circle(0.015);
\fill (0.816,-0.707) circle(0.015);
\fill (0.816,0.707) circle(0.015);
\fill (-0.816,-0.707) circle(0.015);
\fill (-0.816,0.707) circle(0.015);
\draw (0.000,0.000) node[anchor=north] {$v_{6},v_{7},v_{8},v_{9}$};
\draw (-0.816,-0.707) node[anchor=north] {$v_{12}$};
\draw (-1.000,0.000) node[anchor=east] {$v_{1}$};
\draw (0.816,-0.707) node[anchor=north] {$v_{10}$};
\draw (1.000,0.000) node[anchor=west] {$v_{0}$};
\draw (0.000,-0.707) node[anchor=north] {$v_{2},v_{4}$};
\draw (0.816,0.707) node[anchor=south] {$v_{11}$};
\draw (0.000,0.707) node[anchor=south] {$v_{3},v_{5}$};
\draw (-0.816,0.707) node[anchor=south] {$v_{13}$};
\draw (0,-1.7) node {$a=1/\rt$};
\end{scope}
\begin{scope}[xshift=7.0cm]
\draw[cyan] plot[smooth] coordinates{ (0.000,-0.625) (0.000,-0.193) (0.000,0.506) (0.000,0.506) (0.000,-0.193) (0.000,-0.625) (0.000,-0.193) (0.000,0.506) (0.000,0.506) (0.000,-0.193) (0.000,-0.625) };
\draw[green] plot[smooth] coordinates{ (1.000,0.000) (0.809,0.000) (0.309,0.000) (-0.309,0.000) (-0.809,0.000) (-1.000,0.000) (-0.809,0.000) (-0.309,0.000) (0.309,0.000) (0.809,0.000) (1.000,0.000) };
\draw[green] plot[smooth] coordinates{ (1.000,0.000) (0.809,0.000) (0.309,0.000) (-0.309,0.000) (-0.809,0.000) (-1.000,0.000) (-0.809,0.000) (-0.309,0.000) (0.309,0.000) (0.809,0.000) (1.000,0.000) };
\draw[magenta] plot[smooth] coordinates{ (0.883,0.530) (0.715,0.577) (0.273,0.620) (-0.273,0.620) (-0.715,0.577) (-0.883,0.530) (-0.715,0.577) (-0.273,0.620) (0.273,0.620) (0.715,0.577) (0.883,0.530) };
\draw[magenta] plot[smooth] coordinates{ (0.883,-0.530) (0.715,-0.577) (0.273,-0.620) (-0.273,-0.620) (-0.715,-0.577) (-0.883,-0.530) (-0.715,-0.577) (-0.273,-0.620) (0.273,-0.620) (0.715,-0.577) (0.883,-0.530) };
\draw[blue] plot[smooth] coordinates{ (1.000,0.000) (0.988,0.051) (0.958,0.183) (0.921,0.347) (0.894,0.480) (0.883,0.530) (0.894,0.480) (0.921,0.347) (0.958,0.183) (0.988,0.051) (1.000,0.000) };
\draw[blue] plot[smooth] coordinates{ (1.000,0.000) (0.988,-0.051) (0.958,-0.183) (0.921,-0.347) (0.894,-0.480) (0.883,-0.530) (0.894,-0.480) (0.921,-0.347) (0.958,-0.183) (0.988,-0.051) (1.000,-0.000) };
\draw[blue] plot[smooth] coordinates{ (-1.000,0.000) (-0.988,0.051) (-0.958,0.183) (-0.921,0.347) (-0.894,0.480) (-0.883,0.530) (-0.894,0.480) (-0.921,0.347) (-0.958,0.183) (-0.988,0.051) (-1.000,0.000) };
\draw[blue] plot[smooth] coordinates{ (-1.000,0.000) (-0.988,-0.051) (-0.958,-0.183) (-0.921,-0.347) (-0.894,-0.480) (-0.883,-0.530) (-0.894,-0.480) (-0.921,-0.347) (-0.958,-0.183) (-0.988,-0.051) (-1.000,-0.000) };
\fill (1.000,0.000) circle(0.015);
\fill (-1.000,0.000) circle(0.015);
\fill (0.000,-0.625) circle(0.015);
\fill (0.000,0.625) circle(0.015);
\fill (0.000,-0.625) circle(0.015);
\fill (0.000,0.625) circle(0.015);
\fill (0.000,0.000) circle(0.015);
\fill (0.000,0.000) circle(0.015);
\fill (0.000,0.000) circle(0.015);
\fill (0.000,0.000) circle(0.015);
\fill (0.883,-0.530) circle(0.015);
\fill (0.883,0.530) circle(0.015);
\fill (-0.883,-0.530) circle(0.015);
\fill (-0.883,0.530) circle(0.015);
\draw (0.000,0.625) node[anchor=south] {$v_{3},v_{5}$};
\draw (0.883,0.530) node[anchor=south] {$v_{11}$};
\draw (0.000,0.000) node[anchor=north] {$v_{6},v_{7},v_{8},v_{9}$};
\draw (0.000,-0.625) node[anchor=north] {$v_{2},v_{4}$};
\draw (-0.883,-0.530) node[anchor=north] {$v_{12}$};
\draw (0.883,-0.530) node[anchor=north] {$v_{10}$};
\draw (1.000,0.000) node[anchor=west] {$v_{0}$};
\draw (-1.000,0.000) node[anchor=east] {$v_{1}$};
\draw (-0.883,0.530) node[anchor=south] {$v_{13}$};
\draw (0,-1.7) node {$a=4/5$};
\end{scope}
\end{tikzpicture}
\end{center}
Here is a more detailed version for the case $a=1/\rt$:
\begin{center}
\begin{tikzpicture}[scale=4]
\draw[cyan] plot[smooth] coordinates{ (0.000,-0.707) (0.000,-0.572) (0.000,-0.219) (0.000,0.219) (0.000,0.572) (0.000,0.707) (0.000,0.572) (0.000,0.219) (0.000,-0.219) (0.000,-0.572) (0.000,-0.707) (0.000,-0.572) (0.000,-0.219) (0.000,0.219) (0.000,0.572) (0.000,0.707) (0.000,0.572) (0.000,0.219) (0.000,-0.219) (0.000,-0.572) (0.000,-0.707) };
\draw[green] plot[smooth] coordinates{ (1.000,0.000) (0.951,0.000) (0.809,0.000) (0.588,0.000) (0.309,0.000) (-0.000,0.000) (-0.309,0.000) (-0.588,0.000) (-0.809,0.000) (-0.951,0.000) (-1.000,0.000) (-0.951,0.000) (-0.809,0.000) (-0.588,0.000) (-0.309,0.000) (0.000,0.000) (0.309,0.000) (0.588,0.000) (0.809,0.000) (0.951,0.000) (1.000,0.000) };
\draw[green] plot[smooth] coordinates{ (1.000,0.000) (0.951,0.000) (0.809,0.000) (0.588,0.000) (0.309,0.000) (-0.000,0.000) (-0.309,0.000) (-0.588,0.000) (-0.809,0.000) (-0.951,0.000) (-1.000,0.000) (-0.951,0.000) (-0.809,0.000) (-0.588,0.000) (-0.309,0.000) (0.000,0.000) (0.309,0.000) (0.588,0.000) (0.809,0.000) (0.951,0.000) (1.000,0.000) };
\draw[magenta] plot[smooth] coordinates{ (0.816,0.707) (0.777,0.707) (0.661,0.707) (0.480,0.707) (0.252,0.707) (-0.000,0.707) (-0.252,0.707) (-0.480,0.707) (-0.661,0.707) (-0.777,0.707) (-0.816,0.707) (-0.777,0.707) (-0.661,0.707) (-0.480,0.707) (-0.252,0.707) (0.000,0.707) (0.252,0.707) (0.480,0.707) (0.661,0.707) (0.777,0.707) (0.816,0.707) };
\draw[magenta] plot[smooth] coordinates{ (0.816,-0.707) (0.777,-0.707) (0.661,-0.707) (0.480,-0.707) (0.252,-0.707) (-0.000,-0.707) (-0.252,-0.707) (-0.480,-0.707) (-0.661,-0.707) (-0.777,-0.707) (-0.816,-0.707) (-0.777,-0.707) (-0.661,-0.707) (-0.480,-0.707) (-0.252,-0.707) (0.000,-0.707) (0.252,-0.707) (0.480,-0.707) (0.661,-0.707) (0.777,-0.707) (0.816,-0.707) };
\draw[blue] plot[smooth] coordinates{ (1.000,0.000) (0.994,0.017) (0.977,0.068) (0.952,0.146) (0.923,0.244) (0.894,0.354) (0.868,0.463) (0.846,0.561) (0.830,0.640) (0.820,0.690) (0.816,0.707) (0.820,0.690) (0.830,0.640) (0.846,0.561) (0.868,0.463) (0.894,0.354) (0.923,0.244) (0.952,0.146) (0.977,0.068) (0.994,0.017) (1.000,0.000) };
\draw[blue] plot[smooth] coordinates{ (1.000,0.000) (0.994,-0.017) (0.977,-0.068) (0.952,-0.146) (0.923,-0.244) (0.894,-0.354) (0.868,-0.463) (0.846,-0.561) (0.830,-0.640) (0.820,-0.690) (0.816,-0.707) (0.820,-0.690) (0.830,-0.640) (0.846,-0.561) (0.868,-0.463) (0.894,-0.354) (0.923,-0.244) (0.952,-0.146) (0.977,-0.068) (0.994,-0.017) (1.000,-0.000) };
\draw[blue] plot[smooth] coordinates{ (-1.000,0.000) (-0.994,0.017) (-0.977,0.068) (-0.952,0.146) (-0.923,0.244) (-0.894,0.354) (-0.868,0.463) (-0.846,0.561) (-0.830,0.640) (-0.820,0.690) (-0.816,0.707) (-0.820,0.690) (-0.830,0.640) (-0.846,0.561) (-0.868,0.463) (-0.894,0.354) (-0.923,0.244) (-0.952,0.146) (-0.977,0.068) (-0.994,0.017) (-1.000,0.000) };
\draw[blue] plot[smooth] coordinates{ (-1.000,0.000) (-0.994,-0.017) (-0.977,-0.068) (-0.952,-0.146) (-0.923,-0.244) (-0.894,-0.354) (-0.868,-0.463) (-0.846,-0.561) (-0.830,-0.640) (-0.820,-0.690) (-0.816,-0.707) (-0.820,-0.690) (-0.830,-0.640) (-0.846,-0.561) (-0.868,-0.463) (-0.894,-0.354) (-0.923,-0.244) (-0.952,-0.146) (-0.977,-0.068) (-0.994,-0.017) (-1.000,-0.000) };
\fill (1.000,0.000) circle(0.010);
\fill (-1.000,0.000) circle(0.010);
\fill (0.000,-0.707) circle(0.010);
\fill (0.000,0.707) circle(0.010);
\fill (0.000,-0.707) circle(0.010);
\fill (0.000,0.707) circle(0.010);
\fill (0.000,0.000) circle(0.010);
\fill (0.000,0.000) circle(0.010);
\fill (0.000,0.000) circle(0.010);
\fill (0.000,0.000) circle(0.010);
\fill (0.816,-0.707) circle(0.010);
\fill (0.816,0.707) circle(0.010);
\fill (-0.816,-0.707) circle(0.010);
\fill (-0.816,0.707) circle(0.010);
\draw (0.000,-0.707) node[anchor=north] {$v_{2},v_{4}$};
\draw (-0.816,0.707) node[anchor=south] {$v_{13}$};
\draw (0.000,0.000) node[anchor=north] {$v_{6},v_{7},v_{8},v_{9}$};
\draw (-0.816,-0.707) node[anchor=north] {$v_{12}$};
\draw (0.000,0.707) node[anchor=south] {$v_{3},v_{5}$};
\draw (1.000,0.000) node[anchor=west] {$v_{0}$};
\draw (-1.000,0.000) node[anchor=east] {$v_{1}$};
\draw (0.816,0.707) node[anchor=south] {$v_{11}$};
\draw (0.816,-0.707) node[anchor=north] {$v_{10}$};
\draw (0.000,-0.354) node[anchor=east] {$c_{0}$};
\draw (-0.666,0.000) node[anchor=north] {$c_{1}$};
\draw (-0.505,0.000) node[anchor=north] {$c_{2}$};
\draw (-0.412,0.707) node[anchor=south] {$c_{3}$};
\draw (-0.412,-0.707) node[anchor=north] {$c_{4}$};
\draw (0.901,0.329) node[anchor=west] {$c_{5}$};
\draw (0.901,-0.329) node[anchor=west] {$c_{6}$};
\draw (-0.901,0.329) node[anchor=east] {$c_{7}$};
\draw (-0.901,-0.329) node[anchor=east] {$c_{8}$};
\end{tikzpicture}
\end{center}
The above picture shows the images of all the points $v_i$, but it is
also useful to restrict attention to those that lie in $F_4$:
\begin{center}
\begin{tikzpicture}[scale=4]
\draw[cyan] plot[smooth] coordinates{ (0.000,-0.707) (0.000,-0.672) (0.000,-0.572) (0.000,-0.416) (0.000,-0.219) (0.000,0.000) (0.000,0.219) (0.000,0.416) (0.000,0.572) (0.000,0.672) (0.000,0.707) };
\draw[green] plot[smooth] coordinates{ (1.000,0.000) (0.951,0.000) (0.809,0.000) (0.588,0.000) (0.309,0.000) (-0.000,0.000) (-0.309,0.000) (-0.588,0.000) (-0.809,0.000) (-0.951,0.000) (-1.000,0.000) };
\draw[magenta] plot[smooth] coordinates{ (0.816,0.707) (0.777,0.707) (0.661,0.707) (0.480,0.707) (0.252,0.707) (-0.000,0.707) (-0.252,0.707) (-0.480,0.707) (-0.661,0.707) (-0.777,0.707) (-0.816,0.707) };
\draw[magenta] plot[smooth] coordinates{ (-0.816,-0.707) (-0.777,-0.707) (-0.661,-0.707) (-0.480,-0.707) (-0.252,-0.707) (0.000,-0.707) (0.252,-0.707) (0.480,-0.707) (0.661,-0.707) (0.777,-0.707) (0.816,-0.707) };
\draw[blue] plot[smooth] coordinates{ (1.000,0.000) (0.994,0.017) (0.977,0.068) (0.952,0.146) (0.923,0.244) (0.894,0.354) (0.868,0.463) (0.846,0.561) (0.830,0.640) (0.820,0.690) (0.816,0.707) };
\draw[blue] plot[smooth] coordinates{ (1.000,0.000) (0.994,-0.017) (0.977,-0.068) (0.952,-0.146) (0.923,-0.244) (0.894,-0.354) (0.868,-0.463) (0.846,-0.561) (0.830,-0.640) (0.820,-0.690) (0.816,-0.707) };
\draw[blue] plot[smooth] coordinates{ (-1.000,0.000) (-0.994,0.017) (-0.977,0.068) (-0.952,0.146) (-0.923,0.244) (-0.894,0.354) (-0.868,0.463) (-0.846,0.561) (-0.830,0.640) (-0.820,0.690) (-0.816,0.707) };
\draw[blue] plot[smooth] coordinates{ (-1.000,0.000) (-0.994,-0.017) (-0.977,-0.068) (-0.952,-0.146) (-0.923,-0.244) (-0.894,-0.354) (-0.868,-0.463) (-0.846,-0.561) (-0.830,-0.640) (-0.820,-0.690) (-0.816,-0.707) };
\draw[cyan,arrows={-angle 90}] (0.0000,-0.5000) -- (0.0000,-0.4899);
\draw[cyan,arrows={-angle 90}] (0.0000,0.5000) -- (0.0000,0.5099);
\draw[green,arrows={-angle 90}] (0.7071,0.0000) -- (0.7000,0.0000);
\draw[green,arrows={-angle 90}] (-0.7071,0.0000) -- (-0.7141,0.0000);
\draw[magenta,arrows={-angle 90}] (0.5774,0.7071) -- (0.5715,0.7071);
\draw[magenta,arrows={-angle 90}] (-0.5774,0.7071) -- (-0.5831,0.7071);
\draw[magenta,arrows={-angle 90}] (-0.5774,-0.7071) -- (-0.5715,-0.7071);
\draw[magenta,arrows={-angle 90}] (0.5774,-0.7071) -- (0.5831,-0.7071);
\draw[blue,arrows={-angle 90}] (0.8944,0.3536) -- (0.8935,0.3571);
\draw[blue,arrows={-angle 90}] (0.8944,-0.3536) -- (0.8935,-0.3571);
\draw[blue,arrows={-angle 90}] (-0.8944,0.3536) -- (-0.8935,0.3571);
\draw[blue,arrows={-angle 90}] (-0.8944,-0.3536) -- (-0.8935,-0.3571);
\fill (1.000,0.000) circle(0.010);
\fill (-1.000,0.000) circle(0.010);
\fill (0.000,-0.707) circle(0.010);
\fill (0.000,0.707) circle(0.010);
\fill (0.000,0.000) circle(0.010);
\fill (0.816,-0.707) circle(0.010);
\fill (0.816,0.707) circle(0.010);
\fill (-0.816,-0.707) circle(0.010);
\fill (-0.816,0.707) circle(0.010);
\draw (0.000,-0.707) node[anchor=north] {$v_{2}$};
\draw (-0.816,0.707) node[anchor=south] {$v_{13}$};
\draw (0.000,0.000) node[anchor=north] {$v_{6}$};
\draw (-0.816,-0.707) node[anchor=north] {$v_{12}$};
\draw (0.000,0.707) node[anchor=south] {$v_{3}$};
\draw (1.000,0.000) node[anchor=west] {$v_{0}$};
\draw (-1.000,0.000) node[anchor=east] {$v_{1}$};
\draw (0.816,0.707) node[anchor=south] {$v_{11}$};
\draw (0.816,-0.707) node[anchor=north] {$v_{10}$};
\draw (0.000,-0.354) node[anchor=east] {$c_{0}(0\dotsb\ppi)$};
\draw (-0.589,0.000) node[anchor=north] {$c_{1}(0\dotsb\pi)$};
\draw (-0.412,0.707) node[anchor=south] {$c_{3}(0\dotsb\pi)$};
\draw (-0.412,-0.707) node[anchor=north] {$c_{4}(\pi\dotsb2\pi)$};
\draw (0.901,0.329) node[anchor=west] {$c_{5}(0\dotsb\pi)$};
\draw (0.901,-0.329) node[anchor=west] {$c_{6}(0\dotsb\pi)$};
\draw (-0.901,0.329) node[anchor=east] {$c_{7}(0\dotsb\pi)$};
\draw (-0.901,-0.329) node[anchor=east] {$c_{8}(0\dotsb\pi)$};
\end{tikzpicture}
\end{center}
The annotation $c_1(0\dotsb\pi)$ indicates that $c_1$ sends the
interval $[0,\pi]$ to $F_4$. For values $t\in(\pi,2\pi)$, the
point $c_1(t)$ lies outside $F_4$, but it has the same $H_4$-orbit as
some point $c_1(t')$ with $t'\in[0,\pi]$, so $p_4c_1(t)$ will still
lie on the middle horizontal line in the above diagram. The other
annotations should be interpreted in the same way.
We now define a retractive fundamental domain for the full group $G$.
\begin{definition}\label{defn-F-sixteen}
We put
\begin{align*}
F_{16} &= \{x\in EX(a) \st x_1,x_2,y_1,y_2\geq 0\} \\
r_{16}(x) &=
\begin{cases}
(|x_1|,|x_2|,|x_3|,-|x_4|) & \text{ if } y_2\geq 0 \\
(|x_2|,|x_1|,|x_3|,-|x_4|) & \text{ if } y_2\leq 0.
\end{cases} \\
F_{16}^* &= \{(z_1,z_2)\in\R^2\st z_1,z_2,u_3,u_4\geq 0\}
\end{align*}
Here $y_1,y_2,z_1,z_2,u_3$ and $u_4$ are as in
Definition~\ref{defn-yzu}. Note that if $y_2=0$ then
$x_4=-y_1y_2=0$ so the relation $y_2=0$ becomes $x_2^2=x_1^2$, so
$|x_1|=|x_2|$; this shows that $r_{16}$ is well-defined.
\end{definition}
\begin{proposition}\label{prop-F-sixteen}
$F_{16}$ is a retractive fundamental domain for $G$, with retraction
$r_{16}$. Moreover, there is a map $p_{16}\:EX(a)\to F_{16}^*$ given by
\[ p_{16}(x) = (x_3^2,(x_2^2 - x_1^2 - (a^{-1}+a)x_3x_4)^2/(4a^2))
= (z_1,z_2),
\]
and a map $s_{16}\:F_{16}^*\to F_{16}$ given by
\[ s_{16}(z_1,z_2) = s_4(\sqrt{z_1},\sqrt{z_2}) \]
and these satisfy $p_{16}s_{16}=1$ and $s_{16}p_{16}=r_{16}$. Thus,
$p_{16}$ restricts to give a homeomorphism $F_{16}\to F_{16}^*$ with
inverse $s_{16}$.
\end{proposition}
Maple notation for $p_{16}(x)$ and $s_{16}(t)$ is \mcode+z_proj(x)+ and
\mcode+z_lift(t)+.
\begin{proof}
First, a straightforward check of cases shows that
$r_{16}(EX(a))\sse F_{16}$ and that $r_{16}$ is the identity on $F_{16}$,
so $r_{16}$ is a retraction. We also claim that
$r_{16}(\gm(x))=r_{16}(x)$ for all $x\in EX(a)$ and $\gm\in G$. If
$\gm\in\ip{\lm^2,\mu,\nu}$ then $|\gm(x)_i|=|x_i|$ for all $i$ and
$y_2(\gm(x))=y_2(x)$ so everything is easy. This just leaves the
case $\gm=\lm$. Here $\gm^*$ exchanges $|x_1|$ and $|x_2|$, and
changes the sign of $y_2$, so we again have
$r_{16}(\gm(x))=r_{16}(x)$. It follows that $r_{16}$ induces a
surjective map $EX(a)/G\to F_{16}$.
We now show that any point $x\in EX(a)$ can be moved into $F_{16}$
by the action of $G$. First, after applying an element of $H_4$ we may
assume that $x\in F_4$, so $x_1,x_2\geq 0$. Now note that
\begin{align*}
\lm\mu(x) &= (x_2,x_1, -x_3,\pp x_4) & y_2(\lm\mu(x)) &= -y_2(x) \\
\lm\nu(x) &= (x_2,x_1,\pp x_3, -x_4) & y_2(\lm\nu(x)) &= -y_2(x) \\
\mu\nu(x) &= (x_1,x_2, -x_3, -x_4) & y_2(\mu\nu(x)) &= \pp y_2(x).
\end{align*}
We can thus apply one of the maps $1,\lm\mu,\lm\nu,\mu\nu$ to move
into $F_{16}$, as required. Note also that if $\gm(x)=a\in F_{16}$
then we can apply $r_{16}$ to deduce that $a=r_{16}(x)$, so
$x=\gm^{-1}(r_{16}(x))$. It follows that $r_{16}(x)=r_{16}(x')$ iff
$Gx=Gx'$, so the induced map $EX(a)/G\to F_{16}$ is a bijective
retraction and therefore a homeomorphism.
Now put
\[ E = \{x\in F_{16}\st x_1=0 \text{ or } x_2 = 0 \text{ or }
y_1 = 0 \text{ or } y_2 = 0\}.
\]
The set $F_{16}\sm E$ is defined by strict inequalities and so is
contained in the interior of $F_{16}$. To understand the structure
of $E$, it is helpful to recall that $c_k(\R)=C_k=C_k^*$ for all $k$,
where $C_k^*$ was defined in Definition~\ref{defn-C-star}. Using
this, we see that
\[ E =
c_0([\tfrac{\pi}{4},\ppi]) \cup
c_1([0,\ppi]) \cup
c_3([0,\ppi]) \cup
c_5([0,\pi]).
\]
Using this, we can show that arbitrarily close to every point in $E$
there are points outside $F_{16}$. For example, if
$x=c_0(t)=(\cos(t),\sin(t),0,0)$ with $\pi/4\leq t\leq\pi/2$ then the
vector $(0,0,-1,-\cos(2t))$ is tangent to $EX(a)$ at $x$, and after
moving a small distance in this direction we reach the region where
$y_1=x_3<0$. Similar arguments work for the other cases, so we see
that $E$ is precisely the boundary of $F_{16}$.
Now consider a point $x\in F_{16}$ and an element $\gm\in G\sm\{1\}$
such that $\gm(x)$ also lies in $F_{16}$. By applying $r_{16}$, we
see that $\gm(x)=x$. In Section~\ref{sec-E-curves} we
discussed the fixed sets $EX(a)^\gm$ for all orientation-reversing
elements of $G$, and in particular we saw that $EX(a)^\gm$ is always
contained in $\bigcup_{i=0}^8C_i$. Thus, if $\gm$ reverses
orientation then $x\in E$. On the other hand, if $\gm$ preserves
orientation then Proposition~\ref{prop-no-more-isotropy} tells us
that $x=v_i$ for some $i$ with $0\leq i\leq 13$. The only points of
this type lying in $F_{16}$ are $v_0,v_3,v_6$ and $v_{11}$, but we
have
\begin{align*}
v_0 &= c_1(0) = c_5(0) \\
v_3 &= c_0(\pi/2) = c_3(\pi/2) \\
v_6 &= c_0(\pi/4) = c_1(\pi/2) \\
v_{11} &= c_3(\pi/2) = c_5(\pi)
\end{align*}
so these points are also in $E$. This proves that
$\text{int}(F_{16})\cap\gm(F_{16})=\emptyset$. We conclude that
$F_{16}$ is a retractive fundamental domain, as claimed.
We now need to show that $p_{16}(EX(a))\sse F_{16}^*$, or equivalently
that $z_1,z_2,u_3,u_4\geq 0$ as functions on $EX(a)$. This is clear from
the identities $z_i=y_i^2$ and $u_3=4u_1u_2=4x_1^2x_2^2$ and
$u_4=u_1+u_2=x_1^2+x_2^2$.
Now suppose we start with a point $z\in F_{16}^*$, and define
$y_i=\sqrt{z_i}$ for $i=1,2$, and then
\begin{align*}
u_1 &= \half(1-y_2) - \half(y_2-a)(y_2-a^{-1})y_1^2 \\
u_2 &= \half(1+y_2) - \half(y_2+a)(y_2+a^{-1})y_1^2.
\end{align*}
We find that $u_1+u_2=1-z_1-z_1z_2$ and
$4u_1u_2=(1-z_1-z_1z_2)^2 - z_2((a+a^{-1})z_1-2a)^2$; these are the
quantities $u_3$ and $u_4$ that are assumed to be nonnegative by the
definition of $F_{16}^*$. It follows from this by a check of cases
that both $u_1$ and $u_2$ are nonnegative, so $y\in F_4^*$ and
the point $x=s_4(y)=(\sqrt{u_1},\sqrt{u_2},y_1,-y_1y_2)$ is a
well-defined element of $F_4$. It is clear by construction that
in fact $x\in F_{16}$, so we have a well-defined map
$s_{16}\:F_{16}^*\to F_{16}$ with the claimed properties.
\begin{checks}
embedded/EX_check.mpl: check_E_F16()
\end{checks}
\end{proof}
The following picture shows the set $F_{16}^*$ (where $a=1/\rt$)
together with the images under $p_{16}$ of the curves $c_i(t)$ (for
$0\leq i\leq 8$) and the points $v_j$ (for $0\leq j\leq 13$).
\begin{center}
\begin{tikzpicture}[scale=8]
\draw[cyan] plot[smooth] coordinates{ (0.000,0.500) (0.000,0.327) (0.000,0.048) (0.000,0.048) (0.000,0.327) (0.000,0.500) (0.000,0.327) (0.000,0.048) (0.000,0.048) (0.000,0.327) (0.000,0.500) (0.000,0.327) (0.000,0.048) (0.000,0.048) (0.000,0.327) (0.000,0.500) (0.000,0.327) (0.000,0.048) (0.000,0.048) (0.000,0.327) (0.000,0.500) };
\draw[green] plot[smooth] coordinates{ (1.000,0.000) (0.905,0.000) (0.655,0.000) (0.345,0.000) (0.095,0.000) (0.000,0.000) (0.095,0.000) (0.345,0.000) (0.655,0.000) (0.905,0.000) (1.000,0.000) (0.905,0.000) (0.655,0.000) (0.345,0.000) (0.095,0.000) (0.000,0.000) (0.095,0.000) (0.345,0.000) (0.655,0.000) (0.905,0.000) (1.000,0.000) };
\draw[green] plot[smooth] coordinates{ (1.000,0.000) (0.905,0.000) (0.655,0.000) (0.345,0.000) (0.095,0.000) (0.000,0.000) (0.095,0.000) (0.345,0.000) (0.655,0.000) (0.905,0.000) (1.000,0.000) (0.905,0.000) (0.655,0.000) (0.345,0.000) (0.095,0.000) (0.000,0.000) (0.095,0.000) (0.345,0.000) (0.655,0.000) (0.905,0.000) (1.000,0.000) };
\draw[magenta] plot[smooth] coordinates{ (0.667,0.500) (0.603,0.500) (0.436,0.500) (0.230,0.500) (0.064,0.500) (0.000,0.500) (0.064,0.500) (0.230,0.500) (0.436,0.500) (0.603,0.500) (0.667,0.500) (0.603,0.500) (0.436,0.500) (0.230,0.500) (0.064,0.500) (0.000,0.500) (0.064,0.500) (0.230,0.500) (0.436,0.500) (0.603,0.500) (0.667,0.500) };
\draw[magenta] plot[smooth] coordinates{ (0.667,0.500) (0.603,0.500) (0.436,0.500) (0.230,0.500) (0.064,0.500) (0.000,0.500) (0.064,0.500) (0.230,0.500) (0.436,0.500) (0.603,0.500) (0.667,0.500) (0.603,0.500) (0.436,0.500) (0.230,0.500) (0.064,0.500) (0.000,0.500) (0.064,0.500) (0.230,0.500) (0.436,0.500) (0.603,0.500) (0.667,0.500) };
\draw[blue] plot[smooth] coordinates{ (1.000,0.000) (0.988,0.000) (0.954,0.005) (0.907,0.021) (0.853,0.060) (0.800,0.125) (0.753,0.214) (0.716,0.315) (0.689,0.409) (0.672,0.476) (0.667,0.500) (0.672,0.476) (0.689,0.409) (0.716,0.315) (0.753,0.214) (0.800,0.125) (0.853,0.060) (0.907,0.021) (0.954,0.005) (0.988,0.000) (1.000,0.000) };
\draw[blue] plot[smooth] coordinates{ (1.000,0.000) (0.988,0.000) (0.954,0.005) (0.907,0.021) (0.853,0.060) (0.800,0.125) (0.753,0.214) (0.716,0.315) (0.689,0.409) (0.672,0.476) (0.667,0.500) (0.672,0.476) (0.689,0.409) (0.716,0.315) (0.753,0.214) (0.800,0.125) (0.853,0.060) (0.907,0.021) (0.954,0.005) (0.988,0.000) (1.000,0.000) };
\draw[blue] plot[smooth] coordinates{ (1.000,0.000) (0.988,0.000) (0.954,0.005) (0.907,0.021) (0.853,0.060) (0.800,0.125) (0.753,0.214) (0.716,0.315) (0.689,0.409) (0.672,0.476) (0.667,0.500) (0.672,0.476) (0.689,0.409) (0.716,0.315) (0.753,0.214) (0.800,0.125) (0.853,0.060) (0.907,0.021) (0.954,0.005) (0.988,0.000) (1.000,0.000) };
\draw[blue] plot[smooth] coordinates{ (1.000,0.000) (0.988,0.000) (0.954,0.005) (0.907,0.021) (0.853,0.060) (0.800,0.125) (0.753,0.214) (0.716,0.315) (0.689,0.409) (0.672,0.476) (0.667,0.500) (0.672,0.476) (0.689,0.409) (0.716,0.315) (0.753,0.214) (0.800,0.125) (0.853,0.060) (0.907,0.021) (0.954,0.005) (0.988,0.000) (1.000,0.000) };
\fill (1.000,0.000) circle(0.005);
\fill (1.000,0.000) circle(0.005);
\fill (0.000,0.500) circle(0.005);
\fill (0.000,0.500) circle(0.005);
\fill (0.000,0.500) circle(0.005);
\fill (0.000,0.500) circle(0.005);
\fill (0.000,0.000) circle(0.005);
\fill (0.000,0.000) circle(0.005);
\fill (0.000,0.000) circle(0.005);
\fill (0.000,0.000) circle(0.005);
\fill (0.667,0.500) circle(0.005);
\fill (0.667,0.500) circle(0.005);
\fill (0.667,0.500) circle(0.005);
\fill (0.667,0.500) circle(0.005);
\draw (0.000,0.000) node[anchor=north] {$v_{6},v_{7},v_{8},v_{9}$};
\draw (0.000,0.500) node[anchor=south] {$v_{2},v_{3},v_{4},v_{5}$};
\draw (1.000,0.000) node[anchor=north] {$v_{0},v_{1}$};
\draw (0.667,0.500) node[anchor=south] {$v_{10},v_{11},v_{12},v_{13}$};
\draw (0.000,0.272) node[anchor=east] {$c_{0}$};
\draw (0.444,0.000) node[anchor=north] {$c_{1}$};
\draw (0.544,0.000) node[anchor=north] {$c_{2}$};
\draw (0.296,0.500) node[anchor=south east] {$c_{3}$};
\draw (0.362,0.500) node[anchor=south east] {$c_{4}$};
\draw (0.765,0.188) node[anchor=west] {$c_{5}$};
\draw (0.745,0.235) node[anchor=west] {$c_{6}$};
\draw (0.727,0.283) node[anchor=west] {$c_{7}$};
\draw (0.711,0.331) node[anchor=west] {$c_{8}$};
\end{tikzpicture}
\end{center}
If we show only the curves and vertices lying in $F_{16}$, we obtain the
following picture:
\begin{center}
\begin{tikzpicture}[scale=8]
\draw[cyan] plot[smooth] coordinates{ (0.000,0.000) (0.000,0.012) (0.000,0.048) (0.000,0.103) (0.000,0.173) (0.000,0.250) (0.000,0.327) (0.000,0.397) (0.000,0.452) (0.000,0.488) (0.000,0.500) };
\draw[green] plot[smooth] coordinates{ (1.000,0.000) (0.976,0.000) (0.905,0.000) (0.794,0.000) (0.655,0.000) (0.500,0.000) (0.345,0.000) (0.206,0.000) (0.095,0.000) (0.024,0.000) (0.000,0.000) };
\draw[magenta] plot[smooth] coordinates{ (0.667,0.500) (0.650,0.500) (0.603,0.500) (0.529,0.500) (0.436,0.500) (0.333,0.500) (0.230,0.500) (0.137,0.500) (0.064,0.500) (0.016,0.500) (0.000,0.500) };
\draw[blue] plot[smooth] coordinates{ (1.000,0.000) (0.988,0.000) (0.954,0.005) (0.907,0.021) (0.853,0.060) (0.800,0.125) (0.753,0.214) (0.716,0.315) (0.689,0.409) (0.672,0.476) (0.667,0.500) };
\draw[cyan,arrows={-angle 90}] (0.0000,0.2500) -- (0.0000,0.2600);
\draw[green,arrows={-angle 90}] (0.5000,0.0000) -- (0.4900,0.0000);
\draw[magenta,arrows={-angle 90}] (0.3333,0.5000) -- (0.3267,0.5000);
\draw[blue,arrows={-angle 90}] (0.7385,0.2507) -- (0.7373,0.2539);
\fill (1.000,0.000) circle(0.005);
\fill (0.000,0.500) circle(0.005);
\fill (0.000,0.000) circle(0.005);
\fill (0.667,0.500) circle(0.005);
\draw (0.000,0.000) node[anchor=north] {$v_{6}$};
\draw (0.000,0.500) node[anchor=south] {$v_{3}$};
\draw (1.000,0.000) node[anchor=north west] {$v_{0}$};
\draw (0.667,0.500) node[anchor=south] {$v_{11}$};
\draw (0.000,0.250) node[anchor=east] {$c_{0}(\qpi\dotsb\ppi)$};
\draw (0.500,0.000) node[anchor=north] {$c_{1}(0\dotsb\ppi)$};
\draw (0.333,0.500) node[anchor=south] {$c_{3}(0\dotsb\ppi)$};
\draw (0.739,0.251) node[anchor=west] {$c_{5}(0\dotsb\pi)$};
\end{tikzpicture}
\end{center}
It is clear from these pictures that $F_{16}^*$ is homeomorphic to the
unit square. It is convenient to have an explicit homeomorphism,
which is provided by the following result. (Later we will give other
homeomorphisms that are specific to the case $a=1/\rt$, and have
better properties.)
\begin{proposition}\label{prop-square}
Put
\begin{align*}
w_1 &= \frac{z_1(1+(a+a^{-1})\sqrt{z_2}+z_2)}{1+2a\sqrt{z_2}} &
w_2 &=
\sqrt{z_2}\frac{2a(1-z_1)+z_1\sqrt{z_2}}{1-z_1+(a^{-1}-a)z_1\sqrt{z_2}},
\end{align*}
with the convention that $w_2=0$ at points where
$1-z_1+(a^{-1}-a)z_1\sqrt{z_2}=0$. Then the map
$q\:(z_1,z_2)\mapsto(w_1,w_2)$ gives a homeomorphism $F_{16}^*\to[0,1]^2$.
\end{proposition}
\begin{proof}
In view of Proposition~\ref{prop-F-sixteen} we can identify
$F_{16}^*$ with $F_{16}\sse EX(a)$, so we have nonnegative functions
$y_1,y_2,u_1,u_2$ as discussed previously and $\sqrt{z_2}=y_2$.
It is clear that $w_1$ is continuous and nonnegative. Next,
note that the functions $1-z_1=u_4+z_1z_2$ and $z_1\sqrt{z_2}$ are
nonnegative on $F_{16}^*$, and the only place where they both vanish
is $(1,0)$. Away from that point we deduce that $w_2$ is continuous
with $0\leq w_2\leq\sqrt{z_2}$, and these inequalities imply that
$w_2$ is continuous at the exceptional point as well.
One can also check that $1-w_1=2u_2/(1+2ay_2)\geq 0$ and
$1-w_2=2u_1/(1-y_1^2+(a^{-1}-a)y_1^2y_2)\geq 0$ so
$w_1,w_2\leq 1$. Thus, we have a well-defined and continuous map
$q\:F_{16}^*\to [0,1]^2$.
Now suppose we start with a point $w\in[0,1]^2$. We will assume that
$w_1<1$; the case $w_1=1$ requires only minor modifications and is
left to the reader. Consider the functions
\begin{align*}
p_0(s) &= s((1+w_1)s^2+(a^{-1}+a+((2a)^{-1}-2a)w_1)s+(1-w_1)) \\
p_1(s) &= ((2a)^{-1}+(a^{-1}-a)w_1)s^2+
\half((a^{-2}-1)(w_1+1)+2(1-w_1))s+(2a)^{-1}(1-w_1) \\
p(s) &= p_0(s)/p_1(s).
\end{align*}
All coefficients of powers of $s$ in $p_0(s)$ and $p_1(s)$
are strictly positive, and $p_0(s)$ is cubic whereas $p_1(s)$ is
quadratic. It follows that $p$ defines a continuous function
$[0,\infty)\to[0,\infty)$, with $p(0)=0$ and $p(s)\to\infty$ as
$s\to\infty$. We claim that this is strictly increasing. Indeed, by
standard algebra one can check that
$p'(s)=p_1(s)^{-2}\sum_{i=0}^4m_is^i$, where
\begin{align*}
m_0 &= \half a^{-1}(1-w_1)^2 \\
m_1 &= (1-w_1)((\tfrac{3}{2}a^{-2}-1)w_1+(a^{-2}+1)(1-w_1)) \\
m_2 &= \half a^{-3}(1-a^2)(1-w_1) +
a^{-3}(1-a^2)(\tfrac{3}{2}-a^2)w_1^2 + \\ & \qquad
\tfrac{1}{4}a^{-3}(1+2a^2)(5-a^2)w_1(1-w_1) +
\half a^{-1}(5+a^2)(1-w_1)^2 \\
m_3 &= a^{-2}(1+w_1)(2(1-a^2)w_1+(1+a^2)(1-w_1)) \\
m_4 &= \half a^{-1}(1+w_1)(1+2(1-a^2)w_1).
\end{align*}
We have written these coefficients in a form that makes it clear that
they are positive. It follows that $p'(s)>0$ for $s\geq 0$, as
claimed. It follows that there is a unique number $y_2\geq 0$ with
$p(y_2)=w_2$. We put
\[ y_1 = \sqrt{\frac{w_1(1+2ay_2)}{(y_2+a)(y_2+a^{-1})}} \]
and $y=(y_1,y_2)\in [0,\infty)^2$. We then define $u_1$ and $u_2$ in
terms of $y$ in the usual way. If we substitute the above value for
$y_1$, then straightforward algebra gives
\begin{align*}
u_1 &= (1-p(y_2))
\frac{((1-a^2)w_1+\half)y_2^2+(\half(1-w_1)(a+a^{-1})+w_1(a+a^{-1}))y_2+(1-w_1)/2}{(y_2+a)(y_2+a^{-1})} \\
u_2 &= (1+2ay_2)(1-w_1)/2.
\end{align*}
After recalling that $y_2\geq 0$ and $p(y_2)=w_2\in [0,1]$ it follows that
$u_1,u_2\geq 0$, so $y$ lies in $F_4^*$ and the point
$z=(y_1^2,y_2^2)$ lies in $F_{16}^*$. Now note that $y_2\geq 0$ and put
\begin{align*}
w'_1 &= \frac{z_1(1+(a+a^{-1})\sqrt{z_2}+z_2)}{1+2a\sqrt{z_2}} &
w'_2 &=
\sqrt{z_2}\frac{2a(1-z_1)+z_1\sqrt{z_2}}{1-z_1+(a^{-1}-a)z_1\sqrt{z_2}},
\end{align*}
so $q(z)=(w'_1,w'_2)$. If we substitute in our definition of $y_1$
and simplify we get $w'_1=w_1$ and $w'_2=p(y_2)$, but $p(y_2)=w_2$,
so $q(z)=w$. We leave it to the reader to check that all steps in this
construction are forced, so $z$ is the unique point in $F_{16}^*$ with
$q(z)=w$. This means that $q\:F_{16}^*\to [0,1]^2$ is a continuous
bijection between compact Hausdorff spaces, so it is a
homeomorphism.
\begin{checks}
embedded/invariants_check.mpl: check_invariants()
\end{checks}
\end{proof}
The following picture shows the images under $q\circ p_{16}$ of the
curves $c_i(t)$ and the points $v_j$.
\begin{center}
\begin{tikzpicture}[scale=6]
\draw[cyan] plot[smooth] coordinates{ (0.000,0.000) (0.000,0.156) (0.000,0.309) (0.000,0.454) (0.000,0.588) (0.000,0.707) (0.000,0.809) (0.000,0.891) (0.000,0.951) (0.000,0.988) (0.000,1.000) };
\draw[green] plot[smooth] coordinates{ (1.000,0.000) (0.976,0.000) (0.905,0.000) (0.794,0.000) (0.655,0.000) (0.500,0.000) (0.345,0.000) (0.206,0.000) (0.095,0.000) (0.024,0.000) (0.000,0.000) };
\draw[magenta] plot[smooth] coordinates{ (1.000,1.000) (0.977,1.000) (0.911,1.000) (0.805,1.000) (0.668,1.000) (0.514,1.000) (0.357,1.000) (0.214,1.000) (0.099,1.000) (0.026,1.000) (0.000,1.000) };
\draw[blue] plot[smooth] coordinates{ (1.000,0.000) (1.000,0.025) (1.000,0.097) (1.000,0.208) (1.000,0.348) (1.000,0.502) (1.000,0.656) (1.000,0.795) (1.000,0.905) (1.000,0.976) (1.000,1.000) };
\draw[cyan,arrows={-angle 90}] (0.0000,0.5403) -- (0.0000,0.5234);
\draw[green,arrows={-angle 90}] (0.5000,0.0000) -- (0.4900,0.0000);
\draw[magenta,arrows={-angle 90}] (0.5138,1.0000) -- (0.5037,1.0000);
\draw[blue,arrows={-angle 90}] (1.0000,0.5665) -- (1.0000,0.5715);
\fill (1.000,0.000) circle(0.005);
\fill (0.000,1.000) circle(0.005);
\fill (0.000,0.000) circle(0.005);
\fill (1.000,1.000) circle(0.005);
\draw (1.000,1.000) node[anchor=south] {$v_{11}$};
\draw (0.000,1.000) node[anchor=south] {$v_{3}$};
\draw (1.000,0.000) node[anchor=north west] {$v_{0}$};
\draw (0.000,0.000) node[anchor=north] {$v_{6}$};
\draw (0.000,0.540) node[anchor=east] {$c_{0}(\qpi\dotsb\ppi)$};
\draw (0.500,0.000) node[anchor=north] {$c_{1}(0\dotsb\ppi)$};
\draw (0.514,1.000) node[anchor=south] {$c_{3}(0\dotsb\ppi)$};
\draw (1.000,0.567) node[anchor=west] {$c_{5}(0\dotsb\pi)$};
\end{tikzpicture}
\end{center}
Some relevant formulae are as follows:
\begin{align*}
qp_{16}(v_i) &= (1,0) && \text{ for } 0\leq i \leq 1 \\
qp_{16}(v_i) &= (0,1) && \text{ for } 2\leq i \leq 5 \\
qp_{16}(v_i) &= (0,0) && \text{ for } 6\leq i \leq 9 \\
qp_{16}(v_i) &= (1,1) && \text{ for } 10\leq i \leq 13
\end{align*}
\begin{align*}
qp_{16}(c_0(t)) &=\! (0,\;|\cos(2t)|) \\
qp_{16}(c_1(t))=qp_{16}(c_2(t)) &=\! (\cos(t)^2,\;0) \\
qp_{16}(c_{\alg}(t)) &=\! \begin{cases}
\left(a^{-1}\frac{(1-2at)(t+a)(t+a^{-1})}{(1+2at)(t-a)(t-a^{-1})},\;1\right)
& \text{ if } t\in T_{\alg}^+ \\
\left(1,\;-at\frac{3-2a^2-4at}{2-2a^2-(3a+2a^3)t}\right)
& \text{ if } t\in T_{\alg}^-.
\end{cases}
\end{align*}
\begin{proposition}\label{prop-E-cromulent}
$EX(a)$ is a cromulent surface.
\end{proposition}
\begin{proof}
Condition~(a) in Definition~\ref{defn-precromulent} is proved in
Section~\ref{sec-E-G}, and conditions~(b) and~(c) are
proved in Section~\ref{sec-E-isotropy}. For condition~(d),
we can take $F'$ to be the interior of $F_{16}$.
\end{proof}
\subsection{Additional points and curves}
\label{sec-extra-curves}
The surface $EX(a)$ has some additional points and curves that are
not part of the precromulent structure but are nevertheless useful for
some purposes (such as our analysis of torus quotients of $EX^*$ in
Section~\ref{sec-torus-quotients}). All claims in this section are checked as
follows:
\begin{checks}
embedded/extra_vertices_check.mpl: check_extra_vertices()
embedded/extra_curves_check.mpl: check_extra_curves()
\end{checks}
\begin{definition}\label{defn-c-nine}
We put
\[ c_9(t) = \left(
\sqrt{\frac{1-a^2}{2(1+a^2)}}\sin(t),\;
\sqrt{\frac{1-a^2}{2(1+a^2)}}\sin(t),\;
\sqrt{\frac{2a^2}{1+a^2}},\;
-\sqrt{\frac{1-a^2}{1+a^2}}\cos(t)
\right),
\]
then
\[ c_{10}(t) = \lm(c_9(t)) \qquad
c_{11}(t) = \mu(c_9(t)) \qquad
c_{12}(t) = \lm\mu(c_9(t)).
\]
\end{definition}
It is straightforward to check that $c_9$ lands in $EX(a)$, so the
same is true for $c_{10}$, $c_{11}$ and $c_{12}$. One can also check
that $\lm^2c_9(t)=c_9(-t)$ and $\lm\nu(c_9(t))=c_9(\pi-t)$. It
follows that we cannot get anything interestingly new by applying
further group elements to the above curves. In the case $a=1/\rt$, we
have
\[ c_9(t) = \left(
\frac{\sin(t)}{\sqrt{6}},\;
\frac{\sin(t)}{\sqrt{6}},\;
\sqrt{\frac{2}{3}},\;
-\frac{\cos(t)}{\sqrt{3}}
\right).
\]
\begin{definition}\label{defn-c-thirteen}
We put
\[\begin{split}
c_{13}(t) = &\left(
\frac{\cos(t)}{\rt}\left(1-\frac{\sin(t)}{\sqrt{2/(1-a^2)-\cos(t)^2}}\right),\quad
\frac{\cos(t)}{\rt}\left(1+\frac{\sin(t)}{\sqrt{2/(1-a^2)-\cos(t)^2}}\right),
\right.\\ & \qquad\left.
\frac{\rt a}{\sqrt{1+a^2}}\,\sin(t),\quad
\frac{-\rt\sin(t)^2}{\sqrt{1+a^2}\sqrt{2/(1-a^2)-\cos(t)^2}}
\right),
\end{split}\]
then
\[
c_{14}(t) = \lm(c_{13}(t)) \qquad
c_{15}(t) = \nu(c_{13}(t)) \qquad
c_{16}(t) = \lm\nu(c_{13}(t)).
\]
\end{definition}
In the case $a=1/\rt$ this becomes
\[ c_{13}(t) = \left(
\frac{\cos(t)}{\rt}\left(1-\frac{\sin(t)}{\sqrt{4-\cos(t)^2}}\right),
\frac{\cos(t)}{\rt}\left(1+\frac{\sin(t)}{\sqrt{4-\cos(t)^2}}\right),
\sqrt{2/3}\,\sin(t),
\frac{-2\sin(t)^2}{\sqrt{3}\sqrt{4-\cos(t)^2}}
\right).
\]
It is straightforward to check that $c_{13}$ lands in $EX(a)$, so the
same is true for $c_{14}$, $c_{15}$ and $c_{16}$. One can also check
that $\lm\mu(c_{13}(t))=c_{13}(-t)$ and
$\lm^2(c_{13}(t))=c_{13}(\pi-t)$. It again follows that we cannot get
anything interestingly new by applying further group elements.
\begin{definition}\label{defn-c-seventeen}
If $a\leq 1/\rt$, we put
\[ c_{17}(t) = \left(
\sqrt{\frac{1-2a^2}{3(1+2a^2)}}\sin(t),\;
-\cos(t),\;
-\frac{4a}{\sqrt{6(1+2a^2)}}\sin(t),\;
\frac{2}{\sqrt{6(1+2a^2)}}\sin(t)
\right),
\]
then
\[ c_{18}(t) = \lm(c_{13}(t)) \qquad
c_{19}(t) = \lm^2(c_{13}(t)) \qquad
c_{20}(t) = \lm^3(c_{13}(t)).
\]
\end{definition}
One can again check that this produces curves in $EX(a)$ satisfying
$\nu c_{17}(t)=c_{17}(\pi-t)$ and $\lm^2\mu(c_{17}(t))=c_{17}(-t)$.
In the case $a=1/\rt$, we find that
$c_{17}(t)=c_3(-\tfrac{\pi}{2}-t)$, and similarly $c_{18}$, $c_{19}$
and $c_{20}$ are just reparametrisations of the lower numbered curves.
Note that for $k\in\{0,1,2,17,\dotsc,20\}$, the image $C_k=c_k(\R)$ is
a great circle. The homogeneous polynomial $g(x)$ defines a cubic
surface in the projective space $\mathbb{P}^3$, which is smooth except
when $a=1/\rt$. A famous theorem of Cayley and Salmon says that any
smooth cubic surface contains precisely 27 linearly embedded copies of
$\mathbb{P}^1$ (when counted with appropriate multiplicities);
see~\cite[Theorem 9.1.13]{do:cag} for a modern treatment. In our
case, the above great circles give seven copies of $\mathbb{P}^1$.
One can check that the remaining copies come in ten complex conjugate
pairs, and so do not correspond to great circles in the real variety
$EX(a)$. In the case $a=1/\rt$ everything degenerates and we have
only five great circles and two additional conjugate pairs of
$\mathbb{P}^1$'s. Some or all of these must have multiplicity greater
than one, but we have not investigated this.
\begin{checks}
embedded/cayley_check.mpl: check_cayley()
\end{checks}
For $k\in\{9,\dotsc,12\}$ the image is again the intersection of $S^3$
with a two-dimensional subspace of $\R^4$, but in these cases it is an
affine subspace rather than a vector subspace. We suspect that again
there are no more curves of this type contained in $EX(a)$, but we
have not proved this.
\begin{remark}
The curves $c_i(t)$ for $0\leq i\leq 16$ are represented in Maple as
\mcode+c_E[i](t)+. However, for $17\leq i\leq 20$, the curves
$c_i(t)$ are not defined when $a>1/\rt$, and this makes it
inconvenient to use the same framework. Instead, these curves are
represented in Maple by the functions \mcode+c_cayley[j](t)+ for
$1\leq j\leq 4$.
\end{remark}
We next introduce some additional points $v_i$ for $14\leq i\leq 45$.
For this, it is convenient to enumerate the elements of $G$ as
follows:
\begin{align*}
\gm_0 &= 1 & \gm_1 &= \lm &
\gm_2 &= \lm^2 & \gm_3 &= \lm^3 \\
\gm_4 &= \mu & \gm_5 &= \lm\mu &
\gm_6 &= \lm^2\mu & \gm_7 &= \lm^3\mu \\
\gm_8 &= \nu & \gm_9 &= \lm\nu &
\gm_{10} &= \lm^2\nu & \gm_{11} &= \lm^3\nu \\
\gm_{12} &= \mu\nu & \gm_{13} &= \lm\mu\nu &
\gm_{14} &= \lm^2\mu\nu & \gm_{15} &= \lm^3\mu\nu.
\end{align*}
\begin{definition}
We put
\[
v_{14} = \left(\sqrt{\frac{1-a^2}{2(1+a^2)}},\;
\sqrt{\frac{1-a^2}{2(1+a^2)}},\;
\sqrt{\frac{2a^2}{1+a^2}},\;
0\right),
\]
then $v_{14+i}=\gm_i(v_{14})$ for $0\leq i<8$. If
$a\leq 1/\rt$ we also put
\begin{align*}
v_{22} &= \left(
\sqrt{\frac{1-2a^2}{3(1+2a^2)}},\;
0,\;
\sqrt{\frac{8a^2}{3(1+2a^2)}},\;
-\sqrt{\frac{2}{3(1+2a^2)}}
\right) \\
v_{30} &= \left(
\sqrt{\frac{1-2a^2}{4(1+a^2)}},\;
\sqrt{\frac{1-2a^2}{4(1+a^2)}},\;
\sqrt{\frac{2a^2}{1+a^2}},\;
\sqrt{\frac{1}{2(1+a^2)}}
\right),
\end{align*}
then
\begin{align*}
v_{22+i} &= \gm_i(v_{22}) && (0\leq i<8) \\
v_{30+i} &= \gm_i(v_{30}) && (0\leq i<16).
\end{align*}
\end{definition}
Straightforward calculations show that these points lie in $EX(a)$.
We have $\lm\nu(v_{14})=v_{14}$ and $\nu(v_{22})=v_{22}$, which
implies that $\{v_0,\dotsc,v_{45}\}$ is closed under the action of
$G$. One can check that
\begin{align*}
C_1\cap C_9 &= \{v_{14},v_{16}\} \\
C_5\cap C_{19} &= \{v_{22}\} \\
C_9\cap C_{20} &= \{v_{30}\}.
\end{align*}
This is the main justification for considering these extra points.
\subsection{Charts}
\label{sec-E-charts}
In this section we discuss three different kinds of charts for $EX(a)$.
Our first construction is simple and works at every point of $EX(a)$. It
only gives an approximate chart, but that is sufficient for many
purposes.
\begin{definition}\label{defn-quadratic-chart}
Let $x$ be a point in $EX(a)$, and let $T$ be the tangent space to
$EX(a)$ at $x$. We define $\phi\:T\to\R^4$ by
\[ \phi(t) = \left(1 - \frac{\|t\|^2}{2}\right)x + t
- \frac{n(t).x}{\|n(x)\|^2}n(x).
\]
(Here $n(x)$ is the gradient of $g$, as in Definition~\ref{defn-nx}.)
We call this the \emph{quadratic approximate chart} at $x$. Maple
notation for $\phi(t)$ is \mcode+quadratic_chart(x,t)+ (defined in
\fname+embedded/geometry.mpl+).
\end{definition}
\begin{proposition}\label{prop-quadratic-chart}
We have
\begin{align*}
\phi(t) &= x + t + O(\|t\|^2) \\
g(\phi(t)) &= O(\|t\|^3) \\
\rho(\phi(t)) &= 1 + O(\|t\|^4).
\end{align*}
\end{proposition}
\begin{proof}
As the map $n\:\R^4\to\R^4$ is homogeneous quadratic, we see that
$\phi(t)=x+t+O(\|t\|^2)$. Next, as $x$, $n(x)$ and $t$ are mutually
orthogonal we have
\begin{align*}
\rho(\phi(t)) &=
\left(1 - \frac{\|t\|^2}{2}\right)^2 + \|t\|^2
+ \left(\frac{n(t).x}{\|n(x)\|^2}\right)^2\|n(x)\|^2 \\
&= 1 + \frac{\|t\|^4}{4} + \frac{(n(t).x)^2}{\|n(x)\|^2}.
\end{align*}
Using the fact that $n$ is homogeneous quadratic again, we see that
this is $1+O(\|t\|^4)$.
Next, using the fact that $g$ is homogeneous cubic we find that
\[ g(a+b) = g(a)+n(a).b + n(b).a + g(b). \]
We apply this with $a=\left(1 - \frac{\|t\|^2}{2}\right)x$ and
$b=t-\frac{n(t).x}{\|n(x)\|^2}n(x)$, neglecting terms of order
$\|t\|^3$ everywhere. As $g(x)=0$ we have $g(a)=0$. Moreover, $b$
is $O(\|t\|)$, so $g(b)$ is negligible, and when calculating $n(a).b$
we can neglect terms in $n(a)$ that are quadratic in $t$. This
leaves $n(a)\simeq n(x)$, and $n(x)$ is normal to $t$, so
\[ n(x).b \simeq -\frac{n(t).x}{\|n(x)\|^2} n(x).n(x) =
-n(t).x.
\]
Similarly, as $n$ is quadratic, we can neglect terms in $b$ that are
$O(\|t\|^2)$ when calculating $n(b)$. This gives $n(b)\simeq n(t)$
and so $n(b).a\simeq n(t).a\simeq n(t).x$. Altogether, we have
\[ g(\phi(t)) = g(a) + n(a).b + n(b).a + g(b) \simeq
0 -n(t).x + n(t).x + 0 = 0.
\]
\begin{checks}
embedded/geometry_check.mpl: check_quadratic_chart()
\end{checks}
\end{proof}
Next, recall from Section~\ref{sec-holomorphic-curves} that each of
the maps $c_k\:\R\to EX(a)$ can be extended in a canonical way to give
a holomorphic map $\tc_k$ defined on a neighbourhood of $\R$ in $\C$.
The file \fname+embedded/annular_charts.mpl+ gives a formula for
$\tc_0(t+iu)$ modulo $u^4$, and formulae for $\tc_1(t+iu)$ and
$\tc_2(t+iu)$ modulo $u^3$, but we will not reproduce them here.
Given a fixed value of $a$ and $t_0\in\R$ it is also not hard to
compute power series for $\tc_k(t_0+t+iu)$ to reasonably high order,
and similar methods can be used to produce series for conformal charts
centred at points that do not lie on any of the curves $C_k$. We
postpone a more detailed discussion to
Section~\ref{sec-roothalf-charts}, where we focus on the case
$a=1/\rt$.
We next describe a different class of charts that will be useful for
triangulating $EX(a)$. It is inspired by the definition of barycentric
coordinates for spherical triangles described in~\cite{labese:sbc}.
Related code is in the file \fname+embedded/barycentric.mpl+.
\begin{definition}\label{defn-barycentric}
Let $a_0$, $a_1$ and $a_2$ be distinct points on $EX(a)$. For any
$x\in EX(a)$ we let $n(x)$ denote the gradient of $g$ at $EX(a)$, and we put
\[ \widetilde{p}(x) =
(\det(x,n(x),a_1,a_2),\;
\det(x,n(x),a_2,a_0),\;
\det(x,n(x),a_0,a_1)) \in \R^3.
\]
If $\sum_i\widetilde{p}(x)_i\neq 0$, we put
\[ p(x) = \widetilde{p}(x)/\sum_i\widetilde{p}(x)_i. \]
This clearly lies in the set $\R^3_1=\{t\in\R^3\st\sum_it_i=1\}$,
which contains the simplex $\Dl_2$. If we need to emphasise the
dependence on the points $a_i$, we will write $p_{a_0,a_1,a_2}(x)$
rather than $p(x)$. We call the components of $p(x)$ the
\emph{barycentric coordinates} of $x$ (with respect to the $a_i$).
\end{definition}
\begin{definition}
For $x\in EX(a)$ we put $T'_xEX(a)=x+T_xEX(a)\subset\R^4$, and call this the
\emph{affine tangent space} to $EX(a)$ at $x$. In a small neighbourhood
of $x$, this is of course a good approximation to $EX(a)$ itself. We
write $\pi'_x(y)$ for the closest point in $T'_xEX(a)$ to $y$. This can
be computed as
\[ \pi'_x(y) = x+y-\ip{y,x}x-\ip{y,n(x)}n(x)/\|n(x)\|^2. \]
\end{definition}
The next result motivates the term ``barycentric coordinates''.
\begin{lemma}
If $p(x)$ is defined, then it is the unique element $t\in\R^3_1$ such
that $x=\sum_it_i\pi'_x(a_i)$.
\end{lemma}
\begin{proof}
First, we write $\tu=\widetilde{p}(x)$, so
\begin{align*}
\tu_0 &= \det(x,n(x),a_1,a_2) \\
\tu_1 &= \det(x,n(x),a_2,a_0) \\
\tu_2 &= \det(x,n(x),a_0,a_1).
\end{align*}
As $p(x)$ is assumed to be defined, we must have
$\tu_0+\tu_1+\tu_2\neq 0$, and in particular
$(\tu_0,\tu_1,\tu_2)\neq(0,0,0)$.
Next, the list $a_0,a_1,a_2,x,n(x)$ consists of five vectors in
$\R^4$, so there must be a linear relation
\[ s_0a_0 + s_1a_1 + s_2a_2 + s_3x + s_4n(x) = 0 \]
where not all of the coefficients $s_i$ are zero. As $x$ and $n(x)$
are nonzero and orthogonal, it follows easily that
$(s_0,s_1,s_2)\neq(0,0,0)$.
Now apply the map $y\mapsto\det(x,n(x),a_0,y)$ to our linear
relation. Only the second and third terms contribute anything, and
we deduce that $s_1\tu_2-s_2\tu_1=0$. Similarly, we can apply the
map $y\mapsto\det(x,n(x),a_1,y)$ to see that $s_0\tu_2-s_2\tu_0=0$,
and we can apply the map $y\mapsto\det(x,n(x),a_2,y)$ to see that
$s_0\tu_1-s_1\tu_0=0$. We have already seen that the vectors
$(\tu_0,\tu_1,\tu_2)$ and $(s_0,s_1,s_2)$ are nonzero, and the above
relations imply that they are unit multiples of each other. It
follows that the vector $u=p(x)=\tu/\sum_i\tu_i$ is also a unit
multiple of $(s_0,s_1,s_2)$. The definition of the coefficients
$s_i$ implies that
\[ s_0a_0+s_1a_1+s_2a_2\in\text{span}(x,n(x)) = (T_xEX(a))^\perp, \]
and we now see that $\sum_iu_ia_i\in(T_xEX(a))^\perp$ as well. From this
it follows easily that $x=\sum_iu_i\pi'_x(a_i)$ as claimed.
All that is left is to check that the numbers $u_i$ are uniquely
characterised by the above property. Suppose there is another
vector $u'\in\R^3_1$ with $x=\sum_iu'_i\pi'_x(a_i)$. It
follows that the vector $r=u'-u$ has $\sum_ir_i=0$ and
$\sum_ir_ia_i\in(T_xEX(a))^\perp=\text{span}(x,n(x))$, so there exist
scalars $r_3,r_4$ such that
\[ r_0a_0 + r_1a_1 + r_2a_2 + r_3x + r_4n(x) = 0. \]
Just as before we deduce that $r_i\tu_j-r_j\tu_i=0$ for all
$i,j\in\{0,1,2\}$, so $(r_0,r_1,r_2)$ is a multiple of
$(\tu_0,\tu_1,\tu_2)$. As $\sum_ir_i=0$ but $\sum_i\tu_i\neq 0$, we
see that the multiplier must be zero, so $u'=u$ as claimed.
\end{proof}
From the above characterisation, we can show that barycentric
coordinates for adjacent triangles are equal on the shared edge, in
the following sense.
\begin{corollary}\label{cor-edge-barycentric}
Suppose we have points $a_0,a_1,b,b',x\in EX(a)$ such that the vectors
$u=p_{a_0,a_1,b}(x)$ and $u'=p_{a_0,a_1,b'}(x)$ are both defined, and
that $u_2=0$. Then $u'=u$ (and in particular, $u'_2$ is also zero).
\end{corollary}
\begin{proof}
We can apply the lemma to see that
$x=u_0\pi'_x(a_0)+u_1\pi'_x(a_1)$. On the other hand, $u'$ is
uniquely characterised by the fact that
$x=u'_0\pi'_x(a_0)+u'_1\pi'_x(a_1)+u'_2\pi'_x(b')$. The claim is
clear from this.
\end{proof}
\begin{remark}
A similar uniqueness argument shows that if $p_{a_0,a_1,a_2}(a_i)$ is
defined, then it must be equal to the $i$'th standard basis vector
$e_i$. More precisely, it is clear from the definitions that
$\widetilde{p}_{a_0,a_1,a_2}(a_i)_j=0$ for all $j\neq i$, so either
$\widetilde{p}_{a_0,a_1,a_2}(a_i)_i\neq 0$ (and
$p_{a_0,a_1,a_2}(a_i)=e_i$) or $\widetilde{p}_{a_0,a_1,a_2}(a_i)_i=0$
(and $p_{a_0,a_1,a_2}(a_i)$ is undefined).
\end{remark}
\begin{definition}
We put
\[ T_0(a_0,a_1,a_2) = \{x\in EX(a)\st p_{a_0,a_1,a_2}(x)
\text{ is defined and lies in } \Dl_2 \}.
\]
Typically, if the $a_i$ are close together and in general position,
there will be a single connected component
$T(a_0,a_1,a_2)\sse T_0(a_0,a_1,a_2)$ that contains all the points
$a_i$, and the map $p_{a_0,a_1,a_2}$ will restrict to give a
diffeomorphism $T(a_0,a_1,a_2)\to\Dl_2$. We can use maps of this
form to give a triangulation of $EX(a)$, with compatibility along edges
given by Corollary~\ref{cor-edge-barycentric}.
\end{definition}
\begin{remark}\label{rem-barycentric-inverse}
The inverse of the map $p\:T(a_0,a_1,a_2)\to\Dl_2$ can be computed
efficiently by a kind of Newton-Raphson method. If $x$ is reasonably
close to $p^{-1}(t)$ then the first order Taylor approximation at $x$
to the map $p\:EX(a)\to\R^3_1$ will be an affine isomorphism
$p_*\:T'_xEX(a)\to\R^3_1$, so we can define
$\kp(x)=\sg^{\infty}(p_*^{-1}(t))$, where $\sg^\infty$ is as in
Remark~\ref{rem-move-to-X}. As an initial approximation we can take
$x_0=\sg^{\infty}(\sum_it_ia_i)$, and then the sequence
$(\kp^n(x_0))_{n\geq 0}$ converges rapidly to $p^{-1}(t)$. If we
need to do this for a large number of different points $t$, we can
also speed up the method by precomputing various coefficients that
depend only on the points $a_i$.
\begin{checks}
embedded/barycentric_check.mpl: check_barycentric()
\end{checks}
\end{remark}
One could attempt to triangulate $EX^*$ using the points $v_i$ as
vertices, but it turns out that the resulting simplices are too large
for barycentric coordinates to work properly. We have tried several
different triangulations with smaller simplices. In one of these, we
introduce a new set of points $a_{ij}$ for $0\leq i\leq 6$ and $0\leq
j\leq 4$, and use them as vertices for a triangulation of $F_{16}$
with $48$ triangles. We can then use the group action to obtain a
triangulation of all of $EX^*$, with $768$ triangles. It is
convenient to use points with good rationality properties, as
described in Section~\ref{sec-rational}. This ensures that at least
the first steps of our calculations can be specified exactly. It is
also convenient to choose points such that the various edges have
lengths that do not differ by too large a factor. We use the
following corner points:
{\tiny \[ \hspace{-6em}
\renewcommand{\arraystretch}{1.5} \begin{array}{llll}
\left(\tfrac{1}{2}\sqrt{2},\tfrac{1}{2}\sqrt{2},0,0\right) &
\left(0,1,0,0\right) & \left(0,0,1,0\right) &
\left(0,0,\tfrac{1}{3}\sqrt{6},-\tfrac{1}{3}\sqrt{3}\right)
\end{array} \]}
Plus the following additional edge points:
{\tiny \[ \hspace{-6em} \renewcommand{\arraystretch}{1.5} \begin{array}{lll}
\left(\tfrac{407}{745},\tfrac{624}{745},0,0\right) &
\left(\tfrac{5}{13},\tfrac{12}{13},0,0\right) &
\left(\tfrac{9}{41},\tfrac{40}{41},0,0\right) \\
\left(\tfrac{1}{60}\sqrt{195},0,\tfrac{1}{4}\sqrt{15},-\tfrac{1}{60}\sqrt{30}\right) &
\left(\tfrac{3}{70}\sqrt{55},0,\tfrac{2}{7}\sqrt{10},-\tfrac{9}{70}\sqrt{5}\right) &
\left(\tfrac{1}{68}\sqrt{238},0,\tfrac{1}{12}\sqrt{102},-\tfrac{7}{102}\sqrt{51}\right)
\end{array} \]}
{\tiny \[ \hspace{-6em} \renewcommand{\arraystretch}{1.5} \begin{array}{lllll}
\left(\tfrac{23}{34},\tfrac{23}{34},\tfrac{7}{34}\sqrt{2},0\right) &
\left(\tfrac{79}{130},\tfrac{79}{130},\tfrac{47}{130}\sqrt{2},0\right) &
\left(\tfrac{1}{2},\tfrac{1}{2},\tfrac{1}{2}\sqrt{2},0\right) &
\left(\tfrac{79}{202},\tfrac{79}{202},\tfrac{119}{202}\sqrt{2},0\right) &
\left(\tfrac{7}{34},\tfrac{7}{34},\tfrac{23}{34}\sqrt{2},0\right) \\
\left(0,\tfrac{59}{62},\tfrac{11}{62}\sqrt{2},-\tfrac{11}{62}\right) &
\left(0,\tfrac{11}{13},\tfrac{4}{13}\sqrt{2},-\tfrac{4}{13}\right) &
\left(0,\tfrac{61}{86},\tfrac{35}{86}\sqrt{2},-\tfrac{35}{86}\right) &
\left(0,\tfrac{1}{2},\tfrac{1}{2}\sqrt{2},-\tfrac{1}{2}\right) &
\left(0,\tfrac{11}{38},\tfrac{21}{38}\sqrt{2},-\tfrac{21}{38}\right)
\end{array} \]}
\medskip
Plus the following points in the interior:
{\tiny \[ \hspace{-6em} \renewcommand{\arraystretch}{1.5} \begin{array}{lll}
\left(\tfrac{92124}{152207},\tfrac{232883}{304414},\tfrac{31}{202}\sqrt{2},-\tfrac{11005}{304414}\right) &
\left(\tfrac{15}{38},\tfrac{86}{95},\tfrac{1}{10}\sqrt{2},-\tfrac{13}{190}\right) &
\left(\tfrac{4793}{21846},\tfrac{72098}{76461},\tfrac{23}{154}\sqrt{2},-\tfrac{20585}{152922}\right) \\
\left(\tfrac{1287}{2425},\tfrac{344}{485},\tfrac{8}{25}\sqrt{2},-\tfrac{248}{2425}\right) &
\left(\tfrac{214099}{533478},\tfrac{217940}{266739},\tfrac{73}{274}\sqrt{2},-\tfrac{91469}{533478}\right) &
\left(\tfrac{1121}{4510},\tfrac{1864}{2255},\tfrac{25}{82}\sqrt{2},-\tfrac{237}{902}\right) \\
\left(\tfrac{7267}{15458},\tfrac{8565}{15458},\tfrac{125}{262}\sqrt{2},-\tfrac{1000}{7729}\right) &
\left(\tfrac{37323}{100798},\tfrac{29390}{50399},\tfrac{95}{202}\sqrt{2},-\tfrac{28595}{100798}\right) &
\left(\tfrac{22828}{111879},\tfrac{1572959}{2461338},\tfrac{907}{2046}\sqrt{2},-\tfrac{975025}{2461338}\right) \\
\left(\tfrac{92619}{240218},\tfrac{224276}{600545},\tfrac{359}{610}\sqrt{2},-\tfrac{166217}{1201090}\right) &
\left(\tfrac{5389}{14982},\tfrac{27671}{74910},\tfrac{1301}{2270}\sqrt{2},-\tfrac{10408}{37455}\right) &
\left(\tfrac{342221}{1435238},\tfrac{4932765}{10046666},\tfrac{1549}{3038}\sqrt{2},-\tfrac{2143816}{5023333}\right) \\
\left(\tfrac{143}{486},\tfrac{298}{1701},\tfrac{83}{126}\sqrt{2},-\tfrac{415}{3402}\right) &
\left(\tfrac{42732}{137557},\tfrac{20435}{137557},\tfrac{280}{457}\sqrt{2},-\tfrac{49720}{137557}\right) &
\left(\tfrac{23}{119},\tfrac{4}{17},\tfrac{4}{7}\sqrt{2},-\tfrac{60}{119}\right)
\end{array} \]}
To create a triangulation of $F_{16}$, we link the above points according to the
following combinatorial scheme:
\begin{center}
\begin{tikzpicture}
\draw[cyan] (0,0) -- (0,4);
\draw[green] (0,0) -- (6,0);
\draw[blue] (6,0) -- (6,4);
\draw[magenta] (0,4) -- (6,4);
\draw (1,0) -- (1,4);
\draw (2,0) -- (2,4);
\draw (3,0) -- (3,4);
\draw (4,0) -- (4,4);
\draw (5,0) -- (5,4);
\draw (0,1) -- (6,1);
\draw (0,2) -- (6,2);
\draw (0,3) -- (6,3);
\draw (0,1) -- (1,2);
\draw (0,0) -- (2,2);
\draw (1,0) -- (3,2);
\draw (2,0) -- (3,1);
\draw (0,3) -- (1,2);
\draw (0,4) -- (2,2);
\draw (1,4) -- (3,2);
\draw (2,4) -- (3,3);
\draw (6,1) -- (5,2);
\draw (6,0) -- (4,2);
\draw (5,0) -- (3,2);
\draw (4,0) -- (3,1);
\draw (6,3) -- (5,2);
\draw (6,4) -- (4,2);
\draw (5,4) -- (3,2);
\draw (4,4) -- (3,3);
\end{tikzpicture}
\end{center}
This has been arranged to ensure that no edge in the triangulation has
endpoints on two different sides of $F_{16}$, which would cause
trouble in certain numerical algorithms. We have also used a finer
triangulation obtained by subdividing each of the above triangles in
the following pattern:
\begin{center}
\begin{tikzpicture}[scale=2]
\begin{scope}
\fill (-0.50,0.00) circle(0.02);
\fill ( 0.50,0.00) circle(0.02);
\fill ( 0.00,0.86) circle(0.02);
\draw (-0.50,0.00) -- (0.5,0) -- (0,0.86) -- cycle;
\end{scope}
\begin{scope}[xshift=1.5cm]
\draw (0,0.43) -- (1,0.43);
\draw (0.9,0.53) -- (1,0.43) -- (0.9,0.33);
\end{scope}
\begin{scope}[xshift=4cm]
\fill (-0.50,0.00) circle(0.02);
\fill ( 0.50,0.00) circle(0.02);
\fill ( 0.00,0.86) circle(0.02);
\fill (-0.25,0.43) circle(0.02);
\fill ( 0.25,0.43) circle(0.02);
\fill ( 0.00,0.00) circle(0.02);
\draw (-0.50,0.00) -- (0.5,0) -- (0,0.86) -- cycle;
\draw (-0.25,0.43) -- (0.25,0.43) -- (0,0) -- cycle;
\end{scope}
\end{tikzpicture}
\end{center}
This gives a grid with 192 faces in $F_{16}$.
Information about this kind of triangulation can be encoded in an
object of the class \mcode+E_grid+, which is declared in the file
\fname+embedded/E_domain.mpl+. This class extends the \mcode+grid+
class, which is declared in \fname+domain/grid.mpl+. (There used to
be parallel classes \mcode+H_grid+ and \mcode+P_grid+, but we found that
various algorithms based on triangulations were not very effective, so
we have not maintained that code.) In particular, the 192 face
triangulation described above is encoded in this form and stored in
the file
\fname+embedded/roothalf/split_rational_grid_wx_30.mpl+ in the
\mcode+data+ directory. One can
thus enter
\begin{mcodeblock}
read(cat(data_dir,"/embedded/roothalf/split_rational_grid_wx_30.mpl"));
G := eval(split_rational_grid_wx_30):
G["num_points"];
\end{mcodeblock}
This will print $117$, indicating that the triangulation has $117$
vertices lying in $F_{16}$. As well as the obvious data about the
vertices, edges and faces of the triangulation, the object \mcode+G+
also contains extensive information about $175$ sample points in each of
the $192$ faces. This can be used for computing integrals over
$EX^*$, as will be explained in Section~\ref{sec-integration}. All of
this information is computed to $100$ decimal places. Because of
this, the file is rather large (about 53MB).
One can also regenerate the object \mcode+G+ using the function
\begin{mcodeblock}
build_data["grid"]();
\end{mcodeblock}
defined in \fname+build_data.mpl+. See
Section~\ref{sec-build} for more discussion of this framework.
\subsection{Curvature and the Laplacian}
\label{sec-E-curvature}
We now discuss the Gaussian curvature of $EX(a)$. Most treatments of
this invariant are formulated in terms of local coordinates on the
manifold, but for us it is more useful to have a formula in terms of
the coordinates $x_i$ for the ambient space $\R^4$. We have not been
able to find a reference for the formula given below, although it
would be surprising if it did not appear somewhere. Most of our work
will be valid for $EX(a)$ for all $a$, but we will focus on the case
$a=1/\rt$ for simplicity.
Let $n(x)\in\R^4$ be the gradient of $g$ at $x$, and let
$m(x)\in M_4(\R)$ be the Hessian, so
\begin{align*}
n(x)_i &= \partial g(x)/\partial x_i \\
m(x)_{ij} &= \partial^2 g(x)/\partial x_i\,\partial x_j.
\end{align*}
Explicitly, for $a=1/\rt$ we have
\begin{align*}
n(x) &= \left[\begin{array}{c}
-4 x_1 x_4+2\rt x_1 x_3 \\
-4 x_2 x_4-2\rt x_2 x_3 \\
2 x_3 x_4+\rt x_1^2-\rt x_2^2 \\
-2 x_1^2-2 x_2^2+x_3^2-6 x_4^2
\end{array}\right] \\
m(x) &= \left[ \begin {array}{cccc}
-4 x_4+2\rt x_3&0&2\rt x_1&-4 x_1 \\
0&-4 x_4-2\rt x_3&-2\rt x_2&-4 x_2\\
2\rt x_1&-2\rt x_2&2 x_4&2 x_3 \\
-4 x_1&-4 x_2&2 x_3&-12 x_4
\end {array} \right] .
\end{align*}
Next, we let $\xi$ be the usual isomorphism $\Lm^3(\R^4)\to\R^4$,
given by
\begin{align*}
\xi(e_1\wedge e_2\wedge e_3) &= \pp e_4 \\
\xi(e_1\wedge e_2\wedge e_4) &= -e_3 \\
\xi(e_1\wedge e_3\wedge e_4) &= \pp e_2 \\
\xi(e_2\wedge e_3\wedge e_4) &= -e_1,
\end{align*}
so
\[ \xi^{-1}(u)\wedge v =
\ip{u,v} e_1\wedge e_2\wedge e_3\wedge e_4.
\]
We then let $p(x)\in M_4(\R)$ denote the unique matrix such that
\[ p(x)y = \xi(x\wedge n(x)\wedge m(x)y) \]
for all $y\in\R^4$.
\begin{theorem}\label{thm-curvature}
The Gaussian curvature of $EX^*$ at $x$ is
$1+\text{trace}(p(x)^2)/\|n(x)\|^2$.
\end{theorem}
The proof will follow after some preliminaries.
\begin{definition}
We put
\[ g_{ijk} = \frac{1}{6}
\frac{\partial^3g(x)}{\partial x_i\partial x_j\partial x_k}.
\]
It is clear that $g_{ijk}$ is invariant under permutations of the
three indices. As $g$ is a homogeneous cubic, we see that $g_{ijk}$
is constant, and that $g(x)=\sum_{ijk}g_{ijk}x_ix_jx_k$. By
differentiating this we obtain $n(x)_p=3\sum_{j,k}g_{pjk}x_jx_k$ and
$m(x)_{pq}=6\sum_kg_{pqk}x_k$.
\end{definition}
\begin{remark}
Some of our calculations in this section are most easily understood
in terms of Penrose diagrams. The paper~\cite{jost:gtci} is a good
reference for these from a mathematical perspective. The above
expressions for $g(x)$, $n(x)$ and $m(x)$ can be expressed
graphically as follows:
\begin{center}
\begin{tikzpicture}[scale=3]
\begin{scope}
\draw[red] (-30:0.2) -- (90:0.2) -- (210:0.2) -- cycle;
\draw (0,0) node {$g$};
\draw (-30:0.2) -- (-30:0.6);
\draw ( 90:0.2) -- ( 90:0.6);
\draw (210:0.2) -- (210:0.6);
\draw[blue] (-30:0.7) circle(0.1);
\draw[blue] ( 90:0.7) circle(0.1);
\draw[blue] (210:0.7) circle(0.1);
\draw (-30:0.7) node {$x$};
\draw ( 90:0.7) node {$x$};
\draw (210:0.7) node {$x$};
\draw (270:0.7) node {$g(x)$};
\end{scope}
\begin{scope}[xshift=2cm]
\draw[red] (-30:0.2) -- (90:0.2) -- (210:0.2) -- cycle;
\draw (0,0) node {$g$};
\draw (-30:0.2) -- (-30:0.6);
\draw ( 90:0.2) -- ( 90:0.6);
\draw (210:0.2) -- (210:0.6);
\draw[blue] (-30:0.7) circle(0.1);
\draw[blue] ( 90:0.7) circle(0.1);
\draw[blue] (210:0.7) circle(0.1);
\draw (-30:0.7) node {$x$};
\draw ( 90:0.7) node {$u$};
\draw (210:0.7) node {$x$};
\draw (270:0.7) node {$n(x).u/3$};
\end{scope}
\begin{scope}[xshift=4cm]
\draw[red] (-30:0.2) -- (90:0.2) -- (210:0.2) -- cycle;
\draw (0,0) node {$g$};
\draw (-30:0.2) -- (-30:0.6);
\draw ( 90:0.2) -- ( 90:0.6);
\draw (210:0.2) -- (210:0.6);
\draw[blue] (-30:0.7) circle(0.1);
\draw[blue] ( 90:0.7) circle(0.1);
\draw[blue] (210:0.7) circle(0.1);
\draw (-30:0.7) node {$x$};
\draw ( 90:0.7) node {$u$};
\draw (210:0.7) node {$v$};
\draw (270:0.7) node {$(m(x)u).v/6$};
\end{scope}
\end{tikzpicture}
\end{center}
\end{remark}
We next need a little exterior algebra. To keep everything straight,
we need to spell out some conventions.
\begin{definition}\label{defn-exterior}
For any vector space $V$, we will identify $\lm^k(V)$ with a subspace
of $V^{\ot k}$ in such a way that $v_1\wedge\dotsb\wedge v_k$ becomes
\[ \sum_{\sg\in\Sg_k}
\ep(\sg) v_{\sg(1)}\ot\dotsb\ot v_{\sg(k)}.
\]
If $V$ has an inner product, we give $V^{\ot k}$ the inner product
such that
\[ \ip{v_1\ot\dotsb\ot v_k,\;w_1\ot\dotsb\ot w_k} =
\prod_i \ip{v_i,w_i}.
\]
However, we give the subspace $\lm^k(V)$ the alternative inner
product $\ip{\al,\bt}'=\ip{\al,\bt}/k!$; this has the property that
\[ \ip{v_1\wedge\dotsb\wedge v_k,\;w_1\wedge\dotsb\wedge w_k}' \]
is the determinant of the matrix of inner products $\ip{v_i,w_j}$.
Now suppose that $V$ has dimension $d$ and we have a given volume
form $\om\in\lm^d(V)$ with $\ip{\om,\om}'=1$. We define the Hodge
operator $*\:\lm^k(V)\to\lm^{d-k}(V)$ by the property
\[ \al\wedge *\bt = \ip{\al,\bt}'\om. \]
\end{definition}
\begin{lemma}\label{lem-hodge}
Suppose that $x\in EX(a)$, and let $(u,v)$ be any oriented
orthonormal basis for $T_xEX^*$. Let $\ep$ be the usual totally
antisymmetric tensor:
\[ \ep_{ijkl} = \begin{cases}
+1 & \text{ if $(i,j,k,l)$ is an even permutation of $(1,2,3,4)$} \\
-1 & \text{ if $(i,j,k,l)$ is an odd permutation of $(1,2,3,4)$} \\
0 & \text{ if $(i,j,k,l)$ is not a permutation of $(1,2,3,4)$}.
\end{cases}
\]
Then
\[ u\wedge v = u\ot v - v\ot u =
\frac{1}{2\|n\|}\sum_{ijkl}\ep_{ijkl}x_in(x)_je_k\wedge e_l.
\]
\end{lemma}
\begin{proof}
This is standard, except that we need to check that our conventions
give the indicated factor of two.
The vectors $x$ and $w=n(x)/\|n(x)\|$ form an orthonormal basis for
$T_xEX(a)^\perp$, and it follows that the map
$\phi\:\al\mapsto x\wedge w\wedge\al$ gives an isomorphism
$\lm^2(T_xEX(a))\to\lm^4(\R^4)=\R.\om_4$. As $(u,v)$ is an oriented
orthonormal basis for $T_xEX(a)$, we have $\phi(u\wedge v)=\om_4$.
Now put
\[ \tht =
\frac{1}{2\|n\|}\sum_{ijkl}\ep_{ijkl}x_in(x)_je_k\wedge e_l =
\frac{1}{2}\sum_{ijkl}\ep_{ijkl}x_iw_je_k\wedge e_l.
\]
From the definitions we have
\[ x\wedge w\wedge e_k\wedge e_l =
\sum_{p,q} \ep_{pqkl} x_pw_q\om_4.
\]
This gives
\[ \phi(\tht) = x\wedge w\wedge\tht =
\frac{1}{2}\sum_{ijklpq}
\ep_{ijkl}\ep_{pqkl} x_ix_pw_jw_q\om_4.
\]
One can also check that
\[ \sum_{kl} \ep_{ijkl}\ep_{pqkl} =
2\dl_{ip}\dl_{jq} - 2\dl_{iq}\dl_{jp}.
\]
(For example, if $i=p=1$ and $j=q=2$ then the terms for $(k,l)=(3,4)$
and $(k,l)=(4,3)$ both contribute $+1$, and all other terms are zero
so the sum is $+2$. On the other hand, if $i=q=1$ and $j=p=2$ then
the terms for $(k,l)=(3,4)$ and $(k,l)=(4,3)$ both contribute $-1$,
and all other terms are zero so the sum is $-2$.) This gives
\begin{align*}
\phi(\tht) &=
\sum_{ijpq}
(\dl_{ip}\dl_{jq} - \dl_{iq}\dl_{jp}) x_ix_pw_jw_q\om_4 \\
&= \left(\sum_{ij} x_i^2w_j^2 - \sum_{ij} x_iw_ix_jw_j\right)\om_4 \\
&= (\|x\|^2\|w\|^2 - \ip{x,w}^2)\om_4 = \om_4.
\end{align*}
As $\phi$ is an isomorphism, this means that $u\wedge v=\tht$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm-curvature}]
Put $r=\|n(x)\|$. Let $N$ be the span of $x$ and $n(x)$, so the
tangent space to $EX^*$ at $x$ is the space $T=N^\perp$. Define a
quadratic map $\psi_0\:T\to N$ by
\[ \psi_0(t) = (1 - \|t\|^2/2)x - (n(t).x)n(x)/r^2, \]
and put
\[ X' = \text{ graph of } \psi_0 = \{t+\psi_0(t)\st t\in T\}. \]
Proposition~\ref{prop-quadratic-chart} tells us that $X'$ agrees with
$EX^*$ to second order near $x$. As curvature depends only on second
derivatives, the curvature of $EX^*$ at $x$ is the same as that of
$X'$.
Now choose an orthonormal basis $(u,v)$ for $T$, oriented so that
$x\wedge n(x)\wedge u\wedge v$ is a positive multiple of
$e_1\wedge e_2\wedge e_3\wedge e_4$. Define $\psi\:\R^2\to N$ by
$\psi(a,b)=\psi_0(au+bv)$. It will be enough to calculate the
curvature of the graph of $\psi$ at the point where $a=b=0$.
There is a well-known formula for the curvature of the graph of a
function $\psi(a,b)$ from $\R^2$ to $\R$ (rather than $N$): we put
\begin{align*}
E &= 1 + \frac{\partial \psi}{\partial a} . \frac{\partial \psi}{\partial a} &
F &= \frac{\partial \psi}{\partial a} . \frac{\partial \psi}{\partial b} &
G &= 1 + \frac{\partial \psi}{\partial b} . \frac{\partial \psi}{\partial b} \\
L &= \frac{\partial^2\psi}{\partial^2 a} &
M &= \frac{\partial^2\psi}{\partial a\;\partial b} &
N &= \frac{\partial^2\psi}{\partial^2 b},
\end{align*}
where everything is evaluated at $(0,0)$. The curvature is then
\[ K = \frac{LN-M^2}{EG-F^2}. \]
Essentially the same argument works in our case, except that now $L$,
$M$ and $N$ are vectors, and the formula is
\[ K = \frac{L.N-M.M}{EG-F^2}. \]
In our case $\psi$ is constant plus quadratic and so the first order
derivatives vanish at $(a,b)=(0,0)$, which gives $E=G=1$ and $F=0$,
so $K=L.N-M.M$.
Next, using the fact that $\|t\|^2$ and $n(t)$ are homogeneous
quadratic functions of $t$, we see that
\begin{align*}
L &= -x - 2(x.n(u))n(x)/r^2 \\
N &= -x - 2(x.n(v))n(x)/r^2 \\
M &= -(x.m(u)v)n(x)/r^2.
\end{align*}
Here $x.n(x)=3g(x)=0$, and $n(u)$ can be written as $m(u)u/2$, and
similarly for $v$. This gives
\begin{align*}
K &= L.N - M.M \\
&= x.x +
\frac{n(x).n(x)}{r^4}
\left(4\;x.n(u)\;x.n(v)-(x.m(u)v)^2\right) \\
&= 1 + \frac{1}{r^2}
\left((x.m(u)u)(x.m(v)v)-(x.m(u)v)^2\right).
\end{align*}
We put $P=(x.m(u)u)(x.m(v)v)-(x.m(u)v)^2$ so that $K=1+P/r^2$. Note
that $P/36$ can be expressed as the following difference of Penrose
diagrams:
\begin{center}
\begin{tikzpicture}[scale=2]
\draw (-0.95,-0.7) -- (-1,-0.7) -- (-1,1) -- (-0.95,1);
\begin{scope}
\draw[red] (-30:0.2) -- (90:0.2) -- (210:0.2) -- cycle;
\draw (0,0) node {$g$};
\draw (-30:0.2) -- (-30:0.6);
\draw ( 90:0.2) -- ( 90:0.6);
\draw (210:0.2) -- (210:0.6);
\draw[blue] (-30:0.7) circle(0.1);
\draw[blue] ( 90:0.7) circle(0.1);
\draw[blue] (210:0.7) circle(0.1);
\draw ( 90:0.7) node {$x$};
\draw (-30:0.7) node {$u$};
\draw (210:0.7) node {$u$};
\end{scope}
\begin{scope}[xshift=1.8cm]
\draw[red] (-30:0.2) -- (90:0.2) -- (210:0.2) -- cycle;
\draw (0,0) node {$g$};
\draw (-30:0.2) -- (-30:0.6);
\draw ( 90:0.2) -- ( 90:0.6);
\draw (210:0.2) -- (210:0.6);
\draw[blue] (-30:0.7) circle(0.1);
\draw[blue] ( 90:0.7) circle(0.1);
\draw[blue] (210:0.7) circle(0.1);
\draw ( 90:0.7) node {$x$};
\draw (-30:0.7) node {$v$};
\draw (210:0.7) node {$v$};
\end{scope}
\draw (2.75,-0.7) -- (2.8,-0.7) -- (2.8,1) -- (2.75,1);
\draw (3.1,0) node{$-$};
\draw (3.55,-0.7) -- (3.5,-0.7) -- (3.5,1) -- (3.55,1);
\begin{scope}[xshift=4.4cm]
\draw[red] (-30:0.2) -- (90:0.2) -- (210:0.2) -- cycle;
\draw (0,0) node {$g$};
\draw (-30:0.2) -- (-30:0.6);
\draw ( 90:0.2) -- ( 90:0.6);
\draw (210:0.2) -- (210:0.6);
\draw[blue] (-30:0.7) circle(0.1);
\draw[blue] ( 90:0.7) circle(0.1);
\draw[blue] (210:0.7) circle(0.1);
\draw ( 90:0.7) node {$x$};
\draw (210:0.7) node {$u$};
\draw (-30:0.7) node {$v$};
\end{scope}
\begin{scope}[xshift=6.2cm]
\draw[red] (-30:0.2) -- (90:0.2) -- (210:0.2) -- cycle;
\draw (0,0) node {$g$};
\draw (-30:0.2) -- (-30:0.6);
\draw ( 90:0.2) -- ( 90:0.6);
\draw (210:0.2) -- (210:0.6);
\draw[blue] (-30:0.7) circle(0.1);
\draw[blue] ( 90:0.7) circle(0.1);
\draw[blue] (210:0.7) circle(0.1);
\draw ( 90:0.7) node {$x$};
\draw (210:0.7) node {$u$};
\draw (-30:0.7) node {$v$};
\end{scope}
\draw (7.15,-0.7) -- (7.2,-0.7) -- (7.2,1) -- (7.15,1);
\end{tikzpicture}
\end{center}
We now consider various elements of the space $(\R^4)^{\ot 4}$. We
give this the obvious inner product so that elements of the form
$e_i\ot e_j\ot e_k\ot e_l$ give an orthonormal basis. Any
permutation $\sg\in\Sg_4$ gives an automorphism $\al_\sg$ of
$(\R^4)^{\ot 4}$ by permuting the tensor factors; for example, we have
\[ \al_{(2;3)}.(u_1\ot u_2\ot u_3\ot u_4) =
u_1\ot u_3\ot u_2\ot u_4.
\]
We put
\begin{align*}
A &= \sum_ix_ig_{ijk}e_j\ot e_k\in(\R^4)^{\ot 2} \\
B &= u\ot v - v\ot u\in(\R^4)^{\ot 2} \\
C &= u\ot u\ot v\ot v-u\ot v\ot u\ot v \in (\R^4)^{\ot 4}.
\end{align*}
It is now not hard to see that $P=36\ip{A\ot A,C}$. We claim that
also
\[ P = 18\ip{A\ot A,\al_{(2\;3)}(B\ot B)}
= 18\ip{\al_{(2\;3)}(A\ot A),B\ot B}.
\]
Indeed, we have
\begin{align*}
\al_{(2\;3)}(B\ot B) &=
\al_{(2\;3)}(u\ot v\ot u\ot v -
u\ot v\ot v\ot u -
v\ot u\ot u\ot v +
v\ot u\ot v\ot u) \\
&= u\ot u\ot v\ot v -
u\ot v\ot v\ot u -
v\ot u\ot u\ot v +
v\ot v\ot u\ot u.
\end{align*}
Now $A\ot A$ is invariant under the permutations $(1\;2)$, $(3\;4)$
and $(1\;3)(2\;4)$. Using this we see that
\begin{align*}
\ip{A\ot A,u\ot v\ot v\ot u} &= \ip{A\ot A,u\ot v\ot u\ot v} \\
\ip{A\ot A,v\ot u\ot u\ot v} &= \ip{A\ot A,u\ot v\ot u\ot v} \\
\ip{A\ot A,v\ot v\ot u\ot u} &= \ip{A\ot A,u\ot u\ot v\ot v}.
\end{align*}
The claim follows easily from this. It can be rewritten as
\[ P = 18\sum_{i,j,k,l} A_{ij}A_{kl}B_{ik}B_{jl}. \]
We now define a matrix $Q$ with $Q_{jk}=\sum_iA_{ij}B_{ik}$. After
noting that $A_{kl}=A_{lk}$ and $B_{jl}=-B_{lj}$, the
above expression can be rewritten again as
\[ P=-18\sum_{j,k}Q_{jk}Q_{kj} = -18\text{trace}(Q^2). \]
Some of the above can be represented graphically as follows:
\begin{center}
\begin{tikzpicture}[scale=2]
\begin{scope}
\draw[green] (-1.1, 0.8) circle(0.13);
\draw[green] ( 1.1, 0.8) circle(0.13);
\draw[magenta] (-1.25,-0.65) -- (-1.25,-0.95) -- (-0.95,-0.8) -- cycle;
\draw[magenta] ( 0.95,-0.65) -- ( 0.95,-0.95) -- ( 1.25,-0.8) -- cycle;
\draw[rounded corners] (-1.23,0.8) -- (-1.5,0.8) -- (-1.5,-0.8) -- (-1.25,-0.8);
\draw[rounded corners] ( 1.23,0.8) -- ( 1.5,0.8) -- ( 1.5,-0.8) -- ( 1.25,-0.8);
\draw[rounded corners] (-0.97, 0.8) -- (-0.8,0.8) -- (0.8,-0.8) -- (0.95,-0.8);
\draw[rounded corners] (-0.95,-0.8) -- (-0.8,-0.8) -- (-0.1,-0.1);
\draw[rounded corners] (0.1,0.1) -- (0.8,0.8) -- (0.97,0.8);
\draw(-1.10, 0.8) node {$A$};
\draw( 1.10, 0.8) node {$A$};
\draw(-1.16,-0.8) node {$B$};
\draw( 1.04,-0.8) node {$B$};
\draw(0,-1.1) node{$P/18$};
\end{scope}
\begin{scope}[xshift=5cm]
\draw (-0.85,0) -- (-0.53,0);
\draw[green] (-0.4, 0.0) circle(0.13);
\draw (-0.27,0) -- (0.25,0);
\draw[magenta] ( 0.25,0.15) -- ( 0.25,-0.15) -- (0.55,0.0) -- cycle;
\draw (0.55,0) -- (0.85,0);
\draw(-0.40, 0.0) node {$A$};
\draw( 0.34, 0.0) node {$B$};
\draw(0,-1.1) node{$Q$};
\end{scope}
\end{tikzpicture}
\end{center}
Note that our choice of symbols reflects the fact that $A$ is
symmetric and $B$ is not.
Now put
\[ D = x\wedge n(x) = x\ot n(x) - n(x)\ot x, \]
so
\[ D_{ij} = 3\sum_{k,l} (g_{jkl} x_ix_kx_l - g_{ikl} x_jx_kx_l).
\]
Lemma~\ref{lem-hodge} tells us that
\[ B_{ik} = \sum_{l,m}\ep_{iklm}D_{lm}/(2r). \]
Now
\[ Q_{jk} = \sum_iA_{ij}B_{ik}
= \frac{1}{2r}\sum_{i,l,m} \ep_{iklm}A_{ij}D_{lm}
= \frac{3}{2r}\sum_{h,i,l,m,n,p}
\ep_{iklm} g_{hij}x_h(g_{lnp}x_mx_nx_p-g_{mnp}x_lx_nx_p).
\]
Note that the two terms in brackets are essentially the same except
that $l$ and $m$ are exchanged, but $\ep_{iklm}$ is also antisymmetric
in $l$ and $m$. We can thus drop one of the terms and introduce a
factor of $2$ giving
\begin{align*}
Q_{jk}
&= \frac{3}{r}\sum_{h,i,l,m,n,p}
\ep_{iklm} g_{hij}g_{lnp}x_hx_mx_nx_p \\
&= \frac{1}{6r}\sum_{i,l,m}\ep_{iklm}n(x)_l\;m(x)_{ij}
= \pm p(x)_{jk}/(6r).
\end{align*}
Equivalently, we have the following Penrose diagram for $rQ/3$:
\begin{center}
\begin{tikzpicture}[scale=2]
\draw[rounded corners] (-0.9,0) -- (0,0) -- (0,0.4);
\draw[red] (0,0.4) -- (0,0.7) -- (-0.27,0.55) -- cycle;
\draw (-0.09,0.55) node {$g$};
\draw (-0.27,0.55) -- (-0.57,0.55);
\draw[blue] (-0.67,0.55) circle(0.1);
\draw (-0.67,0.55) node {$x$};
\draw(0,0.7) -- (0,1);
\draw(-0.2,1) -- (2.0,1) -- (2.0,1.5) -- (-0.2,1.5) -- cycle;
\draw (0.9,1.25) node{$\epsilon$};
\draw[rounded corners] (2.6,0) -- (0.5,0) -- (0.5,1);
\draw[blue] (1.0,0.3) circle(0.1);
\draw (1.0,0.3) node {$x$};
\draw[blue] (1.3,0.3) circle(0.1);
\draw (1.3,0.3) node {$x$};
\draw[blue] (1.8,0.3) circle(0.1);
\draw (1.8,0.3) node {$x$};
\draw[red] (1.0,0.6) -- (1.3,0.6) -- (1.15,0.87) -- cycle;
\draw (1.15,0.7) node {$g$};
\draw (1.0,0.4) -- (1.0,0.6);
\draw (1.3,0.4) -- (1.3,0.6);
\draw (1.15,0.87) -- (1.15,1);
\draw (1.8,0.4) -- (1.8,1);
\end{tikzpicture}
\end{center}
We now have
\[ P=-18\text{trace}(Q^2)=-\text{trace}(p(x)^2)/(2r^2), \]
and so
\[ K = 1+P/r^2 = 1 - \text{trace}(p(x)^2)/(2r^4). \]
\begin{checks}
embedded/curvature_check.mpl: check_EX_curvature()
\end{checks}
\end{proof}
\begin{remark}\label{rem-curvature-z}
The full formula for $K$ in terms of the variables $x_i$ is too large
to be given here. However, $K$ is invariant under the group action,
and so can be expressed in terms of the functions $z_1$ and $z_2$
from Section~\ref{sec-E-functions}. Even that expression is somewhat
unwieldy for general $a$, but when $a=1/\rt$ one can check that
the formula is as follows:
\[ K = 1+ 8 \frac{2z_2-1}{(2-z_1)^2(1+z_2)^2}. \]
\end{remark}
It will also be useful for us to have an expression for the
Laplace-Beltrami operator $\Dl$ on $EX^*$ in terms of the ambient
coordinates $x_i$.
\begin{definition}
As before, we write $r=\|n(x)\|$, so
\[ r^2 = n(x).n(x) = 9\sum_{ijklm}g_{ijm}g_{klm}x_ix_jx_kx_l. \]
We also define
\begin{align*}
r' &= \text{trace}(m(x)) = 6\sum_{i,j} g_{ijj}x_i \\
r'' &= n(x)^T m(x) n(x)
= 54 \sum_{ijklmnp} g_{ijk}g_{klm}g_{mnp}x_ix_jx_lx_nx_p
\end{align*}
\end{definition}
\begin{definition}\label{defn-Delta-prime}
Let $U'$ be an open subset of $\R^4\sm\{0\}$. We define a
differential operator $\Dl'\:C^\infty(U')\to C^\infty(U')$ by
\[ \Dl'(p) =
\sum_i \frac{\partial^2p}{\partial x_i^2}
- \sum_{i,j} x_ix_j \frac{\partial^2p}{\partial x_i\partial x_j}
- \frac{1}{r^2} \sum_{i,j} n(x)_i n(x)_j
\frac{\partial^2p}{\partial x_i\partial x_j}
- 2\sum_i x_i\frac{\partial p}{\partial x_i}
+ \left(\frac{r''}{r^4}-\frac{r'}{r^2}\right)
\sum_i n(x)_i\frac{\partial p}{\partial x_i}
\]
\end{definition}
\begin{proposition}\label{prop-Delta-prime}
If we put $U=U'\cap EX^*$, then the following diagram commutes:
\[ \xymatrix{
C^\infty(U') \ar[r]^{\Dl'} \ar[d]_{\text{res}} &
C^\infty(U') \ar[d]^{\text{res}} \\
C^\infty(U) \ar[r]_\Dl &
C^\infty(U).
} \]
\end{proposition}
\begin{proof}
Consider a smooth function $p\in C^\infty(U')$ and a point $x\in EX^*$.
Choose a chart $\phi\:\R^2\to EX^*$ with $\phi(0,0)=x$. We will write
$a$ and $b$ for coordinates on $\R^2$. The chart gives a Riemannian
metric on $\R^2$, corresponding to the matrix $M=\bbm E&F\\ F&G\ebm$,
where
\begin{align*}
E &= \frac{\partial \phi}{\partial a} . \frac{\partial \phi}{\partial a} &
F &= \frac{\partial \phi}{\partial a} . \frac{\partial \phi}{\partial b} &
G &= \frac{\partial \phi}{\partial b} . \frac{\partial \phi}{\partial b}.
\end{align*}
We will use the standard formula
\[ (\Delta p)\circ\phi =
\det(M)^{-1/2}
\text{div}\left(\det(M)^{1/2} M^{-1}\text{grad}(p\circ\phi)\right).
\]
We will only be using this formula at the point $(a,b)=(0,0)$, so it
will be harmless to replace $M$ by an approximation involving only
terms that are constant or linear in $a$ and $b$. Similarly, $\phi$
need not be an exact chart, so long as it is quadratically close to
$EX^*$. We can thus follow Proposition~\ref{prop-quadratic-chart} and
define $\phi$ as below:
\begin{align*}
\psi_0(t) &= (1-\|t\|^2/2)x - (n(t).x)n(x)/r^2 &
\phi_0(t) &= t + \psi_0(t) \\
\psi(a,b) &= \psi_0(au+bv) &
\phi(a,b) &= \phi_0(au+bv).
\end{align*}
By routine calculation we have
\begin{align*}
\frac{\partial\phi}{\partial a} &=
u - ax - (Aa+Cb) n(x) \\
\frac{\partial\phi}{\partial b} &=
v - bx - (Bb+Ca) n(x)
\end{align*}
where
\begin{align*}
A &= \frac{6}{r^2} \sum_{ijk} g_{ijk}x_iu_ju_k \\
B &= \frac{6}{r^2} \sum_{ijk} g_{ijk}x_iv_jv_k \\
C &= \frac{6}{r^2} \sum_{ijk} g_{ijk}x_iu_jv_k.
\end{align*}
Because $x$, $n(x)$, $u$ and $v$ are orthogonal, it follows that we
have $E=G=1$ and $F=0$ up to quadratic corrections. We may thus take
$M$ to be the identity matrix, and deduce that $(\Delta\ p)(x)$ is
the value at $(0,0)$ of
$\left(\frac{\partial^2}{\partial a^2}+
\frac{\partial^2}{\partial b^2}\right)(p\circ\phi)$. After
using the chain rule twice, we see that this is the same as $P+Q$,
where
\begin{align*}
P &= \sum_{i,j} \frac{\partial^2p}{\partial x_i\,\partial x_j}
\left(\frac{\partial\phi_i}{\partial a}
\frac{\partial\phi_j}{\partial a}+
\frac{\partial\phi_i}{\partial b}
\frac{\partial\phi_j}{\partial b}\right) \\
Q &= \sum_{i} \frac{\partial p}{\partial x_i}
\left(\frac{\partial^2\phi_i}{\partial a^2}+
\frac{\partial^2\phi_i}{\partial b^2}\right).
\end{align*}
Previously we recorded formulae for
$\partial\phi/\partial a$ and $\partial\phi/\partial b$; when $a=b=0$
they just give $u$ and $v$. Thus, we have
\[ P = \sum_{i,j} \frac{\partial^2\phi}{\partial x_i\,\partial x_j}
\left(u_iu_j+v_iv_j\right).
\]
Now, the numbers $u_iu_j+v_iv_j$ are the matrix entries for the
orthogonal projection onto the tangent space $T$, which is the
identity minus the projection onto the normal space
$N=T^\perp=\text{span}(x,n(x))$. We thus have
\[ u_iu_j+v_iv_j=\dl_{ij} - x_ix_j - n(x)_in(x)_j/r^2. \]
Using this we obtain
\[ P = \sum_i \frac{\partial^2p}{\partial x_i^2}
- \sum_{i,j} x_ix_j \frac{\partial^2p}{\partial x_i\partial x_j}
- \frac{1}{r^2} \sum_{i,j} n(x)_i n(x)_j
\frac{\partial^2p}{\partial x_i\partial x_j}.
\]
This accounts for the first three terms in $\Dl'(p)$.
We now turn to $Q$. We have
\begin{align*}
\frac{\partial^2\phi}{\partial a^2} &= - x - A\,n(x) \\
\frac{\partial^2\phi}{\partial b^2} &= - x - B\,n(x)
\end{align*}
and it follows that
\[ Q = (-2x-(A+B)n(x)) . \nabla(p). \]
Now
\[ A+B = \frac{6}{r^2} \sum_{i,j,k} g_{ijk}x_i(u_ju_k+v_jv_k), \]
and we can again use the relation
$u_ju_k+v_jv_k=\dl_{jk}-x_jx_k-n(x)_jn(x)_k/r^2$ to eliminate $u$ and
$v$, giving
\[ A+B = \frac{6}{r^2} \left(
\sum_{i,j} g_{ijj}x_i -
\sum_{i,j,k} g_{ijk}x_ix_jx_k -
\frac{1}{r^2} \sum_{i,j,k} g_{ijk}x_in(x)_jn(x)_k
\right).
\]
Here $\sum_{i,j,k} g_{ijk}x_ix_jx_k$ is $g(x)$, which is zero
because $x\in EX^*$. We also have
\begin{align*}
\sum_{i,j} g_{ijj}x_i &= r'/6 \\
\sum_{i,j,k} g_{ijk}x_in(x)_jn(x)_k &= r''/6.
\end{align*}
Putting this together gives
\[ A+B = \frac{r'}{r^2} - \frac{r''}{r^4}, \]
and this identifies $Q$ with the remaining terms in $\Dl'(p)$.
\begin{checks}
embedded/geometry_check.mpl: check_laplacian_a()
embedded/geometry_check.mpl: check_laplacian_b()
embedded/geometry_check.mpl: check_laplacian_z()
\end{checks}
\end{proof}
\begin{remark}
The uniformization theorem for Riemann surfaces says that $EX^*$ is
conformally equivalent to the quotient of the open unit disc by a
discrete group of automorphisms that preserve the hyperbolic metric.
Using this, we deduce that the original metric on $EX^*$ can be
conformally rescaled so that the new metric has constant curvature
$-1$. Let $g$ denote the original metric, with curvature $K$, and
consider a rescaled metric $g^*=e^{2f}g$. It is then known that the
corresponding curvature is $K^*=(K-\Dl(f))/e^{2f}$. We therefore
want to find a $G$-invariant function $f$ such that
$K=\Dl(f)-e^{2f}$. There is a lot of freedom to do this locally, but
not globally. Thus, the most natural approach is to try to minimize
$\int_{EX^*}(1+(K-\Dl(f))/e^{2f})^2$ as $f$ ranges over some
finite-dimensional space of invariant functions. To carry this
forward, we need some theory of integration on $EX^*$, which will be
treated in the next section.
\end{remark}
\section{The surface \texorpdfstring{$EX^*$}{EX*}}
\label{sec-roothalf}
We now focus on the surface $EX^*=EX(1/\rt)$. We start by recording
explicitly a number of formulae that are obtained by substituting
$a=1/\rt$ in the results of Section~\ref{sec-E}. Maple
notation for all these things is obtained by appending the character
\mcode+0+ to the corresponding notation in Section~\ref{sec-E}:
the points $v_i$ are \mcode+v_E0[i]+, the curves $c_j(t)$ are
\mcode+c_E0[j](t)+ and so on.
We have
\[ EX^* = \{x\in \R^4\st \rho(x)=1,\;g(x)=0\}
= \{x\in \R^4\st \rho(x)=1,\;g_0(x)=0\},
\]
where
\begin{align*}
g(x) &= x_3^2x_4-2x_4^3-(2(x_1^2+x_2^2))x_4+\rt(x_1^2-x_2^2)x_3 \\
g_0(x) &= (3x_3^2-2)x_4+\rt(x_1^2-x_2^2)x_3.
\end{align*}
The gradient of $g(x)$ is
\[ n(x) = \left(
2x_1(\rt x_3-2x_4),\;
-2x_2(\rt x_3+2x_4),\;
2x_3x_4+\rt(x_1^2-x_2^2),\;
-2x_1^2-2x_2^2+x_3^2-6x_4^2
\right).
\]
We put $y_1=x_3$ and $y_2=(x_2^2-x_1^2)/\rt-\tfrac{3}{2}x_3x_4$ and
$z_i=y_i^2$. We find that
\begin{align*}
x_1^2 &= u_1= \half(1-\rt y_2)(1-y_1^2(1-y_2/\rt)) \\
x_2^2 &= u_2= \half(1+\rt y_2)(1-y_1^2(1+y_2/\rt)) \\
4x_1^2x_2^2 &= u_3 = (1-2z_2)((1-z_1)^2-z_1^2z_2/2) \\
x_1^2+x_2^2 &= u_4 = 1-z_1-z_1z_2.
\end{align*}
The ring of $G$-invariant polynomial functions on $EX^*$ is $\R[z_1,z_2]$.
In particular, we have
\[ \|n(x)\|^2 = 4 (1-z_1/2)^2(1+z_2), \]
and the curvature is
\[ K = 1 - 2 \frac{1-2z_2}{(1-z_1/2)^2(1+z_2)^2}. \]
We can reduce polynomials to normal form using the functions
\mcode+NF_x0+, \mcode+NF_y0+ and \mcode+NF_z0+. There are also
functions \mcode+FNF_y0+ and \mcode+FNF_z0+ which deal intelligently
with rational functions as well as polynomials.
The isotropy points are
\begin{align*}
v_{ 0} &= (\pp 0,\pp 0,\pp 1,\pp 0) &
v_{ 6} &= (\pp 1,\pp 1,\pp 0,\pp 0)/\rt \\
v_{ 1} &= (\pp 0,\pp 0, -1,\pp 0) &
v_{ 7} &= ( -1,\pp 1,\pp 0,\pp 0)/\rt \\
v_{ 2} &= (\pp 1,\pp 0,\pp 0,\pp 0) &
v_{ 8} &= ( -1, -1,\pp 0,\pp 0)/\rt \\
v_{ 3} &= (\pp 0,\pp 1,\pp 0,\pp 0) &
v_{ 9} &= (\pp 1, -1,\pp 0,\pp 0)/\rt \\
v_{ 4} &= ( -1,\pp 0,\pp 0,\pp 0) &
v_{10} &= (0,0,\pp\sqrt{2/3},\pp\sqrt{1/3}) \\
v_{ 5} &= (\pp 0,-1,\pp 0,\pp 0) &
v_{11} &= (0,0,\pp\sqrt{2/3}, -\sqrt{1/3}) \\
&&
v_{12} &= (0,0, -\sqrt{2/3}, -\sqrt{1/3}) \\
&&
v_{13} &= (0,0, -\sqrt{2/3},\pp\sqrt{1/3}).
\end{align*}
The curve system is as follows:
\begin{align*}
c_0(t) &= \left(\cos(t),\sin(t),0,0\right) \\
c_1(t) &= \left(\sin(t)/\rt,\sin(t)/\rt,\cos(t),0\right) \\
c_2(t) &= \lm(c_1(t)) \\
c_3(t) &= \left(0,\sin(t),\sqrt{2/3}\cos(t),-\sqrt{1/3}\cos(t)\right) \\
c_4(t) &= \lm(c_3(t)) \\
c_5(t) &= \left(-\sin(t),0,2\rt,\cos(t)-1\right)/\sqrt{10-2\cos(t)} \\
c_6(t) &= \lm(c_5(t)) \\
c_7(t) &= \mu(c_5(t)) \\
c_8(t) &= \lm\mu(c_5(t)).
\end{align*}
In this case, the curves $c_5,\dotsc,c_8$ have some additional
properties. If we put
\[ h(x) = x_3x_4/\rt+x_1^2+x_4^2, \]
we find that $h=0$ on $C_5\cup C_7$. Note here that $h(x)$ is a
homogeneous quadratic. It can be diagonalised as
$h(x)=\sum_{i=1}^4m_i\ip{u_i,x}^2$, where
\begin{align*}
m_1 &= 1 &
u_1 &= (1,0,0,0) \\
m_2 &= 0 &
u_2 &= (0,1,0,0) \\
m_3 &= \frac{2-\sqrt{6}}{4} \simeq -0.11 &
u_3 &= \left(0,0,\pp\sqrt{1/2+1/\sqrt{6}},\quad -\sqrt{1/2-1/\sqrt{6}}\right) \\
m_4 &= \frac{2-\sqrt{6}}{4} \simeq -0.11 &
u_4 &= \left(0,0,\pp\sqrt{1/2-1/\sqrt{6}},\quad\pp\sqrt{1/2+1/\sqrt{6}}\right).
\end{align*}
The vectors $u_i$ here form an oriented orthonormal basis for $\R^4$.
\begin{checks}
embedded/roothalf/E_roothalf_check.mpl: check_oval()
\end{checks}
One can also check that there is an alternative parametrisation of
$C_5$ as follows:
\[
c_5^{\text{alt}}(t) =
\left(\frac{\sin(t)}{\bt},\;0,\;
\frac{1+\bt^2}{12}\left(\sqrt{1-\sin(t)^2/\bt^4} + \cos(t)/\bt^2\right),\;
\frac{\cos(t)-\sqrt{1-\sin(t)^2/\bt^4}}{2\sqrt{3}}\right),
\]
where $\bt=\sqrt{2}+\sqrt{3}\simeq 3.15$. Note here that
$\bt^4\simeq 98$, so we have a good approximation
\[
c_5^{\text{alt}}(t) \simeq
\left(\frac{\sin(t)}{\bt},\;0,\;
\frac{1+\bt^2}{12}\left(1 + \cos(t)/\bt^2\right),\;
\frac{\cos(t)-1}{2\sqrt{3}}\right),
\]
showing that $C_5$ is close to an ellipse. If we put
\[ c_6^{\text{alt}}(t) = \lm(c_5^{\text{alt}}(t)) \qquad
c_7^{\text{alt}}(t) = \mu(c_5^{\text{alt}}(t)) \qquad
c_8^{\text{alt}}(t) = \lm\mu(c_5^{\text{alt}}(t)),
\]
and $c_k^{\text{alt}}=c_k$ for $0\leq k\leq 4$, then one can check
that this gives an alternative curve system.
\begin{checks}
embedded/roothalf/E_roothalf_check.mpl: check_c_alt()
\end{checks}
\begin{proposition}\label{prop-roothalf-fundamental}
In the case $a=1/\rt$, we have
\begin{align*}
F_4^* &= \{y\in\R^2\st |y_2|\leq 1/\rt,\;
|y_1|\leq (1+|y_2|/\rt)^{-1/2}\} \\
F_{16}^* &= \{z\in \R^2\st 0\leq z_2\leq 1/2,\;
0\leq z_1\leq (1+\sqrt{z_2/2})^{-1}\}.
\end{align*}
\end{proposition}
\begin{proof}
From the definitions it is clear that
\[ F_4^*=\{y\in\R^2 \st (y_1^2,y_2^2)\in F_{16}^*\}; \]
using this, we can reduce the first claim to the second one. Put
\[ F'_{16} = \{z\in \R^2\st 0\leq z_2\leq 1/2,\;
0\leq z_1\leq (1+\sqrt{z_2/2})^{-1}\},
\]
so we need to show that $F^*_{16}=F'_{16}$.
Recall that $F_{16}^*=\{z\in\R^2\st z_1,z_2,u_3,u_4\geq 0\}$, and
that in the present case $a=1/\rt$ we have
\begin{align*}
u_3 &= (1-2z_2)((1-z_1)^2-z_1^2z_2/2) \\
u_4 &= 1-z_1-z_1z_2.
\end{align*}
We assume implicitly throughout that $z_1,z_2\geq 0$. Put
$w_1=1-2z_2$ and $w_2=(1-z_1)^2-z_1^2z_2/2$, so $u_3=w_1w_2$. We
will leave to the reader all the cases where any of the quantities
$z_1,z_2,w_1,w_2,u_3$ or $u_4$ are zero.
First suppose that $z\in F^*_{16}$, so $u_3,u_4>0$. As
$u_3=w_1w_2$ we see that $w_1$ and $w_2$ have the same sign.
One can check that
\[ \half z_1^2z_2w_1 + w_2 = u_4(u_4+2z_1z_2) > 0, \]
so $w_1$ and $w_2$ must both be positive. As $w_1>0$ we have
$0<z_2<1/2$. As $u_4>0$ we have $1-z_1>z_1z_2>0$, and as $w_2>0$ we
have $(1-z_1)^2>z_1^2z_2/2$; it follows that $1-z_1>z_1\sqrt{z_2/2}$,
and thus that $0<z_1<1/(1+\sqrt{z_2/2})$ as claimed.
Conversely, suppose that $0<z_1<1/(1+\sqrt{z_2/2})$ (so in particular
$z_1<1$) and $0<z_2<1/2$. We can reverse the above arguments to see
that $w_1,w_2>0$ and so $u_3=w_1w_2>0$. Next, one can check that
\[ ((1-\half z_2)(1+z_2)(1-z_1)+\half z_2(4+z_2))u_4 =
\half z_2 w_1 + (1+z_2)^2 w_2.
\]
The coefficient of $u_4$ is strictly positive, as are all the terms
on the right hand side, so we deduce that $u_4>0$, which means that
$z\in F^*_{16}$.
\end{proof}
\subsection{Linear projections}
\label{sec-disc}
In this section we study the images of our points and curves under
three different orthogonal projections $\R^4\to\R^2$. These do not
have any great theoretical significance, but they provide some insight
into the geometry. Many additional details are given in the Maple
code, especially the files \fname+embedded/disc_proj.mpl+ and
\fname+embedded/roothalf/zeta.mpl+ and
\fname+embedded/roothalf/crease.mpl+.
The projections that we consider are
\begin{align*}
\pi(x) &= (x_1,x_2) \\
\dl(x) &= ((x_1-x_2)/\rt,x_3) \\
\zt(x) &= ((x_3-x_4)/\rt,x_2).
\end{align*}
The picture for $\pi$ is as follows:
\begin{center}
\begin{tikzpicture}[scale=4]
\draw[cyan] plot[smooth] coordinates{ (1.000,0.000) (0.988,0.156) (0.951,0.309) (0.891,0.454) (0.809,0.588) (0.707,0.707) (0.588,0.809) (0.454,0.891) (0.309,0.951) (0.156,0.988) (-0.000,1.000) (-0.156,0.988) (-0.309,0.951) (-0.454,0.891) (-0.588,0.809) (-0.707,0.707) (-0.809,0.588) (-0.891,0.454) (-0.951,0.309) (-0.988,0.156) (-1.000,-0.000) (-0.988,-0.156) (-0.951,-0.309) (-0.891,-0.454) (-0.809,-0.588) (-0.707,-0.707) (-0.588,-0.809) (-0.454,-0.891) (-0.309,-0.951) (-0.156,-0.988) (0.000,-1.000) (0.156,-0.988) (0.309,-0.951) (0.454,-0.891) (0.588,-0.809) (0.707,-0.707) (0.809,-0.588) (0.891,-0.454) (0.951,-0.309) (0.988,-0.156) (1.000,0.000) };
\draw[green] plot[smooth] coordinates{ (0.000,0.000) (0.111,0.111) (0.219,0.219) (0.321,0.321) (0.416,0.416) (0.500,0.500) (0.572,0.572) (0.630,0.630) (0.672,0.672) (0.698,0.698) (0.707,0.707) (0.698,0.698) (0.672,0.672) (0.630,0.630) (0.572,0.572) (0.500,0.500) (0.416,0.416) (0.321,0.321) (0.219,0.219) (0.111,0.111) (-0.000,-0.000) (-0.111,-0.111) (-0.219,-0.219) (-0.321,-0.321) (-0.416,-0.416) (-0.500,-0.500) (-0.572,-0.572) (-0.630,-0.630) (-0.672,-0.672) (-0.698,-0.698) (-0.707,-0.707) (-0.698,-0.698) (-0.672,-0.672) (-0.630,-0.630) (-0.572,-0.572) (-0.500,-0.500) (-0.416,-0.416) (-0.321,-0.321) (-0.219,-0.219) (-0.111,-0.111) (0.000,0.000) };
\draw[green] plot[smooth] coordinates{ (0.000,0.000) (-0.111,0.111) (-0.219,0.219) (-0.321,0.321) (-0.416,0.416) (-0.500,0.500) (-0.572,0.572) (-0.630,0.630) (-0.672,0.672) (-0.698,0.698) (-0.707,0.707) (-0.698,0.698) (-0.672,0.672) (-0.630,0.630) (-0.572,0.572) (-0.500,0.500) (-0.416,0.416) (-0.321,0.321) (-0.219,0.219) (-0.111,0.111) (0.000,-0.000) (0.111,-0.111) (0.219,-0.219) (0.321,-0.321) (0.416,-0.416) (0.500,-0.500) (0.572,-0.572) (0.630,-0.630) (0.672,-0.672) (0.698,-0.698) (0.707,-0.707) (0.698,-0.698) (0.672,-0.672) (0.630,-0.630) (0.572,-0.572) (0.500,-0.500) (0.416,-0.416) (0.321,-0.321) (0.219,-0.219) (0.111,-0.111) (-0.000,0.000) };
\draw[magenta] plot[smooth] coordinates{ (0.000,0.000) (0.000,0.156) (0.000,0.309) (0.000,0.454) (0.000,0.588) (0.000,0.707) (0.000,0.809) (0.000,0.891) (0.000,0.951) (0.000,0.988) (0.000,1.000) (0.000,0.988) (0.000,0.951) (0.000,0.891) (0.000,0.809) (0.000,0.707) (0.000,0.588) (0.000,0.454) (0.000,0.309) (0.000,0.156) (0.000,-0.000) (0.000,-0.156) (0.000,-0.309) (0.000,-0.454) (0.000,-0.588) (0.000,-0.707) (0.000,-0.809) (0.000,-0.891) (0.000,-0.951) (0.000,-0.988) (0.000,-1.000) (0.000,-0.988) (0.000,-0.951) (0.000,-0.891) (0.000,-0.809) (0.000,-0.707) (0.000,-0.588) (0.000,-0.454) (0.000,-0.309) (0.000,-0.156) (0.000,0.000) };
\draw[magenta] plot[smooth] coordinates{ (0.000,0.000) (-0.156,0.000) (-0.309,0.000) (-0.454,0.000) (-0.588,0.000) (-0.707,0.000) (-0.809,0.000) (-0.891,0.000) (-0.951,0.000) (-0.988,0.000) (-1.000,0.000) (-0.988,0.000) (-0.951,0.000) (-0.891,0.000) (-0.809,0.000) (-0.707,0.000) (-0.588,0.000) (-0.454,0.000) (-0.309,0.000) (-0.156,0.000) (0.000,0.000) (0.156,0.000) (0.309,0.000) (0.454,0.000) (0.588,0.000) (0.707,0.000) (0.809,0.000) (0.891,0.000) (0.951,0.000) (0.988,0.000) (1.000,0.000) (0.988,0.000) (0.951,0.000) (0.891,0.000) (0.809,0.000) (0.707,0.000) (0.588,0.000) (0.454,0.000) (0.309,0.000) (0.156,0.000) (-0.000,0.000) };
\draw[blue] plot[smooth] coordinates{ (0.000,0.000) (0.055,0.000) (0.109,0.000) (0.158,0.000) (0.203,0.000) (0.241,0.000) (0.272,0.000) (0.295,0.000) (0.310,0.000) (0.317,0.000) (0.316,0.000) (0.308,0.000) (0.292,0.000) (0.270,0.000) (0.242,0.000) (0.209,0.000) (0.172,0.000) (0.132,0.000) (0.090,0.000) (0.045,0.000) (-0.000,0.000) (-0.045,0.000) (-0.090,0.000) (-0.132,0.000) (-0.172,0.000) (-0.209,0.000) (-0.242,0.000) (-0.270,0.000) (-0.292,0.000) (-0.308,0.000) (-0.316,0.000) (-0.317,0.000) (-0.310,0.000) (-0.295,0.000) (-0.272,0.000) (-0.241,0.000) (-0.203,0.000) (-0.158,0.000) (-0.109,0.000) (-0.055,0.000) (0.000,0.000) };
\draw[blue] plot[smooth] coordinates{ (0.000,0.000) (0.000,0.055) (0.000,0.109) (0.000,0.158) (0.000,0.203) (0.000,0.241) (0.000,0.272) (0.000,0.295) (0.000,0.310) (0.000,0.317) (0.000,0.316) (0.000,0.308) (0.000,0.292) (0.000,0.270) (0.000,0.242) (0.000,0.209) (0.000,0.172) (0.000,0.132) (0.000,0.090) (0.000,0.045) (0.000,-0.000) (0.000,-0.045) (0.000,-0.090) (0.000,-0.132) (0.000,-0.172) (0.000,-0.209) (0.000,-0.242) (0.000,-0.270) (0.000,-0.292) (0.000,-0.308) (0.000,-0.316) (0.000,-0.317) (0.000,-0.310) (0.000,-0.295) (0.000,-0.272) (0.000,-0.241) (0.000,-0.203) (0.000,-0.158) (0.000,-0.109) (0.000,-0.055) (0.000,0.000) };
\draw[blue] plot[smooth] coordinates{ (0.000,0.000) (0.055,0.000) (0.109,0.000) (0.158,0.000) (0.203,0.000) (0.241,0.000) (0.272,0.000) (0.295,0.000) (0.310,0.000) (0.317,0.000) (0.316,0.000) (0.308,0.000) (0.292,0.000) (0.270,0.000) (0.242,0.000) (0.209,0.000) (0.172,0.000) (0.132,0.000) (0.090,0.000) (0.045,0.000) (-0.000,0.000) (-0.045,0.000) (-0.090,0.000) (-0.132,0.000) (-0.172,0.000) (-0.209,0.000) (-0.242,0.000) (-0.270,0.000) (-0.292,0.000) (-0.308,0.000) (-0.316,0.000) (-0.317,0.000) (-0.310,0.000) (-0.295,0.000) (-0.272,0.000) (-0.241,0.000) (-0.203,0.000) (-0.158,0.000) (-0.109,0.000) (-0.055,0.000) (0.000,0.000) };
\draw[blue] plot[smooth] coordinates{ (0.000,0.000) (0.000,0.055) (0.000,0.109) (0.000,0.158) (0.000,0.203) (0.000,0.241) (0.000,0.272) (0.000,0.295) (0.000,0.310) (0.000,0.317) (0.000,0.316) (0.000,0.308) (0.000,0.292) (0.000,0.270) (0.000,0.242) (0.000,0.209) (0.000,0.172) (0.000,0.132) (0.000,0.090) (0.000,0.045) (0.000,-0.000) (0.000,-0.045) (0.000,-0.090) (0.000,-0.132) (0.000,-0.172) (0.000,-0.209) (0.000,-0.242) (0.000,-0.270) (0.000,-0.292) (0.000,-0.308) (0.000,-0.316) (0.000,-0.317) (0.000,-0.310) (0.000,-0.295) (0.000,-0.272) (0.000,-0.241) (0.000,-0.203) (0.000,-0.158) (0.000,-0.109) (0.000,-0.055) (0.000,0.000) };
\fill (0.000,0.000) circle(0.010);
\fill (0.000,0.000) circle(0.010);
\fill (1.000,0.000) circle(0.010);
\fill (0.000,1.000) circle(0.010);
\fill (-1.000,0.000) circle(0.010);
\fill (0.000,-1.000) circle(0.010);
\fill (0.707,0.707) circle(0.010);
\fill (-0.707,0.707) circle(0.010);
\fill (-0.707,-0.707) circle(0.010);
\fill (0.707,-0.707) circle(0.010);
\fill (0.000,0.000) circle(0.010);
\fill (0.000,0.000) circle(0.010);
\fill (0.000,0.000) circle(0.010);
\fill (0.000,0.000) circle(0.010);
\draw (0.000,0.000) node[anchor=south] {$v_{0},v_{1},v_{10},v_{11},v_{12},v_{13}$};
\draw (-1.000,0.000) node[anchor=east] {$v_{4}$};
\draw (0.000,-1.000) node[anchor=north] {$v_{5}$};
\draw (1.000,0.000) node[anchor=west] {$v_{2}$};
\draw (0.707,0.707) node[anchor=south west] {$v_{6}$};
\draw (0.000,1.000) node[anchor=south] {$v_{3}$};
\draw (-0.707,0.707) node[anchor=south east] {$v_{7}$};
\draw (-0.707,-0.707) node[anchor=north east] {$v_{8}$};
\draw (0.707,-0.707) node[anchor=north west] {$v_{9}$};
\draw (0.924,0.383) node[anchor=north] {$c_{0}$};
\draw (0.595,0.595) node[anchor=north west] {$c_{1}$};
\draw (-0.500,0.500) node[anchor=north east] {$c_{2}$};
\draw (0.000,0.707) node[anchor=west] {$c_{3}$};
\draw (-0.707,0.000) node[anchor=north] {$c_{4}$};
\draw (0.276,0.000) node[anchor=north] {$c_{5}$};
\draw (0.000,0.276) node[anchor=west] {$c_{6}$};
\draw (-0.300,0.000) node[anchor=north] {$c_{7}$};
\draw (0.000,-0.300) node[anchor=west] {$c_{8}$};
\draw[dotted] plot[smooth] coordinates{ (0.000,0.318) (-0.182,0.331) (-0.247,0.343) (-0.291,0.354) (-0.324,0.365) (-0.348,0.374) (-0.367,0.383) (-0.381,0.390) (-0.392,0.396) (-0.399,0.401) (-0.405,0.405) (-0.407,0.407) (-0.408,0.408) (-0.407,0.407) (-0.405,0.405) (-0.401,0.399) (-0.396,0.392) (-0.390,0.381) (-0.383,0.367) (-0.374,0.348) (-0.365,0.324) (-0.354,0.291) (-0.343,0.247) (-0.331,0.182) (-0.318,0.000) };
\draw[dotted] plot[smooth] coordinates{ (0.000,0.318) (0.182,0.331) (0.247,0.343) (0.291,0.354) (0.324,0.365) (0.348,0.374) (0.367,0.383) (0.381,0.390) (0.392,0.396) (0.399,0.401) (0.405,0.405) (0.407,0.407) (0.408,0.408) (0.407,0.407) (0.405,0.405) (0.401,0.399) (0.396,0.392) (0.390,0.381) (0.383,0.367) (0.374,0.348) (0.365,0.324) (0.354,0.291) (0.343,0.247) (0.331,0.182) (0.318,0.000) };
\draw[dotted] plot[smooth] coordinates{ (0.000,-0.318) (-0.182,-0.331) (-0.247,-0.343) (-0.291,-0.354) (-0.324,-0.365) (-0.348,-0.374) (-0.367,-0.383) (-0.381,-0.390) (-0.392,-0.396) (-0.399,-0.401) (-0.405,-0.405) (-0.407,-0.407) (-0.408,-0.408) (-0.407,-0.407) (-0.405,-0.405) (-0.401,-0.399) (-0.396,-0.392) (-0.390,-0.381) (-0.383,-0.367) (-0.374,-0.348) (-0.365,-0.324) (-0.354,-0.291) (-0.343,-0.247) (-0.331,-0.182) (-0.318,0.000) };
\draw[dotted] plot[smooth] coordinates{ (0.000,-0.318) (0.182,-0.331) (0.247,-0.343) (0.291,-0.354) (0.324,-0.365) (0.348,-0.374) (0.367,-0.383) (0.381,-0.390) (0.392,-0.396) (0.399,-0.401) (0.405,-0.405) (0.407,-0.407) (0.408,-0.408) (0.407,-0.407) (0.405,-0.405) (0.401,-0.399) (0.396,-0.392) (0.390,-0.381) (0.383,-0.367) (0.374,-0.348) (0.365,-0.324) (0.354,-0.291) (0.343,-0.247) (0.331,-0.182) (0.318,0.000) };
\end{tikzpicture}
\end{center}
The set of singular values of $\pi$ is the union of the unit circle
with the dotted curve. Points inside the dotted curve have six
preimages in $EX^*$, whereas those between the dotted curve and the
unit circle have two preimages, but there is no tidy formula for
these.
The effect of $\dl$ on the curves $c_i$ and vertices $v_j$ can be
displayed as follows:
\begin{center}
\begin{tikzpicture}[scale=4]
\draw[cyan] plot[smooth] coordinates{ (0.707,0.000) (0.588,0.000) (0.454,0.000) (0.309,0.000) (0.156,0.000) (0.000,0.000) (-0.156,0.000) (-0.309,0.000) (-0.454,0.000) (-0.588,0.000) (-0.707,0.000) (-0.809,0.000) (-0.891,0.000) (-0.951,0.000) (-0.988,0.000) (-1.000,0.000) (-0.988,0.000) (-0.951,0.000) (-0.891,0.000) (-0.809,0.000) (-0.707,0.000) (-0.588,0.000) (-0.454,0.000) (-0.309,0.000) (-0.156,0.000) (0.000,0.000) (0.156,0.000) (0.309,0.000) (0.454,0.000) (0.588,0.000) (0.707,0.000) (0.809,0.000) (0.891,0.000) (0.951,0.000) (0.988,0.000) (1.000,0.000) (0.988,0.000) (0.951,0.000) (0.891,0.000) (0.809,0.000) (0.707,0.000) };
\draw[green] plot[smooth] coordinates{ (0.000,1.000) (0.000,0.988) (0.000,0.951) (0.000,0.891) (0.000,0.809) (0.000,0.707) (0.000,0.588) (0.000,0.454) (0.000,0.309) (0.000,0.156) (0.000,-0.000) (0.000,-0.156) (0.000,-0.309) (0.000,-0.454) (0.000,-0.588) (0.000,-0.707) (0.000,-0.809) (0.000,-0.891) (0.000,-0.951) (0.000,-0.988) (0.000,-1.000) (0.000,-0.988) (0.000,-0.951) (0.000,-0.891) (0.000,-0.809) (0.000,-0.707) (0.000,-0.588) (0.000,-0.454) (0.000,-0.309) (0.000,-0.156) (0.000,0.000) (0.000,0.156) (0.000,0.309) (0.000,0.454) (0.000,0.588) (0.000,0.707) (0.000,0.809) (0.000,0.891) (0.000,0.951) (0.000,0.988) (0.000,1.000) };
\draw[green] plot[smooth] coordinates{ (0.000,1.000) (-0.156,0.988) (-0.309,0.951) (-0.454,0.891) (-0.588,0.809) (-0.707,0.707) (-0.809,0.588) (-0.891,0.454) (-0.951,0.309) (-0.988,0.156) (-1.000,-0.000) (-0.988,-0.156) (-0.951,-0.309) (-0.891,-0.454) (-0.809,-0.588) (-0.707,-0.707) (-0.588,-0.809) (-0.454,-0.891) (-0.309,-0.951) (-0.156,-0.988) (0.000,-1.000) (0.156,-0.988) (0.309,-0.951) (0.454,-0.891) (0.588,-0.809) (0.707,-0.707) (0.809,-0.588) (0.891,-0.454) (0.951,-0.309) (0.988,-0.156) (1.000,0.000) (0.988,0.156) (0.951,0.309) (0.891,0.454) (0.809,0.588) (0.707,0.707) (0.588,0.809) (0.454,0.891) (0.309,0.951) (0.156,0.988) (-0.000,1.000) };
\draw[magenta] plot[smooth] coordinates{ (0.000,0.883) (-0.102,0.873) (-0.204,0.840) (-0.304,0.787) (-0.399,0.715) (-0.486,0.625) (-0.562,0.519) (-0.624,0.401) (-0.670,0.273) (-0.698,0.138) (-0.707,-0.000) (-0.698,-0.138) (-0.670,-0.273) (-0.624,-0.401) (-0.562,-0.519) (-0.486,-0.625) (-0.399,-0.715) (-0.304,-0.787) (-0.204,-0.840) (-0.102,-0.873) (0.000,-0.883) (0.102,-0.873) (0.204,-0.840) (0.304,-0.787) (0.399,-0.715) (0.486,-0.625) (0.562,-0.519) (0.624,-0.401) (0.670,-0.273) (0.698,-0.138) (0.707,0.000) (0.698,0.138) (0.670,0.273) (0.624,0.401) (0.562,0.519) (0.486,0.625) (0.399,0.715) (0.304,0.787) (0.204,0.840) (0.102,0.873) (-0.000,0.883) };
\draw[magenta] plot[smooth] coordinates{ (0.000,0.883) (-0.102,0.873) (-0.204,0.840) (-0.304,0.787) (-0.399,0.715) (-0.486,0.625) (-0.562,0.519) (-0.624,0.401) (-0.670,0.273) (-0.698,0.138) (-0.707,-0.000) (-0.698,-0.138) (-0.670,-0.273) (-0.624,-0.401) (-0.562,-0.519) (-0.486,-0.625) (-0.399,-0.715) (-0.304,-0.787) (-0.204,-0.840) (-0.102,-0.873) (0.000,-0.883) (0.102,-0.873) (0.204,-0.840) (0.304,-0.787) (0.399,-0.715) (0.486,-0.625) (0.562,-0.519) (0.624,-0.401) (0.670,-0.273) (0.698,-0.138) (0.707,0.000) (0.698,0.138) (0.670,0.273) (0.624,0.401) (0.562,0.519) (0.486,0.625) (0.399,0.715) (0.304,0.787) (0.204,0.840) (0.102,0.873) (-0.000,0.883) };
\draw[blue] plot[smooth] coordinates{ (0.000,1.000) (0.027,0.999) (0.053,0.997) (0.078,0.993) (0.101,0.988) (0.121,0.982) (0.138,0.975) (0.151,0.967) (0.161,0.958) (0.166,0.949) (0.167,0.939) (0.163,0.930) (0.156,0.921) (0.145,0.913) (0.131,0.906) (0.114,0.899) (0.094,0.894) (0.072,0.889) (0.049,0.886) (0.025,0.884) (-0.000,0.883) (-0.025,0.884) (-0.049,0.886) (-0.072,0.889) (-0.094,0.894) (-0.114,0.899) (-0.131,0.906) (-0.145,0.913) (-0.156,0.921) (-0.163,0.930) (-0.167,0.939) (-0.166,0.949) (-0.161,0.958) (-0.151,0.967) (-0.138,0.975) (-0.121,0.982) (-0.101,0.988) (-0.078,0.993) (-0.053,0.997) (-0.027,0.999) (0.000,1.000) };
\draw[blue] plot[smooth] coordinates{ (0.000,1.000) (-0.027,0.999) (-0.053,0.997) (-0.078,0.993) (-0.101,0.988) (-0.121,0.982) (-0.138,0.975) (-0.151,0.967) (-0.161,0.958) (-0.166,0.949) (-0.167,0.939) (-0.163,0.930) (-0.156,0.921) (-0.145,0.913) (-0.131,0.906) (-0.114,0.899) (-0.094,0.894) (-0.072,0.889) (-0.049,0.886) (-0.025,0.884) (0.000,0.883) (0.025,0.884) (0.049,0.886) (0.072,0.889) (0.094,0.894) (0.114,0.899) (0.131,0.906) (0.145,0.913) (0.156,0.921) (0.163,0.930) (0.167,0.939) (0.166,0.949) (0.161,0.958) (0.151,0.967) (0.138,0.975) (0.121,0.982) (0.101,0.988) (0.078,0.993) (0.053,0.997) (0.027,0.999) (-0.000,1.000) };
\draw[blue] plot[smooth] coordinates{ (0.000,-1.000) (0.027,-0.999) (0.053,-0.997) (0.078,-0.993) (0.101,-0.988) (0.121,-0.982) (0.138,-0.975) (0.151,-0.967) (0.161,-0.958) (0.166,-0.949) (0.167,-0.939) (0.163,-0.930) (0.156,-0.921) (0.145,-0.913) (0.131,-0.906) (0.114,-0.899) (0.094,-0.894) (0.072,-0.889) (0.049,-0.886) (0.025,-0.884) (-0.000,-0.883) (-0.025,-0.884) (-0.049,-0.886) (-0.072,-0.889) (-0.094,-0.894) (-0.114,-0.899) (-0.131,-0.906) (-0.145,-0.913) (-0.156,-0.921) (-0.163,-0.930) (-0.167,-0.939) (-0.166,-0.949) (-0.161,-0.958) (-0.151,-0.967) (-0.138,-0.975) (-0.121,-0.982) (-0.101,-0.988) (-0.078,-0.993) (-0.053,-0.997) (-0.027,-0.999) (0.000,-1.000) };
\draw[blue] plot[smooth] coordinates{ (0.000,-1.000) (-0.027,-0.999) (-0.053,-0.997) (-0.078,-0.993) (-0.101,-0.988) (-0.121,-0.982) (-0.138,-0.975) (-0.151,-0.967) (-0.161,-0.958) (-0.166,-0.949) (-0.167,-0.939) (-0.163,-0.930) (-0.156,-0.921) (-0.145,-0.913) (-0.131,-0.906) (-0.114,-0.899) (-0.094,-0.894) (-0.072,-0.889) (-0.049,-0.886) (-0.025,-0.884) (0.000,-0.883) (0.025,-0.884) (0.049,-0.886) (0.072,-0.889) (0.094,-0.894) (0.114,-0.899) (0.131,-0.906) (0.145,-0.913) (0.156,-0.921) (0.163,-0.930) (0.167,-0.939) (0.166,-0.949) (0.161,-0.958) (0.151,-0.967) (0.138,-0.975) (0.121,-0.982) (0.101,-0.988) (0.078,-0.993) (0.053,-0.997) (0.027,-0.999) (-0.000,-1.000) };
\fill (0.000,1.000) circle(0.010);
\fill (0.000,-1.000) circle(0.010);
\fill (0.707,0.000) circle(0.010);
\fill (-0.707,0.000) circle(0.010);
\fill (-0.707,0.000) circle(0.010);
\fill (0.707,0.000) circle(0.010);
\fill (0.000,0.000) circle(0.010);
\fill (-1.000,0.000) circle(0.010);
\fill (0.000,0.000) circle(0.010);
\fill (1.000,0.000) circle(0.010);
\fill (0.000,0.883) circle(0.010);
\fill (0.000,0.883) circle(0.010);
\fill (0.000,-0.883) circle(0.010);
\fill (0.000,-0.883) circle(0.010);
\draw (0.000,1.000) node[anchor=south] {$v_{0}$};
\draw (0.000,-1.000) node[anchor=north] {$v_{1}$};
\draw (0.000,0.000) node[anchor=north] {$v_{6},v_{8}$};
\draw (0.000,0.883) node[anchor=north] {$v_{10},v_{11}$};
\draw (-0.707,0.000) node[anchor=north] {$v_{3},v_{4}$};
\draw (1.000,0.000) node[anchor=west] {$v_{9}$};
\draw (0.707,0.000) node[anchor=north] {$v_{2},v_{5}$};
\draw (-1.000,0.000) node[anchor=east] {$v_{7}$};
\draw (0.000,-0.883) node[anchor=south] {$v_{12},v_{13}$};
\draw (0.383,0.000) node[anchor=north] {$c_{0}$};
\draw (0.000,0.540) node[anchor=west] {$c_{1}$};
\draw (-0.707,0.707) node[anchor=south east] {$c_{2}$};
\draw (-0.486,0.625) node[anchor=north west] {$c_{3}$};
\draw (-0.581,0.488) node[anchor=north west] {$c_{4}$};
\draw (0.148,0.915) node[anchor=west] {$c_{5}$};
\draw (-0.148,0.915) node[anchor=east] {$c_{6}$};
\draw (0.148,-0.915) node[anchor=west] {$c_{7}$};
\draw (-0.148,-0.915) node[anchor=east] {$c_{8}$};
\end{tikzpicture}
\end{center}
Here again the image of $\dl$ is the full unit disc. For most points
in the disc, the preimage consists of two points that are exchanged by
the action of $\lm^{-1}\nu$, and one can give a nice formula for these
points.
For the map $\zt$, we have the following picture:
\begin{center}
\begin{tikzpicture}[scale=4]
\draw[cyan] plot[smooth] coordinates{ (0.000,0.000) (0.000,0.156) (0.000,0.309) (0.000,0.454) (0.000,0.588) (0.000,0.707) (0.000,0.809) (0.000,0.891) (0.000,0.951) (0.000,0.988) (0.000,1.000) (0.000,0.988) (0.000,0.951) (0.000,0.891) (0.000,0.809) (0.000,0.707) (0.000,0.588) (0.000,0.454) (0.000,0.309) (0.000,0.156) (0.000,-0.000) (0.000,-0.156) (0.000,-0.309) (0.000,-0.454) (0.000,-0.588) (0.000,-0.707) (0.000,-0.809) (0.000,-0.891) (0.000,-0.951) (0.000,-0.988) (0.000,-1.000) (0.000,-0.988) (0.000,-0.951) (0.000,-0.891) (0.000,-0.809) (0.000,-0.707) (0.000,-0.588) (0.000,-0.454) (0.000,-0.309) (0.000,-0.156) (0.000,0.000) };
\draw[green] plot[smooth] coordinates{ (0.707,0.000) (0.698,0.111) (0.672,0.219) (0.630,0.321) (0.572,0.416) (0.500,0.500) (0.416,0.572) (0.321,0.630) (0.219,0.672) (0.111,0.698) (-0.000,0.707) (-0.111,0.698) (-0.219,0.672) (-0.321,0.630) (-0.416,0.572) (-0.500,0.500) (-0.572,0.416) (-0.630,0.321) (-0.672,0.219) (-0.698,0.111) (-0.707,-0.000) (-0.698,-0.111) (-0.672,-0.219) (-0.630,-0.321) (-0.572,-0.416) (-0.500,-0.500) (-0.416,-0.572) (-0.321,-0.630) (-0.219,-0.672) (-0.111,-0.698) (0.000,-0.707) (0.111,-0.698) (0.219,-0.672) (0.321,-0.630) (0.416,-0.572) (0.500,-0.500) (0.572,-0.416) (0.630,-0.321) (0.672,-0.219) (0.698,-0.111) (0.707,0.000) };
\draw[green] plot[smooth] coordinates{ (0.707,0.000) (0.698,0.111) (0.672,0.219) (0.630,0.321) (0.572,0.416) (0.500,0.500) (0.416,0.572) (0.321,0.630) (0.219,0.672) (0.111,0.698) (-0.000,0.707) (-0.111,0.698) (-0.219,0.672) (-0.321,0.630) (-0.416,0.572) (-0.500,0.500) (-0.572,0.416) (-0.630,0.321) (-0.672,0.219) (-0.698,0.111) (-0.707,-0.000) (-0.698,-0.111) (-0.672,-0.219) (-0.630,-0.321) (-0.572,-0.416) (-0.500,-0.500) (-0.416,-0.572) (-0.321,-0.630) (-0.219,-0.672) (-0.111,-0.698) (0.000,-0.707) (0.111,-0.698) (0.219,-0.672) (0.321,-0.630) (0.416,-0.572) (0.500,-0.500) (0.572,-0.416) (0.630,-0.321) (0.672,-0.219) (0.698,-0.111) (0.707,0.000) };
\draw[magenta] plot[smooth] coordinates{ (0.986,0.000) (0.973,0.156) (0.937,0.309) (0.878,0.454) (0.797,0.588) (0.697,0.707) (0.579,0.809) (0.447,0.891) (0.305,0.951) (0.154,0.988) (-0.000,1.000) (-0.154,0.988) (-0.305,0.951) (-0.447,0.891) (-0.579,0.809) (-0.697,0.707) (-0.797,0.588) (-0.878,0.454) (-0.937,0.309) (-0.973,0.156) (-0.986,-0.000) (-0.973,-0.156) (-0.937,-0.309) (-0.878,-0.454) (-0.797,-0.588) (-0.697,-0.707) (-0.579,-0.809) (-0.447,-0.891) (-0.305,-0.951) (-0.154,-0.988) (0.000,-1.000) (0.154,-0.988) (0.305,-0.951) (0.447,-0.891) (0.579,-0.809) (0.697,-0.707) (0.797,-0.588) (0.878,-0.454) (0.937,-0.309) (0.973,-0.156) (0.986,0.000) };
\draw[magenta] plot[smooth] coordinates{ (0.169,0.000) (0.167,0.000) (0.161,0.000) (0.151,0.000) (0.137,0.000) (0.120,0.000) (0.099,0.000) (0.077,0.000) (0.052,0.000) (0.026,0.000) (-0.000,0.000) (-0.026,0.000) (-0.052,0.000) (-0.077,0.000) (-0.099,0.000) (-0.120,0.000) (-0.137,0.000) (-0.151,0.000) (-0.161,0.000) (-0.167,0.000) (-0.169,0.000) (-0.167,0.000) (-0.161,0.000) (-0.151,0.000) (-0.137,0.000) (-0.120,0.000) (-0.099,0.000) (-0.077,0.000) (-0.052,0.000) (-0.026,0.000) (0.000,0.000) (0.026,0.000) (0.052,0.000) (0.077,0.000) (0.099,0.000) (0.120,0.000) (0.137,0.000) (0.151,0.000) (0.161,0.000) (0.167,0.000) (0.169,0.000) };
\filldraw[draw=blue,fill=gray!20] plot[smooth] coordinates{ (0.707,0.000) (0.709,0.000) (0.715,0.000) (0.725,0.000) (0.737,0.000) (0.753,0.000) (0.771,0.000) (0.791,0.000) (0.812,0.000) (0.834,0.000) (0.856,0.000) (0.877,0.000) (0.898,0.000) (0.917,0.000) (0.934,0.000) (0.949,0.000) (0.962,0.000) (0.972,0.000) (0.980,0.000) (0.984,0.000) (0.986,0.000) (0.984,0.000) (0.980,0.000) (0.972,0.000) (0.962,0.000) (0.949,0.000) (0.934,0.000) (0.917,0.000) (0.898,0.000) (0.877,0.000) (0.856,0.000) (0.834,0.000) (0.812,0.000) (0.791,0.000) (0.771,0.000) (0.753,0.000) (0.737,0.000) (0.725,0.000) (0.715,0.000) (0.709,0.000) (0.707,0.000) };
\filldraw[draw=blue,fill=gray!20] plot[smooth] coordinates{ (0.707,0.000) (0.703,0.055) (0.691,0.109) (0.671,0.158) (0.644,0.203) (0.612,0.241) (0.575,0.272) (0.535,0.295) (0.493,0.310) (0.451,0.317) (0.409,0.316) (0.368,0.308) (0.330,0.292) (0.294,0.270) (0.262,0.242) (0.235,0.209) (0.211,0.172) (0.193,0.132) (0.180,0.090) (0.172,0.045) (0.169,-0.000) (0.172,-0.045) (0.180,-0.090) (0.193,-0.132) (0.211,-0.172) (0.235,-0.209) (0.262,-0.242) (0.294,-0.270) (0.330,-0.292) (0.368,-0.308) (0.409,-0.316) (0.451,-0.317) (0.493,-0.310) (0.535,-0.295) (0.575,-0.272) (0.612,-0.241) (0.644,-0.203) (0.671,-0.158) (0.691,-0.109) (0.703,-0.055) (0.707,0.000) };
\filldraw[draw=blue,fill=gray!20] plot[smooth] coordinates{ (-0.707,0.000) (-0.709,0.000) (-0.715,0.000) (-0.725,0.000) (-0.737,0.000) (-0.753,0.000) (-0.771,0.000) (-0.791,0.000) (-0.812,0.000) (-0.834,0.000) (-0.856,0.000) (-0.877,0.000) (-0.898,0.000) (-0.917,0.000) (-0.934,0.000) (-0.949,0.000) (-0.962,0.000) (-0.972,0.000) (-0.980,0.000) (-0.984,0.000) (-0.986,0.000) (-0.984,0.000) (-0.980,0.000) (-0.972,0.000) (-0.962,0.000) (-0.949,0.000) (-0.934,0.000) (-0.917,0.000) (-0.898,0.000) (-0.877,0.000) (-0.856,0.000) (-0.834,0.000) (-0.812,0.000) (-0.791,0.000) (-0.771,0.000) (-0.753,0.000) (-0.737,0.000) (-0.725,0.000) (-0.715,0.000) (-0.709,0.000) (-0.707,0.000) };
\filldraw[draw=blue,fill=gray!20] plot[smooth] coordinates{ (-0.707,0.000) (-0.703,0.055) (-0.691,0.109) (-0.671,0.158) (-0.644,0.203) (-0.612,0.241) (-0.575,0.272) (-0.535,0.295) (-0.493,0.310) (-0.451,0.317) (-0.409,0.316) (-0.368,0.308) (-0.330,0.292) (-0.294,0.270) (-0.262,0.242) (-0.235,0.209) (-0.211,0.172) (-0.193,0.132) (-0.180,0.090) (-0.172,0.045) (-0.169,-0.000) (-0.172,-0.045) (-0.180,-0.090) (-0.193,-0.132) (-0.211,-0.172) (-0.235,-0.209) (-0.262,-0.242) (-0.294,-0.270) (-0.330,-0.292) (-0.368,-0.308) (-0.409,-0.316) (-0.451,-0.317) (-0.493,-0.310) (-0.535,-0.295) (-0.575,-0.272) (-0.612,-0.241) (-0.644,-0.203) (-0.671,-0.158) (-0.691,-0.109) (-0.703,-0.055) (-0.707,0.000) };
\fill (0.707,0.000) circle(0.010);
\fill (-0.707,0.000) circle(0.010);
\fill (0.000,0.000) circle(0.010);
\fill (0.000,1.000) circle(0.010);
\fill (0.000,0.000) circle(0.010);
\fill (0.000,-1.000) circle(0.010);
\fill (0.000,0.707) circle(0.010);
\fill (0.000,0.707) circle(0.010);
\fill (0.000,-0.707) circle(0.010);
\fill (0.000,-0.707) circle(0.010);
\fill (0.169,0.000) circle(0.010);
\fill (0.986,0.000) circle(0.010);
\fill (-0.169,0.000) circle(0.010);
\fill (-0.986,0.000) circle(0.010);
\draw (0.000,0.707) node[anchor=south] {$v_{6},v_{7}$};
\draw (-0.169,0.000) node[anchor=east] {$v_{12}$};
\draw (0.000,0.000) node[anchor=south] {$v_{2},v_{4}$};
\draw (0.707,0.000) node[anchor=east] {$v_{0}$};
\draw (0.986,0.000) node[anchor=west] {$v_{11}$};
\draw (-0.707,0.000) node[anchor=west] {$v_{1}$};
\draw (0.000,-0.707) node[anchor=north] {$v_{8},v_{9}$};
\draw (0.000,-1.000) node[anchor=north] {$v_{5}$};
\draw (-0.986,0.000) node[anchor=east] {$v_{13}$};
\draw (0.169,0.000) node[anchor=west] {$v_{10}$};
\draw (0.000,1.000) node[anchor=south] {$v_{3}$};
\draw (0.000,0.383) node[anchor=west] {$c_{0}$};
\draw (0.500,0.500) node[anchor=south] {$c_{1}$};
\draw (0.500,0.500) node[anchor=east] {$c_{2}$};
\draw (0.697,0.707) node[anchor=west] {$c_{3}$};
\draw (0.120,0.000) node[anchor=north] {$c_{4}$};
\draw (0.856,0.000) node[anchor=north] {$c_{5}$};
\draw (0.409,0.316) node[anchor=north] {$c_{6}$};
\draw (-0.856,0.000) node[anchor=north] {$c_{7}$};
\draw (-0.409,0.316) node[anchor=north] {$c_{8}$};
\end{tikzpicture}
\end{center}
In this case, the shaded regions are not part of the image, so the
image is homeomorphic to a disc with two holes. It can be identified
with $EX^*/\ip{\lm^2\nu}$, and is combinatorially equivalent to the
space $\Net_4^+$ discussed in Section~\ref{sec-fundamental}.
\subsection{Homeomorphisms with the square}
\label{sec-roothalf-square}
\begin{proposition}
There is a homeomorphism $\tau\:EX^*/G\to [0,1]^2$ given by
\[ \tau(x) = (2z_1-z_1^2+\half z_1^2z_2,\;2z_2). \]
Moreover, we can define a homeomorphism $\tau^*\:[0,1]^2\to F_{16}$ by
\[ \tau^*(t_1,t_2) = \left(
\sqrt{\frac{(1-\sqrt{t_2})(\sqrt{t_3}+\sqrt{t_2})}{2(2+\sqrt{t_2})}},\;
\sqrt{\frac{(1+\sqrt{t_2})(\sqrt{t_3}-\sqrt{t_2})}{2(2-\sqrt{t_2})}},\;
\rt\sqrt{\frac{2-\sqrt{t_3}}{4-t_2}},\;
-\sqrt{t_2}\sqrt{\frac{2-\sqrt{t_3}}{4-t_2}}
\right),
\]
where $t_3=t_1t_2+4(1-t_1)$. This is inverse to $\tau|_{F_{16}}$.
\end{proposition}
Maple notation for $\tau(x)$ and $\tau^*(t)$ is \mcode+t_proj(x)+ and
\mcode+t_lift(t)+.
\begin{proof}
First, recall that the map $p_{16}\:x\mapsto z$ gives a homeomorphism
from $F_{16}\simeq EX^*/G$ to the set
\[ F^*_{16} = \{z\in \R^2\st 0\leq z_2\leq 1/2,\;
0\leq z_1\leq (1+\sqrt{z_2/2})^{-1}\}.
\]
We have $\tau=\sg\circ p_{16}$, where
\[ \sg(z) = (2z_1-z_1^2+\half z_1^2z_2,\;2z_2). \]
An elementary exercise shows that for fixed $z_2\in[0,1/2]$, the map
$z_1\mapsto 2z_1-z_1^2+\half z_1^2z_2$ gives a bijection from the
interval $[0,(1+\sqrt{z_2/2})^{-1}]$ to $[0,1]$. This implies that
$\sg$ gives a continuous bijection from $F^*_{16}$ to $[0,1]^2$. It
follows that $\tau$ gives a continuous bijection from $F_{16}$ to
$[0,1]^2$. All the spaces involved are compact and Hausdorff, so
continuous bijections are homeomorphisms.
Next, it is clear that the definition of $\tau^*$ involves only
strictly positive denominators, and square roots of nonnegative
quantities, so it gives a well-defined and continuous map from
$[0,1]^2$ to $\R^4$. Routine simplification gives
$\rho(\tau^*(t))=1$ and $g(\tau^*(t))=0$, so $\tau^*(t)\in EX^*$.
Recall also that $F_{16}=\{x\in EX^*\st x_1,x_2,y_1,y_2\geq 0\}$,
where $y_1=x_3$ and $y_2$ is given generically by $-x_4/x_3$. This
makes it clear that the image of $\tau^*$ is contained in $F_{16}$.
It is now easy to check that $\tau\tau^*=1$, so $\tau^*$ is the
inverse of $\tau$.
\begin{checks}
embedded/roothalf/E_roothalf_check.mpl: check_t_proj()
\end{checks}
\end{proof}
The map $\tau$ is clearly smooth, but the inverse map fails to be
smooth on the boundary of $[0,1]^2$. This is a necessary consequence
of the fact that $\tau$ comes from a $G$-invariant smooth function
defined on all of $EX^*$, but it is often awkward. For example, we
can try to use $\tau$ to convert integrals over $F_{16}$ to integrals
over $[0,1]^2$, but the singular boundary behaviour makes it difficult
to obtain accurate results, even with adaptive quadrature methods. We
will therefore describe a different map, which has a different set of
advantages and disadvantages.
\begin{definition}
We define $\dl\:EX^*\to\R^2$ by
\begin{align*}
\al_0(x) &= x_3 - x_4/\rt + x_1^2 + x_4^2 + x_3(x_4 - x_2)/\rt \\
\al_1(x) &= (\tfrac{3}{\sqrt{8}}-1)x_1+x_2-x_3-\rt x_4 \\
\al_2(x) &= x_1 - \tfrac{3}{4}\sqrt{3}x_3 +
(3-\tfrac{3}{4}\sqrt{6})x_4 \\
\dl(x) &=
\left(x_3\al_0(x) - x_2^2x_4,\; (x_2-x_1)\al_1(x) + x_4\al_2(x)\right).
\end{align*}
\end{definition}
\begin{proposition}\label{prop-square-diffeo}
The map $\dl$ gives a diffeomorphism $F_{16}\to[0,1]^2$, which satisfies
\[ \dl(C_0)\sse 0 \tm\R \hspace{3em}
\dl(C_1)\sse \R\tm 0 \hspace{3em}
\dl(C_3)\sse \R\tm 1 \hspace{3em}
\dl(C_5)\sse 1 \tm\R.
\]
\end{proposition}
In order to prove this, we will need to consider the Jacobian of
$\dl$. It will be convenient to formulate the required discussion
more generally.
\begin{definition}
Consider a map $f\:EX^*\to\R^2$, and a point $a\in EX^*$. Choose an
oriented orthonormal basis $(u,v)$ for $T_aEX^*$, giving vectors
$f_*(u),f_*(v)\in\R^2$. It is easy to see that the determinant
$\det(f_*(u),f_*(v))$ is independent of the choice of $(u,v)$. We
write $j(f)(a)$ for this determinant, and we call $j(f)$ the
\emph{Jacobian} of $f$.
\end{definition}
\begin{lemma}\label{lem-jacobian}
Suppose that there are functions $f_1,f_2$ defined on some
neighbourhood of $EX^*$ in $\R^4$, such that $f(x)=(f_1(x),f_2(x))$.
Put
\[ \tj(f) =
\det\left(x,n,\nabla f_1,\nabla f_2\right)
\]
(where as usual $n=\nabla g$). Then $j(f)=\tj(f)/\|n\|$.
\end{lemma}
\begin{proof}
Fix a point $a\in EX^*$, and choose an oriented orthonormal basis
$(u,v)$ for the corresponding tangent space, as before. We write
$n$ for $n(a)$, and $w_i$ for the value of $\nabla f_i$ at $a$. Now
$(a,n/\|n\|,u,v)$ is an oriented orthonormal basis for $\R^4$, so
we can write
\[ w_i = \al_i a + \bt_i n/\|n\| + \gm_i u + \dl_i v \]
for some scalars $\al_i,\dotsc,\dl_i$. We have
\begin{align*}
f_*(u) &= (u.w_1,\;u.w_2) = (\gm_1,\gm_2) \\
f_*(v) &= (v.w_1,\;v.w_2) = (\dl_1,\dl_2),
\end{align*}
so $j(f)(a)=\gm_1\dl_2-\gm_2\dl_1$. We need to show that this is the
same as
\[ D = \det(a,n/\|n\|,w_1,w_2). \]
Because $(a,n/\|n\|,u,v)$ is an oriented orthonormal basis for
$\R^4$, we find that
\[ D = \det
\bbm
1 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 \\
\al_1 & \bt_1 & \dl_1 & \gm_1 \\
\al_2 & \bt_2 & \dl_2 & \gm_2
\ebm = \gm_1\dl_2-\gm_2\dl_1
\]
as required.
\end{proof}
\begin{proof}[Proof of Proposition~\ref{prop-square-diffeo}]
Put $h(x)=x_3x_4/\rt+x_1^2+x_4^2$. As we have noted previously, this
vanishes on $C_5$. By direct expansion of polynomials, we find the
following:
\begin{itemize}
\item If $x=(x_1,x_2,0,0)$ then $\dl(x)_1=0$.
\item If $x=(x_1,x_1,x_3,x_4)$ then $\dl(x)_2=0$.
\item If $x=(0,x_2,x_3,-x_3/\rt)$ then $\dl(x)_2=\rho(x)$.
\item If $x=(x_1,0,x_3,x_4)$ then $\dl(x)_1=\rho(x)-(1-x_3)h(x)$.
\end{itemize}
We can compare this with the definitions of the curves $c_k(t)$ for
$k\in\{0,1,3,5\}$, remembering that $\rho(x)=1$ for $x\in EX^*$; we
find that the images $\dl(C_k)$ are as claimed. It is also
straightforward to check that $\dl$ sends $v_6$, $v_0$, $v_3$ and
$v_{11}$ to $(0,0)$, $(1,0)$, $(0,1)$ and $(1,1)$ respectively.
Next, we can use Lemma~\ref{lem-jacobian} to obtain a formula for the
Jacobian $j(\dl)(x)$. By numerical evaluation and plotting, we find
that $j(f)>0.1$ everywhere in $F_{16}$. (More precisely, the minimum
is approximately $0.1079$, attained at a point $c_1(t_0)$ for some
$t_0$ with $0<|t_0-5\pi/16|<0.002$.)
Now put $E_0=C_0\cap F_{16}$, and note that $\dl_2$ gives a map from
$E_0$ to $\R$ with $\dl_2(v_6)=0$ and $\dl_2(v_3)=1$. As $j(\dl)>0$
on $F_{16}$ we see that this restricted map has no critical points,
so it must give a diffeomorphism $E_0\to[0,1]$. By applying the same
line of argument to the other edges of $F_{16}$, we find that $\dl$
gives a bijection $\partial F_{16}\to\partial [0,1]^2$.
Now consider a point $b\in (0,1)^2$, and put
$A=\{x\in F_{16}\st\dl(x)=b\}$. As
$\dl(\partial F_{16})=\partial[0,1]^2$, we see that
$A\cap\partial F_{16}=\emptyset$. Using the fact that $j(\dl)>0$ on
$F_{16}$, we see that $A$ is discrete in $F_{16}$, and therefore
finite, say $A=\{a_1,\dotsc,a_n\}$. The fact that $j(\dl)>0$ on
$F_{16}$ also means that $\dl$ gives an orientation preserving
homeomorphism from some neighbourhood of $a_i$ to some neighbourhood
of $b$, and therefore induces an isomorphism
\[ H_2(F_{16},F_{16}\sm\{a_i\})\to H_2(\R^2,\R^2\sm\{b\})\simeq\Z \]
of homology groups. We also have a commutative diagram
\[ \xymatrix{
H_2(F_{16},\partial F_{16}) \ar[r]\ar[d]_{\dl_*} &
H_2(F_{16},F_{16}\sm A) \ar[d]^{\dl_*} \\
H_2(\R^2,\partial[0,1]^2) \ar[r]_\simeq &
H_2(\R^2,\R^2\sm\{b\})
} \]
Because $F_{16}$ and $\R^2$ are contractible, and
$\dl\:\partial F_{16}\to\partial [0,1]^2$ is a homeomorphism, we see
that the left hand map is an isomorphism. Standard methods also show
that the bottom map is an isomorphism, with both groups being
isomorphic to $\Z$. On the other hand, $H_2(F_{16},A^c)$ splits as
the sum of the groups $H_2(F_{16},F_{16}\sm\{a_i\})\simeq\Z$
indexed by the points of $A$, and the map
\[ H_2(F_{16},F_{16}\sm A)\to
\bigoplus_{i=1}^n (F_{16},F_{16}\sm\{a_i\})
\]
is just the diagonal map $\Z\to\Z^n$. We have seen that $\dl_*$ acts
as the identity on each summand, and this can only be consistent if
$n=1$. It follows that $\dl$ gives a bijection $F_{16}\to [0,1]^2$,
as claimed.
\begin{checks}
embedded/roothalf/square_diffeo_check.mpl: check_square_diffeo_E0()
\end{checks}
\end{proof}
\subsection{Charts}
\label{sec-roothalf-charts}
Recall from Section~\ref{sec-holomorphic-curves} that each of
the maps $c_k\:\R\to EX^*$ can be extended in a canonical way to give
a holomorphic map defined on a neighbourhood of $\R$ in $\C$. There
is no case where we know a closed formula for such a holomorphic
extension. However, it is not too hard to calculate high order power
series approximations. For example, we have found a map
$c_0^*\:\C\to\R^4$ such that
\begin{itemize}
\item Each component $c_0^*(t+iu)_n$ (for $1\leq n\leq 4$) is a
polynomial of total degree at most $44$ in $t$ and $u$, with
coefficients in $\Q(\rt,\sqrt{3})$.
\item The polynomials $\rho(c_0^*(t+iu))-1$ and $g(c_0^*(t+iu))$ lie
in $(t,u)^{45}$, so $c_0^*(t+iu)$ lies very close to $EX^*$ when
$(t,u)$ is small.
\item If we put $a=\partial c_0^*(t+iu)/\partial t$ and
$b=\partial c_0^*(t+iu)/\partial u$ then $\ip{a,b}$ and
$\ip{a,a}-\ip{b,b}$ lie in $(t,u)^{44}$, so $c_0^*$ is very close to
being conformal.
\item $c^*_0|_{\R}$ is the 44th order Taylor approximation to $c_0$
at $t=0$.
\end{itemize}
These conditions imply that $c^*_0(z)$ agrees with the holomorphic
extension $\tc_0(z)$ to order 44 at $z=0$.
Calculations of this kind are implemented by methods of the class
\mcode+E_chart+, which is defined in the file
\fname+embedded/roothalf/E_atlas.mpl+. Specifically, we can
calculate the above chart as follows:
\begin{mcodeblock}
C := `new/E_chart`():
C["curve_set_exact",0,0]:
C["curve_set_degree_exact",45]:
C["p"]([t,u]);
\end{mcodeblock}
Proposition~\ref{prop-chart} can also be used to define charts at
points that do not lie on the curves $C_k$:
\begin{proposition}\label{prop-frame-chart}
Suppose that $a\in EX^*$, and that $(u,v)$ is an oriented orthonormal
basis for the tangent space $T_aEX^*$. Then there is a unique local
conformal chart $\phi$ with $u.\phi(t)=t$ and $v.\phi(t)=0$ for
small $t\in\R$.
\end{proposition}
\begin{proof}
We can define a real analytic map $\pi\:EX^*\to\C$ by
$\pi(x)=u.x+iv.x$ (so $\pi(a)=0$). This induces an isomorphism
$T_aEX^*\to\C$, so it is locally invertible, and we can define
$c(t)=\pi^{-1}(t)$ for small $t\in\R$. Proposition~\ref{prop-chart}
now gives a holomorphic map $\phi$ that is defined on a neighbourhood
of $0$ in $\C$ and extends $c$.
\end{proof}
It is again fairly straightforward to find polynomial approximations
to $\phi$, of any desired order. We can start with
$\phi_1(t+is)=a+tu+sv$. Suppose we have defined $\phi_d$ of degree
$d$ such that
\begin{align*}
\rho(\phi_d(t+is)) &= 1 \pmod{(s,t)^{d+1}} \\
g(\phi_d(t+is)) &= 0 \pmod{(s,t)^{d+1}} \\
\ip{u,\phi_d(t)} &= t \pmod{(s,t)^{d+1}} \\
\ip{v,\phi_d(t)} &= 0 \pmod{(s,t)^{d+1}} \\
\ip{\partial_t\phi_d(t+is),\partial_u\phi_d(t+is)}
&= 0 \pmod{(s,t)^d} \\
\ip{\partial_t\phi_d(t+is),\partial_t\phi_d(t+is)} -
\ip{\partial_u\phi_d(t+is),\partial_u\phi_d(t+is)}
&= 0 \pmod{(s,t)^d}.
\end{align*}
(Note that we assume a lower degree of accuracy in the last two
conditions, which is natural because they involve a derivative.) We
then take
\[ p_{d+1}(t+is)=p_d(t+is) + \sum_{j=0}^{d+1}\al_jt^js^{d+1-j}, \]
where $\al_j\in\R^4$. It is not hard to see that $p_{d+1}$ satisfies
the required conditions with one more degree of accuracy iff the
coefficients $\al_j$ satisfy a certain system of inhomogeneous linear
equations. The abstract theory tells us that these equations must be
uniquely solvable, and of course that is easily verified in any
explicit computation. We have implemented a version of this using
numerical approximations for the power series coefficients. (As
usual, we work with 100 digit precision by default.) If \mcode+x0+ is a
point in $EX^*$, we can enter the following to find a chart of
polynomial degree $20$:
\begin{mcodeblock}
C := `new/E_chart`():
C["centre_set_numeric",x0]:
C["centre_set_degree_numeric",20]:
C["p"]([t,u]);
\end{mcodeblock}
We have found charts centred at many different points of $EX^*$. The
real problem is to patch them together by some kind of analytic
continuation. The only way we have succeeded in doing this is via a
hyperbolic rescaling of the metric, as we will discuss in
Section~\ref{sec-rescaling}. (We have attempted various more direct
approaches to numerical analytic continuation, but the results were
not robust, and the literature suggests that we should not expect
otherwise.)
\subsection{Torus quotients}
\label{sec-torus-quotients}
We saw in Section~\ref{sec-quotients} that for any cromulent surface
$X$, the quotients $X/\ip{\mu}$ and $X/\ip{\lm\mu}$ are tori. In
Section~\ref{sec-ellquot}, we gave a detailed analysis of these
quotients for the projective family. In the present section, we study
$EX^*/\ip{\mu}$ and $EX^*/\ip{\lm\mu}$. If we were very optimistic we
might hope for explicit conformal isomorphisms between these quotients
and suitable elliptic curves, but we have not achieved that. However,
we will write down reasonably simple formulae for homeomorphisms from
$EX^*/\ip{\mu}$ and $EX^*/\ip{\lm\mu}$ to $S^1\tm S^1$, which have all
the expected equivariance properties and homological properties, and
which do not deviate too far from being conformal.
\begin{definition}\label{defn-AR}
We recall from Proposition~\ref{prop-roothalf-fundamental} that
$|y_1|\leq 1$ and $|y_2|\leq 1/\rt$ on $EX^*$, so we can define functions
$r_1,r_2\:EX^*\to\R^+$ by $r_1=\sqrt{1-y_2/\rt}$ and
$r_2=\sqrt{1+y_2/\rt}$. We write $AR$ for the extension of
$A=\CO_{EX^*}$ generated by $r_1$ and $r_2$, and note that $AR$ is
freely generated by the set
\[ \{x_1^ix_2^jr_1^kr_2^l\st 0\leq i,j,k,l\leq 1\} \]
as a module over $\R[y_1,y_2]$. We also write $KR$ for the field of
fractions of $AR$, which is freely generated by the same set as a
module over $\R(y_1,y_2)$.
\end{definition}
\begin{remark}\label{rem-KR-subfields}
The field $KR$ has automorphisms $\al_1$ and $\al_2$ which act as the
identity on the subring $A$ and satisfy
\begin{align*}
\al_1(r_1) &= -r_1 & \al_1(r_2) &= \pp r_2 \\
\al_2(r_1) &= \pp r_1 & \al_2(r_2) &= -r_2.
\end{align*}
The group $G'=\ip{G,\al_1,\al_2}$ has order $64$, and it acts on
$KR$. Some of the work in this section and the following section can
be interpreted in terms of the Galois theory of this action. There
is code related to this in the files
\fname+embedded/roothalf/group64.mpl+ and
\fname+embedded/roothalf/KR_subfields.mpl+.
\end{remark}
\begin{definition}\label{defn-E-to-TTC}
We define $\tau_1\:EX^*\to S^1\subset\C$ by
\[ \tau_1(x) =
\frac{y_1(1-y_2/\rt)-1/\rt+ix_1}{(1-y_1/\rt)r_1}
\]
A straightforward calculation, using the relations in
Section~\ref{sec-E-functions}, shows that $|\tau_1(x)|^2=1$.
We then define $\tau_i\:EX^*\to S^1$ for $2\leq i\leq 4$ by
\[ \tau_2 = \tau_1\lm^{-1} \hspace{4em}
\tau_3 = \tau_1\mu^{-1} \hspace{4em}
\tau_4 = \tau_1(\lm\mu)^{-1}.
\]
We define $q\:EX^*\to(S^1)^4$ by
\[ q(x) = (\tau_1(x),\tau_2(x),\tau_3(x),\tau_4(x)). \]
\end{definition}
\begin{remark}
Although we have not succeeded in formulating a precise theorem in
this direction, extensive experimental investigation suggests that
the map $\tau_1$ is much simpler and better behaved than any other
map in the same homotopy class.
\end{remark}
\begin{remark}\label{rem-q-denom}
We can define a homeomorphism $S^1\to\R_\infty$ by
$x+iy\mapsto(1-x)/y$, or equivalently
$e^{i\tht}\mapsto\tan(\tht/2)$. Composing $\tau_1$ with this gives
the map $\tau_1^*\:EX^*\to\R_\infty$ with formula
\[ \tau_1^*(x) = (1/\rt+r_1)(1-r_1y_1)/x_1. \]
More explicitly, if $x\in EX^*$ is such that the numerator and
denominator of the above fraction are not both zero, then the
fraction can be interpreted in an obvious way as an element of
$\R_\infty$, and that element is $\tau_1^*(x)$. However, for
$x\in C_6$, the numerator and denominator both vanish, so we can only
evaluate $\tau_1^*(x)$ by first simplifying $\tau_1(x)$, or by taking
a limit over nearby points. For some purposes it is convenient to
work with $\tau_1^*$ instead of $\tau_1$, but these kinds of
degenerate cases cause significant trouble. We also put
\[ q^*(x) = (\tau_1^*(x),\tau_2^*(x),\tau_3^*(x),\tau_4^*(x))
\in (\R_\infty)^4.
\]
\end{remark}
\begin{remark}
In Maple, we also need to distinguish explicitly between the circle
in $\C$ and the circle in $\R^2$, and thus between the $4$-torus in
$\C^4$ and the $4$-torus in $\R^8$. We thus have two versions of
$q$, namely \mcode+E_to_TTC+ (with values in $\C^4$) and
\mcode+E_to_TT+ (with values in $\R^8$). We also have
\mcode+E_to_TTP+, corresponding to $q^*$. There are functions
\mcode+TT_to_TTC+ and so on, which convert between these
representations.
\end{remark}
\begin{proposition}\label{prop-E-to-TTC}
The map $q\:EX^*\to(S^1)^4$ is equivariant with
respect to the $G$-action on $(S^1)^4$ given by
\begin{align*}
\lm(z) &= (\ov{z_2},z_1,\ov{z_4},z_3) \\
\mu(z) &= (z_3,\ov{z_4},z_1,\ov{z_2}) \\
\nu(z) &= (z_1,\ov{z_2},z_3,\ov{z_4}).
\end{align*}
Moreover, the induced map
\[ q_* \: H_1EX^* \to H_1((S^1)^4) = \Z^4 \]
is the same as the isomorphism $\psi$ from
Proposition~\ref{prop-homology}.
\end{proposition}
\begin{proof}
First, using the formulae
\begin{align*}
\nu(x) &= (x_1,-x_2,x_3,x_4) \\
\lm^2(x) &= (-x_1,-x_2,x_3,x_4)
\end{align*}
We see that $\tau_1(\nu(x))=\tau_1(x)$
and $\tau_1(\lm^2(x))=\ov{\tau_1(x)}=\tau_1(x)^{-1}$. Using this and
the structure of $G$ we deduce that $q$ is equivariant. Next, recall
that the classes $\{[c_k]\st 5\leq k\leq 8\}$ give a basis for
$H_1EX^*$, whereas the inclusions of the axes give a basis
$\{e_k\st 1\leq k\leq 4\}$ for $H_1((S^1)^4)$. Recall also that if
$u,v\:S^1\to S^1$ have $|u-v|<2$ everywhere, then $u$ and $v$ are
homotopic (by a straight line homotopy) in $\C^\tm$, so $u$ and $v$
have the same winding numbers. By simplification and plotting, one
can check that
\begin{align*}
|\tau_1(c_5(t))-e^{it}| &\leq 0.14 \\
|\tau_1(c_6(t))-1| &= 0 \\
|\tau_1(c_7(t))+1| &\leq 0.23 \\
|\tau_1(c_8(t))+1| &= 0.
\end{align*}
It follows that the winding numbers of $\tau_1$ composed with
$c_5,\dotsc,c_8$ are $1,0,0,0$. Using the group action, we deduce
that $q_*[c_5]=e_1$, and then that $q_*[c_{4+k}]=e_k$ for
$1\leq k\leq 4$. This proves that $q_*\:H_1EX^*\to H_1((S^1)^4)$ is
an isomorphism.
\begin{checks}
embedded/roothalf/torus_quotients_check.mpl: check_torus_T()
\end{checks}
\end{proof}
\begin{proposition}\label{prop-q-inj}
The map $q\:EX^*\to(S^1)^4$ is injective.
\end{proposition}
\begin{proof}
Let $Q$ denote the image of the map
\[ q^*\:C((S^1)^4,\R) \to C(EX^*,\R), \]
and let $Q^+$ denote the set of strictly positive functions in $Q$.
Note that if $f\in Q^+$ we have $f=f_0\circ q$ for some
$f_0\in C((S^1)^4,\R)$, and by compactness there exists constant
$\ep>0$ such that $f\geq\ep$. If we put
$f_1=\max(\ep,f_0)\in C((S^1)^4,\R)$ then we still have
$f=f_1\circ q$, and from this it is clear that the functions
$1/f=(1/f_1)\circ q$ and $\sqrt{f}=\sqrt{f_1}\circ q$ also lie in
$Q^+$.
We regard $x_1,\dotsc,x_4$ and $y_1,y_2,z_1,z_2,r_1,r_2$ as functions
on $EX^*$; we need to show that they lie in $Q$. We write
\[ q(x) = (u_1+iu_2,u_3+iu_4,u_5+iu_6,u_7+iu_8); \]
this defines elements $u_1,\dotsc,u_8\in Q$. The argument can be
summarised by the following list of equations.
\[
\begin{array}{rclcl}
a_1 &=& (2-u_1u_5-u_2u_6-u_3u_7-u_4u_8)/4 &=& z_1/(2-z_1) \\
z_1 &=& 2a_1/(1+a_1) \in Q \\
a_2 &=& (u_1^2+u_3^2+u_5^2+u_7^2)/4 \\
a_3 &=& (2a_1+a_2)/((1+a_1)(1+2a_1)) &=& 1/(2-z_2) \\
z_2 &=& 2 - 1/a_3 \in Q \\
r_1r_2 &=& \sqrt{1-z_2/2} \in Q^+ \\
r_1+r_2 &=& \sqrt{2(1+r_1r_2)} \in Q^+ \\
a_4 &=& 1-z_2+r_1r_2 \in Q^+ \\
a_5 &=& (u_1+u_3-u_5-u_7)(r_1+r_2)r_1r_2/(2(1+a_1)a_4) &=& y_1 = x_3\\
a_6 &=& (u_3u_5-u_1u_7)r_1r_2(1-z_1/2)/2 &=& -y_1y_2 = x_4 \\
a_7 &=& (1-z_2/2)((u_1^2-u_3^2)(1-y_1/\rt)^2+(u_5^2-u_7^2)(1+y_1/\rt)^2) \\
a_8 &=& (u_1u_5+u_2u_6-u_3u_7-u_4u_8)/(1+a_1) \\
a_9 &=& (a_7+a_8)/(\rt(1+z_1z_2)) &=& y_2 \\
r_1 &=& \sqrt{1-y_2/\rt} \in Q^+ \\
r_2 &=& \sqrt{1+y_2/\rt} \in Q^+ \\
a_{10} &=& u_2r_1(1-y_1/\rt) &=& x_1 \\
a_{11} &=& u_4r_2(1-y_1/\rt) &=& x_2.
\end{array}
\]
The equations with $a_i$ on the left are definitions. In each case
they define $a_i$ in terms of functions that are already known to lie
in $Q$, so $a_i\in Q$. All other equations are claims that can be
verified by straightforward (but sometimes lengthy) calculation in
the ring $AR$. Along the way, we need to verify that certain
denominators are strictly positive. By applying the Cauchy-Schwartz
inequality to the unit vectors $(u_{2i-1},u_{2i})$, we see that $a_1$
takes values in $[0,1]$ (as does $a_2$); this validates the
definition of $a_3$. We know that $0\leq z_1=y_1^2\leq 1$ and
$0\leq z_2=y_2^2\leq 1/2$, and thus that $r_1,r_2>0$; this validates
all other denominators. We also see that $r_1+r_2>0$, and it is
straightforward to check that $(r_1+r_2)^2=2(1+r_1r_2)$, so
$r_1+r_2=\sqrt{2(1+r_1r_2)}$. At the end of the chain of equations we
have seen that the functions $x_i$ all lie in $Q$, and this clearly
implies that $q$ is injective.
\begin{checks}
embedded/roothalf/torus_quotients_check.mpl: check_torus_T()
\end{checks}
\end{proof}
\begin{remark}\label{rem-q-inj}
We can give simpler formulae if we are willing to use denominators
that sometimes vanish. Generically, one can check that
\begin{align*}
x_1 &= -\rt u_2u_6/(u_1u_6+u_2u_5) \\
x_2 &= -\rt u_4u_8/(u_3u_8+u_4u_7) \\
x_3 &= \rt(u_2-u_6)/(u_2+u_6) = \rt(u_4-u_8)/(u_4+u_8) \\
y_2 &= \rt(\al-\bt)/(\al+\bt),
\end{align*}
where
\begin{align*}
\al &= (u_1-u_5)^2(1-u_3u_7-u_4u_8)^2 \\
\bt &= (u_3-u_7)^2(1-u_1u_5-u_2u_6)^2.
\end{align*}
One can then check that $q$ is injective by doing some additional
work to cover the cases where one or more of the above denominators
are zero. However, this is unpleasant.
\end{remark}
\begin{remark}
The proof of Proposition~\ref{prop-q-inj} implicitly gives a map from
the image of $q$ back to $EX^*$. This is implemented in Maple as
\mcode+TTC_to_E+. The simpler function defined in
Remark~\ref{rem-q-inj} is \mcode+TTC_to_E_generic+.
\end{remark}
\begin{definition}\label{defn-qp-qm}
We define $q_+,q_-\:EX^*\to S^1\tm S^1$ by
\begin{align*}
q_+(x) &= ( -\tau_1(x)\tau_3(x),\; -\tau_2(x)\tau_4(x)^{-1}) \\
q_-(x) &= ( -\tau_1(x)\tau_4(x),\; -\tau_2(x)\tau_3(x))
\end{align*}
\end{definition}
In Maple these are \mcode+E_to_TCp+ and \mcode+E_to_TCm+.
In terms of the variables $x_1,x_2,y_1,y_2$ one can check that
\begin{align*}
q_+(x)_1 &= \frac{2ix_1 + y_1^2(1-\rt y_2)/\rt - y_2(1-y_1^2/2)}{
\rt(1-y_1^2/2)(1-y_2/\rt)} \\
q_+(x)_2 &= \frac{2ix_2y_1 + y_1^2(1+\rt y_2) - (1-y_1^2/2)}{
1-y_1^2/2} \\
q_-(x)_1 &= -\frac{(ix_1+y_1(1-y_2/\rt) - 1/\rt)
(ix_2-y_1(1+y_2/\rt) - 1/\rt)}{
(1-y_1^2/2)\sqrt{1-y_2^2/2}} \\
q_-(x)_2 &= -\frac{(ix_1-y_1(1-y_2/\rt) - 1/\rt)
(ix_2+y_1(1+y_2/\rt) - 1/\rt)}{
(1-y_1^2/2)\sqrt{1-y_2^2/2}}.
\end{align*}
We will prove the following result:
\begin{proposition}\label{prop-qp-qm}
The map $q_+$ induces a homeomorphism $EX^*/\ip{\mu}\to(S^1)^2$,
and the map $q_-$ induces a homeomorphism $EX^*/\ip{\lm\mu}\to(S^1)^2$.
\end{proposition}
We can get most of the way by a fairly straightforward argument. We
will show that the Jacobian of $q_+$ (suitably interpreted) is
strictly positive away from the fixed points of $\mu$, and that the
Jacobian of $q_-$ is strictly positive away from the fixed points of
$\lm\mu$. If the Jacobian of $q_+$ was strictly positive everywhere,
we would be able to conclude that $q_+$ was a covering map, and
everything would follow quite easily from the general theory of
coverings. In reality we have only a branched covering, and we do not
have complex structures with respect to which $q_+$ is conformal, so
we cannot use the analytic theory of branched coverings. We will need
some digressions to deal with this.
We first define the version of the Jacobian which we will use.
\begin{definition}
Suppose we have a smooth map $u\:EX^*\to S^1$. We then have a real
vector field
\[ D(u)=(\nabla u)/(iu) = \nabla(\text{arg}(u)) \]
on $EX^*$. Now suppose we have a smooth map $u\:EX^*\to(S^1)^2$. We
then define $\tj(u)\:EX^*\to\R$ by
$\tj(u)(x)=\det(x,n(x),D(u_1),D(u_2))$, and we call this the
\emph{Jacobian} of $u$.
\end{definition}
\begin{remark}\label{rem-jacobian-formula}
By a tiny adaptation of Lemma~\ref{lem-jacobian}, we see that the
induced map $u_*\:T_xEX^*\to T_{u(x)}(S^1)^2$ is an isomorphism
provided that $\tj(u)(x)\neq 0$, and that it preserves orientations
provided that $\tj(u)(x)>0$. If we have an expression for $u_i$ as a
function of the variables $x_i$, then we can calculate $\nabla u_i$
by taking the vector of partial derivatives, and then projecting it
orthogonally into the tangent space. However, this orthogonal
projection will just alter our vector by multiples of $x$ and $n(x)$,
and this will leave the determinant $\tj(u)(x)$ unchanged. Thus, we
can just work with the original vector of partial derivatives.
\end{remark}
\begin{remark}
Suppose that $u=v+iw$ with $v^2+w^2=1$, and put
$u^*=(1-v)/w\:EX^*\to\R_\infty$, as in Remark~\ref{rem-q-denom}.
Differentiating the relation $v^2+w^2=1$ gives
$v\nabla(v)+w\nabla(w)=0$. Using this one can check that
\[ D(u) = \frac{\nabla u}{iu} =
\frac{2\nabla u^*}{1+(u^*)^2}.
\]
This form is sometimes easier to use.
\end{remark}
\begin{proposition}\label{prop-qp-J}
The Jacobian of $q_+$ is
\[ \frac{4\rt(1-x_1^2)}{(1-y_1^2/2)(1-y_2/\rt)}. \]
This is zero at the points $v_2=(1,0,0,0)$ and $v_4=-v_2$
(which are precisely the fixed points of $\mu$). It is strictly
positive everywhere else in $EX^*$.
\end{proposition}
\begin{proof}
The formula can be checked by computer calculation following the
recipe described above. (It is somewhat miraculous that the final
answer is so simple, as the intermediate calculations are enormous.)
It is clear from the formula that the Jacobian vanishes iff
$x_1=\pm 1$, which forces $x_2=x_3=x_4=0$ because
$\sum_ix_i^2=\rho(x)=1$.
\begin{checks}
embedded/roothalf/torus_quotients_check.mpl: check_torus_jacobian()
\end{checks}
\end{proof}
\begin{proposition}\label{prop-qm-J}
The Jacobian of $q^*_-$ is
\[ 4\rt\frac{3/2-(1-y_1^2/2)(1-y_2^2/2)-x_1x_2}{(1-y_1^2/2)(1-y_2^2/2)}.
\]
This is zero at the points $v_6=(1,1,0,0)/\rt$ and $v_8=-v_6$ (which
are precisely the fixed points of $\lm\mu$). It is strictly positive
everywhere else in $EX^*$.
\end{proposition}
\begin{proof}
The formula for the Jacobian can be checked by computer calculation
following the recipe described above. The conclusion is that $j$ is
a positive multiple of $a-x_1x_2$, where
$a=3/2-(1-y_1^2/2)(1-y_2^2/2)$. Now $x_1^2$ and $x_2^2$ can be
rewritten as polynomials in $y_1$ and $y_2$, and using this we obtain
\[ a^2-(x_1x_2)^2 = y_2^2(1+y_2^2/4) +
y_1^2(1-y_2^2)(1+y_2^2/4) +
\tfrac{3}{8}y_1^4y_2^2(1-y_2^2/2).
\]
It is visible that the right hand side is nonnegative, and it
vanishes only where $y_1=y_2=0$. It is easy to see that the only
points with these values of $y$ are $v_6,v_7,v_8$ and $v_9$. By
going back to the original formula, we see that the Jacobian is zero
at $v_6$ and $v_8$, but $8\rt$ at $v_7$ and $v_9$. Moreover, the
Jacobian is nowhere zero on the path-connected space
$EX^*\sm\{v_6,v_8\}$, so it cannot change sign; it is positive at
$v_7$, so it must be positive everywhere.
\begin{checks}
embedded/roothalf/torus_quotients_check.mpl: check_torus_jacobian()
\end{checks}
\end{proof}
We next need to understand the preimages of a few points under the
maps $q_+$ and $q_-$.
\begin{proposition}\label{prop-q-preimages}
\begin{align*}
q_+^{-1}\{(\pp 1,\pp 1)\} &= \{v_0,v_1\} \\
q_+^{-1}\{((1+2\rt i)/3,-1)\} &= \{v_2\} \\
q_+^{-1}\{((1-2\rt i)/3,-1)\} &= \{v_4\} \\
q_-^{-1}\{(\pp 1,\pp 1)\} &= \{v_0,v_1\} \\
q_-^{-1}\{(\pp i,\pp i)\} &= \{v_6\} \\
q_-^{-1}\{( -i, -i)\} &= \{v_8\}.
\end{align*}
\end{proposition}
\begin{proof}
First recall that
\[ q_+(x)_2 =
\frac{2ix_2y_1 + y_1^2(1+\rt y_2) - (1-y_1^2/2)}{1-y_1^2/2}.
\]
By inspecting the imaginary part, we see that $q_+(x)$ can only be
equal to $(1,1)$ or $((1+2\rt i)/3,-1)$ if $x_2y_1=0$. The results
in Section~\ref{sec-E-curves} show that this is only possible
if $x\in C_0\cup C_4\cup C_5\cup C_7$. One can check from the
definitions that
\begin{align*}
q_+(c_{ 0}(t)) &=
\left(\frac{\cos(t)+i/\rt}{\cos(t)-i/\rt},\;-1\right) \\
q_+(c_{ 4}(t)) &=
\left(\frac{i\rt+\sin(t)}{i\rt-\sin(t)} ,\; -1\right) \\
q_+(c_{ 5}(t)) = q_+(c_{ 7}(t)) &=
\left( \frac{\sin(t)^2+8\cos(t)+4\sin(t)\sqrt{5-\cos(t)}i}{9-\cos(t)^2},\; 1\right).
\end{align*}
This gives
\[
\text{Im}(q_+(c_5(t))_1) =
\text{Im}(q_+(c_7(t))_1) =
\frac{4\sin(t)\sqrt{5-\cos(t)}}{9-\cos(t)^2}.
\]
It follows easily that
\[ q_+^{-1}\{(1,1)\} \sse \{c_5(0),c_5(\pi),c_7(0),c_7(\pi)\} =
\{v_0,v_1,v_{10},v_{11}\}.
\]
By inspecting the definitions, we find that $q_+(v_0)=q_+(v_1)=(1,1)$
but $q_+(v_{10})=(1,-1)$ and $q_+(v_{11})=(-1,1)$. It follows that
$q_+^{-1}\{(1,1)\}=\{v_0,v_1\}$ as claimed. Similarly, if
$q_+(x)=((1+2\rt i)/3,-1)$ then we must have $x=c_0(s)$ for some $s$,
or $x=c_4(t)$ for some $t$. Solving
$(\cos(s)+i/\rt)/(\cos(s)-i/\rt)=(1+2\rt i)/3$ gives $\cos(s)=1$, and
solving $(i\rt+\sin(t))/(i\rt-\sin(t))=(1+2\rt i)/3$ gives
$\sin(s)=-1$. We must therefore have $x=c_0(0)$ or $x=c_4(-\pi/2)$,
and both of these are equal to $v_2$ as expected. A very similar
argument gives $q_+^{-1}\{((1+2\rt i)/3,-1)\}=\{v_4\}$.
Next, if $q_-(x)$ is $(1,1)$ or $(i,i)$ or $(-i,-i)$ then we have
$q_-(x)_1-q_-(x)_2=0$. One can check from the definitions that
\[ q_-(x)_1 - q_-(x)_2 =
2\frac{(x_1y_1(1+y_2/\rt)-x_2y_1(1-y_2/\rt))i-y_1y_2}{
(1-y_1^2/2)\sqrt{1-y_2^2/2}}.
\]
By inspecting the real part, we see that $y_1y_2=0$, but
$y_1y_2=-x_4$, so Proposition~\ref{prop-slices} tells us that
$x\in C_0\cup C_1\cup C_2$. Now put
\begin{align*}
m_0(t) &=
\frac{i \rt (\sin(t)+\cos(t))+\sin(2t)-1}{\sqrt{4-\cos(2 t)^2}} \\
m_1(t) &=
\frac{i - \sin(t)}{i + \sin(t)} \\
m_2(t) &=
\frac{i \rt \cos(t)+2 \sin(t)}{i \rt \cos(t)-2 \sin(t)}.
\end{align*}
One can directly that $q_-(c_k(t))=(m_k(t),m_k(t))$
for $k=0,1$, but $q_-(c_2(t))=(m_2(t),\ov{m_2(t)})$. We therefore
need to solve $m_k(t)=1$ and $m_k(t)=\pm i$.
It is easy to see that $\text{Re}(m_0(t))\leq 0$ for all $t$, so
$m_0^{-1}\{1\}=\emptyset$. It is also easy to see that
$m_1^{-1}\{1\}=m_2^{-1}\{1\}=\{0,\pi\}$, so
\[ q_+^{-1}\{(1,1)\} = \{c_1(0),c_1(\pi),c_2(0),c_2(\pi)\}
= \{v_0,v_1,v_1,v_0\} = \{v_0,v_1\}
\]
as expected.
Next, for $m_0(t)=\pm i$ we need $\text{Re}(m_0(t))=0$, which gives
$\sin(2t)=-1$, so $t=\pi/4\pmod{\pi}$. In fact we have
$m_0(\pi/4)=i$ and $m_0(5\pi/4)=-i$, whereas $c_0(\pi/4)=v_6$ and
$c_0(5\pi/4)=v_8$. Similarly, for $m_1(t)=\pm i$ we need
$\sin(t)=\pm 1$, so $t=\pm\pi/2\pmod{2\pi}$, whereas $c_1(\pi/2)=v_6$
and $c_1(-\pi/2)=v_8$. On the other hand, the relation
$q_-(c_2(t))=(m_2(t),\ov{m_2(t)})$ shows that $q_-(c_2(t))$ can never
be equal to $(i,i)$ or $(-i,-i)$. Putting this together, we see that
$q_-^{-1}\{(i,i)\}=\{v_6\}$ and $q_-^{-1}\{(-i,-i)\}=\{v_8\}$, as
expected.
\end{proof}
\begin{proof}[Proof of Proposition~\ref{prop-qp-qm}]
First, it is straightforward to check that $q_+\mu=q_+$, so that
$q_+$ induces a map $EX^*/\ip{\mu}\to(S^1)^2$.
Next, put
\begin{align*}
w_2 = q_+(v_2) &= ((1+2\rt i)/3,-1) \\
w_4 = q_+(v_4) &= ((1-2\rt i)/3,-1).
\end{align*}
For any $u\in(S^1)^2\sm\{w_2,w_4\}$, Proposition~\ref{prop-qp-J}
tells us that $q_+^{-1}\{u\}$ consists of points where the Jacobian
is strictly positive. It follows (using the standard theory of
degrees of maps of compact oriented manifolds) that the set
$q_+^{-1}\{u\}$ is finite, of cardinality equal to the degree of
$q_+$. This cardinality is two in the case $u=(1,1)$, so it must be
two for all $u\not\in\{w_2,w_4\}$. In these cases $q_+^{-1}\{u\}$ is
contained in the set $EX^*\sm\{v_2,v_4\}$ where $\mu$ acts freely,
so $q_+^{-1}\{u\}$ must be a $\mu$-orbit. Moreover, if $u=w_2$ or
$u=w_4$ then Proposition~\ref{prop-q-preimages} again tells us that
$q_+^{-1}\{u\}$ is a (singleton) $\mu$-orbit. It follows that the
induced map $EX^*/\ip{\mu}\to(S^1)^2$ is a continuous bijection, and
thus a homeomorphism (because the domain and codomain are compact and
Hausdorff).
The proof for $q_-$ is essentially the same.
\end{proof}
\begin{remark}
The Maple code contains a formula for the inverse of the map
$q_+\:EX^*/\ip{\mu}\to(S^1)^2$. It also contains a method for
computing the inverse of the map $q_-\:EX^*/\ip{\lm\mu}\to(S^1)^2$,
which is not quite a formula because it involves solutions of a
polynomial of degree four in one variable. These are given by the
functions \mcode+TCp_to_E+ and \mcode+TCm_to_E+, defined in
\fname+embedded/roothalf/torus_quotients.mpl+.
\end{remark}
\begin{remark}\label{rem-not-smooth}
Recall from Remark~\ref{rem-smooth-branch} that the smooth structures
on $EX^*/\ip{\mu}$ and $EX^*/\ip{\lm\mu}$ are subtle, so we cannot
assume that the induced maps $EX^*/\ip{\mu}\to(S^1)^2$ and
$EX^*/\ip{\lm\mu}\to(S^1)^2$ are smooth. In fact, one can check that
they are not. To do this, we need a chart $\phi$ centred at the
branch point $v_2$ as in Section~\ref{sec-roothalf-charts}. One of
the relevant functions is only implemented for vertices in $F_{16}$, so
we find a chart centred at $v_3$ and apply $\lm^{-1}$:
\begin{mcodeblock}
C := `new/E_chart`():
C["vertex_set_exact",3]:
C["curve_set_degree_exact",11]:
x0 := act_R4[LLL](C["p"]([t,u])):
s0 := simplify(multi_series(E_to_TCp(x0)[1],7,t,u));
\end{mcodeblock}
This sets \mcode+s0+ to a Taylor approximation to $q_+(\phi(t,u))_1$.
If $q_+$ was smooth, it is not hard to see that \mcode+s0+ would be
expressible as a polynomial in the quantities
\begin{align*}
m &= \text{Re}((t+iu)^2) = t^2 - u^2 \\
n &= \text{Im}((t+iu)^2) = 2tu,
\end{align*}
and thus that the coefficients of $t^6u^0$ and $t^0u^6$ in \mcode+s0+
would be negatives of each other. However, the above calculation
gives the real parts of the relevant coefficients as $-8/405$ and
$-8/243$, so the map is not in fact smooth. The same argument works
for $q_-$, using a chart based at $v_6$, but in that case we already
see a contradiction from the coefficients of $t^2u^0$ and $t^0u^2$.
\end{remark}
\subsection{Sphere quotients}
\label{sec-sphere-quotients}
Remark~\ref{rem-p-hat} gives us a canonical conformal isomorphism
$\hp\:EX^*/\ip{\lm^2}\to S^2$, but we do not know an exact formula for
that. However, we will define a different homeomorphism
$m\:EX^*/\ip{\lm^2}\to S^2$ which has many of the same properties as
$\hp$. Specifically, $m$ and $\hp$ are both equivariant for the same
action of $G/\ip{\lm^2}$ on $S^2$, and $m(v_i)=\hp(v_i)$ for
$0\leq i\leq 9$.
\begin{definition}\label{defn-sphere-quotient-a}
For $x\in EX^*$, we put
\begin{align*}
\mt(x) &= \left(\rt (1-y_1^2)y_2,\;2x_1x_2,\;-2y_1\right)/
(1+y_1^2)\in\R^3 \\
s(x) &= z_1^2z_2(\half-z_2)/(1+z_1)^2 \\
m(x) &= \mt(x)/\sqrt{1-s(x)}.
\end{align*}
(Maple notation for $m(x)$ is \mcode+E_to_S2(x)+.)
\end{definition}
\begin{proposition}\label{prop-sphere-quotient-a}
The above formula gives a map $m\:EX^*/\ip{\lm^2}\to S^2$. It
satisfies
\begin{align*}
m(v_0) &= (\pp 0,\pp 0, -1) \\
m(v_1) &= (\pp 0,\pp 0,\pp 1) \\
m(v_2) &= m(v_4) = ( - 1,\pp 0,\pp 0) \\
m(v_3) &= m(v_5) = (\pp 1,\pp 0,\pp 0) \\
m(v_6) &= m(v_8) = (\pp 0,\pp 1,\pp 0) \\
m(v_7) &= m(v_9) = (\pp 0, - 1,\pp 0) \\
m(v_{10}) &= ( - 1,0, -2\sqrt{6})/5 \\
m(v_{11}) &= (\pp 1,0, -2\sqrt{6})/5 \\
m(v_{12}) &= ( - 1,0,\pp 2\sqrt{6})/5 \\
m(v_{13}) &= (\pp 1,0,\pp 2\sqrt{6})/5.
\end{align*}
Moreover, we have
\begin{align*}
m_1(\lm(x)) &= -m_1(x) & m_2(\lm(x)) &= -m_2(x) & m_3(\lm(x)) &= \pp m_3(x) \\
m_1(\mu(x)) &= \pp m_1(x) & m_2(\mu(x)) &= -m_2(x) & m_3(\mu(x)) &= -m_3(x) \\
m_1(\nu(x)) &= \pp m_1(x) & m_2(\nu(x)) &= -m_2(x) & m_3(\nu(x)) &= \pp m_3(x).
\end{align*}
\end{proposition}
\begin{proof}
Follows directly from the definitions.
\begin{checks}
embedded/roothalf/sphere_quotients_check.mpl: check_E_to_S2()
\end{checks}
\end{proof}
\begin{remark}\label{rem-m-tilde}
One can check that the function $s(x)$ is zero at all the points
$v_i$, and on $\bigcup_{i=0}^4C_i$. Moreover, it is nonnegative and
small everywhere, with a maximum value of about $0.0114$. (An exact
expression is recorded as \mcode+E_to_S2_s_max+.) Thus, the simpler
function $\mt(x)$ is a good approximation to $m(x)$.
\end{remark}
\begin{remark}\label{rem-m-a}
Recall that if $EX^*\simeq PX(a)$ then we have
$\hp(v_{11})=(2a,0,a^2-1)/(a^2+1)$. If
$a_0=(\sqrt{3}-\sqrt{2})^2\simeq 0.10102$ then we find that
$(2a_0,0,a_0^2-1)/(a_0^2+1)=(1,0,-2\sqrt{6})/5=m(v_{11})$. Thus, if
we believed that $m$ was close $\hp$ then we would expect that
$EX^*\simeq PX(a)$ for some $a$ that is close to $a_0$. In fact, the
correct value of $a$ is approximately $0.09836$. It is perhaps
surprising that this is so close to $a_0$, as $m$ is quite far from
being conformal.
\end{remark}
\begin{proposition}\label{prop-sphere-quotient-b}
The map $m\:EX^*/\ip{\lm^2}\to S^2$ is a homeomorphism.
\end{proposition}
As with Proposition~\ref{prop-qp-qm}, the main ingredient is the
calculation of the Jacobian of $m$. Here we need a slightly different
version of the Jacobian, because the codomain is $S^2$ rather than
$\R^2$. Suppose that $m(a)=b$, so $m$ gives a linear map
$T_aEX^*\to T_bS^2$. Let $(u,v)$ is an oriented orthonormal basis for
$T_aEX^*$, and let $(u',v')$ is an oriented orthonormal basis for
$T_aS^2$. We can then form the matrix of $m_*$ with respect to these
bases, and $j(m)(a)$ is defined to be the determinant of that matrix.
\begin{lemma}\label{lem-m-J}
Put $m_{i,j}=\partial m_i/\partial x_j$ and $n=\nabla(g)$ and
\[ \tj(m) = -\frac{1}{2}\sum_{\sg\in\Sg_4}\sum_{\tau\in\Sg_3}
\ep(\sg)\ep(\tau) x_{\sg(1)}n_{\sg(2)}
m_{\tau(1),\sg(3)}m_{\tau(2),\sg(4)}m_{\tau(3)}.
\]
Then $j(m)=\tj(m)/\|n\|$.
\end{lemma}
\begin{proof}
Fix a point $x$, and choose orthonormal bases $(u,v)$
and $(u',v')$ for $T_xEX(a)$ and $T_{m(x)}S^2$ as in the definition
of $j(m)$. Note that with
conventions as spelled out in Remark~\ref{rem-p-hat}, the orientation
conditions are that $\det(x,n/\|n\|,u,v)=1$ and $\det(u',v',m)=-1$.
Given this, it is not hard to see that
$j(m)=-\det(m_*(u),m_*(v),m)$. Equivalently, if $\om_d$ denotes
the standard volume form for $\R^d$ and $\bt=u\wedge v$, then
$j(m)$ is characterised by $m_*(\bt)\wedge m=-j(m)\,\om_3$.
Now Lemma~\ref{lem-hodge} tells us that
\[ \bt = \frac{1}{2\|n\|}
\sum_{ijkl}\ep_{ijkl}x_in_je_k\wedge e_l,
\]
so
\begin{align*}
m_*(\bt) &=
\frac{1}{2\|n\|} \sum_{ijkl}\ep_{ijkl}
x_in_jm_*(e_k)\wedge m_*(e_l) \\
&=
\frac{1}{2\|n\|} \sum_{ijklpq}\ep_{ijkl}
x_in_jm_{p,k}m_{q,l}e_p\wedge e_q \\
m_*(\bt)\wedge m &=
\frac{1}{2\|n\|} \sum_{ijklpqr}\ep_{ijkl}
x_in_jm_{p,k}m_{q,l}m_r e_p\wedge e_q\wedge e_r \\
&=
\frac{1}{2\|n\|} \sum_{ijklpqr}\ep_{ijkl}\ep_{pqr}
x_in_jm_{p,k}m_{q,l}m_r \om_3.
\end{align*}
The claim is clear from this.
\end{proof}
\begin{corollary}\label{cor-m-J}
$\tj(m)=8j_1/j_2^{3/2}$, where
\begin{align*}
j_1 &= (1+z_1)((1-z_1)^2-z_1^2z_2/2) + z_1^2z_2(\half-z_2)(3+z_1) \\
j_2 &= 1+2z_1+z_1^2(1-z_2/2)+z_1^2z_2^2 \geq 1.
\end{align*}
Moreover, we have $j_1\geq 0$ everywhere, with $j_1=0$ only if
\[ x \in (EX^*)^{\lm^2} = \{v_0,v_1,v_{10},v_{11},v_{12},v_{13}\}. \]
\end{corollary}
\begin{proof}
The lemma reduces the formula to a direct calculation, which can be
checked by Maple. Next, we can write $j_1=j_3j_4+j_5$, where
\begin{align*}
j_3 &= 1 + z_1 \geq 1 \\
j_4 &= (1-z_1)^2-z_1^2z_2/2 \\
j_5 &= z_1^2z_2(\half-z_2)(3+z_1) \geq 0.
\end{align*}
Another standard calculation gives
\[ (1/2-z_2)j_4 = 2x_1^2x_2^2 \geq 0. \]
Recall that $1/2-z_2\geq 0$ everywhere, and $1/2-z_2>0$ on a dense
subset of $EX^*$; it follows that $j_4\geq 0$ everywhere. It follows
that $j_1\geq 0$ everywhere, and that if $j_1=0$ then we must have
$j_4=j_5=0$. Note that $j_4=1$ when $z_1=0$, so we must have
$z_1>0$ (and also $z_1\leq 1$ as always). This lets us rearrange
$j_4=0$ to give $z_2=2(z_1^{-1}-1)^2$. After substituting this in
the relation $j_5=0$ we see that $z_1\in\{2/3,1\}$ and so
$z\in\{(2/3,1/2),(1,0)\}$. Each point in the $z$-plane corresponds
to a $G$-orbit in $EX^*$, and it is straightforward to check that the
relevant $G$-orbits are $\{v_{10},v_{11},v_{12},v_{13}\}$ and
$\{v_0,v_1\}$.
\begin{checks}
embedded/roothalf/sphere_quotients_check.mpl: check_E_to_S2()
\end{checks}
\end{proof}
\begin{lemma}\label{lem-sphere-quotient-preimages}
For all $i$ we have $m^{-1}\{m(v_i)\}=\{v_i,\lm^2(v_i)\}$. In
particular, for $i\in\{0,1,10,11,12,13\}$, we have
$\lm^2(v_i)=v_i$ and $m^{-1}\{m(v_i)\}=\{v_i\}$.
\end{lemma}
\begin{proof}
Recall that formulae for $m(v_i)$ were given in
Proposition~\ref{prop-sphere-quotient-a}.
Now suppose that $x\in EX^*$ with $m(x)=u$.
\begin{itemize}
\item[(a)] We have $u_1=0$ iff $(1-y_1^2)y_2=0$ iff $x_3=y_1=\pm 1$
or $y_2=0$. Note that if $x_3=\pm 1$ we must have
$x\in\{e_3,-e_3\}=\{v_0,v_1\}$. On the other hand, we have $y_2=0$
iff $x\in C_1\cup C_2$.
\item[(b)] We have $u_2=0$ iff $x_1=0$ or $x_2=0$, which means that
$x\in\bigcup_{i=3}^8C_i$.
\item[(c)] We have $u_3=0$ iff $x_3=0$ iff $x\in C_0$.
\end{itemize}
If $u=m(v_i)$ for some $i<10$ then two of the coordinates $u_p$ are
zero, and it follows that $x\in C_r\cap C_s$ for some $r\neq s$, so
$x=v_j$ for some $j$. A check of cases then shows that
$x\in\{v_i,\lm^2(v_i)\}$. Suppose instead that
\[ m(x) = m(v_{11})=(1,0,-2\sqrt{6})/5. \]
As $m(x)_2=0$ we have $x_1x_2=0$ and so $x\in\bigcup_{i=3}^8C_i$.
From the form of $m(x)_1$ and $m(x)_3$ it is also clear that
$y_1,y_2>0$. Also, we have
\[ \frac{\rt(1-y_1^2)y_2}{2y_1} = -\frac{m(x)_1}{m(x)_3} =
\frac{1}{2\sqrt{6}},
\]
so $y_2=(2\sqrt{3}(y_1^{-1}-y_1))^{-1}$. Substituting this in the
relation $m(x)_1^2=1/25$ and factoring leads to a relation
\[ (2y_1^2-3)(3y_1^2-2)^2(8y_1^6-27y_1^4+30y_1^2-12) = 0. \]
One can check that the only root in the required range $0<y_1\leq 1$
is $y_1=\sqrt{2/3}$, and this in turn gives
$y_2=(2\sqrt{3}(y_1^{-1}-y_1))^{-1}=1/\rt$. This gives
$(x_3,x_4)=(y_1,-y_1y_2)=(\sqrt{2/3},-\sqrt{1/3})$, and as
$x_3^2+x_4^2=1$ we must have $x_1=x_2=0$, so $x=v_{11}$ as required.
The remaining cases $i\in\{10,12,13\}$ now follow using the group
action.
\end{proof}
\begin{proof}[Proof of Proposition~\ref{prop-sphere-quotient-b}]
Put
\[ U=S^2\sm\{m(v_i)\st i\in\{0,1,10,11,12,13\}\}. \]
We now see that for $u\in U$, the preimage $m^{-1}\{u\}$ contains
only points where the Jacobian of $m$ is strictly positive, so the
number of points is equal to the degree of $m$. Taking $u=m(v_2)$ we
see that $m^{-1}\{u\}=\{v_2,\lm^2(v_2)\}=\{v_2,v_4\}$, so the degree
is two. It follows that for all $u\in U$, the preimage consists of
two points and is closed under the action of $\lm^2$. The only
points fixed by $\lm^2$ are
$\{v_0,v_1,v_{10},v_{11},v_{12},v_{13}\}$, and these cannot lie in
$m^{-1}\{u\}$, so $m^{-1}\{u\}$ must consist of a single
$\lm^2$-orbit. The same holds by
Lemma~\ref{lem-sphere-quotient-preimages} in the exceptional cases
where $u\in S^2\sm U$. It follows that the induced map
$EX^*/\ip{\lm^2}\to S^2$ is a continuous bijection of compact
Hausdorff spaces, so it is a homeomorphism.
\end{proof}
\begin{remark}
Again put
\[ U=S^2\sm\{m(v_i)\st i\in\{0,1,10,11,12,13\}\}. \]
In Section~\ref{sec-energy} we will explain how to define a function
$u=C_+(Dm)/(C_+(Dm)+C_-(Dm))\:U\to [0,1]$ which is zero at points
where $m$ is conformal, and one at points where $m$ is anticonformal.
We have not found a formula for $u$, but it is not hard to
evaluate it at any given point. We find that $0\leq u<1/2$
everywhere in $U$, with $u(x)\to 1/2$ as $x\to v_0$ or $x\to v_1$.
For $i\in\{10,11,12,13\}$ we find that $u$ oscillates between about
$0.01$ and $0.08$ on any small circle surrounding $v_i$, so it does
not extend continuously to $v_i$. On about $90\%$ of the area of
$EX^*$ we have $u < 0.075$.
\end{remark}
There are two other maps $EX^*\to S^2$ that have natural geometric
descriptions, so it is reasonable to ask whether they are related to
$\hp$. The first is the Hopf fibration $\eta\:S^3\to S^2$, which we
can restrict to $EX^*$. (In Maple this is \mcode+hopf_map+, which
is defined in \fname+Rn.mpl+.) Note that $H_2(EX^*)\simeq
H_2(S^2)\simeq\Z$, but $H_2(S^3)=0$, which implies that the map
$\eta_*\:H_2(EX^*)\to H_2(S^2)$ is zero. In other words, the
restricted map $\eta\:EX^*\to S^2$ has degree zero, whereas $\hp$ has
degree $2$, and any other nonconstant conformal map has strictly
positive degree. This means that $\eta$ cannot be closely related to
$\hp$.
\begin{checks}
Rn_check.mpl: check_hopf_map()
\end{checks}
The second possibility is a variant of the Gauss map. We can identify
$\R^4$ with the algebra $\H$ of quaternions, and $\R^3$ with the
subspace $\H_0$ of purely imaginary quaternions. This allows us to
interpret conjugation and multiplication of elements of $\R^4$. For
$x\in X$ we recall that $n(x)$ is orthogonal to $x$, so the element
$\gm(x)=n(x)\ov{x}/\|n(x)\|$ is a purely imaginary quaternion of norm
one. This defines a map $\gm\:X\to S^2$ (which is \mcode+gauss_map+
in Maple). It depends on our conventions for identifying $\R^4$ with
$\H$, but the dependence is easy to analyse. It is known that every
special orthogonal automorphism of $\H$ has the form
$x\mapsto ux\ov{v}$ for some $u,v\in S^3$, and every special
orthogonal automorphism of $\H_0$ has the form $x\mapsto wx\ov{w}$.
Changing conventions will replace $\gm$ by a map of the form
$\gm'(x)=w\gm(ux\ov{v})\ov{w}$. By choosing paths in $S^3$ from $u$,
$v$ and $w$ to the identity, we can produce a homotopy between $\gm$
and $\gm'$, so they at least have the same degree. One can calculate
the degree by counting preimages of a regular value, with signs
determined by the orientation behaviour. One can check that
$\gm^{-1}\{(1,1,0)/\rt\}=\{v_6\}$, and that $\gm$ gives an
orientation-reversing isomorphism of tangent spaces at this point. It
follows that $\gm$ has degree $-1$, and cannot be homotopic to $\hp$.
\begin{checks}
embedded/geometry_check.mpl: check_gauss_map();
\end{checks}
\subsection{Rational points}
\label{sec-rational}
In this section we study points in $EX^*$ where the coordinates are
rational or lie in some small extension of $\Q$. As well as being
interesting for its own sake, it is useful to have a supply of points
where we can easily do exact calculations rather than relying on
numerical approximation.
\begin{proposition}\label{prop-rational}
The set $EX^*(\Q)=EX^*\cap\Q^4$ is as follows:
\begin{itemize}
\item[(a)] For every $t\in\Q$, the point $(t^2-1,2t,0,0)/(1+t^2)$ lies in
$EX^*(\Q)\cap C_0$.
\item[(b)] The point $v_2=(1,0,0,0)$ (corresponding to $t=\infty$)
also lies in $EX^*(\Q)\cap C_0$.
\item[(c)] For any pair $(s,t)\in\Q^2$ with $2s^2+t^2=1$, the point
$(s,s,t,0)$ lies in $EX^*(\Q)\cap C_1$, and the point $(s,-s,t,0)$ lies
in $EX^*(\Q)\cap C_2$.
\item[(d)] All points in $EX^*(\Q)$ are of type~(a), (b) or~(c).
\end{itemize}
\end{proposition}
\begin{proof}
Recall that the cubic defining equation is
\[ (\tfrac{3}{2}x_3^2-2)x_4+(x_1^2-x_2^2)x_3/\rt = 0. \]
Note here that $1$ and $1/\sqrt{2}$ are linearly independent over
$\Q$, and also that $\tfrac{3}{2}x_3^2-2\neq 0$ for $x_3\in\Q$.
It follows that all rational solutions have $x_4=0$ and (either
$x_3=0$ or $x_1=\pm x_2$). If $x_3=x_4=0$ then $x_1^2+x_2^2=1$. If
$x_1=1$ then we have case~(a), otherwise we have case~(b) with
$t=x_2/(1-x_1)$. If $x_4=0$ and $x_1=\pm x_2$ then we have
case~(c).
\end{proof}
The pairs $(s,t)$ as in~(c) are well-understood in terms of the
arithmetic of the field $\Q(\sqrt{-2})$, as we now recall briefly.
For any element $x=t+s\sqrt{-2}\in\Q(\sqrt{-2})$, we put
\[ N(x) = |x|^2 = 2s^2+t^2 \in \Q. \]
This gives a homomorphism $\Q(\sqrt{-2})^\tm\to\Q^\tm$, and
$EX^*(\Q)\cap C_1$ bijects with $\ker(N)$. If $p$ is a prime
congruent to $1$ or $3$ mod $8$, then it is well-known that there is a
unique pair of positive integers $(a,b)$ such that $2a^2+b^2=p$. We
put $\pi_p=b+a\sqrt{-2}$ and
\[ u_p=\frac{\pi_p}{\ov{\pi_p}} = ((b^2-2a^2)+2ab\sqrt{-2})/p. \]
Standard methods of algebraic number theory show that $\ker(N)$ is the
product of $\{\pm 1\}$ with the free abelian group generated by these
elements $u_p$. Moreover, the denominator of a product
$\prod_pu_p^{n_p}$ is $\prod_pp^{|n_p|}$. Using this, we can
enumerate all the solutions for which the denominator is less than
some specified bound, and thus produce rational points that are
closely spaced around $C_1$ and $C_2$.
\begin{remark}
The map in part~(a) of Proposition~\ref{prop-rational} is
\mcode+c_rational[0](t)+, and the map in part~(c) is
\mcode+c_rational[1]([s,t])+. The function \mcode+two_circle(n)+
returns the list of all pairs $(s,t)\in\Q^2$ where $2s^2+t^2=1$ and
the denominators of $s$ and $t$ are less than or equal to $n$.
\end{remark}
As the rational points do not cover much of $EX^*$, we instead
consider the set
\[ Q = \{x\in EX^*\st x_1,x_2,x_4,\;x_3/\rt\in\Q\}. \]
This set is again arithmetically simple, but has a richer structure
than $EX^*(\Q)$. We call the points in $Q$ \emph{quasirational}.
We first consider quasirational points which have nontrivial isotropy
(and so lie in $\bigcup_{i=0}^8C_i$).
\begin{proposition}\label{prop-quasirational-isotropy}
\begin{itemize}
\item[(a)] For every $t\in\Q\cup\{\infty\}$ we have a point
\[ \left(t^2-1,2t,0,0\right)/(1+t^2) \in Q\cap C_0 = EX^*(\Q)\cap C_0. \]
\item[(b)] For every $t\in\Q\cup\{\infty\}$ we have points
\begin{align*}
\left(1-t^2-2t,\;1-t^2-2t,\;\rt(1-t^2+2t),0\right)/(2(1+t^2))
& \in Q\cap C_1 \\
\left(-(1-t^2-2t),\;1-t^2-2t,\;\rt(1-t^2+2t),0\right)/(2(1+t^2))
& \in Q\cap C_2.
\end{align*}
\item[(c)] For each $(s,t)\in\Q^2$ with $3s^2+t^2=1$, we have points
\begin{align*}
\left(0,t,\rt s,-s\right) &\in Q\cap C_3 \\
\left(-t,0,\rt s,s\right) &\in Q\cap C_4.
\end{align*}
\item[(d)] All quasirational points with nontrivial isotropy are
accounted for by~(a) to~(c). In particular, we have
$Q\cap\bigcup_{i=5}^8C_i=\emptyset$.
\end{itemize}
\end{proposition}
\begin{proof}
First, it is straightforward to check that the constructions in~(a)
to~(c) do in fact give quasirational points on the indicated curves.
Next, points in $C_0$ have $x_3=0$ so they are quasirational if and
only if they are rational. Thus, we have seen already that all
quasirational points on $C_0$ are as in~(a).
Now let $x$ be a quasirational point in $C_1$. We then have
$x_1=x_2=m$ and $x_3=\rt n$ and $x_4=0$ for some rational numbers $m$
and $n$. Put $p=n-m\in\Q$ and $q=n+m\in\Q$ so
\[ x = \left((q-p)/2,\;(q-p)/2,\;(q+p)/\rt,\;0\right). \]
For such points we have $g(x)=0$ automatically, and
$\rho(x)=p^2+q^2$. It follows that $(p,q)=(2t,1-t^2)/(1+t^2)$ for
some $t\in\Q\cup\{\infty\}$, and we can use this to get the first
formula in~(b). The $C_2$ case follows from the $C_1$ case by the
group action.
Next, one can check that (in the current case where $a=1/\rt$) we
have
\[ g(x_1,0,x_3,x_4) = \rt(x_3-\rt x_4)(x_1^2+x_4^2+x_3x_4/\rt). \]
The first factor vanishes on $C_4$. The claim about $C_4$ in~(c)
follows easily from this, and the claim about $C_3$ can be deduced
using the group action.
Now consider a point $x\in C_5$, so $x_2=0$ and the functions
$r_0=x_1^2+x_3^2+x_4^2-1$ and $r_1=x_1^2+x_4^2+x_3x_4/\rt$ also
vanish at $x$. (Here $r_1$ is the second term in the above
factorisation of $g$.) Now put $t=x_1^2$ and
\[ u = x_1(x_4^2-5x_3x_4/\rt-1). \]
We claim that $u^2=t^3-10t^2+t$. In fact, one can check by direct
expansion that
\[ u^2 - (t^3-10t^2+t) = a_0r_0 + a_1r_1 + a_2r_0r_1, \]
where
\begin{align*}
a_0 &= -6 \rt (x_3+\rt x_4) x_4^3 \\
a_1 &= (1-x_3^2-x_4^2) (x_3^2-10 x_4^2+9+x_3 x_4/\rt) \\
a_2 &= x_3^2+2 x_4^2+x_3 x_4/\rt-x_1^2+9.
\end{align*}
As $r_0=r_1=0$, the claim follows. Note also that if $x$ is
quasirational then $t$ and $u$ will be rational. Now, the equation
$u^2=t^3-10t^2+t$ describes an elliptic curve $E$ over $\Q$, and
algorithms to determine rational points on such curves are built in
to the symbolic mathematics system Sage. Our curve can be
described in extended Weierstrass form as
\[ u^2 + \al_1ut + \al_3u = t^3 + \al_2t^2 + \al_4t + \al_6, \]
where
\[ (\al_1,\al_2,\al_3,\al_4,\al_6) = (0,-10,0,1,0). \]
We can thus enter the following in Sage:
\begin{mcodeblock}
E = EllipticCurve([0,-10,0,1,0])
E.cremona_label()
E.rank()
E.torsion_points()
\end{mcodeblock}
We learn that $E$ is isomorphic to the curve labelled 96b2 in
Cremona's database of elliptic curves~\cite{cr:ecd}, and that the
rank is zero, so the rational points are all torsion points. We also
learn that the only rational torsion points in the projective closure
are $[0:0:1]$ and $[0:1:0]$, so the only rational point on the
original affine curve is $(0,0)$. Thus, if $x$ is quasirational then
we must have $t=u=0$, but that gives $x_1=0$. Substituting $x_1=0$
in the relation $r_1=0$ gives $x_4(x_4+x_3/\rt)=0$, so $x_4=0$ or
$x_4=-x_3/\rt$. Putting $x_1=x_4=0$ in $r_0$ gives $x_3^2=1$;
putting $x_1=0$ and $x_4=-x_3/\rt$ instead gives $x_3^2=2/3$.
Neither of these is possible with $x_3/\rt\in\Q$, so we see that
there are no quasirational points in $C_5$. It follows using the
group action that there are no quasirational points in
$\bigcup_{i=5}^8C_i$.
\begin{checks}
embedded/roothalf/rational_check.mpl: check_rational_elliptic()
\end{checks}
\end{proof}
We can understand rational solutions to $3s^2+t^2=1$ in terms of the
arithmetic of the field $\Q(\sqrt{-3})$, and thus produce
quasirational points that are closely spaced around $C_3$ and $C_4$.
The story similar to that for $\Q(\sqrt{-2})$, and we will not give
the details here.
\begin{remark}
The map in part~(a) of Proposition~\ref{prop-quasirational-isotropy}
is \mcode+c_rational[0]+, and the maps in~(b) and~(c) are
\mcode+c_quasirational[i]+ for $i\in\{1,2,3,4\}$. The function
\mcode+three_circle(n)+ returns the list of all pairs $(s,t)\in\Q^2$
where $3s^2+t^2=1$ and the denominators of $s$ and $t$ are less than
or equal to $n$.
\end{remark}
We now consider quasirational points with trivial isotropy. We do not
have a very good theory for these, but we have a reasonably efficient
method for exhaustive search, as we now explain.
\begin{proposition}\label{prop-quasirational-lift}
For any $x\in Q\cap F_{16}$, the numbers $s_1=y_1/\rt=x_3/\rt$ and
\[ s_2 = \rt y_2 = x_2^2-x_1^2-3x_3x_4/\rt \]
lie in $\Q\cap [0,1]$. Conversely, suppose that
$s\in(\Q\cap [0,1])^2$, and put
\[ t_1 = s_1^2 \hspace{4em}
t_2 = 6s_1^2 - 2 \hspace{4em}
t_3 = -2s_2^2-4
\]
\[ p_1 = 2+t_1t_3 \hspace{4em}
p_2 = s_2t_2 \hspace{4em}
p_3 = p_1+p_2 \hspace{4em}
p_4 = p_1-p_2.
\]
Then $s$ arises from a quasirational point $x\in F_{16}$ if and only
if $p_3$ and $p_4$ are perfect squares (and therefore nonnegative).
If so, then
\[ x = (\sqrt{p_3}/2,\;\sqrt{p_4}/2,\;\rt s_1,\;-s_1s_2). \]
\end{proposition}
\begin{proof}
This is essentially a reformulation of Proposition~\ref{prop-F-four},
in the special case $a=1/\rt$. The formulae have been reorganised
slightly to make rationality questions more visible, and to allow for
efficient calculation as discussed below.
\end{proof}
We wrote code in C to look for solutions using the above proposition.
We ran the program on a cluster of machines with 64 bit processors.
The native 64 bit integers are not large enough for our intermediate
calculations, but fortunately the GCC compiler provides built in
support for 128 bit integers encoded as pairs of native integers, and
similarly for floating point numbers. (If we needed larger integers
than 128 bits we would need to use an arbitrary precision library,
which would come with a significant performance penalty.) We took
$N=2^{12}=4096$ and enumerated the rational numbers of denominator at
most $N$ as a Farey sequence (of length $5100021\simeq 5\tm 10^6$).
Note that the numbers $t_i$ in
Proposition~\ref{prop-quasirational-lift} depend only on a single
rational number in the sequence, so we precomputed them in a single
pass. We then looped through all possible pairs $(s_1,s_2)$ and
computed the numbers $p_i$. One can check that if $p_1<0$ at
$(s_1,s_2)$, then all pairs $(s'_1,s'_2)$ with $s'_1\geq s_1$ and
$s'_2\geq s_2$ can be disregarded. Using this we can cut down the
number of pairs to be considered, but the order of magnitude is still
$10^{13}$, so the remaining steps must be highly optimised.
To check whether a rational number $p=a/b$ (possibly not in lowest
terms) is a perfect square, we first test whether the $2$-adic
valuations of $a$ and $b$ are the same mod $2$. Here the $2$-adic
valuation is the number of trailing zeros in the binary
representation; this can be calculated in a single processor
instruction for $64$ bit integers, and is only slightly slower for
$128$ bit pairs. If this first test is passed, we divide $a$ and $b$
by appropriate powers of $2$ to make them odd. This can be done as a
bitwise shift rather than a division, so again it is very fast. Next,
with the new $a$ and $b$, it is not hard to check that $a/b$ can only
be a square if $a=b\pmod{8}$. This can again be checked by bit
masking rather than division, so it is very fast. If these fast tests
are passed for both $p_3$ and $p_4$ (which is already relatively
rare), we then start using some slower tests. We calculate the gcd of
$a$ and $b$ and then divide by it to make $a$ and $b$ coprime. It
seems that no algorithm is known that is usefully more efficient
than the obvious one, so a significant number of divisions may be
required. Now $a/b$ will be a perfect square if and only if $a$ and
$b$ are individually perfect squares. They are still odd, so we first
test that they are congruent to $1$ mod $8$. If so, we reduce them
modulo $3\tm 5\tm 7\tm 11\tm 13\tm 17=255255$, and look up in a
precomputed table whether the result is a quadratic residue. Only
about $0.63\%$ of numbers pass these modular tests. If we get to this
stage, we calculate the square root as a $128$ bit floating point
number, take the nearest integer, square it, and check whether the
result is the same as the number we first thought of. If so, our
number is obviously a perfect square. The converse is less obvious
because of the possibility of rounding errors. However, one can check
that our numerators and denominators are not much bigger than
$N^7=2^{84}$ so $128$ bit accuracy should be ample to avoid problems.
We found precisely $130$ pairs $s$ corresponding to quasirational
points in the interior of $F_{16}$:
\[ \renewcommand{\arraystretch}{1.5} \setlength{\arraycolsep}{4pt}
\begin{array}{cccccccc}
\left(\frac{ 103}{ 154},\frac{ 65}{ 449}\right) &
\left(\frac{ 103}{ 554},\frac{2945}{3041}\right) &
\left(\frac{ 104}{1073},\frac{ 105}{ 233}\right) &
\left(\frac{ 1}{ 10},\frac{ 13}{ 19}\right) &
\left(\frac{ 1}{ 10},\frac{ 16}{ 17}\right) &
\left(\frac{ 1}{ 10},\frac{ 245}{ 267}\right) &
\left(\frac{ 11}{ 118},\frac{ 819}{1331}\right) &
\left(\frac{ 11}{1190},\frac{1664}{2057}\right)\\
\left(\frac{1127}{2194},\frac{ 581}{ 731}\right) &
\left(\frac{1135}{2114},\frac{1088}{1137}\right) &
\left(\frac{1147}{2390},\frac{ 97}{ 961}\right) &
\left(\frac{1147}{2830},\frac{ 640}{1369}\right) &
\left(\frac{1151}{3722},\frac{ 11}{ 139}\right) &
\left(\frac{1223}{2330},\frac{ 16}{ 17}\right) &
\left(\frac{ 12}{ 29},\frac{ 247}{ 297}\right) &
\left(\frac{ 125}{ 262},\frac{ 16}{ 59}\right)\\
\left(\frac{1255}{2146},\frac{1547}{1675}\right) &
\left(\frac{1260}{3599},\frac{ 559}{1009}\right) &
\left(\frac{ 127}{ 450},\frac{ 416}{ 417}\right) &
\left(\frac{1301}{2270},\frac{ 16}{ 33}\right) &
\left(\frac{ 13}{ 30},\frac{ 145}{1521}\right) &
\left(\frac{ 13}{ 30},\frac{ 16}{ 33}\right) &
\left(\frac{1389}{3038},\frac{ 176}{1049}\right) &
\left(\frac{1424}{3913},\frac{1571}{2179}\right)\\
\left(\frac{ 148}{ 215},\frac{ 29}{ 323}\right) &
\left(\frac{ 151}{ 338},\frac{ 279}{1129}\right) &
\left(\frac{1549}{3038},\frac{2768}{3307}\right) &
\left(\frac{1603}{2350},\frac{ 19}{ 147}\right) &
\left(\frac{1611}{3878},\frac{ 355}{ 643}\right) &
\left(\frac{1648}{2875},\frac{2295}{3113}\right) &
\left(\frac{1672}{3329},\frac{1257}{2057}\right) &
\left(\frac{ 169}{ 322},\frac{ 475}{ 507}\right)\\
\left(\frac{ 17}{ 106},\frac{1760}{2601}\right) &
\left(\frac{ 172}{ 511},\frac{ 15}{ 17}\right) &
\left(\frac{ 175}{ 298},\frac{2464}{2825}\right) &
\left(\frac{ 175}{ 362},\frac{ 112}{ 163}\right) &
\left(\frac{1792}{3635},\frac{2623}{3649}\right) &
\left(\frac{1809}{3050},\frac{ 32}{ 41}\right) &
\left(\frac{1829}{2750},\frac{ 97}{ 961}\right) &
\left(\frac{1864}{2843},\frac{ 145}{2993}\right)\\
\left(\frac{ 191}{3490},\frac{1211}{1243}\right) &
\left(\frac{1939}{3350},\frac{ 19}{ 147}\right) &
\left(\frac{ 196}{ 335},\frac{2719}{3553}\right) &
\left(\frac{1964}{3085},\frac{ 31}{ 97}\right) &
\left(\frac{ 19}{ 70},\frac{ 16}{ 33}\right) &
\left(\frac{2093}{3534},\frac{ 115}{ 147}\right) &
\left(\frac{ 209}{ 338},\frac{ 819}{1331}\right) &
\left(\frac{ 211}{ 350},\frac{ 80}{ 107}\right)\\
\left(\frac{ 23}{ 154},\frac{ 895}{ 993}\right) &
\left(\frac{ 23}{ 34},\frac{ 481}{2881}\right) &
\left(\frac{ 23}{ 50},\frac{ 32}{ 41}\right) &
\left(\frac{2435}{3854},\frac{ 16}{ 59}\right) &
\left(\frac{ 248}{ 401},\frac{ 95}{ 449}\right) &
\left(\frac{ 248}{ 665},\frac{ 29}{ 323}\right) &
\left(\frac{ 256}{ 377},\frac{ 35}{ 387}\right) &
\left(\frac{ 25}{ 82},\frac{ 237}{ 275}\right)\\
\left(\frac{2636}{3887},\frac{ 445}{4003}\right) &
\left(\frac{ 280}{ 457},\frac{1243}{2107}\right) &
\left(\frac{ 28}{ 45},\frac{ 559}{1617}\right) &
\left(\frac{ 307}{ 830},\frac{ 267}{ 779}\right) &
\left(\frac{ 31}{ 202},\frac{ 355}{1507}\right) &
\left(\frac{ 313}{ 466},\frac{ 560}{2651}\right) &
\left(\frac{ 3}{ 14},\frac{ 5}{ 27}\right) &
\left(\frac{ 31}{ 50},\frac{ 355}{1507}\right)\\
\left(\frac{ 316}{ 487},\frac{ 835}{2243}\right) &
\left(\frac{ 32}{ 49},\frac{ 767}{2433}\right) &
\left(\frac{ 341}{1070},\frac{ 97}{ 961}\right) &
\left(\frac{ 359}{ 610},\frac{ 463}{1969}\right) &
\left(\frac{ 364}{ 687},\frac{ 409}{ 441}\right) &
\left(\frac{ 36}{ 77},\frac{2575}{2673}\right) &
\left(\frac{ 367}{ 986},\frac{2176}{3401}\right) &
\left(\frac{ 37}{ 190},\frac{ 688}{1369}\right)\\
\left(\frac{ 373}{ 630},\frac{ 16}{ 33}\right) &
\left(\frac{ 37}{ 70},\frac{ 5}{ 27}\right) &
\left(\frac{ 401}{ 650},\frac{ 16}{1067}\right) &
\left(\frac{ 403}{ 590},\frac{ 97}{ 961}\right) &
\left(\frac{ 404}{ 685},\frac{ 31}{ 97}\right) &
\left(\frac{ 415}{ 706},\frac{ 341}{ 459}\right) &
\left(\frac{ 427}{ 694},\frac{ 128}{ 803}\right) &
\left(\frac{ 43}{ 166},\frac{ 224}{ 251}\right)\\
\left(\frac{ 432}{1681},\frac{1055}{3553}\right) &
\left(\frac{ 43}{ 94},\frac{ 480}{1849}\right) &
\left(\frac{ 4}{ 47},\frac{ 65}{ 193}\right) &
\left(\frac{ 455}{ 778},\frac{ 973}{1075}\right) &
\left(\frac{ 463}{1970},\frac{ 16}{ 17}\right) &
\left(\frac{ 468}{ 775},\frac{ 385}{ 673}\right) &
\left(\frac{ 47}{ 106},\frac{ 80}{ 97}\right) &
\left(\frac{ 4}{ 7},\frac{ 15}{ 17}\right)\\
\left(\frac{ 48}{ 235},\frac{ 107}{ 459}\right) &
\left(\frac{ 501}{ 742},\frac{ 5}{ 27}\right) &
\left(\frac{ 52}{ 205},\frac{ 137}{ 169}\right) &
\left(\frac{ 555}{ 974},\frac{ 163}{ 675}\right) &
\left(\frac{ 560}{3163},\frac{ 951}{1225}\right) &
\left(\frac{ 56}{ 107},\frac{ 605}{ 931}\right) &
\left(\frac{ 569}{1042},\frac{2735}{2993}\right) &
\left(\frac{ 57}{ 130},\frac{ 32}{ 41}\right)\\
\left(\frac{ 577}{1546},\frac{1680}{1873}\right) &
\left(\frac{ 600}{ 913},\frac{ 7}{ 25}\right) &
\left(\frac{ 61}{ 718},\frac{ 581}{ 731}\right) &
\left(\frac{ 65}{ 122},\frac{ 32}{ 507}\right) &
\left(\frac{ 65}{ 122},\frac{ 973}{1075}\right) &
\left(\frac{ 65}{ 274},\frac{2777}{2873}\right) &
\left(\frac{ 700}{1303},\frac{ 931}{3075}\right) &
\left(\frac{ 704}{1163},\frac{1435}{2299}\right)\\
\left(\frac{ 7}{ 106},\frac{ 245}{ 267}\right) &
\left(\frac{ 71}{ 130},\frac{ 13}{ 19}\right) &
\left(\frac{ 7}{ 194},\frac{1855}{3777}\right) &
\left(\frac{ 72}{ 115},\frac{ 217}{ 729}\right) &
\left(\frac{ 73}{ 274},\frac{1253}{1947}\right) &
\left(\frac{ 735}{1594},\frac{1024}{2401}\right) &
\left(\frac{ 739}{1318},\frac{ 249}{ 601}\right) &
\left(\frac{ 7}{ 466},\frac{ 656}{ 931}\right)\\
\left(\frac{ 75}{ 134},\frac{ 341}{ 459}\right) &
\left(\frac{ 75}{ 134},\frac{ 656}{ 675}\right) &
\left(\frac{ 76}{ 143},\frac{ 327}{ 473}\right) &
\left(\frac{ 767}{1834},\frac{ 235}{ 779}\right) &
\left(\frac{ 777}{2050},\frac{ 224}{ 513}\right) &
\left(\frac{ 7}{ 90},\frac{ 19}{ 147}\right) &
\left(\frac{ 80}{ 187},\frac{ 799}{2049}\right) &
\left(\frac{ 8}{ 25},\frac{1007}{1041}\right)\\
\left(\frac{ 8}{ 25},\frac{ 31}{ 97}\right) &
\left(\frac{ 827}{2590},\frac{1963}{3531}\right) &
\left(\frac{ 83}{ 126},\frac{ 5}{ 27}\right) &
\left(\frac{ 839}{2170},\frac{ 736}{ 857}\right) &
\left(\frac{ 881}{1538},\frac{ 160}{ 209}\right) &
\left(\frac{ 907}{2046},\frac{1075}{1203}\right) &
\left(\frac{ 92}{ 381},\frac{1127}{2073}\right) &
\left(\frac{ 95}{ 202},\frac{1264}{1411}\right)\\
\left(\frac{ 95}{ 202},\frac{ 301}{ 499}\right) &
\left(\frac{ 965}{2086},\frac{1072}{1075}\right) &
\end{array}
\]
\begin{checks}
embedded/roothalf/rational_check.mpl: check_rational()
\end{checks}
There do not appear to be any discernable patterns in the prime
factorisations of these rational numbers. The corresponding points in
$EX^*$ are stored in the variable
\mcode+inner_quasirational_points+.
The corresponding points in
$F_4^*$ can be displayed as follows:
\begin{center}
\begin{tikzpicture}[scale=8]
\draw[cyan ] (0.000,0.000) -- (0.000,0.707);
\draw[green ] (0.000,0.000) -- (1.000,0.000);
\draw[magenta] (0.000,0.707) -- (0.816,0.707);
\draw[blue,smooth] (1.000,0.000) -- (0.996,0.012) -- (0.984,0.047) -- (0.965,0.104) -- (0.943,0.177) -- (0.919,0.262) -- (0.894,0.354) -- (0.872,0.445) -- (0.853,0.530) -- (0.837,0.604) -- (0.826,0.660) -- (0.819,0.695) -- (0.816,0.707);
\fill(0.946,0.102) circle(0.004);
\fill(0.263,0.685) circle(0.004);
\fill(0.137,0.319) circle(0.004);
\fill(0.141,0.484) circle(0.004);
\fill(0.141,0.666) circle(0.004);
\fill(0.141,0.649) circle(0.004);
\fill(0.132,0.435) circle(0.004);
\fill(0.013,0.572) circle(0.004);
\fill(0.726,0.562) circle(0.004);
\fill(0.759,0.677) circle(0.004);
\fill(0.679,0.071) circle(0.004);
\fill(0.573,0.331) circle(0.004);
\fill(0.437,0.056) circle(0.004);
\fill(0.742,0.666) circle(0.004);
\fill(0.585,0.588) circle(0.004);
\fill(0.675,0.192) circle(0.004);
\fill(0.827,0.653) circle(0.004);
\fill(0.495,0.392) circle(0.004);
\fill(0.399,0.705) circle(0.004);
\fill(0.811,0.343) circle(0.004);
\fill(0.613,0.067) circle(0.004);
\fill(0.613,0.343) circle(0.004);
\fill(0.647,0.119) circle(0.004);
\fill(0.515,0.510) circle(0.004);
\fill(0.974,0.063) circle(0.004);
\fill(0.632,0.175) circle(0.004);
\fill(0.721,0.592) circle(0.004);
\fill(0.965,0.091) circle(0.004);
\fill(0.587,0.390) circle(0.004);
\fill(0.811,0.521) circle(0.004);
\fill(0.710,0.432) circle(0.004);
\fill(0.742,0.662) circle(0.004);
\fill(0.227,0.478) circle(0.004);
\fill(0.476,0.624) circle(0.004);
\fill(0.830,0.617) circle(0.004);
\fill(0.684,0.486) circle(0.004);
\fill(0.697,0.508) circle(0.004);
\fill(0.839,0.552) circle(0.004);
\fill(0.941,0.071) circle(0.004);
\fill(0.927,0.034) circle(0.004);
\fill(0.077,0.689) circle(0.004);
\fill(0.819,0.091) circle(0.004);
\fill(0.827,0.541) circle(0.004);
\fill(0.900,0.226) circle(0.004);
\fill(0.384,0.343) circle(0.004);
\fill(0.838,0.553) circle(0.004);
\fill(0.874,0.435) circle(0.004);
\fill(0.853,0.529) circle(0.004);
\fill(0.211,0.637) circle(0.004);
\fill(0.957,0.118) circle(0.004);
\fill(0.651,0.552) circle(0.004);
\fill(0.894,0.192) circle(0.004);
\fill(0.875,0.150) circle(0.004);
\fill(0.527,0.063) circle(0.004);
\fill(0.960,0.064) circle(0.004);
\fill(0.431,0.609) circle(0.004);
\fill(0.959,0.079) circle(0.004);
\fill(0.866,0.417) circle(0.004);
\fill(0.880,0.244) circle(0.004);
\fill(0.523,0.242) circle(0.004);
\fill(0.217,0.167) circle(0.004);
\fill(0.950,0.149) circle(0.004);
\fill(0.303,0.131) circle(0.004);
\fill(0.877,0.167) circle(0.004);
\fill(0.918,0.263) circle(0.004);
\fill(0.924,0.223) circle(0.004);
\fill(0.451,0.071) circle(0.004);
\fill(0.832,0.166) circle(0.004);
\fill(0.749,0.656) circle(0.004);
\fill(0.661,0.681) circle(0.004);
\fill(0.526,0.452) circle(0.004);
\fill(0.275,0.355) circle(0.004);
\fill(0.837,0.343) circle(0.004);
\fill(0.748,0.131) circle(0.004);
\fill(0.872,0.011) circle(0.004);
\fill(0.966,0.071) circle(0.004);
\fill(0.834,0.226) circle(0.004);
\fill(0.831,0.525) circle(0.004);
\fill(0.870,0.113) circle(0.004);
\fill(0.366,0.631) circle(0.004);
\fill(0.363,0.210) circle(0.004);
\fill(0.647,0.184) circle(0.004);
\fill(0.120,0.238) circle(0.004);
\fill(0.827,0.640) circle(0.004);
\fill(0.332,0.666) circle(0.004);
\fill(0.854,0.405) circle(0.004);
\fill(0.627,0.583) circle(0.004);
\fill(0.808,0.624) circle(0.004);
\fill(0.289,0.165) circle(0.004);
\fill(0.955,0.131) circle(0.004);
\fill(0.359,0.573) circle(0.004);
\fill(0.806,0.171) circle(0.004);
\fill(0.250,0.549) circle(0.004);
\fill(0.740,0.460) circle(0.004);
\fill(0.772,0.646) circle(0.004);
\fill(0.620,0.552) circle(0.004);
\fill(0.528,0.634) circle(0.004);
\fill(0.929,0.198) circle(0.004);
\fill(0.120,0.562) circle(0.004);
\fill(0.753,0.045) circle(0.004);
\fill(0.753,0.640) circle(0.004);
\fill(0.335,0.683) circle(0.004);
\fill(0.760,0.214) circle(0.004);
\fill(0.856,0.441) circle(0.004);
\fill(0.093,0.649) circle(0.004);
\fill(0.772,0.484) circle(0.004);
\fill(0.051,0.347) circle(0.004);
\fill(0.885,0.210) circle(0.004);
\fill(0.377,0.455) circle(0.004);
\fill(0.652,0.302) circle(0.004);
\fill(0.793,0.293) circle(0.004);
\fill(0.021,0.498) circle(0.004);
\fill(0.792,0.525) circle(0.004);
\fill(0.792,0.687) circle(0.004);
\fill(0.752,0.489) circle(0.004);
\fill(0.591,0.213) circle(0.004);
\fill(0.536,0.309) circle(0.004);
\fill(0.110,0.091) circle(0.004);
\fill(0.605,0.276) circle(0.004);
\fill(0.453,0.684) circle(0.004);
\fill(0.453,0.226) circle(0.004);
\fill(0.452,0.393) circle(0.004);
\fill(0.932,0.131) circle(0.004);
\fill(0.547,0.607) circle(0.004);
\fill(0.810,0.541) circle(0.004);
\fill(0.627,0.632) circle(0.004);
\fill(0.341,0.384) circle(0.004);
\fill(0.665,0.633) circle(0.004);
\fill(0.665,0.427) circle(0.004);
\fill(0.654,0.705) circle(0.004);
\end{tikzpicture}
\end{center}
(It may appear that some of the points lie on the boundary, but in
fact they are just inside, by a distance of only about $5\tm 10^{-4}$ in
some cases.)
\subsection{Integration}
\label{sec-integration}
Later on we will need to integrate various functions on $EX^*$.
Integration is most naturally defined for differential $2$-forms.
However, the metric and orientation on $EX^*$ gives a volume form $\om$
in a standard way, and we define the integral of a function $f$ to be
the integral of the form $f\om$. In most cases of interest, $f$ will
be $G$-invariant so we can just integrate over $F_{16}$ and multiply
by $16$.
Unfortunately, it seems to be difficult to compute such integrals
accurately. Given an explicit smooth embedding
$\phi\:[0,1]^2\to EX^*$, we can compute the Jacobian and then use
standard numerical integration techniques to evaluate integrals over
the image of $\phi$. However, we have not succeeded in finding a
family of such maps $\phi$ for which the Jacobians are explicitly
computable and the images cover $EX^*$ without overlap. We do know
several different homeomorphisms $[0,1]^2\to F_{16}$ that are
diffeomorphisms away from the boundary, but it seems that the singular
boundary behaviour destroys any possibility of using these maps for
accurate integration. The best that we can do along these lines is to
construct a barycentric triangulation of $EX^*$ as in
Section~\ref{sec-E-charts}. This gives us a decomposition of $EX^*$
into triangles $T$ and diffeomorphisms $\phi\:T\to\Dl_2$ where $\phi$
and its Jacobian are simple and explicit, but $\phi^{-1}$ is not.
Nonetheless, we can compute $\phi^{-1}(a)$ and associated quantities
for a large number of points $a\in T$. This leads to an approximate
integration rule of the form $I(f)=\sum_{i=1}^nw_i\,f(a_i)$ for some
points $a_1,\dotsc,a_n\in EX^*$ and weights $w_i\in\R$. We have
carried out this process and obtained a rule for which we believe that
$|I(f)-\int_{EX^*}f|/\|f\|_2$ is at most $10^{-25}$ or so for typical
functions that we need to consider. The basis for this estimate will
be explained after we have discussed some more theoretical ideas.
Unfortunately, for this rule the number $n=33600$ of sample points is
very large, so we cannot compute integrals quickly. We will also
discuss a method that gives less accurate integrals much more easily.
\subsubsection{A characterisation of the integration functional}
\label{sec-int-props}
\begin{definition}\label{defn-stokes}
For any smooth one-form $\al=\sum_{j=1}^4u_j\,dx_j$ on $\R^4$, we
define a function $D(\al)\:EX^*\to\R$ by
\[ D(\al) =
\|n\|^{-1} \sum_{ijkl} \ep_{ijkl}
\frac{\partial u_j}{\partial x_i} x_k n_l.
\]
Here $n$ is the gradient of the function $g$ in the definition of
$EX^*$, and $\ep$ is the totally antisymmetric tensor. The operator
$D$ is called \mcode+stokes+ in Maple; the definition is in
\fname+embedded/roothalf/forms.mpl+.
\end{definition}
\begin{lemma}\label{lem-stokes}
For any smooth one-form $\al$ on $\R^4$, we have
\[ d(\al|_{EX^*})=(d\al)|_{EX^*}=D(\al)\om. \]
It follows that $D(\al)$ depends only on $\al|_{EX^*}$, and that
$\int_{EX^*}D(\al)\om=0$.
\end{lemma}
\begin{proof}
We will freely use the metric to identify one-forms with vectors, so
$dx_i$ becomes the $i$'th basis vector $e_i$. We then have
$d\al=\sum_{i,j}\frac{\partial u_j}{\partial x_i}e_i\wedge e_j$.
Now consider a point $x\in EX^*$. The vectors $x$ and $n/\|n\|$ form
an orthonormal basis for $(T_xEX^*)^\perp$, so multiplication by the
form $\bt_x=x\wedge(n/\|n\|)$ gives an isometric isomorphism from
$\Lm^2T_xEX^*$ to $\Lm^4\R^4$. Thus, if we put
$\ep=e_1\wedge\dotsb\wedge e_4$, we must have
$\om_x\wedge\bt_x=\pm\ep$, and a glance at our orientation
conventions shows that the sign is positive. However, it is clear
from the definitions that $d\al\wedge\bt_x=D(\al)\ep$, so we must
have $(d\al)|_{EX^*}=D(\al)\om$ as claimed. It is standard that
$(d\al)|_{EX^*}=d(\al|_{EX^*})$ so we can apply Stokes's Theorem to
$\al|_{EX^*}$ to see that $\int_{EX^*}D(\al)\om=0$.
\end{proof}
\begin{definition}\label{defn-antiinvariant}
We say that a one-form $\al$ is \emph{antiinvariant} if
$\lm^*(\al)=\mu^*(\al)=\al$ and $\nu^*(\al)=-\al$. Equivalently, we
must have $\gm^*(\al)=\chi_7(\gm)\al$ for all $\gm\in G$, where
$\chi_7$ is as in Proposition~\ref{prop-characters}, so
$\gm^*(\al)=\al$ whenever $\gm$ preserves orientation, and
$\gm^*(\al)=-\al$ whenever $\gm$ reverses orientation.
\end{definition}
\begin{proposition}\label{prop-integration}
The integration functional $I\:C^\infty(EX^*)\to\R$ is the unique
$\R$-linear map with the following properties.
\begin{itemize}
\item[(a)] For all $f\in C^\infty(EX^*)$ and all $\gm\in G$ we
have $I(\gm^*f)=I(f)$.
\item[(b)] For the curvature map $K$ we have $I(K)=-4\pi$.
\item[(c)] For all antiinvariant one-forms $\al$ we have
$I(D(\al))=0$.
\end{itemize}
\end{proposition}
\begin{proof}
First consider the following property:
\begin{itemize}
\item[(d)] For all one-forms $\al$ we have $I(D(\al))=0$.
\end{itemize}
This clearly implies~(c), and in fact~(a) and~(c) imply~(d). To see
this, note that the operator $D$ involves division by $\om$; because
of this, it satisfies $D(\gm^*(\al))=\chi_7(\gm)\gm^*(D(\al))$.
Thus, for any one-form $\al$, the form
$\al'=|G|^{-1}\sum_\gm\chi_7(\gm)\gm^*(\al)$ is antiinvariant, and
if~(a) holds then $I(D(\al))=I(D(\al'))$, so if~(c) holds then
$I(D(\al))=0$.
For the integration functional property~(a) is clear, and~(b) is the
Gauss-Bonnet theorem, and~(d) follows from Lemma~\ref{lem-stokes}.
Now let $I'$ be another functional with properties~(a), (b) and (c),
and therefore also~(d). The de Rham Theorem tells us that the space
of two-forms on $EX^*$ modulo the image of $d$ is isomorphic to the
cohomology group $H^2(EX^*;\R)$, which has dimension one. Integration
gives a well-defined and nontrivial map from this quotient to $\R$,
which is therefore an isomorphism. It is clear from this that $I'$
must be equal to $I$.
\end{proof}
To use the above proposition, we need to have a good understanding of
the antiinvariant forms.
\begin{definition}
We write $\Xi$ for the set of antiinvariant polynomial $1$-forms on
$EX^*$. This is a module over the ring $B=A^G=\R[z_1,z_2]$. We will
also consider the ring $B'=B[(2-z_1)^{-1}]$ and the module
$\Xi'=B'\ot_B\Xi$.
\end{definition}
Recall that $0\leq z_1\leq 1$ on $EX^*$, so $2-z_1$ is everywhere
positive. It follows that $B'$ can still be regarded as a ring of
real analytic functions on $EX^*$.
\begin{proposition}\label{prop-alpha-basis}
$\Xi'$ is freely generated over $B'$ by the following forms:
\begin{align*}
\al_1 &= y_1(x_2\,dx_1-x_1\,dx_2) \\
\al_2 &= z_1(1+z_2)y_1y_2(x_2\,dx_1+x_1\,dx_2)
-2z_1y_2x_1x_2\,dx_3 + 2(1+z_1z_2)x_1x_2\,dx_4.
\end{align*}
\end{proposition}
\begin{proof}
First, let $\Om^*$ denote the free module over $A$ on generators
$dx_1,\dotsc,dx_4$. Let $\Tht$ be the submodule generated by the
elements
\begin{align*}
\tht_1 &= dg_0 \\
\tht_2 &= \half d(\rho-1) = \sum_ix_i\,dx_i;
\end{align*}
then $\Om=\Om^*/\Tht$. There is an evident action of $G$ on $\Om^*$
that preserves $\Tht$ and is compatible with the standard action on
$\Om$.
Next, for any $\R[G]$-module $V$ we define $\Pi\:V\to V$ by
\[ \Pi(v)=|G|^{-1}\sum_{\gm\in G} \chi_7(\gm)\gm^*(v). \]
It is standard that $\Pi$ is $B$-linear with $\Pi^2=\Pi$, and the
image of $\Pi$ is the subspace
\[ V[\chi_7] = \{v\in V\st \gm^*(v) = \chi_7(\gm)v
\text{ for all } \gm\in G\}.
\]
We thus have $\Xi=\Om[\chi_7]=\Pi(\Om^*)/\Pi(\Tht)$.
Now put
\[ M = \{x_1^ix_2^jy_1^ky_2^l\st 0\leq i,j,k,l \leq 1\}, \]
and recall that this is a basis for $A=\CO_{EX^*}$ over $B$. It follows
that the group $\Om^*[\chi_7]=\Pi(\Om^*)$ is generated by elements of the
form $\Pi(m\,dx_i)$ with $m\in M$ and $1\leq i\leq 4$. A computer
calculation shows that every nonzero element of this form is a
constant multiple of one of the following forms:
\begin{align*}
\bt_1 &= y_1(x_2\,dx_1 - x_1\,dx_2) \\
\bt_2 &= y_1y_2(x_2\,dx_1 + x_1\,dx_2) \\
\bt_3 &= x_1x_2y_2\,dx_3 \\
\bt_4 &= x_1x_2\,dx_4.
\end{align*}
Thus, these form a basis for $\Om^*[\chi_7]$ over $B$.
Similarly, $\Tht[\chi_7]$ is generated by elements of the form
$\Pi(m\,\tht_i)$ with $m\in M$ and $i\in\{1,2\}$. Another computer
calculation shows that every nonzero element of this form is a
constant multiple of one of the following forms:
\begin{align*}
\al_3 &= x_1x_2\tht_1 \\
\al_4 &= x_1x_2y_1y_2\tht_2.
\end{align*}
Thus, these give a basis for $\Tht[\chi_7]$ over $B$.
Now consider the matrix
\[ P = \bbm
1 & 0 & 0 & 0 \\
0 & z_1(1+z_2) & -2z_1 & 2(1+z_1z_2) \\
\rt(1-z_1)(1-2z_2) & z_1(1-2z_2) & z_1-2 & -2+3z_1-4z_1z_2 \\
(3z_1-2)z_2/(2\rt) & (1-z_1-z_1z_2)/2 & z_1 & -z_1z_2
\ebm.
\]
By straightforward calculation in $A$, we have
$\al_i=\sum_{j=1}^4P_{ij}\bt_j$ for $i=1,\dotsc,4$. We also have
$\det(P)=2-z_1$, which means that $\al_1,\dotsc,\al_4$ give a basis
for $A'\ot_A\Om^*[\chi_7]$. It follows that $\al_1$ and $\al_2$ give a
basis for $A'\ot_A\Om[\chi_7]$, as claimed.
\begin{checks}
embedded/roothalf/forms_check.mpl: check_forms()
\end{checks}
\end{proof}
The forms $\al_i$ and $\bt_j$ are \mcode+alpha_form[i]+ and
\mcode+beta_form[j]+ in Maple. The matrix $P$ is \mcode+alpha_beta+.
\begin{proposition}\label{prop-d-alpha}
The forms $\al_i$ satisfy
\begin{align*}
d\al_1 &= (9z_1^2 z_2-9 z_1^2+2z_1z_2+9z_1-2)\|n\|^{-1}\om \\
d\al_2 &= (45 z_1^3 z_2^2-45 z_1^3 z_2+78 z_1^2 z_2-20 z_1 z_2^2
-12 z_1^2-28 z_1 z_2+12 z_1-8 z_2)/\rt \|n\|^{-1}\om\\
dz_1\wedge\al_1 &= 2 z_1 (3 z_1-2) (z_1 z_2-z_1+1)\|n\|^{-1}\om \\
dz_1\wedge\al_2 &= \rt z_1
(9 z_1^3 z_2^2-9 z_1^3 z_2-12 z_1^2 z_2^2+
26 z_1^2 z_2+4 z_1 z_2^2-4 z_1^2-24 z_1 z_2+
8 z_1+8 z_2-4)\|n\|^{-1}\om \\
dz_2\wedge\al_1 &= 4 z_1 z_2 (2 z_2-1)\|n\|^{-1}\om \\
dz_2\wedge\al_2 &= 2 \rt (3 z_1^2 z_2-2 z_1 z_2+2 z_1-2) z_2 (2 z_2-1)\|n\|^{-1}\om,
\end{align*}
where
\[ \|n\|=\left(\sum_i(\partial g/\partial x_i)^2\right)^{1/2} =
(2-z_1)\sqrt{1+z_2}.
\]
\end{proposition}
\begin{proof}
The ring $A=\CO_{EX^*}$ is an integral domain, with field of fractions
\[ K = \R(y_1,y_2)[x_1,x_2]/(x_1^2-u_1,x_2^2-u_2). \]
It will suffice to verify the above identities in $K\ot_A\Om^2$,
which is the exterior square over $K$ of the space $K\ot_A\Om^1$.
From the above description of $K$, we have
\[ dx_i = \frac{1}{2x_i} du_i = \frac{x_i}{2u_i} du_i \in
K\{dy_1,\;dy_2\}.
\]
Using this, it is not hard to see that $K\ot_A\Om^1$ is freely
generated over $K$ by $dy_1$ and $dy_2$. After rewriting everything
in terms of this basis, all the above equations become
straightforward.
\begin{checks}
embedded/roothalf/forms_check.mpl: check_forms()
\end{checks}
\end{proof}
Equations for rewriting forms in terms of $dy_1$ and $dy_2$ are in the
list \mcode+forms_to_y+. The functions $(d\al_i)/\om$ are
\mcode+D_alpha[i]+, and the functions $(dz_i\wedge\al_j)/\om$ are
\mcode+dz_cross_alpha[i,j]+.
Using Proposition~\ref{prop-d-alpha}, we can calculate
$D(f_1\al_1+f_2\al_2)$ for any invariant smooth functions $f_1$ and
$f_2$. This is implemented in Maple as
\mcode+stokes_alpha([f1,f2])+, and it gives us a supply of functions $f$ with
$\int_{EX^*}f=0$. Given an approximate integration functional $I$, we
can test the accuracy of $I$ by evaluating $I(f)/\sqrt{I(f^2)}$ for
these functions $f$.
\subsubsection{Integration over triangles}
We next discuss integration over $2$-simplices. Given any continuous
function $f$ on $\Dl_2$, we write
\[ \int_{\Dl_2} f =
2\int_{t_1=0}^1\int_{t_2=0}^{1-t_1} f(t_1,t_2,1-t_1-t_2)
\,dt_2\,dt_1.
\]
In other words, this is the integral with respect to the ordinary
Lebesgue measure on $\Dl_2$ normalised in such a way that the total
area of $\Dl_2$ is one. A standard exercise shows that
\[ \int_{\Dl_2} t_1^it_2^jt_3^k = 2\frac{i!\,j!\,k!}{(i+j+k+2)!}. \]
By an \emph{$n$'th order quadrature rule} we mean a pair $Q=(a,w)$
with $a\in(\Dl_2)^n$ and $w\in\R^n$. Given any function $f$ on
$\Dl_2$ we put $I_Q(f)=\sum_i w_i\,f(a_i)$. Given $n$ distinct points
$a_i$ in $\Dl_2$, and an $n$-dimensional space of functions $P$, there
will typically be a unique vector $w$ of weights such that
$I_{(a,w)}(f)=\int_{\Dl_2}f$ for all $f\in P$, which can be found by
solving a system of linear equations. It may or may not be
the case that $I_{(a,w)}(f)$ is close to $\int_{\Dl_2}f$ for functions
$f$ not lying in $P$. In particular, if we take the points $a_i$ to
form a regularly spaced grid, then $I_{(a,w)}(f)$ is a rather poor
approximation to $\int_{\Dl_2}f$ for general $f$. This was a surprise
to the author, but is apparently well-known to numerical analysts. It
is better to allow the points $a_i$ to vary as well as the weights
$w_i$. A naive dimension count then suggests that for any
$3n$-dimensional space $P$ there should be a unique $n$'th order
quadrature rule $Q$ that agrees with integration on $P$, but one has
to solve a complex system of nonlinear equations to find $Q$. Things
become somewhat simpler if $P$ is preserved by the action of the
symmetric group $\Sg_3$ on $P$. Dunavant~\cite{du:hde} developed an
elegant theory for this case, which made it tractable to solve the
relevant equations by computer. He found a quadrature rule of order
$73$ that integrates all polynomials of degree at most $20$ exactly.
It appears that rules obtained in this way have much better behaviour
outside the space $P$ on which they are exact by construction.
Later, Wandzurat and Xiao~\cite{waxi:sqr} gave a rule of order $175$
that is exact to degree 30, and Xiao and Gimbutas gave a rule that is
exact to degree 50. However, we have not been able to get correct
answers from this last rule, so either we are misunderstanding the
conventions or there is some kind of transcription error in the tables
in the paper. We have therefore used the Wandzurat-Xiao rule
instead.
Rules as above are represented in Maple by instances of the class
\mcode+triangle_quadrature_rule+, which is declared in the file
\fname+quadrature/quadrature.mpl+. In the \fname+quadrature+
subdirectory of the \mcode+data+ directory there is a file
\mcode+wandzurat_xiao_30.mpl+. Reading this file creates an object
representing the Wandzurat-Xiao rule, and assigns it to the variable
\mcode+wandzurat_xiao_30+. There is also another file
\mcode+dunavant_19.mpl+ which implements the Dunavant rule (which is
less accurate but faster).
Next, recall that in Section~\ref{sec-E-charts} we discussed a
triangulation of the fundamental domain $F_{16}$ using certain
barycentric coordinate maps $p:T\xra{\simeq}\Dl_2$ for certain subsets
$T\sse X$. For each point $a_i$ in our quadrature rule, we can use
Remark~\ref{rem-barycentric-inverse} to find $a'_i=p^{-1}(a_i)\in T$.
The components of $p$ are rational functions in the
coordinates $x_i$, so it is straightforward to differentiate them and
calculate the Jacobian of $p$ at $a'_i$, say $u_i$. If $w_i$ is the
$i$'th weight of our quadrature rule, then $\sum_iw_iu_i^{-1}f(a_i)$
is a good approximation to $\int_Tf$, and we can take the sum over all
simplices to get a functional $J(f)$ approximating $\int_{F_{16}}f$.
The Gauss-Bonnet theorem says that $J$ of the curvature function $K$
should be $(-4\pi)/|G|=-\pi/4$. With our $192$-simplex triangulation
we in fact have $|J(K)+\pi/4|<10^{-27}$. Similarly, for a function
$f$ of the form $D(z_1^iz_2^j\al_k)$ we would ideally have $J(f)=0$,
and in fact we have $|J(f)|\leq 3\tm 10^{-27}\sqrt{J(f^2)}$ provided
that $i+j\leq 10$. Subdivision makes a big difference here; with the
original $48$-simplex triangulation, errors are larger by a factor of
$10^8$ or so.
\subsubsection{Faster integration on \texorpdfstring{$EX^*$}{EX*}}
After constructing a quadrature rule for $EX^*$, we can tabulate the
integrals of monomials $z_1^iz_2^j$, which then gives a fast way of
integrating arbitrary invariant polynomials. It is also useful to
extend this slightly and include monomials $z_1^iz_2^j\|n\|^{-k}$ for
$0\leq k\leq 4$ say. This is enough for some purposes, but in other
cases we need to integrate more general functions (such as
exponentials of polynomials) which cannot easily be expressed as
linear combinations of some standard basis. It is therefore desirable
to have an approximate integration functional of the form
$I(f)=\sum_{i=1}^n w_if(a_i)$ where $n$ is not too large, but the
approximation is reasonably accurate.
Suppose we fix $n$, and choose an $n$-dimensional subspace $P$ of
smooth invariant functions on $EX^*$. Let $p_1,\dotsc,p_n$ be a basis
for $P$. For any $n$-tuple of points $a_i\in EX^*$, we let
$\dl(\un{a})$ denote the determinant of the matrix with entries
$p_i(a_j)$. The literature on other quadrature problems suggests that
we should aim to choose $\un{a}$ so that $|\dl(\un{a})|$ is as large
as possible. Note that this problem is independent of the choice of
basis $\{p_i\}$, because a different choice would just change $\dl$ by
a constant factor. With the obvious kind of monomial bases, it works
out that $\dl(\un{a})$ is always extremely small, but it can be
increased by many orders of magnitude if we choose the points $a_i$
appropriately.
In our largest calculation of this kind, we took $P$ to be the span of
$256$ monomials in $z_1$ and $z_2$, including all monomials of degree
at most $21$, plus some monomials of degree $22$. For a randomly chosen
set of $256$ points it is typical that $\log_{10}|\dl|<-2800$ or
so, but by numerical optimization we found a set with
$\log_{10}|\dl|\simeq -2539$.
We next want to choose the weights $w_i$. One option is to set these
weights such that $I(p_j)=\int_{EX^*}p_j$ for all $j$, which can be
done by solving a system of linear equations. We then find that some
of the weights are negative. This is an undesirable feature, leading
to reduced accuracy when integrating functions outside the space $P$.
We therefore extended our list of monomials to include all $300$
monomials of degree at most $23$, and then chose the weights $w_i$ to
minimise $\sum_j(I(p_j)-\int p_j)^2$ subject to the constraints
$w_i\geq 0$. (This was done using Maple's \texttt{LSSolve()}
command.) It turns out that there are 18 indices $i$ such that
$w_i=0$, so we really only use 238 sample points. With these weights
we have $|I(K)+\pi/4|\simeq 10^{-17.1}$, and if $f=D(z_1^iz_2^j\al_k)$
with $i+j\leq 10$ then $|I(f)|\leq 10^{-14.5}\sqrt{I(f^2)}$.
Quadrature rules as above are represented by instances of the class
\mcode+E_quadrature_rule+, which is declared in the file
\fname+embedded/E_quadrature.mpl+. The specific rule described above
is stored in the file \fname+quadrature_frobenius_256a.m+ in the
directory \fname+data/embedded/roothalf+. After reading that file,
one can enter the following to integrate $1/(1+z_1)$ (for example):
\begin{mcodeblock}
Q := eval(quadrature_frobenius_256a):
Q["int_z",1/(1+z[1])];
\end{mcodeblock}
One can test the accuracy of \mcode+Q+ (as described above) using the
methods \mcode+Q["curvature_error"]+ and
\mcode+Q["stokes_error",10]+. Various other methods are documented in
the code.
One can regenerate the object \mcode+Q+ using the function
\mcode+build_data["E_quadrature_rule"]()+ defined in the file
\fname+build_data.mpl+. However, there is not a very compelling
reason to do this, as we can check that \mcode+Q+ has the desired
properties without needing to regenerate it. Also, the process is
very slow, and may take several days to run.
\section{Classifying \texorpdfstring{$EX^*$}{EX*}}
\label{sec-classify-roothalf}
Theorems~\ref{thm-classify-cromulent} and~\ref{thm-H-universal} tell
us that there are cromulent isomorphisms
$HX(b)\xra{q}EX^*\xra{r}PX(a)$ for suitable values $a,b\in(0,1)$.
In this section, we discuss numerical methods that enable us to
calculate approximations to $a$, $b$, $q$ and $r$. Note that the
methods of Section~\ref{sec-P-H} allow us to compute $a$ and $r$ from
$b$ and $q$, or \emph{vice versa}. We have tried several different
approaches. The most successful will be described first, in
Section~\ref{sec-rescaling}. We will then outline one other approach
in Section~\ref{sec-energy}.
Our current estimates are $a\simeq 0.0983562$ and $b\simeq 0.8005319$.
We have some reason to hope that all the quoted digits are accurate,
but we have not performed a rigorous error analysis.
\subsection{Hyperbolic rescaling}
\label{sec-rescaling}
Theorem~\ref{thm-H-universal} tells us that there is a conformal
covering map $q\:\Dl\to EX^*$, which induces an isomorphism
$HX(b)\to EX^*$ for some $b$. Let $g$ denote the Riemannian metric
that $EX^*$ inherits from $\R^4$, and let $g_{\hyp}$ denote the
standard hyperbolic metric $ds^2=4|dz|^2/(1-|z|^2)^2$ on $\Dl$. Recall
that Remark~\ref{rem-curvature-z} gives a formula for the Gaussian
curvature $K(g)$, and it is a standard fact that
$K(g_{\hyp})=-1$.
\begin{proposition}\label{prop-rescaling}
There is a unique real analytic function $f$ on $EX^*$ such that
$K(e^{2f}g)=-1$. Moreover, this function is $G$-invariant, and it
satisfies $q^*(e^{2f}g)=g_{\hyp}$. The curves
$C_0,\dotsc,C_8\subset EX^*$ are geodesics with respect to the metric
$e^{2f}g$.
\end{proposition}
\begin{remark}
Note here that when we multiply the metric tensor $g$ by $e^{2f}$,
this multiplies lengths by $e^f$, and areas by $e^{2f}$.
\end{remark}
The proof depends on the following formula:
\begin{lemma}\label{lem-rescaled-curvature}
Let $Z$ be a smooth oriented surface equipped with a Riemannian
metric $g$. Let $\Delta=\Dl_g$ denote the associated Laplacian
operator, and let $K(g)$ denote the Gaussian curvature. Then for any
smooth function $f$ on $Z$, we have
\[ K(e^{2f}g) = (K(g) - \Dl(f))/e^{2f}. \]
\end{lemma}
\begin{proof}
See Chapter~V of~\cite{scya:ldg}, for example.
\end{proof}
\begin{proof}[Proof of Proposition~\ref{prop-rescaling}]
Because $q$ is a conformal covering, we see that $q^*(g)$ is a
positive multiple of $g_{\hyp}$, say
$q^*(g)=e^{-2\tf}g_{\hyp}$ for some real analytic function
$\tf$ on $\Dl$. Note that $\tPi$ acts isometrically on $\Dl$, and
also (via $\tP/\Pi=G$) on $EX^*$, and $q$ is equivariant for these
actions. It follows that $\tf$ is invariant under $\tPi$, so it has
the form $\tf=q^*(f)$ for some $G$-invariant real analytic function
$f$ on $EX^*$. Now put $g_1=e^{2f}g$. We have
$q^*K(g_1)=K(e^{2\tf}q^*(g))=K(g_{\hyp})=-1$, but $q$ is
surjective so $K(g_1)=-1$ as required. Now $q$ is a local isometry
from $(\Dl,g_{\hyp})$ to $(EX^*,g_1)$, and it carries the
geodesics $C_i\subset\Dl$ to the curves $C_i\subset EX^*$, so the
latter must also be geodesic.
Now suppose we have a function $u$ with $K(e^{2u}g_1)=-1$; we
claim that $u=0$. In the proof we will use the gradient operator
$\nabla$, the Laplacian operator $\Dl$, and the integration operator
$\int_{EX^*}$; these are all defined using the metric $g_1$. Using
Lemma~\ref{lem-rescaled-curvature} we obtain $e^{2u}-1-\Dl(u)=0$.
It is a standard fact that for any functions $a,b\in C^\infty(EX^*)$
we have
\[ \int_{EX^*} a\Dl(b) = -\int_{EX^*} \ip{\nabla(a),\nabla(b)}. \]
Using this, we get
\[ \int_{EX^*}\left(u(e^{2u}-1)+\|\nabla(u)\|^2\right)
= \int_{EX^*}\left(u(e^{2u}-1+\Delta(u))\right) = 0.
\]
By considering the cases $u\geq 0$ and $u\leq 0$ separately, we see
that $u(e^{2u}-1)\geq 0$, with equality only where $u=0$. In view of
this, the above integral formula shows that $u=0$ everywhere. This
shows that $f$ is uniquely characterised by the fact that
$K(e^{2f}g)=-1$.
\end{proof}
To go further, we will need to discuss the curves $c_k\:\R\to\Dl$ as
well as the curves $c_k\:\R\to EX^*$. We will distinguish between
them by writing $c_{Hk}$ for the former, and $c_{Ek}$ for the latter.
Similarly, we will write $v_{Hj}$ and $v_{Ej}$ for the usual points in
$\Dl$ and $EX^*$.
\begin{corollary}\label{cor-side-lengths}
Let $f$ be the unique function such that $K(e^{2f}g)=-1$, and put
\begin{align*}
L_0 &= \int_{\pi/4}^{\pi/2} e^{f(c_{E0}(t))}\|c'_{E0}(t)\|\,dt \\
L_1 &= \int_{0}^{\pi/2} e^{f(c_{E1}(t))}\|c'_{E1}(t)\|\,dt \\
L_3 &= \int_{0}^{\pi/2} e^{f(c_{E3}(t))}\|c'_{E3}(t)\|\,dt \\
L_5 &= \int_{0}^{\pi} e^{f(c_{E5}(t))}\|c'_{E5}(t)\|\,dt.
\end{align*}
If $b$ is the parameter such that $EX^*\simeq HX(b)$, and the points
$v_{Hi}\in\Dl$ are defined in terms of $b$ as in
Definition~\ref{defn-v-H}, then we have
\begin{align*}
L_0 &= \dhyp(v_{H3},v_{H6}) = 2\arctanh((b_+-\rt b)/b_-) \\
L_1 &= \dhyp(v_{H0},v_{H6}) = 2\arctanh((\rt-b_-)/b_+) \\
L_3 &= \dhyp(v_{H3},v_{H11}) = 2\arctanh((1-b_-)/b) \\
L_5 &= \dhyp(v_{H0},v_{H11}) = 2\arctanh(b_+-b).
\end{align*}
\end{corollary}
\begin{proof}
The domain $HF_{16}(b)\subset\Dl$ is a hyperbolic polygon, with
geodesic edges, and vertices $v_{H0},v_{H3},v_{H6}$ and $v_{H11}$.
The map $q$ is a local isometry, which sends $v_{Hi}$ to $v_{Ei}$.
It follows that the geodesic distance from $v_{Hi}$ to $v_{Hj}$ in
$\Dl$ is the same as the geodesic distance from $v_{Ei}$ to $v_{Ej}$
in $EX^*$. As the curves $C_i\subset EX^*$ are geodesic, the
distances in $EX^*$ are given by the indicated integrals.
The distances in $\Dl$ are given by the standard formula
$\dhyp(z,w)=2\arctanh(m(z,w))$, where
\[ m(z,w) = \left|\frac{z-w}{1-\ov{z}w}\right|. \]
We therefore need to show that $m(v_{H3},v_{H6})=(b_+-\rt b)/b_-$ and
so on. It is not hard to see that $(b_+-\rt b)/b_-\geq 0$, so it
will suffice to show that $m(v_{H3},v_{H6})^2=((b_+-\rt b)/b_-)^2$,
and this can be done by straightforward algebraic manipulation. The
same method works for the other three cases.
\begin{checks}
hyperbolic/HX_check.mpl: check_side_lengths()
\end{checks}
\end{proof}
We now describe an algorithm based on the above results. The full
algorithm can be carried out by executing the function
\mcode+build_data["EH_atlas",0]()+ defined in \mcode+build_data.mpl+,
which in turn invokes the functions \mcode+build_data["EH_atlas",i]()+
for $i$ from $1$ to $4$, each of which carries out a subset of the
steps described below. All of these steps are implemented by methods
of the class \mcode+EH_atlas+ (which is declared in the file
\fname+embedded/roothalf/EH_atlas.mpl+) and some other related
classes. We can enter
\begin{mcodeblock}
EHA := `new/EH_atlas`():
\end{mcodeblock}
to create a new object of the required type.
First, we need a quadrature rule $Q(f)=\sum_iw_if(a_i)$,
which is intended to approximate $\int_{EX^*}f$ when $f$ is
$G$-invariant. It is not important for this approximation to be
accurate, we just want the seminorm $\|f\|_Q=\sqrt{Q(f^2)}$ to be a
reasonable measure of the size of $f$, at least for the functions $f$
that arise in our calculations. After constructing a suitable
instance \mcode+Q+ of the class \mcode+E_quadrature_rule+, we can
enter
\begin{mcodeblock}
EHA["quadrature_rule"] := eval(Q):
\end{mcodeblock}
to attach the quadrature rule to the atlas.
Next, we need to choose a finite-dimensional subspace $F$ of
invariant, real analytic functions on $EX^*$, in which we will search
for an approximation to the rescaling function $f$. Two possibilities
are as follows:
\begin{align*}
F_{\text{poly}}(d) &=
\{\text{ polynomials } p(z_1,z_2)
\text{ of total degree at most } d\} \\
F_{\text{pade}}(d) &=
\left\{\text{ rational functions } \frac{p(z_1,z_2)}{q(z_1,z_2)},\;
\max(\deg(p),\deg(q)) \leq d,\; q(0,0)=1\right\}.
\end{align*}
We have found that $F_{\text{pade}}$ works better than
$F_{\text{poly}}$ even if the degrees are chosen so that the total
number of parameters is the same. We do not fully understand why,
although it is reminiscent of the situation with Pad\'e approximations
in one variable.
One can search through the above spaces using one of the following
methods:
\begin{mcodeblock}
EHA["find_rescale_poly",d];
EHA["find_rescale_poly_alt",d];
EHA["find_rescale_pade",d];
EHA["find_rescale_pade_alt",d];
\end{mcodeblock}
It is best to call these methods repeatedly starting with $d=2$. A
measure of success is stored in the field
\mcode+EHA["rescaling_error"]+, and this should decrease on each
iteration. When this measure has settled down, one can increase $d$
by one.
The first two of the above methods use $F_{\text{poly}}(d)$, and the
second two use $F_{\text{pade}}(d)$. The alternative methods
\mcode+find_rescale_poly+ and \mcode+find_rescale_pade+ are coded
using a very direct translation of the mathematical problem that we
seek to solve, so it is easy to check their correctness. The methods
\mcode+find_rescale_poly_alt+ and \mcode+find_rescale_pade_alt+ are
harder to understand but much more efficient: they precompute various
vectors and matrices, and thereby convert the problem to linear
algebra, as far as possible. The polynomial case works as follows:
\begin{itemize}
\item We enumerate the monomials of degree at most $d$ as
$m_1,\dotsc,m_r$.
\item The sample points for the quadrature rule are $u_1,\dotsc,u_n$,
with weights $w_i\geq 0$. We set \mcode+rweights+ to be the vector
with entries $\sqrt{w_i}$, and we set \mcode+kpoints+ to be the
vector whose entries are the values of the curvature at the points
$u_i$ (computed using Remark~\ref{rem-curvature-z}). We also set
\mcode+ones+ to be the vector of length $n$ whose entries are all
one.
\item Similarly, we set \mcode+mpoints+ and \mcode+lpoints+ to be
matrices with entries $m_i(u_j)$ and $\Delta(m_i)(u_j)$ (computed
using Proposition~\ref{prop-Delta-prime}).
\item Now if $f=\sum_ja_jm_j$, then the values of $f$ and
$K(g)-\Delta(f)$ are given by \mcode+mpoints.a+ and
\mcode+kpoints-lpoints.a+. This makes it easy to compute the
objective function $FT(a)=Q(((K(g)-\Delta(f))/e^{2f}+1)^2)$.
\item More precisely, $FT(a)$ is $\sum_iF_i(a)^2$, where $F_i(a)$ is
$\sqrt{w_i}$ times the value of $(K(g)-\Delta(f))/e^{2f}+1$ at
$u_i$. This is useful, because there are special algorithms (used
by Maple's \mcode+LSSolve+ command) for minimising a sum of
squares.
\item The above framework gives an efficient method for calculating
the vector $F(a)$ and also the matrix of derivatives
$\partial F_i(a)/\partial a_j$. These are the ingredients that we
need in order to use the \mcode+LSSolve+ command with the
\mcode+objectivejacobian+ option.
\end{itemize}
The rational case is more complicated. Here we have $f=g/h$, and the
values of $f$ and $\Delta(f)$ do not depend linearly on the
coefficients of $g$ and $h$. However, one can construct linear
differential operators $P_i$ and $Q_j$ such that
\[ \Delta(f) =
\frac{P_1(g)}{h} +
\frac{P_2(g)Q_2(h)+P_3(g)Q_3(h)+P_4(g)Q_4(h)}{h^2} +
P_5(g)\frac{Q_2(h)Q_5(h)+Q_3(h)Q_6(h)}{h^3}.
\]
(This is just a version of the quotient rule for second derivatives.)
One can again precompute the values $P_i(m_k)(u_l)$ and
$Q_j(m_k)(u_l)$ and thereby streamline the calculation of the
objective function, and of its derivatives with respect to the
coefficients of $g$ and $h$.
After using these methods to find the rescaling function
$f$, we can enter \mcode+EHA["log_rescale_z"](z)+ to see $f$ as an
expression in $z_1$ and $z_2$, or \mcode+EHA["log_rescale_x"](x)+ to
see it as an expression in $x_1,\dotsc,x_4$.
Once we have found $f$, we define lengths $L_i$ by the integrals
specified in Corollary~\ref{cor-side-lengths}. We then use numerical
methods to find $b_5$ such that
$L_5=2\arctanh(2b_5^2+1-2b_5\sqrt{1+b_5^2})$, and similarly for $b_3$,
$b_1$ and $b_0$, using the formulae in
Corollary~\ref{cor-side-lengths}. If our approximations are good,
then $b_0,b_1,b_3$ and $b_5$ should all be close to the parameter $b$
such that $EX^*\simeq HX(b)$. We can thus get an imperfect measure of
the accuracy of our approximations from the differences $|b_i-b_j|$;
these are at most $10^{-7.4}$ in our best attempt.
The above algorithm is implemented by the method
\mcode+EHA["find_a_H"]+. The length $L_k$ is stored as
\mcode+EHA["curve_lengths"][k]+, and $b_k$ is stored as
\mcode+EHA["curve_a_H_estimates"][k]+. The average of these is
\mcode+EHA["a_H"]+, and the maximum discrepancy between them is
\mcode+EHA["a_H_discrepancy"]+.
Having found $b$, we can construct an isomorphism $HX(b)\to PX(a)$ by
the methods of Sections~\ref{sec-a-from-b} and~\ref{sec-b-from-a}.
This gives objects of class \mcode+H_to_P_map+ and
\mcode+P_to_H_map+. These can be assigned to the fields
\mcode+EHA["H_to_P_map"]+ and \mcode+EHA["P_to_H_map"]+, in order to
keep everything packaged together in a single object. These steps are
not included in the function \mcode+build_data["EH_atlas",0]()+, but
are instead in the functions \mcode+build_data["H_to_P_map"]()+ and
\mcode+build_data["P_to_H_map"]()+.
We next want to approximate the map $q\:\Dl\to EX^*$.
\begin{remark}\label{rem-H-to-E-method}
The broad outline of our method is as follows:
\begin{itemize}
\item[(a)] It is not hard to see that $q(c_{kH}(t))=c_{kE}(u_k(t))$
for a certain function $u_k(t)$, and to find Fourier series
approximations to $u_k(t)-t$ by numerical integration of arc lengths.
\item[(b)] Given a point $a\in EX^*$, it is not too hard to find a
polynomial map $p_a\:\Dl\to\R^4$ which satisfies $p_a(0)=a$ and is a
good approximation to an isometry $\Dl\to EX^*$, at least if we
consider points close to the origin in $\Dl$. (We will call these
maps \emph{hyperbolic charts}.) If $a\in C_k$ for some $k$ then we
can exploit information from~(a) to find $p_a$; otherwise we use a
more general method with power series. Experiment suggests that
our value of $p_a(z)$ can only be trusted for fairly small values
of $|z|$, perhaps $|z|<0.1$ or so.
\item[(c)] We then show that there exists a M\"obius map
$m_a\in\Aut(\Dl)$ such that $p_a(z)\simeq q(m_a(z))$ for small $z$.
Equivalently, for $w$ close to $m_a(0)$ we have
$q(w)\simeq p_a(m_a^{-1}(w))$. Thus, to find $q$, we should try to
find $p_a$ and $m_a$ for a reasonable supply of points
$a\in F_{16}$. Methods for $p_a$ were discussed above, but we still
need to consider $m_a$.
\item[(d)] Given nearby points $a,b\in EX^*$, we can find
$p_a^{-1}(b)$ and $p_b^{-1}(a)$ numerically, and the values
$\dhyp(0,p_a^{-1}(b))$ and $\dhyp(p_b^{-1}(a),0)$ give two different
estimates for the geodesic distance between $a$ and $b$. (The
discrepancy between them gives a check on the accuracy of our
methods.)
\item[(e)] We now choose a reasonably fine grid of points
$a_1,\dotsc,a_N$ in $F_{16}$, and try to find the points
$\bt_i=q^{-1}(a_i)\in HF_{16}(b)$. For any points $a_i$ that lie in
$\partial F_{16}$, we can do this using~(a). For the remaining
points, we note that when $a_i$ is a neighbour of $a_j$, we can
estimate the geodesic distance between them as in~(d), and we should
then have $\dhyp(\bt_i,\bt_j)=d(a_i,a_j)$. We therefore choose the
points $\bt_i$ to minimize a suitable measure of the overall
discrepancy between the lengths $\dhyp(\bt_i,\bt_j)$ and $d(a_i,a_j)$.
\item[(f)] We now need the M\"obius maps $m_i$ such that
$p_i=qm_i$. It is not hard to see that these must have the form
\[ m_i(z)=\lm_i\frac{z+\ov{\lm_i}\bt_i}{1-\lm_i\ov{\bt_i}z}
\hspace{4em}
m_i^{-1}(w) = \ov{\lm_i}\frac{w-\bt_i}{1-\ov{\bt_i}w}.
\]
If $a_j$ is adjacent to $a_i$ then we find that $m_i^{-1}(\bt_j)$
should be equal to $p_i^{-1}(a_j)$. We can again calculate
$p_i^{-1}(a_j)$ numerically, and this gives
\[ \lm_i = (\bt_j-\bt_i)/(1 - \ov{\bt_i}\bt_j)/p_i^{-1}(a_j). \]
We can perform this calculation for every $j$ such that $a_j$ is
adjacent to $a_i$, and then take a kind of average to get a final
estimate for $\lm_i$. (Of course, in the averaging process we
impose the constraint $|\lm_i|=1$.)
\item[(g)] Now given a point $z\in\Dl$, we can approximate $q(z)$ as
follows: we find $\gm\in\tPi$ such that $\gm(z)\in F_{16}$, then
find $i$ such that $\gm(z)$ is as close as possible to $\bt_i$,
then take $q(z)=\gm^{-1}(p_i(m_i^{-1}(\gm(z))))$ (using the action
of $\tPi$ on $EX^*$ via $\tPi/\Pi=G$). We can use this method to
calculate $q(z)$ for a large sample of points $z\in\Dl$, and then
use numerical techniques to find an approximation to $q(x+iy)$ using
rational functions in $x$ and $y$.
\end{itemize}
\end{remark}
We now discuss the above points~(a) to~(e) in more detail.
First, as $q$ gives a cromulent isomorphism $HX(b)\to EX^*$, we must
have $q(v_{Hi})=v_{Ei}$ for all $i$, and $q(c_{Hk}(\R))=C_{Ek}$. As
$q\circ c_{Hk}\:\R\to C_{Ek}$ and $c_{Ek}\:\R\to C_{Ek}$ are both
$2\pi$-periodic universal coverings, it is not hard to see
that we must have $q(c_{Hk}(t))=c_{Ek}(u_k(t))$ for some strictly
increasing diffeomorphism $u_k\:\R\to\R$ with
$u_k(t+2\pi)=u_k(t)+2\pi$. This in turn means that the function
$u_k(t)-t$ is $2\pi$-periodic, so it can be represented by a
Fourier series. The maps $c_k\:\R\to\Dl$ were defined so as to have
constant speed with respect to the hyperbolic metric; let that speed
be $s_k$. As $q$ is locally isometric, we can differentiate the
relation $q(c_{Hk}(u_k^{-1}(t)))=c_{Ek}(t)$ to get
\[ \frac{d}{dt}u_k^{-1}(t) =
s_k^{-1}\|c'_{Ek}(t)\|e^{f(c_{Ek}(t))}.
\]
We can integrate this numerically to find a Fourier series
approximation to $u_k^{-1}(t)-t$. From this we can obtain a Fourier
approximation to $u_k(t)-t$, and thus approximate formulae for
$q(c_{Hk}(t))$. To give an idea of the size of the dominant terms, we
have
\begin{align*}
u_0(t) - t &\simeq 0.017 \sin(4t) \\
u_1(t) - t &\simeq -0.169 \sin(2t) + 0.010 \sin(4t) - 0.001 \sin(6t) \\
u_3(t) - t &\simeq -0.074 \sin(2t) + 0.002 \sin(4t) \\
u_5(t) - t &\simeq -0.362 \sin(t) + 0.026\sin(2t) - 0.001\sin(3t).
\end{align*}
These are calculated by the method \mcode+EHA["find_u",d]+ (where $d$
controls the number of terms in the various Fourier series). After
invoking this method, one can calculate $u_k(t)$ and $u_k^{-1}(t)$ as
\mcode+EHA["u"][k](t)+ and \mcode+EHA["u_inv"][k](t)+.
We next discuss point~(b) in Remark~\ref{rem-H-to-E-method}.
\begin{proposition}\label{prop-local-isometry}
Let $Z$ be an oriented surface with a Riemannian metric of curvature
$-1$. Let $a$ be a point in $Z$, and let $v$ be a nonzero tangent
vector at $a$. Then there is a unique germ of an oriented local
isometry $p\:\Dl\to Z$ such that $p(0)=a$ and $p'(0)\in\R^+v$.
\end{proposition}
\begin{proof}
Local existence of $p$ is a theorem of Riemann; a convenient
reference is~\cite[Theorem 2.4.11]{wo:scc}. For uniqueness, it will
suffice to prove the following: if $u$ is a germ of an oriented local
isometry $\Dl\to\Dl$ with $u(0)=0$ and $u'(0)>0$, then $u$ is the
identity. It is clear that $u$ must act as a rotation on the tangent
space $T_0\Dl$, so the condition $u'(0)>0$ forces $u'(0)=1$. The
exponential map $\exp\:T_0\Dl\to\Dl$ is characterised by its metric
properties, so it commutes with $u$, and $T_0u=1$ so $u=1$.
\end{proof}
\begin{corollary}\label{cor-local-isometry}
Let $p\:U\to EX^*$ be a hyperbolic chart (where $U$ is a disc around
$0$ in $\Dl$). Then there is a M\"obius map $m\in\Aut(\Dl)$ such
that $p=qm$ on $U$.
\end{corollary}
\begin{proof}
As $q\:\Dl\to EX^*$ is a covering map and $U$ is simply connected, we
can choose $m\:U\to\Dl$ such that $p=qm$. As both $p$ and $q$ are
orientation-preserving isometries, the same is true of $m$. Now put
$\al=m(0)\in\Dl$, and let $\lm$ denote the unit complex number such
that $m'(0)$ is a positive multiple of $\ov{\lm}$. Put
$m_1(z)=(z+\lm\al)/(\lm+\ov{\al}z)$ (so
$m_1^{-1}(z)=(z-\al)/(1-\ov{\al}z)$). We find that $m$ and $m_1$ are
both orientation-preserving isometries of $U$ into $\Dl$ such that
$m'(0)$ and $m'_1(0)$ are positive multiples of each other. It
follows (by the uniqueness clause in the Proposition) that $m=m_1$.
\end{proof}
Charts $p$ as above are represented by instances of the class
\mcode+EH_chart+, which is declared in the file
\fname+embedded/roothalf/EH_atlas.mpl+. It extends the class
\mcode+E_chart+, which was discussed in
Section~\ref{sec-roothalf-charts}. The algorithm to make a chart
isometric is actually coded in the \mcode+isometrize+ method of the
\mcode+E_chart+ class; this is invoked automatically by the methods
that initialize instances of the \mcode+EH_chart+ class. In more
detail, the algorithm is as follows. We start with an approximate
polynomial conformal chart $p_0\:\C\to EX^*$ as in
Proposition~\ref{prop-frame-chart}, which can be constructed by
methods of the \mcode+E_chart+ class. It is then not hard to show
that there are unique numbers $a_1\in\R^+$ and $a_2\in\C$ such that
the map $p_2(z)=p_0(a_1z+a_2z^2)$ is isometric to first order. We can
then try to find $a_3$ such that the map $p_3(z)=p_2(z+a_3z^3)$ is
isometric to second order. This involves solving a system of
inhomogeneous linear equations for the real and imaginary parts of
$a_3$. As the curvature of $g_1$ is not exactly equal to $-1$, these
equations will not usually be solvable. However, we can choose $a_3$
to minimize the mean square error in these equations, and then proceed
to find coefficients $a_4$, $a_5$ and so in in the same way.
As mentioned previously, there a different method that is available
for charts centred on one of the curves $C_k$. Suppose that
$a=c_{Ek}(t_0)$, and that we have found a good approximation to the
conformal chart $p_0$ with $p_0(t)=c_{Ek}(t_0+t))$ for small $t\in\R$
(as discussed in Section~\ref{sec-roothalf-charts}). Put
$t_1=u_k^{-1}(t_0)$. The function $u_k\:\R\to\R$ is real analytic, so
it can be extended (using power series, for example) to give a
holomorphic function on a neighbourhood of the point
$t_1=u_k^{-1}(t_0)$. Now put
\begin{align*}
\lm &= c'_{Hk}(t_1)/|c'_{Hk}(t_1)|\in S^1 \\
\al &= -c_{Hk}(t_1)/\lm \\
\bt &= -\lm\al \\
m(z) &= \lm(z-\al)/(1-\ov{\al}z),
\end{align*}
so $m\in\Aut(\Dl)$ with $m(0)=c_{Hk}(t_1)$ and
$m'(0)\in\R^+.c'_{Hk}(t_1)$. Finally, recall that $s_k$ denotes the
(constant) speed of the map $c_{Hk}\:\R\to\Dl$.
\begin{proposition}\label{prop-curve-chart}
With notation as above, the map
\[ p(z) = p_0(u_k(t_1+2s_k^{-1}\arctanh(z))-t_0) \]
is a hyperbolic chart at $a$. More specifically, we have
$p(z)=q(m(z))$. In particular, we have $q^{-1}p(0)=m(0)=-\lm\al=\bt$.
\end{proposition}
\begin{proof}
Note that $m^{-1}(c_{Hk}(\R))$ is a geodesic in $\Dl$ which is
tangent to $\R$ at $0$; it follows that
$m^{-1}(c_{Hk}(\R))=(-1,1)$. The standard speed one parametrisation
of $(-1,1)$ is $t\mapsto\tanh(t/2)$. It follows that
$c_{Hk}(t_1+t)=m(\tanh(s_kt/2))$.
We now claim that $p(z)=q(m(z))$. Both $p(z)$ and $q(m(z))$ are
holomorphic, so it will suffice to prove this for small real values
of $z$. When $z$ is real we see that
$u_k(t_1+2s_k^{-1}\arctanh(z))-t_0$ is also real, so
\begin{align*}
p(z) &= c_{Ek}(u_k(t_1+2s_k^{-1}\arctanh(z))) \\
&= q(c_{Hk}(t_1+2s_k^{-1}\arctanh(z))) = q(m(t))
\end{align*}
as required.
\end{proof}
For $j\in\{0,3,6,11\}$ one can add an isometric chart centred at the
point $v_j$ to the atlas using the method
\mcode+EHA["add_vertex_chart",j]+. Now suppose that
$k\in\{0,1,3,5\}$, so the set $c_k^{-1}(F_{16})$ is a closed interval
$[a,b]$ for some $a$ and $b$. Then the method
\mcode+EHA["add_curve_chart",k,t]+ adds an isometric chart centred at
$a+t(b-a)$ (so the natural domain for $t$ is $[0,1]$). Finally, for a
point $x_0$ in the interior of $F_{16}$, we can use the method
\mcode+EHA["add_centre_chart",x0]+ to add a chart centred at $x_0$.
The function \mcode+build_data["EH_atlas",2]()+ adds a total of 119
charts to the atlas created by \mcode+build_data["EH_atlas",1]()+.
They are chosen so that the centres form an approximately equilateral
triangular grid with respect to the rescaled hyperbolic metric. (To
make everything fit, some triangles on the edge of $F_{16}$ have to
deviate strongly from being equilateral, but the ones in the interior
are quite regular.) The corresponding points in $HF_{16}(b)$ can be
displayed as follows:
\begin{center}
\begin{tikzpicture}[scale=12]
\draw[maplegreen] (0.000,0.000) -- (0.020,0.020);
\draw[maplegreen] (0.000,0.000) -- (0.032,0.032);
\draw[mapleblue] (0.000,0.000) -- (0.034,0.000);
\draw[maplecyan] (0.568,0.364) -- (0.547,0.375);
\draw[maplemagenta] (0.568,0.364) -- (0.557,0.341);
\draw[maplegrey] (0.568,0.364) -- (0.550,0.346);
\draw[maplecyan] (0.450,0.450) -- (0.467,0.434);
\draw[maplegreen] (0.450,0.450) -- (0.422,0.422);
\draw[maplegrey] (0.450,0.450) -- (0.434,0.418);
\draw[maplemagenta] (0.480,-0.000) -- (0.481,0.023);
\draw[mapleblue] (0.480,-0.000) -- (0.470,0.000);
\draw[maplecyan] (0.467,0.434) -- (0.493,0.412);
\draw[maplegrey] (0.467,0.434) -- (0.434,0.418);
\draw[maplegrey] (0.467,0.434) -- (0.463,0.398);
\draw[maplecyan] (0.493,0.412) -- (0.519,0.393);
\draw[maplegrey] (0.493,0.412) -- (0.463,0.398);
\draw[maplegrey] (0.493,0.412) -- (0.492,0.380);
\draw[maplecyan] (0.519,0.393) -- (0.547,0.375);
\draw[maplegrey] (0.519,0.393) -- (0.492,0.380);
\draw[maplegrey] (0.519,0.393) -- (0.521,0.362);
\draw[maplegrey] (0.547,0.375) -- (0.521,0.362);
\draw[maplegrey] (0.547,0.375) -- (0.550,0.346);
\draw[maplegreen] (0.422,0.422) -- (0.403,0.403);
\draw[maplegrey] (0.422,0.422) -- (0.434,0.418);
\draw[maplegreen] (0.352,0.352) -- (0.378,0.378);
\draw[maplegreen] (0.352,0.352) -- (0.331,0.331);
\draw[maplegrey] (0.352,0.352) -- (0.365,0.350);
\draw[maplegreen] (0.378,0.378) -- (0.403,0.403);
\draw[maplegrey] (0.378,0.378) -- (0.365,0.350);
\draw[maplegrey] (0.378,0.378) -- (0.399,0.367);
\draw[maplegrey] (0.403,0.403) -- (0.434,0.418);
\draw[maplegrey] (0.403,0.403) -- (0.399,0.367);
\draw[maplegrey] (0.403,0.403) -- (0.432,0.383);
\draw[maplegreen] (0.275,0.275) -- (0.304,0.304);
\draw[maplegreen] (0.275,0.275) -- (0.253,0.253);
\draw[maplegrey] (0.275,0.275) -- (0.291,0.273);
\draw[maplegreen] (0.304,0.304) -- (0.331,0.331);
\draw[maplegrey] (0.304,0.304) -- (0.291,0.273);
\draw[maplegrey] (0.304,0.304) -- (0.327,0.293);
\draw[maplegrey] (0.331,0.331) -- (0.365,0.350);
\draw[maplegrey] (0.331,0.331) -- (0.327,0.293);
\draw[maplegrey] (0.331,0.331) -- (0.363,0.311);
\draw[maplegreen] (0.191,0.191) -- (0.223,0.223);
\draw[maplegreen] (0.191,0.191) -- (0.166,0.166);
\draw[maplegrey] (0.191,0.191) -- (0.211,0.186);
\draw[maplegreen] (0.223,0.223) -- (0.253,0.253);
\draw[maplegrey] (0.223,0.223) -- (0.211,0.186);
\draw[maplegrey] (0.223,0.223) -- (0.251,0.209);
\draw[maplegrey] (0.253,0.253) -- (0.291,0.273);
\draw[maplegrey] (0.253,0.253) -- (0.251,0.209);
\draw[maplegrey] (0.253,0.253) -- (0.289,0.231);
\draw[maplegreen] (0.020,0.020) -- (0.059,0.059);
\draw[maplegreen] (0.020,0.020) -- (0.032,0.032);
\draw[maplegreen] (0.059,0.059) -- (0.097,0.097);
\draw[maplegreen] (0.059,0.059) -- (0.032,0.032);
\draw[maplegrey] (0.059,0.059) -- (0.080,0.060);
\draw[maplegreen] (0.097,0.097) -- (0.132,0.132);
\draw[maplegrey] (0.097,0.097) -- (0.080,0.060);
\draw[maplegrey] (0.097,0.097) -- (0.126,0.088);
\draw[maplegreen] (0.132,0.132) -- (0.166,0.166);
\draw[maplegrey] (0.132,0.132) -- (0.126,0.088);
\draw[maplegrey] (0.132,0.132) -- (0.169,0.114);
\draw[maplegrey] (0.166,0.166) -- (0.211,0.186);
\draw[maplegrey] (0.166,0.166) -- (0.169,0.114);
\draw[maplegrey] (0.166,0.166) -- (0.211,0.139);
\draw[maplegrey] (0.032,0.032) -- (0.034,0.000);
\draw[maplegrey] (0.032,0.032) -- (0.082,0.000);
\draw[maplegrey] (0.032,0.032) -- (0.080,0.060);
\draw[maplemagenta] (0.545,0.316) -- (0.557,0.341);
\draw[maplemagenta] (0.545,0.316) -- (0.533,0.287);
\draw[maplegrey] (0.545,0.316) -- (0.521,0.329);
\draw[maplegrey] (0.545,0.316) -- (0.550,0.346);
\draw[maplegrey] (0.545,0.316) -- (0.523,0.294);
\draw[maplegrey] (0.557,0.341) -- (0.550,0.346);
\draw[maplemagenta] (0.524,0.260) -- (0.533,0.287);
\draw[maplemagenta] (0.524,0.260) -- (0.514,0.228);
\draw[maplegrey] (0.524,0.260) -- (0.493,0.275);
\draw[maplegrey] (0.524,0.260) -- (0.523,0.294);
\draw[maplegrey] (0.524,0.260) -- (0.496,0.239);
\draw[maplegrey] (0.533,0.287) -- (0.523,0.294);
\draw[maplemagenta] (0.506,0.200) -- (0.514,0.228);
\draw[maplemagenta] (0.506,0.200) -- (0.498,0.166);
\draw[maplegrey] (0.506,0.200) -- (0.466,0.217);
\draw[maplegrey] (0.506,0.200) -- (0.496,0.239);
\draw[maplegrey] (0.506,0.200) -- (0.470,0.179);
\draw[maplegrey] (0.514,0.228) -- (0.496,0.239);
\draw[maplemagenta] (0.498,0.166) -- (0.491,0.130);
\draw[maplegrey] (0.498,0.166) -- (0.470,0.179);
\draw[maplegrey] (0.498,0.166) -- (0.475,0.141);
\draw[maplemagenta] (0.487,0.101) -- (0.491,0.130);
\draw[maplemagenta] (0.487,0.101) -- (0.483,0.063);
\draw[maplegrey] (0.487,0.101) -- (0.444,0.115);
\draw[maplegrey] (0.487,0.101) -- (0.475,0.141);
\draw[maplegrey] (0.487,0.101) -- (0.451,0.075);
\draw[maplegrey] (0.491,0.130) -- (0.475,0.141);
\draw[maplemagenta] (0.483,0.063) -- (0.481,0.023);
\draw[maplegrey] (0.483,0.063) -- (0.451,0.075);
\draw[maplegrey] (0.483,0.063) -- (0.460,0.034);
\draw[maplegrey] (0.481,0.023) -- (0.470,0.000);
\draw[maplegrey] (0.481,0.023) -- (0.460,0.034);
\draw[mapleblue] (0.034,0.000) -- (0.082,0.000);
\draw[mapleblue] (0.082,0.000) -- (0.131,0.000);
\draw[maplegrey] (0.082,0.000) -- (0.080,0.060);
\draw[maplegrey] (0.082,0.000) -- (0.128,0.035);
\draw[mapleblue] (0.131,0.000) -- (0.175,0.000);
\draw[maplegrey] (0.131,0.000) -- (0.128,0.035);
\draw[mapleblue] (0.175,0.000) -- (0.222,0.000);
\draw[maplegrey] (0.175,0.000) -- (0.128,0.035);
\draw[maplegrey] (0.175,0.000) -- (0.171,0.063);
\draw[maplegrey] (0.175,0.000) -- (0.217,0.042);
\draw[mapleblue] (0.222,0.000) -- (0.269,0.000);
\draw[maplegrey] (0.222,0.000) -- (0.217,0.042);
\draw[maplegrey] (0.222,0.000) -- (0.262,0.022);
\draw[mapleblue] (0.269,0.000) -- (0.307,0.000);
\draw[maplegrey] (0.269,0.000) -- (0.262,0.022);
\draw[mapleblue] (0.307,0.000) -- (0.351,0.000);
\draw[maplegrey] (0.307,0.000) -- (0.262,0.022);
\draw[maplegrey] (0.307,0.000) -- (0.300,0.051);
\draw[maplegrey] (0.307,0.000) -- (0.343,0.034);
\draw[mapleblue] (0.351,0.000) -- (0.395,0.000);
\draw[maplegrey] (0.351,0.000) -- (0.343,0.034);
\draw[maplegrey] (0.351,0.000) -- (0.386,0.018);
\draw[mapleblue] (0.395,0.000) -- (0.428,0.000);
\draw[maplegrey] (0.395,0.000) -- (0.386,0.018);
\draw[mapleblue] (0.428,0.000) -- (0.470,0.000);
\draw[maplegrey] (0.428,0.000) -- (0.386,0.018);
\draw[maplegrey] (0.428,0.000) -- (0.419,0.047);
\draw[maplegrey] (0.428,0.000) -- (0.460,0.034);
\draw[maplegrey] (0.470,0.000) -- (0.460,0.034);
\draw[maplegrey] (0.434,0.418) -- (0.432,0.383);
\draw[maplegrey] (0.434,0.418) -- (0.463,0.398);
\draw[maplegrey] (0.365,0.350) -- (0.399,0.367);
\draw[maplegrey] (0.365,0.350) -- (0.363,0.311);
\draw[maplegrey] (0.365,0.350) -- (0.397,0.329);
\draw[maplegrey] (0.399,0.367) -- (0.432,0.383);
\draw[maplegrey] (0.399,0.367) -- (0.397,0.329);
\draw[maplegrey] (0.399,0.367) -- (0.430,0.347);
\draw[maplegrey] (0.432,0.383) -- (0.463,0.398);
\draw[maplegrey] (0.432,0.383) -- (0.430,0.347);
\draw[maplegrey] (0.432,0.383) -- (0.461,0.363);
\draw[maplegrey] (0.463,0.398) -- (0.461,0.363);
\draw[maplegrey] (0.463,0.398) -- (0.492,0.380);
\draw[maplegrey] (0.291,0.273) -- (0.327,0.293);
\draw[maplegrey] (0.291,0.273) -- (0.289,0.231);
\draw[maplegrey] (0.291,0.273) -- (0.326,0.251);
\draw[maplegrey] (0.327,0.293) -- (0.363,0.311);
\draw[maplegrey] (0.327,0.293) -- (0.326,0.251);
\draw[maplegrey] (0.327,0.293) -- (0.362,0.272);
\draw[maplegrey] (0.363,0.311) -- (0.397,0.329);
\draw[maplegrey] (0.363,0.311) -- (0.362,0.272);
\draw[maplegrey] (0.363,0.311) -- (0.396,0.291);
\draw[maplegrey] (0.397,0.329) -- (0.430,0.347);
\draw[maplegrey] (0.397,0.329) -- (0.396,0.291);
\draw[maplegrey] (0.397,0.329) -- (0.429,0.310);
\draw[maplegrey] (0.430,0.347) -- (0.461,0.363);
\draw[maplegrey] (0.430,0.347) -- (0.429,0.310);
\draw[maplegrey] (0.430,0.347) -- (0.461,0.328);
\draw[maplegrey] (0.461,0.363) -- (0.492,0.380);
\draw[maplegrey] (0.461,0.363) -- (0.461,0.328);
\draw[maplegrey] (0.461,0.363) -- (0.491,0.345);
\draw[maplegrey] (0.492,0.380) -- (0.491,0.345);
\draw[maplegrey] (0.492,0.380) -- (0.521,0.362);
\draw[maplegrey] (0.211,0.186) -- (0.251,0.209);
\draw[maplegrey] (0.211,0.186) -- (0.211,0.139);
\draw[maplegrey] (0.211,0.186) -- (0.251,0.163);
\draw[maplegrey] (0.251,0.209) -- (0.289,0.231);
\draw[maplegrey] (0.251,0.209) -- (0.251,0.163);
\draw[maplegrey] (0.251,0.209) -- (0.290,0.187);
\draw[maplegrey] (0.289,0.231) -- (0.326,0.251);
\draw[maplegrey] (0.289,0.231) -- (0.290,0.187);
\draw[maplegrey] (0.289,0.231) -- (0.327,0.209);
\draw[maplegrey] (0.326,0.251) -- (0.362,0.272);
\draw[maplegrey] (0.326,0.251) -- (0.327,0.209);
\draw[maplegrey] (0.326,0.251) -- (0.362,0.231);
\draw[maplegrey] (0.362,0.272) -- (0.396,0.291);
\draw[maplegrey] (0.362,0.272) -- (0.362,0.231);
\draw[maplegrey] (0.362,0.272) -- (0.397,0.252);
\draw[maplegrey] (0.396,0.291) -- (0.429,0.310);
\draw[maplegrey] (0.396,0.291) -- (0.397,0.252);
\draw[maplegrey] (0.396,0.291) -- (0.429,0.272);
\draw[maplegrey] (0.429,0.310) -- (0.461,0.328);
\draw[maplegrey] (0.429,0.310) -- (0.429,0.272);
\draw[maplegrey] (0.429,0.310) -- (0.461,0.292);
\draw[maplegrey] (0.461,0.328) -- (0.491,0.345);
\draw[maplegrey] (0.461,0.328) -- (0.461,0.292);
\draw[maplegrey] (0.461,0.328) -- (0.492,0.310);
\draw[maplegrey] (0.491,0.345) -- (0.521,0.362);
\draw[maplegrey] (0.491,0.345) -- (0.492,0.310);
\draw[maplegrey] (0.491,0.345) -- (0.521,0.329);
\draw[maplegrey] (0.521,0.362) -- (0.521,0.329);
\draw[maplegrey] (0.521,0.362) -- (0.550,0.346);
\draw[maplegrey] (0.080,0.060) -- (0.126,0.088);
\draw[maplegrey] (0.080,0.060) -- (0.128,0.035);
\draw[maplegrey] (0.126,0.088) -- (0.169,0.114);
\draw[maplegrey] (0.126,0.088) -- (0.128,0.035);
\draw[maplegrey] (0.126,0.088) -- (0.171,0.063);
\draw[maplegrey] (0.169,0.114) -- (0.211,0.139);
\draw[maplegrey] (0.169,0.114) -- (0.171,0.063);
\draw[maplegrey] (0.169,0.114) -- (0.213,0.091);
\draw[maplegrey] (0.211,0.139) -- (0.251,0.163);
\draw[maplegrey] (0.211,0.139) -- (0.213,0.091);
\draw[maplegrey] (0.211,0.139) -- (0.253,0.117);
\draw[maplegrey] (0.251,0.163) -- (0.290,0.187);
\draw[maplegrey] (0.251,0.163) -- (0.253,0.117);
\draw[maplegrey] (0.251,0.163) -- (0.292,0.142);
\draw[maplegrey] (0.290,0.187) -- (0.327,0.209);
\draw[maplegrey] (0.290,0.187) -- (0.292,0.142);
\draw[maplegrey] (0.290,0.187) -- (0.328,0.167);
\draw[maplegrey] (0.327,0.209) -- (0.362,0.231);
\draw[maplegrey] (0.327,0.209) -- (0.328,0.167);
\draw[maplegrey] (0.327,0.209) -- (0.364,0.190);
\draw[maplegrey] (0.362,0.231) -- (0.397,0.252);
\draw[maplegrey] (0.362,0.231) -- (0.364,0.190);
\draw[maplegrey] (0.362,0.231) -- (0.398,0.212);
\draw[maplegrey] (0.397,0.252) -- (0.429,0.272);
\draw[maplegrey] (0.397,0.252) -- (0.398,0.212);
\draw[maplegrey] (0.397,0.252) -- (0.431,0.234);
\draw[maplegrey] (0.429,0.272) -- (0.461,0.292);
\draw[maplegrey] (0.429,0.272) -- (0.431,0.234);
\draw[maplegrey] (0.429,0.272) -- (0.463,0.255);
\draw[maplegrey] (0.461,0.292) -- (0.492,0.310);
\draw[maplegrey] (0.461,0.292) -- (0.463,0.255);
\draw[maplegrey] (0.461,0.292) -- (0.493,0.275);
\draw[maplegrey] (0.492,0.310) -- (0.521,0.329);
\draw[maplegrey] (0.492,0.310) -- (0.493,0.275);
\draw[maplegrey] (0.492,0.310) -- (0.523,0.294);
\draw[maplegrey] (0.521,0.329) -- (0.550,0.346);
\draw[maplegrey] (0.521,0.329) -- (0.523,0.294);
\draw[maplegrey] (0.128,0.035) -- (0.171,0.063);
\draw[maplegrey] (0.171,0.063) -- (0.213,0.091);
\draw[maplegrey] (0.171,0.063) -- (0.217,0.042);
\draw[maplegrey] (0.213,0.091) -- (0.253,0.117);
\draw[maplegrey] (0.213,0.091) -- (0.217,0.042);
\draw[maplegrey] (0.213,0.091) -- (0.257,0.070);
\draw[maplegrey] (0.253,0.117) -- (0.292,0.142);
\draw[maplegrey] (0.253,0.117) -- (0.257,0.070);
\draw[maplegrey] (0.253,0.117) -- (0.295,0.097);
\draw[maplegrey] (0.292,0.142) -- (0.328,0.167);
\draw[maplegrey] (0.292,0.142) -- (0.295,0.097);
\draw[maplegrey] (0.292,0.142) -- (0.332,0.123);
\draw[maplegrey] (0.328,0.167) -- (0.364,0.190);
\draw[maplegrey] (0.328,0.167) -- (0.332,0.123);
\draw[maplegrey] (0.328,0.167) -- (0.367,0.148);
\draw[maplegrey] (0.364,0.190) -- (0.398,0.212);
\draw[maplegrey] (0.364,0.190) -- (0.367,0.148);
\draw[maplegrey] (0.364,0.190) -- (0.401,0.172);
\draw[maplegrey] (0.398,0.212) -- (0.431,0.234);
\draw[maplegrey] (0.398,0.212) -- (0.401,0.172);
\draw[maplegrey] (0.398,0.212) -- (0.434,0.195);
\draw[maplegrey] (0.431,0.234) -- (0.463,0.255);
\draw[maplegrey] (0.431,0.234) -- (0.434,0.195);
\draw[maplegrey] (0.431,0.234) -- (0.466,0.217);
\draw[maplegrey] (0.463,0.255) -- (0.493,0.275);
\draw[maplegrey] (0.463,0.255) -- (0.466,0.217);
\draw[maplegrey] (0.463,0.255) -- (0.496,0.239);
\draw[maplegrey] (0.493,0.275) -- (0.523,0.294);
\draw[maplegrey] (0.493,0.275) -- (0.496,0.239);
\draw[maplegrey] (0.217,0.042) -- (0.257,0.070);
\draw[maplegrey] (0.217,0.042) -- (0.262,0.022);
\draw[maplegrey] (0.257,0.070) -- (0.295,0.097);
\draw[maplegrey] (0.257,0.070) -- (0.262,0.022);
\draw[maplegrey] (0.257,0.070) -- (0.300,0.051);
\draw[maplegrey] (0.295,0.097) -- (0.332,0.123);
\draw[maplegrey] (0.295,0.097) -- (0.300,0.051);
\draw[maplegrey] (0.295,0.097) -- (0.337,0.079);
\draw[maplegrey] (0.332,0.123) -- (0.367,0.148);
\draw[maplegrey] (0.332,0.123) -- (0.337,0.079);
\draw[maplegrey] (0.332,0.123) -- (0.372,0.105);
\draw[maplegrey] (0.367,0.148) -- (0.401,0.172);
\draw[maplegrey] (0.367,0.148) -- (0.372,0.105);
\draw[maplegrey] (0.367,0.148) -- (0.406,0.131);
\draw[maplegrey] (0.401,0.172) -- (0.434,0.195);
\draw[maplegrey] (0.401,0.172) -- (0.406,0.131);
\draw[maplegrey] (0.401,0.172) -- (0.438,0.156);
\draw[maplegrey] (0.434,0.195) -- (0.466,0.217);
\draw[maplegrey] (0.434,0.195) -- (0.438,0.156);
\draw[maplegrey] (0.434,0.195) -- (0.470,0.179);
\draw[maplegrey] (0.466,0.217) -- (0.496,0.239);
\draw[maplegrey] (0.466,0.217) -- (0.470,0.179);
\draw[maplegrey] (0.262,0.022) -- (0.300,0.051);
\draw[maplegrey] (0.300,0.051) -- (0.337,0.079);
\draw[maplegrey] (0.300,0.051) -- (0.343,0.034);
\draw[maplegrey] (0.337,0.079) -- (0.372,0.105);
\draw[maplegrey] (0.337,0.079) -- (0.343,0.034);
\draw[maplegrey] (0.337,0.079) -- (0.378,0.062);
\draw[maplegrey] (0.372,0.105) -- (0.406,0.131);
\draw[maplegrey] (0.372,0.105) -- (0.378,0.062);
\draw[maplegrey] (0.372,0.105) -- (0.411,0.089);
\draw[maplegrey] (0.406,0.131) -- (0.438,0.156);
\draw[maplegrey] (0.406,0.131) -- (0.411,0.089);
\draw[maplegrey] (0.406,0.131) -- (0.444,0.115);
\draw[maplegrey] (0.438,0.156) -- (0.470,0.179);
\draw[maplegrey] (0.438,0.156) -- (0.444,0.115);
\draw[maplegrey] (0.438,0.156) -- (0.475,0.141);
\draw[maplegrey] (0.470,0.179) -- (0.475,0.141);
\draw[maplegrey] (0.343,0.034) -- (0.378,0.062);
\draw[maplegrey] (0.343,0.034) -- (0.386,0.018);
\draw[maplegrey] (0.378,0.062) -- (0.411,0.089);
\draw[maplegrey] (0.378,0.062) -- (0.386,0.018);
\draw[maplegrey] (0.378,0.062) -- (0.419,0.047);
\draw[maplegrey] (0.411,0.089) -- (0.444,0.115);
\draw[maplegrey] (0.411,0.089) -- (0.419,0.047);
\draw[maplegrey] (0.411,0.089) -- (0.451,0.075);
\draw[maplegrey] (0.444,0.115) -- (0.475,0.141);
\draw[maplegrey] (0.444,0.115) -- (0.451,0.075);
\draw[maplegrey] (0.386,0.018) -- (0.419,0.047);
\draw[maplegrey] (0.419,0.047) -- (0.451,0.075);
\draw[maplegrey] (0.419,0.047) -- (0.460,0.034);
\draw[maplegrey] (0.451,0.075) -- (0.460,0.034);
\end{tikzpicture}
\end{center}
All the charts are based on polynomials of degree $16$. We find that,
on the disc of radius $0.1$, the entries in $p^*(g_1)-g_{\hyp}$ are of
absolute value less than $10^{-6}$.
We next need to record the combinatorial structure of the above grid.
If there is an edge joining the centre of chart $i$ to the centre of
chart $j$, we need to invoke the method \mcode+EHA["add_edge",i,j]+.
(Charts are numbered from $0$, in the order that they were added to
the atlas.) This is also done by the function
\mcode+build_data["EH_atlas",2]()+.
Next, for each chart $p_i$, we need to find the point
$\bt_i=q^{-1}p_i(0)\in\Dl$. For charts that are centred on one of the
curves $C_k$, this is given by Proposition~\ref{prop-curve-chart}.
For the remaining charts, it is useful to start with a crude
approximation, obtained by applying an essentially arbitrary
diffeomorphism between the fundamental domains for $EX^*$ and
$HX(b)$. This is done using the method
\mcode+EHA["set_beta_approx"]+. We then invoke the method
\mcode+EHA["set_edge_lengths"]+. This calculates various quantities
for each edge $(i,j)$. In particular, it calculates the average of
$d_{\hyp}(0,q_i^{-1}(q_j(0)))$ and $d_{\hyp}(0,q_j^{-1}(q_i(0)))$,
which is an estimate of the hyperbolic distance in $EX^*$ between the
centres $q_i(0)$ and $q_j(0)$. Each edge is actually represented by
an object \mcode+E+ of class \mcode+EH_atlas_edge+, and this distance
is stored as \mcode+E["EH_length"]+. On the other hand,
\mcode+E["H_length"]+ is set equal to $d_{\hyp}(\bt_i,\bt_j)$, using the
approximate values of $\bt_i$ and $\bt_j$, which may be quite
inaccurate at this stage. One can then invoke the method
\mcode+EHA["optimize_beta"]+ to adjust the values of $\bt_i$ so as
optimize the match between the edge lengths measured in $\Dl$ and in
$EX^*$. The same method also calculates appropriate values for the
parameters $\lm_i$, and thus also for the M\"obius maps $m_i$.
Now all the maps $p_im_i^{-1}$ are approximations to $q$, and it is
useful to test how well they agree with each other. The method
\mcode+EHA["make_H_samples",N]+ sets \mcode+EHA["H_samples"]+ to be
the list of all numbers $z=(s+it)/N$ (with $s,t\in\Z$) that lie in
$HF_{16}(b)$. The method \mcode+EHA["max_patching_error",r]+ then
does the following. For each point $z_i$ in \mcode+EHA["H_samples"]+,
it looks for charts $p_j$ where $|z_i-\bt_j|<r$. Let $k_i$ be the
number of such charts. For each such chart, the method
calculates $x_{ij}=p_jm_j^{-1}(z_i)\in EX^*$. These points should
all be the same, so we let $d_i$ denote the maximum euclidean
distance between any two of them. The return value of the method is
a triple $(z_i,m_i,d_i)$, where $d_i$ is maximal. If we take
$r=0.12$, we find that $m_i\geq 3$ and $d_i<10^{-10.4}$ for all $i$.
Thus, for an arbitrary point $z\in HF_{16}(b)$, it is safe to
calculate $q(z)$ as $p_jm_j^{-1}(z)$, where $j$ is chosen to minimize
$d_{\hyp}(z,\bt_j)$. We can then extend this over all of $\Dl$ by
using the group action, as discussed earlier. This is implemented by
the methods \mcode+EHA["q",[x,y]]+ or \mcode-EHA["q_c",x+I*y]-.
We now want to find a function given by a single formula which is a
good approximation to $q$ on a reasonably large part of $\Dl$, such as
the disc of radius $0.9$ centred at the origin. An obvious approach
would be to approximate $q(x+iy)_k$ (for $1\leq k\leq 4$) by a
polynomial or rational function in $x$ and $y$. This is implemented
by the methods \mcode+set_q_approx_poly+ and \mcode+set_q_approx_pade+
of the class \mcode+EH_atlas+. However, results from this
approach are poor. The approximating polynomials have extremely large
coefficients (of different signs), even though $|q(x+iy)_k|\leq 1$,
and the errors are fairly large even if we use polynomials or rational
functions of high degree. It is better to consider the Fourier series
on circles of fixed radius. To understand how this works, we first
recall that $q$ is equivariant with respect to $\ip{\lm,\nu}$, which
gives
\begin{align*}
q_1(re^{i(\tht+\pi/2)}) &= -q_2(re^{i\tht}) &
q_1(re^{-i\tht}) &= \pp q_1(re^{i\tht}) \\
q_2(re^{i(\tht+\pi/2)}) &= \pp q_1(re^{i\tht}) &
q_2(re^{-i\tht}) &= -q_2(re^{i\tht}) \\
q_3(re^{i(\tht+\pi/2)}) &= \pp q_3(re^{i\tht}) &
q_3(re^{-i\tht}) &= \pp q_3(re^{i\tht}) \\
q_4(re^{i(\tht+\pi/2)}) &= -q_4(re^{i\tht}) &
q_4(re^{-i\tht}) &= \pp q_4(re^{i\tht}).
\end{align*}
From this it follows that there are functions $a_{k,m}(r)$ such that
\begin{align*}
q_1(r e^{i\tht}) &= \sum_m a_{1,m}(r) \cos((2m+1)\tht) \\
q_2(r e^{i\tht}) &= \sum_m (-1)^m a_{1,m}(r) \sin((2m+1)\tht) \\
q_3(r e^{i\tht}) &= \sum_m a_{3,m}(r) \cos(4m\tht) \\
q_4(r e^{i\tht}) &= \sum_m a_{4,m}(r) \cos((4m+2)\tht).
\end{align*}
We can find approximations to the coefficients $a_{j,l}(r)$ by fixing
$k\geq 0$, then calculating $q(re^{2\pi ij/2^k})$ for $0\leq j<2^k$, then
taking a discrete Fourier transform. This algorithm is implemented by the
method \mcode+EHA["set_q_approx_fourier",r_max,m,k]+, which calculates
Fourier coefficients for \mcode+m+ different radii, the largest being
\mcode+r_max+.
It seems experimentally that for $r\leq 0.9$ we have $|a_{j,m}(r)|\leq
2$ for all $j$ and $m$, and that $|a_{j,m}(r)|$ decreases quite
rapidly with $m$. Thus, for a fixed value of $r$, the above
representation is quite satisfactory. Now fix $k$ and $m$, and
consider $a_{k,m}(r)$ as a function of $r$. The following picture
shows a typical sample of these functions, for $0\leq r\leq 0.9$.
\begin{center}
\begin{tikzpicture}[scale=8]
\draw[->] (-0.05, 0.00) -- (0.95,0.00);
\draw[->] ( 0.00,-0.30) -- (0.00,0.30);
\draw[smooth,red] (0.028,0.000) -- (0.057,0.000) -- (0.085,0.000) -- (0.113,0.000) -- (0.141,0.000) -- (0.169,0.000) -- (0.196,0.000) -- (0.224,0.000) -- (0.251,0.000) -- (0.278,0.000) -- (0.305,0.000) -- (0.331,0.000) -- (0.357,0.000) -- (0.383,-0.000) -- (0.409,-0.000) -- (0.434,-0.000) -- (0.458,0.001) -- (0.482,0.003) -- (0.506,0.007) -- (0.529,0.012) -- (0.552,0.021) -- (0.574,0.032) -- (0.595,0.046) -- (0.616,0.063) -- (0.636,0.081) -- (0.656,0.102) -- (0.675,0.124) -- (0.693,0.145) -- (0.711,0.166) -- (0.728,0.184) -- (0.744,0.199) -- (0.760,0.211) -- (0.775,0.217) -- (0.789,0.218) -- (0.802,0.214) -- (0.814,0.205) -- (0.826,0.192) -- (0.837,0.176) -- (0.847,0.159) -- (0.856,0.141) -- (0.864,0.123) -- (0.872,0.107) -- (0.878,0.093) -- (0.884,0.080) -- (0.889,0.069) -- (0.893,0.061) -- (0.896,0.054) -- (0.898,0.049) -- (0.900,0.046) -- (0.900,0.046) ;
\draw[smooth,red] (0.028,-0.000) -- (0.057,-0.000) -- (0.085,-0.000) -- (0.113,-0.000) -- (0.141,-0.000) -- (0.169,-0.000) -- (0.196,-0.000) -- (0.224,-0.000) -- (0.251,-0.000) -- (0.278,-0.000) -- (0.305,-0.000) -- (0.331,-0.000) -- (0.357,-0.000) -- (0.383,-0.000) -- (0.409,-0.000) -- (0.434,-0.001) -- (0.458,-0.001) -- (0.482,-0.001) -- (0.506,-0.001) -- (0.529,-0.001) -- (0.552,0.000) -- (0.574,0.002) -- (0.595,0.005) -- (0.616,0.009) -- (0.636,0.015) -- (0.656,0.024) -- (0.675,0.034) -- (0.693,0.048) -- (0.711,0.064) -- (0.728,0.083) -- (0.744,0.103) -- (0.760,0.125) -- (0.775,0.146) -- (0.789,0.166) -- (0.802,0.184) -- (0.814,0.198) -- (0.826,0.208) -- (0.837,0.214) -- (0.847,0.217) -- (0.856,0.218) -- (0.864,0.216) -- (0.872,0.214) -- (0.878,0.211) -- (0.884,0.208) -- (0.889,0.205) -- (0.893,0.203) -- (0.896,0.201) -- (0.898,0.200) -- (0.900,0.199) -- (0.900,0.198) ;
\draw[smooth,red] (0.028,0.000) -- (0.057,0.000) -- (0.085,-0.000) -- (0.113,-0.000) -- (0.141,0.000) -- (0.169,0.000) -- (0.196,0.000) -- (0.224,0.000) -- (0.251,0.000) -- (0.278,-0.000) -- (0.305,-0.000) -- (0.331,-0.000) -- (0.357,-0.000) -- (0.383,-0.000) -- (0.409,-0.000) -- (0.434,-0.000) -- (0.458,-0.000) -- (0.482,-0.000) -- (0.506,-0.001) -- (0.529,-0.001) -- (0.552,-0.002) -- (0.574,-0.003) -- (0.595,-0.004) -- (0.616,-0.006) -- (0.636,-0.009) -- (0.656,-0.012) -- (0.675,-0.017) -- (0.693,-0.022) -- (0.711,-0.028) -- (0.728,-0.036) -- (0.744,-0.044) -- (0.760,-0.055) -- (0.775,-0.068) -- (0.789,-0.082) -- (0.802,-0.099) -- (0.814,-0.117) -- (0.826,-0.136) -- (0.837,-0.156) -- (0.847,-0.174) -- (0.856,-0.191) -- (0.864,-0.207) -- (0.872,-0.220) -- (0.878,-0.232) -- (0.884,-0.242) -- (0.889,-0.250) -- (0.893,-0.256) -- (0.896,-0.261) -- (0.898,-0.265) -- (0.900,-0.267) -- (0.900,-0.267) ;
\draw[smooth,red] (0.028,-0.000) -- (0.057,-0.000) -- (0.085,0.000) -- (0.113,-0.000) -- (0.141,-0.000) -- (0.169,0.000) -- (0.196,-0.000) -- (0.224,-0.000) -- (0.251,-0.000) -- (0.278,-0.000) -- (0.305,-0.000) -- (0.331,-0.000) -- (0.357,-0.000) -- (0.383,0.000) -- (0.409,0.000) -- (0.434,0.000) -- (0.458,0.000) -- (0.482,0.000) -- (0.506,0.000) -- (0.529,0.000) -- (0.552,0.000) -- (0.574,0.000) -- (0.595,0.000) -- (0.616,0.001) -- (0.636,0.001) -- (0.656,0.002) -- (0.675,0.002) -- (0.693,0.003) -- (0.711,0.003) -- (0.728,0.003) -- (0.744,0.002) -- (0.760,0.000) -- (0.775,-0.003) -- (0.789,-0.009) -- (0.802,-0.016) -- (0.814,-0.023) -- (0.826,-0.032) -- (0.837,-0.041) -- (0.847,-0.049) -- (0.856,-0.057) -- (0.864,-0.064) -- (0.872,-0.070) -- (0.878,-0.075) -- (0.884,-0.081) -- (0.889,-0.085) -- (0.893,-0.089) -- (0.896,-0.093) -- (0.898,-0.095) -- (0.900,-0.097) -- (0.900,-0.098) ;
\draw ( 0.900, 0.046) node[anchor=west] {$a_{1,3}$};
\draw ( 0.900, 0.198) node[anchor=west] {$a_{1,5}$};
\draw ( 0.900,-0.267) node[anchor=west] {$a_{1,8}$};
\draw ( 0.900,-0.098) node[anchor=west] {$a_{1,10}$};
\draw (-0.020, 0.200) -- ( 0.000, 0.200);
\draw (-0.020, 0.200) node[anchor=east] {$0.2$};
\draw ( 0.900, 0.000) -- ( 0.900,-0.020);
\end{tikzpicture}
\end{center}
These functions are extremely flat for small values of $r$, and grow
to modest size as $r$ approaches $0.9$. They cannot easily be
approximated by polynomials or rational functions in $r$. We have
tried various transformations, such as using the variable
$s=1/\log(1/r)$ instead of $r$, but none of these have yielded
compelling results. It would be of interest to have a better
theoretical understanding of the asymptotics of the functions
$a_{k,m}(r)$, but so far we have not achieved that. However,
approximation by cubic splines (which are also calculated by the
\mcode+set_q_approx_fourier+ method) is quite effective. An
approximation to $q(x+iy)$ using these splines can be calculated by
invoking \mcode+EHA["q_fourier",[x,y]]+ or
\mcode-EHA["q_fourier_c",x+I*y]-.
The following pictures show the images of $q(0.8e^{it})$ under the
linear projections $\pi,\dl,\zt\:EX^*\to\R^2$ that were discussed in
Section~\ref{sec-disc}:
\[ \includegraphics[scale=0.12]{images/pi_ring.jpg} \hspace{4em}
\includegraphics[scale=0.12]{images/delta_ring.jpg} \hspace{4em}
\includegraphics[scale=0.12]{images/zeta_ring.jpg}
\]
The following picture shows the image under the map $p_4\:x\mapsto y$
from Proposition~\ref{prop-F-four}:
\[ \includegraphics[scale=0.2]{images/y_ring.jpg} \]
Finally, we discuss the canonical conformal map $\hp\:EX^*\to S^2$.
The transformation properties of $\hp$ were discussed in
Remark~\ref{rem-p-hat}. By comparing these with the transformation
properties of the basis in Proposition~\ref{prop-OX-basis}, we see
that $\hp(x)_i$ must be given by some $G$-invariant function
$u_i(z_1,z_2)$ multiplied by $p^*(x)_i$, where
\[ p^*(x) = \left(\rt \,y_2,\; 2x_1x_2,\; -x_3) \right). \]
Remark~\ref{rem-p-hat} also records the values of $\hp(v_i)$ for
$0\leq i\leq 9$; these are equivalent to the conditions
\[ u_1(0,1/2) = u_2(0,0) = u_3(1,0) = 1. \]
It turns out that the functions $u_i$ can be approximated effectively
by rational functions, using the method that we will now explain. We
have already seen how to produce a list of points $a^\Dl_i$ in
$HF_{16}(b)$ and calculate the images $a^E_i=q(a^\Dl_i)$ under the map
$q\:\Dl\to EX^*$. By applying the map $x\mapsto z$ to these, we
obtain points $a^Z_i\in F^*_{16}\subset\R^2$. We have also seen how
to calculate the isomorphism $HX(b)\to PX(a)$. By combining this with
the projection $PX(a)\to\C_\infty$ and the stereographic projection
map $\C_\infty\to S^2$ we obtain points $a_i^S\in S^2$. We must have
$\hp(a^E_i)=a^S_i$, and using this, we can find the values
$a^U_{ij}=u_j(a^Z_i)=\hp(a^E_i)_j/p^*(a^E_i)_j$ for $1\leq j\leq 3$.
Our problem is thus to find rational functions $u_j^R$ such that
$u_j^R(a^Z_i)$ is close to $a^U_{ij}$. In principle this should hold
for all $i$, but we have found it best to discard the cases where
$|p^*(a^E_i)_j|<10^{-3}$, in order to avoid numerical instability.
We have found the following general approach to be effective.
\begin{method}
Suppose we have a finite set $S=\{s_1,\dotsc,s_n\}$, and a function
$f\:S\to\R$. We want to find an approximation $f\simeq g_1/g_2$,
where $g_1$ and $g_2$ lie in some vector space $V\leq\Map(S,\R)$. If
$n$ is large then it is not very tractable to minimize
$\|f-g_1/g_2\|^2$, because this is nonlinear in the coefficients of
$g_2$. However, it will often be adequate to minimize $\|fg_2-g_1\|^2$
subject to a positive definite quadratic constraint on the size of
$g_2$. To do this, let $\{v_1,\dotsc,v_m\}$ be a basis for $V$. Let
$M_1$ be the matrix of values $v_j(s_i)$, and let $M_2$ be the matrix
of values $f(s_i)v_j(s_i)$. We find the QR decompositions
$M_i=Q_iR_i$, and put $N=Q_1^TQ_2$. We let $a_2$ denote a singular
vector for $N$ of maximal singular value with $\|a_2\|=1$, and put
$a_1=Na_2$. We then put $b_i=R_i^{-1}a_i$ and
$g_i=\sum_jb_{ij}v_j$. The function $g_1/g_2$ is then the desired
approximation.
\end{method}
The above algorithm is implemented by the \mcode+find_p+ method of the
class \mcode+E_to_S_map+, which is declared in
\fname+embedded/roothalf/E_to_S.mpl+. This method must be passed an
object of class \mcode+EH_atlas+, which encodes information about the
points $a^\Dl_i$ and the maps $\Dl\to EX^*$ and $\Dl\to\C_\infty$. We
have used this method to find rational approximations $u_j^R$ where the
numerator and denominator have total degree eight in $z_1$ and $z_2$.
It turns out that these functions are quite tame, as shown in the
graphs below.
\[ \includegraphics[scale=0.16]{images/E_to_S_u1.jpg}
\includegraphics[scale=0.16]{images/E_to_S_u2.jpg}
\includegraphics[scale=0.16]{images/E_to_S_u3.jpg}
\]
All values lie between about $0.3$ and $2.3$. All coefficients in the
numerators and denominators have absolute value at most one, and they
appear to decrease quite rapidly with the total degree of the
corresponding monomials. The full calculation can be carried out using
the function \mcode+build_data["E_to_S_map"]()+, which is defined in
\fname+build_data.mpl+.
\subsection{Energy minimisation}
\label{sec-energy}
As explained in Remark~\ref{rem-p-hat}, there is a canonical conformal
map $\hp\:EX^*\to S^2$. If we can find $\hp$, then all other
information can easily be derived from that. One approach is to start
with the map $\hp\:EX^*\to S^2$ from
Definition~\ref{defn-sphere-quotient-a}. This has the right
equivariance properties and the right homotopy class, but it is not
conformal. We can hope to adjust it by a numerical minimisation
algorithm to make it conformal. For this, we need to recall the
theory of Dirichlet energy.
\begin{definition}
Given matrices $P,Q\in M_2(\R)$ we put
\[ \ip{P,Q} = \sum_{i,j=1}^2 P_{ij}Q_{ij} = \text{trace}(P^TQ). \]
This is an inner product, with associated norm
$\|P\|^2=\sum_{i,j}P_{ij}^2$. We also put
\begin{align*}
C_+(P) &= (P_{11}-P_{22})^2 + (P_{12}+P_{21})^2 \\
C_-(P) &= (P_{11}+P_{22})^2 + (P_{12}-P_{21})^2.
\end{align*}
\end{definition}
\begin{remark}
It is clear that $C_+(P)=0$ iff $P=\bbm a&b\\ -b&a\ebm$ for some
$a,b\in\R$, or in other words $P$ is a conformal matrix. Similarly,
we have $C_-(P)=0$ iff $P$ is anticonformal.
\end{remark}
\begin{lemma}
For all $P\in M_2(\R)$ we have
\begin{align*}
\|P\|^2 &= C_+(P) + 2\det(P) \geq 2\det(P) \\
\|P\|^2 &= C_-(P) - 2\det(P) \geq -2\det(P),
\end{align*}
so $\|P\|^2\geq 2|\det(P)|$.
\end{lemma}
\begin{proof}
The equalities are direct calculations, and it is clear that
$C_+(P)\geq 0$ and $C_-(P)\geq 0$.
\end{proof}
\begin{corollary}\label{cor-invariants}
For $A,B\in SO(2)$ we have $\|APB\|^2=\|P\|^2$ and
$C_{\pm}(APB)=C_{\pm}(P)$ and $\det(APB)=\det(P)$.
\end{corollary}
\begin{proof}
We have $\det(A)=\det(B)=1$ so
$\det(APB)=\det(A)\det(P)\det(B)=\det(P)$. We also have
$A^TA=B^TB=1$ and $\text{trace}(XY)=\text{trace}(YX)$ so
\[ \|APB\|^2 = \text{trace}(B^TP^TA^TAPB)
= \text{trace}(B^TP^TPB)
= \text{trace}(BB^TP^TP)
= \text{trace}(P^TP) = \|P\|^2.
\]
We can now use $C_{\pm}(P)=\|P\|^2\mp 2\det(P)$ to deduce that
$C_{\pm}(APB)=C_{\pm}(P)$.
\end{proof}
\begin{definition}
Let $V$ and $W$ be oriented two-dimensional inner product spaces over
$\R$, and let $\phi\:V\to W$ be a linear map. We then choose
oriented orthonormal bases $v_1,v_2$ for $V$ and $w_1,w_2$ for $W$,
and let $P$ be the matrix such that
\begin{align*}
\phi(v_1) &= P_{11}w_1 + P_{21}w_2 \\
\phi(v_2) &= P_{12}w_1 + P_{22}w_2.
\end{align*}
We then put $\|\phi\|^2=\|P\|^2$ and $\det(\phi)=\det(P)$ and
$C_{\pm}(\phi)=C_{\pm}(P)$. This is independent of the choice of
bases, by Corollary~\ref{cor-invariants}.
\end{definition}
\begin{definition}\label{defn-energy}
Let $X$ and $Y$ be connected oriented smooth closed surfaces with
given Riemannian metrics. We give them the measures derived from the
metric in the usual way. Consider a smooth map $f\:X\to Y$. For
each $x\in X$ we have a linear map $D_xf\:T_xX\to T_{f(x)}Y$ between
oriented two-dimensional inner product spaces, so we can define
$\|D_xf\|^2$ and $\det(D_xf)$ and $C_{\pm}(D_xf)$. The
\emph{Dirichlet energy} of $f$ is the integral over $X$ of the
scalar-valued function $x\mapsto\half\|D_xf\|^2$. We write this as
$E(f)=\int_X\half\|Df\|^2$. We also define the \emph{area} of $f$ to
be $A(f)=\int_X\det(Df)$.
\end{definition}
\begin{remark}
In terms of differential forms, we can let $\om_X$ and $\om_Y$ denote
the volume forms for $X$ and $Y$, and then
$f^*(\om_Y)=\det(Df)\om_X$. We can regard $\om_X$ as a generator of
the de Rham cohomology group $H^2(X)$, and similarly for $Y$.
Integration gives an isomorphism from each of these cohomology groups
to the reals. From this point of view it is clear that $A(f)$
depends only on the homotopy class of $f$. Moreover, if $f$ is an
orientation-preserving diffeomorphism then the standard
change-of-variables formula shows that $A(f)$ is just the area of
$Y$.
\end{remark}
\begin{proposition}\label{prop-dirichlet}
For any $f\:X\to Y$ as above, we have $E(f)\geq A(f)$, with equality
iff $f$ is conformal.
\end{proposition}
\begin{proof}
This is clear from the identity $\|Df\|^2=2\det(Df)+C_+(Df)$.
\end{proof}
One way to exploit this is via a discretised version. We triangulate
the fundamental domain $F_{16}\subset EX^*$, then use the group action
to obtain an equivariant triangulation of $EX^*$, with vertex set
$K_0$ say. Taking the convex hulls of vertices of simplices gives a
piecewise linear surface $X'\subset\R^4$ that lies close to $EX^*$.
If we have a map $f\:K_0\to S^2$, then we can extend it linearly to
give a map $f'\:X'\to\R^3$, which will usually land in $\R^3\sm\{0\}$.
There is an obvious retraction of $\R^3\sm\{0\}$ onto $S^2$, and one
can also construct a map from $EX^*$ to $X'$ that is close to the
identity. After composing with these we get a map $f''\:EX^*\to S^2$.
One could attempt to minimise $E(f'')$, but that is analytically
intractable. However, a slight modification of
Definition~\ref{defn-energy} defines a quantity $E'(f')$ that is
analogous to $E(f)$, and we can attempt to minimise that instead. We
find that the rate of convergence is slow, and the resulting
approximation is inaccurate close to the points $v_{10},\dotsc,v_{13}$
where the equivariance properties force the derivative of $\hp$ to be
zero. One can improve the accuracy of the method by subdividing the
triangulation, but this makes everything much slower. One could also
use an approximation scheme that is better than linear interpolation,
perhaps based on the various types of splines that are popular in
computer graphics. However, we did not find an approach of this type
that worked well for our purposes.
We could also avoid discretisation, and instead attempt to minimise
the energy over some finite-dimensional space $M$ of maps
$EX^*\to S^2$. This should ideally be chosen so that it is easy to
calculate $\|Df\|^2$ for $f\in M$, together with the derivatives of
$\|Df\|^2$ with respect to suitable coordinates on $M$. We have not
found a space $M$ for which this works nicely.
\section{Overview of the Maple code}
\label{sec-maple}
\subsection{Directory structure}
The main directory for this project has subdirectories as follows:
\begin{itemize}
\item \fname+latex:+ \LaTeX code for this document. The subdirectory
\fname+tikz_includes+ contains some files that were generated by
Maple and are included in the main \LaTeX document by \mcode+\input+
commands. The code in the file \fname+maple/plots.mpl+ is relevant
here.
\item \fname+images:+ Image files (with extensions \mcode+.png+ or
\mcode+.jpg+), all generated by Maple. The code in the file
\fname+maple/plots.mpl+ is relevant here.
\item \fname+plots:+ Image files (with extensions \mcode+.m+) in
Maple's internal format for plots. The code in the file
\fname+maple/plots.mpl+ is relevant here.
\item \fname+maple:+ This contains Maple code in various
subdirectories, which will be described in more detail below. The
files contain plain text, and have extension \fname+.mpl+. Many
files occur in pairs like \fname+projective/ellquot.mpl+ (which
defines various functions related to elliptic curve quotients of
$PX(a)$) an \fname+projective/ellquot_check.mpl+ (which defines
procedures to check various assertions about those functions).
\item \fname+doc:+ This contains various kinds of documentation of the
Maple code, in HTML format. Some of the Maple code is object
oriented (as will be discussed in Section~\ref{sec-oo-maple} below)
and some is not. For the object oriented code, there is
automatically generated documentation of classes, fields and
methods, similar to the standard javadoc framework for Java code.
For the remaining code, there is an index of definitions of all
defined symbols, with links to the defining files.
\item \fname+worksheets:+ This contains Maple worksheets, with
extension \fname+.mw+. They all start with the following block:
\begin{mcodeblock}
restart;
interface(quiet=true):
olddir := currentdir("../maple"):
read("genus2.mpl"):
currentdir(olddir):
interface(quiet=false):
\end{mcodeblock}
Executing this block will read in the file \fname+genus2.mpl+, which
will in turn read in many other files from the \fname+maple+
directory and its subdirectories. We have mostly used worksheets
for development, and have moved code to the \fname+.mpl+ files when
it has become stable.
\item \fname+data:+ This contains files generated by Maple recording
the results of certain complex calculations. There is a hierarchy
of subdirectories parallel to those in the \fname+maple+ directory.
Some are plain text files with extension \fname+.mpl+, but most are
in Maple's internal format and have extension \fname+.m+. See
Section~\ref{sec-build} for information about how to recalculate
these results.
\end{itemize}
The subdirectories of the \fname+maple+ directory are as follows:
\begin{itemize}
\item The top level directory contains some code about the general
theory of precromulent surfaces (not tied to any of the three
families), and some general utility code.
\item \fname+projective+: for code related to the projective family.
\item \fname+hyperbolic+: for code related to the hyperbolic family,
and code for isomorphisms between hyperbolic and projective surfaces.
\item \fname+embedded+: for code related to the embedded family.
\item \fname+embedded/roothalf+: for code related to the special case
$EX^*=EX(1/\rt)$.
\item \fname+domain:+ for object oriented code dealing with
triangulations of cromulent surfaces; this can be specialised to
each of the three families. (At an earlier stage, we planned to do
various substantial calculations using triangulations, but we
eventually switched to different methods.)
\item \fname+quadrature:+ object oriented code for quadrature rules
on triangles.
\end{itemize}
\subsection{Checks}
\label{sec-checks}
Maple code that checks the correctness of various assertions is
contained in files whose names end with \fname+_check+. Unlike the
other Maple files, these are not loaded automatically by the standard
block at the top of the worksheets. One can read an individual file
by entering a command like
\begin{mcodeblock}
read("../maple/projective/ellquot_check.mpl"):
\end{mcodeblock}
Alternatively, one can enter
\begin{mcodeblock}
read("../maple/check_all.mpl"):
\end{mcodeblock}
to load and run all possible checks (which takes a long time). The
global variable \mcode+checklist+ contains a list of all the checking
functions that have been loaded. One can execute all of them by
invoking the \mcode+check_all()+ function. These functions will
usually stop running if they encounter an assertion that fails, but
this can be prevented by setting the global variable
\mcode+assert_stop_on_fail+ to \mcode+false+. Each checking function
will print its name when it starts to run. Most functions will check
a large number of assertions; if the global variable
\mcode+assert_verbosely+ is set to \mcode+true+, then a brief identifier
will be printed for each assertion. (This happens after Maple has
done the work of checking the assertion, but before it prints an error
message if the assertion has failed.) The general framework for all
this is set up by the files \fname+util.mpl+ and \fname+checks.mpl+ in
the top Maple directory. The basic claim that the embedded,
projective and hyperbolic families are precromulent has some special
features; see the function \mcode+check_precromulent()+ and associated
comments in the file \fname+cromulent.mpl+.
One might ask about the reliability, rigour and completeness of these
checks.
First, we should explain that almost all checks cover assertions that
are claimed to be exact. For the parts of the code that involve
numerical approximation, we have also performed many checks, but we
have not encapsulated them in a systematic framework.
\begin{itemize}
\item[(a)] Some claims are of the form $u=0$, where $u$ is a constant
expression. Usually $u$ will be built from rational numbers by
algebraic operations and by extraction of square roots of positive
quantities. In some cases we evaluate trigonometric functions and
rational multiples of $\pi$, and there are a few examples involving
roots of polynomials of degree greater than two. In many cases, we
can just use the command \mcode+simplify(u)+ and the result will be
zero. In some cases we need to use a more complicated command like
\begin{mcodeblock}
simplify(factor(expand(rationalize(u))))
\end{mcodeblock}
We have been willing to assume that if a procedure like this returns
zero, then $u$ is genuinely equal to zero. One can check this by
numerical evaluation of $u$. We have used 100 digit precision by
default, and have not found any examples where the symbolic
simplification functions seem to be incorrect.
\item[(b)] Some other claims are of the form $u=0$, where $u$ is an
expression involving several constants and variables, which may be
subject to certain constraints. The constants are of the type
discussed in~(a). Variables may be constrained to be real (using
a command like \mcode+assume(t::real)+) or to lie in the unit
interval (using a command like \mcode+assume(a_H>0 and a_H<1)+).
Most expressions are built using algebraic operations, square roots
and logarithms of quantities that can be shown to be positive,
trigonometric functions and exponentials, and extraction of real
parts of complex numbers. There are also some derivatives and
integrals. There are obvious algorithms to deal with most of these
things, and it seems unlikely that any bugs would have escaped
detection. However, there are some expressions where Maple
implicitly uses a significant amount of logic to determine that
various terms are positive, and uses this to justify manipulations
with roots and logarithms. We do not know what algorithms are used
for this, but we have not detected any problems. All the relevant
symbolic simplifications can be tested by numerical evaluation,
either by plotting or by setting parameters to randomly chosen values.
\item[(c)] There is another kind of constraint that we did not
mention under~(b): variables can be subject to certain polynomial
relations, which can be encoded using a Gr\"obner basis for the
corresponding ideal. We only have examples where the polynomials
have coefficients in $\Q(\rt)$, which is easy to handle. Often,
we only need to check expressions of the form $u=0$, where $u$ is a
rational function of the constrained variables. There are
very standard algorithms for working with Gr\"obner bases, and it is
highly unlikely that there could be any problems with expressions of
this type. It is also common to have expressions $u$ that involve
square roots of polynomials that can be shown to be positive, and
the roots are sometimes nested. The algorithms for this case are
not quite as standard, but again we have detected no problems.
\item[(d)] A few expressions involve more sophisticated functions
such as elliptic integrals and the Weierstrass $\wp$ function. It
is here that we encountered the only significant bug that we have
seen: Maple's numerical evaluation of $\wp'(z)$ is incorrect for
certain ranges of arguments. On the other hand, when we first
started working with $\wp'(z)$, it immediately became clear that
something was causing inconsistent results, although it took time to
locate the precise source of trouble. This raised our confidence
that other bugs would also quickly become visible.
\item[(e)] There are a few cases where we have an indirect reason to
know that an expression $u$ should be zero, but we have not been
able to persuade Maple to simplify $u$ to zero. In these cases, we
have written checking functions that rely completely on numerical
methods, by evaluating $u$ to 100 decimal places at various points
in the parameter space. Alternatively, if $u$ is a function on the
surface $EX^*$, we can evaluate $u$ exactly at all the quasirational
points of $EX^*$ (as discussed in Section~\ref{sec-rational}).
\item[(f)] As well as the kinds of claims discussed in~(a) to~(e), we
have various claims about more combinatorial structures, such as the
groups $G$, $\Pi$ and $\tPi$. For these we have mostly written our
own code, both to implement the definitions and to check the claimed
properties. Thus, very little is hidden in the internals of Maple,
and the sceptical reader can inspect all the relevant code.
\end{itemize}
\subsection{Object oriented Maple}
\label{sec-oo-maple}
For some of our work, it is natural to use an object oriented style of
programming. For example, it is natural to have a class whose objects
represent conformal charts on $EX^*$, and another class for atlases,
and a class for quadrature rules, and so on. We have described a
complex algorithm for calculating the canonical covering
$\Dl\to EX^*$, and it is natural to implement the steps in this
algorithm as methods of various classes. Maple does not natively
support object oriented programming, but we have implemented our own
framework using Maple's system of tables with user-defined indexing
functions. Our framework was in fact developed some years ago for a
rather different project, and adapted slightly for our current
purposes. The relevant code is in the file \fname+class.mpl+ in the
top Maple directory. Typical syntax is as follows:
\begin{itemize}
\item \mcode+Q := `new/E_quadrature_rule`();+ sets $Q$ to be a new
quadrature rule on $F_{16}\subset EX^*$.
\item \mcode+Q["int_z",z[1]];+ given a quadrature rule $Q$,
returns the estimated integral of $z_1$ over $F_{16}$.
\item \mcode+Q["curvature_error"];+ given a quadrature rule $Q$,
returns the difference between the estimated integral of the
curvature and the correct value of $-\pi/4$.
\item \mcode+A["num_charts"];+ given an atlas $A$ on $EX^*$, returns
the number of charts.
\end{itemize}
Classes are declared using the \mcode+`Class/Declare`+ function.
There is one notable place where we have chosen not to use the above
framework. We have a lot of parallel structures for our three
families of cromulent surfaces, for example the functions
\mcode+c_E[k](t)+, \mcode+c_H[k](t)+ and \mcode+c_P[k](t)+ which encode
the three curve systems. In some respects it would be natural to
encapsulate these using a system of classes. However, that would lead
to unwieldy notation for objects that we need to use extremely
frequently, so we chose to avoid it.
\subsection{Building the data}
\label{sec-build}
This project involves some numerical computations, the results of
which are stored in the \fname+data+ directory and its
subdirectories. The file \fname+build_data.mpl+ (in the top Maple
directory) defines various functions that can be used to perform these
calculations. For example, the function
\mcode+build_data["HP_table"]()+ can be used to perform all the
calculations described in Section~\ref{sec-P-H}, relating the
projective and hyperbolic families. The result is encoded as an
object of the class \mcode+HP_table+ (declared in
\fname+hyperbolic/HP_table.mpl+). It can be saved in an appropriate
place using the function \mcode+save_data["HP_table"]()+, and then
reloaded later using the function \mcode+load_data["HP_table"]()+.
The functions \mcode+build_data["all"]()+, \mcode+save_data["all"]()+
and \mcode+load_data["all"]()+ work in the obvious way, building,
saving or loading all of the required data. A full build of all data
will take several days of computer time, at least. However, one can
enter \mcode+set_toy_version(true)+ before invoking
\mcode+build_data["all"]+. This will cause Maple to do all
calculations to lower accuracy, and finish in an hour or two. In this
context, results will be saved to or loaded from the
\fname+data_toy+ directory instead of the \fname+data+ directory.
The build process will generate a fairly large number of messages
about the progress of the calculation. One can reduce the volume by
setting \mcode+infolevel[genus2]+ to a number less than the default
value of $7$, before invoking \mcode+build_data["all"]+.
|
\section{Introduction}
Symmetry-protected topological (SPT) phases of matter
are gapped phases of quantum many-body systems with short-range entanglement.
They are topologically distinct from topologically trivial states in the presence of symmetries.
In other words, SPT phases are separated from topologically trivial phases by
quantum critical points.
Here, ``short-range entanglement'' means in particular the absence of topological order,
and hence the uniqueness of the ground state even when
the system is put on an arbitrary spatial manifold.
(This property is often called ``invertible'', and
hence SPT phases are said to have an invertible topological order.)
Bosonic SPT phases in (1+1) dimensions are
known to be classified by the second group cohomology $H^2(G,U(1))$.
\cite{ChenGuLiuWen2013}
For quantum many-body systems defined on one-dimensional lattices,
this can be most easily seen from the matrix product state (MPS) representations of
the quantum ground states of SPT phases.
\cite{Chen2011, Schuch2011, PollmannBergTurner2012,PollmannTurner2012}
On the other hand,
deep inside a gapped phase where
the correlation length is very short (order of a few lattice constant),
one could expect that the universal properties of the system
can be described in terms of a topological quantum field theory (TQFT or TFT).
The canonical examples include
Chern-Simons theories,
which describe various fractional quantum Hall liquids,
and
the BF theory,
which describes the topological limit of the $\mathbb{Z}_n$ lattice gauge theory.
\cite{wen2004quantum, fradkin2013field}
In this paper, we will undertake the task of bridging
the descriptions of (1+1)d bosonic SPT phases using MPSs,
and those using (1+1)d TFTs.
For (1+1)d bosonic SPT phases protected by symmetry $G$,
where $G$ is a symmetry group,
the relevant TFTs are $G$-equivariant TFTs discussed by
Turaev and Moore-Segal.
\cite{Turaev, turaev2010homotopy, Moore_lecture, Moore-Segal, Kapustin-Turzillo}
We in particular address the following two issues.
The first issue is about topological invariants (SPT invariants) of (1+1)d bosonic SPT phases.
These are quantities (numbers) which one can compute for a given
quantum ground state of a gapped (1+1)d system,
and take the same value anywhere in a given gapped phase.
I.e., they are stable and remain unchanged against adiabatic deformations
of Hamiltonians so far as one stays within a given gapped phase.
On the one hand, several topological invariants for bosonic
SPT phases have been constructed so far by using MPSs.
\cite{PollmannTurner2012}
In this paper, we will rederive these invariants by using
$G$-equivariant TFTs in (1+1)d.
The topological invariants are nothing but the partition functions of TFTs.
Second,
the hallmark of (1+1)d SPT phases
is the presence of boundary degrees of freedom that appear
when the SPT phases are terminated by boundaries.
The canonical example is the spin 1/2 that appears at the end of
the spin 1 Haldane spin chain.
In terms of MPSs,
these physical boundary degrees of freedom are captured by
degrees of freedom living in the auxiliary bond (entanglement) Hilbert space.
On the TFT side,
a natural framework to discuss the boundary degrees of freedom
is an ``open'' TFT.
\cite{lazaroiu2001structure, Moore-Segal}
Open TFTs are TFTs defined on the (1+1)d spacetime which has (1+0)d boundaries.
In this paper,
we will make an attempt to
make a dictionary between
MPSs with boundaries and open TFTs.
The rest of the paper is organized as follows:
In Sec.\ \ref{Classification and topological invariants of MPSs},
we introduce the descriptions of bosnic SPT phases in (1+1)d using MPSs.
In particular, we review the known construction of
various topological invariants for (1+1)d SPT phases built out of MPSs.
By using the fixed point MPSs,
we confirm that these topological invariants
characterizes the elements of the group cohomology $H^2(G,U(1))$,
and describe the procedure to extract these from a given quantum ground state.
(All results in Sec.\ \ref{Classification and topological invariants of MPSs} are
known in the literature,
so readers who are familiar with the MPS descriptions of (1+1)d bosonic SPT phases
and their topological invariants can skip this section.)
In Sec.\ \ref{$G$-equivariant Topological field theory},
we introduce $G$-equivariant TFTs following Moore and Segal.
\cite{Moore-Segal}
We discuss both closed and open TFTs;
In closed TFTs we consider the (1+1)d spacetime which has no
boundary,
whereas in open TFTs the (1+1)d spacetime has (1+0)d boundaries.
In Sec.\ \ref{Fukuma-Hosono-Kawai state sum construction},
we will derive
the topological invariants from the point of view of (1+1)d TFTs.
To this end, we evaluate the partition functions of
(1+1)d TFTs by using the so-called state sum construction~\cite{FHK}.
Introducing an orientation reversing operation on Frobenius
algebras enables us to define partition function on the real
projective plane $\mathbb{R} P^2$~\cite{Karimipour-Mostafazadeh}.
Finally, Appendices are devoted to an introduction to
the group cohomology,
and
projective representations,
the relation to orbifolded theories ($(1+1)$d Dijkgraaf-Witten theories),
and
the derivations of algebraic relations in $G$-equivariant open and closed TFTs.
\section{Classification and topological invariants of SPT phases using MPSs}
\label{Classification and topological invariants of MPSs}
In this section,
we briefly review the topological classification of bosonic SPT phases
in (1+1)d~\cite{PollmannBergTurner2012, ChenGuLiuWen2013, Schuch2011},
and their topological invariants~\cite{PollmannTurner2012}.
From the field theoretical point of view,
bosonic SPT phases are described by $G$-equivariant TFTs~\cite{Moore-Segal},
which will be introduced in the next section.
\subsection{Symmetry and group cohomology classification}
Let us consider a short range entangled pure state $\ket{\Psi}$ on a closed chain of length $L$,
which is represented by a MPS
\begin{align}
\ket{\Psi}
&= \sum_{\{m_i \}} {\rm Tr}\, (A_{m_1} \cdots A_{m_L} ) \ket{m_1\cdots m_L},
\quad
\ket{m_1\cdots m_L} = \ket{m_1}_1 \otimes \cdots \otimes \ket{m_L}_L,
\end{align}
where
$\ket{m}_j$ represents a state in the {\it physical} Hilbert space at the $j$-th site and $A_{m_j}$ is a $\chi \times \chi$ matrix which acts on the
{\it auxiliary} Hilbert space or
the {\it entanglement} Hilbert space living on the bonds;
The trace is taken over the $\chi\times \chi$ dimensional auxiliary Hilbert space.
Here and henceforth, we assume the translational symmetry for simplicity.
Let $G$ be a symmetry group.
The symmetry group $G$ possibly includes orientation reversing symmetries
(time-reversal or inversion symmetry, say).
For the purpose of specifying the orientation reversing symmetries,
let us introduce a homomorphism $\phi: G \to G_0$,
where $G_0$ is a group consisting of the orientation preserving symmetries.
The symmetry action $g \in G$ is defined on the basis $\ket{m}_j$
by a linear representation of $G$,
\begin{align}
&\hat g (\ket{m}_j) = \ket{n}_j [U_g]_{nm}, && U_g U_h = U_{gh},
&& (\mbox{$g$ is on-site unitary symmetry}), \\
&\hat T (\ket{m}_j) = \ket{n}_{j} [U_T]_{nm}, && U_T U_g^* = U_{Tg},
&& (\mbox{$T$ is time-reversal symmetry}), \\
&\hat P (\ket{m}_j) = \ket{n}_{L-j} [U_P]_{nm}, && U_P U_g = U_{Pg},
&& (\mbox{$P$ is inversion symmetry}),
\end{align}
for any $h \in G$.
By choosing different $A$, one can construct the ground states of gapped phases in (1+1) dimensions.
One can consider to classify these gapped phase topologically,
in the presence of a prescribed symmetry $G$.
The topological classification of (1+1)d SPT phases of bosons is given by the classification of the symmetry action on $A_{m}$.
Under the assumption that $\ket{\Psi}$ is a pure state, one can show~\cite{Cirac2008}
\begin{align}
&[U_g]_{mn} A_n = e^{i \theta_g} V_g^{\dag} A_m V_g,
&& (\mbox{$g$ is on-site unitary symmetry}), \\
&[U_T]_{mn} A^*_n = e^{i \theta_T} V_T^{\dag} A_m V_T,
&& (\mbox{$T$ is time-reversal symmetry}), \label{Eq:Sym_A_TRS} \\
&[U_P]_{mn} A_n^T = e^{i \theta_P} V_P^{\dag} A_m V_P,
&& (\mbox{$P$ is inversion symmetry}),
\end{align}
where $e^{i \theta_g}, e^{i \theta_T}$, and $e^{i \theta_P}$ are 1-dimensional linear representations of $G$,
and $V_g, V_T$, and $V_P$ act on the entanglement Hilbert space,
and obey, for any $h\in G$,
\begin{align}
\left\{
\begin{array}{ll}
V_g V_h = b(g,h) V_{gh} & (\mbox{$g$ is on-site unitary symmetry})
\\
\\
V_g V_h^* = b(g,h) V_{gh} & (\mbox{$g$ is time-reversal or inversion symmetry}) \\
\end{array} \right.
\label{Eq:ProjRep_V}
\end{align}
with a $U(1)$ phase $b(g,h) \in U(1)$ (2-cocycle).
These symmetry actions on $A_m$ are diagrammatically represented
in Fig.\ \ref{Fig:Symmetry_MPS}.
(In the figure, we neglect the 1-dimensional representation $e^{i \theta_g}$.)
From the associativity condition of $V_g$,
it follows that $b(g,h)$ is a representative of
$\phi$-twisted second group cohomology $H^2(G,U(1)_{\phi})$.
(Here, ``$\phi$-twisted'' means
the $g \notin G_0$ action on $U(1)$ group is defined by complex conjugate.
See Appendix \ref{Group cohomology}.)
The factor group $[b(g,h)] \in H^2(G,U(1)_{\phi})$ classifies how symmetry $G$ acts on the
short-range entangled pure state $\ket{\Psi}$ on the 1-dimensional closed chain.
\begin{figure}[!]
\begin{center}
\includegraphics[width=0.6\linewidth, trim=0cm 0cm 0cm 0cm]{figs/mps/Symmetry_MPS.pdf}
\end{center}
\caption{The symmetry property of $A_m$
under [a] on-site unitary, [b] time-reversal, and [c] inversion symmetries, respectively.}
\label{Fig:Symmetry_MPS}
\end{figure}
\subsection{Edge degrees of freedom}
In a non-trivial SPT phase
(specified by a nontrivial group cohomology
$[b(g,h)] \in H^2(G, U(1)_{\phi})$, $[b(g,h)] \neq 0$)
on an open chain $I = \{1, \dots, L\}$,
there emerge edge degrees of freedom where
the symmetry action of $G$ is ``fractionalized''.
The boundary degrees of freedom can be discussed
by using the MPS of the open chain:
\begin{align}
\Psi(\bra{v_L} \otimes \ket{v_R}) = \sum_{\{m_i\}}
v_L^{\dag} A_{m_1} \cdots A_{m_L} v_R \ket{m_1 \cdots m_L},
\qquad
\bra{v_L} \in V^*,
\quad
\ket{v_R} \in V,
\end{align}
where $\bra{v_L}$ and $\ket{v_R}$ specify boundary conditions
and belong to the edge Hilbert space $V^*$ and $V$, respectively.
$V$ is a $b(g,h)$-projective representation defined by (\ref{Eq:ProjRep_V})
and $V^*$ is the its conjugate representation.
Here, the symmetry fractionalization is realized in the following sense:
A symmetry action $g \in G_0$ on the MPS on the open chain is given by
\begin{align}
\hat g \Psi(\bra{v_L} \otimes \ket{v_R}) = \sum_{\{m_i\}} v_L^{\dag} V_g^{\dag} A_{m_1} \cdots A_{m_L} V_g v_R \ket{m_1 \cdots m_L}
= \Psi(\bra{v_L} V_g^{\dag} \otimes V_g \ket{v_R}),
\end{align}
where a symmetry operation $g \in G_0$ is projectively represented at the edges,
as opposed to the $g$ action on the bulk physical degrees of freedom, which is
a linear representation.
For an on-site unitary symmetry $g \in G_0$, we can introduce
a $g$-twisted MPS
\begin{align}
\ket{\Psi_g} := \sum_{\{m_i\}}
{\rm Tr}\, (A_{m_1} \cdots A_{m_L} V_g) \ket{m_1\cdots m_L}, \quad (g \in G_0).
\end{align}
From the perspective of Hamiltonians,
$\ket{\Psi_g}$ is a ground state of a Hamiltonian with a $g \in G_0$ symmetry defect.
For our purpose to make a comparison between MPSs and TFTs,
it is useful to introduce an open to closed map $\imath^{g}$
and a closed to open map $\imath_{g}$~\cite{Moore-Segal}
as
\begin{align}
&\imath^{g} \big( \Psi(\bra{v_L} \otimes \ket{v_R}) \big) := \braket{v_L| V_g^{\dag} |v_R} \ket{\Psi_g},
\quad (g \in G_0), \\
&\imath_{g}(\ket{\Psi_g}) := \sum_a \Psi(\bra{a} \otimes V_g \ket{a}),
\quad (g \in G_0).
\end{align}
Here $\{ \ket{a} \}_{a=1}^{{\rm dim} V}$ is a basis of $V$.
Finally, we also introduce a formal ``gluing'' operation of two open MPSs by
\begin{align}
\Psi(\bra{v_L} \otimes \ket{v_R}) \cdot \Psi(\bra{w_L} \otimes \ket{w_R})
:= \braket{w_L | v_R} \Psi(\bra{v_L} \otimes \ket{w_R}).
\end{align}
This will be also useful when we make a comparison between MPSs and TFTs.
\subsection{Simple and fixed point MPSs}
\label{Fixed point MPSs}
To study ground states deep inside a gapped phase,
or to study SPT phases in general,
it is useful and convenient to introduce a simple and fixed point MPS.
``Simple'' here means that the transfer matrix
$T_{ab,cd} = \sum_{m} [A_m]_{ab} [A^*_m]_{cd}$
has an only one eigenstate with unit magnitude of eigenvalue $|\nu|=1$,
i.e., unique ground state.~\cite{fidkowski2011topological}
``Fixed point'' means that we are in the limit of zero correlation length.
It is in this limit where we expect
SPT phases and the corresponding MPSs
are faithfully described by TQFTs.
In the following, we will construct fixed point MPSs
$\ket{\Psi}$
with a nontrivial group cohomology $H^2(G,U(1)_{\phi})$.
Let $G$ be a symmetry group with nontrivial group cohomology $H^2(G,U(1)_{\phi}) \neq 0$.
We fix a nontrivial 2-cocycle $b(g,h) \in Z^2(G,U(1)_{\phi}), [b(g,h)] \neq 0$.
We choose two $b(g,h)$-projective representations
$V$ which satisfy (\ref{Eq:ProjRep_V}).
We use the tensor product representation
$V^* \otimes V$ as a
physical Hilbert space, where
$V^*$ is the complex representation of
$V$.
Note that in the product representation,
the effect of the 2-cocycle cancels, $b^*(g,h) b(g,h)=1$.
For each site $j$, the basis is given by $\{ \ket{a}^L_j \otimes \ket{b}^R_j \}$,
where $\ket{a}^L_j$ ($\ket{b}^R_j$) is the basis of $V^*$ ($V$)
and transformed as
\begin{align}
&\left\{\begin{array}{l}
\hat g (\ket{a}^L_j \otimes \ket{b}^R_j) = \ket{c}^L_j \otimes \ket{d}^R_j {[V^*_g]}_{ca} {[V_g]}_{db}, \\
{[U_g]}_{ab,cd} = {[V^*_g]}_{ac} {[V_g]}_{bd}, \\
\end{array}\right. && \mbox{($g$ is on-site unitary symmetry)},\\
&\left\{\begin{array}{l}
\hat T (\ket{a}^L_j \otimes \ket{b}^R_j) = \ket{c}^L_j \otimes \ket{d}^R_j {[V^*_T]}_{ca} {[V_T]}_{db}, \ \ \hat T i \hat T^{-1} = -i , \\
{[U_T]}_{ab,cd} = {[V^*_T]}_{ac} {[V_T]}_{bd}, \\
\end{array}\right. && \mbox{($T$ is time-reversal symmetry)}, \\
&\left\{\begin{array}{l}
\hat P (\ket{a}^L_j \otimes \ket{b}^R_j) = \ket{c}^L_{L-j} \otimes \ket{d}^R_{L-j} {[V^*_P]}_{cb} {[V_P]}_{da}, \\
{[U_P]}_{ab,cd} = {[V^*_P]}_{ad} {[V_P]}_{bc}, \\
\end{array}\right. && \mbox{($P$ is inversion symmetry)}.
\end{align}
Note that $V$ and $V^*$ representations are exchanged under the inversion transformation.
\begin{figure}[!]
\begin{center}
\includegraphics[width=\linewidth, trim=0cm 0cm 0cm 0cm]{figs/mps/MPS_Fixed_Point.pdf}
\end{center}
\caption{[a] A fixed point MPS. The blue bond represents the singlet representation in $V^* \otimes V$. [b] $A$ matrix. [c] The transfer matrix.}
\label{Fig:MPS_Fixed_Point}
\end{figure}
To write down ground state wave functions,
we make use of a singlet included in the decomposition of the product representations,
$
V^* \otimes V = 1 \oplus \cdots.
$
The fixed point MPS can be constructed as the product state of singlet bonds
(Fig.\ \ref{Fig:MPS_Fixed_Point} [a]) as
\begin{align}
\ket{\Psi}
= \cdots \otimes (\sum_{a} \ket{a}^R_j \otimes \ket{a}^L_{j+1}) \otimes (\sum_{b} \ket{b}^R_{j+1} \otimes \ket{b}^L_{j+2}) \otimes \cdots.
\label{eq:fixed_point_MPS_def}
\end{align}
This can be written in the MPS form as
\begin{align}
\ket{\Psi} = \sum \cdots [A_{b_j c_j}]_{a_j a_{j+1}} [A_{b_{j+1} c_{j+1}}]_{a_{j+1} a_{j+2}} \cdots
\ket{ \cdots (b_j c_j) (b_{j+1} c_{j+1}) \cdots}
\end{align}
with
\begin{align}
[A_{ab}]_{cd} = \frac{1}{\sqrt{{\rm dim} V}} \delta_{ac} \delta_{bd}.
\end{align}
(See Fig.\ \ref{Fig:MPS_Fixed_Point} [b]. )
Here $(ab)$ is the physical index whereas $(cd)$ is the entanglement index.
We abbreviated $\ket{b_j}^L_j \otimes \ket{c_j}^R_j$ by $\ket{(b_jc_j)}$.
The prefactor $\frac{1}{\sqrt{{\rm dim} V}}$ is the normalization constant.
The MPS $\ket{\Psi}$ is the AKLT state \cite{affleck1988valence}
without any projection on the site degrees of freedom.~\cite{chen2012symmetry}
For the fixed point MPSs, the transfer matrix $T_{ab, a'b'}$ (Fig.\ \ref{Fig:MPS_Fixed_Point} [c]) is given by
\begin{align}
T_{ab,cd}
= \sum_{ef} [A_{ef}]_{ab} [A^*_{ef}]_{cd}
= \frac{1}{{\rm dim} V} \delta_{ac} \delta_{bd}.
\end{align}
In any symmetries, we have the following group action on $[A_{ab}]_{cd}$ for the fixed point MPS,
\begin{align}
&[U_g]_{ab,ef} [A_{ef}]_{cd}
= \frac{1}{\sqrt{{\rm dim} V}} [V_g^*]_{ac} [V_g]_{bd},
\quad (\mbox{$g$ is on-site unitary symmetry}), \\
&[U_T]_{ab,ef} [A^*_{ef}]_{cd}
= \frac{1}{\sqrt{{\rm dim} V}} [V_T^*]_{ac} [V_T]_{bd},
\quad (\mbox{$T$ is time-reversal symmetry}), \\
&[U_P]_{ab,ef} [A^T_{ef}]_{cd}
= \frac{1}{\sqrt{{\rm dim} V}} [V_P^*]_{ac} [V_P]_{bd},
\quad (\mbox{$P$ is inversion symmetry}).
\end{align}
Finally, it is worth pointing out that in fixed point MPSs
the length of the MPS chain is irrelevant because of the zero correlation length.
I.e., since they are at a renormalization group fixed point,
increasing/decreasing the number of cites does not change the essential properties of
the state.
For this reason, we always identify MPS chains with different lengths as
\begin{align}
{\rm Tr} ( A_{m_1} \cdots A_{m_L} ) \ket{m_1 \cdots m_L} \sim {\rm Tr} ( A_{m_1} \cdots A_{m_L} A_{m_{L+1}}) \ket{m_1 \cdots m_L m_{L+1}}.
\label{eq:mps_equivalence}
\end{align}
\subsection{Topological invariants}
\label{Topological invariants}
In this section, we will construct and discuss topological invariants
of bosonic SPT phases in (1+1)d using the MPS.
We start by listing topological invariants.
Detailed descriptions of topological invariants will follow shortly.
There are three types of topological invariants,
which are defined in terms of the data of 2-cocycle $\{b(g,h)\}$:
\begin{itemize}
\item The discrete torsion phase (partition function on $T^2$ with twist)
\begin{align}
\label{eq:T^2inv}
\epsilon(g,h) = \frac{b(g,h)}{b(h,g)},
\quad
V_g V_h = \epsilon(g,h) V_h V_g,
\quad
g,h\in G_0,
\quad
gh=hg.
\end{align}
\item The crosscap invariant (partition function on $\mathbb{R} P^2$)
\begin{align}
\label{eq:Corsscapinv}
\theta(g) := b(g,g),
\quad V_g V_g^* = \theta(g),
\quad g \notin G_0,
\quad g^2 = 1.
\end{align}
\item The Klein bottle invariant (partition function on the Klein bottle with twist)
\begin{align}
\label{eq:KBinv}
\kappa(g;h) = \frac{b(g,h^{-1}) b(h,h^{-1})}{b(h,g)},
\quad
V_g V_h^T = \kappa(g;h) V_h V_g,
\quad
g \notin G_0, h \in G_0,
\quad g h^{-1} = h g.
\end{align}
\end{itemize}
Several comments are in order.
--
First, in \eqref{eq:T^2inv}, \eqref{eq:Corsscapinv} and \eqref{eq:KBinv},
and throughout this subsection,
we omit the 1-dimensional representation
$\{ e^{i \theta_g}\}$ for simplicity.
--
One can check easily that
these quantities are left unchanged under the 1-coboundary $b(g,h) \mapsto b(g,h) a_g a_g^{\phi(g)} a_{gh}^{-1}$, $a_g \in U(1)$.
--
One can give interpretations to
these topological invariants
in terms of spacetime path integrals.
We will mention these interpretations
later in this subsection,
and also in Sec.\
\ref{$G$-equivariant Topological field theory}
-
\ref{Fukuma-Hosono-Kawai state sum construction}
from the TFT point of view.
In short,
these three topological invariants are interpreted as
the partition function on the torus,
the projective plane,
and
the Klein bottle respectively.
For this reason, we will often refer the topological invariants
as the partition functions
(on the torus, the projective plane, and the Klein bottle).
--
Finally, the above three SPT invariants are not independent.
One can show
\begin{align}
&\epsilon(g,1) = \epsilon(1,g) = 1 && {\rm for\ \ } g \in G_0, \\
&\epsilon(h,g) = \epsilon(g,h)^{-1} && {\rm for\ \ } g,h \in G_0, gh=hg, \\
&\epsilon(g,hk) = \epsilon(g,h)\epsilon(g,k) && {\rm for\ \ } g,h,k \in G_0, gh=hg, gk=kg, \\
&\kappa(g;1) = 1 && {\rm for\ \ } g \notin G_0, \\
&\kappa(g;hk) = \kappa(g;h) \kappa(g;k) && {\rm for\ \ } g \notin G_0, h,k \in G_0, g h^{-1} = h g, g k^{-1}=k g, \\
&\kappa(kg;h) = \kappa(gk^{-1};h) = \epsilon(k,h) \kappa(g;h) && {\rm for\ \ } g \notin G_0, h,k \in G_0, g h^{-1} = h g, h k=k h, \\
&\theta(h g) = \theta(g h^{-1}) = \kappa(g;h) \theta(g) && {\rm for\ \ } g \notin G_0, h \in G_0, g^2=1, g h^{-1} = h g.
\end{align}
In many cases, the Klein bottle SPT invariant $\kappa(g;h)$
can be written in terms of $\epsilon(g,h)$ and $\theta(g)$.
This is however not always the case.
A simple example in which the Klein bottle invariant does not reduce to the other invariants
is an SPT phase protected by
$G = \mathbb{Z}_4 = \{1,\sigma,\sigma^2, \sigma^3\}$.
Here, the generator $\sigma$ is inversion/time-reversal.
For a nontrivial projective representation generated by
$V_{\sigma} = e^{-i s_y \frac{\pi}{4}}$,
where
$s_y = \begin{pmatrix}
0 & -i \\
i & 0 \\
\end{pmatrix}$
is the $y$-component of the Pauli matrix,
from $V_{\sigma} V_{\sigma^2}^T = - V_{\sigma^2} V_{\sigma}$,
the Klein bottle SPT invariant reads $\kappa(\sigma;\sigma^2) = -1$.
\subsubsection{Topological invariants in terms of ground state
wave functions}
As mentioned,
the topological invariants can be interpreted
by using the path integral formalism.
In the rest of this subsection,
we will instead use the operator formalism,
and in particular
aim to extract the topological invariant
solely by using ground state wave functions.
(Apart from our goal of bridging MPSs and TFTs,
expressing topological invariants solely in terms
of ground state wave functions may have practical (numerical)
merits. \textcolor{blue}{)}
In order to discuss and define these topological invariants,
one important ingredient
is {\it gauging} symmetry.
Here, by gauging, we mean coupling the system to the background flat $G$-bundle.
We will describe how this can be done within MPSs
for on-site unitary symmetry
in Sec.\ \ref{Discrete torsion phase: symmetry action on twisted ground state}.
The same gauging procedure can be introduced by using
the path-integral.
(In addition, one could promote the background gauge field into a dynamical one.
This procedure is often called {\it orbifolding}
to distinguish it from gauging.
In this paper, for the purpose of describing SPT phses,
we will consider gauging but not orbifolding.
Orbifolding leads to the so-called Dijkgraaf-Wittten
theories, which we briefly discuss in Appendix \ref{Orbifolding: Dijkgraaf-Witten theory in (1+1)d}.)
As for the crosscap invariant,
we need to introduce a ``trick''
within the operator formalism
in order to mimic the effect of
putting the theory on $\mathbb{R}P^2$
in the path integral formalism.
This can be done in two different ways,
depending on whether the symmetry
group includes spatial inversion or time-reversal.
Following Pollmann and Turner
\cite{PollmannTurner2012}
we will introduce two operations,
``partial inversion'' and ``adjacent partial transposition'',
for spatial inversion and time-reversal, respectively.
When interpreted in the path integral formalism,
these operations effectively create
$\mathbb{R} P^2$ as the spacetime manifold.
Introducing such a {\it partial space-time twist operator}
is a useful way to detect SPT topological invariant
which cannot be represented by the partition function
on a mapping torus.~\footnote{
A mapping torus is space-time manifold which takes
the form
$M \times_f S^1 := M \times [0,1] / \big\{ (x,0) \sim (f(x), 1) \big\}$,
where $f : M \to M$ is a diffeomorphism.
In (invertible) TFTs, the partition function $Z(M \times_f S^1)$
on $M \times_f S^1$ is given by the expectation value of
operator $\hat f$ representing diffeomorphism $f$ on
the ground state wave function $\ket{\Psi_M}$ on $M$
as $Z(M \times_f S^1) = \braket{\Psi_M | \hat f | \Psi_M}$.
}
This prescription can also be applied to
fermionic SPT invariants
\cite{Shapourian-Shiozaki-Ryu}
and SPT phases in more general space dimensions.
\cite{Shiozaki-Shapourian-Ryu}
For all topological invariants, the fact that they can be extracted from
ground state wave functions
can be easily proven if we use the fixed point MPS.
\paragraph{Discrete torsion phase (1): symmetry action on twisted ground state}
\label{Discrete torsion phase: symmetry action on twisted ground state}
Let us first express the torus topological invariant
\eqref{eq:T^2inv} by using ground state wave functions.
\begin{figure}[t]
\begin{center}
\includegraphics[width=\linewidth, trim=0cm 0cm 0cm 0cm]{torus_discrete_torsion.pdf}
\end{center}
\caption{
[a] MPS expression of the symmetry action on the twisted ground state.
[b] The equivalent path integral on the 2-torus $T^2$.
The blue and red lines express the symmetry defect lines.}
\label{Fig:torus_discrete_torsion}
\end{figure}
From the MPS and the projective representation $\{V_g\}_{g \in G}$,
we can construct the MPS $\ket{\Psi_h}$ with
boundary condition twisted by an on-site unitary symmetry $h \in G_0$ as
\begin{align}
\ket{\Psi_h}
&= \sum_{\{m_i\}} {\rm Tr}\, (A_{m_1} \cdots A_{m_L} V_h) \ket{m_1\cdots m_L}.
\end{align}
Then, for
a global unitary symmetry,
$g \in G_0$,
\begin{align}
\hat g \ket{\Psi_h}
= \sum_{\{m_i\}}
{\rm Tr}\, (A_{m_1} \cdots A_{m_L} V_g V_{h} V_g^{-1}) \ket{m_1\cdots m_L}
= \frac{b(g,h)}{b(g h g^{-1}, g)} \ket{\Psi_{ghg^{-1}}}.
\label{Eq:G-action_MPS}
\end{align}
The MPS diagram is shown in Fig.~\ref{Fig:torus_discrete_torsion}[a].
If $g h g^{-1} = h$, the $U(1)$ factor is well-defined which is invariant under the 1-coboundary.
From the TFT point of view,
this invariant is nothing but the partition function $Z_{T^2}(g,h)$ on the torus $T^2$ with the background $g$ and $h$ twist
\begin{align}
\epsilon(g,h) = Z_{T^2}(g,h) = \braket{\Psi_h | \hat g | \Psi_h},
\quad (gh=hg).
\end{align}
In the spacetime path integral,
it would be useful to introduce symmetry defect lines to express the background $G$ field.
The matter field is transformed by $U_g$ when it passes through the symmetry defect line of $g$.
Fig.~\ref{Fig:torus_discrete_torsion}[b] shows the symmetry defect lines corresponding to the partition function $Z_{T^2}(g,h)$ twisted by $g$ and $h$.
The discrete torsion phase $\epsilon(g,h)$ arises from the intersection of two symmetry defect lines of $g$ and $h$ with $[g,h]=0$.
\paragraph{Discrete torsion phase (2): partial symmetry action and swapping}
\label{Discrete torsion phase: partial symmetry action and swapping}
There is an alternative way to detect the discrete torsion phase invariant.
It is given by the combination of the partial symmetry action and the swapping operator, described as follows.
Let $\ket{\Psi}$ be the ground state on $S^1$ with no flux.
(In the TFT path integral, this state is obtained/defined by the path-integral over the disc.)
We introduce three adjacent intervals $I_1 \cup I_2 \cup I_3$ with $I_1$ and $I_3$ having the same number of sites.
The discrete torsion phase $\epsilon(g,h)$ is then extracted as the complex $U(1)$ phase of the quantity~\cite{PhysRevLett.109.050402}
\begin{align}
Z = \Braket{\Psi | \prod_{j \in I_1 \cup I_2} (U_h)_j \cdot {\rm Swap}(I_1,I_3) \cdot \prod_{j \in I_1 \cup I_2} (U_g)_j | \Psi}, \qquad
gh=hg,
\label{eq:symmetry_action_and_swap}
\end{align}
in the limit $|I_1|,|I_2|, |I_3| \gg \xi$, where $\xi$ is the correlation length of the bulk.
Here, ${\rm Swap}(I_1,I_3)$ is the operator swapping the two intervals $I_1$ and $I_3$, which is defined by
\begin{align}
{\rm Swap}(I_1,I_3) \ket{m_j}
= \left\{\begin{array}{ll}
\ket{m_{j+|I_1|+|I_2|}} & (j \in I_1), \\
\ket{m_{j-|I_1|-|I_2|}} & (j \in I_3), \\
\ket{m_j} & ({\rm otherwise}).
\end{array}\right.
\end{align}
For the MPS representation of the ground state $\ket{\Psi}$, the MPS diagram of $Z$ is written as Fig.~\ref{Fig:symmetry_action_and_swap}.
For the fixed point MPS (\ref{eq:fixed_point_MPS_def}), it is easy to show that $Z = |Z| \epsilon(g,h)$.
The path-integral picture also verifies that $Z$ gives the discrete torsion phase.
See Fig.~\ref{Fig:symmetry_action_and_swap_path_integral}.
Topologically, the swapping operator ${\rm Swap}(I_1,I_3)$ with the intermediate region $I_2$ is equivalent to adding a genus.
The background $G$ field obtained by the partial symmetry actions $U_g$ and
$U_h$ on the adjacent intervals $I_1 \cup I_2$
has an intersection between two symmetry defect lines of $g$ and $h$, which leads to the discrete torsion phase $\epsilon(g,h)$.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.7\linewidth, trim=0cm 0cm 0cm 0cm]{symmetry_action_and_swap.pdf}
\end{center}
\caption{
MPS expression of the partial symmetry action with swapping defined in (\ref{eq:symmetry_action_and_swap}).
}
\label{Fig:symmetry_action_and_swap}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width=\linewidth, trim=0cm 0cm 0cm 0cm]{symmetry_action_and_swap_path_integral.pdf}
\end{center}
\caption{
The geometry of path integral for the partial symmetry action with swapping operator defined in (\ref{eq:symmetry_action_and_swap}).
The red and blue lines express the symmetry defect lines.
The intervals with arrows are identified with ones having the same number of arrows, which results in the 2-torus.
}
\label{Fig:symmetry_action_and_swap_path_integral}
\end{figure}
\paragraph{Crosscap from inversion symmetry: ``partial inversion''}
\begin{figure}[t]
\begin{center}
\includegraphics[width=\linewidth, trim=0cm 0cm 0cm 0cm]{figs/mps/Partial_Inversion.pdf}
\end{center}
\caption{[a] MPS representation of crosscap. [b] The geometry of path integral for the partial inversion.}
\label{Fig:Partial_Inversion}
\end{figure}
The crosscap topological invariant
can be defined when the symmetry group $G$
includes spatial inversion or time-reversal.
The procedures to extract the invariant
from ground state wave functions
are different for spatial inversion and time-reversal.
Let us first discuss the crosscap topological
invariant when $G$ includes
spatial inversion.
Let $\ket{\Psi}$ be a MPS on the closed chain $L$.
To create the real projective plane $\mathbb{R} P^2$, we take the {\it partial inversion} on the interval $I = \{1, \dots, N\}$,
as shown in Fig.\ \ref{Fig:Partial_Inversion} [a],
as
\begin{align}
\hat P_I \ket{\Psi}
= \sum_{\{m_i\}} {\rm Tr}\, (A_{m_1} \cdots A_{m_L}) \ket{\widetilde{m}_1 \cdots \widetilde{m}_N m_{N+1} \cdots m_L},
\end{align}
where $\widetilde{\ket{m}}_j$ is the partially inverted physical degrees of freedom,
\begin{align}
\widetilde{\ket{m}}_j = \ket{n}_{N-j} [U_P]_{nm}.
\end{align}
We assume the length of the interval $I$ is sufficiently larger than the correlation length.
The fact that this operation
creates $\mathbb{R}P^2$ as the spacetime manifold
can be easily understood from Fig.\ \ref{Fig:Partial_Inversion}-[b];
In the path integral representation, the partial inversion is equivalent to inserting
a one crosscap on the time slice at $\tau=0$.
One can show that for $P^2 = 1$ the amplitude $\hat P_I \ket{\Psi}$ gives the crosscap invariant~\cite{PollmannTurner2012}
\begin{align}
\theta(P) = Z_{\mathbb{R} P^2}(P) = \frac{\braket{\Psi| \hat P_I |\Psi}}{|\braket{\Psi| \hat P_I |\Psi}|} = b(P,P),
\quad P^2 = 1,
\end{align}
in the limit $L \to \infty$ and $N \to \infty$.
This formula can easily be proven
for fixed point MPSs introduced in the previous section.
Here we give a proof by using the cut and glue construction.~\cite{Qi2012}
To illustrate the proof, we use the Haldane chain protected by the inversion symmetry with $H^2(\mathbb{Z}_2;U(1)_{\phi}) = \mathbb{Z}_2$.
First, we cut the chain $L$ by the interval $I$.
There are four effective low energy degrees of freedom localized at the boundary of $I$,
\begin{align}
\ket{e^R_0} \otimes \ket{e^L_1} \otimes \ket{e^R_N} \otimes \ket{e^L_{N+1}},
\quad
e^L_j, e^R_j = \uparrow, \downarrow.
\end{align}
Next, we glue these degrees of freedom to get the original ground state by forming the singlet bond as
sites $0$-$1$ and $N$-$(N+1)$ as
\begin{align}
\ket{\Psi} = \frac{1}{\sqrt{2}} (\ket{\uparrow^R_0} \otimes \ket{\downarrow^L_1} - \ket{\downarrow^R_0} \otimes \ket{\uparrow^L_1} )
\otimes \frac{1}{\sqrt{2}} (\ket{\uparrow^R_N} \otimes \ket{\downarrow^L_{N+1}} - \ket{\downarrow^R_N} \otimes \ket{\uparrow^L_{N+1}} ).
\end{align}
The reduced density matrix $\rho_I$ is
\begin{align}
\rho_I
&= {\rm tr}_{0,{N+1}} (\ket{\Psi} \bra{\Psi})
= \frac{1}{4} \left[
(\ket{\uparrow^L_1}\bra{\uparrow^L_1} +\ket{\downarrow^L_1}\bra{\downarrow^L_1}) \otimes
(\ket{\uparrow^R_N}\bra{\uparrow^R_N} +\ket{\downarrow^R_N}\bra{\downarrow^R_N})
\right]
= \frac{1}{4} {\rm Id}^L_{1} \otimes {\rm Id}^R_N.
\end{align}
The partial inversion $P_I$ acts as
\begin{align}
\hat P_I \ket{\uparrow^L_1} = \ket{\downarrow^R_N},
\quad
\hat P_I \ket{\downarrow^L_1} = -\ket{\uparrow^R_N},
\quad
\hat P_I \ket{\uparrow^R_N} = \ket{\downarrow^L_1},
\quad
\hat P_I \ket{\downarrow^R_N} = -\ket{\uparrow^L_1},
\end{align}
from which we read off
\begin{align}
\braket{\Psi|\hat P_I|\Psi} = {\rm tr}_I( \hat P_I \rho_I) = - \frac{1}{2}.
\end{align}
We thus obtained the topological invariant $b(P,P) = -1$,
which, as expected, is non-trivial (differs from $b(P,P)=1$),
and is the $\mathbb{Z}_2$ invariant.
\paragraph{Crosscap from time-reversal symmetry: ``adjacent partial transposition''}
\begin{figure}[!]
\begin{center}
\includegraphics[width=\linewidth, trim=0cm 0cm 0cm 0cm]{figs/mps/Partial_Transpose.pdf}
\end{center}
\caption{[a] MPS representation of the crosscap. [b] The path integral representation of partial transposition. }
\label{Fig:Partial_Transpose}
\end{figure}
\begin{figure}[!]
\begin{center}
\includegraphics[width=\linewidth, trim=0cm 0cm 0cm 0cm]{TRS_to_RP2_v2.pdf}
\end{center}
\caption{Equivalence between
adjacent partial transpose and inserting a cross cap.
The black line is identified. }
\label{Fig:TRS_to_RP2}
\end{figure}
Next, we describe the extraction of the crosscap invariant
when $G$ includes time-reversal.
To this end, we will consider the so-called partial transposition.
The partial transposition has been used, for example, to define
the entanglement negativity.~\cite{VidalWerner}
Pollmann and Turner~\cite{PollmannTurner2012}
showed the MPS network
(Fig.\ \ref{Fig:Partial_Transpose} [a]) of
the partial transposition on adjacent intervals $I= I_1 \cup I_2$
is nothing but the $\mathbb{Z}_2$ topological invariant $b(T,T)$
associated with the time-reversal symmetry $T \notin G_0$.
(Here, to be concrete, let
$I_1 = \{1, \dots, N\}$ and $I_2 = \{N+1, \dots, M\}$
be two adjacent intervals in the chain of length $L$.)
In the following, we will review the topological invariant of
Pollmann and Turner.
We will also note, with an eye toward the TFT descriptions of
SPT phases and their topological invariants, that
the Pollmann-Turner can be interpreted as a space-time path integral
on the real projective plane.
Below, we will express the Pollmann-Turner invariant in terms of a density matrix.
We introduce the reduced density matrix for the interval $I$ by taking the
partial trace of the degrees of freedom on living on the compliment of $I$, $L\backslash I$,
as
\begin{align}
\rho_I = {\rm tr}_{L \backslash I} (\ket{\Psi} \bra{\Psi}).
\end{align}
The reduced density matrix
$\rho_I$ is expanded in the basis of $I_1$ and $I_2$ as
\begin{align}
\rho_I = \sum_{ijkl} \ket{e^1_i, e^2_j} \braket{e^1_i, e^2_j | \rho_I | e^1_k, e^2_l} \bra{e^1_k, e^2_l},
\end{align}
where $\ket{e^1_i}$ and $\ket{e^2_j}$ are the basis on the intervals $I_1$ and $I_2$, respectively.
We introduce the partial transposition $\rho_I^{T_1}$ for the interval $I_1$ which is defined by~\cite{VidalWerner}
\begin{align}
\rho_I^{T_1} := \sum_{ijkl} \ket{e^1_i, e^2_j} \braket{e^1_k, e^2_j | \rho_I | e^1_i, e^2_l} \bra{e^1_k, e^2_l}.
\end{align}
In addition to the partial transposition,
we also consider the ``unitary part'' of time-reversal,
and consider
partial time-reversal transformation $\hat T'_{I_1} = \prod_{j \in I_1} (U_T)_j$
action only on $I_1$.
Note that $\hat T'_{I_1}$ is unitary, i.e.,\ it consists of the only unitary part of the time-reversal transformation $\hat T = (\prod_j (U_T)_j) K$,
where $K$ is complex conjugation.
Putting everything together,
we consider
\begin{align}
{\rm tr}_I \big( \rho_I \hat T'_{I_1} \rho_I^{T_1} [\hat T'_{I_1}]^{\dag} \big), \quad \hat T'_{I_1} = \prod_{j \in I_1} (U_T)_j.
\label{path int PR2}
\end{align}
Finally, the Pollmann-Turner topological invariant, i.e.,
the crosscap topological invariant
is given by the phase
of \eqref{path int PR2},
\begin{align}
\theta(T) = \frac{{\rm tr}_I \big( \rho_I \hat T'_{I_1} \rho_I^{T_1} [\hat T'_{I_1}]^{\dag} \big)}{|{\rm tr}_I \big( \rho_I \hat T'_{I_1} \rho_I^{T_1} [\hat T'_{I_1}]^{\dag} \big)|}
\to b(T,T) , \quad (N,M \to \infty).
\label{P-T invariant}
\end{align}
In the limit $N,M \gg \xi$, where $\xi$ the correlation length,
the phase $\theta(T)$
is a quantized topological invariant.
The path integral representation of the quantity
\eqref{path int PR2}
is shown in Fig.\ \ref{Fig:Partial_Transpose} [b],
which is topologically equivalent to a sphere with one crosscap as shown in Fig.\ \ref{Fig:TRS_to_RP2}.
That \eqref{P-T invariant} is indeed a quantized and topological invariant
can be proven within the MPS framework.
Here, we demonstrate this by again using the cut and glue construction~\cite{Qi2012},
and by taking the Haldane chain with time-reversal symmetry as an example.
Within the cut and glue construction,
there are six active degrees of freedom at low energies
in the reduced density matrix,
\begin{align}
\ket{e^R_0} \otimes \ket{e^L_1} \otimes \ket{e^R_N} \otimes \ket{e^L_{N+1}}\otimes \ket{e^R_M} \otimes \ket{e^L_{M+1}},
\quad
e^L_j, e^R_j = \uparrow, \downarrow.
\end{align}
The ground state is a singlet formed from $(\ket{e^R_0} ,\ket{e^L_1})$, $(\ket{e^R_N} ,\ket{e^L_{N+1}})$ and $(\ket{e^R_M} ,\ket{e^L_{M+1}})$ as
\begin{equation}\begin{split}
\ket{\Psi}
&= \frac{1}{\sqrt{2}} (\ket{\uparrow^R_0} \otimes \ket{\downarrow^L_1} - \ket{\downarrow^R_0} \otimes \ket{\uparrow^L_1} )
\otimes \frac{1}{\sqrt{2}} (\ket{\uparrow^R_N} \otimes \ket{\downarrow^L_{N+1}} - \ket{\downarrow^R_N} \otimes \ket{\uparrow^L_{N+1}} ) \\
&\qquad \otimes \frac{1}{\sqrt{2}} (\ket{\uparrow^R_M} \otimes \ket{\downarrow^L_{M+1}} - \ket{\downarrow^R_M} \otimes \ket{\uparrow^L_{M+1}} ).
\end{split}\end{equation}
The reduced density matrix $\rho_I$ reads
\begin{align}
\rho_I
&= {\rm tr}_{0,{M+1}} (\ket{\Psi} \bra{\Psi})
\nonumber \\
&= \frac{1}{2} {\rm Id}^L_1
\otimes \frac{1}{2} \Big[
\ket{\uparrow^R_N} \bra{\uparrow^R_N} \otimes \ket{\downarrow^L_{N+1}} \bra{\downarrow^L_{N+1}}
+\ket{\downarrow^R_N} \bra{\downarrow^R_N} \otimes \ket{\uparrow^L_{N+1}} \bra{\uparrow^L_{N+1}}
\nonumber \\
&\ \ \ \ \ \ \ \ \ \ \ -\ket{\uparrow^R_N} \bra{\downarrow^R_N} \otimes \ket{\downarrow^L_{N+1}} \bra{\uparrow^L_{N+1}}
-\ket{\downarrow^R_N} \bra{\uparrow^R_N} \otimes \ket{\uparrow^L_{N+1}} \bra{\downarrow^L_{N+1}}
\Big]
\otimes \frac{1}{2} {\rm Id}^R_M.
\end{align}
By taking the partial transposition on $I_1 = \{1,N\}$ and
noting that the unitary part of the time-reversal transformation is given by
\begin{align}
U_T \ket{\uparrow} = \ket{\downarrow},
\quad
U_T \ket{\downarrow} = -\ket{\uparrow},
\end{align}
we have
\begin{align}
\hat T'_{I_1} \rho^{T_1}_I [\hat T'_{I_1}]^{\dag}
&= \frac{1}{2} {\rm Id}^L_1
\otimes \frac{1}{2} \Big[
\ket{\downarrow^R_N} \bra{\downarrow^R_N} \otimes \ket{\downarrow^L_{N+1}} \bra{\downarrow^L_{N+1}}
+\ket{\uparrow^R_N} \bra{\uparrow^R_N} \otimes \ket{\uparrow^L_{N+1}} \bra{\uparrow^L_{N+1}}
\nonumber \\
&\ \ \ +\ket{\uparrow^R_N} \bra{\downarrow^R_N} \otimes \ket{\downarrow^L_{N+1}} \bra{\uparrow^L_{N+1}}
+\ket{\downarrow^R_N} \bra{\uparrow^R_N} \otimes \ket{\uparrow^L_{N+1}} \bra{\downarrow^L_{N+1}}
\Big]
\otimes \frac{1}{2} {\rm Id}^R_M ,
\end{align}
which leads to
\begin{align}
{\rm tr}_I(\rho_I \hat T'_{I_1} \rho^{T_1}_I [\hat T'_{I_1}]^{\dag})
= - \frac{1}{8}.
\end{align}
The minus sign $(-1)$ is the proper $\mathbb{Z}_2$ invariant for the Haldane chain with time-reversal symmetry.
\paragraph{Klein bottle partition function from inversion symmetry}
Similar to the crosscap topological invariant,
the Klein bottle topological invariant can be
defined both for spatial inversion and
time-reversal.
Let us start with the case of spatial inversion.
We act with an inversion transformation $P \notin G_0$ on the twisted MPS $\ket{\Psi_g}$,
\begin{align}
\hat P \ket{\Psi_g}
= \sum_{ \{m_i\} } {\rm Tr}\, (A_{m_1} \cdots A_{m_L} V_P V^T_{g} V_P^{-1}) \ket{m_1\cdots m_L}
= \frac{b(P,g^{-1}) b(g,g^{-1})}{b(P g^{-1} P^{-1},P)} \ket{\Psi_{P g^{-1} P^{-1}}}.
\end{align}
If $P g^{-1} P^{-1} = g$, the $U(1)$ prefactor is well-defined which is invariant under the 1-coboundary.
This invariant is nothing but the partition function $Z_{KB}(P;g)$ over the Klein bottle ($KB$) with the background $P$ and $g$ twists
\begin{align}
\kappa(P;g) = Z_{KB}(P;g) = \braket{\Psi_g | \hat P | \Psi_g},
\quad (P \notin G_0, \ g \in G_0, \ P g^{-1}= g P).
\end{align}
An example is
a $\mathbb{Z}_2 (= \{1,\sigma\})$ paramagnet with inversion symmetry $\mathbb{Z}_2^P (= \{1,P\})$
where $\mathbb{Z}_2$ charge is preserved under the inversion.
The topological classification is given by $H^2(\mathbb{Z}_2 \times \mathbb{Z}_2^P,U(1)_{\phi}) = \mathbb{Z}_2 \times \mathbb{Z}_2$.
The two topological invariants can be seen in $V_P V_P^* = \theta(P)$ and $V_P V^T_{\sigma} = \kappa(P;\sigma) V_{\sigma} V_P$.
\paragraph{Klein bottle partition function from time-reversal symmetry: ``disjoint partial transposition with intermediate twist''}
\begin{figure}[!]
\begin{center}
\includegraphics[width=0.9\linewidth, trim=0cm 0cm 0cm 0cm]{figs/mps/Partial_Transpose_Klein.pdf}
\end{center}
\caption{[a] MPS representation of the Klein bottle partition function with twist. [b] The path integral representation. }
\label{Fig:Partial_Transpose_Klein}
\end{figure}
Next, as for
time-reversal symmetry,
the Klein bottle partition function
obtained from from time-reversal
can be represented in terms of MPSs in a way similar to the crosscap partition function.
First, we divide the closed chain $L$ into three adjacent intervals $I_1 \cup I_2 \cup I_3$,
$I_1 = \{1, \dots, N_1\}$, $I_2 = \{N_1+1, \dots, N_2\}$, $I_3 = \{N_2+1, \dots, N_3\}$.
In addition, we introduce one replica.
We trace out the region except for $I_1 \cup I_3$ {\it with symmetry twist in the interval $I_2$} as
\begin{align}
\rho_{I_1 \cup I_3}(g):= {\rm tr}_{L \backslash (I_1 \cup I_3)} \Big( \hat g_{I_2} \ket{\Psi} \bra{\Psi} \Big),
\quad
\hat g_{I_2} = \prod_{j \in I_2} (U_g)_{j}.
\end{align}
Then, we consider the following quantity
\begin{align}
{\rm tr}_{I_1 \cup I_3} \Big( \rho_{I_1 \cup I_3}(g) \hat T'_{I_1} \rho_{I_1 \cup I_3}(g^{-1}) [\hat T'_{I_1}]^{\dag} \Big),
\quad (T g^{-1} T^{-1} = g),
\quad
T'_{I_1} = \prod_{j \in I_1} (U_T)_j.
\end{align}
The corresponding MPS network and path integral are shown in Fig.\ \ref{Fig:Partial_Transpose_Klein} [a] and [b], respectively.
One can show this quantity approaches the Klein bottle partition function
\begin{align}
\frac{{\rm tr}_{I_1 \cup I_3} \Big( \rho_{I_1 \cup I_3}(g) \hat T'_{I_1} \rho_{I_1 \cup I_3}(g^{-1}) [\hat T'_{I_1}]^{\dag} \Big)}
{\Big| {\rm tr}_{I_1 \cup I_3} \Big( \rho_{I_1 \cup I_3}(g) \hat T'_{I_1} \rho_{I_1 \cup I_3}(g^{-1}) [\hat T'_{I_1}]^{\dag} \Big) \Big|}
\to Z_{KB}(T;g) = \kappa(T;g)
\end{align}
in the limit $N_1, N_2-N_1, N_3-N_2, L-N_3 \gg \xi$, where $\xi$ is the correlation length.
It is easy to show the above formula
for fixed point MPSs
by using the symmetry properties of $A$ matrix.
\section{$G$-equivariant topological field theories and MPSs}
\label{$G$-equivariant Topological field theory}
Having discussed the MPS description
of (1+1)d bosonic SPT phases,
we now move on (1+1)d $G$-equivaliant TFTs.
In the following sections,
Sec.\ \ref{Some basics of TFTs} to Sec.\ \ref{G-equivariant open and closed TFTs},
we briefly summarize necessary ingredients of
open and closed $G$-equivariant oriented $(1+1)$d TFTs
following Moore-Segal\cite{Moore-Segal}.
There are some overlaps with Ref.\ \cite{Kapustin-Turzillo},
where they also discuss closed $G$-equivariant $(1+1)$d unoriented TFTs.
Ref.\ \cite{Moore-Segal} also discusses
$(1+1)$d open and closed TFTs with spin structure,
which can describe fermionic SPT phases such as
class D topological superconductors.
Here we restrict ourselves to $(1+1)$d bosonic SPT phase protected by on-site unitary $G$-symmetry where $G$ is a finite group.
In short,
a $G$-equivariant TFT is a TFT couped with the background $G$-gauge field.
(Integrating out the background $G$-gauge field, i.e. orbifolding the $G$-symmetry,
leads to an orbifolded theory which is a TFT without $G$-symmetry.)
In the following, first, we introduce some general properties of TFTs.
Next, we summarize
$(1+1)$d $G$-equivariant closed TFTs
with an eye toward $(1+1)$d SPT phases.
Our notations closely follow Moore-Segal\cite{Moore-Segal}.
Next, we will summarize $(1+1)$d $G$-equivariant open and closed TFTs.
\subsection{Some basics of TFTs}
\label{Some basics of TFTs}
\begin{figure}[!]
\begin{center}
\includegraphics[width=0.7\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/ex.pdf}
\end{center}
\caption{An example of cobordism.}
\label{fig:tft_ex}
\end{figure}
In the axiomatic definition,
a TFT in $(d+1)$ dimensions
is a functor $Z$ from a cobordism category
$\mathbf{Bord}_{<d,d+1>}$
to the category of finite dimensional complex vector spaces $\mathbf{Vect}$
equipped with tensor product.~\cite{Atiyah:1989vu, Quinn}
In $\mathbf{Bord}_{<d,d+1>}$,
objects are $d$-dimensional manifolds $X_1,X_2,\dots $,
and morphism is a cobordism $Y: X_1 \to X_2$
which is a manifold of dimension $d+1$
and has $X_1$ and $X_2$ as its boundary components,
$\partial Y = (-X_1) \sqcup X_2$,
where $(-X)$ is $X$ with opposite orientation.
In general, we can associate a structure
(e.g. spin structure for spin TFTs, background gauge field for equivariant TFTs, etc.)
with manifolds.
For each $d$-dimensional manifold $X$, we associate a
Hilbert space ${\cal H}_X$ by a functor $Z$.
A direct sum of manifolds $X_1 \sqcup X_2 \sqcup \cdots $ is
mapped into a tensor product ${\cal H}_{X_1} \otimes {\cal H}_{X_2} \otimes \cdots$.
A cobordism $Y$ between $X$ and $X'$ leads to a linear map
$Z(Y) : {\cal H}_{X} \to {\cal H}_{X'}$.
See Fig.~ \ref{fig:tft_ex}, for an example.
In any TFT, the cylinder cobordism $X \times I$ leads
to the identity map $Z(X \times I) = {\rm id}$
\begin{align}
&\vcenter{\hbox{\includegraphics[width=0.2\linewidth]{figs/tft/id}}}
\ \Longrightarrow \ \
{\cal H}_X \overset{\rm id}{\longrightarrow} {\cal H}_X,
\end{align}
which is equivalent to the fact that the Hamiltonian of TFTs is zero.
In addition, we have a bilinear form $Q$ and a coform $\Delta$:
\begin{align}
&\vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/tft/Q}}} &&
\vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/tft/Delta}}} \notag \\
&{\cal H}_X \otimes {\cal H}_X \overset{Q}{\longrightarrow} \mathbb{C}, &&
\mathbb{C} \overset{\Delta}{\longrightarrow} {\cal H}_X \otimes {\cal H}_X.
\end{align}
Let $\{\phi_i \}$ be a basis of ${\cal H}_X$ and write
$Q(\phi_i, \phi_j) = Q_{ij}, \Delta(1) = \sum_{ij} \Delta_{ij} \phi_i \otimes \phi_j$.
The equivalence between the ``S-tube'' and the cylinder,
$$
\includegraphics[width=0.7\linewidth]{figs/tft/S}
$$
implies
\begin{align}
\phi \mapsto \phi \otimes \sum_{jk} \Delta_{jk} \phi_j \otimes \phi_k \mapsto \sum_i Q(\phi, \phi_j) \Delta_{jk} \phi_k = \phi , && \phi \in {\cal H}_X.
\label{eq:S-condition}
\end{align}
By setting $\phi = \phi_k$, we have $\sum_{j} Q_{ij} \Delta_{jk} = \delta_{ik}$,
which means $Q$ is nondegenerate.
Choosing the basis so that $Q_{ij} = \delta_{ij}$,
in this basis the coform $\Delta$ is simply $\Delta = \sum_{i} \phi_i \otimes \phi_i$.
\subsection{$G$-equivariant oriented closed TFTs}
A $G_0$-equivariant oriented $(1+1)$d TFT
is a functor $Z$ from
a cobordism category with a background $G_0$ gauge field
to the category of complex vector spaces.
(To distinguish on-site unitary symmetries from orientation-reversing symmetries,
here we use a notation $G_0$ to denote on-site unitary symmetries.)
For $(1+1)$d TFTs, the minimum object is an oriented circle
$(S^1,pt,g)$ with background $g \in G_0$ flux
together with a trivialization at a base point $pt \in S^1$,
which is specified by a twisted boundary condition by an element $g \in G_0$ at $pt$.
We denote the Hilbert space associated with $(S^1, pt, g \in G_0)$ by ${\cal C}_g$:
\begin{align}
&\vcenter{\hbox{\includegraphics[width=0.3\linewidth]{figs/tft/S1_g_comment}}}
\ \Longrightarrow \ \ \ \
{\cal C}_g = {\cal H}_{(S^1,pt,g)}, \ (g \in G_0).
\end{align}
For an unpointed circle $S^1$,
a background flux inserted in the circle $S^1$ is characterized by
a conjugacy class $[g] = \{ h g h^{-1} | h \in G\}$
rather than an element $g \in G_0$.
For point circles $(S^1,pt)$ with trivialization of background $G_0$ gauge field at $pt$,
Hilbert spaces are labeled by elements $g \in G_0$.
We have thus a $G_0$-graded Hilbert space
\begin{align}
{\cal C} = \bigoplus_{g \in G_0} {\cal C}_g.
\end{align}
\begin{figure}[!]
\begin{center}
\includegraphics[width=0.8\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/fusion_detail.pdf}
\end{center}
\caption{
[a] A fusion process ${\cal C}_g \otimes {\cal C}_h \to {\cal C}_{gh}$.
Dashed lines with group elements represent holonomies.
The figure [b] shows a holonomy around the boundary of simply connected space
that is obtained by cutting the surface in [a] at the lines connecting base points.
}
\label{fig:tft_fusion_detail}
\end{figure}
\begin{table}[!]
\caption{Building blocks of $G$-equivariant closed $(1+1)$d TFTs.
Building blocks (a)-(f) define $G$-equivariant oriented TFTs in $(1+1)$d.
$G$-equivariant unoriented TFTs in $(1+1)$d
are defined by including (g) and (h), in addition to (a)-(f).
$G_0 \subset G$ represents orientation preserving symmetries.
The fourth column shows corresponding simple and fixed point MPS representations (see Sec.~\ref{sec:tft-mps}). }
\begin{center}
\scalebox{0.8}{
\hspace*{-2cm}
\begin{tabular}{| >{\centering\arraybackslash}m{1cm} | >{\centering\arraybackslash}m{3cm} | >{\centering\arraybackslash}m{3.5cm} | >{\centering\arraybackslash}m{5cm} | >{\centering\arraybackslash}m{5cm} | }
\hline
& Manifolds & Hilbert spaces & Simple and fixed point MPS & Comment \\
\hline
(a) & \includegraphics[width=0.5\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/S1_g.pdf} &
${\cal C}_g \ (g \in G_0)$ & Hilbert space generated by $\ket{\Psi_g} = {\rm tr}(A_m V_g) \ket{m}, g \in G_0$ & Hilbert space over a space circle with $g$-flux \\
\hline
(b) & \includegraphics[width=0.5\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/closed_trace.pdf} &
$$
\begin{array}{l}
\theta_{\cal C}: {\cal C}_1 \to \mathbb{C}, \\
\phi \mapsto \theta_{\cal C}(\phi) \\
\end{array}
$$
& $\theta_{\cal C} \big( {\rm tr} (A_m) \ket{m} \big) = 1$ &
$Q(\phi_1, \phi_2) := \theta_{\cal C}(\phi_1 \phi_2), (\phi_1 \in {\cal C}_g , \phi_2 \in {\cal C}_{g^{-1}})$ is a bilinear nondegenerate form. \\
\hline
(c) & \includegraphics[width=0.5\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/disc.pdf} &
$$
\begin{array}{l}
\mathbb{C} \to {\cal C}_1, \\
1 \mapsto 1_{\cal C} \\
\end{array}
$$
& $1_{\cal C} = {\rm tr} (A_m) \ket{m}$ & State on the boundary of disc
$1_{\cal C} \phi = \phi 1_{\cal C} = \phi$. \\
\hline
(d) & \includegraphics[width=0.7\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/Cylinder.pdf} &
$$
\begin{array}{ll}
\alpha_{g\in G_0} : & {\cal C}_h \to{\cal C}_{ghg^{-1}}, \\
& \phi \mapsto \alpha_g(\phi) \\
\end{array}
$$
&
$$
\begin{array}{l}
{\rm tr} (A_m V_h) \ket{m} \\
\mapsto {\rm tr} (A_m V_h) \ \hat g (\ket{m}) \\
= {\rm tr} (A_m V_g V_h V_g^{\dag}) \ket{m} \\
\end{array}
$$
&
On site unitary $g \in G_0$ symmetry action \\
\hline
(e) & \includegraphics[width=\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/pants.pdf} &
$$
\begin{array}{l}
{\cal C}_g \otimes {\cal C}_h \to {\cal C}_{gh}, \\
\phi_1 \otimes \phi_2 \mapsto \phi_1 \phi_2 \\
\end{array}
$$
&
$$
\begin{array}{l}
{\rm tr} (A_{m_1} V_g) \ket{m_1} \\
\otimes {\rm tr} (A_{m_2} V_h) \ket{m_2} \\
\mapsto {\rm tr} (A_m V_g V_h) \ket{m} \\
\end{array}
$$
& ``Fusion'' of two closed chains \\
\hline
(f) & \includegraphics[width=\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/coform.pdf} &
$$
\begin{array}{l}
\Delta_g : \mathbb{C} \to {\cal C}_g \otimes {\cal C}_{g^{-1}}, \\
\Delta_g(1) = \sum_i \xi^g_i \otimes \xi_i^{g^{-1}} \\
\end{array}
$$
& ${\rm tr} (A_{m_1} V_g) \ket{m_1} \otimes {\rm tr} (A_{m_2} V_{g^{-1}}) \ket{m_2}$ &
$\xi^g_{i} \in {\cal C}_{g}$ are basis of ${\cal C}_{g}$ and
$\xi_{i}^{g^{-1}} \in {\cal C}_{g^{-1}}$
are their dual basis of ${\cal C}_{g^{-1}}$ that satisfy
$\theta_{\cal C} (\xi^g_{i} \xi_{j}^{g^{-1}}) = \theta_{\cal C} (\xi_{j}^{g^{-1}} \xi^g_{i}) = \delta_{i j}$. \\
\hline
\hline
(g) & \includegraphics[width=0.6\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/mobius} &
$$
\begin{array}{l}
\mathbb{C} \to {\cal C}_{g^2}, g \notin G_0, \\
1 \mapsto \theta_{g} \\
\end{array}
$$
& $b(g,g) {\rm tr} (A_m V_{g^2}) \ket{m}$ &
State on the boundary state of M\"{o}bius strip, ``cross cap state''. \\
\hline
(h) & \includegraphics[width=0.7\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/cylinder_P.pdf} &
$\alpha_{g\notin G_0} : {\cal C}_h \to{\cal C}_{gh^{-1}g^{-1}}$
&
$$
\begin{array}{l}
{\rm tr} (A_m V_h) \ket{m} \\
\mapsto {\rm tr} (A_m V_h) \ \hat g (\ket{m}) \\
= {\rm tr} (A^T_m V_g V_h V_g^{\dag}) \ket{m} \\
\end{array}
$$
&
$g$ reflection \\
\hline
\end{tabular}
\hspace*{-2cm}
}
\end{center}
\label{tab:functor_ori_tft}
\end{table}
In graphical representations of morphisms,
we specify the background gauge field by
holonomies connecting base points on initial and mapped circles.
For example, the fusion process of two circles with $g$ and $h$ fluxes
is represented in Fig.\ \ref{fig:tft_fusion_detail} [a].
The $G_0$ flux of mapped pointed circle $(S^1,pt)$
is determined by holonomies along base points as
shown in Fig.~\ref{fig:tft_fusion_detail} [b].
Recall that a holonomy around a boundary of simply connected spaces is trivial.
In short, we simply write the bordsim of the fusion process as
\begin{align}
&\vcenter{\hbox{\includegraphics[width=0.4\linewidth]{figs/tft/fusion_short}}}
\ \Longrightarrow \ \ \ \
{\cal C}_g \otimes {\cal C}_h \to {\cal C}_{gh}, \quad (g,h \in G_0).
\end{align}
In Table \ref{tab:functor_ori_tft},
we show building blocks of $G_0$-equivariant oriented (1+1)d TFTs.
All other cobordisms and partition functions can be constructed by processes in Table \ref{tab:functor_ori_tft}.
For example,
the
``branching'' process
is given by
\begin{align}
&\vcenter{\hbox{\includegraphics[width=0.8\linewidth]{figs/tft/separate}}}
\label{fig:tft/separate}
\end{align}
\begin{align}
\Longrightarrow Z \Big( \vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/tft/separate_2}}} \Big):
{\cal C}_g \to {\cal C}_{gh} \otimes {\cal C}_{h^{-1}}, &&
\phi \mapsto \sum_i \phi \xi_i^h \otimes \xi_i^{h^{-1}}
= \sum_i \xi_i^{gh} \otimes \xi_i^{(gh)^{-1}} \phi, &&
(g,h \in G_0).
\end{align}
Here,
we made use of Items (e) and (f) in Table\ref{tab:functor_ori_tft},
and
$\Delta_g(1) = \sum_i \xi_i^g \otimes \xi_i^{g^{-1}} \in {\cal C}_g \otimes {\cal C}_{g^{-1}}$
is the coform defined in Item (f).
The fusion process (e),
which, by an axiom of TFTs,
is associated to a map $\mathcal{C}_g \otimes \mathcal{C}_h \to \mathcal{C}_{gh}$,
makes the Hilbert space $\mathcal{C}$ into an algebra.
There are several constraints on the algebra,
which are obtained, e.g.,
by considering different factorizations
of surfaces into building blocks in Table \ref{tab:functor_ori_tft}.
Due to Turaev,~\cite{Turaev}
we have the minimum defining algebraic relations~\cite{Moore-Segal, Kapustin-Turzillo}:
To give a $G_0$-equivariant oriented TFT is
equivalent to give a $G_0$-graded algebra ${\cal C} = \bigoplus_{g \in G_0} {\cal C}_g$
together with a group homomorphism $\alpha : G_0 \to {\rm Aut}({\cal C})$
such that ${\rm Aut}({\cal C}) \ni \alpha_g : {\cal C}_h \to {\cal C}_{g h g^{-1}}$, and
\begin{itemize}
\item[(1)] There is a $G_0$-invariant trace $\theta_{\cal C} : {\cal C}_1 \to \mathbb{C}, \ \theta_{\cal C} \circ \alpha_g = \theta_{\cal C}$,
such that the induced paring ${\cal C}_g \otimes {\cal C}_{g^{-1}} \to \mathbb{C}$ is nondegenerate.
\item[(2)] For $\phi \in {\cal C}_g$, \ $\alpha_g (\phi) = \phi$.
\item[(3)] For $\phi_1 \in {\cal C}_{g_1}, \phi_2 \in {\cal C}_{g_2}$, $\alpha_{g_2}(\phi_1) \phi_2 = \phi_2 \phi_1$.
\item[(4)] (Punctured Torus) $\sum_i \alpha_{h}(\xi^g_i) \xi_i^{g^{-1}} = \sum_i \xi^h_i \alpha_{g}(\xi_i^{h^{-1}}) \in {\cal C}_{hg h^{-1} g^{-1}}$.
\end{itemize}
The non-degenerate property in (1)
is followed by the same way as (2).
Derivations of (1-4)
are summarized in Appendix \ref{app:Closed TFT}.
The state in (4) is
the handle adding operator
\begin{align}
\vcenter{\hbox{\includegraphics[width=0.2\linewidth]{figs/tft/handle}}}
= \sum_i \alpha_{h}(\xi^g_i) \xi_i^{g^{-1}},
\label{eq:tft_handle}
\end{align}
which enables us to
compute all possible partition functions on surfaces of genus $g$ with twist.
For example,
the partition function on torus $T^2$ with twist is given by
\begin{align}
\vcenter{\hbox{\includegraphics[width=0.2\linewidth]{figs/tft/Torus}}}
= Z_{T^2}(h,g)
= \sum_i \theta_{\cal C} (\alpha_h(\xi^g_i) \xi_i^{g^{-1}}), \ \ (hg=gh).
\end{align}
\subsubsection{General solution for semi simple cases}
If ${\cal C}_1$, the untwisted sector Hilbert space,
is semisimple,
${\cal C}_1 \cong \bigoplus_{x \in X} \mathbb{C} \epsilon_x$,
$\epsilon_x \epsilon_y = \delta_{x,y} \epsilon_x$,
we have general solutions for the
algebraic constraints
(1-4) as follows.~\cite{Turaev, Moore-Segal}
Here, $X$ is a finite set equipped with $G_0$-action
$g \cdot (h \cdot x) = (gh) \cdot x$.
For a given $G_0$-set $X$,
the twisted sector Hilbert space ${\cal C}_g$
consists of little group at $x$ as
${\cal C}_g = \bigoplus_{x \in X, g \cdot x = x} L_{g,x}$,
where $L_{g,x} \cong \mathbb{C}$ are lines.
The multiplication of ${\cal C} = \bigoplus_{g \in G_0} {\cal C}_g$ is determined by
a given group cocycle $b_x(g,h) \in Z^2(G_0,C(X,U(1))) \big( \cong Z^2_{G_0}(X,U(1)) \big)$ as
\footnote{$C(X,U(1))$ is the $G_0$-module consisting of
$U(1)$-valued functions on $X$.
The $G_0$-structure is defined by $(g \cdot f)(x) := f(g \cdot x)$, $g \in G_0, x \in X$.
The group cohomology $H^2(G_0,C(X,U(1)))$ classifies the following extension
$$
1 \to C(X,U(1)) \to \hat G \to G_0 \to 1.
$$
}
\begin{align}
\ell_{g_2,x_2} \ell_{g_1,x_1} =
\left\{\begin{array}{ll}
b_{x_1}(g_2, g_1) \ell_{g_2 g_1,x_1} & (x_2 = g_1 \cdot x_1) \\
\\
0 & (\mbox{otherwise})
\end{array}\right. &&
(\ell_{g,x} \in L_{g,x}).
\end{align}
The associativity condition $\ell_{g_3,x_3} (\ell_{g_2,x_2} \ell_{g_1,x_1}) = (\ell_{g_3,x_3} \ell_{g_2,x_2}) \ell_{g_1,x_1}$
corresponds to the 2-cocycle condition
\begin{align}
b_{g_1 \cdot x}(g_3,g_2) b_x(g_3 g_2, g_1) = b_x(g_2,g_1) b_x(g_3,g_2 g_1).
\end{align}
In short, $G_0$-equivariant TFTs are classified by
the group cohomology $H^2(G_0,C(X,U(1))) \big( \cong H^2_{G_0}(X,U(1)) \big)$.
To make a contact with physics of SPT phases,
let us specialize to the case where ${\cal C}_1$ is simple
${\cal C}_1 \cong \mathbb{C}$.
In this case, the ground state in the untwisted sector is unique,
and the classification is reduced into group cohomology with
$U(1)$ coefficient $H^2(G_0,U(1))$.
On the other hand, in semisimple cases,
we have a combination of symmetry breaking
and symmetry fractionalization discussed in
Refs.~\cite{Schuch2011, chen2011complete}.
Since the group cohomology $H^2(G_0, C(X,U(1)))$
splits into $G_0$-orbits,
we can simply assume that $X$ consists of a single $G_0$-orbit.
Let subgroup $G' \subset G_0$ be an unbroken symmetries,
then, we have a bijection $X \cong G_0/G'$ as a set,
which is a ``Nambu-Goldstone manifold''.
Each element $x \in G_0/G'$ represents
a vacuum which partially breaks $G_0$ symmetry and
retains $G'$ symmetry.
All the elements $G_0/G'$ are permuted by broken symmetries
in $G_0$.
The topological classification is given by
\begin{align}
H^2_{G_0}(G_0/G', U(1)) \cong H^2_{G'}(pt, U(1)) \cong H^2(G',U(1)),
\end{align}
says, the group cohomology classification for
unbroken symmetries.
\begin{figure}[!]
\begin{center}
\includegraphics[width=0.3\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/spec}
\end{center}
\caption{Example of a $G_0$-set $X$ for a combination of symmetry broken and
symmetry fractionalization.
In this example, $X$ consists of two points $X = \{x_1, x_2\}$
which are permuted by the broken symmetry $\sigma_3 \in G_0$.
}
\label{fig:tft/spec}
\end{figure}
For example,
let full symmetry be $G_0 = \mathbb{Z}_2[\sigma_1] \times \mathbb{Z}_2[\sigma_2] \times \mathbb{Z}_2[\sigma_3]$ and
unbroken symmetry be $G' = \mathbb{Z}_2[\sigma_1] \times \mathbb{Z}_2[\sigma_2]$,
where $\sigma_i (i=1,2,3)$ are generators of $\mathbb{Z}_2$.
In this case, $X$ consists of two points $\{x_1, x_2\}$ which
are exchanged by the broken symmetry as $x_2 = \sigma_3 \cdot x_1$
as shown in Fig.~\ref{fig:tft/spec}.
The topological classification is given by
that for the unbroken symmetry as
$H^2(\mathbb{Z}_2[\sigma_1] \times \mathbb{Z}_2[\sigma_2],U(1)) = \mathbb{Z}_2$.
See Fig.~\ref{fig:tft/spec}.
\subsection{$G$-equivariant unoriented closed TFTs}
$(1+1)$d oriented (closed) TFTs were extended to
unoriented TFTs by Turaev-Turner~\cite{turaev2006unoriented}
and equivariant unoriented TFTs by Kapsutin-Turzillo~\cite{Kapustin-Turzillo}.
See also Refs.~\cite{tagami2012unoriented, sweet2013equivariant}.
Here we review $G$-equivariant unoriented $(1+1)$d TFTs.
As before,
let $G$ be a full symmetry group including orientation-reversing symmetries
and $G_0 \subset G$ be the orientation-preserving subgroup.
There are two new ingredients to define
(equivariant) unoriented $(1+1)$d TFTs:
the crosscap state and reflection transformation
((g) and (h) in Table \ref{tab:functor_ori_tft}, respectively).
As for Item (g),
the boundary state of the M\"obius strip defines the
crosscap state $\theta_g \in {\cal C}_{g^2} (g \notin G_0)$
\begin{align}
\vcenter{\hbox{\includegraphics[width=0.4\linewidth]{figs/tft/crosscap}}}
\ = \theta_{g} \in {\cal C}_{g^2}, \quad (g \notin G_0).
\label{eq:tft_crosscap}
\end{align}
Notice that the crosscap state $\theta_{g}$ belongs to
the twisted sector of $g^2 \in G_0$.
As for Item (h),
the presence of an orientation-reversing symmetry $g \notin G_0$
can be used to
consider reflection of the circle
\begin{align}
\vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/tft/cylinder_P}}}
\ \Rightarrow \alpha_g : {\cal C}_h \to {\cal C}_{g h^{-1} g^{-1}},
\quad (g \notin G_0).
\end{align}
In a way similar to $G_0$-equivariant oriented TFTs,
we have several constraints on the algebraic category.
Kapustin-Turzillo\cite{Kapustin-Turzillo} showed that
to give a $G$-equivariant unoriented $(1+1)$d TFT
is equivalent to give a $G_0$-graded algebra
${\cal C} = \bigoplus_{g \in G_0} {\cal C}_g$
together with a group homomorphism
$\alpha : G \to {\rm Aut}({\cal C})$
such that $\alpha_{g \in G_0} : {\cal C}_h \to {\cal C}_{g h g^{-1}}$,
$\alpha_{g \notin G_0} : {\cal C}_h \to {\cal C}_{g h^{-1} g^{-1}}$.
They must satisfy
(1)-(4), and
\begin{itemize}
\item[(5)] $\alpha_g(\phi_1 \phi_2) = \alpha_g(\phi_2) \alpha_g(\phi_1)$, $g \notin G_0$.
\item[(6)] $\alpha_{h \in G_0}(\theta_g) = \theta_{h g h^{-1}}$ and $\alpha_{h \notin G_0}(\theta_g) = \theta_{h g^{-1} h^{-1}}$.
\item[(7)] (Punctured M\"obius strip) $\theta_g \phi = \alpha_g(\phi) \theta_{gh}$, $\phi \in {\cal C}_h$.
\item[(8)] (Punctured Klein bottle) $\sum_i \alpha_{g}(\xi^{(gh)^{-1}}_i) \xi_i^{gh} = \theta_g \theta_h$, $g,h \notin G_0$.
\end{itemize}
Derivations of these constraints\cite{Kapustin-Turzillo} are summarized in Appendix \ref{app:Closed TFT}.
All possible partition functions are constructed from
the handle adding operator (\ref{eq:tft_handle})
and crosscap adding operator (\ref{eq:tft_crosscap}).
For example,
the partition function on real projective plane reads
\begin{align}
Z_{\mathbb{R} P^2}(g)
= \vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/tft/RP2}}}
= \theta_{\cal C}(\theta_g), \qquad (g \notin G_0, g^2 = 1).
\end{align}
The Klein bottle partition function is
\begin{align}
Z_{KB}(g;h)
= \vcenter{\hbox{\includegraphics[width=0.2\linewidth]{figs/tft/KB}}}
= \theta_{\cal C}\big( \sum_i \alpha_g(\xi_i^h) \xi_i^{h^{-1}} \big)
= \theta_{\cal C}\big( \theta_g \theta_{g^{-1} h^{-1}} \big), \quad (g \notin G_0, h \in G_0).
\end{align}
\subsubsection{General solution for simple cases}
In the cases where ${\cal C}_1$ is simple ${\cal C}_1 \cong \mathbb{C}$,
i.e., there is a unique ground state,
Kapustin-Turzillo~\cite{Kapustin-Turzillo} showed general solutions
of the algebraic constraints (1) - (8).
They showed that to give a $G$-equivariant
unoriented $(1+1)$d simple TFT is to give a
2-group cycle $b(g,h) \in Z^2(G,U(1)_{\phi})$.~\footnote{
$U(1)_{\phi}$ is equipped with $G$-action defined by (\ref{eq:app_g_action}). }
This is consistent with the group cohomology classification of
bosonic $(1+1)$d SPT phases with reflection or time-reversal symmetry~\cite{chen2011complete}.
\subsubsection{Relation to MPS}
\label{sec:tft-mps}
In the SPT context, the spatial circle $S^1$ with $g$-flux in TFTs is
identified with a bulk SPT phase with $g$-twisted boundary condition.
The uniqueness condition of
the ground state in SPT phases
implies that the corresponding TFTs are invertible,
i.e.,
we have a simple algebra of untwisted sector ${\cal C}_1 \cong \mathbb{C}$.
In TFTs,
there is no excited state and the Hilbert space consists only of ground states.
The correlation length of the bulk is zero,
so a TFT is represented by a fixed point MPS
introduced in Sec.~\ref{Fixed point MPSs}
\begin{align}
\ket{\Psi_g}
= \sum_m {\rm Tr}\, (A_{m_1} \cdots A_{m_L} V_g) \ket{m_1 \cdots m_L}
\sim \sum_m {\rm Tr}\, (A_m V_g) \ket{m}.
\end{align}
Here we used the equivalence relation of fixed point MPSs (\ref{eq:mps_equivalence}).
Only one physical site is sufficient to describe the MPS representation of a TFT's ground state.
The correspondence between MPSs and equivariant TFTs,
can be pictorially represented as
\begin{align}
\vcenter{\hbox{\includegraphics[width=0.3\linewidth]{figs/tft/s1_mps}}}
\end{align}
Cobordisms in $G$-equivariant TFTs correspond to
various ``adiabatic deformations'' of closed chains, e.g.,
``fusion'' and ``separating'', and symmetry operations.
The fourth column in Table \ \ref{tab:functor_ori_tft}
summarizes correspondences between cobordisms in $G$-equivariant TFTs
and MPS representations.
For example, the fusion process of two
closed chain is {\it formally} represented in MPS networks as follows.
For two MPSs
\begin{align}
\ket{\Psi_g} = {\rm Tr} \big[ A_1 V_g \big],
\quad
\ket{\Psi_h} = {\rm Tr} \big[ A_2 V_h \big],
\end{align}
the fusion $\ket{\Psi_g} \cdot \ket{\Psi_h}$ is given by
\begin{align}
&\vcenter{\hbox{\includegraphics[width=0.6\linewidth]{figs/tft/fusion_mps}}} \\
&\Rightarrow
\ket{\Psi_g} \cdot \ket{\Psi_h} = {\rm Tr} \big[ A_2 A_1 V_g V_h \big] \ket{m_2 m_1} \sim b(g,h) {\rm Tr} \big[ A V_{gh} \big] \ket{m} = b(g,h) \ket{\Psi_{gh}}.
\end{align}
Here we used the equivalence relation of fixed point MPSs (\ref{eq:mps_equivalence}).
\subsection{$G$-equivariant open and closed TFTs}
\label{G-equivariant open and closed TFTs}
\begin{table}[h!]
\caption{Building blocks of $G$-equivariant open and closed $(1+1)$d TFTs.
In the forth column, MPS representations are shown.
In figures, dashed lines with group elements represent holonomies.
(r) is the definition of the boundary state for a boundary condition $a$.
}
\begin{center}
\scalebox{0.8}{
\hspace*{-2cm}
\begin{tabular}{| >{\centering\arraybackslash}m{1cm} | >{\centering\arraybackslash}m{3cm} | >{\centering\arraybackslash}m{3cm} | >{\centering\arraybackslash}m{6cm} | >{\centering\arraybackslash}m{5cm} | }
\hline
& Manifolds & Hilbert spaces & Simple and fixed point MPS & Comment \\
\hline
(i) & \includegraphics[width=0.2\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/Interval.pdf} &
${\cal O}_{ab}$ &
$$
\begin{array}{c}
\mbox{Hilbert space spanned by} \\
\Big\{ \big( L_i^T A_m R_j \big) \ket{m} \Big\}, \\
L_i \in V_a^*, R_j \in V_b.
\end{array}
$$ & Open chain. $L_i$ ($R_j$) are basis of
$V_a^*$ ($V_b$). \\
\hline
(j) & \includegraphics[width=0.3\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/Interval_trace.pdf} &
$$
\begin{array}{c}
\theta_{a}: {\cal O}_{aa} \to \mathbb{C}, \\
\psi \mapsto \theta_{a} (\psi) \\
\end{array}
$$
&
$$
\begin{array}{c}
\theta_a \Big( \big( v_L^T A_m v_R \big) \ket{m} \Big) = (v_L,v_R), \\
v_L \in V_a^* , v_R \in V_a \\
\end{array}
$$ &
$(v_L,v_R) = \sum_i [v_L]_i [v_R]_i$.
Notice that $\theta_a (1_a) = {\rm dim} V_a$ for simple and fixed point MPS.
\\
\hline
(k) & \includegraphics[width=0.3\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/Interval_create.pdf} &
$$
\begin{array}{c}
\mathbb{C} \to {\cal O}_{aa}, \\
1 \mapsto 1_{a} \\
\end{array}
$$
& $\sum_i \big( L_i^T A_m R_i \big) \ket{m}$ &
$1_{a}$ is the unit satisfying $1_{a} \psi = \psi 1_{b} = \psi, \psi \in {\cal O}_{ab}$. \\
\hline
(l) & \includegraphics[width=0.6\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/Interval_action.pdf} &
$$
\begin{array}{c}
\rho_{g \in G_0}: {\cal O}_{ab} \to {\cal O}_{ab}, \\
\psi \mapsto \rho_g \psi \\
\end{array}
$$
&
$$
\begin{array}{l}
\big( v_L^T A_m v_R \big) \ket{m} \\
\mapsto \big( v_L^T V_{g,a}^{\dag} A_m V_{g,b} v_R \big) \ket{m}, \\
v_L \in V_a^*, v_R \in V_b.
\end{array}
$$
& $g$-action on open chain. $V_{g,a}$ is representation matrix of $V_a$. \\
\hline
(m) & \includegraphics[width=0.6\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/Interval_action_ref.pdf} &
$$
\begin{array}{c}
\rho_{g \notin G_0}: {\cal O}_{ab} \to {\cal O}_{ab}, \\
\psi \mapsto \rho_g \psi \\
\end{array}
$$
&
$$
\begin{array}{l}
\big( v_L^T A_m v_R \big) \ket{m} \\
\mapsto \big( v_R^T V_{g,b}^{\dag} A_m V_{g,a} v_L \big) \ket{m}, \\
v_L \in V_a^*, v_R \in V_b.
\end{array}
$$
& $g$-reflection on an open chain \\
\hline
(n) & \includegraphics[width=0.6\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/Interval_Fusion.pdf} &
$$
\begin{array}{c}
{\cal O}_{ab} \otimes {\cal O}_{bc} \to {\cal O}_{ac}, \\
\psi_1 \otimes \psi_2 \mapsto \psi_1 \psi_2 \\
\end{array}
$$
&
$$
\begin{array}{c}
\big( v_L^T A_{m_1} v_R \big) \ket{m_1} \otimes \big( w_L^T A_{m_2} w_R \big) \ket{m_2} \\
\mapsto (w_L, v_R) \big( v_L^T A_m w_R \big) \ket{m}, \\
v_L \in V_a^*, v_R \in V_b, w_L \in V^*_b, w_R \in V_c.
\end{array}
$$
& Fusion of two open chains \\
\hline
(o) & \includegraphics[width=0.5\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/coform_interval.pdf} &
$$
\begin{array}{c}
\mathbb{C} \to {\cal O}_{ab} \otimes {\cal O}_{ba}, \\
1 \mapsto \sum_{\mu} \psi_{\mu} \otimes \psi^{\mu} \\
\end{array}
$$
& $\sum_{ij} \big( L_i^T A_{m_1} R_j \big) \ket{m_1} \otimes \big( L_j^T A_{m_2} R_i \big) \ket{m_2}$ &
$\psi_{\mu} \in {\cal O}_{ab}$ are basis of ${\cal O}_{ab}$ and $\psi^{\mu} \in {\cal O}_{ba}$ are their dual of ${\cal O}_{ba}$ that satisfy $\theta_{a} (\psi_{\mu} \psi^{\nu}) = \delta_{\mu}^{\nu}$. \\
\hline
(p) & \includegraphics[width=0.66\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/OpenToClose.pdf} &
$$
\begin{array}{c}
\imath^{g,a} : {\cal O}_{aa} \to {\cal C}_g, \\
\psi \mapsto \imath^{g,a}(\psi) \\
\end{array}
$$
& $\big( v_L^T A_m v_R \big) \ket{m} \mapsto (V_g^* v_L, v_R) {\rm Tr}(A_m V_g) \ket{m}$ & Open to closed map \\
\hline
(q) & \includegraphics[width=0.6\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/CloseToOpen.pdf} &
$$
\begin{array}{c}
\imath_{g,a} : {\cal C}_{g} \to {\cal O}_{aa}, \\
\phi \mapsto \imath_{g,a}(\phi) \\
\end{array}
$$
&
${\rm Tr}(A_m V_g) \ket{m} \mapsto \sum_i \big( L_i^T A_m V_g R_i \big) \ket{m}$ & Closed to open map \\
\hline \hline
(r) & \includegraphics[width=0.7\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/bs} &
$B_a = \imath^{g,a}(1_a) \in {\cal C}_g$
&
${\rm tr}\big( V_{g,a}^{\dag} \big) {\rm Tr}(A_m V_{g,a}) \ket{m}$ &
Boundary state for boundary condition $a$.
\\
\hline
\end{tabular}
\hspace*{-2cm}
}
\end{center}
\label{Tab:Functor_Open}
\end{table}
\clearpage
Next, we extend closed TFTs to include open chains (intervals).
A new object is an oriented interval $I_{ab} = [0,1]$ with boundary conditions
$a,b$ as shown in Table.\ \ref{Tab:Functor_Open} (i).~\footnote{
In string theory,
the boundary conditions $a, b, \dots$ are
Chan-Paton factors associated with the
endpoints of open strings. }
We denote the Hilbert space associated with the interval $I_{ab}$ by ${\cal O}_{ab}$.
An element $\psi \in {\cal O}_{ab}$ represents
a state living in the open chain with boundary conditions $a$ and $b$.
(Note that the boundary conditions $a, b$ do not represent
some states in the open chain. )
Similar to closed TFTs,
we have several cobordisms in open and closed TFTs.
Table \ref{Tab:Functor_Open} summarizes the building blocks.
We have some remarks in order.
\begin{itemize}
\item
We use the same notation as Moore-Segal~\cite{Moore-Segal}.
The fusion process is represented as
\begin{align}
\vcenter{\hbox{\includegraphics[width=0.2\linewidth]{figs/tft/fusion_open}}}
\Rightarrow \
{\cal O}_{ab} \otimes {\cal O}_{bc} \to {\cal O}_{ac}, \ \ \psi_1 \otimes \psi_2 \mapsto \psi_1 \psi_2.
\end{align}
Note the order of two intervals $I_{ab}$ and $I_{bc}$.
\item
In addition to on-site symmetry transformation
$\rho_{g \in G_0} : {\cal O}_{ab} \to {\cal O}_{ab}$,
we have reflection on an open chain
$\rho_{g \notin G_0}: {\cal O}_{ab} \to {\cal O}_{ba}$
which exchanges the boundary conditions $a,b$.
$\rho_g$ satisfies $\rho_g \circ \rho_h = \rho_{gh}$ $(g,h \in G)$.
\item
Essentially new ingredients are
the open-to-closed map $\imath^{g,a}$
and the closed-to-open map $\imath_{g,a}$
which connect closed chains and open chains as~\cite{Moore-Segal}
\begin{align}
\imath^{g,a} : {\cal O}_{aa} \to {\cal C}_g, &&
\imath_{g,a} : {\cal C}_g \to {\cal O}_{aa}, &&
(g \in G_0).
\end{align}
Here, to glue back to a closed chain from a open chain,
the boundary conditions should agree.
\end{itemize}
All bordsims can be constructed by using building blocks listed in Table \ref{Tab:Functor_Open}.
For example,
a ``branching'' process of an open chain
is given by the same way as (\ref{fig:tft/separate}),
\begin{align}
&
\vcenter{\hbox{\includegraphics[width=0.6\linewidth]{figs/tft/separate_interval}}}
\nonumber \\
& \Rightarrow \ \
{\cal O}_{ab} \to {\cal O}_{ac} \otimes {\cal O}_{cb}, \ \
\psi \mapsto
\sum_{\mu} \psi \psi_{\mu} \otimes \psi^{\mu}
= \sum_{\mu} \tilde \psi_{\mu} \otimes \tilde \psi^{\mu} \psi.
\end{align}
Here, $\sum_{\mu} \psi_{\mu} \otimes \psi^{\mu} (\psi_{\mu} \in {\cal O}_{bc}, \psi^{\mu} \in {\cal O}_{cb})$
and $\sum_{\mu} \tilde \psi_{\mu} \otimes \tilde \psi^{\mu} (\tilde \psi_{\mu} \in {\cal O}_{ac}, \tilde \psi^{\mu} \in {\cal O}_{ca})$
are coform defined in (o) of Table \ref{Tab:Functor_Open}.
In the target algebraic category,
there are constraints from the open and closed cobordism category.
We have the following constraints for oriented open and closed TFTs by Moore-Segal~\cite{Moore-Segal}:
\begin{itemize}
\item[(9)] \
$\imath_{1,a}(1_{\cal C}) = 1_{{\cal O}_{aa}}$.
\item[(10)]\
$\rho_g (\psi_1 \psi_2) = (\rho_g \psi_1) (\rho_g\psi_2), \ g \in G_0, \psi_1 \in {\cal O}_{ab}, \psi_2 \in {\cal O}_{bc}$.
\item[(11)]\
$\imath_{g_1,a}(\phi_1) \imath_{g_2,a}(\phi_2) = \imath_{g_2 g_1,a}(\phi_2 \phi_1), \ \phi_1 \in {\cal C}_{g_1}, \phi_2 \in {\cal C}_{g_2}$.
\item[(12)]\
$\imath_{g,a}(\phi) (\rho_g \psi) = \psi \imath_{g,a}(\phi), \ \phi \in {\cal C}_g, \psi \in {\cal O}_{aa}$.
\item[(13)]\
$\theta_{a} ( \psi \imath_{g^{-1},a}(\phi) ) = \theta_{\cal C} ( \imath^{g,a}(\psi) \phi ), \ \phi \in {\cal C}_{g^{-1}}, \psi \in {\cal O}_{aa}$.
\item[(14)]\ ($G$-equivariant Cardy condition)
$\pi_{g,b}^a = \imath_{g,b} \circ \imath^{g,a} \mbox{ with } \pi_{g,b}^a(\psi) = \sum_{\mu} \psi^{\mu} \psi (\rho_g \psi_{\mu}),\ g \in G_0$.
\end{itemize}
For unoriented open and closed TFTs,
one can find the following additional constraints:
\begin{itemize}
\item[(15)]\
$\rho_g (\psi_1 \psi_2) = (\rho_g \psi_2) (\rho_g \psi_1), \ g \notin G_0, \psi_1 \in {\cal O}_{ab}, \psi_2 \in {\cal O}_{bc}$.
\item[(16)]\
$\sum_{\mu} (\rho_g \psi_{\mu} ) \psi^{\mu} = \imath_{g^2,a} (\theta_g), \ \ g \notin G_0, \ \psi_{\mu}, \psi^{\mu} \in {\cal O}_{aa}$.
\end{itemize}
In Appendix \ref{app:open TFT}, we summarize the derivations
of these constraints.
By solving these constraints, we can determine the
general properties of the target algebraic category for a
given $G$-equivariant closed TFT $b \in Z^2_G(X,U(1)_{\phi})$
with $G$-set $X$.
In the cases where ${\cal C}_1$ is semisimple, i.e.,
combination of symmetry breaking and symmetry fractionalization,
and there is no orientation-reversing symmetry,
Moore-Segal~\cite{Moore-Segal} gives the complete solution:
$b$-twisted equivariant vector bundles over $X$.
Here, for simplicity, we assume the ground state of closed chain is unique,
${\cal C}_1 \cong \mathbb{C}$,
and there are only on-site symmetries $G_0$.
We have~\cite{Moore-Segal}
\begin{itemize}
\item
The category of boundary conditions $\{a,b,\dots \}$
is equivalent to the category of $b$-projective
representations $\{V_a, V_b, \dots \}$.
\item
${\cal O}_{ab} \cong {\rm Hom}(V_b^*,V_a^*) = V_a^* \otimes V_b$.
\end{itemize}
This is precisely the boundary degrees of freedom that appear when one
introduce a boundary in SPT phases.
In the next section,
we describe
how to represent elements of ${\cal O}_{ab}$ and cobordisms
by using simple and fixed point MPS for open chains.
\subsubsection{Relation to open MPS}
In the SPT context, an interval $I_{ab}$ is
identified with an open SPT phase with
boundary condition $a$ and $b$.
An element of $\psi \in {\cal O}_{ab} \cong V_a^* \otimes V_b$
is identified with a state of the open chain Hilbert space~\footnote{
Note that $a,b$ do not specify a state
in the representation space of the $b(g,h)$-projective representations.
For example, the dihedral group $D_4 = \{1, C_4, C_2, C_4^{-1}, \sigma_x, \sigma_y, \sigma_d, \sigma_d'\}$ has two
$b$-irreps $E_{\frac{1}{2}}$ and $E_{\frac{3}{2}}$ for the nontrivial two cocycle $[b] \in H^2(D_4,U(1)) \cong \mathbb{Z}_2$.
In this case, $a, b$ specify $E_{\frac{1}{2}}$ or $E_{\frac{3}{2}}$.
}
\begin{align}
\psi
= \sum_m \big( v_L^T A_m v_R \big) \ket{m}
= \sum_m [v_L]_i [A_m]_{ij} [v_R]_j \ket{m},
\quad
v_L \in V_a^*,
\quad
v_R \in V_b.
\end{align}
The correspondence between MPSs and equivariant TFTs,
can be pictorially represented as
\begin{align}
\vcenter{\hbox{\includegraphics[width=0.2\linewidth]{figs/tft/interval_mps}}}
\end{align}
Cobordisms in $G$-equivariant open and closed TFTs correspond to
various ``adiabatic deformations'' of open chains and closed chains.
The fourth column in Table \ \ref{Tab:Functor_Open}
summarizes MPS representations,
which satisfy algebraic constraints (9) - (16).
For example, the fusion process of two
open chains is represented in MPS networks as follows.
For two open MPSs
\begin{align}
\psi_1 &= \sum_m v_L^T A^{(1)}_m v_R \ket{m}, \quad v_L \in V_a^*, v_R \in V_b, \\
\psi_2 &= \sum_m w_L^T A^{(2)}_m w_R \ket{m}, \quad w_L \in V_b^*, w_R \in V_c,
\end{align}
the fusion $\psi_1 \psi_2$ is given by
\begin{align}
\vcenter{\hbox{\includegraphics[width=0.4\linewidth]{figs/tft/interval_fusion_mps}}} &&
\Longrightarrow &&
\begin{array}{ll}
\psi_1 \psi_2 & = (w_L,v_R) \sum_{m_1 m_2} v_L^T A^{(1)}_{m_1} A^{(2)}_{m_2} w_R \ket{m_1 m_2} \\
& \\
& \sim (w_L,v_R) \sum_{m} v_L^T A_{m} w_R \ket{m} \\
\end{array}
\end{align}
Here we introduced a notation $(w_L,v_R) = \sum_i [w_L]_i [v_R]_i$
and used an equivalence relation of fixed point MPSs (\ref{eq:mps_equivalence}).
\subsubsection{Equivariant Cardy conditions and boundary states}
\begin{figure}[!]
\begin{center}
\includegraphics[width=0.7\linewidth, trim=0cm 2cm 0cm 0cm]{figs/tft/G_Cardy.pdf}
\end{center}
\caption{
$G$-equivariant Cardy condition.}
\label{fig:G_Cardy}
\end{figure}
In the derivation of the category of boundary conditions by
Moore-Segal~\cite{Moore-Segal},
the (generalized) $G$-equivariant Cardy condition (14) plays an essential role.
Here, we show MPS representations listed in the fourth column
in Table \ref{Tab:Functor_Open} satisfy the $G$-equivariant Cardy condition.
The Cardy condition comes from the equivalence between
(i) the double twist diagram shown in the left of Fig. \ref{fig:G_Cardy}
and
(ii) closed string channel shown in the right of Fig. \ref{fig:G_Cardy}.
These diagram can be interpreted in the context of
opne SPT chains :
\begin{itemize}
\item[(i)]
Cutting an open SPT chain $I_{aa}$ into two open chains $I_{ab} \sqcup I_{ba}$
and taking the $G_0$-action on the left chain ${\cal O}_{ab}$ and
exchanging two open chains and gluing back at $a$.
This process is written as
\begin{align}
\psi
= \big( v_L^T A_m v_R \big) \ket{m}
&\mapsto \sum_{j} \big( v_L^T A_{m_1} R_j \big) \ket{m_1} \otimes \big( L_j^T A_{m_2} v_R \big) \ket{m_2}
\nonumber \\
&\mapsto \sum_{j} \big( v_L^T V_g^{\dag} A_{m_1} V_g R_j \big) \ket{m_1} \otimes \big( L_j^T A_{m_2} v_R \big) \ket{m_2}
\nonumber \\
&\mapsto \sum_{j} \big( L_j^T A_{m_2} v_R \big) \ket{m_2} \otimes \big( v_L^T V_g^{\dag} A_{m_1} V_g R_j \big) \ket{m_1}
\nonumber \\
&\mapsto (V_g^* v_L, v_R) \sum_{j} \big( L_j^T A_m R_j \big) \ket{m}.
\end{align}
\item[(ii)]
Gluing the both ends of open SPT chain $I_{aa}$ to the $g$-twisted closed chain $(S^1,g)$ and cutting into the open chain $I_{bb}$.
This process is expressed as
\begin{align}
\psi
= \big( v_L^T A_m v_R \big) \ket{m}
&\mapsto (V_g^* v_L, v_R) {\rm tr} \big( A_m V_g \big) \ket{m}
\mapsto (V_g^* v_L, v_R) \sum_{j} \big( L_j^T A_m R_j \big) \ket{m}.
\end{align}
\end{itemize}
\begin{figure}[!]
\begin{center}
\includegraphics[width=1\linewidth, trim=0cm 2cm 0cm 0cm]{figs/tft/bs_def}
\end{center}
\caption{
The definition of $G$-equivariant boundary state $B_{g,a}$.
[a] Correlation functions on the upper half plane with boundary condition $a$.
[b] Amplitude of closed sector with insertion of the boundary state.
}
\label{fig:bs_def}
\end{figure}
It is useful to introduce the equivariant boundary state
$B_{g,a} \in {\cal C}_g$ in a way similar to
usual boundary state $B_a$ for non-equivariant TFTs.
Defining property of boundary state is that
the correlation functions on upper half plane with boundary condition $a$ (Fig.~\ref{fig:bs_def} [a])
is the same as the closed string amplitude with insertion of the boundary state (Fig.~\ref{fig:bs_def} [b]):
\begin{align}
\theta_a \big( \imath^{h,a}(\phi_1 \phi_2 \cdots \phi_n) \big)
= \theta_{\cal C} \big( B_{h^{-1}, a} \phi_1 \phi_2 \cdots \phi_n \big), &&
\phi_i \in {\cal C}_{g_i}, h = g_1 g_2 \cdots g_n.
\end{align}
From the algebraic constraint (13),
the $G$-equivariant boundary state
is given by
the image of open to closed map on the unit element $1_a$ of the open chain ${\cal O}_{aa}$,
\begin{align}
B_{g,a}
:= \vcenter{\hbox{\includegraphics[width=0.1\linewidth]{figs/tft/bs_2}}}
= \imath^{g,a}(1_{a}) \in {\cal C}_g.
\end{align}
For simple and fixed point MPSs, we have
\begin{align}
B_{g,a}
= {\rm tr} \big( V_{g,a}^{\dag} \big) {\rm Tr} \big( A_m V_{g,a} \big) \ket{m},
\end{align}
where $V_{g,a}$ is the representation matrix of the $V_a$ representation.
Notice that $\chi_a(g)^* = {\rm tr} \big( V_{g,a}^{\dag} \big)$
is the character of $V_a$ representation, which is
vacuous if there is a group element element $h \in G_0, [g,h]=0$
with nontrivial discrete torsion phase $b(g,h) \neq b(h,g)$.~\cite{cho2016relationship}
If we insert the boundary states in the Cardy condition (14),
we get a more familiar form
\begin{align}
\Braket{B_{g,b} | B_{g,a}}
= \theta_{\cal C} \big( B_{g^{-1},b} B_{g,a} \big)
= \theta_b \big( \imath_{g,b} \circ \imath^{g,a} (1_a) \big)
= \theta_b \big( \pi_{g,b}^a (1_a) \big)
= {\rm Tr}_{{\cal O}_{ab}} (\rho_g),
\end{align}
which is the character of $G$-action on the
open chain Hilbert space ${\cal O}_{ab}$,
\begin{align}
\chi_{{\cal O}_{ab}}(g)
= {\rm Tr}_{{\cal O}_{ab}} (\rho_g)
= \chi_a(g)^* \chi_b(g) .
\end{align}
\subsubsection{Crosscap invariant in open chain}
The partition function on $\mathbb{R} P^2$,
$Z_{\mathbb{R} P^2}(g) = \theta(g), g \notin G_0, g^2 = 1$,
can be detected in open chains.
Making use of the algebraic relation (16),
one can find the M\"obius strip with boundary condition $a$
is equivalent to the closed string amplitude from
crosscap $\theta_g$ to boundary state $B_{g^2,a}$,
\begin{align}
&\vcenter{\hbox{\includegraphics[width=0.5\linewidth]{figs/tft/mobius_a}}}
\nonumber \\
&\Rightarrow \
\braket{B_{g^2,a} | \theta_g}
= \theta_{\cal C}(B_{g^{-2},a} \theta_g)
= \theta_a(\imath_{g^2,a}(\theta_g))
= {\rm Tr}_{{\cal O}_{aa}} (\rho_g) , \ \ g \notin G_0.
\end{align}
For $g^2 = 1$ and unique ground state ${\cal C}_1 \cong \mathbb{C}$,
we have the topological invariant on the real projective plane,
which can be confirmed in simple and fixed point MPS as
\begin{align}
\theta_{a} \Big( \sum_{ij} \big( R_j^T V_g^{\dag} A_{m_1} V_g L_i \big) \ket{m_1} \cdot \big( L_j^T A_{m_2} R_i \big) \ket{m_2} \Big)
= [V_g]_{ji} [V_g^{\dag}]_{ij}
= {\rm dim}(V_a) \ \theta(g).
\end{align}
\subsection{State sum construction}
\label{Fukuma-Hosono-Kawai state sum construction}
In this section, we discuss the so-called state sum construction of TFTs.
Compared with the axiomatic approaches discussed previously,
the state sum construction exploits specific discretizations (triangulations)
of spacetime.
We will first review this construction for standard (non-equivariant) TFTs
following Fukuma-Hosono-Kawai \cite{FHK}.
We will then consider the state sum construction of $G$-equivariant TFTs,
and compute, among others, the partition functions on the torus, Klein bottle, and real
projective plane.
As promised earlier, we will confirm that they match precisely with the
topological (SPT) invariants derived from MPSs.
\subsubsection{Fukuma-Hosono-Kawai state sum construction}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.5 \linewidth, trim=0cm 0cm 0cm 0cm]{figs/state_sum/triangulation}
\end{center}
\caption{Triangulation of two-dimensional spacetime
and its dual (represented by double lines).}
\label{fig:state_sum/triangulation}
\end{figure}
Let us start by briefly reviewing the Fukuma-Hosono-Kawai
state sum construction.
\cite{FHK}
In the state sum construction of oriented 2d TFTs,
we consider a triangulation of 2d spacetime.
For a given triangulation, we can consider its dual, the dual
triangulation -- see Fig.\ \ref{fig:state_sum/triangulation}.
For faces and edges of the triangulation,
we associate
$\mathbb{C}$ numbers $C_{\mu \nu \rho}$ and $g^{\mu \nu}$ $(\mu,\nu,\rho = 1, \dots, N)$ as
\begin{align}
&
\vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/state_sum/triangle}}} \
= \vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/state_sum/dual_triangle}}} \
= \ C_{\mu \nu \rho},
\\
&
\vcenter{\hbox{\includegraphics[width=0.1\linewidth]{figs/state_sum/metric}}} \
= \vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/state_sum/dual_metric}}} \
= \ g^{\mu \nu}.
\end{align}
We demand that
$C_{\mu \nu \rho}$ is cyclically symmetric
$C_{\mu \nu \rho} = C_{\nu \rho \mu} = C_{\rho \mu \nu}$,
and $g^{\mu \nu}$ is symmetric $g^{\mu \nu} = g^{\nu \mu}$.
$g_{\mu \nu}$ is defined as the inverse of $g^{\mu \nu}$,
$g^{\mu \nu} g_{\nu \rho} = \delta^{\mu}_{\rho}$.
$g_{\mu \nu}$ and $g^{\mu \nu}$ are used for raising and lowering indices.
For example,
we introduce ${C_{\mu \nu}}^{\rho} = C_{\mu \nu \sigma} g^{\sigma \rho}$.
For a given triangulation $\Sigma_T$
of a surface $\Sigma$,
the partition function on $\Sigma_T$
is given by
\begin{align}
Z(\Sigma_T) = \sum_{\rm faces} C_{\mu \nu \rho} \sum_{\rm edges} g^{\eta \epsilon}.
\end{align}
In order to make $Z(\Sigma_T)$ independent of triangulations,
$C$ and $g$ have to satisfy the fusion and bubble conditions
\begin{align}
\label{eq:fusion_cond}
&\vcenter{\hbox{\includegraphics[width=0.3\linewidth]{figs/state_sum/fusion}}} : \qquad
{C_{\mu \nu}}^{\eta} {C_{\eta \rho}}^{\sigma} = {C_{\nu \rho}}^{\eta} {C_{\mu \eta}}^{\sigma},
\\
\label{eq:bubble_cond}
&\vcenter{\hbox{\includegraphics[width=0.3\linewidth]{figs/state_sum/bubble}}} : \qquad
g_{\mu \nu} = {C_{\mu \rho}}^{\sigma} {C_{\nu \sigma}}^{\rho}.
\end{align}
From the data of $C_{\mu \nu \rho}$ and $g^{\mu \nu}$,
one can introduce an algebra ${\cal C} = \oplus_{\mu=1}^N \mathbb{C} \phi_{\mu}$
as
\begin{align}
\phi_{\mu} \phi_{\nu} = C_{\mu \nu \rho} \phi_{\rho}.
\end{align}
The fusion condition (\ref{eq:fusion_cond}) means ${\cal C}$ is associative
$(\phi_{\mu} \phi_{\nu}) \phi_{\rho} = \phi_{\mu} (\phi_{\nu} \phi_{\rho})$.
We define a bilinear form by $Q(\phi_{\mu}, \phi_{\nu}) = g_{\mu \nu}$.
Existence of inverse of $g_{\mu \nu}$ ensures that $Q$ is non-degenerate and
the algebra ${\cal C}$ is semi simple.
The cyclicity condition of $C_{\mu \nu \rho}$
leads to the Frobenius condition
$Q(\phi_{\mu} \phi_{\nu}, \phi_{\rho}) = Q(\phi_{\mu}, \phi_{\nu} \phi_{\rho})$,
i.e.,\ ${\cal C}$ is a semi simple Frobenius algebra.
One can show that all the physical observables
constructed from the data $C_{\mu \nu \rho}$ and $g^{\mu \nu}$
depend only on the
center of ${\cal C}$,
$Z({\cal C}) = \{\phi \in {\cal C} | \phi \phi' = \phi' \phi , \forall \phi' \in {\cal C} \}$.~\cite{FHK}
In other words,
the Fukuma-Hosono-Kawai state sum construction
describes 2d oriented TFTs
which are equivariant to commutative semisimple Frobenius algebras.
For example, for a matrix algebra ${\cal C} = {\rm Mat}(\mathbb{C}^N)$ with
$Q(A,B) := {\rm Tr} A B$,
the center is trivial: $Z({\rm Mat}(\mathbb{C}^N)) = \mathbb{C} 1_{N \times N}$.
\subsubsection{$G$-equivariant state sum construction}
The state sum construction of $G$-equivariant closed TFTs
(both oriented and unoriented)
can be formulated in a way analogous to
the Fukuma-Hosono-Kawai construction of 2d oriented TFTs.~\cite{Turaev}
In the following, we
will discuss this within the context of TFTs describing SPT phases.
As before,
let $G$ be a symmetry group
which possibly includes orientation-reversing symmetries.
We specify orientation-preserving elements by subgroup $G_0 \subset G$.
We fix a group 2-cocycle $b(g,h) \in Z^2(G,U(1)_{\phi})$ and
assume $[b(g,h)] \in H^2(G,U(1)_{\phi})$ is nontrivial.
Let $V$ be a $b$-projective $N$-dimensional irrep.\ and $V^*$ be its dual.
Recall that $V$ represents the ``bond Hilbert space'' in MPSs.
$V$ also play an analogous role in the state sum construction,
which will be developed in the following.
The $G$ symmetry is projectively represented in the bond Hilbert space as
\begin{align}
&\hat g (\ket{i})
= \ket{j} [V_g]_{ji},
\quad V_g V_h = b(g,h) V_{gh},
\quad
\ket{i}, \ket{j} \in V,
\end{align}
in the same way as Sec.\ \ref{Fixed point MPSs}.
As in the Fukuma-Hosono-Kawai construction, we need
the input data -- the Frobenius algebra -- to boot-strap
a $G$-equivariant TFT.
To describe SPT phases (i.e., invertible TFTs),
we take the matrix algebra of $V$
as the algebra ${\cal C}$,
${\cal C} := {\rm End}(V) \cong V \otimes V^*$.
The bilinear non-degenerate form is defined
by the matrix trace $Q(X,Y) = N {\rm tr} (X Y) = N \sum_{ij} X_{ij} Y_{ji}$.
A canonical basis of ${\cal C}$ can be given as
\begin{align}
\left\{ E_{ij} = \ket{i} \bra{j} \right\}.
\end{align}
In this basis, $C_{ij, kl, mn} = C(E_{ij}, E_{kl}, E_{mn})$ and $g = g^{ij,kl} E_{ij} \otimes E_{kl}$
are given by
\begin{align}
C_{ij,kl,mn}& =
N \
\vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/state_sum/G_triangle}}} \
= N \delta_{ml} \delta_{kj} \delta_{in},
\\
g^{ij,kl} &=
\frac{1}{N} \ \vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/state_sum/G_metric}}} \
=\frac{1}{N} \delta^{il} \delta^{kj}.
\end{align}
One can show
$Q(E_{ij},E_{kl}) = N g_{ij,kl} = N \delta_{il} \delta_{kj}$,
${C_{ij,kl}}^{mn} = \delta_{kj} \delta_i^m \delta_l^n$,
$E_{ij} E_{kl} = {C_{ij,kl}}^{mn} E_{mn}$,
and can check
(\ref{eq:fusion_cond}),
and (\ref{eq:bubble_cond}).
Form the construction,
the algebra ${\cal C}$ has $G$ action
\begin{align}
&\hat g (X) = V_g X V^{\dag}_g, \quad X \in {\cal C}, g \in G_0, \\
&\hat P (X) = V_P X^T V^{\dag}_P, \quad X \in {\cal C}, P \notin G_0.
\end{align}
Here, observe that the orientation-reversing symmetry $P \notin G_0$ exchanges left and right.
This $G$ action can be
used to
to incorporate the background $G_0$ gauge field
in the networks of the state sum construction.
We introduce a symmetry twisted metric
by
\begin{align}
[T_g]^{ij,kl}
:= g^{ij,pq} Q\big( E_{pq}, \hat g(E_{rs}) \big) g^{rs,kl}
= \frac{1}{N} [V_g]_{il} [V^{\dag}_g]_{kj}
= \vcenter{\hbox{\includegraphics[width=0.25\linewidth]{figs/state_sum/G_twisted_metric}}}
\quad (g \in G_0).
\end{align}
We replace $g^{ij,kl}$ by $[T_g]^{ij,kl}$ on a nontrivial 1-cycle of the triangulation.
On the other hand, an orientation reversing symmetry $g \notin G_0$ induces
the exchange of indices $i$ and $j$.
We introduce the orientation reversing twisted metric~\cite{Karimipour-Mostafazadeh} by
\begin{align}
[T_P]^{ij,kl}
:= g^{ij,pq} Q\big( E_{pq}, \hat P(E_{rs}) \big) g^{rs,kl}
= \frac{1}{N} [V_P]_{ik} [V^{\dag}_P]_{lj}
= \vcenter{\hbox{\includegraphics[width=0.3\linewidth]{figs/state_sum/reflection_metric}}}
\quad (P \notin G_0).
\end{align}
\subsubsection{Partition functions}
Let us now construct, by using the state sum,
the partition functions on $T^2$, the Klein bottle, and $\mathbb{R}P^2$
(with symmetry twist).
We will show that these match precisely with the
topological invariants discussed and constructed by using MPSs
in Sec.\ \ref{Topological invariants}.
\paragraph{Partition function on $T^2$ with twist}
A background $G_0$ gauge field
on a torus $T^2$ is specified by
two commuting elements $g,h \in G_0, [g,h] = 0, g,h \in G_0$.
From the twisted metrics $T_g, T_h$
we have the torus partition function with twist
\begin{align}
Z_{T^2}(g,h)
&= \vcenter{\hbox{\includegraphics[width=0.2\linewidth]{figs/state_sum/Torus}}}
= {C_{ij,kl}}^{mn} C_{mn,pq,rs} [T_g]^{ij,pq} [T_h]^{rs,kl} \\
&= \frac{1}{N} {\rm tr}(V_g V_h V^{\dag}_g V^{\dag}_h) = \epsilon(g,h), \ \
(g,h \in G_0, h g h^{-1} = g).
\end{align}
This is the discrete torsion phase (\ref{eq:T^2inv}),
a topological invariant that characterizes $H^2(G,U(1)_{\phi})$.
\paragraph{Partition function on the Klein bottle with twist}
Similar to the torus partition function with twist,
the Klein bottle partition function with twist is computed in
the state sum construction.
Let $P \notin G_0$ be an orientation reversing symmetry and
$g \in G_0$ be a orientation preserving symmetry.
We have
\begin{align}
Z_{KB}(P;g)
&= \vcenter{\hbox{\includegraphics[width=0.2\linewidth]{figs/state_sum/KB}}}
= {C_{ij,kl}}^{mn} C_{mn,pq,rs} [T_P]^{ij,pq} [T_g]^{rs,kl} \\
&= \frac{1}{N} {\rm tr}(V_P V_g^T V^{\dag}_P V^{\dag}_g) = \kappa(P;g) , \ \
(g \in G_0, P \notin G_0, P g^{-1} P^{-1} = g).
\end{align}
Here, $\kappa(P;g)$ is the Klein bottle invariant of $H^2(G,U(1)_{\phi})$
introduced in (\ref{eq:KBinv}).
\paragraph{Partition function on $\mathbb{R} P^2$}
By using the orientation-reversing symmetry $P \in G_0$,
we can construct the partition funciton on
the real projective plane $\mathbb{R} P^2$ as
\begin{align}
Z_{\mathbb{R} P^2}(P)
&= \vcenter{\hbox{\includegraphics[width=0.2\linewidth]{figs/state_sum/RP2}}}
= \vcenter{\hbox{\includegraphics[width=0.2\linewidth]{figs/state_sum/RP2_2}}} \\
&= {C_{ij,kl}}^{mn} C_{mn,pq,rs} [T_P]^{ij,pq} [T_P]^{rs,kl}
= \frac{1}{N^2} {\rm tr}(V_P V_P^*) {\rm tr}(V_P V_P^{\dag}) = \theta(P),
\quad
(P \notin G_0, P^2 = 1).
\end{align}
This is the cross cap invariant (\ref{eq:Corsscapinv}).
\subsubsection{Cobordisms}
In addition to the closed surfaces considered above,
we can also consider surfaces with boundaries by using
the state sum construction.
From the generalities of TFTs,
a surface with boundary represents a state of the Hilbert space.
Here, we will construct various states that can be obtained
by considering state sum with open boundary/boundaries.
For our TFTs that describe SPT phases,
the physical Hilbert space ${\cal C}$ is spanned by a
basis of algebra $\{E_{ij}\}_{i,j = 1}^N$.
\paragraph{Disc (cap state)}
By the path integral on the disc,
we define a state associated to the disc (the cap state).
The cap state is the vacuum state on untwisted sector.
By triangulating the disc, the path-integral can be evaluated
explicitly as
\begin{align}
1_{\cal C}
&= \vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/tft/disc}}}
= \vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/state_sum/disc_1}}}
= \vcenter{\hbox{\includegraphics[width=0.1\linewidth]{figs/state_sum/disc_2}}}
\nonumber \\
&
= {C_{kl,mn}}^{mn} g^{kl,ij} E_{ij}
= \delta^{ij} E_{ij}
= \sum_i \ket{i} \bra{i}.
\end{align}
This is nothing but the simple and fixed point MPS representation of the ground state of SPT phases
$\ket{\Psi_1} = {\rm tr}(A_{ij}) \ket{i^L} \otimes \ket{j^R} = \frac{1}{\sqrt{N}} \sum_{i} \ket{i^L} \otimes \ket{i^R}$
introduced in Sec.\ \ref{Fixed point MPSs} up to a normalization.
\paragraph{M\"obius strip (cross cap state)}
By the path integral on the M\"obius strip,
we define a state associated to the M\"obius strip (the cross cap state).
By triangulating the M\"obius strip, the path-integral can be evaluated
explicitly as
\begin{align}
\theta_P
&= \vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/state_sum/mobius}}}
= \vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/state_sum/Mobius_1}}}
= \vcenter{\hbox{\includegraphics[width=0.1\linewidth]{figs/state_sum/Mobius_2}}}
\nonumber \\
&
= {C_{mn,pq}}^{ij} [T_P]^{mn,pq} E_{ij}
= \frac{1}{N} [V_P V_P^*]_{ij} E_{ij}
= \frac{1}{N} b(P,P) \sum_{ij} [V_{P^2}]_{ij} \ket{i} \bra{j},
\quad
(P \notin G_0).
\end{align}
\paragraph{Coform $\Delta_g$}
For a cylinder with two outgoing circles, following the axiom of TFTs,
we associate a coform $\Delta_g$.
By triangulating the cylinder,
we have
\begin{align}
\Delta_g
&= \vcenter{\hbox{\includegraphics[width=0.2\linewidth]{figs/tft/coform}}}
= \vcenter{\hbox{\includegraphics[width=0.2\linewidth]{figs/state_sum/coform}}}
\nonumber \\
&= {C^{ij}}_{mn,pq} {C^{kl,mn}}_{rs} [T_g]^{pq,rs} E_{ij} \otimes E_{kl}
= \frac{1}{N^2} \sum_{ijkl} [V^{\dag}_g]_{ij} [V_g]_{kl} E_{ij} \otimes E_{kl},
\quad (g \in G_0).
\end{align}
From this, we can read off the twisted ground state $\ell_g$ as
\begin{align}
\ell_g = \frac{1}{N} \sum_{ij} [V_g]_{ij} \ket{i} \bra{j}.
\end{align}
This is the same as the fixed point MPS ground state with twist up to a normalization.
\paragraph{Cylinder $\alpha_g$}
Since the Hilbert space with twist ${\cal C}_h (h \in G_0)$
is defined on the circle with $h$-flux,
we have to associate cylinder with twist by $T_h$.
For orientation preserving action $g \in G_0$, $\alpha_g$ reads
\begin{align}
\alpha_{g \in G_0} |_{{\cal C}_h}
&= \vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/tft/Cylinder}}}
= \vcenter{\hbox{\includegraphics[width=0.1\linewidth]{figs/state_sum/cylinder}}}
\nonumber \\
&= {C_{ij,mn}}^{pq} C_{pq,rs,tu} [T_h]^{tu,mn} [T_g]^{kl,rs} E^{ij} \otimes E_{kl}
= \frac{1}{N} [V^*_h]_{ij} E^{ij} \otimes [V_g V_h V_g^{\dag}]_{kl} E_{kl} ,
\quad (g \in G_0).
\end{align}
Here, $E^{ij}$ is dual basis of $E_{ij}$.
For an orientation reversing action $g \notin G_0$, $\alpha_g$ reads
\begin{align}
\alpha_{g \notin G_0} |_{{\cal C}_h}
&= \vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/tft/cylinder_P}}}
= \vcenter{\hbox{\includegraphics[width=0.1\linewidth]{figs/state_sum/cylinder_P}}}
\nonumber \\
&= {C_{ij,mn}}^{pq} C_{pq,rs,tu} [T_h]^{tu,mn} [T_P]^{kl,rs} E^{ij} \otimes E_{kl}
= \frac{1}{N} [V^*_h]_{ij} E^{ij} \otimes [V_P V^T_h V_P^{\dag}]_{kl} E_{kl} ,
\quad (g \in G_0).
\end{align}
\paragraph{Fusion}
The sphere with three punctures (the pants diagram)
describes a
fusion process ${\cal C}_g \otimes {\cal C}_h \to {\cal C}_{gh}$ ($g,h \in G_0$).
The path integral can be evaluated as
\begin{align}
&\vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/tft/pants}}}
= \vcenter{\hbox{\includegraphics[width=0.25\linewidth]{figs/state_sum/pants}}}
\nonumber \\
&=
{C_{kl,ab}}^{cd} {C_{ef,cd}}^{gh} [T_h]^{ef,ab}
{C_{pq,ij}}^{rs} {C_{rs,tu}}^{vw} [T_g]^{pq,tu}
{C_{vw,gh}}^{mn}
E^{ij} \otimes E^{kl} \otimes E_{mn}
\nonumber \\
&= \frac{1}{N^2} [V^*_g]_{ij} E^{ij} \otimes [V^*_h]_{kl} E^{kl} \otimes [V_g V_h]_{mn} E_{mn}.
\end{align}
\section{Conclusion}
Tensor networks methods have have been employed as an efficient way to represent
correlated, entangled, many-body ground states.
In particular, they are expected to provide a powerful framework to study
gapped quantum many-body systems with (symmetry-protected) topological order.
On the other hand,
topological quantum field theories
have been playing an important guiding
role in topological phases of matter.
Indeed, one definition of a topological phase of matter is simply that it is
described by a TFT.
These two descriptions (methods) are complementary to each other:
The tensor network methods in general
can provide a powerful practical (numerical) framework
to study a given lattice model.
Within the tensor network framework, it is important to
develop a methodology to diagnose topological properties of
a given (ground state) many-body wave function.
E.g., to develop a method to extract topological invariants from
a given many-body wave function (in the tensor network representation).
On the other hand,
TFTs allow us to work directly in the topological limit (the limit of zero-correlation
length), and hence provide a systematic and abstract (axiomatic) framework
to, e.g., systematically classify possible topological phases of matter.
In this paper, focusing on $(1+1)$d bosonic SPT phases,
we bridge MPTs and TFTs in $(1+1)$d.
In particular,
we discuss $(1+1)$d $G$-equivariant (possibly unoriented) TFTs,
which are TFTs coupled with a background gauge field.
Our results are briefly summarized as follows:
-- In Sec.\ \ref{Classification and topological invariants of MPSs},
we summarized the construction of SPT invariants in terms of
MPS networks.~\cite{PollmannTurner2012}
By expressing those by a reduced density matrix,
MPS networks representing SPT
topological invariants can be identified
with path integrals on manifold with a twist.
We showed the partial inversion and
the adjacent partial transpose leads to
partition function on the real projective plane.
-- In Sec.\ \ref{$G$-equivariant Topological field theory}
we reviewed
$(1+1)$d $G$-equivariant open and closed TFTs by Moore-Segal~\cite{Moore-Segal}
which,
in addition to closed chains in $(1+1)$d closed TFTs,
have $(1+0)$d open chains as an object.
We established a
fixed point MPS representation of
$(1+1)$d $G$-equivariant open and closed TFTs
(when TFTs are invertible).
A concrete connection between the MPS and TFT descriptions
is summarized in Table \ref{tab:functor_ori_tft} and \ref{Tab:Functor_Open}.
In particular,
we noted, for example,
that the classification of the $G$-equivariant closed unoriented simple TFTs
is given by the second group cohomology,
which precisely is the known classification of $(1+1)$d SPT phases (without
orientation-reversing symmetry).
We also noted that semisimple TFTs correspond to a combination of
symmetry breaking and symmetry fractionalization discussed in the MPS context.
Furthermore, for $G$-equivariant open TFTs,
the category of boundary conditions is equivalent to
the known boundary degrees of freedom in $(1+1)$d SPT phases.
-- In Sec.\ \ref{Fukuma-Hosono-Kawai state sum construction},
we presented a state sum construction for
$G$-equivariant unoriented closed TFT for $(1+1)$d bosonic SPT phases.
The symmetry twisted metrics play roles of nontrivial holonomy.
Partition functions and correlation functions can be
calculated in a unified framework.
In particular, we showed that the partition functions
on the torus, the real projective plane, and the Klein bottle
match precisely with the SPT invariants constructed from the MPS method.
There is a number of natural extensions of the current paper:
For example, it is natural to speculate that we can make a precise dictionary between
higher-dimensional TFTs
and higher-dimensional tensor networks, such as
projected entangled pair states (PEPS).
Another interesting direction is
to consider topological phases of fermions, and
their descriptions in terms of (fermionic) tensor networks, and
spin TFTs.
(For recent works addressing these issues,
see Refs.\ \cite{barrett2015two, novak2015state}
(the state sum construction of (1+1)d oriented spin TFTs),
Refs.\ \cite{gaiotto2015spin, bhardwaj2016state}
($(1+1)$d and $(2+1)d$ oriented equivariant spin TFTs and the state sum construction),
and
Ref.\ \cite{Shapourian-Shiozaki-Ryu}.
\paragraph{Note added.}
After completing this work, we became aware of an independent work \cite{Kapustin-Turzillo-You},
which established the connection between the state sum construction of $(1+1)$d $G$-equivariant TQFT and MPS representations.
\acknowledgments
We thank
Gil Young Cho,
Kiyonori Gomi,
Andreas W.\ W.\ Ludwig,
Kantaro Ohmori,
Hassan Shapourian,
Tadashi Takayanagi,
Apoorv Tiwari,
Keisuke Totsuka,
Alex Turzillo,
Juven C. Wang,
Xueda Wen,
and
Peng Ye
for useful discussion.
Especially, K.S.\ is grateful to Takahiro Morimoto for pointing out
the equivalence between
a partial transposition and crosscap.
This work was supported in part
by the National Science Foundation grant DMR-1455296,
and by Alfred P. Sloan foundation.
K.S.\ is supported by JSPS Postdoctoral Fellowship for Research Abroad.
\section{Introduction}
Symmetry-protected topological (SPT) phases of matter
are gapped phases of quantum many-body systems with short-range entanglement.
They are topologically distinct from topologically trivial states in the presence of symmetries.
In other words, SPT phases are separated from topologically trivial phases by
quantum critical points.
Here, ``short-range entanglement'' means in particular the absence of topological order,
and hence the uniqueness of the ground state even when
the system is put on an arbitrary spatial manifold.
(This property is often called ``invertible'', and
hence SPT phases are said to have an invertible topological order.)
Bosonic SPT phases in (1+1) dimensions are
known to be classified by the second group cohomology $H^2(G,U(1))$.
\cite{ChenGuLiuWen2013}
For quantum many-body systems defined on one-dimensional lattices,
this can be most easily seen from the matrix product state (MPS) representations of
the quantum ground states of SPT phases.
\cite{Chen2011, Schuch2011, PollmannBergTurner2012,PollmannTurner2012}
On the other hand,
deep inside a gapped phase where
the correlation length is very short (order of a few lattice constant),
one could expect that the universal properties of the system
can be described in terms of a topological quantum field theory (TQFT or TFT).
The canonical examples include
Chern-Simons theories,
which describe various fractional quantum Hall liquids,
and
the BF theory,
which describes the topological limit of the $\mathbb{Z}_n$ lattice gauge theory.
\cite{wen2004quantum, fradkin2013field}
In this paper, we will undertake the task of bridging
the descriptions of (1+1)d bosonic SPT phases using MPSs,
and those using (1+1)d TFTs.
For (1+1)d bosonic SPT phases protected by symmetry $G$,
where $G$ is a symmetry group,
the relevant TFTs are $G$-equivariant TFTs discussed by
Turaev and Moore-Segal.
\cite{Turaev, turaev2010homotopy, Moore_lecture, Moore-Segal, Kapustin-Turzillo}
We in particular address the following two issues.
The first issue is about topological invariants (SPT invariants) of (1+1)d bosonic SPT phases.
These are quantities (numbers) which one can compute for a given
quantum ground state of a gapped (1+1)d system,
and take the same value anywhere in a given gapped phase.
I.e., they are stable and remain unchanged against adiabatic deformations
of Hamiltonians so far as one stays within a given gapped phase.
On the one hand, several topological invariants for bosonic
SPT phases have been constructed so far by using MPSs.
\cite{PollmannTurner2012}
In this paper, we will rederive these invariants by using
$G$-equivariant TFTs in (1+1)d.
The topological invariants are nothing but the partition functions of TFTs.
Second,
the hallmark of (1+1)d SPT phases
is the presence of boundary degrees of freedom that appear
when the SPT phases are terminated by boundaries.
The canonical example is the spin 1/2 that appears at the end of
the spin 1 Haldane spin chain.
In terms of MPSs,
these physical boundary degrees of freedom are captured by
degrees of freedom living in the auxiliary bond (entanglement) Hilbert space.
On the TFT side,
a natural framework to discuss the boundary degrees of freedom
is an ``open'' TFT.
\cite{lazaroiu2001structure, Moore-Segal}
Open TFTs are TFTs defined on the (1+1)d spacetime which has (1+0)d boundaries.
In this paper,
we will make an attempt to
make a dictionary between
MPSs with boundaries and open TFTs.
The rest of the paper is organized as follows:
In Sec.\ \ref{Classification and topological invariants of MPSs},
we introduce the descriptions of bosnic SPT phases in (1+1)d using MPSs.
In particular, we review the known construction of
various topological invariants for (1+1)d SPT phases built out of MPSs.
By using the fixed point MPSs,
we confirm that these topological invariants
characterizes the elements of the group cohomology $H^2(G,U(1))$,
and describe the procedure to extract these from a given quantum ground state.
(All results in Sec.\ \ref{Classification and topological invariants of MPSs} are
known in the literature,
so readers who are familiar with the MPS descriptions of (1+1)d bosonic SPT phases
and their topological invariants can skip this section.)
In Sec.\ \ref{$G$-equivariant Topological field theory},
we introduce $G$-equivariant TFTs following Moore and Segal.
\cite{Moore-Segal}
We discuss both closed and open TFTs;
In closed TFTs we consider the (1+1)d spacetime which has no
boundary,
whereas in open TFTs the (1+1)d spacetime has (1+0)d boundaries.
In Sec.\ \ref{Fukuma-Hosono-Kawai state sum construction},
we will derive
the topological invariants from the point of view of (1+1)d TFTs.
To this end, we evaluate the partition functions of
(1+1)d TFTs by using the so-called state sum construction~\cite{FHK}.
Introducing an orientation reversing operation on Frobenius
algebras enables us to define partition function on the real
projective plane $\mathbb{R} P^2$~\cite{Karimipour-Mostafazadeh}.
Finally, Appendices are devoted to an introduction to
the group cohomology,
and
projective representations,
the relation to orbifolded theories ($(1+1)$d Dijkgraaf-Witten theories),
and
the derivations of algebraic relations in $G$-equivariant open and closed TFTs.
\section{Classification and topological invariants of SPT phases using MPSs}
\label{Classification and topological invariants of MPSs}
In this section,
we briefly review the topological classification of bosonic SPT phases
in (1+1)d~\cite{PollmannBergTurner2012, ChenGuLiuWen2013, Schuch2011},
and their topological invariants~\cite{PollmannTurner2012}.
From the field theoretical point of view,
bosonic SPT phases are described by $G$-equivariant TFTs~\cite{Moore-Segal},
which will be introduced in the next section.
\subsection{Symmetry and group cohomology classification}
Let us consider a short range entangled pure state $\ket{\Psi}$ on a closed chain of length $L$,
which is represented by a MPS
\begin{align}
\ket{\Psi}
&= \sum_{\{m_i \}} {\rm Tr}\, (A_{m_1} \cdots A_{m_L} ) \ket{m_1\cdots m_L},
\quad
\ket{m_1\cdots m_L} = \ket{m_1}_1 \otimes \cdots \otimes \ket{m_L}_L,
\end{align}
where
$\ket{m}_j$ represents a state in the {\it physical} Hilbert space at the $j$-th site and $A_{m_j}$ is a $\chi \times \chi$ matrix which acts on the
{\it auxiliary} Hilbert space or
the {\it entanglement} Hilbert space living on the bonds;
The trace is taken over the $\chi\times \chi$ dimensional auxiliary Hilbert space.
Here and henceforth, we assume the translational symmetry for simplicity.
Let $G$ be a symmetry group.
The symmetry group $G$ possibly includes orientation reversing symmetries
(time-reversal or inversion symmetry, say).
For the purpose of specifying the orientation reversing symmetries,
let us introduce a homomorphism $\phi: G \to G_0$,
where $G_0$ is a group consisting of the orientation preserving symmetries.
The symmetry action $g \in G$ is defined on the basis $\ket{m}_j$
by a linear representation of $G$,
\begin{align}
&\hat g (\ket{m}_j) = \ket{n}_j [U_g]_{nm}, && U_g U_h = U_{gh},
&& (\mbox{$g$ is on-site unitary symmetry}), \\
&\hat T (\ket{m}_j) = \ket{n}_{j} [U_T]_{nm}, && U_T U_g^* = U_{Tg},
&& (\mbox{$T$ is time-reversal symmetry}), \\
&\hat P (\ket{m}_j) = \ket{n}_{L-j} [U_P]_{nm}, && U_P U_g = U_{Pg},
&& (\mbox{$P$ is inversion symmetry}),
\end{align}
for any $h \in G$.
By choosing different $A$, one can construct the ground states of gapped phases in (1+1) dimensions.
One can consider to classify these gapped phase topologically,
in the presence of a prescribed symmetry $G$.
The topological classification of (1+1)d SPT phases of bosons is given by the classification of the symmetry action on $A_{m}$.
Under the assumption that $\ket{\Psi}$ is a pure state, one can show~\cite{Cirac2008}
\begin{align}
&[U_g]_{mn} A_n = e^{i \theta_g} V_g^{\dag} A_m V_g,
&& (\mbox{$g$ is on-site unitary symmetry}), \\
&[U_T]_{mn} A^*_n = e^{i \theta_T} V_T^{\dag} A_m V_T,
&& (\mbox{$T$ is time-reversal symmetry}), \label{Eq:Sym_A_TRS} \\
&[U_P]_{mn} A_n^T = e^{i \theta_P} V_P^{\dag} A_m V_P,
&& (\mbox{$P$ is inversion symmetry}),
\end{align}
where $e^{i \theta_g}, e^{i \theta_T}$, and $e^{i \theta_P}$ are 1-dimensional linear representations of $G$,
and $V_g, V_T$, and $V_P$ act on the entanglement Hilbert space,
and obey, for any $h\in G$,
\begin{align}
\left\{
\begin{array}{ll}
V_g V_h = b(g,h) V_{gh} & (\mbox{$g$ is on-site unitary symmetry})
\\
\\
V_g V_h^* = b(g,h) V_{gh} & (\mbox{$g$ is time-reversal or inversion symmetry}) \\
\end{array} \right.
\label{Eq:ProjRep_V}
\end{align}
with a $U(1)$ phase $b(g,h) \in U(1)$ (2-cocycle).
These symmetry actions on $A_m$ are diagrammatically represented
in Fig.\ \ref{Fig:Symmetry_MPS}.
(In the figure, we neglect the 1-dimensional representation $e^{i \theta_g}$.)
From the associativity condition of $V_g$,
it follows that $b(g,h)$ is a representative of
$\phi$-twisted second group cohomology $H^2(G,U(1)_{\phi})$.
(Here, ``$\phi$-twisted'' means
the $g \notin G_0$ action on $U(1)$ group is defined by complex conjugate.
See Appendix \ref{Group cohomology}.)
The factor group $[b(g,h)] \in H^2(G,U(1)_{\phi})$ classifies how symmetry $G$ acts on the
short-range entangled pure state $\ket{\Psi}$ on the 1-dimensional closed chain.
\begin{figure}[!]
\begin{center}
\includegraphics[width=0.6\linewidth, trim=0cm 0cm 0cm 0cm]{figs/mps/Symmetry_MPS.pdf}
\end{center}
\caption{The symmetry property of $A_m$
under [a] on-site unitary, [b] time-reversal, and [c] inversion symmetries, respectively.}
\label{Fig:Symmetry_MPS}
\end{figure}
\subsection{Edge degrees of freedom}
In a non-trivial SPT phase
(specified by a nontrivial group cohomology
$[b(g,h)] \in H^2(G, U(1)_{\phi})$, $[b(g,h)] \neq 0$)
on an open chain $I = \{1, \dots, L\}$,
there emerge edge degrees of freedom where
the symmetry action of $G$ is ``fractionalized''.
The boundary degrees of freedom can be discussed
by using the MPS of the open chain:
\begin{align}
\Psi(\bra{v_L} \otimes \ket{v_R}) = \sum_{\{m_i\}}
v_L^{\dag} A_{m_1} \cdots A_{m_L} v_R \ket{m_1 \cdots m_L},
\qquad
\bra{v_L} \in V^*,
\quad
\ket{v_R} \in V,
\end{align}
where $\bra{v_L}$ and $\ket{v_R}$ specify boundary conditions
and belong to the edge Hilbert space $V^*$ and $V$, respectively.
$V$ is a $b(g,h)$-projective representation defined by (\ref{Eq:ProjRep_V})
and $V^*$ is the its conjugate representation.
Here, the symmetry fractionalization is realized in the following sense:
A symmetry action $g \in G_0$ on the MPS on the open chain is given by
\begin{align}
\hat g \Psi(\bra{v_L} \otimes \ket{v_R}) = \sum_{\{m_i\}} v_L^{\dag} V_g^{\dag} A_{m_1} \cdots A_{m_L} V_g v_R \ket{m_1 \cdots m_L}
= \Psi(\bra{v_L} V_g^{\dag} \otimes V_g \ket{v_R}),
\end{align}
where a symmetry operation $g \in G_0$ is projectively represented at the edges,
as opposed to the $g$ action on the bulk physical degrees of freedom, which is
a linear representation.
For an on-site unitary symmetry $g \in G_0$, we can introduce
a $g$-twisted MPS
\begin{align}
\ket{\Psi_g} := \sum_{\{m_i\}}
{\rm Tr}\, (A_{m_1} \cdots A_{m_L} V_g) \ket{m_1\cdots m_L}, \quad (g \in G_0).
\end{align}
From the perspective of Hamiltonians,
$\ket{\Psi_g}$ is a ground state of a Hamiltonian with a $g \in G_0$ symmetry defect.
For our purpose to make a comparison between MPSs and TFTs,
it is useful to introduce an open to closed map $\imath^{g}$
and a closed to open map $\imath_{g}$~\cite{Moore-Segal}
as
\begin{align}
&\imath^{g} \big( \Psi(\bra{v_L} \otimes \ket{v_R}) \big) := \braket{v_L| V_g^{\dag} |v_R} \ket{\Psi_g},
\quad (g \in G_0), \\
&\imath_{g}(\ket{\Psi_g}) := \sum_a \Psi(\bra{a} \otimes V_g \ket{a}),
\quad (g \in G_0).
\end{align}
Here $\{ \ket{a} \}_{a=1}^{{\rm dim} V}$ is a basis of $V$.
Finally, we also introduce a formal ``gluing'' operation of two open MPSs by
\begin{align}
\Psi(\bra{v_L} \otimes \ket{v_R}) \cdot \Psi(\bra{w_L} \otimes \ket{w_R})
:= \braket{w_L | v_R} \Psi(\bra{v_L} \otimes \ket{w_R}).
\end{align}
This will be also useful when we make a comparison between MPSs and TFTs.
\subsection{Simple and fixed point MPSs}
\label{Fixed point MPSs}
To study ground states deep inside a gapped phase,
or to study SPT phases in general,
it is useful and convenient to introduce a simple and fixed point MPS.
``Simple'' here means that the transfer matrix
$T_{ab,cd} = \sum_{m} [A_m]_{ab} [A^*_m]_{cd}$
has an only one eigenstate with unit magnitude of eigenvalue $|\nu|=1$,
i.e., unique ground state.~\cite{fidkowski2011topological}
``Fixed point'' means that we are in the limit of zero correlation length.
It is in this limit where we expect
SPT phases and the corresponding MPSs
are faithfully described by TQFTs.
In the following, we will construct fixed point MPSs
$\ket{\Psi}$
with a nontrivial group cohomology $H^2(G,U(1)_{\phi})$.
Let $G$ be a symmetry group with nontrivial group cohomology $H^2(G,U(1)_{\phi}) \neq 0$.
We fix a nontrivial 2-cocycle $b(g,h) \in Z^2(G,U(1)_{\phi}), [b(g,h)] \neq 0$.
We choose two $b(g,h)$-projective representations
$V$ which satisfy (\ref{Eq:ProjRep_V}).
We use the tensor product representation
$V^* \otimes V$ as a
physical Hilbert space, where
$V^*$ is the complex representation of
$V$.
Note that in the product representation,
the effect of the 2-cocycle cancels, $b^*(g,h) b(g,h)=1$.
For each site $j$, the basis is given by $\{ \ket{a}^L_j \otimes \ket{b}^R_j \}$,
where $\ket{a}^L_j$ ($\ket{b}^R_j$) is the basis of $V^*$ ($V$)
and transformed as
\begin{align}
&\left\{\begin{array}{l}
\hat g (\ket{a}^L_j \otimes \ket{b}^R_j) = \ket{c}^L_j \otimes \ket{d}^R_j {[V^*_g]}_{ca} {[V_g]}_{db}, \\
{[U_g]}_{ab,cd} = {[V^*_g]}_{ac} {[V_g]}_{bd}, \\
\end{array}\right. && \mbox{($g$ is on-site unitary symmetry)},\\
&\left\{\begin{array}{l}
\hat T (\ket{a}^L_j \otimes \ket{b}^R_j) = \ket{c}^L_j \otimes \ket{d}^R_j {[V^*_T]}_{ca} {[V_T]}_{db}, \ \ \hat T i \hat T^{-1} = -i , \\
{[U_T]}_{ab,cd} = {[V^*_T]}_{ac} {[V_T]}_{bd}, \\
\end{array}\right. && \mbox{($T$ is time-reversal symmetry)}, \\
&\left\{\begin{array}{l}
\hat P (\ket{a}^L_j \otimes \ket{b}^R_j) = \ket{c}^L_{L-j} \otimes \ket{d}^R_{L-j} {[V^*_P]}_{cb} {[V_P]}_{da}, \\
{[U_P]}_{ab,cd} = {[V^*_P]}_{ad} {[V_P]}_{bc}, \\
\end{array}\right. && \mbox{($P$ is inversion symmetry)}.
\end{align}
Note that $V$ and $V^*$ representations are exchanged under the inversion transformation.
\begin{figure}[!]
\begin{center}
\includegraphics[width=\linewidth, trim=0cm 0cm 0cm 0cm]{figs/mps/MPS_Fixed_Point.pdf}
\end{center}
\caption{[a] A fixed point MPS. The blue bond represents the singlet representation in $V^* \otimes V$. [b] $A$ matrix. [c] The transfer matrix.}
\label{Fig:MPS_Fixed_Point}
\end{figure}
To write down ground state wave functions,
we make use of a singlet included in the decomposition of the product representations,
$
V^* \otimes V = 1 \oplus \cdots.
$
The fixed point MPS can be constructed as the product state of singlet bonds
(Fig.\ \ref{Fig:MPS_Fixed_Point} [a]) as
\begin{align}
\ket{\Psi}
= \cdots \otimes (\sum_{a} \ket{a}^R_j \otimes \ket{a}^L_{j+1}) \otimes (\sum_{b} \ket{b}^R_{j+1} \otimes \ket{b}^L_{j+2}) \otimes \cdots.
\label{eq:fixed_point_MPS_def}
\end{align}
This can be written in the MPS form as
\begin{align}
\ket{\Psi} = \sum \cdots [A_{b_j c_j}]_{a_j a_{j+1}} [A_{b_{j+1} c_{j+1}}]_{a_{j+1} a_{j+2}} \cdots
\ket{ \cdots (b_j c_j) (b_{j+1} c_{j+1}) \cdots}
\end{align}
with
\begin{align}
[A_{ab}]_{cd} = \frac{1}{\sqrt{{\rm dim} V}} \delta_{ac} \delta_{bd}.
\end{align}
(See Fig.\ \ref{Fig:MPS_Fixed_Point} [b]. )
Here $(ab)$ is the physical index whereas $(cd)$ is the entanglement index.
We abbreviated $\ket{b_j}^L_j \otimes \ket{c_j}^R_j$ by $\ket{(b_jc_j)}$.
The prefactor $\frac{1}{\sqrt{{\rm dim} V}}$ is the normalization constant.
The MPS $\ket{\Psi}$ is the AKLT state \cite{affleck1988valence}
without any projection on the site degrees of freedom.~\cite{chen2012symmetry}
For the fixed point MPSs, the transfer matrix $T_{ab, a'b'}$ (Fig.\ \ref{Fig:MPS_Fixed_Point} [c]) is given by
\begin{align}
T_{ab,cd}
= \sum_{ef} [A_{ef}]_{ab} [A^*_{ef}]_{cd}
= \frac{1}{{\rm dim} V} \delta_{ac} \delta_{bd}.
\end{align}
In any symmetries, we have the following group action on $[A_{ab}]_{cd}$ for the fixed point MPS,
\begin{align}
&[U_g]_{ab,ef} [A_{ef}]_{cd}
= \frac{1}{\sqrt{{\rm dim} V}} [V_g^*]_{ac} [V_g]_{bd},
\quad (\mbox{$g$ is on-site unitary symmetry}), \\
&[U_T]_{ab,ef} [A^*_{ef}]_{cd}
= \frac{1}{\sqrt{{\rm dim} V}} [V_T^*]_{ac} [V_T]_{bd},
\quad (\mbox{$T$ is time-reversal symmetry}), \\
&[U_P]_{ab,ef} [A^T_{ef}]_{cd}
= \frac{1}{\sqrt{{\rm dim} V}} [V_P^*]_{ac} [V_P]_{bd},
\quad (\mbox{$P$ is inversion symmetry}).
\end{align}
Finally, it is worth pointing out that in fixed point MPSs
the length of the MPS chain is irrelevant because of the zero correlation length.
I.e., since they are at a renormalization group fixed point,
increasing/decreasing the number of cites does not change the essential properties of
the state.
For this reason, we always identify MPS chains with different lengths as
\begin{align}
{\rm Tr} ( A_{m_1} \cdots A_{m_L} ) \ket{m_1 \cdots m_L} \sim {\rm Tr} ( A_{m_1} \cdots A_{m_L} A_{m_{L+1}}) \ket{m_1 \cdots m_L m_{L+1}}.
\label{eq:mps_equivalence}
\end{align}
\subsection{Topological invariants}
\label{Topological invariants}
In this section, we will construct and discuss topological invariants
of bosonic SPT phases in (1+1)d using the MPS.
We start by listing topological invariants.
Detailed descriptions of topological invariants will follow shortly.
There are three types of topological invariants,
which are defined in terms of the data of 2-cocycle $\{b(g,h)\}$:
\begin{itemize}
\item The discrete torsion phase (partition function on $T^2$ with twist)
\begin{align}
\label{eq:T^2inv}
\epsilon(g,h) = \frac{b(g,h)}{b(h,g)},
\quad
V_g V_h = \epsilon(g,h) V_h V_g,
\quad
g,h\in G_0,
\quad
gh=hg.
\end{align}
\item The crosscap invariant (partition function on $\mathbb{R} P^2$)
\begin{align}
\label{eq:Corsscapinv}
\theta(g) := b(g,g),
\quad V_g V_g^* = \theta(g),
\quad g \notin G_0,
\quad g^2 = 1.
\end{align}
\item The Klein bottle invariant (partition function on the Klein bottle with twist)
\begin{align}
\label{eq:KBinv}
\kappa(g;h) = \frac{b(g,h^{-1}) b(h,h^{-1})}{b(h,g)},
\quad
V_g V_h^T = \kappa(g;h) V_h V_g,
\quad
g \notin G_0, h \in G_0,
\quad g h^{-1} = h g.
\end{align}
\end{itemize}
Several comments are in order.
--
First, in \eqref{eq:T^2inv}, \eqref{eq:Corsscapinv} and \eqref{eq:KBinv},
and throughout this subsection,
we omit the 1-dimensional representation
$\{ e^{i \theta_g}\}$ for simplicity.
--
One can check easily that
these quantities are left unchanged under the 1-coboundary $b(g,h) \mapsto b(g,h) a_g a_g^{\phi(g)} a_{gh}^{-1}$, $a_g \in U(1)$.
--
One can give interpretations to
these topological invariants
in terms of spacetime path integrals.
We will mention these interpretations
later in this subsection,
and also in Sec.\
\ref{$G$-equivariant Topological field theory}
-
\ref{Fukuma-Hosono-Kawai state sum construction}
from the TFT point of view.
In short,
these three topological invariants are interpreted as
the partition function on the torus,
the projective plane,
and
the Klein bottle respectively.
For this reason, we will often refer the topological invariants
as the partition functions
(on the torus, the projective plane, and the Klein bottle).
--
Finally, the above three SPT invariants are not independent.
One can show
\begin{align}
&\epsilon(g,1) = \epsilon(1,g) = 1 && {\rm for\ \ } g \in G_0, \\
&\epsilon(h,g) = \epsilon(g,h)^{-1} && {\rm for\ \ } g,h \in G_0, gh=hg, \\
&\epsilon(g,hk) = \epsilon(g,h)\epsilon(g,k) && {\rm for\ \ } g,h,k \in G_0, gh=hg, gk=kg, \\
&\kappa(g;1) = 1 && {\rm for\ \ } g \notin G_0, \\
&\kappa(g;hk) = \kappa(g;h) \kappa(g;k) && {\rm for\ \ } g \notin G_0, h,k \in G_0, g h^{-1} = h g, g k^{-1}=k g, \\
&\kappa(kg;h) = \kappa(gk^{-1};h) = \epsilon(k,h) \kappa(g;h) && {\rm for\ \ } g \notin G_0, h,k \in G_0, g h^{-1} = h g, h k=k h, \\
&\theta(h g) = \theta(g h^{-1}) = \kappa(g;h) \theta(g) && {\rm for\ \ } g \notin G_0, h \in G_0, g^2=1, g h^{-1} = h g.
\end{align}
In many cases, the Klein bottle SPT invariant $\kappa(g;h)$
can be written in terms of $\epsilon(g,h)$ and $\theta(g)$.
This is however not always the case.
A simple example in which the Klein bottle invariant does not reduce to the other invariants
is an SPT phase protected by
$G = \mathbb{Z}_4 = \{1,\sigma,\sigma^2, \sigma^3\}$.
Here, the generator $\sigma$ is inversion/time-reversal.
For a nontrivial projective representation generated by
$V_{\sigma} = e^{-i s_y \frac{\pi}{4}}$,
where
$s_y = \begin{pmatrix}
0 & -i \\
i & 0 \\
\end{pmatrix}$
is the $y$-component of the Pauli matrix,
from $V_{\sigma} V_{\sigma^2}^T = - V_{\sigma^2} V_{\sigma}$,
the Klein bottle SPT invariant reads $\kappa(\sigma;\sigma^2) = -1$.
\subsubsection{Topological invariants in terms of ground state
wave functions}
As mentioned,
the topological invariants can be interpreted
by using the path integral formalism.
In the rest of this subsection,
we will instead use the operator formalism,
and in particular
aim to extract the topological invariant
solely by using ground state wave functions.
(Apart from our goal of bridging MPSs and TFTs,
expressing topological invariants solely in terms
of ground state wave functions may have practical (numerical)
merits. \textcolor{blue}{)}
In order to discuss and define these topological invariants,
one important ingredient
is {\it gauging} symmetry.
Here, by gauging, we mean coupling the system to the background flat $G$-bundle.
We will describe how this can be done within MPSs
for on-site unitary symmetry
in Sec.\ \ref{Discrete torsion phase: symmetry action on twisted ground state}.
The same gauging procedure can be introduced by using
the path-integral.
(In addition, one could promote the background gauge field into a dynamical one.
This procedure is often called {\it orbifolding}
to distinguish it from gauging.
In this paper, for the purpose of describing SPT phses,
we will consider gauging but not orbifolding.
Orbifolding leads to the so-called Dijkgraaf-Wittten
theories, which we briefly discuss in Appendix \ref{Orbifolding: Dijkgraaf-Witten theory in (1+1)d}.)
As for the crosscap invariant,
we need to introduce a ``trick''
within the operator formalism
in order to mimic the effect of
putting the theory on $\mathbb{R}P^2$
in the path integral formalism.
This can be done in two different ways,
depending on whether the symmetry
group includes spatial inversion or time-reversal.
Following Pollmann and Turner
\cite{PollmannTurner2012}
we will introduce two operations,
``partial inversion'' and ``adjacent partial transposition'',
for spatial inversion and time-reversal, respectively.
When interpreted in the path integral formalism,
these operations effectively create
$\mathbb{R} P^2$ as the spacetime manifold.
Introducing such a {\it partial space-time twist operator}
is a useful way to detect SPT topological invariant
which cannot be represented by the partition function
on a mapping torus.~\footnote{
A mapping torus is space-time manifold which takes
the form
$M \times_f S^1 := M \times [0,1] / \big\{ (x,0) \sim (f(x), 1) \big\}$,
where $f : M \to M$ is a diffeomorphism.
In (invertible) TFTs, the partition function $Z(M \times_f S^1)$
on $M \times_f S^1$ is given by the expectation value of
operator $\hat f$ representing diffeomorphism $f$ on
the ground state wave function $\ket{\Psi_M}$ on $M$
as $Z(M \times_f S^1) = \braket{\Psi_M | \hat f | \Psi_M}$.
}
This prescription can also be applied to
fermionic SPT invariants
\cite{Shapourian-Shiozaki-Ryu}
and SPT phases in more general space dimensions.
\cite{Shiozaki-Shapourian-Ryu}
For all topological invariants, the fact that they can be extracted from
ground state wave functions
can be easily proven if we use the fixed point MPS.
\paragraph{Discrete torsion phase (1): symmetry action on twisted ground state}
\label{Discrete torsion phase: symmetry action on twisted ground state}
Let us first express the torus topological invariant
\eqref{eq:T^2inv} by using ground state wave functions.
\begin{figure}[t]
\begin{center}
\includegraphics[width=\linewidth, trim=0cm 0cm 0cm 0cm]{torus_discrete_torsion.pdf}
\end{center}
\caption{
[a] MPS expression of the symmetry action on the twisted ground state.
[b] The equivalent path integral on the 2-torus $T^2$.
The blue and red lines express the symmetry defect lines.}
\label{Fig:torus_discrete_torsion}
\end{figure}
From the MPS and the projective representation $\{V_g\}_{g \in G}$,
we can construct the MPS $\ket{\Psi_h}$ with
boundary condition twisted by an on-site unitary symmetry $h \in G_0$ as
\begin{align}
\ket{\Psi_h}
&= \sum_{\{m_i\}} {\rm Tr}\, (A_{m_1} \cdots A_{m_L} V_h) \ket{m_1\cdots m_L}.
\end{align}
Then, for
a global unitary symmetry,
$g \in G_0$,
\begin{align}
\hat g \ket{\Psi_h}
= \sum_{\{m_i\}}
{\rm Tr}\, (A_{m_1} \cdots A_{m_L} V_g V_{h} V_g^{-1}) \ket{m_1\cdots m_L}
= \frac{b(g,h)}{b(g h g^{-1}, g)} \ket{\Psi_{ghg^{-1}}}.
\label{Eq:G-action_MPS}
\end{align}
The MPS diagram is shown in Fig.~\ref{Fig:torus_discrete_torsion}[a].
If $g h g^{-1} = h$, the $U(1)$ factor is well-defined which is invariant under the 1-coboundary.
From the TFT point of view,
this invariant is nothing but the partition function $Z_{T^2}(g,h)$ on the torus $T^2$ with the background $g$ and $h$ twist
\begin{align}
\epsilon(g,h) = Z_{T^2}(g,h) = \braket{\Psi_h | \hat g | \Psi_h},
\quad (gh=hg).
\end{align}
In the spacetime path integral,
it would be useful to introduce symmetry defect lines to express the background $G$ field.
The matter field is transformed by $U_g$ when it passes through the symmetry defect line of $g$.
Fig.~\ref{Fig:torus_discrete_torsion}[b] shows the symmetry defect lines corresponding to the partition function $Z_{T^2}(g,h)$ twisted by $g$ and $h$.
The discrete torsion phase $\epsilon(g,h)$ arises from the intersection of two symmetry defect lines of $g$ and $h$ with $[g,h]=0$.
\paragraph{Discrete torsion phase (2): partial symmetry action and swapping}
\label{Discrete torsion phase: partial symmetry action and swapping}
There is an alternative way to detect the discrete torsion phase invariant.
It is given by the combination of the partial symmetry action and the swapping operator, described as follows.
Let $\ket{\Psi}$ be the ground state on $S^1$ with no flux.
(In the TFT path integral, this state is obtained/defined by the path-integral over the disc.)
We introduce three adjacent intervals $I_1 \cup I_2 \cup I_3$ with $I_1$ and $I_3$ having the same number of sites.
The discrete torsion phase $\epsilon(g,h)$ is then extracted as the complex $U(1)$ phase of the quantity~\cite{PhysRevLett.109.050402}
\begin{align}
Z = \Braket{\Psi | \prod_{j \in I_1 \cup I_2} (U_h)_j \cdot {\rm Swap}(I_1,I_3) \cdot \prod_{j \in I_1 \cup I_2} (U_g)_j | \Psi}, \qquad
gh=hg,
\label{eq:symmetry_action_and_swap}
\end{align}
in the limit $|I_1|,|I_2|, |I_3| \gg \xi$, where $\xi$ is the correlation length of the bulk.
Here, ${\rm Swap}(I_1,I_3)$ is the operator swapping the two intervals $I_1$ and $I_3$, which is defined by
\begin{align}
{\rm Swap}(I_1,I_3) \ket{m_j}
= \left\{\begin{array}{ll}
\ket{m_{j+|I_1|+|I_2|}} & (j \in I_1), \\
\ket{m_{j-|I_1|-|I_2|}} & (j \in I_3), \\
\ket{m_j} & ({\rm otherwise}).
\end{array}\right.
\end{align}
For the MPS representation of the ground state $\ket{\Psi}$, the MPS diagram of $Z$ is written as Fig.~\ref{Fig:symmetry_action_and_swap}.
For the fixed point MPS (\ref{eq:fixed_point_MPS_def}), it is easy to show that $Z = |Z| \epsilon(g,h)$.
The path-integral picture also verifies that $Z$ gives the discrete torsion phase.
See Fig.~\ref{Fig:symmetry_action_and_swap_path_integral}.
Topologically, the swapping operator ${\rm Swap}(I_1,I_3)$ with the intermediate region $I_2$ is equivalent to adding a genus.
The background $G$ field obtained by the partial symmetry actions $U_g$ and
$U_h$ on the adjacent intervals $I_1 \cup I_2$
has an intersection between two symmetry defect lines of $g$ and $h$, which leads to the discrete torsion phase $\epsilon(g,h)$.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.7\linewidth, trim=0cm 0cm 0cm 0cm]{symmetry_action_and_swap.pdf}
\end{center}
\caption{
MPS expression of the partial symmetry action with swapping defined in (\ref{eq:symmetry_action_and_swap}).
}
\label{Fig:symmetry_action_and_swap}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width=\linewidth, trim=0cm 0cm 0cm 0cm]{symmetry_action_and_swap_path_integral.pdf}
\end{center}
\caption{
The geometry of path integral for the partial symmetry action with swapping operator defined in (\ref{eq:symmetry_action_and_swap}).
The red and blue lines express the symmetry defect lines.
The intervals with arrows are identified with ones having the same number of arrows, which results in the 2-torus.
}
\label{Fig:symmetry_action_and_swap_path_integral}
\end{figure}
\paragraph{Crosscap from inversion symmetry: ``partial inversion''}
\begin{figure}[t]
\begin{center}
\includegraphics[width=\linewidth, trim=0cm 0cm 0cm 0cm]{figs/mps/Partial_Inversion.pdf}
\end{center}
\caption{[a] MPS representation of crosscap. [b] The geometry of path integral for the partial inversion.}
\label{Fig:Partial_Inversion}
\end{figure}
The crosscap topological invariant
can be defined when the symmetry group $G$
includes spatial inversion or time-reversal.
The procedures to extract the invariant
from ground state wave functions
are different for spatial inversion and time-reversal.
Let us first discuss the crosscap topological
invariant when $G$ includes
spatial inversion.
Let $\ket{\Psi}$ be a MPS on the closed chain $L$.
To create the real projective plane $\mathbb{R} P^2$, we take the {\it partial inversion} on the interval $I = \{1, \dots, N\}$,
as shown in Fig.\ \ref{Fig:Partial_Inversion} [a],
as
\begin{align}
\hat P_I \ket{\Psi}
= \sum_{\{m_i\}} {\rm Tr}\, (A_{m_1} \cdots A_{m_L}) \ket{\widetilde{m}_1 \cdots \widetilde{m}_N m_{N+1} \cdots m_L},
\end{align}
where $\widetilde{\ket{m}}_j$ is the partially inverted physical degrees of freedom,
\begin{align}
\widetilde{\ket{m}}_j = \ket{n}_{N-j} [U_P]_{nm}.
\end{align}
We assume the length of the interval $I$ is sufficiently larger than the correlation length.
The fact that this operation
creates $\mathbb{R}P^2$ as the spacetime manifold
can be easily understood from Fig.\ \ref{Fig:Partial_Inversion}-[b];
In the path integral representation, the partial inversion is equivalent to inserting
a one crosscap on the time slice at $\tau=0$.
One can show that for $P^2 = 1$ the amplitude $\hat P_I \ket{\Psi}$ gives the crosscap invariant~\cite{PollmannTurner2012}
\begin{align}
\theta(P) = Z_{\mathbb{R} P^2}(P) = \frac{\braket{\Psi| \hat P_I |\Psi}}{|\braket{\Psi| \hat P_I |\Psi}|} = b(P,P),
\quad P^2 = 1,
\end{align}
in the limit $L \to \infty$ and $N \to \infty$.
This formula can easily be proven
for fixed point MPSs introduced in the previous section.
Here we give a proof by using the cut and glue construction.~\cite{Qi2012}
To illustrate the proof, we use the Haldane chain protected by the inversion symmetry with $H^2(\mathbb{Z}_2;U(1)_{\phi}) = \mathbb{Z}_2$.
First, we cut the chain $L$ by the interval $I$.
There are four effective low energy degrees of freedom localized at the boundary of $I$,
\begin{align}
\ket{e^R_0} \otimes \ket{e^L_1} \otimes \ket{e^R_N} \otimes \ket{e^L_{N+1}},
\quad
e^L_j, e^R_j = \uparrow, \downarrow.
\end{align}
Next, we glue these degrees of freedom to get the original ground state by forming the singlet bond as
sites $0$-$1$ and $N$-$(N+1)$ as
\begin{align}
\ket{\Psi} = \frac{1}{\sqrt{2}} (\ket{\uparrow^R_0} \otimes \ket{\downarrow^L_1} - \ket{\downarrow^R_0} \otimes \ket{\uparrow^L_1} )
\otimes \frac{1}{\sqrt{2}} (\ket{\uparrow^R_N} \otimes \ket{\downarrow^L_{N+1}} - \ket{\downarrow^R_N} \otimes \ket{\uparrow^L_{N+1}} ).
\end{align}
The reduced density matrix $\rho_I$ is
\begin{align}
\rho_I
&= {\rm tr}_{0,{N+1}} (\ket{\Psi} \bra{\Psi})
= \frac{1}{4} \left[
(\ket{\uparrow^L_1}\bra{\uparrow^L_1} +\ket{\downarrow^L_1}\bra{\downarrow^L_1}) \otimes
(\ket{\uparrow^R_N}\bra{\uparrow^R_N} +\ket{\downarrow^R_N}\bra{\downarrow^R_N})
\right]
= \frac{1}{4} {\rm Id}^L_{1} \otimes {\rm Id}^R_N.
\end{align}
The partial inversion $P_I$ acts as
\begin{align}
\hat P_I \ket{\uparrow^L_1} = \ket{\downarrow^R_N},
\quad
\hat P_I \ket{\downarrow^L_1} = -\ket{\uparrow^R_N},
\quad
\hat P_I \ket{\uparrow^R_N} = \ket{\downarrow^L_1},
\quad
\hat P_I \ket{\downarrow^R_N} = -\ket{\uparrow^L_1},
\end{align}
from which we read off
\begin{align}
\braket{\Psi|\hat P_I|\Psi} = {\rm tr}_I( \hat P_I \rho_I) = - \frac{1}{2}.
\end{align}
We thus obtained the topological invariant $b(P,P) = -1$,
which, as expected, is non-trivial (differs from $b(P,P)=1$),
and is the $\mathbb{Z}_2$ invariant.
\paragraph{Crosscap from time-reversal symmetry: ``adjacent partial transposition''}
\begin{figure}[!]
\begin{center}
\includegraphics[width=\linewidth, trim=0cm 0cm 0cm 0cm]{figs/mps/Partial_Transpose.pdf}
\end{center}
\caption{[a] MPS representation of the crosscap. [b] The path integral representation of partial transposition. }
\label{Fig:Partial_Transpose}
\end{figure}
\begin{figure}[!]
\begin{center}
\includegraphics[width=\linewidth, trim=0cm 0cm 0cm 0cm]{TRS_to_RP2_v2.pdf}
\end{center}
\caption{Equivalence between
adjacent partial transpose and inserting a cross cap.
The black line is identified. }
\label{Fig:TRS_to_RP2}
\end{figure}
Next, we describe the extraction of the crosscap invariant
when $G$ includes time-reversal.
To this end, we will consider the so-called partial transposition.
The partial transposition has been used, for example, to define
the entanglement negativity.~\cite{VidalWerner}
Pollmann and Turner~\cite{PollmannTurner2012}
showed the MPS network
(Fig.\ \ref{Fig:Partial_Transpose} [a]) of
the partial transposition on adjacent intervals $I= I_1 \cup I_2$
is nothing but the $\mathbb{Z}_2$ topological invariant $b(T,T)$
associated with the time-reversal symmetry $T \notin G_0$.
(Here, to be concrete, let
$I_1 = \{1, \dots, N\}$ and $I_2 = \{N+1, \dots, M\}$
be two adjacent intervals in the chain of length $L$.)
In the following, we will review the topological invariant of
Pollmann and Turner.
We will also note, with an eye toward the TFT descriptions of
SPT phases and their topological invariants, that
the Pollmann-Turner can be interpreted as a space-time path integral
on the real projective plane.
Below, we will express the Pollmann-Turner invariant in terms of a density matrix.
We introduce the reduced density matrix for the interval $I$ by taking the
partial trace of the degrees of freedom on living on the compliment of $I$, $L\backslash I$,
as
\begin{align}
\rho_I = {\rm tr}_{L \backslash I} (\ket{\Psi} \bra{\Psi}).
\end{align}
The reduced density matrix
$\rho_I$ is expanded in the basis of $I_1$ and $I_2$ as
\begin{align}
\rho_I = \sum_{ijkl} \ket{e^1_i, e^2_j} \braket{e^1_i, e^2_j | \rho_I | e^1_k, e^2_l} \bra{e^1_k, e^2_l},
\end{align}
where $\ket{e^1_i}$ and $\ket{e^2_j}$ are the basis on the intervals $I_1$ and $I_2$, respectively.
We introduce the partial transposition $\rho_I^{T_1}$ for the interval $I_1$ which is defined by~\cite{VidalWerner}
\begin{align}
\rho_I^{T_1} := \sum_{ijkl} \ket{e^1_i, e^2_j} \braket{e^1_k, e^2_j | \rho_I | e^1_i, e^2_l} \bra{e^1_k, e^2_l}.
\end{align}
In addition to the partial transposition,
we also consider the ``unitary part'' of time-reversal,
and consider
partial time-reversal transformation $\hat T'_{I_1} = \prod_{j \in I_1} (U_T)_j$
action only on $I_1$.
Note that $\hat T'_{I_1}$ is unitary, i.e.,\ it consists of the only unitary part of the time-reversal transformation $\hat T = (\prod_j (U_T)_j) K$,
where $K$ is complex conjugation.
Putting everything together,
we consider
\begin{align}
{\rm tr}_I \big( \rho_I \hat T'_{I_1} \rho_I^{T_1} [\hat T'_{I_1}]^{\dag} \big), \quad \hat T'_{I_1} = \prod_{j \in I_1} (U_T)_j.
\label{path int PR2}
\end{align}
Finally, the Pollmann-Turner topological invariant, i.e.,
the crosscap topological invariant
is given by the phase
of \eqref{path int PR2},
\begin{align}
\theta(T) = \frac{{\rm tr}_I \big( \rho_I \hat T'_{I_1} \rho_I^{T_1} [\hat T'_{I_1}]^{\dag} \big)}{|{\rm tr}_I \big( \rho_I \hat T'_{I_1} \rho_I^{T_1} [\hat T'_{I_1}]^{\dag} \big)|}
\to b(T,T) , \quad (N,M \to \infty).
\label{P-T invariant}
\end{align}
In the limit $N,M \gg \xi$, where $\xi$ the correlation length,
the phase $\theta(T)$
is a quantized topological invariant.
The path integral representation of the quantity
\eqref{path int PR2}
is shown in Fig.\ \ref{Fig:Partial_Transpose} [b],
which is topologically equivalent to a sphere with one crosscap as shown in Fig.\ \ref{Fig:TRS_to_RP2}.
That \eqref{P-T invariant} is indeed a quantized and topological invariant
can be proven within the MPS framework.
Here, we demonstrate this by again using the cut and glue construction~\cite{Qi2012},
and by taking the Haldane chain with time-reversal symmetry as an example.
Within the cut and glue construction,
there are six active degrees of freedom at low energies
in the reduced density matrix,
\begin{align}
\ket{e^R_0} \otimes \ket{e^L_1} \otimes \ket{e^R_N} \otimes \ket{e^L_{N+1}}\otimes \ket{e^R_M} \otimes \ket{e^L_{M+1}},
\quad
e^L_j, e^R_j = \uparrow, \downarrow.
\end{align}
The ground state is a singlet formed from $(\ket{e^R_0} ,\ket{e^L_1})$, $(\ket{e^R_N} ,\ket{e^L_{N+1}})$ and $(\ket{e^R_M} ,\ket{e^L_{M+1}})$ as
\begin{equation}\begin{split}
\ket{\Psi}
&= \frac{1}{\sqrt{2}} (\ket{\uparrow^R_0} \otimes \ket{\downarrow^L_1} - \ket{\downarrow^R_0} \otimes \ket{\uparrow^L_1} )
\otimes \frac{1}{\sqrt{2}} (\ket{\uparrow^R_N} \otimes \ket{\downarrow^L_{N+1}} - \ket{\downarrow^R_N} \otimes \ket{\uparrow^L_{N+1}} ) \\
&\qquad \otimes \frac{1}{\sqrt{2}} (\ket{\uparrow^R_M} \otimes \ket{\downarrow^L_{M+1}} - \ket{\downarrow^R_M} \otimes \ket{\uparrow^L_{M+1}} ).
\end{split}\end{equation}
The reduced density matrix $\rho_I$ reads
\begin{align}
\rho_I
&= {\rm tr}_{0,{M+1}} (\ket{\Psi} \bra{\Psi})
\nonumber \\
&= \frac{1}{2} {\rm Id}^L_1
\otimes \frac{1}{2} \Big[
\ket{\uparrow^R_N} \bra{\uparrow^R_N} \otimes \ket{\downarrow^L_{N+1}} \bra{\downarrow^L_{N+1}}
+\ket{\downarrow^R_N} \bra{\downarrow^R_N} \otimes \ket{\uparrow^L_{N+1}} \bra{\uparrow^L_{N+1}}
\nonumber \\
&\ \ \ \ \ \ \ \ \ \ \ -\ket{\uparrow^R_N} \bra{\downarrow^R_N} \otimes \ket{\downarrow^L_{N+1}} \bra{\uparrow^L_{N+1}}
-\ket{\downarrow^R_N} \bra{\uparrow^R_N} \otimes \ket{\uparrow^L_{N+1}} \bra{\downarrow^L_{N+1}}
\Big]
\otimes \frac{1}{2} {\rm Id}^R_M.
\end{align}
By taking the partial transposition on $I_1 = \{1,N\}$ and
noting that the unitary part of the time-reversal transformation is given by
\begin{align}
U_T \ket{\uparrow} = \ket{\downarrow},
\quad
U_T \ket{\downarrow} = -\ket{\uparrow},
\end{align}
we have
\begin{align}
\hat T'_{I_1} \rho^{T_1}_I [\hat T'_{I_1}]^{\dag}
&= \frac{1}{2} {\rm Id}^L_1
\otimes \frac{1}{2} \Big[
\ket{\downarrow^R_N} \bra{\downarrow^R_N} \otimes \ket{\downarrow^L_{N+1}} \bra{\downarrow^L_{N+1}}
+\ket{\uparrow^R_N} \bra{\uparrow^R_N} \otimes \ket{\uparrow^L_{N+1}} \bra{\uparrow^L_{N+1}}
\nonumber \\
&\ \ \ +\ket{\uparrow^R_N} \bra{\downarrow^R_N} \otimes \ket{\downarrow^L_{N+1}} \bra{\uparrow^L_{N+1}}
+\ket{\downarrow^R_N} \bra{\uparrow^R_N} \otimes \ket{\uparrow^L_{N+1}} \bra{\downarrow^L_{N+1}}
\Big]
\otimes \frac{1}{2} {\rm Id}^R_M ,
\end{align}
which leads to
\begin{align}
{\rm tr}_I(\rho_I \hat T'_{I_1} \rho^{T_1}_I [\hat T'_{I_1}]^{\dag})
= - \frac{1}{8}.
\end{align}
The minus sign $(-1)$ is the proper $\mathbb{Z}_2$ invariant for the Haldane chain with time-reversal symmetry.
\paragraph{Klein bottle partition function from inversion symmetry}
Similar to the crosscap topological invariant,
the Klein bottle topological invariant can be
defined both for spatial inversion and
time-reversal.
Let us start with the case of spatial inversion.
We act with an inversion transformation $P \notin G_0$ on the twisted MPS $\ket{\Psi_g}$,
\begin{align}
\hat P \ket{\Psi_g}
= \sum_{ \{m_i\} } {\rm Tr}\, (A_{m_1} \cdots A_{m_L} V_P V^T_{g} V_P^{-1}) \ket{m_1\cdots m_L}
= \frac{b(P,g^{-1}) b(g,g^{-1})}{b(P g^{-1} P^{-1},P)} \ket{\Psi_{P g^{-1} P^{-1}}}.
\end{align}
If $P g^{-1} P^{-1} = g$, the $U(1)$ prefactor is well-defined which is invariant under the 1-coboundary.
This invariant is nothing but the partition function $Z_{KB}(P;g)$ over the Klein bottle ($KB$) with the background $P$ and $g$ twists
\begin{align}
\kappa(P;g) = Z_{KB}(P;g) = \braket{\Psi_g | \hat P | \Psi_g},
\quad (P \notin G_0, \ g \in G_0, \ P g^{-1}= g P).
\end{align}
An example is
a $\mathbb{Z}_2 (= \{1,\sigma\})$ paramagnet with inversion symmetry $\mathbb{Z}_2^P (= \{1,P\})$
where $\mathbb{Z}_2$ charge is preserved under the inversion.
The topological classification is given by $H^2(\mathbb{Z}_2 \times \mathbb{Z}_2^P,U(1)_{\phi}) = \mathbb{Z}_2 \times \mathbb{Z}_2$.
The two topological invariants can be seen in $V_P V_P^* = \theta(P)$ and $V_P V^T_{\sigma} = \kappa(P;\sigma) V_{\sigma} V_P$.
\paragraph{Klein bottle partition function from time-reversal symmetry: ``disjoint partial transposition with intermediate twist''}
\begin{figure}[!]
\begin{center}
\includegraphics[width=0.9\linewidth, trim=0cm 0cm 0cm 0cm]{figs/mps/Partial_Transpose_Klein.pdf}
\end{center}
\caption{[a] MPS representation of the Klein bottle partition function with twist. [b] The path integral representation. }
\label{Fig:Partial_Transpose_Klein}
\end{figure}
Next, as for
time-reversal symmetry,
the Klein bottle partition function
obtained from from time-reversal
can be represented in terms of MPSs in a way similar to the crosscap partition function.
First, we divide the closed chain $L$ into three adjacent intervals $I_1 \cup I_2 \cup I_3$,
$I_1 = \{1, \dots, N_1\}$, $I_2 = \{N_1+1, \dots, N_2\}$, $I_3 = \{N_2+1, \dots, N_3\}$.
In addition, we introduce one replica.
We trace out the region except for $I_1 \cup I_3$ {\it with symmetry twist in the interval $I_2$} as
\begin{align}
\rho_{I_1 \cup I_3}(g):= {\rm tr}_{L \backslash (I_1 \cup I_3)} \Big( \hat g_{I_2} \ket{\Psi} \bra{\Psi} \Big),
\quad
\hat g_{I_2} = \prod_{j \in I_2} (U_g)_{j}.
\end{align}
Then, we consider the following quantity
\begin{align}
{\rm tr}_{I_1 \cup I_3} \Big( \rho_{I_1 \cup I_3}(g) \hat T'_{I_1} \rho_{I_1 \cup I_3}(g^{-1}) [\hat T'_{I_1}]^{\dag} \Big),
\quad (T g^{-1} T^{-1} = g),
\quad
T'_{I_1} = \prod_{j \in I_1} (U_T)_j.
\end{align}
The corresponding MPS network and path integral are shown in Fig.\ \ref{Fig:Partial_Transpose_Klein} [a] and [b], respectively.
One can show this quantity approaches the Klein bottle partition function
\begin{align}
\frac{{\rm tr}_{I_1 \cup I_3} \Big( \rho_{I_1 \cup I_3}(g) \hat T'_{I_1} \rho_{I_1 \cup I_3}(g^{-1}) [\hat T'_{I_1}]^{\dag} \Big)}
{\Big| {\rm tr}_{I_1 \cup I_3} \Big( \rho_{I_1 \cup I_3}(g) \hat T'_{I_1} \rho_{I_1 \cup I_3}(g^{-1}) [\hat T'_{I_1}]^{\dag} \Big) \Big|}
\to Z_{KB}(T;g) = \kappa(T;g)
\end{align}
in the limit $N_1, N_2-N_1, N_3-N_2, L-N_3 \gg \xi$, where $\xi$ is the correlation length.
It is easy to show the above formula
for fixed point MPSs
by using the symmetry properties of $A$ matrix.
\section{$G$-equivariant topological field theories and MPSs}
\label{$G$-equivariant Topological field theory}
Having discussed the MPS description
of (1+1)d bosonic SPT phases,
we now move on (1+1)d $G$-equivaliant TFTs.
In the following sections,
Sec.\ \ref{Some basics of TFTs} to Sec.\ \ref{G-equivariant open and closed TFTs},
we briefly summarize necessary ingredients of
open and closed $G$-equivariant oriented $(1+1)$d TFTs
following Moore-Segal\cite{Moore-Segal}.
There are some overlaps with Ref.\ \cite{Kapustin-Turzillo},
where they also discuss closed $G$-equivariant $(1+1)$d unoriented TFTs.
Ref.\ \cite{Moore-Segal} also discusses
$(1+1)$d open and closed TFTs with spin structure,
which can describe fermionic SPT phases such as
class D topological superconductors.
Here we restrict ourselves to $(1+1)$d bosonic SPT phase protected by on-site unitary $G$-symmetry where $G$ is a finite group.
In short,
a $G$-equivariant TFT is a TFT couped with the background $G$-gauge field.
(Integrating out the background $G$-gauge field, i.e. orbifolding the $G$-symmetry,
leads to an orbifolded theory which is a TFT without $G$-symmetry.)
In the following, first, we introduce some general properties of TFTs.
Next, we summarize
$(1+1)$d $G$-equivariant closed TFTs
with an eye toward $(1+1)$d SPT phases.
Our notations closely follow Moore-Segal\cite{Moore-Segal}.
Next, we will summarize $(1+1)$d $G$-equivariant open and closed TFTs.
\subsection{Some basics of TFTs}
\label{Some basics of TFTs}
\begin{figure}[!]
\begin{center}
\includegraphics[width=0.7\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/ex.pdf}
\end{center}
\caption{An example of cobordism.}
\label{fig:tft_ex}
\end{figure}
In the axiomatic definition,
a TFT in $(d+1)$ dimensions
is a functor $Z$ from a cobordism category
$\mathbf{Bord}_{<d,d+1>}$
to the category of finite dimensional complex vector spaces $\mathbf{Vect}$
equipped with tensor product.~\cite{Atiyah:1989vu, Quinn}
In $\mathbf{Bord}_{<d,d+1>}$,
objects are $d$-dimensional manifolds $X_1,X_2,\dots $,
and morphism is a cobordism $Y: X_1 \to X_2$
which is a manifold of dimension $d+1$
and has $X_1$ and $X_2$ as its boundary components,
$\partial Y = (-X_1) \sqcup X_2$,
where $(-X)$ is $X$ with opposite orientation.
In general, we can associate a structure
(e.g. spin structure for spin TFTs, background gauge field for equivariant TFTs, etc.)
with manifolds.
For each $d$-dimensional manifold $X$, we associate a
Hilbert space ${\cal H}_X$ by a functor $Z$.
A direct sum of manifolds $X_1 \sqcup X_2 \sqcup \cdots $ is
mapped into a tensor product ${\cal H}_{X_1} \otimes {\cal H}_{X_2} \otimes \cdots$.
A cobordism $Y$ between $X$ and $X'$ leads to a linear map
$Z(Y) : {\cal H}_{X} \to {\cal H}_{X'}$.
See Fig.~ \ref{fig:tft_ex}, for an example.
In any TFT, the cylinder cobordism $X \times I$ leads
to the identity map $Z(X \times I) = {\rm id}$
\begin{align}
&\vcenter{\hbox{\includegraphics[width=0.2\linewidth]{figs/tft/id}}}
\ \Longrightarrow \ \
{\cal H}_X \overset{\rm id}{\longrightarrow} {\cal H}_X,
\end{align}
which is equivalent to the fact that the Hamiltonian of TFTs is zero.
In addition, we have a bilinear form $Q$ and a coform $\Delta$:
\begin{align}
&\vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/tft/Q}}} &&
\vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/tft/Delta}}} \notag \\
&{\cal H}_X \otimes {\cal H}_X \overset{Q}{\longrightarrow} \mathbb{C}, &&
\mathbb{C} \overset{\Delta}{\longrightarrow} {\cal H}_X \otimes {\cal H}_X.
\end{align}
Let $\{\phi_i \}$ be a basis of ${\cal H}_X$ and write
$Q(\phi_i, \phi_j) = Q_{ij}, \Delta(1) = \sum_{ij} \Delta_{ij} \phi_i \otimes \phi_j$.
The equivalence between the ``S-tube'' and the cylinder,
$$
\includegraphics[width=0.7\linewidth]{figs/tft/S}
$$
implies
\begin{align}
\phi \mapsto \phi \otimes \sum_{jk} \Delta_{jk} \phi_j \otimes \phi_k \mapsto \sum_i Q(\phi, \phi_j) \Delta_{jk} \phi_k = \phi , && \phi \in {\cal H}_X.
\label{eq:S-condition}
\end{align}
By setting $\phi = \phi_k$, we have $\sum_{j} Q_{ij} \Delta_{jk} = \delta_{ik}$,
which means $Q$ is nondegenerate.
Choosing the basis so that $Q_{ij} = \delta_{ij}$,
in this basis the coform $\Delta$ is simply $\Delta = \sum_{i} \phi_i \otimes \phi_i$.
\subsection{$G$-equivariant oriented closed TFTs}
A $G_0$-equivariant oriented $(1+1)$d TFT
is a functor $Z$ from
a cobordism category with a background $G_0$ gauge field
to the category of complex vector spaces.
(To distinguish on-site unitary symmetries from orientation-reversing symmetries,
here we use a notation $G_0$ to denote on-site unitary symmetries.)
For $(1+1)$d TFTs, the minimum object is an oriented circle
$(S^1,pt,g)$ with background $g \in G_0$ flux
together with a trivialization at a base point $pt \in S^1$,
which is specified by a twisted boundary condition by an element $g \in G_0$ at $pt$.
We denote the Hilbert space associated with $(S^1, pt, g \in G_0)$ by ${\cal C}_g$:
\begin{align}
&\vcenter{\hbox{\includegraphics[width=0.3\linewidth]{figs/tft/S1_g_comment}}}
\ \Longrightarrow \ \ \ \
{\cal C}_g = {\cal H}_{(S^1,pt,g)}, \ (g \in G_0).
\end{align}
For an unpointed circle $S^1$,
a background flux inserted in the circle $S^1$ is characterized by
a conjugacy class $[g] = \{ h g h^{-1} | h \in G\}$
rather than an element $g \in G_0$.
For point circles $(S^1,pt)$ with trivialization of background $G_0$ gauge field at $pt$,
Hilbert spaces are labeled by elements $g \in G_0$.
We have thus a $G_0$-graded Hilbert space
\begin{align}
{\cal C} = \bigoplus_{g \in G_0} {\cal C}_g.
\end{align}
\begin{figure}[!]
\begin{center}
\includegraphics[width=0.8\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/fusion_detail.pdf}
\end{center}
\caption{
[a] A fusion process ${\cal C}_g \otimes {\cal C}_h \to {\cal C}_{gh}$.
Dashed lines with group elements represent holonomies.
The figure [b] shows a holonomy around the boundary of simply connected space
that is obtained by cutting the surface in [a] at the lines connecting base points.
}
\label{fig:tft_fusion_detail}
\end{figure}
\begin{table}[!]
\caption{Building blocks of $G$-equivariant closed $(1+1)$d TFTs.
Building blocks (a)-(f) define $G$-equivariant oriented TFTs in $(1+1)$d.
$G$-equivariant unoriented TFTs in $(1+1)$d
are defined by including (g) and (h), in addition to (a)-(f).
$G_0 \subset G$ represents orientation preserving symmetries.
The fourth column shows corresponding simple and fixed point MPS representations (see Sec.~\ref{sec:tft-mps}). }
\begin{center}
\scalebox{0.8}{
\hspace*{-2cm}
\begin{tabular}{| >{\centering\arraybackslash}m{1cm} | >{\centering\arraybackslash}m{3cm} | >{\centering\arraybackslash}m{3.5cm} | >{\centering\arraybackslash}m{5cm} | >{\centering\arraybackslash}m{5cm} | }
\hline
& Manifolds & Hilbert spaces & Simple and fixed point MPS & Comment \\
\hline
(a) & \includegraphics[width=0.5\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/S1_g.pdf} &
${\cal C}_g \ (g \in G_0)$ & Hilbert space generated by $\ket{\Psi_g} = {\rm tr}(A_m V_g) \ket{m}, g \in G_0$ & Hilbert space over a space circle with $g$-flux \\
\hline
(b) & \includegraphics[width=0.5\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/closed_trace.pdf} &
$$
\begin{array}{l}
\theta_{\cal C}: {\cal C}_1 \to \mathbb{C}, \\
\phi \mapsto \theta_{\cal C}(\phi) \\
\end{array}
$$
& $\theta_{\cal C} \big( {\rm tr} (A_m) \ket{m} \big) = 1$ &
$Q(\phi_1, \phi_2) := \theta_{\cal C}(\phi_1 \phi_2), (\phi_1 \in {\cal C}_g , \phi_2 \in {\cal C}_{g^{-1}})$ is a bilinear nondegenerate form. \\
\hline
(c) & \includegraphics[width=0.5\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/disc.pdf} &
$$
\begin{array}{l}
\mathbb{C} \to {\cal C}_1, \\
1 \mapsto 1_{\cal C} \\
\end{array}
$$
& $1_{\cal C} = {\rm tr} (A_m) \ket{m}$ & State on the boundary of disc
$1_{\cal C} \phi = \phi 1_{\cal C} = \phi$. \\
\hline
(d) & \includegraphics[width=0.7\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/Cylinder.pdf} &
$$
\begin{array}{ll}
\alpha_{g\in G_0} : & {\cal C}_h \to{\cal C}_{ghg^{-1}}, \\
& \phi \mapsto \alpha_g(\phi) \\
\end{array}
$$
&
$$
\begin{array}{l}
{\rm tr} (A_m V_h) \ket{m} \\
\mapsto {\rm tr} (A_m V_h) \ \hat g (\ket{m}) \\
= {\rm tr} (A_m V_g V_h V_g^{\dag}) \ket{m} \\
\end{array}
$$
&
On site unitary $g \in G_0$ symmetry action \\
\hline
(e) & \includegraphics[width=\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/pants.pdf} &
$$
\begin{array}{l}
{\cal C}_g \otimes {\cal C}_h \to {\cal C}_{gh}, \\
\phi_1 \otimes \phi_2 \mapsto \phi_1 \phi_2 \\
\end{array}
$$
&
$$
\begin{array}{l}
{\rm tr} (A_{m_1} V_g) \ket{m_1} \\
\otimes {\rm tr} (A_{m_2} V_h) \ket{m_2} \\
\mapsto {\rm tr} (A_m V_g V_h) \ket{m} \\
\end{array}
$$
& ``Fusion'' of two closed chains \\
\hline
(f) & \includegraphics[width=\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/coform.pdf} &
$$
\begin{array}{l}
\Delta_g : \mathbb{C} \to {\cal C}_g \otimes {\cal C}_{g^{-1}}, \\
\Delta_g(1) = \sum_i \xi^g_i \otimes \xi_i^{g^{-1}} \\
\end{array}
$$
& ${\rm tr} (A_{m_1} V_g) \ket{m_1} \otimes {\rm tr} (A_{m_2} V_{g^{-1}}) \ket{m_2}$ &
$\xi^g_{i} \in {\cal C}_{g}$ are basis of ${\cal C}_{g}$ and
$\xi_{i}^{g^{-1}} \in {\cal C}_{g^{-1}}$
are their dual basis of ${\cal C}_{g^{-1}}$ that satisfy
$\theta_{\cal C} (\xi^g_{i} \xi_{j}^{g^{-1}}) = \theta_{\cal C} (\xi_{j}^{g^{-1}} \xi^g_{i}) = \delta_{i j}$. \\
\hline
\hline
(g) & \includegraphics[width=0.6\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/mobius} &
$$
\begin{array}{l}
\mathbb{C} \to {\cal C}_{g^2}, g \notin G_0, \\
1 \mapsto \theta_{g} \\
\end{array}
$$
& $b(g,g) {\rm tr} (A_m V_{g^2}) \ket{m}$ &
State on the boundary state of M\"{o}bius strip, ``cross cap state''. \\
\hline
(h) & \includegraphics[width=0.7\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/cylinder_P.pdf} &
$\alpha_{g\notin G_0} : {\cal C}_h \to{\cal C}_{gh^{-1}g^{-1}}$
&
$$
\begin{array}{l}
{\rm tr} (A_m V_h) \ket{m} \\
\mapsto {\rm tr} (A_m V_h) \ \hat g (\ket{m}) \\
= {\rm tr} (A^T_m V_g V_h V_g^{\dag}) \ket{m} \\
\end{array}
$$
&
$g$ reflection \\
\hline
\end{tabular}
\hspace*{-2cm}
}
\end{center}
\label{tab:functor_ori_tft}
\end{table}
In graphical representations of morphisms,
we specify the background gauge field by
holonomies connecting base points on initial and mapped circles.
For example, the fusion process of two circles with $g$ and $h$ fluxes
is represented in Fig.\ \ref{fig:tft_fusion_detail} [a].
The $G_0$ flux of mapped pointed circle $(S^1,pt)$
is determined by holonomies along base points as
shown in Fig.~\ref{fig:tft_fusion_detail} [b].
Recall that a holonomy around a boundary of simply connected spaces is trivial.
In short, we simply write the bordsim of the fusion process as
\begin{align}
&\vcenter{\hbox{\includegraphics[width=0.4\linewidth]{figs/tft/fusion_short}}}
\ \Longrightarrow \ \ \ \
{\cal C}_g \otimes {\cal C}_h \to {\cal C}_{gh}, \quad (g,h \in G_0).
\end{align}
In Table \ref{tab:functor_ori_tft},
we show building blocks of $G_0$-equivariant oriented (1+1)d TFTs.
All other cobordisms and partition functions can be constructed by processes in Table \ref{tab:functor_ori_tft}.
For example,
the
``branching'' process
is given by
\begin{align}
&\vcenter{\hbox{\includegraphics[width=0.8\linewidth]{figs/tft/separate}}}
\label{fig:tft/separate}
\end{align}
\begin{align}
\Longrightarrow Z \Big( \vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/tft/separate_2}}} \Big):
{\cal C}_g \to {\cal C}_{gh} \otimes {\cal C}_{h^{-1}}, &&
\phi \mapsto \sum_i \phi \xi_i^h \otimes \xi_i^{h^{-1}}
= \sum_i \xi_i^{gh} \otimes \xi_i^{(gh)^{-1}} \phi, &&
(g,h \in G_0).
\end{align}
Here,
we made use of Items (e) and (f) in Table\ref{tab:functor_ori_tft},
and
$\Delta_g(1) = \sum_i \xi_i^g \otimes \xi_i^{g^{-1}} \in {\cal C}_g \otimes {\cal C}_{g^{-1}}$
is the coform defined in Item (f).
The fusion process (e),
which, by an axiom of TFTs,
is associated to a map $\mathcal{C}_g \otimes \mathcal{C}_h \to \mathcal{C}_{gh}$,
makes the Hilbert space $\mathcal{C}$ into an algebra.
There are several constraints on the algebra,
which are obtained, e.g.,
by considering different factorizations
of surfaces into building blocks in Table \ref{tab:functor_ori_tft}.
Due to Turaev,~\cite{Turaev}
we have the minimum defining algebraic relations~\cite{Moore-Segal, Kapustin-Turzillo}:
To give a $G_0$-equivariant oriented TFT is
equivalent to give a $G_0$-graded algebra ${\cal C} = \bigoplus_{g \in G_0} {\cal C}_g$
together with a group homomorphism $\alpha : G_0 \to {\rm Aut}({\cal C})$
such that ${\rm Aut}({\cal C}) \ni \alpha_g : {\cal C}_h \to {\cal C}_{g h g^{-1}}$, and
\begin{itemize}
\item[(1)] There is a $G_0$-invariant trace $\theta_{\cal C} : {\cal C}_1 \to \mathbb{C}, \ \theta_{\cal C} \circ \alpha_g = \theta_{\cal C}$,
such that the induced paring ${\cal C}_g \otimes {\cal C}_{g^{-1}} \to \mathbb{C}$ is nondegenerate.
\item[(2)] For $\phi \in {\cal C}_g$, \ $\alpha_g (\phi) = \phi$.
\item[(3)] For $\phi_1 \in {\cal C}_{g_1}, \phi_2 \in {\cal C}_{g_2}$, $\alpha_{g_2}(\phi_1) \phi_2 = \phi_2 \phi_1$.
\item[(4)] (Punctured Torus) $\sum_i \alpha_{h}(\xi^g_i) \xi_i^{g^{-1}} = \sum_i \xi^h_i \alpha_{g}(\xi_i^{h^{-1}}) \in {\cal C}_{hg h^{-1} g^{-1}}$.
\end{itemize}
The non-degenerate property in (1)
is followed by the same way as (2).
Derivations of (1-4)
are summarized in Appendix \ref{app:Closed TFT}.
The state in (4) is
the handle adding operator
\begin{align}
\vcenter{\hbox{\includegraphics[width=0.2\linewidth]{figs/tft/handle}}}
= \sum_i \alpha_{h}(\xi^g_i) \xi_i^{g^{-1}},
\label{eq:tft_handle}
\end{align}
which enables us to
compute all possible partition functions on surfaces of genus $g$ with twist.
For example,
the partition function on torus $T^2$ with twist is given by
\begin{align}
\vcenter{\hbox{\includegraphics[width=0.2\linewidth]{figs/tft/Torus}}}
= Z_{T^2}(h,g)
= \sum_i \theta_{\cal C} (\alpha_h(\xi^g_i) \xi_i^{g^{-1}}), \ \ (hg=gh).
\end{align}
\subsubsection{General solution for semi simple cases}
If ${\cal C}_1$, the untwisted sector Hilbert space,
is semisimple,
${\cal C}_1 \cong \bigoplus_{x \in X} \mathbb{C} \epsilon_x$,
$\epsilon_x \epsilon_y = \delta_{x,y} \epsilon_x$,
we have general solutions for the
algebraic constraints
(1-4) as follows.~\cite{Turaev, Moore-Segal}
Here, $X$ is a finite set equipped with $G_0$-action
$g \cdot (h \cdot x) = (gh) \cdot x$.
For a given $G_0$-set $X$,
the twisted sector Hilbert space ${\cal C}_g$
consists of little group at $x$ as
${\cal C}_g = \bigoplus_{x \in X, g \cdot x = x} L_{g,x}$,
where $L_{g,x} \cong \mathbb{C}$ are lines.
The multiplication of ${\cal C} = \bigoplus_{g \in G_0} {\cal C}_g$ is determined by
a given group cocycle $b_x(g,h) \in Z^2(G_0,C(X,U(1))) \big( \cong Z^2_{G_0}(X,U(1)) \big)$ as
\footnote{$C(X,U(1))$ is the $G_0$-module consisting of
$U(1)$-valued functions on $X$.
The $G_0$-structure is defined by $(g \cdot f)(x) := f(g \cdot x)$, $g \in G_0, x \in X$.
The group cohomology $H^2(G_0,C(X,U(1)))$ classifies the following extension
$$
1 \to C(X,U(1)) \to \hat G \to G_0 \to 1.
$$
}
\begin{align}
\ell_{g_2,x_2} \ell_{g_1,x_1} =
\left\{\begin{array}{ll}
b_{x_1}(g_2, g_1) \ell_{g_2 g_1,x_1} & (x_2 = g_1 \cdot x_1) \\
\\
0 & (\mbox{otherwise})
\end{array}\right. &&
(\ell_{g,x} \in L_{g,x}).
\end{align}
The associativity condition $\ell_{g_3,x_3} (\ell_{g_2,x_2} \ell_{g_1,x_1}) = (\ell_{g_3,x_3} \ell_{g_2,x_2}) \ell_{g_1,x_1}$
corresponds to the 2-cocycle condition
\begin{align}
b_{g_1 \cdot x}(g_3,g_2) b_x(g_3 g_2, g_1) = b_x(g_2,g_1) b_x(g_3,g_2 g_1).
\end{align}
In short, $G_0$-equivariant TFTs are classified by
the group cohomology $H^2(G_0,C(X,U(1))) \big( \cong H^2_{G_0}(X,U(1)) \big)$.
To make a contact with physics of SPT phases,
let us specialize to the case where ${\cal C}_1$ is simple
${\cal C}_1 \cong \mathbb{C}$.
In this case, the ground state in the untwisted sector is unique,
and the classification is reduced into group cohomology with
$U(1)$ coefficient $H^2(G_0,U(1))$.
On the other hand, in semisimple cases,
we have a combination of symmetry breaking
and symmetry fractionalization discussed in
Refs.~\cite{Schuch2011, chen2011complete}.
Since the group cohomology $H^2(G_0, C(X,U(1)))$
splits into $G_0$-orbits,
we can simply assume that $X$ consists of a single $G_0$-orbit.
Let subgroup $G' \subset G_0$ be an unbroken symmetries,
then, we have a bijection $X \cong G_0/G'$ as a set,
which is a ``Nambu-Goldstone manifold''.
Each element $x \in G_0/G'$ represents
a vacuum which partially breaks $G_0$ symmetry and
retains $G'$ symmetry.
All the elements $G_0/G'$ are permuted by broken symmetries
in $G_0$.
The topological classification is given by
\begin{align}
H^2_{G_0}(G_0/G', U(1)) \cong H^2_{G'}(pt, U(1)) \cong H^2(G',U(1)),
\end{align}
says, the group cohomology classification for
unbroken symmetries.
\begin{figure}[!]
\begin{center}
\includegraphics[width=0.3\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/spec}
\end{center}
\caption{Example of a $G_0$-set $X$ for a combination of symmetry broken and
symmetry fractionalization.
In this example, $X$ consists of two points $X = \{x_1, x_2\}$
which are permuted by the broken symmetry $\sigma_3 \in G_0$.
}
\label{fig:tft/spec}
\end{figure}
For example,
let full symmetry be $G_0 = \mathbb{Z}_2[\sigma_1] \times \mathbb{Z}_2[\sigma_2] \times \mathbb{Z}_2[\sigma_3]$ and
unbroken symmetry be $G' = \mathbb{Z}_2[\sigma_1] \times \mathbb{Z}_2[\sigma_2]$,
where $\sigma_i (i=1,2,3)$ are generators of $\mathbb{Z}_2$.
In this case, $X$ consists of two points $\{x_1, x_2\}$ which
are exchanged by the broken symmetry as $x_2 = \sigma_3 \cdot x_1$
as shown in Fig.~\ref{fig:tft/spec}.
The topological classification is given by
that for the unbroken symmetry as
$H^2(\mathbb{Z}_2[\sigma_1] \times \mathbb{Z}_2[\sigma_2],U(1)) = \mathbb{Z}_2$.
See Fig.~\ref{fig:tft/spec}.
\subsection{$G$-equivariant unoriented closed TFTs}
$(1+1)$d oriented (closed) TFTs were extended to
unoriented TFTs by Turaev-Turner~\cite{turaev2006unoriented}
and equivariant unoriented TFTs by Kapsutin-Turzillo~\cite{Kapustin-Turzillo}.
See also Refs.~\cite{tagami2012unoriented, sweet2013equivariant}.
Here we review $G$-equivariant unoriented $(1+1)$d TFTs.
As before,
let $G$ be a full symmetry group including orientation-reversing symmetries
and $G_0 \subset G$ be the orientation-preserving subgroup.
There are two new ingredients to define
(equivariant) unoriented $(1+1)$d TFTs:
the crosscap state and reflection transformation
((g) and (h) in Table \ref{tab:functor_ori_tft}, respectively).
As for Item (g),
the boundary state of the M\"obius strip defines the
crosscap state $\theta_g \in {\cal C}_{g^2} (g \notin G_0)$
\begin{align}
\vcenter{\hbox{\includegraphics[width=0.4\linewidth]{figs/tft/crosscap}}}
\ = \theta_{g} \in {\cal C}_{g^2}, \quad (g \notin G_0).
\label{eq:tft_crosscap}
\end{align}
Notice that the crosscap state $\theta_{g}$ belongs to
the twisted sector of $g^2 \in G_0$.
As for Item (h),
the presence of an orientation-reversing symmetry $g \notin G_0$
can be used to
consider reflection of the circle
\begin{align}
\vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/tft/cylinder_P}}}
\ \Rightarrow \alpha_g : {\cal C}_h \to {\cal C}_{g h^{-1} g^{-1}},
\quad (g \notin G_0).
\end{align}
In a way similar to $G_0$-equivariant oriented TFTs,
we have several constraints on the algebraic category.
Kapustin-Turzillo\cite{Kapustin-Turzillo} showed that
to give a $G$-equivariant unoriented $(1+1)$d TFT
is equivalent to give a $G_0$-graded algebra
${\cal C} = \bigoplus_{g \in G_0} {\cal C}_g$
together with a group homomorphism
$\alpha : G \to {\rm Aut}({\cal C})$
such that $\alpha_{g \in G_0} : {\cal C}_h \to {\cal C}_{g h g^{-1}}$,
$\alpha_{g \notin G_0} : {\cal C}_h \to {\cal C}_{g h^{-1} g^{-1}}$.
They must satisfy
(1)-(4), and
\begin{itemize}
\item[(5)] $\alpha_g(\phi_1 \phi_2) = \alpha_g(\phi_2) \alpha_g(\phi_1)$, $g \notin G_0$.
\item[(6)] $\alpha_{h \in G_0}(\theta_g) = \theta_{h g h^{-1}}$ and $\alpha_{h \notin G_0}(\theta_g) = \theta_{h g^{-1} h^{-1}}$.
\item[(7)] (Punctured M\"obius strip) $\theta_g \phi = \alpha_g(\phi) \theta_{gh}$, $\phi \in {\cal C}_h$.
\item[(8)] (Punctured Klein bottle) $\sum_i \alpha_{g}(\xi^{(gh)^{-1}}_i) \xi_i^{gh} = \theta_g \theta_h$, $g,h \notin G_0$.
\end{itemize}
Derivations of these constraints\cite{Kapustin-Turzillo} are summarized in Appendix \ref{app:Closed TFT}.
All possible partition functions are constructed from
the handle adding operator (\ref{eq:tft_handle})
and crosscap adding operator (\ref{eq:tft_crosscap}).
For example,
the partition function on real projective plane reads
\begin{align}
Z_{\mathbb{R} P^2}(g)
= \vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/tft/RP2}}}
= \theta_{\cal C}(\theta_g), \qquad (g \notin G_0, g^2 = 1).
\end{align}
The Klein bottle partition function is
\begin{align}
Z_{KB}(g;h)
= \vcenter{\hbox{\includegraphics[width=0.2\linewidth]{figs/tft/KB}}}
= \theta_{\cal C}\big( \sum_i \alpha_g(\xi_i^h) \xi_i^{h^{-1}} \big)
= \theta_{\cal C}\big( \theta_g \theta_{g^{-1} h^{-1}} \big), \quad (g \notin G_0, h \in G_0).
\end{align}
\subsubsection{General solution for simple cases}
In the cases where ${\cal C}_1$ is simple ${\cal C}_1 \cong \mathbb{C}$,
i.e., there is a unique ground state,
Kapustin-Turzillo~\cite{Kapustin-Turzillo} showed general solutions
of the algebraic constraints (1) - (8).
They showed that to give a $G$-equivariant
unoriented $(1+1)$d simple TFT is to give a
2-group cycle $b(g,h) \in Z^2(G,U(1)_{\phi})$.~\footnote{
$U(1)_{\phi}$ is equipped with $G$-action defined by (\ref{eq:app_g_action}). }
This is consistent with the group cohomology classification of
bosonic $(1+1)$d SPT phases with reflection or time-reversal symmetry~\cite{chen2011complete}.
\subsubsection{Relation to MPS}
\label{sec:tft-mps}
In the SPT context, the spatial circle $S^1$ with $g$-flux in TFTs is
identified with a bulk SPT phase with $g$-twisted boundary condition.
The uniqueness condition of
the ground state in SPT phases
implies that the corresponding TFTs are invertible,
i.e.,
we have a simple algebra of untwisted sector ${\cal C}_1 \cong \mathbb{C}$.
In TFTs,
there is no excited state and the Hilbert space consists only of ground states.
The correlation length of the bulk is zero,
so a TFT is represented by a fixed point MPS
introduced in Sec.~\ref{Fixed point MPSs}
\begin{align}
\ket{\Psi_g}
= \sum_m {\rm Tr}\, (A_{m_1} \cdots A_{m_L} V_g) \ket{m_1 \cdots m_L}
\sim \sum_m {\rm Tr}\, (A_m V_g) \ket{m}.
\end{align}
Here we used the equivalence relation of fixed point MPSs (\ref{eq:mps_equivalence}).
Only one physical site is sufficient to describe the MPS representation of a TFT's ground state.
The correspondence between MPSs and equivariant TFTs,
can be pictorially represented as
\begin{align}
\vcenter{\hbox{\includegraphics[width=0.3\linewidth]{figs/tft/s1_mps}}}
\end{align}
Cobordisms in $G$-equivariant TFTs correspond to
various ``adiabatic deformations'' of closed chains, e.g.,
``fusion'' and ``separating'', and symmetry operations.
The fourth column in Table \ \ref{tab:functor_ori_tft}
summarizes correspondences between cobordisms in $G$-equivariant TFTs
and MPS representations.
For example, the fusion process of two
closed chain is {\it formally} represented in MPS networks as follows.
For two MPSs
\begin{align}
\ket{\Psi_g} = {\rm Tr} \big[ A_1 V_g \big],
\quad
\ket{\Psi_h} = {\rm Tr} \big[ A_2 V_h \big],
\end{align}
the fusion $\ket{\Psi_g} \cdot \ket{\Psi_h}$ is given by
\begin{align}
&\vcenter{\hbox{\includegraphics[width=0.6\linewidth]{figs/tft/fusion_mps}}} \\
&\Rightarrow
\ket{\Psi_g} \cdot \ket{\Psi_h} = {\rm Tr} \big[ A_2 A_1 V_g V_h \big] \ket{m_2 m_1} \sim b(g,h) {\rm Tr} \big[ A V_{gh} \big] \ket{m} = b(g,h) \ket{\Psi_{gh}}.
\end{align}
Here we used the equivalence relation of fixed point MPSs (\ref{eq:mps_equivalence}).
\subsection{$G$-equivariant open and closed TFTs}
\label{G-equivariant open and closed TFTs}
\begin{table}[h!]
\caption{Building blocks of $G$-equivariant open and closed $(1+1)$d TFTs.
In the forth column, MPS representations are shown.
In figures, dashed lines with group elements represent holonomies.
(r) is the definition of the boundary state for a boundary condition $a$.
}
\begin{center}
\scalebox{0.8}{
\hspace*{-2cm}
\begin{tabular}{| >{\centering\arraybackslash}m{1cm} | >{\centering\arraybackslash}m{3cm} | >{\centering\arraybackslash}m{3cm} | >{\centering\arraybackslash}m{6cm} | >{\centering\arraybackslash}m{5cm} | }
\hline
& Manifolds & Hilbert spaces & Simple and fixed point MPS & Comment \\
\hline
(i) & \includegraphics[width=0.2\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/Interval.pdf} &
${\cal O}_{ab}$ &
$$
\begin{array}{c}
\mbox{Hilbert space spanned by} \\
\Big\{ \big( L_i^T A_m R_j \big) \ket{m} \Big\}, \\
L_i \in V_a^*, R_j \in V_b.
\end{array}
$$ & Open chain. $L_i$ ($R_j$) are basis of
$V_a^*$ ($V_b$). \\
\hline
(j) & \includegraphics[width=0.3\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/Interval_trace.pdf} &
$$
\begin{array}{c}
\theta_{a}: {\cal O}_{aa} \to \mathbb{C}, \\
\psi \mapsto \theta_{a} (\psi) \\
\end{array}
$$
&
$$
\begin{array}{c}
\theta_a \Big( \big( v_L^T A_m v_R \big) \ket{m} \Big) = (v_L,v_R), \\
v_L \in V_a^* , v_R \in V_a \\
\end{array}
$$ &
$(v_L,v_R) = \sum_i [v_L]_i [v_R]_i$.
Notice that $\theta_a (1_a) = {\rm dim} V_a$ for simple and fixed point MPS.
\\
\hline
(k) & \includegraphics[width=0.3\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/Interval_create.pdf} &
$$
\begin{array}{c}
\mathbb{C} \to {\cal O}_{aa}, \\
1 \mapsto 1_{a} \\
\end{array}
$$
& $\sum_i \big( L_i^T A_m R_i \big) \ket{m}$ &
$1_{a}$ is the unit satisfying $1_{a} \psi = \psi 1_{b} = \psi, \psi \in {\cal O}_{ab}$. \\
\hline
(l) & \includegraphics[width=0.6\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/Interval_action.pdf} &
$$
\begin{array}{c}
\rho_{g \in G_0}: {\cal O}_{ab} \to {\cal O}_{ab}, \\
\psi \mapsto \rho_g \psi \\
\end{array}
$$
&
$$
\begin{array}{l}
\big( v_L^T A_m v_R \big) \ket{m} \\
\mapsto \big( v_L^T V_{g,a}^{\dag} A_m V_{g,b} v_R \big) \ket{m}, \\
v_L \in V_a^*, v_R \in V_b.
\end{array}
$$
& $g$-action on open chain. $V_{g,a}$ is representation matrix of $V_a$. \\
\hline
(m) & \includegraphics[width=0.6\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/Interval_action_ref.pdf} &
$$
\begin{array}{c}
\rho_{g \notin G_0}: {\cal O}_{ab} \to {\cal O}_{ab}, \\
\psi \mapsto \rho_g \psi \\
\end{array}
$$
&
$$
\begin{array}{l}
\big( v_L^T A_m v_R \big) \ket{m} \\
\mapsto \big( v_R^T V_{g,b}^{\dag} A_m V_{g,a} v_L \big) \ket{m}, \\
v_L \in V_a^*, v_R \in V_b.
\end{array}
$$
& $g$-reflection on an open chain \\
\hline
(n) & \includegraphics[width=0.6\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/Interval_Fusion.pdf} &
$$
\begin{array}{c}
{\cal O}_{ab} \otimes {\cal O}_{bc} \to {\cal O}_{ac}, \\
\psi_1 \otimes \psi_2 \mapsto \psi_1 \psi_2 \\
\end{array}
$$
&
$$
\begin{array}{c}
\big( v_L^T A_{m_1} v_R \big) \ket{m_1} \otimes \big( w_L^T A_{m_2} w_R \big) \ket{m_2} \\
\mapsto (w_L, v_R) \big( v_L^T A_m w_R \big) \ket{m}, \\
v_L \in V_a^*, v_R \in V_b, w_L \in V^*_b, w_R \in V_c.
\end{array}
$$
& Fusion of two open chains \\
\hline
(o) & \includegraphics[width=0.5\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/coform_interval.pdf} &
$$
\begin{array}{c}
\mathbb{C} \to {\cal O}_{ab} \otimes {\cal O}_{ba}, \\
1 \mapsto \sum_{\mu} \psi_{\mu} \otimes \psi^{\mu} \\
\end{array}
$$
& $\sum_{ij} \big( L_i^T A_{m_1} R_j \big) \ket{m_1} \otimes \big( L_j^T A_{m_2} R_i \big) \ket{m_2}$ &
$\psi_{\mu} \in {\cal O}_{ab}$ are basis of ${\cal O}_{ab}$ and $\psi^{\mu} \in {\cal O}_{ba}$ are their dual of ${\cal O}_{ba}$ that satisfy $\theta_{a} (\psi_{\mu} \psi^{\nu}) = \delta_{\mu}^{\nu}$. \\
\hline
(p) & \includegraphics[width=0.66\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/OpenToClose.pdf} &
$$
\begin{array}{c}
\imath^{g,a} : {\cal O}_{aa} \to {\cal C}_g, \\
\psi \mapsto \imath^{g,a}(\psi) \\
\end{array}
$$
& $\big( v_L^T A_m v_R \big) \ket{m} \mapsto (V_g^* v_L, v_R) {\rm Tr}(A_m V_g) \ket{m}$ & Open to closed map \\
\hline
(q) & \includegraphics[width=0.6\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/CloseToOpen.pdf} &
$$
\begin{array}{c}
\imath_{g,a} : {\cal C}_{g} \to {\cal O}_{aa}, \\
\phi \mapsto \imath_{g,a}(\phi) \\
\end{array}
$$
&
${\rm Tr}(A_m V_g) \ket{m} \mapsto \sum_i \big( L_i^T A_m V_g R_i \big) \ket{m}$ & Closed to open map \\
\hline \hline
(r) & \includegraphics[width=0.7\linewidth, trim=0cm 0cm 0cm 0cm]{figs/tft/bs} &
$B_a = \imath^{g,a}(1_a) \in {\cal C}_g$
&
${\rm tr}\big( V_{g,a}^{\dag} \big) {\rm Tr}(A_m V_{g,a}) \ket{m}$ &
Boundary state for boundary condition $a$.
\\
\hline
\end{tabular}
\hspace*{-2cm}
}
\end{center}
\label{Tab:Functor_Open}
\end{table}
\clearpage
Next, we extend closed TFTs to include open chains (intervals).
A new object is an oriented interval $I_{ab} = [0,1]$ with boundary conditions
$a,b$ as shown in Table.\ \ref{Tab:Functor_Open} (i).~\footnote{
In string theory,
the boundary conditions $a, b, \dots$ are
Chan-Paton factors associated with the
endpoints of open strings. }
We denote the Hilbert space associated with the interval $I_{ab}$ by ${\cal O}_{ab}$.
An element $\psi \in {\cal O}_{ab}$ represents
a state living in the open chain with boundary conditions $a$ and $b$.
(Note that the boundary conditions $a, b$ do not represent
some states in the open chain. )
Similar to closed TFTs,
we have several cobordisms in open and closed TFTs.
Table \ref{Tab:Functor_Open} summarizes the building blocks.
We have some remarks in order.
\begin{itemize}
\item
We use the same notation as Moore-Segal~\cite{Moore-Segal}.
The fusion process is represented as
\begin{align}
\vcenter{\hbox{\includegraphics[width=0.2\linewidth]{figs/tft/fusion_open}}}
\Rightarrow \
{\cal O}_{ab} \otimes {\cal O}_{bc} \to {\cal O}_{ac}, \ \ \psi_1 \otimes \psi_2 \mapsto \psi_1 \psi_2.
\end{align}
Note the order of two intervals $I_{ab}$ and $I_{bc}$.
\item
In addition to on-site symmetry transformation
$\rho_{g \in G_0} : {\cal O}_{ab} \to {\cal O}_{ab}$,
we have reflection on an open chain
$\rho_{g \notin G_0}: {\cal O}_{ab} \to {\cal O}_{ba}$
which exchanges the boundary conditions $a,b$.
$\rho_g$ satisfies $\rho_g \circ \rho_h = \rho_{gh}$ $(g,h \in G)$.
\item
Essentially new ingredients are
the open-to-closed map $\imath^{g,a}$
and the closed-to-open map $\imath_{g,a}$
which connect closed chains and open chains as~\cite{Moore-Segal}
\begin{align}
\imath^{g,a} : {\cal O}_{aa} \to {\cal C}_g, &&
\imath_{g,a} : {\cal C}_g \to {\cal O}_{aa}, &&
(g \in G_0).
\end{align}
Here, to glue back to a closed chain from a open chain,
the boundary conditions should agree.
\end{itemize}
All bordsims can be constructed by using building blocks listed in Table \ref{Tab:Functor_Open}.
For example,
a ``branching'' process of an open chain
is given by the same way as (\ref{fig:tft/separate}),
\begin{align}
&
\vcenter{\hbox{\includegraphics[width=0.6\linewidth]{figs/tft/separate_interval}}}
\nonumber \\
& \Rightarrow \ \
{\cal O}_{ab} \to {\cal O}_{ac} \otimes {\cal O}_{cb}, \ \
\psi \mapsto
\sum_{\mu} \psi \psi_{\mu} \otimes \psi^{\mu}
= \sum_{\mu} \tilde \psi_{\mu} \otimes \tilde \psi^{\mu} \psi.
\end{align}
Here, $\sum_{\mu} \psi_{\mu} \otimes \psi^{\mu} (\psi_{\mu} \in {\cal O}_{bc}, \psi^{\mu} \in {\cal O}_{cb})$
and $\sum_{\mu} \tilde \psi_{\mu} \otimes \tilde \psi^{\mu} (\tilde \psi_{\mu} \in {\cal O}_{ac}, \tilde \psi^{\mu} \in {\cal O}_{ca})$
are coform defined in (o) of Table \ref{Tab:Functor_Open}.
In the target algebraic category,
there are constraints from the open and closed cobordism category.
We have the following constraints for oriented open and closed TFTs by Moore-Segal~\cite{Moore-Segal}:
\begin{itemize}
\item[(9)] \
$\imath_{1,a}(1_{\cal C}) = 1_{{\cal O}_{aa}}$.
\item[(10)]\
$\rho_g (\psi_1 \psi_2) = (\rho_g \psi_1) (\rho_g\psi_2), \ g \in G_0, \psi_1 \in {\cal O}_{ab}, \psi_2 \in {\cal O}_{bc}$.
\item[(11)]\
$\imath_{g_1,a}(\phi_1) \imath_{g_2,a}(\phi_2) = \imath_{g_2 g_1,a}(\phi_2 \phi_1), \ \phi_1 \in {\cal C}_{g_1}, \phi_2 \in {\cal C}_{g_2}$.
\item[(12)]\
$\imath_{g,a}(\phi) (\rho_g \psi) = \psi \imath_{g,a}(\phi), \ \phi \in {\cal C}_g, \psi \in {\cal O}_{aa}$.
\item[(13)]\
$\theta_{a} ( \psi \imath_{g^{-1},a}(\phi) ) = \theta_{\cal C} ( \imath^{g,a}(\psi) \phi ), \ \phi \in {\cal C}_{g^{-1}}, \psi \in {\cal O}_{aa}$.
\item[(14)]\ ($G$-equivariant Cardy condition)
$\pi_{g,b}^a = \imath_{g,b} \circ \imath^{g,a} \mbox{ with } \pi_{g,b}^a(\psi) = \sum_{\mu} \psi^{\mu} \psi (\rho_g \psi_{\mu}),\ g \in G_0$.
\end{itemize}
For unoriented open and closed TFTs,
one can find the following additional constraints:
\begin{itemize}
\item[(15)]\
$\rho_g (\psi_1 \psi_2) = (\rho_g \psi_2) (\rho_g \psi_1), \ g \notin G_0, \psi_1 \in {\cal O}_{ab}, \psi_2 \in {\cal O}_{bc}$.
\item[(16)]\
$\sum_{\mu} (\rho_g \psi_{\mu} ) \psi^{\mu} = \imath_{g^2,a} (\theta_g), \ \ g \notin G_0, \ \psi_{\mu}, \psi^{\mu} \in {\cal O}_{aa}$.
\end{itemize}
In Appendix \ref{app:open TFT}, we summarize the derivations
of these constraints.
By solving these constraints, we can determine the
general properties of the target algebraic category for a
given $G$-equivariant closed TFT $b \in Z^2_G(X,U(1)_{\phi})$
with $G$-set $X$.
In the cases where ${\cal C}_1$ is semisimple, i.e.,
combination of symmetry breaking and symmetry fractionalization,
and there is no orientation-reversing symmetry,
Moore-Segal~\cite{Moore-Segal} gives the complete solution:
$b$-twisted equivariant vector bundles over $X$.
Here, for simplicity, we assume the ground state of closed chain is unique,
${\cal C}_1 \cong \mathbb{C}$,
and there are only on-site symmetries $G_0$.
We have~\cite{Moore-Segal}
\begin{itemize}
\item
The category of boundary conditions $\{a,b,\dots \}$
is equivalent to the category of $b$-projective
representations $\{V_a, V_b, \dots \}$.
\item
${\cal O}_{ab} \cong {\rm Hom}(V_b^*,V_a^*) = V_a^* \otimes V_b$.
\end{itemize}
This is precisely the boundary degrees of freedom that appear when one
introduce a boundary in SPT phases.
In the next section,
we describe
how to represent elements of ${\cal O}_{ab}$ and cobordisms
by using simple and fixed point MPS for open chains.
\subsubsection{Relation to open MPS}
In the SPT context, an interval $I_{ab}$ is
identified with an open SPT phase with
boundary condition $a$ and $b$.
An element of $\psi \in {\cal O}_{ab} \cong V_a^* \otimes V_b$
is identified with a state of the open chain Hilbert space~\footnote{
Note that $a,b$ do not specify a state
in the representation space of the $b(g,h)$-projective representations.
For example, the dihedral group $D_4 = \{1, C_4, C_2, C_4^{-1}, \sigma_x, \sigma_y, \sigma_d, \sigma_d'\}$ has two
$b$-irreps $E_{\frac{1}{2}}$ and $E_{\frac{3}{2}}$ for the nontrivial two cocycle $[b] \in H^2(D_4,U(1)) \cong \mathbb{Z}_2$.
In this case, $a, b$ specify $E_{\frac{1}{2}}$ or $E_{\frac{3}{2}}$.
}
\begin{align}
\psi
= \sum_m \big( v_L^T A_m v_R \big) \ket{m}
= \sum_m [v_L]_i [A_m]_{ij} [v_R]_j \ket{m},
\quad
v_L \in V_a^*,
\quad
v_R \in V_b.
\end{align}
The correspondence between MPSs and equivariant TFTs,
can be pictorially represented as
\begin{align}
\vcenter{\hbox{\includegraphics[width=0.2\linewidth]{figs/tft/interval_mps}}}
\end{align}
Cobordisms in $G$-equivariant open and closed TFTs correspond to
various ``adiabatic deformations'' of open chains and closed chains.
The fourth column in Table \ \ref{Tab:Functor_Open}
summarizes MPS representations,
which satisfy algebraic constraints (9) - (16).
For example, the fusion process of two
open chains is represented in MPS networks as follows.
For two open MPSs
\begin{align}
\psi_1 &= \sum_m v_L^T A^{(1)}_m v_R \ket{m}, \quad v_L \in V_a^*, v_R \in V_b, \\
\psi_2 &= \sum_m w_L^T A^{(2)}_m w_R \ket{m}, \quad w_L \in V_b^*, w_R \in V_c,
\end{align}
the fusion $\psi_1 \psi_2$ is given by
\begin{align}
\vcenter{\hbox{\includegraphics[width=0.4\linewidth]{figs/tft/interval_fusion_mps}}} &&
\Longrightarrow &&
\begin{array}{ll}
\psi_1 \psi_2 & = (w_L,v_R) \sum_{m_1 m_2} v_L^T A^{(1)}_{m_1} A^{(2)}_{m_2} w_R \ket{m_1 m_2} \\
& \\
& \sim (w_L,v_R) \sum_{m} v_L^T A_{m} w_R \ket{m} \\
\end{array}
\end{align}
Here we introduced a notation $(w_L,v_R) = \sum_i [w_L]_i [v_R]_i$
and used an equivalence relation of fixed point MPSs (\ref{eq:mps_equivalence}).
\subsubsection{Equivariant Cardy conditions and boundary states}
\begin{figure}[!]
\begin{center}
\includegraphics[width=0.7\linewidth, trim=0cm 2cm 0cm 0cm]{figs/tft/G_Cardy.pdf}
\end{center}
\caption{
$G$-equivariant Cardy condition.}
\label{fig:G_Cardy}
\end{figure}
In the derivation of the category of boundary conditions by
Moore-Segal~\cite{Moore-Segal},
the (generalized) $G$-equivariant Cardy condition (14) plays an essential role.
Here, we show MPS representations listed in the fourth column
in Table \ref{Tab:Functor_Open} satisfy the $G$-equivariant Cardy condition.
The Cardy condition comes from the equivalence between
(i) the double twist diagram shown in the left of Fig. \ref{fig:G_Cardy}
and
(ii) closed string channel shown in the right of Fig. \ref{fig:G_Cardy}.
These diagram can be interpreted in the context of
opne SPT chains :
\begin{itemize}
\item[(i)]
Cutting an open SPT chain $I_{aa}$ into two open chains $I_{ab} \sqcup I_{ba}$
and taking the $G_0$-action on the left chain ${\cal O}_{ab}$ and
exchanging two open chains and gluing back at $a$.
This process is written as
\begin{align}
\psi
= \big( v_L^T A_m v_R \big) \ket{m}
&\mapsto \sum_{j} \big( v_L^T A_{m_1} R_j \big) \ket{m_1} \otimes \big( L_j^T A_{m_2} v_R \big) \ket{m_2}
\nonumber \\
&\mapsto \sum_{j} \big( v_L^T V_g^{\dag} A_{m_1} V_g R_j \big) \ket{m_1} \otimes \big( L_j^T A_{m_2} v_R \big) \ket{m_2}
\nonumber \\
&\mapsto \sum_{j} \big( L_j^T A_{m_2} v_R \big) \ket{m_2} \otimes \big( v_L^T V_g^{\dag} A_{m_1} V_g R_j \big) \ket{m_1}
\nonumber \\
&\mapsto (V_g^* v_L, v_R) \sum_{j} \big( L_j^T A_m R_j \big) \ket{m}.
\end{align}
\item[(ii)]
Gluing the both ends of open SPT chain $I_{aa}$ to the $g$-twisted closed chain $(S^1,g)$ and cutting into the open chain $I_{bb}$.
This process is expressed as
\begin{align}
\psi
= \big( v_L^T A_m v_R \big) \ket{m}
&\mapsto (V_g^* v_L, v_R) {\rm tr} \big( A_m V_g \big) \ket{m}
\mapsto (V_g^* v_L, v_R) \sum_{j} \big( L_j^T A_m R_j \big) \ket{m}.
\end{align}
\end{itemize}
\begin{figure}[!]
\begin{center}
\includegraphics[width=1\linewidth, trim=0cm 2cm 0cm 0cm]{figs/tft/bs_def}
\end{center}
\caption{
The definition of $G$-equivariant boundary state $B_{g,a}$.
[a] Correlation functions on the upper half plane with boundary condition $a$.
[b] Amplitude of closed sector with insertion of the boundary state.
}
\label{fig:bs_def}
\end{figure}
It is useful to introduce the equivariant boundary state
$B_{g,a} \in {\cal C}_g$ in a way similar to
usual boundary state $B_a$ for non-equivariant TFTs.
Defining property of boundary state is that
the correlation functions on upper half plane with boundary condition $a$ (Fig.~\ref{fig:bs_def} [a])
is the same as the closed string amplitude with insertion of the boundary state (Fig.~\ref{fig:bs_def} [b]):
\begin{align}
\theta_a \big( \imath^{h,a}(\phi_1 \phi_2 \cdots \phi_n) \big)
= \theta_{\cal C} \big( B_{h^{-1}, a} \phi_1 \phi_2 \cdots \phi_n \big), &&
\phi_i \in {\cal C}_{g_i}, h = g_1 g_2 \cdots g_n.
\end{align}
From the algebraic constraint (13),
the $G$-equivariant boundary state
is given by
the image of open to closed map on the unit element $1_a$ of the open chain ${\cal O}_{aa}$,
\begin{align}
B_{g,a}
:= \vcenter{\hbox{\includegraphics[width=0.1\linewidth]{figs/tft/bs_2}}}
= \imath^{g,a}(1_{a}) \in {\cal C}_g.
\end{align}
For simple and fixed point MPSs, we have
\begin{align}
B_{g,a}
= {\rm tr} \big( V_{g,a}^{\dag} \big) {\rm Tr} \big( A_m V_{g,a} \big) \ket{m},
\end{align}
where $V_{g,a}$ is the representation matrix of the $V_a$ representation.
Notice that $\chi_a(g)^* = {\rm tr} \big( V_{g,a}^{\dag} \big)$
is the character of $V_a$ representation, which is
vacuous if there is a group element element $h \in G_0, [g,h]=0$
with nontrivial discrete torsion phase $b(g,h) \neq b(h,g)$.~\cite{cho2016relationship}
If we insert the boundary states in the Cardy condition (14),
we get a more familiar form
\begin{align}
\Braket{B_{g,b} | B_{g,a}}
= \theta_{\cal C} \big( B_{g^{-1},b} B_{g,a} \big)
= \theta_b \big( \imath_{g,b} \circ \imath^{g,a} (1_a) \big)
= \theta_b \big( \pi_{g,b}^a (1_a) \big)
= {\rm Tr}_{{\cal O}_{ab}} (\rho_g),
\end{align}
which is the character of $G$-action on the
open chain Hilbert space ${\cal O}_{ab}$,
\begin{align}
\chi_{{\cal O}_{ab}}(g)
= {\rm Tr}_{{\cal O}_{ab}} (\rho_g)
= \chi_a(g)^* \chi_b(g) .
\end{align}
\subsubsection{Crosscap invariant in open chain}
The partition function on $\mathbb{R} P^2$,
$Z_{\mathbb{R} P^2}(g) = \theta(g), g \notin G_0, g^2 = 1$,
can be detected in open chains.
Making use of the algebraic relation (16),
one can find the M\"obius strip with boundary condition $a$
is equivalent to the closed string amplitude from
crosscap $\theta_g$ to boundary state $B_{g^2,a}$,
\begin{align}
&\vcenter{\hbox{\includegraphics[width=0.5\linewidth]{figs/tft/mobius_a}}}
\nonumber \\
&\Rightarrow \
\braket{B_{g^2,a} | \theta_g}
= \theta_{\cal C}(B_{g^{-2},a} \theta_g)
= \theta_a(\imath_{g^2,a}(\theta_g))
= {\rm Tr}_{{\cal O}_{aa}} (\rho_g) , \ \ g \notin G_0.
\end{align}
For $g^2 = 1$ and unique ground state ${\cal C}_1 \cong \mathbb{C}$,
we have the topological invariant on the real projective plane,
which can be confirmed in simple and fixed point MPS as
\begin{align}
\theta_{a} \Big( \sum_{ij} \big( R_j^T V_g^{\dag} A_{m_1} V_g L_i \big) \ket{m_1} \cdot \big( L_j^T A_{m_2} R_i \big) \ket{m_2} \Big)
= [V_g]_{ji} [V_g^{\dag}]_{ij}
= {\rm dim}(V_a) \ \theta(g).
\end{align}
\subsection{State sum construction}
\label{Fukuma-Hosono-Kawai state sum construction}
In this section, we discuss the so-called state sum construction of TFTs.
Compared with the axiomatic approaches discussed previously,
the state sum construction exploits specific discretizations (triangulations)
of spacetime.
We will first review this construction for standard (non-equivariant) TFTs
following Fukuma-Hosono-Kawai \cite{FHK}.
We will then consider the state sum construction of $G$-equivariant TFTs,
and compute, among others, the partition functions on the torus, Klein bottle, and real
projective plane.
As promised earlier, we will confirm that they match precisely with the
topological (SPT) invariants derived from MPSs.
\subsubsection{Fukuma-Hosono-Kawai state sum construction}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.5 \linewidth, trim=0cm 0cm 0cm 0cm]{figs/state_sum/triangulation}
\end{center}
\caption{Triangulation of two-dimensional spacetime
and its dual (represented by double lines).}
\label{fig:state_sum/triangulation}
\end{figure}
Let us start by briefly reviewing the Fukuma-Hosono-Kawai
state sum construction.
\cite{FHK}
In the state sum construction of oriented 2d TFTs,
we consider a triangulation of 2d spacetime.
For a given triangulation, we can consider its dual, the dual
triangulation -- see Fig.\ \ref{fig:state_sum/triangulation}.
For faces and edges of the triangulation,
we associate
$\mathbb{C}$ numbers $C_{\mu \nu \rho}$ and $g^{\mu \nu}$ $(\mu,\nu,\rho = 1, \dots, N)$ as
\begin{align}
&
\vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/state_sum/triangle}}} \
= \vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/state_sum/dual_triangle}}} \
= \ C_{\mu \nu \rho},
\\
&
\vcenter{\hbox{\includegraphics[width=0.1\linewidth]{figs/state_sum/metric}}} \
= \vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/state_sum/dual_metric}}} \
= \ g^{\mu \nu}.
\end{align}
We demand that
$C_{\mu \nu \rho}$ is cyclically symmetric
$C_{\mu \nu \rho} = C_{\nu \rho \mu} = C_{\rho \mu \nu}$,
and $g^{\mu \nu}$ is symmetric $g^{\mu \nu} = g^{\nu \mu}$.
$g_{\mu \nu}$ is defined as the inverse of $g^{\mu \nu}$,
$g^{\mu \nu} g_{\nu \rho} = \delta^{\mu}_{\rho}$.
$g_{\mu \nu}$ and $g^{\mu \nu}$ are used for raising and lowering indices.
For example,
we introduce ${C_{\mu \nu}}^{\rho} = C_{\mu \nu \sigma} g^{\sigma \rho}$.
For a given triangulation $\Sigma_T$
of a surface $\Sigma$,
the partition function on $\Sigma_T$
is given by
\begin{align}
Z(\Sigma_T) = \sum_{\rm faces} C_{\mu \nu \rho} \sum_{\rm edges} g^{\eta \epsilon}.
\end{align}
In order to make $Z(\Sigma_T)$ independent of triangulations,
$C$ and $g$ have to satisfy the fusion and bubble conditions
\begin{align}
\label{eq:fusion_cond}
&\vcenter{\hbox{\includegraphics[width=0.3\linewidth]{figs/state_sum/fusion}}} : \qquad
{C_{\mu \nu}}^{\eta} {C_{\eta \rho}}^{\sigma} = {C_{\nu \rho}}^{\eta} {C_{\mu \eta}}^{\sigma},
\\
\label{eq:bubble_cond}
&\vcenter{\hbox{\includegraphics[width=0.3\linewidth]{figs/state_sum/bubble}}} : \qquad
g_{\mu \nu} = {C_{\mu \rho}}^{\sigma} {C_{\nu \sigma}}^{\rho}.
\end{align}
From the data of $C_{\mu \nu \rho}$ and $g^{\mu \nu}$,
one can introduce an algebra ${\cal C} = \oplus_{\mu=1}^N \mathbb{C} \phi_{\mu}$
as
\begin{align}
\phi_{\mu} \phi_{\nu} = C_{\mu \nu \rho} \phi_{\rho}.
\end{align}
The fusion condition (\ref{eq:fusion_cond}) means ${\cal C}$ is associative
$(\phi_{\mu} \phi_{\nu}) \phi_{\rho} = \phi_{\mu} (\phi_{\nu} \phi_{\rho})$.
We define a bilinear form by $Q(\phi_{\mu}, \phi_{\nu}) = g_{\mu \nu}$.
Existence of inverse of $g_{\mu \nu}$ ensures that $Q$ is non-degenerate and
the algebra ${\cal C}$ is semi simple.
The cyclicity condition of $C_{\mu \nu \rho}$
leads to the Frobenius condition
$Q(\phi_{\mu} \phi_{\nu}, \phi_{\rho}) = Q(\phi_{\mu}, \phi_{\nu} \phi_{\rho})$,
i.e.,\ ${\cal C}$ is a semi simple Frobenius algebra.
One can show that all the physical observables
constructed from the data $C_{\mu \nu \rho}$ and $g^{\mu \nu}$
depend only on the
center of ${\cal C}$,
$Z({\cal C}) = \{\phi \in {\cal C} | \phi \phi' = \phi' \phi , \forall \phi' \in {\cal C} \}$.~\cite{FHK}
In other words,
the Fukuma-Hosono-Kawai state sum construction
describes 2d oriented TFTs
which are equivariant to commutative semisimple Frobenius algebras.
For example, for a matrix algebra ${\cal C} = {\rm Mat}(\mathbb{C}^N)$ with
$Q(A,B) := {\rm Tr} A B$,
the center is trivial: $Z({\rm Mat}(\mathbb{C}^N)) = \mathbb{C} 1_{N \times N}$.
\subsubsection{$G$-equivariant state sum construction}
The state sum construction of $G$-equivariant closed TFTs
(both oriented and unoriented)
can be formulated in a way analogous to
the Fukuma-Hosono-Kawai construction of 2d oriented TFTs.~\cite{Turaev}
In the following, we
will discuss this within the context of TFTs describing SPT phases.
As before,
let $G$ be a symmetry group
which possibly includes orientation-reversing symmetries.
We specify orientation-preserving elements by subgroup $G_0 \subset G$.
We fix a group 2-cocycle $b(g,h) \in Z^2(G,U(1)_{\phi})$ and
assume $[b(g,h)] \in H^2(G,U(1)_{\phi})$ is nontrivial.
Let $V$ be a $b$-projective $N$-dimensional irrep.\ and $V^*$ be its dual.
Recall that $V$ represents the ``bond Hilbert space'' in MPSs.
$V$ also play an analogous role in the state sum construction,
which will be developed in the following.
The $G$ symmetry is projectively represented in the bond Hilbert space as
\begin{align}
&\hat g (\ket{i})
= \ket{j} [V_g]_{ji},
\quad V_g V_h = b(g,h) V_{gh},
\quad
\ket{i}, \ket{j} \in V,
\end{align}
in the same way as Sec.\ \ref{Fixed point MPSs}.
As in the Fukuma-Hosono-Kawai construction, we need
the input data -- the Frobenius algebra -- to boot-strap
a $G$-equivariant TFT.
To describe SPT phases (i.e., invertible TFTs),
we take the matrix algebra of $V$
as the algebra ${\cal C}$,
${\cal C} := {\rm End}(V) \cong V \otimes V^*$.
The bilinear non-degenerate form is defined
by the matrix trace $Q(X,Y) = N {\rm tr} (X Y) = N \sum_{ij} X_{ij} Y_{ji}$.
A canonical basis of ${\cal C}$ can be given as
\begin{align}
\left\{ E_{ij} = \ket{i} \bra{j} \right\}.
\end{align}
In this basis, $C_{ij, kl, mn} = C(E_{ij}, E_{kl}, E_{mn})$ and $g = g^{ij,kl} E_{ij} \otimes E_{kl}$
are given by
\begin{align}
C_{ij,kl,mn}& =
N \
\vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/state_sum/G_triangle}}} \
= N \delta_{ml} \delta_{kj} \delta_{in},
\\
g^{ij,kl} &=
\frac{1}{N} \ \vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/state_sum/G_metric}}} \
=\frac{1}{N} \delta^{il} \delta^{kj}.
\end{align}
One can show
$Q(E_{ij},E_{kl}) = N g_{ij,kl} = N \delta_{il} \delta_{kj}$,
${C_{ij,kl}}^{mn} = \delta_{kj} \delta_i^m \delta_l^n$,
$E_{ij} E_{kl} = {C_{ij,kl}}^{mn} E_{mn}$,
and can check
(\ref{eq:fusion_cond}),
and (\ref{eq:bubble_cond}).
Form the construction,
the algebra ${\cal C}$ has $G$ action
\begin{align}
&\hat g (X) = V_g X V^{\dag}_g, \quad X \in {\cal C}, g \in G_0, \\
&\hat P (X) = V_P X^T V^{\dag}_P, \quad X \in {\cal C}, P \notin G_0.
\end{align}
Here, observe that the orientation-reversing symmetry $P \notin G_0$ exchanges left and right.
This $G$ action can be
used to
to incorporate the background $G_0$ gauge field
in the networks of the state sum construction.
We introduce a symmetry twisted metric
by
\begin{align}
[T_g]^{ij,kl}
:= g^{ij,pq} Q\big( E_{pq}, \hat g(E_{rs}) \big) g^{rs,kl}
= \frac{1}{N} [V_g]_{il} [V^{\dag}_g]_{kj}
= \vcenter{\hbox{\includegraphics[width=0.25\linewidth]{figs/state_sum/G_twisted_metric}}}
\quad (g \in G_0).
\end{align}
We replace $g^{ij,kl}$ by $[T_g]^{ij,kl}$ on a nontrivial 1-cycle of the triangulation.
On the other hand, an orientation reversing symmetry $g \notin G_0$ induces
the exchange of indices $i$ and $j$.
We introduce the orientation reversing twisted metric~\cite{Karimipour-Mostafazadeh} by
\begin{align}
[T_P]^{ij,kl}
:= g^{ij,pq} Q\big( E_{pq}, \hat P(E_{rs}) \big) g^{rs,kl}
= \frac{1}{N} [V_P]_{ik} [V^{\dag}_P]_{lj}
= \vcenter{\hbox{\includegraphics[width=0.3\linewidth]{figs/state_sum/reflection_metric}}}
\quad (P \notin G_0).
\end{align}
\subsubsection{Partition functions}
Let us now construct, by using the state sum,
the partition functions on $T^2$, the Klein bottle, and $\mathbb{R}P^2$
(with symmetry twist).
We will show that these match precisely with the
topological invariants discussed and constructed by using MPSs
in Sec.\ \ref{Topological invariants}.
\paragraph{Partition function on $T^2$ with twist}
A background $G_0$ gauge field
on a torus $T^2$ is specified by
two commuting elements $g,h \in G_0, [g,h] = 0, g,h \in G_0$.
From the twisted metrics $T_g, T_h$
we have the torus partition function with twist
\begin{align}
Z_{T^2}(g,h)
&= \vcenter{\hbox{\includegraphics[width=0.2\linewidth]{figs/state_sum/Torus}}}
= {C_{ij,kl}}^{mn} C_{mn,pq,rs} [T_g]^{ij,pq} [T_h]^{rs,kl} \\
&= \frac{1}{N} {\rm tr}(V_g V_h V^{\dag}_g V^{\dag}_h) = \epsilon(g,h), \ \
(g,h \in G_0, h g h^{-1} = g).
\end{align}
This is the discrete torsion phase (\ref{eq:T^2inv}),
a topological invariant that characterizes $H^2(G,U(1)_{\phi})$.
\paragraph{Partition function on the Klein bottle with twist}
Similar to the torus partition function with twist,
the Klein bottle partition function with twist is computed in
the state sum construction.
Let $P \notin G_0$ be an orientation reversing symmetry and
$g \in G_0$ be a orientation preserving symmetry.
We have
\begin{align}
Z_{KB}(P;g)
&= \vcenter{\hbox{\includegraphics[width=0.2\linewidth]{figs/state_sum/KB}}}
= {C_{ij,kl}}^{mn} C_{mn,pq,rs} [T_P]^{ij,pq} [T_g]^{rs,kl} \\
&= \frac{1}{N} {\rm tr}(V_P V_g^T V^{\dag}_P V^{\dag}_g) = \kappa(P;g) , \ \
(g \in G_0, P \notin G_0, P g^{-1} P^{-1} = g).
\end{align}
Here, $\kappa(P;g)$ is the Klein bottle invariant of $H^2(G,U(1)_{\phi})$
introduced in (\ref{eq:KBinv}).
\paragraph{Partition function on $\mathbb{R} P^2$}
By using the orientation-reversing symmetry $P \in G_0$,
we can construct the partition funciton on
the real projective plane $\mathbb{R} P^2$ as
\begin{align}
Z_{\mathbb{R} P^2}(P)
&= \vcenter{\hbox{\includegraphics[width=0.2\linewidth]{figs/state_sum/RP2}}}
= \vcenter{\hbox{\includegraphics[width=0.2\linewidth]{figs/state_sum/RP2_2}}} \\
&= {C_{ij,kl}}^{mn} C_{mn,pq,rs} [T_P]^{ij,pq} [T_P]^{rs,kl}
= \frac{1}{N^2} {\rm tr}(V_P V_P^*) {\rm tr}(V_P V_P^{\dag}) = \theta(P),
\quad
(P \notin G_0, P^2 = 1).
\end{align}
This is the cross cap invariant (\ref{eq:Corsscapinv}).
\subsubsection{Cobordisms}
In addition to the closed surfaces considered above,
we can also consider surfaces with boundaries by using
the state sum construction.
From the generalities of TFTs,
a surface with boundary represents a state of the Hilbert space.
Here, we will construct various states that can be obtained
by considering state sum with open boundary/boundaries.
For our TFTs that describe SPT phases,
the physical Hilbert space ${\cal C}$ is spanned by a
basis of algebra $\{E_{ij}\}_{i,j = 1}^N$.
\paragraph{Disc (cap state)}
By the path integral on the disc,
we define a state associated to the disc (the cap state).
The cap state is the vacuum state on untwisted sector.
By triangulating the disc, the path-integral can be evaluated
explicitly as
\begin{align}
1_{\cal C}
&= \vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/tft/disc}}}
= \vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/state_sum/disc_1}}}
= \vcenter{\hbox{\includegraphics[width=0.1\linewidth]{figs/state_sum/disc_2}}}
\nonumber \\
&
= {C_{kl,mn}}^{mn} g^{kl,ij} E_{ij}
= \delta^{ij} E_{ij}
= \sum_i \ket{i} \bra{i}.
\end{align}
This is nothing but the simple and fixed point MPS representation of the ground state of SPT phases
$\ket{\Psi_1} = {\rm tr}(A_{ij}) \ket{i^L} \otimes \ket{j^R} = \frac{1}{\sqrt{N}} \sum_{i} \ket{i^L} \otimes \ket{i^R}$
introduced in Sec.\ \ref{Fixed point MPSs} up to a normalization.
\paragraph{M\"obius strip (cross cap state)}
By the path integral on the M\"obius strip,
we define a state associated to the M\"obius strip (the cross cap state).
By triangulating the M\"obius strip, the path-integral can be evaluated
explicitly as
\begin{align}
\theta_P
&= \vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/state_sum/mobius}}}
= \vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/state_sum/Mobius_1}}}
= \vcenter{\hbox{\includegraphics[width=0.1\linewidth]{figs/state_sum/Mobius_2}}}
\nonumber \\
&
= {C_{mn,pq}}^{ij} [T_P]^{mn,pq} E_{ij}
= \frac{1}{N} [V_P V_P^*]_{ij} E_{ij}
= \frac{1}{N} b(P,P) \sum_{ij} [V_{P^2}]_{ij} \ket{i} \bra{j},
\quad
(P \notin G_0).
\end{align}
\paragraph{Coform $\Delta_g$}
For a cylinder with two outgoing circles, following the axiom of TFTs,
we associate a coform $\Delta_g$.
By triangulating the cylinder,
we have
\begin{align}
\Delta_g
&= \vcenter{\hbox{\includegraphics[width=0.2\linewidth]{figs/tft/coform}}}
= \vcenter{\hbox{\includegraphics[width=0.2\linewidth]{figs/state_sum/coform}}}
\nonumber \\
&= {C^{ij}}_{mn,pq} {C^{kl,mn}}_{rs} [T_g]^{pq,rs} E_{ij} \otimes E_{kl}
= \frac{1}{N^2} \sum_{ijkl} [V^{\dag}_g]_{ij} [V_g]_{kl} E_{ij} \otimes E_{kl},
\quad (g \in G_0).
\end{align}
From this, we can read off the twisted ground state $\ell_g$ as
\begin{align}
\ell_g = \frac{1}{N} \sum_{ij} [V_g]_{ij} \ket{i} \bra{j}.
\end{align}
This is the same as the fixed point MPS ground state with twist up to a normalization.
\paragraph{Cylinder $\alpha_g$}
Since the Hilbert space with twist ${\cal C}_h (h \in G_0)$
is defined on the circle with $h$-flux,
we have to associate cylinder with twist by $T_h$.
For orientation preserving action $g \in G_0$, $\alpha_g$ reads
\begin{align}
\alpha_{g \in G_0} |_{{\cal C}_h}
&= \vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/tft/Cylinder}}}
= \vcenter{\hbox{\includegraphics[width=0.1\linewidth]{figs/state_sum/cylinder}}}
\nonumber \\
&= {C_{ij,mn}}^{pq} C_{pq,rs,tu} [T_h]^{tu,mn} [T_g]^{kl,rs} E^{ij} \otimes E_{kl}
= \frac{1}{N} [V^*_h]_{ij} E^{ij} \otimes [V_g V_h V_g^{\dag}]_{kl} E_{kl} ,
\quad (g \in G_0).
\end{align}
Here, $E^{ij}$ is dual basis of $E_{ij}$.
For an orientation reversing action $g \notin G_0$, $\alpha_g$ reads
\begin{align}
\alpha_{g \notin G_0} |_{{\cal C}_h}
&= \vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/tft/cylinder_P}}}
= \vcenter{\hbox{\includegraphics[width=0.1\linewidth]{figs/state_sum/cylinder_P}}}
\nonumber \\
&= {C_{ij,mn}}^{pq} C_{pq,rs,tu} [T_h]^{tu,mn} [T_P]^{kl,rs} E^{ij} \otimes E_{kl}
= \frac{1}{N} [V^*_h]_{ij} E^{ij} \otimes [V_P V^T_h V_P^{\dag}]_{kl} E_{kl} ,
\quad (g \in G_0).
\end{align}
\paragraph{Fusion}
The sphere with three punctures (the pants diagram)
describes a
fusion process ${\cal C}_g \otimes {\cal C}_h \to {\cal C}_{gh}$ ($g,h \in G_0$).
The path integral can be evaluated as
\begin{align}
&\vcenter{\hbox{\includegraphics[width=0.15\linewidth]{figs/tft/pants}}}
= \vcenter{\hbox{\includegraphics[width=0.25\linewidth]{figs/state_sum/pants}}}
\nonumber \\
&=
{C_{kl,ab}}^{cd} {C_{ef,cd}}^{gh} [T_h]^{ef,ab}
{C_{pq,ij}}^{rs} {C_{rs,tu}}^{vw} [T_g]^{pq,tu}
{C_{vw,gh}}^{mn}
E^{ij} \otimes E^{kl} \otimes E_{mn}
\nonumber \\
&= \frac{1}{N^2} [V^*_g]_{ij} E^{ij} \otimes [V^*_h]_{kl} E^{kl} \otimes [V_g V_h]_{mn} E_{mn}.
\end{align}
\section{Conclusion}
Tensor networks methods have have been employed as an efficient way to represent
correlated, entangled, many-body ground states.
In particular, they are expected to provide a powerful framework to study
gapped quantum many-body systems with (symmetry-protected) topological order.
On the other hand,
topological quantum field theories
have been playing an important guiding
role in topological phases of matter.
Indeed, one definition of a topological phase of matter is simply that it is
described by a TFT.
These two descriptions (methods) are complementary to each other:
The tensor network methods in general
can provide a powerful practical (numerical) framework
to study a given lattice model.
Within the tensor network framework, it is important to
develop a methodology to diagnose topological properties of
a given (ground state) many-body wave function.
E.g., to develop a method to extract topological invariants from
a given many-body wave function (in the tensor network representation).
On the other hand,
TFTs allow us to work directly in the topological limit (the limit of zero-correlation
length), and hence provide a systematic and abstract (axiomatic) framework
to, e.g., systematically classify possible topological phases of matter.
In this paper, focusing on $(1+1)$d bosonic SPT phases,
we bridge MPTs and TFTs in $(1+1)$d.
In particular,
we discuss $(1+1)$d $G$-equivariant (possibly unoriented) TFTs,
which are TFTs coupled with a background gauge field.
Our results are briefly summarized as follows:
-- In Sec.\ \ref{Classification and topological invariants of MPSs},
we summarized the construction of SPT invariants in terms of
MPS networks.~\cite{PollmannTurner2012}
By expressing those by a reduced density matrix,
MPS networks representing SPT
topological invariants can be identified
with path integrals on manifold with a twist.
We showed the partial inversion and
the adjacent partial transpose leads to
partition function on the real projective plane.
-- In Sec.\ \ref{$G$-equivariant Topological field theory}
we reviewed
$(1+1)$d $G$-equivariant open and closed TFTs by Moore-Segal~\cite{Moore-Segal}
which,
in addition to closed chains in $(1+1)$d closed TFTs,
have $(1+0)$d open chains as an object.
We established a
fixed point MPS representation of
$(1+1)$d $G$-equivariant open and closed TFTs
(when TFTs are invertible).
A concrete connection between the MPS and TFT descriptions
is summarized in Table \ref{tab:functor_ori_tft} and \ref{Tab:Functor_Open}.
In particular,
we noted, for example,
that the classification of the $G$-equivariant closed unoriented simple TFTs
is given by the second group cohomology,
which precisely is the known classification of $(1+1)$d SPT phases (without
orientation-reversing symmetry).
We also noted that semisimple TFTs correspond to a combination of
symmetry breaking and symmetry fractionalization discussed in the MPS context.
Furthermore, for $G$-equivariant open TFTs,
the category of boundary conditions is equivalent to
the known boundary degrees of freedom in $(1+1)$d SPT phases.
-- In Sec.\ \ref{Fukuma-Hosono-Kawai state sum construction},
we presented a state sum construction for
$G$-equivariant unoriented closed TFT for $(1+1)$d bosonic SPT phases.
The symmetry twisted metrics play roles of nontrivial holonomy.
Partition functions and correlation functions can be
calculated in a unified framework.
In particular, we showed that the partition functions
on the torus, the real projective plane, and the Klein bottle
match precisely with the SPT invariants constructed from the MPS method.
There is a number of natural extensions of the current paper:
For example, it is natural to speculate that we can make a precise dictionary between
higher-dimensional TFTs
and higher-dimensional tensor networks, such as
projected entangled pair states (PEPS).
Another interesting direction is
to consider topological phases of fermions, and
their descriptions in terms of (fermionic) tensor networks, and
spin TFTs.
(For recent works addressing these issues,
see Refs.\ \cite{barrett2015two, novak2015state}
(the state sum construction of (1+1)d oriented spin TFTs),
Refs.\ \cite{gaiotto2015spin, bhardwaj2016state}
($(1+1)$d and $(2+1)d$ oriented equivariant spin TFTs and the state sum construction),
and
Ref.\ \cite{Shapourian-Shiozaki-Ryu}.
\paragraph{Note added.}
After completing this work, we became aware of an independent work \cite{Kapustin-Turzillo-You},
which established the connection between the state sum construction of $(1+1)$d $G$-equivariant TQFT and MPS representations.
\acknowledgments
We thank
Gil Young Cho,
Kiyonori Gomi,
Andreas W.\ W.\ Ludwig,
Kantaro Ohmori,
Hassan Shapourian,
Tadashi Takayanagi,
Apoorv Tiwari,
Keisuke Totsuka,
Alex Turzillo,
Juven C. Wang,
Xueda Wen,
and
Peng Ye
for useful discussion.
Especially, K.S.\ is grateful to Takahiro Morimoto for pointing out
the equivalence between
a partial transposition and crosscap.
This work was supported in part
by the National Science Foundation grant DMR-1455296,
and by Alfred P. Sloan foundation.
K.S.\ is supported by JSPS Postdoctoral Fellowship for Research Abroad.
|
\section{Introduction}
High-altitude balloon platforms are an economical alternative to space missions for testing instruments as well as for specific classes of observations, particularly those that require a rapid response such as comets, or other transients. Telescopic platforms at high altitudes have significant advantages over operations from the ground enabling observations at forbidden wavelengths. A UV telescope (200--400 nm) in stratosphere with aperture of just 6 inch in diameter with sufficient pointing stability/accuracy and a 1K$\times 1$K CCD array could provide wide-field images with FWHM better than $1^{\prime\prime}$ approaching the diffraction limit (Fesen \& Brown, 2015), similar to that of space observatories but at a much lower cost. We have initiated a high-altitude balloon program at the Indian Institute of Astrophysics to develop low-cost instruments for use in atmospheric and astronomical studies (Nayak et al. 2013, Safonova et al. 2016). We have developed a number of payloads which operate in the near-ultraviolet (NUV), but we are limited to weights under 6 kg for regulatory reasons which constrains our payload size (Sreejith et al. 2016a). Our first experiments were of atmospheric lines (Sreejith et al. 2016b) where the pointing stability is less important, but we plan to observe astronomical sources for which a pointing mechanism is required.
Unlike in space missions, in stratospheric balloon pointing systems the payload is attached to the balloon by a long flight train. Besides transmitting the balloon's buoyant force, the flight train is the source of disturbances that the pointing control must reject. A typical stratospheric balloon has a train of several meters in length comprised of a recovery parachute beneath with flight termination systems, which is connected to a gondola consisting of scientific equipment with communication and associated electronics. The balloon flight, planned by the Tata Institute of Fundamental Research Balloon Facility (TIFR-BF) on 6--8th May 2016, provided a timely opportunity to piggyback attitude instrumentation on the gondola solely to measure its natural motion during the float phase of the flight. We expect that this information would be helpful in efforts to characterize disturbances that could be expected on any balloon-borne pointing system.
\section{Pointing Systems and Pointing Disturbances on High-Altitude Balloons}
With a telescope placed on the high-altitude balloon payload, one of the major factor in the design of the pointing system is the transfer of oscillations from the flight train; this constraint is more important if the required stability for observation is within a degree of accuracy. In our first experiments of atmospheric lines (Sreejith et al. 2016) the pointing accuracy was not of a great concern, but we plan to observe astronomical sources for which pointing and stabilization of the order of arcseconds are required. The disturbances can excite
the balloon--payload train dynamics, which at
the float altitude may disrupt pointing. There are two categories of disturbances:
transverse and azimuth. Transverse disturbances include swinging and bouncing
oscillations that mostly affect gondola (or telescope) motion about the two horizontal axes
(elevation and tilt). Azimuth disturbances are
rotation disturbances about the vertical that similarly affect the telescope azimuth
positioning.
Payload teams devise methods for mitigating the effects of these disturbances. The
primary strategy is to place the balloon in as quiet an environment as possible during the
observation, resulting in the so-called `observational windows' --- a period when the winds in
stratosphere are low and steady (see e.g. Manchanda et al. 2011). In addition, the balloon
team at TIFR-BF uses the bifilar load line suspension to mitigate the disturbances
(Robbins \& Martone, 1991).
The second strategy is in the careful design of the pointing system. The most common pointing configuration is the azimuth/elevation configuration, where the telescope is tilted about a horizontal elevation axis, fixed in the payload. We have designed and developed a low-cost lightweight, closed-loop pointing system build completely from off-the-shelf components (Nirmal et al. 2016). The system performance was checked on the ground and in tethered flights with satisfactory results. The system can point to an accuracy of $0.1^{\circ}$, and track objects from the ground with an accuracy of $\pm 0.15^{\circ}$.
It is not certain whether the balloon itself is going to rotate during a flight; however, balloon rotation even during float has been noted in the past (e.g. Gruner et al. 2005).
\section{Instrumentation}
With no requirement to point the gondola on this flight, an inertial measurement unit (IMU)\footnote{From x-io Technologies, UK, http://www.x-io.co.uk/products/x-imu/} with
internal Li-polymer battery (tested previously in the lab to work for up to 12 hrs) and a
4 GB SD card for data logging (Fig.~\ref{figure:IMU_photo} and Table~\ref{table:imu}),
was placed into the main payload (gondola) (Fig.~\ref{figure:IMU_payload}).
\begin{figure}[h!]
\centering
\includegraphics[width=.3\textwidth]{imu_photo.jpg}
\hskip 0.2in
\includegraphics[width=.3\textwidth]{reference_frame.jpg}
\caption{\label{figure:IMU_photo} {\it Left}: The x-IMU with housing and battery.
{\it Right}: The x-IMU reference frame. Axes $X$, $Y$ and $Z$ define Earth-centered inertial (ECI) reference
frame; $X'$, $Y'$ and $Z'$ define body-centered reference frame, and angles
$\alpha$, $\beta$ and $\gamma$ are the Euler angles.}
\end{figure}
\begin{table}[ht!]
\begin{center}
\caption{Technical specifications of the X-IMU sensor}
\begin{tabular}{lc}
\hline
Dimensions & $57\times 38 \times 21$ mm \\
Weight & 49 gms \\
Operating Temperature & -30 C to +85 C \\
Power & internal Li-polymer battery 3.7 V 1000 mAh \\
Data storage & internal 4GB SD card \\
Components & 3-axis accelerometer, 3-axis gyroscope, 3-axis magnetometer, T sensor\\
\hline
\end{tabular}
\label{table:imu}
\end{center}
\end{table}
\begin{figure}[h!]
\centering
\includegraphics[width=.25\textwidth]{fig1.png}
\includegraphics[width=.32\textwidth]{fig3.jpg}
\includegraphics[width=.37\textwidth]{fig2.jpg}
\caption{\label{figure:IMU_payload} {\it Left}: x-IMU before integration with the main
payload. {\it Middle}:
The main payload before closing the panels. The IMU will be placed in the middle tier.
{\it Right}: Final placement inside the main payload. A view through the specially made cut-out in the front panel.}
\end{figure}
The x-IMU uses Euler angle transformation to calculate yaw $\alpha$, pitch $\beta$
and roll $\gamma$ angles (Fig.~1). The roll and pitch angles are derived from the
accelerometer and gyroscope output, and the yaw readings from the magnetometer. According
to the way we place the x-IMU in the payload, the yaw angle is equivalent to the azimuth,
pitch is equivalent to the elevation, and tilt is equivalent to the roll angle. Since we
are only interested in the motion of the payload, not the absolute position, we take these
values of roll, pitch and yaw straight away for azimuth, elevation and tilt without any conversion.
\section{Flight Description}
The initial launch was proposed to be conducted before sunrise on May 6th 2016, however, weather conditions deteriorated rapidly and delayed the launch till the next day. The balloon was launched on May 7th 2016 at 6:43 am from the TIFR-BF, Hyderabad,
India (17.4729N, 78.5785E, altitude 532 m). The weather conditions were fairly clear with thin
clouds and favourable surface winds. Several rubber
weather balloons were tethered-launched prior to the mission to test the direction and strength of the
surface winds. The IMU was switched on at 5:35 am, few minutes before the final integration with
the main payload. Total payload weight was $\sim 200$ kg, connected to the balloon through the bifilar suspension. The main balloon was 20 $\mu$m-thick
plastic zero-pressure balloon of volume 38,211 $m^3$ filled with hydrogen. The average ascent rate was 4.59 m/sec.
The balloon reached the float altitude of 31.2 km at 08:36 am (ascent time 1 hr 53 min). The flight was
terminated by the onboard programmable timer at 12:42 pm IST on May 7th, and the payload with deployed parachute landed 380 km west of Hyderabad in good condition. The IMU was recovered on May 7th at 4:00 pm, with the power
and data logging still on at the landing site. The IMU was
powered off without switching off the data logging, and delivered to the TIFR-BF at 11:30 am on May 8th. In
order to keep the data intact, the IMU was powered on and the data
logger was switched off, followed by the IMU power off. The battery was working continuously for nearly 11 hrs, and
there was still power in the battery even after 22 hrs. In Fig.~\ref{fig:BAT}, we show the trend of the voltage with time for the first $\sim 4.5$ hrs of the flight.
\begin{figure}[h!]
\centering
\includegraphics[width=.6\textwidth]{BAT.jpg}
\caption{Battery voltage variation with time.}
\label{fig:BAT}
\end{figure}
\section{Observations and Results}
We have combined the data from our x-IMU and the data from measurements by the TIFR BF from the main payload. Because the x-IMU was powered off
before turning off the data logging, a lot of data got corrupted. We have recovered the data from only the first
$\sim 4.5$ hours of flight ($\sim 2.4$ hrs of float), which is sufficient for our purposes.
\subsection{Stratospheric Winds}
Despite the lack of atmosphere ($\lesssim 3$ mbar), the stratosphere both directly and indirectly affects the balloon-borne payload via stratospheric wind shears, causing translational and rotational acceleration as well as rotation.
The speed and direction of stratospheric winds obtained by iMet-1 GPS Radiosonde\footnote{InterMet Systems, Inc. South Africa. {\tt http://http://intermetsystems.com/}.}, were provided by the TIFR Balloon Facility. We need the speed of the wind to estimate its influence on the motion of the payload, and compare the stratospheric wind conditions at float with the near-surface conditions. The wind speed in stratosphere is shown in Fig.~\ref{figure:windspeed}.
\begin{figure}[ht!]
\hspace{-0.5in}
\includegraphics[scale=0.56]{windspeed.png}
\hspace{-0.18in}
\includegraphics[scale=0.56]{FloatWinds1.png}
\caption{Wind speed in stratosphere ({\it Left}) and at float ({\it Right}) from the TIFR-BF data. Shaded area shows standard deviation about the mean.}
\label{figure:windspeed}
\end{figure}
It is interesting to note that despite density decreasing with altitude, the lowest winds were reported at altitudes of about 20 km, rising after that (Fesen \& Brown, 2015). We also noticed that at $\sim 20$ km, the wind speeds were the least, at only 2--5 m/sec, while at the float altitude, $\sim 31$ km, the mean speed was higher at $\sim 25.5$ m/sec.
\subsection{Temperature variation}
Figure~\ref{figure:temp} shows the temperature inside (measured by our IMU) and outside
(measured by the TIFR-BF) the payload box for the first $\sim 4.5$ hours of flight.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.27]{temp_in_out.jpg}
\caption{\label{figure:temp} Temperature outside (blue line, TIFR data) and inside the payload box (green, IMU data).}
\end{figure}
\subsection{Payload Motion}
The variation of altitude with time for the first $\sim 4.5$ hours of the flight is shown in Fig.~\ref{figure:height}.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.8\textwidth]{height.jpg}
\caption{\label{figure:height} The variation of altitude with time from launch, obtained by the TIFR-BF.}
\end{figure}
There was no significant surface wind at launch, however, the disturbance from the load train caused the payload to
swing and rotate at the ascent (Fig.~\ref{figure:ascent}). The IMU provides data on the payload motion
in 3-axes. In addition, we can extract the accelerations in 3 directions.
\begin{figure}[ht!]
\centering
\includegraphics[scale=0.173]{fig4.png}
\includegraphics[scale=0.29]{fig5.png}
\caption{\label{figure:ascent} Swinging and rotation of the payload in the first minutes of ascent.}
\end{figure}
The payload motion in roll (tilt), pitch (elevation) and yaw axis (azimuth) measured by our x-IMU is displayed in
Fig.~\ref{figure:payload_rot}. We also obtained data on payload acceleration in three axes (displayed in Fig.~\ref{figure:acc}). We have correlated the motion of the payload at float with the changes in the wind direction. This is shown in Fig.~\ref{fig:cylinder}.
The plot on the left shows the change in the direction of the wind with respect to the North with time. The wind direction data points were generated from the GPS coordinates of the main payload with time resolution of 10 seconds. The wind direction plot is in meteorological convention. That is, for example, an angle of $110\deg$ from the North shows that the wind is blowing from $110\deg$ to its diagonally opposite point which is $250\deg$. The plot on the right shows the changes in the payload azimuth (yaw angle). Time axis shows the time from launch: at 1.5 hours the balloon reached an approximate altitude of 25 km, at $\sim 2$ hours the balloon reached the float altitude (see Fig.~\ref{figure:height} for the plot of altitude with time) till the end of available data from the x-IMU (about 3.1 hrs from launch). In this layer of the atmosphere we observe reduced wind turbulence comparing to the lower layers of the atmosphere and the reduction in the disturbances of the payload at float altitude.
\begin{figure}[hb!]
\vspace{0.3in}
\centering
\hspace{-0.35in}
\includegraphics[scale=0.57]{combined_figure_new.jpg}
\caption{\label{figure:ascent} Wind direction ({\it Left}) and movement of the payload in azimuth ({\it Right}) after reaching the stratosphere. Time axis shows time in hours elapsed since launch. The float altitude was reached at about 2 hrs after launch. The plots clearly show the difference between the rapid variation of payload azimuth before reaching the final float altitude and less variation of payload azimuth after reaching the float altitude. }
\label{fig:cylinder}
\end{figure}
\begin{figure}[ht!]
\includegraphics[width=0.56\textwidth]{tilt.jpg}
\includegraphics[width=0.53\textwidth]{tilt_zoom.jpg}
\includegraphics[width=0.56\textwidth]{azimuth.jpg}
\includegraphics[width=0.53\textwidth]{azi_zoom.jpg}
\includegraphics[width=0.56\textwidth]{elevation.jpg}
\includegraphics[width=0.53\textwidth]{ele_zoom.jpg}
\caption{\label{figure:payload_rot} Movement of payload in tilt (roll) ({\it Top}), azimuth (yaw)
({\it Middle}), and elevation (pitch) ({\it Bottom}). $X$-axis is time in hours from the switch on at 5:35 am IST.
On the left is the total retrieved data, and on the right is the data at the float altitude.}
\end{figure}
\begin{figure}[hb!]
\centering
\includegraphics[scale=0.17]{accx.jpg}
\includegraphics[scale=0.17]{accy.jpg}
\hspace{-0.35in}
\includegraphics[scale=0.17]{accz.jpg}
\caption{\label{figure:acc} The acceleration of payload in $X$ ({\it Top}), $Y$ ({\it Bottom left})
and $Z$ ({\it Bottom right}) axes during the flight.}
\end{figure}
\subsection{Magnetic field}
The variation of magnetic field around x-IMU in $X,Y,Z$ axes is shown in Fig.~\ref{figure:mag}.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.17]{magx.jpg}
\includegraphics[scale=0.17]{magy.jpg}
\hspace{-0.4in}
\includegraphics[scale=0.17]{magz.jpg}
\caption{\label{figure:mag} The variation of magnetic field strength in $X$ (Top),
$Y$ (Middle) and $Z$ (Bottom) axes during the flight.}
\end{figure}
\section{Summary and Conclusions}
The flight data derived from our x-IMU and from measurements by the TIFR-BF main payload are consolidated in Table~\ref{table: flight_summary}.
\begin{table}[hb!]
\caption{Summary of the flight ($\sim 2.5$ hrs at float altitude)}
\label{table: flight_summary}
\begin{center}
\begin{tabular}{|l|c|}
\hline
\rule[-1ex]{0pt}{3.5ex} Maximum height reached by payload & 31.4 Km \\
\hline
\rule[-1ex]{0pt}{3.5ex} Total duration of the flight & 5 hrs 56 mins \\
\hline
\rule[-1ex]{0pt}{3.5ex} Average ascent rate & 4.59 m/s \\
\hline
\rule[-1ex]{0pt}{3.5ex} Float altitude & 31.2 km \\
\hline
\rule[-1ex]{0pt}{3.5ex} Float reached (time) & 08:35 am IST \\
\hline
\rule[-1ex]{0pt}{3.5ex} \begin{tabular}{@{}c@{}} RMS motion of payload \\ during float in azimuth ($^{\circ}/s$) \end{tabular} & 0.5865 \\
\hline
\rule[-1ex]{0pt}{3.5ex} \begin{tabular}{@{}c@{}} RMS motion of payload \\ during float in elevation ($^{\circ}/s$) \end{tabular} & 0.0104 \\
\hline
\rule[-1ex]{0pt}{3.5ex} \begin{tabular}{@{}c@{}} RMS motion of payload \\ during float in tilt ($ ^{\circ}/s$) \end{tabular} & 0.01 \\
\hline
\rule[-1ex]{0pt}{3.5ex} \begin{tabular}{@{}c@{}} Average temperature \\ inside payload\end{tabular} & 34.35$^{\circ}$C \\
\hline
\rule[-1ex]{0pt}{3.5ex} \begin{tabular}{@{}c@{}} Average acceleration \\ of payload (X axis)\end{tabular} & 0.176 g \\
\hline
\rule[-1ex]{0pt}{3.5ex} \begin{tabular}{@{}c@{}} Average acceleration \\ of payload (Y axis)\end{tabular} & 0.0036 g\\
\hline\rule[-1ex]{0pt}{3.5ex} \begin{tabular}{@{}c@{}} Average acceleration \\ of payload (Z axis)\end{tabular} & 1.026 g\\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{enumerate}
\item During the flight the payload reached the maximum height of 31.2 km, where the outside temperature
was $-32^{\circ} $C. The temperature inside the payload stayed above $0^{\circ} $C. This shows that the
electronic components inside the payload were thermally insulated.
\item The stratospheric conditions during the TIFR flight at float are more stable than the near-surface
conditions we have experienced during our previous tethered launches at the IIA, and comparable to our
previous stratospheric flights. The full analysis is presented in the forthcoming paper (Nirmal et al. 2016).
\end{enumerate}
\section{Acknowledgements}
Part of this research has been supported by the Department of Science and Technology (Government of India) under Grant IR/S2/PU-006/2012. We thank Prof.~D.~K.~Ojha for allowing us to use our IMU in their flight. We also thank the staff of the TIFR
Balloon Facility, Hyderabad, for sharing the GPS and flight information.
\section{References}
\noindent
Fesen, R. \& Brown, Y. {\it A method for establishing a long duration, stratospheric platform for astronomical research}. 2015, Experimental Astronomy, 39, 475 \\
\\
\noindent
Gruner.~T.~D., Olney, D.~J. \& Russo, A.~M. {\it Measurements of Load Train Motion on a Stratospheric Balloon Flight}. 2005, NASA Goddard Space Flight Center Technical Reports Server (NTRS) \\
\\
\noindent
Manchanda, R.K., Subba Rao, J.V., Sreenivasan, S. \& Suneelkumar, B. {\it Study of seasonal variation of winds in upper stratosphere over Hyderabad}. 2011, Advances in Space Research, 47, 480–487\\
\\
\noindent
Nayak A., Sreejith A.~G., Safonova M. \& Murthy J. {\it High-Altitude Ballooning Program at the Indian Institute of Astrophysics}. 2013, Current Science, 104: 708-713\\
\\
\noindent
Nirmal, K., Sreejith, A. G., Mathew, J., Sarpotdar, M., Suresh Ambily, M. Safonova \& J. Murthy. {\it Pointing System for the Balloon-Borne Telescope}. 2016, Journal of Astronomical Telescopes, Instruments and Systems, submitted \\
\\
\noindent
Robbins, E. \& Martone, M. {\it Recovery System Termination Load Reduction through the use of a Central Load Core}. 1991, AlAA International Balloon Technology Conference, October 8-10, 1991, Albuquerque, NM, USA\\
\\
\noindent
Safonova, M., Nayak, A., Sreejith, A. G., Joice Mathew, Mayuresh Sarpotdar, S. Ambily, K. Nirmal, Sameer Talnikar, Shripathy Hadigal, Ajin Prakash \& Jayant Murthy. {\it An Overview of High-Altitude Balloon Experiments at the Indian
Institute of Astrophysics}. 2016, Astron. \& Astroph. Trans., Vol. 29, No. 3, in press \\
\\
\noindent
Sreejith, A. G., Mathew, J., Sarpotdar, M., Nirmal, K., Suresh, A., Prakash, A., Safonova, M. \& Murthy, J. {\it Measurement of limb radiance and Trace Gases in UV over Tropical region by Balloon-Borne Instruments -- Flight Validation and Initial Results}. 2016a, Atmos. Meas. Tech. Discuss., in review \\
\\
\noindent
Sreejith, A. G., Mathew, J., Sarpotdar, M., Nirmal, K., Suresh, A., Prakash, A., Safonova, M. \& Murthy, J. {\it Balloon UV Experiments for Astronomical and Atmospheric Observation}. 2016b, Proc. of SPIE, SPIE Astronomical Telescopes + Instrumentation 2016, Edinburgh, UK
\end{document}
|
\section{Introduction}
The widespread adoption of social media is challenging the way traditional media have been used to distribute news, and to discuss top social and political issues~\cite{lerman2010information, metaxas2012social, ferrara2013traveling}.
A large body of \emph{Computational Social Science} research focuses on the study of individuals and their behaviors on such platforms~\cite{lazer2009life, boyd2012critical, pentland2014social}. Various seminal papers investigate social and political conversations on social platforms like Twitter~\cite{ratkiewicz2011detecting, effing2011social, stieglitz2012political, bekafigo2013tweets} and Facebook~\cite{bond201261, carlisle2013social, ellison2014cultivating}. Yet, little work has been devoted to understand how the main actors of political discussion, the politician themselves, adopt and leverage such platforms~\cite{golbeck2010twitter, chi2010twitter, hemphill2013s}.
During the 2008 Presidential Election, Barack Obama used fifteen social media sites to support his campaign.
His successful effort demonstrated the central role of Twitter and other social platforms as integral parts of modern political communication.
Since then, online political discussion and the attention toward political candidates and political figures, and their social media presence, arose.
Politicians are influential figures in the offline world, and surely can acquire a great deal of influence in the social media spheres as well. Their social media activity, in turn, can alter their success and affect their careers, especially during election time.
The online campaigns preceding the 2016 Presidential Election carried out by both parties in support of various potential nominees, including Hillary Clinton, Bernie Sanders, and Donald Trump, further demonstrate the social media power to shape the political scene~\cite{wang2016deciphering, wang2016will}.
A better understanding of politicians' usage of social media channels for political conversation could therefore reveal something about the complex mechanisms of political success in the era of \emph{social politics}.
Yet, social media are not limited to political ``propaganda''.
The effects of social media political communication on the offline world are tangible. Examples of political campaigns that preceded mass mobilizations and civilian protests include the Arab Springs~\cite{howard2011opening, gerbaudo2012tweets}, Occupy Wall Street~\cite{conover2013geospatial, conover2013digital}, and the Gezi Park protest~\cite{varol2014evolution}. Although it is difficult to establish a causality link, we can safely say that the ``Twittersphere'' can be a strong indicator of political and public opinion~\cite{tumasjan2010election}.
The open nature of Twitter\footnote{At least with respect to other platforms like Facebook where ties are mostly formed based on pre-existing offline connections~\cite{demeo2014on}.} probably contributed to determine its \emph{political communicative power}. The ability to communicate interesting political issues yields the opportunity to users to acquire more visibility and influence~\cite{cha2010measuring, bakshy2011everyone, parmelee2011politics}, although Twitter political discussion is plagued by a number of issues related to manipulation and abuse~\cite{ratkiewicz2011detecting, ferrara2014rise, ferrara2015manipulation}.
In this paper we explore how the main actors of political discussion, the politicians, adopt Twitter to cover social and political issues. We focus on U.S. President Obama and all the 50 U.S. State Governors, and adopt the framework of \emph{agenda-setting theory} to identify their main topics of discussion.
The analysis of over one hundred thousand of their tweets reveals how Governors and the President use Twitter, what are the emerging patterns of political discussion, the top issues for each party, and finally who are the politicians who exhibit the most coherent political agenda.
\section{Social Media and Politics}
Twitter was born in 2006. In less than 10 years, it acquired half billion users, 310 million of which are active and produce over 500 million tweets per day as of July 2016.\footnote{Twitter official data: \url{https://about.twitter.com/company}} Twitter suggests that ``each tweet represents an opportunity to show one's voice and strengthen relationships with one's followers".\footnote{Twitter official blog: \url{https://blog.twitter.com/2014/what-fuels-a-tweets-engagement}}
As a modern political toolbox, Twitter has been widely used by various Presidents, Congressmen, Governors, and other politicians all over the world. In particular in the United States, Twitter and other social media have been not only the subject of extensive research, but also the platforms used to run large-scale social experiments to study political mobilization~\cite{bond201261}.
Scholars from various disciplines have investigated the role of these platforms in modern political communication.
Generally, social media research related to politics can be categorized into two fields. The former focuses on the possibility of using social media signals to predict political elections. A large number of papers faced this challenging question, with at times promising results. For example, Gibson and McAllister's study~\cite{gibson2006does} demonstrated a significant relationship between online campaigning and candidate support. Macnamara found evidence of a ``significant online political engagement'' in the 2008 U.S. Presidential Election~\cite{macnamara2010quadrivium}. Other studies covered the U.S. Presidential debate and Twitter sentiment, finding an alignment between popular opinions and votes~\cite{diakopoulos2010characterizing, stieglitz2012political, digrazia2013more}. Despite some promising work, the issue of predicting elections using social data remains debated~\cite{gayo2012wanted}.
The second area of research investigates Twitter users' behaviors, opinions and topics of political interest, at times proposing methods to identify their political alignments~\cite{conover2011predicting, cohen2013classifying}. Some of these studies highlighted interesting socio-political phenomena: for example, Conover \emph{et al.}~\cite{conover2011political} found that the network of political retweets exhibits a highly segregated bipartisan structure, which seems to reflect the users' political leanings, similarly to political blogs~\cite{adamic2005political}.
Shogan's \textit{et al.} research showed that, in recent years, Republican politicians tweeted more than five times as often as Democrats, suggesting that Twitter might be particularly appealing to American opposition politicians, who use it as an instrument for voicing their dissent directly to the public~\cite{glassman2009social, shogan2010blackberries}. A study conducted by Chi and Yang~\cite{chi2010twitter} found that Democratic congressmen tend to release information that citizens want to hear, while Republican congressmen share with the citizens their own agenda. Hemphill's work suggested that Congressmen of opposing parties use very different strategies to choose the hashtags that better reflect their framing efforts~\cite{hemphill2013framing}.
It appears that most literature either focuses on Twitter and elections, especially before and during election time, or focuses on President or Congressmen, even though ``most Americans have more daily contacts with their state and local governments than with the federal government''.\footnote{White House: State and Local Government, 2015 \url{https://www.whitehouse.gov/1600/state-and-local-government}}
Studies on State Governors and their social media presence are absent, and this paper aims at filling this gap. Although some research focuses on how politicians use social media before and during their election, what happens after that? Voters are excited about their party's success, and they are vocal about it. What comes after this initial excitement? We want to shed light on which Governors really follow their agenda after their election, and determine whether a framing of clear intents and goals emerges from their political channels online.
As of April 25, 2015, the 50 U.S. State Governors in charge collectively gathered over 3 million followers and sent out over 150,000 tweets. Though the majority of their Twitter accounts are merely political, some, such as Michigan Governor Rick Snyder's ``OneToughNerd'' account, show some character's personality traits, while others lend a certain intimacy, for example including family pictures like for Maryland Governor Larry Hogan, New Jersey Governor Chris Christie, Maine Governor Paul LePage and Louisiana Bobby Jingdal. Balancing personal lives and public service information makes State Governors' Twitter accounts very interesting objects to study the Governors' political stance in front of the public. This paper tries to dig into this unexplored field to analyze the State Governors' Twitter accounts by using agenda-setting theory, to understand whether the State Governors' activity on Twitter can be used to predict the popularity of parties or coalitions.
\section{Agenda-Setting Theory}
Twitter allows politicians to set their political agenda and reach their audience directly. Studying their behaviors brings the promising opportunity to further our understanding of \emph{agenda setting} in digital media~\cite{russell2014dynamics}. The agenda-setting theory is regarded as a key element to explain mass communication effects and mass media influence in long-term conditions. The primary assumptions of the theory were formulated by Maxwell McCombs and Donald Shaw in 1972~\cite{mccombs1972agenda}. Agenda setting is one of the most widely used theories in communication studies since then~\cite{iyengar1994anyone, iyengar2000new, weaver2004agenda, mccombs2005look, wanta2007effects}.
Agenda setting is the filter mass media perform when selecting certain issues and portraying them frequently and prominently, which leads people to perceive those issues as more important than others. Two levels of agenda-setting theory will be used in this study. The first-level agenda setting focuses on the amount of coverage of an issue, suggesting which issues the public will be more likely to be exposed to. The second-level agenda setting, also called \emph{framing} as suggested by McCombs, Shaw and Weaver~\cite{mccombs1997communication}, examines the influence of attribute salience, or the properties, qualities, characteristics, and relations. By making some political issues salient, agenda setting makes these specific issues more accessible than others.
The first level of agenda setting is the issue level. Though some scholars categorize top issues manually~\cite{russell2014dynamics}, we plan to use top issues listed on the White House's homepage. As of April 2015, the top seven issues listed were: economy, education, foreign policy, health care, immigration, climate change, energy and environment, and civil rights. April 2015 is also the time of our Twitter data collection. We will try to identify whether politicians give attention to these issues by analyzing how often kewords and hashtags related to these issue are mentioned on their Twitter accounts.
In the second level of agenda setting, we will analyze whether Democrats and Republicans highlight different attributes of the same issue by examining the hashtags and keywords they choose when they do discuss an issue.
We will also examine those hashtags and keywords relations by constructing occurrence networks to see how those hashtags and keywords are framed in the Governors' tweets.
Many researchers found different tweeting patterns among Democrats and Republicans Congressman, such as Shogan \textit{et al}.~\cite{glassman2009social, shogan2010blackberries} and Chi and Yang~\cite{chi2010twitter}. Our research as well aims to find whether State Governors' Twitter accounts exhibit different levels of engagement.
Then, we would like to further our understanding of the general patterns of usage, applying the second level agenda-setting theory, or framing, to scrutinize the hashtags and keywords network structure. Hence, we formalize the following three research questions:
\textbf{RQ 1}: How frequently do Governors use Twitter to discuss their political agenda? Do party differences emerge?
\textbf{RQ 2}: How do Governors' Twitter accounts reflect their political agenda, and how similar political agendas are across Governors?
\textbf{RQ 3}: What similarities and differences emerge in hashtag usage among Governors' Twitter accounts?
\section{Data Collection}
We used the Governors' timelines to reference the tweets from the 50 U.S. Governors and the U.S. President Barack Obama. We collected 114,316 tweets from the Governors' timelines. We downloaded the stream of tweets for each account by querying the Twitter Public API for user timeline by using a manually-collected list of account names. This returns the entire stream of tweets for each account, avoiding sampling issues~\cite{morstatter2013sample}. We performed the queries between January 23 and April 26, 2015, for all 51 accounts, in a systematic way and with a 100 second pause between each account. The pause was set to prevent our script from sending queries that exceed the rate limitation of the API. All data were finally stored into a JSON file and later analyzed.
We parsed each tweet to extract words and hashtags using the regular expression package \textit{re} with Python 3. We first removed the URLs by excluding patterns starting with http, https, ftp, and mailto. Then, tweet texts were converted into lowercase for consistency. Finally, we obtained hashtags and words by another set of regular expressions. The hashtags were defined as sets of concatenated characters starting with a pound sign (\#), while the words were defined as concatenated sets that start and end with alpha-numeric characters.
We identified the keywords by manually looking for the most frequent words that could be indicative of specific topics and sound meaningful to ordinary readers. To identify what could be the candidate words associated with each topic, we first manually parsed our collection of tweets and assigned the words that appeared together with the target topic as the candidate word selection for that topic.\footnote{Given the massive size of the dataset, with over one hundred thousand tweets, this procedure required three annotators and countless hours of work.}
For example, when we query for ``health care'' we will assign each of the 17 words (we, will, fight, to, protect, the, healthcare, of, Floridians, their, right, to, be, free, from, federal, overreach) appearing in the tweet ``We will fight to protect the healthcare of Floridians \& their right to be free from federal overreach.'' as a candidate choice of keywords for health care. All the stop-words that were identified by the Python Natural Language Toolkit (NLTK) were removed. In the previous example, the set of candidate words after this further cleansing is reduced to (fight, protect, heathcare, Floridians, right, free, federal, overreach). The next step was to remove the words that are syntactically needed but not contextually meaningful. We identified the words that were a keywords of more than one topic and manually marked them to be further removed or not. Words that were shared by more than one topic were marked to be deleted if we were unable to find a potential topic for them; words that possibly related to any of the topics were marked to be kept. In the example, words to be deleted included: fight, protect, Floridians, right, federal, overreach. These words could not be attached unequivocally to any one topic. For example, the words \textit{fight} and \textit{protect} appeared more often attached to foreign and immigration issues, and the word \emph{right} appeared more often related to civil right issues. Words to be kept included: \textit{healthcare} (as well as \textit{health care} with a space), and \textit{freedom}, which could be assigned to health care, in particular related to the Affordable Care Act (or, ObamaCare). After we identified which words to delete or keep, we then updated the sets of each candidate keywords for each topic. We then ranked each candidate keywords by their overall frequency in our collection. The top seven candidate keywords for each category were used to identify the topic of each tweet. We assigned a tweet to a topic whenever any of the 7 keywords for a topic appeared in a tweet. The topics were not mutually exclusive: in other words, one tweet could be assigned to more than one topic when the top candidate keywords from different categories occurred in a tweet. We counted the numbers of tweets for each topic among the Governors. The agenda was finally recovered by ranking the topics by the numbers of tweets associated to them: the results are displayed in Table~\ref{tab:top_words}. The assessment of the quality of the agenda produced by our semi-automatic method is satisfactory: the seven topics are each clearly identified by a short list of intuitive keywords. By means of the same approach, we varied the number of keywords to include more words, finding that the results (discussed later) were substantially unaltered. Finally, the proposed method to generate the agenda was preferred over traditional topic modeling techniques that we tested, such as LDA, because of the inability of such probabilistic generative models~\cite{blei2012probabilistic} to discriminate between topics related to issues relevant to politics, and other irrelevant (for our purpose) topics that appeared in the Governors' Twitter timelines.
\begin{table*}[!t]
\centering
\caption{Top Words per Category}
\begin{tabular}{@{}c@{} c@{} c@{} c@{} c@{} c@{} c@{}}
Civil Right & Economy & Education & Energy and & Foreign & Health Care & Immigration \\ [0.8ex]
& & & Environment & & & \\
\hline\hline
veterans & economic & education & energy & drug & health & investments\\
\hline
citizens & economy & students & manufacturing & sexual & food & immigration\\
\hline
rights & unemployment & school & water & assault & medicaid & employment\\
\hline
equal & manufacturing & veterans & affordable & campuses & insurance & sustainable\\
\hline
marriage & employees & schools & climate & uniform & transportation & struggling\\
\hline
defense & transportation & kids & tech & foreign & affordable & action\\
\hline
restoration & companies & college & capital & asia & freedom & portfolio\\
\hline
\end{tabular}
\label{tab:top_words}
\end{table*}
\section{Experimental Results}
\subsection{Overall Tweeting Patterns}
To try answer \textbf{RQ 1}, we analyze the 114,316 tweets collected from the Governor's timelines. The amount of tweets produced by each Governor ranged from 30 to 3,242, with a median of 2,838.
These figures demonstrate that the majority of Governors is quite active on the platform. There were 46,125 tweets posted by the 19 Democrat Governors, and 68,047 by the 30 Republican ones: this suggests that, on average, each Democrat produced 2,427 tweets, and each Republican posted 2268 tweets; this difference is not particularly significant.
President Obama contributed 3,242 to the Democrats, and the independent Governor of Arkansas had 144 tweets.
We were able to identify 75,202 hashtags and words from the tweet texts after removing the URLs. Democrat Governors used 50,960 words while Republican governors used 41,263.
The Democrats also tweeted more distinct hashtags, 6,463, while Republicans had only 4,264.
A previous study conducted by Shogan \textit{et al}.~\cite{glassman2009social, shogan2010blackberries} on the House tweeting patterns suggested that Republicans tweet more, and Twitter might be particularly appealing to the American opposition politicians.
Our analysis demonstrates that there are no significant differences in terms of average posting volumes between the two parties, and the larger sheer number of Republican tweets is to be attributed to the significantly greater number of Republican Governors (30 versus 19 Democrats).
However some stylistic differences emerge, in that Democrat Governors seem to make a much more pervasive and diverse use of hashtags than Republicans.
\subsection{Political agenda and keywords usage}
To answer \textbf{RQ 2}, we plan to describe each Governor's posting behavior according to the agenda we defined in Table~\ref{tab:top_words}. For each Governor's account, we calculated the number of times each keyword of Table~\ref{tab:top_words} appeared in any of the Governor's tweets. By sorting this dictionary of keywords and relative usage in descending order, we can obtain a rank of each Governor's keyword usage. We can therefore use the ranked keyword dictionaries to perform pairwise comparisons of Governors and try capture similarities and differences in priorities regarding the categories of political discussion. Note that using rankings is preferable to using simple feature vectors of keyword counts: ranks are more amenable to direct comparisons (for example via Spearman's rank correlation) without data normalization to account for different intensity of activities and other biases.
To measure the correlation of discussion keywords between all pairs of Governors, we use Spearman's correlation applied on their ranked keyword dictionaries. Spearman's rank correlation assigns each pair $<X_i, X_j>$ a similarity score between -1 and 1, with $X_i$ and $X_j$ being the keyword ranks of Governors $i$ and $j$ respectively. Score of 1 and -1 indicate perfect positive and negative correlation, respectively, whereas a score of 0 suggests no correlation.
To understand the distribution of pairwise correlation scores, we plotted Figure~\ref{fig:spearman}. The range of scores spans roughly from $-0.2$ (indicating a slight negative correlation) to very strong positive correlation scores greater than 0.8. The skewness towards positive scores can be attributed to the fact that we have considered only seven words per category, with seven total categories, for determining the rank distributions.
\begin{figure}[t!] \centering
\caption{Distribution of Spearman Rank Correlation scores}
\includegraphics[width=\columnwidth]{weight_distribution.png}
\label{fig:spearman}\vspace*{-.75cm}
\end{figure}
Figure~\ref{fig:matrix} shows the matrix of pairwise Spearman correlations among the 50 U.S. Governors plus the U.S. President Barack Obama. The visual inspection of Figure~\ref{fig:matrix} suggests the presence of a strong block structure, as groups of highly correlated accounts happen to be clearly identifiable. To further inspect this hypothesis, we generated a weighted graph of inter-Governor similarity using the matrix of Figure~\ref{fig:matrix} as adjacency matrix. The resulting graph is displayed in Figure~\ref{fig:governor_network}, where for visual clarity, self-loops have been removed and all edges with weights (i.e., Spearman correlation) less than 0.8 have been filtered out.
Figure~\ref{fig:governor_network} captures the agenda similarity network among Governors.
Its analysis suggests the emergence of a strong community structure, where Republican Governors appear to be strongly aligned on agenda priorities and form two tight clusters: the larger red cluster revolves around Wisconsin Governor Scott Walker, California's Jerry Brown, Maryland's Larry Hogan, Iowa's Terry Branstad and few others.
The second red cluster revolves around (former) Indiana Governor (and current Vice President nominee) Mike Pence, Maine's Paul LePage, New Jersey's Chris Christie, Luisiana's Bobby Jindal and few others.
The similarity, in terms of agenda priorities (as measured by the rank correlation) seems to be much less pronounced for Democrats: Governor of Rhode Island Gina Raimondo and West Virginia's Earl Tomblin, and Virginia's Terry McAuliffe form a small central cluster, whereas Colorado Governor John Hickenlooper, Missouri's Jay Nixon, Kentucky's Steve Beshear, and New Hampshire's Maggie Hassan show some agenda similarity. All the other Democrats Governors somehow sit at the periphery of this network showing spurious alignments with some of their Republican counterparts, and a less pronounces inter-party agenda priority sharing.
\begin{figure}[t!] \centering
\caption{Keyword-based correlation among Governors}
\includegraphics[width=\columnwidth]{heatmap_mod.png}
\label{fig:matrix}\vspace*{-.75cm}
\end{figure}
\begin{figure}[t!] \centering
\caption{Governors Network through the lens of Agenda Setting Theory}
\includegraphics[width=\columnwidth,clip=true,trim=160 80 220 80]{governor_network_filter_084}
\label{fig:governor_network}\vspace*{-.75cm}
\end{figure}
\subsection{The Governor-hashtag graph}
To address \textbf{RQ 3}, we finally explored the similarity among the governors at a hashtag level. We extracted the hashtags from each Governor's timeline and created a Governor-hashtag graph. The nodes in this bipartite graph represent the Governors and the hashtags they used. A Governor node and a hashtag node would be connected if the Governor had used the hashtag in any of his/her tweet. The weight is the number of tweets that contain that hashtag. We only extracted the hashtags that were used more than 10 times among all the Governors and by more than two Governors, to focus specifically on more common hashtags.
We were able to identify 658 common hashtags that occurred more than 10 times and were used by more than two Governors from our collection.
We also tried to recover the community structure by using the Louvain modularity maximization algorithm~\cite{blondel2008fast}.
The result for the Governors' hashtag usage are demonstrated in Figure~\ref{fig:governors_hashtag_network}.
The graph only represents the nodes that were connected with edges with weights larger than four, for visual clarity. The large circles denoted the nodes for Governors, and the small ones were nodes for hashtags.
We were able to identify four communities using the modularity algorithm with the resolution set to $2.0$. Varying the resolution limit parameter~\cite{fortunato2007resolution} provided consistent results. The four communities contained 36, 9, 3 and 3 Governors, respectively, as shown in Figure~\ref{fig:governors_hashtag_network}. We colored the largest community in red to indicate that it's the community with the largest fraction of Republican Governors (24). The second community is colored in blue to indicate that it's the community with the largest fraction of Democrats (8). The other two communities were colored in green and purple, respectively. We believe that the green cluster should belong to the Democrats (it contains Dems like Vermont's Peter Shumlin and New Hampshire's Maggie Hassan); the purple cluster contains several Republican Governors (e.g., Ohio's John Kasich and Maine's Paul LePage). Overall, the clustering algorithm assignment was correct for 32 of the 51 Governors (62.7\%). It generated 24 correct assignments out of the 30 Republicans (80.0\%), 8 correct among the 19 Democrats (42.1\%), and the Independent Governor of Arkansas was assigned to the reds.
In light of the most meaningful keywords for each of the seven categories summarized by Table~\ref{tab:top_words}, we parsed each Governor's timeline to determine to what extent the tweets of each individual were representative of each category. The underlying assumption of this strategy is that the more a State Governor tweets about any particular category, the more he/she is concerned about that particular issue, or at least wants to convey that message to his/her followers.
In general, for both parties, it is quite easy to scrutinize the most recurring topics of discussion of each Governor and identify those who concentrate more or less on politics and policy related topics, or or other types of events.
\begin{figure}[t!] \centering
\caption{Governors and Hashtag network}
\includegraphics[width=\columnwidth,clip=true,trim=0 140 0 140]{governors_hashtag_network_dark50}
\label{fig:governors_hashtag_network}\vspace*{-.75cm}
\end{figure}
Figure~\ref{fig:governors_hashtag_network} illustrates the most commonly occurring hashtags and issues of discussion of the two groups. Its analysis yields a good amount of insights into U.S. political discussion. One can notice the commitment of certain Governors to specific topics: for example, Vermont Governor Peter Shumlin seems pushing an agenda focused on environment, energy, and local economy issues. Other Democrats, like Connecticut Governor Dan Malloy, Arkansas' Asa Huthinson, the U.S. President Obama, focus on issues related to climate change, equality, health care, and education.
The Republican agenda is sufficiently diverse but focuses mostly on issues related to economy (small business, innovation, ``made in USA'', agriculture), immigration and security (human trafficking, Texas), and civil rights (especially veterans', military, and marriage rights).
A number of external events are also discussed (note that we did not remove any hashtags from the Governor-hashtag graph as long as it matched the threshold criteria explained above): some examples include reference to sport events (Nascar, Basket's March Madness, etc.), political events (2012 Elections, the GOP Convention, etc.) and tragedies (the Boston Marathon bombing, the Sandy Hook school shooting, etc.).
\section{Conclusions}
In this article we explored the landscape of U.S. Governors political communication on Twitter using the tool of agenda setting theory. We first collected a sizable amount of tweets (over one hundred thousand) generated by these politicians, and assessed that most of them are quite active Twitter users. Our results clarified some previous research about the usage of social media platforms by Democratic and Republican politicians, showing that Republican and Democrat Governors tend to be more or less equally active on Twitter on average, however they exhibit different styles of communication, with the Democrats significantly more inclined to use hashtags than their counterparts.
We furthered our understanding of Governors' priorities using the agenda-setting theory to identify a set of seven categories of top socio-political issues, by means of a semi-automatic annotation strategy.
After inferring the priorities of each Governor, and computing the pairwise similarity among Governors, we constructed a network that reflects Governor agendas similarity. Its analysis illustrates that President Obama has a distinctive agenda-setting strategy, which has no affinity with either Democrats or Republicans.
The graph also shows that Republican Governors, such as Arkansas Governor Asa Hutchinson, Tennessee's Bill Haslam, Louisiana's Bobby Jindal, Oklahoma's Mary Fallin, and Arizona's Gough Ducey, shared the most similar issue agenda settings. Republican Governors, compared to Democratic Governors, tend to be more clustered, which suggested that they share their party issues and priorities more than their Democrat counterparts, and in turn these priorities can be easily identified through the Governors' messages.
Many Democratic Governors did not align significantly with their party colleagues in top issue agendas. This suggested that Democrats have less cohesive priorities on Twitter as compared their Republicans counterparts. Similar insights were brought by examining the Governor-hashtag graph we built from hashtag usage patterns.
This study displayed the high-level dynamics of adoption of Twitter by U.S. Governors based on how they set their agenda on top political issues and how they frame their conversation around it. Further studies should explore the \textit{public agenda setting}, which means the agenda setting of the public in each State, to see if these share similar trends with their Governors' agendas. This would shed light on the effects of politicians' social media conversation on the public.
\bibliographystyle{splncs03}
|
\section{Introduction}
The recent detection of the gravitational wave (GW) sources GW150914 \citep{2016PhRvL.116f1102A} and GW151226 \citep{2016ApJ...818L..22A} opened the era of gravitational wave astronomy, and has provided the first direct confirmation of the existence of black holes (BHs), and in particular binary BHs (BBH) that merge within the age of the Universe. Based on the rate of BBH mergers inferred from these detections \citep{2016arXiv160203842A,2016arXiv160604856T} many more sources are expected to be discovered in the second and third Advanced LIGO observing runs \citep{2016arXiv160604856T}.
Ground-based interferometers such as Advanced LIGO, which is already gathering data, as well as VIRGO and KAGRA which are expected to become operational in the near future, are sensitive to gravitational waves in the frequency range $\sim 30 - 1000$ Hz, and are designed to detect mergers of BBH and binary neutrons stars (NSs), as well as the gravitational wave background from unresolved mergers of these binary compact objects \citep{2016arXiv160203847T}. Pulsar timing array (PTA) networks \citep{2013CQGra..30v4009K,2015MNRAS.453.2576L} may detect the GW background produced by merging super-massive BHs (SMBH), topological defects such as cosmic strings, and individually resolvable SMBHs in the frequency range $\sim 10^{-9}-10^{-8}$ Hz. The frequency ranges $\sim 10^{-4} - 10^{-1}$ Hz and $\sim 10^{-1} - 10$ Hz will be explored by the space-based eLISA \citep{PhysRevLett.116.231101} and DECIGO \citep{2011CQGra..28i4011K} observatories, respectively, planned to be launched in the next decade. The frequency coverage that will be attained when all of these observatories are operational suggests the possibility of multi-wavelength GW astronomy \citep{2016PhRvL.116w1102S,2016arXiv160501341N}, where the same source can be observed by different observatories as the merger proceeds.
Detections of individual sources, such as GW150914 and GW151226 are invaluable in studying the properties of compact objects and constraining gravity under extreme conditions. The masses and spins of observed BBH already begin to inform astrophysical models of BH formation \citep{2016ApJ...818L..22A,2016arXiv160204531B,2016arXiv160604889A} and future detections may provide information on the equation of state of NSs \citep{2015PhRvD..92b3012A}. Moreover, the waveforms of individual merger events allow to place stringent constraints on extensions to General Relativity \citep{2016arXiv160604856T}.
Another component that can be detected with GW observatories is the gravitational wave background from unresolved, merging and inspiraling sources. This component will allow to study the compact object population from a different viewpoint, in particular by constraining the distribution of the binary parameters and their formation mechanisms \citep{2016arXiv160203847T}.
The background from unresolved binary compact objects has in general three components: (a) the signal emitted by core-collapse supernovae (SNe) \citep{1985PhRvL..55..891S,1999MNRAS.303..247F}; (b) the contribution from objects that are about to merge (usually referring to inspiral, merger and ringdown phases) \citep{2008PhRvD..77j4017A,2011PhRvD..84h4037A} and (c) the contribution from inspiraling binaries which do not merge during a Hubble time but which still emit gravitational radiation, resulting in a circularisation and shrinking of their orbit. Component (b) is perhaps the most extensively studied, both in the domain of stellar-mass BBH and NSs as well as SMBH in view of its importance for predicting the signal of merger events such as GW150914 and GW151226. While the waveform of a single isolated merge is well understood, many uncertainties remain, in particular regarding the merger rates (which are related to the properties of the progenitors). GW background from stellar-mass BBH is expected to be detected by Advanced LIGO \citep{2016arXiv160203847T}, while the signal from SMBH is beginning to be constrained by PTA experiments and will be further probed by the eLISA satellite \citep{2015MNRAS.453.2576L,PhysRevLett.116.231101}. Finally, the GW signal from SN collapse is difficult to estimate due to uncertainties in the collapse mechanism \citep{2013ApJ...766...43M,2013ApJ...768..115O,2015PhRvD..92f3005C}.
While the contribution from merging compact binary systems is dominant, most binaries are not expected to merge within a Hubble time. They will, however, emit gravitational radiation while slowly approaching each other and, depending on the merger rate, the source mass and redshift distribution and the initial orbital parameters, might be detectable with future generations of GW observatories.
It is important to stress that the evolution of massive stars and compact objects is affected by their environment. Interactions with other stars in a dense star cluster might be an inportant channel for creating heavy stellar-mass BHs \citep{2016arXiv160604889A} and the growth rate of SMBH is clearly related to the properties of its host galaxy (e.g. \citep{2012Sci...337..544V}).
The complexity of the different astrophysical processes involved in producing the GW background and their vastly different length and time scales lead to great difficulties in constructing a model that can be easily tested against upcoming data. Moreover, it is often challenging to estimate the relative importance of the various uncertainties involved. In this paper we develop a general framework for calculating the GW background from binary compact objects in an astrophysical context. As will be discussed below, many of the ingredients of this calculation are highly uncertain, therefore we tried to construct a modular approach to the problem, allowing to narrow down on one kind of uncertainty at a time. We then apply this approach to inspiraling stellar-mass BBH and binary NSs that have not merged during a Hubble time.
The core of our method is in describing the number density of binary systems in terms of the continuity equation in the space of orbital parametres of the binary. A similar approach was used by Refs. \citep{1994MNRAS.268..841B,1995MNRAS.274..115M,2001MNRAS.327..531I} to study the GW background from high-mass binary pulsars in our Galaxy. In this work we go beyond the steady-state solution assumed in these studies and treat multiple source classes.
This paper is structured as follows: section \ref{sec:Model} describes our synthetic approach: we start with some basic definitions in section \ref{sec:general}. We then define the number densities and formation rates of single and binary compact objects and derive the equations for the evolution of binary orbital parameters in section \ref{sec:evo_nums}. We discuss our complete synthetic model in section \ref{sec:complete_set}. Section \ref{sec:GWback} is an application of our approach to the calculation of several GW backgrounds in the context of a particular astrophysical model. In section \ref{sec:spectrum} we review the GW energy spectrum from inspiraling and merging sources, in section \ref{sec:astro} we outline our astrophysical model and in sections \ref{sec:ins_BH} and \ref{sec:ins_NS} we calculate the GW background from inspiraling and merging stellar-mass BBH and inspiraling binary NSs. We conclude in section \ref{sec:discussion}.
\section{Description of the model}
\label{sec:Model}
\subsection{Gravitational wave background: general definitions}
\label{sec:general}
The spectrum of gravity waves is characterized by the dimensionless density parameter \citep{1999PhRvD..59j2001A}
\begin{equation}
\Omega_{\rm gw}(f)=\frac{1}{\rho_c}\frac{{\rm d}\rho_{\rm gw}}{{\rm d}\ln f}
\end{equation}
where $\rho_c=3H_0^2/8\pi G$ is the critical density of the Universe and $f$ is the frequency measured by the observer. It is related to the frequency at emission $f_e$ by
\begin{equation}
f_e =f(1+z)
\end{equation}
where $z$ is the redshift. For a single class of sources, the energy density of the emitted gravitational waves can be expressed as \citep{2011RAA....11..369R}
$$
\Omega_{\rm gw}(f)=\frac{1}{\rho_c c^3}fF(f),
$$
in terms of the integrated flux of energy received by the observer at frequency $f$,
\begin{equation}
F(f)=\int p(\theta) \frac{{\rm d} {\cal N}(\theta,z)}{{\rm d} z} {\cal F}(f,\theta,z) {\rm d}\theta{\rm d} z
\end{equation}
where $p(\theta)$ is the probability distribution of the parameters of the sources (such as orbital parameters, masses, etc.). The quantity
\begin{equation}
{\cal F}(f,\theta,z) = \frac{1}{4\pi \chi^2(z)}\frac{{\rm d} E_{\rm GW}}{{\rm d} f_e}(\theta, f_e)
\end{equation}
depends on the GW signal emitted by the source ${{\rm d} E_{\rm GW}}/{{\rm d} f_e}$ and $\chi(z)$ is the comoving radial distance. The number of sources with parameters in the range $[\theta,\theta+{\rm d}\theta]$ per unit time and redshift interval is given by
\begin{equation}
\frac{{\rm d} {\cal N}(\theta,z)}{{\rm d} z} =\frac{{\rm d} n}{{\rm d} t}(\theta,z)\frac{{\rm d} V}{{\rm d} z}(z)
\end{equation}
where the comoving volume element is:
\begin{equation}
{\rm d} V(z) =\frac{c}{H_0}\frac{\chi^2(z)}{E(z)}{\rm d}\Omega^2{\rm d} z\:,
\end{equation}
the Hubble parameter is $H(z)=H_0E(z)$ and ${\rm d}\Omega^2$ is the unit solid angle.
The above analysis can be generalized to account for multiple types of sources, such as binary BHs, binary NSs etc. so that the total contribution is obtained by summing over all the components $i$ and their respective internal parameters $\theta_i$:
\begin{align}
& \Omega_{\rm gw}(f)= \nonumber\\ & \frac{1}{\rho_c c^3}f \sum_i
\int \frac{{\rm d} z}{4\pi\chi^2(z)}\frac{{\rm d} V}{{\rm d} z}(z)
\int {\rm d}\theta_i
\frac{{\rm d} n_i}{{\rm d} t}(\theta_i,z)
\frac{{\rm d} E^{(i)}_{\rm GW}}{{\rm d} f_e}(\theta_i, f_e)
\label{eq:Omega_gw}
\end{align}
where the sum is over different types of sources.
The total background thus depends on the following quantities:
\begin{itemize}
\item the nature of each class of sources $i$ and the relevant set of parameters $\theta_i$ (i.e. masses, spins, binary orbital parameters etc);
\item the evolution of the comoving number density of each source per unit time, $\frac{{\rm d} n_i}{{\rm d} t}(\theta_i,z)$;
\item the evolution of the parameters $\theta_i$ with time. We set $\theta_i(z)\equiv\theta_i[\theta_i^{(0)},z,z_0]$. Among the parameters, the metallicity and orbital parameters evolve with time;
\item the probability distribution function (PDF) of $\theta_i$ at formation, i.e. ${\cal P}(\theta_i^{(0)})$ which determines the density distribution at later times $n_i(\theta_i,z)$. Its form will generally depend on both the relation $\theta_i[\theta_i^{(0)},z,z_0]$ and the time of formation $z_0$ of the source;
\item the energy spectrum ${\rm d} E^{(i)}_{\rm GW}/{\rm d} f_e$ emitted by source $i$ with parameters $\theta_i$.
\end{itemize}
These ingredients are described in detail below.
\subsection{Evolution of the number density of binaries}
\label{sec:evo_nums}
The goal of our analysis is to provide a general description of the formation and merger rates of binary systems while accounting for the variation in the binary orbital parameters.
\subsubsection{Formation of the binaries}
We start by modeling the comoving number density of objects of type $X$. Each of these objects is either single or belongs to a binary system. In the latter case its companion can be another object of type $X$ or an object of different type $Y$, where $X$ and $Y$ can be either a BH or a NS. We thus define
\begin{itemize}
\item $n_X(M,t)$ : the total number density of objects of type $X$ from which the total number density is obtained as
$$
\bar n_X(t) = \int n_X(M,t){\rm d} M.
$$
\item $n^{(1)}_X(M, t)$ : the number density of $X$ that are in a single system with mass $M$;
\item $n^{(2)}_X(M_1,M_2,{\bm w}, t)$ : the number density of $X$ that are in a $XX$ binary system with masses $M_1$ and $M_2$ and orbital parameters ${\bm w}$;
\item $n^{(1,1)}_{XY}(M_X,M_Y,{\bm w}, t)$ : the number density of $X$ that are in a $XY$ binary system with masses $M_X$ and $M_Y$ and orbital parameters ${\bm w}$;
\end{itemize}
It is clear from these definitions that
\begin{align}
n_X(M,t) = n^{(1)}_X(M_, t) & + 2\int n^{(2)}_X(M,M_2,{\bm w}, t) {\rm d} M_2{\rm d}^n{\bm w} \nonumber \\ & + \int n^{(1,1)}_{XY}(M,M_Y,{\bm w}, t) {\rm d} M_Y{\rm d}^n{\bm w}.
\end{align}
In order to describe the evolution of such systems, we need to calculate their rates of formation. We define the following rates:
\begin{itemize}
\item $R_X(M,t)$ : the total formation rate of objects of type $X$ with masses $M$ at time $t$;
\item $R^{(1)}_X(M, t)$ : the formation rate of $X$ that are in a single system with mass $M$ at time $t$;
\item $R^{(2)}_X(M_1,M_2,{\bm w}, t)$ : the formation rate of $X$ that are in a $XX$ binary system with masses $M_1$ and $M_2$ and orbital parameters ${\bm w}$;
\item $R^{(1,1)}_{XY}(M_X,M_Y,{\bm w}, t)$ : formation rate of $X$ that are in a $XY$ binary system with masses $M_X$ and $M_Y$ and orbital parameters ${\bm w}$;
\end{itemize}
These rates are clearly related to each other and depend on the chosen physical model of stellar evolution.\\
If we assume for the sake of simplicity that both components in a $XX$ binary system always have equal masses then
$$
n^{(2)}_X(M,M,{\bm w}, t) =n^{(2)}_X(M,M_2,{\bm w}, t) \delta(M-M_2)
$$
so that
\begin{align}
n_X(M,t) = n^{(1)}_X(M_, t) & + 2 \int n^{(2)}_X(M,M,{\bm w}, t) {\rm d}^n{\bm w} \nonumber \\ & + \int n^{(1,1)}_{XY}(M,M_Y,{\bm w}, t) {\rm d} M_Y{\rm d}^n{\bm w}.
\end{align}
We further assume that a fraction $\gamma_X$ of the $X$ component resides in $XY$ systems. We define this ratio by
\begin{equation}
\gamma_X(M,t) R_X(M,t)\equiv \int R^{(1,1)}_{XY}(M,M_Y,{\bm w}, t) {\rm d} M_Y{\rm d}^n{\bm w},
\label{eq:gamma_definition}
\end{equation}
from which it follows that
$$
\gamma_X(M,t) R_X(M,t) = \gamma_Y(M,t) R_Y(M,t).
$$
We can then assume that a fraction $\alpha_X$ are in $XX$ binaries so that
\begin{eqnarray}
R_X^{(1)}(M,t)&=&(1-\alpha_X-\gamma_X) R_X(M,t)\\ \label{eq:RxRy_definition1}
R_X^{(2)}(M,M,{\bm w}, t) &=&\frac{1}{2}\alpha_X R_X(M,t){\cal P}_X({\bm w}),
\label{eq:RxRy_definition2}
\end{eqnarray}
where ${\cal P}_X({\bm w})$ is the PDF of the orbital parameters at the time of formation normalized such that
$$
\int {\cal P}_X({\bm w}){\rm d}^n{\bm w}=1,
$$
and ${\cal P}_{XY}({\bm w})$ is the PDF of the orbital parameters of the hybrid systems. We can then check that
\begin{align}
R_X(M,t) = R^{(1)}_X(M_, t) & + 2\int R_X^{(2)}(M,M,{\bm w}, t){\rm d}^n{\bm w} \nonumber \\ & + \int R_{XY}^{(1,1)}(M,M_Y,{\bm w}, t){\rm d}^n{\bm w}{\rm d} M_Y.
\end{align}
Note also that $R_X^{(2)}$ and $R_{XY}^{(1,1)}$ do not have the same dimensions.
\subsubsection{Evolution of the densities}
We shall now formulate the equations that describe the evolution of the density of binary compact objects in our model. As a first step, let us consider the case with only one species. During the evolution of the binary its orbital parameters are constantly changing due to perturbations or the emission of gravitational waves so that their time evolution can be expressed as
\begin{equation}\label{eq:ee1}
\frac{{\rm d} {\bm w}}{{\rm d} t} = {\bm f}({\bm w}, M)
\end{equation}
where ${\bm f}({\bm w}, M)$ depends on the physical process at work. A merger occurs when ${\bm w}={\bm w}_{\rm merger}$.
At any given time, single objects $X$ of mass $M$ are formed via two routes: direct formation (i.e. SN collapse) with a rate $R_X^{(1)}$ and from the merger of binary systems. In the latter case the final state has a mass $2M-\Delta M$ where $\Delta M$ is a function of $M$ that corresponds to the energy radiated in gravitational waves. We shall define $S(M,M,t)$ as the rate of mergers of binary systems of masses $(M,M)$ at time $t$ and assume that a fraction $(1-\beta_X)$ of the merger products remains single. Binary systems thus form from newly-born objects $X$ at a rate $R_X^{(2)}(M,{\bm w},t)$, as well as from merger remnants with a fraction $\beta_X$. In addition, their orbital parameters evolve according to Eq.~(\ref{eq:ee1}).
These considerations translate into the following set of equations:
\begin{widetext}
\begin{eqnarray}
\frac{{\rm d} n_X^{(1)}(M,t)}{{\rm d} t} &=& R_X^{(1)}(M,t) +(1-\beta_X)S\left(M',M',t\right)\label{evo2}\\
\frac{{\rm d} n_{X}^{(2)}(M,M,{\bm w}, t)}{{\rm d} t} &=& R_X^{(2)}(M,M,{\bm w},t)+\frac{1}{2}\beta_X S\left(M',M',t\right){\cal P}_X({\bm w})- \frac{\partial }{ \partial {\bm w}}.[{\bm f}\left({\bm w},M\right)n_{X}^{(2)} \left(M,M,{\bm w},t\right)]\label{evo3}\\
M&=&2M'-\Delta M (M').\label{evo4}
\end{eqnarray}
\end{widetext}
In Eq.~(\ref{evo3}), the last term describes the evolution of the density of systems in the 2-dimensional space of their orbital parameters. If one thinks of ${\bm f}$ as a velocity and $n_{XX}^{(2)}$ as the density of the fluid, then one recognizes a continuity equation with a sink and a source term. Eq. (\ref{evo3}) relates the mass of the two merging stars $M'$ to the mass of the final state $M$. The source term $S$ due to the mergers takes the form
\begin{equation}
S\left(M',M',t\right)=\int_{C_m} {\bm f}n^{(2)}_X\left(M',M',{\bm w},t\right).{\rm d}{\bm \ell}
\label{eq:source_term}
\end{equation}
where $C_m$ is a contour in the 2-dimensional parameter space around ${\bm w}_{\rm merger}$ so that ${\rm d}{\bm \ell}$ has dimension $w$. $C_m$ characterizes all the systems that merge in a time step.\\
We note that a similar approach which utilizes the continuity equation was taken by Refs. \citep{1994MNRAS.268..841B,1995MNRAS.274..115M,2001MNRAS.327..531I} to study the GW background from high-mass binary pulsars in our Galaxy. In particular, these studies assumed a \emph{steady-state} solution, which leads to a particular distribution of orbital parameters. In the present analysis our goal is to calculate the source distribution on cosmological timescales, and we therefore relax the steady-state assumption and introduce the source and sink terms.\\
In order to include hybrid systems, we need to determine the final state of a $XY$ merger. In the following we assume that it leads to the formation of a single object $X$ (that is, a BH) of mass $M'_X=M_X+M_Y-\Delta M$. The evolution of the $XY$ systems is then similar to Eq.~(\ref{evo3}),
\begin{widetext}
\begin{eqnarray}
\frac{{\rm d} n_{XY}^{(1,1)}(M_X,M_Y,{\bm w}, t)}{{\rm d} t} = R_{XY}^{(1,1)}(M_X,M_Y,{\bm w},t)
-\frac{\partial }{\partial {\bm w}}.[{\bm f}_{XY}\left({\bm w},M_X,M_Y\right)n_{XY}^{(1,1)} ] \left(M_X,M_Y,{\bm w},t\right),
\label{eq:dnxy}
\end{eqnarray}
\end{widetext}
where ${\bm f}_{XY}\left({\bm w},M_X,M_Y\right)$ describes the evolution of the orbital parameters of $XY$, that may be different from $XX$ systems, and where we assume for simplicity that second generation $X$ do not form new $XY$ systems. At merger, we get a source term
\begin{equation}
S_{XY\rightarrow X}\left(M',t\right)=\int_{C_m} {\bm f}_{XY}n^{(1,1)}_{XY}\left(M_X,M_Y,{\bm w},t\right).{\rm d}{\bm \ell}{\rm d} M_Y
\label{eq:sxy}
\end{equation}
with the constraint $M'_X=M_X+M_Y-\Delta M (M_X,M_Y)$. This source has to be added to Eq. (\ref{evo2}) which describes the evolution of $n^{(1)}_X$, assuming the products of a $XY$ merger remain single (see Eq. (\ref{eq:evo1_full}) below).
\subsubsection{Evolution of the orbital parameters}
Compact binaries undergo orbit circularization due to emission of GW. The evolution of the eccentricity $e$ and the semi-major axis $a$ is given by Ref. \citep{PhysRev.131.435}
\begin{eqnarray}
\frac{{\rm d} a}{{\rm d} t} &=& -\frac{64}{5}\frac{G^3\mu m^2}{c^5a^3}\frac{\left(1+\frac{73}{24}e^2 + \frac{37}{96}e^4\right)}{(1-e^2)^{7/2}} \label{eq:orbital_ev}\\
\frac{{\rm d} e}{{\rm d} t} &=& -\frac{304}{15}\frac{G^3\mu m^2}{c^5a^4}\frac{e\left(1+\frac{121}{304}e^2\right)}{(1-e^2)^{5/2}}
\end{eqnarray}
where $m=M_1+M_2$ is the total mass of the binary and $\mu=M_1M_2/M$ is the reduced mass.
Clearly when $e=0$, ${\rm d} e/{\rm d} t=0$ so that a circular orbit remains circular.
The lifetime of a binary system is given by
\begin{equation}
\tau(a_0,e_0) = \frac{5}{256}\frac{c^5a_0^4}{G^3m^2\mu}F(e_0)
\label{eq:lifetime}
\end{equation}
where
\begin{equation}
F(e_0)=\frac{48}{19}\frac{1}{g^4(e_0)}\int_0^{e_0}\frac{g^4(e)(1-e^2)^{5/2}}{e(1+\frac{121}{304}e^2)}{\rm d} e\:
\end{equation}
and
\begin{equation}
g(e)= \frac{e^{12/19}}{1-e^2} \left(1+\frac{121}{304}e^2\right)^{870/2299}.
\end{equation}
A solution of equations (\ref{eq:orbital_ev}) in the case of a $30M_{\odot}-30M_{\odot}$ binary is shown in Fig. \ref{fig:orbParam}. Note the difference in merging timescale: from $0.2$ Gyr for the case $e_0=0.5,a_0=0.1$ AU to more than the age of the Universe for $e_0=0.5$ but $a_0=0.5$ AU. It is clear from these results that the initial distribution of orbital parameters will have a significant influence on the merger rate of binary compact objects.
Note that the evolution of the orbital parameters can be more complex if the binary is embedded in a dense stellar environment. This is the case for stellar-mass BBH formed in globular clusters \citep{2016arXiv160604889A} and SMBHs at sub-parsec separations \citep{2008ApJ...686..432S}. In both these cases the binary may enter the observable frequency band while still having a non-negligible eccentricity.
The solution to eq. (\ref{eq:lifetime}) for several masses ($5M_{\odot}-5M_{\odot}$, $10M_{\odot}-10M_{\odot}$, $30M_{\odot}-30M_{\odot}$ and $50M_{\odot}-50M_{\odot}$ binaries) is shown in Fig. (\ref{fig:merging_time}): the area to the left of each curve indicates the parameter space for which the merger occures within the age of the Universe. The merger time is clearly very sensitive to the initial semi-major axis, but also to the masses and, to a lesser extent, the eccentricity.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=.9\columnwidth]{fig1.eps}
\caption{Evolution of eccentricity for a $30M_{\odot}-30M_{\odot}$ binary with different initial eccentricities and separations. These different initial conditions induce the difference in merging timescales of several orders of magnitude.}
\label{fig:orbParam}
\end{center}
\end{figure}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=.9\columnwidth]{fig2.eps}
\caption{The region in parameter space (eccentricity and semi-major axis) corresponding to \emph{rapid mergers} (i.e. within the age of the Universe) to the left of the solid lines and \emph{slow mergers} (i.e. taking longer than the age of the Universe) to the right of the solid lines. Each line corresponds to a different mass of the BBH components, as indicated in the legend.}
\label{fig:merging_time}
\end{center}
\end{figure}
\subsection{Complete synthetic model of the evolution of binary compact objects}
\label{sec:complete_set}
Our model framework is shown schematically in Figure~\ref{fig-1}. The basis for our calculation is a cosmic evolution model that follows the growth of galaxies, in particular their stellar and gaseous components, including the evolution of gas-phase metallicity $Z$. An underlying stellar evolution model describes the fate of massive stars based on their initial mass and metallicity and predicts whether they would form a BH or a NS at the end of their life. We note that other parameters, such as stellar rotation, are expected to strongly influence the evolution of massive stars and thus the mass of the BH or NS that forms. A certain fraction of these objects belong to binary systems, as shown on Figure~\ref{fig-1}. These systems emit gravitational radiation and may experience interactions with their surroundings, and as a result their orbits can shrink and they merge (as indicated by the red arrows). All the mergers we consider are assumed to lead to a formation of a BH. We can now formulate the set of differential equations that govern the evolution of this system using eqs. (\ref{eq:ee1})-(\ref{eq:sxy}).
We consider all three types of binary objects, namely binary NSs, BHs and BH-NS, where the evolution of the orbital parameters of each type is governed by eq. (\ref{eq:ee1}). The evolution of the number densities of binary NSs and BHs are given by Eqs. (\ref{evo2})-(\ref{evo4}) with the source terms provided by eqs. (\ref{eq:source_term}) and (\ref{eq:sxy}) while the evolution of the hybrid population NS-BH is provided by Eq. (\ref{eq:dnxy}).
\begin{figure}
\begin{center}
\includegraphics[width=1.\columnwidth]{fig3.eps}
\caption{A schematic representation of our model framework. Stars form from interstellar gas with metallicity $Z$ as described in an underlying model of galactic evolution. Massive stars end their lifes as BHs or NSs, according to their initial mass and metallicity. We assume that part of these objects form binary systems: $BH-BH$, $NS-NS$ and $BH-NS$. Gravitational waves are produced during the inspiral phase and merger of binary systems prior to their collapse into a BH (marked by red arrows) and during the core collapse of massive stars. The latter also affect the metallicity of the interstellar gas.}
\label{fig-1}
\end{center}
\end{figure}
The resulting equations (\ref{Fevo1}-\ref{Fevo4}) can be found in the Appendix. Their solution provides the evolution of the number densities of single and binary compact objects under a specific set of astrophysical assumptions. We note that these equations are generic in the sense that the processes driving this evolution are not implicit in the formulation. For example, the evolution of the orbital parameters can be primariliy driven by the emission of GW, as we assume below, but can also be influenced by interactions with the surrounding medium. This influence can be taken into account by an introduction of an appropriate function ${\bm f}({\bm w}, M)$ describing this evolution. Similarly, the formation rates should be obtained from specific astrophysical models.
\section{Computation of the gravitational wave background}
\label{sec:GWback}
\subsection{Composite spectrum}
\label{sec:spectrum}
An inspiraling system of compact binaries in an elliptical Keplerian orbit emits gravitational radiation in a discrete set of harmonics of the orbital frequency $f_n=nf_0$ for $n \geq 1$ where $(2\pi f_0)^2=Gm/a^3$. Note that, as we have seen, $a$ (and hence $f_0$) evolves with time due energy losses to gravitational wave emission. Instanteneously, the power emitted by the inspiraling binary at frequency $f_n$ is \citep{PhysRev.131.435}:
\begin{equation}
\left. \frac{dE}{dt}\right\vert_n = \frac{32G^4 \mu^2 m^3}{5c^5 a^5}g(n,e)
\end{equation}
where
\begin{equation}
g(n,e) = \frac{n^6}{96 a^4}\left[A^2_n(e)+B^2_n(e)+3C^2_n(e)-A_n(e)B_n(e) \right]\:.
\end{equation}
The functions $A_n,B_n,C_n$ are given in terms of Bessel functions \citep{GravBookMaggiore}:
\begin{eqnarray}
A_n(e) = & \frac{a^2}{n}\lbrace J_{n-2}(ne) -J_{n+2}(ne)\nonumber \\ & -2e\left[J_{n-1}(ne)-J_{n+1}(ne) \right] \rbrace
\end{eqnarray}
\begin{equation}
B_n(e) = \frac{(1-e^2)a^2}{n}\lbrace J_{n+2}(ne) - J_{n-2}(ne) \rbrace
\end{equation}
\begin{eqnarray}
C_n(e) = & \frac{\sqrt{1-e^2}a^2}{n}\lbrace J_{n+2}(ne) + J_{n-2}(ne) \nonumber \\ & - e\left[J_{n+1}(ne) - J_{n-1}(ne) \right] \rbrace .
\end{eqnarray}
The total emitted power is then given by the sum over all the harmonics:
\begin{equation}
\frac{dE}{dt}=\sum_{n=1}^{\infty}\left.\frac{dE}{dt}\right\vert_n\:.
\label{eq:dedt}
\end{equation}
The energy per unit \emph{emitted} frequency $f$ (we omit the subscript in this section) by a system with orbital parameters $(a,e)$ can be expressed as:
\begin{equation}
\frac{dE}{df} = \frac{dE}{dt}\left(\frac{df_0}{dt} \right)^{-1}\frac{df_0}{df}\:.
\label{eq:dedf1}
\end{equation}
The orbital frequency evolves with time as $df_0/dt=-(3f_0/2a)df/dt$ where
$da/dt$ is given by Eq. (\ref{eq:orbital_ev}). Furthermore, $df_0/df=1/n$ for $f=nf_0$ and $0$ otherwise. Combining Eqs. (\ref{eq:dedt}) and (\ref{eq:dedf1})
and defining:
\begin{equation}
\delta(x,y)=
\begin{cases}
1, & \text{if}\ x=y \\
0, & \text{otherwise}
\end{cases}
\end{equation}
we obtain the following expression for the energy spectrum:
\begin{equation}
\frac{dE}{df}=\frac{64G^4\mu^2 m^3}{15c^5 a^4 f_0}\sum_{n=1}^{\infty}\frac{g(n,e)}{n}\left|\left(\frac{da}{dt} \right)^{-1} \right|\delta\left(\frac{f}{f_0},n\right) \:.
\end{equation}
Finally, using Eq. (\ref{eq:orbital_ev}), the usual definition of the chirp mass $M_c=m^{2/5}\mu^{3/5}$ and $(2\pi f_0)^2=Gm/a^3$ we obtain:
\begin{widetext}
\begin{eqnarray}
\frac{dE}{df}(a,e,M_c) =\frac{(2\pi)^{2/3} (GM_c)^{5/3}}{3G}\frac{(1-e^2)^{7/2}}{\left(1+\frac{73}{24}e^2 + \frac{37}{96}e^4 \right)}\frac{1}{f_0^{1/3}} \sum_{n=1}^{\infty}\frac{g(n,e)}{n}\delta\left(\frac{f}{f_0},n\right)
\label{eq:dedf_ecc}
\end{eqnarray}
\end{widetext}
where $f$ is the emitted frequency and the notation emphasizes that the spectrum is a function of the \emph{instantaneous} orbital parameters $(a,e)$ and the chirp mass of the binary (Eq. (\ref{eq:dedf_ecc}) is equivalent to Eq. (15) of Ref. \citep{2015MNRAS.449.2700E}). The sum in Eq. (\ref{eq:dedf_ecc}) can be approximated by an analytical expression (Ref. \citep{2015PhRvD..92f3010H} and see also \citep{2007PThPh.117..241E}).
In the case of a circular orbit $g(n,e=0)=\delta(2,n)$ and we recover the familiar expression:
\begin{equation}
\frac{dE}{df}=\frac{\pi^{2/3}(GM_c)^{5/3}}{3G}\frac{1}{f^{1/3}} \:.
\end{equation}
The energy spectrum is therefore given by a \emph{frequency comb}, which becomes more and more sharply peaked around $n=2$ as the eccentricity decreases. Note also that as the semi-major axis shrinks, the orbital frequency $f_0$ grows and the whole spectrum is shifted to higher frequencies.
Eq. (\ref{eq:dedf_ecc}) describes the inspiraling phase of binaries on eccentric orbits and is valid only up to some frequency $f_{merger}$. Since by this time the orbit had already circularized, we use the expressions from Ref. \citep{2008PhRvD..77j4017A} for the merger and ringdown phases to describe this stage of the binary evolution:
\begin{equation}
\frac{\textrm{d}E}{\textrm{d}f}=\frac{(G\pi)^{2/3}M_c^{5/3}}{3}
\begin{cases}
f^{-1/3},& f\leq f_1\\
\omega_1 f^{2/3}, & f_1 < f < f_2\\
\omega_2 \left(\frac{f}{1+\left(\frac{f-f_2}{\sigma/2} \right)^2} \right)^2, & f_2 \leq f\leq f_3
\end{cases}
\label{eq:dedf_circ}
\end{equation}
The set of
parameters $(f_1,f_2,f_3,\sigma)$, where $f_1,f_2$ correspond to the end of the
inspiral and merger phases, respectively, is taken from Ref.
\citep{2008PhRvD..77j4017A} for the case of non-spinning BHs for each set of
masses. The constants $\omega_1=f_1^{-1}$
and $\omega_2=f_1^{-1}f_2^{-4/3}$ are chosen to make
$\textrm{d} E_{\textrm{gw}}/\textrm{d} f$ continuous.
Eqs. (\ref{eq:dedf_ecc}) and (\ref{eq:dedf_circ}) describe the spectrum of the gravitational wave background as a function of the chirp mass and the orbital parameters for eccentric and circular orbits, respectively. In order to use them in Eq. (\ref{eq:Omega_gw}) we also need to account for the comoving number density of the different sources and their orbital parameter distribution. These quantities need to be computed in the context of an astrophysical model of stellar formation and evolution. Below we describe one particular model, but we stress that our formalism can be applied to a wider class of astrophysical prescriptions.
\subsection{Astrophysical model}
\label{sec:astro}
The birthrate of black holes and neutron stars as a function of mass and redshift (or, equivalently, time) $R_{X}(t,m_{X},{\bm w})$ (in units of events per unit time per unit comoving volume per unit eccentricity per unit semi-major axis) is given by \citep{2016arXiv160404288D}:
\begin{widetext}
\begin{eqnarray}
R_{X}(t,m_{X},{\bm w})=\int\psi[t-\tau(m)]\phi(m)\delta(m-g_{X}^{-1}(m_{X})){\cal P}_X({\bm w})\textrm{d}m
\label{eq:birthrate}
\end{eqnarray}
\end{widetext}
where $\tau(m)$ is the lifetime of a star of mass $m$, $\phi(m)$ is the stellar initial mass function (IMF),
$\psi(t)$ is the cosmic star formation rate (SFR) and $\delta(m)$ is the Dirac delta distribution. ${\cal P}_X({\bm w})$ is the PDF of the orbital parameters at birth. The initial stellar
mass and BH/NS masses are related by the function $m_{X}=g_{X}(m)$ which is implicit in the equation above.
Note that $g_{X}$ also depends on time through its metallicity dependence $Z(t)$. We use the galaxy and stellar evolution models described in Ref. \citep{2016arXiv160404288D} and briefly discuss them below.
We start with a description of the IMF and the SFR. We assume a Salpeter IMF with slope $x=2.35$ in the mass range $0.1-100M_{\odot}$:
\begin{equation}
\phi(m) = \frac{dN}{dm} = Am^{-x}
\end{equation}
where $A$ is a normalization constant. We use the functional form of Ref. \citep{2003MNRAS.339..312S} for the SFR:
\begin{equation}
\psi(z)=\nu\frac{a\exp[b(z-z_m)]}{a-b+b\exp[a(z-z_m)]}
\end{equation}
where $z$ is the redshift. Our fiducial model is a fit to the observations of luminous galaxies compiled by Ref. \citep{2013ApJ...770...57B} and complemented by high-redshift observations from Ref. \citep{2011ApJ...737...90B}. We use the fit parameters given in Ref. \citep{2015MNRAS.447.2575V}, namely $\nu=0.178$ $M_{\odot}$yr$^{-1}$Mpc$^{-3}$, $z_m=2$, $a=2.37$ and $b=1.8$. The evolution of the metallicity of interstellar matter is calculated using the cosmic chemical evolution model described in Refs. \citep{2015MNRAS.447.2575V,2004ApJ...617..693D,2006ApJ...647..773D}. We use metal yields from Ref. \citep{1995ApJS..101..181W} and stellar lifetimes from Ref. \citep{2002A&A...382...28S}. This model reproduces the metallicity evolution of high-redshift damped Ly-$\alpha$ absorbers \citep{2012ApJ...755...89R} as discussed in \citep{2015MNRAS.447.2575V} and is consistent with the optical depth to reionization recently measured by the \emph{Planck} Collaboration \citep{2016arXiv160502985P,2016arXiv160503507P}.
The function $g_{X}(m)$ defines the mass of the BH formed from a star with an initial mass $m$ and in general depends on the metallicity of the star \citep{2008NewAR..52..419V} and its rotation \citep{2009A&A...497..243D,2016A&A...588A..50M,2016MNRAS.458.2634M} which determine the amount of mass loss the star experienced before reaching the core collapse phase. If the star belongs to a binary system its evolution strongly depends on its companion and their possible mass exchange, in particular the common envelope phase. Furthermore, the star, single or in a binary, may belong to a dense star cluster, in which case in can be influenced by dynamical interactions with other stars \citep{2014MNRAS.441.3703Z,2016arXiv160604889A}. In this work we make the simplifying assumption that $g_{X}$ depends only on initial stellar mass and metallicity and use the function computed for models of isolated stellar evolution. We stress, however, that all other dependencies can in principle be expressed by a more general $g_{X}$. In this work we use the models of Ref. \citep{2012ApJ...749...91F} to obtain $g_{X}$, in particular their \emph{delayed} model, as described in Ref. \citep{2016arXiv160404288D}.
The solution to Eqs. (\ref{eq:ee1})-(\ref{evo3}) depends on the initial distribution of the orbital parameters $\textbf{w}=(a,e)$ of the newborn binary system. Unfortunately, there are no direct observations of this quantity. Using observations of massive (O type) stars (BH/NS progenitors) in Galactic stellar clusters, \citep{2012Sci...337..444S} deduced a binary fraction of about $0.5$ and estimated the PDFs of orbital periods, mass ratios and eccentricities. In particular, they found a strong preference for small separations, so that the orbital periods were distributed as $P(\log T)\propto (\log T)^{-0.55}$. We stress, however, that these distributions refer to massive stars and not to binary compact objects, for which the initial separations are expected to be much larger. Indeed, it can be seen from Fig. 2 of Ref. \citep{2012Sci...337..444S} that binaries on very close orbits are expected to merge while still on the main sequence, before forming two separate compact objects. For the purposes of this work we adopt the power-law PDFs deduced by Ref. \citep{2012Sci...337..444S} and keep the lower and upper values as free parameters. We model the joint PDF as the product ${\cal P}_X({\bm w})=P(e)P(a)$ and assume the following distribution of initial eccentricities in the range $e\in [0,1]$:
\begin{equation}
P(e) \propto e^{\kappa}
\label{eq:Pe}
\end{equation}
with $\kappa=-0.42$ \citep{2015ApJ...814...58D} and orbital periods $T$ (in days) for $\log T\in [\log T_{\textrm{min}},\log T_{\textrm{max}}]$:
\begin{equation}
P(\log T) \propto (\log T)^{-0.5}\:.
\label{eq:Pa}
\end{equation}
For the calculations shown below we chose a minimal orbital period corresponding to a semi-major axis of $a_{min}$ in the range $0.1-0.3$ AU and a maximal period corresponding to a semi-major axis of $a_{max}=5000$ AU. Note that while $a_{max}$ has little effect on the results, $a_{min}$ affects the merger rate through eq. (\ref{eq:lifetime}). We chose a range of $a_{min}$ that brackets the uncertainty in the merger rate estimated from Advanced LIGO observations \citep{2016arXiv160203842A}, as can be seen in Figure \ref{fig:gw_bbh_all} and discussed below.
Eq. (\ref{eq:birthrate}) describes the birthrate of neutron stars or black holes under the assumptions outlined above. In order to follow the evolution with time of the number density of \emph{binary} systems we need to account for the fraction of compact objects that form binaries with a given set of orbital parameters and the time evolution of those orbital parameters, given by eqs. (\ref{eq:ee1})-(\ref{evo3}). In the following we assume that all the black holes, as well as all the neutron stars are born in binaries, and that there are no BH-NS binaries. This corresponds to a choice of parameters $\alpha_{BH}=\alpha_{NS}=1$ and $\gamma=0$ in Eqs. (\ref{eq:gamma_definition})-(\ref{eq:RxRy_definition2}). Furthemore, we assume that BHs that formed from a merger of any kind remain single and do not merge with another BH, so that $\beta_{BH}=0$ in eqs. (\ref{evo2})-(\ref{evo3}). We then solve Eqs. (\ref{eq:ee1})-(\ref{evo3}) with the birthrates taken from eq. (\ref{eq:birthrate}) to obtain $n_i(M_{c,i},\textbf{w}_i,z_i)$, the number density of binaries of a given type per unit comoving volume per unit mass per unit $\textbf{w}$, where $\textbf{w}=(a,e)$ signifies the orbital parameters of the binary. The time derivative of this quantity is then used in Eq. (\ref{eq:Omega_gw}) to calculate the gravitational wave background. For simplicity we neglect the energy radiated in GW, setting $\Delta M = 0$ in Eq. (\ref{evo4}). We assume that a binary merges immediately (i.e. in less than one timestep) when the separation is $a<100 R_{*}$ where $R_{*}=2GM/c^2$. For the solution of Eqs. (\ref{eq:ee1})-(\ref{evo3}) we use a grid in the $(e,a)$ space with the corresponding resolution of $\Delta e = 0.025$ and $\Delta \log \tilde{a} = 0.13$ where $\tilde{a}=a/R_{*}$. The timestep was $300$ Myr.
We shall now present the gravitational wave background from inspiraling binary BHs and NSs.
\subsection{Inspiraling and merging binary black holes}
\label{sec:ins_BH}
In this section we will treat two types of sources: inspiraling BBH, i.e. binaries that have not merged during the Hubble time and merging BBH. Their number densities are provided by the solution of Eqs. (\ref{eq:ee1})-(\ref{evo3}) assuming the initial distibution of orbital parameters given by Eqs. (\ref{eq:Pe})-(\ref{eq:Pa}). It can be seen from Eq. (\ref{eq:Pe}) that a large fraction of these binaries are expected to have relatively eccentric orbits, and therefore contribute gravitational radiation in a range of frequencies well above their orbital frequency.
The source density of inspiraling BBH ${{\rm d} n_i}/{{\rm d} t}(M,z)$ (integrated over the orbital paremeters) for our chosen astrophysical model is shown in Figure \ref{fig:astroSourceRates} for $3$ different redshifts. This quantity is the solution of Eqs. (\ref{eq:ee1})-(\ref{evo3}) under the assumptions outlined above and represents the entire population of BBH. The population builds over time, as expected, with the number density of BBH with masses around $\sim 20-25M_{\odot}$ growing considerably faster than that of other masses. The reason for this rapid growth is the evolution in metallicity: whereas at low metallicities stars above $\sim 30M_{\odot}$ undergo direct collapse in the model of Ref. \citep{2012ApJ...749...91F}, these same stars are less efficient in producing massive BHs at higher metallicities and instead end up as BHs with masses around $\sim 20-25M_{\odot}$. This build-up of BHs around $\sim 20-25M_{\odot}$ corresponds to the stagnation in the number density of BHs with $M>30M_{\odot}$.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=.9\columnwidth]{fig4.eps}
\caption{The source density ${{\rm d} n_i}/{{\rm d} t}(M,z)$ for our chosen astrophysical model as a function of BH (source frame) mass at $3$ different redshifts. The population builds over time with the number density of BBH with masses around $\sim 20-25M_{\odot}$ growing considerably faster than that of other masses due to transition of stars above $\sim 30M_{\odot}$ from direct collapse, which occures at low metallicities, to SN explosion (and associated lower remnant masses) at higher metallicities.}
\label{fig:astroSourceRates}
\end{center}
\end{figure}
The energy spectra of inspiraling and merging binaries are given by Eqs. (\ref{eq:dedf_ecc}) and (\ref{eq:dedf_circ}), respectively while Eq. (\ref{eq:Omega_gw}) can be simplified to:
\begin{widetext}
\begin{eqnarray}
\Omega_{\rm gw}(f)&=&\frac{f}{\rho_c c^2 H_0}
\int \frac{{\rm d} z}{E(z)}
\int {\rm d}\textbf{w} \int {\rm d} M_c
\frac{{\rm d} n_i}{{\rm d} t}(M_c,\textbf{w},z)
\frac{{\rm d} E^{(i)}_{\rm GW}}{{\rm d} f_e}(M_c,\textbf{w}, f_e)
\end{eqnarray}
\end{widetext}
Figure \ref{fig:gw_bbh_all} shows the full energy spectrum resulting from BBHs assuming $a_{\textrm{min}}=0.2$ AU (red line). The low-frequency part of the spectrum is due to \emph{inspiraling} binaries and exhibits a characteristic peak at $f\sim 10^{-5}$ Hz which corresponds to typical separations at which the binary merges within a Hubble time or less. In contrast, high frequencies accessible with Advanced LIGO and VIRGO are dominated by the signal from merging binaries. Unfortunately, the transition frequency is beyond the reach of current and planned GW observatories. We stress, however, that the signal from inspiraling stellar-mass BBH, if detected, will provide important constraints on the population of BBH, complementary to the information that can be obtained from merging systems alone.
The same astrophysical model was used in Ref. \citep{2016arXiv160404288D} to calculate the GW background from merging BBH, where the merger rate was calibrated (uniformly for all masses) to the rate based on the observation of GW150914 \citep{2016arXiv160203842A} of $1.02^{+1.98}_{-0.79} 10^{-7}$ Mpc$^{-3}$yr$^{-1}$ (black lines in Figure \ref{fig:gw_bbh_all}).
The red band in Figure \ref{fig:gw_bbh_all} corresponds to the range $a_{\textrm{min}}=0.15-0.3$ AU and was chosen so as to reflect this uncertainty in the merger rate. In fact, the main uncertainty in our calculation stems from the unknown initial distribution of the orbital parameters which affects the merger rate. Note that below $f\sim 10^{-5}$ Hz this uncertainty does not affect the signal which is due entirely to inspiraling systems.
The amplitude of the peak at the transition frequency $f\sim 10^{-5}$ Hz is determined by the lifetime of the binaries (in other words, the merger rate) and by their distribution of orbital parameters. In particular, since the energy spectrum of an inspiraling binary represents a \emph{frequency comb}, as can be seen from Eq. (\ref{eq:dedf_ecc}), only binaries with the corresponding separations contribute to a given frequency.
We also show in Figure \ref{fig:gw_bbh_all} the expected sensitivities of Advanced LIGO, Cosmic Explorer, eLISA and DECIGO observatories, all of which should be able to detect the stochastic background from merging stellar-mass BBH.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=1.\columnwidth]{fig5.eps}
\caption{GW background from inspiraling and merging binary black holes with $a_{\textrm{min}}=0.2$ AU (red line), the range $a_{\textrm{min}}=0.15-0.3$ AU (red shaded area) and the GW background from merging BBH from Ref. \citep{2016arXiv160404288D} where the merger rate was normalized to the observed value (black lines). Also shown are the expected sensitivities of Advanced LIGO during observing run O5 \citep{2016arXiv160203847T} (solid magenta line), eLISA \citep{2016PhRvL.116w1102S} (solid blue line), DECIGO \citep{2015MNRAS.449.1076P} (dashed blue line) and the Cosmic Explorer \citep{2015PhRvD..91h2001D} (dashed magenta line). The last two curves were estimated from the expected strain sensitivities.}
\label{fig:gw_bbh_all}
\end{center}
\end{figure}
\subsection{Inspiraling binary neutron stars}
\label{sec:ins_NS}
Similarly to BBH, binary NSs also produce a GW signal during the inspiraling phase. We calculated the contribution from binary NSs in our model, assuming for simplicity a fixed mass for all NSs of $m=2M_{\odot}$. The result, shown in Figure \ref{fig:gw_bns}, is even weaker than in the case of BBH and, as expected, shifted to lower frequency. The signal we compute is much weaker than in Ref. \citep{2015MNRAS.449.2700E}, although note that they introduced a sharp frequency cutoff by hand and also did not use a physically motivated distribution of eccentricities. We do not expect this signal to be detectable with PTA as it is many orders of magnitude below the GW background predicted for merging SMBH.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=.9\columnwidth]{fig6.eps}
\caption{Gravitational wave background from inspiraling binary neutron stars that did not merge during the Hubble time.}
\label{fig:gw_bns}
\end{center}
\end{figure}
\section{Discussion}
\label{sec:discussion}
In this paper we developed a synthetic approach that allows to model the evolution of binary compact objects and the GW background they produce. We described the evolution of the number density of binaries and their interactions in the space of orbital parameters with a set of continuity equations providing also for a source and sink terms due to formation and mergers of binary systems. While we used a specific astrophysical model to estimate the GW background from inspiraling and merging BBH, our approach is modular and any ingredient can be updated or tested against competing models. In particular, we can study the rate of formation of compact objects resulting from different astrophysical models, the rate of binary formation, the distribution of initial orbital parameters and the time evolution.
We use our approach to calculate for the first time the GW background from inspiraling BBH. The signal we predict is very weak and, moreover, is not in the frequency range of any of the current or planned GW observatories. Nevertheless, the characteristic shape of the transition from inspiral to merger dominated signal might provide very interesting constraints on the entire population of BBH. Since the majority of BBH are not expected to merge within a Hubble time and are thus beyond the observational capabilites of ground-based interferometers such as Advanced LIGO and VIRGO, the signal we predict offers a unique handle on the properties of this population.
We stress that this work does not include an exhaustive treatment of the various astrophysical effects, such as BH production mechanisms, the effects of rotation, binary co-evolution and possible influence of dense environments. Our treatment of the various GW backgrounds is also far from complete, as we did not include the contribution from SN collapse nor merging NSs. Moreover, we assumed that all the binaries consist of equal-mass objects and did not calculate the background due to inspiraling and merging BH-NS binaries. An extensive study of these topics and the estimate of the associated uncertainties are left for future work.
Finally, we expect this framework to be useful for different classes of GW sources not discussed here, such as SMBH binaries during the later stages of the merger. The signal from inspiraling SMBH which take longer than the age of the Universe to merge falls in the frequency range accesible with PTA ($10^{-9}-10^{-8}$ Hz), while merging SMBH will be observable with eLISA. The evolution of the eccentricity of these systems may be affected by their environment and some of the SMBH binaries may enter the observable frequency band while still retaining a non-negligible eccentricity which will have an imprint on the GW background \citep{2007PThPh.117..241E,2008ApJ...686..432S,2016ApJ...817...70T}. These questions can be treated within the formalism described in this paper and we plan to study them in future work.
\section*{Acknowledgments}
This work has been done within the Labex ILP (reference ANR-10-LABX-63), part of the Idex SUPER, and received financial state
aid managed by the Agence Nationale de la Recherche, as part of the programme Investissements d'avenir under the reference ANR-11-IDEX-0004-02.. The work of ID and JS was supported by the ERC Project No. 267117 (DARK) hosted by Universit\'{e} Pierre et Marie Curie (UPMC) - Paris 6, PI J. Silk. JS acknowledges the support of the JHU by NSF grant OIA-1124403.
|
\section{Introduction}
In general, the description of quantum mechanical scattering becomes simpler when the relative kinetic energy between the collision partners is negligibly small. Specifically, in this limit the de Broglie wavelength of relative motion greatly exceeds the length scale over which the two particles exert forces on one another, rendering details of their interaction, if not insignificant, at least fairly simple to account for~\cite{LandauQuantum}. This concept was first articulated by Enrico Fermi in the context of collisional broadening of spectral lines in a gas of Rydberg atoms \cite{Fermi34_NC}; was instrumental in the understanding of low-energy scattering of neutrons from nuclei \cite{Bethe}; and has found expression in modern times as the bedrock upon which our understanding of ultracold gases is based \cite{Chin10_RMP}. Ordinarily, the key aspect of this simplicity is that the scattered state contains only a single (or very few) eigenstate(s) of angular momentum, i.e., one or only a few partial waves. Ultracold identical bosons have a differential cross
section which is isotropic and independent of the collision energy such that a single length scale, referred to as the scattering length, suffices for its description. The situation is only marginally more complicated for ultracold identical fermions in which the differential cross section is $\propto k^4|\hat{k}\cdot\hat{k}'|^2$, where $k$ is the magnitude of the relative wave vector and $\hat{k}$ ($\hat{k}'$) is a unit vector along the direction of relative incoming (outgoing) momentum~\cite{TaylorBook}. Once again, a single scalar quantity, the scattering volume, sufficiently characterises the underlying potential. These simple results dictate many important dynamical properties of quantum gases such as the efficiency of evaporative cooling~\cite{PhysRevLett.80.3419,PhysRevA.59.1500,DeMarcoJin1999_PRL,DeMarcoJin_Science1999} and the speed of sound~\cite{Pethick2002}.
However, this situation is radically altered when the potential energy between the collision partners does not decay suitably fast. If the interaction potential decays as $1/r^n$, where $r$ is the distance separating the two particles, then for $n\leq3$ a different scenario emerges~\cite{Shakeshaft_r3potential,DipolarCollisions_OMalley}. Ultracold gases containing atoms or molecules which possess magnetic~\cite{Chromium_BEC,Dysprosium_BEC,Erbium_BEC,Dysprosium_DegenerateFermiGas,Erbium_FermiGas,Chromium_FermiGas} or electric~\cite{KRb,RbCs,NaK,NaRb} dipole moments therefore provide a magnificent opportunity to observe a novel regime of low-energy quantum scattering. When the dipoles are aligned along a chosen axis by an externally applied field, the overall magnitude of the cross section is still determined by a characteristic length scale, the dipole length, which depends on the dipole moment~\cite{Bohn09_NJP}. However, the dipole alignment direction and the $1/r^3$ asymptotic decay of the interaction potential conspire to create a differential cross section which now involves a very large set of partial waves in an essential way due to the coupling between different angular momentum states. Moreover, it
presents a novelty in that such a differential cross section depends explicitly on the relative momenta {\it before} the collision (as well as after) with respect to the polarization axis of the dipoles~\cite{Bohn14_PRA}. The consequence of this dependence has been measured indirectly through its effect on the equilibration rate of a dipolar gas, which is taken out of equilibrium by diabatically changing the trap along a certain direction~\cite{Aikawa14_preprint,TangSykes2015_PRA}. It was seen that the equilibration rate can vary by as much as a factor of four, depending on the angle between the dipole alignment direction and the dynamic axis of the trap.
\begin{figure}[t!]
\centering
\includegraphics[width=\columnwidth]{schematicv2.png}
\caption{(a) Schematic of BEC lattice diffraction used to produce scattered halos. Two 741-nm beams propagating along $\hat{x}+\hat{y}$ and $\hat{y}-\hat{x}$ produce a lattice along $\hat{x}$ which undergoes a precisely calculated two-pulse sequence to optimally split the BEC in half, with each piece travelling along $\pm\hat{x}$. The 3D scattered halo is measured via absorption imaging along a $\hat{y}$-projection. The dipole alignment is set by a bias magnetic field $B$. (b) and (c) 3D scattered halos and projections with the bias field along $\hat{y}$ and $\hat{z}$, respectively. (d) 2D projection along $\hat{y}$ of the differential cross section calculated from Eq.~(\ref{EQdifferential}) with dipole alignment $\hat{\epsilon}$ along $\hat{x}$. (e) and (f) Same as (d), but with dipole alignment $\hat{\epsilon}$ along $\hat{y}$ and $\hat{z}$, respectively. The 2D projections presented in (d)-(f) correspond to the halos in panels (a)-(c), respectively.}
\label{FIG:schematic}
\end{figure}
We report a direct measurement of the differential scattering cross section of $^{162}$Dy, which has an exceptionally large magnetic moment of 9.93$\mu_{\rm B}$, where $\mu_{\rm B}$ is the Bohr magneton. We obtain this measurement by colliding two Bose-Einstein condensates (BECs) at a relative velocity significantly greater than the width of their respective internal velocity distributions. Furthermore, most of these measurements are made in the dilute limit where the majority of particles pass through the opposing condensate without experiencing a collision.
In this situation, scattered particles emerge on a spherical halo defined by the constraints of conserved energy and momentum. The angular distribution of the particles on the sphere is almost entirely determined by the differential cross section.
We vary the dipole alignment direction relative to the collision axis and take absorption images of the post-collision number density after a long time-of-flight (TOF). To obtain a detailed theoretical understanding of these images, we perform Monte-Carlo simulations, which capture the essential features of the experiment.
Related experiments on bosonic alkali gases which also observed halo-like structures due to atomic collisions were reported in Refs.~\cite{swaveHalo_1995,ImpurityScattering_2000,sanddwaveHalo_2004_Niels,sanddwaveHalo_2004_Jook} and recently in fermionic $^{40}$K~\cite{Genkina:2016cb,Niels_Nature}. Remarkably beautiful experiments were reported in Ref.~\cite{PairCorrelations_MetastableHelium} in which metastable helium was employed. In addition to the halo structure, the internal energy of metastable helium allowed for pair correlations to be measured via spatially and temporally resolved single-atom counting. This stimulated theoretical interest in the strength of such correlations~\cite{PairCorrelations_Theory1,PairCorrelations_Theory2,PairCorrelations_Theory3}. All effects from dipolar interactions are negligible in these experiments involving alkali and He gases, and the reported halo is well described by a very small set of partial waves. A unique experiment reported in Ref.~\cite{Spielman_SyntheticPartialWaves} studied the creation of artificial partial waves in $^{87}$Rb BEC collisions mediated by an optical potential which modifies the constraints due to energy and momentum conservation. Using this approach they were able to engineer differential scattering with several partial waves, in spite of the fact that a bare collision would have been entirely $s$-wave. Our work presents a natural and intriguing extension of these experiments, demonstrating the nature of differential scattering in the presence of dipolar interactions which are anisotropic and unavoidably couple to a large number of angular momentum states during the collision.
\section{Experimental details}
We produce a single BEC of $^{162}$Dy spin polarized in the $J=8$, $m_J = -8$ absolute ground state with $6.3(4)\times10^{4}$ atoms as described in Ref.~\cite{Tang2015}. An optical lattice is used to diffract the condensate into the $2n\hbar \mathbf{k}_{L}$ diffraction orders, where the lattice wave vector $\mathbf{k}_{L}=2\pi\sin(\theta/2)/\lambda$ depends on the lattice lasers wavelength $\lambda$ and alignment angle $\theta$, and $n$ is an integer which labels the different diffraction orders. The optical lattice is formed by two fiber-coupled beams derived from the same Ti:Sapphire laser with $\lambda=741$~nm that propagate along $\hat{x}+\hat{y}$ and $\hat{y}-\hat{x}$ such that $\theta=\pi/2$, as shown in Fig.~\ref{FIG:schematic}(a). Both beams are linearly polarized along $\hat{z}$.
The condensate is split into the $\pm2\hbar \mathbf{k}_{L}$ diffraction orders with high efficiency by using a precisely timed two-pulse sequence~\cite{Wu2005}. Thus, two condensates are produced and collide with relative momentum $\mathbf{p}_{\mathrm{rel}}=4\hbar {k}_{L}\hat{x}$.
Immediately after applying the lattice grating, all trapping potentials are removed, and the two condensates collide and expand. After a 22-ms TOF, a column-integrated density profile of the atoms in the $\hat{x}$--$\hat{z}$ plane is measured via absorption imaging along $\hat{y}$. We manipulate the dipole alignment direction by applying a bias magnetic field relative to both the collision axis and the imaging axis. In all configurations, the magnetic field magnitude is held at 1.58~G and away from any Feshbach resonance~\cite{Dy_FeshbachSpectrum}. Three instructive cases are shown in Fig.~\ref{FIG:schematic}, which provides a schematic of our experiment. The differential cross section is isotropic when the dipole alignment is parallel to the relative momentum, as in Fig.~\ref{FIG:schematic}(a). However, with the dipole alignment perpendicular to the relative momentum, as in Fig.~\ref{FIG:schematic}(b) and (c), the differential cross section is clearly anisotropic. While the cross section is the same in panels (b) and (c), the rotation with respect to the imaging direction allows different projections of the 3D halo to be imaged.
\section{Theory and simulation}
\subsection{Two-body scattering theory}~\label{SUBSEC:TwoBody}
The central theoretical element is the differential scattering cross section for identical bosons interacting via a combination of dipolar and short-range interactions. This is found by solving the two-body Schr\"odinger equation in the center-of-mass rest frame,
\begin{equation}
\left[-\frac{\hbar^2}{m}\nabla_{\mathbf{r}}^2+V_{\rm d}(\hat{\epsilon},\mathbf{r})+V_{\rm sr}(\mathbf{r})\right]\psi(\mathbf{r})=\frac{\hbar^2k^2}{m}\psi(\mathbf{r}),
\end{equation}
where $m$ is the mass of a single particle and $\hbar$ is the reduced Planck's constant.
The dipole-dipole potential energy for aligned dipoles is $V_{\rm d}(\hat{\epsilon},\mathbf{r})=(2\hbar^2 a_d/m)\left[1-3(\hat{\epsilon}\cdot\hat{r})^2\right]/{r^3}
$,
where $\hat{\epsilon}$ is the direction of alignment. The potential energy from short-range van der Waals interactions can be approximated by $V_{\rm sr}(\mathbf{r})=(4\pi\hbar^2 a/m)\delta^{(3)}(\mathbf{r})$, provided one stays within the first-order Born approximation.
The strength of the dipole interaction is determined by the length scale $a_d=\mu_0\mu^2m/8\pi\hbar^2$, where $\mu_0$ is the vacuum permeability and $\mu$ is the magnetic moment of a single particle~\cite{DipolarBosonsReview}. Similarly, the strength of the short-range interaction is determined by the $s$-wave scattering length $a$. This problem is solved in detail in Ref.~\cite{Bohn14_PRA}, see also Ref.~\cite{DipolarCollisions_OMalley}. Briefly, in the limit where $k\rightarrow0$, a solution can be found which takes the asymptotic form $\psi(\mathbf{r})=e^{i\mathbf{k}\cdot\mathbf{r}}+f(\hat{\epsilon},\hat{k},\hat{k}')e^{ikr}/r$,
where $e^{i\mathbf{k}\cdot\mathbf{r}}$ is an incoming wave and $f$ is the scattering amplitude that depends on the dipole alignment direction, the incoming wave vector $\mathbf{k}=k\hat{k}$, and the outgoing wave vector $\mathbf{k}'=k\hat{k}'$. The differential scattering cross section for identical bosons, given by $\frac{d\sigma}{d\Omega}=\frac{1}{2}|f(\hat{\epsilon},\hat{k},\hat{k}')+f(\hat{\epsilon},\hat{k},-\hat{k}')|^2$ in the low energy limit as $k\rightarrow0$ and within the first-order Born approximation, is found to be
\begin{equation}\label{EQdifferential}
\frac{d\sigma}{d\Omega}=2a_d^2\left[
\frac{(\hat{k}\cdot\hat{\epsilon})^2+(\hat{k}'\cdot\hat{\epsilon})^2-2(\hat{k}\cdot\hat{\epsilon})(\hat{k}'\cdot\hat{\epsilon})(\hat{k}\cdot\hat{k}')}{1-(\hat{k}\cdot\hat{k}')^2}
-\frac{2}{3}+\frac{a}{a_d}
\right]^2,
\end{equation}
where $\hat{k}$ ($\hat{k}'$) is a unit vector along the direction of incoming (outgoing) relative momentum. The total cross section, found by integrating Eq.~\eqref{EQdifferential} over all possible outgoing directions, is given by~\cite{Bohn14_PRA}
\begin{equation}\label{EQsigma}
\sigma(\eta) = a_d^2\frac{\pi}{9}\left\{72\frac{a^2}{a_d^2}-24\frac{a}{a_d}\left[1-3\cos^2(\eta)\right]+11-30\cos^2(\eta)+27\cos^4(\eta)\right\},
\end{equation}
where $\eta$ is the angle between $\hat{k}$ and $\hat{\epsilon}$. We note that the spatial anisotropy of this cross section is caused by $\hat{\epsilon}$. Equations~\eqref{EQdifferential} and~\eqref{EQsigma} are universal in the sense that they are insensitive to any details of the short-range potential.
\subsection{Many-body considerations: Monte-Carlo simulation}~\label{SUBSEC:ManyBody}
When two atoms from different condensates collide, their relative velocity is much larger than the width of each condensate's internal velocity distribution. Therefore, these primary collisions effectively occur at a fixed angle between $\hat{k}$ and $\hat{\epsilon}$. However, each atom then has a finite probability of a secondary scattering event. These secondary scattering events occur with an essentially random angle between $\hat{k}$ and $\hat{\epsilon}$ and corrupt the direct correspondence between the observed halo and the differential scattering cross section. For this reason we use a direct-simulation Monte Carlo (DSMC) algorithm, which is able to keep track of such multiple collision events, in order to quantitatively understand the experimental data as accurately as possible.
The system dynamics can be separated into a low energy part (relevant to the two condensates) and a high energy part (relevant to the halo). These two parts have high and low phase space densities, respectively. Because we are primarily interested in the halo, we focus on the classical kinetic equation for the phase space distribution function, which is given by
\begin{equation}\label{EQ:KineticEq}
\partial_t f(\mathbf{r},\mathbf{p},t)+\frac{\mathbf{p}}{m}\cdot\nabla_{\mathbf{r}}f(\mathbf{r},\mathbf{p},t)=C[f],
\end{equation}
where the left-hand side contains free-streaming terms and the right-hand side includes effects from two-body collisions. This is given by
\begin{equation}\label{CollisionIntegral}
C[f]=\int\frac{d^3 \mathbf{p}_1}{h^3}\int d^2\hat{\Omega}\frac{d\sigma}{d\Omega}v_r\left[
f'f_1'-ff_1
\right],
\end{equation}
where $f=f(\mathbf{r},\mathbf{p})$, $f_1=f(\mathbf{r},\mathbf{p}_1)$, $f'=f(\mathbf{r},\mathbf{p}')$, and $f_1'=f(\mathbf{r},\mathbf{p}_1')$ account for the four momenta (two incoming, $\mathbf{p}$, $\mathbf{p}_1$, and two outgoing, $\mathbf{p}'$, $\mathbf{p}_1'$) associated with a two-body collision, and $v_r=|\mathbf{p}-\mathbf{p}_1|/m$ is the relative velocity. Note that $\mathbf{p}=\hbar \mathbf{k}$ connects the momentum to the wavevector, which was used in Sec.~\ref{SUBSEC:TwoBody}. The momenta in Eq.~\eqref{CollisionIntegral} are related by energy and momentum conservation such that $\mathbf{p}+\mathbf{p}_1=\mathbf{p}'+\mathbf{p}_1'$ and $|\mathbf{p}-\mathbf{p}_1|=|\mathbf{p}'-\mathbf{p}_1'|$. The integration variable is $\hat{\Omega}=(\mathbf{p}'-\mathbf{p}_1')/|\mathbf{p}'-\mathbf{p}_1'|$. Equation~\eqref{EQ:KineticEq} does not provide an accurate description of the atoms within the high-phase-space-density regions, i.e., the condensates. These would presumably be described within the context of a Gross-Pitaevskii-type theory of BEC evolution~\cite{Norrie_TWA_Collisions1,Norrie_TWA_Collisions2,Deuar_Hybrid}. However, the goal of simulating both the high and low-phase-space-density components of the gas is beyond the scope of our current work. We simply wish to approximately capture the density evolution of the condensates. This in turn yields a reasonable prediction for the scattering rates and therefore the formation and deformation of the scattering halo.
We solve Eq.~\eqref{EQ:KineticEq} numerically using a DSMC algorithm, details of which have been published in Ref.~\cite{Sykes14_preprint}. We use an initial condition corresponding to both spatial and momentum densities proportional to
\begin{equation}\label{ProbDistFun}
P(\mathbf{R};\mathbf{x})={\rm max}\left\{
\frac{15}{8\pi\prod_\alpha R_\alpha}\left[1-\sum_{\alpha}\frac{x_\alpha^2}{R_\alpha^2}\right],0
\right\},\qquad \alpha = \{1,2,3\},
\end{equation}
where $\mathbf{x}$ ($R_\alpha$) can denote either spatial or momentum coordinates (widths).
The spatial part corresponds to a Thomas-Fermi condensate density, with the widths calculated in the manner prescribed by Refs.~\cite{Eberlein2005,Griesmaier2006}. The momentum widths are found by fitting to the experimental image of the expanded condensate. The system is then divided into two halves, which propagate along the positive and negative $x$-axis, respectively.
For a detailed account of the simulation method, we refer the reader to Ref.~\cite{Sykes14_preprint}, which adjusts the original DSMC approach~\cite{BirdBook} into a version appropriate to dipolar gases. Similar methods have been employed in the study of ultracold gases, see for instance Refs.~\cite{WadeBlakie2011_PRA,Lobo_Boltzmann,Urban_Boltzmann}. In particular, Ref.~\cite{WadeBlakie2011_PRA} also studied halo formation in the case of $s$- and $d$-wave collisional cross sections. Briefly, our computational algorithm uses test-particles with phase-space coordinates which are sampled from the initial distribution in Eq.~\eqref{ProbDistFun}. The test particles move classically from one time-step to the next. At each time-step particles are binned in position space. The bin-size represents the finite resolution of the delta-function within the numerics. Within each bin, the collision probability for each pair of particles is evaluated according to Eq.~\eqref{EQsigma} using the correct value of $\hat{k}$ corresponding to each pair. Collisions are then chosen to occur stochastically in accordance with these probabilities. The post-collision velocities are also chosen stochastically, in accordance with the differential scattering cross section in Eq.~\eqref{EQdifferential}.
\begin{figure}
\centering
\includegraphics[width=0.7\columnwidth]{AllFiveExperimentsOnOneImageCB3.pdf}
\caption{Absorption images in the $\hat{x}$--$\hat{z}$ plane after 22-ms TOF and averaged forty times each. The larger images labeled (a1), (b1), etc.~show the experimental data and the dipole alignment configuration, while the smaller images (a2)--(a9), (b2)--(b9), etc.~show simulation images, each with a different contribution from the $s$-wave (short-range) interaction. This contribution varies from $a=40a_0$ in (a2), (b2), etc.~to $a=180a_0$ in (a9), (b9), etc., with uniform steps of $20a_0$ in between. Dipoles are aligned along the direction: $(1,0,0)$ in (a), $(0,1,0)$ in (b), $(0,0,1)$ in (c), $(1,0,1)/\sqrt{2}$ in (d), and $(0,1,1)/\sqrt{2}$ in (e). The colour scale to the right shows the number of atoms per pixel for all data sets.}
\label{FIG:MainComparison}
\end{figure}
\section{Results}
The experimental absorption images are presented in the first column of Fig.~\ref{FIG:MainComparison}, i.e., in panels (a1), (b1), $\ldots$ (e1). Each image corresponds to the average of forty experimental runs at a fixed dipole alignment direction, as stated in the figure caption.
Imperfections in the diffraction leaves residual atoms near $0\hbar \mathbf{k}_{L}$. This was also observed in Ref.~\cite{QuantumNewtonsCradle}. These atoms contribute noise to the measurement both near $0\hbar \mathbf{k}_{L}$ and in the $|2\hbar \mathbf{k}_{L}|$ halos from collisions with the $\pm2\hbar \mathbf{k}_{L}$ BECs. However, the effect of the dipole alignment direction is clear from the remarkable variation between images (a1), (b1), $\ldots$ (e1). We run a family of simulations for each value of the dipole alignment direction, but with varying $s$-wave scattering length~\footnote{We allow scattering length to vary because the large number of internal degrees of freedom within open-shell lanthanide atoms~\cite{Kotochigova_Dysprosium1,Kotochigova_Dysprosium2} restricts our ability to microscopically determine the scattering lengths of these atoms, though previous measurements have restricted the range of likely values~\cite{TangSykes2015_PRA,Maier2015,Tang2016}.}. The test particles in the DSMC are initialized as follows: We estimate the initial density of the condensate by using the exact solution to the Hartree-Fock theory for a dipolar condensate in the Thomas-Fermi limit~\cite{Eberlein2005}, and we use this to define spatial widths $R_\alpha^{\rm (spat.)}$ in Eq.~\eqref{ProbDistFun}. TOF expansion images are used to estimate the in situ momentum distribution, and we use this measurement to define momentum widths $R_\alpha^{\rm (mom.)}$ in Eq.~\eqref{ProbDistFun}. We have also looked at Gaussian distribution functions in place of Eq.~\eqref{ProbDistFun}, and also small variations in momentum and spatial widths $R_\alpha$. However, we found that such variations have very little effect on our conclusions. These families of simulation results are shown as the smaller images in Fig.~\ref{FIG:MainComparison} panels (a2)--(a9), (b2)--(b9), $\ldots$ (e2)--(e9).
\begin{figure}[t]
\centering
\includegraphics[width=.75\textwidth]{pixelbypixel2.pdf}
\caption{(a)--(e) Pixel-by-pixel comparison of the experimental and simulated absorption images for different dipole alignment directions: $(1,0,0)$ in (a), $(0,1,0)$ in (b), $(0,0,1)$ in (c), $(1,0,1)/\sqrt{2}$ in (d), and $(0,1,1)/\sqrt{2}$ in (e). The points denote the value of the weighted least squares cost function for each simulated $s$-wave scattering length $a$ (see text for details). (f) Example of the masking used for pixel-by-pixel comparisons. The grey regions to the left and right mask the unscattered BECs, and the central grey region masks atoms not diffracted by the lattice. }
\label{FIG:PixelByPixelComparison}
\end{figure}
\begin{figure}[b]
\centering
\includegraphics[width=.75\textwidth]{MultipleCollisionStatisticsDipolesAlongXAxisV3.pdf}
\caption{Analysis of multiple collision effects in the case where dipoles are aligned along the $\hat{x}$ axis. (a) The experimental absorption image. (b)--(f) Analysis of simulation data: (b) shows the population of atoms separated by the number of scattering events each atom incurred; (c) shows the full simulation of the absorption image; (d) an absorption image with {\it only} the atoms which scattered once, and is therefore in perfect correspondence with the two-body differential cross section; (e) shows an absorption image with only the atoms which scattered twice; (f) shows an absorption image with only the atoms which scattered three times. The simulations were done with an $s$-wave scattering length $a=140a_0$. The colour-scale (defining the number of atoms per pixel) is the same in panels (a) and (c) and shown beneath panel (a). The remaining panels correspond to the colour-scale shown on the right.
}
\label{FIG:MultipleCollisionsX}
\end{figure}
We perform a pixel-by-pixel comparison between the experimental and simulated scattering images. Each pixel corresponds to an area of 2.6~$\mu$m $\times$ 2.6~$\mu$m. For each dipole alignment direction, the average experimental image $\mathcal{E}$ is obtained as well as the standard error for each pixel $\mathcal{E}^{\sigma}$. The experimental image and each simulation image $\mathcal{S}$ (including unscattered atoms) is normalized to suppress the effects of atom number variations, and a mask is then applied to the images to exclude the unscattered BECs and atoms not diffracted by the lattice. An example of a masked image is shown in Fig.~\ref{FIG:PixelByPixelComparison}(f). The weighted sum of the squared residuals is calculated for each simulated image ${S=\sum_{ij}(\mathcal{E}_{ij}-\mathcal{S}_{ij})^{2}/(\mathcal{E}^{\sigma}_{ij})^{2}}$,
where the subscript $ij$ denotes the pixel in the $i$th row and $j$th column of an image. This corresponds to the cost function for a weighted least squares regression. The results of this analysis are shown in Fig.~\ref{FIG:PixelByPixelComparison}(a)--(e). Though a particular choice of scattering length, $a$, minimizes $S$ for each dipole alignment direction, the analysis is not sensitive enough to define a 1$\sigma$ confidence interval. However, as seen in Fig~\ref{FIG:PixelByPixelComparison}, all the minima fall within the range of $120a_0$--$160a_0$, which is consistent with measurements reported in our previous work~\cite{TangSykes2015_PRA,Tang2016}. We believe that the fluctuations in the location of the minima in Fig.~\ref{FIG:PixelByPixelComparison} are primarily due to an incomplete knowledge of the initial density of the BEC. Recent results suggest that dipolar interactions in strongly dipolar BECs can have surprising consequences which require beyond-mean-field effects to be quantitatively understood, see Refs.~\cite{PfauDroplet1,PfauDroplet2,SantosWachtler}. We have run additional simulations with a variety of initial densities and note that variance in initial density leads to deformations in the halo due to the additional (fewer) multiple scattering events in the case of an increased (decreased) density. These deformations are capable of shifting the minima in Fig.~\ref{FIG:PixelByPixelComparison}.
We also analyze the collision frequency and, through a detailed comparison between simulation and experiment, establish the relevance of multiple collisions. These are collisions involving atoms that have already been scattered out of the original condensates. The nature of the DSMC method allows one to label particles and keep track of collisions as they occur.
The results are shown in Fig.~\ref{FIG:MultipleCollisionsX} for dipoles aligned along $\hat{x}$, a case in which the presence of multiple collisions is strong. Figure~\ref{FIG:MultipleCollisionsXZ_and_Z} shows results where the dipoles are aligned along both $(\hat{x}+\hat{z})/\sqrt{2}$ and $\hat{z}$, and the presence of multiple collisions is less relevant. The data can be reasonably well understood starting from Eq.~\eqref{EQsigma} and noting that, when $a=140a_0$, the total cross section has a maximum at $\eta=0^\circ$ (corresponding to dipoles aligned along $\hat{x}$) and a minimum at $\eta=90^\circ$ (corresponding to dipoles aligned anywhere in the $\hat{y}$--$\hat{z}$ plane). Specifically, we have $\sigma(0^\circ)/\sigma(90^\circ)\approx2.6$ and $\sigma(0^\circ)/\sigma(45^\circ)\approx1.6$. For this reason, we observe the unscattered fraction of the gas increases from $\sim$40\% in Fig.~\ref{FIG:MultipleCollisionsX} (b), to $\sim$60\% in Fig.~\ref{FIG:MultipleCollisionsXZ_and_Z} (b), and to $\sim$80\% in Fig.~\ref{FIG:MultipleCollisionsXZ_and_Z} (f). From this, we conclude that the experimental absorption image with dipoles aligned along the $\hat{x}$ axis is strongly affected by atoms which have undergone multiple collisions. Indeed, these atoms appear to constitute the majority of the absorption image in this case. Reducing the initial density to suppress multiple collisions is not practicable due to the sharp loss in signal-to-noise. However, for all other alignment directions the experimental absorption image (away from the condensate regions) is dominated by a single scattering event, and is therefore in close correspondence with the differential scattering cross section of Eq.~\eqref{EQdifferential}.
\begin{figure}[t]
\centering
\includegraphics[width=.77\textwidth]{MultipleCollisionStatisticsDipolesAlongXZandZ.pdf}
\caption{Similar to Fig.~\ref{FIG:MultipleCollisionsX}, except with dipoles aligned along the $(\hat{x}+\hat{z})/\sqrt{2}$ axis [shown in the top row, (a)--(d)] and the $\hat{z}$ axis [shown in the bottom row, (e)--(h)]. (a) and (e) The experimental absorption images associated with their own colour-scale at left, indicating number of atoms per pixel. (b)--(d) and (f)--(h) Analysis of simulation data: (b) and (f) show the population of atoms separated by the number of scattering events each atom incurred; (c) and (g) show an absorption image with {\it only} the atoms which scattered once, and is therefore in perfect correspondence with the two-body differential cross section; (d) and (h) show an absorption image with only the atoms which scattered twice. The simulations were done with an $s$-wave scattering length $a=140a_0$. The presence of multiple collisions is less pronounced for these dipole alignment directions than in Fig.~\ref{FIG:MultipleCollisionsX}.
}
\label{FIG:MultipleCollisionsXZ_and_Z}
\end{figure}
\section{Conclusion and discussion}
In conclusion, we have measured the differential scattering between identical dipolar bosons in the low-energy (universal) regime by imaging the halos of atoms scattering from two colliding Dy BECs. The results depend strongly on the angle between the dipole alignment direction and the collision axis and are well described by the analytic formula in Eq.~\eqref{EQdifferential}, derived under the first-order Born approximation. Although it is not intended to be entirely quantitative, a classical Monte-Carlo simulation of the many-body dynamics provides a reasonable higher-order approximation to account for finite-density effects and finite-momentum distribution widths. We allow the $s$-wave scattering length to vary within the simulations, and note that discrepancies between simulation and experiment are minimized at a scattering length which remains consistent with previous measurements~\cite{TangSykes2015_PRA,Tang2016}.
The measurements provide a beautiful demonstration of the theoretical prediction for a low energy scattering amplitude, which dates back to O'Malley in 1964~\cite{DipolarCollisions_OMalley}. The envisaged physical system under consideration at that time was not dipolar collisions, but rather the scattering of an electron by a non-spherical atom, such as atomic oxygen. However, we see that the interaction potential is asymptotically equivalent in the two cases. Future work may use such experiments to probe the complex collisional physics of dipolar condensates and degenerate Fermi gases near Feshbach resonances in and among the dense and ultradense spectra observed in the dysprosium and erbium systems~\cite{Erbium_FermiGas,Dy_FeshbachSpectrum,Frisch2014,Maier2015Chaotic,Maier2015,Burdick:2016vv}.
\section{Acknowledgements}
We gratefully acknowledge John Bohn for early contributions to this work. We thank Jack DiSciacca and Wil Kao for experimental assistance and Ian Spielman for a stimulating discussion. We are grateful to the NSF and AFOSR for funding support. Y.T.~acknowledges partial support from a Stanford Graduate Fellowship. The research leading to these results received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 658311.
\medskip
\providecommand{\newblock}{}
|
\section{Introduction}
Most of the epidemiological models in the literature are continuous models. In spite of this, recently, there has been a growing interest in discrete-time models~\cite{Diekmann-Heesterbeek-W-2000,Tan-Gao-Fan-DDNS-2015,Wang-Teng-Rehim-DDNS-2014,Wei-Le-DDNS-2015,Liu-Peng-Zhang-AML-2015,Hua-Teng-Zhang-MCS-2014,Hu-Teng-Jia-Zhang-Zhang-ADE-2014}.
In this work, we will use Mickens nonstandard difference (NSFD) scheme to achieve a discretization of a family of continuous epidemiological models with vaccination and general incidence function considered in~\cite{Pereira-Silva-Silva}. We have multiple objectives: firstly, we want to obtain conditions for extinction and permanence of the disease for the discrete family; next, having a continuous and a corresponding discrete family of models, we wish to discuss the problem of consistency of the discrete models with the corresponding continuous ones; finally, we intend to present some simulation results.
The dynamical consistency of a numerical scheme with the associated continuous system is not a precise definition. By the expression ''dynamical consistency'' it is meant that the numerical solutions replicate some of the properties of the continuous systems solutions. For us, dynamical consistency means: whenever there is extinction (respectively permanence) of the disease for the continuous-time model the same holds for the discrete-time one. Several papers~\cite{Cui-Yang-Zhang-JDEA-2014,Cui-Zhang-JDEA-2015,Ding-Ding-JDEA-2013,Ding-Ding-JDEA-2014,Jang-Elaydi-CAMQ-2003,Mickens-JCAM-1999,Mickens-Washington-JDEA-2012} discuss the dynamical consistency with respect to some particular properties of discrete epidemiological models obtained from continuous models by some NSFD scheme~\cite{Mickens-JDEA-2002}. We note that while the papers cited above consider autonomous models, in the present work we discuss dynamical consistency for a non-autonomous model. To the best of our knowledge, this is the first work where the consistency of a discretized epidemiological model with the original continuous model is discussed in the non-autonomous context.
Regarding our simulation results, we considered two different types of computational experiments. Our first set of simulation results are designed to compare different possible discretizations of our continuous models. After this discussion, we apply our model to a real situation, considering data from the incidence of measles in France in the period 2012-2016. To the possible extent, this data is used to estimate our model parameters and the computational results obtained are compared with real data.
The law of mass action, states that the rate of change in the disease incidence is directly proportional to the product of the number of susceptible and infective individuals, and was the paradigm in the classic models in epidemiology. This is why classical models usually consider a bilinear incidence rate $\beta SI$, where $S$ and $I$ denote respectively the number of susceptible and infective individuals, to model the disease transmission. In spite of this, it is sometimes important to consider other forms of incidence functions. Another usual assumption is the time independence of the parameters model parameters: in fact, the majority of the epidemiological models in the literature are given by a system of autonomous differential or difference equations. Nevertheless, the assumption that the parameters are independent of time is not very realistic in many situations and it is useful to consider non-autonomous models that, for instance, allow the discussion of environmental and demographic effects that change with time~\cite{Khasnis-Nettleman-AMR-2005, Kloeden-Potzsche-Springer-2013}. In this work the family of models considered is non-autonomous and the incidence rates are taken from a large class of functions.
Our model generalizes one obtained by Mickens nonstandard finite difference method from the continuous model~\cite{Pereira-Silva-Silva} (see section~\ref{section:discretization}). In~\cite{Zhang-AMC-2015}, a discrete non-autonomous epidemic model with vaccination and mass action incidence was obtained by Mickens method. We emphasize that, in the particular mass-action case, our model is not exactly similar to the model in~\cite{Zhang-AMC-2015}, although Mickens rules were considered in both. We briefly compare computationally these two slightly different models in Section~\ref{section:simulation}.
The model we will consider is the following
\small{
\begin{equation}\label{eq:ProblemaPrincipal}
\begin{cases}
S_{n+1}-S_n=\Lambda_n-\beta_n\varphi(S_{n+1},I_n)-(\mu_n+p_n)S_{n+1}+\eta_nV_{n+1} \\ I_{n+1}-I_n=\beta_n\varphi(S_{n+1},I_n)+\sigma_n\psi(V_{n+1},I_n)-(\mu_n+\alpha_n+\gamma_n)I_{n+1} \\
R_{n+1}-R_n=\gamma_nI_{n+1} -\mu_nR_{n+1} \\
V_{n+1}-V_n=p_nS_{n+1}-(\mu_n+\eta_n)V_{n+1}-\sigma_n\psi(V_{n+1},I_n)
\end{cases},
\end{equation}
}
$n \in \ensuremath{\mathds N}$, where the classes $S$, $I$, $R$, and $V$ correspond, respectively, to susceptible, infective, recovered and vaccinated individuals and the parameter functions have the following meanings: $\Lambda_n$ denotes the inflow of newborns in the Susceptible class; the function $\beta_n \varphi$ is the incidence (into the Infective class) from the susceptible individuals; the function $\sigma_n \psi$ is the incidence (into the Infective class) from the vaccinated individuals; $\mu_n$ are the natural deaths; $p_n$ represents the susceptibles vaccination; $\eta_n$ represents the immunity loss and consequence influx in the susceptible class; $\alpha_n$ are the deaths occurring in the infective class; $\gamma_n$ is the recovery.
We will assume that $(\Lambda_n), (\mu_n), (p_n), (\eta_n), (\alpha_n)$, $(\beta_n)$, $(\sigma_n)$ and $(\gamma_n)$ are bounded and nonnegative sequences and that there are positive constants
$w_\mu$, $w_\Lambda$, $w_p$, $k_{\varphi}$ and $k_{\psi}$ such that:
\begin{enumerate}[H1)]
\item \label{cond-H1} the functions $\varphi:\ensuremath{\mathds R}^2 \to \ensuremath{\mathds R}$ and $\psi:\ensuremath{\mathds R}^2 \to \ensuremath{\mathds R}$ are nonnegative and differentiable in $(\ensuremath{\mathds R}_0^+)^2$ and the functions $\ensuremath{\mathds R}^+_0 \ni x \rightarrow \partial_2 \varphi(x,0)$ and $ \ensuremath{\mathds R}^+_0 \ni x \rightarrow \partial_2 \psi(x,0)$ are non-decreasing and Lipschitz, with Lipschitz constants $k_{\varphi}$ and $k_{\psi}$.
\item \label{cond-H2} we have $\varphi(x,0)=\psi(x,0)=\varphi(0,y)=\psi(0,y)=0$ for all $x, y \in \ensuremath{\mathds R}_0^{+}$.
\item \label{cond-P1}
$\displaystyle \limsup_{n \to +\infty} \prod_{k=n}^{n+\omega_\mu} \dfrac{1}{1+\mu_k} <1$;
\item \label{cond-P2} $\displaystyle \liminf_{n \to +\infty} \sum_{k=n+1}^{n+\omega_\Lambda} \Lambda_k > 0 \quad \text{and} \quad \liminf_{n \to +\infty} \sum_{k=n+1}^{n+\omega_p} p_k > 0$;
\item \label{cond-P3}Functions $\ensuremath{\mathds R}^+ \ni y \mapsto \phi(x,y)/y$ and $\ensuremath{\mathds R}^+ \ni y \mapsto \psi(x,y)/y$ are non-increasing.
\end{enumerate}
In this work, we prove, when our conditions prescribe extinction (respectively permanence) for the continuous model we also have extinction (respectively permanence) for the corresponding discrete model as long as the time step is smaller than some constant (that depends on some model parameters and on the threshold condition). We also consider a family of examples of the periodic system of period $1$ such that the continuous and the discrete time system with time step $h=1/L$ is not consistent, highlighting the importance of knowing that for time steps smaller than some explicit value we have consistency.
\section{Discretization of the continuous model} \label{section:discretization}
We start with a non-autonomous SIRVS model that is slightly less general than the one considered in~\cite{Pereira-Silva-Silva} and generalizes the one in~\cite{Zhang-Teng-Gao-AA-2008}. Namely, we consider the model:
\begin{equation}\label{eq:ProblemaPrincipal-cont}
\begin{cases}
S'=\Lambda(t)-\beta(t)\varphi(S)I-(\mu(t)+p(t))S+\eta(t)V \\
I'=\left[ \beta(t)\varphi(S)+\sigma(t)\psi(V)-\mu(t)-\alpha(t)-\gamma(t) \right]I \\
R'=\gamma(t)I -\mu(t)R \\
V'=p(t)S-(\mu(t)+\eta(t))V-\sigma(t)\psi(V)I
\end{cases}.
\end{equation}
We assume that the functions $\Lambda, \mu, p, \eta, \alpha$, $\beta$, $\sigma$ and $\gamma$ belong to the class $C^1(\ensuremath{\mathds R}_0^+)$, are nonnegative and bounded. We also require that:
\begin{enumerate}[C1)]
\item \label{cond-C1} the functions $\varphi:\ensuremath{\mathds R} \to \ensuremath{\mathds R}$ and $\psi:\ensuremath{\mathds R} \to \ensuremath{\mathds R}$ are nonnegative, non decreasing, differentiable and Lipschitz with Lipschitz constants $k_{\varphi}$ and $k_{\psi}$ respectively;
\item \label{cond-C2} $\varphi(0)=\psi(0)=0$;
\item \label{cond-C3} there is $\omega>0$ such that $\displaystyle \liminf_{t \to +\infty} \int_t^{t+\omega} \mu(s) \, ds > 0$.
\end{enumerate}
In order to obtain threshold conditions for model~\eqref{eq:ProblemaPrincipal-cont}, it was considered in~\cite{Pereira-Silva-Silva} the following auxiliary system:
\begin{equation}\label{eq:SistemaAuxiliar-Pereira-Silva-Silva}
\begin{cases}
x'=\Lambda(t) - [\mu(t) + p(t)] x + \eta(t)y \\
y'=p(t)x - [\mu(t)+\eta(t)] y.
\end{cases}
\end{equation}
and for each solution $(x^*(t),y^*(t))$ of~\eqref{eq:SistemaAuxiliar-Pereira-Silva-Silva} with positive initial conditions, it was shown that the numbers
\[
R^\ell_C(\lambda) = \liminf_{t \to \infty} \int_t^{t+\lambda} \beta(s)\varphi(x^*(s))+\sigma(s)\psi(y^*(s))-\mu(s)-\alpha(s)-\gamma(s) \, ds
\]
and
\[
R^u_C(\lambda) = \limsup_{t \to \infty} \int_t^{t+\lambda} \beta(s)\varphi(x^*(s))+\sigma(s)\psi(y^*(s))-\mu(s)-\alpha(s)-\gamma(s) \, ds
\]
are independent of the particular solution.
Using the above numbers, the following results are contained in results obtained in~\cite{Pereira-Silva-Silva}:
\begin{theorem}[Theorem 1 of~\cite{Pereira-Silva-Silva}] \label{teo:Permanence-Pereira-Silva-Silva}
Assume that conditions~C\ref{cond-C1}),~C\ref{cond-C2}) and~C\ref{cond-C3}) hold. Then, if there is a constant $\lambda>0$ such that $R^\ell_C (\lambda) >0$ then the infectives $I$ are permanent
in system~\eqref{eq:ProblemaPrincipal-cont}.
\end{theorem}
\begin{theorem}[Theorem 2 of~\cite{Pereira-Silva-Silva}] \label{teo:Extinction-Pereira-Silva-Silva}
Assume that conditions~C\ref{cond-C1}),~C\ref{cond-C2}) and~C\ref{cond-C3}) hold. Then
if there is a constant $\lambda>0$ such that $R^u_C(\lambda) <0$ then the infectives $I$ go to extinction in system~\eqref{eq:ProblemaPrincipal-cont}.
\end{theorem}
In the literature, several models were discretized using Mickens NSFD schemes~\cite{Allen-Jones-Martin-MB-1994, Allen-MB-1975, Cooke-TPB-1975,DeJong-Diekmann-Heesterbeek-MB-1994, Hattaf-Lashari-Boukari-Yousfi-DEDS-2015, Kaitala-Heino-Getz-BMB-1997, Lena-Serio-MB-1982, Lefevre-Malice-MM-1986, Lefevre-Picard-MB-1989, Longini-MB-1986, Mickens-B-1982, Spicer-BMB-1979, Viaud-MB-1993, West-Thompson-MB-1997}. Next, we will apply Micken's non-standard method to obtain a discrete version of system~\eqref{eq:ProblemaPrincipal-cont}.
Let $\phi:\ensuremath{\mathds R}_0^+ \to \ensuremath{\mathds R}$ be a positive continuous function such that
\begin{equation}\label{eq:derivative-denominator}
\lim_{h \to 0} \phi(h)=0.
\end{equation}
Given $h \in \ensuremath{\mathds R}^+$, we let $t=nh$, with $n \in \ensuremath{\mathds N}$, and identify $S'(t)$ with
$$
\dfrac{S(nh+h)-S(nh)}{\phi(h)}.
$$
After deciding a non-local representation for the incidence function and that terms that do not correspond to an interaction will be considered in the $n+1$ time, the first equation in~\eqref{eq:ProblemaPrincipal-cont} becomes
\[
\begin{split}
S((n+1)h)-S(nh)
& =\phi(h)\left[\Lambda(nh)-\beta(nh)\varphi(S((n+1)h))I(nh)\right.\\
& \quad \left. -(\mu(nh)+p(nh))S((n+1)h)+\eta(nh)V(nh+h)\right].
\end{split}
\]
Writing $S_n=S(nh)$, $I_n=I(nh)$, $V_n=V(nh)$, $\Lambda_n=\phi(h)\Lambda(nh)$, $\beta_n=\phi(h)\beta(nh)$,
$\mu_n=\phi(h)\mu(nh)$, $p_n=\phi(h)p(nh)$ and $\eta_n=\phi(h)\eta(nh)$, we have
$$S_{n+1}-S_n=\Lambda_n-\beta_n\varphi(S_{n+1})I_n-(\mu_n+p_n)S_{n+1}+\eta_nV_{n+1}.$$
Proceeding similarly for the other equations, we obtain the following discrete model
\small{
\begin{equation}\label{eq:ProblemaPrincipal-disc}
\begin{cases}
S_{n+1}-S_n=\Lambda_n-\beta_n\varphi(S_{n+1})I_n-(\mu_n+p_n)S_{n+1}+\eta_nV_{n+1} \\ I_{n+1}-I_n=\beta_n\varphi(S_{n+1})I_n+\sigma_n\psi(V_{n+1})I_n-(\mu_n+\alpha_n+\gamma_n)I_{n+1} \\
R_{n+1}-R_n=\gamma_nI_{n+1} -\mu_nR_{n+1} \\
V_{n+1}-V_n=p_nS_{n+1}-(\mu_n+\eta_n)V_{n+1}-\sigma_n\psi(V_{n+1})I_n
\end{cases},
\end{equation}
}
$n \in \ensuremath{\mathds N}_0$. We will consider a model that contains this one to obtain some of our results. Namely, based on model~\eqref{eq:ProblemaPrincipal-disc}, in sections~\ref{section:PE} and~\ref{section:SR} we will study model~\eqref{eq:ProblemaPrincipal} that has a more general form for the incidence function.
Now, we need to make some definitions. We say that:
\begin{enumerate}[i)]
\item the infectives $(I_n)$ are \emph{permanent} if for any solution $(S_n,I_n,R_n,V_n)$ of~\eqref{eq:ProblemaPrincipal} with initial conditions $S_0,I_0,R_0,V_0>0$ there are constants $0<m<M$ such that
\[
m < \liminf_{n \to \infty} I_n \le \limsup_{n\to \infty} I_n < M;
\]
\item the infectives $(I_n)$ go to \emph{extinction} if for any solution $(S_n,I_n,R_n,V_n)$ of~\eqref{eq:ProblemaPrincipal} with initial conditions $S_0,I_0,R_0,V_0\ge 0$ we have
$\displaystyle \lim_{n \to \infty} I_n = 0$.
\end{enumerate}
Similar definitions can be made for the other compartments. For instance, if
there exists constants $0<m<M$ such that for any solution $(S_n,I_n,R_n,V_n)$ of~\eqref{eq:ProblemaPrincipal}
with initial conditions $S_0,I_0,R_0,V_0 > 0$ we have
\[
m < \liminf_{n \to \infty} S_n \le \limsup_{n \to \infty} S_n < M
\]
we say that the susceptibles are permanent.
\section{Permanence and Extinction in the Discrete Model} \label{section:PE}
In this section, we will extend the results obtained for the model with the usual mass action incidence in~\cite{Zhang-AMC-2015} to our generalized family of models. Namely, suitable thresholds are defined and conditions for persistence and extinction of the disease are obtained. As a corollary of our results, we consider the periodic case where we have a unique number that establishes the boundary between the regions of permanence and extinction. Although the proofs of our results are inspired in~\cite{Zhang-AMC-2015}, some difficulties must be dealt with. In particular, it was necessary to understand the right conditions to impose to the incidence functions in order to overcome the technical difficulties.
To lighten the reading, the proofs of our results are presented in appendix~\ref{Appendix:A}.
\subsection{Auxiliary results} \label{subsection:AR}
Consider the auxiliary system,
\begin{equation}\label{eq:SistemaAuxiliar}
\begin{cases}
x_{n+1}=\dfrac{\Lambda_n + \eta_n y_{n+1} +x_n}{1+\mu_n+p_n} \\[2mm]
y_{n+1}=\dfrac{p_n x_{n+1} +y_n}{1+\mu_n+\eta_n}
\end{cases}.
\end{equation}
Note that the auxiliary system describes the behaviour of the system in the absence of infection. If $(\Lambda_n)$, $(\mu_n)$, $(p_n)$, $(\eta_n)$, $(\alpha_n)$, $(\mu_n)$, $(\sigma_n)$ and $(\beta_n)$ are constant sequences then~the linear system \eqref{eq:SistemaAuxiliar} becomes autonomous and corresponds to the linearization of the equations for $(S_n)$ and $(V_n)$ in the classical (autonomous) SIRVS model.
In order to proceed we need to recall some notions. A solution $(u_n)$ of some system of difference equations $u_{n+1}=f_n(u_n)$ is said to be \emph{attractive} if for all $n_0 \in \ensuremath{\mathds N}$ and all $\varepsilon>0$ there is $\sigma(n_0)>0$ and $T(\varepsilon,n_0,u_0) \in \ensuremath{\mathds N}$ such that if $(\bar u_n)$ is a solution with $\|u_0-\bar u_0\|<\sigma(n_0)$ then $\|u_n-\bar u_n\|<\varepsilon$, for all $n \ge n_0+T(\varepsilon,n_0,u_0)$. Additionally, if some solution is attractive and we can take $T$ to be only dependent on $\varepsilon$, we say that it is \emph{uniformly attractive}.
The following theorem furnishes some simple properties of system~\eqref{eq:SistemaAuxiliar}.
\begin{lemma}[Lemma 2.2 of~\cite{Zhang-AMC-2015}]\label{lema:auxSystem}
Assume that conditions~H\ref{cond-P1}) and~H\ref{cond-P2}) hold. Then
\begin{enumerate}[i)]
\item
all solutions $(x_n,y_n)$ of system ~\eqref{eq:SistemaAuxiliar} with initial condition $x_0 \ge 0$ and $y_0 \ge 0$ are nonnegative for all
$n \in \ensuremath{\mathds N}_0$;
\item \label{cond-2-aux} each fixed solution $(x_n,y_n)$ of~\eqref{eq:SistemaAuxiliar}
is bounded and globally uniformly attractive for all $n \in \ensuremath{\mathds N}_0$;
\item \label{cond-3-aux} if $(x_n,y_n)$ is a solution of~\eqref{eq:SistemaAuxiliar} and $(\tilde x_n, \tilde y_n)$
is a solution of the system
\begin{equation}\label{eq:SistemaAuxiliar2}
\begin{cases}
x_{n+1}=\dfrac{\Lambda_n + \eta_n y_{n+1} +x_n+f_n}{1+\mu_n+p_n} \\[2mm]
y_{n+1}=\dfrac{p_n x_{n+1} +y_n+g_n}{1+\mu_n+\eta_n}
\end{cases}.
\end{equation}
with $(\tilde
x_0, \tilde y_0)=(x_0,y_0)$ then there is a constant $L>0$,
only depending on $\mu_n$, satisfying
\[
\sup_{n \in \ensuremath{\mathds N}_0} \left\{ |\tilde x_n - x_n| + |\tilde y_n -y_n|\right\} \le L \sup_{n \in \ensuremath{\mathds N}_0} \left( |f_n|+|g_n| \right);
\]
\item \label{cond-4-aux}
there exists constants $m,M>0$ such that,
for each solution $(x_n,y_n)$ of~\eqref{eq:SistemaAuxiliar}, we have
\[
\begin{split}
& m \le \liminf_{n \to \infty} x_n \le \limsup_{n \to \infty} x_n \le M, \\
& m \le \liminf_{n \to \infty} y_n \le \limsup_{n \to \infty} y_n \le M.
\end{split}
\]
\item \label{cond-5-aux} when the system~\eqref{eq:SistemaAuxiliar} is $\omega$--periodic, it has a unique positive $\omega$--periodic solution which is globally uniformly attractive.
\end{enumerate}
\end{lemma}
We have the following lemma
\begin{lemma}\label{lema:system}
Assume that condition ~H\ref{cond-P3}) holds. Then we have the following:
\begin{enumerate}[i)]
\item \label{cond-1-bs} all solutions $(S_n,I_n,R_n,V_n)$ of~\eqref{eq:ProblemaPrincipal}
with nonnegative initial conditions are nonnegative for all $n \in \ensuremath{\mathds N}_0$;
\item \label{cond-2-bs} all solutions $(S_n,I_n,R_n,V_n)$ of~\eqref{eq:ProblemaPrincipal}
with positive initial conditions are positive for all $n \in \ensuremath{\mathds N}_0$;
\item \label{cond-3-bs} there is a constant $M > 0$ such that, if $(S_n,I_n,R_n,V_n)$ is a solution
of~\eqref{eq:ProblemaPrincipal} with nonnegative initial conditions then
\[
\limsup_{n \to +\infty} S_n +I_n + R_n + V_n < M.
\]
\end{enumerate}
\end{lemma}
\begin{proof}
See Appendix~\ref{Appendix:A}.
\end{proof}
For each $\lambda$ and each particular solution $\xi^{*}_n=(x^{*}_n,y^{*}_n)$ of~\eqref{eq:SistemaAuxiliar} with $x^{*}_0>0$ and $y^{*}_0>0$
we define the numbers
\begin{equation}\label{eq:discrete-threshold-liminf-ell}
\mathcal R_D^\ell(\xi^{*},\lambda) = \liminf_{n \to \infty} \prod_{k=n}^{n+\lambda} \frac{1+\beta_k \partial_2 \varphi(x^{*}_{k+1},0)+\sigma_k \partial_2 \psi(y^{*}_{k+1},0)}{1+\mu_k+\alpha_k+\gamma_k}
\end{equation}
and
\begin{equation}\label{eq:discrete-threshold-limsup-u}
\mathcal R_D^u(\xi^{*},\lambda) = \limsup_{n \to \infty} \prod_{k=n}^{n+\lambda} \frac{1+\beta_k \partial_2 \varphi(x^{*}_{k+1},0)+\sigma_k \partial_2 \psi(y^{*}_{k+1},0)}{1+\mu_k+\alpha_k+\gamma_k},
\end{equation}
where $\partial_i f$ denotes the partial derivative of $f$ with respect to the $i$-th variable. Contrarily to what one could expect, the next lemma shows that
the numbers above do not depend on the particular solution $\xi_n=(x_n,y_n)$ of~\eqref{eq:SistemaAuxiliar} with $x_n(0)>0$ and $y_n(0)>0$.
\begin{lemma} \label{lema:indep}
Assume that~H\ref{cond-H1}),~H\ref{cond-P1}) and~H\ref{cond-P2}) hold. If $(\xi_1^*)_n=((x_1)^*_n,(y_1)^*_n)$ and $(\xi_2^*)_n=((x_2)^*_n,(y_2)^*_n)$ are
two solutions of~\eqref{eq:SistemaAuxiliar} with $x_i^*(0)>0$ and
$y_i^*(0)>0$, $i=1,2$, then
\[
\mathcal R_D^\ell(\xi_1^*,\lambda)=\mathcal R_D^\ell(\xi_2^*,\lambda), \quad \text{and} \quad
\mathcal R_D^u(\xi_1^*,\lambda)=\mathcal R_D^u(\xi_2^*,\lambda).
\]
\end{lemma}
\begin{proof}
See Appendix~\ref{Appendix:A}.
\end{proof}
By lemma ~\ref{lema:indep} we can drop the dependence of the particular solution and simply write $\mathcal R_D^\ell(\lambda)$ and $\mathcal R_D^u(\lambda)$ instead of $\mathcal R_D^\ell(\xi^*,\lambda)$ an $\mathcal R_D^u(\xi^*,\lambda)$ respectively.
\subsection{Extinction and permanence} \label{subsection:EP}
We have the following result about the extinction of the disease:
\begin{theorem}[Extinction of the disease] \label{teo:Extinction}
Assume that conditions~H\ref{cond-H1}) to~H\ref{cond-P3}) hold. Then
\begin{enumerate}[a)]
\item \label{teo:Extinction-a} If there is a constant $\lambda>0$ such that $\mathcal R_D^u (\lambda) <1$, then the infectives $(I_n)$ go to extinction.
\item \label{teo:Extinction-b} Any solution $(x_n^{*}, 0, 0, y_n^{*} )$, where $(x_n^*,y_n^*)$ is a particular solution of system~\eqref{eq:SistemaAuxiliar}, is globally uniformly attractive.
\end{enumerate}
\end{theorem}
\begin{proof}
See Appendix~\ref{Appendix:A}.
\end{proof}
We have the following result about the permanence of the disease:
\begin{theorem}[Permanence of the disease] \label{teo:Permanence}
Assume that conditions~H\ref{cond-H1}) to~H\ref{cond-P3}) hold. If there is a constant $\lambda>0$
such that $\mathcal R_D^\ell(\lambda) >1$ then the infectives $(I_n)$ are permanent in system~\eqref{eq:ProblemaPrincipal}.
\end{theorem}
\begin{proof}
See Appendix~\ref{Appendix:A}.
\end{proof}
We consider now the particular periodic case: assume that all parameters of system~\eqref{eq:ProblemaPrincipal} are periodic with period $\omega \in \ensuremath{\mathds N}$. By~\ref{cond-5-aux}) in Lemma~\ref{lema:auxSystem}, there is an $\omega$-periodic disease-free solution of~\eqref{eq:SistemaAuxiliar}, $\xi^*=(x^*_n,y^*_n)_{n \in \ensuremath{\mathds N}}$. Thus, in the periodic setting,~\eqref{eq:discrete-threshold-liminf-ell} and~\eqref{eq:discrete-threshold-limsup-u} become both equal to
\begin{equation}\label{eq:discrete-threshold-per}
\mathcal R_D^{per}(\xi^{*}) = \prod_{k=0}^{\omega-1} \frac{1+\beta_k \partial_2 \varphi(x^{*}_{k+1},0)+\sigma_k \partial_2 \psi(y^{*}_{k+1},0)}{1+\mu_k+\alpha_k+\gamma_k}.
\end{equation}
Therefore we obtain the corollary:
\begin{corollary}[Periodic case] \label{teo:periodic}
Assume that all coefficients are $\omega$-periodic in~\eqref{eq:ProblemaPrincipal} and that conditions~H\ref{cond-H1}) to~H\ref{cond-P3}) hold. Then
\begin{enumerate}[a)]
\item \label{teo:Extinction-per-a} If $\mathcal R_D^{per}(\xi^{*})<1$ then the infectives $(I_n)$ go to extinction.
\item \label{teo:Extinction-per-b} The disease-free solution $(x_n^{*}, 0, 0, y_n^{*} )$, where $(x^*_n,y^*_n)_{n \in \ensuremath{\mathds N}}$ is an disease-free $\omega$-periodic solution of~\eqref{eq:SistemaAuxiliar}, is globally attractive.
\item \label{teo:Permanence-per} If $\mathcal R_D^{per}(\xi^{*})>1$ then the infectives $(I_n)$ are permanent.
\end{enumerate}
\end{corollary}
\begin{proof}
See Appendix~\ref{Appendix:A}.
\end{proof}
\section{Consistency} \label{section:consistency}
In this section, under the additional assumption that the parameter functions $\Lambda$, $\mu$, $\eta$ and $p$ are constant, we will get a result stating that when our integral conditions prescribe extinction (respectively persistence) for the continuous-time model, then the discrete-time conditions prescribe extinction (respectively persistence) for the corresponding discrete-time models, as long as the time step is less than some constant.
Throughout this section, we assume that the parameter functions $\Lambda$, $\mu$, $\eta$ and $p$ are constant functions and $\phi(h)$ will be the function used in the discretization of the derivative.
We consider the continuous time model~\eqref{eq:ProblemaPrincipal-cont} and, for a given time step $h$, the corresponding discrete time model, that is the discrete time model with parameters $\beta^h_k=\phi(h)\beta(kh)$, $\sigma^h_k=\phi(h)\sigma(kh)$, $\Lambda^h_k=\phi(h)\Lambda$, $\mu^h_k=\phi(h)\mu$, $p^h_k=\phi(h)p$, $\eta^h_k=\phi(h)\eta$, $\alpha^h_k=\phi(h)\alpha(kh)$ and $\gamma^h_k=\phi(h)\gamma(kh)$.
For a given time step $h>0$, the expressions $\mathcal R_D^\ell(\lambda)$ and $\mathcal R_D^u(\lambda)$ in~\eqref{eq:discrete-threshold-liminf-ell} and~\eqref{eq:discrete-threshold-limsup-u} become, in our context
\[
\mathcal R_D^\ell(\lambda,h) = \liminf_{n \to \infty} \prod_{k=n}^{n+\lambda} \frac{1+\beta^h_k \varphi(x^{*}_{k+1})+\sigma^h_k \psi(y^{*}_{k+1})}{1+\mu^h_k+\alpha^h_k+\gamma^h_k}
\]
and
\[
\mathcal R_D^u(\lambda,h) = \limsup_{n \to \infty} \prod_{k=n}^{n+\lambda} \frac{1+\beta^h_k \varphi(x^{*}_{k+1})+\sigma^h_k \psi(y^{*}_{k+1})}{1+\mu^h_k+\alpha^h_k+\gamma^h_k},
\]
where $(x^{*}_k,y^{*}_k)$ is the solution of the (in our context autonomous) system~\eqref{eq:SistemaAuxiliar}.
We have the following result.
\begin{theorem}\label{teo:consitency}
For system~\eqref{eq:ProblemaPrincipal-cont}, assume that $\Lambda(t)=\Lambda$, $\mu(t)=\mu$, $\eta(t)=\eta$ and $p(t)=p$ for all $t\ge 0$ and that the functions $\alpha(t)$, $\gamma(t)$, $\beta(t)$ and $\sigma(t)$ are differentiable, nonnegative, bounded and have bounded derivative. Assume also that conditions~C\ref{cond-C1}) to C\ref{cond-C3}) hold and let
$$h^u_{\text{max}}=-\dfrac{R^u_C(\lambda)}{\sup_{t \ge 0}|f'(t)|(\lambda+1)} \quad \text{and} \quad h^\ell_{\text{max}}=\dfrac{R^\ell_C(\lambda)}{\sup_{t \ge 0}|f'(t)|(\lambda+1)},$$
where
$$f(t)=\beta(t)\varphi\left(\frac{\Lambda(\mu+\eta)}{\mu(\mu+\eta+p)}\right)
+\sigma(t)\psi\left(\frac{p\Lambda}{\mu(\mu+\eta+p)}\right)-\mu-\alpha(t)-\gamma(t).$$ Then:
\begin{enumerate}[a)]
\item \label{teo:consitency-a} If $\mathcal R_C^u(\lambda)<0$ then $\mathcal R_D^u(\lfloor\lambda/h\rfloor,h)<1$ for all $h \in ]0,h^u_{\text{max}}[$;
\item \label{teo:consitency-b} If $\mathcal R_C^\ell(\lambda)>0$ then $\mathcal R_D^\ell(\lfloor\lambda/h\rfloor,h)>1$ for all for all $h \in ]0,h^\ell_{\text{max}}[$.
\end{enumerate}
\end{theorem}
\begin{proof}
Observe that $(x_n,y_n)=(a,b)$, $n \in \ensuremath{\mathds N}$ and $(x(t),y(t))=(a,b)$, $t \in \ensuremath{\mathds R}$, where
$$(a,b)=\left(\Lambda(\mu+\eta)/[\mu(\mu+\eta+p)],p\Lambda/[\mu(\mu+\eta+p)]\right),$$
are respectively solutions of system~\eqref{eq:SistemaAuxiliar} and system~\eqref{eq:SistemaAuxiliar-Pereira-Silva-Silva}. Thus
\[
\mathcal R_C^u(\lambda) = \limsup_{t \to \infty} \int_t^{t+\lambda} \beta(s) \varphi(a)+\sigma(s) \psi(b)-[\mu(s)+\alpha(s)+\gamma(s)] \, ds
\]
and
\[
\mathcal R_D^u(\lfloor\lambda/h\rfloor,h) = \limsup_{n \to \infty} \prod_{k=n}^{n+\lfloor\lambda/h\rfloor} \frac{1+\beta^h_k \varphi(a)+\sigma^h_k \psi(b)}{1+\mu^h_k+\alpha^h_k+\gamma^h_k}.
\]
By contradiction, assume that
\begin{equation}\label{eq:contrad-mRCu<0}
\mathcal R_C^u(\lambda)<0,
\end{equation}
and that there is a sequence $(h_m)_{m \in \ensuremath{\mathds N}}$ such that
$h_m \to 0$ as $m \to +\infty$ and
\begin{equation}\label{eq:contrad}
\mathcal R_D^u(\lfloor\lambda/h_m\rfloor,h_m) =\limsup_{n \to \infty} \prod_{k=n}^{n+\lfloor\lambda/h_m\rfloor} \frac{1+\beta^{h_m}_k \varphi(a)+\sigma^{h_m}_k \psi(b)}{1+\mu^{h_m}_k+\alpha^{h_m}_k+\gamma^{h_m}_k}\ge 1,
\end{equation}
for all $m \in \ensuremath{\mathds N}$. By~\eqref{eq:contrad}, we conclude that, for each $m \in \ensuremath{\mathds N}$, there are sequences $(h_m)_{m \in \ensuremath{\mathds N}}$ and $(n_{m,r})_{r \in \ensuremath{\mathds N}}$ such that
$h_m \to 0$ as $m \to +\infty$, $n_{m,r} \to +\infty$ as $r \to +\infty$ and
\begin{equation}\label{eq:contrad-a}
\begin{split}
& \prod_{k=n_{m,r}}^{n_{m,r}+\lfloor\lambda/h_m\rfloor}(1+\beta^{h_m}_k \varphi(a)+\sigma^{h_m}_k \psi(b))\\
& > (1-h_m) \prod_{k=n_{m,r}}^{n_{m,r}+\lfloor\lambda/h_m\rfloor}(1+\mu^{h_m}_k+\alpha^{h_m}_k+\gamma^{h_m}_k).
\end{split}
\end{equation}
By~\eqref{eq:contrad-a}, we have
\begin{equation}\label{eq:contrad-c}
\begin{split}
& \sum_{k=n_{m,r}}^{n_{m,r}+\lfloor\lambda/h_m\rfloor}(\beta^{h_m}_k \varphi(a)+\sigma^{h_m}_k \psi(b)-\mu^{h_m}_k-\alpha^{h_m}_k-\gamma^{h_m}_k) \\
& >( B_{n_{m,r},\lambda,h_m}-A_{n_{m,r},\lambda,h_m}-C_{n_{m,r},\lambda,h_m})/h_m,
\end{split}
\end{equation}
where
\begin{equation}\label{eq:An-lambda-h}
A_{n,L,h}:=-h+h\prod_{k=n}^{n+\lfloor L/h\rfloor}(1+\beta^h_k \varphi(a)+\sigma^h_k \psi(b))
-h\sum_{k=n}^{n+\lfloor L/h\rfloor}(\beta^h_k \varphi(a)+\sigma^h_k \psi(b)),
\end{equation}
\begin{equation}\label{eq:Bn-lambda-h}
B_{n,L,h}:=-h+h\prod_{k=n}^{n+\lfloor L/h\rfloor}(1+\mu^h_k+\alpha^h_k+\gamma^h_k)
-h\sum_{k=n}^{n+\lfloor L/h\rfloor}(\mu^h_k+\alpha^h_k+\gamma^h_k)
\end{equation}
and
\begin{equation}\label{eq:Cn-lambda-h}
C_{n,L,h}=
h \prod_{k=n}^{n+\lfloor L/h\rfloor}(1+\mu^{h}_k+\alpha^{h}_k+\gamma^{h}_k).
\end{equation}
and, multiplying both sides by $h_m$, we get
\Small{
\begin{equation}\label{eq:contrad-d}
\begin{split}
& \phi(h_m)\sum_{k=n_{m,r}}^{n_{m,r}+\lfloor\lambda/h_m\rfloor}h_m\left[\beta(kh_m) \varphi(a)+\sigma(kh_m) \psi(b)-\mu(kh_m)-\alpha(kh_m)-\gamma(kh_m)\right]\\
& > B_{n_{m,r},\lambda,h_m}-A_{n_{m,r},\lambda,h_m}-C_{n_{m,r},\lambda,h_m}.
\end{split}
\end{equation}
}
We also have
\begin{equation}\label{eq:maj-A-nlh-prev}
\begin{split}
|A_{n_{m,r},\lambda,h_m}|
& \le h_m \sum_{k=2}^{\lfloor\lambda/h_m\rfloor} \binom{\lfloor\lambda/h_m\rfloor}{k}[(\beta^u \varphi(a)+\sigma^u \psi(b))]^k[\phi(h_m)]^k\\
& \le h_m \sum_{k=0}^{\lfloor\lambda/h_m\rfloor} \binom{\lfloor\lambda/h_m\rfloor}{k}[(\beta^u \varphi(a)+\sigma^u \psi(b))]^k[\phi(h_m)]^k\\
& = h_m \left[1+(\beta^u \varphi(a)+\sigma^u \psi(b))\phi(h_m)\right]^{\lfloor\lambda/h_m\rfloor}.
\end{split}
\end{equation}
Noting that, by~\eqref{eq:derivative-denominator}, we have
\[
\begin{split}
& \lim_{m \to +\infty} \left[1+(\beta^u \varphi(a)+\sigma^u \psi(b))\phi(h_m)\right]^{\lfloor\lambda/h_m\rfloor}\\
& = \lim_{m \to +\infty} \left[\left(1+\frac{\beta^u \varphi(a)+\sigma^u \psi(b)}{1/\phi(h_m)}\right)^{1/\phi(h_m)}\right]^{\phi(h_m)\lfloor\lambda/h_m\rfloor}\\
& = \e^{(\beta^u \varphi(a)+\sigma^u \psi(b))\lambda}\\
\end{split}
\]
and that a convergent sequence is bounded, by~\eqref{eq:maj-A-nlh-prev} there is $C_1>0$ such that
\begin{equation}\label{eq:maj-A-nlh}
|A_{n_{m,r},\lambda,h_m}|
\le C_1 h_m.
\end{equation}
Similarly, we have
\begin{equation}\label{eq:maj-B-nlh-prev}
\begin{split}
|B_{n_{m,r},\lambda,h_m}|
& \le \sum_{k=2}^{\lfloor\lambda/h_m\rfloor} \binom{\lfloor\lambda/h_m\rfloor}{k}[(\mu^u+\alpha^u+\gamma^u)]^k[\phi(h_m)]^k\\
& \le \sum_{k=0}^{\lfloor\lambda/h_m\rfloor} \binom{\lfloor\lambda/h_m\rfloor}{k}[(\mu^u+\alpha^u+\gamma^u)]^k[\phi(h_m)]^k\\
& = h_m \left[1+(\mu^u+\alpha^u+\gamma^u)\phi(h_m)\right]^{\lfloor\lambda/h_m\rfloor}.
\end{split}
\end{equation}
Using~\eqref{eq:derivative-denominator} again, we get
\[
\lim_{m \to +\infty} \left[1+(\mu^u+\alpha^u+\gamma^u)\phi(h_m)\right]^{\lfloor\lambda/h_m\rfloor}
\le \e^{(\mu^u+\alpha^u+\gamma^u)\lambda},
\]
there is $C_2>0$ such that
\begin{equation}\label{eq:maj-B-nlh}
|B_{n_{m,r},\lambda,h_m}| \le C_2 h_m.
\end{equation}
Finally, we have
\[
\begin{split}
|C_{n_{m,r},\lambda,h_m}|
& = h_m \prod_{k=n_{m,r}}^{n_{m,r}+\lfloor \lambda/h_m\rfloor}(1+\mu^{h_m}_k+\alpha^{h_m}_k+\gamma^{h_m}_k)\\
& = h_m (1+3\phi(h_m)\max\{\mu^u,\alpha^u,\gamma^u\})^{\lfloor \lambda/h_m\rfloor+1}.
\end{split}
\]
According~\eqref{eq:derivative-denominator}, we obtain
\[
\begin{split}
& \lim_{m \to +\infty} (1+3\phi(h_m)\max\{\mu^u,\alpha^u,\gamma^u\})^{\lfloor \lambda/h_m\rfloor+1}\\
& =\lim_{m \to +\infty} \left[\left(1+\frac{3\max\{\mu^u,\alpha^u,\gamma^u\}}{1/\phi(h_m)}\right)^{1/\phi(h_m)}\right]^{\phi(h_m)(\lfloor \lambda/h_m\rfloor+1)}\\
& = \e^{3\max\{\mu^u,\alpha^u,\gamma^u\}\lambda},
\end{split}
\]
there is $C_3>0$ such that
\begin{equation}\label{eq:maj-C-nlh}
|C_{n_{m,r},\lambda,h_m}| \le C_3 h_m.
\end{equation}
Thus
\begin{equation}\label{eq:to-zero-uniformly}
\begin{split}
& B_{n_{m,r},\lambda,h_m}-A_{n_{m,r},\lambda,h_m}-C_{n_{m,r},\lambda,h_m}\\
& \le |A_{n_{m,r},\lambda,h_m}|+|B_{n_{m,r},\lambda,h_m}|+|C_{n_{m,r},\lambda,h_m}|\\
& \le (C_1+C_2+C_3)h_m,
\end{split}
\end{equation}
for all $m \ge M$. Since the right hand side of~\eqref{eq:to-zero-uniformly} is independent of $n_{m,r}$, we conclude that
\begin{equation}\label{eq:(An-Bn)/phi-to-0}
B_{n_{m,r},\lambda,h_m}-A_{n_{m,r},\lambda,h_m}-C_{n_{m,r},\lambda,h_m} \to 0,
\end{equation}
as $m \to +\infty$, uniformly in $r$.
On the other hand we note that the $C^1$ function $f:\ensuremath{\mathds R}^+_0 \to \ensuremath{\mathds R}$ given by
\begin{equation}\label{eq:funct-integ}
f(t)=\beta(t) \varphi(a)+\sigma(t) \psi(b)-\mu(t)-\alpha(t)-\gamma(t)
\end{equation}
is Riemann-integrable on any bounded interval $I \subset \ensuremath{\mathds R}_0^+$.
We have that
\[
\begin{split}
\sum_{k=n_{m,r}}^{n_{m,r}+\lfloor\lambda/h_m\rfloor} h_mf(kh_m)
+(\lambda-\lfloor\lambda/h_m\rfloor h_m)f(n_{m,r} h_m+\lfloor\lambda/h_m\rfloor h_m),
\end{split}
\]
is a Riemann sum of
$$\int_{n_{m,r} h_m}^{n_{m,r} h_m+\lambda} \beta(s) \varphi(a)+\sigma(s) \psi(b)-[\mu(s)+\alpha(s)+\gamma(s)] \, ds$$
with respect to the partition
$$\{n_{m,r}h_m,n_{m,r}h_m+h_m,\ldots,n_{m,r}h_m+\lfloor\lambda/h_m\rfloor h_m,n_{m,r}h_m+\lambda\}$$
of size $h_m$ of the interval $[n_{m,r},n_{m,r}+\lambda]$. Note that
\[
s_{m,r}:=(\lambda-\lfloor\lambda/h_m\rfloor h_m)f(n_{m,r}h_m+\lfloor\lambda/h_m\rfloor h_m) \le h_mf^u :=s_m
\]
and $s_m \to 0$ as $m \to +\infty$, uniformly in $r$.
Since $f$ is $C^1$ with bounded derivative, for any $h>0$ we have
$$|f(x)-f(x+h)|\le Ch,$$
where $\displaystyle C=\sup_{t\ge 0} |f'(t)|$. We conclude that
\begin{equation}\label{eq:sum-int-to-0}
\begin{split}
&\left|\sum_{k=n_{m,r}}^{n_{m,r}+\lfloor\lambda/h_m\rfloor} h_mf(kh_m)+s_{m,r} - \int_{n_{m,r} h_m}^{n_{m,r} h_m+\lambda} f(s)\, ds\right|\\
& < Ch_m^2\lfloor\lambda/h_m \rfloor+Ch_m^2\\
& < C(\lambda+1)h_m,
\end{split}
\end{equation}
thus
\[
\sum_{k=n_{m,r}}^{n_{m,r}+\lfloor\lambda/h_m\rfloor} h_mf(kh_m)<\int_{n_{m,r} h_m}^{n_{m,r} h_m+\lambda} f(s)\, ds-s_{m,r}+C(\lambda+1)h_m,
\]
and therefore
\Small{
\begin{equation}\label{eq:aprox-int-sum}
\sum_{k=n_{m,r}}^{n_{m,r}+\lfloor\lambda/h_m\rfloor} \phi(h_m)h_mf(kh_m)<\phi(h_m)\left[\int_{n_{m,r} h_m}^{n_{m,r} h_m+\lambda} f(s)\, ds+C(\lambda+1)h_m\right].
\end{equation}
}
By~\eqref{eq:aprox-int-sum} we conclude that, given $\delta >0$, there is $r_m \in \ensuremath{\mathds N}$ such that, for all $r\ge r_m$,
\begin{equation}\label{eq:contradiction}
\phi(h_m) \sum_{k=n_{m,r_m}}^{n_{m,r_m}+\lfloor\lambda/h_m\rfloor} h_mf(kh_m)<\phi(h_m)\left[\mathcal R_C^u(\lambda)+\delta+C(\lambda+1)h_m\right].
\end{equation}
Finally, recalling that $\mathcal R_C^u(\lambda)<0$, by assumption, by the arbitrariness of $\delta>0$ and the fact that $h_m \to 0$ as $m \to +\infty$, we obtain for sufficiently large $m \in \ensuremath{\mathds N}$,
$$0 \le \phi(h_m) \sum_{k=n_{m,r_m}}^{n_{m,r_m}+\lfloor\lambda/h_m\rfloor} h_mf(kh_m)<0,$$
which is a contradiction. We obtain~\ref{teo:consitency-a}).
A similar argument allow us to prove~\ref{teo:consitency-b}). In fact, assuming by contradiction that
\begin{equation}\label{eq:contrad-mRCu>0}
\mathcal R_C^\ell(\lambda)>0,
\end{equation}
and that there is a sequence $(h_m)_{m \in \ensuremath{\mathds N}}$ such that
$h_m \to 0$ as $m \to +\infty$ and
\[
\mathcal R_D^\ell(\lfloor\lambda/h_m\rfloor,h_m) =\liminf_{n \to \infty} \prod_{k=n}^{n+\lfloor\lambda/h_m\rfloor} \frac{1+\beta^{h_m}_k \varphi(a)+\sigma^{h_m}_k \psi(b)}{1+\mu^{h_m}_k+\alpha^{h_m}_k+\gamma^{h_m}_k} \le 1,
\]
it is possible to conclude that
\begin{equation}\label{eq:contrad-d--per}
\begin{split}
& \phi(h_m)\sum_{k=n_{m,r}}^{n_{m,r}+\lfloor\lambda/h_m\rfloor}h_m(\beta(kh_m) \varphi(a)+\sigma(kh_m) \psi(b)-\mu(kh_m)-\alpha(kh_m)-\gamma(kh_m))\\
& < B_{n_{m,r},\lambda,h_m}-A_{n_{m,r},\lambda,h_m}+C_{n_{m,r},\lambda,h_m},
\end{split}
\end{equation}
where $A_{n_{m,r},\lambda,h_m}$, $B_{n_{m,r},\lambda,h_m}$ and $C_{n_{m,r},\lambda,h_m}$ are given respectively by~\eqref{eq:An-lambda-h},~\eqref{eq:Bn-lambda-h} and~\eqref{eq:maj-C-nlh} and still satisfy~\eqref{eq:maj-A-nlh},~\eqref{eq:maj-B-nlh} and~\eqref{eq:maj-C-nlh}. Consequently, given $\delta >0$, there is $r_m \in \ensuremath{\mathds N}$ such that, for all $r\ge r_m$,
\begin{equation}\label{eq:aprox-int-sum-per}
\phi(h_m)\sum_{k=n_{m,r}}^{n_{m,r}+\lfloor\lambda/h_m\rfloor} h_mf(kh_m)>\phi(h_m)\left[\int_{n_{m,r} h_m}^{n_{m,r}h_m+\lambda} f(s)\, ds-\delta+C(\lambda+1)h_m\right].
\end{equation}
Recalling that $\mathcal R_C^\ell(\lambda)>0$, by assumption and since $\delta>0$ is arbitrary, we obtain for sufficiently large $m \in \ensuremath{\mathds N}$,
$$0 \ge \phi(h_m)\sum_{k=n_{m,r_m}}^{n_{m,r_m}+\lfloor\lambda/h_m\rfloor} f(kh_m)>0,$$
which is a contradiction. We obtain~\ref{teo:consitency-b}) and the theorem follows.
\end{proof}
Next, for each $L \in \ensuremath{\mathds N}$, we give an example of a periodic system of period $1$ such that the continuous and the discrete time system with time step $h=1/L$ are not consistent, namely we will have persistence for the continuous time model and extinction for the discrete time model with time step $h=1/L$.
\begin{example}
Let $L \in \ensuremath{\mathds N}$. Consider in system~\eqref{eq:ProblemaPrincipal-cont} that $\phi(x)=\psi(x)=x$, that, with the exception of $\sigma$ and $\beta$, all parameters are constant, that $\Lambda=\mu$ and that
$$\sigma(t)=\beta(t)=d[1+c\sin^2(2\pi L t)(1+\cos(2\pi t))].$$
We obtain a periodic system of period $1$.
In this context, $(x_n,y_n)=(a,b)$, $n \in \ensuremath{\mathds N}$, and $(x(t),y(t))=(a,b)$, $t \in \ensuremath{\mathds R}$, where
$$(a,b)=\left((\mu+\eta)/(\mu+\eta+p),p/\mu+\eta+p)\right),$$
are respectively solutions of system~\eqref{eq:SistemaAuxiliar} and system~\eqref{eq:SistemaAuxiliar-Pereira-Silva-Silva}.
It is now possible to compute the number $\mathcal R_C^\ell(1)$. In fact, noting that $x^*(t)+y^*(t)=1$, we get
\[
\begin{split}
\mathcal R_C^\ell(1)
& =\int_0^1 \beta(s)x^*(s)+\sigma(s)y^*(s)-\mu-\alpha-\gamma \, ds \\
& =\int_0^1 d[1+c\sin^2(2\pi L t)(1+\cos(2\pi t))] \, ds -\mu-\alpha-\gamma\\
& =d(1+c/2)-\mu-\alpha-\gamma.
\end{split}
\]
We can also compute $\mathcal R_D^\ell(1,1/L)$. Namely we have
$$\mathcal R_D^\ell(1,1/L)=\frac{1+d/L}{1+\mu+\alpha+\gamma}.$$
If we let $d$ be sufficiently small so that $d<(\mu+\alpha+\gamma)L$, or in other words, $d<(\mu+\alpha+\gamma)$ and $c$ be sufficiently large so that $c>\frac{2}{d}(\mu+\gamma+\alpha-d)$, we obtain
$$\mathcal R_C^\ell(1)>1 \quad \Leftrightarrow \quad \frac{1+d(1+c/2)}{1+\mu+\alpha+\gamma}>1$$
and
$$\mathcal R_D^\ell(1,1/L)<1 \quad \Leftrightarrow \quad \frac{1+d/L}{1+\mu+\alpha+\gamma}<1.$$
so we conclude that we don't have consistency for time step $1/L$.
Let $L=6$ and consider the continuous model with the following parameters $\mu=\Lambda=0.25$, $\gamma=0.3$, $\alpha=0.05$, $\eta=0.05$, $p=2/3$, $d=0.6$ and $c=1.5$. In figure~\ref{fig-exemplo}, we plot function $\beta$ (or similarly $\sigma$) and the component $I(t)$ of the solution of system~\eqref{eq:ProblemaPrincipal-cont} given by the solver of Mathematica$^\circledR$ (that we take to represent the solution of the continuous-time model) and the solution of the discrete-time model~\eqref{eq:ProblemaPrincipal-disc} with time step $1/6$. As can be seen, the infectives are persistent in the continuous-time model but go to extinction in the discrete-time model. We have inconsistency in this case.
\begin{figure}[h]
\begin{minipage}[b][2.7cm]{.4\linewidth}
\includegraphics[scale=0.4]{beta.eps}
\end{minipage}
\begin{minipage}[b][2.7cm]{.2\linewidth}
\includegraphics[scale=0.4]{EXEMPLO.eps}
\end{minipage}
\caption{Left: function $\beta$; right: inconsistency (time-step=1/6).}
\label{fig-exemplo}
\end{figure}
\end{example}
Note that, changing $\beta(t)$ and $\sigma(t)$ slightly, we can construct an example of a periodic system with period $1$ where the infectives in the continuous time model goes to extinction but, in the discrete time model with time step $h=1/L$, the infectives are persistent.
Furthermore, we emphasize that this lack of consistency is not a result of the discretization method used but simply a result of the fact that the time steps lead to a situation where the points $n/L$ where the functions $\beta$ and $\sigma$ are evaluated (in order to obtain the discrete time parameters) correspond to minimums of $\beta$ and $\sigma$.
\section{Simulation} \label{section:simulation}
Our objective in this section is twofold. On the one hand, we want to consider different incidence functions $\varphi$, corresponding to different discretizations of our continuous model, and compare the several discrete models obtained. We do this in the first subsection. On the other hand, we want to use our model to describe a real situation. We do this in the second subsection where we consider data from the incidence of measles in France in the period 2012-2016.
\subsection{Simulation with several NSFD schemes} \label{subsection:simulation1}
In this subsection we do some simulation to illustrate our results. To begin, we compare our model~\eqref{eq:ProblemaPrincipal} with mass action incidence ($\varphi(S,I)=SI$ and $\psi(V,I)=VI$) with Zhang's model~\cite{Zhang-AMC-2015}. We use the following set of parameters: $\phi(h)=h+0.2 h^2$,
$\Lambda=0.5$, $\mu(t)=\gamma(t)=\delta(t)=0.3$, $\alpha(t)=0.05$, $\eta=0.05$, $p=2/3$ and
$$\beta(t)=\sigma(t)=b(1+0.3\cos(t\pi/2)).$$
Setting $b=0.3$ we obtain $\mathcal R^u_C(4)=-0.6<0$ and thus we conclude that we have extinction for the continuous model. Taking time-steps equal to $4$, $1$ and $0.5$, we get $\mathcal R_D^u(0,4)=\mathcal R_D^\ell(0,4)=1$, $\mathcal R_D^u(3,1)=0.644<1$ and $\mathcal R_D(7,0.5)=0.601<1$ and we conclude that we have extinction for time steps $1$ and $0.5$. For these parameters, we have consistency in the sense of Theorem~\ref{teo:consitency} as long as the time step is less than $0.05$. Clearly, there is numerical evidence that there is consistency even for higher time steps.
Figure~\ref{fig-compare-Mick-MickZhang-ext} illustrates this situation.
\begin{figure}[h]
\begin{minipage}[b][2.7cm]{.3\linewidth}
\includegraphics[scale=0.4]{OMELHOR_Per_E_4.eps}
\end{minipage}
\begin{minipage}[b][2.7cm]{.3\linewidth}
\includegraphics[scale=0.4]{OMELHOR_Per_E_1.eps}
\end{minipage}
\begin{minipage}[b][2.7cm]{.3\linewidth}
\includegraphics[scale=0.4]{OMELHOR_Per_E_05.eps}
\end{minipage}
\caption{$SI$; $\phi(h)=h+0.2 h^2$; step-size: $4$, $1$, $0.5$.}
\label{fig-compare-Mick-MickZhang-ext}
\end{figure}
Changing $b$ to $0.9$ we obtain $\mathcal R^\ell_C(4)=3.4>0$ and thus we conclude that we have persistence.
Taking time-steps equal to $2$, $1$ and $0.5$, we get $\mathcal R_D^\ell(1,2)=3.201>1$, $\mathcal R_D^\ell(3,1)=5.9>1$ and $\mathcal R_D^\ell(7,0.5)=10.2>1$ and we conclude that we have persistence for all these time steps. Figure~\ref{fig-compare-Mick-MickZhang-per} illustrates this situation. Figure~\ref{fig-compare-Mick-MickZhang-ext} and figure~\ref{fig-compare-Mick-MickZhang-per} suggest that numerically our model is slightly better that Zhang's model, at least for large time steps.
\begin{figure}[h]
\begin{minipage}[b][2.7cm]{.3\linewidth}
\includegraphics[scale=0.4]{OMELHOR_Per_P_2.eps}
\end{minipage}
\begin{minipage}[b][2.7cm]{.3\linewidth}
\includegraphics[scale=0.4]{OMELHOR_Per_P_1.eps}
\end{minipage}
\begin{minipage}[b][2.7cm]{.3\linewidth}
\includegraphics[scale=0.4]{OMELHOR_Per_P_05.eps}
\end{minipage}
\caption{$SI$; $\phi(h)=h+0.2 h^2$; step-size: $2$, $1$, $0.5$.}
\label{fig-compare-Mick-MickZhang-per}
\end{figure}
Next, we compare our model with the discretized model obtained by Euler method and the output of the Mathematica$^\circledR$ solver ODE (that uses a Runge-Kutta method). Considering $b=0.3$, we get extinction for the continuous time model, as we already saw. Taking time steps equal to $2$, $1$ and $0.5$, we can see in figure~\ref{fig-compare-Mick-Euler-RK-ext} that for all methods considered and all time steps we have extinction, although the behaviour of our model shadows better the behaviour given by Mathematica's solver, at least for these time steps.
\begin{figure}[h]
\begin{minipage}[b][2.7cm]{.3\linewidth}
\includegraphics[scale=0.4]{Eul_MM_Per_E_2.eps}
\end{minipage}
\begin{minipage}[b][2.7cm]{.3\linewidth}
\includegraphics[scale=0.4]{Eul_MM_Per_E_1.eps}
\end{minipage}
\begin{minipage}[b][2.7cm]{.3\linewidth}
\includegraphics[scale=0.4]{Eul_MM_Per_E_05.eps}
\end{minipage}
\caption{$SI$; $\phi(h)=h+0.2 h^2$; step-size: $2$, $1$, $0.5$.}
\label{fig-compare-Mick-Euler-RK-ext}
\end{figure}
Changing $b$ to $0.9$ we already saw that we get persistence for the continuous model.
Figure~\ref{fig-compare-Mick-Euler-RK-per} illustrates this situation.
\begin{figure}[h]
\begin{minipage}[b][2.7cm]{.3\linewidth}
\includegraphics[scale=0.4]{Eul_MM_Per_P_2.eps}
\end{minipage}
\begin{minipage}[b][2.7cm]{.3\linewidth}
\includegraphics[scale=0.4]{Eul_MM_Per_P_1.eps}
\end{minipage}
\begin{minipage}[b][2.7cm]{.3\linewidth}
\includegraphics[scale=0.4]{Eul_MM_Per_P_05.eps}
\end{minipage}
\caption{$SI$; $\phi(h)=h+0.2 h^2$; step-size: $2$, $1$, $0.5$.}
\label{fig-compare-Mick-Euler-RK-per}
\end{figure}
Next, we change our incidence function and consider $\phi(S,I)=SI/(1+0.7I)$, maintaining the set of parameters. Letting $b=0.3$ we have extinction for the continuous model and letting $b=0.9$ we have persistence for the continuous time model. Note that the thresholds $\mathcal R_C^u$, $\mathcal R_C^\ell$, $\mathcal R_D^u$ and $\mathcal R_D^\ell$ are similar to the mass action case. Figures~\ref{fig-compare-Mick-Euler-RK-ext-Hol} and~\ref{fig-compare-Mick-Euler-RK-per-Hol} illustrate this situation.
\begin{figure}[h]
\begin{minipage}[b][2.7cm]{.3\linewidth}
\includegraphics[scale=0.4]{Eul_MM_H_Per_E_2.eps}
\end{minipage}
\begin{minipage}[b][2.7cm]{.3\linewidth}
\includegraphics[scale=0.4]{Eul_MM_H_Per_E_1.eps}
\end{minipage}
\begin{minipage}[b][2.7cm]{.3\linewidth}
\includegraphics[scale=0.4]{Eul_MM_H_Per_E_05.eps}
\end{minipage}
\caption{$\frac{SI}{1+0.7I}$; $\phi(h)=h+0.2 h^2$; step-size: $2$, $1$, $0.5$.}
\label{fig-compare-Mick-Euler-RK-ext-Hol}
\end{figure}
\begin{figure}[h]
\begin{minipage}[b][2.7cm]{.3\linewidth}
\includegraphics[scale=0.4]{Eul_MM_H_Per_P_2.eps}
\end{minipage}
\begin{minipage}[b][2.7cm]{.3\linewidth}
\includegraphics[scale=0.4]{Eul_MM_H_Per_P_1.eps}
\end{minipage}
\begin{minipage}[b][2.7cm]{.3\linewidth}
\includegraphics[scale=0.4]{Eul_MM_H_Per_P_05.eps}
\end{minipage}
\caption{$\frac{SI}{1+0.7I}$; $\phi(h)=h+0.2h^2$; step-size: $2$, $1$, $0.5$.}
\label{fig-compare-Mick-Euler-RK-per-Hol}
\end{figure}
Doing corresponding simulations and comparisons for our model with $\phi(h)=(1-\e^{-0.002h})/(0.002)$ instead of $\phi(h)=h+0.2h^2$ we can draw the same conclusions regarding extinction/persistence, relation to Zhang's model and the model obtained by Euler method.
\subsection{Simulation with real data} \label{subsection:simulation2}
In this subsection, we present some simulation regarding measles. This disease is endemic in some countries such as France. In that country, with the measles outbreak in 2011, it was introduced a vaccination policy that lowered the number of reported cases. We will focus on measles in France, between 2012-2016. For a study concerning the period before 2012 see \cite{Bacaer-2014}. For our parameters estimation, we gathered information from several websites. We considered standard incidence functions $\psi(V_{n+1}, I_n) = {V_{n+1}I_{n}}/{P_n}$ and $\varphi(S_{n+1}, I_n)={S_{n+1}I_n}/{P_n}$, where $P_n$ is the total population. Inspired in the time series for the infectives (https://ecdc.europa.eu), we considered $\sigma_n=0.03$ and $\beta_n$ given by
\[
\beta_n=\left\{
\begin{array}{ll}
3.8+10 \sin\left(\frac{(n+1)\pi}{6}\right), & \hbox{if $\Bigl \lfloor\frac{n}{12} \Bigr \rfloor \leq 5$} \\
2.7, & \hbox{otherwise}
\end{array}.
\right.
\]
The remaining parameters were considered time independent and were inspired in data contained in the websites www.worldbank.org, https://data.oecd.org and www.geoba.se. Namely, we took the mortality rate $\mu_n=0.0007$, the newborns $\Lambda_n=50000$, the disease induced mortality $\alpha_n=0.000375$, the immunity loss $\eta_n=0.001$, the vaccination rate $p_n=0.001$ and the recovery rate $\gamma_n=0.957$. We used the initial conditions $S_0=7.20428\times10^6$, $I_0=106$, $V_0=5.84372\times10^7$ and $R_0=1.81918\times10^4$. In Figure~\ref{fig-Measles-Simul} we plot the real data for the infectives and the output given by our model.
\begin{figure}[h]
\centering
\includegraphics[scale=0.7]{finally.eps}
\caption{Measles (2012-2016), Simulation}
\label{fig-Measles-Simul}
\end{figure}
Can be seen, in a general way, that our model behaves in the same manner as the real data. It seems that if the vaccination policy in France, continues to be very strict, it may decrease the number of cases
\section{conclusions} \label{section:conclusions}
We considered a discretization procedure, based on Mickens NSFD scheme, to get a discrete-time model from a continuous time with vaccination and incidence given by a general function. For a family of models containing the previous discrete-time model, we achieved results on the persistence and the extinction of the disease (Theorems~\ref{teo:Extinction} and~\ref{teo:Permanence}). They contain the results of Zhang~\cite{Zhang-AMC-2015} as a particular case. Our threshold conditions depend on the parameters of the model and of the incidence function derivative, with respect to the infectives, computed on some disease-free solution. This agrees with the continuous counterparts of these results~\cite{Pereira-Silva-Silva}.
We also considered the problem of establishing the consistency of the continuous-time model and the discrete-time model for small time-steps, in the sense that if the time step is small enough when we have persistence (respectively extinction) for the continuous-time model we also have persistence (respectively extinction) for the discrete-time model (at least for situations where Theorems~\ref{teo:Permanence-Pereira-Silva-Silva} and~~\ref{teo:Extinction-Pereira-Silva-Silva} allow us to conclude that we have persistence or extinction). Assuming the differentiability of parameters, our result on this direction, Theorem~\ref{teo:consitency}, furnishes an interval $[0, a]$, where $a$ depend only on the parameters of the model and their derivatives, where there is consistency.
We present an example of a periodic system of period $1$ where the continuous and the discrete-time system with time-step $h=1/L$ are not consistent. Namely, for that time-step, we will have persistence for the continuous time model and extinction for the discrete-time model. These examples show the importance of knowing that for time steps smaller than some explicit value we have consistency, a type of result like the one in Theorem~\ref{teo:consitency}.
Finally, we carried out some simulations to illustrate our results. As one might expect our simulations furnish evidence that we may have consistency in intervals whose length are several times bigger than the length of the given interval in Theorem~\ref{teo:consitency}. Additionally, we used our model to describe a real situation, namely the case of measles incidence in France in the period 2012-2016, and compared our results with the real-time series for the infectives. We found in general, the predictive behaviour of our model very similar to the real data.
|
\section{Effective Hamiltonian}
To calculate the energies and eigenstates of the rotational levels in a particular vibrational state, we diagonalize an effective Hamiltonian. See, for example, ref.~\cite[Eq.~10.114--10.115]{Sbrown_carrington}. We include the electronic and vibrational state energy $T_v$, spin--orbit coupling $A_v$, and rigid-body rotation $B_v$. As discussed below, higher-order terms such as centrifugal distortion $D_v$ or $\Lambda$-doubling are not necessary at our precision.
Eigenstates in both $X\,^2\Pi_g$ and $a\,^4\Pi_u$ are written in the Hund's case-(a) basis:
\begin{align}
&c_{3/2}\left|^2\Pi_{3/2}\right\rangle + c_{1/2}\left|^2\Pi_{1/2}\right\rangle \\
c_{5/2}\left|^4\Pi_{5/2}\right\rangle + &c_{3/2}\left|^4\Pi_{3/2}\right\rangle + c_{1/2}\left|^4\Pi_{1/2}\right\rangle + c_{-1/2}\left|^4\Pi_{-1/2}\right\rangle .
\end{align}
In these bases, the effective Hamiltonians are given by
\begin{equation}
H(^2\Pi) = \begin{pmatrix}
T_v + \tfrac{A_v}{2} + B_v\left[J(J+1)-\tfrac{7}{4}\right] & -B_v\sqrt{J(J+1)-\tfrac{3}{4}} \\
-B_v\sqrt{J(J+1)-\tfrac{3}{4}} & T_v - \tfrac{A_v}{2}+B_v\left[J(J+1)+\tfrac{1}{4}\right]
\end{pmatrix}
\end{equation}
and
\begin{equation}
H(^4\Pi) = \left(\begin{smallmatrix}
T_v + \tfrac{3A_v}{2}+B_v\left[J(J+1)-\tfrac{19}{4}\right] & -\sqrt{3}B_v\sqrt{J(J+1)-\tfrac{15}{4}} & 0 & 0 \\
-\sqrt{3}B_v\sqrt{J(J+1)-\tfrac{15}{4}} & T_v + \tfrac{A_v}{2} + B_v\left[J(J+1)+\tfrac{5}{4}\right] & -2B_v\sqrt{J(J+1)-\tfrac{3}{4}} & 0 \\
0 & -2B_v\sqrt{J(J+1)-\tfrac{3}{4}} & T_v - \tfrac{A_v}{2} + B_v \left[J(J+1)+\tfrac{13}{4}\right] & -\sqrt{3}B_v(J+\tfrac{1}{2}) \\
0 & 0 & -\sqrt{3}B_v(J+\tfrac{1}{2}) & T_v - \tfrac{3A_v}{2}+B_v\left[J(J+1)+\tfrac{5}{4}\right]
\end{smallmatrix}\right) .
\end{equation}
The top-left component is the one with $\Omega = 3/2$ and $5/2$, respectively.
The parameters used in these Hamiltonians are listed in Table~\ref{tab:coef}. Although some identified transitions occur at fairly high $J$, the contributions of the $D_v$ coefficients are not important at the few-cm$^{-1}$ scale. The $D_v$ coefficients for the $|a\,^4\Pi_u, v^\ensuremath{\prime}=0, 1\rangle$ states are $5.0455(189)\times10^{-6}~{\rm cm}^{-1}$ and $5.0567(176)\times10^{-6}~{\rm cm}^{-1}$, respectively~\cite{ShansenJMS1983}. We extrapolate the $D_v$ coefficients of the $X\,^2\Pi_g$ state from merged parameters in ref.~\cite{ScoxonJMS1984} to obtain $D_{21} = 5.43(85)\times10^{-6}~{\rm cm}^{-1}$ and $D_{22} = 5.34(92)\times10^{-6}~{\rm cm}^{-1}$. Even at the higher $J$'s, the contributions from $D_v$ cancel to be consistent with zero with uncertainties of a few times $0.1~{\rm cm}^{-1}$.
\begingroup
\begin{table}[h]
\caption{Coefficients used in energy calculations. Uncertainties are shown in parentheses.}\label{tab:coef}
\begin{ruledtabular}
\begin{tabular}{clll}
state & $T_v / {\rm cm}^{-1}$ & $A_v / {\rm cm}^{-1}$ & $B_v / {\rm cm}^{-1}$ \\
\hline
$X\,^2\Pi_g, v=21$ & $129896.9(2.0)$\footnotemark[1] & $177.0(1.0)$\footnotemark[2]
& 1.25(3)\footnotemark[2] \\
$X\,^2\Pi_g, v=22$ & $131075(5)$\footnotemark[2] & $174.0(1.0)$\footnotemark[2]
& 1.25(1)\footnotemark[2] \\
$a\,^4\Pi_u, v=0$ & $129892(2)$\footnotemark[3] & $-47.7927(19)$\footnotemark[4]
& $1.096990(26)$\footnotemark[4] \\
$a\,^4\Pi_u, v=1$ & $130904(2)$\footnotemark[3] & $-47.7997(21)$\footnotemark[4]
& $1.081532(18)$\footnotemark[4]
\end{tabular}
\end{ruledtabular}
\footnotetext[1]{Ref.~\cite{SkongCJP1994}}
\footnotetext[2]{Ref.~\cite{SsongJCP1999}}
\footnotetext[3]{Ref.~\cite{SkongIJMSIP1996}}
\footnotetext[4]{Ref.~\cite{ShansenJMS1983}}
\end{table}
\endgroup
\newpage
\section{Tables of near-degeneracies}
Below are tables listing every pair of energy levels with $|\Delta E|<10~{\rm cm}^{-1}$ and $\Delta J = 0, \pm 1$. Also provided are the estimated linewidths and the eigenstate superposition coefficients from diagonalizing the effective Hamiltonians above. The uncertainties in $\Delta E$ are 3--6~${\rm cm}^{-1}$. They are highly correlated, however, such that even if these particular transitions are no longer within $10~{\rm cm}^{-1}$, others likely will be. By convention~\cite{SherzbergVolI1950,Sbrown_carrington}, the $F_i$ indicate the energy order of the eigenstates for a given $J$ with $F_1$ having the lowest energy. In the case (a) limit, the \OOp~$X\,^2\Pi_g$ state has $\Omega=\tfrac{1}{2}$ in $F_1$ and $\tfrac{3}{2}$ in $F_2$, while the $a\,^4\Pi_u$ state has $\Omega=\tfrac{5}{2}, \tfrac{3}{2}, \tfrac{1}{2}, -\tfrac{1}{2}$ in $F_{1,2,3,4}$.
\begin{longtable}{@{\extracolsep{\fill}}ccccrrrrrrrr}
\caption{The near-degeneracies $|X\,^2\Pi_g, v^\pp=21, J^\pp\rangle$ and $|a\,^4\Pi_u, v^\ensuremath{\prime}=0, J^\ensuremath{\prime}\rangle$, including the eigenstate superposition coefficients.}\\
\hline\hline
$X\,^2\Pi_g$ & $a\,^4\Pi_u$ & $J^\pp$ & $J^\ensuremath{\prime}$ & $\Delta E / {\rm cm}^{-1}$ & $\frac{\Gamma}{2\pi} / {\rm Hz}$ & $c_{3/2}^{\pp}$ & $c_{1/2}^\pp$ & $c_{5/2}^\ensuremath{\prime}$ & $c_{3/2}^\ensuremath{\prime}$ & $c_{1/2}^\ensuremath{\prime}$ & $c_{-1/2}^\ensuremath{\prime}$ \Tstrut\Bstrut\\
\hline \endfirsthead
\caption{(continued)}\Tstrut\\
\hline
$X\,^2\Pi_g$ & $a\,^4\Pi_u$ & $J^\pp$ & $J^\ensuremath{\prime}$ & $\Delta E / {\rm cm}^{-1}$ & $\frac{\Gamma}{2\pi} / {\rm Hz}$ & $c_{3/2}^{\pp}$ & $c_{1/2}^\pp$ & $c_{5/2}^\ensuremath{\prime}$ & $c_{3/2}^\ensuremath{\prime}$ & $c_{1/2}^\ensuremath{\prime}$ & $c_{-1/2}^\ensuremath{\prime}$ \Tstrut\Bstrut\\
\hline \endhead
$F_2$ & $F_4$ & $\frac{7}{2}$ & $\frac{9}{2}$ & $-8.88$ & $0.45$ & $-1.00$ & $0.03$ & $-0.00$ & $0.02$ & $-0.20$ & $0.98$ \Tstrut\\
$F_2$ & $F_4$ & $\frac{9}{2}$ & $\frac{11}{2}$ & $-7.30$ & $0.63$ & $-1.00$ & $0.04$ & $-0.00$ & $0.03$ & $-0.24$ & $0.97$ \Tstrut\\
$F_2$ & $F_4$ & $\frac{11}{2}$ & $\frac{13}{2}$ & $-5.91$ & $0.83$ & $-1.00$ & $0.04$ & $-0.00$ & $0.04$ & $-0.27$ & $0.96$ \Tstrut\\
$F_2$ & $F_4$ & $\frac{13}{2}$ & $\frac{15}{2}$ & $-4.71$ & $1.05$ & $-1.00$ & $0.05$ & $-0.01$ & $0.05$ & $-0.30$ & $0.95$ \Tstrut\\
$F_2$ & $F_4$ & $\frac{15}{2}$ & $\frac{17}{2}$ & $-3.71$ & $1.28$ & $-1.00$ & $0.06$ & $-0.01$ & $0.07$ & $-0.34$ & $0.94$ \Tstrut\\
$F_2$ & $F_4$ & $\frac{17}{2}$ & $\frac{19}{2}$ & $-2.91$ & $1.52$ & $-1.00$ & $0.06$ & $-0.01$ & $0.08$ & $-0.36$ & $0.93$ \Tstrut\\
$F_2$ & $F_4$ & $\frac{19}{2}$ & $\frac{21}{2}$ & $-2.32$ & $1.77$ & $-1.00$ & $0.07$ & $-0.01$ & $0.09$ & $-0.39$ & $0.92$ \Tstrut\\
$F_2$ & $F_4$ & $\frac{21}{2}$ & $\frac{23}{2}$ & $-1.95$ & $2.02$ & $-1.00$ & $0.08$ & $-0.01$ & $0.11$ & $-0.42$ & $0.90$ \Tstrut\\
$F_2$ & $F_4$ & $\frac{23}{2}$ & $\frac{25}{2}$ & $-1.80$ & $2.27$ & $-1.00$ & $0.08$ & $-0.02$ & $0.12$ & $-0.44$ & $0.89$ \Tstrut\\
$F_2$ & $F_4$ & $\frac{25}{2}$ & $\frac{27}{2}$ & $-1.88$ & $2.52$ & $-1.00$ & $0.09$ & $-0.02$ & $0.13$ & $-0.46$ & $0.88$ \Tstrut\\
$F_2$ & $F_4$ & $\frac{27}{2}$ & $\frac{29}{2}$ & $-2.19$ & $2.76$ & $-1.00$ & $0.10$ & $-0.03$ & $0.15$ & $-0.48$ & $0.86$ \Tstrut\\
$F_2$ & $F_4$ & $\frac{29}{2}$ & $\frac{31}{2}$ & $-2.74$ & $3.00$ & $-0.99$ & $0.11$ & $-0.03$ & $0.16$ & $-0.50$ & $0.85$ \Tstrut\\
$F_2$ & $F_4$ & $\frac{31}{2}$ & $\frac{33}{2}$ & $-3.54$ & $3.23$ & $-0.99$ & $0.11$ & $-0.03$ & $0.18$ & $-0.51$ & $0.84$ \Tstrut\\
$F_2$ & $F_4$ & $\frac{33}{2}$ & $\frac{35}{2}$ & $-4.60$ & $3.45$ & $-0.99$ & $0.12$ & $-0.04$ & $0.19$ & $-0.53$ & $0.83$ \Tstrut\\
$F_2$ & $F_4$ & $\frac{35}{2}$ & $\frac{37}{2}$ & $-5.90$ & $3.66$ & $-0.99$ & $0.13$ & $-0.04$ & $0.20$ & $-0.54$ & $0.81$ \Tstrut\\
$F_2$ & $F_4$ & $\frac{37}{2}$ & $\frac{39}{2}$ & $-7.47$ & $3.87$ & $-0.99$ & $0.13$ & $-0.05$ & $0.21$ & $-0.55$ & $0.80$ \Tstrut\\
$F_2$ & $F_4$ & $\frac{39}{2}$ & $\frac{41}{2}$ & $-9.30$ & $4.07$ & $-0.99$ & $0.14$ & $-0.05$ & $0.23$ & $-0.57$ & $0.79$ \Tstrut\Bstrut\\
\hline
$F_1$ & $F_1$ & $\frac{7}{2}$ & $\frac{5}{2}$ & $-3.90$ & $0.07$ & $-0.03$ & $-1.00$ & $1.00$ & $0.08$ & $0.00$ & $0.00$ \Tstrut\\
$F_1$ & $F_1$ & $\frac{9}{2}$ & $\frac{7}{2}$ & $-7.85$ & $0.16$ & $-0.04$ & $-1.00$ & $0.99$ & $0.12$ & $0.01$ & $0.00$ \Tstrut\Bstrut\\
\hline
$F_1$ & $F_1$ & $\frac{5}{2}$ & $\frac{5}{2}$ & $4.79$ & $0.07$ & $-0.02$ & $-1.00$ & $1.00$ & $0.08$ & $0.00$ & $0.00$ \Tstrut\\
$F_1$ & $F_1$ & $\frac{7}{2}$ & $\frac{7}{2}$ & $3.32$ & $0.16$ & $-0.03$ & $-1.00$ & $0.99$ & $0.12$ & $0.01$ & $0.00$ \Tstrut\\
$F_1$ & $F_1$ & $\frac{9}{2}$ & $\frac{9}{2}$ & $1.43$ & $0.27$ & $-0.04$ & $-1.00$ & $0.99$ & $0.16$ & $0.02$ & $0.00$ \Tstrut\\
$F_1$ & $F_1$ & $\frac{11}{2}$ & $\frac{11}{2}$ & $-0.87$ & $0.41$ & $-0.04$ & $-1.00$ & $0.98$ & $0.19$ & $0.02$ & $0.00$ \Tstrut\\
$F_1$ & $F_1$ & $\frac{13}{2}$ & $\frac{13}{2}$ & $-3.59$ & $0.57$ & $-0.05$ & $-1.00$ & $0.97$ & $0.23$ & $0.03$ & $0.00$ \Tstrut\\
$F_1$ & $F_1$ & $\frac{15}{2}$ & $\frac{15}{2}$ & $-6.71$ & $0.74$ & $-0.06$ & $-1.00$ & $0.97$ & $0.26$ & $0.04$ & $0.00$ \Tstrut\Bstrut\\
\hline
$F_1$ & $F_1$ & $\frac{17}{2}$ & $\frac{19}{2}$ & $9.42$ & $1.12$ & $-0.06$ & $-1.00$ & $0.95$ & $0.31$ & $0.06$ & $0.01$ \Tstrut\\
$F_1$ & $F_1$ & $\frac{19}{2}$ & $\frac{21}{2}$ & $7.57$ & $1.33$ & $-0.07$ & $-1.00$ & $0.94$ & $0.34$ & $0.08$ & $0.01$ \Tstrut\\
$F_1$ & $F_1$ & $\frac{21}{2}$ & $\frac{23}{2}$ & $5.34$ & $1.55$ & $-0.08$ & $-1.00$ & $0.93$ & $0.37$ & $0.09$ & $0.01$ \Tstrut\\
$F_1$ & $F_1$ & $\frac{23}{2}$ & $\frac{25}{2}$ & $2.72$ & $1.76$ & $-0.08$ & $-1.00$ & $0.92$ & $0.39$ & $0.10$ & $0.02$ \Tstrut\\
$F_1$ & $F_1$ & $\frac{25}{2}$ & $\frac{27}{2}$ & $-0.28$ & $1.99$ & $-0.09$ & $-1.00$ & $0.90$ & $0.41$ & $0.11$ & $0.02$ \Tstrut\\
$F_1$ & $F_1$ & $\frac{27}{2}$ & $\frac{29}{2}$ & $-3.67$ & $2.21$ & $-0.10$ & $-1.00$ & $0.89$ & $0.43$ & $0.13$ & $0.02$ \Tstrut\\
$F_1$ & $F_1$ & $\frac{29}{2}$ & $\frac{31}{2}$ & $-7.42$ & $2.43$ & $-0.11$ & $-0.99$ & $0.88$ & $0.45$ & $0.14$ & $0.03$ \Tstrut\Bstrut\\
\hline
$F_1$ & $F_2$ & $\frac{25}{2}$ & $\frac{23}{2}$ & $4.95$ & $9.38$ & $-0.09$ & $-1.00$ & $-0.36$ & $0.82$ & $0.44$ & $0.10$ \Tstrut\\
$F_1$ & $F_2$ & $\frac{27}{2}$ & $\frac{25}{2}$ & $-1.75$ & $9.16$ & $-0.10$ & $-1.00$ & $-0.39$ & $0.79$ & $0.46$ & $0.11$ \Tstrut\\
$F_1$ & $F_2$ & $\frac{29}{2}$ & $\frac{27}{2}$ & $-8.77$ & $8.93$ & $-0.11$ & $-0.99$ & $-0.41$ & $0.76$ & $0.48$ & $0.13$ \Tstrut\Bstrut\\
\hline
$F_1$ & $F_2$ & $\frac{35}{2}$ & $\frac{35}{2}$ & $5.91$ & $8.04$ & $-0.13$ & $-0.99$ & $-0.48$ & $0.66$ & $0.55$ & $0.18$ \Tstrut\\
$F_1$ & $F_2$ & $\frac{37}{2}$ & $\frac{37}{2}$ & $-0.21$ & $7.82$ & $-0.13$ & $-0.99$ & $-0.50$ & $0.63$ & $0.56$ & $0.19$ \Tstrut\\
$F_1$ & $F_2$ & $\frac{39}{2}$ & $\frac{39}{2}$ & $-6.63$ & $7.61$ & $-0.14$ & $-0.99$ & $0.51$ & $-0.60$ & $-0.58$ & $-0.21$ \Tstrut\Bstrut\\
\hline
$F_1$ & $F_2$ & $\frac{51}{2}$ & $\frac{53}{2}$ & $5.66$ & $6.33$ & $-0.18$ & $-0.98$ & $-0.58$ & $0.44$ & $0.62$ & $0.29$ \Tstrut\\
$F_1$ & $F_2$ & $\frac{53}{2}$ & $\frac{55}{2}$ & $-0.72$ & $6.18$ & $-0.18$ & $-0.98$ & $-0.59$ & $0.42$ & $0.62$ & $0.30$ \Tstrut\\
$F_1$ & $F_2$ & $\frac{55}{2}$ & $\frac{57}{2}$ & $-7.39$ & $6.04$ & $-0.19$ & $-0.98$ & $0.60$ & $-0.40$ & $-0.63$ & $-0.30$ \Tstrut\Bstrut\\
\hline
$F_1$ & $F_3$ & $\frac{41}{2}$ & $\frac{39}{2}$ & $8.89$ & $7.13$ & $-0.15$ & $-0.99$ & $-0.19$ & $0.57$ & $-0.57$ & $-0.56$ \Tstrut\\
$F_1$ & $F_3$ & $\frac{43}{2}$ & $\frac{41}{2}$ & $1.14$ & $6.94$ & $-0.15$ & $-0.99$ & $0.21$ & $-0.58$ & $0.55$ & $0.57$ \Tstrut\\
$F_1$ & $F_3$ & $\frac{45}{2}$ & $\frac{43}{2}$ & $-6.87$ & $6.76$ & $-0.16$ & $-0.99$ & $-0.22$ & $0.59$ & $-0.52$ & $-0.58$ \Tstrut\Bstrut\\
\hline
$F_1$ & $F_3$ & $\frac{57}{2}$ & $\frac{57}{2}$ & $2.66$ & $5.71$ & $-0.20$ & $-0.98$ & $0.29$ & $-0.62$ & $0.37$ & $0.63$ \Tstrut\\
$F_1$ & $F_3$ & $\frac{59}{2}$ & $\frac{59}{2}$ & $-5.06$ & $5.58$ & $-0.20$ & $-0.98$ & $-0.30$ & $0.62$ & $-0.35$ & $-0.63$ \Tstrut\Bstrut\\
\hline
$F_1$ & $F_3$ & $\frac{75}{2}$ & $\frac{77}{2}$ & $8.43$ & $4.74$ & $-0.25$ & $-0.97$ & $0.37$ & $-0.62$ & $0.21$ & $0.66$ \Tstrut\\
$F_1$ & $F_3$ & $\frac{77}{2}$ & $\frac{79}{2}$ & $0.36$ & $4.67$ & $-0.25$ & $-0.97$ & $-0.37$ & $0.62$ & $-0.20$ & $-0.66$ \Tstrut\\
$F_1$ & $F_3$ & $\frac{79}{2}$ & $\frac{81}{2}$ & $-7.99$ & $4.60$ & $-0.26$ & $-0.97$ & $0.38$ & $-0.62$ & $0.18$ & $0.66$ \Tstrut\Bstrut\\
\hline
$F_1$ & $F_4$ & $\frac{61}{2}$ & $\frac{59}{2}$ & $1.74$ & $5.45$ & $-0.21$ & $-0.98$ & $-0.09$ & $0.32$ & $-0.63$ & $0.70$ \Tstrut\\
$F_1$ & $F_4$ & $\frac{63}{2}$ & $\frac{61}{2}$ & $-6.89$ & $5.57$ & $-0.21$ & $-0.98$ & $0.09$ & $-0.33$ & $0.63$ & $-0.69$ \Tstrut\Bstrut\\
\hline
$F_1$ & $F_4$ & $\frac{79}{2}$ & $\frac{79}{2}$ & $3.89$ & $6.38$ & $-0.26$ & $-0.97$ & $0.13$ & $-0.39$ & $0.66$ & $-0.63$ \Tstrut\\
$F_1$ & $F_4$ & $\frac{81}{2}$ & $\frac{81}{2}$ & $-4.98$ & $6.45$ & $-0.26$ & $-0.96$ & $0.13$ & $-0.39$ & $0.66$ & $-0.63$ \Tstrut\Bstrut\\
\hline
$F_1$ & $F_4$ & $\frac{101}{2}$ & $\frac{103}{2}$ & $6.82$ & $7.01$ & $-0.31$ & $-0.95$ & $0.16$ & $-0.44$ & $0.67$ & $-0.58$ \Tstrut\\
$F_1$ & $F_4$ & $\frac{103}{2}$ & $\frac{105}{2}$ & $-2.97$ & $7.05$ & $-0.31$ & $-0.95$ & $0.17$ & $-0.44$ & $0.67$ & $-0.57$ \Tstrut\Bstrut\\
\hline\hline
\end{longtable}
\begin{longtable}{@{\extracolsep{\fill}}ccccrrrrrrrr}
\caption{The near-degeneracies $|X\,^2\Pi_g, v^\pp=22, J^\pp\rangle$ and $|a\,^4\Pi_u, v^\ensuremath{\prime}=1, J^\ensuremath{\prime}\rangle$, including the eigenstate superposition coefficients.}\\
\hline\hline
$X\,^2\Pi_g$ & $a\,^4\Pi_u$ & $J^\pp$ & $J^\ensuremath{\prime}$ & $\Delta E / {\rm cm}^{-1}$ & $\frac{\Gamma}{2\pi} / {\rm Hz}$ & $c_{3/2}^{\pp}$ & $c_{1/2}^\pp$ & $c_{5/2}^\ensuremath{\prime}$ & $c_{3/2}^\ensuremath{\prime}$ & $c_{1/2}^\ensuremath{\prime}$ & $c_{-1/2}^\ensuremath{\prime}$ \Tstrut\Bstrut\\
\hline \endfirsthead
\caption{(continued)}\Tstrut\\
\hline
$X\,^2\Pi_g$ & $a\,^4\Pi_u$ & $J^\pp$ & $J^\ensuremath{\prime}$ & $\Delta E / {\rm cm}^{-1}$ & $\frac{\Gamma}{2\pi} / {\rm Hz}$ & $c_{3/2}^{\pp}$ & $c_{1/2}^\pp$ & $c_{5/2}^\ensuremath{\prime}$ & $c_{3/2}^\ensuremath{\prime}$ & $c_{1/2}^\ensuremath{\prime}$ & $c_{-1/2}^\ensuremath{\prime}$ \Tstrut\Bstrut\\
\hline \endhead
$F_1$ & $F_4$ & $\frac{1}{2}$ & $\frac{3}{2}$ & $-7.84$ & $0.08$ & $0.00$ & $1.00$ & $0.00$ & $0.00$ & $-0.08$ & $1.00$ \Tstrut\\
$F_1$ & $F_4$ & $\frac{3}{2}$ & $\frac{5}{2}$ & $-5.77$ & $0.17$ & $-0.01$ & $-1.00$ & $-0.00$ & $0.01$ & $-0.12$ & $0.99$ \Tstrut\\
$F_1$ & $F_4$ & $\frac{5}{2}$ & $\frac{7}{2}$ & $-3.87$ & $0.30$ & $-0.02$ & $-1.00$ & $-0.00$ & $0.01$ & $-0.16$ & $0.99$ \Tstrut\\
$F_1$ & $F_4$ & $\frac{7}{2}$ & $\frac{9}{2}$ & $-2.15$ & $0.45$ & $-0.03$ & $-1.00$ & $-0.00$ & $0.02$ & $-0.20$ & $0.98$ \Tstrut\\
$F_1$ & $F_4$ & $\frac{9}{2}$ & $\frac{11}{2}$ & $-0.61$ & $0.63$ & $-0.04$ & $-1.00$ & $-0.00$ & $0.03$ & $-0.23$ & $0.97$ \Tstrut\\
$F_1$ & $F_4$ & $\frac{11}{2}$ & $\frac{13}{2}$ & $0.75$ & $0.84$ & $-0.04$ & $-1.00$ & $-0.00$ & $0.04$ & $-0.27$ & $0.96$ \Tstrut\\
$F_1$ & $F_4$ & $\frac{13}{2}$ & $\frac{15}{2}$ & $1.92$ & $1.06$ & $-0.05$ & $-1.00$ & $-0.00$ & $0.05$ & $-0.30$ & $0.95$ \Tstrut\\
$F_1$ & $F_4$ & $\frac{15}{2}$ & $\frac{17}{2}$ & $2.90$ & $1.30$ & $-0.06$ & $-1.00$ & $-0.01$ & $0.06$ & $-0.33$ & $0.94$ \Tstrut\\
$F_1$ & $F_4$ & $\frac{17}{2}$ & $\frac{19}{2}$ & $3.67$ & $1.55$ & $-0.06$ & $-1.00$ & $-0.01$ & $0.08$ & $-0.36$ & $0.93$ \Tstrut\\
$F_1$ & $F_4$ & $\frac{19}{2}$ & $\frac{21}{2}$ & $4.24$ & $1.80$ & $-0.07$ & $-1.00$ & $-0.01$ & $0.09$ & $-0.39$ & $0.92$ \Tstrut\\
$F_1$ & $F_4$ & $\frac{21}{2}$ & $\frac{23}{2}$ & $4.59$ & $2.06$ & $-0.08$ & $-1.00$ & $-0.01$ & $0.10$ & $-0.41$ & $0.91$ \Tstrut\\
$F_1$ & $F_4$ & $\frac{23}{2}$ & $\frac{25}{2}$ & $4.73$ & $2.31$ & $-0.09$ & $-1.00$ & $-0.02$ & $0.12$ & $-0.44$ & $0.89$ \Tstrut\\
$F_1$ & $F_4$ & $\frac{25}{2}$ & $\frac{27}{2}$ & $4.63$ & $2.57$ & $-0.09$ & $-1.00$ & $-0.02$ & $0.13$ & $-0.46$ & $0.88$ \Tstrut\\
$F_1$ & $F_4$ & $\frac{27}{2}$ & $\frac{29}{2}$ & $4.31$ & $2.82$ & $-0.10$ & $-0.99$ & $-0.02$ & $0.14$ & $-0.48$ & $0.87$ \Tstrut\\
$F_1$ & $F_4$ & $\frac{29}{2}$ & $\frac{31}{2}$ & $3.74$ & $3.06$ & $-0.11$ & $-0.99$ & $-0.03$ & $0.16$ & $-0.49$ & $0.85$ \Tstrut\\
$F_1$ & $F_4$ & $\frac{31}{2}$ & $\frac{33}{2}$ & $2.93$ & $3.30$ & $-0.11$ & $-0.99$ & $-0.03$ & $0.17$ & $-0.51$ & $0.84$ \Tstrut\\
$F_1$ & $F_4$ & $\frac{33}{2}$ & $\frac{35}{2}$ & $1.87$ & $3.53$ & $-0.12$ & $-0.99$ & $-0.04$ & $0.18$ & $-0.53$ & $0.83$ \Tstrut\\
$F_1$ & $F_4$ & $\frac{35}{2}$ & $\frac{37}{2}$ & $0.56$ & $3.75$ & $-0.13$ & $-0.99$ & $-0.04$ & $0.20$ & $-0.54$ & $0.82$ \Tstrut\\
$F_1$ & $F_4$ & $\frac{37}{2}$ & $\frac{39}{2}$ & $-1.01$ & $3.97$ & $-0.13$ & $-0.99$ & $0.04$ & $-0.21$ & $0.55$ & $-0.81$ \Tstrut\\
$F_1$ & $F_4$ & $\frac{39}{2}$ & $\frac{41}{2}$ & $-2.85$ & $4.17$ & $-0.14$ & $-0.99$ & $0.05$ & $-0.22$ & $0.56$ & $-0.79$ \Tstrut\\
$F_1$ & $F_4$ & $\frac{41}{2}$ & $\frac{43}{2}$ & $-4.95$ & $4.36$ & $-0.15$ & $-0.99$ & $-0.05$ & $0.23$ & $-0.57$ & $0.78$ \Tstrut\\
$F_1$ & $F_4$ & $\frac{43}{2}$ & $\frac{45}{2}$ & $-7.32$ & $4.55$ & $-0.15$ & $-0.99$ & $-0.06$ & $0.25$ & $-0.58$ & $0.77$ \Tstrut\\
$F_1$ & $F_4$ & $\frac{45}{2}$ & $\frac{47}{2}$ & $-9.97$ & $4.73$ & $-0.16$ & $-0.99$ & $-0.06$ & $0.26$ & $-0.59$ & $0.76$ \Tstrut\Bstrut\\
\hline\hline
\end{longtable}
\section{Other degeneracies}
\begin{figure}[h]
\includegraphics[width=0.75\columnwidth]{higher_degeneracy.eps}%
\caption{Overlap of the $X\,^2\Pi_g$, $a\,^4\Pi_u$, and $A\,^2\Pi_u$ states near $|X, v=27\textnormal{--}30\rangle$. The levels are calculated from refs.~\cite{SsongJCP1999,SsongJCP2000,SsongJCP2000a}. Note that the $A$ state's spin-orbit constant is small enough that the doublet splitting is not visible at this scale.\label{fig:moreDegen}}
\end{figure}
\input{supplemental.bbl}
\end{document}
|
\section{Introduction}
\label{intro}
\vspace{-0.1in}
Audio plays a critical role in understanding the environment around us. This makes audio content analysis research important for tasks related to multimedia \cite{cheng2012sri,jiang2015fast}, and human computer interaction~\cite{chu2006scene,janvier:hal-00768767} to mention a pair. However, unlike the field of computer vision which has a variety of standard publicly available datasets such as Imagenet, audio event/scene analysis lacks such large dataset. This makes it difficult to compare different approaches and establishing the state of art. The second iteration of DCASE~\cite{Mesaros2016_EUSIPCO}, occurring in 2016, offers an opportunity to compare approaches on a standard public dataset. This edition it includes four different tasks: acoustic scene classification, sound event detection-- real and synthetic audio, and audio tagging.
The state-of-the-art of the previous DCASE challenge, for both acoustic scenes~\cite{roma2013recurrence,rakotomamonjy2015histogram,schroder2013use} and sound event detection~\cite{roma2013recurrence,schroder2013use,gemmeke2013exemplar}, attributed their success mainly to features and audio representations rather than classifiers. Hence, an important aspect in our work is to emphasize on classifier exploration along with features. In this paper we present our work performed on Task 1 and Task 3. We proposed a variety of methods for both tasks and we obtained significant improvement over the baseline methods.
\vspace{-0.2in}
\section{Tasks and Data}
\vspace{-0.1in}
\label{sec:data}
The goal of Task 1, Acoustic Scene Classification, is to classify a test recording into one of predefined classes that characterizes the environment in which it was recorded — for example \textit{park}, \textit{home}, \textit{office}. TUT Acoustic Scenes 2016 dataset is used for this task. It consists of recordings from various acoustic scenes. For each recording location, a 3-5 minute long audio recording was captured. The original recordings were then split into 30-second segments for the challenge. There are 15 acoustic scenes for the task.
Task 3, Sound Event Detection in Real Life Recordings, evaluates performance of sound event detection in multi-source conditions similar to our everyday life. There is no control over the number of overlapping sound events at each time, not in the training nor in the audio data. TUT Sound Events 2016 dataset is used for Task 3, which consists of recordings from two acoustic scenes: \textit{Home} and \textit{Residential Area}. There are 18 selected sound event classes, 11 for Home and 7 for Residential Area.
\vspace{-0.15in}
\section{Task 1: Acoustic Scene Classification}
\vspace{-0.1in}
\label{sec:task1}
From machine learning perspective, we treated Task 1 as a multi-class classification problem. The first step is to use a suitable method for characterizing acoustic scenes in the audio segments. An effective approach for characterizing audio events is bag-of-audio-words based feature representation \cite{bow}, which is usually built over low-level features such as MFCCs. Acoustic scenes, however, are more complex mixtures of different audio events and a more robust representation is required. To obtain a more robust representation we use Gaussian Mixture Models (GMMs) for feature representations of audio segments. Broadly, we employed two high-level feature representations to represent audio scenes. On the classification front we used Support Vector Machines (SVMs) as our primary classifier and in combination with other classifiers.
\vspace{-0.1in}
\subsection{Feature Representations}
\vspace{-0.1in}
Let $D$-dimensional MFCCs vectors for a recording be represented as $\vec{x}_t$, where $t=1\,\,$to$\,\,T$, $T$ is the total number of MFCCs vectors for the recording. The major idea behind both high-level feature representation is to capture the distribution of MFCCs vectors of a recording. We will refer to these features as $\vec{\alpha}$ and $\vec{\beta}$ features and the sub-types will be represented using appropriate subscripts and superscripts.
The first step in obtaining high-level fixed dimensional feature representation for audio segments is to train a GMM on MFCC vectors of the training data. Let us represent this GMM by $\mathcal{G} = \{w_k,N(\vec{\mu}_k, \Sigma_k), k = 1 \,\,to \,\,M\}$, where $w_k$, $\vec{\mu}_k$ and $\Sigma_k$ are the mixture weight, mean and covariance parameters of the $k^{th}$ Gaussian in $\mathcal{G}$. We will assume diagonal covariance matrices for all Gaussians and $\vec{\sigma}_k$ will represent the diagonal vector of $\Sigma_k$. Given the MFCCs vectors $\vec{x}_t$ of a recording, we computed the probabilistic assignment of $\vec{x}_t$ to the $k^{th}$ Gaussian. These soft assignments are added over all $t$ to obtain the total mass of MFCCs vectors belonging to the $k^{th}$ Gaussian (Eq \ref{eq:addms}). Normalization by $T$ is used to remove the effect of the duration of recordings.
\begin{equation}
\label{eq:addms}
\resizebox{0.90\columnwidth}{!}{$ Pr(k | \vec{x}_{t}) = \frac{w_{k}N(\vec{x}_{t} ; \vec{\mu}_k, \Sigma_k)}{\sum\limits_{j=1}^M w_jN(\vec{x}_{t} ; \vec{\mu}_k, \Sigma_k)},P(k) = \frac{1}{T}\sum\limits_{i=1}^T Pr(k | \vec{x}_{t})$}
\vspace{-0.05in}
\end{equation}
The soft count histogram features referred to as $\vec{\alpha}$ is, $\vec{\alpha}^M=[P(1),..P(k)..P(M)]^T$. $\vec{\alpha}^M$ is an $M$-dimensional feature representation for a given recording. It captures how the MFCC vectors of a recording are distributed across the Guassians in $\mathcal{G}$. $\vec{\alpha}^M$ is normalized to sum to $1$ before using it for classifier training.
The next feature ($\vec{\beta}$), also based on the GMM $\mathcal{G}$, tries to capture the actual distribution of the MFCC vectors of a recording. This is done by adapting the parameters of $\mathcal{G}$ to the MFCC vectors of the recording. We employ maximum {\em a posteriori} (MAP) estimation to for the adaptation \cite{gauvain1994} \cite{bimbot}. Parameter adaptation for $k^{th}$ Gaussian follows the following steps. First we compute,
\begin{equation}
\resizebox{1\columnwidth}{!}{$n_{k}=\sum\limits_{t=1}^T Pr(k | \vec{x}_{t}),\,\,\,E_{k}(\vec{x})=\frac{1}{n_{k}}\sum\limits_{t=1}^T Pr(k | \vec{x_{t}})\vec{x}_{t},\,\,E_{k}(\vec{x}^2)=\frac{1}{n_{k}}\sum\limits_{t=1}^T Pr(k | \vec{x_{t}})\vec{x}_{t}^2$}
\end{equation}
Finally, the updated mean and variances are obtained as
\begin{align}
\hat{\vec{\mu}}_k=&\frac{n_k}{n_k+r}E_{k}(\vec{x})+\frac{r}{n_k+r}\vec{\mu}_k \\
\hat{\vec{\sigma}}_k=&\frac{n_k}{n_k+r}E_{k}(\vec{x}^2)+\frac{r}{n_k+r}(\vec{\sigma}_k^2+\vec{\mu}_k^2) - \hat{\vec{\mu}}_k^2
\end{align}
The relevance factor $r$ controls the effect of the original parameters on the new estimates. We obtain $2$ different feature representation using the adapted means ($\hat{\vec{\mu}}_k$) and variances ($\hat{\vec{\sigma}}_k$). The first one denoted by $\vec{\beta}^M$ is an $M \times D$ dimensional feature obtained by concatenating the adapted means $\hat{\vec{\mu}}_k$ for all $k$, that is $\vec{\beta}^M=[\hat{\vec{\mu}}_1^T,...\hat{\vec{\mu}}_K^T]^T$. In the second $\vec{\beta}$ features adapted $\hat{\vec{\sigma}}_k$ are concatenated along with $\hat{\vec{\mu}}_k$ to obtain a $2 \times M \times D$ dimensional features. This form of $\vec{\beta}$ features are denoted by $\vec{\beta}^M_{\sigma}$.
\vspace{-0.15in}
\subsection{Classification}
\vspace{-0.1in}
Once the feature representation for audio segments have been obtained, Task 1 essentially becomes a multi-class classification problem. Our primary classifiers are SVMs where we explore a variety of kernels. For the $\vec{\beta}$ features, we use Linear Kernel (LK) and RBF Kernel (RK). For soft-count histogram $\vec{\alpha}$ features we explore a panoply of kernels. Along with LK and RK we explored the following kernels.
\begin{itemize}
\item Exponential $\chi^2$ Distance (ECK): the kernel is computed as $K(\vec{x},\vec{y})=\exp^{-\gamma D(\vec{x},\vec{y})}$, where $D(\vec{x},\vec{y}) = \sum_i (x_i-y_i)^2/(x_i+y_i)$ is $\chi^2$ distance.
\item $\chi^2$ Kernel (CK): In this case $K(\vec{x},\vec{y})=\sum_i \frac{2 x_i y_i}{x_i+y_i}$
\item Intersection Kernel (IK): $K(\vec{x},\vec{y})=\sum_i \min(x_i,y_i)$
\item Exponential Hellinger Distance Kernel (EHK): $K(\vec{x},\vec{y}) = \exp^{-\gamma D(\vec{x},\vec{y})}$ where $D(\vec{x},\vec{y}) = \sum_i (\sqrt{x_i} - \sqrt{y_i})^2$
\item Hellinger Kernel (HK): $(\vec{x},\vec{y}) = \sum_i \sqrt{x_i y_i}$
\end{itemize}
The details of these kernels can be found in \cite{zhang2007local,vedaldi2012efficient,li2013sign}. For kernels where $\gamma$ term appears, the optimal value of $\gamma$ value can be obtained by cross validation over training data. However, setting $\gamma$ equal to the inverse of average distance $D(\vec{x},\vec{y})$ between training data points works well in general as well. We use \cite{LIBSVM}\cite{fanliblinear} for SVM implementation.
Finally, we have a classifier fusion step where we combined the output of the different classifiers. We combined multiple classifiers by taking prediction vote from each classifier and the final predicted class is the one which gets the maximum vote. We call it the \emph{Fused Classifier} and we observed that the fused classifier can give significant improvement for several acoustic scenes.
\vspace{-0.15in}
\subsection{Results}
\vspace{-0.1in}
Our experimental setup with the folds structure, is same as the one provided by DCASE. We extracted 20 dimensional MFCC features using $30ms$ window and $50\%$ overlap. MFCCs are augmented with their delta and acceleration features. For our final feature representation we experimented with $4$ different values of GMM component size $M$, $64,128,256$ and $512$. The relevance factor $r$ for $\vec{\beta}$ is set to $20$. Due to space constraints we cannot present fold-and-scene specific results for all cases and hence overall accuracy for all $4$ folds is shown. Table \ref{tab:task11} shows overall accuracy results for different cases. The accuracy for the MFCC-GMM \emph{baseline} method provided in the challenge is $72.6\%$.
We can observe from Table \ref{tab:task11} that $\vec{\alpha}$ features in general do not perform better than the baseline method for any SVM kernel. However, $\vec{\beta}$ features clearly outperformed baseline method. In the best case, with $M=128$ and $\vec{\beta}_\sigma^M$ our method outperformed the baseline by an absolute $5\%$.
Table \ref{tab:task12} shows results for the fused classifiers. For the fusion step we did not consider classifiers built over $\vec{\alpha}$ since these classifiers are inferior compared to those using $\vec{\beta}$ features. We can observe that our proposed method beats the baseline method by an absolute $\mathbf{6.3\%}$. Moreover, for scenes such as \emph{Park, Train, Library} where the baseline method gives very poor results, we improved the accuracy by an absolute $\mathbf{16-30\%}$. We also obtained superior overall accuracy on all folds which suggests that our proposed method is fairly robust. This is further supported by the fact that on DCASE evaluation set, We achieved an overall accuracy of $\mathbf{85.9\%}$.
\begin{table}[t]
\centering
\caption{Task 1 Accuracy for different cases (Single Classifier)}
\label{tab:task11}
\resizebox*{1.0\columnwidth}{!}{
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
&\multicolumn{7}{c|}{$\vec{\alpha}^M$} & \multicolumn{2}{c|}{$\vec{\beta}^M$} & \multicolumn{2}{c|}{$\vec{\beta}^M_{\sigma}$}\\
\cline{2-12}
$M$&LK&RK&ECK&CK&IK&EHK&HK&LK&RK&LK&RK\\
\hline
64&62.8&60.6&66.2&66.3&66.0&64.7&65.3&76.8&76.6&75.5&$\mathbf{76.7}$\\
\hline
128&63.6&62.3&67.5&67.1&66.4&67.4&66.5&76.5&75.3&77.5&$\mathbf{77.5}$\\
\hline
256&63.9&63.9&67.3&67.8&66.5&68.7&67.7&76.5&71.9&$\mathbf{76.6}$&75.9\\
\hline
512&65.0&62.9&67.8&67.8&67.1&68.9&69.3&$\mathbf{76.4}$&72.2&76.2&75.9\\
\hline
\end{tabular}
}
\vspace{-0.20in}
\end{table}
\begin{table}[t]
\centering
\caption{Overall Task 1 Accuracy (Fused Classifier)}
\label{tab:task12}
\resizebox*{1.0\columnwidth}{!}{
\begin{tabular}{|c|c|c|c|c|c?c|c|c|c|c|}
\hline
&\multicolumn{5}{c?}{\textbf{Baseline}} & \multicolumn{5}{c|}{\textbf{Proposed}}\\
\cline{2-11}
Scene&Fold 1&Fold 2&Fold 3&Fold 4& \textbf{Avg.} & Fold 1&Fold 2&Fold 3&Fold 4 & \textbf{Avg.}\\
\hline
Beach&84.2&66.7&78.9&47.4&69.3&100&71.4&89.5&52.6&78.4\\
\hline
Bus&68.4&65.0&100&85.0&79.6&68.4&50.0&100&95.0&78.4\\
\hline
Cafe/Restaurant&66.7&94.7&71.4&100&83.2&88.9&63.2&76.2&95.0&80.8\\
\hline
Car&70.0&89.5&89.5&100&87.3&80.0&100&100&100&95.0\\
\hline
City Center&83.3&73.7&89.5&95.5&85.5&88.9&84.2&100&95.5&92.1\\
\hline
Forest Path&57.1&100&66.7&100&81.0&81.0&100&100&100&95.2\\
\hline
Grocery Store&52.6&81.0&89.5&36.8&65&89.5&81.0&94.7&84.2&87.3\\
\hline
Home&100&55.6&95.0&77.8&82.1&100&61.1&80.0&44.4&71.4\\
\hline
Library&47.6&38.9&15.0&100&50.4&47.6&33.3&85.0&100&66.5\\
\hline
Metro Station&84.2&94.4&100&100&94.7&94.7&94.4&100&100&97.3\\
\hline
Office&100&100&94.4&100&98.6&78.9&100&72.2&83.3&83.6\\
\hline
Park&10.0&5.6&0&40.0&13.9&65.0&33.3&50.0&30.0&44.6\\
\hline
Residential&78.9&47.6&100&84.2&77.7&84.2&42.9&94.7&57.9&69.9\\
\hline
Train&16.7&31.6&30.4&61.1&34.9&50.0&63.2&34.8&88.9&59.2\\
\hline
Tram&88.9&88.9&63.6&100&85.3&83.3&88.9&63.6&100&84.0\\
\hline
\textbf{Overall}&\textbf{67.2}&\textbf{68.9}&\textbf{72.3}&\textbf{81.9}&\textbf{72.6}&\textbf{80.0}&\textbf{71.1}&\textbf{82.7}&\textbf{81.8}&\textbf{78.9}\\
\hline
\end{tabular}
}
\vspace{-0.10in}
\end{table}
\vspace{-0.1in}
\section{Task 3: Sound Event Detection in Real Life Recordings}
\label{sec:task3}
\vspace{-0.1in}
\begin{table}[t]
\small
\centering
\resizebox*{1.0\columnwidth}{!}{
\begin{tabular}{ | l | c | c | }
\hline
\textbf{Feature Type} & \textbf{Accuracy}\% & \textbf{Classifier} \\ \hline
MFCCs & 67.7 & Logistic Regression \\\hline
GBFB & 52.4 & Gradient Boosting \\\hline
SGBFB & 61.5 & Gradient Boosting \\\hline
Scatnet & 62.1 & Random Forest \\\hline
Stacked & 66.68 & Random Forest \\\hline
Stacked + PCA& 66.06 & Random Forest \\\hline
\end{tabular}
}
\caption{Sound-event classification accuracy for different feature types using the 18 sounds and the 75 training - 25 testing ratio. Stacked included normalized MFCCs +SGBFB +Scatnet.}
\label{featureTable}
\vspace{-1em}
\end{table}
Detection of sound events in scenes and long recordings have been treated as a multi-class classification problem before in\cite{elizaldeaudio,elizalde2013lost,Elizaldeacm} where a classifier is trained with the sound segments. For testing, the classifier outputs segment/frame-level predictions for all the classes. In order to follow a similar approach, first we wanted to analyze features' performance for sound events regardless of the scene. This way, we could have an intuition of performance on the harder scenario of Task 3 where not every segment of the scene corresponds to a labeled sound event.
\vspace{-0.1in}
\subsection{Features and Classifiers Optimization}
\label{sec:features}
\vspace{-0.1in}
For the features we tried the conventional MFCCs with standard parameters such as 12 coefficients plus energy, delta and double delta for a total of 39 dimensions. Moreover, we explored three features addressing the time-frequency acoustic characteristics. The Gabor Filter Bank (GBFB) in ~\cite{schadler2012spectro} have 2D-filters arranged by spectral and temporal modulation frequencies in a filter bank. The Separable Gabor filter bank (SGBFB) features extract spectro-temporal patterns with two separate 1D GBFBs, a spectral one and a temporal one. This approach reduces the complexity of the spectro-temporal feature extraction and further improves robustness as demonstrated in ~\cite{schadler2015separable}. Both features have the default parameters from the toolbox\footnote{http://www.uni-oldenburg.de/mediphysik-akustik/mediphysik/downloads/gabor-filter-bank-features/} for a total dimension of 1,020 each. The Scatnet~\cite{sifre2013rotation} features are generated by a scattering architecture which computes invariants to translations, rotations, scaling and deformations, while keeping enough discriminative information. It can be interpreted as a deep convolution network, where convolutions are performed along spatial, rotation and scaling variables. As opposed to standard convolution networks, the filters are not learned but are scaled and rotated wavelets. The features were extracted with a toolbox\footnote{http://www.di.ens.fr/data/software/scatnet/} using 0.25 second segments. The dimensionality of the three Scatnet components are 2, 84, 435 for a total of 521. Additionally, we included the normalized (mean and variance) Stacked (MFCCs+ SGBFB+ Scatnet) with PCA and also the normalized (mean and variance) Stacked without PCA. For the PCA we used Scikit's~\cite{scikit-learn} and used the full dimensionality of 1,580 as the number of input components and the resultant automatic reduction was 909 dimensions. For all the feature types and for the sake of avoiding the length variability of the temporal dimension, we averaged the vectors across time to end up with one single vector per sound event file.
Then, for the classifiers we considered Tpot~\cite{Olson2016EvoBIO}, built on top of Scikit~\cite{scikit-learn}, which is a Python tool that automatically creates and optimizes machine learning pipelines using genetic programming. This toolbox (version 4) considers 12 classifiers such as Decision Tree, Random Forest, Xtreme Gradient Boosting, SVMs, K-Neighbors and Logistic Regression. The main Tpot parameter is ``number of generations", which corresponds to the number of iterations carried to tune the classifier, we set it to 15. An example of the best classifier for each feature type can be seen in Table ~\ref{featureTable}. Interestingly, decision tree-based algorithms and logistic regression outperformed others like SVMs.
For our experiments, we extracted the 18 sound events from the two scenes using the annotations, and then we extracted different feature types from these isolated sounds. For each feature type experiment, the sound events' feature files were fed to Tpot in a randomly selected ratio of 75\% training and 25\% testing, each set with different files. We kept the same partitions across our experiments for consistency. The performance was measured in terms of accuracy and is displayed in Table ~\ref{featureTable}.
\begin{table}[t]
\small
\centering
\begin{tabular}{| l | c | c | }
\hline
\textbf{Feature Type} & \textbf{Accuracy}\% & \textbf{Classifier} \\ \hline
Home & 56.4 & Random Forest \\\hline
Home + G & 55.2 & Random Forest \\\hline
Home + G + P & 55.7 & Random Forest \\\hline
Residential & 53.3 & Gradient Boosting \\\hline
Residential + G & 57.8 & Decision Tree \\\hline
Residential + G + P & 56.7 & Random Forest \\\hline
\end{tabular}
\caption{Sound-event classification accuracy using the DCASE set up with four-folds partitions. The inclusion of the [G]eneric class improved performance for both scenes, whereas the inclusion of the [P]erturbed audio improved only the Home performance.}
\label{genericTable}
\vspace{-0.15in}
\end{table}
The features with the best performance were MFCCs with 67.7\% and thus we keep them for our DCASE evaluation set up. The other features have shown better results than MFCCs on audio classification, but it wasn't the case for this particular dataset. Results for Scatnet was 62.1\%, for GBF was 52.4\%, and for SGBF was 61.5\%. Moreover, the two normalized stacked features performed almost as good as MFCCs with 66.68\% for the stacked without PCA and 66.08\% for the stacked with PCA. In principle the stacked version contains more information about the acoustics and thus they were expected to perform better. Nevertheless, they didn't outperform MFCCs which is designed for speech and focus on lower frequencies rather than on a wider frequency range. We cannot draw a fundamental conclusion on the performance of these features for sound event classification since the amount of data and classes are determinant.
\begin{figure}[t]
\centering
\includegraphics[width=0.45\textwidth]{mfcc_features_18_normalized_confusion_matrix.png}
\caption{[H]ome and [R]esidential without the generic class. \emph{Object impact} and \emph{bird singing} capture most of ambiguities.}
\label{fig:mfcc_features_18_normalized_confusion_matrix}
\vspace{-0.25in}
\end{figure}
\vspace{-0.1in}
\subsection{Inclusion of Generic Sound Event Class}
\vspace{-0.1in}
\label{sec:generic}
In the annotated scenes, not every segment of audio corresponds to a labeled sound. Hence, it cannot be assured that any of our sound event classes have to be present on every test segment. To handle out-of-vocabulary segments, we proposed a generic sound event class.
For the first experiment we wanted to analyze the impact of the generic class together with the 18 sounds in the multi-class classification set up described in Section~\ref{sec:features} using MFCCs. To create such class, we used the sound events annotations and trimmed out the audio between the labeled segments, which are unlabeled. Then, we randomly selected from both scenes, 60 audio files which is about the average number of sound event samples per class. In order to visualize the performance, we included the normalized confusion matrices (CMs) in Figures ~\ref{fig:mfcc_features_18_normalized_confusion_matrix} and ~\ref{fig:mfcc_features_18_plus_generic_normalized_matrix}. The accuracy performance without the generic class was 67.7\% and with the generic class was 60.94\%. The performance dropped with the inclusion of the new class, but we can also observe how although the generic class shared the background acoustics with the other sound classes, it didn't significantly ambiguate with them.
The second set of experiments used the DCASE setup of separate scenes and four folds, and utilized the sound events with and without the generic class, but this time the generic class will have files particular to the scene. The results can be seen in Table ~\ref{genericTable} showing benefit of including the generic class. Moreover, the CMs not included due to space limitations, had cleaner diagonals. The reasons for performance improvements on the DCASE set up are suggested by the utilization of less sound classes, which reduces class ambiguity. The utilization of the generic class built with same-scene files as opposed to a mix of both scenes. As well as the optimization per scene of the classifier using Tpot.
\vspace{-0.1in}
\subsection{Generation of Data Through Perturbation}
\vspace{-0.1in}
\label{sec:perturbation}
The scarcity of labeled data per event is a common issue as discussed in~\cite{kumar2016audio,anuragweakly}. Annotations are costly, sounds don't occur with the same frequency and in general it's hard to capture enough variations of the same sound to train robust models. To address this problem, multiple techniques have been explored in the literature such as perturbation of the audio signal as in~\cite{chen2015noise,kanda2013elastic}. The authors presented multiple types of perturbations resulting in improvements of speech separation. For Task 3 we performed time-based perturbation by speeding up and slowing down the sound event samples. We empirically analyzed multiple combinations of speed up-down values for different events. We concluded that speeding up more than 30\% the original signal resulted in unintelligible audio and speeding down the signal more than 100\% would be unlikely to occur. The range included 13 different speed values and the original version.
The set of experiments used the DCASE setup of separate scenes and four folds, and utilized the time-based perturbed audio. For training, we added to the original files the 13 versions of the perturbed audio files, whereas for testing, the set remained intact. The results can be seen in Table~\ref{genericTable}, where the performance for \textit{Home} improved, but not for \textit{Residential}. Thus, we decided to use perturbation for the DCASE evaluation.
\vspace{-0.1in}
\subsection{Sound Event Detection and Submission Systems}
\vspace{-0.1in}
\label{sec:sed}
For Task 3, we used the DCASE setup of separate scenes and four folds in a similar setup as the experiments from Table~\ref{genericTable}. For each scene, we extracted the sound events from the recordings using the annotations from the train set, followed by the extraction of MFCCs features. After, we trained the Tpot optimized multi-class classifier with the event samples. For testing, instead of using sound event files only, we segmented the scene recordings from the test set into one-second consecutive segments. This number was selected due to the metric schema of the DCASE evaluation, which considers one-second segments. After, we extracted audio features from the test segments and evaluate them with the classifier to obtain scores for each trained sound event class. The label corresponding to the highest score was chosen for the segments and then were written down into the DCASE format output file and fed to the official scoring scripts~\footnote{http://www.cs.tut.fi/sgn/arg/dcase2016/sound-event-detection-metrics} along with the ground truth to compute performance.
\begin{figure}[t]
\centering
\includegraphics[width=0.45\textwidth]{mfcc_features_18_plus_generic_normalized_matrix.png}
\caption{[H]ome and [R]esidential with the generic class. Although this class shared the background acoustics with the 18 sounds, it didn't cause major confusion.}
\label{fig:mfcc_features_18_plus_generic_normalized_matrix}
\vspace{-1em}
\end{figure}
We utilized the pipeline for three experiments, without generic class \& without perturbation, with generic class \& without perturbation and with generic class \& with perturbation. The results using the development-test set are shown in Table~\ref{SBER}. The inclusion of the generic class and the perturbation outperformed the baseline method by a significant margin for both \emph{Home} and \emph{Residential} scenes. Our submission consisted on the runs using G and G+P but using the evaluation set. The eval results were SBER of 0.9613 and Fscore of 33.6\% given by the G+P version.
\begin{table}[t]
\small
\centering
\begin{tabular}{ | l | c | c | }
\hline
\textbf{Acoustic Scene} & \textbf{SBER} & \textbf{F-score}\\ \hline
Home & 1.05 & 25.1 \\\hline
Home + G & 0.91 & 23.7 \\\hline
Home + G + P & 0.9 & 24.7 \\\hline
Residential & 0.64 & 54.6 \\\hline
Residential + G & 0.72 & 45.9 \\\hline
Residential + G + P & 0.63 & 52.2 \\\hline
\end{tabular}
\caption{Our Segment-based Error Rate, using [G]eneric and [P]erturbation, outperformed the baseline.}
\label{SBER}
\vspace{-1em}
\end{table}
\vspace{-0.15in}
\section{Conclusion}
\vspace{-0.1in}
\label{sec:conc}
In this paper we showed different approaches for both acoustic scene classification (Task 1) and sound event detection (Task 3) of the 2016 DCASE challenge. On both tasks we were able to obtain significant improvement over the baseline method. For Task 1 we observed that the $\vec{\beta}$ features performed much better than $\vec{\alpha}$ features. Although, linear and RBF kernels with $\vec{\beta}$ features can outperform the baseline by considerable margin on its own, we make note of the fact that a multiple classifier system can give further improvements. For Task 3, we tested different features and classifiers and significantly improved the baseline. Moreover, we explored a way of handling out-of-vocabulary sound segments with the generic class and the inclusion of perturbed audio to add robustness.
\pagebreak
\ninept
\bibliographystyle{IEEEtran}
|
\section*{References}
\section*{References}
\renewcommand\bibsection{%
\paragraph*{}%
}%
|
\section{Calculation of the neutralino nucleon cross section}
\label{Technical}
\subsection{Composition of direct detection rate}
\label{DDFormulas}
In this subsection, we briefly review the standard formulas for the calculation of neutralino direct detection rates. The desired quantity is the rate of events $\mathrm{d}R$ per energy interval $\mathrm{d}E$. This differential event rate is typically expressed in terms of counts per kg and day and keV. It can be written as
\begin{equation}
\label{DDRate2}
\frac{\mathrm{d}R}{\mathrm{d}E} = \sum_i c_i \frac{\sigma_i}{2m_{\tilde{\chi}^0_1}\mu_i^2}\rho_0 \eta_i.
\end{equation}
The sum runs over all detector nuclides $i$, and the factor $c_i$ denotes the mass fraction of the nuclear species $i$ in the detector. Let $m_i$ be the mass of the nucleus of species $i$. Then $\mu_i$ is the reduced mass
\begin{equation}
\mu_i = \frac{m_{\tilde{\chi}^0_1}m_i}{m_{\tilde{\chi}^0_1} + m_i}.
\end{equation}
The local dark matter density is described by $\rho_0$. Before using the canonical value of 0.3 GeV/cm$^3$, one should calculate the neutralino relic density to ensure that its value is in agreement with the experimental constraints and that the neutralinos can solely account for dark matter. $\eta_i$ contains the integration over the dark matter velocity relative to the detector $\vec{v}$,
\begin{equation}
\eta_i = \int_{v_{\mathrm{min},i}}^{v_{\mathrm{esc}}}\mathrm{d}^3v\frac{f(\vec{v})}{v}\quad\mathrm{with}\quad v_{\mathrm{min},i} = \sqrt{\frac{m_iE}{2\mu_i^2}}.
\end{equation}
The lower integration limit $v_{\mathrm{min},i}$ is given by the minimal neutralino velocity, which can cause a recoil energy $E$. The upper integration limit is fixed by the galactical escape speed $v_{\mathrm{esc}}$, which is usually set to 544 km/s. Faster particles are not gravitationally bound in the Milky Way. More details on the integration limits can be found in Refs.\ \cite{SmithRAVE, vesc, CirelliTools}. $f(\vec{v})$ is the local velocity distribution, which is typically assumed to be Maxwellian. However, several studies have unveiled that this simplification might not describe the situation properly, see e.g. \cite{vDistrib, GreenDistribution, DD_ProtonNeutron}. All the particle physics is contained in the cross sections for elastic nucleus-neutralino scattering $\sigma_i$, where we distinguish between spin-independent and spin-dependent contributions.
The spin-independent cross section can be written as
\begin{equation}
\label{sigmaSI}
\sigma_i^{\mathrm{SI}} = \frac{\mu_i^2}{\pi}\left|Z_i g_p^{\mathrm{SI}} +(A_i-Z_i)g_n^{\mathrm{SI}}\right|^2|F_i^{\mathrm{SI}}(Q_i)|^2,
\end{equation}
where $F_i^{\mathrm{SI}}(Q_i)$ is the spin-independent structure function for the nucleus $i$. It depends on the momentum transfer $Q_i = \sqrt{2m_iE}$, can be understood as the Fourier transform of the nucleon density, and is normalised to $F_i^{\mathrm{SI}}(0) = 1$. The nucleus $i$ consists of $Z_i$ protons and $A_i -Z_i$ neutrons, where $Z_i$ is its atomic number and $A_i$ is its mass number. To enable a comparison of direct detection results, that is independent of the detector material and technology, the experimental collaborations typically publish constraints on the cross section of the dark matter particle and a single nucleon\footnote{At this point, the typical assumption is that the interaction strength of neutralinos is the same for protons and neutrons. This is not necessarily fulfilled in a non-minimal model like the MSSM. Therefore we keep our calculations general and distinguish between protons and neutrons.} $N$, which simply reads
\begin{equation}
\label{sigmaSINucleon}
\sigma_N^{\mathrm{SI}} = \frac{\mu_N^2}{\pi}\left|g_N^{\mathrm{SI}}\right|^2.
\end{equation}
Here, the neutralino-nucleus reduced mass $\mu_i$ is replaced by the neutralino-nucleon reduced mass $\mu_N$ in complete analogy. The nucleon masses $m_N$ are given by
\begin{equation}
\label{Nucleonmasses}
m_p = 0.9383\ \mathrm{GeV}\quad\mathrm{and}\quad m_n = 0.9396\ \mathrm{GeV}.
\end{equation}
The effective spin-independent four-fermion couplings among neutralinos and protons $p$ or neutrons $n$ are denoted by $g_p^{\mathrm{SI}}$ and $g_n^{\mathrm{SI}}$. They can be determined via
\begin{equation}
g_N^{\mathrm{SI}} = \sum_q \langle N |\bar{q}q| N\rangle \alpha_{q}^{\mathrm{SI}},
\end{equation}
where the nucleon index $N$ stands either for a proton or a neutron and where the sum runs over all quark types $q$.\footnote{We are summing over all quark types, as we do not include gluon operators yet. Alternatively, one could replace the heavy-quark contributions by loop-induced gluon processes including heavy quarks as virtual particles.} The spin-independent interaction between quarks and neutralinos is denoted by $\alpha_{q}^{\mathrm{SI}}$. The quark matrix element $\langle N |\bar{q}q| N\rangle$ can be qualitatively understood as the probability to find a quark $q$ in the nucleon $N$. We write it as
\begin{equation}
\label{fTDef}
\langle N |m_q \bar{q}q| N\rangle = f_{Tq}^N m_N,
\end{equation}
where $m_N$ denotes the nucleon mass and $m_q$ the quark mass. The scalar coefficients $f_{Tq}^N$ are determined experimentally or via lattice QCD. We point out that especially $f_{Ts}^N$ is affected by experimental uncertainties, which mainly stem from the determination of the pion-nucleon sigma term \cite{NuclearUncertainties, EllisHadron, BottinoNucleon}. We use the values given in Refs.\ \cite{CrivellinNucleon, Hoferichter, Junnarkar} which differ from the ones implemented in {\tt DarkSUSY}\ \cite{DarkSUSY} or {\tt micrOMEGAs}\ \cite{micrOMEGAs}. We list all values for comparison in Tab.\ \ref{fTTable}.\footnote{We are working with {\tt micrOMEGAs}\ \texttt{2.4.1} to benefit from our established relic density interface. However, we have updated the nuclear input values to the most recent version manually. Hence the values given in Tab.\ \ref{fTTable} correspond to {\tt micrOMEGAs}\ \texttt{4.2.5}.} The factors $f_{Tq}^N$ of the heavy quarks are linked to those of the light quarks via \cite{Shifman}
\begin{equation}
f_{Tc}^N = f_{Tb}^N = f_{Tt}^N= \frac{2}{27}\left(1-\sum_{q=u,d,s} f_{Tq}^N\right).
\end{equation}
\begin{table}
\caption{Scalar coefficients $f_{Tq}^N$ used in different codes.}
\begin{center}
\begin{tabular}{|c|ccc|}
\hline
Scalar coefficient & {\tt DM@NLO} & {\tt DarkSUSY} & {\tt micrOMEGAs}\\
\hline
$f_{Tu}^p$ & 0.0208 & 0.023 & 0.0153 \\
$f_{Tu}^n$ & 0.0189 & 0.019 & 0.0110 \\
$f_{Td}^p$ & 0.0411 & 0.034 & 0.0191\\
$f_{Td}^n$ & 0.0451 & 0.041 & 0.0273\\
$f_{Ts}^p = f_{Ts}^n$ & 0.043 & 0.14 & 0.0447\\
$f_{Tc}^p = f_{Tb}^p = f_{Tt}^p$ & 0.0663 & 0.0595 & 0.0682\\
$f_{Tc}^n = f_{Tb}^n = f_{Tt}^n$ & 0.0661 & 0.0592 & 0.0679\\
\hline
\end{tabular}
\end{center}
\label{fTTable}
\end{table}
The spin-dependent cross section can be cast into the form
\begin{eqnarray}
\label{sigmaSD}
\sigma_i^{\mathrm{SD}} & = & \frac{4\mu_i^2}{2J +1}\big(|g_p^{\mathrm{SD}}|^2S_{\mathrm{pp},i}(Q_i) + |g_n^{\mathrm{SD}}|^2S_{\mathrm{nn},i}(Q_i)\nonumber\\
&& + |g_p^{\mathrm{SD}}g_n^{\mathrm{SD}}|S_{\mathrm{pn},i}(Q_i)\big),
\end{eqnarray}
where $J$ denotes the nuclear spin. Details on the spin structure functions $S_{\mathrm{pp},i}(Q_i)$, $S_{\mathrm{nn},i}(Q_i)$ and $S_{\mathrm{pn},i}(Q_i)$ can be found in Ref.\ \cite{VogelStructure}. The spin-dependent cross section for a neutralino and a single nucleon $N$ reads
\begin{equation}
\label{sigmaSDNucleon}
\sigma_N^{\mathrm{SD}} = \frac{3\mu_N^2}{\pi}|g_N^{\mathrm{SD}}|^2.
\end{equation}
The effective spin-dependent four-fermion couplings among neutralinos and protons $p$ ($g_p^{\mathrm{SD}}$) or neutrons $n$ ($g_n^{\mathrm{SD}}$) are given by
\begin{equation}
g_N^{\mathrm{SD}} = \sum_{q= u,d,s} (\Delta q)_N \alpha_{q}^{\mathrm{SD}}.
\end{equation}
In contrast to the spin-independent case, we sum only over the light quarks $u$, $d$ and $s$, as mainly these flavors contribute to the spin of the nucleon.\footnote{Note, however, that it was recently claimed that bottom quarks may also contribute to the spin-dependent interaction \cite{LiBottom}.} $(\Delta q)_N$ can be seen as the fraction of the nucleon spin carried by the quark $q$. More precisely, it describes the second moment of the polarized quark density and is related to the nucleon spin vector $s_\mu$ via
\begin{equation}
\langle N | \bar{q}\gamma_\mu\gamma_5 q| N\rangle = 2s_\mu(\Delta q)_N.
\end{equation}
We choose the default values of {\tt micrOMEGAs}\ for the polarized quark densitites
\begin{eqnarray}
(\Delta u)_p = (\Delta d)_n & = & 0.842,\label{Delta1}\\
(\Delta d)_p = (\Delta u)_n & = & -0.427,\label{Delta2}\\
(\Delta s)_p = (\Delta s)_n & = & -0.085,
\end{eqnarray}
constrained by isospin symmetry, i.e.\ $(\Delta u)_p = (\Delta d)_n$ and $(\Delta d)_p = (\Delta u)_n$.
\subsection{Renormalization scheme}
\label{Renormalization}
Our QCD calculations at next-to-leading order (NLO) and beyond are performed within a hybrid on-shell/$\overline{\rm DR}$ renormalization scheme, described in detail in Refs.\ \cite{ChiChi2qq3, NeuQ2qx1, NeuQ2qx2}. In the quark sector, the top and bottom quark masses are defined on-shell and in the $\overline{\rm DR}$ scheme, respectively. Note that through the Yukawa coupling to (in particular the neutral pseudoscalar) Higgs boson resonances, the bottom quark mass can have a sizeable influence on the dark matter annihilation cross section and must therefore be treated with particular care. We obtain it from the SM $\overline{\rm MS}$ mass $m_b(m_b)$, determined in an analysis of $\Upsilon$ sum rules, through evolution to the scale $\mu_R$, transformation to the SM $\overline{\rm DR}$ and then MSSM $\overline{\rm DR}$ scheme \cite{ChiChi2qq3, NeuQ2qx1}. In the squark sector, we have five independent parameters
\begin{equation}
m_{\tilde{t}_1}, \quad m_{\tilde{b}_1}, \quad m_{\tilde{b}_2}, \quad A_t \quad\mathrm{and}\quad A_b=0.
\label{eq:RenInput}
\end{equation}
The lighter stop mass and the two sbottom masses are taken to be on-shell, while the stop and sbottom trilinear coupling parameters are taken in the $\overline{\rm DR}$ scheme. From these parameters, we compute as dependent quantities the stop and sbottom mixing angles $\theta_{\tilde{t}}$ and $\theta_{\tilde{b}}$ and $m_{\tilde{t}_2}$ for the heavier stop \cite{NeuQ2qx1}. The masses of the first- and second-generation squarks are taken on-shell. The strong coupling constant $\alpha_s(\mu_R)$ is renormalized in the MSSM $\overline{\rm DR}$ scheme with six active flavors and obtained after evolution of the world-average, five-flavor SM $\overline{\rm MS}$ value at the $Z^0$-boson mass to the renormalization scale $\mu_R$ and an intermediate transformation to the SM $\overline{\rm DR}$ scheme \cite{NeuQ2qx2}.
Although EFT calculations are usually performed in a minimal scheme such as {$\overline{\mathrm{MS}}$}\ or its SUSY equivalent {$\overline{\mathrm{DR}}$}, we continue to use the hybrid scheme presented above for three main reasons: First, we want to combine our direct detection calculations with our relic density analysis, where this scheme has proven very reliable. In particular, the on-shell description of the top quark leads to improved perturbative stability and better fits our supersymmetric processes and top quark final states in comparison to a definition in the {$\overline{\mathrm{DR}}$}\ scheme \cite{Scalepaper}. A second reason is that the hybrid scheme also leads to improved perturbative stability for direct detection as described below in Sec.\ \ref{ScenarioC}. The last reason is that using this hybrid scheme allows for simpler comparison of the leading order result with {\tt micrOMEGAs}. This is due to fact that both, our calculation and {\tt micrOMEGAs}, use the same on-shell squark masses calculated by {\tt SPheno}\ as described in section \ref{Numerics}.
\subsection{Matching of the full and effective theory}
\begin{figure}
\includegraphics[width=0.49\textwidth]{DDTree.pdf}
\caption{Tree-level processes in the full theory.}
\label{fig:DDTree}
\end{figure}
\begin{figure}
\includegraphics[width=0.18\textwidth]{DDEFTTree2.pdf}
\includegraphics[width=0.18\textwidth]{DDEFTVertex.pdf}
\caption{Tree-level process (left) and virtual correction (right) in the effective theory.}
\label{fig:DDEFT}
\end{figure}
This subsection is devoted to the matching of the full theory, namely the MSSM, valid at high energies ($\mu_\mathrm{high}\sim 1$ TeV) onto the effective energy valid at low energies ($\mu_\mathrm{low}\sim 5$ GeV). The tree-level diagrams of the scattering process $\tilde{\chi}^0_1q\rightarrow\tilde{\chi}^0_1q$ within the MSSM are shown in Fig.\ \ref{fig:DDTree}. The corresponding amplitudes have to be evaluated at vanishing relative velocity and mapped onto the yet unknown Wilson coefficients $c_1$ and $c_2$ of the effective Lagrangian
\begin{equation}
\label{Leff}
\mathcal{L}_\mathrm{eff} = c_1Q_1 + c_2Q_2 = c_1\bar{\chi}\chi\bar{q}q + c_2\bar{\chi}\gamma_\mu\gamma_5\chi\bar{q}\gamma^\mu\gamma_5q.
\end{equation}
We stress that in this convention a factor $m_q$ has to be factored out of $c_1$ when replacing the nuclear matrix elements via Eq.\ (\ref{fTDef}). Both of the operators $Q_1$ and $Q_2$ given above lead to an effective four-fermion interaction as shown in the left diagram of Fig.\ \ref{fig:DDEFT}. The Higgs processes contribute solely to the scalar operator and the $Z^0$ processes solely to the axial-vector operator. We include only the scalar Higgs bosons $h^0$ and $H^0$ and not the pseudoscalar Higgs boson $A^0$, since the latter leads to the kinematically suppressed operator $\bar{\chi}\gamma_5\chi\bar{q}\gamma_5q$. The squark processes contribute to both operators. To bring the spinor fields into the desired order, a Fierz transformation has to be performed in this case.
The aforementioned mapping onto the Wilson coefficients is governed by the matching condition. This condition demands that the amplitude of the full theory $\mathcal{M}_\mathrm{full}$ is reproduced by the effective theory at the high scale $\mu_\mathrm{high}$. At tree level we have
\begin{equation}
\mathcal{M}_\mathrm{full}^\mathrm{tree} \overset{!}{=} \mathcal{M}_\mathrm{eff}^\mathrm{tree} = c_1^\mathrm{tree}Q_1^\mathrm{tree} + c_2^\mathrm{tree}Q_2^\mathrm{tree}\label{TreeMatching}
\end{equation}
which leads to
\begin{eqnarray}
c_1^\mathrm{tree} = \alpha_q^{\mathrm{SI}} & = & \sum_{\phi = h^0,H^0}\frac{g_{\tilde{\chi}\tilde{\chi}\phi}^Rg^L_{qq\phi}}{m_\phi^2} - \frac{1}{4}\sum_{i=1}^2\frac{g_{\tilde{\chi}\tilde{q}_iq}^Lg_{\tilde{\chi}\tilde{q}_iq}^{R*}}{m_{\tilde{q}_i}^2 - s}\nonumber\\
&& - \frac{1}{4}\sum_{i=1}^2\frac{g_{\tilde{\chi}\tilde{q}_iq}^Lg_{\tilde{\chi}\tilde{q}_iq}^{R*}}{m_{\tilde{q}_i}^2 - u},\\
c_2^\mathrm{tree} = \alpha_q^{\mathrm{SD}} & = & \frac{1}{2}\frac{g_{\tilde{\chi}\tilde{\chi}Z^0}^R(g^L_{qqZ^0} - g^R_{qqZ^0})}{m_{Z^0}^2}\nonumber\label{CSITree}\\
&& + \frac{1}{8}\sum_{i=1}^2\frac{|g_{\tilde{\chi}\tilde{q}_iq}^L|^2 + |g_{\tilde{\chi}\tilde{q}_iq}^{R}|^2}{m_{\tilde{q}_i}^2 - s}\nonumber\\
&& + \frac{1}{8}\sum_{i=1}^2\frac{|g_{\tilde{\chi}\tilde{q}_iq}^L|^2 + |g_{\tilde{\chi}\tilde{q}_iq}^{R}|^2}{m_{\tilde{q}_i}^2 - u}.\label{CSDTree}
\end{eqnarray}
In the limit of vanishing relative velocity, the Mandelstam variables $s$ and $u$ simplify to $(m_{\tilde{\chi}^0_1} \pm m_q)^2$, respectively. The elementary couplings between three particles $a$, $b$ and $c$ are denoted by $g_{abc}$. Using the chirality projectors $P_{L/R} = ({\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l}{\rm 1\mskip-4.5mu l} {\rm 1\mskip-5mu l}} \mp \gamma_5)/2$, they can be decomposed into left- and right-handed parts via
\begin{equation}
g_{abc} = g_{abc}^LP_L + g_{abc}^RP_R.
\end{equation}
Explicit expressions for the couplings can be found, e.g., in Ref.\ \cite{DreesBook}. The tree-level results have been analytically compared with those implemented in {\tt DarkSUSY}. Taking into account that {\tt DarkSUSY}\ does not distinguish between $s$- and $u$-channels, we find perfect agreement.
\begin{figure}
\includegraphics[width=0.49\textwidth]{DDProp.pdf}
\includegraphics[width=0.49\textwidth]{DDVertex.pdf}
\includegraphics[width=0.49\textwidth]{DDBox.pdf}
\caption{Virtual corrections in the full theory.}
\label{fig:DDCorrections}
\end{figure}
So far we have basically reproduced already available results. The next step is to improve on the tree-level calculation by including all $\mathcal{O}(\alpha_s)$ corrections to the leading operators. The corresponding diagrams within the full theory are shown in Fig.\ \ref{fig:DDCorrections}. We distinguish between propagator corrections (the first row), vertex corrections (the second and the third row) and box contributions (the last row).
We have calculated all of the loop amplitudes in full generality using dimensional reduction. The gluon propagator correction shown as the third diagram in the first row then vanishes, as it is proportional to the scaleless scalar integral $A_0(0) = 0$. In the case of the other propagator and the vertex corrections, we were able to benefit from previous loop calculations performed in the context of Ref.\ \cite{ChiChi2qq3}. The box amplitudes were calculated from scratch. These amplitudes lead to a plethora of effective operators. We keep only the most relevant, namely those of Eq.\ (\ref{Leff}). In case of the gluon boxes, Fierz transformations are necessary again. Explicit expressions for all involved loop amplitudes will be given in Ref.\ \cite{myPhD}.
In contrast to our relic density calculations, the loops are evaluated at zero relative velocity in the context of direct detection. This leads to additional problems, namely vanishing Gram determinants. We illustrate this technical issue separately in App.\ \ref{GramDet}.
The propagator and vertex corrections give rise to ultraviolet divergences. These divergences are removed via renormalization (cf.\ Sec.\ \ref{Renormalization}), i.e.\ by adding the corresponding counterterms. A detailed description of the counterterms involved here is given in Ref.\ \cite{NeuQ2qx1}. As we always distinguish ultraviolet and infrared poles ($\epsilon_{UV}$ and $\epsilon_{IR}$) when evaluating loop integrals, we were able to explicitly check the ultraviolet safety of our calculation.
Having the renormalized amplitudes of the full theory at hand, we can start with the matching procedure at NLO. The matching condition remains basically unchanged and reads
\begin{eqnarray}\hspace*{-1cm}
\mathcal{M}_\mathrm{full}^\mathrm{NLO} & \overset{!}{=} & \mathcal{M}_\mathrm{eff}^\mathrm{NLO}.\\
\hspace*{-1cm}\Leftrightarrow\mathcal{M}_\mathrm{full}^\mathrm{tree} + \mathcal{M}_\mathrm{full}^\mathrm{1loop} & \overset{!}{=} & c_1^\mathrm{NLO}Q_1^\mathrm{NLO} + c_2^\mathrm{NLO}Q_2^\mathrm{NLO}.\label{NLOMatching}
\end{eqnarray}
In this convention, the full NLO result consists of the tree-level result and its $\mathcal{O}(\alpha_s)$ one-loop correction. The latter includes all the virtual corrections depicted in Fig.\ \ref{fig:DDCorrections}. The crucial point in Eq. (\ref{NLOMatching}) is that there is a one-loop correction to the Wilson coefficients \textit{and} the effective operators. We neglect terms of $\mathcal{O}(\alpha_s^2)$ and write
\begin{widetext}
\begin{eqnarray}
\mathcal{M}_\mathrm{full}^\mathrm{tree} + \mathcal{M}_\mathrm{full}^\mathrm{1loop} & \overset{!}{=} & (c_1^\mathrm{tree} + c_1^\mathrm{1loop})(Q_1^\mathrm{tree}+ Q_1^\mathrm{1loop}) + (c_2^\mathrm{tree} + c_2^\mathrm{1loop})(Q_2^\mathrm{tree} + Q_2^\mathrm{1loop})\nonumber\\
& = & c_1^\mathrm{tree}Q_1^\mathrm{tree} + c_2^\mathrm{tree}Q_2^\mathrm{tree} + c_1^\mathrm{1loop}Q_1^\mathrm{tree} + c_2^\mathrm{1loop}Q_2^\mathrm{tree} + c_1^\mathrm{tree}Q_1^\mathrm{1loop} + c_2^\mathrm{tree}Q_2^\mathrm{1loop}.
\end{eqnarray}
\end{widetext}
At $\mathcal{O}(\alpha_s^0)$, we reproduce the tree-level matching condition Eq. (\ref{TreeMatching}). At $\mathcal{O}(\alpha_s)$, we obtain
\begin{eqnarray}
&&\mathcal{M}_\mathrm{full}^\mathrm{1loop} - c_1^\mathrm{tree}Q_1^\mathrm{1loop} - c_2^\mathrm{tree}Q_2^\mathrm{1loop}\nonumber\\
&=& c_1^\mathrm{1loop}Q_1^\mathrm{tree} + c_2^\mathrm{1loop}Q_2^\mathrm{tree}.
\end{eqnarray}
Before we can calculate the $\mathcal{O}(\alpha_s)$ corrections to the Wilson coefficients, i.e.\ determine $c_1^\mathrm{1loop}$ and $c_2^\mathrm{1loop}$, we have to identify the one-loop corrections to the effective operators $Q_1^\mathrm{1loop}$ and $Q_2^\mathrm{1loop}$. These can be written as
\begin{eqnarray}
Q_1^\mathrm{1loop} & = & (\mathcal{K}_\mathrm{EFTV1} + \mathcal{K}_\mathrm{EFTVC1})Q_1^\mathrm{tree}\quad\mathrm{and}\quad\\
Q_2^\mathrm{1loop} & = & (\mathcal{K}_\mathrm{EFTV2} + \mathcal{K}_\mathrm{EFTVC2})Q_2^\mathrm{tree},
\end{eqnarray}
i.e.\ they can be expressed as the tree-level operators multiplied with correction factors describing vertex corrections and vertex counterterms in the effective field theory. The vertex correction in the effective field theory is depicted on the right of Fig.\ \ref{fig:DDEFT}. This allows us to explicitly write down the one-loop Wilson coefficients as
\begin{eqnarray}
c_1^\mathrm{1loop} & = & \alpha^\mathrm{SI}_{q,\mathrm{P}} + \alpha^\mathrm{SI}_{q,\mathrm{PC}} + \alpha^\mathrm{SI}_{q,\mathrm{V}} + \alpha^\mathrm{SI}_{q,\mathrm{VC}}\nonumber\\
& + & \alpha^\mathrm{SI}_{q,\mathrm{B}} - c_1^\mathrm{tree}(\mathcal{K}_\mathrm{EFTV1} + \mathcal{K}_\mathrm{EFTVC1}).\\
c_2^\mathrm{1loop} & = & \alpha^\mathrm{SD}_{q,\mathrm{P}} + \alpha^\mathrm{SD}_{q,\mathrm{PC}} + \alpha^\mathrm{SD}_{q,\mathrm{V}} + \alpha^\mathrm{SD}_{q,\mathrm{VC}}\nonumber\\
& + & \alpha^\mathrm{SD}_{q,\mathrm{B}} - c_2^\mathrm{tree}(\mathcal{K}_\mathrm{EFTV2} + \mathcal{K}_\mathrm{EFTVC2}).
\end{eqnarray}
Here $\alpha^\mathrm{SI}_{q,\mathrm{P}}$, $\alpha^\mathrm{SI}_{q,\mathrm{PC}}$, $\alpha^\mathrm{SI}_{q,\mathrm{V}}$, $\alpha^\mathrm{SI}_{q,\mathrm{VC}}$ and $\alpha^\mathrm{SI}_{q,\mathrm{B}}$ denote the contributions to the spin-independent four-fermion coupling stemming from the propagator corrections, propagator counterterm, vertex correction, vertex counterterms and box diagrams, respectively. The spin-dependent contributions are labeled analogous. All of these terms will be given explicitly in Ref. \cite{myPhD}.
We stress that $\alpha^\mathrm{SI}_{q,\mathrm{P}} + \alpha^\mathrm{SI}_{q,\mathrm{PC}}$, $\alpha^\mathrm{SI}_{q,\mathrm{V}} + \alpha^\mathrm{SI}_{q,\mathrm{VC}}$, $\alpha^\mathrm{SI}_{q,\mathrm{B}}$ and $\mathcal{K}_\mathrm{EFTV1} + \mathcal{K}_\mathrm{EFTVC1}$ are separately ultraviolet finite, and the same holds for the spin-dependent case and the associated correction factors $\mathcal{K}_\mathrm{EFTV2} + \mathcal{K}_\mathrm{EFTVC2}$. However, there are also infrared divergences involved, which have not been discussed yet. Although most of the individual terms given above are infrared divergent, $c_1^\mathrm{1loop}$ and $c_2^\mathrm{1loop}$ as a whole are infrared finite, which is an essential feature of the matching procedure. The appearance of infrared divergences is connected with massless particles like gluons. These particles are likewise degrees of freedom in the full and the effective theory. In other words: The infrared regime of both theories is the same. Whenever there occurs an infrared divergence in the full theory, the very same infrared divergence occurs in the effective theory as well, and both cancel during the matching procedure. In our calculation, this cancellation is due to the correction factors $\mathcal{K}_\mathrm{EFTV1}$, $\mathcal{K}_\mathrm{EFTVC1}$, $\mathcal{K}_\mathrm{EFTV2}$ and $\mathcal{K}_\mathrm{EFTVC2}$ which we list now.
The vertex correction factor $\mathcal{K}_\mathrm{EFTV1}$ is obtained by calculating the diagram shown on the right of Fig.\ \ref{fig:DDEFT} involving the effective operator $Q_1^\mathrm{tree}$. We get
\begin{eqnarray}
\mathcal{K}_\mathrm{EFTV1} & = & \frac{\alpha_sC_F}{4\pi}\Big(4B_0 - 2 + 4p_bp_2(C_0 + C_1 + C_2)\Big),\nonumber\\
\end{eqnarray}
where the two- and three-point functions possess the arguments $B = B(p_b - p_2, m_q^2, m_q^2)$ and $C = C(p_2, p_b, 0, m_q^2, m_q^2)$. Here the four-momentum of the ingoing quark is denoted by $p_b$ and that of the outgoing quark by $p_2$. In the limit of vanishing relative velocity, we simply have $p = p_b = p_2$. Moreover, $C_F = 4/3$ denotes the usual color factor. This vertex correction is algebraically identical to the Higgs-gluon vertex shown on the very left in the second row of Fig.\ \ref{fig:DDEFT}, which has two important consequences. On the one hand, the Higgs-gluon vertex completely cancels in the matching procedure. The gluon is likewise a degree of freedom in the full and the effective theory and therefore the corresponding vertex correction occurs in both theories. It is included in the effective operator, not the Wilson coefficient. Moreover the correction factor $\mathcal{K}_\mathrm{EFTV1}$ is ultraviolet divergent, as it includes the two-point function $B_0$. To allow for a consistent matching procedure, we have to renormalize the effective theory in the same way as the full theory. This means that we have to add a counterterm $\delta c_1$ to the four-fermion coupling. This counterterm has to be of the same form as $\delta g_{\phi qq}$ (with $\phi = h^0,H^0$) and reads
\begin{equation}
\delta c_1^L = c_1^{\mathrm{tree},L}\left(\frac{\delta Z_m}{m_q} + \frac{1}{2}\delta Z_q^L + \frac{1}{2}\delta Z_q^{R*}\right),
\end{equation}
where $\delta Z_m$ denotes the mass and $\delta Z_q$ the wave function counterterm. For more details on these counterterms we refer the reader again to Ref.\ \cite{NeuQ2qx1}. The associated right-handed part of $\delta c_1$ is obtained by the substitution $L\leftrightarrow R$. The correction factor $\mathcal{K}_\mathrm{EFTVC1}$ is then simply given by
\begin{equation}
\mathcal{K}_\mathrm{EFTVC1} = \frac{\delta c_1^L/c_1^{\mathrm{tree},L} + \delta c_1^R/c_1^{\mathrm{tree},R}}{2}.
\end{equation}
Remember that $c_1^\mathrm{tree}$ does not only incorporate Higgs contributions, but that squark processes contribute as well (cf.\ Eq.\ (\ref{CSITree})). Whereas the Higgs-gluon vertex correction and its associated counterterm completely vanish in the matching procedure, this is not true for the vertex corrections to the squark processes shown in the third row of Fig.\ \ref{fig:DDCorrections} and their counterterms. However, the infrared divergences of these corrections and the ones stemming from the boxes shown in the last row of Fig.\ \ref{fig:DDCorrections} are precisely cancelled by the correction factors. This is an important consistency check of the whole calculation. Thanks to our generic implementation of loop integrals and the discrimination between ultraviolet and infrared poles, we could verify this cancellation explicitly.
We continue with the determination of $\mathcal{K}_\mathrm{EFTV2}$, i.e.\ the vertex correction factor for the spin-dependent operator $Q_2$. The associated diagram is shown on the right of Fig. \ref{fig:DDEFT} again, the only difference to the previous case is the included four-fermion coupling. Keeping only the relevant effective operators, we obtain
\begin{eqnarray}
\mathcal{K}_\mathrm{EFTV2} & = & \frac{\alpha_sC_F}{4\pi}\Big(2B_0 + 4p_bp_2(C_0 + C_1 + C_2)\nonumber\\
&& - 4C_{00} -1\Big)
\end{eqnarray}
where the two- and three-point functions possess the same arguments as before. The missing piece is the counterterm $\delta c_2$, which renders the vertex correction given above ultraviolet finite. This counterterm is constructed in analogy to $\delta g_{Z^0qq}$ and reads
\begin{equation}
\delta c_2^L = c_2^{\mathrm{tree},L}\left(\frac{1}{2}\delta Z_q^{\mathrm{SM},L} + \frac{1}{2}\delta Z_q^{\mathrm{SM},L*} + \frac{\alpha_sC_F}{\pi}\right).
\label{EFTVertexCSD}
\end{equation}
As before, the correction factor $\mathcal{K}_\mathrm{EFTVC2}$ is obtained via
\begin{equation}
\mathcal{K}_\mathrm{EFTVC2} = \frac{\delta c_2^L/c_2^{\mathrm{tree},L} + \delta c_2^R/c_2^{\mathrm{tree},R}}{2}.
\end{equation}
Note that we have included the additional finite part $\frac{\alpha_sC_F}{\pi}$ to retain a conventional axial current divergence which is in agreement with Refs. \cite{Hill2} and \cite{Larin}.\footnote{The results given in Ref. \cite{Larin} were obtained using the {$\overline{\mathrm{MS}}$}\ scheme and dimensional regularization. Transferring results from this scheme to the {$\overline{\mathrm{DR}}$}\ scheme and dimensional reduction -- which we are using -- is nontrivial in general. Discrepancies may arise due to the treatment of $\gamma_5$ in $D$ dimensions. However, these problems should occur at the three-loop order for the first time and do neither affect the finitite contribution included in Eq. (\ref{EFTVertexCSD}) nor the running of the axial-vector operator presented in the next section \cite{LarinMail}.} Moreover we incorporate just Standard Model contributions to $\delta Z_q$ in this case. The reason is as follows: In case of the Higgs vertex corrections including the gluon and the gluino, only the former is ultraviolet divergent. The whole counterterm $\delta g_{\phi qq}$ is responsible for the cancellation of this divergence. As the gluon vertex correction occurs likewise in the effective theory, we have constructed its associated counterterm $\delta c_1$ in complete analogy to $\delta g_{\phi qq}$. In case of the $Z^0$ vertex corrections including the gluon and the gluino, both are ultraviolet divergent. The divergences of the first diagram are removed by the Standard Model part of $\delta g_{Z^0 qq}$ and the latter by the SUSY part of $\delta g_{Z^0 qq}$. During the matching procedure, the gluon vertex correction and its corresponding counterterm has to cancel, whereas the vertex correction including the gluino and its counterterm contributes to the Wilson coefficient. Hence we only include Standard Model contributions to the spinor field counterterms in $\delta c_2$. This completes our matching calculation at NLO.
\subsection{Running of effective operators and associated Wilson coefficients}
\label{RunningSection}
The matching calculation presented in the last subsection is performed at the high scale $\mu_\mathrm{high}\sim 1$ TeV. In contrast, the nuclear matrix elements are defined at a low scale $\mu_\mathrm{low}\sim 5$ GeV. This is the energy regime we finally aim to describe with our effective field theory. To connect the two energy regimes, we have to evolve the effective operators and associated Wilson coefficients from the high scale down to the low scale by solving the corresponding renormalization group equations (RGEs). This part of the calculation is briefly referred to as ``running'' and is presented in this subsection.
The scale dependence of the Wilson coefficients is inverse to that of the corresponding operators. Therefore it cancels in the product, which is an essential feature of any operator product expansion. In the effective Lagrangian introduced in Eq.\ (\ref{Leff}), we have neglected higher-dimensional operators in our operator product expansion, i.e.\
\begin{equation}
\mathcal{L}_\mathrm{eff} = \sum_{i=1}^\infty c_iQ_i \approx c_1\bar{\chi}\chi\bar{q}q + c_2\bar{\chi}\gamma_\mu\gamma_5\chi\bar{q}\gamma^\mu\gamma_5q +\ldots
\end{equation}
As we are interested only in QCD effects, the running of the two operators given above is solely determined by their respective quark parts.
The scalar operator $m_q\bar{q}q$ is scale independent. As a consequence, the running calculation in the spin-indepependent case is rather simple. We have to factor out the quark mass $m_q(\mu_\mathrm{high})$ from the coefficient $c_1$. This quark mass has to be evolved down to the low scale $\mu_\mathrm{low}$ in the usual way, i.e.\ by solving its RGE. We then replace the combination $m_q(\mu_\mathrm{low})\bar{q}q$ via Eq.\ (\ref{fTDef}).
In contrast to that, the renormalization and the resulting running of the axial-vector operator is not trivial. This calculation has first been performed in Ref.\ \cite{Larin}. The relevant renormalization constant reads
\begin{eqnarray}
Z_A^\mathrm{Singlet} & = & 1 + \frac{\alpha_s}{\pi}C_F - \frac{1}{\epsilon_{UV}}\left(\frac{\alpha_s}{4\pi}\right)^2\left(\frac{20}{9}n_f + \frac{88}{3}\right)\nonumber\\
&& + \mathcal{O}(\alpha_s^3),
\end{eqnarray}
where $n_f$ denotes the number of active flavors and an additional finite term has been included to cure the axial anomaly. It is precisely this term which has been included in Eq.\ (\ref{EFTVertexCSD}) as well. Finite terms of order $\mathcal{O}(\alpha_s^2)$ have been neglected, as they are irrelevant for the running up to the desired order. Given this constant, we can calculate the corresponding anomalous dimension via
\begin{equation}
\gamma_A^\mathrm{Singlet} = (Z_A^\mathrm{Singlet})^{-1}\frac{\mathrm{d}}{\mathrm{d}\log\mu}Z_A^\mathrm{Singlet}
\end{equation}
and obtain
\begin{equation}
\gamma_A^\mathrm{Singlet} = \left(\frac{\alpha_s}{4\pi}\right)^2 16n_f + \mathcal{O}(\alpha_s^3).
\end{equation}
To arrive at this result one has to insert the RGE of the strong coupling constant including its divergent part, namely
\begin{equation}
\frac{\mathrm{d}g}{\mathrm{d}\log\mu} = -\epsilon_{UV}g + \beta(g),
\end{equation}
where $\beta(g)$ is the usual QCD beta function
\begin{eqnarray}
\frac{\beta(g)}{g} & = & -\beta_0\frac{\alpha_s}{4\pi} + \mathcal{O}(\alpha_s^2) = -(11-\frac{2}{3}n_f)\frac{\alpha_s}{4\pi} + \mathcal{O}(\alpha_s^2).\nonumber\\
\end{eqnarray}
The remaining step is to determine the running of the Wilson coefficient $c_2$ via
\begin{equation}
\frac{\mathrm{d}}{\mathrm{d}\log\mu}c_2(\mu) = \gamma_A^\mathrm{Singlet}c_2(\mu).
\end{equation}
We finally obtain
\begin{equation}
\frac{c_2(\mu_\mathrm{low})}{c_2(\mu_\mathrm{high})} = \exp\left(\frac{2n_f(\alpha_s(\mu_\mathrm{high}) - \alpha_s(\mu_\mathrm{low}))}{\beta_0\pi}\right),
\end{equation}
which agrees with the result given in Ref.\ \cite{Hill2}. Note that in general different operators may mix under renormalization. This is fortunately not the case here, but it will happen when one includes e.g.\ the gluon operator $G_{\mu\nu}G^{\mu\nu}$ \cite{Hill2}.
\section{Tensor reduction for vanishing Gram determinant}
\label{GramDet}
In the course of the {\tt DM@NLO}\ project, we have computed a large collection of loop integrals and associated special cases in generic form. In addition, we always distinguish between infrared and ultraviolet divergences. As these divergences have to vanish when determining physical observables, this discrimination allows for powerful checks of our calculations. Hence it is desirable to use the same thoroughly tested routines for the new direct detection calculation. However, in the context of direct detection, all the amplitudes are evaluated at zero momentum transfer.\footnote{To determine the relic density, only cross sections including a finite relative velocity are needed, as those with zero relative velocity are weighted by zero in the thermal averaging procedure.} This causes problems for the tensor reduction of loop amplitudes which we employ \cite{PVIntegrals}.
In this appendix we present our alternative approach, which is partially based on Ref.\ \cite{Ganesh}. To keep the discussion transparent and to stress the general idea, we restrict ourselves to the simple case of determining the tensor coefficients $C_1$ and $C_2$. All other necessary tensor coefficients can be worked out analogously.
We start by setting up our notation. The scalar and tensor integrals relevant for our discussion are defined via
\begin{eqnarray}
B_0(p_1, m_0^2, m_1^2) & = & \frac{(2\pi\mu_R)^{4-D}}{i\pi^2}\int\mathrm{d}^Dq\frac{1}{\mathcal{D}_0\mathcal{D}_1},\label{B0}\nonumber\\
&&\\
B_\mu(p_1, m_0^2, m_1^2) & = & \frac{(2\pi\mu_R)^{4-D}}{i\pi^2}\int\mathrm{d}^Dq\frac{q_\mu}{\mathcal{D}_0\mathcal{D}_1},\label{Bmu}\nonumber\\
&&\\
C_0(p_1, p_2, m_0^2, m_1^2, m_2^2) & = & \frac{(2\pi\mu_R)^{4-D}}{i\pi^2}\int\mathrm{d}^Dq\frac{1}{\mathcal{D}_0\mathcal{D}_1\mathcal{D}_2},\label{C0}\nonumber\\
&&\\
C_\mu(p_1, p_2, m_0^2, m_1^2, m_2^2) & = & \frac{(2\pi\mu_R)^{4-D}}{i\pi^2}\int\mathrm{d}^Dq\frac{q_\mu}{\mathcal{D}_0\mathcal{D}_1\mathcal{D}_2},\label{Cmu}\nonumber\\
&&\\
C_{\mu\nu}(p_1, p_2, m_0^2, m_1^2, m_2^2) & = & \frac{(2\pi\mu_R)^{4-D}}{i\pi^2}\int\mathrm{d}^Dq\frac{q_\mu q_\nu}{\mathcal{D}_0\mathcal{D}_1\mathcal{D}_2}.\label{Cmunu}\nonumber\\
\end{eqnarray}
Here $\mu_R$ denotes the renormalisation scale which has been introduced to fix the mass dimension of the integrals. The denominators are given by $\mathcal{D}_i = (q+p_i)^2 - m_i^2 + i\epsilon$ with $p_0 = 0$. The idea of the tensor reduction method is to decompose the tensor integrals into a linear combination of all possible Lorentz structures accompanied by yet unknown tensor coefficients. Omitting the arguments, we have
\begin{eqnarray}
B_\mu & = & p_{1,\mu}B_1,\\
C_\mu & = & p_{1,\mu}C_1 + p_{2,\mu}C_2,\\
C_{\mu\nu} &= & g_{\mu\nu}C_{00} + p_{1,\mu}p_{1,\nu}C_{11} + p_{2,\mu}p_{2,\nu}C_{22} \nonumber\\
&& + (p_{1,\mu}p_{2,\nu} + p_{2,\mu}p_{1,\nu})C_{12}.\label{StandardDecomposition}
\end{eqnarray}
The tensor coefficents are obtained by multiplying both sides with the available Lorentz invariants. In this way, the tensor integrals are reduced to a combination of scalar integrals. When determining $C_1$ and $C_2$ we have to solve a set of linear equations which results in
\begin{eqnarray}
\begin{pmatrix} C_1 \\ C_2 \end{pmatrix} & = & A^{-1}\begin{pmatrix} R_1 \\ R_2 \end{pmatrix} = \frac{1}{\det{A}}\begin{pmatrix} p_2^2 & -p_1p_2\\ -p_1p_2 & p_1^2\end{pmatrix}\begin{pmatrix} R_1 \\ R_2 \end{pmatrix}\label{C1-C2},\nonumber\\
&&
\end{eqnarray}
where we have introduced
\begin{eqnarray}
\det(A) & = & p_1^2p_2^2 -(p_1p_2)^2,\\
R_1 & = & \frac{1}{2}\left(B_0(0,2) - B_0(1,2) - f_1C_0\right),\label{R1}\\
R_2 & = &\frac{1}{2}\left(B_0(0,1) - B_0(1,2) - f_2C_0\right),\label{R2}\\
f_i & = & p_i^2 -m_i^2 + m_0^2\quad\mathrm{with}\quad i=1,2.\label{fi}
\end{eqnarray}
Furthermore we define the shorthand notation $B_0(i,j) = B_0(p_j-p_i, m_i^2, m_j^2)$, which we use analogously for $B_1$.
The method illustrated above breaks down when the matrix $A$ is not invertible, i.e.\ when $\det{(A)}$ vanishes. However, instead of using Eq. (\ref{C1-C2}) for determining the tensor coefficients $C_1$ and $C_2$, we can assume that these coefficients still exist and express $C_0$ in terms of two-point functions by writing $p_2^2R_1 -p_1p_2R_2 = 0$ and $-p_1p_2R_1 + p_1^2R_2 = 0$ and solving these (equivalent) equations for $C_0$.
The main idea of Ref.\ \cite{Ganesh} is to repeat this procedure for every tensor rank successively. We can write down the expressions determining the tensor coefficients of second rank, i.e.\ $C_{00}$, $C_{11}$, $C_{12}$ and $C_{22}$. The ultraviolet divergent coefficient $C_{00}$ is not directly\footnote{The coefficient $C_{00}$ is indirectly plagued by problems in the limit of a vanishing Gram determinant, as it is composed of the problematic coefficients $C_1$ and $C_2$.} affected by problems of vanishing Gram determinants and found to be
\begin{equation}
C_{00} = \frac{m_0^2C_0}{D-2} + \frac{B_0(1,2) + f_1C_1 + f_2C_2}{2(D-2)}.
\end{equation}
\\
\noindent
In contrast, the remaining tensor coefficients can not be obtained via standard tensor reduction for vanishing Gram determinant. Instead of that, the corresponding equations can be used to determine the tensor equations of rank one, i.e.\ $C_1$ and $C_2$. The result can be written in compact form as
\begin{equation}
\begin{pmatrix} C_1 \\ C_2 \end{pmatrix} = Z_i^{-1}\begin{pmatrix} R_{3,i} \\ R_{4,i}\end{pmatrix}\quad\mathrm{with}\quad i=1,2.\label{C1C2Alternative}
\end{equation}
\\
\noindent
The abbreviations used here are
\begin{widetext}
\begin{eqnarray}
R_{3,i} & = & x_{i1}\left(B_1(1,2) + B_0(1,2) -\frac{2m_0^2}{D-2}C_0 - \frac{1}{D-2}B_0(1,2)\right) + x_{i2}\left(B_1(0,1) + B_1(1,2) + B_0(1,2)\right),\\
R_{4,i} & = & x_{i1}\left(B_1(0,2) - B_1(1,2)\right) + x_{i2}\left(-B_1(1,2) -\frac{2m_0^2}{D-2}C_0 - \frac{1}{D-2}B_0(1,2)\right),\\
Z_i & = &\begin{pmatrix} Y_i + \frac{x_{i1}}{D-2}f_1 & \frac{x_{i1}}{D-2}f_2 \\ \frac{x_{i2}}{D-2}f_1 & Y_i + \frac{x_{i2}}{D-2}f_2 \end{pmatrix}\quad\mathrm{with}\quad Y_i = x_{i1}f_1 + x_{i2}f_2\quad\mathrm{and}\quad \begin{pmatrix} p_2^2 & -p_1p_2\\ -p_1p_2 & p_1^2\end{pmatrix} = \begin{pmatrix} x_{11} & x_{12}\\ x_{21} & x_{22}\end{pmatrix}.
\end{eqnarray}
\end{widetext}
To summarize, the presented method allows to determine the tensor coefficients of rank $n$ by investigating the equations for tensor coefficients of rank $n+1$ in the limit of vanishing Gram determinant. This works in an algorithmic manner. In comparison to the standard tensor reduction method, the expressions are more lengthy. However, note that the algebraic form of Eqs.\ (\ref{C1-C2}) and (\ref{C1C2Alternative}) is the same. One might ask what happens when $\det(Z_i)$ vanishes. This is of interest, as we precisely run into this situation in the course of our direct detection calculations when evaluating e.g.\ the three-point function $C_0(p,p,m_0^2,m_1^2,m_2^2)$.
\begin{figure*}
\centering
\includegraphics[width=0.49\textwidth]{C1-C2.pdf}
\includegraphics[width=0.49\textwidth]{Ca.pdf}
\includegraphics[width=0.49\textwidth]{C00.pdf}
\includegraphics[width=0.49\textwidth]{Caa.pdf}
\caption{Numerical stability of the three-point tensor coefficients in the limit $v\rightarrow 0$ or equivalently $t\rightarrow 0$.}
\label{fig:AltReduction}
\end{figure*}
There are basically three ways to proceed. First, note that there are actually two sets of equations hiding behind Eq.\ (\ref{C1C2Alternative}). In some lucky cases it might happen that only one variant fails while the other is still working. The second possibility is to apply l'H\^opital's rule without encountering $\det(Z_i)$ again. This is an improvement in comparison to the standard tensor reduction method, where $\det(A)$ usually reappears when taking the limit.
We took a closer look at the problematic cases involved in our calculation and found a third way out of this dilemma. We illustrate this by referring to the three-point function $C_0(p,p,m_0^2,m_1^2,m_2^2)$. Remember that $p_1 = p_2 = p$ is a stronger condition than just $p_1^2 = p_2^2 = p^2$ which happens frequently for identical external particles. The crucial observation is that the tensor coefficients $C_1$ and $C_2$ are no longer uniquely defined in this situation; only their sum $C_a = C_1 + C_2$ is. Instead of Eq.\ (\ref{StandardDecomposition}), we get
\begin{equation}
C_\mu = p_{1,\mu}C_1 + p_{2,\mu}C_2 \rightarrow p_\mu(C_1 + C_2) = p_\mu C_a.
\end{equation}
It is precisely this combination which remains in all the amplitudes in the limit $p_1\rightarrow p_2$. Hence we replace this sum by $C_a$. This coefficient can be easily obtained in the usual way and reads
\begin{equation}
C_a = \frac{1}{2p^2}\left(B_0(0,2) - B_0(1,2) - f_1C_0\right) \label{Ca}.
\end{equation}
By taking into account that tensor coefficients may coalesce under certain kinematical circumstances and applying the method of Ref.\ \cite{Ganesh} as illustrated above, we were able to stabilize the tensor reduction method for vanishing Gram determinant for all loop amplitudes occuring in our direct detection analysis. This is particularly true for the four-point functions needed for the box contributions. Although the basic idea remains unchanged, the corresponding expressions become very large and were therefore calculated with the help of \texttt{Mathematica}.
All tensor coefficients obtained in this way have been tested extensively. We have numerically compared them with the corresponding coefficients resulting from the standard tensor reduction method for small, but non-vanishing Gram determinant. Some examples are shown in Fig.\ \ref{fig:AltReduction}.
The upper left plot of Fig.\ \ref{fig:AltReduction} shows the numerical stability of the tensor coefficients $C_1$ and $C_2$ in the limit of equal momenta or equivalently $(p_b - p_2)^2 = t \rightarrow 0$. More precisely, we show the real parts of $C_1$ (in red) and $C_2$ (in blue) obtained by the regular tensor reduction method divided by $C_a/2$ obtained via Eq.\ (\ref{Ca}) and subtracted by one. In this representation, the black null line corresponds directly to $C_a/2$. We observe, as expected, that both $C_1$ and $C_2$ are relatively stable for $t \leq -0.5$ GeV$^2$ and marginally differ from $C_a/2$. The regular tensor reduction is still working here and the small, but finite velocity leads to a small shift relative to the black reference line. However, when we approach the limit $t\rightarrow0$ the regular tensor method fails and both of the coefficients become numerically unstable.
As explained before, the individual tensor coefficients $C_1$ and $C_2$ are no longer uniquely defined in this limit, only their sum is. The real part of this sum divided by $C_a$ and subtracted by one is shown in the upper right plot of Fig.\ \ref{fig:AltReduction}. It is more stable than the individual coefficients, but still becomes noisy at very small relative velocities. For larger (but still small) relative velocities, the agreement between $C_1 + C_2$ and $C_a$ is excellent, which justifies our approach.
We show analogous plots for the tensor coefficient $C_{00}$ and the combination $C_{11} + 2C_{12} + C_{22}$ in the lower part of Fig.\ \ref{fig:AltReduction}. The main features are similar to the ones discussed before. Using the original tensor reduction method, the tensor coefficients become numerically unstable at very small relative velocities. When using the alternative approach described in this section, we obtain a stable result for $v = 0$ which is in perfect agreement with the standard method for small, but non-zero relative velocities. The black reference line in the lower right plot of Fig.~\ref{fig:AltReduction} is defined by
\begin{eqnarray}
C_b & = &\frac{1}{3p^2}\big(B_0(1,2) - m_0^2C_0 + 2B_1(0,1)\nonumber\\
&& - 2f_2C_a - \frac{1}{2}\big)
\end{eqnarray}
Although of minor importance for the tensor reduction itself, we list all the masses used in the plots above for completeness. They are $m_b = 2.3$ GeV, $m_t = 148.0$ GeV, $m_{\tilde{g}} = 1170.7$ GeV, $m_{\tilde{b}_1} = 1007.3$ GeV, $m_{\tilde{b}_2} = 1071.9$ GeV, $m_{\tilde{t}_1} = 827.9$ GeV and finally $m_{\tilde{t}_2} = 1042.6$ GeV. Note that we have $p_2^2 = p_b^2 = m_b^2$ in the first three plots, whereas $p_2^2 = p_b^2 = m_t^2$ in the lower right plot. More details on the presented alternative tensor reduction method can be found in Ref.\ \cite{myPhD}.
\section{Conclusion}
\label{Conclusion}
In this paper, we presented a NLO SUSY-QCD calculation for the scattering of neutralino dark matter off of the partonic constituents of nucleons, which required a novel tensor reduction method of loop integrals with vanishing relative velocities and Gram determinants. We consistently matched these one-loop corrections to the scalar and axial-vector operators, which govern the spin-independent and spin-dependent scattering proceses in the effective field theory approach. As a result, the operators and Wilson coefficients aquired a scale dependence, which was taken into account by applying renormalization group running to the Wilson coefficients. Our formalism is valid for general compositions of bino, wino, or higgsino dark matter.
We investigated three benchmark scenarios, which satisfy current Higgs mass, relic density, flavor-changing neutral current and direct SUSY particle search constraints from the LHC, but which were not tuned to be particularly sensitive to the new NLO corrections for direct detection. Despite the fact that the first- and second generation squark masses were at the TeV scale, we observed corrections that were of similar size or in some cases larger than the currently estimated nuclear uncertainties. This could be explained by small neutralino-squark mass differences governing the propagator denominators at low velocity. In general, large corrections can be expected in the spin-independent case for Higgs bosons coupling to winos and heavy quarks, in the spin-dependent case for $Z$-bosons coupling to higgsinos and light (potentially also heavy) quark flavors, and in both cases from squarks with small masses or mass differences or scenarios with destructive interference at tree level. In the first case, our calculation is complementary to the explicit generation of heavy quarks from gluon operators at one loop, similarly to the complementarity of variable and fixed flavor schemes that are both employed in deep-inelastic scattering. The calculation for gluon operators has been performed previously elsewhere; its implementation in DM@NLO and a comparison of the two approaches is left for future work, as is a numerical study for light or nearly neutralino mass-degenerate squarks.
Through the implementation of direct detection at NLO as a second dark matter observable in DM@NLO, consistent investigations of correlations between direct detection and the relic density at NLO are now possible. First examples have been given in this paper in the three mentioned reference scenarios. Systematically, shifts in the extracted dark matter mass from NLO corrections to the relic density implied different NLO corrections to be expected in direct detection experiments.
\section{Introduction}
\label{Intro}
Nowadays, the existence of dark matter is well established by experimental observations on many different length scales. In particular, on cosmological length scales, measurements of the temperature anisotropies of the Cosmic Microwave Background (CMB) allow a very precise determination of the relic density of dark matter. The most recent value obtained by the Planck collaboration \cite{Planck}, including polarization data from the Wilkinson Microwave Anisotropy Probe \cite{WMAP9}, is
\begin{equation}
\Omega_{\mathrm{CDM}}h^2 = 0.1199 \pm 0.0022,
\label{Planck}
\end{equation}
where $h$ denotes the present Hubble expansion rate in units of 100 km s$^{-1}$ Mpc$^{-1}$. Even though the quantity of dark matter in the Universe is known very accurately, its nature remains concealed. The reason for this unfortunate situation is that so far all experimental evidence for dark matter stems exclusively from its gravitational interaction.
Among the numerous attempts to explain dark matter, postulating the existence of a yet unknown Weakly Interacting Massive Particle (WIMP) is a widely adopted paradigm. This approach is attractive because a WIMP with typical weak scale interactions and a mass of $\sim$ 100 GeV naturally leads to the observed relic density via thermal freeze-out \cite{Klasen:2015uma}. The canonical example for a WIMP is the lightest neutralino $\tilde{\chi}^0_1$, which is the lightest supersymmetric particle in many scenarios of the Minimal Supersymmetric Standard Model (MSSM). In the following, we refer to it simply by ``the neutralino''. Remember that giving rise to a suitable dark matter candidate is only a positive byproduct of introducing Supersymmetry (SUSY) as the most general space-time symmetry, which is furthermore motivated by its elegant solution to the hierarchy problem and the possible unification of gauge and Yukawa couplings. Alternatively, more minimal extensions to the SM with additional Higgs doublets \cite{Barbieri:2006dq}, neutrinos \cite{Ma:2006km} or other scalars and fermions \cite{Esch:2013rta} may be considered.
Assuming that dark matter actually consists of WIMPs, additional non-gravitational detection possibilities open up. First, one can try to directly produce WIMPs at a collider. As the WIMPs themselves are not detectable with current collider detectors, the typical observable of such a process consists of a single jet or gauge boson and missing transverse energy. The second possibility is to look for Standard Model annihilation products of WIMPs in very dense astronomical objects such as the Sun or the center of the Galaxy, where the WIMPs might have accumulated. The observational challenge of this indirect detection approach is to distinguish between the astrophysical background and a possible WIMP signal. Finally, one can try to observe the rare interactions of a WIMP with a nucleus by detecting its recoil in the so-called direct detection experiments. The technical difficulty here is to detect a very weak signal, while simultaneously excluding all non-dark matter sources \cite{Klasen:2015uma}.
The direct detection rate, i.e.\ the number of events per time and per detector mass, depends on the dark matter-nucleus interaction. On the microscopic level, this corresponds to the interaction of the WIMP with the quarks and gluons inside the nucleons of the nucleus. However, as the typical process energies are much smaller than the mediator masses\footnote{As this condition is not necessarily fulfilled at a collider, EFT methods are under debate in this context, and so-called simplified models should be used \cite{Buchmueller, Busoni, DeSimone, Matsumoto}.} of the microscopic theory, it is customary to calculate the corresponding cross sections in the framework of effective field theories (EFT) \cite{Hill1, Hill2, HisanoEFT, JijiEFT}. In the EFT approach, the heavy particles which mediate the interaction between dark matter and the constituents of the nucleus are integrated out.
Integrating-out heavy particles translates different Lorentz structures of the microscopic theory to different effective contact interactions expressed in terms of effective operators. Not all of the effective interactions contribute in the non-relativistic limit which is relevant for direct detection. In the MSSM, the dominant effective operators for neutralino dark matter are the scalar operator $m_q\bar{\chi}\chi\bar{q}q$ and the axial-vector operator $\bar{\chi}\gamma_\mu\gamma_5\chi\bar{q}\gamma^\mu\gamma_5q$, as the vector and tensor operators vanish in the case of a Majorana fermion. These operators lead to coherent spin-independent (SI) and spin-dependent (SD) contributions, respectively.
The tree-level contributions of neutralino dark matter to these operators have first been calculated in Ref.\ \cite{Griest}. Since then, several improvements have been made by either including additional operators like e.g.\ gluon operators \cite{Drees, HisanoGluon} or by calculating electroweak radiative corrections for pure wino, higgsino or bino dark matter \cite{Hisanoelw, Berlin1, Berlin2}.
In this paper, we perform a full $\mathcal{O}(\alpha_s)$ calculation for the two dominant operators listed above. In contrast to previous approaches, we allow for a general neutralino admixture and calculate the radiative corrections using fully general loop integrals. By doing so, we implement a second, loop-improved dark matter observable in our numerical package {\tt DM@NLO}, the first one being the relic density \cite{DMNLOPage, AFunnel, ChiChi2qq1, ChiChi2qq2, ChiChi2qq3, NeuQ2qx1, NeuQ2qx2, QQ2xx, Scalepaper}. Combining these calculations allows to effectively constrain the MSSM parameter space and precisely predict the direct detection rate.
The remainder of this paper is organized as follows: In Sec.\ \ref{Technical}, we briefly remind the reader how the direct detection rate is computed in general, and we describe our renormalization scheme. We present the underlying EFT calculation, specify the matching of full and effective theory, and describe the running of the operators and their associated Wilson coefficients. In order to use the same tensor reduction method for our direct detection and relic density calculation, the tensor reduction method had to be modified to account for vanishing Gram determinants. As this technical aspect might be interesting on its own, we illustrate it separately in App.\ \ref{GramDet}. Our numerical results are then given in Sec.\ \ref{Numerics}. We analyze the impact of the radiative corrections and contrast them with the nuclear uncertainties. We also study the influence of the neutralino composition on the resulting neutralino-nucleus cross sections. Furthermore, we combine our direct detection and relic density routines to obtain precise predictions for the neutralino-nucleon cross section in a given scenario. Finally, we conclude in Sec.\ \ref{Conclusion}. We do not present any technical details of our relic density calculations here, but instead refer the reader to our previous papers and in particular Ref.\ \cite{ChiChi2qq3}.
\section{Numerical results}
\label{Numerics}
\begin{table*}
\caption{pMSSM input parameters for three selected reference scenarios. All parameters except $\tan\beta$ are given in GeV. }
\begin{tabular}{|c|ccccccccccc|}
\hline
$\quad$ & $\quad\tan\beta\quad$ & $\quad\mu\quad$ & $\quad m_A\quad$ & $\quad M_1\quad$ & $\quad M_2\quad$ & $\quad M_3\quad$ & $\quad M_{\tilde{q}_{1,2}}\quad$ & $\quad M_{\tilde{q}_3}\quad$ & $\quad M_{\tilde{u}_3}\quad$ & $\quad M_{\tilde{\ell}}\quad$& $\quad A_t\quad$ \\
\hline
A & 13.4 & 1286.3 & 1592.9 & 731.0 & 766.0 & 1906.3 & 3252.6 & 1634.3 & 1054.4 & 3589.6 & -2792.3\\
B & 13.7 & 493.0 & 500.8 & 270.0 & 1123.4 & 1020.3 & 479.9 & 1535.5 & 836.7 & 3469.4 & -2070.9\\
C & 7.0 & 815.0 & 1452.8 & 675.3 & 1423.4 & 1020.3 & 809.9 & 1835.5 & 1436.7 & 3469.4 & -2670.9\\
\hline
\end{tabular}
\label{ScenarioList}
\end{table*}
\begin{table*}
\caption{Gaugino and squark masses and other selected observables corresponding to the reference scenarios of Tab.\ \ref{ScenarioList}. All masses are given in GeV.}
\begin{tabular}{|c|cc|cc|ccccc|ccc|}
\hline
$\quad$ & ~~ $m_{\tilde{\chi}^0_1}$~~ & ~~$m_{\tilde{\chi}^0_2}$~~ & ~~$m_{\tilde{\chi}^{\pm}_1}$~~ & ~~$m_{\tilde{\chi}^{\pm}_2}$~~ & ~~$m_{\tilde{u}_1}$~~ & ~~$m_{\tilde{d}_1}$~~ & ~~$m_{\tilde{t}_1}$~~ & ~~$m_{\tilde{b}_1}$~~ & ~~$m_{\tilde{g}}$~~ & ~~$m_{h^0}$~~ & ~~$\Omega_{\tilde{\chi}^0_1} h^2$~~ & $\mathrm{BR}(b\rightarrow s\gamma)$ \\
\hline
A & 738.1 & 802.4 & 802.3 & 1295.1 & 3270.9 & 3271.6 & 993.9 & 1622.9 & 2049.9 & 126.3 & 0.1244 & $3.0\cdot 10^{-4}$ \\
B & 265.7 & 498.4 & 495.7 & 1135.3 & 549.5 & 555.7 & 802.9 & 1531.0 & 1061.2 & 124.8 & 0.1199 & $3.6\cdot 10^{-4}$\\
C & 669.2 & 826.6 & 819.6 & 1438.9 & 865.0 & 868.4 & 1389.1 & 1832.3 & 1090.7 & 125.2 & 0.1179 & $3.3\cdot 10^{-4}$\\
\hline
\end{tabular}
\label{ScenarioProps}
\end{table*}
\begin{table}
\caption{Most relevant (co)annihilation channels in the reference scenarios of Tab.\ \ref{ScenarioList}. Channels which contribute less than 1\% to the thermally averaged cross section and/or are not implemented in our code are not shown.}
\begin{tabular}{|rl|cccc|}
\hline
& & ~~~~ A ~~~~ & ~~~~ B ~~~~ & ~~~~ C ~~~~ & \\
\hline
$\tilde{\chi}^0_1 \tilde{\chi}^0_1 \to$ & $t\bar{t}$ & 1\% & 10\% & 52\% & \\
& $b\bar{b}$ & 9\% & 78\% & 40\% & \\
$\tilde{\chi}^0_1 \tilde{\chi}^0_2 \to$ & $t\bar{t}$ & 3\% & & & \\
& $b\bar{b}$ & 23\% & & & \\
$\tilde{\chi}^0_1 \tilde{\chi}^{\pm}_1 \to$ & $t\bar{b}$ & 43\% & & & \\
\hline
\multicolumn{2}{|c|}{Total} & 79\% & 88\% & 92\% & \\
\hline
\end{tabular}
\label{ScenarioChannels}
\end{table}
In this section we describe our numerical setup and present numerical results for three selected reference scenarios. These scenarios are defined in a phenomenological MSSM (pMSSM) with eleven free parameters, which we have already used in our previous analyses. This setup was designed for relic density calculations including light stops \cite{NeuQ2qx1, NeuQ2qx2, QQ2xx}. As it has proven sufficient for finding interesting direct detection scenarios, we stick to it for consistency and keep in mind, that a more specific pMSSM setup may lead to considerably larger loop contributions.
The aforementioned eleven free parameters are as follows: The Higgs sector is fixed by the higgsino mass parameter $\mu$, the ratio of the vacuum expectation values of the two Higgs doublets $\tan\beta$, and the pole mass of the pseudoscalar Higgs boson $m_A$. The gaugino sector is defined by the bino ($M_1$), wino ($M_2$) and gluino ($M_3$) mass parameters, which in our setup are not related through any assumptions stemming from Grand Unified Theories. Moreover we define a common soft SUSY-breaking mass parameter $M_{\tilde{q}_{1,2}}$ for the first- and second-generation squarks. The third-generation squark masses are controlled by the parameter $M_{\tilde{q}_3}$ associated with sbottoms and left-handed stops and by the parameter $M_{\tilde{u}_3}$ for right-handed stops. The trilinear coupling in the stop sector is given by $A_t$, while the trilinear couplings of the other sectors, including $A_b$, are set to zero. Since the slepton sector is not at the center of our attention, it is parametrized by a single soft parameter $M_{\tilde{\ell}}$. The most interesting parameters for the following discussion are those determining the neutralino decomposition ($\mu$, $M_1$ and $M_2$) and $M_{\tilde{q}_{1,2}}$.
These eleven pMSSM input parameters are defined in the {$\overline{\mathrm{DR}}$}\ scheme at the scale $\tilde{M} = 1$ TeV according to the SPA convention \cite{Spa}. We identify this scale with our renormalization scale $\mu_{R}$, which simultaneously corresponds to the high scale $\mu_{\mathrm{high}}$ of our EFT calculation. The input parameters are handed over to the numerical package {\tt SPheno}\ \cite{SPheno} to calculate the associated physical spectrum.
We neglect the masses of the quarks of the first two generations in the kinematics to improve numerical stability. On the other hand, we keep those masses in the Yukawa couplings to allow for Higgs exchange processes. Remember that the Yukawa masses are basically factored out of the amplitudes and replaced by the nuclear matrix elements via Eq.\ (\ref{fTDef}). It has been checked explicitly that the effect of this simplification on the final results is negligible.
Our three reference scenarios are listed in Tab.\ \ref{ScenarioList}. Table \ref{ScenarioProps} contains the corresponding relevant gaugino and squark masses\footnote{We are not showing the squark masses $m_{\tilde{u}_2}$, $m_{\tilde{d}_2}$, $m_{\tilde{c}_1}$, $m_{\tilde{c}_2}$, $m_{\tilde{s}_1}$ and $m_{\tilde{s}_2}$. However, as we are working with a common soft mass parameter $M_{\tilde{q}_{1,2}}$, all squark masses of the first two generations are roughly the same.} as well as the obtained mass of the lightest neutral (and thus SM-like) Higgs boson, the neutralino relic density computed at tree level with {\tt micrOMEGAs}\ and the important branching ratio of the rare $B$-meson decay $b\to s\gamma$ computed with {\tt SPheno}. Moreover, Tab.\ \ref{ScenarioChannels} lists the most relevant (co)annihilation channels for determining the relic density. Other important parameters are the neutralino mixing angles, i.e.\ its bino, wino and higgsino admixture. As the phenomenology of the three reference scenarios is to a large extent driven by these parameters, we explore them in more detail in the following. We devote an individual subsection to each scenario.
\subsection*{Scenario A -- Bino-wino dark matter}
We start by investigating scenario A. This scenario has been introduced in Ref.\ \cite{ChiChi2qq3} and studied again in Ref.\ \cite{Scalepaper}. Its main feature are sizeable gaugino coannihilation contributions to the relic density calculation as listed in Tab.\ \ref{ScenarioChannels}. The direct detection in this scenario is in no way special and we include this scenario as an arbitrary conservative case.
The decomposition of the neutralino in dependence of the pMSSM input parameter $M_1$ is shown in Fig.\ \ref{fig:MixingScenA}. As long as $M_1<M_2$, the neutralino is mostly bino. It turns into mostly wino when $M_1 > M_2 = 766$ GeV while the higgsino content always stays small due to $M_1, M_2 < \mu$. Note that scenario A itself, i.e the cosmologically preferred region, sits near the turnover ($M_1 = $ 731 GeV). This situation is encountered in many pMSSM scenarios and clearly calls for a general treatment of the neutralino admixture. We also show the associated neutralino mass on the top of each plot as a derived parameter. This connects our theoretical predictions to experimental exclusion limits, which are usually given in dependence of the WIMP mass. Note that the correspondance between $M_1$ and $m_{\tilde{\chi}^0_1}$ is basically 1:1 for $M_1$ up to 800 GeV, but for larger values of $M_1$ the neutralino becomes mostly wino, so that its mass is almost independent of $M_1$.
\begin{figure}
\includegraphics[width=0.49\textwidth]{Mixing1.pdf}
\caption{Neutralino decomposition in scenario A.}
\label{fig:MixingScenA}
\end{figure}
\begin{figure*}
\includegraphics[width=0.49\textwidth]{SI-p1.pdf}
\includegraphics[width=0.49\textwidth]{SI-n1.pdf}
\includegraphics[width=0.49\textwidth]{SD-p1.pdf}
\includegraphics[width=0.49\textwidth]{SD-n1.pdf}
\caption{Spin-independent (top) and spin-dependent (bottom) neutralino-nucleon cross sections in scenario A for protons (left) and neutrons (right).}
\label{fig:CrossSectionsScenA}
\end{figure*}
We continue with the discussion of the neutralino-nucleon cross sections, which are displayed in Fig.\ \ref{fig:CrossSectionsScenA}. The upper left plot of Fig.\ \ref{fig:CrossSectionsScenA} illustrates the spin-independent neutralino-proton cross section. This quantity has been calculated by {\tt micrOMEGAs}\ (orange solid line), our code at tree level (black solid line) and our code including full $\mathcal{O}(\alpha_s)$ corrections to the dominant effective operators (blue solid line). The shift between our tree-level calculation and {\tt micrOMEGAs}\ is mainly due to different nuclear input values (cf.\ Tab.\ \ref{fTTable}). After adjusting the nuclear input, our tree-level calculation agrees quite well with {\tt micrOMEGAs}, which is shown by the dotted black line. In absolute numbers, as expected, the neutralino-proton cross section is rather small (10$^{-47}$ - 10$^{-46}$ cm$^{2}$), as long as the neutralino is mostly bino. The tree-level couplings to Higgs bosons are supressed in this case, and so are the squark processes because of the heavy squark masses (cf.\ Tab.\ \ref{ScenarioProps}). The shift between our tree-level and our full NLO calculation is of similar size as the shift between our tree level and {\tt micrOMEGAs}. In the present case, the first shift is mainly caused by SUSY-QCD corrections to the Higgs exchange process including third generation squarks as the other squarks are much heavier.
Furthermore we show the improved\footnote{More precisely, the green dotted line corresponds to the choice \texttt{MSSMDDTest(loop=1, ...)}, whereas the orange solid line corresponds to \texttt{MSSMDDTest(loop=0, ...)}.} tree-level calculation of {\tt micrOMEGAs}\ as the green dotted line. Among other improvements, this choice is supposed to replace the heavy quark contributions by the gluon one-loop processes as given in Ref.\ \cite{Drees}. However, we could not find a significant difference in comparison to the pure tree-level calculation in any scenario. Therefore, the green dotted and orange full lines are indistinguishable also in this plot.
We also show the resulting relic density obtained with {\tt micrOMEGAs}\ as the dashed orange line (right ordinate). Note that this curve is roughly inverse to the cross section curves. This correlation is not completely unexpected. Larger gaugino (co)annihilation cross sections into final quark states leading to a smaller relic density are linked to larger neutralino-nucleon cross sections. The crucial condition for this correlation is that the neutralinos annihilate dominantly into quark final states. In the present case this is given for $M_1 > 200$ GeV. For smaller $M_1$, neutralinos prefer to annihilate into electroweak final states, and the resulting bump in the relic density has no counterpart in the neutralino-nucleon cross section. The orange vertical band marks the region of $M_1$ leading to a relic density compatible with the Planck limits as given in Eq.\ (\ref{Planck}). We investigate this region in greater detail later.
The upper right plot of Fig.\ \ref{fig:CrossSectionsScenA} shows the spin-independent neutralino-neutron cross section. No major difference in comparison to the proton case is found in this scenario, since the isospin-dependent contributions from first-generation quarks are suppressed by large squark masses.
We continue with the lower left plot of Fig.\ \ref{fig:CrossSectionsScenA} where the spin-dependent neutralino-proton cross section is given. Here the blue and black solid lines completely overlap, signalizing that the NLO corrections are negligible. This is indeed the case in this scenario. Remember that only light quarks ($u,d,s$) and corresponding squarks contribute to the spin-dependent cross section (cf.\ subsection \ref{DDFormulas}). These squarks are very heavy in this scenario (cf.\ Tab.\ \ref{ScenarioProps}) and loops including them are strongly suppressed. The small shift ($\sim + 7\%$) between our results and {\tt micrOMEGAs}\ is not due to the nuclear input values this time -- by default we are using the same input in the spin-dependent case. It is rather due to the running of the operator and associated Wilson coefficient described in subsection \ref{RunningSection}, which is not implemented in {\tt micrOMEGAs}. If we deactivate the running in our code, we find perfect agreement with {\tt micrOMEGAs}. The spin-dependent neutralino-neutron cross section is shown in the lower right plot of Fig.\ \ref{fig:CrossSectionsScenA}. As before, no major difference in comparison to the proton case is found in this scenario.
\begin{figure}
\includegraphics[width=0.49\textwidth]{SI-p_Zoom1.pdf}
\caption{Combined relic density and direct detection calculation in scnenario A.}
\label{fig:ZoomScenA}
\end{figure}
We take a closer look at the cosmologically preferred region now, i.e.\ we zoom into the region 700 GeV $< M_1 <$ 800 GeV of the upper left plot of Fig.\ \ref{fig:CrossSectionsScenA}. The result is shown in Fig.\ \ref{fig:ZoomScenA}. Apart from the previously introduced three solid lines, we depict the relic density obtained with {\tt micrOMEGAs}\ (orange dashed line), our code at tree level (black dashed line) and our code at NLO (blue dashed line). These three calculations lead to different cosmologically preferred regions as indicated by the orange, black and blue vertical band, respectively. Assuming that the neutralinos solely account for dark matter, we can combine these calculations to constrain the pMSSM parameter space and to precisely predict the resulting neutralino-nucleon cross section. This corresponds to identifying the intersections of the vertical bands and solid lines of the same color. The results are given in Tab.\ \ref{PredictionsScenA} where we also list the relative shifts of the {\tt micrOMEGAs}\ and our full NLO result with respect to our tree-level calculation. The shifts are in opposite directions and of similar size in this case.
\begin{table}
\caption{Resulting $M_1$ and spin-independent neutralino-proton cross section when combining direct detection and relic density routines in scenario A.}
\begin{tabular}{|c|ccc|}
\hline
$\quad$ & $M_1$ [GeV] & $\sigma^{\mathrm{SI}}_p$ [$10^{-46}$cm$^2$]& Shift of $\sigma^{\mathrm{SI}}_p$\\
\hline
{\tt micrOMEGAs}\ & 731 & $1.68$ & $-15\%$ \\
Tree level & 734 & $1.98$ & \\
Full NLO & 733 & $2.26$ & $+14\%$ \\
\hline
\end{tabular}
\label{PredictionsScenA}
\end{table}
\subsection*{Scenario B -- Bino-higgsino dark matter}
\begin{figure}
\includegraphics[width=0.49\textwidth]{Mixing2.pdf}
\caption{Neutralino decomposition in scenario B.}
\label{fig:MixingScenB}
\end{figure}
When varying $M_1$ in scenario B, the neutralino decomposition changes again, this time from mostly bino into mostly higgsino as shown in Fig.\ \ref{fig:MixingScenB}. The turning point is at $M_1 \sim \mu \sim 500$ GeV. The neutralino mass depends only weakly on $M_1$ for larger values of $M_1$. In comparison to the previous scenario, the remaining dependence is larger which is in agreement with the softer admixture transition (compare Figs.\ \ref{fig:MixingScenA} and \ref{fig:MixingScenB}). This decomposition and the relatively light squarks (cf.\ Tab.\ \ref{ScenarioProps}) are the essential phenomenological properties of this scenario.
\begin{figure*}
\includegraphics[width=0.49\textwidth]{SI-p2.pdf}
\includegraphics[width=0.49\textwidth]{SI-n2.pdf}
\includegraphics[width=0.49\textwidth]{SD-p2.pdf}
\includegraphics[width=0.49\textwidth]{SD-n2.pdf}
\caption{Spin-independent (top) and spin-dependent (bottom) neutralino-nucleon cross sections in scenario B for protons (left) and neutrons (right).}
\label{fig:CrossSectionsScenB}
\end{figure*}
The neutralino-nucleon cross sections for scenario B are shown in Fig.\ \ref{fig:CrossSectionsScenB}. The first thing to note is that there are three vertical orange bands now, corresponding to three regions which lead to a relic density compatible with Eq.\ (\ref{Planck}). Apart from scenario B itself ($M_1 = 270$ GeV), there is a second line on the other side of the peak of the dashed orange line and a third one at $M_1 \sim 475$ GeV. The peak is due to a Higgs resonance caused by $2m_{\tilde{\chi}^0_1} \sim m_{H^0}, m_{A^0} \sim 500$ GeV, which heavily increases the neutralino cross section into bottom quarks and in turn heavily reduces the resulting relic density. Bottom quarks are favored over top quarks, as $\tan\beta = 13.7$ is rather large here. The peak does not show up in the neutralino-nucleon cross sections. This is as expected, as the Higgs process has turned from a resonant $s$-channel to a non-resonant $t$-channel. The third vertical band lies precisely in the region where the neutralino admixture changes from bino to higgsino, stressing again the necessity to treat the general neutralino admixture.
The spin-independent nucleon cross sections are shown in the upper plots of Fig.\ \ref{fig:CrossSectionsScenB}. Once again, no major difference is found between the proton and neutron case. The relative shifts between our tree-level calculation (black solid line) and {\tt micrOMEGAs}\ (orange solid line) or our NLO calculation (blue solid line) are roughly as before. No significant change is found when activating the improved tree-level calculation of {\tt micrOMEGAs}\ (green dotted line). The agreement between our tree-level calculation using the nuclear input values of {\tt micrOMEGAs}\ (black dotted line) and the {\tt micrOMEGAs}\ result is slightly worse. The remaining discrepancy is mainly due to the use of effective couplings in {\tt micrOMEGAs}\ and a different treatment of the top quark mass and of the associated stop sector (cf.\ subsection \ref{Renormalization}). Moreover {\tt micrOMEGAs}\ does not kinematically distinguish between the $s$- and the $u$-channels shown in Fig.\ \ref{fig:DDTree}. Although these differences are present in general, the resulting discrepancy depends on the concrete scenario. In this scenario they lead to a small, but visible shift, whereas they do not in the other two scenarios.
New features show up in the spin-dependent case, i.e.\ in the lower plots of Fig.\ \ref{fig:CrossSectionsScenB}. Here, the proton and neutron cross sections differ by almost one order of magnitude in the small $M_1$ regime. Moreover the tree-level and NLO results clearly separate in the proton case for small $M_1$. This large splitting is absent in the neutron case. The reason is as follows: In the small $M_1$ regime, the neutralino is mostly bino (cf.\ Fig.\ \ref{fig:MixingScenA}). Moreover, the squarks of the first two generations are rather light in this scenario (cf.\ Tab.\ \ref{ScenarioProps}). The former leads to a suppression of the usually dominant $Z^0$ processes, while the latter kinematically favors the squark processes. As a result, the squark processes contribute sizeably in the small $M_1$ regime. In contrast to the $Z^0$ processes, these processes strongly depend on the involved quark flavor and are sensitive to different choices of $(\Delta_q)_N$ as given in Eqs.\ (\ref{Delta1}) and (\ref{Delta2}). In the case of the proton, this leads to a partial cancellation of the individually large squark contributions, which is much less pronounced in the neutron case. This explains the difference between the proton and neutron cross sections. The rather large impact of the NLO corrections on the proton cross section has a related origin. As the leading squark contributions cancel in the proton case, the cross section becomes more sensitive to the subleading virtual corrections. Due to the rather light squark masses in this scenario, these virtual corrections are not negligible. For large $M_1$, the $Z^0$ processes dominate and the virtual corrections are less important.
\begin{figure}
\includegraphics[width=0.49\textwidth]{SD-p_Zoom2.pdf}
\caption{Combined relic density and direct detection calculation in scnenario B.}
\label{fig:ZoomScenB}
\end{figure}
We take a closer look at the cosmologically preferred region around the Higgs resonance in the case of the spin-dependent neutralino-proton cross section in Fig.\ \ref{fig:ZoomScenB}. As before, we are showing the resulting relic density obtained with {\tt micrOMEGAs}\ (orange dashed line), our tree-level calculation (black dashed line) and our NLO calculation (blue dashed line). The vertical bands of the respective colors correspond to the $M_1$ regions leading to a relic density compatible with Eq.\ (\ref{Planck}). These bands are very thin here, as the relic density is changing rapidly near the resonance, which allows to effectively constrain the pMSSM parameter space. Subsequently we can read off the predicted cross section. The results are shown in Tab.\ \ref{PredictionsScenB}.
\begin{table}
\caption{Resulting $M_1$ and spin-dependent neutralino-proton cross section when combining direct detection and relic density routines in scenario B.}
\begin{tabular}{|c|ccc|}
\hline
$\quad$ & $M_1$ [GeV] & $\sigma^{\mathrm{SD}}_p$ [$10^{-43}$cm$^2$]& Shift of $\sigma^{\mathrm{SD}}_p$\\
\hline
{\tt micrOMEGAs}\ & 226 & $2.78$ & $+3\%$ \\
Tree level & 228 & $2.70$ & \\
Full NLO & 227 & $1.65$ & $-39\%$ \\
\hline
{\tt micrOMEGAs}\ & 270 & $4.14$ & $+8\%$\\
Tree level & 267 & $3.84$ & \\
Full NLO & 269 & $2.47$ & $-36\%$ \\
\hline
\end{tabular}
\label{PredictionsScenB}
\end{table}
As we are using the same nuclear input as {\tt micrOMEGAs}\ in the spin-dependent case, the shift between our tree-level prediction and {\tt micrOMEGAs}\ is smaller than in scenario A, where we investigated the spin-independent neutralino-proton cross section. Note that the relative position of the vertical bands, i.e.\ the relic density constraint, can influence this shift in both directions. The effect of reading off the cross section at different $M_1$ reduces the shift in the first case ($M_1= 226$ GeV and $M_1 = 228$ GeV) and increases the shift in the second case ($M_1 = 270$ GeV and $M_1 = 267$ GeV), as the order of the bands has changed. The exact opposite occurs when comparing our tree-level and our NLO results. Here both relative shifts are large, reaching almost $-40\%$.
\subsection*{Scenario C -- Higgsino-bino dark matter}
\label{ScenarioC}
\begin{figure}
\includegraphics[width=0.49\textwidth]{Mixing3.pdf}
\caption{Neutralino decomposition in scenario C.}
\label{fig:MixingScenC}
\end{figure}
In scenario C, we vary the higgsino mass parameter $\mu$, which changes the neutralino decomposition from higgs\-ino to bino as shown in Fig.\ \ref{fig:MixingScenC}. The turning point is at $\mu \sim M_1 \sim 675$ GeV. Concerning the neutralino admixture, scenario C can be understood as a mirrored version of scenario B (cf.\ Fig.\ \ref{fig:MixingScenB}).
\begin{figure*}
\includegraphics[width=0.49\textwidth]{SI-p3.pdf}
\includegraphics[width=0.49\textwidth]{SI-n3.pdf}
\includegraphics[width=0.49\textwidth]{SD-p3.pdf}
\includegraphics[width=0.49\textwidth]{SD-n3.pdf}
\caption{Spin-independent (top) and spin-dependent (bottom) neutralino-nucleon cross sections in scenario C for protons (left) and neutrons (right).}
\label{fig:CrossSectionsScenC}
\end{figure*}
The neutralino-nucleon cross sections in scenario C are shown in Fig.\ \ref{fig:CrossSectionsScenC}. As the slope of the relic density (orange dashed line) is smaller than in the previous scenarios, the region compatible with the Planck limits is larger, which leads to a thicker vertical orange band. No essential new features are found in the spin-independent cross sections shown in the upper plots of Fig.\ \ref{fig:CrossSectionsScenC}. In particular, the relative shift between our tree-level calculation (black solid line) and {\tt micrOMEGAs}\ (orange solid line) for a given $M_1$ is roughly as big as the shift between our tree-level and our NLO calculation (blue solid line) amounting to $\sim-16\%$ and $\sim+13\%$, respectively. No significant difference is found between the proton and the neutron case.
In contrast to that, the spin-dependent cross sections shown in the lower plots of Fig.\ \ref{fig:CrossSectionsScenC} obviously depend on the nucleon type. This difference is caused by a similar phenomenon as the one described in the previous subsection. In the large $\mu$ region, the neutralino becomes mostly bino (cf.\ Fig.\ \ref{fig:MixingScenC}), which suppresses the $Z^0$ processes. Although the squarks are not as light as in scenario B here (cf.\ Tab.\ \ref{ScenarioProps}), the squark processes are kinematically favored again. Remember that the squark processes occur in the $s$- and $u$-channel. The denominators of the tree-level processes read $s/u - m^2_{\tilde{q}_i}$ which simplifies to $(m_{\tilde{\chi}^0_1} \pm m_q)^2 - m^2_{\tilde{q}_i}$ in the limit of vanishing relative velocity. Hence it is not the total squark mass, but the neutralino-squark mass difference that matters. This difference decreases with increasing $\mu$. As a result, the squark processes contribute sizeably to the spin-dependent cross sections for large $\mu$. These processes depend on the involved flavor and in turn the chosen nuclear input values as given in Eqs.\ (\ref{Delta1}) and (\ref{Delta2}). In the case of the proton, we encounter a destructive interference of the individual terms, which leads to the drop observed at $\mu\sim 850$ GeV. Here the associated four-fermion coupling changes its sign and the resulting cross section vanishes. A similar situation would be encountered in the neutron case for larger values of $\mu$. However, as this region leads to a too large relic density, we are not investigating this in more detail.
\begin{figure}
\includegraphics[width=0.49\textwidth]{SD-p-Zoom3.pdf}
\caption{Combined relic density and direct detection calculation in scnenario C.}
\label{fig:ZoomScenC}
\end{figure}
Instead, we zoom into the region $700 < \mu < 900$ GeV and analyze the spin-dependent neutralino-proton cross section in Fig.\ \ref{fig:ZoomScenC}. As before, we are showing the resulting relic density obtained with {\tt micrOMEGAs}\ (orange dashed line), our tree-level (black dashed line) and our NLO routines (blue dashed line). These three calculations lead to different regions compatible with the Planck limits, as indicated by the vertical bands of the corresponding colors. The bands are broader than before, as the relic density is increasing less rapidly when changing $\mu$. Note that the blue and orange bands overlap to a large extent, which signals that the effective couplings used in the relic routines of {\tt micrOMEGAs}\ are able to approximate the dominant NLO contributions quite well in this scenario. This may happen, but is not necessarily the case, as studied e.g. in Ref. \cite{ChiChi2qq3}. On the other hand, {\tt micrOMEGAs}\ does not include radiative corrections to the spin-dependent cross section. Hence the orange solid line is closely following the black solid line. As mentioned before, the remaining difference is due to the running of the operator and associated Wilson coefficient.
It is again interesting to combine the relic density and direct detection calculations. Our tree-level and NLO routines lead to different preferred regions along the $\mu$ axis. Simultaneosly the shift between the cross section obtained at tree level and at NLO is very large for a given $\mu$ (more than $-50\%$ near the drop). However, when combining both calculations these effects cancel each other. This is not the case for the comparison of {\tt micrOMEGAs}\ with our tree-level result, where {\tt micrOMEGAs}\ predicts a larger cross section.
The aforementioned regions of $\mu$ and the corresponding cross sections are listed in Tab.\ \ref{PredictionsScenC}. The broader vertical bands result in a range of allowed $\mu$ values and an associated range of cross sections. Note that these ranges exist in principle in every scenario. However, as they are very small in the previous scenarios we omitted them for simplicity. The shifts given in Tab.\ \ref{PredictionsScenC} are exemplary and have been obtained by combining the mean values of the cross sections.
\begin{table}
\caption{Resulting $\mu$ and spin-dependent neutralino-proton cross section when combining direct detection and relic density routines in scenario C.}
\begin{tabular}{|c|ccc|}
\hline
$\quad$ & $\mu$ [GeV] & $\sigma^{\mathrm{SD}}_p$ [10$^{-43}$cm$^2$]& Shift of $\sigma^{\mathrm{SD}}_p$\\
\hline
{\tt micrOMEGAs}\ & 815 - 821 & 1.80 - 2.43 & $+63\%$ \\
Tree level & 823 - 829 & 1.06 - 1.53 & \\
Full NLO & 813 - 819 & 1.08 - 1.62 & $+4\%$ \\
\hline
\end{tabular}
\label{PredictionsScenC}
\end{table}
\begin{figure}
\includegraphics[width=0.49\textwidth]{SD-nDRBar.pdf}
\caption{Spin-dependent neutralino-neutron cross section in scenario C using a pure {$\overline{\mathrm{DR}}$}\ scheme}
\label{fig:DRbar}
\end{figure}
Before concluding, we take a small detour and briefly comment on the renormalization scheme dependence. As we have described in subsection \ref{Renormalization}, we are working with a hybrid on-shell/{$\overline{\mathrm{DR}}$}\ scheme. In particular the squark masses of the first two generations are treated on-shell just like in {\tt micrOMEGAs}. Our code optionally also supports a pure {$\overline{\mathrm{DR}}$}\ scheme. When studying the differences between the two schemes, the spin-dependent neutralino-neutron cross section shown in the lower right plot of Fig.\ \ref{fig:CrossSectionsScenC} has proven very useful. This plot is shown again in Fig.\ \ref{fig:DRbar}, this time using a pure {$\overline{\mathrm{DR}}$}\ scheme. No visible differences between the two plots occur for small $\mu$. The virtual corrections to the spin-dependent cross section are negligible in this regime which is not affected by the choice of the scheme. For larger values of $\mu$, our tree-level result (black solid line) -- now using {$\overline{\mathrm{DR}}$}\ squark masses -- clearly separates from {\tt micrOMEGAs}\ (orange solid line) -- still using on-shell squark masses -- which has previously not been the case (cf.\ Fig.\ \ref{fig:CrossSectionsScenC}).
Remember that the cross section in this region is heavily influenced by the squark processes as explained in the beginning of this subsection. These processes benefit from the decreasing neutralino-squark mass difference appearing in the denominators of the corresponding propagators. This mass difference is sensitive to the choice of the scheme. To investigate this in greater detail, we write the scale-independent on-shell squark mass $m_{\tilde{q}}^{\mathrm{OS}}$ as a sum of two individually scale-dependent terms, the scale-dependent {$\overline{\mathrm{DR}}$}\ mass $m_{\tilde{q}}^{\mathrm{\overline{DR}}}(\mu_R)$ and an additional finite term resumming virtual corrections $\Delta m_{\tilde{q}}(\mu_R)$,
\begin{equation}
m_{\tilde{q}}^{\mathrm{OS}} = m_{\tilde{q}}^{\mathrm{\overline{DR}}}(\mu_R) + \Delta m_{\tilde{q}}(\mu_R).
\end{equation}
If we replace the on-shell squark masses by their smaller {$\overline{\mathrm{DR}}$}\ masses, i.e.\ if we discard the finite $\Delta m_{\tilde{q}}(\mu_R)$ terms, the neutralino-squark mass difference decreases even further. This leads to the observed steep drop of our tree-level result. However, at NLO this effect diminishes again and the blue lines shown in the lower right plot of Fig.\ \ref{fig:CrossSectionsScenC} and in Fig.\ \ref{fig:DRbar} roughly agree. The reason behind that is that the leading corrections incorporated in $\Delta m_{\tilde{q}}(\mu_R)$ reappear again, this time as virtual corrections to the squark propagators. In other words, the tree-level result heavily depends on the definition of the squark mass in this special situation, but the NLO result is much more stable. The main difference between the two schemes is that the virtual corrections are partially included at tree level in the on-shell mass in the first case, whereas they show up as large propagator corrections in the second case. We prefer the first scheme, which leads to smaller virtual corrections and an improved perturbative stability. Let us finally mention that the resulting differences between the two schemes in other cases, i.e.\ for other cross sections, are less pronounced. A similar study in the context of the relic density can be found in Ref.\ \cite{Scalepaper}.
|
\section{Introduction}
The study of collective behavior of connected systems has drawn significant amount of attention in variety of disciplines spanning from theoretical sciences to engineering. As the study found more applications in smart grids, biological systems, etc. more scientists put effort to understand and solve the related problems in such systems \cite{Wiener65, Winfree67, Pecora98, Chen07, Manaffam.ACC13, Yu09, DeLellis11, Manaffam.TCAS13}.
Synchronization in the networked systems as one manifestation of such collective behavior was first introduced by Wiener \cite{Wiener65}. Pursued by Winfree in his pioneering work \cite{Winfree67}, the problem of synchronization of network of identical system was recognized as relevant in many fields of research \cite{Pecora98, Pirani.TAC16, Chen07, Yu09, DeLellis11, Manaffam.TCAS13, Manaffam.ACC13}. Introduction of master stability function framework by Pecora and Carroll \cite{Pecora98}, made it possible to separate the topological impact of the network from the dynamical properties of individual nodes on synchronizability of the networked systems \cite{Pecora98}. This type of master stability function uses Lyapunov exponents as a measure of stability which provides necessary conditions on the stability of the network \cite{Pecora98}. However, most recently, the sufficient conditions the stability of the network has been provided by Lyapunov direct method and assuming that the nonlinear systems satisfy a very general assumption, namely, Quad$-$condition or its equivalents \cite{Manaffam.ACC13, Yu09, DeLellis11}.
As experimental studies have shown, similar to the network of identical systems, the network of semi-similar systems also exhibits certain collective behaviors \cite{Restrepo04, Sun09, Sorrentino11,Acharyya12,Manaffam.IET16}. The semi-similarity in these studies implies identical structure for systems while the parameters of the systems in the network can slightly differ from one another \cite{Restrepo04, Sun09, Sorrentino11, Acharyya12, Manaffam.IET16}. In \cite{Restrepo04}, it is reported that if in the network of semi-similar systems, if the parameters of couplings and isolated systems are slightly different, the states of all the system although cannot be absolutely synchronized, however, they can approach to a close vicinity of each other as the states evolve. The results of this work have been provided mostly on experimental merit. Following \cite{Restrepo04} and similar experimental works, a sensitivity analysis for mismatch systems and concept of $\varepsilon$-synchronization have been given in \cite{Sun09, Sorrentino11, Acharyya12}. In \cite{Sun09}, by assuming the parameter mismatch only in isolated systems, an approximate master stability function for the radius of the neighborhood which the trajectories in the network converges has been calculated. The results in \cite{Sun09} are generalized by \cite{Sorrentino11} by introducing mismatches in the inner coupling as well as weights of the connections. In \cite{Acharyya12}, a new master stability function is given by including higher terms in Taylor series of states around the average trajectories of the network. Additionally, coupling optimization to achieve ``best synchronization properties have been given \cite{Acharyya12}. The results of previous work are generalized in \cite{Manaffam.IET16} for weighted directed systems where it has shown that the center of the neighborhood for the trajectories is the weighted average of trajectories where the weights belong to left null space of the Laplacian matrix of the network. For symmetric networks, this weighted average reduces to simple average as assumed by \cite{Restrepo04, Sun09, Sorrentino11, Acharyya12}. Also probability of $\varepsilon$-stability is used as a measure to study the phase transition of the network from desynchronization to $\varepsilon$-synchronization \cite{Manaffam.IET16}.
As the analysis of the synchronization in networks of the identical systems with no regulations seems to be advanced, the problem of \textit{pinning} has emerged in various fields of researches \cite{ Chen07}\cite{Manaffam.ACC13}\cite{ Yu09}\cite{Pirani.TAC16}. The objective of pinning is to have the network synchronize to a reference trajectory/state, where the reference trajectory is only available in fraction of the locations in the network, where objective of the problem is to locate the systems which would stabilizes the netwok by providing the reference in minimal number of pinned nodes \cite{Porfiri06}. This problem has been studied in many literature such as \cite{ Chen07}\cite{Manaffam.ACC13}\cite{ Yu09}\cite{Pirani.TAC16} \cite{Bapat.Cyb16}\cite{ DeLellis13} and references therein. This method also has been used in cooperative control schemes where the network is spatially distributed and providing the reference trajectory to all the systems is not desirable \cite{DeLellis13, Manaffam.TCST16}.
In this paper, first, we investigate the problem of $\varepsilon-$synchronization in the symmetric network of mismatched oscillators. Using Lyapunov direct method, we find an upper bound on the error of trajectories from the average of trajectories, where the network converges in finite time. The stated conditions on achieving $\varepsilon-$synchronization in finite time are sufficient and it also applies to time varying mismatches. Note that the bounds given in \cite{Sun09, Sorrentino11,Acharyya12,Manaffam.IET16} are asymptotic bounds and only true for constant parameter mismatches not time varying ones. Then, we devise decentralized and distributed mismatch estimators to compensate for the parameter mismatches of the oscillators. It is shown that if in decentralized control method, the reference trajectory is provided for all the systems, the network of mismatched systems can asymptotically converge to the reference. Since, in most applications, the availability of reference model and/or trajectory in all locations is not desirable, we cooperative/distributed control via pinning to synchronize the mismatched network to the reference trajectory. In the distributed scheme, we assume that there is a connected communication/cooperation network between the systems in the network, and the reference trajectory is available in a fraction of the locations, \textit{i. e.,} pinning locations. Finally, we consider a network of Lorenz oscillators with parameter mismatches to numerically verify our analytical results.
\section{Preliminaries}
\subsection{Notations and Background}
The set of real $n$-vectors is denoted by $\mathbb{{R}}^{n}$ and the set of real $m\times n$ matrices is denoted by $\mathbb{{R}}^{m\times n}$. We refer to the set of non-negative real numbers by $\mathbb{{R}}_{+}$. Matrices and vectors are denoted by capital and lower-case bold letters, respectively. Identity matrix is shown by \bI. The Euclidean ($\mathcal{L}_{2}$) vector norm is represented by $\lVert\cdot\rVert $. Symmetric part of matrix, \bA, is denoted as $\bA^{(s)}\triangleq (\bA+\bA^T)/2$.
As it is known, the mismatched network (in general, mismatched systems), without compensators, cannot be absolutely synchronized; hence, the synchronization for these networks reduces to neighborhood synchronization, where the network trajectories will converge to a certain vicinity of each other and continue to stay there \cite{Sun09, Sorrentino11, Acharyya12, Manaffam.IET16}. To analyze this type of synchronization, the objective is to find the center and the radius of that neighborhood. In \cite{ Manaffam.IET16}, it has been shown that this center for \textit{undirected} networks is simple average of all the trajectories. This has also been used in \cite{Sun09, Sorrentino11, Acharyya12}. Consequently, the error of system $i$ from the average trajectory, $\bar{\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}} \triangleq \sum_{i=1}^{N}\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_{i}/N$, yields, $\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i\triangleq\sum_{j=1}^N(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i-\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j ) / N$. Therefore, we can formalize the definition of $\varepsilon$-synchronization as follow.
\begin{definition}[$\varepsilon$-synchronization]
Let $\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$} \triangleq [\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_1^T \, \cdots \, \text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_N^T]^T$ denote the state of the network, where $N$ is the number of systems in the network. Then, the undirected network is $\varepsilon-$synchronized, iff
\begin{equation} \label{def: e-sync}
\frac1N\lim_{ t \to \infty} \|(\bR_N \otimes \bI_n) \, \text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$} \| = \varepsilon,
\end{equation}
where
\begin{equation} \label{eq: Globally Connected Net}
\bR_N\triangleq\left[\begin{array}{ccccc}
N-1 &-1&-1&\cdots&-1\\
-1 & N-1&-1&\cdots&-1\\
\vdots&\ddots&\cdots& &\vdots\\
-1& -1&\cdots &-1 & N-1
\end{array}\right]_{N\times N}.
\end{equation}
\end{definition}
Please note that asymptotic absolute-synchronization is achieved if \cite{Restrepo04, Sun09, Sorrentino11, Acharyya12, Manaffam.IET16}\[\lim_{t\to\infty}\|\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}-\bar{\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}}\| =0.\]
Here are several lemma's, which will be used later in the analysis.
\begin{lemma}\label{lem: commute} Any $N \times N$ symmetric Laplacian, \bP, and $\bR_N$ commute and moreover,
\begin{equation}
\bR_N \bP = \bP \bR_N = N \bP.
\end{equation}
\begin{IEEEproof}
See Appendix \ref{proof: commute}.
\end{IEEEproof}
\end{lemma}
\begin{lemma}\label{Lemma: Vector Product Inequality}
Let \text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}~and \text{$\mathbf{y}$}} \def\bY{\text{$\mathbf{Y}$}~to be any arbitrary vectors and \bK~to be a positive definite matrix and \bP~a matrix of proper dimensions. Then
\begin{align*}
\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}^T\bP\text{$\mathbf{y}$}} \def\bY{\text{$\mathbf{Y}$}+\text{$\mathbf{y}$}} \def\bY{\text{$\mathbf{Y}$}^T\bP^T\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}=2\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}^T\bP\text{$\mathbf{y}$}} \def\bY{\text{$\mathbf{Y}$}&\le\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}^T\bP\bK^{-1}\bP^T\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}+\text{$\mathbf{y}$}} \def\bY{\text{$\mathbf{Y}$}^T\bK\text{$\mathbf{y}$}} \def\bY{\text{$\mathbf{Y}$}.
\end{align*}
\end{lemma}
\begin{lemma}\cite{Prasolov}\label{Lemma: Joint Diagonalization}
If \bM~and \bK~commute, \textit{i. e.}, $\bM\bK=\bK\bM$, then they can be jointly diagonalized by a unitary matrix, \bQ~such that
\begin{align*}
\bM&=\bQ\bJ_M\bQ^T,\\
\bK&=\bQ\bJ_K\bQ^T
\end{align*}
where superscript $T$ denotes Hermitian transpose. The diagonal entries of $\bJ_M$ and $\bJ_K$ are eigenvalues of $\bM$ and \bK, respectively.
\end{lemma}
\begin{lemma}\cite[{Theorem 4. 8}]{Khalil02}\label{Lemma: asymptotic W}
Suppose that $\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$})$ is continuous and satisfies \eqref{eq: boundF} and it is uniform in $t$. Let $V:\mathbb{{R}}^m\to\mathbb{{R}}$ be continuously differentiable function and continuous function $W(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$})$ such that
\begin{align}
k_1\|\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}\|^{c_1}\le V(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$})\le k_2\|\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}\|^{c_2}\\
\dot{V}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$})=\frac{\partial V}{\partial \text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}}\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$})\le-W(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$})\le 0
\end{align}
where $k_i$ and $c_i$ are positive constants. Then all solutions of
\begin{align*}
&\text{$\dot{\mathbf{x}}$}=\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$})\\
&\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}(t_0)=\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}^0,
\end{align*}
are ultimately bounded and
\[\lim_{t\to\infty}W(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$})=0.\]
\end{lemma}
\begin{lemma}\cite[{Theorem 8. 2}]{Khalil02}\label{Lemma: Bounded x}
Suppose that $\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$})$ is continuous and satisfies \eqref{eq: boundF} and it is uniform in $t$. Let $V:\mathbb{{R}}^m\to\mathbb{{R}}$ be continuously differentiable function such that
\begin{align}\begin{array}{ll}
k_1\|\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}\|^{c_1}\le V(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$})\le k_2\|\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}\|^{c_2}& \\
\dot{V}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$})=\frac{\partial V}{\partial \text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}}\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$})\le-k_3\|\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}\|^{c_3} &\forall \|\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}\|\ge r\end{array}
\end{align}
where $k_i$ and $c_i$ are positive constants. Then there exists $t_1>t_0$ such that
\begin{align*}\begin{array}{ll}
\|\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}\|\le k_4 \|\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_0\| \exp(-c_4(t-t_0)),&\forall t_0\le t\le t_1\\
\|\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}\|\le \left(\frac{k_2}{k_1}\right)^{1/c_1}r^{c_2/c_1}&\forall t>t_1.\end{array}
\end{align*}
\end{lemma}
\subsection{Systems Model}
Let the dynamics of networked systems be given as $ \forall i$
\begin{align}\label{eq: NetworkEq1}
&\text{$\dot{\mathbf{x}}$}_i=\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)+\bG(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)\text{$\boldsymbol{\gamma}$}_i+\sum_{j=1}^N a_{ij}\bH(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j-\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)+\text{$\mathbf{u}$}} \def\bU{\text{$\mathbf{U}$}_i\\
&\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i(t_0)=\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i^0. \nonumber
\end{align}
where $\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i\in\Omega$ is the state vector of the system $i$, $\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i^0$ is the initial state of the system $i$,
$\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}:\Omega \to \mathbb{{R}}^n$ describes the dynamics of the nominal system, and $\text{$\mathbf{u}$}} \def\bU{\text{$\mathbf{U}$}_i\in \mathbb{{R}}^n$ is the input vector. $\bG(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i) \text{$\boldsymbol{\gamma}$}_i$ represents the uncertainty in the dynamics of system $i$. More precisely, $\text{$\boldsymbol{\gamma}$}_i\in \text{${\mathcal{P}}$}$ is the uncertainty/mismatch vector corresponding to the system $i$ and it is limited to the set $\text{${\mathcal{P}}$}$ with dimension $|\text{${\mathcal{P}}$}|=m$, where the uncertainties affect the individual systems according to the function $\bG:\mathbb{{R}}^n\to\mathbb{{R}}^{n\times m}$. The adjacency matrix of the network is denoted by $\bA=[a_{ij}]$, where $a_{ij}\in\mathbb{{R}}$ indicates the weight of the connection from node $j$ to node $i$. There is no connection if $a_{ij}=0$. The term $a_{ij} \bH(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i-\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j)$ indicates that the system $i$ is coupled to the system $j$, where $\bH \in \mathbb{{R}}^{n \times n}$~is the inner coupling matrix.
Define the Laplacian/gradient matrix of the network, $\bL=[l_{ij}]$, as
\begin{align}
l_{ij}=\left\{\begin{array}{ll}
-a_{ij}&i\ne j,\\
\sum_{j=1}^Na_{ij}&i=j.
\end{array}\right.\label{eq: Laplacian}
\end{align}
\bL~is a zero-sum-row matrix and it is positive semidefinite. From this point on, we will represent the network by \bL. With this definition, \eqref{eq: NetworkEq1} can be rewritten as
\begin{align}\label{eq: NetworkEq}
&\text{$\dot{\mathbf{x}}$}_i = \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i) + \bG(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i) \text{$\boldsymbol{\gamma}$}_i - \sum_{j=1}^N l_{ij}\bH\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j +\text{$\mathbf{u}$}} \def\bU{\text{$\mathbf{U}$}_i,
\\
&\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i(t_0) = \text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i^0.
\nonumber
\end{align}
In the remainder of the paper, we will assume that the followings hold.
\begin{assumption}\label{Assumption: Connected}
The plant network represented by the Laplacian matrix, \bL,~is connected and undirected.
\end{assumption}
The connectivity of the network implies that the Laplacian matrix in \eqref{eq: Laplacian}, has only one zero eigenvalue \cite{Mohar91}. The network being undirected implies that the Laplacian is symmetric, \textit{i .e.,} $\bL = \bL^T$. Therefore, all its eigenvalues, $\mu_{i}$, are non-negative real numbers. Please note that we do note require the weights of the connection to be binary, hence, we consider the general class of weighted-undirected networks.
\begin{assumption}\label{Assumption: BoundF}
There exists a positive semidefinite matrix \bF~such that following inequality holds
\begin{align}
(\tilde{\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}}-\tilde{\text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$}})^T[\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\tilde{\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}})-\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\tilde{\text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$}})]\le(\tilde{\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}}-\tilde{\text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$}})^T\bF(\tilde{\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}}-\tilde{\text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$}}),\label{eq: boundF}
\end{align}
for all $(\tilde{\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}},\tilde{\text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$}})\in\Omega\times\Omega$.
\end{assumption}
Note that this assumption is not very restrictive: if all the elements of the Jacobian of $\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$})$ with respect to state vector, \text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}, is bounded, there always exists a positive semidefinite matrix \bF~such that assumption \eqref{eq: boundF} holds \cite{Yu09}. As discussed in \cite{DeLellis11}, this assumption is closely related to QUAD$-$condition.
\begin{assumption}[Bounded uncertainties]\label{Assumption: Bounded Mismatch}
If $(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$},\, \text{$\boldsymbol{\gamma}$})\in \Omega \times \text{${\mathcal{P}}$}$, there exists a symmetric positive semidefinite matrix \text{$\boldsymbol{\Gamma}$}~and vector $\text{$\boldsymbol{\gamma}$}_c$ such that following inequality holds
\begin{align}
\text{$\boldsymbol{\gamma}$}^T \bG(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$})^T \bG(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}) \text{$\boldsymbol{\gamma}$} \le \text{$\boldsymbol{\gamma}$}_c^T \text{$\boldsymbol{\Gamma}$} \gamma_c,\quad \forall (\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$} ,\, \text{$\boldsymbol{\gamma}$})\in \Omega \times \text{${\mathcal{P}}$}.
\end{align}
\end{assumption}
This assumption basically states that the uncertainties, $\text{$\boldsymbol{\gamma}$}_i$'s, are bounded.
\subsection{Problems Statements}
As discussed, since there are uncertainties in the network, in general, the network does not achieve absolute synchronization. Thus, based on the described system model, there are two natural questions to be asked:
\textit{1) what are the conditions that the network in \eqref{eq: NetworkEq} should satisfy to achieve $\varepsilon$-synchronization? And if it $\varepsilon$-synchronizes, what is the bound on the norm total error from the average trajectory, $\varepsilon$? 2) How does the network in \eqref{eq: NetworkEq} can be pinned such that it asymptotically converges to a known reference trajectory?}
Next, we formalize these questions as
\begin{problem}
a) If Assumptions \ref{Assumption: Connected}-\ref{Assumption: Bounded Mismatch} hold. Find conditions on $\mathbf{L}$ such that there exists a positive constant, $\varepsilon > 0$, that in finite time, $ t_\varepsilon > t_0$, we have
\begin{align}
\frac1N \|(\mathbf{R}_N \otimes \mathbf{I}_n) \, \mathbf{x}\| \le \varepsilon \quad \forall t > t_\varepsilon > t_0;
\end{align}
b) If $\varepsilon-$synchronization occurs, what is $\varepsilon$?
\end{problem}
\begin{problem}
For a given reference trajectory, \text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$},
\begin{align}\label{eq: reference}
&\text{$\dot{\mathbf{s}}$}=\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$}),\\
&\text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$}(t_0)=\text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$}^0,\nonumber
\end{align}
find a control law, $\text{$\mathbf{u}$}} \def\bU{\text{$\mathbf{U}$}_i$, such that the network in \eqref{eq: NetworkEq} asymptotically converges to \text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$}, that is,
\begin{equation}
\lim_{t \to \infty} \|\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$} - \mathbf{1}_N \otimes \mathbf{s}\| = 0.
\end{equation}
\end{problem}
\section{Analytical Results}
In this section, first we derive the sufficient conditions on bounded stability networked systems with uncertainties. Then, using decentralized control and assumption of constant uncertainties, the mismatched parameters are compensated and the network is driven to the reference trajectory.
\subsection{Boundedness of The Synchronization Error}
In this section, we will show that the error of the network in \eqref{eq: NetworkEq} from its average trajectory, $\bar{\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}} = \sum\limits_{i = 1}^N \text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i/N$, is ultimately bounded. Additionally, we will derive an upper bound on the norm of that error.
\subsubsection{Error development}
let us define the synchronization error as $\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i \triangleq \text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i - \bar{\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}}$. The error dynamics for $\forall i$ can be expressed as
\begin{align}
\text{$\dot{\mathbf{e}}$}_i=&\frac1N(\bR_N \otimes \bI_n) \, \text{$\dot{\mathbf{x}}$} \nonumber\\
=& \frac1N\sum_{j=1}^N\left[\Big(\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)-\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j)\Big)+\Big(\bG(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)\text{$\boldsymbol{\gamma}$}_i - \bG(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j) \text{$\boldsymbol{\gamma}$}_j\Big)\right]\nonumber\\
& - \sum_{j =1}^N l_{ij}\, \bH \,\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j + \frac1N\sum^N_{i,j =1} l_{ij}\, \bH \,\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j \nonumber \\
=& \frac1N\sum_{j=1}^N\left[\Big(\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)-\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j)\Big)+\Big(\bG(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)\text{$\boldsymbol{\gamma}$}_i - \bG(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j) \text{$\boldsymbol{\gamma}$}_j\Big)\right]\nonumber\\
& - \sum_{j =1} l_{ij}\, \bH \,\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j .\label{eq: NetworkError}
\end{align}
Please note that from Assumption \ref{Assumption: Connected} the Laplacian of the network is symmetric, and henceforth zero column-sum, thus, we have $\sum^N_{i,j =1} l_{ij}\, \bH \,\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j = \text{$\mathbf{0}$}$, and the last equality follows.
\subsubsection{Results and Discussion}
\begin{theorem}\label{Theorem: BoundedError}
Let Assumptions \ref{Assumption: Connected}-\ref{Assumption: Bounded Mismatch} hold, if there exists a positive constant, $\lambda$ such that
\begin{align}
\bF-{\mu_i}\bH^{(s)}+\lambda\bI_{n}\prec\text{$\mathbf{0}$},\quad \forall i = 1, \,\cdots,\,N-1\label{eq: ConditionTheorem1}
\end{align}
where $\mu_i$'s are eigenvalues of the Laplacian of the network sorted descendingly, then the error in \eqref{eq: NetworkEq} is ultimately uniformly bounded around $\bar{\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}} = \sum_{i =1}^N \text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i/N$. Furthermore, the synchronization error in \eqref{eq: NetworkError}, $\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$} \triangleq [ \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_1^T \, \cdots \, \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}^T_N ]^T$, is bounded as
\begin{align}\label{eq: Theorem1}
\|\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}\| \le & \quad \sqrt{ {N}{\text{$\boldsymbol{\gamma}$}_c^T\text{$\boldsymbol{\Gamma}$}\text{$\boldsymbol{\gamma}$}_c}/{\lambda^\star}^2},\\
\lambda^{\star}=&~\max \quad\lambda \label{eq: lambda_star}\\
~&~\text{s. t.} \quad \bF-{\mu_i}\bH^{(s)}+\lambda\bI_n\prec\text{$\mathbf{0}$},\quad \forall i = 1, \,\cdots,\,N-1\nonumber
\end{align}
in finite time.
\begin{IEEEproof}
See appendix \ref{Proof: BoundedError}.
\end{IEEEproof}
\end{theorem}
\begin{remark}
The existence of positive $\lambda>0$ that satisfies \eqref{eq: ConditionTheorem1} is a sufficient condition on existence and convergence to the bound in \eqref{eq: Theorem1}. Additionally, it should be noted that according to Lemma \ref{Lemma: Bounded x}, the network reaches the bound \eqref{eq: Theorem1} in finite time.
\end{remark}
\begin{remark}
Since Assumption \ref{Assumption: Bounded Mismatch} does not require the uncertainties to be constant, the results in Theorem \ref{Theorem: BoundedError} hold if the uncertainties are time varying but bounded.
\end{remark}
In contrast to the previous results reported in \cite{Restrepo04, Sun09, Sorrentino11, Acharyya12, Manaffam.IET16}, Theorem \ref{Theorem: BoundedError} guarantees convergence in finite time. Although conservative, the conditions in \eqref{eq: ConditionTheorem1} does not require linearization around average trajectory which leads to calculation of transition matrix of the network error or Lyapunov exponents. The same is true for the bound on the error in \eqref{eq: Theorem1}. Consequently, this theorem renders the stability analysis of the networked systems much simpler and straightforward.
\subsection{Compensation for constant uncertainties}
In this part, first, we employ decentralized control to stabilize the network and compensate for the constant uncertainties in the network.
\subsubsection{Error development}
if the dynamics of the reference trajectory is given by \eqref{eq: reference}, then the error dynamics from the reference, \text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$}, can be expressed as $\forall i = 1, \cdots, N$
\begin{align}\label{eq: Error_Ref}
\text{$\dot{\mathbf{e}}$}_i = \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i) - \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$}) + \bG(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i) \text{$\boldsymbol{\gamma}$}_i - \sum_{j =1}^N l_{ij} \bH\,\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j + \text{$\mathbf{u}$}} \def\bU{\text{$\mathbf{U}$}_i.
\end{align}
\subsubsection{Results and discussion}
In the rest of the paper, we will assume that all the nodes have compensators and all the uncertainties are constant. The results for asymptotic convergence of the synchronization error from the reference signal will be presented in two fold: decentralized and distributed.
\begin{theorem}[Decentralized Compensation]\label{Theorem: MismatchEstimation}
Let Assumptions \ref{Assumption: Connected} and \ref{Assumption: BoundF} hold and the uncertainty vectors, $\text{$\boldsymbol{\gamma}$}_i$'s, be constant. Then, the network in \eqref{eq: NetworkEq} with $\forall i = 1,\,\cdots,\, N$
\begin{align}\label{eq: input}
\text{$\mathbf{u}$}} \def\bU{\text{$\mathbf{U}$}_i&=-z_i\bH(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i-\text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$})-\bG(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)\hat{\text{$\boldsymbol{\gamma}$}}_i,\\
\label{eq: Estimation}
\dot{\hat{\text{$\boldsymbol{\gamma}$}}}_i&=k_i\bG^T(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i-\text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$}),
\end{align}
asymptotically uniformly converges to the reference signal, \text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$}, if there exists positive constants, $k_i > 0$, and a diagonal matrix, $\bZ=\mbox{diag}([z_1,\,\cdots,\,z_N])$, $\forall z_i>0$ , such that
\begin{align}\label{eq: Theorem2}
\bF-\mu_i\bH^{(s)}\prec\text{$\mathbf{0}$} ,\quad i = 1,\, \cdots, \, N
\end{align}
where $\mu_i$'s are eigenvalues of $\bL+\bZ$ and $\bZ \triangleq \text{diag}([z_1 \, \cdots \, z_N]^T)$.\\
\textbf{Proof:} See appendix \ref{Proof: MismatchEstimation}.
\end{theorem}
The first term in \eqref{eq: input} is a common feedback control used in pinning control of identical networked systems\cite{Chen07, Manaffam.TCAS13, Manaffam.ACC13,Yu09}. The second term in conjunction with \eqref{eq: Estimation} estimates and compensates for the parameter mismatches of the systems. The detail of choosing the estimator in \eqref{eq: Estimation}, is given in Appendix \ref{Proof: MismatchEstimation}.
\begin{remark}\label{remark: Conv_Est}
From Theorem \ref{Theorem: MismatchEstimation}, we have
\[ \lim_{t \to \infty} [ \text{$\dot{\mathbf{x}}$}_i - \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)] = \text{$\dot{\mathbf{s}}$} - \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$}) = \text{$\mathbf{0}$},\]
and
\[ \lim_{t \to \infty} \bG (\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i) = \bG(\text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$}).\]
Now, if there exists $T \ge t_0$ such that $\bG(\text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$})$ is not singular for $t > T$, then from \eqref{eq: NetworkEq} we can conclude that \[\lim_{t \to \infty} \hat{\text{$\boldsymbol{\gamma}$}_i} = \text{$\boldsymbol{\gamma}$}_i.\]
\end{remark}
As it is clear from Theorem \ref{Theorem: MismatchEstimation}, all the nodes are required to have the reference trajectory \eqref{eq: reference}, however, in spatially distributed networks to provide the reference trajectory to all the systems is costly and impractical. To alleviate this issue, it is more convenient to have the systems receive and send information to their neighboring systems and provide the reference trajectory to small fraction of the network. The idea is to \textit{cooperatively} estimate the mismatches. Next theorem is devised to address this drawback of Theorem \ref{Theorem: MismatchEstimation}.
\begin{theorem}[Distributed Compensation]\label{Theorem: Distributed Estimation}
Let Assumptions \ref{Assumption: Connected} and \ref{Assumption: BoundF} hold and the uncertainty vectors, $\text{$\boldsymbol{\gamma}$}_i$'s, be constant. Let $\bB = [b_{ij}]$ and $\bC = [c_{ij}]$ be the Laplacian matrices corresponding of two graphs on $1,\, \cdots,\,N$, where $\bC$ is connected and undirected and $ \forall i, \, j: c_{ij} \le 0$, if the input of the system $i$ is selected as
\begin{eqnarray}
\text{$\mathbf{u}$}} \def\bU{\text{$\mathbf{U}$}_i &= &-\sum_{j =1}^N b_{ij} \bH \text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j - g_i \bH (\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i - \text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$}) - \bG(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)\hat{\text{$\boldsymbol{\gamma}$}}_i \label{eq: Distributed_input}\\
\dot{\hat{\text{$\boldsymbol{\gamma}$}}}_i &= &k_i \bG^T(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i) \left(\sum_{j = 1}^N c_{ij} \text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j \,+\, z^\prime_i (\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i - \text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$}) \right),\label{eq: Distributed_Estimation}
\end{eqnarray}
then, the network in \eqref{eq: NetworkEq} asymptotically uniformly converges to the reference signal, if there exists positive constants, $k_i>0$, and nonnegative constants $z_i,z_i^\prime \ge0$ such that at least one $z^\prime_i >0$ and $z_i >0$ and
\begin{align}\label{eq: Theorem3}
\bF- \mu_i \bH^{(s)}\prec\text{$\mathbf{0}$}, \quad i = 1\,\cdots, \, N
\end{align}
where $\mu_i$'s are eigenvalues of $\bL+\bB+\bZ$ and $\bZ \triangleq \text{diag}([z_1 \, \cdots \, z_N]^T)$.\\
\textbf{Proof:} See Appendix \ref{proof: 3}.
\end{theorem}
\begin{remark}
It should be noted that the connectivity of the Laplacian $\bB$ is not required in Theorem \ref{Theorem: Distributed Estimation}, in fact, in sufficiently connected/coupled networks the condition \eqref{eq: Theorem3} can be satisfied with $\bB = \text{$\mathbf{0}$}$, as will be shown in the numerical example. In this case, the problem of synchronization of the network to the reference trajectory, is reduced to the well-known pinning problem. However, Theorem \ref{Theorem: Distributed Estimation} requires the communication/feedback network of mismatch estimation, \bC, in \eqref{eq: Distributed_Estimation} to be connected. Although this condition is related to the validity of the chosen Lyapunov condition, intuitively, it also seems required as to estimate the mismatches, each system should have some level of information about the reference system which can collect from its neighbors on the feedback network, \bC, and if the network is not connected this might not be satisfied. Furthermore, Theorem \ref{Theorem: Distributed Estimation} requires $ \forall i, \, j: c_{ij}\le 0$ which implies that the feedback from the neighboring systems should be positive while there is an equivalent negative self-feedback. This is standard condition in cooperative control.
\end{remark}
\begin{remark}For practical purposes, if condition \eqref{eq: Theorem3} allows, the communication network for the linear part of the controller corresponding to the Laplacian $\bB$ can be assumed to be a subnetwork of the communication network of distributed mismatch estimators corresponding to $\bC$. Also in application such as microgrid for distributed generators, due to availability of communication protocols such as power line communication (PLC), it is reasonable to assume that the communication network corresponding to \bC, is a subnetwork of plant network corresponding to \bL.
\end{remark}
\begin{remark}
The condition of $\bZ\succeq \text{$\mathbf{0}$} $ and $\bZ \ne \text{$\mathbf{0}$}$, is a necessary condition in pinning problems and there are many literatures dealing with the design of $\bZ$, please see \cite{Chen07}\cite{Yu09}\cite{ Manaffam.ACC13}\cite{Pirani.TAC16}\cite{ Bapat.Cyb16}\cite{ DeLellis13} and references therein. The same applies to $\bZ^\prime$.
\end{remark}
\begin{remark}\label{remark: Conv_Est2}
Similar to Remark \ref{remark: Conv_Est}, if there exists a $T>t_0$ such that $\bG(\text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$})$ is not singular, \textit{i. e.,} $\text{det}(\bG(\text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$}))\ne0:~ \forall t \ge T$, then $\lim_{t \to \infty} \hat{\text{$\boldsymbol{\gamma}$}}_i = \text{$\boldsymbol{\gamma}$}_i$.
\end{remark}
\section{Numerical Example}
Consider the Lorenz system
\begin{align*}
\text{$\dot{\mathbf{x}}$}=\left[\begin{array}{c}
a(x_2-x_1)\\
bx_1-x_2-x_1x_3\\
x_1x_2-cx_3
\end{array}\right]+\left[\begin{array}{ccc}
x_2-x_1 & 0& 0\\
0 & x_1 & 0\\
0 & 0& -x_3
\end{array}\right]\text{$\boldsymbol{\gamma}$},
\end{align*}
where $\text{$\boldsymbol{\gamma}$}=[\gamma_1~\gamma_2~\gamma_3]^T$, $(a,b,c)=(10,\,28,\,8/3)$. We assume a network of size $N = 50$ which is globally connected, \textit{i. e.,} $\bL=\bR_{50}$, where $\bR_N$ is defined in \eqref{eq: Globally Connected Net}. The inner coupling matrix is assumed to be $\bH= 10 \, \bI_3$. The mismatch parameters are assumed to be constant and satisfy $|\gamma_{1,i}|\le 0.1\,a$, $|\gamma_{2,i}|\le 0.1\,b$, $|\gamma_{3,i}|\le 0.1\,c$.
The matrices \bF~and \text{$\boldsymbol{\Gamma}$}, in Assumptions \ref{Assumption: BoundF} and \ref{Assumption: Bounded Mismatch} are \cite{Manaffam.Allerton15}
\begin{align*}
\bF=&\left[\begin{array}{ccc}
21 & 10 & 0 \\
28 & 23 & 0 \\
0 & 0 & 40
\end{array}\right],
\text{$\boldsymbol{\Gamma}$}=&\left[\begin{array}{ccc}
213 & 0& 0\\
0 & 400 & 0\\
0 & 0& 2500
\end{array}\right].
\end{align*}
\begin{figure}
\includegraphics[width =8.9 cm]{Fig1}
\caption{The evolution of the networked systems' error from its average trajectory, and the bound given in \eqref{eq: Theorem1}.}\label{fig: Bounded_Error}
\end{figure}
Fig. \ref{fig: Bounded_Error} shows the error of the networked system from its average trajectory. As it can be seen the error from the average trajectory reaches the bound in \eqref{eq: Theorem1} in finite time as expected from Lemma \ref{Lemma: Bounded x} and Theorem \ref{Theorem: BoundedError}. The solid plot corresponds to norm of total error, \textit{i. e.,} $\|\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}\|$, as defined in Theorem \ref{Theorem: BoundedError}, and dashed plots corresponds to several samples of system trajectory errors, $\|\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i\|$. The bound in \eqref{eq: Theorem1} is $0.91$.
The norm of network error, $\|\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}\|$ and norm of total error for estimation, $\|\hat{\text{$\boldsymbol{\gamma}$}}-\text{$\boldsymbol{\gamma}$}\|$, are shown in fig.s \ref{fig: Decentral_Esti_States} and \ref{fig: Decentral_Esti_Parameters}, respectively, where the compensator \eqref{eq: Theorem2} in Theorem \ref{Theorem: MismatchEstimation} is used on all the locations in the network. The controller parameters are set as follow: $ z_i = 10,\, k_i =1, \,\forall \,i $. The dashed plots provide several examples of the norm of synchronization error, $\|\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i\|$, and estimation error, $\|\hat{\text{$\boldsymbol{\gamma}$}}_i-\text{$\boldsymbol{\gamma}$}_i\|$, at each system, respectively. As it can be seen and predicted by Theorem \ref{Theorem: MismatchEstimation}, the synchronization error from the reference trajectory asymptotically vanishes. Moreover, since $\bG(\text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$})$ is nonsingular, all the estimated uncertainties converge to their actual values, this has been predicted in Remark \ref{remark: Conv_Est} and shown in fig. \ref{fig: Decentral_Esti_Parameters}.
\begin{figure}[t]
\includegraphics[width = 8.9cm]{Fig2} \caption{The evolution of norm of network 's error from the reference trajectory, $\| \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}\| = \|\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$} - \text{$\mathbf{1}$}_N\otimes \text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$}\|$, and sample error trajectories for systems 1 and 2, \textit{i. e.,} $\|\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_1\| = \|\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_1 - \text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$}\| $ and $ \|\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_2 \| = \| \text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_2 - \text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$}\|$ using decentralized controller \eqref{eq: Theorem2} in Theorem \ref{Theorem: MismatchEstimation}.}\label{fig: Decentral_Esti_States}\vspace{+ 0.5 cm}
\includegraphics[width = 8.9cm]{Fig3}\caption{The evolution of norm of network 's estimation error from the mismatches, $\| \tilde{\text{$\boldsymbol{\gamma}$}}\| = \|\hat{\text{$\boldsymbol{\gamma}$}} - \text{$\boldsymbol{\gamma}$}\|$, and sample estimation errors for systems 1 and 2, \textit{i. e.,} $\|\tilde{\text{$\boldsymbol{\gamma}$}}_1\| = \|\hat{\text{$\boldsymbol{\gamma}$}}_1 - \text{$\boldsymbol{\gamma}$}_1\| $ and $ \|\tilde{\text{$\boldsymbol{\gamma}$}}_2 \| = \| \hat{\text{$\boldsymbol{\gamma}$}}_2 - \text{$\boldsymbol{\gamma}$}_2\|$ using decentralized estimator \eqref{eq: input} in Theorem \ref{Theorem: MismatchEstimation}.}\label{fig: Decentral_Esti_Parameters}
\end{figure}
Fig.s \ref{fig: Dist_Esti_States} and \ref{fig: Dist_Esti_Parameters} show the evolution of the norm of synchronization error, $\|\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}\|$ and norm of total error for estimation, $\|\hat{\text{$\boldsymbol{\gamma}$}}-\text{$\boldsymbol{\gamma}$}\|$, respectively, where the compensator \eqref{eq: Theorem3} in Theorem \ref{Theorem: Distributed Estimation} is used. The communication network, $\bC = [c_{ij}]$, assumed to be a path graph, \textit{i. e.,}
\begin{equation*}
c_{ij} = \left\{\begin{array}{ll}
-1 & j = i +1\, \& \, i = 1, \,\cdots \, N - 1\\
-1 & j = i -1\, \& \, i = 2, \,\cdots \, N\\
2 & j = i , \& \, i = 2, \,\cdots \, N -1\\
1 & j = i , \& \, i = 1, \, N\\
0 & \text{o. w.}
\end{array}\right.
\end{equation*}
Considering the calculated $\bF$ in Assumption \ref{Assumption: BoundF} and given $\bH = 10 \, \bI_n$, the condition \eqref{eq: Theorem3} in Theorem \ref{Theorem: Distributed Estimation} is satisfied iff $\mu_i \ge 4.0\, \forall\, i$. Here, we assume $\bB = \text{$\mathbf{0}$}$, hence the stability condition $\mu_i \ge 4.0\, \forall\, i$ implies that the network at least should be pinned on $5$ locations \cite{Pirani.TAC16, Manaffam.CDC13}. To increase the convergence of the estimator in \eqref{eq: Distributed_Estimation}, we have chosen to optimally pin the communication network, \bC, which results in pinning the systems $i = 5 ,\, 16 , \,26 , \, 35 , \, 46$, \textit{i. e.,} $z_i^\prime = z_i = 1,\,\forall \,i\in\{5 ,\, 16 , \,26 , \, 35 , \, 46\}$ and otherwise $z^\prime_i =z_i =0$. The estimator gains are set to $k_i =10, \forall \,i$.
\begin{figure}[t]
\includegraphics[width =8.9cm]{Fig4}
\caption{The evolution of norm of network 's error from the reference trajectory, $\| \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}\| = \|\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$} - \text{$\mathbf{1}$}_N\otimes \text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$}\|$, and sample error trajectories for systems 1 and 2, \textit{i. e.,} $\|\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_1\| = \|\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_1 - \text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$}\| $ and $ \|\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_2 \| = \| \text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_2 - \text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$}\|$ using distributed controllers \eqref{eq: Distributed_input} in Theorem \ref{Theorem: Distributed Estimation}.}\label{fig: Dist_Esti_States}\vspace{+ 0.5 cm}
\includegraphics[width =8.9cm]{Fig5}
\caption{The evolution of norm of network 's estimation error from the mismatches, $\| \tilde{\text{$\boldsymbol{\gamma}$}}\| = \|\hat{\text{$\boldsymbol{\gamma}$}} - \text{$\boldsymbol{\gamma}$}\|$, and sample estimation errors for systems 1 and 2, \textit{i. e.,} $\|\tilde{\text{$\boldsymbol{\gamma}$}}_1\| = \|\hat{\text{$\boldsymbol{\gamma}$}}_1 - \text{$\boldsymbol{\gamma}$}_1\| $ and $ \|\tilde{\text{$\boldsymbol{\gamma}$}}_2 \| = \| \hat{\text{$\boldsymbol{\gamma}$}}_2 - \text{$\boldsymbol{\gamma}$}_2\|$ using distributed estimators \eqref{eq: Distributed_Estimation} in Theorem \ref{Theorem: Distributed Estimation}.}\label{fig: Dist_Esti_Parameters}
\end{figure}
The dashed plots give several examples of the norm of synchronization errors, $\|\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i\|$, and estimation errors, $\|\hat{\text{$\boldsymbol{\gamma}$}}_i-\text{$\boldsymbol{\gamma}$}_i\|$, for each system, respectively. As it can be observed from fig. \ref{fig: Dist_Esti_States} and predicted by Theorem \ref{Theorem: Distributed Estimation}, the synchronization error from the reference trajectory asymptotically vanishes. Moreover, As predicted in Remark \ref{remark: Conv_Est2}, since $\bG(\text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$})$ is nonsingular, all the estimation errors for mismatched parameters asymptotically go to zero, this has been shown in fig. \ref{fig: Dist_Esti_Parameters}.
\section{Conclusion}
{Here, we have derived sufficient conditions on the ultimately bounded stability of network of mismatched systems. A bound on the error from the average trajectory of the networked system has been calculated, which the network achieves that bound in finite time. To pin the network to a reference trajectory/state, two adaptive controllers, with decentralized and distributed structures, have been proposed. It has been shown that the compensated network will achieve absolute synchronization in presence of constant parameter uncertainties.}
\appendices
\section{Proof of Lemma \ref{lem: commute}}\label{proof: commute}
Let $\bY = [y_{ij}] = \bR_N \bP$, then for $ \forall i \ne j$
\begin{eqnarray}
y_{ij} & = & \sum_{ k =1}^N r_{ik} p_{kj} = r_{ii} p_{ij} + r_{ij} p_{jj} + \sum_{k =1,k \ne i,\,j}
^N r_{ik} p_{kj}\nonumber\\
& = & (N - 1) p_{ij} - p_{jj} - \sum_{k =1,k \ne i,\,j}
^N p_{kj},\nonumber
\end{eqnarray}
considering that \bP~is \textit{symmetric} Laplacian, and henceforth, zero column-sum: $- \sum_{k =1,k \ne i,\,j}^N p_{kj} = p_{jj} + p_{ij}$, we have
\[ y_{ij} = N p_{ij}.\]
For $ i = j$
\begin{eqnarray}
y_{ii} & = & \sum_{ k =1}^N r_{ik} p_{ki} = r_{ii} p_{ii} + \sum_{k =1,k \ne i}
^N r_{ik} p_{ki}\nonumber\\
& = & (N - 1) p_{ii} - \sum_{k =1,k \ne i}
^N p_{ki},\nonumber
\end{eqnarray}
using the fact that for \textit{symmetric} Laplacian matrices, $p_{ii} = - \sum_{k =1,k \ne i,\,j}^N p_{ki} $, we have
\[ y_{ii} = N p_{ii},\]
which means $\bR_N\bP = N \bP$. Showing that $\bP\bR_N =N\bP$ is very similar to the above, except that $- \sum_{k =1,k \ne i,\,j}^N p_{jk} = p_{jj} + p_{ij}$ which it is true for any Laplacian.
\section{Proof of Theorem \ref{Theorem: BoundedError}}\label{Proof: BoundedError}
\begin{IEEEproof}
Let $\text{$\mathbf{u}$}} \def\bU{\text{$\mathbf{U}$} = \text{$\mathbf{0}$}$, $\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$} \triangleq [\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_1^T \, \cdots \, \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_N^T]^T$ and the candidate Lyapunov function be $V(\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$})=1/2\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}^T\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}=1/2\sum_{i=1}^N\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i$, then,
\begin{align*}
\dot{V}=& 1/2\sum_{i=1}^N\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\text{$\dot{\mathbf{e}}$}_i+1/2\sum_{i=1}^N\text{$\dot{\mathbf{e}}$}_i^T\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i \\
=& \frac1{N^2}\sum_{i=1}^N\left(\sum_{j=1}^N(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i-\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j)^T\sum_{k=1}^N[\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)-\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_k)]\right) \nonumber \\
&-\sum_{i,j=1}^Nl_{ij}\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\bH^{(s)}\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_j+\sum_{i=1}^N\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\bG(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)\text{$\boldsymbol{\gamma}$}_i \nonumber
\\
=&\frac1{N^2}\sum_{i,j,k=1}^N(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i-\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j)^T[\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)-\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_k)]\nonumber \\
&-\sum_{i,j=1}^Nl_{ij}\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\bH^{(s)}\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_j+\sum_{i=1}^N\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\bG(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)\text{$\boldsymbol{\gamma}$}_i.
\end{align*}
Since the $\sum_{i,j=1}^{N}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_{i}-\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_{j})=\text{$\mathbf{0}$}$, the first sum, referred to as $V_1$, can be rewritten as
\begin{align*}
V_1=&\frac1N\sum_{i,j=1}^N(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i-\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j)^T\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i) \\
=&\frac1{2N}\sum_{i,j=1}^N(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i-\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j)^T\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)-\frac 1{2N} \sum_{i,j=1}^N (\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j-\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)^T \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j) \\
=&\frac1{2N}\sum_{i,j=1}^N(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i-\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j)^T[\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)-\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j)] \\
\stackrel{(a)}{\le}&\frac1{2N}\sum_{i,j=1}^N(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i-\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j)^T\bF(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i-\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j) \\
\le&\frac1{2N}\sum_{i,j=1}^N(\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i-\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_j)^T\bF(\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i-\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_j)=\frac1{N}\sum_{i,j=1}^N\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\bF\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i-\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\bF\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_j \\
\le&\frac1N\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}^T(\bR_N\otimes\bF)\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}.
\end{align*}
where inequality $(a)$ is due to Assumption \ref{Assumption: BoundF}. Using Lemma \ref{Lemma: Vector Product Inequality} and Assumption \ref{Assumption: Bounded Mismatch},
\begin{align*}
\sum_{i=1}^N\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\bG(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)\text{$\boldsymbol{\gamma}$}_i&\le \sum_{i=1}^N\frac{\beta}2\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i+ \frac1{2\beta}\text{$\boldsymbol{\gamma}$}_i^T\bG(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)^T\bG(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)\text{$\boldsymbol{\gamma}$}_i \\
&\le\frac{\beta}2\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}^T\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}+\frac N{2\beta}\text{$\boldsymbol{\gamma}$}_c^T\text{$\boldsymbol{\Gamma}$}\text{$\boldsymbol{\gamma}$}_c,
\end{align*}
where $\beta$ is an arbitrary positive constant. Therefore,
\begin{align*}
\dot{V}\le\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}^T\Big(\frac1N\bR_N\otimes\bF-\bL\otimes\bH^{(s)}+\frac{\beta}2\bI_{Nn}\Big)\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}
+\frac N{2\beta}\text{$\boldsymbol{\gamma}$}_c^T\text{$\boldsymbol{\Gamma}$}\text{$\boldsymbol{\gamma}$}_c.
\end{align*}
Since $\bR_N$ and \bL~ are both symmetric and Laplacian, they commute, \textit{i.e.}, $\bL\bR_N=\bR_N\bL$. From Lemma \ref{Lemma: Joint Diagonalization} there exists a unitary matrix, \bQ, such that $\bR_N$ and \bL~are jointly diagonalizable
\begin{align*}\begin{array}{lr}
\bL=\bQ\bJ_L\bQ^T& \bR_N=\bQ\bJ_R\bQ^T, \end{array}
\end{align*}
where $\bJ_R=\mbox{diag}([0,~N,\cdots,~N])$ and $\bJ_L$ has the eigenvalues of the Laplacian $\bL$ as its diagonal entries.
Define
\[\text{$\boldsymbol{\eta}$}\triangleq (\bQ\otimes\bI_n)\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$},\]
then
\begin{align*}
\dot{V}\le&\text{$\boldsymbol{\eta}$}^T\left(\frac1N\bJ_R\otimes\bF-\bJ_L\otimes{\bH^{(s)}}+\frac{\beta}{2}\bI_{Nn}\right)\text{$\boldsymbol{\eta}$}\\
&+\frac N{2\beta}\text{$\boldsymbol{\gamma}$}_c^T\text{$\boldsymbol{\Gamma}$}\text{$\boldsymbol{\gamma}$}_c.
\end{align*}
{As any Laplacian matrix of a connected network has one zero eigenvalue, $\bJ_R^{(N)}=\bJ_L^{(N)}=0$, with eigenvector $\text{$\mathbf{q}$}} \def\bQ{\text{$\mathbf{Q}$}_N=\text{$\mathbf{1}$}_N/\sqrt{N}$; Thus,}
\[(\text{$\mathbf{q}$}} \def\bQ{\text{$\mathbf{Q}$}_N^T\otimes\bI_n)\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}=\sum_{i=1}^N\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i/\sqrt{N}={\text{$\boldsymbol{\eta}$}_N}.\]
Since $\sum_{i=1}^N\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i=\text{$\mathbf{0}$}$, the last component of $\text{$\boldsymbol{\eta}$}$ is zero, that is, $\text{$\boldsymbol{\eta}$}_N=\text{$\mathbf{0}$}$. Now, if there exists a constant $\rho>0$ such that
\begin{align*}
\bF-{\mu_i}\bH^{(s)}+(\rho-{\beta}/{2})\bI_{n}\prec\text{$\mathbf{0}$},\quad \forall i\in\{1,\,\cdots,\,N-1\}
\end{align*}
where $\mu_i$'s are nonzero eigenvalue of the Laplacian matrix, \bL~(if the network is connected, \bL~has $N-1$ positive eigenvalues which is a direct result of employing Assumption \ref{Assumption: Connected}); Hence,
\begin{align*}
\dot{V}(\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$})\le -(\rho-\beta/{2})\|\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}\|^2 + N\text{$\boldsymbol{\gamma}$}_c^T\text{$\boldsymbol{\Gamma}$}\text{$\boldsymbol{\gamma}$}_c / (2\beta).
\end{align*}
Using lemma \ref{Lemma: Bounded x} and setting $\varepsilon^{2}= N\text{$\boldsymbol{\gamma}$}_c^T\text{$\boldsymbol{\Gamma}$}\text{$\boldsymbol{\gamma}$}_c/(2 \beta \rho-2 \beta \epsilon-\beta^2)$, we have
\begin{align*}
\begin{array}{lc}
\dot{V}(\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$})\le -\epsilon\|\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}\|^2, &\forall \|\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}\|\ge\sqrt{N\text{$\boldsymbol{\gamma}$}_c^T\text{$\boldsymbol{\Gamma}$}\text{$\boldsymbol{\gamma}$}_c/(2 \beta \rho-2 \beta \epsilon-\beta^2)}.\end{array}
\end{align*}
Define $\lambda^\star$ as \eqref{eq: lambda_star}; If $\epsilon>0$ is arbitrarily small number, $\beta=\lambda^\star-\epsilon$ maximizes the denominator subject to the stability condition, $\lambda^\star-{\beta}/{2}-\epsilon \ge 0$. Therefore,
\begin{align*}
\begin{array}{lc}
\dot{V}(\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$})\le -{\epsilon}\|\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}\|^2, &\forall \|\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}\|\ge\sqrt{ {N}{\text{$\boldsymbol{\gamma}$}_c^T\text{$\boldsymbol{\Gamma}$}\text{$\boldsymbol{\gamma}$}_c}/{(\lambda^\star-\epsilon)}^2}.
\end{array}
\end{align*}
Applying Lemma \ref{Lemma: Bounded x}, the error of the network will be bounded by \eqref{eq: Theorem1} in finite time.
\end{IEEEproof}
\section{Proof of Theorem \ref{Theorem: MismatchEstimation}}\label{Proof: MismatchEstimation}
\begin{IEEEproof}
Let
\[\mbox{$\tilde{\bgamma}$}_i\triangleq\text{$\boldsymbol{\gamma}$}_i-\mbox{$\hat{\bgamma}$},\]
\[V=\frac12\sum_{i=1}^N\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i+\sum_{i=1}^N\frac1{2k_i}\mbox{$\tilde{\bgamma}$}_i^T\mbox{$\tilde{\bgamma}$}_i.\]
Then,
\begin{align*}
\dot{V}=&\frac12\sum_{i=1}^N\text{$\dot{\mathbf{e}}$}_i^T\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i+\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\text{$\dot{\mathbf{e}}$}_i+\sum_{i=1}^N\frac1{k_i}\dot{\mbox{$\tilde{\bgamma}$}}_i^T\mbox{$\tilde{\bgamma}$}_i\\
=&\sum_{i=1}^N\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T[\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)-\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$})]+\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\bG(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)\mbox{$\tilde{\bgamma}$}_i- c_i\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\bH^{(s)}\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i\\
&-\sum_{i,j=1}^Nl_{ij}\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\bH^{(s)}\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_j-\sum_{i=1}^N\frac1{k_i}\dot{\mbox{$\hat{\bgamma}$}}^T_i\mbox{$\tilde{\bgamma}$}_i,
\end{align*}
substituting $\dot{\mbox{$\hat{\bgamma}$}}$ from \eqref{eq: input}
\begin{align*}
\dot{V}=&\sum_{i=1}^N\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T[\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)-\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$})]+(\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\bG(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)-\frac1{k_i}\dot{\mbox{$\hat{\bgamma}$}}^T_i)\mbox{$\tilde{\bgamma}$}_i\\
&-\sum_{i,j=1}^Nl_{ij}\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\bH^{(s)}\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_j- \sum_{i=1}^Nc_i\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\bH^{(s)}\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i\\
\le&\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}^T(\bI_N\otimes\bF-(\bL+\bC)\otimes\bH^{(s)})\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}.
\end{align*}
If \eqref{eq: Theorem2} holds, then using Lemma \ref{Lemma: asymptotic W}, we conclude that $\|\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}\| $ uniformly goes to zero, $\|\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}\|\to 0$, and $\|\mbox{$\tilde{\bgamma}$}_i\|$'s are bounded.
\end{IEEEproof}
\section{Proof of Theorem 3}\label{proof: 3}
\begin{IEEEproof}
Substituting \eqref{eq: Distributed_input} in \eqref{eq: Error_Ref}, we have
\begin{align}
\text{$\dot{\mathbf{e}}$}_i = \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i) - \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$}) - \sum_{j = 1}^N (l_{ij}+b_{ij}) \bH \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_j - g_i \bH \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i + \bG(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i) \tilde{\text{$\boldsymbol{\gamma}$}}_i,\label{eq: proof0}
\end{align}
where $\tilde{\text{$\boldsymbol{\gamma}$}}_i \triangleq \text{$\boldsymbol{\gamma}$}_i - \hat{\text{$\boldsymbol{\gamma}$}}_i$. Now let $\bC = [c_{ij}]$ be the Laplacian of connected undirected graph on $1,\,\cdots,\, N$, then $\bC + \bZ^\prime$ is symmetric positive definite matrix if there exists at least one $g_i^\prime>0$ where $\bZ^\prime \triangleq \text{diag}([z^\prime_1 \, \cdots \, z^\prime_N]^T)$ \cite{Chen07}. Hence
\[ V = \frac12 \sum_{i ,j =1}^N c_{ij} \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_j + \frac12 \sum_{i =1}^N z^\prime_i \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i + \frac12\sum_{i=1}^N \tilde{\text{$\boldsymbol{\gamma}$}}_i^T\tilde{\text{$\boldsymbol{\gamma}$}}_i,\]
is a candidate Lyapunov function.
\begin{align}
\dot{V} = \underbrace{\sum_{i ,j =1}^N c_{ij} \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T \text{$\dot{\mathbf{e}}$}_j }_{\triangleq V_1}+ \underbrace{\sum_{i =1}^N z^\prime_i \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\text{$\dot{\mathbf{e}}$}_i }_{\triangleq V_2} - \sum_{i=1}^N \tilde{\text{$\boldsymbol{\gamma}$}}_i^T\dot{\hat{\text{$\boldsymbol{\gamma}$}}}_i. \label{eq: proof1}
\end{align}
By substituting \eqref{eq: proof0} in the first term in \eqref{eq: proof1}, we have
\begin{align}
V_1 =& \sum_{i,j =1}^N c_{ij} \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T (\text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j) - \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{s}$}} \def\bS{\text{$\mathbf{S}$}))- \!\! \!\! \sum_{i,j,k =1}^N c_{ij}(l_{jk}+b_{jk})\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\bH\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_k\nonumber\\
&+\sum_{i,j =1}^N g_jc_{ij} \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\bH \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_j + \sum_{i,j =1}^Nc_{ij}\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i\bG(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j)\tilde{\text{$\boldsymbol{\gamma}$}}_j\nonumber\\
{=}& \sum_{i,j =1}^N c_{ij} \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j)-\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}^T\left(\bC(\bL+\bB+\bG)\otimes\mbox{{\bf H}}^{\scriptsize\mbox{(s)}} \right)\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}\nonumber\\
&+ \sum_{i,j =1}^Nc_{ij}\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\bG(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j)\tilde{\text{$\boldsymbol{\gamma}$}}_j,
\end{align}
where last equality is due to zero row-sum property of $\bC$. Also $\sum_{i,j =1}^N c_{ij} \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i) = \sum_{i=1}^N \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)\sum_{j = 1}^N c_{ij} = \text{$\mathbf{0}$}$, hence, we have
\begin{align*}
\sum_{i,j =1}^N c_{ij} \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j) & = \sum_{i, \, j =1}^N c_{ij} \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j) - \sum_{i,j =1}^N c_{ij} \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i)\\
& = \sum_{i, \, j =1}^N c_{ij} \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\Big( \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j) - \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i) \Big)\\
& = \frac12 \sum_{i,\, j =1}^N c_{ij} \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T \Big( \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j) - \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i) \Big) \\
&\quad \quad +\frac12 \sum_{i,j =1}^N c_{ij} \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T \Big( \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j) - \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i) \Big) \\
& = \frac12 \sum_{i,\, j =1}^N c_{ij} \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T \Big( \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j) - \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i) \Big) \\
&\quad \quad +\frac12 \sum_{i,j =1}^N c_{ji} \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_j^T \Big( \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i) - \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j) \Big)
\\
&= \frac12 \sum_{i,\, j =1}^N c_{ij} (\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i - \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_j)^T \Big( \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_i) - \text{$\mathbf{f}$}} \def\bF{\text{$\mathbf{F}$}(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j) \Big) \\
& \le \frac12 \sum_{i,\, j =1}^N |c_{ij}|(\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i - \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_j)^T \bF (\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i- \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_j) \\
& \le \sum_{\tiny\begin{array}{l} i,\,j =1, \\ j\ne \, i\end{array}}^N \!\! \!\! |c_{ij}| ~\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i ^T\, \bF \, \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i - \!\! \!\! \!\! \sum_{\tiny\begin{array}{l} i,\,j =1, \\ j\ne \, i\end{array}}\!\! \!\! |c_{ij}| \, \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T \, \bF \, \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_j
\end{align*}
\begin{align*}
& \le \sum_{i=1}^N ~\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i \, \bF \, \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i \left(\sum_{\tiny\begin{array}{l} j =1, \\ j\ne \, i\end{array}}^N \!\! \!\! |c_{ij}| \right)\\
& \quad \quad- \, \sum_{\tiny\begin{array}{l} i,\,j =1, \\ j\ne \, i\end{array}}|c_{ij}| \, \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i \, \bF \, \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_j \\
& \le \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}^T \left(\bC \otimes \bF \right) \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}.
\end{align*}
Hence,
\begin{align}
V_1 \le ~ & \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}^T\left(\bC \otimes \bF - \bC(\bL+\bB+\bG)\otimes\mbox{{\bf H}}^{\scriptsize\mbox{(s)}}\right)\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}\nonumber\\
&+ \sum_{i,j =1}^Nc_{ij}\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\bG(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j)\tilde{\text{$\boldsymbol{\gamma}$}}_j.
\end{align}
Therefore,
\begin{align}
\dot{V} \le \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}^T\left((\bC+\bZ^\prime) \otimes \bF - (\bC+\bZ^\prime)(\bL+\bB+\bG)\otimes\mbox{{\bf H}}^{\scriptsize\mbox{(s)}} \right)\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$} \nonumber\\
+\sum_{i =1}^Nz^\prime_{i}\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\bG(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j)\tilde{\text{$\boldsymbol{\gamma}$}}_j + \sum_{i,j =1}^Nc_{ij}\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i^T\bG(\text{$\mathbf{x}$}} \def\bX{\text{$\mathbf{X}$}_j)\tilde{\text{$\boldsymbol{\gamma}$}}_j - \sum_{i =1}^N \dot{\hat{\text{$\boldsymbol{\gamma}$}}}_i^T\tilde{\text{$\boldsymbol{\gamma}$}}_i.\nonumber
\end{align}
Substituting $\dot{\hat{\text{$\boldsymbol{\gamma}$}}}_i$ from \eqref{eq: Distributed_Estimation}, yields
\begin{align*}
\dot{V} \le \text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}^T \!\! \left([(\bC+\bZ^\prime) \otimes \bI_n ]\!\! \left[\bI_N \otimes \bF - (\bL+\bB+\bZ)\otimes\mbox{{\bf H}}^{\scriptsize\mbox{(s)}} \right]\right)\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}.
\end{align*}
If conditions in \eqref{eq: Theorem3} hold, Lemma \ref{Lemma: asymptotic W} implies that $\text{$\mathbf{e}$}} \def\bE{\text{$\mathbf{E}$}_i$'s are asymptotically stable and $\tilde{\text{$\boldsymbol{\gamma}$}}_i$'s are bounded. It should be noted that the product of two symmetric positive definite matrices is a positive definite matrix (not necessarily symmetric).This can be shown by the Weyl's inequalities for product of two symmetric matrices.
\end{IEEEproof}
\bibliographystyle{IEEETran}
|
\section{Introduction}
A wireless sensor network (WSN) consists of a number of embedded sensor devices, where the nodes\footnote{A sensor device and a node are synonymous.} collect data using their sensors and report it to a central collection point, called the sink node. If the sink is connected to the Internet, the data can be accessed over the Internet, i.e., the WSN deployment region can be monitored remotely. The ease of deployment and management broadened the use of WSN and makes it a very useful tool for the Internet of Things (IoT).
Most of the WSN devices are battery operated with the sink node as a possible exception. Thus, WSN related research is primarily focused on energy-efficient network operations as energy is one of the most precious resources. As a result, the routing protocols for other types of ad-hoc networks are not suitable for WSN. The routing protocols for WSN are primarily focused towards reducing the duty cycle (radio on time) of the nodes.
Over the years, many routing protocols have been proposed and developed. Recently, the low-power wireless bus (LWB)~\cite{ferrari2012low} has been proposed by Ferrari \textit{et al.}. Using extensive experiments, it is shown to be the most energy efficient communication mechanism till its inception. The design of LWB is very generic and simple. Also, there is a drastic difference in its design as compared to traditional communication mechanism in WSN. This makes it a very attractive choice as a communication protocol. However, there is a lack of open implementation of LWB. This restricts the usage and further development of protocols in the academia and industry. This article provides an open implementation of the protocol for Tmote Sky platform. The reason behind choosing the Sky platform is that it is one of the most common sensor node platform among in the WSN research community. Most of the online testbed (e.g., Indriya~\cite{doddavenkatappa2012indriya}, Flocklab~\cite{lim2013flocklab}) also consists of Sky-based devices. The goal of this article is to enable development of communication mechanisms based on LWB as well as provide a comparison platform for the new protocols.
As mentioned earlier, LWB is a very generic protocol and it can cater a vast range of application scenarios. This also provides the scope of improvement for a particular use-case. Forwarder selection low-power wireless bus (FS-LWB)~\cite{forwarderdcoss2013} is one such optimization of LWB for data collection applications. This implementation also provides a plug-in to enable forwarder selection on top of LWB. The code is available in~\cite{lwb_code}.
\section{Background }
Before providing the implementation details, we describe the basic concept of LWB and the optimization achieved by FS-LWB. However, we highly recommend to read the original articles by Ferrari \textit{et al.}~\cite{ferrari2012low} and Carlsen \textit{et al.}\cite{forwarderdcoss2013}, respectively.
The development of LWB stands on the shoulder of a fast and efficient network-wide flooding mechanism, called Glossy~\cite{ferrari2011efficient}. In Glossy, the master clock holder (the sink node) sends a periodic sync packet. Every node, after receiving the sync packet, rectifies its clock offset and becomes synchronized with the sink node. In contrast with traditional clock synchronization protocol where the process of resynchronization takes a long time, glossy can quickly resynchronize the whole network. This fast synchronization is an effect of the fast flooding mechanism. In glossy, nodes that are far away from the sync, they too receive the sync packet within a small time-bound. This delay is compensated based on hop distance of a node from the sink. The sync packet contains a counter that is initiated to 1 by the sink. Every node after receiving the packet increases the counter before forwarding it. Thus, based on the counter value of the received packet, a node determines its hop distance from the sink.
Glossy achieves an incredibly fast network-wide flooding by utilizing a phenomenon called constructive interference (CI). When multiple nodes transmit the same content simultaneously, the signals interfere constructively at the receiver. As a result, the receiver can successfully decode the content of the transmitted packet. In Glossy, every node wakes up (turns on the radio) just before the start of a new flood. The sink sends the sync packet and all its first hop neighbors receive its. Packet reception at a node triggers the packet transmission in by the node. Glossy ensures a fixed delay for receive-to-transmit switching delay at every node. Thus, when the first hop nodes transmit the packet (immediately or after a small but fixed delay), they transmit concurrently. This ensures CI at the second hop nodes. As a result, all the second-hop nodes are able to receive the packet at the same time. Then they also forward the packet immediately, which takes the packet further away from the sink. This ripple effect immediately covers the whole network and the flooding gets completed. As mentioned earlier, nodes implicitly become synchronized with the sink based on receiving the sync packet.
By utilizing Glossy as a communication primitive, the low-power wireless bus (LWB) protocol provides a fast and efficient communication protocol. In Glossy, the sink node is the only node that can start a flood, and the sync packet need not carry any useful payload. However, not only some payload can be attached to the packet to be flooded, the flooding can be started by any other node apart from the sink node. However, if two nodes start the flooding in an overlapping time, the flooding may fail partially or fully. This is because Glossy utilizes CI where one of the requirement is transmitting the same content by the simultaneous transmitters. Thus, the nodes need to coordinate their respective flooding time among themselves such that everyone floods at a non-overlapping time. Moreover, every node should know the number of flooding nodes in the network, and flooding time of each node. LWB arranges these by using a centralized scheduler (can be the sink node). LWB uses a slotted communication, where a slot length is the amount of time required to flood the whole network. Every node requests for a data slot from the sink (scheduler). In return, the scheduler assigns a unique slot for every node.
Once the slot assignment is done, the data transmission is done using Glossy-based flooding, where every node initiates a Glossy flood in its assigned slot. LWB defines an LWB round. At the beginning of a round, the sink sends a sync packet, and the whole network becomes synchronized with the sink. Additionally, the sync packet contains the number of data slots that will follow in that LWB round. At the beginning of every slot, all node wakes up as a receiver except the node that is supposed to start the flood in that slot. Once a node performs it receive-and-forward operation in the flooding process, it goes back to sleep mode until the beginning of the next slot. Then the node either discards the packet or forwards the packet to the application layer depending on whether it is the intended recipient or not. As every communication is based on network-wide flooding, every packet is available to every node. Thus, LWB inherently supports all the traffic patterns, i.e., one-to-one, one-to-many, many-to-one, and all-to-all. One of the major benefits of LWB is that it does not require any control packet exchange to monitor the link quality or the network condition.
As every node participates in every data delivery slot in LWB, it unnecessarily wastes energy when every data packet need not be delivered to every node. For data collection scenarios, where data is only needed to be delivered to the sink, an optimization is proposed by Carlson \textit{et al.}~\cite{forwarderdcoss2013}. For every data source, it finds a set of forwarder nodes along the path from the source to the destination. If a node is on the way to the destination for or a particular source node, it participates in the corresponding flooding slot. Otherwise, it just avoids participating in that particular flooding slot. The forwarder set is discovered in a distributed way without any additional control overhead.
Similarly, another approach that improves the energy efficiency by dynamically adjusting the transmission power is described in~\cite{rao2016dipa}. However, this is out of the scope of this article. Thus, we do not discuss this article any further.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{lwb-round.pdf}
\caption{Slotted communication in LWB: showing different types of slots that are repeated in every LWB round.}
\label{fig:lwb-round}
\end{figure}
\section{Implementation details}
This implementation is done using the Contiki operating system~\cite{dunkels2004contiki}. LWB uses synchronous and slot-based communication scheme. The slot structure of LWB is shown in Fig.~\ref{fig:lwb-round}.
\subsection{LWB slot structure}
As mentioned earlier, every slot is equal to the length of a network-wide Glossy flood. There is three basic type of slots - (i) sync slot, (ii) request/reply slots, and (iii) data slots. Apart from the slots, LWB defines round, which is the periodicity of the sync packet. An LWB round starts with a sync packet, which is followed by other types of slots. However, it is not strict to have either the request/reply or the data slots in every LWB rounds. Also, the number of these slots varies among rounds. Details of each slot are described in the following.
\subsubsection{Sync slot}
The sync slot is exclusively used by the sink node. By receiving this packet, every node becomes synchronized with the sink. The content of this packet indicates the number of request/reply and data slots in this LWB round. Additionally, it also indicates the length of the LWB round, i.e., when the next sync packet will be sent.
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{lwb-process.pdf}
\caption{Working process of LWB from bootstrapping.}
\label{fig:lwb-process}
\end{figure*}
\subsubsection{Request/reply (RR) slot}
These slots are used to obtain a unique data slots. In these slots, multiple nodes can contend to request a data slot from the sink.
In the first slot, nodes contend to send a request. If sink receives a request successfully, it sends a reply in the second slot. The node who already acquires a data slot stops contending. In the third slots again the remaining nodes send a request. The process continues for the specified number of RR slots in an LWB round. If sink mentions $N_{rr}$ RR slots in an LWB round, there can be a maximum of $N_{rr}/2$ requests processed in that round.
\subsubsection{Data slots}
After the intense request/reply period (see Section~\ref{sec:stabilization-period}), the sink node assumes that most of the nodes have acquired a data slot. Thus, the sink node indicates the number of data slots (that it assigned so far), in an LWB round. In the successive LWB rounds, availability of a certain number of data slots ($N_d$) is indicated by the sync packet. In a particular data slot, the node who owns the slot, starts the flooding, while the remaining nodes just participate.
\subsection{LWB operation in a long run}
The process of LWB operation is shown in Fig.~\ref{fig:lwb-process}. The whole process is divided into three phases/periods. The behavior of the process can be manipulated by changing the durations of these periods. Various parameters including the length of these periods can be customized in the specification file (slot-def.h). The three main phases of LWB operation are described below.
\subsubsection{Cool-off period}
The first step of LWB process is to get synchronized with the sync. As the nodes do not have any idea about the LWB slot structure, they can learn this only by receiving a sync packet. At the beginning, a node keeps its radio on to receive the first sync packet. During the cool-off period, the sink node sends a sync packet every second. During this period, an LWB round contains a sync slot only. Once the sync packet is received, the node knows the length of the LWB round, i.e., when the next sync packet will be sent. Thus, the node keeps it radio off until the beginning of the next round. This process is same as Glossy, except the content of the sync packet, which determines the time until the next sync packet and the number of other slots that may follow the sync packet. The idea behind cool-off period is to provide some time so that every node can join the network by synchronizing with the sink nodes.
\subsubsection{Stabilization period}
\label{sec:stabilization-period}
After some cooling off period (LWB round with only sync packets), the sink indicates a number of RR slots that will be followed. Please note, only the sink node knows the duration of the cool-off period. For other nodes, it does not matter, as they just perceive it as an LWB round with just a sync slot.
In the stabilization period, the length of an LWB round is also fixed to 1 seconds. The sink indicates an availability of as many RR slots as it is possible to fit within one second. Hence, the nodes start requesting for a data slot. Please note at this point still there are no data slots in the LWB round. As the time progresses in the stabilization period, more nodes acquire a data slot and stops contending. Thus, if a high number of RR slots are mentioned, a lot of energy will be wasted to participate in the RR slot. We adopt a dynamic approach to adjust the number of RR slots in a round.
Initially, it is set to the maximum that can fit in a second. As soon as the sink sees two request slots within any request, it reduces the number of RR slots to the minimum, i.e., 2 in LWB, and 3 in FS-LWB (see Section~\ref{sec:fs-lwb}). The authors of LWB described 2 RR slots in every LWB rounds. That means in every round only one node can acquire a data slot, and it takes a long time before every node acquires a data slot. In this implementation, data slot acquisition speeds up significantly.
\begin{table*}
\renewcommand{\arraystretch}{1.3}
\centering
\caption{Table defining the application defined parameters (defined in file ``slot-def.h''). These parameters change the behavior of the LWB process and need to be adapted according to application requirements.}
\begin{tabular}{|r|l|l|}
\hline
\textbf{Parameter} & \textbf{Function} & \textbf{Default value} \\ \hline
IPI & sensing frequency & 10\,s \\ \hline
MINIMUM\_LWB\_ROUND & duration of minimum LWB rounds in operational period & 5\,s \\ \hline
COOLOFF\_PERIOD & duration of the cool-off period & 10\,s \\ \hline
STABILIZATION\_PERIOD & duration of the stabilization period & 10\,s \\ \hline
MAX\_PAYLOAD\_LEN & defines maximum payload of an LWB data packet & 40 bytes \\ \hline
SINK\_NODE\_ID & defines who is the sink node & 1 \\ \hline
MAX\_NODE\_NUMBER & highest node-id should not cross this number & 150 \\ \hline
FORWARDER\_SELECTION & enable forwarder selection & 0 (disabled) \\ \hline
\end{tabular}%
\label{table:params}
\end{table*}
\subsubsection{Operational period}
In the stabilization period, most of the nodes have acquired a data slot (if not all). In case there are some nodes who do not have a data slot yet or a node joins the network at a later stage, the minimum number of RR slots are maintained in every LWB round such that these nodes can acquire a data slot. Thus, during the operational period, an LWB round contains 1 sync packet, $N_d$ data slots (equals to the number of data sources in the network), and a minimum number of RR slots.
The length of an LWB round is equal to the periodicity of the sensing application. If the sensing periodicity of the application is too high, the LWB round are cut into smaller length. This ensures that the nodes remain synchronized in between two round. Glossy can tolerate a certain clock drift by the nodes between the rounds, which is defined as the \textit{GLOSSY\_GUARD\_TIME}. It is set to 2\,ms. So the minimum length of an LWB round should be set in a way that the clock drift is contained within this bound. By increasing this guard time, a larger drift can be tolerated and a more infrequent LWB round is possible. However, larger guard time increase energy consumption by the nodes. If the minimum LWB round length is less than the data sensing frequency, the intermediate round would not contain any RR or data slots, but just a sync slot to keep the network synchronous. The relevant parameters that controls the behavior of the LWB process is summarized in Table~\ref{table:params} (defined in file ``slot-def.h'').
\subsection{Forwarder set determination}
\label{sec:fs-lwb}
In the case of traditional LWB, the RR slots are used in pairs, where odd numbered slots are used for requesting a data slot (via contention), and even numbered slots are used by the sink to reply with an assigned data slot. If the forwarder selection mechanism is enabled, the RR slots are used in a triplet. In that case, the every third slot is used by the source node who just acquired the data slot. The source node announces his hop-distance from the sink using this slot. The additional operations are as following.
\begin{itemize}
\item In the second RR slot, when the sink replies with the data slot assigned to the requester, every node calculates its hop distance from the sink along with forwarding the packet. Let's assume this distance is $h_u$.
\item Like any other nodes, the source itself also calculates $h_u$ after receiving the reply.
\item In the third RR slot, the source announces its $h_u$ as $h$ along with the slot number $s$.
\item Upon receiving this announcement, every node calculates its hop distance from the source. Let's assume this distance is $h_d$.
\item After this, a node add slot $s$ to its forwarder set, if it is on the way from the source to the sink, i.e., $h = h_u + h_d$. Otherwise, during the operational phase, the node do not participate in data slot $s$.
\end{itemize}
\section{Conclusions}
The idea behind this document is to help developers to build and evaluate new routing protocols. The development can be an improvement over LWB or FS-LWB, or it can be a completely new approach. In both the cases, this implementation provides an easy way to compare the performance of the new protocols with respect to LWB and FS-LWB.
\section*{Ackowledgement}
This code is primarily based on Glossy~\cite{ferrari2011efficient} implementation and some parts from Chaos~\cite{landsiedel2013chaos}. I cordially thank the authors and the developers of these works.
\tiny
\bibliographystyle{abbrv}
|
\section{Introduction} \label{sec-intro}
Magnetism of diluted and low-dimensional systems such as adatoms,
clusters or monolayers is one of strongly pursued research areas.
Magnetization of these systems often cannot be measured by macroscopic
methods. It can, however, be probed indirectly by making use of
spectroscopy. One of the most powerful methods in this respect is
x-ray magnetic circular dichroism (XMCD). It consists in measuring
the difference in the absorption of left- and right-circularly
polarized x-rays in a magnetized sample while the energy of the
incident x-rays is varied. Analysis of XMCD spectra is often done
with the help of sum rules, which link integrals of XMCD and x-ray
absorption spectral peaks to local spin and orbital magnetic moments.
Most of recent progress in magnetism of atomic-sized systems is
associated with the application of the XMCD sum rules
\cite{GRV+03,DSS+14,DRS+16}.
The strength of the sum rules is that they provide, in the case of
$L_{2,3}$ edge spectra, separate
information about the orbital magnetic moment \mbox{$\mu_{\mathrm{orb}}$}\ and the spin
magnetic moment \mbox{$\mu_{\mathrm{spin}}$}\ of the photoabsorbing atom
\cite{TCSvdL92,CTAW93}. However, extracting values of \mbox{$\mu_{\mathrm{orb}}$}\ and,
especially, of \mbox{$\mu_{\mathrm{spin}}$}\ from the spectra is not straightforward. Considering
the most common case of the \mbox{$L_{2,3}$\ edge}\ spectra and a sample magnetized
along the $\alpha$ direction ($\alpha=x,y,z$), the spin magnetic
moment sum rule can be written as \cite{CTAW93}
\begin{equation}
\frac{3}{I} \, \int \left( \Delta \mu_{{L}_3}
-2\Delta \mu_{{L}_2} \right) \, \mathrm{d}E \, = \,
\frac{\mu_{\text{spin}} + 7T_{\alpha}}{n_{h}}
\enspace ,
\label{eq-spin}
\end{equation}
where \mm{\Delta \mu_{{L}_{2,3}}} are the differences
\mm{\Delta\mu=\mu^{(+)}-\mu^{(-)}} between absorption coefficients for
the left and right circularly polarized light propagating along the
$\alpha$ direction, $I$~is the integrated isotropic absorption
spectrum, $\mu_{\mathrm{spin}}$\ is the local spin magnetic moment
(its $d$\ component, to be precise), and $n_{h}$\ is the number of
holes in the $d$\ band. The term $T_{\alpha}$ is the expectation value
of the intra-atomic spin dipole operator for the valence $d$ electrons. It
is often called the magnetic dipole term in the literature dealing
with XMCD. As the magnetization is typically in the $\alpha=z$
direction, one often speaks simply about the \mbox{$T_{z}$}\ term.
This magnetic dipole \mbox{$T_{\alpha}$}\ term can be written as \cite{Sto95,Sto99}
\begin{align}
T_{\alpha} & \: = \: - \frac{\mu_{B}}{\hbar} \,
\langle \hat{T}_{\alpha} \rangle
\;\; \notag \\
& \: = \: -\frac{\mu_{B}}{\hbar} \,
\left\langle \, \sum_{\beta} Q_{\alpha \beta} S_{\beta} \,
\right\rangle
\enspace ,
\raisetag{1.0\baselineskip} \label{eq-exact}
\end{align}
with
\begin{equation}
Q_{\alpha \beta} \: = \: \delta_{\alpha \beta} \, - \,
3 r^{0}_{\alpha} r^{0}_{\beta}
\end{equation}
being the quadrupole moment operator and $S_{\alpha}$ being the spin
operator. The \mbox{$T_{\alpha}$}\ term cannot be easily determined by experiment and
its occurrence in Eq.~(\ref{eq-spin}) thus poses a serious
problem. For bulk systems, it can be often neglected (provided that
the spin-orbit coupling is not very strong \cite{CLT+95}). However,
for low-dimensional systems it can be significant
\cite{WF94,KEDF02,SMC+10}. Moreover, the \mbox{$T_{\alpha}$}\ term cannot be
considered just as an additive correction that for similar systems simply
shifts the values of \mbox{$\mu_{\mathrm{spin}}$}\ by approximately the same amount. It was
demonstrated that neglecting \mbox{$T_{\alpha}$}\ for a sequence of supported magnetic
clusters could lead to erroneous conclusions regarding the dependence
of the average \mbox{$\mu_{\mathrm{spin}}$}\ on the cluster size \cite{SME09b}. Likewise,
neglecting \mbox{$T_{\alpha}$}\ and its angular dependence could introduce spurious
anisotropy of \mbox{$\mu_{\mathrm{spin}}$}\ for low-dimensional systems
\cite{Sto95,WSN+95,SBE+13}.
In principle, the \mbox{$T_{\alpha}$}\ term can be calculated and inserted into
Eq.~(\ref{eq-spin}). However, one would really have to make the
calculation for each system which is studied,
because the \mbox{$T_{\alpha}$}\ term is quite sensitive to details of the electronic
structure \cite{WF94,KEDF02,SME09b} and taking its values from
calculations for only similar systems might not be reliable. At the
same time, performing calculations for exactly the system one is
interested in may be difficult or impractical.
Fortunately, there appears to be a way to eliminate the \mbox{$T_{\alpha}$}\ term from
Eq.~(\ref{eq-spin}) relying solely on experiment by performing a
series of angle-dependent XMCD measurements. The key here lies in
decoupling the quadrupole moment operator $\hat{Q}$ in
Eq.~(\ref{eq-exact}) from the spin operator $\hat{S}$. This can be
done provided that the influence of the spin-orbit coupling (SOC) on
\mbox{$T_{\alpha}$}\ can be neglected. Then, for a sample magnetically saturated
along the direction $\alpha$, one can express the \mbox{$T_{\alpha}$}\ term as
\cite{SK95a}
\begin{equation}
T_{\alpha} \: = \: \sum_{m} \,
\frac{1}{2} \,
\langle Y_{2m} | \hat{Q}_{\alpha \alpha} | Y_{2m} \rangle \;
\mu_{\text{spin}}^{(m)}
\;\; ,
\label{eq-tzspin}
\end{equation}
where $\mu_{\text{spin}}^{(m)}$ is the spin magnetic moment resolved
into components according to the magnetic quantum number $m$. The
matrix elements $\langle Y_{2m} | \hat{Q}_{\alpha \alpha} | Y_{2m}
\rangle$ can be found in St\"{o}hr and K\"{o}nig \cite{SK95a} and more
elaborate discussion of Eq.~(\ref{eq-tzspin}) can be found in
St\"{o}hr \cite{Sto99} or \v{S}ipr {\it et al.}\ \cite{SBE+13}. Elimination of
the \mbox{$T_{\alpha}$}\ term from the sum rule (\ref{eq-spin}) can then be achieved
by performing three XMCD measurements and making use of the relation
\cite{SK95a}
\begin{equation}
T_{x} + T_{y} + T_{z} = 0
\label{eq-sum}
\; .
\end{equation}
Furthermore, if the system has higher than twofold symmetry around the
$z$ axis, the magnetic dipole term depends on the polar angle $\theta$
as \cite{SK95a,Laan98}
\begin{equation}
T_{\theta} \: \approx \: 3\cos^{2}\theta - 1
\label{eq-magic}
\; .
\end{equation}
The magnetic dipole term, which we will denote $T_{\theta}$ for a
general direction in which the sample is magnetically saturated, can
thus be eliminated by a single XMCD measurement with circularly
polarized x-rays coming in the direction of the magic angle
54.7$^{\circ}$. This approach was employed, e.g., for studying Co
thin films and nanoclusters \cite{WSN+95,KMO+01}.
The important point is that eliminating \mbox{$T_{\alpha}$}\ from the sum rule
analysis is possible only if the effect of SOC on \mbox{$T_{\alpha}$}\ can be
neglected. The question is whether this happens in common
circumstances. Namely, there are theoretical indications that the
effect of SOC on \mbox{$T_{\alpha}$}\ may be sometimes important. It was found that
Eq.~(\ref{eq-sum}) is strongly violated for free-standing Co wires
\cite{EKDF03} (provided that correlation effects beyond the LDA are
included via the Brooks orbital polarization term \cite{Bro85}). For
more realistic materials, violation of Eq.~(\ref{eq-magic}) was
predicted for a Pt monolayer with magnetization induced from an Fe
substrate \cite{OS+04}. Not surprisingly, this violation is even more
serious for systems with very strong SOC such as US
\cite{CLT+95,OS+04}. Recently, there have been also experimental
indications that the SOC may be important for the \mbox{$T_{\alpha}$}\ term: violation of
Eq.~(\ref{eq-sum}) was observed for low-temperature monoclinic phase
of magnetite nanoparticles \cite{SSW+14}.
The most typical situation when XMCD sum rules are used is studying
magnetism of 3$d$ metals in multicomponent systems, and the \mbox{$T_{\alpha}$}\ term
has to be considered especially for thin films, adatoms or clusters.
One should thus explore to what extent the SOC is important for
\mbox{$T_{\alpha}$}\ in these systems so that one knows whether
Eqs.~(\ref{eq-sum})--(\ref{eq-magic}) can be applied to eliminate the
\mbox{$T_{\alpha}$}\ term from the XMCD analysis or not.
To get a comprehensive view, we focus on a sequence of systems
comprising Co monolayers and Co adatoms on Cu, Pd, Ag, Pt, and Au
(111) surfaces. In that way we account for effects connected with the
change of dimensionality and for effects connected with the changes of
chemical environment as well. It should be noted in this context
that the substrate may have a crucial influence on some SOC-induced
properties such as the magnetocrystalline anisotropy \cite{SBE+14}. There
is also theoretical evidence that the substrate has a decisive
influence on \mbox{$T_{\alpha}$}\ of supported systems \cite{SME09b}.
The outline of the paper is the following. We start by describing our
computational framework. Then we present results that are in line
with
Eqs.~(\ref{eq-sum})--(\ref{eq-magic}) for a series of Co monolayers
and adatoms. Here we demonstrate that while for Co monolayers the
effect of SOC on \mbox{$T_{\alpha}$}\ can be neglected for any of the investigated
substrates, the situation is complicated for Co adatoms, where for some
substrates Eqs.~(\ref{eq-sum})--(\ref{eq-magic}) are valid while for
others they are not. This outcome is reinforced by inspection of the
validity of the approximate relation (\ref{eq-tzspin}) for \mbox{$T_{\alpha}$}.
Finally, we investigate the density of states (DOS) to get an
understanding of why for adatoms on some substrates
Eqs.~(\ref{eq-sum})--(\ref{eq-magic}) are valid while for adatoms on
other substrates they are not.
\section{Computational scheme}
\label{sec-comput}
The calculations were performed within the ab-initio spin density
functional theory framework, relying on the local spin density
approximation (LSDA) with the Vosko, Wilk and Nusair parameterization
for the exchange and correlation potential \cite{VWN80}. The
electronic structure is described, including all relativistic effects,
by the Dirac equation, which is solved using the spin polarized
relativistic multiple-scattering or Korringa-Kohn-Rostoker (KKR) Green
function formalism \cite{EKM11} as implemented in the {\sc spr-tb-kkr}
code \cite{tbkkr-code}. The potentials were treated within the atomic
sphere approximation (ASA) and for the multipole expansion of the
Green function, an angular momentum cutoff
\mm{\ell_{\mathrm{max}}}=3 was used. The energy integrals were
evaluated by contour integration on a semicircular path within the
complex energy plane using a logarithmic mesh of 32 points. The
integration over the $\mathbf{k}$ points was done on a regular mesh,
using 10000 points in the full surface Brillouin zone.
The electronic structure of Co monolayers on noble metals surfaces was
calculated by means of the tight-binding or screened KKR method
\cite{ZDU+95}. The substrate was modeled by a slab of 16 layers, the
vacuum was represented by 4 layers of empty sites. The adatoms were
treated as embedded impurities: first the electronic structure of the
host system (clean surface) was calculated and then a Dyson equation
for an embedded impurity cluster was solved \cite{BMP+05}. The
impurity cluster contains 131 sites; this includes one Co atom, 70
substrate atoms and the rest are empty sites. This cluster defines
the zone in which the electrons are allowed to react to the presence
of the adatom; there is an unperturbed host beyond this zone.
We investigate a series of Co adatoms and Co monolayers on the (111)
surface of the noble metals Cu, Ag, Au, Pd, Pt. In this way we include
in our study substrates which are hard to magnetically polarize (Cu,
Ag, Au) and
substrates that are easy to polarize (Pd, Pt), as well as substrates
with weak SOC (Cu), with moderate SOC (Pd, Ag), and with strong SOC
(Pt, Au). We assume that all atoms are located on ideal lattice sites
of the underlying bulk fcc lattice; no structural optimization was
attempted. While this would affect comparison of our data with
experiment, we do not expect this to have a significant influence on
the conclusions.
\section{Results} \label{sec-res}
\subsection{Sum over magnetic dipole term components \mbox{$T_{\alpha}$}}
\label{sec-tzsum}
\begin{table}
\caption{Sum $7(T_{x} + T_{y} + T_{z})$ devided by \mbox{$\mu_{\mathrm{spin}}$}\ for Co
monolayers and Co adatoms on noble metals surfaces.}
\label{tab-aver}
\begin{ruledtabular}
\begin{tabular}{ldd}
\multicolumn{1}{c}{substrate} &
\multicolumn{1}{c}{monolayer} &
\multicolumn{1}{c}{adatom} \\
\hline
Cu & 0.011 & 0.206 \\
Pd & 0.015 & 0.072 \\
Ag & 0.021 & 0.372 \\
Pt & 0.008 & 0.098 \\
Au & 0.009 & 0.284
\end{tabular}
\end{ruledtabular}
\end{table}
The first test of the influence of SOC on the \mbox{$T_{\alpha}$}\ term is checking
the validity of Eq.~(\ref{eq-sum}). Our motivation comes from the spin
moment sum rule Eq.~(\ref{eq-spin}), in which \mbox{$\mu_{\mathrm{spin}}$}\ appears only in
combination with $7T_{\alpha}$, as $\mu_{\text{spin}}+7T_{\alpha}$.
The key indicator is thus the ratio $7T_{\alpha}/\mu_{\text{spin}}$.
Tab.~\ref{tab-aver} shows this ratio summed over all three
coordinates, $\sum_{\alpha=x,y,z}7T_{\alpha}/\mu_{\text{spin}}$. It
should be zero if the influence of SOC on \mbox{$T_{\alpha}$}\ can be neglected.
One can see that for Co monolayers the condition (\ref{eq-sum}) is
fulfilled with a high accuracy. However, the situation changes for Co
adatoms. It is obvious from Tab.~\ref{tab-aver} that the ratio
$\sum_{\alpha}7T_{\alpha}/\mu_{\text{spin}}$ is significantly larger
for adatoms than for the corresponding monolayers. For Pd and Pt
substrates the breakdown of Eq.~(\ref{eq-sum}) is modest. However,
for Cu, Ag, and Au substrates this breakdown is substantial.
The breakdown of Eq.~(\ref{eq-sum}) for adatoms is not related to any
specific choice of the coordinate system. Similar numbers as those
shown in Tab.~\ref{tab-aver} are obtained if the sum over three
coordinate axes is substituted by a corresponding integral over the
full space angle (cf.\ also Fig.~\ref{fig-tz-theta} below).
It should be also noted that the dependence of the spin moment alone
on the magnetization direction is
negligible: the variations do not exceed 0.03~\% for monolayers and
0.4~\% for adatoms.
\subsection{Angular dependence of magnetic dipole term}
\label{sec-tzmagic}
\begin{figure}
\includegraphics[viewport=0.6cm 0.2cm 9.2cm 16.5cm]{fig1-tz-theta-fits.eps}%
\caption{(Color online) Dependence of the magnetic dipole term
\mbox{$T_{\theta}$}\ on the magnetization angle $\theta$ for Co monolayers (left
panels) and Co adatoms (right panels) on different substrates.
Ab-initio results are shown by red marks, fits to
$A(3\cos^{2}\theta-1+B)$ are shown by full blue lines, fits to
$A(3\cos^{2}\theta-1)$ are shown by dashed green lines. Both fits
are practically undistinguishable except for the cases of adatoms on
Cu, Ag, or Au.}
\label{fig-tz-theta}
\end{figure}
Another view on the same problem can be obtained by inspecting the
angular dependence of the magnetic dipole term \mbox{$T_{\theta}$}.
Fig.~\ref{fig-tz-theta} shows the \mbox{$T_{\theta}$}\ term calculated while varying
the angle $\theta$ between the magnetization direction and the surface
normal. The azimuthal angle $\phi$ was kept at \ang{0}, with the $x$
axis parallel to the [10$\bar{1}$] direction. If the influence of SOC
can be neglected, the \mbox{$T_{\theta}$}\ dependence should satisfy
Eq.~(\ref{eq-magic}). Therefore, we tried to fit our ab-initio data
to the expression
\[
A \, (3\cos^{2}\theta-1)
\]
(dashed green lines in Fig.~\ref{fig-tz-theta}). This fit is
quite accurate except for Co adatoms on Cu, Ag, and Au. In these
cases the \mbox{$T_{\theta}$}\ dependence can be fitted with the function
\[
A\, (3\cos^{2}\theta-1+B)
\]
(full blue lines in Fig.~\ref{fig-tz-theta}).
The fact that the \mbox{$T_{\theta}$}\ dependence can be fitted by
Eq.~(\ref{eq-magic}) only if a rigid shift (represented by the
constant $B$) is introduced presents another evidence that the
magnetic dipole term sum rule (\ref{eq-sum}) is not universally valid
for supported 3$d$ systems. Likewise, \mbox{$T_{\theta}$}\ does not vanish at the
magic angle 54.7$^{\circ}$ for systems where $B$ is important.
Rather, it vanishes for a magnetization tilt angle of \ang{45} for a
Co adatom on Cu, \ang{13} for an adatom on Ag, and \ang{42} for an
adatom on Au.
\subsection{Approximate relation for \mbox{$T_{\alpha}$}\ in terms of
$\mu_{\text{spin}}^{(m)}$}
\label{sec-mspin}
\begin{table}
\caption{Magnetic dipole term for $\bm{M}\|x$ ($T_{x}$) and
$\bm{M}\|z$ ($T_{z}$) evaluated using the exact expression
(\ref{eq-exact}) and using the approximate relation
(\ref{eq-tzspin}).}
\label{tab-tz-approx}
\begin{ruledtabular}
\begin{tabular}{lcdddd}
& &
\multicolumn{2}{c}{Co monolayer} &
\multicolumn{2}{c}{Co adatom} \\
\multicolumn{2}{l}{substrate} &
\multicolumn{1}{c}{exact} &
\multicolumn{1}{c}{approx.} &
\multicolumn{1}{c}{exact} &
\multicolumn{1}{c}{approx.} \\
\hline
Cu & $T_{x}$ & 0.020 & 0.021 & 0.057 & 0.031 \\
& $T_{z}$ & -0.037 & -0.042 & -0.052 & -0.061 \\ [0.7ex]
Pd & $T_{x}$ & 0.028 & 0.027 & 0.099 & 0.093 \\
& $T_{z}$ & -0.051 & -0.055 & -0.173 & -0.187 \\ [0.7ex]
Ag & $T_{x}$ & 0.025 & 0.024 & 0.059 & 0.008 \\
& $T_{z}$ & -0.043 & -0.048 & -0.004 & -0.016 \\ [0.7ex]
Pt & $T_{x}$ & 0.028 & 0.028 & 0.109 & 0.098 \\
& $T_{z}$ & -0.053 & -0.055 & -0.184 & -0.196 \\ [0.7ex]
Au & $T_{x}$ & 0.032 & 0.032 & 0.080 & 0.040 \\
& $T_{z}$ & -0.061 & -0.064 & -0.066 & -0.079
\end{tabular}
\end{ruledtabular}
\end{table}
Getting an intuitive insight into the \mbox{$T_{\alpha}$}\ term by relying on the
exact Eq.~(\ref{eq-exact}) is not easy. The approximate
Eq.~(\ref{eq-tzspin}) is far better suited for this purpose. It
presents \mbox{$T_{\alpha}$}\ as a linear combination of orbitally-projected
components of the spin magnetic moment $\mu_{\text{spin}}^{(m)}$,
illustrating thus the frequently used interpretation of the magnetic
dipole term as manifestation of the anisotropy of spin density
distribution. Indeed, if all $m$-components of \mbox{$\mu_{\mathrm{spin}}$}\ are identical,
\mbox{$T_{\alpha}$}\ is zero.
However, this view is transparent only if the effect of SOC on
\mbox{$T_{\alpha}$}\ can be neglected. Therefore we present in
Tab.~\ref{tab-tz-approx} a comparison between values of \mbox{$T_{\alpha}$}\ obtained
by evaluating the exact Eq.~(\ref{eq-exact}) and by evaluating the
approximate Eq.~(\ref{eq-tzspin}). We focus on two magnetization
directions, $\bm{M}\|x$ and $\bm{M}\|z$. One can see that as concerns
Co monolayers, the approximate equation yields similar values as the
exact equation. For Co adatoms, the agreement is worse and, again, it
depends on the substrate. For adatoms on Pd and Pt, the validity of
the approximate equation is worse than for corresponding monolayers
but it is still acceptable. However, for adatoms on Cu, Ag, and Au
the error of the approximate Eq.~(\ref{eq-tzspin}) reaches 50--100~\%.
Comparison of the exact and approximate values of $T_{x}$ and $T_{z}$
in Tab.~\ref{tab-tz-approx} can serve as another indicator of the role
of SOC for the magnetic dipole term. The outcome of this analysis is
consistent with the conclusions based on inspection of
Eq.~(\ref{eq-sum}) in Sec.~\ref{sec-tzsum} and Eq.~(\ref{eq-magic}) in
Sec.~\ref{sec-tzmagic}. Namely, the influence of the SOC on the
\mbox{$T_{\alpha}$}\ term can be neglected for monolayers on any substrate and for
adatoms on Pd and Pt, while it has to be taken into account when
dealing with \mbox{$T_{\alpha}$}\ for adatoms on Cu, Ag, and Au.
\subsection{Density of states} \label{sec-dos}
\subsubsection{Total spin-polarized DOS} \label{sec-totdos}
\begin{figure*}
\includegraphics[viewport=0.2cm 0.2cm 17.5cm 13.5cm]{fig2-DOS-overall.eps}%
\caption{(Color online) Spin-polarized DOS for Co adatoms (upper
panels) and monolayers (lower panels) on noble metals. Blue solid
lines show the DOS for Co atoms (in states per eV), dashed brown
lines show the DOS for those substrate atoms which are nearest
neighbors to Co atoms.}
\label{fig-DOS-overall}
\end{figure*}
To summarize, we found two trends concerning the impact of SOC on \mbox{$T_{\alpha}$}.
First, the dimensionality or perhaps better the size of the system is
crucial: the effect of SOC can be always neglected for monolayers but
only sometimes for adatoms. Second, there is a big variance depending
on the substrate but the nominal strength of the substrate SOC does
not seem to be important.
Thinking about the explanation, one should recall that the SOC
strength $\xi$ should be compared to the crystal field splitting
$\Delta_{\text{CF}}$ \cite{SK95a,SMP+16} --- it is the
\mbox{$\xi/\Delta_{\text{CF}}$} ratio that matters. The splitting
$\Delta_{\text{CF}}$ is a model Hamiltonian parameter that is not
directly accessible by LDA calculations. It can be seen as a measure
how electronic states around an atom are affected by the crystal field
due to its neighbors. An idea how the influence of the crystal field
varies across our systems can be obtained by inspecting the DOS.
Therefore we present in Fig.~\ref{fig-DOS-overall} the spin-polarized
DOS for all the systems we investigate. Apart from the DOS for Co
atoms we show also the DOS for the nearest substrate atoms, so that
hybridization between them can be studied.
One can see that (not surprisingly) the bandwidth for monolayers is
always significantly larger than the bandwidth for adatoms, no matter
what is the substrate. This clarifies why the influence of SOC on the
\mbox{$T_{\alpha}$}\ term is negligible for the monolayers: in that case, the effect
of the crystal field always overwhelms the effect of SOC.
\subsubsection{DOS overlap integrals}
\begin{table}
\caption{Comparing the importance of SOC for \mbox{$T_{\alpha}$}\ (characterized by
sums over three $T_{\alpha}$ components, the second column) to the
degree of atomic-like character of states associated with the adatom
(characterized by reciprocal values of the DOS overlap integrals,
the third colum).}
\label{tab-overlap}
\begin{ruledtabular}
\begin{tabular}{ldd}
&
\multicolumn{1}{c}{relative weight of} &
\multicolumn{1}{c}{relative weight of} \\
&
\multicolumn{1}{c}{$\sum_{\alpha}7T_{\alpha}/\mu_{\text{spin}}$} &
\multicolumn{1}{c}{$1 / \int \! \mathrm{d} E \,
n^{\downarrow}_{\text{Co}}(E) \, n^{\downarrow}_{\text{subs}}(E)$} \\
\hline
Cu & 0.181 & 0.197 \\
Pd & 0.061 & 0.091 \\
Ag & 0.390 & 0.324 \\
Pt & 0.092 & 0.117 \\
Au & 0.276 & 0.269 \\
\end{tabular}
\end{ruledtabular}
\end{table}
What is not clear is why there are so big differences for the adatoms
when going from one substrate to another. The bandwidth is
approximately the same for all substrates. One can, nevertheless,
quantify the importance of hybridization between the adatom and the
substrate by evaluating the DOS overlap integral, i.e., the integral
of the product of the DOS for the adatom $n^{(s)}_{\text{Co}}$ and for
the nearest substrate atom $n^{(s)}_{\text{subs}}$,
\begin{equation}
h^{(s)} \: \equiv \: \int \! \mathrm{d} E \,
n^{(s)}_{\text{Co}}(E) \, n^{(s)}_{\text{subs}}(E)
\; ,
\label{eq-dosint}
\end{equation}
where $s$ stands for the spin. Here an interesting relation between
the integrals $h^{(s)}$ and the importance of the SOC for \mbox{$T_{\alpha}$}\ appears
if we focus on the minority-spin states ($s=\downarrow$). Namely, the
relative importance of SOC for \mbox{$T_{\alpha}$}\ (quantified as the ratio
$\sum_{\alpha}7T_{\alpha}/\mu_{\text{spin}}$, cf.~Sec.~\ref{sec-tzsum}
and Tab.~\ref{tab-aver}) can be linked to the degree of atomic-like
character of minority-spin adatom states (quantified as
$1/h^{(\downarrow)}$). This emerges from Tab.~\ref{tab-overlap} where
relative weights of both quantities are shown. It follows from
Tab.~\ref{tab-overlap} that if adatom states are less hybridized with
the substrate, indicating thus that the crystal field splitting is
smaller, the importance of SOC increases --- in agreement with
intuition.
The only caveat here is that this correspondence holds only for
minority-spin states; if majority-spin states are included in the
analysis, the correspondence between
$\sum_{\alpha}7T_{\alpha}/\mu_{\text{spin}}$ and $1/h^{(s)}$
disappears. However, there is a reason for focusing on minority-spin
states only. If majority-spin states are mostly occupied (as it is
the case for our systems), it is the incomplete occupancy of
minority-spin states which induces asphericity. The importance of
partially-filled minority-spin states for \mbox{$T_{\alpha}$}\ is also emphasized by
the fact that the value of \mbox{$T_{\alpha}$}\ strongly depends on the position of
\mbox{$E_{F}$}, which cuts through minority-spin states
\cite{WF94,KEDF02,SME09b}.
\subsubsection{Orbitally-resolved DOS for Co adatoms}
\begin{figure*}
\includegraphics[viewport=0.2cm 0.2cm 17.5cm 9.5cm]{fig3-DOS-orbital-detail.eps}%
\caption{(Color online) Orbitally-resolved DOS (in states per eV) for
Co adatoms on noble metals surfaces. Majority-spin $d$ states are
shown in the upper panels, minority-spin $d$ states are shown in the
lower panels. Meaning of the lines is shown in the two upper left
panels.}
\label{fig-mDOS}
\end{figure*}
Yet another view on hybridization of adatom states with substrate
states can be obtained from the orbitally-resolved DOS for the Co adatom.
This is shown in Fig.~\ref{fig-mDOS}: majority-spin states are
inspected in the upper panels, minority-spin states in the lower
panels. Several features in this plot are worth commenting. First,
the
individual orbital-resolved peaks are broader for the majority-spin states
than for this minority-spin states. This is because for the majority-spin
states there is a considerable overlap with substrate states while for
the minority-spin states there is practically no overlap (see
Fig.~\ref{fig-DOS-overall}). The Ag substrate with a deep lying $d$
band is an exception, in this
case the majority-spin adatom states have no overlap with substrate
states either.
The DOS peaks for $m=\pm2$ resemble broadened energy levels, as for an
isolated atom. This is because the orbital lobes for $m=\pm2$ lie
parallel to the surface, where there are no other atoms to hybridize
with. The situation for $m=0$ is similar --- here the orbital lobe
points to the void between three nearest substrate atoms. The
influence of the substrate is most pronounced for the $m=\pm1$
orbitals whose lobes are directed toward neighboring atoms. Apart
from that, states for $m=\pm1$ and $m=\pm2$ are split by the SOC.
A more formal discussion about resolving the DOS according to spin and
orbital quantum numbers as well as about the role of the SOC-induced
splitting for the magnetocrystalline anisotropy was recently presented
by \v{S}ipr {\it et al.}~\cite{SMP+16}.
Here our attention is on the hybridization and, in particular, on the
difference between Cu, Ag, and Au substrates on the one hand and Pd
and Pt substrates on the other hand. This difference is apparent for
the minority-spin states with $m=\pm1$ (lower panels of
Fig.~\ref{fig-mDOS}): while there is only one single peak for each of
the $m=\pm1$ components for the Cu, Ag, and Au substrates, there are
two peaks for the Pd and Pt substrates. We can infer from this that
the crystal field splitting $\Delta_{\text{CF}}$ is small in the case
of a Co adatom on Cu, Ag, and Au, enabling thus the SOC to have a
large role for the \mbox{$T_{\alpha}$}\ term, while it is large in the case of a Co
adatom on Pd and Pt, suppressing thus the role of SOC for \mbox{$T_{\alpha}$}. Anc
analysis of the orbital-resolved DOS thus reinforces the message
obtained by analyzing the overlap integrals (\ref{eq-dosint}) in
Tab.~\ref{tab-overlap}.
\section{Discussion} \label{sec-discuss}
The purpose of this work was to study systematically the conditions
under which the influence of SOC on the \mbox{$T_{\alpha}$}\ term can or cannot be
neglected for 3$d$ systems and, in this way, to explore possibilities
to eliminate the \mbox{$T_{\alpha}$}\ term from the spin moment sum rule
(\ref{eq-spin}). We found that even for atoms with low SOC such as Co,
the influence of SOC on \mbox{$T_{\alpha}$}\ in certain environments can be so large
that Eqs.~(\ref{eq-sum})--(\ref{eq-magic}) cannot be used. The
crucial factor turns out to be the ratio between SOC and crystal
field splitting, $\xi/\Delta_{\text{CF}}$. This subsequently
translates itself into the dependence on the dimensionality. It turns
out that for Co monolayers the influence of SOC on \mbox{$T_{\alpha}$}\ can be
neglected for any of the Cu, Pd, Ag, Pt, or Au substrates. We assume
that this is true for any 3$d$ monolayer on any substrate.
For adatoms the situation is more complicated. The decrease of
$\Delta_{\text{CF}}$ caused by the decrease of the dimensionality
appears to be just of that size which is required for SOC to become
important for \mbox{$T_{\alpha}$}. Hence details of the electronic structure of the
substrate begin to matter; for some substrates (Pd, Pt)
Eqs.~(\ref{eq-sum})--(\ref{eq-magic}) still can be used while for
others (Cu, Ag, Au) they cannot. The hybridization between adatom and
substrate states around \mbox{$E_{F}$}\ seems to be the deciding factor. We
expect that for systems with considerable overlap between adatom and
substrate DOS around \mbox{$E_{F}$}\ (minority-spin states in our case, see
Fig.~\ref{fig-DOS-overall}) the influence of SOC on \mbox{$T_{\alpha}$}\ can be
neglected even for adatoms. Otherwise
Eqs.~(\ref{eq-sum})--(\ref{eq-magic}) should rather not be used.
To find more about when the size of the system gets so small that
Eqs.~(\ref{eq-sum})--(\ref{eq-magic}) cannot be used any more, we
performed calculations also for a Co wire on Au(111). The wire was
built along the [1$\bar{1}$0] direction, we modelled it by a
2$\times1$ surface supercell. To test whether Eq.~(\ref{eq-sum})
could be applied for such system, we evaluated the ratio
$\sum_{\alpha}7T_{\alpha}/\mu_{\text{spin}}$ and found it to be
0.058 \footnote{Our finding that SOC is not very important for
\mbox{$T_{\alpha}$}\ for a Co wire on Au(111) is not in contradiction with results
of Ederer {\it et al.}\ \cite{EKDF03} because they studied a {\em
free-standing} Co wire and, additionally, employed the Brooks
orbital polarization correction.}. This is to be compared with 0.284
for a monolayer and 0.009 for an adatom (see Tab.~\ref{tab-aver}). We
conclude, therefore, that the borderline between systems which satisfy
Eqs.~(\ref{eq-sum})--(\ref{eq-magic}) and which do not is somewhere
between the wire and the adatom. When analyzing XMCD spectra for
small 3$d$ clusters of just few atoms, one should not rely on
Eqs.~(\ref{eq-sum})--(\ref{eq-magic}). When analyzing XMCD spectra of
clusters of hundreds of atoms (as was the case, e.g., in the study of
Koide {\it et al.}\ \cite{KMO+01}), reliance on
Eqs.~(\ref{eq-sum})--(\ref{eq-magic}) is justified.
\section{Conclusions} \label{sec-zaver}
The influence of spin-orbit coupling on the magnetic dipole term
\mbox{$T_{\alpha}$}\ can be neglected for 3$d$ transition metal systems as long as
they are sufficiently large. If the system contains just a few 3$d$
atoms (as is the case of adatoms or small supported clusters), the
influence of SOC on \mbox{$T_{\alpha}$}\ may be significant. This further depends on
the hybridization between states of the 3$d$ atoms and of the
substrate, especially around the Fermi level: if the hybridization is
only weak, the role of the SOC is enhanced while if the hybridization
is strong, the role of the SOC is suppressed. For systems where the
influence of SOC on \mbox{$T_{\alpha}$}\ cannot be neglected, the \mbox{$T_{\alpha}$}\ term cannot be
eliminated from the XMCD spin sum rule --- neither by relying on the
$T_{x}+T_{y}+T_{z}=0$ relation, nor by making use of the magic angle
$\theta=54.7^{\circ}$.
\begin{acknowledgments}
This work was supported by the Ministry of Education, Youth and Sport
(Czech Republic) within the projects LD15097 (O.\v{S}.) and LO1402
(J.M.) and by the Deutsche Forschungsgemeinschaft within the project
SFB 689 ``Spinph\"{a}nomene in reduzierten Dimensionen''.
\end{acknowledgments}
|
\section{Mechanism}
This letter describes a new mechanism, dubbed ``$N$naturalness,'' which solves the hierarchy problem. It predicts no new particles at the LHC, but does yield a variety of experimental signatures for the next generation of CMB and large scale structure experiments~\cite{Dodelson:2013pln, Allison:2015qca}. Well-motivated supersymmetric incarnations of this model predict superpartners beneath the scale $m_W \times M_{\text{pl}}/M_{\text{GUT}} \sim 10$~TeV, accessible to a future 100 TeV collider~\cite{Arkani-Hamed:2015vfh, Golling:2016gvc}.
The first step is to introduce $N$ sectors which are mutually non-interacting. The detailed particle content of these sectors is unimportant, with the exception that the Standard Model (SM) should not be atypical; many sectors should contain scalars, chiral fermions, unbroken gauge groups, etc. For simplicity, we imagine that they are exact copies of the SM, with the same gauge and Yukawa structure.
It is crucial that the Higgs mass parameters are allowed to take values distributed between $-\Lambda_H^2$ and $\Lambda_H^2$, where $\Lambda_H$ is the (common) scale that cuts off the quadratic divergences. Then for a wide range of distributions, the generic expectation is that some sectors are accidentally tuned at the $1/N$ level, $\left|m_H^2\right|_{\text{min}} \sim \Lambda_H^2 / N$. We identify the sector with the smallest non-zero Higgs vacuum expectation value (vev), $\vev{H} = v$, as ``our" SM. This picture is illustrated schematically in Fig.~\ref{fig:Sketch}.
\begin{figure}[!t]
\centering
\includegraphics[width=0.48\textwidth]{Figures/N_Copies_Sketch.pdf}
\caption{A sketch of the $N$naturalness setup. The sectors have been ordered so that they range from $m_H^2 \sim \Lambda_H^2$ to $-\Lambda_H^2$. The sector with the smallest vacuum expectation value contains our copy of the SM.}
\label{fig:Sketch}
\end{figure}
In order for small values of $m_H^2$ to be populated, the distribution of the mass parameters must pass through zero. For concreteness, we take a simple uniform distribution of mass squared parameters, indexed by an integer label $i$ such that
\begin{equation}
\left(m_H^2\right)_i = -\frac{\Lambda_H^2}{N} \big(2\, i+ r \big),\quad\quad -\frac{N}{2} \leq i \leq \frac{N}{2},
\label{eq:mHsqiScaling}
\end{equation}
where $i=0 = \text{``us"}$ is the lightest sector with a non-zero vev: $\left(m_H^2\right)_\text{us} = -r\times\Lambda_H^2/N \simeq -(88 \text{ GeV})^2$ is the Higgs mass parameter inferred from observations. The parameter $r$ can be seen as a proxy for fine-tuning,\footnote{There are a variety of other ways one might choose to implement a measure of fine-tuning in this model. For example, one could assume the distribution of Higgs mass squared parameters is random with some (arbitrary) prior, and then ask statistical questions regarding how often the resulting theory is compatible with observations.} since it provides a way to explore how well the naive relation between the cutoff and the mass scale of our sector works in a detailed analysis. Specifically, $r = 1$ corresponds to uniform spacing, while $r<1$ models to an accidentally larger splitting between our sector and the next one. A simple physical picture for this setup is that the new sectors are localized to branes which are displaced from one another in an extra dimension. In this scenario, the lack of direct coupling is clear, and the variation of the mass parameters can be explained geometrically: the $m_H^2$ parameters may be controlled by the profile of a quasi-localized field shining into the bulk.
As a consequence of the existence of a large number of degrees of freedom, the hierarchy between $\Lambda_H$ and the scale $\Lambda_G$ where gravity becomes strongly coupled is reduced. The renormalization of the Newton constant implies $\Lambda_G^2 \sim M_{\text{pl}}^2 / N$. If perturbative gauge coupling unification is to be preserved $\Lambda_G \gtrsim M_{\text{GUT}}$, implying that $N \lesssim 10^4$. This gives a cutoff no greater than $\Lambda_H \sim 10$~TeV, thus predicting a little hierarchy that mirrors the GUT-Planck splitting in the UV. At the scale $\Lambda_H$, new dynamics (\emph{e.g.}, SUSY) must appear to keep the Higgs from experiencing sensitivity to even higher scales. Alternatively, the full hierarchy problem can be solved with $N \sim 10^{16}$, so that $\Lambda_H \sim \Lambda_G \sim 10^{10}$~GeV. Note that this number of copies, while sufficient, is unnecessary for a complete solution. There may be two classes of new degrees of freedom: the $N$ copies that participate directly in the $N$naturalness picture, and another completely sterile set of degrees of freedom that still impact the renormalization of $\Lambda_G$.
So far we have described a theory with a $S_N$ permutation symmetry, broken softly by the $m_H^2$ parameters, such that each of the sectors is SM-like.
Sectors for which $m_H^2 < 0$ are similar to our own, with the exception that particle masses scale with the Higgs vev, $v_i \sim v\,\sqrt{i}$. In addition, once $i \gtrsim 10^{8}$ the quarks are all heavier than their respective QCD scales. Those sectors do not exhibit chiral symmetry breaking, nor do they contain baryons. Sectors with $m_H^2>0$ are dramatically different from ours. In these sectors, electroweak symmetry is broken at low scales due to the QCD condensate $\Lambda_\text{QCD}$. Fermion masses are generated by the four-fermion interactions that are induced by integrating out the complete $SU(2)$ Higgs multiplet. Thus, $m_f \sim y_f\,y_t\,\Lambda_{\text{QCD}}^3 / \left(m_{H}^2\right)_i \lesssim 100$~eV, where $y_t$ is the top Yukawa coupling. All fermionic and gauge degrees of freedom are extremely light relative to the ones in our sector.
With so many additional degrees of freedom, the naive cosmological history is dramatically excluded. In particular, if all sectors have comparable temperatures in the early Universe, then one expects $\Delta N_{\text{eff}} \sim N$ (see Eq.~(\ref{eq:dNeff})). Thus, the hierarchy problem gets transmuted into the question of how to predominantly reheat only those sectors with a tuned Higgs mass.
To accomplish this, we need to introduce a last ingredient into the story, the ``reheaton" field, so named because it is responsible for reheating the Universe via its decays. We call this field $S^c$ for models where the reheaton is a fermion, and $\phi$ if the reheaton is a scalar. The cosmological history of the model begins in a post-inflationary phase where the energy density of the Universe is dominated by the reheaton. As stated multiple times we can not be unique, therefore we assume that the reheaton couples universally to all sectors. Note that the scalars must be near their true minimum when reheating occurs. This can be accomplished by having either low scale inflation, or else a coupling of the Higgses to the Ricci scalar.
In the next section, we present a set of models in which the reheaton dynamically selects and populates only the lightest sectors, despite preserving the aforementioned softly broken $S_N$ symmetry. Sec.~\ref{sec:Signals} then provides constraints on these models, and Sec.~\ref{sec:Discussion} contains our conclusions and highlights potential signals.
\section{Models}
\label{Models}
We have argued that the hierarchy problem can be solved by invoking a large number of copies of the SM, along with some dynamical mechanism which dominantly populates the lightest sector with a non-zero Higgs vev. This section details some simple explicit models that realize a viable cosmological history.
As anticipated in the previous section, we imagine that at a post-inflationary stage the energy density of the Universe is dominated by a reheaton that couples universally to all the new sectors. Its decays populate the SM and its copies. The goal is to deposit as much energy as possible into the sector with the smallest Higgs vev. This may be accomplished by arranging the decays of the reheaton such that the branching fraction into the $i^\text{th}$ sector scales as $\text{BR}_i \sim \left(m_H\right)_i^{-\alpha}$ for some positive exponent $\alpha$. To this end, we construct models that share three features:
\vspace{2pt}
\begin{enumerate}[label=(\roman*)]
\item The reheaton is a gauge singlet;
\vspace{3pt}
\item It is parametrically lighter than the naturalness cutoff, $m_\text{reheaton}\lesssim \Lambda_H/\sqrt{N}$;
\vspace{3pt}
\item Its couplings are the most relevant ones possible that involve the Higgs boson of each sector.
\end{enumerate}
While the requirement of a light reheaton field may appear to require an additional coincidence, it can be easily accommodated in an extra-dimensional picture. In order to couple to all the sectors, the reheaton must be a bulk field. Then, before canonical normalization, its kinetic term carries a factor of $N$. If the reheaton enjoys a shift symmetry that is respected in the bulk, it will receive a $\Lambda_H$-sized mass from each brane on which the shift symmetry is violated. Here we assume that the dynamics above $\Lambda_H$ respect the shift symmetry. As long as the shift symmetry is only violated on the boundaries, the reheaton mass will be parametrically the same as the weak scale after canonical normalization. In the case of a fermionic reheaton, this simple picture corresponds to the brane-localization of its Dirac partner.
The two simplest models, which we denote $\ell$ and $\phi$, are
\begin{equation}
\mathcal{L}_\ell \supset -\lambda\,S^c\,\sum_i \ell_i\,H_i - m_S\,S\, S^c\,,
\end{equation}
if the reheaton is a fermion $S^c$, and
\begin{align}
\mathcal{L}_\phi \supset -a \,\phi\,\sum_i \,|H_i|^2 -\frac{1}{2}\, m^2_\phi\,\phi^2,
\end{align}
if the reheaton is a scalar $\phi$. For the theory to be perturbative, we need the coupling $\lambda$ to obey a `t Hooft-like scaling $\lambda \sim 1/\sqrt{N}$. Naively we would expect the same scaling for $a$, but we find that a stronger condition needs to be imposed ($a \sim 1/N$) to insure that the loop induced mass for $\phi$ is not much larger than $\Lambda_H/\sqrt{N}$. Even with this scaling, the loop-induced tadpole for $\phi$ will be too large unless the sign of $a$ is taken to be arbitrary for each sector. Note that $a$ breaks a $\mathbb{Z}_2$ symmetry on $\phi$, so that this choice is consistent with technical naturalness. Including the arbitrary sign, the sum over tadpole contributions only grows as $\sqrt{N}$, and so the natural range of $\phi$ is restricted to $\Lambda_H \sqrt{N}$. The Higgses will then receive a contribution to their $m_H^2$ parameters of order $a \langle \phi \rangle \sim \Lambda_H^2 / \sqrt{N}$. While these contributions may be large compared to our weak scale, as long as they are smaller than $\mathcal{O}(\Lambda_H^2)$, they can be safely absorbed into the quadratically-divergent contributions to $m_H^2$. Of course, these are upper bounds on the couplings; as we will discuss later in the section, they can be consistently taken smaller, so long as the reheat temperature is sufficiently high.
Before moving on to discuss the details of reheating, we remark on the existence of cross-quartics of the form $\kappa\,|H_i|^2 \,|H_j|^2$. Even if these are absent in the UV theory, they will be induced radiatively. After electroweak symmetry breaking in the various sectors, these can potentially affect the spectrum, and so it is critical to the $N$naturalness mechanism that they be sufficiently suppressed. Given an arbitrary, $S_N$ symmetric cross-quartic, $\kappa$, the $m_H^2$ parameters will shift by approximately $ -\kappa\, \Lambda_H^2 \,N / 8 + \mathcal{O} (\kappa^2 N)$, while the mixing effects are subdominant. Thus, the general picture of hierarchical weak scales remains intact so long as $\kappa \lesssim 1 / N$.
At a minimum, cross-quartics of this form will be induced gravitationally, regardless of the reheaton dynamics. These quartically-divergent gravitational couplings arise at three loops, giving $(16 \,\pi^2)^3 \kappa_g \sim \lambda_h^2 (\Lambda_H / M_{\text{pl}})^4 \sim (\lambda_h / N)^2 (\Lambda_H / \Lambda_G)^4$, where $\lambda_H$ is the SM-like Higgs self quartic. Here we have taken the scale that cuts off these divergences to be $\Lambda_H$, as would be appropriate for a supersymmetric UV completion (for which these quartics are absent). In either case, these gravitational couplings are parametrically safe, since they scale as $(1 / N)^2$.
In addition, potentially dangerous cross-quartics can be generated by reheaton exchange. In the $\ell$ model, the cross-quartic is generated at one loop: $\kappa_{\ell} \sim \lambda^4 / 16 \,\pi^2 \lesssim 1 / N^2$, after enforcing the large-$N$ scaling of $\lambda$. In the $\phi$ model, these quartics are generated at tree-level, $\kappa_{\phi} \sim a^2 / m_{\phi}^2$. Naively this appears borderline problematic, since $\kappa_{\phi}$ scales as $1 / N$. However, the arbitrary sign of $a$, which was necessary to mitigate the tadpole of $\phi$, will once again soften the sum over sectors, so that $\sum a_i \,v_i^2 \sim a\, \Lambda_H^2\, \sqrt{N}$. Combined with the large-$N$ scaling of $a$, these quartics are rendered safely negligible.
\subsection{Reheating}
If the reheaton is sufficiently light, then we may analyze the leading reheaton decay operators using an effective Lagrangian computed by integrating out $H_i$. This immediately makes it clear why we we want the reheaton to be coupled with the most relevant coupling possible, since these will suffer the fastest suppression as $|m_H|\rightarrow \infty$. Integrating out the Higgs and gauge bosons in the $\ell$ model, the leading decays of $S^c$ are given by, \emph{e.g.}
\begin{equation}
\begin{array}{l}
\mathcal{L}^{\vev{H}\neq 0}_\ell \supset \mathcal{C}^\ell_1\, \lambda\,\frac{v}{m_Z^2 \,m_S}\nu^\dag \bar{\sigma}^\mu S^c \, f^\dagger \bar{\sigma}_\mu f \,;\\[10pt]
\mathcal{L}^{\vev{H}= 0}_\ell \supset \mathcal{C}^\ell_2\,\lambda\,\frac{y_t}{m_H^2}S\, \ell\, Q_3^\dag \,u^{c\dag}_3\,,
\end{array}
\label{Eq: fermion Dim 1}
\end{equation}
where $m_{Z}$ is the relevant $Z^0$-boson mass and the $\mathcal{C}^\ell_i$ are numerical coefficients. We have omitted decays through $W$ and Higgs bosons for sectors with $\vev{H} \neq 0$ as they scale in the same way. We include them in all numerical computations.
From this low energy Lagrangian we can easily infer that a light reheaton dominantly populates the lightest negative Higgs mass sector. Denoting with $m_{h_i}$ the physical Higgs mass in sectors with $\langle H\rangle \neq 0$, the reheaton decay widths scale as $\Gamma_{m_H^2<0} \sim 1/m_{h_i}^2$ and $\Gamma_{m_H^2>0} \sim 1/m_{H_i}^4$ in sectors with and without electroweak symmetry breaking, respectively. Thus the reheaton preferentially decays into sectors with light Higgs bosons and non-zero vevs. If, instead, the reheaton were heavy enough to decay directly to on-shell Higgs or gauge bosons, the branching fractions would be democratic into those sectors, and the energy density in our sector would not come to dominate the energy budget of the Universe.
In the scalar case the decays are different, but the scaling of the decay widths is exactly the same. This can be seen once more by integrating out the Higgs and gauge bosons in all the sectors:
\begin{equation}
\begin{array}{l}
\mathcal{L}^{\vev{H}\neq 0}_\phi \supset \mathcal{C}_1^\phi\, a\,y_q\frac{v}{m_h^2} \phi\,q\, q^c \,;\\[10pt]
\mathcal{L}^{\vev{H}= 0}_\phi \supset \mathcal{C}_3^\phi\,a\,\frac{g^2}{16\,\pi^2}\frac{1}{m_H^2}\,\phi\, W_{\mu\nu}W^{\mu\nu} \,,
\end{array}
\label{Eq: scalar Dim}
\end{equation}
where again the $\mathcal{C}^\phi_i$ are numerical coefficients, and $W_{\mu\nu}$ is the $SU(2)$ field strength. As in the fermionic case, this Lagrangian leads to decay widths that scale as $\Gamma_{m_H^2<0} \sim 1/m_{h_i}^2$ and $\Gamma_{m_H^2>0} \sim 1/m_{H_i}^4$ in sectors with and without electroweak symmetry breaking, respectively, through the diagrams shown in Fig.~\ref{fig:PhiDiagrams}. We have not included the one-loop decay $\phi \rightarrow \gamma \,\gamma$ in Eq.~\eref{Eq: scalar Dim} for sectors with $\vev{H}\neq0$. This operator scales as $1/m_h^2$ and is important for sectors with $N \gtrsim 10^8$; we find that this is never the leading decay once the bounds on $N$ discussed in Sec.~\ref{sec:Signals} are taken into account.
\begin{figure}[t!]
\centering
\includegraphics[width = .45\textwidth]{Figures/phi_Diagrams.pdf}
\caption{Feynman diagrams for the most important decays in the $\phi$ model. The left (right) column is for $\vev{H} \neq 0$ $\big(\!\vev{H} = 0\big)$. The top (bottom) row is for $m_\phi \gg |m_H|$ $\big(m_\phi \ll |m_H|\big)$. }
\label{fig:PhiDiagrams}
\end{figure}
Before moving to a more detailed discussion of signals and constraints it is worth pointing out two important differences between the $\phi$ and $\ell$ models that will lead us to modify the latter. Given the scaling of the widths we can approximately neglect the contributions to cosmological observables from the $\vev{H} = 0$ sectors. In the simple case that the vevs squared are equally spaced, $v^2_i \sim 2\, i\times v_\text{us}^2$, as in Eq.~(\ref{eq:mHsqiScaling}) with $r=1$, we find that the branching ratio into the other sectors is $\sum 1/i \sim \log N$.
In the $\phi$ model, this logarithmic sensitivity to $N$ is not realized. Since the reheaton decays into sectors with non-zero vevs via mixing with the Higgs, the decays become suppressed by smaller and smaller Yukawa couplings as $h_i$ becomes heavy. After the charm threshold is crossed $m_\phi < 2 \,m_{c_i}$ we can neglect the contribution of the new sectors to cosmological observables (with one exception that we discuss in the next section). This behavior is displayed in the left panel of Fig.~\ref{fig:TRH}, where we show the fraction of energy density deposited in each sector.
\begin{figure*}[t!]
\centering
\includegraphics[width = .47\textwidth]{Figures/T_phi.pdf} \quad\quad
\includegraphics[width = .47\textwidth]{Figures/T_L4.pdf}
\caption{Energy density deposited in each sector as a function of sector number, normalized to the energy density in our sector. The left panel is for the $\phi$ model with $a=1$~MeV. The right panel is for the $L_4$ model with $\lambda\times \mu_E=1$~MeV, $M_L=400$~GeV, $M_{E, N}=500$~GeV, $Y_E=Y_N=0.2$, and $Y^c_E=Y^c_N=-0.5$. The solid lines are the result of a full numerical calculation. The dashed lines show the expected scalings. As discussed in the text, the steps in the $\phi$ model are proportional to Yukawa couplings due to the fact that $\phi$ decays via mixing with the Higgs. When $i \gtrsim 10^9$ in the $L_4$ model, the process $S^c \rightarrow 2\,e+\nu$ cannot proceed on-shell, which results in the deviation from the naive scaling as denoted by $m_S = 2\,m_e + m_\nu$. Both figures were made using the zero temperature branching ratios of the reheaton; thermal corrections are under control so long as $T_\text{RH}$ is smaller than the weak scale in our sector, as discussed at the end of Sec.~\ref{Models}.}
\label{fig:TRH}
\end{figure*}
The second important difference is that in the $\ell$ model the reheaton couples directly to neutrinos and, in the sectors with electroweak symmetry breaking, it mixes with them. This leads to two effects. First, the physical reheaton mass grows with $N$, implying that the structure of the $\ell$ model forces the reheaton to be heavy at large $N$, and can be inconsistent depending on the value of $\lambda$. Additionally, this mixing can generate a freeze-in abundance~\cite{Hall:2009bx} of neutrinos in the other sectors from the process $\nu_\text{us}\,\nu_\text{us} \rightarrow \nu_\text{us}\,\nu_i$ via an off-shell $Z^0$. Tension with neutrino overclosure and overproduction of hot dark matter leads to an upper bound on the maximum number of sectors. In practice, it is hard to go beyond $N\simeq 10^3$.
However, there is a simple extension of the $\ell$ model that at once mitigates its UV, \emph{i.e.}, large $N$, sensitivity and solves the problems arising from a direct coupling to neutrinos. If the reheaton couples to each sector only through a massive portal (whose mass grows with $v_i$), then the branching ratios will scale with a higher power of the Higgs vev after integrating out the portal states. As an example, consider introducing a 4$^\text{th}$ generation of vector-like leptons $(L_4, L^c_4)$, $(E_4, E^c_4)$, and $(N_4, N^c_4)$ to each sector. Then relying on softly broken $U(1)$ symmetries, we can couple the reheaton to $L_4$ only via the Lagrangian
\begin{align}
&\mathcal{L}_{L_4} \supset \mathcal{L}_{\rm mix} +\mathcal{L}_{Y} + \mathcal{L}_{M}\, , \\
&\mathcal{L}_{\rm mix} = -\lambda\,S^c\, \sum_i \big(L_{4}\,H\big)_i - \mu_E\, \sum_i \big(e^c\, E_4\big)_i\; , \nonumber \\
&\mathcal{L}_{Y} = -\sum_i\Big[Y_E \big(H^\dag\,L_4\,E^c_4\big)_i + Y_E^c \big(H\,L^c_4\,E_4\big)_i \nonumber\\
&\qquad\qquad\qquad +Y_N \big(H\,L_4\,N_4^c\big)_i + Y_N^c \big(H^\dag\,L_4^c\,N_4\big)_i \Big]\, , \nonumber \\
&\mathcal{L}_{M} = -\sum_i\Big[M_E \big(E_4^c\,E_4\big)_i + M_L \big(L_4^c\, L_4\big)_i \nonumber \\
&\qquad\qquad\qquad+ M_N\big(N_4^c\,N_4\big)_i\Big] - m_S\, S \,S^c\,,\nonumber
\label{eq:L4lagr}
\end{align}
where we have assumed universal masses and couplings across all the sectors for simplicity. We again need $\lambda\sim 1/\sqrt{N}$ for perturbativity. Note that we are assuming that the bilinear $\mu_E\,e^c\,E$ only couples a single flavor of right handed lepton to the new 4$^\text{th}$ generation fields, in order to avoid flavor violation bounds in the charged lepton sector. The predictions relevant to cosmology (see Fig.~\ref{fig:Neff}) are insensitive to the choice of flavor; we choose couplings involving the $\tau$ for the additional constraints discussed in Sec.~\ref{sec:MixingBetweenSectors} below since this choice yields the strongest bounds.
To explore the differences between the $L_4$ and $\ell$ models let us again consider the limit in which the reheaton is light. If we integrate out the Higgs and gauge bosons along with the new vector-like leptons, the leading operators for the decays of $S^c$ are given by
\begin{equation}
\begin{array}{l}
\!\!\!\!\mathcal{L}^{\vev{H}\neq 0}_{L_4}\! \supset \mathcal{C}^{L_4}_1 \lambda'\,\frac{g^2}{m_W^2}\, \Big(e^{c\dag}\bar{\sigma}^\mu S^c\Big)\Big(f^\dagger \bar{\sigma}_\mu f^\prime\Big)\,;\\[10pt]
\!\!\!\!\mathcal{L}^{\vev{H}= 0}_{L_4}\! \supset \mathcal{C}^{L_4}_2\lambda\,\frac{y_t \,y_b}{16\,\pi^2}\frac{Y_E \,M_E\, \mu_E}{m_H^4}\,\Big(e^{c\dagger}\bar{\sigma}^{\mu} S^c\Big)\!\Big(u^{c\dagger}_3 \bar{\sigma}_{\mu} d^c_3\Big) \,,
\end{array}
\label{Eq:L4eff}
\end{equation}
where once more the $\mathcal{C}^{L_4}_i$ are numerical coefficients, $M_4$ is used to represent the physical mass of the relevant heavy lepton, and for convenience we have defined $\lambda'_i \equiv \left(\lambda \,v^2_i\,\mu_E/ M_{4i}^4\right) f(Y, M)$. Here $f$ is a function of dimension one that depends on the Yukawa couplings and vector-like masses in Eq.~(\ref{Eq:L4eff}), but not on the Higgs vev. The $M_{4 i}$ masses receive a contribution from $v_i$ that eventually dominates. When this happens $S^c$ decays become suppressed by large powers of the Higgs vev. From the effective Lagrangian above, it is easy to conclude that the widths scale as $\Gamma_{m_H^2<0}\sim {\rm const}$ for the first few sectors, since $M_{4 i}$ is approximately independent of $v_i$. When the Yukawa contribution to the masses begins to dominate, such that $M_{4i} \sim v_i$, the scaling becomes $\Gamma_{m_H^2<0}\sim 1/v_i^8$. Contributions to observables from the sectors with positive Higgs mass squared are negligible: the decay is both three-body and loop-suppressed, and the width scales as $1/v_i^8$ in all the sectors.
The diagrams that lead to these decays are shown in Fig.~\ref{fig:L4Diagrams}, and the energy density deposited in each sector is depicted in the right panel of Fig.~\ref{fig:TRH}. It is obvious that in this model cosmological observables are sensitive only to the few sectors for which the vector-like masses dominate over the Higgs vev, making it insensitive to the UV. This comes at the price of introducing new degrees of freedom near the weak scale. As we will discuss in the following section, the vector-like masses cannot be arbitrarily decoupled, but they must be large enough to avoid tension with direct searches and the measured properties of our Higgs.
\begin{figure}[!t]
\centering
\includegraphics[width = .48\textwidth]{Figures/L4_Diagrams.pdf}
\caption{Feynman diagrams for the most important decays in the $L_4$ model. The left (right) column is for $\vev{H}\! \neq\! 0\, \big(\!\!\vev{H}\! =\! 0\big)$. The top (bottom) row is for $m_S \gg |m_H|$ $\big(m_S \ll |m_H|\big)$.}
\label{fig:L4Diagrams}
\end{figure}
Finally, we end this section by briefly commenting on the presence of an upper bound for the reheating temperature $T_\text{RH}$ such that the mechanism is preserved. Specifically, $T_\text{RH}$ should be at most of order of the weak scale. If the temperature were larger, our Higgs mass would be dominated by thermal corrections resulting in a change in the scalings of the branching ratios. Our Higgs would obtain a large positive thermal mass and no longer be preferentially reheated over the other sectors. Noting that
\begin{eqnarray}
T_\text{RH} \simeq 100 \text{ GeV } \sqrt{\frac{\vev{\Gamma_\text{reheaton}}_T}{10^{-14} \text{ GeV}}},
\label{eq:TRH}
\end{eqnarray}
where $\vev{\Gamma_\text{reheaton}}_T$ denotes a thermal average of the reheaton width that incorporates the effect of time dilation. Then Eq.~\eref{eq:TRH} places an upper bound on the couplings of the reheaton. In the $\phi$ model, the $\phi - h$ mixing angle is bounded to be $\theta_{\phi h} \sim \left(a \,v / m_h^2\right)_{\text{us}} \lesssim 10^{-6}~\left(100\text{ GeV} / m_{\phi}\right)^{1/2}$. In the $L_4$ model, most of the viable region of parameter space predicts on-shell decays to our $W$ boson (see Fig.~\ref{fig:Neff} below). Therefore, the width of $S^c$ is dominated by this two-body decay and the constraint on $T_\text{RH}$ translates into a rough bound of $\lambda'_\text{us} \lesssim 10^{-7}$ when $m_S \simeq 100\text{ GeV}$. For the benchmark values used for the figures below, this in turn translates into a bound $\lambda\times \mu_E \lesssim 10^{-2}~\text{GeV}$.
Finally, we note that at large $N$ there is a more stringent upper bound on the reheating temperature determined by the perturbativity of $\lambda$. Requiring $\lambda~\lesssim~4\,\pi/\sqrt{N}$ and $m_S~\sim~100$~GeV, we find that it is still possible to reheat to a few GeV even with $N \sim 10^{16}$, where this estimate has been done using the complete numerical implementation of the mixings.
In principle, we must also ensure that other sectors are not overly heated by scattering from our own plasma after reheating. However, the aforementioned constraints on the reheaton couplings sufficiently suppress this contribution to their energy density.
\subsection{Baryogenesis}
A viable mechanism for baryogenesis is an even more crucial part of our mechanism for solving the hierarchy problem than in typical natural theories for new physics, where it can be treated in a modular way. One challenge is that our reheating temperature should be near or below the electroweak phase transition. Additionally, baryogenesis cannot occur in all of the copies of the SM, or there would be too much matter in the Universe.
One simple approach, which makes use of features intrinsic to the model, is to imagine that the reheaton $S^c$ carries a lepton number asymmetry. This asymmetry is distributed to the various sectors through the decays of $S^c$. Only in the sectors nearest ours is this lepton asymmetry converted into a baryon asymmetry. The small number abundance of baryons results from the low reheat temperature. At temperatures just below the electroweak phase transition, the sphaleron rate is exponentially suppressed, and only a small fraction of the lepton asymmetry is converted into a baryon asymmetry. The baryon asymmetry in sectors with $m_H^2 > 0$ is even further suppressed; since $m_W \lesssim \Lambda_{\text{QCD}}$, the sphalerons remain active at temperatures below the baryon masses. Any asymmetry in these sectors will eventually be redistributed back into the leptons. We have now laid out the necessary ingredients of our mechanism and we are ready to explore their phenomenology in more detail.
\section{Signals and Constraints}
\label{sec:Signals}
The signals and experimental constraints for $N$naturalness come from two sources: mixing between the sectors and energy density deposited in the new sectors by the reheaton decays. The cosmological observables sensitive to the energy density in each sector can be further divided into two categories.
First we discuss measurements that can detect new light particles. These signatures are dominated by the sectors closest to us and can not be avoided by changing the UV scalings of the model. They provide the most characteristic signatures of the theory. Then we study the impact of stable massive particles from the new sectors. This last set of constraints is dominated by sectors with the largest Higgs masses and can be ignored in the $L_4$ model, where the large $i$ physics is decoupled. In the last two subsections we discuss the bounds arising from mixing between the sectors, followed by possible collider signatures.
\subsection{Massless degrees of freedom}
\label{subsec:massless}
As discussed previously, our models have a large number of massless or nearly massless degrees of freedom. For example, all additional sectors contain photons and neutrinos. There are several kinds of cosmological observations that are sensitive to new relativistic particles. For instance the measurement of the Hubble parameter during either Big Bang Nucleosynthesis or at the epoch of photon decoupling, and bounds on hot dark matter from the matter power spectrum.
The sensitivity of the expansion of the Universe to new relativistic degrees of freedom is usually phrased in terms of the number of effective neutrinos
\begin{equation}
\Delta N_{\rm eff} = \frac{1}{\rho_\nu^\text{us}} {\displaystyle \sum_{i\neq \text{us}} \rho_i}\, .
\label{eq:dNeff}
\end{equation}
Current bounds are $\Delta N_{\rm eff} \lesssim 1$ during BBN~\cite{Cooke:2013cba} and $\Delta N_{\rm eff} \lesssim 0.6$ at photon decoupling~\cite{Ade:2015xua}. In both cases we quote an approximate $95\%$ C.L. constraint. The CMB bound applies to free-streaming radiation~\cite{Bell:2005dr}. However, the photons in some of the new sectors are still in equilibrium with or have just decoupled from electrons at that time and might be more similar to a perfect fluid. Until recently it was impossible to distinguish between the two types of radiation, as they affect the CMB damping tail in the same way~\cite{Hou:2011ec}. The detection of a phase shift in the CMB anisotropies~\cite{Follin:2015hya} has broken this degeneracy, and it is now possible to set a $95\%$ C.L. bound: $N_{\rm fluid}\lesssim 1$ for $\Delta N_{\rm eff}=0$~\cite{Baumann:2015rya}. Here we have defined $N_{\rm fluid}$ in the same way as $\Delta N_{\rm eff}$, normalizing the energy density of non-free-streaming radiation to that of a neutrino in our sector.
In the following, we do not distinguish between the two types of radiation. We use $\Delta N_{\rm eff}$ to denote the sum of the two components. Given the bounds discussed above and the two dimensional exclusions in~\cite{Baumann:2015rya}, this is sufficient to show that the model has large areas of parameter space consistent with current data. In the future, it would be interesting to explore CMB observations in more detail, as it is a generic prediction of this type of theories to have roughly comparable amounts of free-streaming and non-free-streaming extra radiation.
Having set $N_{\text{fluid}}$ to zero, it is straightforward to estimate the contribution to $\Delta N_{\rm eff}$ from our new sectors, since the ratio of energy densities $\rho_i/\rho_\text{us}$ is determined by the decay widths of the reheaton: $\rho_i/\rho_\text{us}\simeq\Gamma_i/\Gamma_\text{us}$.
\begin{figure*}[!t]
\begin{center}
\includegraphics[width=0.47\textwidth]{Figures/Neff_phi.pdf} \quad\quad
\includegraphics[width=0.47\textwidth]{Figures/Neff_L4.pdf}
\caption{$\Delta N_{\rm eff}$ contours as a function of reheaton mass and the $r$ parameter defined in Eq.~(\ref{eq:mHsqiScaling}). $\Delta N_{\rm eff}\simeq 0.03$ corresponds to the sensitivity of CMB stage 4 experiments. The current upper bound at the CMB epoch is around $0.6$.
The left panel is for the $\phi$ model with $a=1$~MeV. The right panel is for the $L_4$ model with $\lambda\times \mu_E=1$~MeV, $M_L=400$~GeV, $M_{E, N}=500$~GeV, $Y_E=Y_N=0.2$, and $Y^c_E=Y^c_N=-0.5$. As discussed in the text, the $L_4$ result is valid for a large range of $N$, namely $30 \lesssim N\lesssim 10^{9}$. Both figures were made using the zero temperature branching ratios of the reheaton; see the end of Sec.~\ref{Models} for a discussion.}
\label{fig:Neff}
\end{center}
\end{figure*}
For example, assume that the reheaton is lighter than the lightest Higgs across all the sectors; then we have
\begin{widetext}
\begin{eqnarray}
\Delta N_{\rm eff}^{\phi} &\sim& {\displaystyle \sum_{i=1}^{N_b} \frac{1}{2\,i+1}} + \frac{y_c^2}{y_b^2}{\displaystyle \sum_{i=N_b+1}^{N_c} \frac{1}{2\,i+1}} \simeq \frac{1}{2}\left(\log 2N_b + \frac{y_c^2}{y_b^2}\log \frac{N_c}{N_b}\right) \, , \quad\quad N_{b,c}=\left(\frac{m_{\phi}^2}{8\,m^2_{b,c}} -\frac{1}{2}\right)\, , \label{eq:Neff} \\
\Delta N_{\rm eff}^{L_4} &\sim& {\displaystyle \sum_{i=1}^{N_V} i^0} + {\displaystyle \sum_{i=N_V+1}^{N/2}}\frac{1}{(2\,i+1)^4} \simeq N_V\, , \quad\quad N_V\simeq\left(\frac{M^2}{Y^2\, v^2} -\frac{1}{2}\right) \, , \nonumber
\end{eqnarray}
\end{widetext}
where $M$ represents one of the vector-like masses in the $L_4$ model and $Y$ one of the new Yukawas. In this estimate we have neglected the contribution from $m_H^2>0$ sectors and the effect of $g_*$ in each sector, to highlight the scaling of $\Delta N_{\rm eff}$. From this simple exercise we see that $\Delta N_{\rm eff}$ is dominated by the bottom of the spectrum. The sectors past $i=N_{b,c}$ or $i=N_V$ receive a negligible fraction of the total energy density and do not contribute to $\Delta N_{\rm eff}$. Using Eq.~(\ref{eq:Neff}) to go beyond a simple parametric estimate gives results that are in tension with current bounds. For example $m_\phi \simeq 50$~GeV implies $N_b \simeq 17$ and $\Delta N_{\rm eff} \simeq 2$. However, these estimates are only qualitative, and break down in a large fraction of the parameter space of the models. The results from a full numerical computation are shown in Fig.~\ref{fig:Neff}.
There are two main messages that can be extracted from this calculation. First, we can satisfy current constraints for a range of reheaton masses up to a few hundred GeV. Second, the models predict values of $\Delta N_{\rm eff}$ within the range of sensitivity relevant for CMB stage 4 experiments~\cite{Dodelson:2013pln}. These next generation detectors, which should start taking data within the next five years, will probe $\Delta N_{\rm eff} \gtrsim 0.03$. If no beyond the SM discovery is made, then the only way to suppress this signal is to introduce ``fine tuning," which in the context of these models is the limit $r \lesssim 0.1$. Alternatively, we could imagine alleviating this tension by taking the vector-like masses in the $L_4$ model far below the weak scale, in potential conflict with electroweak/Higgs measurements.
A few additional features of the $\Delta N_{\rm eff}$ calculation are worth discussing. In the $L_4$ case the plot is valid for a large range of $N$, namely $30 \lesssim N\lesssim 10^{9}$. The upper bound is determined by requiring $\lambda \lesssim 4\,\pi/\sqrt{N}$ and mixing between $e^c$ and the vector-like leptons less than 1\%. It is trivial to go beyond $N=10^{9}$, and even possible to reach $N = 10^{16}$, by lowering the reheaton coupling -- this comes at the expense of an overall decrease in reheating temperature, even though the result for $\Delta N_{\rm eff}$ would not change. For $N < 30$, $\Delta N_{\rm eff}$ is smaller than shown in the figure. In the $\phi$ case, the results are more sensitive to $N$, as shown in Eq.~(\ref{eq:Neff}). We chose the largest $N$ that is both compatible with overclosure (see the next subsection) and also interesting from a model building perspective, given the relation to the Planck/GUT hierarchy ($N=10^4$).
The shapes of the $\Delta N_{\rm eff}$ contours are easy to explain in terms of kinematics. In $L_4$ the allowed region corresponds to the reheaton decaying to our sector via a two-body channel, versus a three-body decay into all the other $m_H^2<0$ sectors. This is highlighted by the $m_S = m_{W_2}$ line in the plot. In the $\phi$ model the situation is different. The mixing with the Higgs naturally introduces a number of mass thresholds that reduce $\Delta N_{\rm eff}$. At very low $\phi$ masses, decays to a pair of $b$-quarks are kinematically allowed only in our sector. As the $\phi$ mass increases, the reheaton can mix resonantly with our Higgs and subsequently decay to a pair of $W$ or $Z$ bosons. The last aspect of these results that is not captured by the simple estimate in Eq.~(\ref{eq:Neff}) is the fact that $\left(\Delta N_{\rm eff}\right)_{\rm CMB} > \left(\Delta N_{\rm eff}\right)_{\rm BBN}$. It is easy to show that this must be the case by appealing to conservation of entropy in each of the sectors. If we compute the ratio of $\Delta N_{\rm eff}$ in sector $i$ at the two different epochs, we obtain
\begin{widetext}
\begin{eqnarray}
\frac{\left(\Delta N_{\rm eff}^i\right)_{\rm CMB}}{\left(\Delta N_{\rm eff}^i\right)_{\rm BBN}} =\frac{g_{*}^i\left(T_{\rm CMB}^i\right)}{g_{*}^i\left(T_{\rm BBN}^i\right)} \left(\frac{g_{*S}^{\rm us}\left(T_{\rm BBN}^{\rm us}\right)}{g_{*S}^{\rm us}\left(T_{\rm CMB}^{\rm us}\right)}\right)^{4/3}
\left(\frac{g_{*S}^i\left(T_{\rm BBN}^i\right)}{g_{*S}^i\left(T_{\rm CMB}^i\right)}\frac{g_{*S}^{\rm us}\left(T_{\rm CMB}^{\rm us}\right)}{g_{*S}^{\rm us}\left(T_{\rm BBN}^{\rm us}\right)}\right)^{4/3}
\simeq \left(\frac{g_{*S}^i\left(T_{\rm BBN}^i\right)}{g_{*S}^i\left(T_{\rm CMB}^i\right)}\right)^{1/3} \geq 1\, .
\end{eqnarray}
\end{widetext}
The first term in the first equality counts the number of relativistic degrees of freedom in sector $i$ at the two different temperatures. The second factor accounts for the fact that neutrinos in our sector are decoupled after BBN, so their temperature during the CMB epoch is lower than that of photons. The last term comes from entropy conservation in our sector and sector $i$. In the last equality we have used $g_{*}\simeq g_{*S}$.
To conclude the discussion of $\Delta N_{\rm eff}$, recall that the result depends almost exclusively on the reheaton branching ratios and that it is largely insensitive to the value of its overall coupling. A single choice of $\lambda$ and $a$ is sufficient to understand the complete parameter space. In contrast, the precise value of the vector masses and Yukawa couplings in the $L_4$ model can change the results considerably, as it is already clear from Eq.~(\ref{eq:Neff}). When the vector-like masses are around the TeV scale or above, the models are excluded, while $M\simeq 500$~GeV yields predictions that are consistent with current data, as shown in Fig.~\ref{fig:Neff}. We leave a more detailed exploration of the parameter space and a discussion of possible collider signatures to future work.
The second class of light particles that can impact our cosmological history are those that are non-relativistic at matter radiation equality, but might have free-streamed enough to suppress the matter power spectrum.
Particles that become non-relativistic at a time $t_\text{NR}<t_\text{EQ}$ suppress structure up to scales $\lambda_\text{FS} =c\, \sqrt{t_\text{EQ}^{\,} \,t_\text{NR}}(2+\log t_\text{EQ}/t_\text{NR})/a(t_\text{EQ})$. The neutrinos from many of the new sectors would have $\lambda_\text{FS}$ larger than one Mpc. At these scales the matter power spectrum can be computed reliably in the linear regime and can be used to infer another upper bound on their energy density. To roughly estimate current constraints we compute the energy density in particles that can suppress structure at one Mpc or above. We find that for Dirac neutrinos the energy density is well below $1\%$ of the total dark matter energy density in all the plane of Fig.~\ref{fig:Neff} for both the $\phi$ and $L_4$ models, while for Majorana neutrinos this is true within the $\left(\Delta N_{\rm eff}\right)_\text{CMB}=0.5$ contours.
The hot dark matter population may provide another signal. The tower of sterile neutrinos results in a characteristic impact on the matter power spectrum. Furthermore, the hot dark matter signal is primarily determined by the reheaton branching ratios (and hence the spacing between the lightest sectors), so once a value of $\Delta N_{\rm eff}\neq 0$ is measured it is possible to make predictions for the distortion of the matter power spectrum and vice versa. In general our theories produce non-trivially related modifications in several CMB observables and we leave to future work a more detailed study. Our generic expectation is that neutrino cosmology is modified at the $\mathcal{O}(1)$ level due to slightly heavier albeit less abundant neutrinos in the closest sectors with electroweak symmetry breaking.
\subsection{Massive stable particles}
Relic neutrinos account for a fraction $\Omega_\nu^\text{us} h^2 \simeq \sum m_\nu({\rm eV})/91.5 \gtrsim 10^{-3}$ of the energy density in the Universe. It is natural to ask if the heavier neutrinos in the sectors with $\vev{H}\neq 0$ can lead to overclosure problems. Furthermore, electrons and protons can be similarly problematic. This is perhaps surprising, since in the standard picture their symmetric component is completely negligible today. However, in the other sectors their masses are $\sqrt{i}$ larger and subsequently their annihilation cross-sections decrease as $1/i$.\footnote{For protons, this scaling is only valid once the quark masses exceed $\Lambda_{\text{QCD}}$; the scaling is slower for the nearer sectors.}
In all cases, the relic density of the new stable particles comes from two different sources. There is a contribution that grows with $i$ from the sectors where the stable particles reach thermal equilibrium (including a possible freeze-in abundance from our sector) and a second contribution that decreases with $i$ from sectors where the particles never thermalize. Let us focus on this first contribution (for the moment we will neglect the freeze-in abundance from our sector):
\begin{eqnarray}
\Omega\, h^2 &=& \frac{s_0}{\rho_c^0}\sum_{i=-1}^{-N_d} m_i\, Y_i^{\text{fo}} +... = a \left(N_d\right)^p+...\, .
\end{eqnarray}
Here we use $\Omega\, h^2$ to indicate the relic density of either neutrinos, electrons or protons; $\rho_c^0$ is the critical energy density today; $s_0$ is the entropy density; $m_i$ is the mass of the stable particle; $Y_i^{\text{fo}}$ is its yield at freeze-out; $N_d$ is the sector after which the stable particles are not ever in thermal equilibrium with the other particles in their sector; and $a$ and $p$ are positive numbers. In general $a \sim \Omega^{\text{us}} h^2$ and $p>1$. The reason for $p>1$ is that $m_i\sim \sqrt{i}$ (or $\sim i$ for Majorana neutrinos) and up to a certain sector number $Y_i^{\text{fo}}$ also grows with $i$, since neutrinos, electrons and protons all freeze-out earlier and earlier.
In the $\phi$ model this thermal abundance turns out to be the only relevant one. Specifically, electrons and positrons provide the dominant constraint. Once the bound on the reheating temperature is taken into account, the freeze-in abundance from our sector is negligible. Furthermore the overclosure bound on $N$ kicks in before including heavy enough sectors where electrons would not thermalize. Therefore the bound arises only from thermal freeze-out ($n_{e}^i \sim 1/\langle \sigma_e v\rangle_i$) and it is straightforward to estimate:
\begin{eqnarray}
\Omega_e^\phi\, h^2 &=& \sum_{i=1}^{N_\text{th}} \frac{m_{e}^i\, n_{e}^i}{\rho_c^0} \simeq \frac{\big(m_e^{\rm us}\, T_0^{\rm us}\big)^3}{\rho_c^0}\frac{N^{5/2}_\text{th}}{M_\text{pl}\, v_{\text{us}} \,\alpha^2}\, \notag\\
&\lesssim& 0.1\times \Omega_\text{DM}\, h^2 \quad \Longrightarrow \quad N_\phi\lesssim 10^5 \, ,
\end{eqnarray}
where the sum runs up to the heaviest sector where the electrons have thermalized as denoted by $N_\text{th}$, $T_0^{\rm us}$ is the photon temperature in our sector today, $m_e^{\rm us}$ is our electron mass, $M_\text{pl}$ is the Planck mass, and all other quantities were defined previously. For this estimate we have assumed that our sector dominates the energy density of the Universe when electron-positron annihilations freeze-out, \emph{i.e.}, the $\Delta N_{\rm eff}$ constraint is satisfied. Furthermore we have conservatively assumed freeze-out happens just after reheating (at $T_\text{RH}^{\rm us}\simeq v_{\text{us}}$) in all the sectors.
Finally, note that we have required that electrons and positrons make up only $10\%$ of dark matter, the rough bound for particles that behave very differently from cold, collisionless dark matter. To be more conservative we could require them to make up only $1\%$ of dark matter, which would reduce the maximum allowed value of $N_\phi$ by $60\%$, still leaving open $N_\phi=10^4$.
To conclude this section we note that the rapid scaling of the energy density with sector number in $L_4$ protects the model from overclosure. This implies that $N$ can be taken all the way to $10^{16}$ and still be consistent with data, at the expense of a low reheat temperature.
\subsection{Mixing Between Sectors}
\label{sec:MixingBetweenSectors}
Upon integrating out the reheaton, the low energy theory will contain cross-couplings between the sectors. Stringent bounds from stellar and supernova cooling place limits on the size of these mixings.
\paragraph{\boldmath $\epsilon_i \,F_{\mu \nu}\, F_i^{\mu \nu}$}
In the presence of kinetic mixing, the electrically charged particles of other sectors will have milli-charge couplings to our photon. The most stringent bound on this coupling is derived from energy loss in stars~\cite{Davidson:1991si, Davidson:2000hf}. In sectors with $m_H^2 > 0$, the charged particles are all extremely light -- much lighter than stellar temperatures -- so that democratic kinetic mixing leads to $\mathcal{O}(N)$ accessible final states for plasmon decay. Thus we require $\sqrt{\sum_i \epsilon_i^2} \lesssim 10^{-14}$, in which $\epsilon_i$ is the coefficient of kinetic mixing between our photon and that of sector $i$, and $i$ runs over all sectors with $m_H^2 > 0$.
Accordingly, there must be no bi-fundamental matter in the UV, the inclusion of which would generate kinetic mixing at one loop. Even in the absence of those states, kinetic mixing may be generated in the IR through the coupling to the reheaton.\footnote{Kinetic mixing is not generated at any order if the coupling between sectors is mediated by a single, real scalar field, since there can be no effective coupling of such a field to the electromagnetic field-strength tensor. Therefore, this effect can be safely neglected in the $\phi$ model.} In this case the bounds may be easily avoided by the smallness of the coupling. As described in Sec.~II, the portal couplings must decrease with increasing $N$ in order to have a consistent large-$N$ limit.
For example, in the $L_4$ model, we must have $\lambda \sim \lambda_0 / \sqrt{N}$. The kinetic mixing parameter is generated only at three loops with four powers of the portal coupling:
\begin{align}
\epsilon_i \sim \frac{\alpha}{4\, \pi}\left(\frac{\lambda_0}{4\,\pi}\right)^4 \frac{1}{N^2 \,i}.
\end{align}
Note that $\epsilon_i$ decreases with $i$ due to the scaling of $\big(m_H^2\big)_i$, and so kinetic mixing is dominated by the sectors nearest to our own. Then the stellar bounds may be avoided as long as $\lambda_0 / 4\pi \lesssim 10^{-3} \sqrt{N}$, and no suppression is required for $N \gtrsim 10^6$ beyond the natural large-$N$ scaling.
\paragraph{\boldmath $\epsilon_i^n\, \nu_i^{\dagger}\, \bar{\sigma}^{\mu} D_{\mu}\, \nu$}
At one loop in the $L_4$ model, the reheaton mass-mixes with neutrinos. After integrating out the reheaton, this induces kinetic mixing between neutrinos of different sectors. However, because the vector-like leptons only couple to the charged leptons, the effective coupling to the neutrinos is Yukawa-suppressed:
\begin{align}
\epsilon^n_i \sim (m_{\ell})_{\text{us}} \,(m_{\ell})_i \left(\frac{\lambda\, Y_E^c\, \mu_E}{16\, \pi^2 \,(M_4)_i \,m_S} \right)^2.
\end{align}
For sectors with $(M_4)_i \sim M_L$, $\epsilon_i \sim \sqrt{i}$. Once $Y_E \,v_i \gtrsim M_L$, the kinetic mixing decreases as $\epsilon_i \sim 1/\sqrt{i}$.
Energy loss in SN1987a~\cite{Raffelt:1987yt} limits the size of the kinetic mixing. The neutrino production rate from neutral-current bremsstrahlung requires $\sqrt{\sum_i (\epsilon^n_i)^2} \lesssim 10^{-4}$. Due to the growth of $\epsilon^n_i$ with $i$ for small values of $i$, the sum is dominated by those sectors for which the vector-like lepton masses are larger than their chiral masses. For typical parameters, such as those shown in Fig.~\ref{fig:Neff}, this is the case for only $\mathcal{O}(10)$ sectors. Taking into account the bound on $T_{\text{RH}}$ from Sec.~\ref{Models}, this gives $\sqrt{\sum_i (\epsilon^n_i)^2} \lesssim 10^{-13} \left(M_L / 4\, \pi\, v \right)^4$, so that there is no constraint as long as the vector-like masses are taken to be sufficiently close to the weak scale.
\paragraph{\boldmath $\epsilon^c_i \,G_F\, (\nu_i^{\dagger}\, \bar{\sigma}^{\mu} \,e^c)\left(p^{\dagger}\, \bar{\sigma}_{\mu}\, n\right)$}
There is a somewhat more powerful constraint from SN1987a due to charge-current neutrino production. The mass-mixing of the reheaton with neutrinos leads to an effective four-Fermi operator, with
\begin{align}
\epsilon^c_i \sim (m_{\tau} \,m_{\nu})_i \left(\frac{\lambda\, Y_E^c \,v_\text{us}\, \mu_E\, M_L}{4\,\pi\, m_S \,(M_4^2)_{\text{us}} \,(M_4)_i}\right)^2.
\end{align}
In the case of Majorana neutrino masses, $\epsilon_i^c$ grows like $\sqrt{i}$, so that $\sqrt{\sum_i (\epsilon_i^{c})^2} \sim N$. Once again taking into account the limit on $T_{\text{RH}}$, we have $\sqrt{\sum_i (\epsilon_i^{c})^2}~\lesssim~10^{-24}\ N\, \big(M_L / \sqrt{4\, \pi\, Y_E} \,v\big)^4$. The supernova bound is only $\sim 10^{-5}$, so that even for $N \sim 10^{16}$, the coupling is unconstrained for $M_L$ near the weak scale.
Finally, limits on active-sterile neutrino oscillations can also bound $\epsilon^c_i$, both from cosmological measurements as well as active neutrino disappearance~\cite{Cirelli:2004cz}. However, due to the Yukawa suppression of the neutrino mixing, the most relevant limits are those involving the tau neutrino, which are comparatively weak. Absent resonant mixing due to accidental degeneracies, which we expect to be atypical in our parameter space, the bounds from neutrino oscillations are negligible.
\subsection{Colliders}
\label{sec:Colliders}
Models of $N$naturalness can provide collider signatures through both direct production of the reheaton as well as rare decays of SM particles. However, the smallness of the reheaton couplings, due to both large-$N$ suppression and $T_{\text{RH}}$ constraints, precludes these signatures from being a generic feature of our models. In the $\phi$ model, for example, rare Higgs decays proceed through $\phi - h$ mixing. The dominant signature in this case is are invisible decays of the SM-like Higgs boson, with $\text{BR}_\text{inv} \sim \theta_{\phi h}^2 \,\Delta N_{\text{eff}} / (1 + \Delta N_{\text{eff}}) \lesssim 10^{-12} $ after requiring sufficiently low $T_{\text{RH}}$. Even using optimistic estimates, future colliders such as TLEP or a 100~TeV machine (with 10 ab$^{-1}$ of luminosity) will only produce $\sim 10^6$~\cite{Gomez-Ceballos:2013zzn} or $10^{10}$~\cite{Arkani-Hamed:2015vfh, Golling:2016gvc} Higgses, respectively, rendering such decays unobservable.
Direct production of $\phi$ is similarly suppressed, since it must proceed through the same mixing angle. Even for $m_{\phi} < m_h$, in which case the production cross sections are somewhat larger, direct production of $\phi$ will be unobservable. For example, a SM-like Higgs with a mass of 10~GeV gives a cross section only 2.5 times larger at TLEP and approximately 14 times larger at a 100~TeV $p-p$ machine than a Higgs at 125~GeV. Nevertheless, if $\phi$ is sufficiently light, there may be a variety of interesting signatures; see, {\emph e.g.},~\cite{Clarke:2013aya} for a study of current constraints which probe mixing angles down to $\theta_{\phi h} \sim 10^{-5}$. We leave a detailed study of detection prospects for $m_{\phi} \lesssim 10$~GeV to future work, see \emph{e.g.}, \cite{Krnjaic:2015mbs} for an analysis of a similar scenario.
In the $L_4$ model, the new sectors and the reheaton are similarly difficult to observe; however, the vector-like leptons of our own sector may be accessible. As discussed above, the vector-like mass parameters should all be of order the weak scale. This implies they would likely be observable, both directly and through $h\to \gamma\, \gamma$ and precision electroweak measurements. For a recent study of these bounds, see~\cite{Altmannshofer:2013zba}. In particular, attempting to evade these constraints by raising the mass of the vector-like leptons will re-introduce tension with $\Delta N_{\text{eff}}$, as described in the Section~\ref{subsec:massless}.
\subsection{A Heavy Axion}
\label{sec:HAxion}
An outstanding puzzle within the SM is the strong CP problem.
If we have $N$ copies of the SM, then naively they all have their own theta angles $\theta_i$. However, if the $S_N$ symmetry is only softly broken by Higgs mass terms, then all of these angles would be equal. A shared axion would be able to set to zero all the $\theta_i$'s at the same time. The only difference between the version proposed here and the single sector story is that here there are $N$ contributions to its mass from each $\Lambda_\text{QCD}$. The potential for the axion has three contributions
\begin{align}
V(a) \ni
\left\{
\begin{array}{ll}
\frac{\Lambda_{\text{QCD},i}^6}{m_{H,i}^2} \left(\frac{a}{f_a} - \theta_i\right)^2 & \text{for } i < 0 \\[8pt]
m_{\pi,i}^2\, f_{\pi,i}^2 \left(\frac{a}{f_a} - \theta_i\right)^2 & \text{for } 0 \leq i < N_u\\[8pt]
\Lambda_{\text{QCD},i}^4 \left(\frac{a}{f_a} - \theta_i\right)^2 & \text{for } i \geq N_u
\end{array}
\right.\,,
\label{Eq: axion}
\end{align}
where $\Lambda_{\text{QCD},i}$ is the QCD scale for the $i^\text{th}$ sector, $i<0$ corresponds to sectors with $\vev{H_i} = 0$, and $N_u \sim 10^5$ is the sector with the smallest vev for which $m_u > \Lambda_\text{QCD}$. The contribution from the sectors with $\vev{H_i} = 0$ are due to higher dimensional operators from integrating out the Higgs doublet, the sectors with $m_u < \Lambda_{\text{QCD}}$ yield the familiar contribution to the axion potential, and the final term is the result for pure QCD with no light quarks. Numerically, the first term can always be neglected, the second term dominates as long as $N < N_u$, and only the third term is relevant for $N \gg N_u$.
In order to estimate how much heavier this state will be as compared to the standard case, we have calculated Eq.~(\ref{Eq: axion}) numerically including the one-loop running of $\Lambda_\text{QCD}$. For the first two sums in Eq.~(\ref{Eq: axion}), we used chiral perturbation theory to calculate the contribution to the axion mass. For the last sum, we normalized it such that it is equal to the chiral perturbation theory result when $m_u = \Lambda_\text{QCD}$. A numerical fit to the axion mass gives approximately
\begin{align}
\frac{m_a(N)}{m_a(1)} \simeq \left\{
\begin{array}{ll}
4 \times 10^3 \left(\frac{N}{10^4}\right)^{1} \quad\quad\quad& \text{for } N < N_u \\[2pt]
2 \times 10^{14} \left(\frac{N}{10^{16}}\right)^{0.9}\, & \text{for } N \gg N_u \\
\end{array}
\right.\, .
\end{align}
It is critical that the soft-breaking of the $S_N$ symmetry by the different Higgs vevs does not lead to any issues via higher dimensional operators. For example, one class of operators that leads to a change in $\theta_i$ between the sectors are
\begin{eqnarray}
\mathcal{O}_{\Delta\theta_\text{QCD}} \sim Y_u\, H_i\, Q_i\, u_i^c \frac{|H^i|^2}{\Lambda_G^2}\,.
\end{eqnarray}
Because the different sectors have different Higgs vevs, a chiral rotation shows that the theta angles all differ by $\sim |H_i|^2/\Lambda_G^2$. Plugging this into Eq.~(\ref{Eq: axion}), solving for the axion vev, and requiring that our theta angle is smaller than $10^{-10}$, we find that $N < 10^{10}$ if a shared axion is the solution to the strong CP problem. This approach requires the important assumption that whatever resolves the hierarchy problem between $\Lambda_H$ and $\Lambda_G$ does not introduce these operators or any other Higgs dependent phases.
\section{Discussion}
\label{sec:Discussion}
In this paper we have proposed a new solution to the hierarchy problem. The need for a huge integer $N$ is obviously the least appealing feature of our setup. It is perhaps not entirely unreasonable to have the mild $N \sim 10^4$ compatible with the existence of a supersymmetric GUT scale, but this seems outlandish in the $N \sim 10^{16}$ limit. At the moment it is difficult to see how such a large integer can be explained dynamically, in the same way as we usually explain hierarchies by, \emph{e.g.} dimensional transmutation. On the other hand, this is simply another large set of degrees of freedom, and we do not deeply understand where the even vaster number of degrees of freedom in a macroscopic expanding universe comes from, so perhaps the large $N$ may eventually find a different sort of natural explanation. The theoretical consistency of the proposal also makes a number of demands on the UV theory, such as the absence of sizable cross-couplings between the sectors, which may be technically natural but may again strain credulity. However, we find it fascinating that huge values of $N$ are experimentally viable. This is highly non-trivial, and indeed in the simplest models we did find significant constraints on $N$. While we have examined all the zeroth-order phenomenological constraints we know of, it is important to continue to look for constraints on (and signals of!)\! the scenarios with high values of $N \big(\gg 10^4\big)$.
It is also interesting to compare $N$naturalness with other approaches. It bears a superficial resemblance to large extra dimensions, which add $10^{32}$ degrees of freedom in the form of KK gravitons, as well as the scenario of Dvali~\cite{Dvali:2007hz} which invokes $10^{32}$ copies of the SM. In each of these cases, $M_{\text{pl}}$ is renormalized down to the TeV~scale. Of course this predicts (as yet unseen) new particles accessible to the LHC~\cite{Dvali:2009ne}. By contrast, $N$naturalness solves the hierarchy problem with cosmological dynamics; the weak scale is parametrically removed from the cutoff, and so it does not demand new physics to be accessible at colliders.
$N$naturalness has some features in common with low-energy SUSY as well. Both models invoke a softly broken symmetry: SUSY is broken by soft terms, and the $S_N$ symmetry is broken by varying Higgs masses. Also in both cases, the most obvious implementations of the idea are experimentally excluded. If SUSY is directly broken in the MSSM sector, we have the famous difficulties with charge and color breaking; in the case of $N$naturalness, direct reheating of all $N$ sectors is grossly excluded by $N_\text{eff}$. Thus in both cases we need to have ``mediators.'' SUSY must be dominantly broken in another sector and have its effects mediated to the MSSM. Similarly, reheating must be dominantly communicated to the reheaton, which subsequently dumps its energy density into the other sectors. Finally, both models have additional scales that are not, on the face of it, tied to the physics responsible for naturalness. In SUSY there is a ``$\mu$ problem'' in that the vector-like Higgsino mass must be comparable to the soft scalar masses, while in $N$naturalness the reheaton mass must be close to the bottom of the spectrum of Higgs masses. While in both cases there are simple pictures for how this can come about, these coincidences do not emerge automatically.
Moving beyond purely field theoretic mechanisms, there is the recent proposal of the relaxion~\cite{Graham:2015cka}, which invokes an extremely long period of inflation coupled with axionic dynamics to relax to a low weak scale. While both the relaxion and $N$naturalness mechanisms are cosmological, the physical mechanism of the relaxation, associated with the huge number of $e$-foldings of inflation, is {\it in principle} unobservable given our current accelerating Universe, much like the vast regions of the multiverse outside our cosmological horizon are imperceptible. By contrast, the cosmological dynamics associated with reheaton decay in $N$naturalness are sharply imprinted on the particle number abundance in all the sectors. They are not only in principle observable but, as we have stressed (at least for a small number of sectors ``close'' to ours), are detectable in practice within our Universe.
It is also interesting to contrast $N$naturalness with the picture of an eternally inflating multiverse, with environmental selection explaining the smallness of the cosmological constant, as well as potentially at least part of the hierarchy problem. This picture is, after all, the first cosmological approach to fine-tuning puzzles. While it is very far from well-understood and has yet to make internal theoretical sense, it is the only cartoon we have for understanding the cosmological constant problem and does not involve any model-building gymnastics. Furthermore, fine-tuning for the Higgs mass also has a plausible environmental explanation. Especially in the context of minimal split SUSY~\cite{ArkaniHamed:2004fb}, these ideas give us a picture which simultaneously accounts for the apparent fine-tuning of the cosmological constant and the Higgs mass, while maintaining the striking quantitative successes of natural SUSY theories in the form of gauge coupling unification and dark matter. Nonetheless, it is important to continue to look for alternatives, minimally as a foil to the landscape paradigm. $N$naturalness is a concrete example of an entirely different cosmological approach to tuning puzzles, and in particular relies on the existence of only a single vacuum.
We note that there is no obstacle to augmenting $N$naturalness with an anthropic solution to the cosmological constant problem. The presence of extra sectors exponentially increases the number of available vacua. For example we could add to the SM a sector with $m$ vacua and end up with $m^N$. Already $N\simeq 10^4$ with two vacua per sector is more than enough to scan the cosmological constant without relying on string theory landscapes. When solving the entire hierarchy problem with $N \simeq 10^{16}$, the vacua utilized to scan the cosmological constant can even be the two minima of the Higgs potential; this requires a high cutoff so that the second minimum is below $\Lambda_H$ and the difference in the potential energy of the two minima is $\mathcal{O}\big(\Lambda_G\big)$.
To conclude, we would like to comment on the nature of the signals that we have discussed in this paper. For concreteness, three models that make $N$naturalness cosmologically viable were presented. However, it is easy to imagine a broader class of theories that realizes the same mechanism. We can relax the assumption that the Higgs masses are uniformly spaced (or even pulled from a uniformly distribution) or that all the new sectors are exact copies of the SM. It is also possible to construct different models of reheating, with new physics near the weak scale to modify the UV behavior of the theory.
Nonetheless our sector can not be special in any way. There will always be a large number of other sectors with massless particles and with matter and gauge contents similar to ours, leading to the following signatures:
\mbox{} \\
\begin{itemize}[leftmargin=12pt]
\item We expect extra radiation to be observable at future CMB experiments.
\item The neutrinos in the closest $m_H^2<0$ sectors are slightly heavier and slightly less abundant than ours. This implies $\mathcal{O}(1)$ changes in neutrino cosmology, which will start to be probed at this level in the next generation of CMB experiments~\cite{Allison:2015qca}.
\item If the strong CP problem is solved by an axion, its mass will be much larger than the standard prediction.
\item If $N\lesssim 10^4$ as motivated by grand unification, supersymmetry or new natural dynamics should appear beneath $10$~TeV.
\end{itemize}
\mbox{} \\
The natural parameter space is being probed now, and soon we may know if the $N$naturalness paradigm explains how the hierarchy problem has been solved by nature.
\vspace{.3in}
\section*{Acknowledgments}
\vspace{-0.1in}
We are grateful to Bob Holdom, Jared Kaplan, Marilena LoVerde and Kris Sigurdson for useful discussions. We also want to thank Arka Banerjee and Neal Dalal for a preliminary investigation of the neutrino cosmology. NAH is supported in part by U.S. Department of Energy grant de-sc0009988. TC is supported by an LHC Theory Initiative Postdoctoral Fellowship, under the National Science Foundation grant PHY-0969510. RTD acknowledges support from the Institute for Advanced Study Marvin L. Goldberger Membership and DOE grant de-sc0009988. AH is supported by the Department of Energy grant DE-SC0012012 and National Science Foundation grant 1316699. HDK is supported by National Research Foundation of Korea(NRF) No. 0426-20140009 and No. 0409-20150110. TC and AH acknowledge the Aspen Center for Physics where this work was completed, which is supported by National Science Foundation grant PHY-1066293.
\end{spacing}
|
\section{Introduction}
Over the course of several decades, a standard view has emerged in which galaxies produced the dominant contribution of H{\sc~i}\ ionizing photons at redshifts $z>4$ \citep[e.g.][]{1987ApJ...321L.107S,2009ApJ...703.1416F,2012ApJ...746..125H,2013MNRAS.436.1023B}. This view is based on numerous measurements of the AGN luminosity function which show a steep decline in the AGN abundance at $z>3$ \citep[e.g.][]{2006AJ....132..117F,2010AJ....139..906W,2013ApJ...768..105M,2015MNRAS.453.1946G}. Recently, the standard view has been challenged by some authors citing a large sample of faint AGN candidates reported by \citet[][G2015 hereafter]{2015A&A...578A..83G}. These objects were selected by searching for $X$-ray flux coincident with $z>4$ galaxy candidates in the CANDELS GOODS-South field -- a technique that in principle allows the detection of fainter AGN compared to prior selection methods. If confirmed, this large population of faint AGN could make necessary a substantial revision of our understanding of the high-$z$ ionizing background, and possibly even of cosmological reionization. Indeed, \citet{2015ApJ...813L...8M} showed that AGN with ionizing emissivities consistent with the G2015 measurements could reionize intergalactic H{\sc~i}\ by $z\approx6$ without any contribution from galaxies (see also \citealt{2016MNRAS.457.4051K} and \citealt{2016arXiv160204407Y}). A distinguishing feature of their model is that He{\sc~ii}\ reionization ends by $z \approx 4$, at least 500 million years earlier than in the standard scenario in which it ends at $z\approx 3$ \cite[see e.g.][and references therein]{2015arXiv151200086M}.
An AGN-sourced ionizing background at $z>4$ potentially explains three puzzling observations of the IGM. First, H{\sc~i}\ Ly$\alpha$ forest\footnote{From here on, the term ``Ly$\alpha$ forest" refers to H{\sc~i}.} measurements show that the H{\sc~i}\ photoionization rate, $\Gamma_{\mathrm{HI}}$, is remarkably flat over the redshift range $2<z<5$ \citep[see e.g.][]{2013MNRAS.436.1023B}. This flatness is traditionally explained by invoking a steep increase in the escape fraction of galaxies with redshift, coinciding with the decline of the AGN abundance at $z>3$ \citep[e.g.][]{2012ApJ...746..125H}. \citet{2015ApJ...813L...8M} showed that the slower evolution of the AGN emissivity claimed by G2015 can more naturally account for the observed flatness of $\Gamma_{\mathrm{HI}}$ without appealing to a coincidental transition between the AGN and galaxy populations. Second, measurements of the mean opacity of the He{\sc~ii}\ Ly$\alpha$ forest at $z=3.1-3.3$ are lower than predictions from existing simulations of He{\sc~ii}\ reionization, which use standard quasar emisisivity models with He{\sc~ii}\ reionization ending at $z\approx3$ \citep{2014arXiv1405.7405W}. \citet{2015ApJ...813L...8M} suggested that such low opacities are more consistent with an earlier onset of He{\sc~ii}\ reionization driven by a large population of high-$z$ AGN. Lastly, recent observations by \citet{2015MNRAS.447.3402B} show that the dispersion of opacities among coeval $50h^{-1}\mathrm{Mpc}$ segments of the Ly$\alpha$ forest increases rapidly above $z>5$, significantly exceeding the dispersion predicted by models that assume a uniform ionizing background. In a companion paper \citep[][Paper I hereafter]{2016arXiv161102711D}, we show that accounting for this dispersion with spatial fluctuations in the ionizing background, under the standard assumption that galaxies are the dominant sources, requires that the mean free path of H{\sc~i}\ ionizing photons be significantly shorter than observations and simulations indicate \citep[see also][]{2015MNRAS.447.3402B,2015arXiv150907131D}. Alternatively, models in which AGN source the ionizing background naturally lead to large fluctuations owing to the brightness and rarity of AGN. \citet{2015arXiv150501853C} showed with a ``proof-of-concept" model that rare sources with a space density of $\sim 10^{-6}~\mathrm{Mpc}^{-3}$, similar to the space density of $>L_*$ AGN in G2015, could generate large-scale ($\sim 50h^{-1}~\mathrm{Mpc}$) opacity variations substantial enough to account for the observed dispersion at $z=5.8$. During the final preparation of this manuscript, \citet{2016arXiv160608231C} used more realistic models to show that a $\gtrsim 50\%$ contribution from AGN to the ionizing background is sufficient to account for the observed dispersion. We reach a similar conclusion in this paper.
The purpose of this paper is to further elucidate the implications of a large AGN population at high redshifts. To this end we will discuss three observational probes of AGN-dominated models of the high-$z$ ionizing background: (1) We will develop empirically motivated models of the Ly$\alpha$ forest in scenarios where AGN constitute a significant fraction of the background. We will then use these models to assess the possible contribution of AGN to the $z>5$ Ly$\alpha$ forest opacity fluctuations; (2) We will quantify the implications of these models for He{\sc~ii}\ reionization and for the thermal history of the IGM -- a facet that has yet to be discussed in the literature; (3) We will discuss the interpretation of recent He{\sc~ii}\ Ly$\alpha$ forest opacity measurements in the context of these models. Foreshadowing, we will show that, while AGN-dominated models are indeed a viable explanation for the Ly$\alpha$ forest opacity measurements of \citet{2015MNRAS.447.3402B}, the models that best match the measurements are qualitatively inconsistent with constraints on the thermal history of the IGM under standard assumptions about the spectra of faint AGN. We will further argue that the discrepancy between the opacities observed in the $z \approx3.1 - 3.3$ He{\sc~ii}\ Ly$\alpha$ forest and those in current He{\sc~ii}\ reionization simulations may reflect deficiencies in the simulations rather than favor AGN-dominated models.
The remainder of this paper is organized as follows. In Section \ref{SEC:lumfunc} we present a comparison of the AGN luminosity function of G2015 to other measurements in the literature. Section \ref{SEC:opacityflucs} is dedicated to models of the Ly$\alpha$ forest opacity fluctuations, while \S \ref{SEC:thermalhistory} explores the impact of high-$z$ AGN on He{\sc~ii}\ reionization and the thermal history of the IGM. Section \ref{SEC:HeIIforest} discusses the interpretation of recent He{\sc~ii}\ Ly$\alpha$ forest opacity measurements. Finally, in \S \ref{SEC:conclusion} we offer closing remarks. All distances are reported in comoving units unless otherwise noted. We assume a flat $\Lambda$CDM cosmology with $\Omega_m=0.31$, $\Omega_b=0.048$, $h=0.68$, $\sigma_8=0.82$, $n_s=0.97$, and $Y_{\mathrm{He}}=0.25$, consistent with recent measurements \citep{2015arXiv150201589P}.
\section{The AGN Luminosity Function}
\label{SEC:lumfunc}
\begin{figure}
\begin{center}
\resizebox{8.5cm}{!}{\includegraphics{Lum_func.eps}}
\end{center}
\caption{A comparison of the G2015 AGN luminosity function measurements to previously published measurements. The luminosity functions are expressed in terms of $M_{\rm AB,1450}$, the AB magnitude at a wavelength of 1450 \AA. We adjust the redshifts of the previous measurements to the central redshifts of the G2015 bins by assuming that the luminosity function scales as $10^{-0.47z}$, a common approximation that is motivated by the observed evolution of the bright end of the luminosity function at $z=3-6$ \citep{2001AJ....121...54F}. The G2015 luminosity function measurements are highly discrepant with previous measurements based on optically selected samples.}
\label{FIG:lumfunc}
\end{figure}
\begin{figure}
\resizebox{8.5cm}{!}{\includegraphics{emissivity.eps}}
\caption{Comoving emissivity of AGN at $h\nu = 1~\mathrm{Ry}~( \approx 13.6~\mathrm{eV})$. For reference, the magenta/dot-dashed curve shows the model of \citet{2015ApJ...813L...8M}, which is motivated by the \citet{2015A&A...578A..83G} emissivity measurements. In this model, AGN emissions alone reionize H{\sc~i}\ by $z\approx5.5$. See \citet{2015ApJ...813L...8M} for a similar comparison. }
\label{FIG:emissivity}
\end{figure}
We begin by comparing the AGN luminosity function measurements of G2015 to a compilation of other recent measurements based on optically selected samples (Fig. \ref{FIG:lumfunc}). We adjust the redshifts of the other measurements to the central redshifts of the G2015 bins by assuming that the luminosity function scales as $10^{-0.47z}$, a common approximation that is motivated by the observed evolution of the bright end of the luminosity function over $z=3-6$ \citep{2001AJ....121...54F}. Fig. \ref{FIG:lumfunc} shows that G2015's $X$-ray selected sample contains a much larger number of faint AGN than is expected from extrapolations of the past measurements.\footnote{We note that G2015's claims of a large number density of faint AGN ($M_{\mathrm{AB},1450} \approx -19$ to $-21$) may be supported by the AGN abundance found in the recent BlueTides simulation \citep{2016MNRAS.455.2778F}. However, a direct comparison is not possible because the simulation does not extend below $z=8$.} To reconcile this discrepancy, G2015 argued that previous optical surveys missed a large fraction of high-$z$ AGN at intermediate luminosities ($M_{\mathrm{AB},1450} \approx -22$ to $-25$). In fact, the G2015 fits shown in Fig. \ref{FIG:lumfunc} (red/solid lines) use only their faint AGN sample and a select subset of previous optical measurements at the bright end, which they argue are less prone to incompleteness. Because of this, Fig. \ref{FIG:lumfunc} shows that extrapolating the G2015 luminosity function to intermediate luminosities yields a much larger AGN population than was found previously by the other surveys.
G2015 assumed that the escape fractions of H{\sc~i}\ ionizing photons for faint and intermediate luminosity AGN are similar to the measured values for quasars. Under this assumption, their luminosity function translates to roughly a factor of ten greater contribution to the $z>4$ H{\sc~i}\ ionizing background than previous measurements. The data points in Fig. \ref{FIG:emissivity} translate the G2015 luminosity function, as well as others in the literature, to an ionizing emissivity of 1 Ry photons,
\begin{equation}
\epsilon^{\rm AGN}_{912}= \int\mathrm{d} L_\nu~\Phi(L_\nu)L_{912}(L_{\nu}),
\label{EQ:em912}
\end{equation}
where $L_\nu$ is the specific luminosity, $\Phi(L_\nu)$ is the luminosity function, and $L_{912}(L_\nu)$ is the specific luminosity at 912 \AA\ (see \citealt{2015ApJ...813L...8M} for a similar comparison). We assume a power-law specific luminosity $\propto\nu^{-0.61}$ at wavelengths $>912$~\AA, consistent with the stacked quasar spectrum of \citet{2015MNRAS.449.4204L}. As in G2015, equation (\ref{EQ:em912}) assumes an escape fraction of unity, and we integrate integrate down to a specific luminosity of $L_* / 100$. We note that much of the ionizing emissivity comes from the faint AGN population for the G2015 measurements. For example, in their highest redshift bin, we find that $88\%, 71\%, 49\%$, and $30\%$ of the ionizing emissivity comes from AGN with $M_{\mathrm{AB},1450} > -24, -23, -22$, and $-21$, respectively. If correct, the G2015 emissivity measurements could change the widely accepted view that galaxies were the chief sources of H{\sc~i}\ reionization. The magenta dot-dashed curve shows the recently published model of \citet{2015ApJ...813L...8M}, in which AGN reionize H{\sc~i}\ by $z\approx 5.5$ without {\it any} contribution from galaxies.
We conclude this section by noting that \citet{2015MNRAS.448.3167W} searched for $z\gtrsim5$ AGN in the GOODS-South field -- the same field considered in G2015 -- using nearly the same data set and found no convincing candidates. In the ensuing sections, we will adopt the point of view that the contribution of AGN to the $z>5$ ionizing background remains highly uncertain. For the sake of exploring implications for the Ly$\alpha$ forest, we will consider several scenarios in which AGN contributed significantly to $z>5$ ionizing background.
\begin{figure*}
\begin{center}
\resizebox{6.cm}{!}{\includegraphics{AGN_source_field.eps}}
\hspace{-0.28cm}
\resizebox{6.cm}{!}{\includegraphics{GammaHI_z5.6_AGN50_MFP30Mpch.eps}}
\hspace{-0.28cm}
\resizebox{6.cm}{!}{\includegraphics{MFP_z5.6_MFP30Mpch_AGN50.eps}}
\end{center}
\caption{Fluctuations in the H{\sc~i}\ ionizing background in one of our galaxies$+$AGN source models. Here we show results from a model in which AGN emissions produce $50\%$ of the global average H{\sc~i}\ photoionization rate. Left: a slice of thickness $20h^{-1}~\mathrm{Mpc}$ through the AGN source field. We draw AGN from the luminosity function of \citet{2015A&A...578A..83G} and randomly distribute them among the most massive galactic halos in our hydro simulation. To improve the statistics of our simulations of the ionizing background, we have tiled together eight of our hydro simulation boxes to achieve an effective $L_{\rm box}=400 h^{-1}~\mathrm{Mpc}$ before drawing AGN. Middle: the H{\sc~i}\ photoionization rate in a slice of thickness $6.25h^{-1}~\mathrm{Mpc}$, situated in the middle of the slice in the left panel. Right: the mean free path of 1 Ry photons in the same slice as in the middle panel. The clustering of the AGN leads to large-scale variations in the mean free path that amplify fluctuations in the ionizing background. }
\label{FIG:visB}
\end{figure*}
\section{Opacity Fluctuations in the High-Redshift Ly$\alpha$ Forest}
\label{SEC:opacityflucs}
The Ly$\alpha$ forest is the foremost observational probe of the metagalactic ionizing background and of the thermal state of the IGM. At a given location of the forest, the Ly$\alpha$ optical depth, $\tau_{\rm Ly \alpha} \equiv -\ln F$ (where $F$ is the transmitted fraction of the quasar's flux) scales as $\tau_{\rm Ly \alpha} \propto T^{-0.7} \Delta_b/\Gamma_{\mathrm{HI}}$. Here, $T$ is the gas temperature, $\Delta_b$ is the gas density in units of the mean, and $\Gamma_{\mathrm{HI}}$ is the photoionization rate which, in turn, scales in proportion to the strength of the local ionizing background. Recently, \citet{2015MNRAS.447.3402B} showed that the observed transmission averaged over $50 h^{-1}~\mathrm{Mpc}$ segments of the forest displays an increasingly large scatter from sightline to slightline at redshifts approaching $z=6$. At $z\gtrsim 5.5$ this scatter is too large to be explained by variations in the IGM density alone, indicating the presence of large spatial fluctuations in the ionizing background and/or in the IGM temperature \citep{2015arXiv150501853C, 2015ApJ...813L..38D, 2015arXiv150907131D, 2016arXiv160503183G}.
\citet{2015ApJ...813L..38D} proposed a model for the latter in which the $z\sim 5.5$ opacity variations are generated largely by relic temperature fluctuations from patchy reionization. In this scenario, the observed opacity fluctuation amplitude favors an extended but late-ending reionization process that was approximately half-complete by $z\approx 9$ and that ended at $z\approx 6$. \citet{2015arXiv150907131D} argued that the opacity variations might not be tied to the reionization process at all; they might instead be a consequence of large fluctuations in the ionizing background well after the end of reionization. In their model, these fluctuations are driven by the clustering of galaxies and by spatial variations in mean free path of ionizing photons. The latter owes to the enhancement (suppression) of optically thick absorbers in voids (overdensities), where the local ionizing background is weaker (stronger).
{ In paper I, we considered the model of \citet{2015arXiv150907131D} in further detail. Under the standard assumption that galaxies dominate the background, we showed that accounting for the $z\approx5.5$ opacity fluctuations requires a short spatially averaged mean free path of $\langle \lambda_{\mathrm{mfp}}^{912} \rangle \lesssim 15 h^{-1}~\mathrm{Mpc}$, a factor of $\approx$ 2 lower than is expected from extrapolations of observations at $z \leq 5.2$ \citep{2014MNRAS.445.1745W}. We further showed that connecting this value with the measured mean free path at $z=5.2$ ($44 \pm 7 h^{-1}~\mathrm{Mpc}$, \citealt{2014MNRAS.445.1745W}) requires an unnatural factor of $\approx2$ decrease in the emissivity of the galaxy population in the $\approx100$ million years between $z=5.6$ and $z=5.2$. Such a rapid evolution in the galaxy population would be surprising because the Hubble time scale is of order one billion years at these redshifts. This is (to within a factor of a few) the time scale over which we should expect such a large change in the galaxy emissivity.\footnote{Recently, \citet{2016arXiv160503183G} claimed that the observed opacity fluctuations are well reproduced by ionizing background fluctuations in fully-coupled radiative transfer$+$hydrodynamics simulations with a standard galactic source model, and without the need for a small $\langle \lambda_{\mathrm{mfp}}^{912} \rangle$. However, in Paper I, some of us argue that their methodology is likely to overestimate the amplitude of $\tau_{\mathrm{eff}}$ fluctuations. } However, we also identified a plausible solution to this problem. We showed that, at $z\gtrsim 5$, the enhanced ionizing flux in the proximity zones of quasars can bias direct measurements of the mean free path higher than $\langle \lambda_{\mathrm{mfp}}^{912} \rangle$ by up to a factor of two. Such a large bias would reconcile the short values of $\langle \lambda_{\mathrm{mfp}}^{912} \rangle$ that are required in the model of \citet{2015arXiv150907131D}. }
{ Given the large uncertainty in the magnitude of this bias (see discussion in Paper I), we explore an alternative model here. } One way to obviate the need for a short mean free path is to make the sources of the ionizing background in the post-reionization IGM much rarer and brighter than the faint, sub-$L_*$ galaxies that dominate in standard models. As \citet{2015arXiv150501853C} pointed out, the AGN of G2015, if they exist, are natural candidates for these rare sources. In this section we shall consider an AGN-sourced ionizing background as a potential solution the problem of large opacity variations in the $z \approx 5.5$ Ly$\alpha$ forest. In what follows, we focus on the effect of background fluctuations from AGN under the implicit assumption that residual temperature fluctuations from the H{\sc~i}\ reionization process are weak at $z\lesssim 5.5$. This assumption is appropriate for scenarios in which reionization ends significantly earlier than $z\approx 6$ and/or occurs rapidly.
\subsection{Numerical Methodology}
\label{SEC:methodology}
Much of the numerical methodology for this work is described in paper I. Here we briefly summarize the methodology and its extension to include AGN sources. We used a modified version of the code of \citet{2004NewA....9..443T} to run a cosmological hydrodynamics simulation with box size $L_{\rm box} = 200 h^{-1}~\mathrm{Mpc}$, with $N_{\rm dm}= 2048^3$ dark matter particles and $N_{\rm gas}=2048^3$ gas cells. This simulation does not include residual temperature fluctuations from H{\sc~i}\ reionization (as noted above), nor does it include temperature fluctuations from He{\sc~ii}\ reionization. In Appendix \ref{SEC:model_uncertainties}, we discuss the impact of this on our conclusions. To model the galaxy population, all dark mater halos with masses $M_{200}\geq 2\times 10^{10}h^{-1}~M_{\odot}$ are populated with galaxies by abundance matching to the observed luminosity function of \citet{2015ApJ...803...34B}, using the scheme described in \citet{2015ApJ...813...54T}. This minimum mass, which was chosen for completeness of the halo mass function, translates to a lower magnitude limit of $M_{\rm AB,1600} \approx -17.5$ at $z=5.5$, above the detection threshold for current observations. We assume a constant escape fraction, which we tune in each model to match the observed mean transmission in the Ly$\alpha$ forest (see below for further details).
Since our models in this paper also include AGN as sources, which are much rarer than typical galaxies, we tile together eight of our hydro simulation boxes for an effective box size of $L_{\rm box}=400 h^{-1}~\mathrm{Mpc}$ before laying down AGN sources. We randomly populate the most massive halos\footnote{For reference, the single most massive halo in our box at $z=5.2, 5.4$ and $5.6$ has $M_{200} = 8, 6,$ and $5.5\times10^{12}~h^{-1}M_{\odot}$, respectively.} in this composite box with luminosities drawn from a rescaled version of the G2015 luminosity function, where the normalization is chosen to yield the desired level of contribution from AGN. (See Appendix \ref{SEC:simLFs} for more details.) We draw AGN down to a magnitude limit of $M_{\mathrm{AB},1450}=-19$, roughly the magnitude of the faintest candidates in G2015's sample. We assume an AGN spectrum with $L_{\nu} \propto v^{-0.6}$ for $\lambda>912$ \AA, and $L_{\nu} \propto \nu^{-1.7}$ for $\lambda\leq912$ \AA, consistent with the quasar stack of \citet{2015MNRAS.449.4204L}. We note that, while the escape fraction of H{\sc~i}\ ionizing photons from quasars is likely close to unity, it is unclear whether this also holds for fainter AGN. One of our main goals is to assess the maximum possible contribution of AGN to the ionizing background, so we follow \citet{2015A&A...578A..83G} and \citet{2015ApJ...813L...8M} in assuming a constant escape fraction of unity for all AGN.
After populating the sources, we post-process the hydro simulation with the model of \citet{2015arXiv150907131D} for the fluctuating ionizing background. This model self-consistently includes the effect of spatial variations in the mean free path. We compute $\Gamma_{\mathrm{HI}}$ on a uniform grid with $N=64^3$, yielding a cell size of $\Delta x = 6.25 h^{-1}~\mathrm{Mpc}$. In all models we set the spatially averaged mean free path to $\langle \lambda_{\mathrm{mfp}}^{912} \rangle = 40, 35$, and $30~h^{-1}\mathrm{Mpc}$ at $z=5.2, 5.4$ and $5.6$, respectively, consistent with the extrapolation of the measurements of \citet{2014MNRAS.445.1745W}. In what follows, we will explore galaxies$+$AGN source models in which AGN constitute different fractions (i.e. $25, 50$ and $90 ~\%$) of the global average photoionization rate, $\langle \Gamma_{\mathrm{HI}} \rangle$. The left, middle, and right panels of Fig. \ref{FIG:visB} show slices through the AGN distribution, photoionization rate, and mean free path field at $z=5.6$ for our model where AGN contribute 50\% of $\langle \Gamma_{\mathrm{HI}} \rangle$. In the left panel we show all of the AGN in the slice, which has a thickness of $20h^{-1}~\mathrm{Mpc}$. The right two panels have thicknesses of $6.25h^{-1}~\mathrm{Mpc}$, and are situated in the middle of the slice in the left panel. The fluctuations in $\Gamma_{\mathrm{HI}}$ are enhanced by large-scale variations in the mean free path of ionizing photons, which is significantly longer in regions surrounding an overdensity of galaxies and/or several bright AGN. { For reference, Fig. \ref{FIG:MFPdist} shows the distribution of $\lambda_{\mathrm{mfp}}^{912}(\boldsymbol{x})$ at $z=5.6$ in our $25\%$, $50\%$, and $90\%$ AGN models (from narrowest to widest, respectively). Fluctuations in $\Gamma_{\mathrm{HI}}$ are enhanced not only by a larger contribution from AGN, but also from the correspondingly larger fluctuations in $\lambda_{\mathrm{mfp}}^{912}(\boldsymbol{x})$. }
\begin{figure}
\begin{center}
\resizebox{8.8cm}{!}{\includegraphics{MFPhistograms.eps}}
\end{center}
\caption{ { Distribution of mean free paths at $z=5.6$ in our $25\%$, $50\%$, and $90\%$ AGN models (narrowest to widest, respectively). Fluctuations in the ionizing background are enhanced not only by a larger contribution from AGN, but also from the correspondingly larger fluctuations in the mean free path. } }
\label{FIG:MFPdist}
\end{figure}
Under the assumption of photoionization equilibrium (a good approximation in the post-reionization IGM), we combine these simulations of the fluctuating $\Gamma_{\mathrm{HI}}$ field with our hydro simulations to produce synthetic Ly$\alpha$ forest spectra. A typical way to quantify opacity fluctuations in the Ly$\alpha$ forest is to measure the distribution of effective optical depths, $\tau_{\mathrm{eff}} \equiv -\ln \langle F \rangle_{L}$, for segments of the forest of length $L$, where $F$ is the continuum normalized flux. The most recent measurements by \citet{2015MNRAS.447.3402B} adopted the convention of $L=50 h^{-1}~\mathrm{Mpc}$, so we will assume the same hereafter. We construct the distribution of $\tau_{\mathrm{eff}}$ from 4000 randomly oriented lines of sight. The spatially averaged photoionization rate, $\langle \Gamma_{\mathrm{HI}} \rangle$, is a free parameter in our models that we fix by matching the observed transmission in the forest. Specifically, we rescale $\Gamma_{\mathrm{HI}}$ by a constant factor such that the mean value of $\langle F \rangle_{50}$ in our models is equal to the observed value in \citet{2015MNRAS.447.3402B}. For a given thermal state of the IGM, this normalization of $\langle \Gamma_{\mathrm{HI}} \rangle$ also fixes the emissivities of the sources in our models.
\subsection{Results}
\begin{figure}
\begin{center}
\resizebox{8.8cm}{!}{\includegraphics{taueffCDF_AGN.eps}}
\end{center}
\caption{Opacity fluctuations in the $z>5$ Ly$\alpha$ forest in models where AGN contribute to the H{\sc~i}\ ionizing background. The green/long-dashed, red/solid, and magenta/dot-dashed curves correspond to models in which AGN emissions produce $25\%$, $50\%$ and $90\%$ of the global average H{\sc~i}\ photoionization rate, respectively. The black histograms show the cumulative probability distribution of $\tau_{\mathrm{eff}}$ measured in $50h^{-1}\mathrm{Mpc}$ segments of the forest by \citet{2015MNRAS.447.3402B}. For reference, the blue/short-dashed curves correspond to a model in which only galaxies contribute to the ionizing background. The light blue shading shows $90\%$ confidence regions estimated by bootstrap sampling of this model. For all models we assume our fiducial values for the spatially averaged mean free path, $\langle \lambda_{\mathrm{mfp}}^{912} \rangle = 40, 35$, and $30~h^{-1}\mathrm{Mpc}$ at $z=5.2, 5.4$, and $5.6$, consistent with the extrapolation of measurements by \citet{2014MNRAS.445.1745W}.}
\label{FIG:taueff_AGN}
\end{figure}
\begin{figure}
\begin{center}
\resizebox{8.8cm}{!}{\includegraphics{AGNemissivityB.eps}}
\end{center}
\caption{The H{\sc~i}\ ionizing emissivities of AGN in our AGN-driven models (see Fig. \ref{FIG:taueff_AGN}). From bottom to top, the black Xs show the $z=5.6$ ionizing emissivities in our $25\%$, $50\%$, and $90\%$ AGN models, respectively, while the corresponding error bars bracket the effects of the highly uncertain thermal state of the IGM. For clarity, we have added a small horizontal offset between these data points. The other data points show some of the high-$z$ measurements from Fig. \ref{FIG:emissivity}. The curves correspond to parametric models that we use in \S \ref{SEC:thermalhistory} to explore the implications of an AGN-dominated ionizing background for He{\sc~ii}\ reionization and the thermal history of the IGM (see main text for details). }
\label{FIG:AGNemissivityB}
\end{figure}
\label{SEC:AGNmodels}
Fig. \ref{FIG:taueff_AGN} shows the cumulative probability distribution of the effective optical depth, $P(<\tau_{\mathrm{eff}})$, in three models where AGN emissions constitute a significant fraction of the ionizing background. The green/long-dashed, red/solid and magenta/dot-dashed curves respectively correspond to scenarios in which AGN contribute $25\%$, $50\%$, and $90\%$ of $\langle \Gamma_{\mathrm{HI}} \rangle$.\footnote{We note that the ionizing background fluctuations in these models are greatly enhanced by large-scale spatial variations in the mean free path. To illustrate this, we have also considered models in which the mean free path is assumed to be uniform in space, with $\lambda_{\mathrm{mfp}}^{912}=30h^{-1}~\mathrm{Mpc}$ at $z=5.6$. (Note, all studies prior to \citealt{2015arXiv150907131D} that implemented a fluctuating ionizing background did so under the assumption of a uniform mean free path.) In this case, the $50\%$ AGN model yields a $P(<\tau_{\mathrm{eff}})$ that is very similar to the green/long-dashed curve in Fig. \ref{FIG:taueff_AGN}, and the $90\%$ AGN model is nearly identical to the red/solid curve. Thus, in these models it is crucial to account for spatial variations in the mean free path. } The black histograms show the observational measurements of \citet{2015MNRAS.447.3402B} while, for reference, the blue/short-dashed curves show our fiducial model from Paper I in which only galaxies source the ionizing background.\footnote{In paper I we show that the $\tau_{\mathrm{eff}}$ distribution in this ``galaxies-only" model is nearly identical to that in a model with spatially uniform $\Gamma_{\mathrm{HI}}$. } The blue shading corresponds to $90\%$ confidence regions estimated by bootstrap sampling of this galaxies-only model (where we fix the mean $\langle F \rangle_{50}$ of each sample
to the observed value). The $\tau_{\mathrm{eff}}$ distribution in the $25\%$ AGN model is only mildly wider than in the galaxies-only model, and is consistently narrower than the observed distribution. The $50\%$ AGN model provides a better match for much of the observed distribution at all redshifts, but the inset in the bottom panel of Fig. \ref{FIG:taueff_AGN} shows that even this model struggles to account for the highest opacity measurement at $z=5.6$. On the other hand, while the $90\%$ AGN model can better account for the highest opacities, the distributions in this model are otherwise consistently wider than the observed distributions.
{ An attractive feature of AGN-dominated models is that they do not require a short $\langle \lambda_{\mathrm{mfp}}^{912} \rangle$ to account for (most of) the observed $\tau_{\mathrm{eff}}$ dispersion. (Indeed, recall that we have assumed a $\langle \lambda_{\mathrm{mfp}}^{912} \rangle(z)$ consistent with extrapolating the $z\lesssim 5.2$ measurements of \citealt{2014MNRAS.445.1745W}.) Thus, unlike the galaxies-only model considered in paper I, they do not require an unnaturally rapid evolution in the emissivity of the galaxy population to be consistent with the mean free path measurements of \citealt{2014MNRAS.445.1745W}. In fact, we find that the redshift evolution of the galaxy emissivity is quite flat. For reference, in our $25\%$ AGN model, the mean emissivity from galaxies is $\langle \epsilon^{\rm gal}_{912}\rangle = 4.7, 4.7$, and $5.0\times10^{24}~\mathrm{erg~s^{-1}}~\mathrm{Mpc}^{-3}~\mathrm{Hz^{-1}}$ at $z=5.2, 5.4$ and $5.6$, respectively - a decrease of only $\approx 6\%$ between $z=5.6$ and $z=5.2$. In Paper I, we argued that the mean free path measurements of \citealt{2014MNRAS.445.1745W} may be biased significantly by the enhanced ionizing flux in quasar proximity zones. We note that such a bias could reduce the contribution from AGN that is required to account for the observed width of $P(<\tau_{\mathrm{eff}})$. }
Let us now quantify in more detail the AGN contribution to the ionizing emissivities in these models, so we can compare them to the measurements of G2015. The value of the emissivity (which is determined by the observed mean transmission in the forest) depends on the thermal state of the IGM, which is highly uncertain at $z \gtrsim 5$. In the AGN-dominated models, these uncertainties are made larger by the fact that the AGN population may begin reionizing He{\sc~ii}\ earlier than in the standard scenario. For a simple estimate of the upper and lower limits of these uncertainties, we parameterize the thermal state of the IGM with a temperature-density relation of the form $T(\Delta) = T_0 \Delta^{\gamma-1}$, where $T_0$ is the temperature at the cosmic mean density, and $\gamma$ specifies the logarithmic slope of the relation \citep{1997MNRAS.292...27H}. Here we bracket the range of plausible thermal states with $(T_0,\gamma)=(20,000~\mathrm{K},1.2)$, and $(T_0,\gamma)=(5,000~\mathrm{K},1.6)$. The former would be expected during or shortly after a reionization event, while the latter represent approximately the asymptotic values expected well after afterwards \citep{1997MNRAS.292...27H,2015MNRAS.450.4081P,2015arXiv150507875M,2015arXiv151105992U}. Using these values, we rescale the AGN emissivity according to the scaling relations derived in \citet{2013MNRAS.436.1023B},
\begin{equation}
\epsilon_{912}(T_0,\gamma) \propto T_0^{-0.575} e^{ 0.7 \gamma},
\label{EQ:GammaRescale}
\end{equation}
to obtain lower and upper limits, respectively. At $z=5.6$, the ionizing emissivities in our $25 \%$, $50 \%$, and $90 \%$ AGN models are $\epsilon^{\rm AGN}_{912} = 1.5^{+1.2}_{-0.6}, 2.8^{+2.3}_{-1.1},$ and $5.3^{+4.2}_{-2.0} \times 10^{24}~\mathrm{erg~s^{-1} Hz^{-1} Mpc^{-1}}$, respectively.\footnote{See Paper I for a discussion of how we correct our emissivities for the effects of finite simulation resolution. Here we also apply a crude correction to account for the fact that $\Gamma_{\mathrm{HI}}$ receives a contribution from ionizing radiation produced during recombinations to the ground state of hydrogen \citep{1996ApJ...461...20H}. At temperatures of $T\sim10,000$ K, the fraction of recombinations that produce an H{\sc~i}\ ionizing photon is $(\alpha_{\rm A} - \alpha_{\rm B})/\alpha_{\rm A} \approx 40\%$, where $\alpha_{\rm A}$ and $\alpha_{\rm B}$ are the case A and B recombination coefficients, respectively. However, just a fraction of these photons are added to the ionizing background, since some of the recombinations will occur in optically thick absorbers, whose abundance is set by the H{\sc~i}\ column density distribution. For standard assumptions about the column density distribution, this fraction is roughly one half. Thus we reduce our emissivities by $20\%$. Previous studies that considered this radiative transfer effect in more detail have obtained similar corrections ranging from $\approx 10 - 30\%$ at these redshifts \citep{2009ApJ...703.1416F}.}
{ Fig. \ref{FIG:AGNemissivityB} compares the emissivities in our AGN-driven models to recent measurements in the literature. (The curves will be described in the next section.) From bottom to top, the black Xs show the $z=5.6$ ionizing emissivities in our $25\%$, $50\%$, and $90\%$ AGN models, respectively. For clarity, we have added a small horizontal offset between these data points. Note that the emissivities in these models are of order the emissivity reported by G2015. Even the 25\% AGN model has an emissivity that is a factor of $5-14$ higher than the value obtained by interpolating in redshift between the measurements of \citet{2013ApJ...768..105M} and \citet{2010AJ....139..906W}. In the next section, we will discuss the implications of such a large ionizing emissivity from AGN for He{\sc~ii}\ reionization and for the thermal history of the IGM. }
\section{Helium Reionization and the Thermal History of the IGM}
\label{SEC:thermalhistory}
Compared to models in which galaxies dominate the ionizing background, the large AGN populations that we considered in the last section would likely emit many more photons with energies in excess of $4$ Ry. These emissions would begin the reionization of intergalactic He{\sc~ii}\ earlier than in the standard scenario. In this section we explore the signature of this earlier He{\sc~ii}\ reionization in the thermal history of the IGM.
Below we consider four models with H{\sc~i}\ ionizing emissivities of AGN given by the curves in Fig. \ref{FIG:AGNemissivityB}. The form of the blue/short-dashed curve is given by equation (\ref{EQ:AGNemStandard}) in Appendix \ref{SEC:emissivitymodels}. This model is representative of the standard scenario in which galaxies dominate $\epsilon_{912}$ at $z\gtrsim4$. In contrast, the magenta/dot-dashed curve shows the model of \citet{2015ApJ...813L...8M}, in which AGN provide all of the ionizing emissivity necessary to reionize H{\sc~i}\ by $z\approx5.5$. { This model is representative of the $z=5.6$ emissivity in the $90\%$ AGN model from the last section. The red/solid and green/long-dashed models are representative of the emissivities in the $50\%$ and $25\%$ AGN models, respectively. } The form of these models is given by equation (\ref{EQ:modMH2015}).
We solve the ionization balance equations to obtain the H{\sc~i}\ and He{\sc~ii}\ reionization histories, tuning the escape fraction of galaxies such that H{\sc~i}\ reionization ends at $z\approx6$ in the three models that include galaxies. For stellar sources we use the star formation rate density compiled by \citet{2015ApJ...802L..19R}. The thin and thick sets of curves in Fig. \ref{FIG:temphistories}a show $\langle x_{\mathrm{HII}}\rangle$ and $\langle x_{\mathrm{HeIII}}\rangle$, the volume-weighted mean ionized fractions of H{\sc~ii}\ and He{\sc~iii}\ in these models. Our fiducial calculations assume constant clumping factors of $C_{\mathrm{HII}}=C_{\mathrm{HeIII}}=2$, approximately consistent with the values found in hydrodynamical simulations \citep[e.g.][]{2009MNRAS.394.1812P,mcquinn-Xray,2015ApJ...810..154K}, but we will discuss the effect of varying $C_{\mathrm{HeIII}}$ below.
\begin{figure}
\begin{center}
\resizebox{8.2cm}{!}{\includegraphics{T_histories.eps}}
\resizebox{8.2cm}{!}{\includegraphics{Thist_optimum_deltas.eps}}
\end{center}
\caption{ The effect of early He{\sc~ii}\ reionization on the thermal history of the IGM. Panel (a): global H{\sc~i}\ and He{\sc~ii}\ reionization histories (thin and thick, respectively) for models with H{\sc~i}\ ionizing emissivities of AGN shown in Figure \ref{FIG:AGNemissivityB}. The cyan/dot-dashed curves correspond to the model of \citet{2015ApJ...813L...8M} in which AGN are assumed to be the only sources of H{\sc~i}\ reionization. For all other models, we tune the escape fraction of galaxies such that H{\sc~i}\ reionization ends at $z\approx6$. The dotted curve corresponds to a version of the green/long-dashed model in which He{\sc~ii}\ reionization has been extended by increasing the recombination rate of He{\sc~iii}. Panel (b): Temperature at the mean density of the universe. The data points show the Ly$\alpha$ forest temperature measurements of \citet{2011MNRAS.410.1096B} (triangles) and \citet{2014MNRAS.441.1916B} (circles). The blue and red data points show these measurements extrapolated to the mean density using the temperature-density relation of the blue/short-dashed and red/solid models, respectively. { For reference, panel (c) shows the comparison at the redshift-dependent ``optimal density," $\bar{\Delta}$, at which the temperature measurements were made. } }
\label{FIG:temphistories}
\end{figure}
To model the thermal history of the IGM, we use a modified version of the two-zone approach of \citet{2015arXiv151105992U}. In summary, we follow the thermal histories of an ensemble of gas parcels at different initial densities, using the Zel'dovich approximation for the parcels' dynamical histories. Starting at $z^{\mathrm{HI}}_{\mathrm{start}}=12$, the redshift at which H{\sc~i}\ reionization begins in our models, the parcels are photoheated by a uniform background of hard photons produced by the AGN population. To delineate this hard background from the softer photons that are absorbed more locally at the boundary of ionization fronts, we include in the hard background only photons with mean free paths $>20~\mathrm{Mpc}$, but our results are insensitive to this choice. We adopt a spectral index of $\alpha_{\rm QSO}=1.7$ for the hard background, and an index of $\alpha_{\rm bk}=1$ for the H{\sc~i}\ ionizing background. As H{\sc~i}\ reionization progresses, the parcels are each instantaneously heated (in theory by a passing ionization front) to a temperature of $T=20,000$ K at a unique redshift, $z=z^{\mathrm{HI}}_{\mathrm{reion}}$, with cumulative probability distribution $\langle x_{\mathrm{HII}}\rangle$. After $z^{\mathrm{HI}}_{\mathrm{reion}}$, the hard background continues heating by He{\sc~ii}\ photoionizations until $z=z^{\mathrm{HeII}}_{\mathrm{reion}}$, at which times the parcels are instantaneously heated by $\approx8,000$ K [see eq. 12 of \citet{2015arXiv151105992U}]. This two zone model approximates the heating found in radiative transfer simulations of He{\sc~ii}\ reionization \citep{2009ApJ...694..842M,2013MNRAS.435.3169C, 2016arXiv161002047L}, and is motivated by the steep dependence of the mean free path on photon energy.
In models with a large AGN emissivity at high redshifts, $z^{\mathrm{HI}}_{\mathrm{reion}}$ and $z^{\mathrm{HeII}}_{\mathrm{reion}}$ are likely to be strongly correlated. To account for this correlation, we assume as an approximation the deterministic relation
\begin{equation}
z^{\mathrm{HeII}}_{\mathrm{reion}}(z^{\mathrm{HI}}_{\mathrm{reion}})=z^{\mathrm{HI}}_{\mathrm{start}}\exp\left[-B_1(z^{\mathrm{HI}}_{\mathrm{start}}-z^{\mathrm{HI}}_{\mathrm{reion}})\right]-B_2.
\end{equation}
Given the exact solution for $\langle x_{\mathrm{HII}}\rangle(z)$ in our models, we find that the values $B_1=(0.035,0.068,0.089,0.142) $ and $B_2=(6.8,4.5,3.2,0.05)$ yield excellent approximations to $\langle x_{\mathrm{HeIII}}\rangle(z)$ for the blue/short-dashed, green/long-dashed, red/solid, and magenta/dot-dashed models, respectively. We find only minor differences in the resulting thermal histories if we assume the opposite limit in which $z^{\mathrm{HI}}_{\mathrm{reion}}$ and $z^{\mathrm{HeII}}_{\mathrm{reion}}$ are completely uncorrelated.
The curves in Fig. \ref{FIG:temphistories}b show the IGM temperature at the mean density in our models. (The line styles match the emissivity histories in Fig. \ref{FIG:AGNemissivityB}.) We compare them against the Ly$\alpha$ forest temperature measurements of \citet{2011MNRAS.410.1096B} (triangles) and \citet{2014MNRAS.441.1916B} (circles). { It is important to note that these measurements are not at the mean density, but at a redshift-dependent optimal density, $\bar{\Delta}$ (see discussion in \citealt{2011MNRAS.410.1096B}). The blue and red symbols show these measurements extrapolated to the mean density using the temperature-density relation of the blue/short-dashed and red/solid models, respectively. For reference, Fig. \ref{FIG:temphistories}c shows the comparison at the original $\bar{\Delta}$ of the measurements. These results show that the AGN-driven models considered in the last section are qualitatively inconsistent with the temperature measurements owing to the earlier reionization of He{\sc~ii}\ and its associated photoheating. Only the standard scenario (blue/short-dashed curve), in which He{\sc~ii}\ reionization ends at $z\approx3$, provides a reasonable match to the high-redshift temperatures (see also \citealt{2015arXiv151105992U}). }
\begin{figure}
\begin{center}
\resizebox{8.4cm}{!}{\includegraphics{T_histories_earlyHIreion.eps}}
\end{center}
\caption{ { Same as in Fig. \ref{FIG:temphistories}b, except we assume that H{\sc~i}\ reionization ends at $z\approx8$, much earlier than in our fiducial model (in which it ends at $z\approx6$). Regardless of the details of H{\sc~i}\ reionization, the earlier onset of He{\sc~ii}\ reionization in AGN-dominated models renders them discrepant with the high-redshift temperature measurements. } }
\label{FIG:earlyHIreion}
\end{figure}
One way to reconcile the AGN-driven scenarios with observations is if He{\sc~ii}\ reionization were significantly more extended than in our models. To explore this possibility, we increased $C_{\mathrm{HeIII}}$ in the green/long-dashed model while holding H{\sc~i}\ reionization fixed. We found that delaying the end of He{\sc~ii}\ reionization to $z\approx3$ in this model requires $C_{\mathrm{HeIII}}=4$, consistent with the highest values bracketed by the radiative transfer simulations of \citet{mcquinn-Xray}. This model is depicted as the green/dotted curves in Fig. \ref{FIG:temphistories}. However, panel \ref{FIG:temphistories}b shows that the longer duration of He{\sc~ii}\ reionization results in larger temperatures owing to the longer build-up of the hard radiation background from AGN and its associated photoheating. Thus extending the duration of He{\sc~ii}\ reionization appears to be in tension with the temperature measurements. We note that this issue can be avoided if the fainter AGN, which account for most of the contribution to the H{\sc~i}\ ionizing emissivity (see \S \ref{SEC:lumfunc}), have a much smaller escape fraction of He{\sc~ii}\ ionizing photons. However, a large escape fraction of H{\sc~i}\ ionizing radiation would still be necessary to account for the observed Ly$\alpha$ forest opacity fluctuations considered in the last section.
{ Finally, Fig. \ref{FIG:earlyHIreion} considers the impact of our assumptions about H{\sc~i}\ reionization on our conclusions. There, we adopt a scenario in which H{\sc~i}\ reionization ends at $z\approx 8$, much earlier than in our fiducial model, such that the IGM starts out significantly colder at $z\approx6$. (Note that we do not consider the AGN-only model of \citet{2015ApJ...813L...8M}, since its H{\sc~i}\ reionization history is fixed by the AGN emissivity.) We find that, regardless of the details of H{\sc~i}\ reionization, the earlier onset of He{\sc~ii}\ reionization still renders AGN-dominated models ($\gtrsim 50\%$ contribution from AGN to the ionizing background) discrepant with the high-redshift temperature measurements. }
We conclude that, under standard assumptions about the spectra of AGN, it is difficult to reconcile AGN-driven models of the high-$z$ ionizing background -- especially scenarios with AGN emissivities as high as those claimed by G2015 -- with current observational constraints on the thermal history of the IGM. On the other hand, recent observations of the He{\sc~ii}\ Ly$\alpha$ forest indicate a lower effective optical depth at $z=3.1-3.3$ than is found in existing simulations of He{\sc~ii}\ reionization \citep{2014arXiv1405.7405W}. Some authors have argued that this low opacity is evidence for a more extended or earlier He{\sc~ii}\ reionization \citep{2014arXiv1405.7405W,2015ApJ...813L...8M}. In the next section we will address these claims.
\section{The He{\sc~ii}\ Ly$\alpha$ Forest}
\label{SEC:HeIIforest}
Observations of the He{\sc~ii}\ Ly$\alpha$ forest have been used to probe the timing and duration of He{\sc~ii}\ reionization. The He{\sc~ii}\ Ly$\alpha$ forest displays Gunn-Peterson troughs at $z\geq2.7$ coeval with cosmic voids observed in the H{\sc~i}\ Ly$\alpha$ forest. The presence of these He{\sc~ii}\ Ly$\alpha$ absorption troughs imply a He{\sc~ii}\ fraction $\gtrsim10\%$ at the mean density, indicating that He{\sc~ii}\ reionization was likely still in progress at $z\geq2.7$ \citep{2009ApJ...704L..89M}. We note that this is qualitatively consistent with the blue/short-dashed model in Fig \ref{FIG:temphistories}, representative of the standard scenario in which AGN are subdominant at $z>4$. On the other hand, \citet{2015ApJ...813L...8M} argued that a more extended (or earlier) He{\sc~ii}\ reionization may be more consistent with the observed evolution in the He{\sc~ii}\ Ly$\alpha$ effective optical depth reported by \citet{2014arXiv1405.7405W}. At the highest redshifts probed by the data ($z=3.1-3.3$), \citet{2014arXiv1405.7405W} found lower He{\sc~ii}\ Ly$\alpha$ optical depths than existing simulations of He{\sc~ii}\ reionization (which use standard quasar emissivity models and hence have He{\sc~ii}\ reionization end at $z\sim3$).
Existing He{\sc~ii}\ reionization simulations are likely to under-predict the optical depths of these measurements. Early on during He{\sc~ii}\ reionization, the regions that show significant transmission are likely the proximity regions of quasars. The $z\gtrsim 3$ He{\sc~ii}\ Ly$\alpha$ forest appears consistent with this picture, having large spans of no transmission punctuated by rare transmission peaks. In this limit in which all of the He{\sc~ii}\ Ly$\alpha$ transmission comes from proximity regions, the transmission in the forest is given by
\begin{equation}
{\cal T} \approx \int \mathrm{d} L \; \Phi(L) V_{\rm eff}(L),
\label{eqn:HeII_trans}
\end{equation}
where $V_{\rm eff}(L)$ is the effective volume encapsulated by the average transmission profile around an individual quasar of luminosity $L$. Using the 1D radiative transfer code presented in \citet{2015arXiv151103659K}, we find $V_{\rm eff} \propto L^{1.5}$, regardless of whether the quasar turns on in a He{\sc~ii}\ region or in a He{\sc~iii}\ region, the same scaling expected for the classical proximity effect assuming ionization equilibrium. However, we find that the proportionality changes by a factor of ten if quasar lifetimes vary from $10-100\;$Myr because of equilibration effects \citep{2015arXiv151103659K} and by an order unity factor if the quasar goes off in a He{\sc~ii}\ region rather than in a He{\sc~iii}\ region. This super-linear scaling means that $>L_*$ quasars contribute almost all of the transmission.
This simple model for the He{\sc~ii}\ Ly$\alpha$ transmission illustrates two potential problems with the He{\sc~ii}\ reionization simulations. First, the simulations are typically limited to one or two quasar models, and the amount of transmission should be very sensitive to the quasar model. We find that the transmission varies by a factor of ten depending on the lifetime that is assumed. Indeed, most of the simulations in \citet{2009ApJ...694..842M} took the brightest quasars to have the shortest lifetimes, following \citet{2006ApJS..163....1H}, a choice that reduces the transmission relative to the more standard lightbulb models. Second, simulations tend to miss quasars that occur at lower number densities than one per simulation volume. For the volumes of previous simulations and using our model for the transmission, this omission leads to an underestimate of the transmission. To mock up this effect in our model, we set the integrand in equation~(\ref{eqn:HeII_trans}) to zero at $ L \Phi= V^{-1}$, where $V$ is the simulation volume. At $z=3.3$, we find that this cutoff reduces the transmission, ${\cal T}$, by a factor of $1.8$ for the $130h^{-1}\;$Mpc simulation box used in \citet{2009ApJ...694..842M} and by a factor of $2.0$ for the $100 h^{-1}~$Mpc box used in \citet{2013MNRAS.435.3169C}. These factors increase to $2.3$ and $2.7$ respectively at $z=3.6$. Thus, because of their limited volumes, existing He{\sc~ii}\ reionization simulations underestimate the observed transmission (and hence $\tau_{\rm eff}$) in the He{\sc~ii}\ forest in the limit that the transmission owes to isolated proximity effects. While this limit does not apply at the end of He{\sc~ii}\ reionization, when most locations see multiple quasars, it is likely to apply at the highest redshifts probed by the data, redshifts where the discrepancy reported by \citet{2014arXiv1405.7405W} arises. Without larger volume simulations spanning a range of quasar lifetime models, the discrepancy between simulations and the \citet{2014arXiv1405.7405W} measurements should not be taken as evidence for a more extended or earlier He{\sc~ii}\ reionization epoch. Our simple model further suggests that the transmission in the high-redshift He{\sc~ii}\ forest is less sensitive to the He{\sc~ii}\ fraction than to the properties of quasar lifetimes.
\section{Conclusion}
\label{SEC:conclusion}
Several recent studies have proposed models in which AGN contribute significantly to (or even dominate) the $z\gtrsim 5$ H{\sc~i}\ ionizing background. Among the main motivations for these models is that they may reconcile recent observations showing large opacity fluctuations in the $z\approx5.5$ H{\sc~i}\ Ly$\alpha$ forest \citep{2015MNRAS.447.3402B}, and a slow evolution in the mean opacity of the $z\approx 3.1-3.3$ He{\sc~ii}\ Ly$\alpha$ forest. Here we have considered several facets of AGN-driven models of the ionizing background: (1) We have quantified the amplitude of $z\approx 5.5$ H{\sc~i}\ Ly$\alpha$ forest opacity fluctuations in models with varying levels of contribution from AGN; (2) We have investigated the implications of these models for cosmological He{\sc~ii}\ reionization and for the thermal history of the IGM; (3) We have discussed the interpretation of recent $\approx 3.1-3.3$ He{\sc~ii}\ Ly$\alpha$ forest opacity measurements in the context of these models.
We found that a model in which $\approx 50\%$ of the H{\sc~i}\ ionizing background is sourced by AGN generally provides a better fit to the observed Ly$\alpha$ forest opacity fluctuation distribution across $z\approx 5-5.7$ compared to a model in which only galaxies source the background. However, even this $50\%$ AGN model struggles to account for the highest opacity measurements. We found that doing so requires that essentially all ($\gtrsim 90\%$) of the ionizing background be produced by AGN, in a similar vein to the model proposed recently by \citet{2015ApJ...813L...8M}. These results are generally consistent with the findings of \citet{2016arXiv160608231C}, which appeared during the final preparations of this manuscript. { An important caveat to this work is that we have adopted a model for the mean free path that is consistent with the recent measurements of \citealt{2014MNRAS.445.1745W}. In Paper I, we argued that these measurements may be biased significantly by the enhanced ionizing flux in quasar proximity zones. We note that such a bias could reduce the contribution from AGN that is required to account for the observed amplitude of Ly$\alpha$ opacity fluctuations.}
Since AGN are expected to have harder spectra than galaxies at He{\sc~ii}\ ionizing energies of $\geq 4$ Ry, a unique prediction of AGN-driven models is that cosmological He{\sc~ii}\ reionization occurs earlier than in the standard scenario \citep{2015ApJ...813L...8M}. Thus the impact of this reionization event on the thermal state of the IGM provides an independent way to constrain these models. We showed that, under standard assumptions about the spectra of AGN, the earlier He{\sc~ii}\ reionization heats the IGM well above the most recent temperature measurements, even for models in which AGN contribute modestly ($\approx 25\%$) to the H{\sc~i}\ ionizing background. We are thus led to conclude that AGN-dominated models are disfavored by the temperature measurements. Our results strongly disfavor the scenario proposed by \citet{2015ApJ...813L...8M} in which AGN emissions alone are enough to reionize intergalactic H{\sc~i}\ by $z\approx5.5$.
Finally, some authors have argued that the slow evolution in the mean opacity of the $z\approx 3.1-3.3$ He{\sc~ii}\ Ly$\alpha$ forest implies that He{\sc~ii}\ reionization ended earlier or was more extended than is expected in standard quasar source models \citep{2014arXiv1405.7405W,2015ApJ...813L...8M}. We argued that the He{\sc~ii}\ reionization simulations that were used to establish this discrepancy should over-predict the opacity at high redshifts. We also argued that the opacity found in these simulations should be most dependent on the assumed quasar lightcurve model. Thus the measurements of \citet{2014arXiv1405.7405W} are not necessarily in conflict with standard quasar models (and a late He{\sc~ii}\ reionization). Therefore, they should not yet be interpreted as evidence in favor of the abundant high-$z$ AGN models considered here.
Combined with direct searches to better characterize the high-$z$ faint AGN population, future measurements of the Ly$\alpha$ forest opacity and of the IGM temperature will more tightly constrain the contribution of AGN to the ionizing background, as well as their contribution to reionization.
\section*{Acknowledgements}
The authors acknowledge support from NSF grants AST1312724 and AST-1312724, from NASA through the Space Telescope
Science Institute grant HST-AR-13903.00, and from the NSF XSEDE allocations TG-AST140087 and TG-AST150004. H.T. also acknowledges support from NASA grant ATP-NNX14AB57G. P.R.S. acknowledges the support of U.S. NSF grant AST-1009799, NASA grant NNX11AE09G, NASA/JPL grant RSA Nos. 1492788 and 1515294. A.D. thanks Fred Davies, Steven Furlanetto, Adam Lidz, Martin Haehnelt, and George Becker for useful discussions/correspondence. The authors also thank the anonymous referee for comments and suggestions that improved this paper.
\nocite{2012ApJ...755..169M,2011ApJ...728L..26G,2007A&A...472..443B,2013A&A...551A..29P,2009A&A...507..781S,2006AJ....131.2766R,2011ApJ...728L..25I,2012ApJ...755..169M,2009AJ....138..305J,2015ApJ...798...28K,2012ApJ...756..160I, 2015ApJ...813L..35K}
|
\section{Introduction}\label{INT}
Consider the general nonlinear partial differential equation
\begin{equation}\label{int_eq_1}
F\left(x, \{\partial ^\alpha u\}_{\vert \alpha\vert \leq m}\right)=0,
\end{equation}
where $F(x,\zeta)\in C^\infty (\mathbb R^n\times \mathbb C^N)$ for suitable positive integer $N$.\\
In order to investigate local and microlocal regularity of the solutions, it is quite natural to reduce the study to the linearized equation, obtained by differentiation with respect to $x_j$
\begin{equation}\label{int_eq_2}
\sum_{\vert \alpha\vert\leq m}\frac{\partial F}{\partial \zeta^\alpha} \left(x, \{\partial^\beta u\}_{\vert \beta\vert\leq m}\right) \partial ^\alpha\partial_{x_j}u=-\frac{\partial F}{\partial x_j}\left(x, \{\partial^\beta u\}_{\vert \beta\vert\leq m}\right).
\end{equation}
Notice that the regularity of the coefficients $a_\alpha(x)=\dfrac{\partial F}{\partial \zeta^\alpha} \left(x, \{\partial^\beta u\}_{\vert \beta\vert\leq m}\right)$, depends on the solution $u$ and the function $F(x,\zeta)$. We need then to study as first step the algebra properties in the function spaces in which we are intended to operate, as well as the behaviour of the pseudodifferential operators with symbols in such spaces.
\\
When working in H\"older spaces and Sobolev spaces $H^{s,2}$, we can refer to the paradifferential calculus, developed by J.M. Bony and Y. Meyer, \cite{BO1}, \cite{ME1}, \cite{TA1}. Generalizations of these arguments to the symbols of quasi homogeneous, or completely inhomogeneous type can be found in \cite{YA1}, \cite{YA2}, \cite{G}, \cite{G0}, \cite{G3}.\\
In this paper we fix the attention on pseudodifferential operators with symbols in weighted Fourier Lebesgue spaces $\mathcal F L^p_\omega(\mathbb R^n)$, following an idea of S. Pilipovi\'c - N. Teofanov - J. Toft, \cite{PTT2}, \cite{PTT1}.\\ Passing now to consider the microlocal regularity properties, let us notice that the H\"ormander wave front set, introduced in \cite{HOR1} for smooth singularities and extended to the Sobolev spaces $H^{s,2}$ in \cite{HOR_hyp}, uses as basic tool the conic neighborhoods in $\mathbb R^n\setminus\{0\}$.
Thus the homogeneity properties of the symbol $p(x,\xi)$ and the characteristic set Char $P$ of the (pseudo) differential operator $P=p(x,D)$, play a key role. In order to better adapt the study to a wider class of equations, starting from the fundamental papers of R. Beals \cite{BE1}, L. H\"ormander \cite{HO3}, an extensive literature about weighted pseudodifferential operators has been developed, see e.g. \cite{YA1}, \cite{CM1}, \cite{LR1}, \cite{BBR}.\\
We are particularly interested here in the generalizations of the wave front set not involving the use of conic neighborhoods and consequently the homogeneity properties. In some cases, for example in the study of propagation of singularities of the Schroedinger operator, $i\partial_t-\Delta$, we can use the quasi-homogeneous wave front set, introduced in \cite{LA1}, see further \cite{SA1}, \cite{YA2}. More generally, failing of any homogeneity properties, the propagation of the microlocal singularities are described in terms of filter of neighborhoods, introduced in \cite{RO1} and further developed in \cite{GA}, \cite{G}, \cite{G0}, \cite{G1}, \cite{G3}, \cite{GM2.1}, \cite{GM2.2} \cite{GM3}.\\
In some previous works of the authors continuity and microlocal properties are considered in Sobolev spaces in $L^p$ setting, see \cite{GM1}, \cite{GM2}, \cite{GMqh2}, \cite{GM2.1}, \cite{GM2.2}, \cite{GM3}.\\
In the present paper we prove a result of propagation of singularities of Fourier Lebesgue type, for partial (pseudo) differential equations, whose symbol satisfies generalized elliptic properties. Namely we obtain an extension of the well known propagation of singularities given by H\"ormander \cite{HOR_hyp} for the Sobolev wave front set $WF_{H^{s,2}}$ and operators of order $m$:
\begin{equation*}
WF_{H^{s-m,2}}(Pf)\subset WF_{H^{s,2}}(f)\subset WF_{H^{s-m,2}}(Pf)\cup {\rm Char}\, (P),
\end{equation*}
given in terms of filter of microlocal singularities and quasi-homogeneous wave front set.\\
Applications to semilinear partial differential equations are given at the end.
The plan of the paper is the following: in \S \ref{prel_sct} the weight funtions $\omega$ and the Fourier Lebesgue spaces $\mathcal F L^p_\omega(\mathbb R^n)$ are introduced and their properties studied. In \S \ref{algebra_sct}, \S \ref{w_FL_pdo_sct}, under suitable additional conditions on the weight function, algebra properties in $\mathcal F L^p_\omega(\mathbb R^n)$ and continuity of pseudodifferential operators with symbols in Fourier Lebesgue spaces are studied. The microlocal regularity, in terms of inhomogeneous neighborhoods, is introduced and studied. In \S \ref{micro_w_FL_sct} the microlcal properties of Fourier Lebesgue spaces are defined, while the propagation of microlocal singularities is given in \S \ref {appl_sct}, namely in Proposition \ref{filter_inclusions}. In \S \ref {appl_sct} applications to semilinear equations are studied, with specific examples in the field of quasi-homogeneous partial differential equations.
\section{Preliminaries}\label{prel_sct}
\subsection{Weight functions}\label{wf_sct}
Throughout the paper, we call {\it weight function} any positive measurable map $\omega:\mathbb R^n\rightarrow ]0,+\infty[$ satisfying the following {\it temperance} condition
\begin{equation*}
(\mathcal T)\qquad\qquad\qquad\omega(\xi)\le C(1+\vert\xi-\eta\vert)^N\omega(\eta)\,,\qquad\forall\,\xi\,,\eta\in\mathbb R^n\,,\qquad\qquad\qquad\,\,\,
\end{equation*}
for suitable positive constants $C$ and $N$.
In the current literature, a positive function $\omega$ obeying condition ($\mathcal T$) is said to be either {\it temperated} (see \cite{G3}, \cite{HOR1}) or, in the field of Modulation Spaces, {\it polynomially moderated} (cf. \cite{F1}, \cite{GRO1}, \cite{PTT1}, \cite{PTT2}).
For $\omega, \omega_1$ weight functions; we write $\omega\preceq\omega_1$ to mean that, for some $C>0$
\begin{equation*}
\omega(x)\le C\omega_1(x)\,,\qquad\forall\,x\in\mathbb R^n\,;
\end{equation*}
moreover we say that $\omega, \omega_1$ are {\it equivalent}, writing $\omega \asymp\omega_1$ in this case, if
\begin{equation}\label{equiv}
\omega\preceq\omega_1\qquad\mbox{and}\qquad\omega_1\preceq\omega\,.
\end{equation}
Applying ($\mathcal T$) it yields at once that
$\omega(\xi)\le C(1+\vert\xi\vert)^N\omega(0)\,$ and
$\omega(0)\le C(1+\vert-\xi\vert)^N\omega(\xi)=C(1+\vert\xi\vert)^N\omega(\xi)\,$, for any $\xi\in\mathbb R^n$.
Thus, for every weight function $\omega$ there exist constants $C\ge 1$ and $N>0$ such that
\begin{equation}\label{pg_cond}
\frac1{C}(1+\vert\xi\vert)^{-N}\le\omega(\xi)\le C(1+\vert\xi\vert)^N\,,\qquad\forall\,\xi\in\mathbb R^n\,.
\end{equation}
\begin{proposition}\label{pm_prop1}
Let $\omega, \omega_1$ be two weight functions and $s\in\mathbb R$. Then $\omega\omega_1$, $1/\omega$ and $\omega^s$ are again weight functions.
\end{proposition}
\begin{proof}
Assume that for suitable constants $C, C_1, N, N_1$
\begin{equation}\label{PMomega1}
\begin{split}
&\omega(\xi)\le C(1+\vert\xi-\eta\vert)^N\omega(\eta)\,,\\
&\omega_1(\xi)\le C_1(1+\vert\xi-\eta\vert)^{N_1}\omega(\eta)\qquad\forall\,\xi\,,\eta\in\mathbb R^n\,.
\end{split}
\end{equation}
Then we deduce that
\begin{equation*}
\begin{split}
&\omega(\xi)\omega_1(\xi)\le CC_1(1+\vert\xi-\eta\vert)^{N+N_1}\omega(\eta)\omega_1(\eta)\,,\\
&1/\omega(\eta)\le C(1+\vert\eta-\xi\vert)^N 1/\omega(\xi)\,,\qquad\forall\,\xi\,,\eta\in\mathbb R^n\,,
\end{split}
\end{equation*}
which show that $\omega\omega_1$ and $1/\omega$ are temperate.
If $s\ge 0$ then condition ($\mathcal{T}$) for $\omega^s$ follows at once from \eqref{PMomega1}. If $s<0$ it suffices to observe that $\omega^s=1/\omega^{-s}$ and then combining the preceding results.
\end{proof}
We introduce now some further onditions on the weight function $\omega$ which will be repeatedly used in the following.
\begin{itemize}
\item[($\mathcal{SV}$)] {\bf Slowly varying condition}: there exist positive constants $C\ge 1$, $N$ such that
\begin{equation}\label{sv_cond}
\frac1{C}\le\frac{\omega(\eta)}{\omega(\xi)}\le C\,,\qquad\mbox{when}\,\,\vert\eta-\xi\vert\le\frac1{C}\omega(\xi)^{1/N}\,;
\end{equation}
\item[($\mathcal{SA}$)] {\bf Sub additive condition}: for some positive constant $C$
\begin{equation}\label{ring_cond}
\omega(\xi)\le C\left\{\omega(\xi-\eta)+\omega(\eta)\right\}\,,\qquad\forall\,\xi\,,\eta\in\mathbb R^n\,;
\end{equation}
\item[($\mathcal{SM}$)] {\bf Sub multiplicative condition}: for some positive constant $C$
\begin{equation}\label{subm_cond}
\omega(\xi)\le C\omega(\xi-\eta)\omega(\eta)\,,\qquad\forall\,\xi\,,\eta\in\mathbb R^n\,;
\end{equation}
\item[($\mathcal{G}$)] {\bf $\delta$ condition}: for some positive constants $C$ and $0<\delta<1$
\begin{equation}\label{delta_cond}
\omega(\xi)\le C\left\{\omega(\eta)\omega(\xi-\eta)^\delta+\omega(\eta)^\delta\omega(\xi-\eta)\right\}\,,\qquad\forall\,\xi\,,\eta\in\mathbb R^n\,;
\end{equation}
\item[($\mathcal{B}$)] {\bf Beurling's condition}: for some positive constant $C$
\begin{equation}\label{B_cond}
\sup\limits_{\xi\in\mathbb R^n}\int_{\mathbb R^n}\frac{\omega(\xi)}{\omega(\xi-\eta)\omega(\eta)}\,d\eta\le C\,.
\end{equation}
\end{itemize}
For a thorough account on the relations between the properties introduced above, we refer to \cite{G3}. For reader's convenience, here we quote and prove only the following result.
\begin{proposition}\label{wf_relationship}
For the previous conditions the following relationships are true.
\begin{itemize}
\item[i.] Assume that $\omega$ is uniformly bounded from below in $\mathbb R^n$, that is
\begin{equation}\label{bb_cond}
\inf_{\xi\in\mathbb R^n}\omega(\xi)=c >0\,.
\end{equation}
Then
$(\mathcal{SV})\,\,\Rightarrow\,\,(\mathcal{T})$ and $(\mathcal{G})\,\,\Rightarrow\,\,(\mathcal{SM})$.
\item[ii.] Assume that
\begin{equation}\label{integrability}
\frac{1}{\omega}\in L^1(\mathbb R^n)\,.
\end{equation}
Then $(\mathcal{SA})\,\,\Rightarrow\,\,(\mathcal{B})$ and $(\mathcal{G})\,\,\Rightarrow\,\,\omega^{\frac1{1-\delta}}\,\,\mbox{satisfies}\,\,(\mathcal{B})$.
\end{itemize}
\end{proposition}
\begin{proof}
{\it Statement i:} Let the constants $C$, $N$ be fixed as in \eqref{sv_cond}. For $\xi, \eta\in\mathbb R^n$ such that $\vert\xi-\eta\vert\le\frac1{C}\omega(\xi)^{1/N}$, it follows directly from \eqref{sv_cond} that $\omega(\xi)\le C\omega(\eta)\le C\omega(\eta)(1+\vert\xi-\eta\vert)^N
$. On the other hand, when $\vert\xi-\eta\vert>\frac1{C}\omega(\xi)^{1/N}$ from \eqref{bb_cond} we deduce at once
$\omega(\xi)\le C^N\vert\xi-\eta\vert^N\le\frac{C^N}{c}\omega(\eta)(1+\vert\xi-\eta\vert)^N\,.
$
This shows the validity of the first implication. As for the second one, it is sufficient to observe that $\omega(\xi)\ge\varepsilon>0$ and $0<\delta<1$ yield at once
\begin{equation}\label{stima_delta}
\omega(\xi)^\delta\le c^{\delta-1} \omega(\xi)\,, \qquad\forall\,\xi\in\mathbb R^n\,.
\end{equation}
Then the result follows from estimating by \eqref{stima_delta} the function $\omega^\delta$ in the right-hand side of \eqref{delta_cond}.
{\it Statement ii:} For every $\xi\in\mathbb R^n$, using \eqref{ring_cond} we get
\begin{equation*}
\frac{\omega(\xi)}{\omega(\xi-\eta)\omega(\eta)}\le C\left\{\frac{1}{\omega(\eta)}+\frac{1}{\omega(\xi-\eta)}\right\}\,;
\end{equation*}
hence the first implication follows observing that, by a suitable change of variables, the right-hand side is an integrable function on $\mathbb R^n$, whose integral is independent of $\xi$.
Concerning the second implication, for every $\xi\in\mathbb R^n$, from \eqref{delta_cond} we get
\begin{equation*}
\begin{split}
\left(\frac{\omega(\xi)}{\omega(\xi-\eta)\omega(\eta)}\right)^{\frac1{1-\delta}}&\le C_\delta\left(\omega(\xi-\eta)^{\delta-1}+\omega(\eta)^{\delta-1}\right)^{\frac1{1-\delta}}\\
&\le C_\delta\left\{\left(\omega(\xi-\eta)^{\delta-1}\right)^{\frac1{1-\delta}}+\left(\omega(\eta)^{\delta-1}\right)^{\frac1{1-\delta}}\right\}\\
&=C_\delta\left\{\frac1{\omega(\xi-\eta)}+\frac1{\omega(\eta)}\right\}\,,
\end{split}
\end{equation*}
for a suitable constant $C_\delta>0$ depending on $\delta$. Now we conclude as in the proof of the first implication.
\end{proof}
{\it Examples}
\begin{itemize}
\item[1.] The {\it standard homogeneous weight}
\begin{equation}\label{ell_wf}
\langle\xi\rangle^m:=\left(1+\vert\xi\vert^2\right)^{m/2}\,,\qquad\xi\in\mathbb R^n\,,\,\,\,m\in\mathbb R\,,
\end{equation}
is a weight function according to the definition given at the beginning of this section.
The well-known Peetre inequality
\begin{equation}\label{peetre}
\langle\xi\rangle^m\le 2^{\vert m\vert}\langle\xi-\eta\rangle^{\vert m\vert}\langle\eta\rangle^m\,,\qquad\forall\,\xi\,,\eta\in\mathbb R^n\,,
\end{equation}
shows that $\langle\cdot\rangle^m$ satisfies the condition ($\mathcal{T}$) for every $m\in\mathbb R$ (with $N=\vert m\vert$) as well as the condition ($\mathcal{SM}$) for $m\ge 0$. For every $m\ge 0$, the function $\langle\cdot\rangle^m$ also fulfils ($\mathcal{SV}$) (where $N=m$) as a consequence of a Taylor expansion, and ($\mathcal{SA}$). Finally $1/\langle\cdot\rangle^m$ satisfies the integrability condition \eqref{integrability} as long as $m>n$; hence $\langle\cdot\rangle^m$ satisfies condition ($\mathcal{B}$) for $m>n$, in view of the statement ii of Proposition \ref{wf_relationship}.
\item[2.] For $M=(m_1,\dots,m_n)\in\mathbb N^n$, the {\it quasi-homogeneous weight} is defined as
\begin{equation}\label{quasi_ell_wf}
\langle\xi\rangle_M:=\left(1+\sum\limits_{j=1}^n\xi_j^{2m_j}\right)^{1/2}\,,\qquad\forall\,\xi\in\mathbb R^n\,.
\end{equation}
The quasi-homogeneous weight obeys the polynomial growth condition
\begin{equation}\label{pg_q_hom}
\frac1{C}\langle\xi\rangle^{m_\ast}\le \langle\xi\rangle_M\le C\langle\xi\rangle^{m^\ast}\,,\qquad\forall\,\xi\in\mathbb R^n\,,
\end{equation}
for some positive constant $C$ and $m_\ast:=\min\limits_{1\le j\le n}m_j$, $m^\ast:=\max\limits_{1\le j\le n}m_j$. Moreover, for all $s\in\mathbb R$, the derivatives of $\langle\cdot\rangle_M^s$ decay according to the estimates below
\begin{equation}\label{q_hom_der}
\left\vert\partial^\alpha_\xi\langle\xi\rangle_M^s\right\vert\le C_\alpha\langle\xi\rangle_M^{s-\langle\alpha\,,\frac1{M}\rangle}\,,\qquad\forall\,\xi\in\mathbb R^n\,,\, \forall \alpha\in\mathbb Z^n_+,
\end{equation}
where $\langle\alpha\,, \,\frac1{M}\rangle:=\sum\limits_{j=1}^n\frac{\alpha_j}{m_j}$ and $C_\alpha>0$ is a suitable constant. Using \eqref{q_hom_der} with $s=\frac1{m^\ast}$ we may prove that $\langle\cdot\rangle_M$ fulfils condition ($\mathcal{SV}$) with $N=m^\ast$; indeed from the trivial identities
\begin{equation}\label{taylor1}
\langle\xi\rangle_M^{1/m^\ast}-\langle\eta\rangle_M^{1/m^\ast}\!=\sum\limits_{j=1}^n(\xi_j-\eta_j)\int_0^1\partial_j\left(\langle\cdot\rangle_M^{1/m^\ast}\right)(\eta+t(\xi-\eta))\,dt
\end{equation}
and \eqref{q_hom_der}, we deduce
\begin{equation*}\label{sv_q_hom1}
\begin{split}
\left\vert\langle\xi\rangle_M^{1/m^\ast}\right.&\left.-\langle\eta\rangle_M^{1/m^\ast}\right\vert\le\sum\limits_{j=1}^n C_j\vert\xi_j-\eta_j\vert\int_0^1\langle \eta+t(\xi-\eta)\rangle_M^{1/m^\ast-1/m_j}\,dt\\
&\le\sum\limits_{j=1}^n C_j\vert\xi_j-\eta_j\vert\,,
\end{split}
\end{equation*}
since $m^\ast\ge m_j$ for every $j$. Now for $\xi$, $\eta$ satisfying
$
\vert\xi-\eta\vert\le\varepsilon \langle\xi\rangle_M^{1/m^\ast}
$,
from the previous inequality we deduce
$
\left\vert\langle\xi\rangle_M^{1/m^\ast}-\langle\eta\rangle_M^{1/m^\ast}\right\vert\!\le\!\widehat C \varepsilon \langle\xi\rangle_M^{1/m^\ast}
$,
with $\widehat C:=\sum\limits_{j=1}^n C_j$, that is
$
(1-\widehat C \varepsilon)\langle\xi\rangle_M^{1/m^\ast}\le\langle\eta\rangle_M^{1/m^\ast}\le(1+\widehat C \varepsilon)\langle\xi\rangle_M^{1/m^\ast}
$,
from which we get the conclusion, if we assume for instance $0<\varepsilon\le\frac1{2\widehat C}$.
From ($\mathcal{SV}$) and the trivial inequality $\langle\xi\rangle_M\ge 1$, using the statement i of Proposition \ref{wf_relationship} we obtain that ($\mathcal{T}$) is also satisfied with $N=m^\ast$.
Also, the weight $\langle\cdot\rangle_M$ satisfies condition ($\mathcal{SA}$) and, because of the left inequality in \eqref{pg_q_hom}, $1/\langle\cdot\rangle_M^s$ satisfies condition \eqref{integrability} provided that $s>\frac{n}{m_\ast}$. Then from the statement ii of Proposition \ref{wf_relationship}, $\langle\cdot\rangle_M^s$ satisfies condition ($\mathcal{B}$) for $s>\frac{n}{m_\ast}$.
At the end, let us observe that for $M=(m,\dots,m)$, with a given $m\in\mathbb N$, $\langle\xi\rangle_M \asymp\langle\xi\rangle^m$.
\item[3.] Let $\mathcal P$ be a {\it complete polyhedron} of $\mathbb{R}^n$ in the sense of Volevich-Gindikin, \cite{VG2}. The {\em multi-quasi-elliptic weight functions} is defined by
\begin{equation}\label{mqe_wf}
\lambda_{\cal P}(\xi):=\left(\sum_{\alpha\in V(\mathcal P)}\xi^{2\alpha}\right)^{1/2},\qquad\xi\in\mathbb R^n\,,
\end{equation}
where $V(\mathcal P)$ denotes the set of {\it vertices} of $\mathcal P$.
We recall that a {\em convex polyhedron} $\mathcal{P}\subset\mathbb{R}^n$ is the convex hull of a finite set $V(\mathcal{P})\subset \mathbb{R}^n$ of convex-linearly independent points, called {\em vertices} of $\mathcal{P}$, and univocally determined by $\mathcal {P}$ itself. Moreover, if $\mathcal{P}$ has non empty interior, it is completely described by
\begin{displaymath}\label{WF6}
\mathcal{P}=\{\xi\in\mathbb{R}^n;\nu\cdot\xi\geq 0, \forall \nu\in \mathcal{N}_0(\mathcal{P})\}\cap\{\xi\in\mathbb{R}^n;\nu\cdot\xi\leq 1,\forall\nu\in\mathcal{N}_1(\mathcal{P})\};
\end{displaymath}
where $\mathcal{N}_0(\mathcal{P})\subset \{\nu\in \mathbb{R}^n;\vert\nu\vert =1\}$, $\mathcal{N}_1(\mathcal{P})\subset \mathbb{R}^n$ are finite sets univocally determined by $\mathcal{P}$ and, as usual, $\nu\cdot\xi=\sum_{j=1}^n\nu_j\xi_j$. The boundary of $\mathcal{P}$, $\mathcal {F}(\mathcal{P})$, is made of faces $\mathcal{F}_{\nu}(\mathcal{P})$ which are the convex hulls of the vertices of $\mathcal{P}$ lying on the hyper-planes $H_\nu$ orthogonal to $\nu\in \mathcal{N}_0(\mathcal{P})\cup\mathcal{N}_1(\mathcal{P})$, of equation :
\begin{equation*}
\nu\cdot\xi=0\quad \mbox{if}\quad \nu\in\mathcal{N}_0(\mathcal{P}),\quad\quad \nu\cdot\xi=1\quad\mbox{if}\quad \nu\in\mathcal{N}_1(\mathcal{P}).
\end{equation*}
A {\it complete polyhedron} is a convex polyhedron $\mathcal{P}\subset\mathbb{R}^n_+=:\{\xi\in\mathbb{R}^n:\ \xi_j\geq 0, \ j=1,...,n\}$ such that:
\begin{itemize}
\item[i)] $V(\mathcal{P})\subset\mathbb{N}^n$;
\item[ii)] $(0,\dots,0)\in V(\mathcal{P})$, and $V(\mathcal{P})\neq\{(0,\dots,0)\}$;
\item[iii)] $\mathcal{N}_0(\mathcal{P})=\{e_1,\dots,e_n\}$ with $e_j=(0,\dots, 1_{j-entry},\dots 0)\in\mathbb{R}^n_+$;
\item[iv)] every $\nu\in \mathcal{N}_1(\mathcal{P})$ has strictly positive components $\nu_j$, $j=1,...,n$.
\end{itemize}
On can prove that the multi-quasi-elliptic weight growths at infinity according to the following estimates
\begin{equation}\label{pg_mqe}
\frac1{C}\langle\xi\rangle^{\mu_0}\le \lambda_{\mathcal P}(\xi)\le C\langle\xi\rangle^{\mu_1}\,,\qquad\forall\,\xi\in\mathbb R^n\,,
\end{equation}
for a suitable positive constant $C$ and where
\begin{equation}\label{min_max_order}
\mu_0:=\min\limits_{\gamma\in V(\mathcal P)\setminus\{0\}}\vert\gamma\vert\quad\mbox{and}\quad\mu_1:=\max\limits_{\gamma\in V(\mathcal P)}\vert\gamma\vert
\end{equation}
are called {\it minimum} and {\it maximum order} of $\mathcal P$ respectively. Moreover it can be proved that for all $s\in\mathbb R$ the derivatives of $\lambda_{\mathcal P}^s$ decay according to the estimates below
\begin{equation}\label{mqe_der}
\left\vert\partial^\alpha_\xi\lambda_{\mathcal P}^s(\xi)\right\vert\le C_\alpha\lambda_{\mathcal P}(\xi)^{s-\frac1{\mu}\vert\alpha\vert}\,,\qquad\forall\,\xi\in\mathbb R^n\,,
\end{equation}
where
\begin{equation}\label{formal_order}
\mu:=\max\left\{1/\nu_j\,,\,\,j=1,\dots,n\,,\,\,\nu\in\mathcal N_1(\mathcal P)\right\}\,,
\end{equation}
satisfying $\mu\ge\mu_1$, is the so-called {\it formal order} of $\mathcal P$, see e.g. \cite{BBR} (see also \cite{GM1} where more general decaying estimates for $\lambda_{\mathcal P}^s$ are established).
Representing $\lambda_{\mathcal P}(\xi)^{1/\mu}-\lambda_{\mathcal P}(\eta)^{1/\mu}$ as in \eqref{taylor1}, for arbitrary $\xi, \eta\in\mathbb R^n$, and using \eqref{mqe_der} (with $s=1/\mu$), we deduce that $\lambda_{\mathcal P}$ satisfies ($\mathcal{SV}$) with $N=\mu$. Using also that $\lambda_{\mathcal P}(\xi)\ge 1$ (recall that $0\in V(\mathcal P)$), in view of Proposition \ref{wf_relationship} it follows that $\lambda_{\mathcal P}$ also satisfies ($\mathcal{T}$) with $N=\mu$, hence it is a weight function agreeing to the definition given at the beginning of this section.
The weight function $\lambda_{\mathcal P}$ does not satisfy condition ($\mathcal{SA}$); on the other hand it can be shown (see \cite{G0}) that condition ($\mathcal G$) is verified taking
\begin{equation}\label{delta}
\delta=\max\limits_{\beta\in\mathcal P\setminus\mathcal F(\mathcal P)}\max\limits_{\nu\in\mathcal N_1(\mathcal P)}\left\{\nu\cdot\beta\right\}\,.
\end{equation}
Since, from the left inequality in \eqref{pg_mqe} we also derive that $\lambda_{\mathcal P}^{-s}$ satisfies \eqref{integrability} for $s>\frac{n}{\mu_0}$, we conclude from the statement ii of Proposition \ref{wf_relationship} that $\lambda_{\mathcal P}^{r}$ satisfies condition ($\mathcal{B}$) if $r>\frac{n}{(1-\delta)\mu_0}$ for $\delta$ defined above.
In the end, we notice that $\lambda_{\mathcal P}$ verifies ($\mathcal{SM}$) as a consequence of ($\mathcal{G}$), since $\lambda_{\mathcal P}(\xi)\ge 1$ and $0<\delta<1$, cf. Proposition \ref{wf_relationship}, statement i.
\end{itemize}
\begin{remark}\label{rmk:5}
{\rm We notice that the quasi-homogeneous weight $\langle\cdot\rangle_M$, $M=(m_1,\dots,m_n)\in\mathbb N^n$, considered in the example 2 is just the multi-quasi-elliptic weight $\lambda_{\mathcal P}$ introduced in the example 3 corresponding to the complete polyhedron $\mathcal P$ defined by the convex hull of the finite set $V(\mathcal P)=\{0\,,\,\, m_je_j\,,\,\,\,j=1,\dots,n\}$; in particular, the growth estimates \eqref{pg_q_hom} are the particular case of \eqref{pg_mqe} corresponding to the previous polyhedron $\mathcal P$ (in which case $\mu_0=m_\ast$ and $\mu_1=m^\ast$). Notice however that the decaying estimates \eqref{q_hom_der} satisfied by the quasi-homogeneous weight $\langle\cdot\rangle_M$ do not admit a counterpart in the case of the general multi-quasi-elliptic weight $\lambda_{\mathcal P}$. Estimates \eqref{q_hom_der} give a precise decay in each coordinate direction: the decrease of $\langle\xi\rangle_M$ corresponding to one derivative with respect to $\xi_j$ is measured by $\langle\xi\rangle_M^{-1/m_j}$\footnote{In other words the decay of the derivatives is measured here by the {\it vector weight} $(\langle\xi\rangle_M^{1/m_1},\dots,\langle\xi\rangle_M^{1/m_n})$, in the sense of {\it vector weighted} symbol classes, see \cite{G}, \cite{GM3}, \cite{RO1}.}. The lack of homogeneity in the weight associated to a general complete polyhedron $\mathcal P$ in \eqref{mqe_wf}, prevents from extending to derivatives of $\lambda_{\mathcal P}(\xi)$ the decay properties in \eqref{q_hom_der}: estimates \eqref{mqe_der} do not take account of the decay corresponding {\it separately to each coordinate direction}.}
\end{remark}
\subsection{Weighted Lebesgue and Fourier--Lebesgue spaces}\label{FL_spaces_sct}
Let $\omega:\mathbb R^n\rightarrow ]0,+\infty[$ be a weight function.
\begin{definition}\label{w_L_spaces}
For every $p\in[1,+\infty]$, the weighted Lebesgue space $L^p_{\omega}(\mathbb R^n)$ is defined as the set of the (equivalence classes of) measurable functions $f:\mathbb R^n\rightarrow\mathbb C$ such that
\begin{equation}\label{w_L_cond}
\begin{split}
&\int_{\mathbb R^n}\omega(x)^p\vert f(x)\vert^p\,dx<+\infty\,,\quad\mbox{if}\,\,p<+\infty\,,\\
&\\
&\omega f\,\,\mbox{is essentially bounded in}\,\,\mathbb R^n\,,\quad\mbox{if}\,\,p=+\infty\,.
\end{split}
\end{equation}
\end{definition}
For every $p\in[1,+\infty]$, $L^p_{\omega}(\mathbb R^n)$ is Banach space with respect to the natural norm
\begin{equation}\label{norma_w_L}
\Vert f\Vert_{L^p_{\omega}}:=
\begin{cases}
\left(\int_{\mathbb R^n}\omega(x)^p\vert f(x)\vert^p\,dx\right)^{1/p}\,,\quad\mbox{if}\,\,p<+\infty\,,\\
\\
{\rm ess\,sup}_{x\in\mathbb R^n}\omega(x)\vert f(x)\vert\,,\quad\mbox{if}\,\,p=+\infty\,.
\end{cases}
\end{equation}
\begin{remark}\label{rmk:1}
{\rm It is easy to see that for all $p\in[1,+\infty]$
\begin{equation}\label{imbedding1}
L^p_{\omega_2}(\mathbb R^n)\hookrightarrow L^p_{\omega_1}(\mathbb R^n)\,,\qquad\mbox{if}\,\,\omega_1\preceq\omega_2\,.
\end{equation}
If in particular $\omega_1\asymp\omega_2$ then $L^p_{\omega_1}(\mathbb R^n)\equiv L^p_{\omega_2}(\mathbb R^n)$, and the norms defined in \eqref{norma_w_L} corresponding to $\omega_1$ and $\omega_2$ are equivalent. When the weight function $\omega$ is constant the related weighted space $L^p_{\omega}(\mathbb R^n)$ reduces to the standard Lebesgue space of order $p$, denoted as usual by $L^p(\mathbb R^n)$.}
\end{remark}
\begin{remark}\label{rmk:2}
{\rm For an arbitrary $p\in[1,+\infty]$, $f\in L^p_{\omega}(\mathbb R^n)$ and every $\varphi\in\mathcal S(\mathbb R^n)$ we obtain
\begin{equation}\label{holder}
\int_{\mathbb R^n}f(x)\varphi(x)\,dx\le\left\Vert\frac{\varphi}{\omega}\right\Vert_{L^q}\Vert\omega f\Vert_{L^p}\,,\qquad\frac1{p}+\frac1{q}=1\,.
\end{equation}
From \eqref{holder} and the estimates \eqref{pg_cond}, we deduce at once that
\begin{equation*}
\mathcal S(\mathbb R^n)\hookrightarrow L^p_{\omega}(\mathbb R^n)\hookrightarrow\mathcal S^\prime(\mathbb R^n)\,.
\end{equation*}
Moreover, $C^\infty_0(\mathbb R^n)$ is a dense subspace of $L^p_{\omega}(\mathbb R^n)$ when $p<+\infty$, see \cite{G2}.}
\end{remark}
\begin{remark}\label{rmk:8}
{\rm For $\Omega$ open subset of $\mathbb R^n$, $L^p_{\omega}(\Omega)$, for any $p\in[1,+\infty]$, is the set of (equivalence classes of) measurable functions on $\Omega$ such that
\begin{equation}\label{Lpomega}
\Vert f\Vert^p_{L^p_{\omega}(\Omega)}:=\int_{\Omega}\omega(x)^p\vert f(x)\vert^p\,dx<+\infty
\end{equation}
(obvious modification for $p=+\infty$). $L^p_{\omega}(\Omega)$ is endowed with a structure of Banach space with respect to the natural norm defined by \eqref{Lpomega}.}
\end{remark}
\begin{definition}\label{w_FL_spaces}
For every $p\in[1,+\infty]$ and $\omega(\xi)$ temperate weight function, the weighted Fourier Lebesgue space $\mathcal F L^p_{\omega}(\mathbb R^n)$ is the vector space of all distributions $f\in\mathcal S^\prime(\mathbb R^n)$ such that
\begin{equation}\label{w_FL_cond}
\widehat f\in L^p_{\omega}(\mathbb R^n)\,,
\end{equation}
equipped with the natural norm
\begin{equation}\label{w_FL_norm}
\Vert f\Vert_{\mathcal FL^p_{\omega}}:=\Vert\widehat f\Vert_{L^p_{\omega}}\,.
\end{equation}
\end{definition}
Here $\widehat f$ is the Fourier transform $\widehat f(\xi)=\int e^{-i\xi\cdot x}f(x)\,dx$, defined in $\mathcal S(\mathbb R^n)$ and extended to $\mathcal S'(\mathbb R^n)$.
The spaces $\mathcal FL^p_{\omega}(\mathbb R^n)$ were introduced in H\"ormander \cite{HOR1}, with the notation $\mathcal B_{p,k}$, $k(\xi)$ weight function, for the study of the regularity of solutions to hypoelliptic partial differential equations with constant coefficients, see also \cite{G2}, \cite{G1}.
From the mapping properties of the Fourier transform on $\mathcal S(\mathbb R^n)$ and $\mathcal S^\prime(\mathbb R^n)$ and the above stated properties of weighted Lebesgue spaces we can conclude, see again \cite{G2}, that for all $p\in[1,+\infty]$ and $\omega$ temperate weight function
\begin{itemize}
\item[(a)] $\mathcal F L^p_{\omega}(\mathbb R^n)$ is a Banach space with respect to the norm \eqref{w_FL_norm};
\item[(b)] $\mathcal S(\mathbb R^n)\hookrightarrow \mathcal F L^p_{\omega}(\mathbb R^n)\hookrightarrow\mathcal S^\prime(\mathbb R^n)$;
\item[(c)] $C^\infty_0(\mathbb R^n)$ is a dense subspace of $\mathcal F L^p_{\omega}(\mathbb R^n)$ when $p<+\infty$;
\item[(d)] $\mathcal F L^p_{\omega_2}(\mathbb R^n)\hookrightarrow\mathcal F L^p_{\omega_1}(\mathbb R^n)$ if $\omega_1\preceq\omega_2$; in particular, we obtain that $\mathcal F L^p_{\omega_2}(\mathbb R^n)\equiv\mathcal F L^p_{\omega_1}(\mathbb R^n)$ as long as $\omega_1\asymp\omega_2$ and the norms corresponding to $\omega_1$ and $\omega_2$ by \eqref{w_FL_norm} are equivalent in this case.
\end{itemize}
When $\omega$ is a positive constant the weighted space $\mathcal F L^p_{\omega}(\mathbb R^n)$ is simply denoted by $\mathcal F L^p(\mathbb R^n)$. Moreover we will adopt the shortcut notations $L^p_s(\mathbb R^n):= L^p_{\langle\cdot\rangle^s}(\mathbb R^n)$, $\mathcal F L^p_s(\mathbb R^n):=\mathcal F L^p_{\langle\cdot\rangle^s}(\mathbb R^n)$ for the corresponding Lebesgue and Four\-ier Lebesgue spaces.
Analogously, when $\omega(\xi)=\langle\xi\rangle_M^s$ or $\omega(\xi)=\lambda_{\mathcal P}(\xi)^s$, for $s\in\mathbb R$, the corresponding Lebesgue and Fourier Lebesgue spaces will be denoted $L^p_{s,M}(\mathbb R^n)$, $\mathcal F L^p_{s,M}(\mathbb R^n)$ and $L^p_{s,\mathcal P}(\mathbb R^n)$, $\mathcal F L^p_{s, \mathcal P}(\mathbb R^n)$ respectively.
\vspace{.5cm}
A local counterpart of Fourier Lebesgue spaces can be introduced in the following natural way (see \cite{PTT2}).
\begin{definition}\label{loc_w_FL_spaces}
For $\omega$ weight function, $\Omega$ open subset of $\mathbb R^n$ and any $p\in[1,+\infty]$, $\mathcal FL^p_{\omega,\,{\rm loc}}(\Omega)$ is the class of all distributions $f\in\mathcal D^\prime(\Omega)$ such that $\varphi f\in\mathcal FL^p_{\omega}(\mathbb R^n)$ for every $\varphi\in C^\infty_0(\Omega)$.
For $x_0\in\Omega$, $f\in\mathcal FL^p_{\omega,\,{\rm loc}}(x_0)$ if there exists $\phi\in C^\infty_0(\Omega)$, with $\phi(x_0)\neq 0$, such that $\phi f\in\mathcal FL^p_\omega(\mathbb R^n)$.
\end{definition}
The family of semi-norms
\begin{equation}\label{loc_w_FL_norm}
f\mapsto\Vert \varphi f\Vert_{\mathcal F L^p_{\omega}}\,,\qquad\varphi\in C^\infty_0(\Omega)\,,
\end{equation}
provides $\mathcal FL^p_{\omega\,,{\rm loc}}(\Omega)$ with a natural Fr\'echet space topology. Moreover the following inclusions hold true with continuous embedding
\begin{equation}\label{inclusioni1}
C^\infty(\Omega)\hookrightarrow\mathcal F L^p_{\omega\,,{\rm loc}}(\Omega)\hookrightarrow\mathcal D^\prime(\Omega)
\end{equation}
and for $\Omega_1\subset\Omega_2$ open sets
\begin{equation}\label{inclusioni2}
\mathcal F L^p_{\omega}(\mathbb R^n)\hookrightarrow\mathcal F L^p_{\omega\,,{\rm loc}}(\Omega_2)\hookrightarrow \mathcal F L^p_{\omega\,,{\rm loc}}(\Omega_1)\,.
\end{equation}
\begin{remark}\label{rmk:3}
{\rm It is worth noticing that, as it was proved in \cite{RSTT}, {\it locally} the weighted Fourier Lebesgue spaces $\mathcal F L^q_{\omega}(\mathbb R^n)$ are the same as the weighted {\it modulation spaces} $M^{p,q}_{\omega}(\mathbb R^n)$ and the {\it Wiener amalgam spaces} $W^{p,q}_{\omega}(\mathbb R^n)$, in the sense that
\begin{equation*}
\mathcal F L^q_{\omega}(\mathbb R^n)\cap\mathcal E^\prime(\mathbb R^n)=M^{p,q}_{\omega}(\mathbb R^n)\cap\mathcal E^\prime(\mathbb R^n)=W^{p,q}_{\omega}(\mathbb R^n)\cap\mathcal E^\prime(\mathbb R^n)\,.
\end{equation*}
We refer to Feichtinger \cite{F1} and Gr\"ochenig \cite{GRO1} for the definition and basic properties of modulation and amalgam spaces.}
\end{remark}
\vspace{.5cm}
Agreeing with the previous notations, when the weight function $\omega$ reduces to those considered in the examples 1, 2, 3 above, the corresponding local Fourier Lebesgue spaces will be denoted respectively by $\mathcal F L^p_{s,{\rm loc}}(\Omega)$, $\mathcal F L^p_{s,M,{\rm loc}}(\Omega)$, $\mathcal F L^p_{s,\mathcal P,{\rm loc}}(\Omega)$.
\vspace{.5cm}
Notice at the end that, from Plancherel Theorem, when $p=2$ the global and local weighted Fourier Lebesgue spaces $\mathcal F L^2_{\omega}(\mathbb R^n)$, $\mathcal F L^2_{\omega\,,{\rm loc}}(\Omega)$ coincide with weighted spaces of Sobolev type, see Garello \cite{G3} for an extensive study of such spaces.
\section{Algebra conditions in spaces $\mathcal FL^p_{\omega}(\mathbb R^n)$}\label{algebra_sct}
In order to seek conditions on the weight function $\omega$ which allow the Fourier Lebesgue space $\mathcal F L^p_{\omega}(\mathbb R^n)$ to be an algebra with respect to the point-wise product, let us first state a general continuity result in the framework of suitable mixed-norm spaces of Lebesgue type.
Following \cite{F1}, \cite{GRO1} and in particurar\cite{PTT2}, for $p, q\in [1,+\infty]$ we denote by $\mathcal L_1^{p,q}(\mathbb R^{2n})$ the space of all (equivalence classes of) measurable functions $F=F(\zeta,\eta)$ in $\mathbb R^{n}\times\mathbb R^n$ such that the mixed norm
\begin{equation}\label{mixed_norm1}
\Vert F\Vert_{\mathcal L_1^{p,q}}:=\left(\int\left(\int\vert F(\zeta,\eta)\vert^p\,d\zeta\right)^{q/p}d\eta\right)^{1/q}
\end{equation}
is finite (with obvious modifications if $p$ or $q$ equal $+\infty$).
We also define $\mathcal L_2^{p,q}(\mathbb R^{2n})$ to be the space of measurable functions $F=F(\xi,\eta)$ in $\mathbb R^{n}\times\mathbb R^n$ such that the norm
\begin{equation}\label{mixed_norm2}
\Vert F\Vert_{\mathcal L_2^{p,q}}:=\left(\int\left(\int\vert F(\zeta,\eta)\vert^q\,d\eta\right)^{p/q}d\zeta\right)^{1/p}
\end{equation}
is finite.
\begin{lemma}\label{lemma_cont2}
For $p, q\in [1,+\infty]$ satisfying $\frac1{p}+\frac1{q}=1$, let $f=f(\zeta,\eta)\in\mathcal L^{p,\infty}_1(\mathbb R^{2n})$ and $F=F(\zeta,\eta)\in\mathcal L_2^{\infty,q}(\mathbb R^{2n})$. Then the linear map
\begin{equation}\label{TF2}
\begin{split}
T: C_0^\infty(\mathbb R^n)&\rightarrow\mathcal S^\prime(\mathbb R^n)\\
&g\mapsto Tg:=\int F(\xi,\eta)f(\xi-\eta,\eta)g(\eta)\,d\eta\,.
\end{split}
\end{equation}
extends uniquely to a continuous map from $L^p(\mathbb R^n)$ into itself, still denoted by $T$; moreover its operator norm is bounded as follows
\begin{equation}\label{estimate_TF}
\Vert T\Vert_{\mathcal L(L^p)}\le \Vert f\Vert_{\mathcal L_1^{p,\infty}}\Vert F\Vert_{\mathcal L_2^{\infty,q}}\,.
\end{equation}
\end{lemma}
The proof is given in \cite[Lemma 2.1]{G2}, where the statement reads in quite different formulation. The reader can find a restricted version, independently proved, in \cite[Proposition 3.2]{PTT2}.
\begin{proposition}\label{algebra_prop}
Assume that $\omega$, $\omega_1$, $\omega_2$ are weight functions such that
\begin{equation}\label{B_Lq}
C_q:=\sup\limits_{\xi\in\mathbb R^n}\left\Vert\frac{\omega(\xi)}{\omega_1(\xi-\cdot)\omega_2(\cdot)}\right\Vert_{L^q}<+\infty\,,
\end{equation}
for some $q\in[1,+\infty]$, and let $p\in [1,+\infty]$ be the conjugate exponent of $q$. Then
\begin{itemize}
\item[(1)] the point-wise product map $(f_1,f_2)\mapsto f_1 f_2$ from $\mathcal S(\mathbb R^n)\times\mathcal S(\mathbb R^n)$ to $\mathcal S(\mathbb R^n)$ extends uniquely to a continuous bilinear map from $\mathcal F L^p_{\omega_1}(\mathbb R^n)\times \mathcal F L^p_{\omega_2}(\mathbb R^n)$ to $\mathcal F L^p_{\omega}(\mathbb R^n)$. Moreover for all $f_i\in\mathcal F L^p_{\omega_i}(\mathbb R^n)$, $i=1,2$, the following holds:
\begin{equation}\label{prod_est_1}
\Vert f_1 f_2\Vert_{\mathcal F L^p_{\omega}}\le C_q\Vert f_1\Vert_{\mathcal F L^p_{\omega_1}}\Vert f_2\Vert_{\mathcal F L^p_{\omega_2}}\,.
\end{equation}
\item[(2)] for every open set $\Omega\subseteq\mathbb R^n$ the point-wise product map $(f_1,f_2)\mapsto f_1 f_2$ from $C^\infty_0(\Omega)\times C^\infty_0(\Omega)$ to $C^\infty_0(\Omega)$ extends uniquely to a continuous bilinear map from $\mathcal F L^p_{\omega_1,{\rm loc}}(\Omega)\times \mathcal F L^p_{\omega_2,{\rm loc}}(\Omega)$ to $\mathcal F L^p_{\omega,{\rm loc}}(\Omega)$.
\end{itemize}
\end{proposition}
\begin{proof}
The proof of statement (2) follows at once from that of statement (1).
As for the proof of statement (1), for given $f_1, f_2\in\mathcal S(\mathbb R^n)$ one easily computes:
\begin{equation}\label{representation}
\begin{split}
\omega(\xi)\widehat{f_1 f_2}(\xi)&=(2\pi)^{-n}\int\omega(\xi)\widehat{f_1}(\xi-\eta)\widehat{f_2}(\eta)\,d\eta\\
&=\int F(\xi,\eta)f(\xi-\eta)g(\eta)\,d\eta\,,
\end{split}
\end{equation}
where
\begin{equation*}
F(\zeta,\eta)=\frac{\omega(\zeta)}{\omega_1(\zeta-\eta)\omega_2(\eta)}\,,\,\,\, f(\zeta)=\omega_1(\zeta)\widehat{f_1}(\zeta)\,,\,\,\, g(\zeta)=\omega_2(\zeta)\widehat{f_2}(\zeta)\,.
\end{equation*}
The right-hand side of \eqref{representation} provides a representation of $\omega\widehat{f_1 f_2}$ as an integral operator of the form \eqref{TF2}. Condition \eqref{B_Lq} just means that the function $F(\zeta,\eta)\in\mathcal L_2^{\infty, q}(\mathbb R^{2n})$ (cf. \eqref{mixed_norm2}) and of course the $\eta-$independent function $f=f(\zeta)\in\mathcal S(\mathbb R^n)$ also belongs to $\mathcal L^{p,\infty}_1(\mathbb R^{2n})$. Then applying to \eqref{representation} the result of Lemma \ref{lemma_cont2}, together with the definition of the norm in Fourier Lebesgue spaces, we obtain that the point-wise product $f_1f_2$ satisfies the estimates in \eqref{prod_est_1} and the proof is concluded.
\end{proof}
When the weight functions $\omega$, $\omega_1$ and $\omega_2$ in the statement of Proposition \ref{algebra_prop} coincide, condition \eqref{B_Lq} provides a sufficient condition for $\mathcal F L^p_{\omega}(\mathbb R^n)$ (or its localized counterpart $\mathcal F L^p_{\omega, {\rm loc}}(\Omega)$) is an algebra for the point-wise product. Then we have the following
\begin{corollary}\label{algebra_prop2}
Let $\omega$ be a weight function such that
\begin{equation}\label{B_q}
C_q:=\sup\limits_{\xi\in\mathbb R^n}\left\Vert\frac{\omega(\xi)}{\omega(\xi-\cdot)\omega(\cdot)}\right\Vert_{L^q}<+\infty\,,
\end{equation}
for $q\in[1,+\infty]$, and $p\in[1,+\infty]$ the conjugate exponent of q. Then
\begin{itemize}
\item[(1)] $\left(\mathcal F L^p_{\omega}(\mathbb R^n),\cdot\right)$ is an algebra and for $f_1, f_2\in \mathcal F L^p_{\omega}(\mathbb R^n)$
\begin{equation}\label{prod_est_2}
\Vert f_1 f_2\Vert_{\mathcal F L^p_{\omega}}\le C_q\Vert f_1\Vert_{\mathcal F L^p_{\omega}}\Vert f_2\Vert_{\mathcal F L^p_{\omega}}\,.
\end{equation}
\item[(2)] for every open set $\Omega\subseteq\mathbb R^n$, $\left(\mathcal F L^p_{\omega,{\rm loc}}(\Omega),\cdot\right)$ is an algebra.
\end{itemize}
\end{corollary}
The algebra properties of Corollary \ref{algebra_prop2} let us handle the composition of a Fourier Lebesgue distribution with an entire analytic functions; namely we have the following result, see \cite[Corollary 2.1]{G2}.
\begin{corollary}\label{composition_prop}
Under the same assumptions of Corollary \ref{algebra_prop2} on $\omega$ and $p$, let $F:\mathbb{C}\rightarrow \mathbb{C}$ be an entire analytic function such that $F(0)=0$. Then $F(u)\in\mathcal FL^p_{\omega}(\mathbb R^n)$ for every $u\in \mathcal FL^p_{\omega}(\mathbb R^n)$, and
\begin{equation}\label{LPE61}
\Vert F(u)\Vert_{\mathcal FL^{p}_{\omega}}\leq C\Vert u\Vert_{\mathcal FL^{p}_{\omega}}\,,
\end{equation}
with $C=C(p,F,\Vert u\Vert_{\mathcal FL^{p}_{\omega}})$.
\end{corollary}
\begin{remark}\label{rmk:3.1}
{\rm A counterpart of Corollary \ref{composition_prop} for the local space $\mathcal FL^p_{\omega, \rm loc}(\Omega)$ can be obtained by replacing $F=F(z)$ above with a function $F=F(x,\zeta)$ mapping $\Omega\times {\mathbb C}^M$ into $\mathbb{C}$, which is locally smooth with respect to the real variable $x\in\Omega$ and entire analytic in the complex variable $\zeta\in\mathbb{C}^M$ uniformly on compact subsets of $\Omega$; namely:
$$
F(x,\zeta)=\sum_{\beta\in\mathbb{Z}^M_+}c_{\beta}(x)\zeta^{\beta}, \quad \quad c_{\beta}\in C^{\infty}(\Omega), \ \zeta\in \mathbb{C}^M\,,
$$
where, for any compact set $K\subset\Omega$, $\alpha\in\mathbb{Z}^n_+$, $\beta\in\mathbb{Z}^M_+$, $\sup_{x\in K}\vert \partial_x^{\alpha}c_{\beta}(x)\vert\leq c_{\alpha,K}\lambda_{\beta}$ and $F_1(\zeta):=\sum_{\beta\in\mathbb{Z}_+^M}\lambda_{\beta}\zeta^{\beta}$ is entire analytic.
Under the assumptions of Corollaries \ref{algebra_prop2}, \ref{composition_prop}, we have that $F(x,u)\in \mathcal FL^{p}_{\omega,{\rm loc}}(\Omega)$ as long as the components of the vector $u=(u_1,...,u_M)$ belong to $\mathcal FL^{p}_{\omega, {\rm loc}}(\Omega)$.}
\end{remark}
\begin{remark}\label{rmk:4}
{\rm Let us notice that for $1\le q<+\infty$, condition \eqref{B_q} on $\omega$ is nothing but condition ($\mathcal B$) for the weight function $\omega^q$, while for $q=+\infty$ \eqref{B_q} reduces to condition ($\mathcal{SM}$) for $\omega$. The latter case means that $\left(\mathcal FL^1_{\omega}(\mathbb R^n),\cdot\right)$ is an algebra provided that the weight function $\omega$ is sub-multiplicative, which is in agreement with the more general result of \cite[Lemma 1.6]{PTT2}.}
\end{remark}
The next result shows that the {\it sub-multiplicative} condition $(\mathcal{SM})$ on a weight function is a necessary condition for the corresponding scale of weighted Fourier Lebesgue spaces to possess the algebra property.
\begin{proposition}\label{prop_sm}
Let $\omega_1$, $\omega_2$, $\omega$ be weight functions and $p_1, p_2, p\!\in\![1,\!+\infty]$. If we assume that the map $(f_1,f_2)\mapsto f_1f_2$ from $\mathcal S(\mathbb R^n)\times\mathcal S(\mathbb R^n)$ to $\mathcal S(\mathbb R^n)$ extends uniquely to a continuous bilinear map from $\mathcal F L^{p_1}_{\omega_1}(\mathbb R^n)\times\mathcal F L^{p_2}_{\omega_2}(\mathbb R^n)$ to $\mathcal F L^{p}_{\omega}(\mathbb R^n)$, then a positive constant $C$ exists such that
\begin{equation}\label{stima_algebra1}
\omega(\eta+\theta)\le C\omega_1(\eta)\omega_2(\theta)\,,\qquad\forall\,\eta\,,\theta\in\mathbb R^n\,.
\end{equation}
\end{proposition}
\begin{proof}
By assumption, there exists a constant $C^\prime>0$ such that for all $f\in \mathcal F L^{p_1}_{\omega_1}(\mathbb R^n)$, $g\in\mathcal F L^{p_2}_{\omega_2}(\mathbb R^n)$
\begin{equation}\label{stima_algebra2}
\Vert f g\Vert_{\mathcal F L^p_{\omega}}\le C^\prime\Vert f\Vert_{\mathcal F L^{p_1}_{\omega_1}}\Vert g\Vert_{\mathcal F L^{p_2}_{\omega_2}}\,.
\end{equation}
From condition ($\mathcal{T}$) (cf. also \eqref{pg_cond}) we may find some constants $\varepsilon>0$, $C>0$ such that
\begin{equation}\label{pm1}
\varepsilon\le\frac{\omega_i(\eta)}{\omega_i(\xi)}\le\varepsilon^{-1}\,\,(i=1,2)\,\,\mbox{and}\,\,\varepsilon\le\frac{\omega(\eta)}{\omega(\xi)}\le\varepsilon^{-1}\,,\quad\mbox{when}\,\,\vert\xi-\eta\vert\le\frac{\varepsilon}{C}
\end{equation}
We follow here the same arguments of the proof of \cite[Theorem 3.8]{G3}. Let us take a function $\varphi\in\mathcal S(\mathbb R^n)$ such that
\begin{equation*}
\widehat{\varphi}(\xi)\ge 0\qquad\mbox{and}\qquad{\rm supp}\,\widehat\varphi\subseteq\left\{\xi\in\mathbb R^n\,:\,\,\vert\xi\vert\le\frac{\varepsilon}{2C}\right\}\,.
\end{equation*}
For arbitrary points $\eta, \theta\in\mathbb R^n$, let us define
\begin{equation}\label{def_f_g}
f(x)=e^{i\eta\cdot x}\varphi(x)\,,\qquad g(x)=e^{i\theta\cdot x}\varphi(x)\,,
\end{equation}
hence
\begin{equation*}
f(x)g(x)=e^{i(\eta+\theta)\cdot x}\varphi^2(x)\,.
\end{equation*}
In view of the assumption on the support of $\widehat\varphi$ we compute:
\begin{equation}\label{norma_feg}
\begin{split}
&\Vert f\Vert^{p_1}_{\mathcal F L^{p_1}_{\omega_1}}=\int\omega_1(\xi)^{p_1}\widehat{\varphi}(\xi-\eta)^{p_1}d\xi=\int_{\vert\xi-\eta\vert\le\frac{\varepsilon}{2C}}\omega_1(\xi)^{p_1}\widehat{\varphi}(\xi-\eta)^{p_1}d\xi\,,\\
&\Vert g\Vert^{p_2}_{\mathcal F L^{p_2}_{\omega_2}}=\int\omega_2(\xi)^{p_2}\widehat{\varphi}(\xi-\theta)^{p_2}d\xi=\int_{\vert\xi-\theta\vert\le\frac{\varepsilon}{2C}}\omega_2(\xi)^{p_2}\widehat{\varphi}(\xi-\theta)^{p_2}d\xi\,,
\end{split}
\end{equation}
In the domain of the integrals above \eqref{pm1} holds, then we get
\begin{equation*}
\Vert f\Vert^{p_1}_{\mathcal F L^{p_1}_{\omega_1}}\le\varepsilon^{-p_1}\omega_1(\eta)^{p_1}\int_{\vert\xi-\eta\vert\le\frac{\varepsilon}{C}}\widehat{\varphi}(\xi-\eta)^{p_1}d\xi=c_{1}^{p_1}\varepsilon^{-p_1}\omega_1(\eta)^{p_1}\,,
\end{equation*}
hence
\begin{equation}\label{stima_norma_f}
\Vert f\Vert_{\mathcal F L^{p_1}_{\omega_1}}\le c_{1}\varepsilon^{-1}\omega_1(\eta)\,,
\end{equation}
where $c_{1}:=\Vert\widehat\varphi\Vert_{L^{p_1}}$. The same holds true for the norm of $g$ in $\mathcal FL^{p_2}_{\omega_2}(\mathbb R^n)$, by replacing $\eta$ with $\theta$, that is
\begin{equation}\label{stima_norma_g}
\Vert g\Vert_{\mathcal F L^{p_2}_{\omega_2}}\le c_2\varepsilon^{-1}\omega_2(\theta)\,,\quad\mbox{for}\,\,c_2=\Vert\widehat\varphi\Vert_{L^{p_2}}\,.
\end{equation}
The preceding calculations are performed under the assumption that both $p_1$ and $p_2$ are finite; however the same estimates \eqref{stima_norma_f}, \eqref{stima_norma_g} can be easily extended to the case when $p_1$ or $p_2$ equals $+\infty$.
As for the norm in $\mathcal F L^p_{\omega}(\mathbb R^n)$ of $fg$ we compute
\begin{equation}\label{norma_fg}
\Vert fg\Vert^{p}_{\mathcal F L^{p}_{\omega}}=\int_{\vert\xi-\eta-\theta\vert\le\frac{\varepsilon}{C}}\omega(\xi)^{p}\widehat{\varphi^2}(\xi-\eta-\theta)^{p}d\xi\,,
\end{equation}
where we used that ${\rm supp}\,\widehat{\varphi^2}={\rm supp}\,(\widehat\varphi\ast\widehat\varphi)\subseteq\left\{\vert\xi\vert\le\frac{\varepsilon}{C}\right\}$ and it is assumed $p<+\infty$ (to fix the ideas). Then recalling again that \eqref{pm1} holds true for $\omega$ on ${\rm supp}\,\widehat{\varphi^2}$ we obtain
\begin{equation}\label{stima_norma_fg}
\Vert fg\Vert_{\mathcal F L^{p}_{\omega}}\ge c\varepsilon\omega(\eta+\theta)\,,\quad\mbox{with}\,\,c:=\Vert\widehat{\varphi^2}\Vert_{L^p}\,.
\end{equation}
The same estimate \eqref{stima_norma_fg} can be easily recovered in the case $p=+\infty$.
We use now \eqref{stima_norma_f}, \eqref{stima_norma_g} and \eqref{stima_norma_fg} to estimate the right- and left-hand sides of \eqref{stima_algebra2} written for $f$ and $g$ defined in \eqref{def_f_g} to get
\begin{equation*}
c\varepsilon\omega(\eta+\theta)\le C^\prime c_{1}c_2\varepsilon^{-2}\omega_1(\eta)\omega_2(\theta)\,.
\end{equation*}
In view of the arbitrariness of $\eta$, $\theta$ and since the constants $c_1$, $c_2$ and $\varepsilon$ are independent of $\eta$ and $\theta$ the preceding inequality gives \eqref{stima_algebra1} with $C=C^\prime c_{1}c_2c^{-1}\varepsilon^{-3}$.
\end{proof}
\begin{remark}\label{rmk:7}
{\rm It is worth observing that any specific relation is assumed on the exponents $p_1, p_2, p\in [1,+\infty]$ in the statement of Proposition \ref{prop_sm}. Notice also that condition \eqref{stima_algebra1} is just \eqref{B_Lq} for $q=+\infty$. When in particular $\omega_1=\omega_2=\omega$, it reduces to ($\mathcal{SM}$).
Notice also that from the results given by Corollary \ref{algebra_prop2} (see also Remark \ref{rmk:6}) and Proposition \ref{prop_sm}, we derive that condition ($\mathcal{SM}$) is {\it necessary and sufficient} to make the Fourier Lebesgue space $\mathcal F L^1_{\omega}(\mathbb R^n)$ an algebra for the point-wise product\footnote{However condition ($\mathcal{SM}$) is far from being sufficient for $\mathcal F L^p_{\omega}(\mathbb R^n)$ to be an algebra, as long as $p>1$.}.
}
\end{remark}
Combining the results of Corollary \ref{algebra_prop2} with the remarks made about the weight functions quoted in the examples 1-3 at the end of Section \ref{wf_sct} we can easily prove the following result.
\begin{corollary}\label{algebra_prop3}
Let $r\in\mathbb R$, $M=(m_1,\dots,m_n)\in\mathbb N^n$ and $\mathcal P$ a complete polyhedron of $\mathbb R^n$ be given and assume that $p,q\in[1,+\infty]$ satisfy $\frac1{p}+\frac1{q}=1$. Then
\begin{itemize}
\item[i.] $\left(\mathcal F L^p_{r}(\mathbb R^n)\,,\,\cdot\right)$ is an algebra if $r>\frac{n}{q}$;
\item[ii.] $\left(\mathcal F L^p_{r,M}(\mathbb R^n)\,,\,\cdot\right)$ is an algebra if $r>\frac{n}{m_\ast q}$, where $m_\ast=\min\limits_{1\le j\le n}m_j$;
\item[iii.] $\left(\mathcal F L^p_{r,\mathcal P}(\mathbb R^n)\,,\,\cdot\right)$ is an algebra if $r>\frac{n}{(1-\delta)\mu_0 q}$, where $\mu_0$ and $\delta$ are defined in \eqref{min_max_order} and \eqref{delta}.
\end{itemize}
Analogous statements hold true for the localized version of the previous spaces on an open subset $\Omega$ of $\mathbb R^n$, defined according to Definition \ref{loc_w_FL_spaces}.
\end{corollary}
\begin{proof}
Let us prove the statement iii of the Theorem; the proof of the other statements is completely analogous. Assume that $p>1$, thus $q<+\infty$. From \eqref{pg_mqe} we have that $s>\frac{n}{\mu_0 q}$ implies $\lambda_{\mathcal P}^{-sq}\in L^1(\mathbb R^n)$; on the other hand, $\lambda_{\mathcal P}^{sq}$ satisfies ($\mathcal{G}$) with $\delta$ defined as in \eqref{delta}, see Example 3 in Sect. \ref{prel_sct}. From Proposition \ref{wf_relationship} applied to $\lambda_{\mathcal P}^{sq}$ we derive that $\lambda_{\mathcal P}^{rq}$ fulfils condition ($\mathcal{B}$), which amounts to say that $\lambda_{\mathcal P}^{r}$ satisfies \eqref{B_q}, where $r=\frac{s}{1-\delta}$. Then the result of Corollary \ref{algebra_prop2} applies to $\mathcal FL^p_{\lambda_{\mathcal P}^{r}}(\mathbb R^n)=\mathcal F L^p_{r, \mathcal P}(\mathbb R^n)$ and gives the statement iii. Notice that condition $r>\frac{n}{(1-\delta)\mu_0 q}$ reduces to $r>0$ when $q=+\infty$ (corresponding to $p=1$). That $\mathcal F L^1_{r,\mathcal P}(\mathbb R^n)$ with $r>0$ is an algebra for the point-wise product follows again from Corollary \ref{algebra_prop2} by observing that $\lambda_{\mathcal P}^r$ satisfies ($\mathcal{SM}$) (that is \eqref{B_q} for $q=+\infty$).
\end{proof}
\begin{remark}\label{rmk:6}
{\rm In agreement with the observation made at the end of Section \ref{FL_spaces_sct}, for $p=2$ the lower bounds of $r$ given in i-iii of Corollary \ref{algebra_prop3} are exactly the same required to ensure the algebra property for the corresponding weighted Sobolev spaces (see \cite{G3} and \cite{GM2} for the case of a general $1<p<+\infty$).}
\end{remark}
To the end of this section, let us observe that, as a byproduct of Propositions \ref{algebra_prop} and \ref{prop_sm}, the following result can be proved.
\begin{proposition}\label{algebra_prop4}
Assume that $\omega$, $\omega_1$, $\omega_2$ are temperate weight functions satisfying condition \eqref{B_Lq} for some $1\le q<+\infty$. Then $\omega$, $\omega_1$, $\omega_2$ also satisfy condition \eqref{stima_algebra1}. In particular, if $\omega$ is a temperate weight function satisfying condition \eqref{B_q} for some $1\le q<+\infty$ then it also satisfies condition ($\mathcal{SM}$).
\end{proposition}
\begin{remark}\label{rmk:10}
{\rm The second part of Proposition \ref{algebra_prop4} slightly improves the result of \cite[Proposition 2.4]{G3}, where the sub-multiplicative condition ($\mathcal{SM}$) was deduced from Beurling condition ($\mathcal{B}$) (corresponding to \eqref{B_q} with $q=1$) and conditions ($\mathcal{SV}$) and \eqref{bb_cond}; here $\omega$, $\omega_1$, $\omega_2$ are only required to satisfy condition ($\mathcal T$) (included in our definition of weight function), which is implied by ($\mathcal{SV}$) and \eqref{bb_cond} in view of Proposition \ref{wf_relationship}.i.}
\end{remark}
\section{Pseudodifferential operators with symbols in weighted Fourier Lebesgue spaces}\label{w_FL_pdo_sct}
This section is devoted to the study of a class of pseudodifferential operators whose symbols $a(x,\xi)$ have a finite regularity of weighted Fourier Lebesgue type with respect to $x$.
Let us first recall that, under the only assumption $a(x,\xi)\in\mathcal S^\prime(\mathbb R^{2n})$, the pseudodifferential operator defined by
\begin{equation}\label{pdo}
a(x,D)f=(2\pi)^{-n}\int e^{ix\cdot\xi} a(x,\xi)\widehat f(\xi)d\xi\,,\qquad f\in\mathcal S(\mathbb R^n)\,,
\end{equation}
maps continuously $\mathcal S(\mathbb R^n)$ to $\mathcal S^\prime(\mathbb R^n)$\footnote{The integral in the right-hand side of \eqref{pdo} must be understood here in a weak (distributional) sense.}. Similarly, if $\Omega$ is an open subset of $\mathbb R^n$ and $a(x,\xi)\in\mathcal D^\prime(\Omega\times\mathbb R^n)$ is such that $\varphi(x)a(x,\xi)\in\mathcal S^\prime(\mathbb R^n\times\mathbb R^n)$ for every $\varphi\in C^\infty_0(\Omega)$, then \eqref{pdo} defines a linear continuous operator from $\mathcal S(\mathbb R^n)$ to $\mathcal D^\prime(\Omega)$.
Let us also recall that, as a linear continuous operator from $C^\infty_0(\Omega)$ to $\mathcal D^\prime(\Omega)$, every pseudodifferential operator with symbol $a(x,\xi)\in\mathcal D^\prime(\Omega\times\mathbb R^n)$ admits a (uniquely defined) Schwartz kernel $\mathcal K_a(x,y)\in\mathcal D^\prime(\Omega\times\Omega)$ such that
\begin{equation*}
\langle a(x,D)\psi,\varphi\rangle=\langle\mathcal K_a,\varphi\otimes\psi\rangle\,,\quad\forall\,\varphi\,,\psi\in C^\infty_0(\Omega)\,.
\end{equation*}
The operator $a(x,D)$ is said to be {\it properly supported} on $\Omega$ when the support of $\mathcal K_a$ is a {\it proper} subset of $\Omega\times\Omega$, that is ${\rm supp}\,\mathcal K_a\cap(\Omega\times K)$ and ${\rm supp}\,\mathcal K_a\cap (K\times\Omega)$ are compact subsets of $\Omega\times\Omega$, for every compact set $K\subset\Omega$. It is well known that every properly supported pseudodifferential operator continuously maps $C^\infty_0(\Omega)$ into the space $\mathcal E^\prime(\Omega)$ of compactly supported distributions and it extends as a linear continuous operator from $C^\infty(\Omega)$ into $\mathcal D^\prime(\Omega)$. In particular for every function $\phi\in C^\infty_0(\Omega)$ another function $\tilde\phi\in C^\infty_0(\Omega)$ can be found in such a way that
\begin{equation}\label{eq_pr_supp}
\phi(x) a(x,D)u=\phi(x) a(x,D)(\tilde\phi u)\,,\quad \forall\,u\in C^\infty(\Omega)\,.
\end{equation}
Following \cite{G2}, we introduce some local and global classes of symbols with finite Fourier Lebesgue regularity.
\begin{definition}\label{symbols}
Let $\omega=\omega(\xi)$, $\gamma=\gamma(\xi)$ be arbitrary weight functions.
\begin{itemize}
\item[1.]
A distribution $a(x,\xi)\in\mathcal S^\prime(\mathbb R^{2n})$ is said to belong to the class $\mathcal F L^p_{\omega} S_{\gamma}$ if $\xi\mapsto a(\cdot,\xi)$ is a measurable $\mathcal F L^p_{\omega}(\mathbb R^n)-$valued function on $\mathbb R^n_\xi$ such that
\begin{equation}\label{symbols_condt1}
\left\Vert\frac{a(\cdot,\xi)}{\gamma(\xi)}\right\Vert_{\mathcal F L^p_{\omega}}\le C\,,\quad\forall\,\xi\in\mathbb R^n\,,
\end{equation}
for some constant $C>0$. More explicitly, the above means that
\begin{equation*}
\frac{\omega(\eta)\widehat{a(\cdot,\xi)}(\eta)}{\gamma(\xi)}\in L^p(\mathbb R^n_\eta)\,,
\end{equation*}
with norm uniformly bounded with respect to $\xi$.
\item[2.] We say that a distribution $a(x,\xi)\in\mathcal D^\prime(\Omega\times\mathbb R^n)$, where $\Omega$ is an open subset of $\mathbb R^n$, belongs to $\mathcal F L^p_{\omega}S_{\gamma}(\Omega)$ if $\phi(x)a(x,\xi)\in \mathcal F L^p_{\omega} S_{\gamma}$ for every $\phi\in C^\infty_0(\Omega)$ (which amounts to have that $a(\cdot,\xi)/\gamma(\xi)\in\mathcal F L^p_{\omega,{\rm loc}}(\Omega)$ uniformly in $\xi$).
\end{itemize}
\end{definition}
\begin{remark}\label{rmk:11}
{\rm For $\omega$, $\gamma$ as in Definition \ref{symbols}, $\mathcal F L^p_{\omega} S_{\gamma}$ is a Banach space with respect to the norm
\begin{equation}\label{symbol_norm1}
\Vert a\Vert_{\mathcal FL^p_{\omega}S_\gamma}:=\sup\limits_{\xi\in\mathbb R^n}\left\Vert\frac{a(\cdot,\xi)}{\gamma(\xi)}\right\Vert_{\mathcal F L^p_{\omega}}\,,
\end{equation}
while $\mathcal FL^p_{\omega}S_\gamma(\Omega)$ is a Fr\'echet space with respect to the family of semi-norms
\begin{equation}\label{symbol_norm2}
a\mapsto \Vert\phi a\Vert_{\mathcal FL^p_{\omega}S_\gamma}\,,\qquad\phi\in C^\infty_0(\Omega)\,.
\end{equation}}
\end{remark}
Let us point out that any assumption is made about the $\xi-$derivatives of the symbol $a(x,\xi)$ in the above definition: the weight function $\gamma$ only measures the $\xi-$decay at infinity of the symbol itself. It is clear that $\mathcal F L^p_{\omega} S_{\gamma_1}\equiv \mathcal F L^p_{\omega} S_{\gamma_2}$, whenever $\gamma_1\sim\gamma_2$ (the same applies of course to the corresponding local classes on an open set). When the weight function $\gamma$ is an arbitrary positive constant, the related symbol class $\mathcal F L^p_{\omega} S_{\gamma}$ will be simply denoted as $\mathcal F L^p_{\omega} S$, and its symbols (and related pseudo-differential operators) will be referred to as {\it zero-th order symbols} (and {\it zero-th order operators}). Finally, we notice that for every weight function $\omega=\omega(\xi)$ and $p\in[1,+\infty]$, the inclusion $\mathcal F L^p_{\omega}(\mathbb R^n)\subset\mathcal F L^p_{\omega} S$ holds true (the elements of $\mathcal FL^p_{\omega}(\mathbb R^n)$ being regarded as $\xi-$independent symbols).
\begin{proposition}\label{continuity1}
For $p\in[1,+\infty]$ let the weight functions $\omega$, $\omega_1$, $\omega_2$ and $\gamma$ satisfy
\begin{equation}\label{B_q1}
C_q:=\sup\limits_{\xi\in\mathbb R^n}\left\Vert\frac{\omega_2(\xi)\gamma(\cdot)}{\omega_1(\cdot)\omega(\xi-\cdot)}\right\Vert_{L^q}<+\infty\,,
\end{equation}
where $q$ is the conjugate exponent of $p$. Then the following hold true.
\begin{itemize}
\item[i.] For every $a(x,\xi)\in\mathcal FL^p_{\omega}S_\gamma$ the pseudodifferential operator $a(x,D)$ extends to a unique linear bounded operator
\begin{equation*}
a(x,D):\mathcal FL^p_{\omega_1}(\mathbb R^n)\rightarrow\mathcal FL^p_{\omega_2}(\mathbb R^n)\,.
\end{equation*}
\item[ii.] For every $a(x,\xi)\in\mathcal FL^p_{\omega}S_{\gamma}(\Omega)$, with $\Omega$ open subset of $\mathbb R^n$, the pseudodifferential operator $a(x,D)$ extends to a unique linear bounded operator
\begin{equation*}
a(x,D):\mathcal FL^p_{\omega_1}(\mathbb R^n)\rightarrow\mathcal FL^p_{\omega_2,{\rm loc}}(\Omega)\,.
\end{equation*}
If in addition the pseudodifferential operator $a(x,D)$ is properly supported, then it extends to a linear bounded operator
\begin{equation*}
a(x,D):\mathcal FL^p_{\omega_1,{\rm loc}}(\Omega)\rightarrow\mathcal FL^p_{\omega_2,{\rm loc}}(\Omega)\,.
\end{equation*}
\end{itemize}
\end{proposition}
\begin{proof}
The second part of the statement {\it ii} is an immediate consequence of the first one; indeed, since the operator $a(x,D)$ is properly supported, for every function $\phi\in C^\infty_0(\Omega)$ another function $\tilde\phi\in C^\infty_0(\Omega)$ can be chosen in such a way that $\phi a(\cdot,D)u=\phi a(\cdot,D)(\tilde\phi u)$, cf. \eqref{eq_pr_supp}.
The first part of the statement {\it ii} follows, in its turn, from the statement {\it i} by noticing that $\phi(x)a(x,\xi)\in \mathcal FL^p_{\omega}S_\gamma$ for every function $\phi\in C^\infty_0(\Omega)$ (cf. Definition \ref{symbols}).
As for the proof of the statement {\it i}, we first observe that for every $u\in\mathcal S(\mathbb R^n)$ one computes
\begin{equation}\label{representation1}
\widehat{a(\cdot,D)u}(\eta)=(2\pi)^{-n}\int \widehat{a}(\eta-\xi,\xi)\widehat{u}(\xi)\,d\xi\,,
\end{equation}
where $\widehat{a}(\eta,\xi):=\widehat{a(\cdot,\xi)}(\eta)$ denotes the partial Fourier transform of the symbol $a(x,\xi)$ with respect to $x$. From \eqref{representation1} we find the following integral representation
\begin{equation}\label{representation2}
\begin{split}
\omega_2(\eta)&\widehat{a(\cdot,D)u}(\eta)=(2\pi)^{-n}\int \omega_2(\eta)\widehat{a}(\eta-\xi,\xi)\widehat{u}(\xi)\,d\xi\\
&=(2\pi)^{-n}\int \frac{\omega_2(\eta)\gamma(\xi)}{\omega(\eta-\xi)\omega_1(\xi)}\,\frac{\omega(\eta-\xi)\widehat{a}(\eta-\xi,\xi)}{\gamma(\xi)}\,\omega_1(\xi)\widehat{u}(\xi)\,d\xi\\
&=(2\pi)^{-n}\int F(\eta,\xi)f(\eta-\xi,\xi)g(\xi)\,d\xi\,,
\end{split}
\end{equation}
where it is set
\begin{equation}\label{positions}
f(\zeta,\xi)=\frac{\omega(\zeta)\widehat{a}(\zeta,\xi)}{\gamma(\xi)}\,,\,\, F(\zeta,\xi)=\frac{\omega_2(\zeta)\gamma(\xi)}{\omega(\zeta-\xi)\omega_1(\xi)}\,,\,\, g(\xi)=\omega_1(\xi)\widehat{u}(\xi)\,.
\end{equation}
The assumptions of Proposition \ref{continuity1} (see \eqref{B_q1}, \eqref{symbols_condt1}) yield the following
\begin{equation*}
\sup\limits_{\zeta\in\mathbb R^n}\Vert F(\zeta,\cdot)\Vert_{L^q}=C_q\,,\quad\sup\limits_{\xi\in\mathbb R^n}\Vert f(\cdot,\xi)\Vert_{L^p}=\Vert a\Vert_{\mathcal FL^p_{\omega}S_\gamma}\,,\quad g\in L^p(\mathbb R^n)\,.
\end{equation*}
Now we apply to $\omega_2(\eta)\widehat{a(\cdot,D)u}(\eta)$, written as the integral operator in \eqref{representation2}, the result of Lemma \ref{lemma_cont2}. Then we have
\begin{equation*}
\begin{split}
\Vert a(\cdot,D)u\Vert_{\mathcal F L^p_{\omega_2}}&=\Vert\omega_2 \widehat{a(\cdot,D)u}\Vert_{L^p}\le(2\pi)^{-n}C_q \Vert a\Vert_{\mathcal FL^p_{\omega}S_\gamma}\Vert u\Vert_{\mathcal FL^p_{\omega_1}}\,.
\end{split}
\end{equation*}
\end{proof}
\begin{remark}\label{rmk:12}
{\rm It is clear that Proposition \ref{continuity1} provides a generalization of the result given by Proposition \ref{algebra_prop}; indeed the multiplication by a given function $v=v(x)\in \mathcal FL^p_{\omega}(\mathbb R^n)\subset \mathcal FL^p_{\omega}S$ can be thought to as the zero-th order pseudodifferential operator with $\xi-$independent symbol $a(x,\xi)=v(x)$, cf. Remark \ref{rmk:10}.}
\end{remark}
\section{Microlocal regularity in weighted Fourier Lebesgue spaces}\label{micro_w_FL_sct}
This section is devoted to introduce a microlocal counterpart of the weighted Fourier Lebesgue spaces presented in Section \ref{FL_spaces_sct} and to define corresponding classes of pseudodifferential operators, with finitely regular symbols, naturally acting on such spaces.
Because of the lack of homogeneity of a generic weight function $\omega=\omega(\xi)$, in order to perform a microlocal analysis in the framework of weighted Fourier Lebesgue spaces it is convenient to replace the usual conic neighborhoods (used in Pilipovi\'c et al. \cite{PTT1, PTT2}) by a suitable notion of {\it $\varepsilon-$neighborhood} of a set, modeled on the weight function itself, following the approach of Rodino \cite{RO1} and Garello \cite{G}.
In the following, let $\omega:\mathbb R^n\rightarrow ]0,+\infty[$ be a weight function satisfying the subadditivity condition ($\mathcal{SA}$) and
\medskip
($\mathcal{SH}$): for a suitable constant $C\ge 0$
\begin{equation}\label{q_h}
\omega(t\xi)\le C\omega(\xi)\,,\quad\forall\,\xi\in\mathbb R^n\,,\,\,\vert t\vert\le 1\,.
\end{equation}
Every weight function $\omega=\omega(\xi)$ satisfying ($\mathcal{SA}$) and ($\mathcal{SH}$) also obeys the following
\begin{equation}\label{strong_sv}
\frac1{C}\le\frac{\omega(\xi+t\theta)}{\omega(\xi)}\le C\,,\qquad\mbox{when}\,\,\omega(\theta)\le\frac1{C}\omega(\xi)\,,\,\,\vert t\vert\le 1\,,
\end{equation}
for a suitable constant $C>1$, cf.\cite{G}.
Throughout the whole section, the weight function $\omega=\omega(\xi)$ will be assumed to be continuous\footnote{This assumption is not as much restrictive, since it can be shown (see e.g. \cite{T1}) that for any weight function $\omega$ an equivalent weight function $\omega_0$ exists such that $\omega_0\in C^\infty(\mathbb R^n)$ and $\partial^\alpha\omega_0/\omega_0\in L^\infty(\mathbb R^n)$ for each multi-index $\alpha$.}. Then an easy consequence of condition ($\mathcal{SH}$) is that $\omega(\xi)$ satisfies \eqref{bb_cond}.
To every set $X\subset\mathbb R^n$ one may associate a one-parameter family of open sets by defining for any $\varepsilon>0$
\begin{equation}\label{omega_neighb}
X_{[\varepsilon\omega]}:=\bigcup\limits_{\xi_0\in X}\left\{\xi\in\mathbb R^n\,:\,\,\omega(\xi-\xi_0)<\varepsilon\omega(\xi_0)\right\}\,.
\end{equation}
We call $X_{[\varepsilon\omega]}$ the {\it $[\omega]-$neighborhood} of $X$ of size $\varepsilon$.
\begin{remark}\label{rmk:intorni}
{\rm Since $\left\{\xi\in\mathbb R^n\,;\,\,\omega(\xi-\xi_0)<\varepsilon\omega(\xi_0)\right\}=\emptyset$ when $\omega(\xi_0)\le c/\varepsilon$, where $c$ is the constant in \eqref{bb_cond}, we effectively have
\begin{equation*}
X_{[\varepsilon\omega]}=\!\!\!\!\bigcup\limits_{\xi_0\in X\,:\,\,\omega(\xi_0)>\frac{c}{\varepsilon}}\!\!\!\left\{\omega(\xi-\xi_0)<\varepsilon\omega(\xi_0)\right\}\,,
\end{equation*}
and for $X$ bounded a constant $\varepsilon_0=\varepsilon_0(X)>0$ exists such that $X_{[\varepsilon\omega]}=\emptyset$ when $0<\varepsilon<\varepsilon_0$.}
\end{remark}
As a consequence of ($\mathcal{SA}$), ($\mathcal{SH}$) and \eqref{strong_sv}, the $[\omega]$-neighborhoods of a set $X$ fulfil the following lemma.
\begin{lemma}\label{mcl_lemma1}
Given $\varepsilon>0$, there exists $0<\varepsilon^\prime<\varepsilon$ such that for every $X\subset\mathbb R^n$
\begin{eqnarray}
\left(X_{[\varepsilon^\prime\omega]}\right)_{[\varepsilon^\prime\omega]}\subset X_{[\varepsilon\omega]}\,;\label{inc_1}\\
\left(\mathbb R^n\setminus X_{[\varepsilon\omega]}\right)_{[\varepsilon^\prime\omega]}\subset\mathbb R^n\setminus X_{[\varepsilon^\prime\omega]}\,.\label{inc_2}
\end{eqnarray}
Moreover there exist constants $\widehat c>0$ and $0<\widehat\varepsilon<1$ such that for all $X\subset\mathbb R^n$ and $0<\varepsilon\le\widehat\varepsilon$
\begin{equation}\label{mcl_impl}
\xi\in X_{[\varepsilon\omega]}\quad\mbox{yields}\quad \omega(\xi)>\frac{\widehat c}{\varepsilon}\,.
\end{equation}
\end{lemma}
\begin{proof}
\eqref{inc_1} and \eqref{inc_2} are direct consequences of ($\mathcal{SA}$), ($\mathcal{SH}$) and \eqref{strong_sv}, see \cite{G} for details.
If $\xi\in X_{[\varepsilon\omega]}$ then $\xi_0\in X$ exists such that
\begin{equation}\label{mcl_eqt1}
\omega(\xi-\xi_0)<\varepsilon\omega(\xi_0)\,,
\end{equation}
hence $\omega(\xi-\xi_0)\ge c$ implies $\omega(\xi_0)>\displaystyle\frac{c}{\varepsilon}$, cf. \eqref{bb_cond} and Remark \ref{rmk:intorni}.
In view of \eqref{strong_sv}
\begin{equation*}
\omega(\xi)=\omega(\xi_0+(\xi-\xi_0))\ge\frac1{C}\omega(\xi_0)>\frac{c}{C\varepsilon}
\end{equation*}
follows from \eqref{mcl_eqt1}, provided that $\varepsilon\le\displaystyle\frac1{C}$ (where $C$ is the same constant involved in \eqref{strong_sv}).
\end{proof}
We use the notion of $[\omega]-$neighborhood of a set to define a microlocal version of the weighted Fourier Lebesgue spaces.
\begin{definition}\label{micro_FL_space}
We say that a distribution $u\in\mathcal S^\prime(\mathbb R^n)$ belongs microlocally to $\mathcal F L^p_{\omega}$ at $X\subset\mathbb R^n$, writing $u\in\mathcal FL^p_{\omega,{\rm mcl}}(X)$, $p\in[1,+\infty]$, if there exists $\varepsilon>0$ such that
\begin{equation}\label{mcl_FLp_1}
\vert u\vert_{X_{[\varepsilon\omega]}}^p:=\int_{X_{[\varepsilon\omega]}}\omega(\xi)^p\vert\widehat{u}(\xi)\vert^p\,d\xi<+\infty
\end{equation}
(with obvious modification for $p=+\infty$).
For $\Omega$ open subset of $\mathbb R^n$, $x_0\in\Omega$ and $X\subset\mathbb R^n$, we say that a distribution $u\in\mathcal D^\prime(\Omega)$ belongs microlocally to $\mathcal F L^p_{\omega}$ on the set $X$ at the point $x_0$, writing $u\in\mathcal F L^p_{\omega, {\rm mcl}}(x_0\times X)$, if there exists a function $\phi\in C^\infty_0(\Omega)$ such that $\phi(x_0)\neq 0$ and $\phi u\in FL^p_{\omega,{\rm mcl}}(X)$.
\end{definition}
\begin{remark}\label{rmk:13}
{\rm
In view of Remark \ref{rmk:intorni}, condition \eqref{mcl_FLp_1} is meaningful only for {\it unbounded} $X$.}
\end{remark}
We can say that $u\in\mathcal FL^p_{\omega,{\rm mcl}}(X)$ and $u\in\mathcal F L^p_{\omega, {\rm mcl}}(x_0\times X)$ if respectively
\begin{equation}\label{mcl_FLp_2}
\chi_{[\varepsilon\omega]}(\xi)\omega(\xi)\widehat{u}(\xi)\in L^p(\mathbb R^n)
\end{equation}
and
\begin{equation}\label{mcl_FLp_3}
\chi_{[\varepsilon\omega]}(\xi)\omega(\xi)\widehat{\phi u}(\xi)\in L^p(\mathbb R^n)\,,
\end{equation}
where $\chi_{[\varepsilon\omega]}=\chi_{[\varepsilon\omega]}(\xi)$ denotes the characteristic function of the set $X_{[\varepsilon\omega]}$ and $\varepsilon>0$, $\phi=\phi(x)$ are given as in Definition \ref{micro_FL_space}.
\vspace{.5cm}
According to Definition \ref{micro_FL_space} one can introduce the notion of {\it filter of Fourier Lebesgue singularities}, which is in some way the extension of the wave front set of Fourier Lebesgue singularities when we lack the homogeneity properties necessary to use effectively conic neighborhoods.
\begin{definition}\label{FL_sing}
For $u\in\mathcal D^\prime(\Omega)$, $x_0\in\Omega$, $p\in[1,+\infty]$, we call filter of $\mathcal FL^p_\omega-$singularities of $u$ at the point $x_0$ the class of all sets $X\subset\mathbb R^n$ such that $u\in\mathcal FL^p_{\omega, {\rm mcl}}(x_0\times(\mathbb R^n\setminus X))$. It may be easily verified that
\begin{equation}\label{filtroFLp}
\Xi_{\mathcal F L^p_{\omega}\,,\,x_0}u:=\bigcup\limits_{\phi\in C^\infty_0(\Omega)\,,\,\,\phi(x_0)\neq 0}\Xi_{\mathcal F L^p_{\omega}}\phi u\,,
\end{equation}
where for every $v\in\mathcal S^\prime(\mathbb R^n)$, $\Xi_{\mathcal F L^p_{\omega}}v$ is the class of all sets $X\subset\mathbb R^n$ such that $v\in\mathcal FL^p_{\omega,{\rm mcl}}(\mathbb R^n\setminus X)$.
\end{definition}
$\Xi_{\mathcal F L^p_{\omega}}v$ and $\Xi_{\mathcal F L^p_{\omega}\,,\,x_0} u$ defined above are {\it $[\omega]-$filters}, in the sense that they satisfy the standard filter properties and moreover for all $X\in \Xi_{\mathcal F L^p_{\omega}}v$ (respectively $X\in \Xi_{\mathcal F L^p_{\omega}\,,\,x_0} u$) there exists $\varepsilon>0$ such that $\mathbb R^n\setminus(\mathbb R^n\setminus X)_{[\varepsilon\omega]}\in \Xi_{\mathcal F L^p_{\omega}}v$ (respectively $\mathbb R^n\setminus(\mathbb R^n\setminus X)_{[\varepsilon\omega]}\in \Xi_{\mathcal F L^p_{\omega}\,,\,x_0} u$), see e.g. \cite{TR} for the definition and properties of a filter.
\subsection{Symbols with microlocal regularity in spaces of Fourier Lebesgue type}\label{mcl_FLp_symb_sect}
Throughout the whole section, we assume that $\lambda=\lambda(\xi)$ and $\Lambda=\Lambda(\xi)$ are two continuous weight functions, such that $\lambda$ satisfies condition \eqref{bb_cond} and $\Lambda$ conditions ($\mathcal{SA}$) and ($\mathcal{SH}$).
For given $p\in[1,+\infty]$ and $X\subset\mathbb R^n$, the space $\mathcal F L^p_{\lambda}(\mathbb R^n)\cap\mathcal F L^p_{\Lambda, {\rm mcl}}(X)$ is provided with the {\it inductive limit} locally convex topology defined on it by the family of subspaces
\begin{equation*}
\mathcal F L^p_{\lambda}(\mathbb R^n)\cap\mathcal F L^p_{\Lambda, \varepsilon}(X):=\{u\in \mathcal F L^p_{\lambda}(\mathbb R^n)\,:\,\,\vert u\vert_{X_{[\varepsilon\Lambda]}}<+\infty\}
\end{equation*}
(cf. \eqref{mcl_FLp_1}), endowed with their natural semi-norm
\begin{equation*}
\Vert u\Vert_{\mathcal F L^p_{\lambda}}+\vert u\vert_{X_{[\varepsilon\Lambda]}}\,,\quad\varepsilon>0\,.
\end{equation*}
Analogously for every $x_0\in\Omega$, the space $\mathcal F L^p_{\lambda\,,{\rm loc}}(x_0)\cap\mathcal F L^p_{\Lambda\,, {\rm mcl}}(x_0\times X)$ is provided with the inductive limit topology defined by the subspaces
\begin{equation*}
\mathcal F L^p_{\lambda\,,\phi}\cap\mathcal F L^p_{\Lambda, \varepsilon}(X):=\{u\in\mathcal D^\prime(\Omega)\,:\,\,\phi u\in \mathcal F L^p_{\lambda}(\mathbb R^n)\cap\mathcal F L^p_{\Lambda, \varepsilon}(X)\}\,,
\end{equation*}
endowed with the natural semi-norms
\begin{equation*}
\Vert \phi u\Vert_{\mathcal F L^p_{\lambda}}+\vert \phi u\vert_{X_{[\varepsilon\Lambda]}}\,,\quad \phi\in C^\infty_0(\Omega),\,\,\phi(x_0)\neq 0\,,\,\,\varepsilon>0\,.
\end{equation*}
From the general properties of the inductive limit topology (see e.g. \cite{TR}), it follows that a sequence $\{u_\nu\}$ converges to $u$ in $\mathcal F L^p_{\lambda}(\mathbb R^n)\cap\mathcal F L^p_{\Lambda\,, {\rm mcl}}(X)$ (resp. $\mathcal F L^p_{\lambda\,,{\rm loc}}(x_0)\cap\mathcal F L^p_{\Lambda\,, {\rm mcl}}(x_0\times X)$) if and only if there exists some $\varepsilon>0$ such that
\begin{equation*}
\Vert u_\nu-u\Vert_{\mathcal F L^p_{\lambda}}\to 0\,\,\,\mbox{and}\,\,\,\vert u_\nu-u\vert_{X_{[\varepsilon\Lambda]}}\to 0\,,\,\,\mbox{as}\,\,\nu\to +\infty
\end{equation*}
(resp. there exist $\phi\in C^\infty_0(\Omega)$, with $\phi(x_0)\neq 0$, and $\varepsilon>0$ such that
\begin{equation*}
\Vert \phi(u_\nu-u)\Vert_{\mathcal F L^p_{\lambda}}\to 0\,\,\,\mbox{and}\,\,\,\vert \phi(u_\nu-u)\vert_{X_{[\varepsilon\Lambda]}}\to 0\,,\,\,\mbox{as}\,\,\nu\to +\infty).
\end{equation*}
\begin{definition}\label{def_mcl_symb}
Let $\lambda=\lambda(\xi)$, $\Lambda=\Lambda(\xi)$ be two weight functions as above and $\gamma=\gamma(\xi)$ a further continuous weight function, $x_0\in\Omega$, $X\subset\mathbb R^n$ and $p\in[1,+\infty]$. We say that a distribution $a(x,\xi)\in\mathcal D^\prime(\Omega\times\mathbb R^n)$ belongs to $\mathcal F L^p_{\lambda,\, \Lambda}S_\gamma(x_0\times X)$ if the function $\xi\mapsto a(\cdot,\xi)$ takes values in the space $\mathcal F L^p_{\lambda,{\rm loc}}(x_0)\cap\mathcal F L^p_{\Lambda, {\rm mcl}}(x_0\times X)$ and for some $\phi\in C^\infty_0(\Omega)$ such that $\phi(x_0)\neq 0$ and $\varepsilon>0$ there holds
\begin{equation}\label{symb}
\begin{split}
\Vert a\Vert_{\phi,\lambda,\gamma}&:=\sup\limits_{\xi\in\mathbb R^n}\left\Vert\frac{\lambda(\cdot)\widehat{\phi a} (\cdot,\xi)}{\gamma(\xi)}\right\Vert_{L^p}<+\infty\quad\mbox{and}\\
&\vert a\vert_{\phi,\Lambda,\gamma,\varepsilon,X}:=\sup\limits_{\xi\in\mathbb R^n}\left\Vert\frac{\Lambda(\cdot)\chi_{\varepsilon,\Lambda}(\cdot)\widehat{\phi a} (\cdot,\xi)}{\gamma(\xi)}\right\Vert_{L^p}<+\infty\,,
\end{split}
\end{equation}
where $\widehat{\phi a}(\eta,\xi):=\mathcal F_{x\to\eta}\left(\phi(x)a(x,\xi)\right)(\eta)$ denotes the partial Fourier transform of $\phi(x)a(x,\xi)$ with respect to $x$.
\end{definition}
\begin{theorem}\label{continuity2}
For $p\in[1,+\infty]$, $x_0\in\Omega$, $X\subset\mathbb R^n$, let $\lambda=\lambda(\xi)$, $\Lambda=\Lambda(\xi)$, $\gamma=\gamma(\xi)$, $\sigma=\sigma(\xi)$ be weight functions such that $\lambda$ obeys condition \eqref{B_q}, where $q$ is the conjugate exponent of $p$, $\Lambda$ conditions ($\mathcal{SA}$), ($\mathcal{SH}$), $1/\sigma\in L^q(\mathbb R^n)$ and
\begin{equation}\label{cndt:2}
\quad\sigma(\xi)\preceq\lambda(\xi)\preceq\Lambda(\xi)\preceq\frac{\lambda(\xi)^2}{\sigma(\xi)}\,.
\end{equation}
\begin{itemize}
\item[(i)] If $a(x,\xi)\in \mathcal F L^p_{\lambda,\, \Lambda}S_\gamma(x_0\times X)$ then the corresponding pseudodifferential operator $a(x,D)$ extends to a bounded linear operator
\begin{equation}\label{mcl_cont1}
\mathcal F L^p_{\lambda\gamma}(\mathbb R^n)\cap\mathcal F L^p_{\Lambda\gamma, {\rm mcl}}(X)\rightarrow \mathcal F L^p_{\lambda,{\rm loc}}(x_0)\cap\mathcal F L^p_{\Lambda, {\rm mcl}}(x_0\times X)\,.
\end{equation}
\item[(ii)] If in addition $a(x,D)$ is properly supported, then it extends to a bounded linear operator
\begin{equation}\label{mcl_cont2}
\mathcal F L^p_{\lambda\gamma,{\rm loc}}(x_0)\cap\mathcal F L^p_{\Lambda\gamma, {\rm mcl}}(x_0\times X)\rightarrow \mathcal F L^p_{\lambda,{\rm loc}}(x_0)\cap\mathcal F L^p_{\Lambda, {\rm mcl}}(x_0\times X)\,.
\end{equation}
\end{itemize}
\end{theorem}
\begin{proof}
The statement $(ii)$ follows at once from $(i)$ in view of the definition of a properly supported operator. Thus, let us focus on the proof of $(i)$.
In view of Definition \ref{def_mcl_symb}, there exist $\varepsilon>0$ and $\phi\in C_0^\infty(\Omega)$, with $\phi(x_0)\neq 0$, such that conditions in \eqref{symb} are satisfied.
We are going first to prove that
\begin{equation}\label{reg_1}
\phi(x)a(x,D)u\in\mathcal F L^p_{\lambda}(\mathbb R^n)\,,
\end{equation}
as long as $u\in\mathcal F L^p_{\lambda\gamma}(\mathbb R^n)$. Let us denote for short
\begin{equation*}
a_{\phi}(x,\xi):=\phi(x)a(x,\xi)\,.
\end{equation*}
In order to check \eqref{reg_1} it is enough to apply the result of Proposition \ref{continuity1} to the symbol $a_{\phi}(x,\xi)\in\mathcal F L^p_{\lambda}S_\gamma$ (cf. Definition \ref{symbols}) where, restoring the notations used there, we set
\begin{equation*}
\omega_1(\zeta)=\lambda(\zeta)\gamma(\zeta)\,,\quad\omega(\zeta)=\omega_2(\zeta)=\lambda(\zeta)\,.
\end{equation*}
Under the previous positions, the condition \eqref{B_q1} of Proposition \ref{continuity1} reduces to require that $\lambda=\lambda(\zeta)$ satisfies \eqref{B_q}. From Proposition \ref{continuity1} we also deduce the continuity of $a(x,D)$ as a linear map from $\mathcal F L^p_{\lambda\gamma}(\mathbb R^n)$ into $\mathcal F L^p_{\lambda,{\rm loc}}(x_0)$.
It remains to show that
\begin{equation}\label{reg_2}
a_\phi(x,D)u\in\mathcal F L^p_{\Lambda,{\rm mcl}}(X)\,,
\end{equation}
when $u\in\mathcal F L^p_{\lambda\gamma}(\mathbb R^n)\cap\mathcal F L^p_{\Lambda\gamma,{\rm mcl}}(X)$, as well as the continuity of $a(x,D)$ as an operator acting on the aforementioned spaces. Throughout the rest of the proof, we will denote by $C$ some positive constant that is independent of the symbol $a(x,\xi)$ and the function $u(x)$ and may possibly differ from an occurrence to another.
In view of Lemma \ref{mcl_lemma1}, there exists some $0<\varepsilon^\prime<\varepsilon$ such that
\begin{equation*}
\left(\mathbb R^n\setminus X_{[\varepsilon\Lambda]}\right)_{[\varepsilon^\prime\Lambda]}\subset\mathbb R^n\setminus X_{[\varepsilon^\prime\Lambda]}\,.
\end{equation*}
Let us denote for short
\begin{equation}\label{char}
\chi(\zeta):=\chi_{[\varepsilon^\prime\Lambda]}(\zeta)\,,\quad \chi_1(\zeta):=\chi_{[\varepsilon\Lambda]}(\zeta)\,,\quad \chi_2(\zeta):=1-\chi_{[\varepsilon\Lambda]}(\zeta)
\end{equation}
and write
\begin{equation*}
\widehat{a_\phi}(\zeta,\xi)\widehat u(\xi)=\sum\limits_{i,j=1,2}\chi_i(\zeta)\widehat{a_\phi}(\zeta,\xi)\chi_j(\xi)\widehat{u}(\xi)\,.
\end{equation*}
Then in view of \eqref{representation1} and condition ($\mathcal{SA}$) for $\Lambda$, we find
\begin{equation}\label{stima_cont_2_1}
\begin{split}
&\vert\chi(\eta)\Lambda(\eta)\widehat{a_\phi(\cdot,D)u}(\eta)\vert\\
& \le C(2\pi)^{-n}\int\chi(\eta)\left\{\Lambda(\eta-\xi)+\Lambda(\xi)\right\}\vert\widehat{a_\phi}(\eta-\xi,\xi)\vert\,\vert\widehat{u}(\xi)\vert d\xi\\
& \le C(2\pi)^{-n}\int\chi(\eta)\sum\limits_{i,j=1,2}\chi_i(\eta-\xi)\Lambda(\eta-\xi)\vert\widehat{a_\phi}(\eta-\xi,\xi)\vert\chi_j(\xi)\vert\widehat{u}(\xi)\vert d\xi\\
&+C(2\pi)^{-n}\int\chi(\eta)\sum\limits_{i,j=1,2}\chi_i(\eta-\xi)\vert\widehat{a_\phi}(\eta-\xi,\xi)\vert\chi_j(\xi)\Lambda(\xi)\vert\widehat{u}(\xi)\vert d\xi\\
&=\mathcal I_1u(\eta)+\mathcal I_2u(\eta)\,.
\end{split}
\end{equation}
Let us set
\begin{equation}\label{positions1}
\begin{split}
& g_1(\zeta,\xi)=\chi_1(\zeta)\Lambda(\zeta)\gamma(\xi)^{-1}\vert\widehat{a_\phi}(\zeta,\xi)\vert\,;\\
& g_2(\zeta,\xi)=\chi_2(\zeta)\sigma(\zeta)\Lambda(\zeta)\gamma(\xi)^{-1}\Lambda(\xi)^{-1}\vert\widehat{a_\phi}(\zeta,\xi)\vert\,;\\
& \tilde{g}_2(\zeta,\xi)=\chi_2(\zeta)\sigma(\zeta)^{1/2}\Lambda(\zeta)^{1/2}\gamma(\xi)^{-1}\vert\widehat{a_\phi}(\zeta,\xi)\vert\,;\\
& v_1(\xi)=\chi_1(\xi)\gamma(\xi)\sigma(\xi)\vert\widehat{u}(\xi)\vert\,;\quad \tilde v_1(\xi)=\chi_1(\xi)\gamma(\xi)\Lambda(\xi)\vert\widehat{u}(\xi)\vert\,;\\
& v_2(\xi)=\chi_2(\xi)\gamma(\xi)\sigma(\xi)\vert\widehat{u}(\xi)\vert\,;\\
& \tilde v_2(\zeta,\xi)=\chi_2(\xi)\sigma(\xi)^{1/2}\Lambda(\zeta)^{1/2}\gamma(\xi)\vert\widehat{u}(\xi)\vert\,.
\end{split}
\end{equation}
Then the first integral in the right-hand side of \eqref{stima_cont_2_1} can be rewritten as
\begin{equation}\label{decomp1}
\begin{split}
&\mathcal I_1u(\eta)=\int\chi(\eta)\frac1{\sigma(\xi)}g_1(\eta-\xi,\xi)v_1(\xi)d\xi\\
&+\!\!\int\chi(\eta)\frac1{\sigma(\xi)}g_1(\eta-\xi,\xi)v_2(\xi)d\xi+\int\chi(\eta)\frac1{\sigma(\eta-\xi)}g_2(\eta-\xi,\xi)\tilde v_1(\xi)d\xi\\
&+\!\!\int\chi(\eta)\frac1{\sqrt{\sigma(\xi)\sigma(\eta-\xi)}}\tilde g_2(\eta-\xi,\xi)\tilde v_2(\eta-\xi,\xi)d\xi\,.
\end{split}
\end{equation}
In view of the assumptions in \eqref{cndt:2} it is easy to see that all the above functions $v_1$, $v_2$, $\tilde v_1$ defined in \eqref{positions1} belong to $L^p(\mathbb R^n)$, if $u\in\mathcal F L^p_{\lambda\gamma}(\mathbb R^n)\cap\mathcal F L^p_{\Lambda\gamma,{\rm mcl}}(X)$, with the following estimates
\begin{equation}\label{stima_cont_2_2}
\Vert v_1\Vert_{L^p}\le\vert u\vert_{X_{[\varepsilon\Lambda\gamma]}}\,,\quad \Vert \tilde v_1\Vert_{L^p}\le\vert u\vert_{X_{[\varepsilon\Lambda\gamma]}}\,,\quad\Vert v_2\Vert_{L^p}\le\Vert u\Vert_{\mathcal F L^p_{\lambda\gamma}}\,.
\end{equation}
Moreover the functions
\begin{equation}\label{Fi}
F_1(\eta,\xi):=\frac{\chi(\eta)}{\sigma(\xi)}\,,\,\,\,F_2(\eta,\xi):=\frac{\chi(\eta)}{\sigma(\eta-\xi)}\,,\,\,\, F_3(\eta,\xi):=\frac{\chi(\eta)}{\sqrt{\sigma(\xi)\sigma(\eta-\xi)}}
\end{equation}
belong to the space $\mathcal L_2^{\infty, q}(\mathbb R^{2n})$, with the estimates
\begin{equation}\label{est_Fi}
\Vert F_i\Vert_{\mathcal L_2^{\infty,q}}\le\Vert 1/\sigma\Vert_{L^q}\,,\quad i=1,2,3\,.
\end{equation}
Again from \eqref{cndt:2} and \eqref{symb} we easily obtain that $g_1(\zeta,\xi)\in\mathcal L^{p,\infty}_1(\mathbb R^{2n})$ and satisfies the estimate
\begin{equation*}
\Vert g_1\Vert_{\mathcal L^{p,\infty}_1}\le\vert a\vert_{\phi,\Lambda,\gamma,\varepsilon,X}\,.
\end{equation*}
In view of the previous analysis, the first two integral operators involved in \eqref{decomp1} have the form of the operator considered in Lemma \ref{lemma_cont2}. Thus from Lemma \ref{lemma_cont2} and the estimates collected above we get
\begin{equation}\label{cont_1}
\begin{split}
\left\Vert\int\chi(\cdot)\right.&\left.\!\!\frac1{\sigma(\xi)}g_1(\cdot-\xi,\xi)v_1(\xi)d\xi\right\Vert_{L^p}\!\!\!+\left\Vert\int\chi(\cdot)\frac1{\sigma(\xi)}g_1(\cdot-\xi,\xi)v_2(\xi)d\xi\right\Vert_{L^p}\\
&\le \Vert 1/\sigma\Vert_{L^q}\vert a\vert_{\phi,\Lambda,\gamma,\varepsilon,X}\left\{\vert u\vert_{X_{[\varepsilon\Lambda\gamma]}}+\Vert u\Vert_{\mathcal F L^p_{\lambda\gamma}}\right\}\,.
\end{split}
\end{equation}
Concerning the third integral operator in \eqref{decomp1}, we notice that the involved function $g_2(\zeta,\xi)$ vanishes when $\zeta+\xi\notin X_{[\varepsilon^\prime\Lambda]}$, due to the presence of the characteristic function $\chi$. For $\zeta+\xi\in X_{[\varepsilon^\prime\Lambda]}$ and $\zeta\in\mathbb R^n\setminus X_{[\varepsilon\Lambda]}$ it follows that $\Lambda(\zeta)\le\frac1{\varepsilon^\prime}\Lambda(\xi)$; indeed the converse inequality $\Lambda((\zeta+\xi)-\zeta)=\Lambda(\xi)<\varepsilon^\prime\Lambda(\zeta)$ should mean that $\zeta+\xi\in (\mathbb R^n\setminus X_{[\varepsilon\Lambda]})_{[\varepsilon^\prime\Lambda]}\subset\mathbb R^n\setminus X_{[\varepsilon^\prime\Lambda]}$. Hence we get
\begin{equation}\label{est_g2}
\vert g_2(\zeta,\xi)\vert\le\frac1{\varepsilon^\prime}\chi_2(\zeta)\sigma(\zeta)\gamma(\xi)^{-1}\vert\widehat{a_\phi}(\zeta,\xi)\vert\,,
\end{equation}
and, using also $\sigma\preceq\lambda$,
\begin{equation*}
\Vert g_2(\cdot,\xi)\Vert_{L^p}\le \frac1{\varepsilon^\prime}\Vert\chi_2(\cdot)\sigma(\cdot)\gamma(\xi)^{-1}\vert\widehat{a_\phi}(\cdot,\xi)\Vert_{L^p}\le \frac{C}{\varepsilon^\prime}\Vert a\Vert_{\phi,\lambda,\gamma}\,.
\end{equation*}
This yields that $g_2(\zeta,\xi)\in\mathcal L_1^{p,\infty}(\mathbb R^{2n})$ with norm bounded by
\begin{equation*}
\Vert g_2\Vert_{\mathcal L_1^{p,\infty}}\le\frac{C}{\varepsilon^\prime}\Vert a\Vert_{\phi,\lambda,\gamma}\,.
\end{equation*}
Hence we may apply again Lemma \ref{lemma_cont2} to the third operator in \eqref{decomp1}, and using also the estimates \eqref{est_Fi}, \eqref{stima_cont_2_2} we find
\begin{equation}\label{cont_2}
\left\Vert\int\chi(\cdot)\frac1{\sigma(\cdot-\xi)}g_2(\cdot-\xi,\xi)\tilde v_1(\xi)d\xi\right\Vert_{L^p}\le \frac{C}{\varepsilon^\prime}\Vert 1/\sigma\Vert_{L^q}\Vert a\Vert_{\phi,\lambda,\gamma}\vert u\vert_{X_{[\varepsilon\Lambda\gamma]}}\,.
\end{equation}
Let us consider now the fourth integral operator in \eqref{decomp1}. Applying the same argument used to provide the estimate \eqref{est_g2}, we obtain
\begin{equation}\label{est_tildev2}
\vert \tilde v_2(\zeta,\xi)\vert\le\frac1{\varepsilon^\prime}\chi_2(\xi)\sigma(\xi)^{1/2}\Lambda(\xi)^{1/2}\gamma(\xi)\vert\widehat{u}(\xi)\vert\,.
\end{equation}
Thanks to \eqref{cndt:2}, $\sigma^{1/2}\Lambda^{1/2}\preceq\lambda$, then
\begin{equation}\label{est_int_oprt}
\begin{split}
\left\vert\int \right.&\left.\frac{\chi(\eta)}{\sqrt{\sigma(\xi)\sigma(\eta-\xi)}}\tilde g_2(\eta-\xi,\xi)\tilde v_2(\eta-\xi,\xi)d\xi\right\vert\\
&\le \frac{C}{\varepsilon^\prime}\int\frac{\chi(\eta)}{\sqrt{\sigma(\xi)\sigma(\eta-\xi)}}\vert\tilde g_2(\eta-\xi,\xi)\vert\lambda(\xi)\gamma(\xi)\vert\widehat{u}(\xi)\vert\,d\xi\,.
\end{split}
\end{equation}
On the other hand, using again $\sigma^{1/2}\Lambda^{1/2}\preceq\lambda$ and $a_\phi(\cdot,\xi)/\gamma(\xi)\in\mathcal F L^p_{\lambda}(\mathbb R^n)$, uniformly with respect to $\xi$, we establish that $\tilde{g}_2(\zeta,\xi)$ belongs to $\mathcal L_1^{p,\infty}(\mathbb R^{2n})$ and satisfies the estimate
\begin{equation}\label{est_tildeg2}
\Vert\tilde g_2\Vert_{\mathcal L_1^{p,\infty}}\le\Vert\sigma(\cdot)^{1/2}\Lambda(\cdot)^{1/2}\gamma(\xi)^{-1}\widehat{a_\phi}(\cdot,\xi)\Vert_{L^p}\le C\Vert a\Vert_{\phi,\lambda,\gamma}\,.
\end{equation}
Since $\lambda(\xi)\gamma(\xi)\vert\widehat{u}(\xi)\vert\in L^p(\mathbb R^n)$ (as $u\in\mathcal FL^p_{\lambda\gamma}(\mathbb R^n)$) and $F_3(\eta,\xi)=\frac{\chi(\eta)}{\sqrt{\sigma(\xi)\sigma(\eta-\xi)}}$ belongs to $\mathcal L_2^{\infty,q}(\mathbb R^{2n})$, the integral operator in the right-hand side of \eqref{est_int_oprt} satisfies the assumptions of Lemma \ref{lemma_cont2}, then from \eqref{est_Fi} and \eqref{est_g2} we find
\begin{equation}\label{cont_3}
\begin{split}
\left\Vert\int \right.&\left.\frac{\chi(\cdot)}{\sqrt{\sigma(\xi)\sigma(\cdot-\xi)}}\tilde g_2(\cdot-\xi,\xi)\tilde v_2(\cdot-\xi,\xi)d\xi\right\Vert_{L^p}\\
&\le \frac{C}{\varepsilon^\prime}\Vert 1/\sigma\Vert_{L^q}\Vert a\Vert_{\phi,\lambda,\gamma}\Vert u\Vert_{\mathcal FL^p_{\lambda\gamma}}\,.
\end{split}
\end{equation}
Summing up the estimates \eqref{cont_1}, \eqref{cont_2}, \eqref{cont_3} the $L^p-$norm of $\mathcal I_1u$ in the right-hand side of \eqref{stima_cont_2_1} is estimated by
\begin{equation}\label{cont_4}
\Vert\mathcal I_1u\Vert_{L^p}\le \frac{C}{\varepsilon^\prime}\Vert 1/\sigma\Vert_{L^q}\left(\vert a\vert_{\phi,\Lambda,\gamma,\varepsilon,X}+\Vert a\Vert_{\phi,\lambda,\gamma}\right)\left(\vert u\vert_{X_{[\varepsilon\Lambda\gamma]}}+\Vert u\Vert_{\mathcal F L^p_{\lambda\gamma}}\right)\,.
\end{equation}
The second integral $\mathcal I_2u(\eta)$ in \eqref{stima_cont_2_1} can be handled similarly as before to provide for its $L^p-$norm the same bound as in \eqref{cont_4}. From \eqref{stima_cont_2_1} we then get
\begin{equation}\label{cont_5}
\begin{split}
\Vert\chi&\Lambda\widehat{a_\phi(\cdot,D)u}\Vert_{L^p}\\
&\le \frac{C}{\varepsilon^\prime}\Vert 1/\sigma\Vert_{L^q}\left(\vert a\vert_{\phi,\Lambda,\gamma,\varepsilon,X}+\Vert a\Vert_{\phi,\lambda,\gamma}\right)\left(\vert u\vert_{X_{[\varepsilon\Lambda\gamma]}}+\Vert u\Vert_{\mathcal F L^p_{(\lambda\gamma)}}\right)
\end{split}
\end{equation}
which proves \eqref{reg_2} and shows the continuity of $a(x,D)$ as a linear map from $\mathcal F L^p_{\lambda\gamma}(\mathbb R^n)\cap\mathcal F L^p_{\Lambda\gamma, {\rm mcl}}(X)$ into $\mathcal F L^p_{\lambda,{\rm loc}}(x_0)\cap\mathcal F L^p_{\Lambda, {\rm mcl}}(x_0\times X)$.
\end{proof}
\begin{remark}\label{rmk:14}
{\rm
Let the same hypotheses of Theorem \ref{continuity2} be satisfied.
Clearly every $v=v(x)\in\mathcal FL^p_{\lambda,{\rm loc}}(x_0)\cap\mathcal FL^p_{\Lambda,{\rm mcl}}(x_0\times X)$ is a $\xi-$independent symbol in the class $\mathcal F L^p_{\lambda,\,\Lambda}S_\gamma(x_0\times X)$ corresponding to the weight function $\gamma(\xi)\equiv 1$, and the product of smooth functions by the multiplier $v$ defines a properly supported ``zeroth order'' operator. Therefore we find that the product of any two elements $u, v\in \mathcal FL^p_{\lambda,{\rm loc}}(x_0)\cap\mathcal FL^p_{\Lambda,{\rm mcl}}(x_0\times X)$ still belongs to the same space (giving a continuous bilinear mapping), as a direct application of Theorem \ref{continuity2}. Similarly as in the proof of Corollary \ref{composition_prop}, see also the subsequent Remark \ref{rmk:3.1}, one can deduce that the composition of a vector-valued distribution $u=(u_1,\dots,u_N)\in\left(\mathcal FL^p_{\lambda,{\rm loc}}(x_0)\cap\mathcal FL^p_{\Lambda,{\rm mcl}}(x_0\times X)\right)^N$ with some nonlinear function $F=F(x,\zeta)$ of $x\in\mathbb R^n$ and $\zeta\in\mathbb C^N$, which is locally smooth with respect to $x$ on some neighborhood of $x_0$ and entire analytic with respect to $\zeta$ in the sense of Remark \ref{rmk:3.1}, is again a distribution in $\mathcal FL^p_{\lambda,{\rm loc}}(x_0)\cap\mathcal FL^p_{\Lambda,{\rm mcl}}(x_0\times X)$.
Let us even point out that in the particular case where $\lambda\equiv\Lambda$ the assumption \eqref{cndt:2} in Theorem \ref{continuity2} reduces to $\sigma\preceq\lambda$. In such a case $\mathcal FL^p_{\lambda,{\rm loc}}(x_0)\cap\mathcal FL^p_{\Lambda,{\rm mcl}}(x_0\times X)\equiv \mathcal FL^p_{\lambda,{\rm loc}}(x_0)$ and $\mathcal FL^p_{\lambda,\Lambda}S_\gamma(x_0\times X)\equiv\mathcal FL^p_{\lambda}S_\gamma(V_{x_0})$ for a suitable neighborhood $V_{x_0}$ of $x_0$, see Definition \ref{symbols}, hence the statement of Theorem \ref{continuity2} reduces to a particular case of the statement of Proposition \ref{continuity1} (where $\omega_1=\gamma\lambda$, $\omega=\omega_2=\lambda$) under slightly more restrictive assumptions; indeed a sub-additive weight function $\lambda$ satisfying $\sigma\preceq\lambda$ for $1/\sigma\in L^q(\mathbb R^n)$ also fulfils condition \eqref{B_q} with the same $q$ (that is the assumption required by Proposition \ref{continuity1}), in view of Proposition \ref{wf_relationship}.{\it ii}.
}
\end{remark}
\section{Propagation of singularities}\label{appl_sct}
In this Section, we give some applications to the local and microlocal regularity of semilinear partial(pseudo)differential equations in weighted Fourier Lebesgue spaces.
The smooth symbols we consider in this Section are related to a suitable subclass of the weight functions introduced in Section \ref{wf_sct}. More precisely, we consider a continuous function $\lambda:\mathbb R^n\rightarrow]0,+\infty[$ satisfying the following:
\begin{equation}\label{p_g_b}
\lambda(\xi)\ge \frac1{C}(1+\vert\xi\vert)^\nu\,,\quad\forall\,\xi\in\mathbb R^n\,;
\end{equation}
\begin{equation}\label{s_v}
\frac1{C}\le\frac{\lambda(\xi)}{\lambda(\eta)}\le C\,,\quad\mbox{as long as}\,\,\vert\xi-\eta\vert\le \frac1{C}\lambda(\eta)^{1/\mu}\,,
\end{equation}
for suitable constants $C\ge 1$, $0<\nu\le\mu$.
Thanks to Proposition \ref{wf_relationship}, it is clear that $\lambda(\xi)$ is a weight function; indeed it also satisfies the temperance condition ($\mathcal T$) for $N=\mu$.
All the weight functions described in the examples 1--3 given in Section \ref{wf_sct} obey the assumptions \eqref{p_g_b}, \eqref{s_v}.
\smallskip
For $r\in\mathbb R$, $\rho\in]0,1/\mu]$, we define $S^r_{\rho,\lambda}$ as the class of smooth functions $a(x,\xi)\in C^\infty(\mathbb R^{2n})$ whose derivatives decay according to the following estimates
\begin{equation}\label{smooth_symb}
\vert\partial^\alpha_\xi\partial^\beta_x a(x,\xi)\vert\le C_{\alpha,\beta}\lambda(\xi)^{r-\rho\vert\alpha\vert}\,,\qquad\forall\,(x,\xi)\in\mathbb R^{2n}\,.
\end{equation}
If $\Omega$ is an open subset of $\mathbb R^n$, the local class $S^r_{\rho,\lambda}(\Omega)$ is the set of functions $a(x,\xi)\in C^\infty(\Omega\times\mathbb R^n)$ such that $\phi(x)a(x,\xi)\in S^r_{\rho,\lambda}$ for all $\phi\in C^\infty_0(\Omega)$. We will adopt the shortcut
\begin{equation*}
S^r_{\lambda}:=S^r_{1/\mu,\lambda}\,,\qquad S^r_{\lambda}(\Omega):=S^r_{1/\mu,\lambda}(\Omega)\,.
\end{equation*}
A symbol $a(x,\xi)\in S^r_{\lambda}(\Omega)$ (and the related pseudodifferential operator) is said to be $\lambda-${\it elliptic} if for every compact subset $K$ of $\Omega$ some positive constants $c_K$ and $R_K>1$ exist such that
\begin{equation}\label{ell}
\vert a(x,\xi)\vert\ge c_K\lambda(\xi)^r\,,\qquad\forall\,x\in K\,\,\,\mbox{and}\,\,\,\vert\xi\vert\ge R_K\,.
\end{equation}
Let us also observe that
$\bigcap\limits_{r\in\mathbb R}S^r_{\rho,\lambda}(\Omega)=S^{-\infty}(\Omega)$, where in the classic terms $S^{-\infty}(\Omega)$ is the class of symbols $a(x,\xi)\in C^\infty(\Omega\times\mathbb R^n)$ such that for arbitrarily large $\theta>0$, for all multi-indices $\alpha,\beta\in\mathbb Z^n_+$ and every compact set $K\subset\Omega$ there holds
\begin{equation*}
\vert\partial^\alpha_\xi\partial^\beta_x a(x,\xi)\vert\le C_{\alpha,\beta,\theta}(1+\vert\xi\vert)^{-\theta}\,,\quad\forall\,x\in K\,,\,\,\,\forall\,\xi\in\mathbb R^n\,.
\end{equation*}
Pseudodifferential operators with symbols $a(x,\xi)\in S^{-\infty}(\Omega)$ are {\it regularizing} operators in the sense that they define linear bounded operators $a(x,D):\mathcal E^\prime(\Omega)\rightarrow C^\infty(\Omega)$.
The weighted symbol classes $S^r_{\rho,\lambda}(\Omega)$ considered above are a special case of the more general classes $S_{m,\Lambda}(\Omega)$, associated to the weight function $m(\xi)=\lambda(\xi)^r$ and the weight vector $\Lambda(\xi)=(\lambda(\xi)^\rho,\dots,\lambda(\xi)^\rho)$, as defined and studied in \cite[Definition 1.1]{GM3}. For the weighted symbol classes $S^r_{\rho,\lambda}(\Omega)$, a complete symbolic calculus is available, cf. \cite[Sect.1]{GM3}; in particular, the existence of a {\it parametrix} of any elliptic pseudodifferential operator is guaranteed.
\begin{proposition}\label{parametrix}
Let $a(x.\xi)$ be a $\lambda-$elliptic symbol in $S^r_{\rho,\lambda}(\Omega)$. Then a symbol $b(x,\xi)\in S^{-r}_{\rho, \lambda}(\Omega)$ exists such that the operator $b(x,D)$ is properly supported and satisfies
\begin{equation*}
b(x,D)a(x,D)=I+c(x,D)\,,
\end{equation*}
where $I$ denotes the identity operator and $c(x,D)$ is a regularizing pseudodifferential operator.
\end{proposition}
The following inclusion
\begin{equation}\label{symb_inclusion}
S^r_{\rho,\lambda}(\Omega)\subset\mathcal FL^p_{\omega}S_{\lambda^r}(\Omega)
\end{equation}
holds true, with continuous imbedding, for all $r\in\mathbb R$, $\rho\in]0,1/\mu]$, $p\in[1,+\infty]$ and any weight function $\omega(\xi)$. As a consequence of Proposition \ref{continuity1} we then obtain the following continuity result.
\begin{proposition}\label{continuity3}
Let $\omega(\xi)$ be any weight function and $p\in[1,+\infty]$. Then every pseudodifferential operator with symbol $a(x,\xi)\in S^r_{\rho,\lambda}(\Omega)$ extends to a linear bounded operator
\begin{equation*}
a(x,D):\mathcal FL^p_{\lambda^r\omega}(\mathbb R^n)\rightarrow\mathcal FL^p_{\omega,{\rm loc}}(\Omega)\,.
\end{equation*}
If in addition $a(x,D)$ is properly supported, then the latter extends to a linear bounded operator
\begin{equation*}
a(x,D):\mathcal FL^p_{\lambda^r\omega,{\rm loc}}(\Omega)\rightarrow\mathcal FL^p_{\omega,{\rm loc}}(\Omega)\,.
\end{equation*}
\end{proposition}
\begin{proof}
In view of \eqref{symb_inclusion}, it is enough to observe that for any weight function $\omega(\xi)$, another weight function $\tilde{\omega}(\xi)$ can be found in such a way that
\begin{equation}\label{B_q2}
\sup\limits_{\xi\in\mathbb R^n}\left\Vert\frac{\omega(\xi)}{\omega(\cdot)\tilde{\omega}(\xi-\cdot)}\right\Vert_{L^q}<+\infty\,,
\end{equation}
where $q\in[1,+\infty]$ is the conjugate exponent of $p$; for instance, one can take $\tilde\omega(\xi)=(1+\vert\xi\vert)^{\tilde N}$, with $\tilde N>0$ sufficiently large. Then the result follows at once, by noticing that $a(x,\xi)$ belongs to $\mathcal FL^p_{\tilde\omega}S_{\lambda^r}(\Omega)$ and \eqref{B_q2} is nothing but condition \eqref{B_q1}, where $\gamma$, $\omega_1$, $\omega_2$ and $\omega$ in Proposition \ref{continuity1} are replaced respectively by $\lambda^r$, $\lambda^r\omega$, $\omega$ and $\tilde\omega$.
\end{proof}
\subsection{Local regularity results}\label{loc_reg_sct}
Let $\lambda=\lambda(\xi)$ be a given continuous weigh function satisfying the assumptions \eqref{p_g_b} and \eqref{s_v}. We consider a nonlinear pseudodifferential equation of the following type
\begin{equation}\label{semilin_eqt}
a(x,D)u+ F(x,b_i(x,D)u)_{1\le i\le M}=f(x)\,,
\end{equation}
where $u=u(x)$ is defined on some open set $\Omega\subseteq\mathbb R^n$ and $a(x,D)$ is a properly supported pseudodifferential operator with symbol $a(x,\xi)\in S^r_{\lambda}(\Omega)$ for given $r>0$. $F(x,b^i(x,D)u)_{1\le i\le M}$ stands for a nonlinear function of $x\in\Omega$ and $b
_1(x,D)u$, $b_2(x,D)u$,..., $b_M(x,D)u$ where $b_i(x,D)$ are still properly supported pseudodifferential operators, and $f=f(x)$ is a given forcing term. We require the equation \eqref{semilin_eqt} to be {\it semilinear} by assuming that the operators involved in the nonlinear part $F(x,b^i(x,D)u)$ have order strictly smaller than the order of the linear part $a(x,D)u$, that is
\begin{equation}\label{lower_order}
b^i(x,\xi)\in S^{r-\varepsilon}_{\lambda}(\Omega)\qquad\mbox{for}\,\, i=1,\dots,M\,,
\end{equation}
for suitable $0<\varepsilon<r$.
For $s\in\mathbb R$, $p\in[1,+\infty]$, let us set
\begin{equation*}
\mathcal FL^p_{s,\lambda}(\mathbb R^n):=\mathcal FL^p_{\lambda^s}(\mathbb R^n)\,,\quad\mathcal FL^p_{s,\lambda,{\rm loc}}(\Omega):=\mathcal FL^p_{\lambda^s,{\rm loc}}(\Omega)\,.
\end{equation*}
The following regularity result can be proved.
\begin{proposition}\label{reg_prop1}
Let the symbol $a(x,\xi)\in S^r_{\lambda}(\Omega)$ be $\lambda-$elliptic and the function $F=F(x,\zeta)$ obey the assumptions collected in Remark \ref{rmk:3.1}. For a given $p\in[1,+\infty]$, take a real number $t$ such that $\lambda^{t-r+\varepsilon}$ fulfils condition \eqref{B_q} with $q$ the conjugate exponent of $p$. If $u\in\mathcal FL^p_{t,\lambda,{\rm loc}}(\Omega)$ is any solution of the equation \eqref{semilin_eqt}, with forcing term $f\in\mathcal F L^p_{s-r,\lambda,{\rm loc}}(\Omega)$ for some $s>t$, then $u\in \mathcal FL^p_{s,\lambda,{\rm loc}}(\Omega)$.
If in particular $u\in\mathcal FL^p_{t,\lambda,{\rm loc}}(\Omega)$ solves the equation \eqref{semilin_eqt} with $f=0$ (that is the equation \eqref{semilin_eqt} is homogeneous) then $u\in C^\infty(\Omega)$.
\end{proposition}
\begin{proof}
Because of Proposition \ref{continuity3} and the assumption \eqref{lower_order}, $b_i(x,D)u\in\mathcal FL^p_{t-r+\varepsilon, \lambda,{\rm loc}}(\Omega)$ for all $i=1,\dots,M$. Since $\lambda^{t-r+\varepsilon}$ satisfies \eqref{B_q}, Corollary \ref{composition_prop} also implies $F(x,b_i(x,D)u)\in\mathcal FL^p_{t-r+\varepsilon,\lambda,{\rm loc}}(\Omega)$ (cf. Remark \ref{rmk:3.1}).
If $t+\varepsilon\ge s$ then $a(x,D)u=-F(x,b_i(x,D)u)+f\in \mathcal F L^p_{s-r,\lambda,{\rm loc}}(\Omega)$ hence $u\in \mathcal F L^p_{s,\lambda,{\rm loc}}(\Omega)$ because of the $\lambda-$ellipticity of $a(x,D)$.
If on the contrary $t+\varepsilon<s$, applying again the $\lambda-$ellipticity of $a(x,D)$, from $a(x,D)u=-F(x,b_i(x,D)u)+f\in \mathcal F L^p_{t-r+\varepsilon,\lambda,{\rm loc}}(\Omega)$ we derive $u\in \mathcal F L^p_{t+\varepsilon,\lambda,{\rm loc}}(\Omega)$. In the latter case, we may repeat the same arguments above, where now $t$ is replaced by $t+\varepsilon$\,\footnote{Let us notice in particular that if the weight function $\lambda^{t-r+\varepsilon}$ satisfies condition \eqref{B_q}, then the same is true for any power of $\lambda$ with exponent greater than $t-r+\varepsilon$, in view of Proposition \ref{algebra_prop4}.}. After that we get $F(x,b_i(x,D)u)\in\mathcal FL^p_{t-r+2\varepsilon,\lambda,{\rm loc}}(\Omega)$ and, provided that $t+2\varepsilon<s$, $u\in \mathcal F L^p_{t+2\varepsilon,\lambda,{\rm loc}}(\Omega)$. It is now clear that the second part of the argument above can be iterated $N$ times, up to get $F(x,b_i(x,D)u)\in\mathcal FL^p_{t-r+N\varepsilon,\lambda,{\rm loc}}(\Omega)$ with $t+N\varepsilon\ge s$; hence $a(x,D)u=-F(x,b_i(x,D)u)+f\in \mathcal F L^p_{s-r,\lambda,{\rm loc}}(\Omega)$ implies $u\in \mathcal F L^p_{s,\lambda,{\rm loc}}(\Omega)$ from the $\lambda-$ellipticity of $a(x,D)$.
The second part of the theorem, concerning the case $f=0$, follows at once from the first one; in this case the argument above can be applied for arbitrarily large $s$, thus $u\in\bigcap\limits_{s\ge t}\mathcal FL^p_{s,\lambda,{\rm loc}}(\Omega)\subset C^\infty(\Omega)$.
\end{proof}
\begin{remark}\label{rmk:100}
{\rm Let us suppose that the weight function $\lambda=\lambda(\xi)$ fulfils condition ($\mathcal{SA}$) (respectively condition ($\mathcal G$)), besides \eqref{p_g_b} and \eqref{s_v}. Then $\lambda^{t-r+\varepsilon}$ satisfies condition \eqref{B_q} if $t>r+\frac{n}{\nu q}-\varepsilon$ (respectively $t>r+\frac{n}{(1-\delta)\nu q}-\varepsilon$) is assumed.}
\end{remark}
\subsection{Microlocal regularity results}\label{mcl_reg_sct}
The results presented in this section apply to a class of weight functions which is smaller than the one considered in Section \ref{loc_reg_sct}. More precisely here we deal with a continuous function $\lambda:\mathbb R^n\rightarrow]0,+\infty[$ which satisfies ($\mathcal{SA}$), ($\mathcal{SH}$) and obeys the following
\begin{itemize}
\item[($\mathcal{PG}$)]{\bf polynomial growth conditions}: for suitable constants $C\ge 1$, $0<\nu\le\mu$.
\begin{equation}\label{p_g}
\frac1{C}(1+\vert\xi\vert)^{\nu}\le\lambda(\xi)\le C(1+\vert\xi\vert)^{\mu}\,,\quad\forall\,\xi\in\mathbb R^n\,.
\end{equation}
\end{itemize}
\begin{remark}\label{rmk:15}
{\rm It is known from the previous section that such a function $\lambda$ also satisfies condition \eqref{strong_sv}. Then it can be shown that \eqref{strong_sv}, together with \eqref{p_g}, also implies that $\lambda$ obeys the slowly varying condition \eqref{s_v}\footnote{More precisely, from Section \ref{loc_reg_sct} we know that conditions \eqref{s_v} and the second inequality in \eqref{p_g} are equivalent under the assumptions \eqref{strong_sv} and \eqref{p_g_b}}. Thus the class of weight functions considered in this Section is a proper subclass of that considered in Section \ref{loc_reg_sct}. It is worthy to be noticed that weight functions described in the examples 1, 2, given in Section \ref{wf_sct}, are included in the class of weight functions that we are considering here, whereas the multi-quasi-elliptic weight function illustrated in the example 3 does not meet all the assumptions required here, precisely the sub-additivity ($\mathcal{SA}$) is not satisfied unless the complete polyhedron $\mathcal P$ gives rise to a quasi-homogeneous weight function of type \eqref{quasi_ell_wf}. Additional examples of weight functions obeying conditions ($\mathcal{SA}$), ($\mathcal{SH}$) and ($\mathcal{PG}$) are provided by the following
\begin{equation*}
\lambda_{r,s}(\xi)=\langle\xi\rangle^s\left[\log(2+\langle\xi\rangle)\right]^r\,,\quad\mbox{for}\,\, r,s\in ]0,+\infty[\,,
\end{equation*}
which were studied by Triebel \cite{TRI_3} (see also \cite{GM2}), or even by such functions as
\begin{equation*}
\begin{split}
\langle\xi\rangle^2_{\mu,\nu}&=1+\sum\limits_{j=1}^n\vert\xi_j\vert^{\mu_j}\left[\log(2+\vert\xi_j\vert)\right]^{\nu_j}\,,\\
&\quad\mbox{for}\,\,\mu=(\mu_1,\dots,\mu_n), \nu=(\nu_1,\dots,\nu_n)\in ]0,+\infty[^n\,,
\end{split}
\end{equation*}
or
\begin{equation*}
\Lambda_{s,\mathcal P}(\xi)=\langle\xi\rangle^s+\log(\lambda_{\mathcal P}(\xi))\,,\quad\mbox{for}\,\, s\in]0,+\infty[\,,
\end{equation*}
being $\lambda_{\mathcal P}(\xi)$ the multi-quasi-elliptic weight associated to a complete polyhedron $\mathcal P$, as it was introduced in Example 3 of Section \ref{wf_sct} (see \eqref{mqe_wf}).
}
\end{remark}
\smallskip
In order to take advantage of the slowly varying condition \eqref{s_v} (which allows in particular the symbolic calculus for smooth classes $S^r_{\rho, \lambda}(\Omega)$, see Section \ref{appl_sct}), it is convenient to introduce here another family of neighborhoods of an arbitrary set $X$ (in the frequency space $\mathbb R^n_{\xi}$), associated to the weight function $\lambda$, besides the $[\lambda]-$neighborhoods $X_{[\varepsilon\lambda]}$ already defined as in \eqref{omega_neighb}. For arbitrary $X\subset\mathbb R^n$ and $\varepsilon>0$ we set
\begin{equation}\label{omega_neighb_eucl}
X_{\varepsilon\lambda}:=\bigcup\limits_{\xi_0\in X}\left\{\xi\in\mathbb R^n\,:\,\,\vert\xi-\xi_0\vert<\varepsilon\lambda(\xi_0)^{1/\mu}\right\}\,,
\end{equation}
where $\mu>0$ is the same exponent involved in \eqref{p_g} (hence in \eqref{s_v} according to Remark \ref{rmk:15}); we will refer to the set $X_{\varepsilon\lambda}$ as the $\lambda-${\it neighborhood} of $X$ of size $\varepsilon$.
In the following for an open set $\Omega\subset\mathbb R^n$ and $x_0\in\Omega$, we also set for short $X_{\varepsilon\lambda}(x_0):=B_\varepsilon(x_0)\times X_{\varepsilon\lambda}$, where $B_\varepsilon(x_0)$ denotes the open ball in $\Omega$ centered at $x_0$ with radius $\varepsilon$.
\medskip
Compared to the case of $[\lambda]-$neighborhoods of a set $X$, to define the corresponding $\lambda-$neighborhoods the weight function $\lambda$ is replaced by the Euclidean norm, as the measure of the distance from points in $X_{\varepsilon\lambda}$ to points in $X$. This reflects into a slightly different behaviour of $\lambda-$neighborhoods: it is clear (just from the definition) that for $\varepsilon>0$ arbitrarily small the set $X_{\varepsilon\lambda}$ is never empty (unless $X=\emptyset$), cf. Remark \ref{rmk:intorni}; it is also clear that $X_{\varepsilon\lambda}$ is open, for it is the union of a family of open balls in $\mathbb R^n$ (centered at points of $X$).
The same set inclusions as given in Lemma \ref{mcl_lemma1} remain true also when the $[\lambda]-$neighborhoods of a set are replaced by the $\lambda-$neighborhoods, see \cite{RO1}, \cite{GM2.1} for the proof.
\begin{lemma}\label{mcl_lemma3}
Given $\varepsilon>0$, there exists $0<\varepsilon^\prime<\varepsilon$ such that for every $X\subset\mathbb R^n$
\begin{itemize}
\item[(1)] $\left(X_{\varepsilon^\prime\lambda}\right)_{\varepsilon^\prime\lambda}\subset X_{\varepsilon\lambda}$;
\item[(2)] $\left(\mathbb R^n\setminus X_{\varepsilon\lambda}\right)_{\varepsilon^\prime\lambda}\subset\mathbb R^n\setminus X_{\varepsilon^\prime\lambda}$.
\end{itemize}
\end{lemma}
A significant relation between $[\lambda]-$ and $\lambda-$neighborhoods is established by the next two results.
\begin{lemma}\label{lemma_inclusione_intorni}
Let $c>0$ be arbitrarily fixed. For every $\varepsilon>0$ there exists $0<\varepsilon^\prime<\varepsilon$ such that the set inclusion
\begin{equation}\label{inclusione_intorni}
\left(X\cap\left\{\lambda(\xi)>c/\varepsilon^\prime\right\}\right)_{\varepsilon^\prime\lambda}\subset X_{[\varepsilon\lambda]}\cap\left\{\lambda(\xi)>c/\varepsilon\right\}
\end{equation}
holds true for every $X\subset\mathbb R^n$.
\end{lemma}
\begin{proof}
Let $0<\varepsilon^\prime<\min\{1,\varepsilon\}$ be such that $X\cap\left\{\lambda(\xi)>c/\varepsilon^\prime\right\}$ be nonempty and take an arbitrary $\xi\in\left(X\cap\left\{\lambda(\xi)>c/\varepsilon^\prime\right\}\right)_{\varepsilon^\prime\lambda}$\footnote{If $X$ is unbounded then $X\cap\left\{\lambda(\xi)>c/\varepsilon^\prime\right\}\neq\emptyset$ for $\varepsilon^\prime>0$ arbitrarily small, because of the left inequality of \eqref{p_g}.}; then there exists some $\xi_0\in X$ such that
\begin{equation}\label{eq:1}
\vert\xi-\xi_0\vert<\varepsilon^\prime\lambda(\xi_0)^{1/\mu}\qquad\mbox{and}\qquad\lambda(\xi_0)>c/\varepsilon^\prime\,.
\end{equation}
From \eqref{p_g} and \eqref{eq:1} we get
\begin{equation}\label{eq:1.1}
\begin{split}
\lambda(\xi-\xi_0)&\le C(1+\vert\xi-\xi_0\vert)^\mu\le C2^{\mu-1}(1+\vert\xi-\xi_0\vert^\mu)\\
&<C2^{\mu-1}(1+\varepsilon^{\prime\,\mu}\lambda(\xi_0))<C2^{\mu-1}(\varepsilon^\prime/c\lambda(\xi_0)+\varepsilon^{\prime\,\mu}\lambda(\xi_0))\\
&<C2^{\mu-1}\varepsilon^\prime(1/c+1)\lambda(\xi_0)\,,
\end{split}
\end{equation}
hence $\lambda(\xi-\xi_0)<\varepsilon\lambda(\xi_0)$ provided that $\varepsilon^\prime$ is such that
\begin{equation*}
C2^{\mu-1}\varepsilon^\prime(1/c+1)<\varepsilon\,.
\end{equation*}
Thus $\xi\in X_{[\varepsilon\lambda]}$ provided that $0<\varepsilon^\prime<\min\left\{1,\frac{\varepsilon}{C2^{\mu-1}(1/c+1)}\right\}$.
Let us now prove that $\lambda(\xi)>c/\varepsilon$ up to a further shrinking of $\varepsilon^\prime$. We use again conditions ($\mathcal{SA}$), ($\mathcal{SH}$), ($\mathcal{PG}$) and \eqref{eq:1.1} to find
\begin{equation*}
\begin{split}
\lambda(\xi)&\ge 1/C\lambda(\xi_0)-\lambda(\xi-\xi_0)\ge 1/C\lambda(\xi_0)-C(1+\vert\xi-\xi_0\vert)^\mu\\
&\ge 1/C\lambda(\xi_0)-C2^{\mu-1}(1+\vert\xi-\xi_0\vert^\mu)>1/C\lambda(\xi_0)-C2^{\mu-1}(1+\varepsilon^{\prime\,\mu}\lambda(\xi_0))\\
&=\left(1/C-C2^{\mu-1}\varepsilon^{\prime\,\mu}\right)\lambda(\xi_0)-C2^{\mu-1}\,,
\end{split}
\end{equation*}
from which we deduce, using also \eqref{eq:1},
\begin{equation*}
\lambda(\xi)>\frac1{2C}\lambda(\xi_0)-C2^{\mu-1}>\frac{c}{2C\varepsilon^\prime}-C2^{\mu-1}>\frac{c}{4C\varepsilon^\prime}>\frac{c}{\varepsilon}\,,
\end{equation*}
provided that $\varepsilon^\prime>0$ is chosen such that
\begin{equation*}
\varepsilon^\prime<\min\left\{\frac{1}{2C^{2/\mu}},\frac{c}{2^{\mu+1}C^2},\frac{\varepsilon}{4C}\right\}\,.
\end{equation*}
This ends the proof that $\xi\in X_{[\varepsilon\lambda]}\cap\left\{\lambda(\xi)>c/\varepsilon\right\}$.
\end{proof}
\begin{remark}\label{rmk:intorni1}
{\rm If $X$ is bounded, the set $X\cap\left\{\lambda(\xi)>c/\varepsilon^\prime\right\}$ (hence the neighborhood $\left(X\cap\left\{\lambda(\xi)>c/\varepsilon^\prime\right\}\right)_{\varepsilon^\prime\lambda}$) is empty for $\varepsilon^\prime>0$ sufficiently small, thus the inclusion \eqref{inclusione_intorni} becomes trivial. However, thanks to \eqref{inclusione_intorni}, this never occurs when $X$ is unbounded; in such a case the set $X\cap\left\{\lambda(\xi)>c/\varepsilon^\prime\right\}$ is nonempty for arbitrarily small $\varepsilon^\prime>0$, since $\lambda$ is unbounded on $X$ as a consequence of the left inequality in \eqref{p_g}. This yields in particular that, for an unbounded set $X$ the $[\lambda]-$neighborhood $X_{[\varepsilon\lambda]}$ is nonempty with size $\varepsilon>0$ arbitrarily small, cf. Remark \ref{rmk:intorni}.}
\end{remark}
\begin{corollary}\label{cor_intorni}
For every $\varepsilon>0$ there exists $0<\varepsilon^\prime<\varepsilon$ such that for all $X\subset\mathbb R^n$
\begin{equation}\label{inclusione_intorni_2}
\left(X_{[\varepsilon^\prime\lambda]}\right)_{\varepsilon^\prime\lambda}\subset X_{[\varepsilon\lambda]}\,.
\end{equation}
\end{corollary}
\begin{proof}
In view of Lemma \ref{mcl_lemma1} we first notice that for arbitrary $\varepsilon>0$ we may find $0<\varepsilon^\ast<\varepsilon$ sufficiently small such that
\begin{equation*}
\left(X_{[\varepsilon^\ast\lambda]}\right)_{[\varepsilon^\ast\lambda]}\subset X_{[\varepsilon\lambda]}\,.
\end{equation*}
Then combining the results of Lemma \ref{mcl_lemma1} and Lemma \ref{lemma_inclusione_intorni}, with $X_{[\varepsilon^\ast\lambda]}$ instead of $X$, another $0<\varepsilon^\prime<\varepsilon^\ast$ sufficiently small can be chosen such that
\begin{equation*}
\begin{split}
\left(X_{[\varepsilon^\prime\lambda]}\right)&_{\varepsilon^\prime\lambda}\equiv\left(X_{[\varepsilon^\prime\lambda]}\cap\{\lambda(\xi)>\hat c/\varepsilon^\prime\}\right)_{\varepsilon^\prime\lambda}\\
&\subset\left(X_{[\varepsilon^\ast\lambda]}\cap\{\lambda(\xi)>\hat c/\varepsilon^\prime\}\right)_{\varepsilon^\prime\lambda}\subset \left(X_{[\varepsilon^\ast\lambda]}\right)_{[\varepsilon^\ast\lambda]}\subset X_{[\varepsilon\lambda]}\,,
\end{split}
\end{equation*}
where $\hat c>0$ is given in Lemma \ref{mcl_lemma1}. The proof is complete.
\end{proof}
In order to perform the subsequent analysis, the next technical lemma will be useful; for its proof, the reader is addressed to \cite[Lemma 1.10]{RO1}, see also \cite[Lemma 1]{GM2.1}.
\begin{proposition}\label{mcl_prop_1}
For arbitrary $\varepsilon>0$ and $X\subset\mathbb R^n$ there exists a symbol $\sigma=\sigma(\xi)\in S^0_{\lambda}$ such that ${\rm supp}\,\sigma\subset X_{\varepsilon\lambda}$ and $\sigma(\xi)=1$ if $\xi\in X_{\varepsilon^\prime\lambda}$, for a suitable $\varepsilon^\prime>0$, with $0<\varepsilon^\prime<\varepsilon$, depending only on $\varepsilon$ and $\lambda$. Moreover for every $x_0\in\Omega$, where $\Omega\subset\mathbb R^n$ is an open set, there exists a symbol $\tau_0(x,\xi)\in S^0_{\lambda}(\Omega)$ such that ${\rm supp}\,\tau_0\subset X_{\varepsilon\lambda}(x_0)$ and $\tau_0(x,\xi)=1$, for $(x,\xi)\in X_{\varepsilon^\ast\lambda}(x_0)$, with a suitable $\varepsilon^\ast$ satisfying $0<\varepsilon^\ast<\varepsilon$.
\end{proposition}
\begin{remark}\label{rmk:17}
{\rm As an application of Corollary \ref{cor_intorni}, one can easily see that a statement similar to Proposition \ref{mcl_prop_1} also holds when $\lambda-$neighborhoods are replaced with the corresponding $[\lambda]-$neighborhoods; indeed for arbitrary $X\subset\mathbb R^n$ and $\varepsilon>0$, take $0<\tilde\varepsilon<\varepsilon$ such that $\left(X_{[\tilde\varepsilon\lambda]}\right)_{\tilde\varepsilon\lambda}\subset X_{[\varepsilon\lambda]}$ and apply the result of Proposition \ref{mcl_prop_1}, where $X$ is replaced by $X_{[\tilde\varepsilon\lambda]}$. Then some numbers $0<\varepsilon^{\prime\prime}<\varepsilon^\prime<\tilde\varepsilon$ and a symbol $\sigma=\sigma(\xi)\in S^0_{\lambda}$ exist such that ${\rm supp}\,\sigma\subset \left(X_{[\tilde\varepsilon\lambda]}\right)_{\varepsilon^\prime\lambda}\subset \left(X_{[\tilde\varepsilon\lambda]}\right)_{\tilde\varepsilon\lambda}\subset X_{[\varepsilon\lambda]}$ and $\sigma\equiv 1$ on $\left(X_{[\tilde\varepsilon\lambda]}\right)_{\varepsilon^{\prime\prime}\lambda}$ (hence on $X_{[\tilde\varepsilon\lambda]}$). As for the construction of a counterpart of the variable coefficients symbol $\tau_0(x,\xi)\in S^0_{\lambda}(\Omega)$ in the second part of the statement above, it comes from the use of the symbol $\sigma(\xi)$ by following the same lines as in Proposition \ref{mcl_prop_1}, see \cite[Lemma 1]{GM2.1}.}
\end{remark}
\begin{definition}\label{mcl_ell}
Let us consider a symbol $a(x,\xi)\in S^r_{\rho,\lambda}(\Omega)$, $x_0\in\Omega$ and $X\subset\mathbb R^n$. We say that $a(x,\xi)$ (or the corresponding pseudodifferential operator) is microlocally $[\lambda]-$elliptic in $X$ at point $x_0$, writing $a(x,\xi)\in{\rm mce}_{r,[\lambda]}X(x_0)$, if there exist constants $c_0>0$ and $\varepsilon>0$ sufficiently small such that
\begin{equation}\label{mcl_ell_ineq}
\vert a(x_0,\xi)\vert\ge c_0\lambda(\xi)^r\,,\qquad\mbox{for}\,\,\xi\in X_{[\varepsilon\lambda]}\,.
\end{equation}
\end{definition}
\begin{remark}\label{rmk:16}
{\rm Let us remark that in the above definition we do not explicitly require that frequencies $\xi$, for which \eqref{mcl_ell_ineq} holds true, are larger than some positive constant (that is usual when defining an ellipticity condition, cf. \eqref{ell}); indeed, because of Lemma \ref{mcl_lemma1}, $\xi\in X_{[\varepsilon\lambda]}$ yields $\lambda(\xi)>\hat c/\varepsilon$ and, for sufficiently small $\varepsilon>0$, the latter turns out to be a largeness condition on $\xi$, in view of the polynomial growth condition ($\mathcal{PG}$).}
\end{remark}
Let us recall the following notion, providing a microlocal counterpart of the notion of regularizing symbol
\begin{definition}
We say that a symbol $a(x,\xi)\in S^r_{\rho,\lambda}(\Omega)$ is rapidly decreasing in $\Theta\subset \Omega\times{\bf R}^n$ if there exists $a_0(x,\xi)\in S^r_{\rho,\lambda}(\Omega)$ such that $a(x,\xi)-a_0(x,\xi)\in S^{-\infty}(\Omega)$ and $a_0(x,\xi)=0$ in $\Theta$.
\end{definition}
The following notion is a natural substitute of that of {\it characteristic set} of a symbol, in the absence of any homogeneity property.
\begin{definition}
We define the characteristic filter of a symbol $a(x,\xi)\in S^r_{\rho,\lambda}(\Omega)$ at a point $x_0\in\Omega$ to be the set
\begin{equation}\label{char_filter}
\Sigma_{[\lambda], x_0}a:=\left\{ X\subset\mathbb R^n\,:\,\,\, a(x,\xi)\in{\rm mce}_{r,[\lambda]}(\mathbb R^n\setminus X)(x_0)\right\}\,.
\end{equation}
\end{definition}
Using Lemma \ref{mcl_lemma1}, it is easy to check that $\Sigma_{[\lambda], x_0}a$ is a $[\lambda]-$filter.
\smallskip
The reader is addressed to \cite{RO1} and \cite{GM3} where analogous notions as above are stated in a more general setting.
\smallskip
Arguing on the properties of $\lambda-$neighborhoods of a set and the slowly varying condition \eqref{s_v} as in the proof of \cite[Lemma 4.3]{GM3}, one can prove that $a(x,\xi)\in S^s_{\rho,\lambda}(\Omega)$ is microlocally $[\lambda]-$elliptic in $X$ at point $x_0$ if and only if
\begin{equation}\label{mcl_ell_ineq_BIS}
\vert a(x,\xi)\vert\ge c^\ast\lambda(\xi)^r\,,\qquad\mbox{for}\,\,(x,\xi)\in B_{\tilde\varepsilon}(x_0)\times \left(X_{[\tilde\varepsilon\lambda]}\right)_{\tilde\varepsilon\lambda}\,,
\end{equation}
for suitable constants $c^\ast>0$ and sufficiently small $\tilde\varepsilon>0$.
Then following the same lines of the proof of \cite[Theorem 4.6]{GM3} one can prove the following
\begin{proposition}\label{mcl_parametrix}
For every symbol $a(x,\xi)\!\in\! S^r_{\rho,\lambda}(\Omega)$ microlocally $[\lambda]\!-$elliptic in $\{x_0
\}\times X$, there exists a symbol $b(x,\xi)\in S^{-r}_{\rho,\lambda}(\Omega)$ such that the associated operator $b(x,D)$ is properly supported and
\begin{equation}\label{MPTEO1a}
b(x,D)a(x,D)={\rm Id}+c(x,D),
\end{equation}
where $c(x,\xi)\in S^0_{\rho,\lambda}(\Omega)$ is rapidly decreasing in $B_{\tilde\varepsilon}(x_0)
\times \left(X_{[\tilde\varepsilon\lambda]}\right)_{\tilde\varepsilon\lambda}$ for a suitable small $\tilde\varepsilon>0$.
\end{proposition}
\medskip
For $s\in\mathbb R$, $p\in[1,+\infty]$, $U$ open neighborhood of $x_0\in\mathbb R^n$ and $X\subset\mathbb R^n$ given, let $\mathcal FL^p_{s,\lambda,{\rm loc}}(U)$ and $\mathcal FL^p_{s,\lambda,{\rm mcl}}(x_0\times X)$ denote the local and microlocal Fourier Lebesgue classes corresponding to the weight function $\lambda^s$, according to Definitions \ref{loc_w_FL_spaces}, \ref{micro_FL_space}. Agreeing with these notations, we denote by $\Xi_{[\lambda],\mathcal FL^p_{s,\lambda},x_0}$ the related $[\lambda]-$filter of Fourier Lebesgue singularities.
\medskip
By resorting to Proposition \ref{mcl_prop_1} and arguing similarly as in the proof of \cite[Proposition 4.5]{GM2.2} and \cite[Proposition 4.10]{GM3}, we are able to prove the following characterization of microlocal Fourier Lebesgue spaces.
\begin{proposition}\label{char_FL}
Let $x_0\in\Omega$, $\Omega\subset\mathbb R^n$ open set, and $X\subset\mathbb R^n$ be given. A distribution $u\in\mathcal D^\prime(\mathbb R^n)$ belongs to $\mathcal FL^p_{s,\lambda,{\rm mcl}}(x_0\times X)$ if and only if one of the following two equivalent conditions is satisfied:
\begin{itemize}
\item[i.] there exist constants $0<\varepsilon^\prime<\varepsilon$ sufficiently small and $\phi\in C^\infty_0(\Omega)$, with $\phi(x_0)\neq 0$, such that
\begin{equation}\label{char_eq1}
\sigma(D)(\phi u)\in \mathcal FL^p_{s,\lambda}(\mathbb R^n)\,,
\end{equation}
where $\sigma=\sigma(\xi)\in S^0_{\lambda}$ is some symbol satisfying ${\rm supp}\,\sigma\subset X_{[\varepsilon\lambda]}$ and $\sigma\equiv 1$ on $X_{[\varepsilon^\prime\lambda]}$;
\item[ii.] There exist an operator $\tau(x,D)\in \widetilde{\rm Op}\,S^0_{\lambda}(\Omega)$ microlocally $[\lambda]-$elliptic in $\{x_0\}\times X$ such that
\begin{equation}\label{char_eq2}
\tau(x,D)u\in\mathcal FL^p_{s,\lambda,{\rm loc}}(\Omega)\,.
\end{equation}
\end{itemize}
\end{proposition}
Following similar arguments to those in \cite[Propositions 5.1, 5.2]{GM2.2} we give the following results.
\begin{proposition}\label{cont1}
Let $s\in\mathbb R$, $r>0$, $x_0\in\Omega$, $X\subset\mathbb{R}^n$, $a(x,D)\in\widetilde{{\rm Op}}S^r_{\rho,\lambda}(\Omega)$ be given. Then for $p\in[1,\infty]$ and $u\in mcl\mathcal FL^p_{s,\lambda}(x_0\times X)$ one has $a(x,D)u\in mcl\mathcal FL^{p}_{s-r,\lambda}(x_0\times X)$.
\end{proposition}
\begin{proposition}\label{cont2}
For $s\in\mathbb R$, $r>0$, $x_0\in\Omega$, $X\subset\mathbb{R}^n$, let $a(x,D)\in\widetilde{\rm Op}S^r_{\rho,\Lambda}(\Omega)$ be microlocally $[\lambda]-$elliptic in $\{x_0\}\times X$. Then for every $p\in[1,\infty]$ and $u\in\mathcal{D}'(\mathbb R^n)$ such that $a(x,D)u\in mcl\mathcal FL^{p}_{s-r,\lambda}(x_0\times X)$ one has $u\in mcl\mathcal FL^{p}_{s}(x_0\times X)$.
\end{proposition}
It is also straightforward to show that the results of Propositions \ref{cont1}, \ref{cont2} can be restated in terms of the filters of Fourier Lebesgue singularities and characteristic filter of a symbol as follows.
\begin{proposition}\label{filter_inclusions}
Let $s\in\mathbb R$, $r>0$ be arbitrary real numbers, $a(x,D)\in\widetilde{\rm Op}\,S_{\rho, \lambda}(\Omega)$, $x_0\in\Omega$ and $p\in[1,\infty]$. Then the following inclusions are satisfied for every $u\in\mathcal{D}'(\mathbb R^n)$:
\begin{equation*}
\Xi_{[\lambda],\mathcal FL^p_{s-r,\lambda},x_0}a(x,D)u\cap\Sigma_{[\lambda],x_0}a\subset \Xi_{[\lambda],\mathcal FL^p_{s,\lambda},x_0}u\subset\Xi_{[\lambda],\mathcal FL^p_{s-r,\lambda},x_0}a(x,D)u\,.
\end{equation*}
\end{proposition}
\subsection{Semilinear equations}\label{semi-linear_sct}
Gathering the results collected in the preceding Sections \ref{w_FL_pdo_sct} and \ref{mcl_reg_sct}, we prove here a result of microlocal regularity in Fourier Lebesgue spaces for solutions to semilinear partial differential equations of type \eqref{semilin_eqt} already considered in Section \ref{loc_reg_sct}. Throughout this Section, we assume that $a(x,D)$, $b_i(x,D)$ for $1\le i\le M$ in \eqref{semilin_eqt} are properly supported operators with symbols $S^r_\lambda(V_{x_0})$ and $S^{r-\varepsilon}_\lambda(V_{x_0})$ on some open bounded neighborhood $V_{x_0}$ of a point $x_0$, where as in Section \ref{w_FL_pdo_sct} $0<\varepsilon<r$ are given, and the nonlinear function $F=F(x,\zeta)$ depending as a $C^\infty-$function on its first argument of $x\in V_{x_0}$ and entire analytic on its second argument $\zeta=(\zeta_i)_{1\le i\le M}\in\mathbb C^M$, satisfying the same requirement made in Section \ref{algebra_sct} (see also Remark \ref{rmk:14}).
The following result was originally proved in \cite{G}.
\begin{theorem}\label{semilin_thm}
For $0<\varepsilon<r$ and $1\le p\le +\infty$ given as above, let $\tau$, $\tilde t$, $s$ be positive real numbers such that
\begin{equation}\label{condizioni_esponenti}
\tau+r-\varepsilon\le\tilde t<s
\end{equation}
and $\lambda^{-\tau}\in L^q(\mathbb R^n)$, where $q$ is the conjugate exponent of $p$. As for the semilinear equation \eqref{semilin_eqt}, let us assume that the pseudodifferential operator $a(x,D)$ is $[\lambda]-$microlocally elliptic in $X\subset\mathbb R^n$ at the point $x_0$ and the source term $f=f(x)$ belongs to $\mathcal FL^p_{s-r,\lambda,{\rm mcl}}(x_0\times X)$. Then every solution $u\in\mathcal FL^p_{\tilde t,\lambda,{\rm loc}}(x_0)$ to the equation \eqref{semilin_eqt} with source term $f$ also satisfies
\begin{equation}\label{reg_finale}
u\in\mathcal FL^p_{t,\lambda,{\rm mcl}}(x_0\times X)\,,\quad\mbox{for all}\,\, t\le\min\left\{s,\tilde t+\left(E\left(\frac{\tilde t-r-\tau}{\varepsilon}\right)+2\right)\varepsilon\right\}\,,
\end{equation}
where $E(\theta)$ is the greatest integer less than or equal to $\theta\in\mathbb R$.
\end{theorem}
\begin{proof}
The proof relies on a bootstrapping argument similar to the one used to prove Proposition \ref{reg_prop1}. So let $u\in\mathcal FL^p_{\tilde t,\lambda,{\rm loc}}(x_0)$ be a solution to equation \eqref{semilin_eqt}. From Propositions \ref{continuity3} we get
\begin{equation}\label{bi}
b_i(x,D)u\in \mathcal FL^p_{\tilde t-r+\varepsilon,\lambda,{\rm loc}}(x_0)\,,\quad i=1,\dots,M\,.
\end{equation}
In view of the assumptions \eqref{condizioni_esponenti}, $\lambda^{-\tau}\in L^q(\mathbb R^n)$ and the sub-additivity of $\lambda$ we may apply the result of Theorem \ref{continuity2} and its consequences stated in Remark \ref{rmk:14}; with reference to the statement of that theorem, here $\lambda^{\tau}$ plays the role of the weight functions $\sigma$ whereas $\lambda^{\tilde t-r+\varepsilon}$ plays the role of both the weight functions $\lambda$, $\Lambda\,\,\,$\footnote{Notice in particular that from the sub-additivity of $\lambda$ and $\lambda^{\tau}\preceq\lambda^{\tilde t-r+\varepsilon}$ (following from \eqref{condizioni_esponenti}), we derive that $\lambda^{\tilde t-r+\varepsilon}$ satisfies condition \eqref{B_q} with the conjugate exponent of $p$ (that is required to apply Theorem \ref{continuity2}.}. Thus it follows from \eqref{bi} that
\begin{equation}\label{Fbi}
F(x,b_i(x,D)u)_{1\le i\le M}\in \mathcal FL^p_{\tilde t-r+\varepsilon,\lambda,{\rm loc}}(x_0)\,.
\end{equation}
If $s\le\tilde t+\varepsilon$, we derive from \eqref{semilin_eqt} that $a(x,D)u\in\mathcal FL^p_{s-r,\lambda,{\rm mcl}}(x_0\times X)$ hence $u\in\mathcal FL^p_{s,\lambda,{\rm mcl}}(x_0\times X)$ in view of the $[\lambda]-$microellipticity of $a(x,D)$ in $X$ at point $x_0$ and Proposition \ref{cont2} (notice that $\tilde t+\varepsilon\le 2\tilde t-r-\tau+2\varepsilon$ under the assumption \eqref{condizioni_esponenti} then $s\le \tilde t+\varepsilon$ implies $s=\min\{s,2\tilde t-r-\tau+2\varepsilon\}$); if on the contrary $s>\tilde t+\varepsilon$ again from \eqref{semilin_eqt} we derive that $a(x,D)u\in\mathcal FL^p_{\tilde t-r+\varepsilon,\lambda,{\rm mcl}}(x_0\times X)$ hence $u\in\mathcal FL^p_{\tilde t+\varepsilon,\lambda,{\rm mcl}}(x_0\times X)$ by the same arguments as before. In this latter case, using once again Propositions \ref{continuity3} and \ref{cont1} we get
\begin{equation*}
b_i(x,D)u\in \mathcal FL^p_{\tilde t-r+\varepsilon,\lambda,{\rm loc}}(x_0)\cap \mathcal FL^p_{\tilde t-r+2\varepsilon,\lambda,{\rm mcl}}(x_0\times X)\,,\quad 1\le i\le M\,.
\end{equation*}
Now we would like to apply Theorem \ref{continuity2} where the role of the weight functions $\sigma$, $\lambda$ and $\Lambda$ is covered respectively by $\lambda^\tau$, $\lambda^{\tilde t-r+\varepsilon}$ and $\lambda^{\tilde t-r+2\varepsilon}$; the only assumption to check is $\Lambda\preceq\lambda^2/\sigma$ which amounts to have that $\tilde t\ge \tau+r$ (cf. \eqref{cndt:2}). If this is the case, Theorem \ref{continuity2} applies to find
\begin{equation*}
F(x,b_i(x,D)u)\in \mathcal FL^p_{\tilde t-r+\varepsilon,\lambda,{\rm loc}}(x_0)\cap \mathcal FL^p_{\tilde t-r+2\varepsilon,\lambda,{\rm mcl}}(x_0\times X)
\end{equation*}
hence from \eqref{semilin_eqt} and Proposition \ref{cont2}
\begin{equation*}
a(x,D)u\in\mathcal FL^p_{s-r,\lambda,{\rm mcl}}(x_0\times X)\,\,\Rightarrow\,\,u\in\mathcal FL^p_{s,\lambda,{\rm mcl}}(x_0\times X) \,,\quad\mbox{if}\,\,s\le\tilde t+2\varepsilon
\end{equation*}
or
\begin{equation*}
a(x,D)u\in\mathcal FL^p_{\tilde t-r+2\varepsilon,\lambda,{\rm mcl}}(x_0\times X)\,\,\Rightarrow\,\,u\in\mathcal FL^p_{\tilde t+2\varepsilon,\lambda,{\rm mcl}}(x_0\times X) \,,\quad\mbox{otherwise}.
\end{equation*}
In the latter case
\begin{equation*}
b_i(x,D)u\in \mathcal FL^p_{\tilde t-r+\varepsilon,\lambda,{\rm loc}}(x_0)\cap \mathcal FL^p_{\tilde t-r+3\varepsilon,\lambda,{\rm mcl}}(x_0\times X)\,,\quad 1\le i\le M
\end{equation*}
and, provided that $\tilde t\ge\tau+r+\varepsilon$, we are still in the position to apply Theorem \ref{continuity2} where $\lambda^{\tau}$ and $\lambda^{\tilde t-r+\varepsilon}$ play again the role of $\sigma$ and $\lambda$, while $\lambda^{\tilde t-r+3\varepsilon}$ plays the role of $\Lambda$. For $\tilde t\ge\tau+r+\varepsilon$ Theorem \ref{continuity2} yields
\begin{equation*}
F(x,b_i(x,D)u)\in \mathcal FL^p_{\tilde t-r+\varepsilon,\lambda,{\rm loc}}(x_0)\cap \mathcal FL^p_{\tilde t-r+3\varepsilon,\lambda,{\rm mcl}}(x_0\times X)
\end{equation*}
and again
\begin{equation*}
a(x,D)u\in\mathcal FL^p_{s-r,\lambda,{\rm mcl}}(x_0\times X)\,\,\Rightarrow\,\,u\in\mathcal FL^p_{s,\lambda,{\rm mcl}}(x_0\times X) \,,\quad\mbox{if}\,\,s\le\tilde t+3\varepsilon
\end{equation*}
or
\begin{equation*}
a(x,D)u\in\mathcal FL^p_{\tilde t-r+3\varepsilon,\lambda,{\rm mcl}}(x_0\times X)\,\,\Rightarrow\,\,u\in\mathcal FL^p_{\tilde t+3\varepsilon,\lambda,{\rm mcl}}(x_0\times X) \,,\quad\mbox{otherwise}.
\end{equation*}
By an iteration of the above procedure we find that if the integer $j\ge 0$ is such that
\begin{equation}\label{condizione_j}
\tilde t<\tau+r+j\varepsilon
\end{equation}
then
\begin{equation*}
u\in\mathcal FL^p_{\min\{\tilde t+(j+1)\varepsilon,s\},\lambda,{\rm mcl}}(x_0\times X)\,.
\end{equation*}
The smallest nonnegative integer $j$ satisfying \eqref{condizione_j} is $\tilde j=E\left(\frac{\tilde t-r-\tau}{\varepsilon}\right)+1$ (from \eqref{condizioni_esponenti} $E\left(\frac{\tilde t-\tau-r}{\varepsilon}\right)\ge -1$ follows), hence $\tilde t+(\tilde j+1)\varepsilon=\tilde t+\left(E\left(\frac{\tilde t-r-\tau}{\varepsilon}\right)+2\right)\varepsilon$. This gives
\begin{equation*}
u\in\mathcal FL^p_{\min\left\{\tilde t+\left(E\left(\frac{\tilde t-r-\tau}{\varepsilon}\right)+2\right)\varepsilon,s\right\},\lambda,{\rm mcl}}(x_0\times X)\,,
\end{equation*}
which completes the proof.
\end{proof}
In terms of the filter of Fourier Lebesgue singularities and characteristic filter of a symbol, the result of Theorem \ref{semilin_thm} can be restated as follows: for every solution $u\in\mathcal FL^p_{\tilde t,\lambda,{\rm loc}}(x_0)$ to equation \eqref{semilin_eqt} one has
\begin{equation*}
\Xi_{[\lambda],\mathcal FL^p_{s-r,\lambda},x_0}f\cap\Sigma_{[\lambda],x_0}a\subset \Xi_{[\lambda],\mathcal FL^p_{t,\lambda},x_0}u\,,
\end{equation*}
for all $t\le\min\left\{s,\tilde t+\left(E\left(\frac{\tilde t-r-\tau}{\varepsilon}\right)+2\right)\varepsilon\right\}$.
\begin{remark}\label{rmk:18}
{\rm Because of the lower estimate of condition ($\mathcal{PG}$), a sufficient condition for $\lambda^{-\tau}\in L^q(\mathbb R^n)$
is $\tau>\frac{n}{\nu q}$.}
\end{remark}
\subsection{The case of quasi-homogeneous equations}\label{q_hom_sct}
In this section, we deal with pseudodifferential operators whose smooth symbols are associated to a quasi-homogeneous weight as defined in the Example 2 of Section \ref{wf_sct}. We recall that for $M=(m_1,\dots,m_n)\in\mathbb N^n$, with $m_\ast:=\min\limits_{1\le j\le n}m_j\ge 1$, the {\it quasi-homogeneous weight} is defined as
\begin{equation}\label{q_hom_w_n}
\langle\xi\rangle_M:=\left(1+\sum\limits_{j=1}^n\xi_j^{2m_j}\right)^{1/2}\,.
\end{equation}
Throughout the rest of this Section, we will make use of the following notations. We set $m^\ast:=\max\limits_{1\le j\le n}m_j$, $\frac 1M:=\left( \frac{1}{m_1},\dots, \frac{1}{m_n}\right)$ and defined the {\it quasi-homogeneous norm} as
\begin{equation}\label{q_hom_norm}
\vert\xi\vert^2_M:=\sum\limits_{j=1}^n\xi_j^{2m_j}\,.
\end{equation}
Clearly the usual Euclidean norm $\vert\xi\vert$ corresponds to the quasi-homogeneous norm in the case of $M=(1,\dots,1)$. For every $\alpha\in\mathbb Z^n_+$, $
\xi\in\mathbb R^n$ and $t>0$ we also set $\langle\alpha,\frac1{M}\rangle:=\sum\limits_{j=1}^n\frac{\alpha_j}{m_j}$ and $t^{1/M}\xi:=(t^{1/m_1}\xi_1,\dots, t^{1/m_n}\xi_n)$. It is worth to notice that, in spite of the terminology, the quasi-homogeneous norm $\vert\cdot\vert_M$ is not a norm; instead of the homogeneity and the triangle inequality, required for norms, the quasi-homogeneous norm enjoys the following properties:
\begin{itemize}
\item[(i)] {\it Quasi-Homogeneity}: for all $t>0$, $\xi\in\mathbb{R}^n$
\begin{equation*}
\vert t^{ 1/M}\xi\vert_M =t\vert\xi\vert_M\,;
\end{equation*}
\item[(ii)] {\it Sub-additivity}: a constant $C\ge 1$ depending only on $M$ exists such that
\begin{equation*}
\vert\xi+\eta\vert_M\leq C(\vert \xi\vert_M+\vert \eta \vert_M)\,,\qquad \forall\,\xi,\eta\in\mathbb R^n\,.
\end{equation*}
\end{itemize}
For $R>0$ and $x_0\in\mathbb R^n$, the {\it $M-$open ball} centered at $x_0$ with radius $R$ is defined to be the set
\begin{equation*}
B_M(x_0;R);=\left\{x\in\mathbb R^n\,:\,\,\vert x-x_0\vert_M<R\right\}\,.
\end{equation*}
The set
\begin{equation}\label{M-unit-ball}
\mathbb S_M:=\left\{x\in\mathbb R^n\,:\,\,\vert x\vert_M=1\right\}
\end{equation}
is the {\it unit $M-$sphere} (centered at the origin). For further details and properties of quasi-homogeneous norm and weight, we address the reader to \cite{GMqh1}, \cite{GMqh2}.
According to the behavior of the weight \eqref{q_hom_w_n} expressed by the estimates \eqref{q_hom_der}, we introduce suitable classes of smooth symbols displaying a decaying behavior of {\it quasi-homogeneous} type.
\begin{definition}\label{def:1}
Given $r\in\mathbb{R}$, $S^r_{M}$ will be the class of functions $a(x,\xi)\in
C^{\infty}(\mathbb{R}^{2n})$ such that for all multi-indices $\alpha,\beta\in\mathbb{Z}^n_+$ there exists $C_{\alpha,\beta}>0$ such that:
\begin{equation}\label{eqh:1}
\vert \partial^{\beta}_x\partial^{\alpha}_{\xi}a(x,\xi)\vert\le
C_{\alpha,\beta}\langle\xi\rangle_M^{r-\langle\alpha,\frac1{M}\rangle},\quad\forall x,\,\xi\in\mathbb{R}^n\,.
\end{equation}
If $\Omega$ is an arbitrary open subset of $\mathbb R^n$, we denote by $S^r_M(\Omega)$ the local class of functions $a(x,\xi)\in C^\infty(\Omega\times\mathbb R^n)$ such that $\phi(x)a(x,\xi)\in S^r_M$ for all $\phi\in C^\infty_0(\Omega)$.
\end{definition}
\medskip
Due to the underlying quasi-homogeneous structure, in the present framework the whole theory of propagation of singularities can be based upon a suitable notion of ``conical'' set in frequency space adapted to this structure.
Let us recall below some basic notions, see \cite{GMqh2} for more details. Later on it is set for short $T^{\circ}\mathbb{R}^n:=\mathbb R^n\times(\mathbb R^n\setminus\{0\})$.
\begin{definition}\label{Mcone}
We say that a set $\Gamma\subset\mathbb{R}^n\setminus\{0\}$ is an $M-$cone (or is $M-$conic), if
\begin{equation*}
\xi\in\Gamma\quad\Rightarrow\quad t^{1/M}\xi\in\Gamma\,\,,\,\,\forall\,t>0\,.
\end{equation*}
\end{definition}
For $\eta\in\mathbb R^n$ and $R>0$ the set
\begin{equation}\label{M-cone}
\Gamma_M(\eta;R):=\left\{t^{1/M}\xi\,:\,\,\xi\in B_M(\eta;R)\,,\,\,t>0\right\}\cap(\mathbb R^n\setminus\{0\})
\end{equation}
is $M-$conic; it is called the {\it $M-$cone generated by $B_M(\eta;R)$}.
\medskip
Since \eqref{q_hom_w_n} also belongs to the class of weight functions considered in Sections \ref{loc_reg_sct}, \ref{mcl_reg_sct}, the results considered there, based upon the notion of $[\lambda]-$filter, could be applied to the quasi-homogeneous setting (that is $\lambda(\xi)=\langle\xi\rangle_M$).
The next results of this Section will provide some evidences that these two alternative approaches are essentially equivalent.
\begin{proposition}\label{equivalence_prop}
There exist constants $\hat c>0$ and $\hat\varepsilon>0$ sufficiently small such that for all $0<\varepsilon\le\hat\varepsilon$ another $0<\varepsilon^\prime<\varepsilon$ exists such that for all $M-$conic set $X\subset\mathbb R^n\setminus\{0\}$
\begin{equation}\label{equiv:1}
X_{[\varepsilon^\prime\langle\cdot\rangle_M]}\subset\bigcup\limits_{\eta\in X\cap\mathbb S_M}\Gamma_M(\eta;\varepsilon)\cap\left\{\langle\xi\rangle_M>\hat c/\varepsilon\right\}
\end{equation}
and, conversely,
\begin{equation}\label{equiv:2}
\bigcup\limits_{\eta\in X\cap\mathbb S_M}\Gamma_M(\eta;\varepsilon^\prime)\cap\left\{\langle\xi\rangle_M>\hat c/\varepsilon^\prime\right\}\subset X_{[\varepsilon\langle\cdot\rangle_M]}\,.
\end{equation}
\end{proposition}
\begin{proof}
For a $M-$conic set $X$ and an arbitrary $\tilde\varepsilon>0$, take $\xi\in X_{[\tilde\varepsilon\langle\cdot\rangle_M]}$ and let $\eta\in X$ such that
\begin{equation}\label{equiv_eq:1}
\langle\xi-\eta\rangle_M<\tilde\varepsilon\langle\eta\rangle_M\,.
\end{equation}
Making use of the trivial inequalities
\begin{equation*}
1/\sqrt 2(1+\vert\zeta\vert_M)\le\langle\zeta\rangle_M\le 1+\vert\zeta\vert_M\,,\qquad\forall\,\zeta\in\mathbb R^n\,,
\end{equation*}
\eqref{equiv_eq:1} implies
\begin{equation*}
1+\vert\xi-\eta\vert_M<\sqrt 2\tilde\varepsilon (1+\vert\eta\vert_M)\,,
\end{equation*}
hence
\begin{equation}\label{equiv_eq:2}
\vert\xi-\eta\vert_M<\sqrt 2\tilde\varepsilon\vert\eta\vert_M
\end{equation}
provided that $0<\tilde\varepsilon<1/\sqrt 2$. Since in particular \eqref{equiv_eq:1} implies $\eta\neq 0$, because of the $M-$homogeneity of the $M-$norm, condition \eqref{equiv_eq:2} can be reformulated as follows
\begin{equation*}
\begin{split}
\vert\xi-\eta\vert_M &<\sqrt 2\tilde\varepsilon\vert\eta\vert_M\quad\Leftrightarrow\quad \vert\vert\eta\vert_M^{1/M}(\zeta-\tilde\eta)\vert_M=\vert\eta\vert_M\vert\zeta-\tilde\eta\vert_M<\sqrt 2\tilde\varepsilon\vert\eta\vert_M\\
&\Leftrightarrow\quad\vert\zeta-\tilde\eta\vert_M<\sqrt 2\tilde\varepsilon\,,
\end{split}
\end{equation*}
where $\zeta:=\vert\eta\vert_M^{-1/M}\xi$ and $\tilde\eta:=\vert\eta\vert_M^{-1/M}\eta\in\mathbb S_M\cap X$ because $X$ is $M-$conic. The last inequality above means that $\zeta$ belongs to the $M-$open ball centered at $\tilde\eta$ with radius $\sqrt 2\tilde\varepsilon$, thus $\xi=\vert\eta\vert_M^{1/M}\zeta\in\Gamma_M(\tilde\eta;\sqrt 2\tilde\varepsilon)$ cf. \eqref{M-cone}. Since in view of Lemma \ref{mcl_lemma1}, $\eta\in X_{[\tilde\varepsilon\langle\cdot\rangle_M]}$ also implies
\begin{equation*}
\langle\eta\rangle_M>\hat c/\tilde\varepsilon\,,
\end{equation*}
for suitable $\hat c>0$ independent of $X$ and $\tilde\varepsilon>0$, the inclusion \eqref{equiv_eq:1} follows taking $\hat{\varepsilon}=\frac{1}{\sqrt 2 C}$ and choosing for each $0<\varepsilon\le\hat\varepsilon$, $0<\varepsilon^\prime\le\varepsilon/\sqrt 2$.
Conversely, let $\xi\in \bigcup\limits_{\eta\in X\cap\mathbb S_M}\Gamma_M(\eta;\tilde\varepsilon)\cap\left\{\langle\xi\rangle_M>\hat c/\tilde\varepsilon\right\}$, where $\tilde\varepsilon>0$ is still arbitrary and is chosen sufficiently small such that
\begin{equation*}
1+\vert\xi\vert_M\ge\langle\xi\rangle_M>\hat c/\tilde\varepsilon\quad\Rightarrow\quad\vert\xi\vert_M>\hat c/\tilde\varepsilon-1\ge\hat c/(2\tilde\varepsilon)\,,
\end{equation*}
that is $0<\tilde\varepsilon\le\hat c/2$. By definition (see \eqref{M-cone}) there exist $t>0$ and $\tilde\eta\in X\cap\mathbb S_M$ such that
\begin{equation*}
\xi=t^{1/M}\zeta\,\,\,\,\,\mbox{for some}\,\,\zeta\in B_M(\tilde\eta;\tilde\varepsilon)\,.
\end{equation*}
Therefore, in view of the $M-$homogeneity, we get
\begin{equation*}
\vert\xi-t^{1/M}\tilde\eta\vert_M=\vert t^{1/M}\zeta-t^{1/M}\tilde\eta\vert_M=t\vert \zeta-\tilde\eta\vert_M<\tilde\varepsilon t=\tilde\varepsilon t\vert\tilde\eta\vert_M=\tilde\varepsilon\vert t^{1/M}\tilde\eta\vert_M\,.
\end{equation*}
Since $X$ is $M-$conic, $\eta:=t^{1/M}\tilde\eta\in X$; hence with such an $\eta\in X$ we have
\begin{equation}\label{equiv_eq:3}
\vert\xi-\eta\vert_M<\tilde\varepsilon\vert\eta\vert_M\,.
\end{equation}
On the other hand, from the sub-additivity (ii), $\vert\xi\vert_M>\hat c/(2\tilde\varepsilon)$ and \eqref{equiv_eq:3} we derive
\begin{equation}\label{equiv_eq:4}
\vert\eta\vert_M\ge\frac1 C\vert\xi\vert_M-\vert\xi-\eta\vert_M\ge \left(\frac1C-\tilde\varepsilon\right)\vert\xi\vert_M\ge\frac1{2C}\vert\xi\vert_M\ge\frac{\hat c}{2C\tilde\varepsilon}\,,
\end{equation}
provided that $0<\tilde\varepsilon\le\frac1{2C}$. Summing up \eqref{equiv_eq:3}, \eqref{equiv_eq:4} then gives
\begin{equation*}
\frac{\hat c}{2C}+\vert\xi-\eta\vert_M<2\tilde\varepsilon\vert\eta\vert_M<2\tilde\varepsilon\langle\eta\rangle_M\,.
\end{equation*}
Combining the latter inequality with
\begin{equation*}
\frac{\hat c}{2C}+\vert\xi-\eta\vert_M\ge (1+\vert\xi-\eta\vert_M)\min\left\{1,\frac{\hat c}{2C}\right\}\ge\frac1{\sqrt 2}\min\left\{1,\frac{\hat c}{2C}\right\}\langle\xi-\eta\rangle_M
\end{equation*}
we finally obtain
\begin{equation*}
\langle\xi-\eta\rangle_M<\widehat C\tilde\varepsilon\langle\eta\rangle_M\,,
\end{equation*}
with a suitable constant $\widehat C>0$ independent of $\tilde\varepsilon$, hence $\xi\in X_{[\widehat C\tilde\varepsilon\langle\cdot\rangle_M]}$. From the previous argument, we conclude that the second inclusion \eqref{equiv:2} holds true.
\end{proof}
Essentially the result above tells that a $[\langle\cdot\rangle_M]-$neighborhood of a $M$-conic set $X$ is made by an arbitrary union of open $M$-cones ``outgoing from points of $X\cap\mathbb S_M$, truncated near their vertex".
\begin{remark}\label{rmk:19}
{\rm It is worthwhile noticing that the quasi-homogeneous symbols considered in Definition \ref{def:1} are related to the weighted smooth symbols introduced in Section \ref{appl_sct} by the following inclusion
\begin{equation*}
S^r_M\subset S^r_{1/m_\ast,\langle\cdot\rangle_M}
\end{equation*}
(a similar inclusion being valid for the corresponding classes of local symbols).}
\end{remark}
\subsection{ Example}\label{EX}
For $M=(1,2)$, let us consider in $\mathbb R^2$ the quasi-homogeneous weight function
\begin{equation}\label{q_h_2}
\langle \xi\rangle_{M}=\left(1+\xi_1^2+\xi^4_2 \right)^{1/2}\,.
\end{equation}
We introduce the following operator
\begin{equation}
\begin{split}
P(x,\partial)= x_1\partial_{x_1}+ i\partial_{x_1}-\partial_{x_2^2}^2\,.
\end{split}
\end{equation}
Its symbol $P(x,\xi)=ix_1\xi_1-\xi_1+\xi_2^2$ belongs to the local class $S^1_M(\Omega)$ where $\Omega=\mathbb R^2$.
Introducing the {\it $M$-characteristic set} of $P(x,\partial)$ as
\begin{equation}\label{eqEX4}
\textup{Char}P=\left\{ (x,\xi)\in\mathbb R^2_x\times\mathbb R^2_\xi\setminus\{0\}\, ,\, P(x,\xi)=0\right\},
\end{equation}
we have
\begin{equation*}
\textup{Char}P=\left\{ (0,x_2,\xi_1,\xi_2);\, x_2\in\mathbb R\,,\,\,\xi_1=\xi_2^2\,,\,\,\xi_2\neq 0 \right\}\!=\!\{0\}\times\mathbb R\times\!\!\bigcap\limits_{0<k<1}\!\!(\mathbb R^2\setminus X_k)\,,
\end{equation*}
where
\begin{equation}\label{eqEX2.1}
X_k=\left\{(\xi_1,\xi_2)\in \mathbb R^2\,\,;\,\, \xi_1\le(1-k)\xi_2^2\,\,\textup{or}\,\, \xi_1\ge\frac{1}{1-k}\xi_2^2 \right\}\,,\,\,0<k<1\,.
\end{equation}
Notice also that $P(x,\xi)$ is {\it quasi-homogeneous of degree one}, in the sense that
\begin{equation}\label{QH}
P(x,t^{1/M}\xi)=P(x,t\xi_1,t^{1/2}\xi_2)=tP(x,\xi)\,,\qquad\forall\,t>0\,.
\end{equation}
The properties collected above yield that the symbol $P(x,\xi)$ is $\langle\cdot\rangle_M-$elliptic at every point $x^0=(x^0_1,x^0_2)\in\mathbb R^2$ with $x^0_1\neq 0$; indeed since $\vert P\vert$ does not vanish at each point of the compact set $\{x^0\}\times\mathbb S_M$, being $\mathbb S_M=\{\eta=(\eta_1,\eta_2)\in\mathbb R^2\,:\,\,\eta_1^2+\eta_2^4=1\}$ is the unit $M-$sphere, by continuity
\begin{equation*}
c_0:=\inf\limits_{\eta\in\mathbb S_M}\vert P(x^0,\eta)\vert>0\,.
\end{equation*}
Hence the quasi-homogeneity of $P$ yields for $\vert\xi\vert_M\ge 1$:
\begin{equation*}
\vert P(x^0,\xi)\vert=\vert\xi\vert_M\vert P(x^0,\eta)\vert\ge c_0\vert\xi\vert_M\ge c_0/\sqrt 2\langle\xi\rangle_M\,,
\end{equation*}
where $\eta:=\vert\xi\vert_M^{1/M}\xi\in\mathbb S_M$.
\medskip
By resorting to Proposition \ref{equivalence_prop}, we show now that at every point $x^0=(0,x^0_2)$, with an arbitrary $x^0_2\in\mathbb R$, $P(x,\xi)$ is $[\langle\cdot\rangle_M]-$microlocally elliptic in any set of the family $\{X_k\}_{0<k<1}$ defined by \eqref{eqEX2.1}. So, let us take arbitrary $x^0=(0,x^0_2)$ and $0<k<1$; since $\vert P\vert$ is different from zero and continuous (hence uniformly continuous) on the compact subset $\{x^0\}\times (X_k\cap\mathbb S_M)$ of $T^\circ\mathbb R^n\setminus\textup{Char}P$, some constants $c_k>0$ and $0<\tilde\varepsilon<1$ sufficiently small can be found such that
\begin{equation}
\vert P(x^0,\eta)\vert\ge c_k\,,
\end{equation}
for $\eta$ ranging on the covering of $X_k\cap\mathbb S_M$ made by the open $M-$balls $B_M(\tilde\eta;\tilde\varepsilon)$ centered at points $\tilde\eta$ of $X_k\cap\mathbb S_M$ with radius $\tilde\varepsilon$. Take now an arbitrary point $\xi\in\bigcup\limits_{\tilde\eta\in X_k\cap\mathbb S_M}\Gamma_M(\tilde\eta;\tilde\varepsilon)$ such that $\vert\xi\vert_M>c/\tilde\varepsilon$ with suitable $c>0$; then $\tilde\eta\in X_k\cap\mathbb S_M$ and $t>0$ exist such that $\xi=t^{1/M}\eta$ for some $\eta\in B_M(\tilde\eta;\tilde\varepsilon)$. Since $\vert\tilde\eta\vert_M=1$, we may take $\tilde\varepsilon$ so small that $\vert\eta\vert_M\le\hat c$ for some positive constant $\hat c$ (independent of $\eta$ and $\tilde\eta$). Exploiting again the quasi-homogeneity of $P$ and the quasi-norm $\vert\cdot\vert_M$ we get
\begin{equation*}
\vert P(x^0,\xi)\vert=t\vert P(x^0,\eta)\vert\ge c_kt=\frac{c_k}{\hat c}t\vert\eta\vert_M=\frac{c_k}{\hat c}\vert\xi\vert_M\ge\tilde c_k\langle\xi\rangle_M\,,
\end{equation*}
with suitable $\tilde c_k>0$. Since the set $X_k$ is $M-$conic, in view of Proposition \ref{equivalence_prop} there exists $0<\varepsilon^\prime<\tilde\varepsilon$ (up to shrink $\tilde\varepsilon$ if necessary) such that
\begin{equation*}
(X_k)_{[\varepsilon^\prime\langle\cdot\rangle_M]}\subset\bigcup\limits_{\tilde\eta\in X_k\cap\mathbb S_M}\Gamma_M(\tilde\eta;\tilde\varepsilon)\cap\left\{\vert\xi\vert_M>c/\tilde\varepsilon\right\}\,.
\end{equation*}
This shows that $P$ is microlocally $[\langle\cdot\rangle_M]-$elliptic in $X_k$ at the point $x^0=(0,x^0_2)$.
\medskip
Since $P(x,\xi)\in S^1_M(\mathbb R^2)$ and in view of Remark \ref{rmk:19}, the results of propagation of Fourier Lebesgue singularities for linear an semilinear equations, collected in the preceding Sections \ref{mcl_reg_sct}, \ref{semi-linear_sct}, can be applied to the operator $P(x,\partial)$.
\medskip
Let $u\in\mathcal D^\prime(\mathbb R^2)$ be a solution to the linear equation
\begin{equation}\label{lin_P}
P(x,\partial)u=f(x)\,,
\end{equation}
with a given forcing term $f$. Applying to \eqref{lin_P} the result of Proposition \ref{filter_inclusions} (with $r=1$), we obtain at once that the following inclusions
\begin{equation*}
\Xi_{[\langle\cdot\rangle_M],\,\mathcal FL^p_{s-1,M},\,x^0}f\cap\Sigma_{[\langle\cdot\rangle_M],\,x^0}P\subset\Xi_{[\langle\cdot\rangle_M],\,\mathcal FL^p_{s,M},\,x^0}u\subset\Xi_{[\langle\cdot\rangle_M],\,\mathcal FL^p_{s-1,M},\,x^0}f
\end{equation*}
hold true for all $s\in\mathbb R$ and $p\in[1,+\infty]$.
\medskip
Consider now the following semilinear equation
\begin{equation}\label{eq_semilineare}
P(x,\partial)u+F(x,u,\partial_{x_2}u)=f(x)\,,
\end{equation}
where $F=F(x,\zeta)$ is a nonlinear function of $x=(x_1,x_2)$ and $\zeta=(\zeta_1,\zeta_2)$ fulfilling the regularity assumptions stated in Theorem \ref{semilin_thm}, and $f=f(x)$ some given forcing term. With respect to the quasi-homogeneous weight \eqref{q_h_2}, the order of derivatives of the unknown function $u$ involved in the nonlinearity in \eqref{eq_semilineare} is easily seen to be $\le 1/2$ (that is such derivatives are properly supported operators in $S_M^{l}$ with order $l\le 1/2$). Then we may apply to \eqref{eq_semilineare} the result of Theorem \ref{semilin_thm} (with $r=1$ and $\varepsilon=1/2$) to prove the following statement.
\begin{proposition}
Given $x^0=(x^0_1,x^0_2)\in\mathbb R^2$, $p\in[1,+\infty]$ and $s>\tilde t>\frac{2}{q}+\frac12$, with $\frac1{p}+\frac1{q}=1$, let $u\in\mathcal FL^p_{\tilde t,\,M,\,{\rm loc}}(x^0)$ be a solution to \eqref{eq_semilineare}.
\begin{itemize}
\item[a.] If $2\tilde t-2-\frac{4}{q}\notin\mathbb Z$ then
\begin{equation}\label{filtro_2}
\Xi_{[\langle\cdot\rangle_M],\,\mathcal FL^p_{s-1,M},\,x^0}f\cap\Sigma_{[\langle\cdot\rangle_M],\,x^0}P\subset\Xi_{[\langle\cdot\rangle_M],\,\mathcal FL^p_{t,M},\,x^0}u\,,
\end{equation}
holds true for all $t\le\min\left\{s,\tilde t+1+\frac12 E\left(2\tilde t-2-\frac{4}{q}\right)\right\}$;
\item[b.] if $2\tilde t-2-\frac{4}{q}\in\mathbb Z$ then the inclusion \eqref{filtro_2} holds true for all $t\le\min\left\{s,\tilde t+\frac12+\frac12 E\left(2\tilde t-2-\frac{4}{q}\right)\right\}$.
\end{itemize}
\end{proposition}
\begin{proof}
For any $\tilde t>\frac{2}{q}+\frac12$, let $\tau>0$ be chosen such that $\tilde t-\frac12\ge\tau>\frac{2}{q}$, then Theorem \ref{semilin_thm} can be directly applied to equation \eqref{eq_semilineare} (where, according to the observations above, it is set $r=1$, $\varepsilon=1/2$, and we also make use of Remark \ref{rmk:18} for $\lambda=\langle\cdot\rangle_M$ and $\tau$ as above) to find that inclusion \eqref{filtro_2} holds true for all $t\le\min\left\{s,\tilde t+1+\frac12E(2\tilde t-2-2\tau)\right\}$. To conclude, it is enough to observe that for $2\tilde t-2-\frac{4}{q}\notin\mathbb Z$ we can take $\tau$ sufficiently close to $\frac{2}{q}$ to have
\begin{equation*}
E\left(2\tilde t-2-2\tau\right)=E\left(2\tilde t-2-\frac{4}{q}\right)\footnote{According to the result of Theorem \ref{semilin_thm}, this corresponds to the best possible range for $t$.};
\end{equation*}
this proves the statement {\it a}.
If, on the contrary, one has $2\tilde t-2-\frac{4}{q}\in\mathbb Z$ then $2\tilde t-2-2\tau<2\tilde t-2-\frac{4}{q}=E(2\tilde t-2-\frac{4}{q})$ whenever $\tau$ is taken as above; then for $\tau$ sufficiently close to $\frac{2}{q}$ we get
\begin{equation*}
E(2\tilde t-2-2\tau)=E\left(2\tilde t-2-\frac{4}{q}\right)-1
\end{equation*}
which gives the statement {\it b}.
\end{proof}
|
\section{Introduction}
\label{Introduction}
We have witnessed a significant revolution in computer vision brought by the deep Convolutional Neural Networks (CNNs).
With powerful computational resources ({\em e.g.}, GPUs)
and a large amount of labeled training data ({\em e.g.}, \cite{Deng_2009_ImageNet}),
a hierarchical structure containing different levels of visual concepts is constructed and trained~\cite{Krizhevsky_2012_ImageNet}
to produce impressive performance on large-scale visual recognition tasks~\cite{Russakovsky_2015_ImageNet}.
A pre-trained deep network is also capable of generating deep features for various tasks,
such as image classification~\cite{Jia_2014_CAFFE}\cite{Donahue_2014_DeCAF},
image retrieval~\cite{Razavian_2014_CNN}\cite{Xie_2015_Image} and object detection~\cite{Girshick_2014_Rich}\cite{Girshick_2015_Fast}.
CNN is composed of several stacked layers, each of which contains a number of neurons.
We argue that modeling the co-occurrence of neuron responses is important, whereas less studied in the previous work.
For this, we define a set of {\em geometric neural phrases} on the basis of the hidden neurons,
and propose the {\bf Geometric Neural Phrase Pooling} (GNPP) algorithm to encode them efficiently.
GNPP can be regarded as a new type of layer, and inserted into a network with little computational overhead
({\em e.g.}, $1.29\%$ and $2.52\%$ extra time and memory costs in the experiments on {\bf ImageNet}).
We explain the behavior of GNPP by noting that it punishes the isolated neuron responses,
and that the isolated responses are often less reliable than clustered ones, especially in the high-level network layers.
Experimental results show that adding GNPP layers boosts image classification accuracy significantly and consistently.
Later, we will discuss the benefits brought by the GNPP layer from different points of view,
showing that GNPP produces better internal representation, builds latent connections, and accelerates the network training process.
The remainder of this paper is organized as follows.
Section~\ref{RelatedWork} briefly introduces related work.
Section~\ref{GNPP} introduces the GNPP layer, and Section~\ref{Experiments} shows experimental results.
We discuss the benefits brought by adding GNPP layers in Section~\ref{Benefits}.
Finally, we conclude this work in Section~\ref{Conclusions}.
\section{Related Work}
\label{RelatedWork}
\subsection{The Bag-of-Visual-Words Model}
\label{RelatedWork:BoVW}
The Bag-of-Visual-Words (BoVW) model~\cite{Csurka_2004_Visual} represents each image as a high-dimensional vector.
It typically consists of three stages, {\em i.e.}, descriptor extraction, feature encoding and feature aggregation.
Due to the limited descriptive ability of raw pixels,
handcrafted descriptors such as SIFT~\cite{Lowe_2004_Distinctive},
HOG~\cite{Dalal_2005_Histograms} or other variants~\cite{Xie_2015_RIDE} are extracted.
The set of local descriptors on an image is denoted as
${\mathcal{D}}={\left\{\left(\mathbf{d}_m,\mathbf{l}_m\right)\right\}_{m=1}^M}$,
where $M$ is the number of descriptors,
$\mathbf{d}_m$ is the description vector and $\mathbf{l}_m$ is the 2D location of the $m$-th word.
A visual vocabulary or codebook is then built to capture the data distribution in feature space.
The codebook is a set of codewords: ${\mathcal{B}}={\left\{\mathbf{c}_b\right\}_{b=1}^B}$,
in which $B$ is the codebook size and each codeword has the same dimension with the descriptors.
Each descriptor $\mathbf{d}_m$ is then quantized onto the codebook as a visual word ${\mathbf{f}_m}\in{\mathbb{R}_{\geqslant0}^B}$.
Effective feature quantization algorithms include sparse coding~\cite{Yang_2009_Linear}\cite{Wang_2010_Locality}
and high-dimensional encoding~\cite{Perronnin_2010_Improving}\cite{Zhou_2010_Image}\cite{Kobayashi_2014_Dirichlet}.
${\mathcal{F}}={\left\{\left(\mathbf{f}_m,\mathbf{l}_m\right)\right\}_{m=1}^M}$ is the set of visual words.
Finally, these words are aggregated as an image-level representation vector~\cite{Lazebnik_2006_Beyond}\cite{Feng_2011_Geometric}.
These Image-level vectors are then normalized and fed into
machine learning algorithms~\cite{Fan_2008_LIBLINEAR} for training and testing,
or used in some training-free image classification algorithms~\cite{Boiman_2008_Defense}\cite{Xie_2015_Image}.
\subsection{Geometric Phrase Pooling}
\label{RelatedWork:GPP}
The basic unit in the BoVW model is a {\em visual word}, {\em i.e.}, a quantized local descriptor.
Dealing with individual visual words does not consider the spatial co-occurrence of visual features.
To this end, researchers propose {\em visual phrase}~\cite{Yuan_2007_Discovery}\cite{Zhang_2011_Image}
as a mid-level data structure connecting low-level descriptors and high-level visual concepts~\cite{Zhang_2009_Descriptive}.
A visual phrase is often defined as a group of neighboring visual words~\cite{Zhang_2009_Descriptive}\cite{Xie_2014_Spatial}.
It can be used to filter out the false matches in object retrieval~\cite{Zhang_2011_Image}\cite{Jiang_2012_Randomized},
or improve the descriptive ability of visual features for image classification~\cite{Xie_2014_Spatial}\cite{Xie_2013_Hierarchical}.
Geometric Phrase Pooling (GPP)~\cite{Xie_2014_Spatial} is an efficient algorithm for extracting and encoding visual phrases.
GPP starts from constructing, for each visual word, a {\em geometric visual phrase}, which is a group of visual words:
${\mathcal{G}_m}={\left(\mathbf{f}_m,\mathbf{l}_m\right)\cup
\left\{\left(\mathbf{f}_m^{\left(k\right)},\mathbf{l}_m^{\left(k\right)}\right)\right\}_{k=1}^K}$.
In $\mathcal{G}_m$, $\left(\mathbf{f}_m,\mathbf{l}_m\right)$ is the {\em central word},
and all the other $K$ words are {\em side words},
located in a small neighborhood $\mathcal{N}_m$ of the central position $\mathbf{l}_m$.
GPP encodes each geometric visual phrase $\mathcal{G}_m$
by adding the maximal response of the side words to the central word:
${\mathbf{p}_m}={\mathbf{f}_m+\max_{k=1}^{K}\left\{s_m^{\left(k\right)}\times\mathbf{f}_m^{\left(k\right)}\right\}}$,
where $\max_{k=1}^{K}\left\{\cdot\right\}$ denotes dimension-wise maximization.
Note that the central word is not included in the maximization.
$s_m^{\left(k\right)}$ is the smoothing weight of the $k$-th side word in $\mathcal{G}_m$.
Most often, $s_m^{\left(k\right)}$ is determined by the Euclidean distance between $\mathbf{l}_m^{\left(k\right)}$ and $\mathbf{l}_m$,
{\em e.g.}, ${s_m^{\left(k\right)}}=
{\exp\!\left\{-\tau\times\left\|\mathbf{l}_m-\mathbf{l}_m^{\left(k\right)}\right\|_2\right\}}$,
where ${\tau}>{0}$ is the pre-defined smoothing parameter.
Note that, at least in theory, the GPP algorithm can be applied to any data with a spatial attribute.
\subsection{Convolutional Neural Networks}
\label{RelatedWork:CNN}
The Convolutional Neural Network (CNN) serves as a hierarchical model for large-scale visual recognition.
It is based on the observation that a network with enough neurons is able to fit any complicated data distribution.
In past years, neural networks were shown effective for simple recognition tasks~\cite{LeCun_1990_Handwritten}.
More recently, the availability of large-scale training data ({\em e.g.}, ImageNet~\cite{Deng_2009_ImageNet}) and powerful GPUs
make it possible to train deep CNNs~\cite{Krizhevsky_2012_ImageNet} which significantly outperform BoVW models.
A CNN is composed of several stacked layers.
In each of them, responses from the previous layer are convoluted with a filter bank and activated by a differentiable non-linearity.
Hence, a CNN can be considered as a composite function,
which is trained by back-propagating error signals defined by the difference between supervision and prediction at the top layer.
Efficient methods were proposed to help CNNs converge faster and prevent over-fitting,
such as ReLU activation~\cite{Krizhevsky_2012_ImageNet},
batch normalization~\cite{Ioffe_2015_Batch} and regularization~\cite{Hinton_2012_Improving}\cite{Xie_2016_DisturbLabel}.
It is believed that deeper networks
produce better recognition results~\cite{Simonyan_2015_Very}\cite{Szegedy_2015_Going}\cite{He_2015_Deep}.
The intermediate responses of CNNs, {\em i.e.}, the so-called deep features,
serve as effective image descriptions~\cite{Donahue_2014_DeCAF}, or a set of latent visual attributes~\cite{Zhang_2014_PANDA}.
They can be used for various types of vision tasks,
including image classification~\cite{Jia_2014_CAFFE}\cite{Xie_2016_InterActive},
image retrieval~\cite{Razavian_2014_CNN}\cite{Xie_2015_Image} and object detection~\cite{Girshick_2014_Rich}.
A discussion of how different CNN configurations impact deep feature performance is available in~\cite{Chatfield_2014_Return}.
\section{Geometric Neural Phrase Pooling}
\label{GNPP}
This section presents the Geometric Neural Phrase Pooling (GNPP) algorithm and its application to improve the CNN model.
\subsection{The GNPP Layer}
\label{GNPP:Layer}
We start with a hidden layer $\mathbf{X}$ in the CNN model.
$\mathbf{X}$ is a 3D neuron cube with $W\times H\times D$ neurons, where $W$, $H$ and $D$ are the width, height and depth of the cube.
The response of each neuron corresponds to
the inner-product of a local patch in the previous layer and a filter (convolutional kernel).
We naturally consider the data as a set of $D$-dimensional {\em visual words} indexed over a 2D spatial domain.
We denote the set as ${\mathcal{X}}={\left\{\mathbf{x}_{w,h}\right\}_{w=1,h=1}^{W,H}}$,
in which ${\mathbf{x}_{w,h}}\in{\mathbb{R}_{\geqslant0}^{D}}$ for each $w$ and $h$.
The spatial domain coordinate $\left(w,h\right)$ is not the same as the pixel coordinate $\left(a,b\right)$ in the original image,
but they are linearly corresponded.
\newcommand{12.0cm}{12.0cm}
\begin{figure}[t]
\begin{center}
\includegraphics[width=12.0cm]{GeometricNeuralPhrases.pdf}
\end{center}
\caption{
Left: the architecture of a toy CNN model.
A geometric neural phrase is defined on the basis of a set of neural words.
Right: two types of neighborhood used in this work (best viewed in color PDF).
The green side words are weighted by $\sigma$ and the blue ones by $\sigma^2$, where $\sigma$ is the smoothing parameter.
}
\label{Fig:GeometricNeuralPhrases}
\end{figure}
A {\em geometric neural phrase} is defined as
${\mathcal{G}_{w,h}}={\left\{\mathbf{x}_{w',h'}\mid{\mathbf{x}_{w',h'}}\in{\mathcal{N}_{w,h}}\right\}}$,
where $\mathcal{N}_{w,h}$ is the neighborhood of $\mathbf{x}_{w,h}$.
Given the number of side words $K$,
we can rewrite it as ${\mathcal{G}_{w,h}}={\mathbf{x}_{w,h}\cup\left\{\mathbf{x}_{w,h}^{\left(k\right)}\right\}_{k=1}^K}$,
where $\mathbf{x}_{w,h}$ is the {\em central word}, and all the others in $\mathcal{G}_{w,h}$ are {\em side words}.
For simplicity, we consider two fixed types of neighborhood, shown in Figure~\ref{Fig:GeometricNeuralPhrases}.
If the central word is located on the boundary of the neuron map, the side words outside the map are simply ignored.
The {\bf Geometric Neural Phrase Pooling} (GNPP) algorithm
computes an updated neural response for each geometric neural phrase $\mathcal{G}_{w,h}$ individually:
\begin{equation}
\label{Eqn:GNPP}
{\mathbf{z}_{w,h}}={\frac{1}{2}\left[\mathbf{x}_{w,h}+
\max_{k=1}^K\left\{s_{w,h}^{\left(k\right)}\times\mathbf{x}_{w,h}^{\left(k\right)}\right\}\right]},
\end{equation}
where $\max_{k=1}^K\left\{\cdot\right\}$ is the maximization over $K$ side words.
Note that the central word is not included in the maximization.
We add the coefficient $\frac{1}{2}$ to approximately preserve the average scale of neuron responses.
We define a smoothing parameter ${\sigma}\in{\left(0,1\right]}$.
A side word is weighted by either ${s_{w,h}^{\left(k\right)}}={\sigma}$ or ${s_{w,h}^{\left(k\right)}}={\sigma^2}$,
according to the relative position to the central word.
Of course, one can modify the definition of both neighborhood and weights,
{\em e.g.}, using a larger neighborhood or assigning smaller weights on side words,
but these minor changes do not impact much on the performance (see Section~\ref{Experiments:Analysis}).
GNPP averages neuron responses over the central word and the maximal candidate among all side words.
Although this looks like the behavior of a local smoother,
we emphasize that GNPP is intrinsically different from other smoothers such as vanilla Gaussian blur.
Gaussian blur can be formulated as convoluting the input data with a fixed kernel.
Applying Gaussian blur after a convolutional layer is similar to using larger kernels,
where some weights are not independent to each other.
As expected, adding Gaussian blur does not obtain accuracy gain.
We add a vanilla Gaussian blur layer before each pooling layer of the {\bf LeNet}, and test it on {\bf CIFAR10}.
The baseline error rate is $17.07\%\pm0.15\%$, and the network with Gaussian blur reports $17.05\%\pm0.13\%$.
On the other side, the network with GNPP reports $14.78\%\pm0.17\%$ (see Section~\ref{Experiments:CIFAR}).
In summary, {\bf GNPP does something that a linear smoother cannot do}.
Since GNPP does not change the dimension ($W$, $H$ and $D$) of the neuron cube,
we can regard GNPP as an intermediate network layer, {\em i.e.}, the {\bf GNPP layer}.
Although the GNPP layer can, at least in theory, appear anywhere,
{\bf we only insert it between a convolutional layer and a pooling layer}, due to the reason to be elaborated in the next subsection.
\subsection{Modeling the Spatial Co-occurrence}
\label{GNPP:Explanation}
We show that GNPP is an implicit way of punishing {\em isolated} neuron responses.
Therefore, GNPP works well on the assumption that {\em clustered} neuron responses are more reliable than {\em isolated} ones.
In this subsection, we will elaborate that such an assumption is better satisfied on the high-level layers of a CNN.
\begin{figure}[t]
\begin{center}
\includegraphics[width=12.0cm]{GNPPIllustration.pdf}
\end{center}
\caption{
A conceptual illustration of GNPP (best viewed in color PDF).
Left: an image is convoluted with a template.
Middle: the original one-dimensional neuron responses.
Right: the responses after GNPP (type 1, ${\sigma}={1.0}$).
The isolated high response (around the bottom-left corner) is decreased and smoothed,
while the clustered high responses (around the bottom-right corner) are preserved.
}
\label{Fig:GNPPIllustration}
\end{figure}
To start, we note that the computation in~\eqref{Eqn:GNPP} is carried out in parallel over the $D$ channels.
Without loss of generality, we only consider a single channel in a hidden layer,
and our conclusion remains valid for the entire layer (containing $D$ channels).
In other words, we can simplify to the situation where we are dealing with $W\times H$ one-dimensional visual words.
In a CNN, neuron responses in one layer are generated by convolution.
Convoluting data with the kernel can be regarded as template matching on different spatial locations.
After ReLU activation~\cite{Krizhevsky_2012_ImageNet},
the preserved positive neuron responses correspond to those local patches with high similarity to the template.
Figure~\ref{Fig:GNPPIllustration} shows a toy example of (ReLU-activated) convolution results,
in which we can find some {\em clustered} high responses and some {\em isolated} ones.
Since GNPP averages the neuron responses over the central word and the maximal candidate of side words,
the {\em clustered} responses are approximately preserved, while the {\em isolated} ones are punished.
A toy example is shown in Figure~\ref{Fig:GNPPIllustration}.
We explain why clustered responses are more reliable, especially on a high-level layer,
where the isolated responses often correspond to unexpected random noise~\cite{Wang_2015_Discovering}.
This is because high-level convolutional kernels are highly ``specialized'',
{\em i.e.}, they often represent concrete visual concepts,
{\em e.g.}, {\em car wheel} or {\em aircraft nose}~\cite{Wang_2015_Discovering}.
Meanwhile, as the network level goes up, the receptive field of a neuron becomes larger
({\em e.g.}, a neuron on the {\em conv-5} layer of the {\bf AlexNet} can ``see'' $163\times163$ pixels on the input image),
and neighboring neurons share more and more common visual information
({\em e.g.}, the overlapping rate of two neighboring neurons on the {\em conv-5} layer is $90.2\%$).
Thus, if a positive neuron response is caused by the correct match of a visual concept,
its neighboring neurons are also likely to be activated, leading to a cluster of positive neuron responses.
Oppositely, if it is caused by some random noise,
its neighboring neurons may not be activated, and this isolated response shall be punished.
In conclusion, the core idea of GNPP is to model the spatial co-occurrence of neurons produced by a convolutional layer,
or find reliable features by punishing the isolated responses which are more likely to be unexpected random noise.
We note that pooling, when applied right after GNPP, is an efficient way of aggregating these rectified neuron responses.
Therefore, in this work, we only insert GNPP between a convolutional layer and a pooling layer.
\subsection{Comparison to Other Work}
\label{GNPP:Comparison}
The GNPP algorithm is inspired by the GPP algorithm which originates from the BoVW model (see Section~\ref{RelatedWork:GPP}).
GPP models the spatial context of visual words, and GNPP models the spatial co-occurrence of neural words.
In the BoVW model, GPP can only be applied before a max-pooling layer, but GNPP can be inserted anywhere into the CNN model.
In the {\bf SVHN} and {\bf CIFAR} experiments, we also show that GNPP cooperates well with the average-pooling layers.
The GNPP layer is related to the Spatial Pyramid Pooling (SPP) layer~\cite{He_2014_Spatial}
and the Region-Of-Interest (ROI) pooling layer~\cite{Ren_2015_Faster}.
However, the motivation and working mechanism of GNPP are quite different from these two layers.
The goal of GNPP is to punish isolated neuron responses and improve the descriptive power of every single neuron,
while the SPP layer and the ROI pooling layer aim at summarizing local neurons into a regional description.
The basic unit in the GNPP layer is a single neuron, and pooling is performed on a small set of its neighboring neurons,
whereas both the SPP layer and the ROI pooling layer work on image regions.
Finally, we point out that GNPP can be integrated with other network layers to further improve the recognition performance.
\section{Experiments}
\label{Experiments}
In this section, we show the experimental results of inserting the GNPP layer into different CNN models.
We first observe the performance by evaluating relatively shallow networks on small datasets,
then use our conclusions to inform the application of GNPP to deeper networks and the large-scale database.
\subsection{The {\bf MNIST} and {\bf SVHN} Datasets}
\label{Experiments:MNISTandSVHN}
{\bf MNIST}~\cite{LeCun_1998_Gradient} is one of the most popular datasets for handwritten digit recognition.
It contains $60\rm{,}000$ training and $10\rm{,}000$ testing samples,
uniformly distributed over $10$ categories (digits $0$--$9$).
All the samples are $28\times28$ grayscale images.
We use a modified version ($2$ convolutional layers) of the {\bf LeNet}~\cite{LeCun_1990_Handwritten} as the baseline.
With abbreviated notation, the network configuration is written as:
\begin{spverbatim}
{C5(S1P0)@20-MP2(S2)}{C5(S1P0)@50-MP2(S2)}{FC500}{FC10}.
\end{spverbatim}
\noindent
Here, {\tt C5(S1P0)@20} denotes a convolutional layer with $20$ kernels of size $5\times5$, spatial stride $1$ and padding width $0$,
{\tt MP2(S2)} is a max-pooling layer with pooling region $2\times2$ and spatial stride $2$,
and {\tt FC500} is a fully-connected layer with $500$ outputs.
All the convolution results are activated by ReLU~\cite{Krizhevsky_2012_ImageNet}, and we use the softmax loss function.
In the later experiments, we will directly use the same notations.
We apply $20$ training epochs with learning rate $10^{-3}$, followed by $4$ epochs with learning rate $10^{-4}$,
and another $1$ epoch with learning rate $10^{-5}$.
We test each network five times individually with different initialization and report the averaged error rate and standard deviation.
The {\bf SVHN} dataset~\cite{Netzer_2011_Reading} is a larger collection of $32\times32$ RGB images,
with $73\rm{,}257$ training samples, $26\rm{,}032$ testing samples, and $531\rm{,}131$ extra training samples.
We split the data as in the previous methods~\cite{Netzer_2011_Reading},
{\em i.e.}, preserving $6\rm{,}000$ images for validation, and using the remainder for training.
We use Local Contrast Normalization (LCN) for data preprocessing,
following~\cite{Sermanet_2012_Convolutional}\cite{Zeiler_2013_Stochastic}\cite{Goodfellow_2013_Maxout}.
We use another version of the {\bf LeNet} with $3$ convolutional layers, abbreviated as:
\begin{spverbatim}
{C5(S1P2)@32-MP3(S2)}{C5(S1P2)@32-AP3(S2)}{C5(S1P2)@64-AP3(S2)}{FC10}.
\end{spverbatim}
\noindent
Here, {\tt AP} indicates an average-pooling layer.
We apply $12$ training epochs with learning rate $10^{-3}$, followed by $2$ epochs with learning rate $10^{-4}$,
and another $1$ epoch with learning rate $10^{-5}$.
Each network is individually tested five times.
\newcommand{1.6cm}{1.6cm}
\begin{table*}[t]
\begin{center}
\subfloat[\normalsize{{\bf MNIST}, with the $2$-layer {\bf LeNet}, no Dropout}]{
\begin{tabular}{|c|c||R{1.6cm}|R{1.6cm}|R{1.6cm}||R{1.6cm}|R{1.6cm}|R{1.6cm}|}
\hline
L$1$ & L$2$ & T1($1.0$) & T1($0.9$) & T1($0.8$) & T2($1.0$) & T2($0.9$) & T2($0.8$) \\
\hline\hline
{} & {} & $ 0.87\pm.02$ & $ 0.87\pm.02$ & $ 0.87\pm.02$ &
$ 0.87\pm.02$ & $ 0.87\pm.02$ & $ 0.87\pm.02$ \\
\hline
\checkmark & {} & $ 0.72\pm.04$ & $ 0.73\pm.03$ & $ 0.70\pm.05$ &
$ 0.71\pm.06$ & $ 0.71\pm.06$ & $ 0.72\pm.04$ \\
\hline
{} & \checkmark & $ 0.75\pm.03$ & $ 0.79\pm.02$ & $ 0.77\pm.05$ &
$ 0.73\pm.04$ & $ 0.75\pm.04$ & $ 0.73\pm.05$ \\
\hline
\checkmark & \checkmark & $\mathbf{ 0.72}\pm.03$ & $\mathbf{ 0.67}\pm.04$ & $\mathbf{ 0.69}\pm.04$ &
$\mathbf{ 0.63}\pm.03$ & $\mathbf{ 0.64}\pm.03$ & $\mathbf{ 0.67}\pm.03$ \\
\hline
\end{tabular}
} \\
\subfloat[\normalsize{{\bf MNIST}, with the $2$-layer {\bf LeNet}, Dropout ratio $0.5$}]{
\begin{tabular}{|c|c||R{1.6cm}|R{1.6cm}|R{1.6cm}||R{1.6cm}|R{1.6cm}|R{1.6cm}|}
\hline
L$1$ & L$2$ & T1($1.0$) & T1($0.9$) & T1($0.8$) & T2($1.0$) & T2($0.9$) & T2($0.8$) \\
\hline\hline
{} & {} & $ 0.72\pm.03$ & $ 0.72\pm.03$ & $ 0.72\pm.03$ &
$ 0.72\pm.03$ & $ 0.72\pm.03$ & $ 0.72\pm.03$ \\
\hline
\checkmark & {} & $ 0.59\pm.02$ & $ 0.61\pm.05$ & $ 0.62\pm.03$ &
$ 0.59\pm.03$ & $ 0.59\pm.02$ & $ 0.63\pm.03$ \\
\hline
{} & \checkmark & $ 0.63\pm.03$ & $ 0.62\pm.07$ & $ 0.64\pm.03$ &
$ 0.62\pm.05$ & $ 0.60\pm.03$ & $ 0.65\pm.03$ \\
\hline
\checkmark & \checkmark & $\mathbf{ 0.58}\pm.05$ & $\mathbf{ 0.55}\pm.02$ & $\mathbf{ 0.57}\pm.02$ &
$\mathbf{ 0.54}\pm.05$ & $\mathbf{ 0.56}\pm.04$ & $\mathbf{ 0.61}\pm.05$ \\
\hline
\end{tabular}
} \\
\subfloat[\normalsize{{\bf SVHN}, with the $3$-layer {\bf LeNet}, no Dropout}]{
\begin{tabular}{|c|c|c||R{1.6cm}|R{1.6cm}|R{1.6cm}||R{1.6cm}|R{1.6cm}|R{1.6cm}|}
\hline
L$1$ & L$2$ & L$3$ & T1($1.0$) & T1($0.9$) & T1($0.8$) & T2($1.0$) & T2($0.9$) & T2($0.8$) \\
\hline\hline
{} & {} & {} & $ 4.63\pm.06$ & $ 4.63\pm.06$ & $ 4.63\pm.06$ &
$ 4.63\pm.06$ & $ 4.63\pm.06$ & $ 4.63\pm.06$ \\
\hline
\checkmark & {} & {} & $ 4.46\pm.06$ & $ 4.47\pm.05$ & $ 4.42\pm.09$ &
$ 4.42\pm.08$ & $ 4.42\pm.07$ & $ 4.43\pm.09$ \\
\hline
{} & \checkmark & {} & $ 4.15\pm.08$ & $ 4.18\pm.01$ & $ 4.17\pm.07$ &
$ 4.08\pm.10$ & $ 4.19\pm.07$ & $ 4.20\pm.05$ \\
\hline
{} & {} & \checkmark & $ 3.76\pm.03$ & $ 3.72\pm.05$ & $ 3.77\pm.06$ &
$ 3.53\pm.07$ & $ 3.64\pm.07$ & $ 3.65\pm.10$ \\
\hline
\checkmark & \checkmark & {} & $ 4.10\pm.05$ & $ 4.07\pm.03$ & $ 4.10\pm.05$ &
$ 4.10\pm.07$ & $ 4.10\pm.03$ & $ 4.14\pm.07$ \\
\hline
\checkmark & {} & \checkmark & $ 3.55\pm.10$ & $ 3.60\pm.03$ & $ 3.67\pm.06$ &
$ 3.47\pm.05$ & $ 3.47\pm.02$ & $ 3.55\pm.09$ \\
\hline
{} & \checkmark & \checkmark & $\mathbf{ 3.43}\pm.06$ & $ 3.52\pm.07$ & $\mathbf{ 3.55}\pm.04$ &
$\mathbf{ 3.41}\pm.03$ & $ 3.42\pm.04$ & $ 3.51\pm.05$ \\
\hline
\checkmark & \checkmark & \checkmark & $ 3.46\pm.07$ & $\mathbf{ 3.47}\pm.06$ & $\mathbf{ 3.55}\pm.06$ &
$ 3.43\pm.05$ & $\mathbf{ 3.39}\pm.01$ & $\mathbf{ 3.46}\pm.03$ \\
\hline
\end{tabular}
}
\caption{
Classification error rates ($\%$) on {\bf MNIST} and {\bf SVHN}.
L1, L2 and L3 are three pooling layers, `\checkmark' denotes that GNPP is added.
T1 and T2 indicate two types of neighborhood (see Figure~\ref{Fig:GeometricNeuralPhrases}).
$1.0$, $0.9$ and $0.8$ are $\sigma$ values.
}
\label{Tab:MNISTandSVHN}
\end{center}
\end{table*}
When the GNPP layer is inserted into the network, it can appear before any subset of the pooling layers.
We enumerate all the possibilities, and summarize the results in Table~\ref{Tab:MNISTandSVHN}.
One can observe that the use of GNPP significantly improves the recognition accuracy.
The relative error rates are decreased by over $20\%$ on both datasets.
Meanwhile, GNPP can be used with Dropout~\cite{Hinton_2012_Improving}
(randomly discarding some neuron responses on the second pooling layer):
on {\bf MNIST}, the error rate is reduced from $0.72\%$ to $0.58\%$.
\subsection{The {\bf CIFAR10} and {\bf CIFAR100} Datasets}
\label{Experiments:CIFAR}
\begin{table*}[t]
\begin{center}
\subfloat[\normalsize{{\bf CIFAR10}, with the $3$-layer {\bf LeNet}, Dropout ratio $0.2$}]{
\begin{tabular}{|c|c|c||R{1.6cm}|R{1.6cm}|R{1.6cm}||R{1.6cm}|R{1.6cm}|R{1.6cm}|}
\hline
L$1$ & L$2$ & L$3$ & T1($1.0$) & T1($0.9$) & T1($0.8$) & T2($1.0$) & T2($0.9$) & T2($0.8$) \\
\hline\hline
{} & {} & {} & $17.07\pm.15$ & $17.07\pm.15$ & $17.07\pm.15$ &
$17.07\pm.15$ & $17.07\pm.15$ & $17.07\pm.15$ \\
\hline
\checkmark & {} & {} & $16.67\pm.22$ & $16.80\pm.25$ & $16.84\pm.12$ &
$16.65\pm.19$ & $17.03\pm.15$ & $17.04\pm.17$ \\
\hline
{} & \checkmark & {} & $15.79\pm.22$ & $16.09\pm.17$ & $15.95\pm.31$ &
$15.69\pm.11$ & $16.07\pm.27$ & $15.90\pm.09$ \\
\hline
{} & {} & \checkmark & $15.49\pm.15$ & $15.31\pm.20$ & $15.51\pm.25$ &
$15.27\pm.10$ & $15.29\pm.14$ & $15.28\pm.16$ \\
\hline
\checkmark & \checkmark & {} & $15.82\pm.23$ & $15.76\pm.18$ & $15.98\pm.14$ &
$16.05\pm.29$ & $15.90\pm.25$ & $15.94\pm.09$ \\
\hline
\checkmark & {} & \checkmark & $15.15\pm.20$ & $15.29\pm.12$ & $15.44\pm.19$ &
$15.29\pm.32$ & $15.19\pm.35$ & $15.20\pm.35$ \\
\hline
{} & \checkmark & \checkmark & $\mathbf{14.92}\pm.18$ & $15.00\pm.18$ & $15.15\pm.15$ &
$\mathbf{14.83}\pm.25$ & $14.93\pm.20$ & $14.92\pm.16$ \\
\hline
\checkmark & \checkmark & \checkmark & $14.97\pm.17$ & $\mathbf{14.83}\pm.23$ & $\mathbf{14.78}\pm.17$ &
$15.22\pm.16$ & $\mathbf{14.79}\pm.26$ & $\mathbf{14.85}\pm.26$ \\
\hline
\end{tabular}
} \\
\subfloat[\normalsize{{\bf CIFAR100}, with the $3$-layer {\bf LeNet}, no Dropout}]{
\begin{tabular}{|c|c|c||R{1.6cm}|R{1.6cm}|R{1.6cm}||R{1.6cm}|R{1.6cm}|R{1.6cm}|}
\hline
L$1$ & L$2$ & L$3$ & T1($1.0$) & T1($0.9$) & T1($0.8$) & T2($1.0$) & T2($0.9$) & T2($0.8$) \\
\hline\hline
{} & {} & {} & $44.99\pm.19$ & $44.99\pm.19$ & $44.99\pm.19$ &
$44.99\pm.19$ & $44.99\pm.19$ & $44.99\pm.19$ \\
\hline
\checkmark & {} & {} & $44.62\pm.17$ & $44.53\pm.45$ & $44.78\pm.06$ &
$44.43\pm.29$ & $44.58\pm.36$ & $44.58\pm.52$ \\
\hline
{} & \checkmark & {} & $43.34\pm.23$ & $43.71\pm.19$ & $43.37\pm.26$ &
$43.21\pm.23$ & $43.03\pm.27$ & $43.37\pm.30$ \\
\hline
{} & {} & \checkmark & $43.11\pm.24$ & $42.77\pm.37$ & $42.99\pm.24$ &
$42.96\pm.32$ & $42.81\pm.38$ & $43.08\pm.39$ \\
\hline
\checkmark & \checkmark & {} & $43.99\pm.07$ & $43.63\pm.11$ & $43.50\pm.26$ &
$43.38\pm.37$ & $43.34\pm.27$ & $43.46\pm.25$ \\
\hline
\checkmark & {} & \checkmark & $42.85\pm.38$ & $42.81\pm.27$ & $42.82\pm.29$ &
$43.08\pm.27$ & $42.79\pm.34$ & $42.93\pm.22$ \\
\hline
{} & \checkmark & \checkmark & $\mathbf{42.35}\pm.30$ & $\mathbf{42.34}\pm.31$ & $\mathbf{42.04}\pm.20$ &
$\mathbf{42.92}\pm.33$ & $\mathbf{42.72}\pm.25$ & $\mathbf{42.54}\pm.29$ \\
\hline
\checkmark & \checkmark & \checkmark & $42.97\pm.29$ & $42.77\pm.36$ & $42.36\pm.18$ &
$43.31\pm.34$ & $42.85\pm.18$ & $42.60\pm.36$ \\
\hline
\end{tabular}
}
\caption{
Classification error rates ($\%$) on the {\bf CIFAR} datasets.
We use the same notations as in Table~\ref{Tab:MNISTandSVHN}.
We apply Dropout on the simpler {\bf CIFAR10} task.
}
\label{Tab:CIFAR}
\end{center}
\end{table*}
Both {\bf CIFAR10} and {\bf CIFAR100} datasets~\cite{Krizhevsky_2009_Learning}
are subsets of the $80$-million tiny image database~\cite{Torralba_2008_80}.
Both of them have $50\rm{,}000$ training and $10\rm{,}000$ testing samples,
uniformly distributed over $10$ or $100$ categories.
We also use the $3$-layer {\bf LeNet} as in {\bf SVHN}, with the fully-connected layer replaced by {\tt FC100} in {\bf CIFAR100}.
We augment the training data by randomly flipping each training image with $50\%$ probability.
We apply $120$ training epochs with learning rate $10^{-3}$, followed by $20$ epochs with learning rate $10^{-4}$,
and another $10$ epochs with learning rate $10^{-5}$.
Each network is individually tested five times.
Results with all possible GNPP settings are summarized in Table~\ref{Tab:CIFAR}.
Once again, GNPP improves the baseline error rate significantly:
the baseline error rates on both {\bf CIFAR10} and {\bf CIFAR100} are reduced by more than $2\%$,
and the relative error rate decrease are $11.25\%$ and $6.56\%$, respectively.
\subsection{Analysis on Small Experiments}
\label{Experiments:Analysis}
Before we go into deeper networks and larger datasets, we conduct some preliminary analysis based on the results we already have.
First, although inserting GNPP before any pooling layers improves the performance,
the most significant accuracy gain brought by a single GNPP layer is obtained by adding GNPP before the last pooling layer.
This reinforces the conclusion drawn in Section~\ref{GNPP:Explanation},
{\em i.e.}, GNPP works better on the high-level neuron responses.
Meanwhile, on the {\bf SVHN} and {\bf CIFAR} datasets,
adding GNPP before all three pooling layers produces inferior results to that adding GNPP before the second and third pooling layers.
In the later experiments, we first add the GNPP layer before each pooling layer individually,
then use the results to inform the design of the final model.
Regarding the scale of neural phrases, {\em i.e.}, $K$,
we find that increasing the scale is not guaranteed to produce better recognition results.
We explain this by noticing that adding a faraway side word to a neural phrase, most often,
does not provide much related information but risks introducing noise to that unit.
This idea can also be used to explain why a proper smoothing parameter, say, ${\sigma}={0.8}$,
helps to reduce the contribution of faraway side words, leading to better recognition performance.
One may certainly try other choices such as a large neighborhood with a very small $\sigma$,
but we note that the time complexity of a GNPP layer is linear to $K$.
In the later experiments, we will directly use the first type of neighborhood (${K}={4}$) with ${\sigma}={0.8}$.
\subsection{Deeper Networks and the State-of-the-Arts}
\label{Experiments:DeeperNetworks}
We adopt two deeper networks on the above four small datasets to compare with the state-of-the-art results.
One of them (we name it as the {\bf BigNet}) is borrowed from~\cite{Nagadomi_2014_Kaggle} in the Kaggle recognition competition,
and other one one is the $16$-layer Wide Residual Network ({\bf WRN})~\cite{Zagoruyko_2016_Wide} with dropout.
Both networks can be used in each of the four small datasets.
In {\bf CIFAR} datasets, we randomly flip the image with $50\%$ probability.
We train the {\bf BigNet} using $6\times10^6$ samples with learning rate $10^{-2}$,
followed by $3\times10^6$ samples with learning rate $10^{-3}$ and $1\times10^6$ samples with learning rate $10^{-4}$, respectively.
We report a $7.80\%$ error rate on {\bf CIFAR10}, comparable to the original version~\cite{Nagadomi_2014_Kaggle},
which uses a very complicated way of data preparation and augmentation to get a $6.68\%$ error rate.
Training our model needs about $1$ hour, while the original version~\cite{Nagadomi_2014_Kaggle} requires $6$ hours.
We train the {\bf WRN} following the original configuration in~\cite{Zagoruyko_2016_Wide}.
\newcommand{1.8cm}{1.8cm}
\begin{table*}[t]
\begin{center}
\begin{tabular}{|l||R{1.8cm}|R{1.8cm}|R{1.8cm}|R{1.8cm}|}
\hline
{} & {\bf MNIST} & {\bf SVHN} & {\bf CIFAR10} & {\bf CIFAR100} \\
\hline\hline
Zeiler {\em et.al}~\cite{Zeiler_2013_Stochastic} & $ 0.47$ & $ 2.80$ & $15.13$ & $42.51$ \\
\hline
Goodfellow {\em et.al}~\cite{Goodfellow_2013_Maxout} & $ 0.45$ & $ 2.47$ & $ 9.38$ & $38.57$ \\
\hline
Lin {\em et.al}~\cite{Lin_2014_Network} & $ 0.47$ & $ 2.35$ & $ 8.81$ & $35.68$ \\
\hline
Lee {\em et.al}~\cite{Lee_2015_Deeply} & $ 0.39$ & $ 1.92$ & $ 7.97$ & $34.57$ \\
\hline
Liang {\em et.al}~\cite{Liang_2015_Recurrent} & $\mathbf{ 0.31}$ & $ 1.77$ & $ 7.09$ & $31.75$ \\
\hline
Lee {\em et.al}~\cite{Lee_2016_Generalizing} & $\mathbf{ 0.31}$ & $ 1.69$ & $ 6.05$ & $32.37$ \\
\hline\hline
{\bf BigNet} (without GNPP) & $ 0.36$ & $ 2.14$ & $ 7.80$ & $31.03$ \\
\hline
{\bf BigNet} (with GNPP) & $\mathbf{ 0.32}$ & $\mathbf{ 1.87}$ & $\mathbf{ 7.14}$ & $\mathbf{29.74}$ \\
\hline
{\bf WRN} (without GNPP) & $ 0.34$ & $ 1.77$ & $ 5.54$ & $25.52$ \\
\hline
{\bf WRN} (with GNPP) & $\mathbf{ 0.31}$ & $\mathbf{ 1.67}$ & $\mathbf{ 5.31}$ & $\mathbf{25.01}$ \\
\hline
\end{tabular}
\caption{
Comparison of the recognition error rate ($\%$) with the state-of-the-arts.
We apply data augmentation on all these datasets, but the competitors do not use it in {\bf CIFAR100}.
Without data augmentation, we report $29.92\%$ and $29.17\%$ error rates (using {\bf WRN}) without and with GNPP, respectively.
}
\label{Tab:Comparison}
\end{center}
\end{table*}
We compare our results with the state-of-the-arts in Table~\ref{Tab:Comparison}.
We add GNPP before the second and the third pooling layers for {\bf BigNet}, and the last pooling layer for {\bf WRN}.
Although the baseline is already pretty high, GNPP still improves it by a margin:
on {\bf BigNet}, the relative error rate drops are $11.11\%$, $12.62\%$, $8.46\%$ and $4.16\%$ on the four datasets, respectively.
Without complicated tricks, our results are very competitive among these recent works.
We believe that GNPP can also be applied to other powerful networks in the future.
\subsection{ImageNet Experiments}
\label{Experiments:ImageNet}
Finally, we evaluate our model on the {\bf ImageNet} large-scale visual recognition task
(the {\bf ILSVRC2012} dataset~\cite{Russakovsky_2015_ImageNet} with $1000$ categories).
We use the {\bf AlexNet} (provided by the {\bf CAFFE} library~\cite{Jia_2014_CAFFE}), abbreviated as:
\begin{spverbatim}
{C11(S4)@96-MP3(S2)}{C5(S1P2)@256-MP3(S2)}{C3(S1P1)@384}{C3(S1P1)@384}
{C3(S1P1)@256-MP3(S2)}{FC4096-D0.5}{FC4096-D0.5}{FC1000}.
\end{spverbatim}
\noindent
The input image is of size $227\times227$, randomly cropped from the original $256\times256$ image.
Following the setting of {\bf CAFFE}, a total of $450\rm{,}000$ mini-batches (approximately $90$ epochs) are used for training,
each of which has $256$ image samples, with the initial learning rate $10^{-2}$, momentum $0.9$ and weight decay $5\times10^{-4}$.
The learning rate is decreased to $1/10$ after every $100\rm{,}000$ mini-batches.
{\bf AlexNet} contains three max-pooling layers, {\em i.e.}, {\em pool-1}, {\em pool-2} and {\em pool-5}.
After individual tests, we only add GNPP before the last one ({\em pool-5}),
since adding GNPP before either {\em pool-1} or {\em pool-2} causes accuracy drop.
With the GNPP layer, the top-$1$ and top-$5$ recognition error rates are $42.16\%$ and $19.24\%$, respectively.
Comparing to the original rates ($43.19\%$ and $19.87\%$), GNPP boosts them by about $1.0\%$ and $0.6\%$, respectively.
We emphasize that the accuracy gain is not so small as it seems,
especially considering that we do not introduce extra parameters and that the overall training time is only increased by $1.29\%$.
Although GNPP is tested on {\bf AlexNet}, we believe it can be applied to other models,
such as {\bf VGGNet}~\cite{Simonyan_2015_Very}, {\bf GoogleNet}~\cite{Szegedy_2015_Going} and Deep Residual Nets~\cite{He_2015_Deep}.
\section{Benefits of GNPP}
\label{Benefits}
This section presents several discussions and diagnostic experiments
that help us understand the side benefits brought by the GNPP layer.
\subsection{Improving Internal Representation}
\label{Benefits:Representation}
\begin{figure}[t]
\begin{center}
\includegraphics[width=12.0cm]{NeuronHeatmaps.pdf}
\end{center}
\caption{
Neuron response heatmaps produced by {\bf AlexNet} and {\bf GNPPNet}.
When the background is relatively simple ({\em e.g.}, first two images), both methods work well.
On those challenging cases, GNPP produces better saliency detection results,
implying that the internal representation of CNN is improved.
}
\label{Fig:NeuronHeatmaps}
\end{figure}
Here we compare the {\em conv-5} layer of the standard {\bf AlexNet} with the corresponding layers in the {\bf GNPPNet}
(defined in Section~\ref{Experiments:ImageNet}).
That is, we compare {\bf AlexNet}'s {\em conv-5} layer with {\bf GNPPNet}'s {\em conv-5} layer and {\em GNPP-5} layer.
Each layer is a $13\times13\times256$ neuron blob corresponding to $256$ convolutional kernels.
We average over the $256$ channels and obtain a $13\times13$ heatmap.
To allow direct comparison with the input image ($227\times227$),
we diffuse each neuron response as a Gaussian distribution over its receptive field on the input image
(the same standard deviation is used on all layers).
Results are shown in Figure~\ref{Fig:NeuronHeatmaps}.
It is observed in~\cite{Wang_2015_Discovering} that
the activation patterns in higher convolutional layers correspond to mid-level parts.
The average over filters is a crude measure that some mid-level parts are detected.
Then Figure~\ref{Fig:NeuronHeatmaps} shows the spatial pattern corresponding to mid-level part detection.
First note that {\bf AlexNet}'s {\em conv-5} layer and {\bf GNPPNet}'s {\em GNPP-5} layer are broadly similar.
This is to be expected as both of them occur at corresponding places in the network architecture,
{\em i.e.}, just before the {\em pool-5} layer and the fully-connected layers.
We might think of the filter averages shown in Figure~\ref{Fig:NeuronHeatmaps}
as spatial summaries of average scores over object parts.
The higher layers in both networks combine spatial co-occurrences of parts into whole object detectors.
For example, {\em car wheels} and {\em car doors} are combined into a whole {\em car}.
Next notice that {\bf GNPPNet}'s {\em conv-5} layer is sparser and more concentrated than {\bf AlexNet}'s {\em conv-5} layer.
Broadly speaking the GNPP operation acts as a smoother
and it is the smoothed {\em conv-5} layer ({\em i.e.}, the {\em GNPP-5} layer) that resembles {\bf AlexNet}'s {\em conv-5} layer.
The difference between {\bf AlexNet}'s {\em conv-5} layer and {\bf GNPPNet}'s {\em GNPP-5} layer is subtle,
but we see that the {\em GNPP-5} layer is more diffuse corresponding to GNPP's action as local smoother.
As a result,
{\bf GNPPNet}'s {\em conv-5} layer produces better saliency detection results compared to {\bf AlexNet}'s {\em conv-5} layer.
This property can be used to extract better {\em deep features}.
We verify our hypothesis on the {\bf Caltech256} dataset~\cite{Griffin_2007_Caltech}.
$256$-dimensional feature vectors are extracted from the {\em conv-5} layer by averaging over $13\times13$ spatial locations.
The classification accuracy using the {\bf AlexNet} is $59.36\%$, and {\bf GNPPNet} improves it to $60.56\%$.
This improvement is significant given that no extra time or memory is required for feature extraction.
In summary, applying GNPP to CNN produces better internal representation.
The deep features extracted from the {\bf GNPPNet} can also benefit other vision applications,
such as image retrieval~\cite{Razavian_2014_CNN} and object detection~\cite{Girshick_2014_Rich}\cite{Girshick_2015_Fast}.
\subsection{Building Latent Connections}
\label{Benefits:LatentConnections}
We show that GNPP builds latent connections between hidden layers in the CNN model.
Consider a geometric neural phrase
${\mathcal{G}_{w,h}}={\mathbf{x}_{w,h}\cup\left\{\mathbf{x}_{w,h}^{\left(k\right)}\right\}}_{k=1}^K$.
Let $\mathcal{S}_{w,h}$ be the set of neurons in the previous layer that are connected to $\mathbf{x}_{w,h}$,
and $\mathcal{S}_{w,h}^{\left(k\right)}$ be the set connected to $\mathbf{x}_{w,h}^{\left(k\right)}$, ${k}={1,2,\ldots,K}$.
If we consider $\mathcal{G}_{w,h}$ as a {\em GNPP neuron},
then the set of neurons in the previous layer that are connected to it is
$\mathcal{S}_{w,h}\cup{\bigcup_{k=1}^{K}}\mathcal{S}_{w,h}^{\left(k\right)}$.
Thus, we are actually building latent neuron connections which do not exist in the original network.
For example, applying GNPP (type 1) before the {\em pool-5} layer of the {\bf AlexNet}
increases the number of neuron connections between {\em conv-4} and {\em conv-5}
from $149.5\mathrm{M}$ (million) to $348.9\mathrm{M}$
(on each neuron in {\em conv-5}, the number of connections to the previous layer increases from $9$ to $21$),
meanwhile the number of learnable parameters remains unchanged.
To verify the benefits of latent connections, we train another version of {\bf AlexNet}, referred to as {\bf AlexNet2},
with the difference that the number of channels on the {\em conv-5} layer increases from $256$ to $512$.
The number of neuron connections between {\em conv-4} and {\em conv-5} increases from $149.5\mathrm{M}$ to $299.0\mathrm{M}$,
comparable to $348.9\mathrm{M}$ in {\bf GNPPNet}.
{\bf AlexNet2} requires $9.97\%$ extra training time and $5.58\%$ extra GNPP memory,
while the numbers for {\bf GNPPNet} are $1.29\%$ and $2.52\%$, respectively.
{\bf AlexNet2} produces $42.45\%$ (top-$5$) and $19.47\%$ (top-$1$) recognition error rates,
which are higher than $43.19\%$ and $19.97\%$ reported by {\bf AlexNet},
but lower than $42.16\%$ and $19.24\%$ reported by {\bf GNPPNet}.
To summarize, GNPP allows latent connections to be built in an efficient manner.
\subsection{Accelerating Network Training}
\label{Benefits:Acceleration}
We show that adding GNPP layers accelerates the network training process,
since GNPP allows visual information to propagate faster, like~\cite{Srivastava_2015_Training}.
Let us investigate the case that training the $3$-layer {\bf LeNet} on the {\bf SVHN} and {\bf CIFAR} datasets.
We are interested in the following question: if the input is a $32\times32$ image,
which is the earliest layer containing a neuron able to ``see'' the entire image?
Without GNPP, we need to wait until the {\em conv-3} layer.
When GNPP is inserted before the second pooling layer, the receptive field of the neurons on the subsequent layers are increased.
Consequently, some neurons in {\em pool-2} can already ``see'' the entire image.
This allows some low-level and mid-level information ({\em e.g.}, object parts) be combined earlier.
\newcommand{5.5cm}{5.5cm}
\newcommand{0.2cm}{0.2cm}
\begin{figure}
\begin{center}
\begin{minipage}{5.5cm}
\centering
\includegraphics[width=5.5cm]{SVHNAccuracy.pdf}
\end{minipage}
\hspace{0.2cm}
\begin{minipage}{5.5cm}
\centering
\includegraphics[width=5.5cm]{CIFAR100Accuracy.pdf}
\end{minipage}
\vspace{0.2cm}
\vspace{0.2cm}
\begin{minipage}{5.5cm}
\centering
\includegraphics[width=5.5cm]{SVHNLoss.pdf}
\end{minipage}
\hspace{0.2cm}
\begin{minipage}{5.5cm}
\centering
\includegraphics[width=5.5cm]{CIFAR100Loss.pdf}
\end{minipage}
\end{center}
\caption{
Error rate and loss function curves on the {\bf SVHN} and {\bf CIFAR100} datasets.
{\bf GNPPNet} refers to the {\bf LeNet} with two GNPP layers inserted before the second and third pooling layers.
The curves in red frames indicate that {\bf GNPPNet} enjoys better convergence, {\em i.e.}, it reaches the plateau sooner.
}
\label{Fig:Curves}
\end{figure}
As a result, GNPP helps the network training process converge faster.
To verify, we plot the testing error rates and the loss function values throughout the training process.
The results on the {\bf SVHN} and {\bf CIFAR100} datasets, using the {\bf LeNet}, are shown in Figure~\ref{Fig:Curves}.
One can see that GNPP causes the error rate and loss function curves drop more quickly, especially in the early epochs.
For example, in the {\bf SVHN} dataset, the network without GNPP requires about $36\rm{,}000$ iterations to reach $6\%$ error rate,
while that with GNPP only needs about $15\rm{,}000$ iterations to get the same rate.
Meanwhile, the training process reaches plateau sooner in the GNPP-equipped networks
(see the error rate curve between $6$--$12$ epochs in {\bf SVHN}, and that between $60$--$120$ epochs in {\bf CIFAR100}).
With the help of GNPP, we can even train a network faster and obtain better performance.
The baseline error rates on {\bf SVHN} and {\bf CIFAR100}, using the {\bf LeNet}, are $4.63\%$ and $44.99\%$, respectively.
We train a GNPP-equipped {\bf LeNet} with half training epochs under each learning rate,
and obtain $3.78\%$ and $43.35\%$ error rates (the full training reports $3.55\%$ and $42.04\%$).
\section{Conclusions}
\label{Conclusions}
In this paper, we demonstrate that constructing and encoding {\em neural phrases} boost the performance of state-of-the-art CNNs.
We insert Geometric Neural Phrase Pooling (GNPP) as an intermediate layer into the network,
and show that it improves the performance of deep networks without requiring much more computational resources.
GNPP can be explained as an implicit way of modeling the spatial co-occurrence of neurons.
We also show that GNPP enjoys the advantage of improving the internal representation of CNN,
building latent connections, and speeding up the network training process.
Our work illustrates that designing deep networks benefits from prior knowledge.
In the case of GNPP, we learn that the isolated neuron responses are less reliable than the clustered ones.
We hope that other prior knowledge can be useful incorporated into the CNN architecture.
Meanwhile, other visual tasks, including detection, segmentation, {\em etc.}, may also benefit from the GNPP algorithm.
The exploration of these topics is left for future works.
\section*{Acknowledgements}
\label{Acknowledgements}
This paper is supported in part to Prof. Alan Yuille by iARPA MICrONS contract D16PC00007 and ONR N00014-12-1-0883.
This work is supported in part to Prof. Qi Tian by ARO grants W911NF-15-1-0290
and Faculty Research Gift Awards by NEC Laboratories of America and Blippar.
This work is also supported in part by National Science Foundation of China (NSFC) 61429201.
We thank Junhua Mao, Cihang Xie and Zhuotun Zhu for insightful discussions.
\bibliographystyle{splncs}
|
\section{Introduction}
A \textit{cubic fourfold} $Y$ is a smooth hypersurface of degree three in $\P^5$. The derived category of $Y$ has the following semiorthogonal decomposition
\vspace{-1 mm}
\begin{equation*}
{\textbf{D}}^b(Y) = \langle \mathscr{T}_Y, \mathscr{O}_Y, \mathscr{O}_Y(1), \mathscr{O}_Y(2) \rangle
\end{equation*}
\noindent
where $\mathscr{O}_Y(i)$ is an exceptional object for $i=0,1,2$ and $\mathscr{T}_Y$, the Kuznetsov component, is the admissible subcategory of ${\textbf{D}}^b(Y)$ right orthogonal to $\langle \mathscr{O}_Y, \mathscr{O}_Y(1), \mathscr{O}_Y(2) \rangle$. Kuznetsov proved that $\mathscr{T}_Y$ has many properties in common with the derived category of a K3 surface, and this is very related with the problem of rationality of cubic fourfolds, as described in \cite{RATPROB}. Hassett, in \cite{HASSETT}, studied the divisors $\mathscr{C}_d$ of the moduli space of all cubic fourfolds that parametrize cubic fourfolds containing a surface non homologous to a complete intersection, also called \textit{special cubic fourfolds}. The integer $d$ is the discriminant of the saturated sublattice of $H^4_{prim}(Y;\mathbb{Z})$ spanned by the square of the hyperplane class and by the class of the surface which is not homologous to a complete intersection. Huybrechts in \cite{HUYK3} studied deeper the relations between properties of the component $\mathscr{T}_Y$ and the divisors $\mathscr{C}_d$.
It is known that $\mathscr{C}_8$ parametrizes cubic fourfolds containing a plane. Let $Y$ be a generic cubic fourfold in $\mathscr{C}_8$ and let $A$ be the plane contained in $Y$. Consider the projection map to another plane $B$ in $\P^5$, disjoint from $A$. The preimage along this map of a point $b \in B$, obtained by intersecting $Y$ with a $\P^3$ spanned by $A$ and $b$, consists of the plane $A$ and, since the degree of $Y$ is three, a quadric surface. By blowing up the plane $A$ one obtains a quadric fibration $\pi: Y^+ \rightarrow B$.
The locus in $B$ which parametrizes the singular fibers is a sextic curve $C$, that turns out to be smooth if the cubic fourfold $Y$ is generic in $\mathscr{C}_8$. In the following, $C$ will be called the sextic associated to the pair $(Y,A)$; notice however that a very general $Y$ in $\mathscr{C}_8$ contains a unique plane, by \cite{VOISIN}. Since $Y$ is smooth, Proposition $1.2$ of \cite{BPrym} applied to the quadric fibration $\pi: Y^+ \rightarrow B$ ensures that this sextic exists and has at most ordinary double points. One can also relate a K3 surface $S$ to the cubic fourfold $Y$ by taking the double covering of the plane $B$ ramified along $C$.
From the homological point of view, Kuznetsov proved the following
\begin{teor}[{\cite[Theorem 4.3]{KUZNETSOV}}] \label{thm:kuznets}
Let $Y$ be a generic cubic fourfold in $\mathscr{C}_8$. Then there exists an exact equivalence of triangulated categories $\mathscr{T}_Y \cong {\textbf{D}}^b(S,\tilde{\mathscr{B}_0})$.
\end{teor}
\noindent
Here, ${\textbf{D}}^b(S,\tilde{\mathscr{B}_0})$ stands for the derived category of coherent sheaves on $S$ twisted by $\tilde{\mathscr{B}_0}$, a sheaf of Azumaya algebras with the property that the pushforward under the map of the double covering is isomorphic to the sheaf of even Clifford algebras $\mathscr{B}_0$ associated with $\pi$.
It is interesting to see how it is possible to attach a Brauer class on a smooth K3 surface to $\mathscr{B}_0$ when the sextic associated to the cubic fourfold and the plane is nodal. It means that some fibres of $\pi$ are the union of two different planes and the related quadratic form is no longer simply degenerate. These fibres correspond exactly to the nodes of the sextic. In this case the double covering $S$ is singular, and one needs to resolve the singularities in order to obtain a K3 surface $S^+$. The Clifford algebra $\mathscr{B}_0$ can still be seen as the pushforward of a certain Azumaya algebra $\tilde{\mathscr{B}_0}$ on $S$ but only on the complement of the singular locus of the sextic, and this is not sufficient to prove Theorem \ref{thm:kuznets}.
It is particularly intriguing to look for a way to make the usual theory work in this singular case as well. More precisely, is it possible to define an Azumaya algebra $\mathscr{A}$ on the K3 surface $S^+$ such that the push forward of $\mathscr{A}$ to the plane $B$ is isomorphic to the Clifford algebra $\mathscr{B}_0$? Proposition \ref{prop:final} answers this question positively and allows us to prove the following a generalization of Theorem \ref{thm:kuznets}.
\begin{teor} \label{teor:extension}
Let $Y$ be a cubic fourfold containing a plane such that the associated sextic curve $C$ on the plane $B$ is nodal. Then there exists an exact equivalence of triangulated categories $\mathscr{T}_Y \cong {\textbf{D}}^b(S^+,\mathscr{A})$, where $S^+$ is the K3 surface obtained by resolving the singularities of the double covering of $B$ ramified along $C$ and $\mathscr{A}$ is an Azumaya algebra on $S^+$ such that the pushforward to the plane $B$ is isomorphic to the Clifford algebra $\mathscr{B}_0$.
\end{teor}
A geometrical meaning to the Kuznetsov component $\mathscr{T}_Y$ has already been established in \cite{T} for cubic forufolds containing a plane and in \cite{BLMS} for all cubic fourfolds.
\textbf{The plan of the paper.} Some preliminaries concerning quadric fibrations and Azumaya algebras are given in Section \ref{Sec:prel}. The geometric context in which the construction takes place is described in Section \ref{Sec:geometriccontext}. Section \ref{Sec:nodalsextic} is devoted to finding the Azumaya algebra $\mathscr{A}$, which is done in Proposition \ref{prop:final} and to giving the proof of Theorem \ref{teor:extension}. In all the paper we will work over the field of the complex numbers.
\section{Preliminaries} \label{Sec:prel}
\subsection{Quadric fibrations and line bundle valued quadratic forms}
Here is a collection of results concerning quadric fibrations of low dimension. A good tool to deal with such objects is provided by quadratic forms with values in a line bundle. One can associate to a line bundle valued quadratic form a particular quadric fibration in such a way that some geometric properties of the fibration reflect on the quadratic form and vice versa.
\begin{defn} \label{Defn:LBVQuadraticForm}
Let $B$ be a scheme. A line bundle valued quadratic form on $B$, often simply called a quadratic form, is a triple $(\mathscr{E},q,\mathscr{L})$ where $\mathscr{E}$ is a vector bundle on $B$, $\mathscr{L}$ is a line bundle on $B$ and $q: \mathscr{E} \rightarrow \mathscr{L}$ is a morphism of sheaves such that
\begin{itemize}
\item $q(av)=a^2q(v)$ where $a$ is a section of $\mathscr{O}_B$ and $v$ is a section of $\mathscr{E}$
\item the morphism $b_q:\mathscr{E} \times \mathscr{E} \rightarrow \mathscr{L}$, defined for every $v$ and $w$ sections of $\mathscr{E}$ by
$$b_q(v,w)=q(v+w)-q(v)-q(w)$$
is $\mathscr{O}_B$-bilinear.
\end{itemize}
The dimension of the quadratic form $(\mathscr{E},q,\mathscr{L})$ is the rank of $\mathscr{E}$. A quadratic form $(\mathscr{E},q,\mathscr{L})$ is called regular if the morphism from $\mathscr{E}$ to $\sfhom(\mathscr{E},\mathscr{L})$ induced by $b_q$ is an isomorphism. Let $B$ now be a noetherian separated integral scheme; a quadratic form $(\mathscr{E},q,\mathscr{L})$ is called generically regular if the form is regular over the generic point of $B$.
\end{defn}
We will call quadric fibration a morphism $p:X \rightarrow S$ such that the fiber $X_s$ over a point $s$ is a quadric. Starting with a quadratic form $(\mathscr{E},q,\mathscr{L})$, one can consider the projection $\pi:\P_S(\mathscr{E}) \rightarrow S$, where $\P_S(\mathscr{E})$ is the projectivization of $\mathscr{E}$ on $S$. By denoting $X \subset \P_S(E)$ as the zero locus of $\sigma$, it is straightforward to prove that the restriction of the projection $\pi$ to $X$ is a quadric fibration. In particular, $X$ is called the quadric fibration associated to $(\mathscr{E},q,\mathscr{L})$. From now on we require all quadric fibrations to be flat, in the same spirit we will use only quadratic form with a flat associated quadric fibration. Auel, Bernardara and Bolognesi in \cite[Lemma 1.1.1]{ABB} proved the equivalence between Definition \ref{Defn:LBVQuadraticForm} and other definitions of line bundle valued quadratic forms. For instance, to give a quadratic form $(\mathscr{E},q,\mathscr{L})$ is equivalent to give a global section of $\sfhom(\mathscr{L}^*,S^2 \mathscr{E}^*)$ on $B$, which corresponds to a morphism of $\mathscr{O}_B$ modules $\sigma:\mathscr{L}^* \rightarrow S^2 \mathscr{E}^*$. This is useful in order to state the following
\begin{defn}[{\cite[Section 3.5]{QUADRIC}}] \label{defn:degenerationlocus}
Let $(\mathscr{E},q,\mathscr{L})$ be a generically regular quadratic form of dimension $n$ on a scheme $B$. The $d$-th degeneration locus of the quadratic form, denoted by $B_d \subset B$, is a closed subscheme defined by the following sheaf of ideals
$$\mathscr{I}_d=\operatorname{Im}(\Lambda^{n+1-d}\mathscr{E}\otimes\Lambda^{n+1-d}\mathscr{E}\otimes (\mathscr{L}^*)^{n+1-d} \xrightarrow{\Lambda^{n+1-d}\sigma} \mathscr{O}_B).$$
\end{defn}
\noindent Notice that $B_{i+1}$ is contained in $B_i$ for all positive integers $i$, and that $B_1$ is a divisor on $B$ that is called the discriminant divisor. Geometrically, the $B_1$ parametrizes the singular fibres of the quadric fibration $(\mathscr{E},q,\mathscr{L})$.
The notion of simple degeneration will be useful to describe the situation arising from a generic cubic fourfold containing a plane. The exact definition in terms of quadratic forms can be found in \cite[Section 1.1]{ABB}, whereas a property of simply degenerate quadratic forms will be taken as a definition in this paper: a quadratic form is simply degenerate if and only if its second degeneracy locus $B_2$ is empty. In general ${\operatorname{Sing}}(B_i) \supseteq B_{i+1}$, so $B_2$ being empty does not imply $B_1$ to be smooth, as can be seen for instance in \cite{APS} Proposition 1.5.
\subsection{Clifford algebra and Azumaya algebra}
A good introduction to Clifford algebras is provided by \cite{KNUS}, where Clifford algebras are defined in the context of modules over commutative rings. In the same spirit, one can associate a sheaf of $\mathbb{Z}$-graded algebras to each line bundle valued quadratic form by following a construction proposed in \cite{BK}. The subalgebra of degree zero, denoted by $\mathscr{B}_0$, has a structure of a sheaf of algebras, and is called the even (part of the) Clifford algebra, see \cite{A}, Section 1.8. These spaces play an important geometric role, described for example in \cite{KUZNETSOV}. The construction proposed in \cite{ABB} makes possible to define directly the even Clifford algebra. Starting from a quadratic form on a scheme $B$, not necessarily regular, consider the tensor algebra $T(\mathscr{E} \otimes \mathscr{E} \otimes \mathscr{L}^*)$, and define the ideals
\begin{align*}
J_1 &= (v \otimes v \otimes f - f(q(v)))\\
J_2 &=(u \otimes v \otimes f \otimes v \otimes w \otimes g - f(q(v))u\otimes w \otimes g)
\end{align*}
where $u,v,w$ are sections of $\mathscr{E}$ and $f,g$ are sections of $\mathscr{L}^*$. The even Clifford algebra of the quadratic form $(\mathscr{E},q,\mathscr{L})$ is defined by the quotient
$$\mathscr{B}_0(\mathscr{E},q,\mathscr{L}) := T(\mathscr{E} \otimes \mathscr{E} \otimes \mathscr{L}^*)/(J_1 + J_2)$$
This is not the only way to construct the even Clifford algebra of a quadratic form; in \cite{QUADRIC}, Kuznetsov defines the sheaf of Clifford algebras, and the description of $\mathscr{B}_0$ and $\mathscr{B}_1$ in terms of the initial quadratic form turns out to be
$$\mathscr{B}_0 \cong \mathscr{O}_B \oplus (\wedge^2 \mathscr{E} \otimes \mathscr{L}^*) \oplus (\wedge^4 \mathscr{E} \otimes (\mathscr{L}^*)^2) \oplus \cdots$$
$$\mathscr{B}_1 \cong \mathscr{E} \oplus (\wedge^3 \mathscr{E} \otimes \mathscr{L}^*) \oplus (\wedge^5 \mathscr{E} \otimes (\mathscr{L}^*)^2) \oplus \cdots$$
as $\mathscr{O}_B$-modules.
Other ways to define the even Clifford algebra are the splitting construction and the gluing construction; see \cite[Appendix A]{ABB}, for further references. These ways turn out to be all equivalent, since the Clifford algebra of a quadratic form over a ring is unique. The splitting construction can be used to describe the even Clifford algebra $\mathscr{B}_0$ as a functor.
\begin{defn} \label{Defn:AzumayaAlgebra}
An algebra $\mathscr{A}$ over a local commutative ring $R$ is said to be Azumaya if $\mathscr{A}$ is a free $R$-algebra of finite rank and such that the map
\begin{eqnarray*}
\mathscr{A} \otimes_R \mathscr{A}^{op} &\rightarrow& {\operatorname{End}}_R(\mathscr{A})\\
a \otimes b &\mapsto& (x \mapsto axb)
\end{eqnarray*}
is an isomorphism.
If $X$ is a scheme, an $\mathscr{O}_X$-algebra $\mathscr{A}$ is Azumaya if it is coherent as an $\mathscr{O}_X$-module and, for every closed point $x$ of $X$, $\mathscr{A}_x$ is Azumaya over $\mathscr{O}_{X,x}$.
\end{defn}
The even Clifford algebra of a generically regular quadratic form with even dimension can be described in terms of an Azumaya algebra.
\begin{prop}[{\cite[Proposition 3.13]{QUADRIC}}] \label{Prop:AzumayaAlgebraEvenFibration}
Let $(\mathscr{E},q,\mathscr{L})$ be a generically regular quadratic form on $B$ of even dimension, and let $f:S \rightarrow B$ be the double cover $S$ of $B$ ramified at the discriminant locus $B_1$, determined by the centre of the Clifford algebra $\mathscr{B}_0$. Then there exists a sheaf of algebras $\tilde{\mathscr{B}}_0$ on $S$ such that $f_*(\tilde{\mathscr{B}}_0) \cong \mathscr{B}_0:=\mathscr{B}_0(\mathscr{E},q,\mathscr{L})$ and the functor
$$f_*:{\operatorname{Coh}}(S,\tilde{\mathscr{B}}_0) \xrightarrow{\sim} {\operatorname{Coh}}(B,\mathscr{B}_0)$$
is an equivalence of categories. Moreover, the restriction of $\tilde{\mathscr{B}}_0$ to the complement of $f^{-1}(B_2) \subset S$ is a sheaf of Azumaya algebras on $S$.
\end{prop}
Some recent results involving cubic fourfolds containing a plane that make use of the tool of quadric fibrations can be found in \cite{ABBV}, \cite{AFIB}, \cite{G}, \cite{LMS}, \cite{MS}, \cite{GO} and \cite{T}.
\subsection{Noncommutative varieties and twisted derived categories}
Let $X$ be an algebraic variety and consider a sheaf of $\mathscr{O}_X$-algebras $\mathscr{B}$ of finite rank as a $\mathscr{O}_X$-module. The pair $(X,\mathscr{B})$ is called a noncommutative variety.
A morphism $f:=(f_0, f_{\operatorname{alg}})$ between $(X_1,\mathscr{B}_1)$ and $(X_2,\mathscr{B}_2)$ is given by a morphism of algebraic varieties $f_0:X_1 \rightarrow X_2$ and a morphism of $\mathscr{O}_{X_1}$-algebras $f_{\operatorname{alg}}:f_0^*\mathscr{B}_2 \rightarrow \mathscr{B}_1$. The morphism $f$ is called strict if $\mathscr{B}_1 \cong f_0^*\mathscr{B}_2$, and is called an extension if $X_1$ is isomorphic to $X_2$, and $f_0$ is the identity. Since the pullback of an Azumaya algebra is still Azumaya, see \cite{KNUS} III.5.1, one can define also Azumaya varieties as noncommutative varieties $(X,\mathscr{B})$ in which the algebra $\mathscr{B}$ is Azumaya.
Given a noncommutative variety $(X,\mathscr{B})$, the so-called twisted derived category ${\textbf{D}}^b(X,\mathscr{B})$ can be defined in the usual way starting with the category of coherent sheaves of right $\mathscr{B}$-modules on $X$ as objects. Some background on noncommutative varieties can be found in \cite{QUADRIC}, and an introduction to Azumaya varieties together with all the definitions of functors involving twisted derived categories can be found in \cite{HYP}.
\begin{prop} \label{prop:pullbackazumvar}
Let $(X,\mathscr{B}_X) \xrightarrow{f} (S,\mathscr{B}_{S})$, and $(Y,\mathscr{B}_Y) \xrightarrow{g} (S,\mathscr{B}_{S})$ be morphisms of noncommutative varieties. Consider the following base change diagram
\begin{equation} \label{dgr:basechange}
\xymatrix{
X \times_S Y\ar[d]^{q_0}\ar[r]^-{p_0}\pullbackcorner & X\ar[d]^{f_0} \\
Y\ar[r]^{g_0} & S
}
\end{equation}
If $f$ is strict, then $(X \times_S Y,q_0^*(\mathscr{B}_Y))$ is a fibre product of $(X,\mathscr{B}_X)$ and $(Y,\mathscr{B}_Y)$ over $(S,\mathscr{B}_{S})$.
If $g$ is strict, then $(X \times_S Y,p_0^*(\mathscr{B}_X))$ is a fibre product of $(X,\mathscr{B}_X)$ and $(Y,\mathscr{B}_Y)$ over $(S,\mathscr{B}_{S})$.
Here the fibre product of noncommutative varieties is defined as in \cite{HYP}, Lemma 10.37.
\end{prop}
\begin{proof}
The proof follows the same lines of Lemma 10.37 in \cite{HYP}, which does not actually require the hypothesis of the algebras being Azumaya.
Assume $f$ to be strict, that is $\mathscr{B}_X = f_0^*\mathscr{B}_S$. First define the morphisms $q:=(q_0,q_{\operatorname{alg}})$ and $p:=(p_0,p_{\operatorname{alg}})$ that makes the following diagram commute
\begin{equation} \label{dgr:basechange2}
\xymatrix{
(X \times_S Y,q_0^*\mathscr{B}_Y)\ar[d]^{q}\ar[r]^-{p} & (X,\mathscr{B}_X)\ar[d]^{f} \\
(Y,\mathscr{B}_Y)\ar[r]^{g} & (S,\mathscr{B}_S)
}
\end{equation}
The map $q_{\operatorname{alg}}:q_0^* \mathscr{B}_Y \rightarrow q_0^* \mathscr{B}_Y$ is the identity. It remains to define $p_{\operatorname{alg}}:p_0^* \mathscr{B}_X \rightarrow q_0^* \mathscr{B}_Y$. Since $f$ is strict $p_0^* \mathscr{B}_X$ is equal to $p_0^* f_0^* \mathscr{B}_S$, and since Square (\ref{dgr:basechange}) is a base change, $p_0^* f_0^* \mathscr{B}_S$ is equal to $q_0^* g_0^* \mathscr{B}_S$. One can then choose as $p_{\operatorname{alg}}$ the morphism induced by $g_{\operatorname{alg}}$. This choice makes Square (\ref{dgr:basechange2}) commute.
Let now $u:(Z,\mathscr{B}_Z) \rightarrow (X,\mathscr{B}_X)$ and $v:(Z,\mathscr{B}_Z) \rightarrow (Y,\mathscr{B}_Y)$ be morphisms of noncommutative varieties such that $f \circ u = g \circ v$.
At the level of varieties, the base change (\ref{dgr:basechange}) gives the existence of a morphism $t_0 : Z \rightarrow X \times_S Y$ such that $p_0 \circ t_0 = u_0$ and $q_0 \circ t_0 = v_0$. The morphism $t_{\operatorname{alg}}: t_0^* q_0^* \mathscr{B}_Y \rightarrow \mathscr{B}_Z$ will be the one induced by $v_{\operatorname{alg}}:v_0^*\mathscr{B}_Y \rightarrow \mathscr{B}_Z$. The two properties $v=q \circ t$ and $u=q \circ p$ follows by the definitions of $q$ and $p$.
The same proof holds in the case $g$ is strict.
\end{proof}
The following definition is the analogous for our case of \cite[Definition 2.18]{HYP}.
\begin{defn} \label{Defn:ExactCartesian}
Let $(X,\mathscr{B}_X) \xrightarrow{f} (S,\mathscr{B}_{S})$, and $(Y,\mathscr{B}_Y) \xrightarrow{g} (S,\mathscr{B}_{S})$ be morphisms of noncommutative varieties. If either $f$ or $g$ is strict, then Proposition \ref{prop:pullbackazumvar} ensures that the fibre product $(X \times_S Y, \mathscr{B}_{X \times_S Y})$ is defined. The square
\begin{equation} \label{Sq:EC}
\xymatrix{
X \times_S Y \ar[r]^q\ar[d]^p & Y\ar[d]^g\\
X\ar[r]^f & S
}
\end{equation} is said to be exact cartesian if the natural morphism of functors
$$g^*f_* \to q_*p^*:{\textbf{D}}^-(X,\mathscr{B}_X) \to {\textbf{D}}^-(Y,\mathscr{B}_Y)$$
is an isomorphism.
\end{defn}
The following lemma follows the same lines of \cite[Corollary 2.27]{HYP}, stated in the context of Azumaya varieties. The difference here is that $\mathscr{B}_0$ and $\bar{\mathscr{B}_0}$ are only Clifford algebras and not Azumaya. For the convenience of the reader we are rewriting the proof.
\begin{lemma} \label{Lemma:ExactCartesian}
In the same notation of Definition \ref{Defn:ExactCartesian}, assume $g$ to be a strict closed embedding, $Y$ a locally complete intersection in $S$ and both $S$ and $X$ to be cohen-Macaulay. If the codimension on $X$ of $X \times_S Y$ is equal to the codimension on $S$ of $Y$, then Square (\ref{Sq:EC}) is exact cartesian.
\end{lemma}
\begin{proof}
Let us first remark the content of \cite[Lemma 2.26]{HYP}. Since the map $g$ is finite, Square (\ref{Sq:EC}) is exact cartesian if the morphism
\begin{equation} \label{Eqn:morphlemmaEC}
f^*g_* \mathscr{B}_Y \to p_*q^* \mathscr{B}_Y=p_* \mathscr{B}_{X \times_S Y}
\end{equation}
is an isomorphism. For any $F \in {\textbf{D}}^-(X,\mathscr{B}_X)$ we get
$$g_*g^*f_*(F) \cong f_*(F) \otimes_{\mathscr{B}_S} g_*\mathscr{B}_Y \cong f_*(F \otimes_{\mathscr{B}_X} p_* \mathscr{B}_{X \times_S Y}) \cong f_* p_* p^* (F) \cong g_*q_*p^*(F).$$
This isomorphism is induced by the pushforward $g_*$ applied to the morphism $g^*f_* \to q_*p^*$, which is a morphism of kernel functors:
$$f_* \cong \Phi_{(\Gamma_{f})_*\mathscr{B}_X} \qquad g^* \cong \Phi_{(\Gamma_g)_*\mathscr{B}_Y} \qquad q_*p^* \cong \Phi_{\mathscr{B}_{X \times_S Y}},$$
where $\Gamma_f: X \to X \times S$ and $\Gamma_g: Y \to S \times Y$ are the graphs of $f$ and $g$ respectively. We can then apply the result of \cite[Lemma 2.8,b]{HYP}, which only involves the varieties and not the sheaf of algebras, to get the desired isomorphism $g^*f_* \cong q_*p^*$.
We are left to check that the homomorphism (\ref{Eqn:morphlemmaEC}) is an isomorphism. Since the claim is local in $S$, we may assume $Y$ to be the zero locus of a regular section $s$ of a vector bundle $\mathscr{V}$ on $S$ having rank equal to the codimension of $Y$ in $S$. Recall that $\operatorname{Kosz}_S(s)$ is the Koszul complex of the section $s$:
$$\operatorname{Kosz}_S(s):=\{0 \to \Lambda^{\operatorname{top}} \mathscr{V}^* \otimes_{\mathscr{O}_X} \mathscr{B}_X \xrightarrow{s} \ldots \xrightarrow{s} \mathscr{V}^* \otimes_{\mathscr{O}_X} \mathscr{B}_X \xrightarrow{s} \mathscr{B}_X \to 0\}$$
Since $s$ is regular and $X$ is Cohen-Macaulay we have $g_* \mathscr{B}_Y \cong \operatorname{Kosz}_S(s)$.
The zero locus of $f^*s$ on $X$ is the fibre product $X \times S Y$, but by the hypothesis on the codimension on $X$ of $X \times_S Y$, we have $\operatorname{codim}_X(X \otimes_S Y) = {\operatorname{rk}} f^* \mathscr{V}$. Therefore the section $f^*s$ is regular and we get an isomorphism and so $f^*g_*\mathscr{B}_Y \cong \operatorname{Kosz}_X(f^*s)$. This leads to the isomorphism $\operatorname{Kosz}_X(f^*s) \cong p_* \mathscr{B}_{X \times_S Y}$ which concludes the proof.
\end{proof}
\section{The geometric context} \label{Sec:geometriccontext}
Let $Y \subset \P(V)$ be a hypersurface of degree $3$ containing a projective space $A$. Let $V$ be a vector space of dimension $n+1$ and $A$ be $\P(A_0)$ for a vector subspace $A_0$ of $V$. Let $B_0$ be the quotient space $V / A_0$, and $B:=\P(B_0)$. As described in the introduction, it is possible to obtain a quadric fibration $Y^+\rightarrow B$ that is associated to a line bundle valued quadratic form $(\mathscr{E},q,\mathscr{L})$. By using the same notation of \cite{KUZNETSOV}, the vector bundle $\mathscr{E}$ is $A_0 \otimes \mathscr{O}_B \oplus \mathscr{O}_B(-1)$, the line bundle $\mathscr{L}$ is $\mathscr{O}_B(1)$ and the map $q$ depends on the equations of $Y$.
\begin{oss}
In the case of a generic cubic fourfold $Y$ in $\mathscr{C}_8$, this quadratic form turns out to be simply degenerate, and then, since $Y$ is also smooth, the vanishing of $B_2$ implies that the sextic curve $C$ on $B$, that coincides with the discriminant divisor $B_1$ of the quadratic form, is also smooth. Hence, one obtains a sheaf of Azumaya algebras on the double cover of the plane $B$ ramified along $C$. Proposition \ref{Prop:AzumayaAlgebraEvenFibration} can also be applied when, for a non generic $Y$, the singular locus $B_2$ of the sextic $C$ turns out to be non empty. The sheaf of algebras obtained is Azumaya only on the complement of $f^{-1}(B_2) \subset S$.
\end{oss}
One can compute the symmetric product of $\mathscr{E}^*$ tensored by $\mathscr{L}$, according to one of the equivalent definitions of a line bundle valued quadratic form.
\begin{align*}
S^2\mathscr{E}^* \otimes \mathscr{L}^* &= S^2\left(A_0^* \otimes \mathscr{O}_B \oplus \mathscr{O}_B(1)\right) \otimes \mathscr{O}(1)\\
&= S^2\left(A_0^*\right) \otimes \mathscr{O}(1) \oplus \left(A_0^* \otimes \mathscr{O}(2)\right) \oplus \mathscr{O}(3)
\end{align*}
A section of this bundle is given by a section of $S^2 A_0^*\otimes \mathscr{O}(1)$, a section of $A_0^* \otimes \mathscr{O}(2)$ and a section of $\mathscr{O}(3)$. The same setting can be described in coordinates as follows. Let the ambient space be $\P^n=\P^n(x_0: \cdots : x_b : y_0 : \cdots : y_a)$, with $a+b+1=n$. The space $A$ of dimension $a$ will be $\P^a(y_0: \cdots : y_a)=\{x_0= \cdots = x_b = 0\}$ and the space $B$ will be $\P^b(x_0: \cdots : x_b)=\{y_0= \cdots = y_a = 0\}$. Notice that $A \cap B = \emptyset$.
Let $\pi$ be the projection map $\P^n \dashrightarrow B$ from the space $A$. One can assume $Y:= \{F=0\}$,
\begin{equation*}
F:= \sum_{i,j=0}^a l_{ij}y_iy_j + 2\sum_{k=0}^a q_k y_k + f
\end{equation*}
where $l_{ij}$, $q_k$ and $f$ are polynomials in $x_0, \cdots x_b$ of degree one, two and three, respectively, and $l_{ij}=l_{ji}$. One can arrange these polynomials in the following symmetric matrix
$$M:=\left( \begin{array}{ccc|c}
l_{00} & \cdots & l_{0a} & q_0\\
\vdots & \ddots & \vdots & \vdots\\
l_{a0} & \cdots & l_{aa} & q_a\\
\hline
q_0 & \cdots & q_a & f\\
\end{array}\right)$$
The discriminant divisor of the quadric fibration $Y^+ \to B$ coincides with the determinant of $M$, which is a hypersurface of degree $a+4$ in $B$.
The matrix $M$ can be seen as an injective map between vector bundles, part of the following exact sequence
\begin{equation} \label{Eqn:MapM}
0 \rightarrow \mathscr{O}_{\P^b}(-2)^{a+1} \oplus \mathscr{O}_{\P^b}(-3) \xrightarrow{M} \mathscr{O}_{\P^b}(-1)^{a+1} \oplus \mathscr{O}_{\P^b} \rightarrow \mathscr{F} \rightarrow 0
\end{equation}
The following proposition is just a recap of \cite{BDet} and relates the matrix $M$ with a line bundle supported on the hypersurface $\{\det M = 0\}$.
\begin{prop}
The cokernel $\mathscr{F}$ of the map induced by $M$ in (\ref{Eqn:MapM}) is an ACM sheaf supported on the hypersurface $C = \{\det M = 0\}$. If $C$ is smooth, $\mathscr{F}$ is a line bundle.
\end{prop}
\begin{proof}
This is just a special case of Theorem A and Corollary $1.8$ in \cite{BDet}.
\end{proof}
In the case of a cubic fourfold $Y$ containing a plane, Beauville and Voisin in \cite{BDet}, \cite{BPrym} and \cite{VOISIN} proved that there is a correspondence between $Y$ and the curve $C$ of degree six defined by $\{\det(M)=0\}$ in the plane $B$, together with a theta-characteristic on $C$. The singular case was studied by Stellari in \cite{STHETA}. Notice that, in coordinates, this corresponds to the case $n=5$, $a=b=2$.
Going back to the derived categories, viewing the cubic fourfold as a quadric fibration is the key for the following proposition, which relates the Kuznetsov component of ${\textbf{D}}^b(Y)$ with a twisted derived category. Notice that since $Y$ is smooth, containing a linear subspace half-dimensional, then the arising quadric fibration will be flat. One can perform the same construction carried out in the introduction obtaining a flat quadric fibration with base a plane $B$ disjoint from $A$; we recall the following
\begin{prop}{\cite[Theorem 4.3]{KUZNETSOV}} \label{prop:stepone}
Let $Y$ be a cubic fourfold containing a plane $A$. There is an equivalence
$$\mathscr{T}_Y \cong {\textbf{D}}^b(B,\mathscr{B}_0)$$
where $\mathscr{T}_Y$ is the Kuznetsov component of ${\textbf{D}}^b(Y)$ and ${\textbf{D}}^b(B,\mathscr{B}_0)$ is the derived category of $B$ twisted by the Clifford algebra $\mathscr{B}_0$ related to the quadric fibration.
\end{prop}
\section{Cubic fourfolds in $\mathscr{C}_8$ with an associated nodal sextic} \label{Sec:nodalsextic}
Let's now consider a cubic fourfold $Y$ with an associated sextic $C$ that is nodal. This hypothesis makes $Y$ non generic in $\mathscr{C}_8$. Since $Y$ is smooth, the fibres over the nodes of the sextic are double planes, see for instance \cite{BPrym}. This is equivalent to fact that $B_2={\operatorname{Sing}}(B_1)$ in terms of the degerenacy locus of the associated line bundle valued quadratic form.
\begin{oss} \label{Rmk:FormSingular}
One can encode the information about the singularity of the sextic in the equation of the cubic fourfold. Up to a change of coordinates, one can assume one node of the sextic to be the point $x:=(1:0:0) \in \P^2(x_0:x_1:x_2)=:B$. The fibre over $x$ is the union of the plane $A$ and two other planes in the space generated by $A$ and $x$, that is $\P^3(x_0:y_0:y_1:y_2)$. One can also assume that, in this $\P^3$, the other two planes have equations $y_1=0$ and $y_2=0$. Hence, the equation of the cubic fourfold $Y$ is $x_0 y_1 y_2 + x_1 k_1 + x_2 k_2 =0$ where $k_1$ and $k_2$ are quadratic polynomials in $x_i$ and $y_i$.
\end{oss}
This geometric picture can be thought as a hyperplane section of a similar picture in $\P^6(x_0:x_1:x_2:x_3:y_0:y_1:y_2)$. Let $\bar{Y}$ be the cubic fivefold defined by the equation $F + x_3 \bar{F}=0$, where $\bar{F}$ is a homogeneous polynomial of degree $2$. $\bar{Y}$ still contains the plane $A$, which in $\P^6$ has equations $\{x_0 = x_1 = x_2 = x_3 = 0\}$ and one can project onto the $\P^3$ defined by $\bar{B} := \{y_0=y_1=y_2=0\}$. Exploiting the construction of $M$ of the previous section, $\bar{Y}$ is given by the equation
\begin{equation*}
\sum_{i,j=0,1,2} l_{ij}y_iy_j + 2\sum_{k=0,1,2} q_k y_k + f + x_3(\sum_{i,j=0,1,2} \bar{l}_{ij}y_iy_j + 2\sum_{k=0,1,2} \bar{q}_k y_k + \bar{f})=0
\end{equation*}
where $\bar{l}_{ij}$, the $\bar{q}_i$ and $\bar{f}$ are polynomials in $x_0$, $x_1$, $x_2$, $x_3$ of degree zero, one and two, respectively, and $\bar{l}_{ij}=\bar{l}_{ji}$.
One can define a symmetric matrix $\bar{M}$, as before:
\begin{equation} \label{Eqn:matrixform}
\left( \begin{array}{c|c}
l_{ij} & q_k\\
\hline
q_k & f\\
\end{array}\right)
+x_3 \left( \begin{array}{c|c}
\bar{l}_{ij} & \bar{q}_k\\
\hline
\bar{q}_k & \bar{f}\\
\end{array}\right)
\end{equation}
\noindent Taking the determinant of this matrix gives rise to a sextic surface $\bar{C}$ in $\P^3(x_0:x_1:x_2:x_3)$.
The idea is to exploit the work of Kuznetsov \cite{K2} to obtain an Azumaya algebra related to the Clifford algebra $\mathscr{B}_0$. The first step consists in proving that is it always possible to find a cubic fivefold as described above with the singularities of the sextic surface $\bar{C}$ being at most isolated nodes. The situation can be reformulated in terms of degeneracy locus of maps between vector bundles. Recall the following
\begin{defn}\label{Defn:DegeneracyLocus}
Let $\mathscr{E}$ and $\mathscr{F}$ be vector bundles on a projective variety $X$, $\phi: \mathscr{E} \rightarrow \mathscr{F}$ a morphism and $k$ a positive integer. Then,
$$D_k(\phi)=\{x \in X \text{ s.t. } {\operatorname{rk}}(\phi_x) \leq k\}$$
is called the $k$-degeneracy locus of $\phi$.
\end{defn}
As in Section \ref{Sec:geometriccontext}, but now for the case of $B = \P^3$, let $\mathscr{E}$ be the vector bundle $3\mathscr{O}_{\P^3} \oplus \mathscr{O}_{\P^3}(-1)$ on $\P^3$. As seen in the previous section, a matrix of homogeneous forms on $\P^3$ with the same degrees as the matrix in Equation (\ref{Eqn:matrixform}) defines a symmetric map $\phi: \mathscr{E} \rightarrow \mathscr{E}^\star(1)$ as in Equation (\ref{Eqn:MapM}), after tensoring the exact sequence with the line bundle $\mathscr{O}_{\P^3}(2)$. This map can be seen as a section of the bundle $S^2 \mathscr{E} \otimes \mathscr{O}_{\P^3}(1)$ and its symmetrical degeneracy locus $D_3(\phi)$ coincides with the locus of $\P^3$ where the determinant of the matrix (\ref{Eqn:matrixform}) is equal to zero, that is the sextic surface $\bar{C}$.
The choice of the second matrix in (\ref{Eqn:matrixform}) defines a linear system $\mathscr{T}$ of sections of $S^2 \mathscr{E} \otimes \mathscr{O}_{\P^3}(1)$.
\begin{oss}
Let $\mathscr{E}=3\mathscr{O}_{\P^3} \oplus \mathscr{O}_{\P^3}(-1)$ as above. The degeneracy locus $D_3(\phi)$, of Definition \ref{Defn:DegeneracyLocus} coincides with the first degeneration locus of the line bundle valued quadratic form of Definition \ref{defn:degenerationlocus}, $B_{1}(\sigma)$, where $\sigma: \mathscr{O}_{\P^3}(-1) \rightarrow S^2 \mathscr{E}$ is a map equivalent to $\phi$.
\end{oss}
The next lemma, following \cite{OTT}, holds in general when $X$ is a projective variety.
\begin{lemma} \label{Lemma:SymmetricCodimension}
Let $\mathscr{E}$ be a vector bundle of rank $n$, $\mathscr{L}$ be a line bundle on $X$ and let $\phi: \mathscr{E} \rightarrow \mathscr{E}^\star \otimes \mathscr{L}$ be a generic symmetric morphism.
If $S^2 \mathscr{E}^\star \otimes \mathscr{L}$ is globally generated, then $D_k(\phi)$ is empty or of codimension $\binom{n-k+1}{2}$. Moreover, ${\operatorname{Sing}}(D_k(\phi))$ is contained in $D_{k-1}(\phi)$.
\end{lemma}
\begin{proof}
The exact sequence $H^0(S^2 \mathscr{E}^\star \otimes \mathscr{L}) \otimes \mathscr{O}_X \rightarrow S^2 \mathscr{E}^\star \otimes \mathscr{L} \rightarrow 0$
induces a projection that is everywhere of maximal rank
$$X \times H^0(S^2 \mathscr{E}^\star \otimes \mathscr{L}) \xrightarrow{p} S^2 \mathscr{E}^\star \otimes \mathscr{L}$$
One can define the variety $\Sigma_k$ inside the total space of $S^2 \mathscr{E}^\star \otimes \mathscr{L}$ that is composed on each fibre of the $n \times n$ matrices with complex coefficients with rank less or equal than $k$. $\Sigma_k$ has codimension equal to $\binom{n-k+1}{2}$.
Now consider the following diagram where $Z$ is the preimage of $\Sigma_k$ by $p$.
$$\xymatrix{
& Z \ar[r]\ar[d]\ar[dl]_q & \Sigma_k\ar@{^{(}->}[d]\\
H^0(S^2 \mathscr{E} \otimes L) & X \times H^0(S^2 \mathscr{E} \otimes L) \ar[r]^-p\ar[l] & S^2 \mathscr{E} \otimes L
}$$
$Z$ is composed of the pairs $(x,\phi)$ where $x$ is a point of $X$, $\phi: \mathscr{E} \rightarrow \mathscr{E}^\star \otimes \mathscr{L}$ is a symmetric map and the rank of the induced map $\phi_x$ is less or equal than $k-1$. Then, the preimage by $q$ of a generic element $\phi_0$ in $H^0(S^2 \mathscr{E} \otimes L)$ coincides with $\{(x,\phi_0) \text{ s.t. } {\operatorname{rk}}(\phi_0|_x) \leq k-1\}$, which is exactly the definition of $D_k(\phi_0)$.
Notice that ${\operatorname{Sing}}(Z) = p^{-1} {\operatorname{Sing}}(\Sigma_k)$ thus, by restricting $p$, one obtains
$$Z \smallsetminus {\operatorname{Sing}}(Z) \xrightarrow{p|_{Z \smallsetminus {\operatorname{Sing}}(Z)}} H^0(S^2 \mathscr{E} \otimes L).$$
If the image of $p|_{Z \smallsetminus {\operatorname{Sing}}(Z)}$ is dense, then $D_k(\phi)$ is smooth by the generic smoothness theorem, and so ${\operatorname{Sing}}(D_k(\phi)) \subset D_{k-1}(\phi)$. If the image is not dense, then $D_k(\phi)$ is empty for generic $\phi$.
\end{proof}
The following proposition ensures that for any cubic fourfold $Y$ with an associated nodal sextic curve $C$ is it possible to find a smooth cubic fivefold $\bar{Y}$ with the associated sextic surface $\bar{C}$ having singularities of codimension $3$ in $\P^3$.
\begin{prop}\label{prop:LocusCodThree}
Let $\mathscr{F}$ be the vector bundle $S^2 \mathscr{E} \otimes \mathscr{O}_{\P^3}(1)$ on $\P^3$, where $\mathscr{E}$ is $3\mathscr{O}_{\P^3} \oplus \mathscr{O}_{\P^3}(-1)$. Let $\mathscr{T}$ be the linear system of sections of $\mathscr{F}$ defined by the choice of the second matrix in (\ref{Eqn:matrixform}). Then a generic section of the linear system $\mathscr{T}$ defines a smooth cubic fivefold and the degeneracy locus $D_2$ of the morphism corresponding to this section is of codimension $3$ in $\P^3$. Moreover, $D_2$ is composed of a finite number of ordinary double points and there is a bijection between $D_2 \cap \{x_3 = 0\}$ and the nodes of the sextic curve $C$.
\end{prop}
\begin{proof}
The generic section of $\mathscr{T}$ defines a smooth fivefold. Notice that the expected codimension of the $2-$degeneracy locus of a section of $\mathscr{F}$ is $\binom{4-2+1}{2}=3$.
Let $x$ be a point in $\P^3 \smallsetminus \{x_3 = 0\}$. The stalk $\mathscr{F}_x$ is globally generated by the sections in $\mathscr{T}$ because $\bar{q}_k$ contains the monomial $\alpha_k x_3$ and $f$ contains the monomial $\beta x_3^2$, $\alpha_k$ and $\beta$ in $\mathbb{C}$. The matrix of complex numbers obtained after evaluating the second matrix in (\ref{Eqn:matrixform}) in the stalk $\mathscr{F}_x$ is generic, since $\bar{l}_{ij}, \alpha_k$ and $\beta$ are arbitrarily chosen. This proves that $S^2 \mathscr{F} \otimes \mathscr{O}(3)$ is globally generated by sections of $\mathscr{T}$ where $x_3\neq 0$.
Then, it is possible to apply Lemma \ref{Lemma:SymmetricCodimension} to the restriction of the bundles to $\P^3 \smallsetminus \{x_3 = 0\}$.
It remains to prove that $D_2 \cap \{x_3=0\}$ has codimension three, but this is a direct consequence of the setting of the problem, in which the sextic curve, which coincides with $D_2 \cap \{x_3=0\}$, has at most singularities of codimension $2$ in $\P^2$.
Since $D_2$ is of dimension $0$ in $\P^3$ then it must be finite. Let $x$ be a point in $D_2$. Proving that $x$ is an ordinary double point of $\bar{C}$ is the same as proving that the fibre over $x$ is composed of $A$ and the singular quadric given by two intersecting planes. Consider the hyperplane section of $\P^6$ given by a hyperplane $H$ containing $A$ and $x$. This gives rise to a cubic fourfold $Y_H$ containing $A$ with a projection to the plane $B_H := \bar{B} \cap H$ still containing the point $x$. By construction the fibre over $x$ is the same, and then the result follows by Proposition $1.2$ of \cite{BPrym}. The same proposition ensures that also the curve $C$ has only nodes. In both cases, each one is determined by the fact that the quadric fibration has a fibre that consists of the union of two planes, and this is sufficient to conclude the proof of the last part of the proposition.
\end{proof}
\begin{cor} \label{cor:pullbackb0}
At the level of noncommutative varieties one has
$$(\P^2,\mathscr{B}_0) = (\P^3,\bar{\mathscr{B}_0}) \times_{\bar{B}} B,$$
where $\bar{\mathscr{B}_0}$ is the even Clifford algebra of the quadratic form over $\P^3$.
\end{cor}
\begin{proof}
By Proposition \ref{prop:LocusCodThree}, the quadric fibration for the cubic fourfold $Y$ is obtained from the one of the cubic fivefold $\bar{Y}$ by base change with respect to the embedding $B \xrightarrow{i} \bar{B}$, where $B$ is the plane $\{x_3=0\}$ in $\P^3$. It follows that $\mathscr{B}_0$ is the pullback along $i$ of $\bar{\mathscr{B}_0}$, and this concludes the proof.
\end{proof}
Let $Y$ be a cubic fourfold containing a plane whose associated sextic curve $C$ is nodal. Let $S$ be the (singular) double cover of $B$ ramified along $C$ and $S^+$ be the minimal resolution of the singularities of $S$. Let $\mathscr{B}_0$ be the Clifford algebra on $B$ of the quadric fibration associated to $Y$
\begin{prop} \label{prop:final}
There exists an Azumaya algebra $\mathscr{A}$ defined on $S^+$ such that the pushforward to $B$ is isomorphic to $\mathscr{B}_0$.
\end{prop}
\begin{proof}
Proposition \ref{prop:LocusCodThree} guarantees the existence of a cubic fivefold $\bar{Y} \subset \P^6$ such that $Y$ is contained in $\bar{Y}$ and the degeneracy locus of the associated quadric fibration is a sextic surface $\bar{C}$ in $\P^3$ with at most ordinary double points. Taking a specific hyperplane section of $\P^6$ gives back the original picture of $Y \subset \P^5$. The quadric fibration on $\P^3$ arising from $\bar{Y}$ satisfies the hypothesis of \cite[Theorem 1.1]{K2}. In particular the construction of sections $4$ and $5$ of \cite{K2} can be carried out to obtain the following diagram
\begin{equation} \label{Eqn:NoPullback}
\xymatrix{
S^+\ar[d] & X^+\ar[d] \\
S\ar[d] & X\ar[d] \\
\P^2\ar@{^{(}->}[r] & \P^3
}
\end{equation}
Here $S$ and $X$ are respectively the double covering of $\P^2$ ramified over the sextic curve $C$ and of $\P^3$ ramified over the sextic surface $\bar{C}$ induced from the centres of the corresponding Clifford algebras. The spaces $S^+$ and $X^+$ are resolutions of the singularities of $S$ and $X$. The structure of the exceptional locus over a singular point of $X$ is described in \cite[Proposition 4.4]{K2} and is isomorphic to $\P^1$. Notice that the projective space $\P^3$ coincides with the smooth base $Y$ in the notation of \cite{K2}.
Lemma \ref{Lemma:FiberProductBranch} and Lemma \ref{Lemma:FiberProductResolution} located below, ensure that the two vertical maps of the diagram (\ref{Eqn:NoPullback}) are actually fibre products, hence the diagram can be completed as follows
\begin{equation} \label{Eqn:Pullback}
\xymatrix{
& S^+\ar[d]^{\tau^+}\ar@{^{(}->}[r]^g \pullbackcorner & X^+\ar[d]^{\sigma^+}\ar@{.}[r] & \mathscr{B}^+\\
\tilde{\mathscr{B}_0}\ar@{.}[r] & S\ar[d]^{\pi}\ar@{^{(}->}[r] \pullbackcorner & X\ar[d]^f &\\
\mathscr{B}_0\ar@{.}[r] & \P^2\ar@{^{(}->}[r] & \P^3\ar@{.}[r] & \bar{\mathscr{B}_0}
}
\end{equation}
where $\mathscr{B}_0$ and $\bar{\mathscr{B}_0}$ are the Clifford algebras associated to the quadric fibrations related to $Y$ and $\bar{Y}$ described in the preliminaries and the dotted arrows denote the fact that the algebras are defined over the corresponding schemes.
Kuznetsov proved that there is a $\P^1$-bundle $M^+ \rightarrow X^+$, where $M^+$ is a flip of the Fano scheme $M$ of lines over the family of quadrics associated to $X$. That is sufficient to obtain an Azumaya algebra $\mathscr{B}^+$ on $X^+$ as done in \cite[Proposition 5.5]{K2}. By Corollary \ref{cor:pullbackb0} one has $\mathscr{B}_0 \cong i^* \bar{\mathscr{B}_0}$. Moreover, \cite[Lemma 5.7]{K2} ensures that the push forward of $\mathscr{B}^+$ along $(f \circ \sigma^+)$ is isomorphic to $\bar{\mathscr{B}_0}$.
The two squares in the diagram (\ref{Eqn:Pullback}) can be composed to obtain the following
\begin{equation} \label{Dgr:comm}
\xymatrix{
g^*(\mathscr{B}^+)\ar@{.}[r] & S^+\ar@{^{(}->}[r]^{g}\ar[d]^k \pullbackcorner & X^+\ar@{.}[r]\ar[d]^{h} & \mathscr{B}^+\\
\mathscr{B}_0\ar@{.}[r] & \P^2\ar@{^{(}->}[r]^{i} & \P^3\ar@{.}[r] & \bar{\mathscr{B}_0}
}
\end{equation}
where $k := \pi \circ \tau^+$ and $h := f \circ \sigma^+$.
The inclusion $i$ is strict, hence Proposition \ref{prop:pullbackazumvar} ensures that square (\ref{Dgr:comm}) gives a base change diagram of noncommutative varieties. Moreover, the pullback $\mathscr{A} := g^*(\mathscr{B}^+)$ is an Azumaya algebra on $S^+$.
It remains to prove that the pushforward $(\pi \circ \tau^+)_*(\mathscr{A})$ is isomorphic to $\mathscr{B}_0$. Recalling that $\mathscr{B}_0 \cong i^* \bar{\mathscr{B}_0}$, $\bar{\mathscr{B}_0} \cong h_* \mathscr{B}^+$ and $\mathscr{A} := g^* \mathscr{B}^+$, one obtains the following base change problem:
prove that $k_* g^* \mathscr{B}^+$ is equivalent to $i^* h_* \mathscr{B}^+$. This depends on the Square (\ref{Dgr:comm}) being exact cartesian, see Definition \ref{Defn:ExactCartesian}. We can apply Lemma \ref{Lemma:ExactCartesian}, since the inclusion $i$ is strict and closed and the codimensions of $S^+$ and $\P^2$ on $X^+$ are the same, and hence the square (\ref{Dgr:comm}) turns out to be exact cartesian giving that the pushforward $(\pi \circ \tau^+)_*(\mathscr{A})$ is isomorphic to $\mathscr{B}_0$. which concludes the proof.
\end{proof}
\noindent Since being Azumaya is a local property, the pushforward of $\mathscr{A}$ to $S$, restricted to the complement of ${\operatorname{Sing}}(C)$, is isomorphic to $\tilde{\mathscr{B}}_0$, the Azumaya algebra on $S \smallsetminus {\operatorname{Sing}}(C)$ provided by Proposition \ref{Prop:AzumayaAlgebraEvenFibration}. The following lemmata, used in the proof of Proposition \ref{prop:final}, hold under the same assumptions.
\begin{lemma} \label{Lemma:FiberProductBranch}
The fibre product along an inclusion of a branched covering is again a covering, ramified on the restriction of the ramification locus.
\end{lemma}
\begin{proof}
One has to prove that, if $V \subset W$ and $\tilde{W}$ is the covering of $W$ ramified over $C$, then the fibre product
\begin{equation*}
\xymatrix{
V \times_W \tilde{W} \ar[r]\ar[d] & \tilde{W}\ar[d]^f\\
V\ar@{^{(}->}[r]^i & W
}
\end{equation*}
is isomorphic to the covering $\tilde{V} \xrightarrow{\pi} V$ of $V$ ramified over $i^{-1}(C)$. This can be proved by using the definition of the fibre product. If $p \in \tilde{V}$ one can consider $f^{-1}(i \circ \pi)(p)$ and this gives a map to $\tilde{W}$. Moreover, if $Z \xrightarrow{a} \tilde{W}$ and $Z \xrightarrow{b} V$ are such that $f \circ a = i \circ b$, then one obtains a map $Z \rightarrow \tilde{V}$ that makes the diagram commute, giving the isomorphism between $\tilde{V}$ and $V \times_W \tilde{W}$.
\end{proof}
\begin{lemma} \label{Lemma:FiberProductResolution}
The fibre product of $X^+ \xrightarrow{\sigma^+} X$ along the inclusion $S \rightarrow X$ is the minimal resolution of singularities of $S$.
\end{lemma}
\begin{proof}
By construction, the inclusion sends singular points of $S$ to singular points of $X$. By the description of the exceptional locus of $X^+$, the pullback map to $S$ is an isomorphism on the smooth points of $S$ and is a contraction of a $\P^1$ on the singular points. The singularities of $X$, in particular the ones in common with $S$, are cones on singular quadrics degenerated to two planes. A small resolution $X^+$, which exists by the work of Kuznetsov, is hence obtained by blowing up one of these planes for each singular point. Since the sextic curve was obtained by taking an hyperplane section of the sextic surface, it follows that $S$ is a Cartier divisor inside $X$. It remains to check that the $\P^1$ contracted by the resolution is not all contained in the hyperplane $\{x_3=0\}$. That is implied by the last part of Proposition \ref{prop:LocusCodThree}, since a node on $X$ remains a node on $S$, hence the fibre product $X^+ \times_X S$ is smooth. This implies that the restriction of the resolution of singularities on $X$ coincides with the resolution of the singularities on $S$.
\end{proof}
\noindent It is now possible to give a proof for Theorem \ref{teor:extension}. The notation is the same used in this whole section. Notice that it is also possible to prove directly the equivalence $\mathscr{T}_Y \cong {\textbf{D}}^b(S^+,\mathscr{A})$ by using the result of Proposition \ref{prop:final} and then retracing the proof of \cite[Theorem 1.1]{K2}.
\begin{proof}[Proof of Theorem \ref{teor:extension}]
The theorem can be proved by a sequence of equivalences; recall first the one given by Proposition \ref{prop:stepone}:
\begin{equation} \label{EStep1}
\mathscr{T}_Y \cong {\textbf{D}}^b(B,\mathscr{B}_0)\text{.}
\end{equation}
Corollary \ref{cor:pullbackb0} gives the following isomorphism of noncommutative varieties:
$$(\P^2,\mathscr{B}_0) = (\P^3,\bar{\mathscr{B}_0}) \times_{\bar{B}} B\text{.}$$
As observed in the proof of \cite[Lemma 3.2]{QUADRIC}, the base change of such algebras can be described in terms of quadric fibrations. At the level of derived categories this gives
\begin{equation} \label{EStep2}
{\textbf{D}}^b(B,\mathscr{B}_0) \cong {\textbf{D}}^b(B,i^*\bar{\mathscr{B}_0}) \cong {\textbf{D}}^b\left((\bar{B},\bar{\mathscr{B}_0}) \times_{\bar{B}} B\right),
\end{equation}
that is the base change of ${\textbf{D}}^b(\bar{B},\bar{\mathscr{B}_0})$ to $\P^2$.
Then, as in the proof of Theorem 1.1 in \cite{K2}, there is an equivalence between the derived category of $\bar{B}$ twisted by $\bar{\mathscr{B}_0}$ and the derived category of $X^+$ twisted by the Azumaya algebra $\mathscr{B}^+$.
\begin{equation} \label{EStep3}
{\textbf{D}}^b(\bar{B},\bar{\mathscr{B}_0}) \cong {\textbf{D}}^b(X^+,\mathscr{B}^+)\text{.}
\end{equation}
Finally, as in the proof of Proposition \ref{prop:final}, since the map $i$ in the square (\ref{Dgr:comm}) is strict, it is possible to apply Proposition \ref{prop:pullbackazumvar}, obtaining
$$(S^+,\mathscr{A}) = (X^+,\mathscr{B}^+) \times_{\bar{B}} B\text{.}$$
In the case of Azumaya algebras, the base change can be described in terms of $\P^1$-fibrations, and for them, a semiorthogonal decomposition is provided in \cite{BER}. Hence, as for the equivalence (\ref{EStep2}),
\begin{equation} \label{EStep4}
{\textbf{D}}^b(S^+,\mathscr{A}) \cong {\textbf{D}}^b(S^+,g^*\mathscr{B}^+) \cong {\textbf{D}}^b\left((X^+,\mathscr{B}^+) \times_{\bar{B}} B\right),
\end{equation}
The theorem is proved by combining, in order, the equivalences (\ref{EStep1}), (\ref{EStep2}), (\ref{EStep3}) and (\ref{EStep4}).
\end{proof}
\medskip
{\small\noindent{\bf Acknowledgements.}
During the preparation of this paper, the author was supported by the Department of Mathematics and Natural Sciences of University of Stavanger in the framework of the grant 230986 of the Research Council of Norway.
It is a pleasure to thank Michal Kapustka, Paolo Stellari and Tom Sutherland for their useful suggestions and helpful discussions. The comments of an anonymous referee helped a lot in correcting many mistakes in the first version of this paper. The author is especially grateful to Alexander Kuznetsov for his valuable comments on a preliminary version of this paper.
}
|
\section{Derivation of the compact KPZ equation (cKPZ) and its effective 2D
static model}
Here we provide some details on the derivation of the compact KPZ
equation in one spatial dimension, and the equivalent equilibrium
description in terms of an effective Hamiltonian.
\subsection{Compact KPZ equation}
Mathematically, the compactness of the phase field can be formulated
as a ``discrete'' local gauge symmetry of the SCGLE in the amplitude-phase
representation of the field $\psi(x,t)\equiv\rho(x,t)e^{i\theta(x,t)}$.
Namely, if $\rho(x,t)e^{i\theta(x,t)}$ is a solution of the SCGLE,
then $\rho(x,t)e^{i\left[\theta(x,t)+2\pi n(x,t)\right]}$ is also
a solution, where $n(x,t)$ is an integer valued function. This symmetry
reflects the simple physical requirement that any choice of the space-time
local definition $\theta(x,t)$ by shifting $2\pi n(x,t)$ must not
change the physical results. Therefore, it is crucial to require the
effective description to respect this symmetry.
The discrete nature of the gauge transformation indicates that both
the temporal and spatial derivative of the phase field have to be
taken with care when one tries to derive the dynamical equation for
the phase field from the SCGLE. This is due to the fact that the existence
of the derivatives for any physical solution of the phase field is
not always guaranteed. To see this point, let us consider a solution
of the SCGLE, $\psi(x,t)=\rho(x,t)e^{i\theta(x,t)}$ with $\theta(x,t)$
having well defined $\partial_{t}\theta$ and $\partial_{x}\theta$
at any space-time point. For a generic gauge transformed solution
$\tilde{\psi}=\rho(x,t)e^{i\tilde{\theta}(x,t)}$ with $\tilde{\theta}(x,t)=\theta(x,t)+2\pi n(x,t)$,
one immediately sees that $\partial_{t}\tilde{\theta}$ or $\partial_{x}\tilde{\theta}$
is not well defined if around a certain space-time point, say $(x',t')$,
$n(x',t'+0^{+})\neq n(x',t'+0^{-})$ or $n(x'+0^{+},t')\neq n(x'+0^{-},t')$.
Nevertheless, due to fact that the exponential function $e^{i\alpha}$
is a smooth function with respect to $\alpha$ for any physical solution
of the phase field, one can always choose a gauge $n(x,t)$ in such
a way that \emph{either} the temporal \emph{or} the spatial derivative
of the transformed phase field exists everywhere in the space-time
domain under consideration, but \emph{not} both.
In the following, we choose the gauge in which the temporal derivative
of the phase field exists everywhere. We choose the value of $n(x,t+0^{+})$
according to $\theta(x,t)$ in such a way that $\theta(x,t+0^{+})-\theta(x,t)$
is infinitesimal. Then, the spatial derivative of the phase field
is not well-defined everywhere. At the price of introducing a short
distance scale, we can circumvent this issue by first rewriting the
SCGLE on a discrete 1D lattice with lattice spacing $\Delta_{x}$,
such that the second order spatial derivative is replaced by the a
second order finite difference. The discretized SCGLE reads
\begin{equation}
\partial_{t}\psi_{i}=(r+u\left|\psi_{i}\right|^{2})\psi_{i}+\frac{K}{\Delta_{x}^{2}}\left(\psi_{i+1}+\psi_{i-1}-2\psi_{i}\right)+\zeta_{i},\label{eq:Discrete_SCGLE}
\end{equation}
with $\psi_{i}(t)\equiv\psi(i\Delta_{x},t)$ and $\langle\zeta_{i}(t)\zeta_{i'}(t')\rangle=2\sigma\delta(t-t')\delta_{ii'}/\Delta_{x}$.
The rest of the derivation goes along the lines of the similar derivation
presented in~\cite{Altman_2d_driven_SF_2015}. Plugging the amplitude-phase
decomposition of $\psi$, i.e. $\psi_{i}(t)=(M_{0}+\chi_{i}(t))e^{i\theta_{i}(t)}$
with $M_{0}$ the homogeneous mean-field solution, into~(\ref{eq:Discrete_SCGLE}),
we arrive at a coupled discretized equation of motion (EOM) of amplitude
and phase fluctuations,\begin{widetext}
\begin{align}
\partial_{t}\chi_{i} & =(r_{d}-u_{d}\left[M_{0}+\chi_{i}\right]^{2})\left[M_{0}+\chi_{i}\right]\\
& +\frac{K_{d}}{\Delta_{x}^{2}}\left(\left[M_{0}+\chi_{i+1}\right]\cos\left(\theta_{i+1}-\theta_{i}\right)+\left[M_{0}+\chi_{i-1}\right]\cos\left(\theta_{i-1}-\theta_{i}\right)-2\left[M_{0}+\chi_{i}\right]\right)\nonumber \\
& -\frac{K_{c}}{\Delta_{x}^{2}}\left(\left[M_{0}+\chi_{i+1}\right]\sin\left(\theta_{i+1}-\theta_{i}\right)+\left[M_{0}+\chi_{i-1}\right]\sin\left(\theta_{i-1}-\theta_{i}\right)\right)+\Re[\zeta_{i}e^{-i\theta_{i}}],\nonumber \\
\partial_{t}\theta_{i} & =(r_{c}-u_{c}\left[M_{0}+\chi_{i}\right]^{2})\\
& +\frac{K_{c}}{\Delta_{x}^{2}}\left(\frac{\left[M_{0}+\chi_{i+1}\right]}{\left[M_{0}+\chi_{i}\right]}\cos\left(\theta_{i+1}-\theta_{i}\right)+\frac{\left[M_{0}+\chi_{i-1}\right]}{\left[M_{0}+\chi_{i}\right]}\cos\left(\theta_{i-1}-\theta_{i}\right)-2\right)\nonumber \\
& +\frac{K_{d}}{\Delta_{x}^{2}}\left(\frac{\left[M_{0}+\chi_{i+1}\right]}{\left[M_{0}+\chi_{i}\right]}\sin\left(\theta_{i+1}-\theta_{i}\right)+\frac{\left[M_{0}+\chi_{i-1}\right]}{\left[M_{0}+\chi_{i}\right]}\sin\left(\theta_{i-1}-\theta_{i}\right)\right)+\frac{\Im[\zeta_{i}e^{-i\theta_{i}}]}{M_{0}+\chi_{i}}.\nonumber
\end{align}
We linearize in $\chi$ and eliminate it adiabatically, made possible
since its evolution is fast compared to the gapless phase degree of
freedom. From the EOM of $\chi_{i}$ we get
\begin{align}
\chi_{i}= & -\frac{1}{2u_{d}M_{0}^{2}}\left\{ -\frac{K_{d}}{\Delta_{x}^{2}}\left(\left[M_{0}+\chi_{i+1}\right]\cos\left(\theta_{i+1}-\theta_{i}\right)+\left[M_{0}+\chi_{i-1}\right]\cos\left(\theta_{i-1}-\theta_{i}\right)-2\left[M_{0}+\chi_{i}\right]\right)\right.\nonumber \\
& \left.+\frac{K_{c}}{\Delta_{x}^{2}}\left(\left[M_{0}+\chi_{i+1}\right]\sin\left(\theta_{i+1}-\theta_{i}\right)+\left[M_{0}+\chi_{i-1}\right]\sin\left(\theta_{i-1}-\theta_{i}\right)\right)-\Re[\zeta_{i}e^{-i\theta_{i}}]\right\} .
\end{align}
\end{widetext}
Plugging the above equation into the EOM of $\theta_{i}$, neglecting
sub-leading and non-linear amplitude fluctuations $\partial_{t}\chi_{i},\chi_{i}^{2},\chi_{i}^{3}$
also here, and keeping only the additive part of the noise, we get
the compact KPZ equation
\begin{equation}
\partial_{t}\theta_{i}=\sum_{j=i\pm1}\left[-\bar{D}\sin\left(\theta_{i}-\theta_{j}\right)+\bar{\lambda}
\sin^2 \! \left(\frac{\theta_{i}-\theta_{j}}{2}\right)
\right]+\bar{\xi}_{i},\label{eq:NE-XY_model}
\end{equation}
with $ $$\theta_{i}(t)\equiv\theta(i\Delta_{x},t)$, $\bar{\xi}_{i}(t)$
being Gaussian white noise with $\langle\bar{\xi}_{i}(t)\bar{\xi}_{i'}(t')\rangle=2\bar{\sigma}_{\mathrm{KPZ}}\delta(t-t')\delta_{ii'}$,
$\bar{D}=D/\Delta_{x}^{2}$, $\bar{\lambda}=\lambda/\Delta_{x}^{2}$,
$\bar{\sigma}_{\mathrm{KPZ}}=\sigma_{\mathrm{KPZ}}/\Delta_{x}$, where
\begin{equation}
D=\frac{u_{c}K_{c}}{u_{d}}+K_{d},\quad\lambda=2\left(\frac{u_{c}K_{d}}{u_{d}}-K_{c}\right),\quad\sigma_{\mathrm{KPZ}}=\sigma\frac{u_{c}^{2}+u_{d}^{2}}{2r_{d}u_{d}}.\label{eq:KPZ_parameter}
\end{equation}
\subsection{2D static phase model $\mathcal{H}[\theta]$}
Here, we construct the effective 2D static equilibrium model for the
compact KPZ equation. Conceptually, the idea is to interpret the compact
KPZ equation as a mapping from the set of stochastic variables $\bar{\xi}(t)=\{\bar{\xi}_{i}(t)\}$
with Gaussian path probability distribution
\begin{equation}
\mathcal{P}[\bar{\xi}]\propto e^{-\frac{1}{4\bar{\sigma}_{\mathrm{KPZ}}}\sum_{i}\int_{t_{0}}^{t_{1}}dt\,\bar{\xi}_{i}(t)^{2}}\label{eq:1}
\end{equation}
to new stochastic variables $\theta(t)=\{\theta_{i}(t)\}$. The distribution
of the new variables can be obtained using the usual relation~\cite{Graham1973}
\begin{equation}
\mathcal{P}[\bar{\xi}]\mathcal{D}[\bar{\xi}]=\mathcal{P}_{\theta}[\theta]\mathcal{D}[\theta],\label{eq:2}
\end{equation}
where $\mathcal{D}[\theta]=\prod_{i}\mathcal{D}[\theta_{i}]$ is the
functional measure. This yields
\begin{equation}
\mathcal{P}_{\theta}[\theta]=\mathcal{P}[\bar{\xi}]\left|\frac{\mathcal{D}[\bar{\xi}]}{\mathcal{D}[\theta]}\right|\propto e^{-\mathcal{H}[\theta]/T},\label{eq:3}
\end{equation}
where $T \equiv 4 \bar{\sigma}_{\mathrm{KPZ}}$ is the effective temperature and
$\mathcal{H}[\theta]$ assumes the form
\begin{multline}
\mathcal{H}[\theta]=\int_{t_{0}}^{t_{1}}dt\Delta_{x}\sum_{i}\left[\partial_{t}\theta_{i} \vphantom{\sin^{2}
\! \left(\frac{\theta_{i}-\theta_{j}}{2}\right)}
\right. \\ \left. +\sum_{j=i\pm1}\left(\bar{D}\sin(\theta_{i}-\theta_{j})-\bar{\lambda}\sin^{2}
\! \left(\frac{\theta_{i}-\theta_{j}}{2}\right)\right)\right]^{2}.\label{eq:H2D}
\end{multline}
In deriving $\mathcal{P}_{\theta}$, we used that the Jacobian determinant
associated with the change of variables from $\bar{\xi}$ to $\theta$ is equal to
unity for a retarded regularization of the time derivative in the compact KPZ
equation. In the continuum limit, $\mathcal{H}[\theta]$ reduces to the form
shown in the main text. Finally, calculating the normalization of
$\mathcal{P}_{\theta}$ corresponds to calculating the partition function $Z$ of
the model defined by Eq.~\eqref{eq:H2D} at temperature $T$,
\begin{equation}
Z=\int\prod_{i}\mathcal{D}[\theta_{i}]\, e^{-\mathcal{H}[\theta]/T}.\label{eq:5}
\end{equation}
\section{Crossover behavior in the temporal correlation function}
A crossover behavior similar to the one seen in the temporal phase
fluctuation $\Delta_{\theta}(t_{1}-t_{2})$ is also observed in the
modulus of the condensate temporal auto correlation function
\begin{eqnarray}
C_{t}^{\psi}(x;t_{1},t_{2}) & \equiv & \langle\psi^{*}(x,t_{1})\psi(x,t_{2})\rangle.
\end{eqnarray}
In practice, by assuming spatial translational invariance of the correlation
function, we calculate the spatially averaged correlation functions,
i.e., $\bar{C}_{t}(t_{1},t_{2})\equiv L^{-1}\int_{0}^{L}dx\, C_{t}^{\psi}(x;t_{1},t_{2})$,
which is equivalent to the corresponding correlation function above
but helps to reduce the statistical error.
In Fig.~\ref{Fig. psi_correlation}, at linear system size $L=2^{10}$, the
dependence of
$-\log\left(\left|\bar{C}_{t}(t_{1},t_{2})\right|/\left|\bar{C}_{t}(t_{2},t_{2})\right|\right)$
on $|t_{1}-t_{2}|$ for 8 different sets of parameters in the SCGLE are shown on
a double-logarithmic scale. The exponent $\beta$ for different parameter choices
is extracted from the slope of a selected portion of the corresponding curve of
$-\log\left(\left|\bar{C}_{t}(t_{1},t_{2})\right|/\left|\bar{C}_{t}(t_{2},t_{2})\right|\right)$
on the double-logarithmic scale (cf. Fig.~\ref{Fig. psi_correlation}). For the
noise levels $\sigma=20^{-1},12^{-1},10^{-1},8^{-1},5^{-1},$ the corresponding
slopes are extracted by performing linear fits to the data points with
$|t_{1}-t_{2}|\in[10^{3},10^{4}]$. This gives rise to $2\beta=0.61$, for all the
first four curves from below with $\sigma=20^{-1},12^{-1},10^{-1},8^{-1}$,
respectively, where we notice the typical KPZ scaling behavior, but no crossover
effect to the anticipated exponential decay. This due to the fact that the
crossover time scales corresponding to these noise levels are much larger than
the time range available in our simulations.
For the fifth curve from below with $\sigma=5^{-1}$, one notices
a considerably increased slope at the right end of the curve, clearly
indicating the crossover effect which sets in within $|t_{1}-t_{2}|\in[10^{3},10^{4}]$.
This is reflected in the extracted exponent with $2\beta=0.74$.
The anticipated exponential decay is expected to manifest itself at
even larger $|t_{1}-t_{2}|$, which is however not accessible for
our current computation resources.
For the first three curves from above with $\sigma=1,2^{-1},4^{-1}$,
since the local gradient of the curve changing noticeably up to the
last accessible data point with the largest value of $|t_{1}-t_{2}|$,
denoted as $\Delta t_{\mathrm{max}}$, the corresponding slopes are
extracted by performing linear fits to the data points of a short
segment of the curve with $|t_{1}-t_{2}|\in[\Delta t_{\mathrm{max}}e^{-1/4},\Delta t_{\mathrm{max}}]$
for $\sigma=4^{-1},2^{-1}$, and an even shorter segment of the curve
with $|t_{1}-t_{2}|\in[\Delta t_{\mathrm{max}}e^{-1/5},\Delta t_{\mathrm{max}}]$,
for $\sigma=1$. This gives rise to $2\beta=1.0$ for all the three
curves, which clearly indicates the anticipated exponential decay.
However, to further numerically verify this exponential decay in a
larger time range is beyond our currently accessible computation resources.
\begin{figure}
\includegraphics[width=2.7in]{fig_1_sup}
\caption{(Color online) The dependence of $-\log\left(\left|\bar{C}_{t}(t_{1},t_{2})\right|/\left|\bar{C}_{t}(t_{2},t_{2})\right|\right)$
on $|t_{1}-t_{2}|$ for 8 different sets of parameters in the SCGLE
at linear system size $L=2^{10}$. $N_{\mathrm{Traj}}=10^{3}$ stochastic
trajectories are used to perform the ensemble average. From right
to left, the different curves corresponds to $\sigma=20^{-1},12^{-1},10^{-1},8^{-1},5^{-1},4^{-1},2^{-1},1$.
Other parameter are $K_{d}=r_{d}=u_{d}=1$, $-r_{c}=u_{c}=0.1$, and
$K_{c}=3$.}
\label{Fig. psi_correlation}
\end{figure}
\section{Phase random walk model}
To see how the finite density of space-time vortices causes the exponential decay of the autocorrelation function, one can consider a simple ``phase random walk'' model in which the time evolution of the phase is assumed to be only determined by the phase jumps due to $N$ uniformly distributed vortex cores at times $nt_{v}$, where $n=1,\dotsc,N,$ with random charges $W_{n}=\pm1$ occurring with equal probability. The model can be formulated as
\begin{eqnarray}
\theta(x,t_{1}+Nt_{v})-\theta(x,t_{1})=\delta\Theta\sum_{n=1}^{N}W_{n},
\end{eqnarray}
where $\delta\Theta$ corresponds to the average amplitude of jumps of the phase field if a vortex core is crossed along the temporal direction. A straightforward calculation shows that $\Delta_{\theta}(t_{1}-t_{2})=\left(\delta\Theta\right)^{2}\left|t_{1}-t_{2}\right|/t_{v}$, i.e., space-time vortices indeed lead to disordered behavior of the phase correlations beyond a time scale $\mathcal{O}(t_{v})$
\section{Technical details in numerical simulations}
In numerical simulations, we discretize the SCGLE in space and time by
$\Delta_{x}$ and $\Delta_{t}$, respectively, and solve the discretized dynamical
equation by using the semi-implicit algorithm developed
in~\cite{Werner_1997}. Each Gaussian white noise sequence $\xi(x,t)$, which is
used in the simulation for each corresponding stochastic trajectory of
$\psi(x,t)$, is generated by using the pseudorandom number generator employing
the Mersenne Twister algorithm~\cite{MT_algorithm}. Typically, we choose
$\Delta_{x}=1$, $\Delta_{t}\in[0.01,0.1]$ in dimensionless units and the spatial
size of the system $L=2^{10}$, which is large enough to avoid finite size
effects in most simulations presented in the main text. However, for results
corresponding to the low noise levels presented in the Fig. 2 in the main text,
we have performed simulations on systems with spatial sizes up to $2^{14}$ to
avoid finite size effects that set in at long time. Typically, we simulate the
SCGLE for a time period $T\in[10^5,10^7]$ in dimensionless units, whose specific
value depends on whether physical quantities of interest have reached their
steady states. With the direct access to the time evolution of $\psi(x,t)$ field
for each stochastic trajectory simulated, the condensate temporal auto
correlation function
$C_{t}^{\psi}(x;t_{1},t_{2})\equiv\langle\psi^{*}(x,t_{1})\psi(x,t_{2})\rangle$
and momentum distribution $n_{q}\equiv\langle\psi^{*}(q)\psi(q)\rangle$ can be
calculated straightforwardly.
In order to calculate the temporal phase fluctuation function $\Delta_{\theta}(t_{1}-t_{2})\equiv\frac{1}{L}\int_{x}\left\langle \left[\theta(x,t_{1})-\theta(x,t_{2})\right]^{2}\right\rangle -\left\langle \theta(x,t_{1})-\theta(x,t_{2})\right\rangle ^{2}$,
we need to further extract the dynamics of the phase field $\theta(x,t)$.
At each time point $t$, we first assign $\theta(x,t)$
to be the complex argument field $\arg\psi(x,t)$ and then add a multiple of $2\pi$ to the
value of $\theta(x,t)$ at each spatial point $x$ in such a way
that the phase difference between the new value of the phase at the current time point
and the value of the phase at previous time step, i.e., $\theta(x,t)-\theta(x,t-\Delta_t)$,
assumes its value within the interval $[-\pi,\pi)$. This procedure corresponds to
a choice of the gauge for the phase field that eliminates the artificial
discontinuous behavior associated with directly assigning $\theta(x,t)$
to be $\arg\psi(x,t)$. Afterwards, the temporal phase fluctuation function
$\Delta_{\theta}(t_{1}-t_{2})$ can be directly calculated.
To calculate the vortex density $P_{v}$, for each stochastic trajectory,
we first construct the discrete phase distribution $\theta_{ij}\equiv(i\Delta x,j\Delta t)$,
choosing $\Delta x=1$ and $\Delta t=1$, on the space-time plane
of the size $L\times\Delta T$ (with $\Delta T$ usually chosen to be
$10^{3}$). and then calculate the winding number of the $\theta_{ij}$ field
on each plaquette, where a vortex is identified whenever the winding number
of the corresponding plaquette, i.e., the charge of the corresponding
vortex, is non-zero. The vortex density $P_{v}$ is obtained by calculating
$\langle N_{v}/(L\times\Delta T)\rangle$, with $N_{v}$ being the
total number of vortices on the space-time plane of the size $L\times\Delta T$
for each stochastic trajectory. In Fig.~\ref{Fig. vx_distri}, a typical snapshot of a vortex charge distribution on the space-time plane of the system in the vortex turbulence phase is shown, from which we notice the vortex charge distribution assumes an irregular structure.
\begin{figure}
\includegraphics[height=1.7in]{fig_2_sup}
\caption{(Color online) A snapshot of space-time vortex charge
distribution for vortex turbulence phase at large $\tilde{\lambda}$
($\tilde{\lambda}\approx4.5$) with $\tilde{\lambda}>\tilde{\lambda}^{*}$
($\tilde{\lambda}^{*}\approx3.2$) in the weak noise regime ($\sigma=0.01$
in this case). Other parameters used in the simulation are $K_{d}=r_{d}=u_{d}=1,\,r_{c}=u_{c}=3,\, K_{c}=0.1$.
Color bar stands for the charge of vortices.}
\label{Fig. vx_distri}
\end{figure}
\section{The first order transition line}
In this section we present technical details concerning the investigation of the properties
of the first order transition line. The dependence of $P_{v}$ on the
nonequilibrium strength $\tilde{\lambda}$ at different noise levels
with $\sigma=0.01,\,0.011,\,0.012,\,0.013,\,0.014,$ in the weak noise
regime are shown in Fig.~\ref{Fig. P_v_vs_lambda_at_different_sigma}(a).
At each noise level, we notice a sudden increase of $P_{v}$ by a
few orders of magnitude when $\tilde{\lambda}$ is tuned across around
$3.2$, signifying a first order transition behavior. We estimate
the critical value $\tilde{\lambda}^{*}$ for the transition at each
noise level (cf. black dashed vertical line in Fig.~\ref{Fig. P_v_vs_lambda_at_different_sigma}(a))
by the arithmetic average of the values of $\tilde{\lambda}$ for
the two data points right before and after the transition. This estimation
gives rise to $\tilde{\lambda}^{*}=3.23$ with an error bar of $\pm0.07$
for each noise level shown in Fig.~\ref{Fig. P_v_vs_lambda_at_different_sigma}(a),
indicating that, within the numerical accuracy of our simulations, $\tilde{\lambda}^{*}$
is insensitive with respect to the noise level in the weak noise regime
as shown in Fig.~3(c) in the main text.
To estimate the space-time vortex density jump at the first order
transition at each noise level, denoted as $\Delta P_{v}\equiv P_{v}(\tilde{\lambda}\rightarrow\tilde{\lambda}^{*+})-P_{v}(\tilde{\lambda}\rightarrow\tilde{\lambda}^{*-})$
with $P_{v}(\tilde{\lambda}\rightarrow\tilde{\lambda}^{*-})$ ($P_{v}(\tilde{\lambda}\rightarrow\tilde{\lambda}^{*+})$)
being the left (right) limit of $P_{v}$ at $\tilde{\lambda}^{*}$,
we first perform a rational function $R(x)\equiv(a_{0}+a_{1}x+a_{2}x^{2})/(1+b_{1}x+b_{2}x^{2})$
fit to the data points lying on the left and the right hand side of the
transition, respectively, which are shown as solid curves in Fig.~\ref{Fig. P_v_vs_lambda_at_different_sigma}(a).
Then, the value of $P_{v}(\tilde{\lambda}\rightarrow\tilde{\lambda}^{*-})$
($P_{v}(\tilde{\lambda}\rightarrow\tilde{\lambda}^{*+})$) is obtained
by performing extrapolation at $\tilde{\lambda}^{*}$ of the corresponding
curve on the left (right) hand side of the transition. From the dependence
of $\Delta P_{v}$ on the noise level $\sigma$ shown in Fig.~\ref{Fig. P_v_vs_lambda_at_different_sigma}(b) (essentially the same plot as the one in Fig. 3(d) in the main text), we notice that $\Delta P_{v}$ decreases with respect to $\sigma$
and vanishes at $\sigma^{*}\approx0.014$ with a diverging derivative of $\Delta P_{v}$ with respect to $\sigma$.
In order to quantitatively estimate the exponent associated with the divergence of the derivative, we further perform a power law fit ($\Delta P_{v}\propto (\sigma-\sigma^*)^{1-\kappa}$) to the five data points on the left in Fig.~\ref{Fig. P_v_vs_lambda_at_different_sigma}(b), from which we extract $\kappa\simeq 0.63$ with a standard error $0.11$ and observe that $\partial \Delta P_{v} / \partial \sigma$ diverges according to $\propto (\sigma-\sigma^*)^{-\kappa}$ with $\kappa\simeq 0.63$ at $\sigma=\sigma^*$.
These observations indicate that the first order trasition line on the $\tilde{\lambda}-\sigma$ plane
terminates at higher noise level at a second order critical point
$(\tilde{\lambda}^{*},\sigma^{*})$ as shown in Fig.~3(c) in the main text.
\begin{figure*}
\includegraphics[width=3in]{fig_3a_sup}\includegraphics[width=3in]{fig_3b_sup}
\caption{(Color online) (a) Dependence of $P_{v}$ on the nonequilibrium strength
$\tilde{\lambda}$ at different noise level in the weak noise regime
when $\tilde{\lambda}$ is tuned across the critical value $\tilde{\lambda}^{*}$
of the first order transition. The black dashed vertical line corresponds
to the estimated value of $\tilde{\lambda}^{*}$. From down to up
(up to down), curves on the left (right) hand side of the black dashed
line correspond to noise levels $\sigma=0.01,\,0.011,\,0.012,\,0.013,\,0.014$,
respectively. The filled circles are data points obtained by numerical
simulations, while the pairs of filled triangles in the same color
at lower and upper locations correspond to the estimated values of
the left and the right limit of $P_{v}$ at $\tilde{\lambda}^{*}$,
i.e., $P_{v}(\tilde{\lambda}\rightarrow\tilde{\lambda}^{*-})$ and
$P_{v}(\tilde{\lambda}\rightarrow\tilde{\lambda}^{*+})$, respectively.
Values of other parameters used in the simulations are $K_{d}=r_{d}=u_{d}=1$,
$K_{c}=0.1$. $\tilde{\lambda}$ is tuned by changing $r_{c}=u_{c}$
from $1.4$ to $3.0$. (b) Space-time vortex density jump $\Delta P_{v}$
at the first order transition at different noise levels $\sigma$ (essentially the same plot as the one in Fig. 3(d) in the main text).
The left part of the black curve before the critical point ($\sigma\leq \sigma^*$) is a power law ($\Delta P_{v}\propto (\sigma-\sigma^*)^{1-\kappa}$) fit to the five data points on the left, which gives rise to $\kappa\simeq 0.63$ with a standard error
$0.11$.
}
\label{Fig. P_v_vs_lambda_at_different_sigma}
\end{figure*}
|
\section{Introduction}
Computing using a limited amount of space is a difficult and intriguing problem. In text-processing, \textit{fingerprinting} and \textit{sampling} are two natural and effective ways to tackle the problem. In this paper we use such techniques to improve on state-of-the-art solutions to the \textit{longest common extension} problem, using properties of Mersenne prime numbers used as moduli in fingerprint computation.
Let $T\in\Sigma^n$ be a length-$n$ text over an alphabet $\Sigma=\{0,\dots,\sigma-1\}$. The longest common extension (LCE) problem asks to build a data structure on $T$ supporting fast LCE queries: given $0\leq i,j<n$, return the length $\llen=LCE(i,j)$ of the longest common prefix of $T[i,\dots, n-1]$ and $T[j,\dots, n-1]$. This problem has several important applications in suffix-sorting, computing palindrome factors~\cite{breslauer1995finding,allouche2003palindrome}, identification of repeats~\cite{landau1993algorithm,kolpakov2014searching}, and string matching\cite{myers1986ano,landau1986introducing,amir2004faster}.
In this paper, we discuss fast and space-efficient LCE data structures. $\Theta(n\log n)$-bits solutions are known, guaranteeing constant-time LCE queries: such techniques include (i) building a $\mathcal{O}(1)$-time lowest common ancestor structure (see, e.g.,~\cite{harel1984fast}) on the suffix tree of the text (ST+LCA) and (ii) combining a longest common prefix structure, a $\mathcal{O}(1)$-time range minimum query structure (see, e.g.~\cite{fischer2006theoretical}), and the inverted suffix array (RMQ+LCP). On another extreme, one could simply store $T$ in $n\lceil\log_2\sigma\rceil$ bits and answer LCE queries in $\mathcal{O}(\llen)$ time by comparing $T[i,\dots, n-1]$ and $T[j,\dots, n-1]$ character by character. In their recent works~\cite{bille2015longest,bille2014time}, Bille et al. propose a number of solutions to this problem that trade query-time for space-usage. Their most practical solution is a Monte Carlo data structure requiring $\mathcal{O}(nw/\tau)$ bits on top of the text, $w$ and $0 <\tau\leq n$ being the memory word size and the block size, respectively, and answering LCE queries w.h.p. (with high probability) in $\mathcal{O}(\tau)$ time. This structure can be built in $\mathcal{O}(n\log(n/\tau))$ time. They also propose deterministic and Las Vegas structures with the same space-time bounds, but their preprocessing times---$\mathcal{O}(n^{2+\epsilon})$ and $\mathcal{O}(n^{3/2})$---make them not practical on big input texts. Deterministic data structures with similar time bounds but considerably improved construction times (of $\mathcal{O}(n\tau)$) have been recently described by Tanimura et al. in~\cite{tanimura2016}. Of interest is also the recent work described in~\cite{nishimoto2015dynamic,inenaga2015faster}, where the authors propose the first compressed LCE data structure based on SLPs (straight-line programs).
In this work we present two LCE data structures based on Karp-Rabin fingerprinting. The first structure takes $n\lceil\log_2\sigma\rceil$ bits of space---i.e. exactly the same space of the plain text---and replaces the text in that it supports $\mathcal{O}\left(\xlen\log\sigma/w\right)$ average-case time extraction of any length-$\xlen$ text substring. LCE queries are answered w.h.p. (with high probability) in $\mathcal{O}(\log\llen)$ average-case time.
In the worst case, text extraction and LCE queries are supported in $\mathcal{O}\left(\xlen\log\sigma + \log n\right)$ and $\mathcal{O}(\log\llen\cdot \log n)$ time, respectively.
This structure exploits particular properties of Mersenne prime numbers when used as moduli of the Karp-Rabin hash function in order to achieve fast query times and the information-theoretic optimal space usage.
We then extend this technique to work with general prime moduli. Our second structure takes $n\lceil\log_2\sigma\rceil + n/w +w\log_2 n$ bits of space and supports---in the worst case---optimal $\mathcal{O}\left(\xlen\log\sigma/w\right)$-time substring extraction and $\mathcal{O}(\log\llen)$-time LCE queries w.h.p.
Both structures can be built in optimal $\mathcal{O}(n)$ time. Our first result improves the state-of-the art~\cite{bille2015longest} in space, query times, and preprocessing times, but requires the modulus to be a Mersenne prime. Our second result removes this assumption and improves query times and preprocessing times of the Monte Carlo structure presented in~\cite{bille2015longest} while using its same space for $\tau=w^2$.
In Table \ref{tab: state of the art} we give an overview of the main solutions for the LCE problem described in literature.
\renewcommand{\arraystretch}{1.5}
\begin{table}
\vspace{.5cm}
\centering
\begin{tabular}{| c | c | c | c | c | c |} \hline
Space (bits) & Query time & correct & build time & Notes & Reference\\\hline
$\mathcal{O}(n\log n)$ & $\mathcal{O}(1)$ & always & $\mathcal{O}(n)$ &---& $ST + LCA$\\\hline
$\mathcal{O}(n\log n)$ & $\mathcal{O}(1)$ & always & $\mathcal{O}(n)$ &---& $RMQ + LCP$\\\hline
$n\lceil\log_2\sigma\rceil + \mathcal{O}(nw/\tau)$ & $\mathcal{O}(\tau)$ & w.h.p. & $\mathcal{O}(n\log(n/\tau))$ &$0<\tau\leq n$& \cite{bille2015longest}\\\hline
$n\lceil\log_2\sigma\rceil + \mathcal{O}(nw/\tau)$ & $\mathcal{O}(\tau)$ & always & $\mathcal{O}(n^{2+\epsilon})$ &$0<\tau\leq n$&\cite{bille2015longest}\\\hline
$n\lceil\log_2\sigma\rceil + \mathcal{O}(nw/\tau)$ & $\mathcal{O}(\tau)$ & always & $\mathcal{O}(n^{3/2})$ w.h.p. &$0<\tau\leq n$&\cite{bille2015longest}\\\hline
$n\lceil\log_2\sigma\rceil + \mathcal{O}(nw/\tau)$ & $\mathcal{O}(\tau\log\tau)$ & always & $\mathcal{O}(n\tau)$ &$0<\tau\leq n$&\cite{tanimura2016}\\\hline
$n\lceil\log_2\sigma\rceil$ & $\mathcal{O}(\llen)$ & always & $\mathcal{O}(1)$ &---& store only $T$\\\hline
$n\lceil\log_2\sigma\rceil$ & $\mathcal{O}(\log\llen)$ avg. & w.h.p. & $\mathcal{O}(n)$ &$q=2^p-1$& this paper\\\hline
$n\lceil\log_2\sigma\rceil$ & $\mathcal{O}(\log\llen\cdot \log n)$ & w.h.p. & $\mathcal{O}(n)$ &$q=2^p-1$& this paper\\\hline
$n\lceil\log_2\sigma\rceil + n/w + w\log_2 n$ & $\mathcal{O}(\log\llen)$ & w.h.p. & $\mathcal{O}(n)$ &---& this paper\\\hline
$\mathcal{O}(zw\log n\log^*n)$ & $\mathcal{O}(\log n\log^*n)$ & always & $\mathcal{O}(n\log z)$ &---& \cite{nishimoto2015dynamic,inenaga2015faster} \\\hline
\end{tabular}
\vspace{.5cm}
\caption{Main solutions for the LCE problem described in literature. Column \emph{correct} indicates whether the structures returns \emph{always} or just \emph{with high probability} (i.e. probability at least $1-2^{-w}$) the correct LCE. We account also for the space to store the plain text if this is required to answer LCE queries. For clarity we omit the $\mathcal{O}(w)$ space term that is always present in all solutions (note that even to store the plain text in practice we need a constant number of pointers to memory). $\llen = LCE(i,j)$ is the output of the LCE query. $q$ is the prime modulus used in the Karp-Rabin hash function, and $q=2^p-1$ means that $q$ must be a Mersenne prime. The ``avg.'' after the query time (second column) means that times are on average on uniformly distributed texts. If not specified, times are worst-case. $z$ is the number of phrases of the LZ77 parsing without self-references. $\log^*n$ is the iterated logarithm function.}\label{tab: state of the art}
\end{table}
\section{Preliminaries}
We assume that our input text is drawn from an integer alphabet: $T\in\{0,\dots,\sigma-1\}^n$. Throughout the description of our strategy we assume without loss of generality that $\sigma = 2$. If this is not the case, then we can simply build a binary text $T'\in\{0,1\}^{n\cdot b}$, where $b = \lceil\log_2\sigma\rceil$, by concatenating the numbers $T[0], T[1], \dots, T[n-1]$ written in binary and answer LCE queries on $T$ as $T.LCE(i,j) = \lfloor T'.LCE(i\cdot b, j\cdot b)/b\rfloor$.
Let $w$ be the memory word size (in bits). We assume, as usual, that $\log n\in\mathcal{O}(w)$. For the above considerations, in our final results the only constraint we will impose on the alphabet size $\sigma$ is that $\log \sigma\in\mathcal{O}(w)$, so that we can manipulate alphabet characters in constant time.
\subsection{Karp-Rabin Fingerprinting}
We adopt the fingerprinting approach of Karp and Rabin~\cite{karp1987efficient} to efficiently compare text substrings. The idea behind Karp-Rabin's technique is to represent a string $S\in\Sigma^n$ as a number in base $\sigma$ modulo a prime $q$. Their hash function is defined as:
$$
\phi_q(S) = \sum_{i=0}^{n-1} S[i]\sigma^{n-i-1}\ \mod q
$$
Under the assumption that $q$ fits in a constant number of memory words, this function can be used to test for string equality in constant time. Although there exists the possibility of incurring into false positives---i.e. $\phi_q(S)=\phi_q(S')$ but $S\neq S'$---, in~\cite{karp1987efficient} the authors show that, if $q$ is a prime chosen uniformly at random, then this probability is only $1/q$.
Function $\phi_q$ has several useful algebraic properties. In our case, we will exploit the fact that $\phi_q(T)$ can be obtained from $\phi_q(ST)$ and $\phi_q(S)$ with just a multiplication and a subtraction modulo $q$ (more details in the next sections).
To simplify notation, for any $i>j\geq 0$ we define $\phi_q(T[i,\dots, j]) = \phi_q(\epsilon) = 0$ ($\epsilon$ is the empty string).
\subsection{Mersenne Primes}\label{sec: Mersenne}
The key point in our first result is the use of Mersenne prime numbers---i.e. primes of the form $q = 2^p-1$---as modulus of the Karp-Rabin function $\phi_q$. First of all, note that $2^p-1$ is a $p$-bits integer and its binary representation is composed exclusively of 1's. This fact implies that \emph{any} combination of $p$ bits except $q$ itself is smaller than the modulus $q$, and will be crucial in order to obtain the information-theoretic optimal space usage in our structure. The second property we use is related to the computation of exponentials modulo $q$. With generic $q$, the computation of $x\cdot2^e\mod q$, with $x<q$ and $e\in\mathbb{N}$, requires an exponentiation---$\mathcal{O}(\log e)$ steps with the fast exponentiation algorithm---and a constant-time multiplication. Here we prove that this operation can be implemented in constant time if $q$ is a Mersenne prime. Let $q=2^p-1$. We have that
$$
\begin{array}{ll}
2^e & = 2^{p\cdot \lfloor e/p \rfloor + e\ mod\ p}\\
& = 2^{p\cdot \lfloor e/p \rfloor}\cdot 2^{e\ mod\ p}\\
& = (2^p)^{\lfloor e/p \rfloor}\cdot 2^{e\ mod\ p}\\
& \equiv_q 1^{\lfloor e/p \rfloor}\cdot 2^{e\ mod\ p}\\
& \equiv_q 2^{e\ mod\ p}\\
\end{array}
$$
Now, let $x_0x_1\dots x_{p-1}$, with $x_i\in\{0,1\}$, be the binary representation of $x<q$: $x = \sum_{i=0}^{p-1} x_i\cdot 2^{p-i-1}$. Then:
$$
\begin{array}{ll}
x\cdot 2^e & = (\sum_{i=0}^{p-1} x_i\cdot 2^{p-i-1})\cdot 2^e\\
& = \sum_{i=0}^{p-1} x_i\cdot 2^{p-i-1+e}\\
& \equiv_q \sum_{i=0}^{p-1} x_i\cdot 2^{(p-i-1+e)\ mod\ p}
\end{array}
$$
The above equality means that $x\cdot 2^e\mod (2^p-1)$ can be computed by left-rotating the binary representation of $x$ by $e\mod p$ positions. This can be done in $\mathcal{O}(1)$ time with a constant number of bitwise shift, OR, and mask operations.
Our first structure will make use of Mersenne primes as moduli of the Karp-Rabin hash function. Note that the number of known Mersenne primes is too small to use them as source of randomness in a randomized algorithm (there are only 9 Mersenne primes fitting in a 64-bits memory word), so our first structure will return the correct LCE with high probability on uniform inputs only. We will remove this assumption in our second structure, which will work with any prime modulus.
\section{A In-Place Monte Carlo LCE structure}
Our approach follows the same general idea described in~\cite{bille2015longest}: we divide the text in blocks of $\tau>0$ bits and we sample by computing the Karp-Rabin fingerprint of text prefixes ending at block boundaries. This structure permits to test (with high probability) the equality of any two text substrings and can therefore be used to compute the LCE of two text suffixes with a binary search. The key difference with the work in~\cite{bille2015longest} is that we show how this sampling of fingerprints can \emph{replace} the text while taking exactly its size ($n$ bits): we achieve this by carefully choosing $\tau$ and the prime modulus $q$ so that each block can be decoded unambiguously. In our first solution, we use primes of the form $q = 2^\tau-1$ (Mersenne primes). With this choice we can decode unambiguously all combinations of $\tau$ bits except the one corresponding to the value $q$ (which is congruent to $0$ modulo $q$). Blocks storing the value $q$ can however be discarded and replaced with an array containing their positions inside the array of blocks. In this way, we do not waste any additional bit other than the $n$ bits required to store the text and we are able to reconstruct it. Finally, since $x\cdot 2^e \mod (2^\tau-1)$ can be computed in constant time for any $x,e\in\mathbb{N}$, we do not need to sample \textit{values} $2^i\mod q$ as done in~\cite{bille2015longest} in order to speed-up fingerprint computations.
\subsection{Fingerprint Sampling}\label{sec:fingerprint sampling}
Let $w$ be the memory word size, $\tau\in\mathcal{O}(w)$ the block size and $q = 2^\tau-1$ the Mersenne modulus. Without loss of generality, we assume that $n$ is a multiple of $\tau$ (the general case can be reduced to this case by left-padding the text with $\tau-(n\mod\tau)$ bits).
Let $B, P'\in [0, q-1]^{n/\tau}$ be the integer arrays defined as
$$
B[i] = \sum_{j=0}^{\tau-1} 2^{\tau-j-1} \cdot T[i\cdot\tau+j],\ \ \ i=0,\dots, n/\tau - 1
$$
and
$$
P'[i] = \sum_{j=0}^{i} 2^{(i-j)\cdot \tau}\cdot B[j] \mod q = \phi_q(T[0,\dots,(i+1)\cdot\tau-1])
$$
To simplify notation, let $P'[-1]=0$. If $B[i]<q$, then $B[i]$ can be retrieved from $P'$ as follows:
\begin{equation}\label{eq:B}
B[i] = P'[i] - 2^\tau\cdot P'[i-1] \mod q
\end{equation}
However, if $B[i]=q$ for some $i$, the above operation yields the value $0$ instead of $q$ so $P'$ alone does not replace the text. We can remove such ambiguity by marking with a bitvector $Q\in \{0,1\}^{n/\tau}$ blocks such that $B[i]=q$:
\begin{equation}\label{eq:Q}
B[i] = q \Leftrightarrow Q[i] = 1,\ \ \ i=0,\dots,n/\tau - 1
\end{equation}
Let $\qbl$ be the number of blocks such that $B[i] = q$. We store $Q$ as an increasing sequence of integers $S_Q=\langle i\ :\ Q[i]=1,\ i=0,\dots,n/\tau-1\rangle$ so that its space usage is of $\qbl\cdot\log\frac{n}{\tau}$ bits.
For simplicity we assume that $B[0]\neq q$ (if this is not the case we can left-pad the text with $\tau$ zeros). At this point, we do not need to store anymore values $P'[i]$ for $i$'s such that $ B[i]=q$ as the following property holds for $i>0$:
$$
B[i] = q\ \ \Rightarrow\ \ P'[i] = P'[i-1]\cdot 2^\tau + B[i] \mod q = P'[i-1]\cdot 2^\tau\mod q$$
Let $B[i] = q$ and $k$ be the smallest integer such that $B[i-k]\neq q$. By applying recursively the above equality, we get
\begin{equation}\label{eq: get P'}
P'[i] = P'[i-k]\cdot 2^{k\cdot \tau} \mod q
\end{equation}
The above considerations imply that we can replace $P'$ with a vector $P\in [0,q-1]^{n/\tau - \qbl}$ containing only the $P'[i]$'s such that $B[i]\neq q$:
$$
P = \langle P'[i]\ :\ B[i] \neq q,\ i=0,\dots,n/\tau-1\rangle
$$
Before showing how we can retrieve $P'$ values using $P$ and $Q$, we describe how to efficiently answer to the following queries on $Q$:
\begin{enumerate}
\item \texttt{Access}: return $Q[i]$
\item \texttt{Rank}: $Q.rank_b(i),\ 0\leq i \leq n/\tau,\ b\in\{0,1\}$. Return the number of bits equal to $b$ before position $i$ excluded
\item \texttt{0-predecessor}: $Q.pred(i)$. Return the rightmost position $j\leq i$ such that $Q[j]=0$
\end{enumerate}
\texttt{access} requires just a binary search on $S_Q$. The same holds for $Q.rank_1(i)$: return, using a binary-search on $S_Q$, the number of elements strictly smaller than $i$. At this point, $Q.rank_0(i)$ is equal to $i - Q.rank_1(i)$. To answer $Q.pred(i)$, where $Q[i]=1$ (if $Q[i]=0$ the answer is $i$), we can use again binary search on $S_Q$. The idea is that we have to find---in the bitvector $Q$---the 0 preceding the run of 1's containing $Q[i]$; in $S_Q$, this run translates to a sub-sequence of integers with the property that the difference between any two of them is equal to their distance in the array $S_Q$. This leads to the following binary search criteria. Let $k=Q.rank_1(i)$. We start the search in the interval $S_Q[l,\dots, r]$, with $l=0,\ r=k$, and we pick the middle $t$ of this interval: $t=(l+r)/2$. First, check if $r=t$ or if $S_Q[r-t]-S_Q[r-t-1]>1$: in these cases, return $S_Q[r-t]-1$. If this is not the case, if $S_Q[r] - S_Q[r-t] = t$ then recurse in $S_Q[l,\dots, r-t-1]$, otherwise in $S_Q[r-t,\dots,r]$.
At this point, let $t = Q.pred(i)$. By combining Equations \ref{eq:B}, \ref{eq:Q}, and \ref{eq: get P'} we obtain the following relations for every $0\leq i < n/\tau$:
\begin{equation}\label{eq: relations}
\left\{\begin{array}{ll}
P'[i] & = P[Q.rank_0(t)]\cdot 2^{\tau\cdot (i-t)} \mod q\\
B[i] & = Q[i]\cdot q + (1-Q[i])(P'[i] - 2^\tau\cdot P'[i-1]\mod q)
\end{array}\right.
\end{equation}
Since multiplication by any power of 2 modulo $q=2^\tau-1$ takes constant time, by using $P$ and $Q$ and applying Equations \ref{eq: relations} we can (remember that $\qbl=|S_Q|$):
\begin{enumerate}
\item Retrieve any $B[i]$ in $\mathcal{O}(\log \qbl)$ time
\item Retrieve any $P'[i]$ in $\mathcal{O}(\log \qbl)$ time
\end{enumerate}
Note that query 1. can be directly used to extract any text substring of length $\xlen$ in $\mathcal{O}(\lceil \xlen/w\rceil\cdot \log \qbl) \subseteq \mathcal{O}((\xlen/w+1)\cdot \log\qbl) \subseteq \mathcal{O}(\xlen + \log n)$ time.
\subsubsection{Space usage}
We add the constraint $\tau \geq \log_2 n > \log_2\frac{n}{\tau}$: a text pointer must fit in $\tau$ bits (we just have to use a big enough Mersenne prime in order to satisfy this requirement).
The size of our structure is:
$$
\begin{array}{ll}
|P|\cdot \tau + |S_Q|\cdot \log\frac{n}{\tau} & = \left(\frac{n}{\tau}-\qbl\right)\tau + \qbl\log\frac{n}{\tau} \\
& \leq \left(\frac{n}{\tau}-\qbl\right)\tau + \qbl\cdot \tau\\
& = n
\end{array}
$$
bits. This is exactly the space required by the naive solution that answer LCE queries by comparing text suffixes character by character.
\subsubsection{Building the structure}
We analyze the construction algorithm assuming a general alphabet of size $\sigma\in\mathcal{O}(2^w)$. In $\mathcal{O}(n)$ time we read the text and pack its binary representation in blocks of $\tau$ bits; this list of blocks is array $B$. Then, array $P'$ can be easily constructed from $B$ in $\mathcal{O}(n/\tau)$ time by performing $|P'|-1$ additions and multiplications by $2^\tau$ modulo $q$. Using $B$ and $P'$, we can compute $P$ and $S_Q$ in time proportional to their size ($\mathcal{O}(n/\tau)$) and then discard $B$ and $P'$.
\subsection{Answering LCE Queries}
First of all, we show how to compute the fingerprint $\phi_q(T[0,\dots, i])$ of any text prefix. Let $j=\lfloor i/\tau\rfloor$. Then,
$$
\begin{array}{ll}
\phi_q(T[0,\dots, i]) & = \phi_q(T[0,\dots, j\cdot\tau-1])\cdot 2^{i-j\cdot\tau+1} + \phi_q(T[j\cdot\tau, \dots, i]) \mod q \\
& = P'[j-1]\cdot 2^{i-j\cdot\tau+1} + \lfloor B[j]/2^{p-i+j\cdot\tau-1}\rfloor \mod q
\end{array}
$$
Since multiplication by any power of 2 takes constant time, we obtain that the computation of $\phi_q(T[0,\dots, i])$ takes $\mathcal{O}(\log\qbl)$ time with our structure for all $0\leq i < n$.
We now have to show how to compute the fingerprint $\phi_q(T[i,\dots, j]),\ j\geq i$ of any text substring. This can be easily achieved by means of the equality:
\begin{equation}\label{eq: rk substring}
\phi_q(T[i,\dots, j]) = \phi_q(T[0,\dots, j]) - \phi_q(T[0,\dots, i-1])\cdot 2^{j-i+1} \mod q
\end{equation}
Again, since multiplication by a power of 2 takes constant time, the above operation takes $\mathcal{O}(\log\qbl)$ time with our structure. At this point, we can easily answer $LCE(i,j)$ by comparing $\phi_q(T[i,\dots, i+k])$ with $\phi_q(T[j,\dots, j+k])$ for $\mathcal{O}(\log n)$ values of $k$ with binary search. We can furthermore improve query times by performing an exponential search before applying the binary search procedure. We compare $\phi_q(T[i,\dots, i+k])$ with $\phi_q(T[j,\dots, j+k])$ for $k=2^0, 2^1, 2^2, \dots$ until the two fingerprints are different. Letting $\llen = LCE(i,j)$, this procedure terminates in $\mathcal{O}(\log\llen)$ steps. We then apply the binary search procedure described above on the interval of size $\mathcal{O}(\llen)$ obtained with the exponential search. Each exponential and binary search step take $\mathcal{O}(\log\qbl)$ time.
\subsubsection{Analysis}
First of all, note that if $\sigma$ is not a power of 2 then the binary representation of any alphabet character contains at least one bit equal to 0. Then, by choosing $\tau\geq 2\cdot \lceil \log_2\sigma\rceil$, we get that $B[i]\neq q$ for every $0\leq i < n/\tau$ (because every block of $\tau$ bits contains at least one bit equal to 0) and therefore $\qbl=0$. We obtain (switching to a alphabet of size $\sigma\in\mathcal{O}(2^w)$):
\begin{theorem}\label{th: alphabet}
If $\sigma\neq 2^e$, $e\in\mathbb{N}$, then our data structure takes $n\lceil \log_2\sigma\rceil$ bits of space and supports extraction of any length-$\xlen$ text substring and LCE queries w.h.p. in
$\mathcal{O}\left(\xlen\log\sigma/w\right)$ and $\mathcal{O}(\log\llen)$ worst-case time, respectively. The structure can be built in optimal $\mathcal{O}(n)$ time.
\end{theorem}
The above time bounds hold also in the average case for any alphabet size. Assuming a uniform distribution of the text, the probability of any block $B[i]$ having value $q$ is only $(1/2)^\tau$. Since $\tau \geq \log_2 n$, this probability is $\mathcal{O}(1/n)$. It follows that, in the average case, the number $\qbl$ of such blocks is $\mathcal{O}(1)$. We obtain:
\begin{theorem}\label{th: avg}
Our data structure takes $n\lceil \log_2\sigma\rceil$ bits of space and supports extraction of any length-$\xlen$ text substring and LCE queries w.h.p. in
$\mathcal{O}\left(\xlen\log\sigma/w\right)$ and $\mathcal{O}(\log\llen)$ average-case time, respectively. The structure can be built in optimal $\mathcal{O}(n)$ time.
\end{theorem}
In the worst case and for any alphabet size, the following bounds hold:
\begin{theorem}
Our data structure takes $n\lceil \log_2\sigma\rceil$ bits of space and supports extraction of any length-$\xlen$ text substring and LCE queries w.h.p. in
$\mathcal{O}\left(\xlen\log\sigma + \log n\right)$ and $\mathcal{O}(\log\llen\cdot \log n)$ worst-case time, respectively. The structure can be built in optimal $\mathcal{O}(n)$ time.
\end{theorem}
Note that in order to guarantee the time bounds of Theorem \ref{th: alphabet} in the worst case for any alphabet size, we can always add an extra character to the alphabet. However, in case $\sigma=2^e$ for some $e>0$, this comes at the price of using $n$ additional bits on top of the optimal $n\cdot e$. The solution we give in the next section works with any alphabet size and reduces this overhead to $n/w+w\log_2n$ bits, while still guaranteeing the time bounds of Theorem \ref{th: alphabet} in the worst case.
%
%
%
%
%
%
%
\section{A Faster Monte Carlo LCE structure}
In this section we generalize our idea to any prime $q$. This generalization comes at the price of using $n/w + w\log_2 n$ additional bits of space.
Let $q$ be our prime modulus and $\tau = \lceil\log_2 q\rceil$ the number of bits required to write numbers modulo $q$. We break the text in blocks $B[0],\dots,B[n/\tau-1]$ of $\tau$ bits each as done in the previous section, and we compute array $P'$ storing the Karp-Rabin fingerprint of text prefixes ending at block boundaries. We then discard $B$ and keep only $P'$. As noted in section \ref{sec:fingerprint sampling}, $P'$ is sufficient to decode any text block $B[i]$ such that $B[i]<q$. If $B[i]\geq q$, note that
$$B[i]-q \leq (2^\tau-1)-2^{\tau-1} < 2^{\tau-1} \leq q$$
so $B[i]\mod q = B[i]-q$ holds. It is therefore sufficient to add a bitvector $D[0,\dots,n/\tau-1]$ marking with a '0' blocks such that $B[i]<q$ and with a '1' blocks such that $B[i]\geq q$ and retrieve $B$'s values as
\begin{equation}\label{eq:B1}
B[i] = \left(P'[i] - 2^\tau\cdot P'[i-1] \mod q\right) + D[i]\cdot q
\end{equation}
Note that this is a generalization of the idea presented in the previous section, the only difference being that with Mersenne primes there is only one block value greater than or equal to $q$ (i.e. $q$ itself). In that case, this difference permits to save the $n/w$ bits of array $D$. Since we store the full $P'$, extracting each $B[i]$ takes constant time\footnote{Note: computing $2^\tau\mod q$ in Equation \ref{eq:B} takes constant time since $2^\tau$ fits in a memory word.}. Computing $LCE(i,j)$, where we assume $i\leq j$ (swapping $i$ and $j$ does not change the result), requires $\mathcal{O}(\log n)$ binary search steps; at this point, since we do not pre-compute and store (i.e. sample) values $2^{i}\mod q,\ i=0,\dots,n-1$, at each binary search step we are forced to spend $\mathcal{O}(\log n)$ additional time to compute powers of 2 modulo $q$ (with the fast exponentiation algorithm) in order to evaluate Equation \ref{eq: rk substring}. We can avoid this overhead by pre-computing and storing (in $\tau\log_2 n$ bits) values $z_i = 2^{2^i}\mod q,\ i=0,\dots,\lfloor\log_2 n\rfloor$ and by performing binary search by splitting interval lengths in correspondence of powers of 2 as follows. First, note that $z_0=2$ and $z_{i+1} = (z_i)^2\mod q$, so these values can be pre-computed in $\mathcal{O}(\log n)$ time. Let the notation $\langle i,j,e,k\rangle$, with $0\leq i,j,e,k<n$ and $e<k$, denote that we already verified that $T[i,\dots, i+e-1] = T[j,\dots, j+e-1]$ and $T[i,\dots, i+k-1] \neq T[j,\dots, j+k-1]$. We use this notation to indicate the state of a binary search step, and start from state $\langle i,j,0,n-j \rangle$ (we assume for simplicity that $T[i,\dots, i+(n-j)-1] \neq T[j,\dots, n-1]$; otherwise, $LCE(i,j)=n-j$). Let $\textit{off}$ be a multiplicative offset. Throughout the procedure, the invariant $\textit{off}=2^e\mod q$ will hold in state $\langle i,j,e,k\rangle$. We start with $\textit{off}=1$. We use a modified version of Equation \ref{eq: rk substring} by adding a parameter (exponential $exp$) to the Karp-Rabin hash function:
\begin{equation}\label{eq: rk substring1}
\phi_q'(T[i,\dots, j],exp) = \phi_q(T[0,\dots, j]) - \phi_q(T[0,\dots, i-1])\cdot exp \mod q
\end{equation}
Note that $\phi_q(T[i,\dots, j]) = \phi_q'(T[i,\dots, j],2^{j-i+1})$. At binary search step $\langle i,j,e, k\rangle$ we still have to compare the last $l = k-e$ characters of $T[i,\dots,i+k-1]$ and $T[j,\dots,j+k-1]$. We split each of these two substrings suffixes in a left part of length $l' = 2^{\lfloor\log_2(l/2)\rfloor}$ (i.e. the closest power of 2 smaller than or equal to $l/2$) and a right part of length $l-l'$. Note that value $2^{l'}\mod q = z_{\log_2l'} = z_{\lfloor\log_2(l/2)\rfloor}$ has been pre-computed, so we can compute and compare in constant time the two values
$$
\phi_q(T[i,\dots,i+e+l'-1]) = \phi_q'(T[i,\dots, i+e+l'-1],\textit{off}\cdot z_{\lfloor\log_2(l/2)\rfloor})
$$
and
$$
\phi_q(T[j,\dots,j+e+l'-1]) = \phi_q'(T[j,\dots, j+e+l'-1],\textit{off}\cdot z_{\lfloor\log_2(l/2)\rfloor})
$$
If the two values differ, then we leave $\textit{off}$ unchanged and recurse on $\langle i,j,e,e+l'\rangle$. If the two values are equal, then we set $\textit{off}\leftarrow \textit{off}\cdot 2^{l'}\mod q = \textit{off}\cdot z_{\lfloor\log_2(l/2)\rfloor} \mod q$ and recurse on $\langle i,j,e+l',k\rangle$. Since $l/4 < l'\leq l/2$, this binary search procedure terminates in $\mathcal{O}(\log n)$ steps, each taking constant time. As done in the previous section, we can perform an exponential search before the binary search in order to reduce the size of the binary search interval from $\mathcal{O}(n)$ to $\mathcal{O}(\llen)$. Note that with our sampling $z_i$ of powers of 2 modulo $q$ it is straightforward to implement each exponential search step in constant time.
In order to minimize space usage, we can choose uniformly a prime $q$ such that $\lceil\log_2q\rceil = w$ (for the prime number theorem, there are $\Theta(2^{w-1}/(w-1))$ such primes). It is easy to see that $P'$ and $D$ can be constructed in $\mathcal{O}(n)$ time, therefore we obtain:
\begin{theorem}
Our data structure takes $n\lceil \log_2\sigma\rceil + n/w + w\log_2 n$ bits of space and supports extraction of any length-$\xlen$ text substring and LCE queries w.h.p. in
$\mathcal{O}\left(\xlen\log\sigma/w\right)$ and $\mathcal{O}(\log\llen)$ worst-case time, respectively. The structure can be built in optimal $\mathcal{O}(n)$ time.
\end{theorem}
\section{Implementation}
Our solutions are very practical to implement. The C++ implementation of our structure based on Mersenne primes---available at~\cite{rklce}---is extremely fast and memory-efficient, supporting (on average) $710\ ns$ LCE queries and $160\ ns$ single-character access queries on a prefix of length $n=3\cdot 10^9$ of the Human genome ($\Sigma=\{A,C,G,T\}$). The space taken by the structure in RAM was of $750000560$ Bytes, i.e. only $560$ Bytes on top of the optimal 2 bits per character required to store the text. As a comparison, accessing at random positions an array of the same size takes---on the same machine---$74\ ns$ per access query. Our implementation took $170\ s$ to build the structure on the input file. We ran the experiment on a \texttt{intel core i7} machine running Ubuntu linux version 15.10. We leave to an extended version of this work a comparison of our software with other available implementations of other solutions for the LCE problem.
\section{Final Remarks}
In this work we showed that $o(n)$-time LCE queries are possible in the same space of the text. We achieve this result by combining Karp-Rabin fingerprinting with properties of Mersenne prime numbers. If a general prime is used in the Karp-Rabin hash function, then our general technique can still be applied but it requires $n/w+w\log_2 n$ additional bits of space. Both our data structures \textit{replace} the text, in that they support fast extraction of any text substring.
With respect to the Monte Carlo solution presented in~\cite{bille2015longest}, our structures are smaller, faster to build, and support faster LCE queries. Our results are also extremely practical. The implementation of our data structure based on Mersenne primes takes the same space of the input text and supports LCE queries with only a 10x slowdown with respect to direct access operations in main memory. The structure is also extremely fast to build.
An interesting extension of this work would be to devise a way of
employing general prime numbers in the Karp-Rabin hash function without using $n/w$ additional bits on top of the space of the plain text. Another interesting line of research would be that of using derandomization techniques such as the ones proposed in~\cite{bille2014time,bille2015longest} in order to get versions of our solutions that always return the correct LCE.
\bibliographystyle{splncs}
|
\section{Introduction}
\label{sec:introduction}
Cosmic rays are high-energy charged particles striking Earth from all directions since time immemorial. They originate mainly in outer space and comprise roughly 99\% of atomic nuclei and 1\% of electrons. They are commonly divided into primary and secondary cosmic rays. Primary cosmic rays, such as electrons, protons, helium, iron and other nuclei synthesized in stars, consist of particles directly accelerated at sources. Secondary cosmic rays, like lithium, beryllium and boron, which are not created by stellar nucleosynthesis, are composed of particles produced by the primary cosmic rays during their propagation in the interstellar medium. A very small fraction of cosmic rays are stable particles of antimatter: positrons and anti-protons. Whether or not cosmic-ray positrons and anti-protons are pure secondaries is still an open question.
Even though electrons\footnote{Unless stated otherwise, we hereafter refer to both e$^-$ and e$^+$ as simply \textit{electrons}.} represent only a tiny part of cosmic rays, they are of great interest, especially at high energy. Indeed, above a few GeV these particles are subject to severe radiative energy losses, mainly by synchrotron radiation in magnetic fields and by inverse Compton scattering in radiation fields \cite{JON65, BLU70}. These processes are so drastic that high-energy cosmic-ray electrons (HECREs) cannot travel far distances from their sources. In stark contrast to hadrons, they can reach Earth only from sources in the local neighborhood \cite{COW79}. Therefore, they can be used as a powerful probe into local cosmic-ray accelerators, which is crucial to the long-standing problem of the origin of cosmic rays. HECREs are also very important to X-ray and $\gamma$-ray astronomies, to dark matter investigation and to many other issues in high-energy astrophysics.
Researchers have taken a keen interest in cosmic-ray electrons for a very long time. The first direct observation of these particles was achieved in 1961 with balloon-borne experiments \cite{EAR61, MEY61}. During the following years many balloon flights using different detectors were carried out \cite{DAN65, SMI68, RUB68, SCH71}. These early experiments measured the flux of cosmic-ray electrons up to several hundred GeV. Experiments then explored higher energy region using more sophisticated devices with a larger geometrical factor, longer exposure and higher proton rejection power. These detectors were chiefly of two kinds: magnetic spectrometers and emulsion chambers. The first category observed negative electrons and positrons separately up to a few tens of GeV. Such was the case with the balloon-borne experiments CAPRICE \cite{BOE00, BOE01}, HEAT \cite{DUV01, BEA04} and MASS-91 \cite{GRI02}, in addition to AMS-01, which was flown on the space shuttle Discovery in 1998 \cite{AGU07}. The second category lacked the ability of discriminating between negative electrons and positrons. However, it had the merit of extending the energy spectrum measurements far beyond the range accessible to the first category. Examples include the balloon-borne experiments ECC \cite{KOB12}, BETS \cite{TOR01}, ATIC-2 \cite{CHA05} and PPB-BETS \cite{YOS08}. Moreover, the ground-based Cerenkov telescopes H.E.S.S. \cite{AHA08} and MAGIC \cite{BOR11} also observed cosmic-ray electrons at very high energy. The launch over the last decade of a new generation of high-precision instruments on-board satellites (PAMELA, the Fermi Gamma-ray Space Telescope, AMS-02 and CALET) opened a new era in the study of HECREs.
The detection of HECREs implies the existence of local sources. These sources are expected to manifest themselves in the energy spectrum, which should display special features. Recent experimental results show that the energy spectrum of cosmic-ray electrons extends well beyond 1~TeV and does have features. Most notably, PAMELA uncovered in the energy range 10-300~GeV a significant increase in the positron fraction (ratio of the positron flux to the combined flux of positrons and negative electrons) \cite{ADR09, ADR10, ADR13}. This remarkable result, confirmed by Fermi-LAT \cite{ACK12} and by AMS-02 \cite{AGU13}, is not consistent with conventional models based on the assumption that positrons arise only from the secondary production of cosmic rays by collision with the interstellar medium. These models rather predict a positron fraction falling smoothly with energy \cite{PRO82, MOS98}. A new precision measurement by AMS-02 extending up to 500~GeV indicates that the positron fraction, on the one hand, levels off at about 200~GeV and, on the other hand, does not show any fine structure or sharp cutoff \cite{ACC14}.
In parallel, ATIC reported an excess of cosmic-ray electrons over conventional model expectations between 300 and 800~GeV followed by a steepening above 1~TeV \cite{CHA08, PAN11}. Although PPB-BETS also showed a bump-like structure between 100 and 700~GeV \cite{YOS08}, Fermi-LAT and H.E.S.S. experiments found no evidence for a prominent peak \cite{ABD09, ACK10b, AHA09}.
There are several possible scenarios, from standard astrophysics to exotic physics, to account for the special features of the electron energy spectrum. They may be ascribed to the contribution from some individual local sources such as supernova remnants (SNRs) (see, e.g., \cite{BLA09, AHL09, STA10}) or pulsars (see, e.g., \cite{HOO09, MAL09, YUK09, DEL10, BLA11, PRO11, DIM14}). However, additional sources such as the annihilation or decay of dark-matter particles cannot be excluded (see, e.g., \cite{KAN09, ARK09, CHO09, POR11, KOP13, BER13, IBA14, FEN14}). The main idea of these models is that equal amounts of negative electrons and positrons are produced by the source, be it an astrophysical object or dark-matter particles. This contribution emerges at high energy from a background formed by electrons coming from distant sources. The observed anomalies may also be due to just propagation effects \cite{SHA09,GAG13,BLU13,COW14}. It must be emphasized here that there are arguments for and against each scenario and none of these approaches is yet conclusive \cite{MOS13}. Recent reviews can be found in \cite{FAN10, SER12, CHO13,PAN13,PIC14,ISR14}.
The energy spectrum of cosmic-ray electrons is typically interpreted within the context of propagation models and the traditional method consists in solving appropriate transport equations \cite{GIN76b, STR07}. Due to its inherent stochastic nature, the diffusive propagation of cosmic rays can also be treated using Monte Carlo simulation \cite{HUA07}. While it is widely believed that this technique is inefficient in this kind of application, the proximity of sources in the case of HECREs, coupled with the absence of hadronic interactions and the simplicity in energy loss processes, simplifies the modeling significantly. Monte Carlo approach comes in handy particularly at very high energy where only a few sources are expected to dominate the cosmic-ray electron flux. It provides very useful information about the electron-by-electron fluctuations, thus complementing the traditional method.
In this respect, we implemented a fully 3-dimensional time-dependent Monte Carlo simulation of the propagation of HECREs in our galaxy. To speed up calculations we employed MPI parallel programming on an HPC cluster system. We restricted ourselves to the energy region above 10~GeV where cosmic rays are not affected by the solar wind modulation. We considered only pure diffusion since convection and reacceleration are negligible above a few GeV \cite{DEL09}. We focused on the most natural way to explain the spectral features of HECREs, assuming that some nearby pulsars and/or SNRs are the sources of such particles. The other possible scenarios were not examined. We used a two-component model, separating the local source contribution from the distant source contribution \cite{ATO95}. The latter, which constitutes the background, was estimated with the public code GALPROP\footnote{http://galprop.stanford.edu/} \cite{VLA11}.
There are two different types of key quantities when interpreting the observed energy spectrum of cosmic-ray electrons. The first ones are the parameters associated with propagation; they are behind the distortion of the injection spectrum. The second ones are the parameters related to the spectral profile at injection; they control the overall form of the measured spectrum. In our study we first examined the propagation effects. Next, we inferred from these initial calculations the most likely astrophysical sources of HECREs. Then, we derived the expected flux from some typical sources. And finally, we addressed the anisotropy issue.
The outline of this paper is as follows. After this introduction (\S\ref{sec:introduction}) we review in \S\ref{sec:propagation} the necessary mathematical background for describing cosmic-ray electron propagation and we detail our Monte Carlo procedure. In the following section (\S\ref{sec:results}) we report the main results achieved: electron energy and lifetime distributions in \S\ref{sec:prop_effects}, potential astrophysical sources in \S\ref{sec:sources}, cosmic-ray electron flux, positron fraction and anisotropy for some typical sources in \S\ref{sec:flux} and \S\ref{sec:anisotropy}. We end our study with a brief conclusion (\S\ref{sec:conclusion}).
\section{Propagation of cosmic-ray electrons}
\label{sec:propagation}
Many aspects of the origin of most cosmic rays can be understood in terms of acceleration at astrophysical sources and propagation in turbulent magnetic fields in the Galaxy (see, e.g., \cite{BER90,GAI90,LON11}). Because of random scattering by the magnetic field irregularities, charged cosmic-ray particles can be considered to diffuse from their sources through the interstellar medium. The principal argument in favor of diffusion comes from the presence in the cosmic radiation of a much greater proportion of secondary nuclei when compared to the elemental abundances of the solar system. Measurements of the ratios of secondary to primary cosmic-ray nuclei, such as B/C, lead to the following conclusions:
\begin{enumerate}
\item Cosmic rays travel distances thousands of times greater than the thickness of the galactic disk between injection and observation. This result suggests diffusion in a confinement region including the galactic disk.
\item Most of propagation occurs after acceleration since the amount of matter traversed by cosmic rays decreases as energy increases.
\end{enumerate}
\noindent Cosmic-ray electrons should not depart from the general rule and should also undergo diffusion between injection and observation.
\subsection{Energy loss processes}
As cosmic-ray electrons propagate through the interstellar medium, they are subject to a number of energy loss processes that cause distortion of their injection spectra. These processes, which involve interaction with matter, with magnetic fields and with radiation, comprise \cite{LON11}:
\begin{itemize}
\item[--] ionization loss, which depends upon energy $E$ logarithmically ($\propto \ln E$);
\item[--] bremsstrahlung (and adiabatic loss when relevant), which varies linearly with energy ($\propto E$);
\item[--] synchrotron radiation in magnetic fields and inverse Compton scattering in radiation fields, which are proportional to the square of energy ($\propto E^2$).
\end{itemize}
The different energy loss processes of cosmic-ray electrons are shown in figure~\ref{fig:energy_loss} for typical values of the interstellar gas density $n$ ($\simeq 1$~cm$^{-3}$), the galactic magnetic field $B$ ($\simeq 3 \, \mu$G) and the energy density of the galactic radiation field $U_{\rm rad}$ ($\simeq 1$~eV~cm$^{-3}$) \cite{LON11}. The galactic radiation is composed of the stellar radiation, the re-emitted radiation from dust grains and the cosmic microwave background (CMB). Except for very low energy, the ionization loss can be ignored. Bremsstrahlung is important in the GeV energy range. Synchrotron radiation and inverse Compton scattering predominate above a few tens of GeV, the latter being about three times stronger than the former.
\begin{figure}[tbp]
\centering
\includegraphics[width=0.58\textwidth]{fig_energy_loss_maxlifetime.eps}
\caption{Energy loss rates as a function of energy of cosmic-ray electrons during their propagation in the Galaxy. The dashed line represents the evolution of the electron maximum lifetime with energy.}
\label{fig:energy_loss}
\end{figure}
For HECREs ($\gtrsim 10$~GeV) it is quite reasonable to consider only bremsstrahlung, synchrotron radiation and inverse Compton scattering. The total energy loss rate by these processes for ultra-relativistic electrons within Thomson approximation is given by \cite{LON11}
\begin{equation}
-\frac{{\rm d}E}{{\rm d}t} = a_0 E + b_0 E^{2} ,
\label{eq:loss_rate}
\end{equation}
\noindent with
\begin{equation}
a_0 = 3.66 \times 10^{-22} \; n \simeq 3.7 \times 10^{-16} \; \; \rm{s}^{-1} ,
\label{eq:a0}
\end{equation}
\noindent and
\begin{equation}
b_0 = \frac{4 \sigma_{\rm T} c}{3 (m_{\rm e}c^{2})^{2}} \left( U_\mathrm{mag} + U_\mathrm{rad} \right)
\simeq 1.3 \times 10^{-16} \; \; \rm{ GeV}^{-1} \; \rm{s}^{-1} .
\label{eq:b01}
\end{equation}
\noindent $\sigma_{\rm T}$ is Thomson scattering cross-section, $c$ the speed of light, $m_{\rm e}$ the rest mass of the electron and $U_\mathrm{mag} (\equiv B^2 / 2 \mu_0)$ the energy density of the galactic magnetic field ($\mu_{0}$ is the permeability of free space). Because of these energy losses, the maximum lifetime of a cosmic-ray electron with an energy $E$ anywhere in the Galaxy is therefore
\begin{equation}
\tau_{\rm max} = \int_E^\infty \frac{{\rm d}E}{(-{\rm d}E/{\rm d}t)} = \frac{1}{a_0} \ln \left( \frac{a_0+b_0 E}{b_0 E} \right) .
\label{eq:tau}
\end{equation}
\noindent Above a few tens of GeV bremsstrahlung becomes negligible and then
\begin{equation}
\tau_{\rm max} \simeq \frac{1}{b_{0}E} \qquad \left( E \gg \frac{a_0}{b_0} \simeq 3 \, \, \, \mathrm{GeV} \right) .
\label{eq:tau2}
\end{equation}
\noindent $\tau_{\rm max}$ decreases very rapidly with energy (figure~\ref{fig:energy_loss}). For 10, $10^2$ and $10^3$~GeV, $\tau_{\rm max}$ is roughly equal to $10^7$, $10^6$ and $10^5$~yr, respectively.
Actually, at very high energy the constant $b_0$ should be corrected according to the full Klein-Nishina cross-section for the inverse Compton scattering interactions of electrons. This correction results in a significant reduction of the electron energy loss rate compared with Thomson approximation. It can be formulated as \cite{SCH10}:
\begin{equation}
b_0 = \frac{4 \sigma_{\rm T} c U_\mathrm{mag}}{3 (m_{\rm e}c^{2})^{2}} \left( 1 + \sum_{i=1}^4 \frac{U_\mathrm{rad}^i}{U_\mathrm{mag}} \frac{E_{K,i}^2}{E^2+E_{K,i}^2} \right) .
\label{eq:b02}
\end{equation}
\noindent The sum relates to the four types of radiation existing in the Galaxy: optical photons from stars of spectral type B, optical photons from stars of spectral type G-K, infrared photons and CMB photons. $U_\mathrm{rad}^i$ and $E_{K,i}$ ($i = 1$-4) are the corresponding energy densities and critical Klein-Nishina energies, respectively. Their values can be found in ref. \cite{SCH10} (Table~1).
\subsection{Description of diffusion}
Diffusion is experienced as a macroscopic phenomenon although the mechanisms of diffusion are founded in microscopic collision processes. The phenomenological theory of cosmic-ray diffusion leads to a well-known transport equation, called the diffusion-loss equation \cite{GIN64}. In the case of homogeneous and isotropic diffusion, it is given for electrons by
\begin{equation}
\frac{\partial \phi(E, \, \mathbf{r}, \, t)}{\partial t} = D(E) \nabla^2 \phi (E, \, \mathbf{r}, \, t)
+ \frac{\partial }{\partial E} \left[ b(E) \phi (E, \, \mathbf{r}, \, t) \right] + Q(E, \, t) ,
\label{eq:transport}
\end{equation}
\noindent where $\phi (E, \, \mathbf{r}, \, t)$ is the density of cosmic-ray electrons with energy between $E$ and $E + \mathrm{d} E$ at position $\mathbf{r}$ and time $t$, $D(E)$ the energy-dependent diffusion coefficient, $Q(E, t)$ the source term, i.e. the rate of injection of cosmic-ray electrons, and $b(E)$ the mean rate at which electrons loose energy:
\begin{equation}
b(E) = -\frac{\mathrm{d} E}{\mathrm{d} \, t} .
\label{eq:be}
\end{equation}
The diffusion coefficient $D$ sets the level of turbulence that scatters cosmic-ray charged particles. It is linked to the diffusion mean free path $\lambda$ \cite{GAI90}:
\begin{equation}
D = \frac{1}{3} \lambda v_{\parallel} ,
\label{eq:d}
\end{equation}
\noindent where $v_{\parallel}$ is the projection of the electron velocity onto the magnetic line direction along which the electron moves. As long as the magnetic field is slowly evolving, cosmic-ray electrons move along the same field lines. However, the random scattering by the magnetic field inhomogeneities results in a uniform distribution of pitch angles over the range [0, $\pi/2$] \cite{LON11}. For relativistic electrons, the quantity $v_\parallel$ is then a random number between 0 and $c$, depending on the value of the pitch angle.
One of the main sources of uncertainty in diffusive propagation models comes from the diffusion coefficient $D$, which is calculated by using the observed ratios of secondary to primary cosmic-ray nuclei (e.g. B/C) in the GeV energy region. But the uncertainties are still large and, moreover, there is some lack of experimental data in the TeV energy region, which makes it difficult to discriminate among different diffusion coefficients. A widely adopted expression for relativistic particles is given by
\begin{equation}
D(E) = D_0 (E/\mathrm{GeV})^\delta ,
\label{eq:de}
\end{equation}
\noindent where $\delta = 0.3$-0.6 and $D_0 = (1$-5)$\times$10$^{28}$ cm$^2$ s$^{-1}$ \cite{STR07}. Accordingly, the diffusion mean free path of cosmic-ray electrons also depends on energy:
\begin{equation}
\lambda(E) = \frac{\lambda_0 (E/\mathrm{GeV})^\delta}{\beta_{\parallel}},
\label{eq:lambda}
\end{equation}
\noindent where $\beta_{\parallel} = v_{\parallel}/c$ and $\lambda_0 = 3 D_0 / c \simeq 1$~pc.
A simple treatment of diffusion (random walk) shows that the root mean square (RMS) of individual displacements, i.e. the average straight-line distance between the origin and the position of particles after time $t$, is given by
\begin{equation}
\sqrt{\langle r^2 \rangle} = \sqrt{6 D t} .
\label{eq:distance}
\end{equation}
\noindent Replacing $t$ with the maximum lifetime ($\tau_\mathrm{max}$) in this purely phenomenological relation, one can readily estimate the maximum distance of cosmic-ray electron sources. For $10$, $10^2$ and $10^3$~GeV, these distances are roughly equal to 7, 3 and 2~kpc, respectively. Thus, only very local sources within a few kpc from the solar system should contribute effectively to the electron spectrum at high energy.
\subsection{Monte Carlo approach}
Astrophysical sources are discrete and time-dependent, unlike dark-matter sources, which are supposed to be continuously distributed and independent of time. The electron acceleration at astrophysical sources, such as pulsars, lasts only a few tens of kyr, while the time scale (typical propagation time) is of the order of several hundred of kyr. Moreover, the sources are typically a few pc in size, while the space scale (typical source distance) is about several hundred pc. As a result, one can use the so-called burst-like approximation \cite{SHE70}, which regards astrophysical sources as point-like and instantaneous. Nonetheless, this approximation is not a crucial element of our model and a delay in electron emission with respect to the source birth can be handled quite easily. We just have to deduct this delay from the arrival time of each observed electron without the need to redo the calculations.
The calculation starts from the selected astrophysical source, where one electron is picked with an energy $E$. Unless $E$ is fixed (monochromatic source), the energy injection spectrum is sampled for more efficiency with the ziggurat algorithm \cite{MAR00}. Assuming homogeneous and isotropic diffusion, the electron is injected into space with a random three-dimensional direction. The path length $l$ is generated according to the probability distribution
\begin{equation}
P(l) \propto \exp \left( -l / \lambda \right) ,
\label{eq:P}
\end{equation}
\noindent where $\lambda$ is calculated with equation (\ref{eq:lambda}). Between two scatterings, the electron is supposed to move along a straight path because the Larmor radius ($\sim 10^{-5}$~pc) is considerably smaller than the mean free path ($\sim 1$~pc). After traveling the distance $l$, the electron energy is adjusted taking into account the energy loss rate given by relation (\ref{eq:loss_rate}). Using the approximation of a continuous energy loss, the new energy $E'$ is given by
\begin{equation}
E' = \frac{a_0 E}{\exp (a_0l/v_{\parallel}) (a_0 + b_0 E) - b_0 E} .
\label{eq:new_energy}
\end{equation}
\noindent The elapsed time is calculated by taking the instant of injection as the time origin, that is an instantaneous release of electrons is supposed. The electron continues to propagate and the whole process is iterated until one of these three events occurs:
\begin{enumerate}
\item The electron reaches the boundaries of the confinement region, in which case it is discarded (free escape) and a new electron is simulated starting from the source;
\item The electron does not reach the solar system within a maximum time (=~2$\times$10$^7$~yr) corresponding to a minimum simulation energy of 10~GeV, in which case it is also discarded and the next electron is considered;
\item The electron crosses a shell of 1~pc radius centered on the Sun, in which case the electron is considered to be observed (or detected) and all its characteristics (energy, lifetime and direction) are recorded.
\end{enumerate}
The geometry of the confinement region is fixed according to the flat halo diffusion model \cite{GIN76a}, which presumes that the Galaxy has the shape of a cylindrical slab with a radius $R$ and total height $2h$. In our case $R = 20$~kpc and $h = 4$~kpc. The distance of the Sun to the galactic center is fixed at 8.5~kpc as usual.
It is obvious that the detection shell should be as small as possible to be considered as point-like (the size of the solar system is less than one thousandth of a pc). However, the smaller the detection shell the smaller the detection efficiency and the larger the computational time. Since the mean free path of cosmic-ray electrons above 10~GeV is greater than 2~pc (\ref{eq:lambda}), the value of 1~pc used here for the radius of the detection shell is a good trade-off. This requirement ensures that no significant diffusion will occur within the detection shell.
We adopted throughout our analysis intermediate values for the diffusion coefficient ($\delta = 0.45$, $D_0 = 3 \! \! \times \! \! 10^{28}$ cm$^2$ s$^{-1}$), and mean values of the interstellar gas density (=~1~cm$^{-3}$) and the galactic radiation field density (=~1~eV~cm$^{-3}$). In addition, we only considered the average value of the galactic magnetic field (=~3~$\mu$G) to keep the model as simple as possible in its first version and save computing time. The fluctuations of the galactic magnetic field undoubtedly play a key role (see, e.g., \cite{HAR16, KIS12}); they will be the subject of a follow-up dedicated study.
For this large scale simulation program developed to run on parallel systems, the quality of the pseudo-random number generator is a crucial factor. We used Tina's Random Number Generator Library \cite{BAU14}, which is especially suited for parallel programming environment. To generate independent streams of random numbers on each processing node, we adopted the leapfrog parallelization technique. Typical calculations take 30-50~hours with 128 cores running at full speed.
\section{Results and discussion}
\label{sec:results}
\subsection{Propagation effects}
\label{sec:prop_effects}
To assess the impact of propagation we first considered monochromatic electron sources located at different distances from the solar system (with null values for the galactic longitude and latitude). We calculated the lifetime and energy distributions of the observed electrons for different values of the injected energy. Examples of such distributions are shown in figure~\ref{fig:time} and \ref{fig:energy}. The calculations were carried out so as to have $10^4$ electrons crossing the detection shell in each run, which corresponds to a total number of simulated events (electrons) in the range $10^8$-$10^9$ per run. It is clear that the farther the source the greater the number of simulated events needed to meet this requirement.
As indicated in figure~\ref{fig:time}, the lifetime distributions are characterized by a dramatic rise followed by a steady decline with a very long tail. Please note that the time scale is logarithmic. The distributions are severely right-skewed (skewness $\sim +12$ for the 300~pc source). Around 50\% of the electrons reach the solar system quite quickly, within a few tens of kyr. The rest arrive later over a much longer period ($\sim 10^4$~kyr). When the distance of the source increases, the lifetime distribution as a whole moves to the right because electrons spend more time when traveling from farther sources (figure~\ref{fig:time1}). On the contrary, when the initial energy increases, the lifetime distribution moves to the left (figure~\ref{fig:time2}). Since the mean free path increases with energy ($\sim E^\delta$), higher-energy electrons spend less time when diffusing from the same source to Earth. However, this effect is partly counterbalanced by the energy loss mechanisms, which are stronger for higher-energy electrons.
\begin{figure*}[t!]
\centering
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{fig_lifetime1.eps}
\caption{}
\label{fig:time1}
\end{subfigure}%
~
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{fig_lifetime2.eps}
\caption{}
\label{fig:time2}
\end{subfigure}
\caption{Lifetime distributions of observed cosmic-ray electrons for: (a)~a monochromatic source ($10^3$~GeV) located at different distances from the solar system; (b)~different monochromatic sources located at the same distance (300~pc) from the solar system.}
\label{fig:time}
\end{figure*}
We have quite the opposite with the energy distributions shown in figure~\ref{fig:energy}. These distributions grow very slowly up to the maximum available energy then fall off very steeply. They are also highly left-skewed (skewness $\sim -3$ for the 300~pc source). The higher-energy part of the spectrum involves the electrons reaching the solar system early, whereas the lower-energy part involves the electrons arriving late. The latter have diffused in the confinement region a much longer time and thus have lost more energy. When the distance of the source increases, the energy distribution flattens because electrons undergo more energy loss when they originate from farther sources (figure~\ref{fig:energy1}). On the other hand, when the initial energy increases, the energy distribution naturally moves to the right (figure~\ref{fig:energy2}). But it also flattens because of the energy loss processes, which are harder at higher energy.
\begin{figure*}[t!]
\centering
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{fig_energy1.eps}
\caption{}
\label{fig:energy1}
\end{subfigure}%
~
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{fig_energy2.eps}
\caption{}
\label{fig:energy2}
\end{subfigure}
\caption{Energy distributions of observed cosmic-ray electrons for: (a) a monochromatic source ($10^3$~GeV) located at different distances from the solar system; (b)~different monochromatic sources located at the same distance (300~pc) from the solar system.}
\label{fig:energy}
\end{figure*}
We can deduce from these initial calculations the following:
\begin{itemize}
\item[--] The electron lifetime distribution depends strongly on the source distance and the initial energy;
\item[--] Assuming the burst-like approximation, there exists a ``right'' timing of electron emission for each position of the source, leading to the most effective contribution to the electron spectrum. Indeed, the contribution is more significant if the source age is around the time corresponding to the peak of the lifetime distribution;
\item[--] Sources with ages beyond the time at the peak of the lifetime distribution also contribute owing to the strong right-skewness of the distributions, but much less significantly; Otherwise, if the age is far below the distribution peak time, the solar system will receive at best a very small fraction of electrons (most electrons haven't arrived yet).
\end{itemize}
\subsection{Possible astrophysical sources}
\label{sec:sources}
It is commonly believed that cosmic rays up to $\sim 10^{15}$~eV, including electrons, are accelerated through diffusive shock acceleration in SNRs \cite{ELL07}. The observation of X-ray synchrotron emission in SNRs, such as SN~1006 \cite{KOY95} and Cassiopeia A \cite{ALL97}, provides compelling evidence for the presence of high-energy electrons. The detection of TeV $\gamma$-ray emission from SN~1006 \cite{TAN98} and G347.3-0.5 \cite{MUR00}, for instance, corroborates the acceleration of electrons by SNRs to about 100~TeV. So it is quite natural to invoke these objects to interpret the cosmic-ray electron spectrum at high energy \cite{KOB01, ERL02, KOB04}. However, if SNRs can accelerate not only negative electrons but also positrons, the ratios B/C and $\bar{\mathrm p}$/p should increase with energy \cite{CHO14}. Such a rise has simply not yet been observed experimentally, with the exception of a very recent observation of an anti-proton excess by AMS-02 \cite{KOU15}.
Furthermore, the idea that the spectrum of cosmic-ray electrons is dominated at very high energy by the contribution from a few local pulsars or pulsar wind nebulae (PWN) has long been known \cite{SHE70, BOU89, AHA95}. These rapidly spinning magnetized neutron stars are believed to be excellent factories of electron-positron pairs. The underlying mechanism is that electrons are first extracted from the surface by the intense electric field induced by the rotation of the star. Then they are transformed into electron-positron pairs through electromagnetic cascades in the pulsar magnetosphere \cite{REE74, ZHA01, AMA14}. Pulsars provide a natural explanation to the peculiar features observed in the cosmic-ray electron spectrum at high energy, especially the rise in the positron fraction. However, the way electrons escape the accelerator environment has yet to be understood \cite{BLA11}.
The curves of figure~\ref{fig:sources} graphically depict the relationship between the elapsed time at the peak of the lifetime distribution and the distance of the source for different values of the initial energy. If we make no presumption in advance about the source nature and suppose a prompt release of electrons, the most likely sources of HECREs should lie very close to these curves. To determine potential astrophysical sources, all we have to do is superimpose on the same graphics pulsars and SNRs by taking their distances and ages as coordinates. As explained previously, only nearby sources within about 1~pc and aged at most $10^7$~yr have relevance. Pulsar data are taken from the catalog of the Australia Telescope National Facility (ATNF)\footnote{http://www.atnf.csiro.au/research/pulsar/psrcat/} \cite{MAN05} and SNR data from \cite{DEL10} (Table C.1) and references therein.
\begin{figure}[tbp]
\centering
\includegraphics[width=0.60\textwidth]{fig_sources.eps}
\caption{Relationship between the source distance and the electron elapsed time at the peak of the lifetime distribution for different values of the initial energy. Nearby pulsars and SNRs with distance and age as coordinates are superimposed on the same graphic.}
\label{fig:sources}
\end{figure}
As can be seen in figure~\ref{fig:sources}, Monogem and Vela are the closest ($\sim 300$~pc) among the potential sources of HECREs and are then the leading candidates. Owing their respective ages, Monogem might well be the main contributor to the electron spectrum below 100~GeV and Vela in the TeV energy range. Interestingly, these objects are SNRs with known pulsar counterparts (PSR B0656+14 and B0833-45). However, it is generally held that the produced e$^+$e$^-$ pairs are first trapped in the PWN or the SNR surrounding the source and released into the interstellar medium a few tens of kyr after the birth of the source \cite{BUS08, HOO09, PRO11}. In this sense, Vela (and PSR B1737-30) appears too young to play any role. Geminga (PSR J0633+1746) and PSR B1822-09 are the ``next-in-line`` candidates and might also contribute below 100~GeV. These results corroborate the thesis of mature SNRs/pulsars already pointed to by many earlier studies \cite{BUS08, HOO09, GRA09, LIN13, ERL13, YIN13}. Other interesting objects include PSR B1742-30 and PSR J1741-2054. The other possible sources located farther away are naturally disadvantaged since the flux obeys the inverse-square law.
A ``right'' position coupled with a ``right'' age is a necessary but certainty not sufficient condition for the sources of HECREs. The spectral profile at injection has also a major impact on the observed flux. According to the theory of cosmic ray acceleration mechanism by non-relativistic expanding shock waves in star explosions, the energy spectrum of electrons injected by the source into the interstellar medium can be expressed as a power law with a characteristic exponential energy cutoff ($E_\mathrm{cut}$):
\begin{equation}
Q(E) = Q_0 E^{-\gamma} \exp \left( - E / E_\mathrm{cut} \right) ,
\end{equation}
\noindent where $\gamma$ is the spectral index and $Q_0$ a normalization factor. This form can be used for SNRs as well as for pulsars. The parameters $\gamma$, $E_\mathrm{cut}$ and $Q_0$ are estimated on the basis of either experimental studies or theoretical considerations. The spectral index $\gamma$ is found to be $\gtrsim 2$ for SNRs and $\lesssim 2$ for PWN \cite{ELL07, ABD13}. The energy cutoff is not well known but there are indications that it lies in the TeV energy range \cite{AHA08, AHA06}. $Q_0$ can be determined from the relation
\begin{equation}
\int_0^\infty Q(E) \ E \ \mathrm{d} E = \eta W,
\end{equation}
\noindent where $W$ is the total spin-down energy of the source and $\eta$ the fraction of energy transmitted into e$^-$e$^+$ pairs. There are arguments that the maximum value of $\eta W$ is on the order of $10^{48}$-$10^{50}$~erg \cite{MAL09, KOB04}.
As a next step of our analysis, we calculated the energy spectra of the observed cosmic-ray electrons from the most interesting objects using a reasonable combination of the free injection parameters ($\gamma = 2$, $E_\mathrm{cut} = 2$~TeV and $\eta W = 10^{46}$~eV). Although it is certainly unrealistic to presume that all sources share the same values of the injection parameters, this calculation allows us to have a better picture of the relative contributions from the different sources. As illustrated in figure~\ref{fig:comparison1}, when assuming the burst-like approximation, the contribution from Vela and B1737-30 outweigh all the others with the notable exception for low energies. In fact, a young source like Vela or B1737-30 mainly contributes at high energy because high-energy electrons are the first to reach Earth (see also figure~\ref{fig:time2}). The contributions from Geminga, B1822-09 and B1742-30 are all weaker than that of Monogem while the contributions from the other objects are utterly insignificant. Monogem contributes to the electron flux mainly at $\lesssim 20$~GeV. When we accept that electrons are confined in the neighborhood of the source a certain time ($\sim 50$~kyr) before release, the signals from Vela and B1737-30 just disappear (figure~\ref{fig:comparison2}). The situation for the other sources barely changes, with Monogem still predominant, but all the signals are somewhat enhanced except for B1822-09.
The steepening at high energy in the different energy spectra is determined by $E_\mathrm{cut}$ and the age of the source. Indeed, the maximum energy $E_\mathrm{max}$ achievable by a source of age $\tau_\mathrm{S}$ can be derived from relation~(\ref{eq:be}):
\begin{equation}
E_\mathrm{max} \simeq \frac{1}{b_0 \tau_\mathrm{S}}.
\end{equation}
\noindent For Vela, $E_\mathrm{max} \simeq 24$~TeV, while for Monogem, Geminga and B1822-09, $E_\mathrm{max} \simeq 1$-2~TeV. The observation by HESS of a steepening in the cosmic-ray electron energy spectrum at $\sim 1$~TeV \cite{AHA08, AHA09} does not support a young source like Vela but rather favors middle-aged objects like Monogem, B1822-09 and Geminga.
\begin{figure*}[t!]
\centering
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=0.80\textwidth]{fig_comparison.eps}
\caption{}
\label{fig:comparison1}
\end{subfigure}%
~
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=0.80\textwidth]{fig_comparison_delay.eps}
\caption{}
\label{fig:comparison2}
\end{subfigure}
\caption{Comparison of the electron energy spectra from the candidate sources using a same set of the injection parameters ($\gamma = 2$, $E_\mathrm{cut} = 2$~TeV and $\eta W = 10^{46}$~eV) within the burst-like approximation (a) and when taking a time delay of 50~kyr before release (b).}
\label{fig:comparison}
\end{figure*}
\subsection{Electron flux and positron fraction}
\label{sec:flux}
Assuming a two-component model, we calculated the flux of cosmic-ray electrons (and the positron fraction) from Monogem. We then superimposed the contribution from this source on the otherwise featureless distant-source spectrum (background), represented here by the outcome of GALPROP. We used one of the GALPROP reference models that gives good fits to cosmic-ray data, namely the plain diffusion model \cite{STR98, PTU06}.
GALPROP is a well-known numerical package for calculating the propagation of cosmic rays (not only electrons) in a realistic distribution of matter. It gives the local interstellar spectra of the different cosmic-ray species by solving numerically the diffusion-loss equations, such as (\ref{eq:transport}). It is the most representative example of conventional models. As already mentioned, in these models it is presumed that equal amounts of secondary positrons and negative electrons are created in collisions between cosmic-ray nuclei and the interstellar medium, in addition to negative electrons directly produced at the same astrophysical sources as cosmic-ray nuclei. Although conventional models fail to match HECRE data, GALPROP is eminently suitable for the background estimation.
\begin{figure*}[t!]
\centering
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{fig_p2er.eps}
\caption{Positron fraction.}
\label{fig:positron_fraction}
\end{subfigure}%
~
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{fig_flux.eps}
\caption{Total e$^+$+e$^-$ flux.}
\label{fig:all_electrons}
\end{subfigure
\vspace{4ex}
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{fig_flux_positron.eps}
\caption{Only e$^+$ flux.}
\label{fig:positrons}
\end{subfigure}%
~
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{fig_flux_electron.eps}
\caption{Only e$^-$ flux.}
\label{fig:negative_electrons}
\end{subfigure}
\caption{Positron fraction and flux of cosmic-ray electrons as a function of energy. The dashed line represents the background (GALPROP) and the histogram in (b), (c) and (d) the contribution from Monogem. ``Monogem + GALPROP'' is intended to mean Monogem contribution on top of the background. Also shown are the experimental data from recent measurements: AMS-02 \cite{ACC14, AGU14b, AGU14a}, PAMELA \cite{ADR13, ADR11}, Fermi-LAT \cite{ACK12, ACK10b}, AMS-01 \cite{AGU07}, ATIC \cite{CHA08}, ECC \cite{KOB12}, HESS \cite{AHA08, AHA09} and PPB-BETS \cite{YOS08}.}
\label{fig:flux}
\end{figure*}
As can seen in figure~\ref{fig:flux}, the contribution from Monogem reproduces simultaneously and satisfactorily all the experimental data: the increase in the positron fraction (\ref{fig:positron_fraction}), the hardening observed in the cosmic-ray electron spectrum (\ref{fig:all_electrons}), as well as the negative electron (\ref{fig:positrons}) and the positron spectra (\ref{fig:negative_electrons}) separately. It should be stressed here that we used in all these calculations the same combination of the free injection parameters ($\gamma = 2$, $E_\mathrm{cut} = 2$~TeV, and $\eta W = 4 \times 10^{46}$~erg). The other sources, like B1822-09 and Geminga, also reproduce the observed spectra after, of course, tuning the injection parameters. Note as well that the traditional methods lead to similar results as shown by many previous studies (see, e.g., \cite{HOO09, MAL09, YUK09, DEL10, BLA11, PRO11, DIM14}). The novelty here, as aforementioned, is that we used a pure Monte Carlo treatment of the propagation of HECREs from nearby single sources, instead of solving transport equations.
\subsection{Anisotropy}
\label{sec:anisotropy}
The anisotropy in the arrival directions of cosmic-ray electrons is thought to be the smoking gun for the astrophysical scenario. High-energy electrons from local sources would indeed experience little deflection by the galactic magnetic field and the distribution of their arrival directions should then show significant anisotropy. Likewise, the observation from any direction in the sky of an excess of cosmic-ray electrons with energies high enough to minimize both geomagnetic and heliospheric effects could potentially be the signature of such nearby sources.
The benchmark method for studying anisotropy involves a multipole expansion of the electron intensity fluctuations:
\begin{equation}
\frac{I_{\mathrm e}(\theta,\, \phi) - \langle I_{\mathrm e} \rangle}{\langle I_{\mathrm e} \rangle} =
\sum_{l=0}^{\infty} \sum_{m=-l}^{l} a_{lm} Y_{lm}(\pi/2-\theta, \, \phi),
\end{equation}
\noindent where $I_{\mathrm e}(\theta,\, \phi)$ is the cosmic-ray electron intensity in the direction defined by the polar angles ($\theta,\, \phi$) and $\langle I_{\mathrm e} \rangle$ the average value over the whole sky map. The functions $Y_{lm}$ are the spherical harmonics and $a_{lm}$ their corresponding weights. The coefficients of the angular power spectrum of the fluctuations are defined as
\begin{equation}
C_l = \frac{1}{2l+1} \sum_{m=-l}^{l} |a_{lm}|^2.
\end{equation}
The amplitude of dipole anisotropy is then
\begin{equation}
\delta = 3 \sqrt{\frac{C_1}{4\pi}}.
\end{equation}
Many theoretical studies show that, even after the diffusive propagation of cosmic-ray electrons, a small dipole anisotropy should be present in the direction of the dominant nearby source at sufficiently high energies \cite{SHE71, PTU95, BUS08}. However, the data from AMS-02, Fermi-LAT and PAMELA are found to be compatible with isotropy at any angular scale \cite{AGU13, ACC14, ACK10a, ADR15}.
The assessment of the degree of anisotropy induced by candidate sources can be handled very efficiently with our Monte Carlo model. That is because we have the arrival direction of each observed electron, besides its energy and lifetime. We analyzed the distribution of arrival directions by using the spherical harmonic method in galactic coordinates within the framework of HEALPix\footnote{http://healpix.sourceforge.net/} \cite{GOR05}. This software is the standard algorithm for pixelization on the sphere, producing a subdivision of a spherical surface in which each pixel covers the same surface area (or solid angle). For all our sky maps we used a grid of 12,288 pixels corresponding to an angular resolution of about $2^\circ$ (angle subtending the centers of 2 adjacent pixels). To maximize the appearance of anisotropy we replaced the content of each pixel with the integrated number of events of all its closest neighbors, the integration extending up to a maximum angular radius defined by the considered angular scale \cite{ACK10a, ADR15}. Such correlated maps enhance sensitivity to week anisotropic signals, especially those distributed through multiple adjacent bins.
We started our analysis by assessing anisotropy in the case of a perfectly isotropic sky. For a set of $10^4$ samples of isotropic cosmic-ray electrons, similarly sized ($10^5$), we obtained an average dipole anisotropy $\langle \delta \rangle = 0.9\%$ with a standard deviation $\sigma = 0.4\%$. Afterwards we investigated the anisotropy induced by the potential sources for electrons with energies above 60~GeV (Table~\ref{tab:anisotropy}). Mollweide views for Vela and Monogem are illustrated in figure~\ref{fig:anisotropy}. We made the calculations in case of the complete absence of any background and in case of the existence of an overwhelming isotropic background (10 times the source signal). As a matter of fact, looking closely at figure~\ref{fig:all_electrons}, it can be seen that the background at 60~GeV is approximately one order of magnitude higher that the contribution from Monogem. This isotropic background may reflect the overall effect of more distant sources, as hinted by the very long tail of the lifetime distributions (see figure~\ref{fig:time}). We used for all sources the same combination of the free injection parameters ($\gamma = 2$ and $E_\mathrm{cut} = 2$~TeV). We also considered a time delay of 50~kyr with respect to the source birth except for Vela. We set the angular scale at $60^\circ$ and the sample size at $10^5$ in each case. Besides the amplitude of dipole anisotropy $\delta$, we calculated the statistical significance of the electron excess, defined here as
\begin{equation}
S = \frac{I_{\mathrm e}^{\mathrm max} - \langle I_{\mathrm e} \rangle}{\sigma}.
\end{equation}
\noindent $I_{\mathrm e}^{\mathrm max}$ is the electron intensity from the hottest pixel. This pixel does not necessarily coincide with the source pixel but always lies very close.
\begin{figure*}[t!]
\centering
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{fig_anisotropy_Vela_bkg.eps}
\end{subfigure}%
~
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{fig_anisotropy_Monogem_bkg.eps}
\end{subfigure}
\caption{Mollweide views of the dipole anisotropy induced by Vela (left) and Monogem (right) for cosmic-ray electrons with energies above 60~GeV at a $60^\circ$ angular scale.}
\label{fig:anisotropy}
\end{figure*}
\begin{table}[tbp]
\centering
\begin{tabular}{|l|c|c|c|c|c|}
\hline
& Vela & Monogem & B1822-30 & Geminga\\
\hline
$\delta$~(\%) & 6.2 (67.8) & 4.6 (47.9) & 3.5 (40.3) & 3.1 (36.1) \\
\hline
$S$~($\sigma$) & 2.15 (2.13) & 2.21 (2.15) & 2.25 (2.13) & 2.32 (2.14) \\
\hline
\end{tabular}
\caption{Amplitude of the dipole anisotropy ($\delta$) and statistical significance of the electron excess ($S$) for the candidate sources in case of the existence of a background one order of magnitude higher than the source signal and, in brackets, in case of the absence of any background.}
\label{tab:anisotropy}
\end{table}
These calculations show in the first place that in no case the electron excess is statistically significant, $S$ always being very low ($\approx 2 \sigma$). This result agrees with experimental data \cite{AGU13, ACC14, ACK10a, ADR15}. However, the obtained amplitudes of dipole anisotropy are in no way consistent with Fermi data, which set the upper limit of $\delta$ above 60~GeV at $0.5\%$ \cite{ACK10a}. The discrepancy is stronger for the young source Vela than for the middle-aged sources Monogem, Geminga and B1822-30. Admittedly, the overwhelming background strongly reduces the dipole anisotropy, but not enough to be in close agreement with experimental data. The idea that the energy spectrum of cosmic-ray electrons is dominated at high energy by one single source is definitely not supported by the observations. Excluding Vela-like sources, there are at least three objects that may play a crucial role, namely Monogem, B1822-30 and Geminga (see figure~\ref{fig:comparison2}). While Monogem and Geminga are located very close together on the sky dome, B1822-30 is roughly diametrically opposite, which would cancel out the anisotropy caused by the two first objects. Indeed, if we neglect the contribution from Geminga and assume that only Monogem and B1822-09 prevail at high energy, each contributing up to 50\%, the dipole anisotropy drops to 6\% in the absence of the background and 0.6\% in its presence. This last value is more in line with Fermi data. In sum, the non-observation of anisotropy, which challenges the astrophysical scenario, can be understood only if we presuppose the existence at high energy of at least two dominant sources in such a configuration that one nullifies the effect of the other, in addition to a significant background induced by more distant sources.
\section{Conclusion}
\label{sec:conclusion}
Regarding the spectral peculiarities observed at high energy for cosmic-ray electrons, which still resist a unified interpretation, there is now a manifest need to explore new avenues in the hope of solving this puzzle. Most often the way of tackling this problem revolves around the resolution of the transport equation describing the galactic propagation of these particles. This work demonstrates the feasibility and the relevance of fully three-dimensional time-dependent Monte Carlo simulation, which can supplement the reference method by providing additional information on the electron-by-electron fluctuations. This approach is proving to be particularly efficient at investigating more deeply the possible astrophysical origin of HECREs from nearby single sources and its benefit appears plainly in addressing the anisotropy issue. The proposed algorithm is quite simple, very flexible and highly scalable.
When looking at the lifetime and energy distributions of observed cosmic-ray electrons, we came to a first list of candidate sources that includes Vela, B1737-30, Monogem, B1822-09, Geminga and B1742-30. When assuming the burst-like approximation, we found out that the signal from a young source like Vela would outweigh all other signals and would produce an exceedingly large anisotropy amplitude, which obviously disagrees with observations. Considering a time delay of 50~kyr with respect to the source birth, we obtained a new list of candidate sources with the middle-aged Monogem taking the lead. We showed then that Monogem is able to reproduce simultaneously all the experimental data, namely the e$^+$/(e$^-$+e$^+$), e$^-$+e$^+$, e$^-$ and e$^+$ energy spectra. However, Monogem is also in conflict with the upper bounds of dipole anisotropy set by Fermi. These calculations show, in fact, that the non-observation of anisotropy does not support the thesis of one single object dominating at high energy, even with the existence of an overwhelming background. But these calculations also indicate that there are other key players, specifically B1822-09 and Geminga. While Monogem and Geminga are located close together on the sky dome, B1822-09 is situated just in the opposite side. In this way, B1822-09 could very well negate the possible anisotropy caused by Monogem and Geminga. But to be in accordance with Fermi data, we had to consider in addition a significant isotropic background likely to arise from distant sources.
Last but not least, we are aware of the shortcomings of this new software. The main line of thinking of this work is to build a workable program with a minimum version to save computing time, and make improvements afterwards. Future refinements will focus on the propagation of HECREs through more realistic magnetic field structures.
\acknowledgments
We gratefully acknowledge computing support generously provided by the Center for High Performance Computing (CHPC) in Cape Town (South Africa), and by the Research Center on Scientific and Technical Information (CERIST) in Algiers (Algeria) through the HPC platform Ibnbadis. We would also like to thank the reviewers for their relevant comments, resulting in a notable improvement of the paper.
|
\section{Introduction and summary}\label{intro}
The most fundamental theory in theoretical physics is probably M-theory, which is an eleven-dimensional theory considered to unify all of the five ten-dimensional perturbative string theories.
M2-branes and M5-branes are respectively fundamental and solitonic excitations in M-theory.
From the fundamental roles it plays in theoretical physics, we naturally expect a large number of duality relations in M-theory.
However, it is difficult to observe these dualities directly, since only little is known for this mysterious theory.
Recently, it was proposed \cite{ABJM,HLLLP2,ABJ} that the ${\mathcal N}=6$ superconformal Chern-Simons theory with gauge group U$(N_1)\times$U$(N_2)$, Chern-Simons levels $(k,-k)$ and two pairs of bifundamental matters describes the worldvolume theory of coincident $\min(N_1,N_2)$ M2-branes and $|N_2-N_1|$ fractional M2-branes on a geometry ${\mathbb C}^4/{\mathbb Z}_k$.
The partition function and vacuum expectation values of the half-BPS Wilson loop on $S^3$, originally defined with the infinite-dimensional path integral, is reduced to a finite-dimensional matrix model \cite{KWY}
\begin{align}
\langle s_Y\rangle_{k}(N_1,N_2)
&=\frac{(-1)^{\frac{1}{2}N_1(N_1-1)+\frac{1}{2}N_2(N_2-1)}}{N_1!N_2!}
\int_{{\mathbb R}^{N_1+N_2}}
\frac{d^{N_1}\mu}{(2\pi)^{N_1}}\frac{d^{N_2}\nu}{(2\pi)^{N_2}}
e^{\frac{ik}{4\pi}(\sum_{m=1}^{N_1}\mu_m^2-\sum_{n=1}^{N_2}\nu_n^2)}\nonumber\\
&\qquad\times\biggl[\frac{\prod_{m<m'}^{N_1}2\sinh\frac{\mu_m-\mu_{m'}}{2}
\prod_{n<n'}^{N_2}2\sinh\frac{\nu_n-\nu_{n'}}{2}}
{\prod_{m=1}^{N_1}\prod_{n=1}^{N_2}2\cosh\frac{\mu_m-\nu_n}{2}}\biggr]^2
s_Y(e^\mu|e^\nu),
\label{vev}
\end{align}
and a hidden super gauge group U$(N_1|N_2)$ was observed \cite{GW,HLLLP2,ABJ,KWY,DT}.
Here $s_Y(x|y)$ is the supersymmetric Schur polynomial or the character of U$(N_1|N_2)$.
We directly observe a relation under the exchange of two sets of integration variables,
\begin{align}
\langle s_Y\rangle_{k}(N_1,N_2)=\langle s_{Y^\text{T}}\rangle_{-k}(N_2,N_1)
=[\langle s_{Y^\text{T}}\rangle_{k}(N_2,N_1)]^*,
\label{conj}
\end{align}
because of $s_Y(x|y)=s_{Y^\text{T}}(y|x)$.
Hereafter we shall fix $k>0$.
Also, we often consider the case of $M=N_2-N_1\ge 0$ unless otherwise stated and denote the expectation value $\langle s_Y\rangle_{k}(N,M+N)$ as $\langle s_Y\rangle_{k,M}(N)$.
It was then exciting to find that the partition function (\cite{MPtop,MP,HMO2,HMO3,CM,HMMO} for the case of equal ranks and \cite{sho,HoOk} for non-equal ranks) and vacuum expectation values of the half-BPS Wilson loop \cite{KMSS,HHMO} are respectively expressed in terms of the free energy of the closed and open topological string theories on local ${\mathbb P}^1\times{\mathbb P}^1$, which implies a certain modular invariance.
Aside from the original computations in the 't Hooft expansion \cite{DMP1,DMP2,FHM}, an important approach that leads to these findings is to rewrite the matrix model into the partition function of a Fermi gas system with $N$ non-interacting particles whose dynamics is governed by a non-trivial density matrix \cite{MP}.
The success in formulating the partition function in terms of that of the Fermi gas system leads to a vast amount of WKB small $k$ expansions \cite{MP,CM} and numerical computations for finite $k$ \cite{KEK,HMO1,PY,HMO2}, from which the relation to the topological strings was found \cite{MPtop,HMO2,HMMO}.
For the partition function in a general background with $M=N_2-N_1$ fractional branes, it was found that
\begin{align}
\frac{\langle 1\rangle_{k,M}(N)}{\langle 1\rangle_{k,M}(0)}
=\frac{1}{N!}\int\frac{d^Nx}{(4\pi k)^N}
\prod_{i<j}^N\biggl(\tanh\frac{x_i-x_j}{2k}\biggr)^2\prod_{i=1}^NV_M(x_i),
\label{pf}
\end{align}
where $V_M(x)$ is defined as
\begin{align}
V_M(x)=\frac{1}{2\cosh\frac{x}{2}}
\prod_{\overline l\in\overline L}\tanh\frac{x+2\pi i\overline l}{2k},
\label{pfV}
\end{align}
with $\overline L=\{M-\frac{1}{2},M-\frac{3}{2},\cdots,\frac{3}{2},\frac{1}{2}\}$.
If we introduce the coordinate and momentum operators satisfying the canonical commutation relation $[\widehat q,\widehat p]=i\hbar$ with $\hbar=2\pi k$, the grand canonical partition function $\langle 1\rangle^\text{GC}_{k,M}(z)=\sum_{N=0}^\infty z^{N}\langle 1\rangle_{k,M}(N)$ is expressed as
\begin{align}
\frac{\langle 1\rangle^\text{GC}_{k,M}(z)}{\langle 1\rangle_{k,M}(0)}
=\det(1+z\widehat\rho_M),
\label{detrho}
\end{align}
with the density matrix $\widehat\rho_M$ given by
\begin{align}
\widehat\rho_M=\sqrt{V_M(\widehat q)}
\frac{1}{2\cosh\frac{\widehat p}{2}}\sqrt{V_M(\widehat q)}.
\label{rho}
\end{align}
The expression \eqref{pf} was first found for the case of equal ranks $M=0$ in \cite{MP} and later extended to the case of non-equal ranks.
Namely, for $M\ne 0$, \eqref{pf} was originally conjectured in \cite{AHS} and proved in \cite{Ho1} with several steps of integrations.
In this paper we shall rederive the result with a more refined presentation motivated by \cite{MS2,MN5} as a byproduct of our analysis.
In the expansion of the determinant in \eqref{detrho} there appear many traces of powers of the density matrix $\widehat\rho_M$.
In introducing a background with $M$ fractional branes, all we have to do is to modify the density matrix $\widehat\rho_M$ \eqref{rho} by changing $V_{M=0}(x)$ into $V_{M}(x)$ \eqref{pfV} without touching the structure of the determinant.
In other words, we express the fractional branes by dressing the density matrix so that the background where the closed strings propagate is changed.
From this viewpoint, this formalism is called ``closed string formalism'' in \cite{PTEP}.
Since some poles of $(2\cosh\frac{x}{2})^{-1}$ in \eqref{pfV} at $x\in 2\pi i({\mathbb Z}\pm\frac{1}{2})$ are cancelled by the zeros of the hyperbolic tangent functions at $x=-2\pi i\overline l$, we can shift the integration contour by $-M\pi i$ \cite{HoOk}.
This shift is essential to make contact with the orthosymplectic Chern-Simons matrix model \cite{MePu,MS1,Ho2,Ok2,MS2,MN5}.
This matrix model is obtained from the localization of the ${\mathcal N}=5$ superconformal Chern-Simons theory with the orthosymplectic gauge group \cite{HLLLP2,ABJ} and the physical interpretation is the introduction of the orientifold plane in the type IIB setup.
In \cite{Ho2,MS2,MN5} it was proved\footnote{The proof for odd $M$ is motivated by the studies in the Chern-Simons matrix models of the $\widehat D$ quiver \cite{CHJ,ADF,MN4}.} that the partition function of the orthosymplectic Chern-Simons matrix model is nothing but the chiral projection of the Chern-Simons matrix model with the super unitary gauge group
(see table 1 in \cite{MN5} for an interesting pattern).
There is another formalism to study the matrix model \cite{sho} called ``open string formalism''.
Here we do not change the expression of the density matrix from the $M=0$ case and instead introduce many extra contributions with endpoints.
In this sense, as the Wilson loop expectation values \cite{HHMO}, we express the fractional brane background with the open string endpoints.
At present we stress that the open string formalism seems superior to the closed string one because it is obtained only from the combinatorics and hence is applicable not only for the partition function but also with the Wilson loop insertion.
In this formalism the expectation values of the half-BPS Wilson loop in the grand canonical ensemble\footnote{The integer $r$ (satisfying $0\le r\le N$) is defined later in \eqref{armlegdef} for a general Young diagram $Y$.
It is known that the supersymmetric Schur polynomial $s_Y(x|y)$ is vanishing for $N<r$.}
\begin{align}
\langle s_Y\rangle^\text{GC}_{k,M}(z)
=\sum_{N=r}^\infty z^{N-r}
\langle s_Y\rangle_{k,M}(N),
\label{sYvevGC}
\end{align}
is reduced to
\begin{align}
\frac{\langle s_Y\rangle^\text{GC}_{k,M}(z)}{\langle 1\rangle^\text{GC}_{k,0}(z)}
=\det\begin{pmatrix}\bigl(H_{l_p,-M+q-\frac{1}{2}}(z)\bigr)_{(M+r)\times M}&
\bigl(\widetilde H_{l_p,a_q}(z)\bigr)_{(M+r)\times r}\end{pmatrix},
\label{open}
\end{align}
where both $H_{l,a}(z)$ and $\widetilde H_{l,a}(z)$ take the form of a certain matrix element of $[1+z\widehat\rho_{M=0}]^{-1}$ (see \cite{sho} for the explicit form\footnote{We have slightly changed the notation from \cite{sho}.
In addition to changing the definition of the arm and leg lengths by $1/2$ as explained later in \eqref{armlegdef}, we also drop the overall factor $z$ from $\widetilde H_{l,a}(z)$.}).
The indices $a_q$, $l_p$ in \eqref{open} are the arm lengths and the leg lengths appearing in the Frobenius notation, which is another description of the Young diagram usually described by listing all of the arm lengths $[\alpha_1,\alpha_2,\cdots]$ or the leg lengths $[\lambda_1,\lambda_2,\cdots]^\text{T}$,
\begin{align}
&a_q=\alpha_q-q-M+1/2,\quad
l_p=\lambda_p-p+M+1/2,\nonumber\\
&r=\max\{q|a_q>0\}=\max\{p|l_p>0\}-M.
\label{armlegdef}
\end{align}
See figure \ref{young} for a pictorial explanation of the Frobenius notation.
Note that here we have deliberately added $\frac{1}{2}$ to the lengths to measure the distances between the midpoints of two segments.
As stressed in \cite{HHMO,HaOk}, the integrations in $H_{l,a}(z)$ and $\widetilde H_{l,a}(z)$ are convergent only for $a_1+l_1<k/2$ (which implies $M\le k/2$).
We shall follow this condition in our analysis.
One advantage of the open string formalism \eqref{open} is that we can prove the Giambelli compatibility for $\langle s_Y\rangle^\text{GC}_{k,M}(z)$ generally when correctly normalized with $\langle 1\rangle^\text{GC}_{k,M}(z)$ \cite{satsuki}.
Although many identities were proved in this context, still a lot of important duality relations await to be proved.
One of them is the miraculous open-closed duality observed recently in \cite{HaOk}.
In \cite{HaOk}, motivated by \cite{ADKMV}, starting from the simplest case with $r=1,M=0$, the authors arrive at a more general relation\footnote{In \cite{HaOk} the absolute values were taken for the expectation values in defining the grand canonical ensemble \eqref{sYvevGC}.
Hence, strictly speaking, the duality relation found in \cite{HaOk} is a consequence of \eqref{openclosed}.
Similarly, the relation \eqref{hook} also needs the modification of the complex conjugation.}
\begin{align}
\langle s_{[(M+r)^r]}\rangle^\text{GC}_{k,M}(z)
\sim\langle 1\rangle^\text{GC}_{k,M+2r}(z),
\label{openclosed}
\end{align}
with numerical computations.
Here $\sim$ means that the relation holds up to a numerical factor independent of $z$.
This duality relates the closed string BPS indices to the open string BPS indices and is another realization of the spirit of the open string formalism \cite{sho}, which expresses the closed string background formed by fractional branes with many open strings in the determinant.
In the same paper, the authors also observe a relation
\begin{align}
\langle s_{(a|l)}\rangle^\text{GC}_{k,M}(z)
\sim[\langle s_{(l+M|a-M)}\rangle^\text{GC}_{k,M}(z)]^*,
\label{hook}
\end{align}
for the hook representation $(a|l)$ with $a>M$.
The complex conjugation applies only for the coefficients of $z$.
\begin{figure}[!ht]
\centering\includegraphics[scale=1]{young.eps}
\caption{Young diagram and Frobenius notation.
The standard Frobenius notation is defined by counting the boxes from the diagonal line.
For the super case U$(N_1|N_2)$, we shift the diagonal line by $M=N_2-N_1$.
For the Young diagram in the figure $[\alpha_1,\alpha_2,\alpha_3]=[5,4,2]$ or $[\lambda_1,\lambda_2,\lambda_3,\lambda_4,\lambda_5]^\text{T}=[3,3,2,2,1]^\text{T}$, the standard Frobenius notation is $(42|21)$ and the shifted one with $M=2$ is $(20|4310)$.
Here we find it useful to define the arm and leg lengths by adding $\frac{1}{2}$, or in other words, measuring the distances between the midpoints of two segments,
$(a_1,a_2|l_1,l_2,l_3,l_4)=(\frac{5}{2},\frac{1}{2}|\frac{9}{2},\frac{7}{2},\frac{3}{2},\frac{1}{2})$.
We sometimes decompose the Young diagram as $([(M+r)^r]+Y)\cup Y'$, which in the current case is $([4^2]+[1])\cup[2]$.
We also display the poles of $(2\cosh\frac{x}{2})^{-1}$ in \eqref{V} in our Fermi gas formalism \eqref{fermigas}.
The green dots denote the poles at $x=2\pi ia$ and $x=-2\pi il$ which are cancelled by the hyperbolic tangent functions in \eqref{V} and hence harmless, while the red dots denote the real poles which are not cancelled.}
\label{young}
\end{figure}
In this paper, we shall generalize the closed string formalism \eqref{pf}, so that it incorporates the Wilson loop insertion.
Namely, contrary to the open string formalism \cite{sho} where we describe the fractional brane as a composite of the Wilson loops, here we propose an opposite formalism, which describes the Wilson loop by changing the closed string backgrounds,
\begin{align}
\frac{\langle s_Y\rangle_{k,M}(N)}{\langle s_Y\rangle_{k,M}(r)}
=\frac{1}{(N-r)!}\int\frac{d^{N-r}x}{(4\pi k)^{N-r}}
\prod_{i<j}^{N-r}\biggl(\tanh\frac{x_i-x_j}{2k}\biggr)^2
\prod_{i=1}^{N-r}V(x_i),
\label{fermigas}
\end{align}
with
\begin{align}
V(x)=\frac{1}{2\cosh\frac{x}{2}}
\prod_{a\in A}\tanh\frac{x-2\pi ia}{2k}
\prod_{l\in L}\tanh\frac{x+2\pi il}{2k}.
\label{V}
\end{align}
Here $A$ and $L$ denote respectively the set of all arm lengths and all leg lengths of the Young diagram $Y$.
Note that $\overline L$ appearing in \eqref{pfV} is the set of all leg lengths in the trivial representation.
From \eqref{V} it is easy to observe that although $x\in 2\pi i({\mathbb Z}\pm\frac{1}{2})$ are poles potentially, poles at $x=2\pi ia$ and $x=-2\pi il$ are cancelled by the zeros of the hyperbolic tangent functions (see figure \ref{young}).
Also, since $a_1+l_1<k/2$, a hyperbolic tangent does not induce a new pole at the zero of another hyperbolic tangent.
Using this new formalism we are able to prove some untouched dualities without difficulties.
Here we prove a generalized open-closed duality\footnote{The generalization without $Y$ was already observed in \cite{HaOk}.}
\begin{align}
\langle s_{([(M+r)^r]+Y)\cup Y'}\rangle^\text{GC}_{k,M}(z)
\sim\langle s_{([(M+r-1)^{r+1}]+Y)\cup Y'}\rangle^\text{GC}_{k,M-2}(z).
\label{generalized}
\end{align}
In fact, from the expression \eqref{fermigas} it is not difficult to find that we can shift the integration contour as long as we do not cross the poles.
Due to this reason, the expectation values of two Young diagrams in the grand canonical ensemble, which share the same set of
\begin{align}
\{a_q|q=1,\cdots,r\}\cup\{-l_p|p=1,\cdots,M+r\},
\label{A-L}
\end{align}
up to an overall shift of an integer, are identical.
This paper is organized as follows.
In the next section, we propose a new formalism to study the Wilson loop expectation values.
In section \ref{openclosedduality}, we use this formalism to prove (a generalized version of) the open-closed duality.
Finally in section \ref{discuss}, we conclude with discussions on the future directions.
\section{Wilson loop as closed string background}
In this section, following the method developed in \cite{MS2,MN5}, we shall rewrite the expectation value of the half-BPS Wilson loop \eqref{vev} into the partition function of a new Fermi gas system, where the density matrix is modified while the determinant structure is kept fixed.
We shall rescale the integration variables by $k$ as
\begin{align}
\langle s_Y\rangle_{k,M}(N)
&=\frac{(-1)^{\frac{1}{2}N_1(N_1-1)+\frac{1}{2}N_2(N_2-1)}}{N_1!N_2!}
\int\frac{d^{N_1}\mu}{\hbar^{N_1}}\frac{d^{N_2}\nu}{\hbar^{N_2}}
e^{\frac{i}{2\hbar}(\sum_{m=1}^{N_1}\mu_m^2-\sum_{n=1}^{N_2}\nu_n^2)}\nonumber\\
&\qquad\times\biggl[\frac{\prod_{m<m'}^{N_1}2\sinh\frac{\mu_m-\mu_{m'}}{2k}
\prod_{n<n'}^{N_2}2\sinh\frac{\nu_n-\nu_{n'}}{2k}}
{\prod_{m=1}^{N_1}\prod_{n=1}^{N_2}2\cosh\frac{\mu_m-\nu_n}{2k}}\biggr]^2
s_Y(e^{\frac{\mu}{k}}|e^{\frac{\nu}{k}}),
\end{align}
with $\hbar=2\pi k$ and $k>0$.
We begin our analysis by assuming $M=N_2-N_1\ge 0$ and denote $N_1=N$, $N_2=M+N$ and $\langle s_Y\rangle_k(N,M+N)=\langle s_Y\rangle_{k,M}(N)$.
As in the open string formalism \cite{sho}, our starting point is the following three determinant formulas; the Cauchy-Vandermonde determinant
\begin{align}
\frac{\prod_{m<m'}^{N_1}(x_m-x_{m'})
\prod_{n<n'}^{N_2}(y_n-y_{n'})}
{\prod_{m=1}^{N_1}\prod_{n=1}^{N_2}(x_m+y_n)}
=(-1)^{N_1(N_2-N_1)}
\det\begin{pmatrix}\biggl[\displaystyle\frac{1}{x_m+y_n}\biggr]
_{(m,n)\in Z_{1}\times Z_{2}}\\
\Bigl[y_n^{\overline l-\frac{1}{2}}\Bigr]
_{(\overline l,n)\in\overline L\times Z_{2}}
\end{pmatrix},
\end{align}
where $m$, $n$ and $\overline l$ are respectively elements of $Z_1=\{1,2,\cdots,N_1\}$, $Z_2=\{1,2,\cdots,N_2\}$ and $\overline L=\{M-\frac{1}{2},M-\frac{3}{2},\cdots,\frac{1}{2}\}$ in this order;
the same determinant
\begin{align}
&(-1)^{\frac{1}{2}N_1(N_1-1)+\frac{1}{2}N_2(N_2-1)}
\frac{\prod_{m<m'}^{N_1}(x_{m'}^{-1}-x_m^{-1})
\prod_{n<n'}^{N_2}(y_{n'}^{-1}-y_n^{-1})}
{\prod_{m=1}^{N_1}\prod_{n=1}^{N_2}(x_m^{-1}+y_n^{-1})}\nonumber\\
&=(-1)^{N_1(N_2-N_1)}
\det\begin{pmatrix}\biggl[\displaystyle\frac{1}{y_n^{-1}+x_m^{-1}}\biggr]
_{(n,m)\in Z_{2}\times Z_{1}}&
\Bigl[y_n^{-\overline l+\frac{1}{2}}\Bigr]
_{(n,\overline l)\in Z_{2}\times\overline L}
\end{pmatrix},
\end{align}
obtained by the substitutions $x_m\to x_m^{-1}$ and $y_n\to y_n^{-1}$;
the determinantal formula for the supersymmetric Schur polynomial due to Moens and Van der Jeugt \cite{MVdJ}
\begin{align}
\frac{s_Y(x|y)}{(-1)^r}
=\det\begin{pmatrix}\biggl[\displaystyle\frac{1}{x_m+y_n}\biggr]
_{(m,n)\in Z_{1}\times Z_{2}}&
\Bigl[x_m^{a-\frac{1}{2}}\Bigr]
_{(m,a)\in Z_{1}\times A}\\
\Bigl[y_n^{l-\frac{1}{2}}\Bigr]
_{(l,n)\in L\times Z_{2}}&
[0]_{L\times A}\end{pmatrix}\bigg/
\det\begin{pmatrix}\biggl[\displaystyle\frac{1}{x_m+y_n}\biggr]
_{(m,n)\in Z_{1}\times Z_{2}}\\
\Bigl[y_n^{\overline l-\frac{1}{2}}\Bigr]
_{(\overline l,n)\in\overline L\times Z_{2}}\end{pmatrix},
\end{align}
with $A=\{a_1,a_2,\cdots,a_r\}$ and $L=\{l_1,l_2,\cdots,l_{M+r}\}$, where, as in figure \ref{young}, the arm and leg lengths measure the distances between the midpoints of two segments and are greater than the standard ones by $\frac{1}{2}$.
After multiplying these three formulas with the substitutions $x_m=e^{\frac{\mu_m}{k}}$ and $y_n=e^{\frac{\nu_n}{k}}$, we find
\begin{align}
&(-1)^{\frac{1}{2}N_1(N_1-1)+\frac{1}{2}N_2(N_2-1)}
\biggl[\frac{\prod_{m<m'}^{N_1}2\sinh\frac{\mu_m-\mu_{m'}}{2k}
\prod_{n<n'}^{N_2}2\sinh\frac{\nu_n-\nu_{n'}}{2k}}
{\prod_{m=1}^{N_1}\prod_{n=1}^{N_2}2\cosh\frac{\mu_m-\nu_n}{2k}}\biggr]^2
\frac{s_Y(e^{\frac{\mu}{k}}|e^{\frac{\nu}{k}})}{(-1)^r}\nonumber\\
&=\det\begin{pmatrix}
\biggl[\displaystyle\frac{1}{2\cosh\frac{\mu_m-\nu_n}{2k}}\biggr]_{N_1\times N_2}&
\bigl[e^{\frac{a\mu_m}{k}}\bigr]_{N_1\times r}\\
\bigl[e^{\frac{l\nu_n}{k}}\bigr]_{(M+r)\times N_2}&
[0]_{(M+r)\times r}
\end{pmatrix}
\det\begin{pmatrix}
\biggl[\displaystyle\frac{1}{2\cosh\frac{\nu_n-\mu_m}{2k}}\biggr]_{N_2\times N_1}
&\bigl[e^{-\frac{\overline l\nu_n}{k}}\bigr]_{N_2\times M}
\end{pmatrix}.
\end{align}
Next, let us rewrite the components of the determinants by using the Fourier transformation of $(2\cosh\frac{\widehat p}{2})^{-1}$ and introducing the formal states $|h\rrangle$ and $\llangle h|$ as in \cite{MS2}, $(h\in{\mathbb Z}\pm\frac{1}{2})$
\begin{align}
\langle\mu|\frac{1}{2\cosh\frac{\widehat p}{2}}|\nu\rangle
=\frac{1}{k}\frac{1}{2\cosh\frac{\mu-\nu}{2k}},\qquad
\llangle h|\mu\rangle=\langle\mu|h\rrangle=e^{\frac{h\mu}{k}}.
\end{align}
Then, we can follow the standard tricks of including the Gaussian factor into the brackets and applying the similarity transformation
\begin{align}
1=\int\frac{dq}{2\pi}|q\rangle\langle q|\quad\Rightarrow\quad
1=\int\frac{dq}{2\pi}e^{-\frac{i}{2\hbar}\widehat p^2}|q\rangle
\langle q|e^{\frac{i}{2\hbar}\widehat p^2},
\end{align}
to the $\mu$ and $\nu$ integrations.
After these manipulations, the expectation value is expressed as
\begin{align}
&\langle s_Y\rangle_{k,M}(N)
=\frac{(-1)^r}{N_1!N_2!}\int\frac{d^{N_1}\mu}{\hbar^{N_1}}
\frac{d^{N_2}\nu}{\hbar^{N_2}}\nonumber\\
&\quad\times\det\begin{pmatrix}
\biggl[\displaystyle k\langle\mu_m|e^{\frac{i}{2\hbar}\widehat p^2}
e^{\frac{i}{2\hbar}\widehat q^2}\frac{1}{2\cosh\frac{\widehat p}{2}}
e^{-\frac{i}{2\hbar}\widehat q^2}e^{-\frac{i}{2\hbar}\widehat p^2}|\nu_n\rangle
\biggr]_{N_1\times N_2}
&\Bigl[\langle\mu_m|e^{\frac{i}{2\hbar}\widehat p^2}
e^{\frac{i}{2\hbar}\widehat q^2}|a\rrangle\Bigr]_{N_1\times r}\\
\Bigl[\llangle l|e^{-\frac{i}{2\hbar}\widehat q^2}
e^{-\frac{i}{2\hbar}\widehat p^2}|\nu_n\rangle\Bigr]_{(M+r)\times N_2}&
\bigl[0\bigr]_{(M+r)\times r}
\end{pmatrix}\nonumber\\
&\quad\times\det\begin{pmatrix}
\biggl[\displaystyle k\langle\nu_n|e^{\frac{i}{2\hbar}\widehat p^2}
\frac{1}{2\cosh\frac{\widehat p}{2}}e^{-\frac{i}{2\hbar}\widehat p^2}
|\mu_m\rangle\biggr]_{N_2\times N_1}&
\Bigl[\langle\nu_n|e^{\frac{i}{2\hbar}\widehat p^2}
|{-\overline l}\rrangle\Bigr]_{N_2\times M}
\end{pmatrix}.
\end{align}
It is magical \cite{MS2} that all of the components in the first determinant reduce to delta functions\footnote{The formal computation in the second and third formulas needs justification \cite{MS2}.
It is important that the remaining factors $(2\cosh\frac{\mu-\nu}{2k})^{-1}$ do not contain poles between $\im\mu=2\pi a_1$ and $\im\nu=-2\pi l_1$ due to the condition $a_1+l_1<k/2$.
We are grateful to Takao Suyama for valuable discussions.}
\begin{align}
k\langle\mu|e^{\frac{i}{2\hbar}\widehat p^2}e^{\frac{i}{2\hbar}\widehat q^2}
\frac{1}{2\cosh\frac{\widehat p}{2}}
e^{-\frac{i}{2\hbar}\widehat q^2}e^{-\frac{i}{2\hbar}\widehat p^2}|\nu\rangle
&=\frac{\hbar}{2\cosh\frac{\mu}{2}}\delta(\mu-\nu),
\nonumber\\
\langle\mu|e^{\frac{i}{2\hbar}\widehat p^2}
e^{\frac{i}{2\hbar}\widehat q^2}|a\rrangle
&=\frac{\hbar}{\sqrt{-ik}}e^{\frac{i}{2\hbar}(2\pi a)^2}\delta(\mu-2\pi ia),
\nonumber\\
\llangle l|e^{-\frac{i}{2\hbar}\widehat q^2}
e^{-\frac{i}{2\hbar}\widehat p^2}|\nu\rangle
&=\frac{\hbar}{\sqrt{ik}}e^{-\frac{i}{2\hbar}(2\pi l)^2}\delta(\nu+2\pi il).
\end{align}
For the second determinant we have
\begin{align}
\langle\nu|e^{\frac{i}{2\hbar}\widehat p^2}|{-\overline l}\rrangle
&=e^{-\frac{i}{2\hbar}(2\pi\overline l)^2}
\langle\nu|{-\overline l}\rrangle.
\end{align}
For the expansion of the first determinant to be non-vanishing we need to choose $r$ rows out of $N$ rows in the upper-right block and $M+r$ columns out of $M+N$ columns in the lower-left block.
Then, the remaining $N-r$ components are chosen from the upper-left block.
In renaming the indices, we have $r!(M+r)!(N-r)!$ identical terms with signs $(-1)^{(M+r)r}$.
After combining these factors, we find
\begin{align}
&\langle s_Y\rangle_{k,M}(N)=\frac{(-1)^{\frac{1}{2}M(M-1)+Mr}}{(N-r)!}
\int\frac{d^N\mu}{\hbar^N}\frac{d^{M+N}\nu}{\hbar^{M+N}}
e^{\frac{i}{2\hbar}(2\pi)^2(\sum a^2-\sum l^2-\sum\overline l^2)}\nonumber\\
&\qquad\times
\prod_{i=1}^{N-r}\frac{\hbar}{2\cosh\frac{\mu_i}{2}}\delta(\mu_i-\nu_i)
\prod_{q=1}^r\frac{\hbar}{\sqrt{-ik}}\delta(\mu_{N-r+q}-2\pi ia_q)
\prod_{p=1}^{M+r}\frac{\hbar}{\sqrt{ik}}\delta(\nu_{N-r+p}+2\pi il_p)
\nonumber\\
&\qquad\times
e^{\frac{M}{2k}(\sum_{m=1}^N\mu_m-\sum_{n=1}^{M+N}\nu_n)}
\frac{\prod_{m<m'}^{N}2\sinh\frac{\mu_m-\mu_{m'}}{2k}
\prod_{n<n'}^{M+N}2\sinh\frac{\nu_n-\nu_{n'}}{2k}}
{\prod_{m=1}^N\prod_{n=1}^{M+N}2\cosh\frac{\mu_m-\nu_n}{2k}},
\end{align}
where we have used $(-1)^{r^2}=(-1)^r$.
After performing the integration of the delta functions by substitutions, we arrive at the expression
\begin{align}
&\frac{\langle s_Y\rangle_{k,M}(N)}
{\langle s_Y\rangle_{k,M}(r)}
=\frac{1}{(N-r)!}\int\frac{d^{N-r}x}{(4\pi k)^{N-r}}
\prod_{i<j}^{N-r}\biggl(\tanh\frac{x_i-x_j}{2k}\biggr)^2\nonumber\\
&\qquad\times\prod_{i=1}^{N-r}\biggl[\frac{1}{2\cosh\frac{x_i}{2}}
\prod_{a\in A}\tanh\frac{x_i-2\pi ia}{2k}
\prod_{l\in L}\tanh\frac{x_i+2\pi il}{2k}\biggr],
\label{closed}
\end{align}
where the normalization is given by\footnote{Note that $N=r$ is the smallest case for the expectation value $\langle s_Y\rangle_{k,M}(N)$ to be non-vanishing.
The absolute value $|\langle s_Y\rangle_{k,M}(r)|$ is coincident with $C_Y(k,M)$ in \cite{HaOk}.}
\begin{align}
\langle s_Y\rangle_{k,M}(r)
&=\frac{i^{\frac{1}{2}M(M-1)+Mr}
e^{\frac{\pi i}{k}(\sum a^2-\sum l^2-\sum\overline l^2)}
e^{\frac{\pi i}{k}M(\sum a+\sum l)}}{\sqrt{-ik}^r\sqrt{ik}^{M+r}}
\frac{\prod_{a>a'}2\sin\frac{\pi(a-a')}{k}
\prod_{l>l'}2\sin\frac{\pi(l-l')}{k}}
{\prod_a\prod_l2\cos\frac{\pi(a+l)}{k}}.
\label{rclosed}
\end{align}
We note that, although we originally start with the situation of $N_1\le N_2$, both of the formulas \eqref{closed} and \eqref{rclosed} are valid for the opposite case $N_1>N_2$ as well\footnote{
Instead of the leg lengths for the trivial representation $\overline L$, we introduce the arm lengths $\overline a\in\overline A=\{-M-\frac{1}{2},-M-\frac{3}{2},\cdots,\frac{1}{2}\}$ for the case $M<0$.
Hence, we need to interpret the phase factor $e^{\frac{\pi i}{k}(-\sum\overline l^2)}$ in \eqref{rclosed} as $e^{\frac{\pi i}{k}(+\sum\overline a^2)}$.
} if we stick to the original notation $N_1=N$, $N_2=M+N$, $\max\{q|a_q>0\}=r$, $\max\{p|l_p>0\}=M+r$ and $\langle s_Y\rangle_{k,M}(N)=\langle s_Y\rangle_k(N,M+N)$ but $M<0$.
Therefore, no special care is needed when crossing the diagonal line and we can extend to $M<0$ directly.
Finally let us comment on the closed string formalism for the partition function \eqref{pf}.
In \cite{Ho1} the proof was done by several steps of integrations.
Following the method of \cite{MS2,MN5}, we have generalized the formalism for the expectation values of the half-BPS Wilson loop.
The result of the partition function can be simply rederived by setting $A=\emptyset, L=L'$ in \eqref{closed}.
\section{Proof of generalized open-closed duality}\label{openclosedduality}
In the previous section we have found that, after suitably normalized, the expectation value $\langle s_Y\rangle_{k,M}(N)$ is given in \eqref{fermigas} with $V(x)$ defined by \eqref{V} and no special care is needed when crossing $M=0$.
From figure \ref{young} we know that $V(x)$ contains poles periodically, though some of them are cancelled by the zeros of the hyperbolic tangent functions.
Therefore, as long as we do not encounter the real poles, we can shift the integration contour or in contrast the position of the poles by $\pm 2\pi i$ freely, so that two expectation values which share the same set of \eqref{A-L} up to an integral shift are identical.
This identity induces the duality relation we want to prove.
Let us see this explicitly for the example in figure \ref{young}.
Here we assume that $k$ is large enough so that we do not have to consider the poles of the hyperbolic tangent functions in our shift of the integration contour.
In figure \ref{shift} we pick up the example in figure \ref{young} and shift the integration contour in the unit of $2\pi i$.
Starting from $(\frac{5}{2},\frac{1}{2}|\frac{9}{2},\frac{7}{2},\frac{3}{2},\frac{1}{2})$, if we move the integration coutour upwards (so that the number of leg lengths increases), we find $(\frac{3}{2}|\frac{11}{2},\frac{9}{2},\frac{5}{2},\frac{3}{2},\frac{1}{2})$, where in this shift we only move across the harmless green pole ($x=\pi i$ in figure \ref{young}) and the expectation value is not changed.
If we move further into $(\frac{1}{2}|\frac{13}{2},\frac{11}{2},\frac{7}{2},\frac{5}{2},\frac{3}{2})$, now we need to cross the real red pole ($x=3\pi i$ in figure \ref{young}) and the expectation value is changed.
Similarly, we can move downwards to $(\frac{7}{2},\frac{3}{2},\frac{1}{2}|\frac{7}{2},\frac{5}{2},\frac{1}{2})$ and so on without changing the expectation values for a while.
We have classified the expectation values by the shaded green backgrounds in figure \ref{shift}.
\begin{figure}[ht!]
\centering\includegraphics[scale=1]{shift.eps}\\
\vspace{-41mm}\hspace{23mm}{\color{white}Figure \ref{young}}
\vspace{41mm}
\caption{Shifting the integration contour across the harmless poles does not change the expectation values and hence induces the duality, while shifting across the poles is not allowed.
We assume that $k$ is large enough.}
\label{shift}
\end{figure}
We can summarize the above duality relation as
\begin{align}
\frac{\langle s_{([(M+r)^r]+Y)\cup Y'}\rangle^\text{GC}_{k,M}(z)}
{\langle s_{([(M+r)^r]+Y)\cup Y'}\rangle_{k,M}(r)}
=\frac{\langle s_{([(M+r-1)^{r+1}]+Y)\cup Y'}\rangle^\text{GC}_{k,M-2}(z)}
{\langle s_{([(M+r-1)^{r+1}]+Y)\cup Y'}\rangle_{k,M-2}(r+1)},
\end{align}
if $\lambda_1\le r$ and $\alpha_1\le M+r-1$ where $\lambda_1=\lambda_1(Y)$ and $\alpha_1=\alpha_1(Y')$ denote the first leg length of $Y$ and the first arm length of $Y'$.
Using this recursively, we find
\begin{align}
\frac{\langle s_{([(\widetilde M-\lambda_1)^{\lambda_1}]+Y)\cup Y'}\rangle
^\text{GC}_{k,\widetilde M-2\lambda_1}(z)}
{\langle s_{([(\widetilde M-\lambda_1)^{\lambda_1}]+Y)\cup Y'}\rangle
_{k,\widetilde M-2\lambda_1}(\lambda_1)}
=\frac{\langle s_{([(\alpha_1)^{\widetilde M-\alpha_1}]+Y)\cup Y'}\rangle
^\text{GC}_{k,-\widetilde M+2\alpha_1}(z)}
{\langle s_{([(\alpha_1)^{\widetilde M-\alpha_1}]+Y)\cup Y'}\rangle
_{k,-\widetilde M+2\alpha_1}(\widetilde M-\alpha_1)},
\label{max}
\end{align}
with $\widetilde M=M+2r$.
For the special case $\lambda_1=\alpha_1=0$ this reduces to \eqref{openclosed}.
In \cite{HaOk} another interesting identity \eqref{hook}
\begin{align}
\frac{\langle s_{(a|l)}\rangle^\text{GC}_{k,M}(z)}
{\langle s_{(a|l)}\rangle_{k,M}(1)}
=\biggl[\frac{\langle s_{(l+M|a-M)}\rangle^\text{GC}_{k,M}(z)}
{\langle s_{(l+M|a-M)}\rangle_{k,M}(1)}\biggr]^*,
\label{armleg}
\end{align}
with $a>M$ was found numerically.
However, after generalizing the open-closed duality into \eqref{max}, we point out that this falls into the same class of the identity.
Namely, since the Young diagram $(a|l)$ is decomposed as $([(M+1)^1]+[a-M-\frac{1}{2}])\cup[1^{l-\frac{1}{2}}]$, we can use our formula \eqref{max} to shift the integration contour to obtain $([1^{M+1}]+[a-M-\frac{1}{2}])\cup[1^{l-\frac{1}{2}}]$, which is $(a-M|l+M)$.
After applying the conjugate relation \eqref{conj}, this reduces to \eqref{armleg}.
Let us comment on the effect of crossing the real poles in shifting the integration contour.
Although this gives different values due to the effect of the poles, we note that the difference is under control by taking care of the residues.
The computation of the difference is similar to \cite{HHMO,sho,MS1}.
\section{Discussions}\label{discuss}
In this paper we have proposed a new Fermi gas formalism to study vacuum expectation values of the half-BPS Wilson loop.
Compared with the open string formalism \cite{sho}, which expresses the partition function with fractional branes using the Wilson loops, our formalism is the opposite of it.
Namely, with the same method which leads to the closed string formalism of the partition function, we end up with a new formalism that expresses the Wilson loop expectation values using the modified density matrix which depends on the set of arm lengths and leg lengths.
With this formalism we can prove some important duality relations proposed previously: the open-closed duality and its generalizations.
We comment that our formalism looks similar to the one conjectured in \cite{HNS}, though the comparison seems difficult.
Also, although we have proved the identity \eqref{generalized} inspiring the open-closed duality, it is not clear to us how this duality relates to those in \cite{GV,OV,ADKMV}.
It would be interesting to clarify the relations\footnote{We are grateful to Shinji Hirano, Takahiro Nishinaka and Masaki Shigemori for valuable discussions.}.
The computation of the half-BPS trace operators in D3-branes \cite{CJR} has a nice interpretation from the fermion droplets \cite{LLM}.
After obtaining a simple formalism for the half-BPS Wilson loop in M2-branes, it is interesting to ask whether we can find a similar interpretation from the supergravity viewpoint as well.
Unfortunately, because of the convergence condition $M\le k/2$, our formalism seems not very helpful in proving the Giveon-Kutasov duality \cite{GK,KWYdual} relating $M$ to $k-M$.
In fact, previously the duality was used to define the formalism for $M\ge k/2$.
It is an interesting open problem to improve this situation.
From a technical viewpoint, we have proposed another nice formalism for general expectation values in the ABJM matrix model.
Since there are some interesting related models \cite{GHN,MN1,MN2,MN3,HHO}, we would like to apply a similar formalism to these models for the numerical computations and find more relations to the topological string theories.
\section*{Acknowledgements}
We are grateful to Heng-Yu Chen, Shun-Jen Cheng, Shinji Hirano, Pei-Ming Ho, Yoshinori Honma, Dharmesh Jain, Takuya Matsumoto, Satsuki Matsuno, Shota Nakayama, Takahiro Nishinaka, Tadashi Okazaki, Takeshi Oota, Masaki Shigemori, Reiji Yoshioka and especially Takao Suyama for valuable discussions.
The work of S.M.\ is supported by JSPS Grant-in-Aid for Scientific
Research (C) \# 26400245.
|
\section{Introduction}
Our aim in this paper is to analyze the following Dirichlet problem:
\begin{equation}\label{prob-prin}
\left\{\begin{array}{ll}
\displaystyle -\hbox{\rm div\,}\left(\frac{Du}{|Du|}\right)+|Du|=f(x)&\hbox{ in }\,\Omega\,,\\[3mm]
u=0 &\hbox{ on }\,\partial\Omega\,,
\end{array}\right.
\end{equation}
where $\Omega$ is a bounded open subset of $\mathbb{R}^N$ with Lipschitz boundary $\partial\Omega$ and $f$ is a non--negative function belonging to $L^1(\Omega)$. As usual when the $1$--Laplacian operator is considered, the natural energy space to study this problem is $BV(\Omega)$, that is, the space of all functions of bounded variation.
The homogeneous problem, in an unbounded domain, arises in the level set formulation of the inverse mean curvature flow, namely,
\begin{equation}\label{IMCF}
\left\{\begin{array}{ll}
\displaystyle -\hbox{\rm div\,}\left(\frac{Du}{|Du|}\right)+|Du|=0&\hbox{ in }\,\Omega\,,\\[3mm]
u=0 &\hbox{ on }\,\partial\Omega\,,\\[3mm]
u(x)\to \infty &\hbox{ as }\, |x| \to \infty\,.
\end{array}\right.
\end{equation}
The inverse mean curvature flow is a one--parameter family of hypersurfaces $\{\Gamma_t\}_{t \geq 0}$ whose normal velocity $V_n(t)$ at each time $t$ equals the inverse of its mean curvature $H(t)$. Given $\Gamma_0$, the problem is to find $F: \Gamma_0 \times [0, T] \rightarrow \mathbb{R}^N$ such that
\begin{equation}\label{e1imc}
\frac{\partial F}{\partial t} = \frac{\nu}{H}\,, \quad \quad t \geq 0\,,
\end{equation}
where $\nu(t)$ denotes the unit outward normal to $\Gamma_t=F(\Gamma_0, t)$. The level set formulation \eqref{IMCF} was introduced in \cite{Huisken} (see also \cite{Huisken2, Moser}); observe that $\hbox{\rm div\,}\left(\frac{Du}{|Du|}\right)$ gives the mean curvature and $|Du|$ yields the inverse of the speed. In the case that $\Omega$ includes a bounded connected component, it produces a sudden phenomenon called fattening by which this component disappear instantaneously.
If a non--negative source is considered, as in \eqref{prob-prin}, then \eqref{e1imc} becomes
\begin{equation*}
\frac{\partial F}{\partial t} = \frac{\nu}{H+\hbox{source}}\le \frac{\nu}{H}, \quad \quad t \geq 0,
\end{equation*}
so that the datum damps the flux. This inhomogeneous inverse mean curvature flow was studied in \cite{MS0}.
Although the homogeneous problem is not interesting in bounded domains because it leads to the trivial solution, this does not occur in the non--homogeneous case since the source can override the fattening phenomenon (at least when $f$ is not very small). Problem \eqref{prob-prin} in bounded domains has been considered in \cite{MS} for data $f\in L^p(\Omega)$, with $p>N$, seeking bounded solutions, and in \cite{LS} when data belong to the Marcinkiewicz space $L^{N,\infty}(\Omega)$, looking for unbounded variational solutions. Existence and uniqueness results have been obtained in both papers for any given non--negative datum.
It is worth mentioning that the gradient term is essential to get existence and uniqueness results. In \cite{K} (see also \cite{CT, MST1} for more general data) it is shown that there exist solutions to problem
\begin{equation*}
\left\{\begin{array}{ll}
\displaystyle -\hbox{\rm div\,}\left(\frac{Du}{|Du|}\right)=f(x)&\hbox{ in }\,\Omega\,,\\[3mm]
u=0 &\hbox{ on }\,\partial\Omega\,,
\end{array}\right.
\end{equation*}
only when data are small enough. On the other hand, uniqueness cannot be expected since if $u$ is a solution and $g$ is a real increasing smooth function, then $v=g(u)$ should be a solution as well. Therefore, the total variation term has a regularizing effect.
Our purpose is to go a step further and study problem \eqref{prob-prin} when data are merely integrable functions. This kind of non--variational problems has extensively been studied for problems involving the $p$--Laplacian ($1<p\le N$). In this framework, there are two different formulations: that of entropy solution introduced in \cite{B-V} (see also \cite{BGO}) and that of renormalized solution, for which we refer to \cite{DMOP}. Both approaches systematically use truncations of solutions. In \cite{ABCM}, in the framework of the $1$--Laplacian, the authors also introduce a notion of solution by means of truncations. We follow the same concept, but adapted to our situation. Indeed, since the regularizing effect of the total variation yields $u\big|_{\partial\Omega}=0$, the boundary condition holds in the sense of traces.
Another feature deriving from the regularizing effect is $u\in BV(\Omega)$ without jump part. Nevertheless, this fact does not allow us to define (following Anzellotti, see \cite{An}) the pairing of a general $L^\infty$--divergence--measure vector field $\z$ and the solution $u$. Hence, truncations must remain in the definition of solution. Instead of products of the form $(\z,Du)$, we have to handle with products such as $(\z,De^{-u})$ and $(e^{-u}\z,Du)$. Beyond these kind of technical complication, the existence theorem holds as it was expected, and we will only make explicit those parts of the proof which are different of that for regular data in \cite[Theorem 4.3]{LS}. Much more interesting is the comparison principle. We point out that, even in the context of bounded solutions, its proof is new and simpler than that of the uniqueness result in \cite{MS}. We also investigate solutions when data belong to $L^p(\Omega)$, with $1<p<N$, finding that the solution lies in $L^{\frac{Np}{N-p}}(\Omega)$. Note that Lebesgue spaces continuously adjust with the known cases $p=1$ (in which $u\in BV(\Omega)\subset L^{\frac N{N-1}}(\Omega)$) and $p=N$ (see \cite[Proposition 4.7]{LS}).
This paper is organized as follows. In section 2 we introduce some definitions and notation and we also give some preliminaries results that we will need. Among these results, we foreground Proposition 2.4 for which we supply a new proof. This proposition is essential to deal with pairings involving functions of the solution (such as truncations) through Proposition 2.7. Section 3 is devoted to prove the existence result and the comparison principle. In section 4 we show the best summability that the solution can get when data belong to $L^p(\Omega)$, with $1<p<N$. Finally, in the last section we show examples of radial solutions, which give evidence that the obtained regularity is optimal.
\section*{Acknowledgements}
This research has been partially supported by the Spanish Mi\-nis\-te\-rio de Econom\'{\i}a y Competitividad and
FEDER, under project MTM2015--70227--P.
The first author was also supported by Ministerio de Econom\'{\i}a y
Competitividad under grant BES--2013--066655.
The authors would like to thank Salvador Moll for some fruitful discussions concerning this paper.
\section{Preliminaries}
In this section we will introduce some notation and auxiliary results which will be used
throughout this paper. In what follows, we will consider $N\ge2$ and, given a set $E$, we will write
$\h^{N-1}(E)$ to denote its $(N - 1)$--dimensional Hausdorff measure and $|E|$ its Lebesgue measure.
In this paper, $\Omega$ will always stands for an open subset of
$\mathbb{R}^N$ with Lipschitz boundary. Thus, an outward normal unit
vector $\nu(x)$ is defined for $\h^{N-1}$--almost every
$x\in\partial\Omega$.
We will make use of the usual Lebesgue and Sobolev
spaces, denoted by $L^q(\Omega)$ and $W_0^{1,p}(\Omega)$,
respectively (see for instance \cite{Br} or \cite{Ev}).
We recall that for a Radon measure $\mu$ in $\Omega$ and a Borel set
$A\subseteq\Omega$, the measure $\mu\res A$ is defined by $(\mu\res A)(B)=\mu(A\cap B)$
for any Borel set $B\subseteq\Omega$.
The truncation function will be use throughout this paper. Given
$k>0$, it is defined by
\begin{equation}\label{trun}
T_k(s)=\min\{|s|, k\}{\rm \; sign} (s)\,,
\end{equation} for all $s\in\mathbb{R}$.
\subsection{Functions of bounded variation}
The natural energy space to study problems involving the $1$--Laplacian is the space of all functions of bounded variation, that is, functions $u\,:\,\Omega\to\mathbb{R}$ belonging to $L^1(\Omega)$ whose derivative in the sense of distributions $Du$ is a Radon measure with finite total variation. This space will be denoted by $BV(\Omega)$.
Let $u \in BV(\Omega)$, we can decompose the Radon measure $Du$ into its absolutely continuous part and its singular part with respect to the Lebesgue measure: $Du=D^au + D^su$. We denote by $S_u$ the set of all $x \in \Omega$ which are not Lebesgue points, that is, $x \not\in S_u$ if there exists $\tilde{u}(x)$ such that
\begin{equation*}
\lim_{\rho \downarrow 0} \frac{1}{|B_{\rho}(x)|} \int_{B_{\rho}(x)} \vert u(y) - \tilde{u}(x) \vert \, dy = 0 \,.
\end{equation*}
We say that $x \in \Omega$ is an approximate jump point of $u$, denoted by $x \in J_u$, if there exist two real numbers $u^+(x)>u^-(x)$ and $\nu_u(x)$ with $|\nu_u(x)|=1$ such that
\begin{equation*}
\lim_{\rho \downarrow 0} \frac{1}{|B_{\rho}^+(x,\nu_u(x))|} \int_{B_{\rho}^+(x,\nu_u(x))} \vert u(y) - u^+(x) \vert \, dy = 0 \,,
\end{equation*}
\begin{equation*}
\lim_{\rho \downarrow 0} \frac{1}{|B_{\rho}^-(x,\nu_u(x))|} \int_{B_{\rho}^-(x,\nu_u(x))} \vert u(y) - u^-(x) \vert \, dy = 0 \,,
\end{equation*}
where
\begin{equation*}
B_{\rho}^+(x,\nu_u(x)) = \{ y \in B_{\rho}(x) \ | \ \langle y - x, \nu_u(x) \rangle >0 \}
\end{equation*}
and
\begin{equation*}
B_{\rho}^-(x,\nu_u(x)) = \{ y \in B_{\rho}(x) \ | \ \langle y - x, \nu_u(x) \rangle <0 \} \,.
\end{equation*}
We know that $S_u$ is countably $\h^{N-1}$--rectifiable and $\h^{N-1}(S_u \backslash J_u) = 0$ by the Federer--Vol'pert Theorem (see \cite[Theorem 3.78]{AFP}). Moreover, we also know
\begin{equation*}
Du \res J_u = (u^+ - u^-) \nu_u \h^{N-1} \res J_u \,.
\end{equation*}
Using $S_u$ and $J_u$, we can split $D^su$ in its jump part $D^j u$ and its Cantor part $D^c u$, defined by
\begin{equation*}
D^ju = D^su \res J_u \qquad {\rm and} \qquad D^c u = D^su \res (\Omega \backslash S_u) \,.
\end{equation*}
Then, we have
\begin{equation*}
D^j u = (u^+ - u^-) \nu_u \h^{N-1} \res J_u \,.
\end{equation*}
In addition, if $x \in J_u$, then $\nu_u(x) = \frac{Du}{| D u |}(x)$ where $\frac{Du}{| D u |}$ is the Radon--Nikod\'ym derivative of $Du$ with respect to its total variation $| D u |$.
The precise representative $u^* : \Omega \setminus(S_u \setminus J_u) \rightarrow \mathbb{R}$ of
$u$ is defined by
\begin{equation*}
u^*(x) = \left\{
\begin{array}{lcc}
\tilde{u}(x) & \mbox{if} & x \in \Omega \setminus S_u \,,\\
\displaystyle\frac{u^-(x) + u^+(x)}{2} & \mbox{if} & x \in J_u \,.
\end{array}
\right.
\end{equation*}
For the sake of simplicity, most of the time we will denote both function and its precise representative by $u$.
We will use the Chain Rule, but only when $u $ is a bounded variation function without jump part.
\begin{Proposition}\label{Chain-Rule}
Let $u \in BV(\Omega)$ with $D^ju=0$ and let $f$ be a Lipschitz function in $\Omega$. Then, $v=f \circ u$ belongs to $BV(\Omega)$ and $Dv=f^\prime(u) Du$, so that $D^j v=0$.
\end{Proposition}
For further information about bounded variation functions we refer to \cite{AFP}, \cite{EG} and \cite{Zi}.
\subsection{$L^\infty$--divergence--measure fields}
We will denote by $\DM(\Omega)$ the set of all vector fields $\z \in L^\infty(\Omega;\mathbb{R}^N)$ such that $\hbox{\rm div\,} \z$ is a Radon measure in $\Omega$ with finite total variation. Following \cite{ABCM}, we will use these vector fields to give a sense to $\displaystyle \frac{Du}{|Du|}$ in our equation, even if $Du$ is a Radon measure and, moreover, if it vanishes in a zone of the domain. More concretely, we seek for a vector field $\z\in L^\infty(\Omega;\mathbb{R}^N)$ satisfying $\|\z\|_\infty\le1$ and $(\z,DT_k(u))=|DT_k(u)|$ for all $k>0$.
Let $\z \in \DM(\Omega)$ and $w \in BV(\Omega)\cap C(\Omega) \cap L^\infty(\Omega)$; for every $\varphi \in C^\infty_0 (\Omega)$ we define the functional
\begin{equation*}
\langle (\z, Dw),\varphi \rangle = - \int_\Omega w \,\varphi \, \hbox{\rm div\,} \z - \int_\Omega w \, \z \cdot \nabla \varphi \,dx \,.
\end{equation*}
It was proved in \cite{An} that this distribution has order $0$ since satisfies
\[
|\langle (\z, Dw),\varphi\rangle|\le\|\varphi\|_\infty\|\z\|_\infty \int_\Omega|Dw|\,.
\]
Thus, it is actually a Radon measure with finite total variation and the following inequality holds
\begin{equation}\label{des-An}
|(\z, Dw)|\le \|\z\|_\infty |Dw|
\end{equation}
as measures in $\Omega$. In particular, the Radon measure $(\z, Dw)$ is absolutely continuous with respect to $|Dw|$.
Denoting by $$\theta(\z, Dw, \cdot) : \Omega \rightarrow \mathbb{R}$$ the
Radon--Nikod\'ym derivative of $(\z,Dw)$ with respect to $|Dw|$,
it follows that
\begin{equation*}
\int_B (\z, Dw) = \int_B \theta(\z, Dw, x) \, |Dw| \qquad \mbox{ for all Borel
sets} \quad B \subset \Omega \,,
\end{equation*}
and
\begin{equation*}
\Vert \theta(\z, Dw, \cdot)\Vert_{L^{\infty}(\Omega, \vert Dw \vert)} \leq \Vert \z \Vert_{\infty}\,.
\end{equation*}
Moreover, if $f : \mathbb{R} \rightarrow \mathbb{R}$ \ is a Lipschitz continuous increasing function, then
\begin{equation}\label{E1paring12}
\theta(\z, D(f \circ w),x) = \theta (\z, Dw, x) \qquad |Dw|\mbox{--a.e.} \quad {\rm in} \quad \Omega \,.
\end{equation}
The Anzellotti theory also provides the definition of a weak trace on $\partial
\Omega$ to the normal component of any vector field $\z\in \DM$, denoted by $[z,\nu]$. This weak trace satisfies $\|[\z,\nu]\|_\infty\le \|\z\|_\infty$. Relating the pairing $(\z, Dw)$ and the weak trace $[z,\nu]$ a Green's formula holds.
\begin{Theorem}\label{Green}
If $\z\in \DM$ and $w\in BV(\Omega)\cap C(\Omega) \cap L^\infty(\Omega)$, then we have
\[
\int_\Omega w\,\hbox{\rm div\,}\z +\int_\Omega (\z, Dw)=\int_{\partial\Omega}w\,[\z,\nu]\, d\h^{N-1}\,.
\]
\end{Theorem}
\bigskip
As mentioned, for a general $\z \in \DM(\Omega)$, Anzellotti's theory assumes that $w \in BV(\Omega)\cap C(\Omega) \cap L^\infty(\Omega)$ in order to define $(\z, Dw)$ and to prove a Green's formula. This theory was generalized to consider $w \in BV(\Omega)\cap L^\infty(\Omega)$ in \cite{CF} using a different approach, and in \cite{C} and \cite{MST2} following the same definitions of Anzellotti. Indeed, given $\z \in \DM(\Omega)$ and $w \in BV(\Omega)\cap L^\infty(\Omega)$; for every $\varphi \in C^\infty_0 (\Omega)$ we may define the functional
\begin{equation*}
\langle (\z, Dw),\varphi \rangle = - \int_\Omega w^* \,\varphi \, \hbox{\rm div\,} \z - \int_\Omega w \, \z \cdot \nabla \varphi \,dx \,.
\end{equation*}
We explicitly mention that the precise representative $w^*$ is summable with respect to $\hbox{\rm div\,}\z$ and that this definition depends on the chosen representative of the function.
Now, we present some results which we use several times in the sequel. Next proposition was proved in \cite{MS}.
\begin{Proposition}\label{Prop2.3}
Let $\z \in \DM(\Omega)$ and let $u , w \in BV(\Omega)\cap L^\infty (\Omega)$ be functions such that $D^ju=D^jw=0$. Then
\begin{equation*}
(w\,\z,Du)= w^*(\z,Du) \quad \mbox{as Radon measures in } \; \Omega \,.
\end{equation*}
\end{Proposition}
In principle, it is not clear that (\ref{E1paring12}) holds in the case that $\z \in \DM(\Omega)$ and
$u \in BV(\Omega) \cap L^{\infty}(\Omega)$. However, we will see that (\ref{E1paring12}) holds if we assume the jump part $D^j u$ vanishes.
This result was proved in \cite{MS} but an extra hypothesis is needed in the proof, namely, the set of discontinuities of $u$ is $\h^{N-1}$--null.
We next prove this result under the general assumption $D^ju=0$.
Following Anzellotti, the main ingredient to prove the above formula is a
``slicing" result that links the measure $(\z, Du)$ with the
measures $(\z, D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}})$, where $E_{u,t}:= \{ x \in \Omega \ : \ u(x) > t \}$.
\begin{Proposition}\label{rebanada} Let $\z \in \DM(\Omega)$ and consider
$u \in BV(\Omega) \cap L^{\infty}(\Omega)$ with $D^j u=0$. Let $E_{u,t}:= \{ x \in \Omega \ : \ u(x) > t \}$. Then
for
all $\varphi \in C^{\infty}_{0}(\Omega)$, the function $t \mapsto
\langle (\z, D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}}), \varphi \rangle$ is $\mathcal L^1$--measurable and
\begin{equation}\label{e2mejora}
\langle (\z, Du), \varphi \rangle = \int_{- \infty}^{+ \infty} \langle (\z, D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}}), \varphi \rangle \, dt\,.
\end{equation}
\end{Proposition}
\begin{proof} First we observe that we may assume $u\ge0$; if not, we consider the function $u+\|u\|_\infty$.
We also point out that for every measurable set $E\subset \Omega$ having finite perimeter, the condition $|\hbox{\rm div\,}\z|(\partial^*E)=0$ implies
\[
{\raise0pt\hbox{{\LARGE $\chi$}}}_E\, \hbox{\rm div\,}\z = {\raise0pt\hbox{{\LARGE $\chi$}}}^*_E\, \hbox{\rm div\,}\z\,.
\]
\bigskip
As a consequence, we obtain the following claim:
\\
{\sl If $E\subset \Omega$ is a measurable set with finite perimeter such that $|\hbox{\rm div\,}\z|(\partial^*E)=0$, then
\[
\langle (\z,D{\raise0pt\hbox{{\LARGE $\chi$}}}_E),\varphi\rangle=-\int_E\varphi\, \hbox{\rm div\,}\z-\int_E\z\cdot\nabla \varphi
\]
for all $\varphi\in C_0^\infty(\Omega)$.}
In what follows, recall that $u$ stands for the precise representative of the $BV$--function.
Observe that, thanks to the coarea formula, the level sets $E_{u,t}$ have finite perimeter for $\mathcal L^1$--almost all $t\in\mathbb{R}$. Moreover, since $D^j u=0$, it follows that
$$\h^{N-1}\left(\partial^* E_{u,t} \cap \partial^* E_{u,s} \right) = 0\quad \hbox{for } s \ne t\,.$$
Then, applying $\vert \hbox{\rm div\,}(\z) \vert \ll \h^{N-1}$ (by \cite[Proposition 3.1]{CF}), we have
\begin{equation*}
|\hbox{\rm div\,}(\z)|\left(\partial^* E_{u,t} \cap \partial^* E_{u,s} \right) = 0 \quad \hbox{ if} \ \ s \ne t\,.
\end{equation*}
Therefore, there exists $A \subset \mathbb{R}$ numerable such that
\begin{equation*}
|\hbox{\rm div\,}(\z)| \left(\partial^* E_{u,t}\right) = 0 \quad \hbox{if} \ \ t \in \mathbb{R} \backslash A\,.
\end{equation*}
In other words, we have seen that $\vert \hbox{\rm div\,}(\z) \vert\left(\partial^* E_{u,t}\right) = 0$ for $ \mathcal{L}^1$--almost all $t>0$.
Thus, our claim implies that if $\varphi\in C_0^\infty(\Omega)$, then
\begin{equation}\label{ec:1}
\langle (\z,D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}}),\varphi\rangle=-\int_{E_{u,t}}\varphi \, \hbox{\rm div\,}\z-\int_{E_{u,t}}\z\cdot\nabla \varphi \,dx\,,\quad \hbox{for } \mathcal{L}^1\hbox{--almost all } t>0\,.
\end{equation}
Considering $\varphi\in C_0^\infty(\Omega)$, we apply the slicing formula for integrable functions (see, for instance, \cite[Lemma 1.5.1]{Zi}) and \eqref{ec:1} to get that the function
$$
t \mapsto -\int_{E_{u,t}} \varphi\, \hbox{\rm div\,}\z\, dt-\int_{E_{u,t}} \z\cdot\nabla\varphi \,dx
$$
is $\mathcal L^1$--measurable and
\begin{multline*}
\langle (\z,Du),\varphi\rangle= -\int_\Omega u^*\varphi\, \hbox{\rm div\,}\z-\int_\Omega u\,\z\cdot\nabla\varphi
\\= \int_0^\infty\left[-\int_{E_{u,t}} \varphi\, \hbox{\rm div\,}\z\, dt-\int_{E_{u,t}} \z\cdot\nabla\varphi \,dx\right]\, dt
\\=\int_0^\infty\langle(\z,D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}}),\varphi\rangle\, dt \,,
\end{multline*}
as desired.
\end{proof}
\begin{Proposition}\label{rebanada1} Let $\z \in \DM(\Omega)$ and consider
$u \in BV(\Omega) \cap L^{\infty}(\Omega)$ with $D^j u=0$. Let $E_{u,t}:= \{ x \in \Omega \ : \ u(x) > t \}$. Then
for
all Borel set $B\subset\Omega$, the function $t \mapsto
\int_B (\z, D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}})$ is $\mathcal L^1$--measurable and
\begin{equation}\label{e3mejora}
\int_B (\z, Du) = \int_{- \infty}^{+ \infty} \left[\int_B (\z, D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}})\right] \, dt.
\end{equation}
\end{Proposition}
\begin{proof} Let $S$ denote a countable set in $C_0^\infty(\Omega)$ which is dense with respect to the uniform convergence.
Then, for every $t\in\mathbb{R}$ such that $E_{u,t}$ has finite perimeter and for every $\varphi\in C_0^\infty(\Omega)$ with $\varphi\ge0$, it yields
\[
\langle (\z, D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}})^+,\varphi\rangle =\sup\{\langle (\z, D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}}),\psi\rangle\>:\>\psi\in S\,,\ 0\le\psi\le\varphi\}\,.
\]
Thus the positive part of the measure $t\mapsto \langle (\z, D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}})^+,\varphi\rangle $ defines a $\mathcal L^1$--measurable function since it is the supremum of a countable quantity of $\mathcal L^1$--measurable functions. Recalling the Riesz Representation Theorem, we may go further considering an open set $B\subset\Omega$: it follows from
\[
\int_B(\z, D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}})^+=\sup\{\langle (\z, D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}})^+,\psi\rangle\>:\>\psi\in S\,,\ 0\le\psi\le{\raise0pt\hbox{{\LARGE $\chi$}}}_B\}\,,
\]
that $t\mapsto \int_B(\z, D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}})^+ $ defines a $\mathcal L^1$--measurable function. The regularity of the measures lead to the same conclusion for an arbitrary Borel set. This function is $\mathcal L^1$--summable since
\[
\int_B(\z, D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}})^+\le\int_B|(\z, D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}})|\le\|\z\|_\infty\int_B|D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}}|\,,
\]
for $\mathcal L^1$--almost all $t\in\mathbb{R}$,
and $t\mapsto \int_B|D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}}|$ defines an $\mathcal L^1$--summable function, due to the coarea formula.
On the other hand, a similar argument can be done for the negative part of the measures $(\z, D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}})$, so that $t\mapsto \int_B(\z, D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}})^-$ defines an $\mathcal L^1$--summable function for every Borel set $B\subset\Omega$. As a consequence, $t\mapsto \int_B(\z, D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}})$ defines an $\mathcal L^1$--summable function for every Borel set $B\subset\Omega$.
Finally, consider a distribution $\mu$ defined by
\[
\langle\mu,\varphi\rangle=\langle(\z,Du),\varphi\rangle-\int_{-\infty}^{+\infty}\langle(\z,D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}}),\varphi\rangle\, dt\,.
\]
Proposition \ref{rebanada} implies that $\langle\mu,\varphi\rangle=0$ for all $\varphi\in C_0^\infty(\Omega)$, wherewith $\mu$ is a Radon measure which vanishes identically. Therefore, \eqref{e3mejora} holds true.
\end{proof}
\begin{Corollary}
Let $\z \in \DM(\Omega)$ and consider $u \in BV(\Omega) \cap L^{\infty}(\Omega)$ with $D^j u=0$. Then
\begin{equation}\label{e1mejora}
\theta(\z, Du, x) = \theta(\z, D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}},x) \ \ \vert D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}} \vert\hbox{--a.e. in } \Omega \; \hbox{ for} \ \mathcal{L}^1\hbox{--almost all} \ t \in \mathbb{R}\,,
\end{equation}
\end{Corollary}
\begin{proof} Let $a,b\in\mathbb{R}$, with $a<b$ and let $B\subset\Omega$ be a Borel set.
Applying \eqref{e3mejora} to the set $\{x\in\Omega\>:\>a\le u(x)\le b\}\cap B$, we obtain
\begin{equation}\label{basic}
\int_{\{a\le u\le b\}\cap B}(\z,Du)=\int_a^b\left[\int_B(\z,D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}})\right]\, dt\,.
\end{equation}
Now we are analyzing both sides of \eqref{basic}. On the one hand, the coarea formula implies
\begin{multline*}
\int_{\{a\le u\le b\}\cap B}(\z,Du)=\int_{\{a\le u\le b\}\cap B}\theta(\z,Du,x)|Du|
\\=\int_a^b\left[\int_B\theta(\z,Du,x)|D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}}|\right]\, dt\,.
\end{multline*}
On the other,
\[
\int_a^b\left[\int_B(\z,D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}})\right]\, dt=\int_a^b\left[\int_B\theta(\z,D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}},x)|D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}}|\right]\, dt\,.
\]
Hence \eqref{basic} becomes
\[
\int_a^b\left[\int_B\theta(\z,Du,x)|D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}}|\right]\, dt=\int_a^b\left[\int_B\theta(\z,D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}},x)|D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}}|\right]\, dt\,.
\]
It follows that, for $\mathcal L^1$--almost all $t\in\mathbb{R}$,
\[\int_B\theta(\z,Du,x)|D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}}|=\int_B\theta(\z,D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}},x)|D{\raise0pt\hbox{{\LARGE $\chi$}}}_{E_{u,t}}|\]
holds for every Borel set $B$. The desired equality \eqref{e1mejora} is proved.
\end{proof}
\begin{Proposition}\label{Prop2.2} Let $\z \in \DM(\Omega)$ and consider
$u \in BV(\Omega) \cap L^{\infty}(\Omega)$ with $D^j u=0$. If $f : \mathbb{R} \rightarrow \mathbb{R}$ is a Lipschitz continuous non--decreasing function, then
\begin{equation}\label{E1paring1200}
\theta(\z, D(f \circ u),x) = \theta (\z, Du, x) \quad \vert D(f \circ u) \vert\hbox{--a.e. in }\ \Omega \,.
\end{equation}
\end{Proposition}
\begin{proof} We may follow Anzellotti (see \cite[Proposition 2.8]{An}) for the case of a increasing function. For the general case, consider $f$ non--decreasing and let $\epsilon>0$. Since the function given by
$t \mapsto f(t)+\epsilon t$ is increasing, it follows that
\begin{multline*}
(\z, D(f \circ u))+\epsilon(\z, Du)=(\z, D((f \circ u)+\epsilon u)) \\ = \theta (\z, Du, x)|D((f \circ u)+\epsilon u)|= \theta (\z, Du, x)(f^\prime( u)+\epsilon) |Du|
\end{multline*}
as measures in $\Omega$. Letting $\epsilon\to0$, we deduce
\begin{equation*}
(\z, D(f \circ u))=\theta (\z, Du, x) |D(f \circ u)|\quad\hbox{as measures in }\Omega\,.
\end{equation*}
Therefore, we have seen that \eqref{E1paring1200} holds.
\end{proof}
\section{Main results}
In this section, we prove our main results, namely the existence theorem and the comparison principle. We begin by stating our concept of solution to problem \eqref{prob-prin}.
The first difficulty we have to deal with is that we are not able to define the distribution $(\z, Du)$ when data are just integrable functions. Following \cite{ABCM}, we will solve this problem introducing truncations in the concept of solution used in \cite{LS}.
\begin{Definition}\label{def-sol}
We say that $u \in BV(\Omega)$ is a solution to problem \eqref{prob-prin} if $D^ju=0$ and there exists a vector field $\z \in \DM(\Omega)$ with $\|\z\|_\infty \le 1$ such that
\begin{equation}\label{cond-ditribucion}
-\hbox{\rm div\,} \z +|Du| =f \,\text{ in }\,\dis (\Omega)\,,
\end{equation}
\begin{equation}\label{cond-medida}
(\z, DT_k(u))=|DT_k(u)| \,\text{ as measures in }\, \Omega \;\; (\mbox{for every } k >0)\,,
\end{equation}
and
\begin{equation}\label{cond-frontera}
u\big|_{\partial \Omega} = 0\,.
\end{equation}
\end{Definition}
\subsection{Existence Theorem}
\begin{Theorem}
Let $\Omega$ be an open and bounded subset of $\mathbb{R}^N$ with Lipschitz boundary and let $f$ be a non--negative function in $L^1(\Omega)$. Then, problem \eqref{prob-prin} has at least one solution.
\end{Theorem}
\begin{proof}
The same proof of \cite[Theorem 4.3]{LS} works with minor modifications. Nevertheless, some remarks are in order.
The first remark is concerning the pairing $(e^{-u}\,\z,Du)$. If $u$ is integrable with respect to the measure $\hbox{\rm div\,}(e^{-u}\z)$ and $\varphi\in C_0^\infty(\Omega)$, then the integrals
\begin{equation*}
\int_\Omega \varphi \, u \, \hbox{\rm div\,}(e^{-u}\z)\quad \hbox{and}\quad \int_\Omega u\,e^{-u}\z \cdot \nabla \varphi \, dx
\end{equation*}
are both finite; notice that the second integral is bounded due to the inequality $u\,e^{-u}\le e^{-1}$. Therefore,
\begin{equation*}
\langle (e^{-u}\,\z,Du),\varphi\rangle =-\int_\Omega \varphi \, u \, \hbox{\rm div\,}(e^{-u}\z)- \int_\Omega u\,e^{-u}\z \cdot \nabla \varphi \, dx
\end{equation*}
is a well--defined distribution (although the distribution $(\z, Du)$ is not). Moreover, we may apply the Anzellotti procedure and obtain a Radon measure.
Taking this fact in mind, we may follow the proof of \cite[Theorem 4.3]{LS}. Starting from suitable approximating problems, we get a limit of the approximate solutions $u\in BV(\Omega)$ such that $D^ju=0$. In addition, we also get a vector field $\z\in \DM (\Omega)$ such that $\|\z\|_\infty\le1$. Moreover, the equation
\eqref{cond-ditribucion} holds and
\begin{equation*}
-\hbox{\rm div\,}(e^{-u}\z)=e^{-u}f\,.
\end{equation*}
This last equality implies that $u$ is integrable with respect to the measure $\hbox{\rm div\,}(e^{-u}\z)$ and so $(e^{-u}\,\z,Du)$ is a Radon measure.
Two conditions of Definition \ref{def-sol} must still be proved, namely, \eqref{cond-medida} and \eqref{cond-frontera}. We begin by seeing
\begin{equation}\label{medidas-T_k}
(\z, DT_k(u))=|DT_k(u)| \,\text{ as measures in }\, \Omega
\end{equation}
for every $k >0$.
To see \eqref{medidas-T_k} we start with the following equality as measures (proved in \cite[Theorem 4.3]{LS}):
\begin{equation}\label{des-previa}
|De^{-u}| \le (e^{-u}\z,Du)\,.
\end{equation}
First, we will show
\begin{equation*}
|De^{-T_k(u)}| \le (e^{-u}\z,DT_k(u)) \,.
\end{equation*}
On the one hand, considering the restriction to the set $\{ u \ge k\}$ we have
\begin{equation*}
|De^{-T_k(u)}|\res{\{ u \ge k\}} = e^{-T_k(u)}|DT_k(u)|\res{\{ u \ge k\}} = 0 \,,
\end{equation*}
and on the other hand
\begin{equation*}
| (e^{-u}\z,DT_k(u))|\res{\{ u \ge k\}}\le |DT_k(u))|\res{\{ u \ge k\}}= 0 \,.
\end{equation*}
Now, we just work with the restriction to the set $\{ u < k\}$. For every $\varphi\in C_0^\infty(\Omega)$ such that $\varphi\ge0$, using the definition of the distribution and applying \eqref{des-previa} we arrive at
\begin{multline*}
\langle (e^{-u}\z,DT_k(u))\res{\{ u < k\}}, \varphi \rangle =-\int_{\{ u < k\}} \varphi\,u\,\hbox{\rm div\,}(e^{-u}\z) - \int_{\{ u < k\}}u\,e^{-u} \z\cdot\nabla \varphi \,dx
\\=\langle (e^{-u}\z,Du)\res{\{ u < k\}}, \varphi \rangle \ge \int_{\{ u < k\}} \varphi \,|De^{-u}|
\\ = \int_{\Omega} \varphi \,e^{-u} |DT_k(u)| = \int_{\Omega}\varphi \, |De^{-T_k(u)}| \,.
\end{multline*}
Now, we have to prove that $(\z,DT_k(u)) = |DT_k(u)|$ as measures in $\Omega$.
We use Proposition \ref{Prop2.3} and the Chain Rule to get
\begin{equation*}
|De^{-T_k(u)}|\le (e^{-u}\z , DT_k(u)) =e^{-u}(\z , DT_k(u)) \le e^{-u}|DT_k(u)| = |De^{-T_k(u)}| \,.
\end{equation*}
Then, the inequality becomes equality and $e^{-u}(\z,DT_k(u)) = e^{-u}|DT_k(u)|$ as measures in $\Omega$. We deduce that
\begin{equation*
(\z,DT_k(u)) = |DT_k(u)|
\end{equation*}
as measures in $\Omega$, since $e^{-u} =0$ yields $T_k(u)=k$ for every $k >0$.
To check the boundary condition \eqref{cond-frontera} we consider the real function defined by
\begin{equation*}
J_1(s)=\int_0^sT_1(\sigma)\, d\sigma\,.
\end{equation*}
Then, in the same way than in \cite[Theorem 4.3]{LS}, we obtain
\begin{multline}\label{des-front}
\int_\Omega|DT_1(u)|+\int_{\partial\Omega}|T_1(u)|\, d\h^{N-1}+\int_\Omega|DJ_1(u)|+\int_{\partial\Omega}|J_1(u)|\, d\h^{N-1}\\
\le \int_\Omega f\,T_1(u)\, dx\,.
\end{multline}
Using the equation and the previous step, and applying the Green's formula and the Chain Rule, we get
\begin{multline*}
\int_\Omega f\,T_1(u)\, dx=-\int_\Omega T_1(u)\,\hbox{\rm div\,}\z+\int_\Omega T_1(u)|Du|
\\=\int_\Omega(\z,DT_1(u))-\int_{\partial\Omega}T_1(u)[\z,\nu]\, d\h^{N-1}+\int_\Omega |DJ_1(u)|
\\=\int_\Omega|DT_1(u)|-\int_{\partial\Omega}T_1(u)[\z,\nu]\, d\h^{N-1}+\int_\Omega |DJ_1(u)|\,.
\end{multline*}
Going back to \eqref{des-front} and simplifying, it follows that
\begin{equation*}
\int_{\partial\Omega}|T_1(u)|+T_1(u)[\z,\nu]\, d\h^{N-1}+\int_{\partial\Omega}|J_1(u)|\, d\h^{N-1}\le0\,.
\end{equation*}
Observe that both integrals are non--negative, so that both vanish. In particular, $J_1(u)=0$ $\h^{N-1}$--a.e. on $\partial\Omega$.
Therefore, the boundary condition holds true.
\end{proof}
\subsection{Comparison principle}
Before proving the comparison principle we need to present some preliminary results.
\begin{Proposition}\label{prop-medidas}
Let $\z$ be a vector field in $\DM(\Omega)$ and let $u$ be a function of bounded variation with $D^ju=0$ and such that $(\z,DT_k(u))=|DT_k(u)|$ for every $k>0$. If $g:\Omega \to \mathbb{R}$ is a bounded, increasing and Lipschitz function, then $(\z,Dg(u))=|Dg(u)|$ holds as measures.
\end{Proposition}
\begin{proof}
Since $(\z,DT_k(u))=|DT_k(u)|$, the Radon--Nikod\'ym derivative of $(\z,DT_k(u))$ with respect its total variation $|DT_k(u)|$ is $\theta(\z,DT_k(u),x) =1$. Moreover, using Proposition \ref{Prop2.2} we get
\begin{equation*}
\theta(\z,Dg(T_k(u)),x)=\theta(\z,DT_k(u),x)=1 \,,
\end{equation*}
that is, $(\z,Dg(T_k(u)))=|Dg(T_k(u))|$ for every $k>0$. Now, by the Dominated Convergence Theorem, we take limits in this expression when $k$ goes to $\infty$ and it leads to
\begin{equation*}
(\z, Dg(u)) = g'(u)\,|Du| = |Dg(u)| \,.
\end{equation*}
\end{proof}
\begin{Proposition}\label{prop-1}
Let $f \in L^1(\Omega)$. If $u \in BV(\Omega)$ is a solution to problem \eqref{prob-prin} and $\z \in \DM(\Omega)$ is the associated vector field, then the following equality holds:
\begin{equation*}
-\hbox{\rm div\,}(e^{-u}\z)=e^{-u}f \,\text{ in }\,\dis (\Omega)\,.
\end{equation*}
\end{Proposition}
\begin{proof}
Let $\varphi \in C_0^\infty(\Omega)$, we take the test function $e^{-u}\varphi$ in problem \eqref{prob-prin} and we obtain
\begin{equation*}
-\int_\Omega e^{-u}\varphi\, \hbox{\rm div\,} \z + \int_\Omega e^{-u}\varphi \,|Du| = \int_\Omega e^{-u}\varphi f \,dx \,.
\end{equation*}
Now, since $e^{-u}$ is bounded, we can use the definition of pairing $(\z,De^{-u})$ and the former equality becomes
\begin{equation*}
\int_\Omega e^{-u} \z \cdot \nabla \varphi \,dx +\int_\Omega \varphi\,(\z , De^{-u})+\int_\Omega e^{-u}\varphi \,|Du| =\int_\Omega e^{-u}\varphi f \,dx \,.
\end{equation*}
Finally, using $(\z,De^{-u})=-e^{-u}|Du|$ (see Proposition \ref{prop-medidas}) and Green's formula we deduce
\begin{equation*}
-\hbox{\rm div\,}(e^{-u}\z)=e^{-u}f \,\text{ in }\,\dis (\Omega)\,.
\end{equation*}
\end{proof}
\begin{Theorem}
Let $f_1$ and $f_2$ be two non--negative functions in $ L^1(\Omega)$ with $f_1\le f_2$, and consider problems
\begin{equation}\label{prob-u1}
\left\{\begin{array}{ll}
\displaystyle -\hbox{\rm div\,}\left(\frac{Du_1}{|Du_1|}\right)+|Du_1|=f_1(x)&\hbox{ in }\,\Omega\,,\\[3mm]
u_1=0 &\hbox{ on }\,\partial\Omega\,,
\end{array}\right.
\end{equation}
and
\begin{equation}\label{prob-u2}
\left\{\begin{array}{ll}
\displaystyle -\hbox{\rm div\,}\left(\frac{Du_2}{|Du_2|}\right)+|Du_2|=f_2(x)&\hbox{ in }\,\Omega\,,\\[3mm]
u_2=0 &\hbox{ on }\,\partial\Omega\,.
\end{array}\right.
\end{equation}
If $u_1$ is a solution to problem \eqref{prob-u1} and $u_2$ is a solution to problem \eqref{prob-u2}, then $u_1\le u_2$.
\end{Theorem}
\begin{proof}
For each $i=1,2$, we know that solution $u_i\in BV(\Omega)$ satisfies $D^ju_i=0$ and there exists a vector field $\z_i \in \DM(\Omega)$ such that $\|\z_i\|_\infty\le 1$. Moreover,
\begin{equation*
-\hbox{\rm div\,} \z_i +|Du_i| =f_i \,\text{ in }\,\dis (\Omega)\,,
\end{equation*}
\begin{equation*
(\z_i, DT_k(u_i))=|DT_k(u_i)| \,\text{ as measures in }\, \Omega\;\; (\mbox{for every } k >0)\,,
\end{equation*}
and
\begin{equation*
u_i\big|_{\partial \Omega} = 0\,.
\end{equation*}
\medskip
We are seeking that $u_1\le u_2$, to this end we divide the proof in several steps.
\bigskip
\textbf{STEP 1: ${\bf (\z_1-\z_2,D(T_k(u_1)-T_k(u_2))^+)}$ is a positive Radon measure for all ${\bf k>0}$.}
Let $\varphi \in C_0^\infty(\Omega)$ with $\varphi \ge 0$. Then, the measure $(\z_1-\z_2,D(T_k(u_1)-T_k(u_2))^+))$ actually is
\begin{multline*}
\int_\Omega \varphi \,(\z_1-\z_2,D(T_k(u_1)-T_k(u_2))^+) = \int_{\{T_k(u_1)> T_k(u_2)\}} \varphi \,(\z_1-\z_2,D(T_k(u_1)-T_k(u_2)))
\\= \int_{\{T_k(u_1)> T_k(u_2)\}} \varphi\bigg[(z_1,DT_k(u_1)) - (z_2,DT_k(u_1)) - (z_1,DT_k(u_2)) + (z_2,DT_k(u_2)) \bigg]
\\=\int_{\{T_k(u_1)> T_k(u_2)\}} \varphi \bigg[ |DT_k(u_1)| - (z_2,DT_k(u_1)) -(z_1,DT_k(u_2)) + |DT_k(u_2)| \bigg]
\\ \ge 0 \,,
\end{multline*}
because $(z_i,Du_j)\le |Du_j|$ for $i,j=1,2$.
Therefore, we conclude that $(\z_1-\z_2,D(T_k(u_1)-T_k(u_2))^+)$ is a positive Radon measure.
\bigskip
\textbf{STEP 2: Prove that ${\bf \displaystyle \int_{\{u_1>u_2\}} (e^{-u_2}- e^{-u_1})(|Du_1| -|Du_2|)\ge 0}$.}
First, we take the test function $(e^{-u_2}-e^{-u_1})^+$ in problem \eqref{prob-u1} and since $(e^{-u_2}-e^{-u_1})^+\not=0$ if $u_1> u_2$, we get
\begin{multline}\label{ec-u1}
\int_{\{u_1>u_2\}} (\z_1,D(e^{-u_2}-e^{-u_1})) + \int_{\{u_1>u_2\}} (e^{-u_2}-e^{-u_1}) \,|Du_1|
\\ = \int_\Omega (e^{-u_2}-e^{-u_1})^+ f_1\,dx\,.
\end{multline}
Moreover, using that $e^{-u_2}-e^{-u_1} = (1-e^{-u_1})-(1-e^{-u_2})$ we also have
\begin{multline}\label{ec2-u1}
\int_\Omega (e^{-u_2}-e^{-u_1})^+ f_1\,dx =\int_{\{u_1>u_2\}}(\z_1,D(1-e^{-u_1}))- \int_{\{u_1>u_2\}}(\z_1,D(1-e^{-u_2}))
\\[2mm]+ \int_{\{u_1>u_2\}} e^{-u_2}\,|Du_1|- \int_{\{u_1>u_2\}} e^{-u_1} \,|Du_1|
\\[2mm]=\int_{\{u_1>u_2\}} |D(1-e^{-u_1})| - \int_{\{u_1>u_2\}}(\z_1,D(1-e^{-u_2}))
\\[2mm]+ \int_{\{u_1>u_2\}} e^{-u_2}\,|Du_1|- \int_{\{u_1>u_2\}} e^{-u_1} \,|Du_1|
\\[2mm] =-\int_{\{u_1>u_2\}}(\z_1,D(1-e^{-u_2})) + \int_{\{u_1>u_2\}} e^{-u_2}\,|Du_1|\,,
\end{multline}
where we have used Proposition \ref{Prop2.2} and the Chain Rule.
Now, taking the same test function $(e^{-u_2}-e^{-u_1})^+$ in problem \eqref{prob-u2} and making similar computations we obtain
\begin{equation}\label{ec2-u2}
\int_\Omega (e^{-u_2}-e^{-u_1})^+ f_2\,dx =\int_{\{u_1>u_2\}}(\z_2,D(1-e^{-u_1})) - \int_{\{u_1>u_2\}} e^{-u_1}\,|Du_2|\,.
\end{equation}
Since $f_1\le f_2$, we can join expressions \eqref{ec2-u1} and \eqref{ec2-u2} to get the following inequality:
\begin{multline*}
\int_{\{u_1>u_2\}} e^{-u_1}\,|Du_2|+ \int_{\{u_1>u_2\}} e^{-u_2}\,|Du_1|
\\ \le \int_{\{u_1>u_2\}}(\z_1,D(1-e^{-u_2})) +\int_{\{u_1>u_2\}}(\z_2,D(1-e^{-u_1}))
\\ \le \int_{\{u_1>u_2\}} |(\z_1,D(1-e^{-u_2}))| +\int_{\{u_1>u_2\}} |(\z_2,D(1-e^{-u_1}))|
\\ \le \int_{\{u_1>u_2\}} |D(1-e^{-u_2})| + \int_{\{u_1>u_2\}} |D(1-e^{-u_1})|
\\= \int_{\{u_1>u_2\}} e^{-u_2}|Du_2| + \int_{\{u_1>u_2\}} e^{-u_1}|Du_1|\,,
\end{multline*}
where we have used that $\|\z_i\|\le 1$ for $i=1,2$ and the Chain Rule.
In conclusion, we have just proved
\begin{equation*}
\int_{\{u_1>u_2\}} e^{-u_2}|Du_2| + \int_{\{u_1>u_2\}} e^{-u_1}|Du_1| -\int_{\{u_1>u_2\}} e^{-u_1}\,|Du_2|-\int_{\{u_1>u_2\}} e^{-u_2}\,|Du_1| \le 0\,,
\end{equation*}
and we are done.
\bigskip
\textbf{STEP 3: The Radon measure ${\bf (\z_1-\z_2,D(T_k(u_1)-T_k(u_2))^+)}$ vanishes for all ${\bf k>0}$.}
Since $u_1$ is a solution to problem \eqref{prob-u1} and $u_2$ is a solution to problem \eqref{prob-u2}, the following equalities hold in $\dis(\Omega)$ (see Proposition \ref{prop-1}):
\begin{equation}\label{dis-u1}
-\hbox{\rm div\,} (e^{-u_1}\z_1)=e^{-u_1}f_1
\end{equation}
and
\begin{equation}\label{dis-u2}
-\hbox{\rm div\,} (e^{-u_2}\z_2)=e^{-u_2}f_2\,.
\end{equation}
Firstly, let $k>0$ we choose the test function $(T_k(u_1)-T_k(u_2))^+$ in equality \eqref{dis-u1}. Applying Green's formula we get
\begin{equation}\label{dis2-u1}
\int_\Omega (e^{-u_1}\z_1,D(T_k(u_1)-T_k(u_2))^+) = \int_\Omega (T_k(u_1)-T_k(u_2))^+ e^{-u_1}f_1\,dx \,,
\end{equation}
and using the same test function but now in equality \eqref{dis-u2} we have
\begin{equation}\label{dis2-u2}
\int_\Omega (e^{-u_2}\z_2,D(T_k(u_1)-T_k(u_2))^+) = \int_\Omega (T_k(u_1)-T_k(u_2))^+ e^{-u_2}f_2\,dx \,.
\end{equation}
Now, we put together \eqref{dis2-u1} and \eqref{dis2-u2} to obtain
\begin{multline}\label{ec-2}
\int_\Omega (T_k(u_1)-T_k(u_2))^+ (e^{-u_1}f_1-e^{-u_2}f_2)\,dx\\
=\int_\Omega (e^{-u_1}\z_1 - e^{-u_2}\z_2,D(T_k(u_1)-T_k(u_2))^+)
\\=\int_{\{ T_k(u_1)>T_k(u_2) \}} (e^{-u_2}\z_2 - e^{-u_1}\z_1,D(T_k(u_2)-T_k(u_1)))\,.
\end{multline}
Observe that the integral on the left hand side is non--positive since $e^{-u_1}f_1-e^{-u_2}f_2\le0$ where
$T_k(u_1)-T_k(u_2)>0$.
Our aim is to prove the following limit:
\begin{equation}\label{limite}
\lim_{k \to \infty} \int_\Omega (e^{-u_1}\z_1 - e^{-u_2}\z_2,D(T_k(u_1)-T_k(u_2))^+) =0 \,,
\end{equation}
which is non--positive because of \eqref{ec-2}. To this end, we write
\begin{multline*}
\int_{\{ T_k(u_1)>T_k(u_2) \}} (e^{-u_2}\z_2 - e^{-u_1}\z_1,D(T_k(u_2)-T_k(u_1)))
\\= \underbrace{\int_{\{ T_k(u_1)>T_k(u_2) \}} ((e^{-u_2}-e^{-u_1})\,\z_2,D(T_k(u_2)-T_k(u_1)))}_{\mbox{(I.1)}}
\\ + \underbrace{\int_{\{ T_k(u_1)>T_k(u_2) \}} (e^{-u_1}(\z_2-\z_1),D(T_k(u_2)-T_k(u_1)))}_{\mbox{(I.2)}} \,,
\end{multline*}
and will see that the limits as $k$ goes to $\infty$ of (I.1) and of (I.2) are non--negative and so \eqref{limite} holds.
On the one hand, we know that
\begin{multline*}
\int_{\{ T_k(u_1)>T_k(u_2) \}} ((e^{-u_2}-e^{-u_1})\,\z_2,D(T_k(u_2)-T_k(u_1)))
\\ \ge \int_{\{ T_k(u_1)>T_k(u_2) \}} (e^{-u_2}-e^{-u_1}){\raise0pt\hbox{{\LARGE $\chi$}}}_{\{u_2<k\}}|Du_2| - \int_{\{ T_k(u_1)>T_k(u_2) \}} (e^{-u_2}-e^{-u_1}){\raise0pt\hbox{{\LARGE $\chi$}}}_{\{u_1<k\}}|Du_1| \,,
\end{multline*}
and when we take limits when $k$ goes to $\infty$, we get
\begin{multline*}
\lim_{k \to \infty} \int_{\{ T_k(u_1)>T_k(u_2) \}} ((e^{-u_2}-e^{-u_1})\,\z_2,D(T_k(u_2)-T_k(u_1)))
\\=\int_{\{u_1>u_2\}} (e^{-u_2}-e^{-u_1}) (|Du_2|-|Du_1|) \ge 0 \,,
\end{multline*}
which is non--negative due to Step 2.
On the other hand, we already know that integral (I.2) is non--negative (because of Step 1), therefore the limit when $k \to \infty$ is non--negative too.
In short, we have proved
\begin{multline*}
\lim_{k \to \infty}\int_{\{ T_k(u_1)>T_k(u_2) \}} (T_k(u_1)-T_k(u_2)) (e^{-u_1}f_1-e^{-u_2}f_2)\,dx
\\ = \lim_{k \to \infty}\int_\Omega (e^{-u_1}\z_1 - e^{-u_2}\z_2,D(T_k(u_1)-T_k(u_2))^+) =0 \,.
\end{multline*}
Furthermore, since \eqref{ec-2}=(I.1)+(I.2) and the limits of integral (I.1) and (I.2) are both non--negative, it follows that both limits vanish.
Now, some remarks on Radon--Nikod\'ym derivatives of these measures are in order.
Let $\theta_k^1(\z_2,DT_k(u_1),x)$ be the Radon--Nikod\'ym derivative of $(\z_2,DT_k(u_1))$ with respect to $|DT_k(u_1)|$:
\begin{equation*}
\theta_k^1(\z_2,DT_k(u_1),x)\,|DT_k(u_1)|=(\z_2,DT_k(u_1)) \,.
\end{equation*}
Since $|(\z_2,DT_k(u_1))| \le |DT_k(u_1)|$, it follows that $|\theta_k^1(\z_2,DT_k(u_1),x)| \le 1$.
We point out that this function is $|DT_k(u_1)|$--measurable and, taking $\theta^1_k(\z_2,DT_k(u_1),x)=0$ in $\{ u_1\ge k\}$, it is $|Du_1|$--measurable.
On the other hand, it holds that $(\z_2,DT_{k+1}(u_1))\res{\{ u_1< k \}}=(\z_2,DT_k(u_1))$. Therefore
\begin{equation*}
\theta^1_{k+1}(\z_2,DT_{k+1}(u_1),x){\raise0pt\hbox{{\LARGE $\chi$}}}_{\{ u_1<k \}}(x)=\theta^1_k(\z_2,DT_k(u_1),x) \,,
\end{equation*}
and $\theta_k^1(\z_2,DT_k(u_1),x)$ defines a non--decreasing sequence of $|Du_1|$--measurable functions.
Likewise, if we denote by $\theta_k^2(\z_1,DT_k(u_2),x)$ the Radon--Nikod\'ym derivative of $(\z_1,DT_k(u_2))$ with respect to $|DT_k(u_2)|$, then we may deduce the inequality $|\theta_k^2(\z_1,DT_k(u_2),x)| \le 1$. Moreover, $\theta_k^2(\z_1,DT_k(u_2),x)$ defines a non--decreasing sequence of $|Du_2|$--measurable functions.
Now, we define the functions $\theta^1(x)$ and $\theta^2(x)$ such that
\begin{equation*}
\theta^1(x)=\theta_k^1(\z_2,DT_k(u_1),x) \quad\mbox{ if }\quad u_1(x)<k \,,
\end{equation*}
and
\begin{equation*}
\theta^2(x)=\theta_k^2(\z_1,DT_k(u_2),x) \quad \mbox{ if }\quad u_2(x)<k \,.
\end{equation*}
We know that $\theta^1$ and $\theta^2$ are $|Du_1|$ and $|Du_2|$--measurable, respectively, and satisfy $|\theta^1| \le 1$ and $|\theta^2| \le 1$.
So let us get back to expression (I.2). We know that
\begin{multline*}
\int_\Omega (e^{-u_1}(\z_1-\z_2),D(T_k(u_1)-T_k(u_2))^+)
\\=\int_{\{ T_k(u_1)>T_k(u_2) \}} e^{-u_1}\bigg[ (\z_1,DT_k(u_1))-(\z_2,DT_k(u_1))-(\z_1,DT_k(u_2))+(\z_2,DT_k(u_2)) \bigg]
\\=\int_\Omega e^{-u_1}{\raise0pt\hbox{{\LARGE $\chi$}}}_{\{ T_k(u_1)>T_k(u_2) \}\cap \{ u_1<k\}} (1-\theta^1(x))|Du_1|
\\+ \int_\Omega e^{-u_1}{\raise0pt\hbox{{\LARGE $\chi$}}}_{\{ T_k(u_1)>T_k(u_2)\}\cap \{ u_2<k\}} (1-\theta^2(x))|Du_2| \,,
\end{multline*}
and using the Convergence Dominated Theorem we can take limit when $k \to \infty$ to arrive at
\begin{equation*}
0=\int_{\{ u_1>u_2 \}} e^{-u_1} (1-\theta^1(x))|Du_1|
+\int_{\{ u_1>u_2 \}} e^{-u_1} (1-\theta^2(x))|Du_2| \,.
\end{equation*}
Since both integrals are non--negative, it yields
\begin{equation*}
0 =\int_{\{ u_1>u_2 \}} e^{-u_1} (1-\theta^1(x))|Du_1|=\int_{\{ u_1>u_2 \}} e^{-u_1} (1-\theta^2(x))|Du_2| \,.
\end{equation*}
Therefore, we deduce that $1-\theta^i(x)=0$ $|Du_i|$--a.e. in $\{ u_1>u_2 \}$ for $i=1,2$ and then, the Radon--Nikod\'ym derivative is $\theta_k^i=1$ $|Du_i|$--a.e. in $\{ u_1>u_2 \} \cap \{u_i<k\}$ with $i=1,2$ for every $k>0$. That is, we have the following equalities as measures:
\begin{equation}\label{med-1}
|DT_k(u_1)|\res{\{ u_1>u_2 \}}=(\z_2,DT_k(u_1))\res{\{ u_1>u_2 \}} \,,
\end{equation}
and
\begin{equation}\label{med-2}
|DT_k(u_2)|\res{\{ u_1>u_2 \}}=(\z_1,DT_k(u_2)) \res{\{ u_1>u_2 \}} \,.
\end{equation}
Finally, noting that $\{T_k(u_1)>T_k(u_2)\}\subseteq \{u_1>u_2\}$ and the measure $(\z_1-\z_2,D(T_k(u_1)-T_k(u_2))^+) $ is non--negative:
\begin{multline*}
(\z_1-\z_2,D(T_k(u_1)-T_k(u_2))^+)
\\=\bigg[|DT_k(u_1)|-(\z_2,DT_k(u_1))-(\z_1,DT_k(u_2))+|DT_k(u_2)|\bigg]\res{\{T_k(u_1)>T_k(u_2)\}}
\\ \le \bigg[|DT_k(u_1)|-(\z_2,DT_k(u_1))-(\z_1,DT_k(u_2))+|DT_k(u_2)|\bigg]\res{\{u_1>u_2\}}
\\=0 \,.
\end{multline*}
\bigskip
\textbf{STEP 4: ${\bf (\z_i,DT_k(u_j))\res{\{T_k(u_1)>T_k(u_2)\}}=|DT_k(u_j)|\res{\{T_k(u_1)>T_k(u_2)\}}}$ as measures for ${\bf i,j=1,2}$ and ${\bf k>0}$.}
Since $\{T_k(u_1)>T_k(u_2)\}\subseteq \{u_1>u_2\}$ and we have proved equalities \eqref{med-1} and \eqref{med-2}, Step 4 is straightforward.
\bigskip
\textbf{STEP 5: If ${\bf u_1>u_2}$, then ${\bf f_1=f_2 =0}$.}
In Step 3 we have proved that the limit of expression \eqref{ec-2} when $k$ goes to $\infty$ is 0. Then, using the Monotone Convergence Theorem, we get
\begin{equation*}
0=\int_\Omega (u_1-u_2)^+(e^{-u_1}f_1-e^{-u_2}f_2) \,dx \,.
\end{equation*}
Notice that if $u_1>u_2$, then $e^{-u_1}f_1= e^{-u_2}f_2$ and $f_1 =e^{-(u_2-u_1)} f_2 > f_2$ when $f_2\ne0$. We conclude that $u_1>u_2$ implies $f_2=f_1=0$.
\bigskip
\textbf{STEP 6: Prove that ${\bf \displaystyle \int_{\{ u_1>u_2 \}} |Du_1| = \int_{\{ u_1>u_2 \}} |Du_2|}$.}
Firstly, we take $T_{\e}((T_k(u_1)-T_k(u_2))^+)$ as a test function in problems \eqref{prob-u1} and \eqref{prob-u2} and using the previous step we get the following equalities:
\begin{multline}\label{ec3-u1}
0 = \int_{\{ T_k(u_1)>T_k(u_2) \}} (\z_1,DT_{\e}(T_k(u_1)-T_k(u_2))) \\+ \int_{\{ T_k(u_1)>T_k(u_2) \}} T_{\e}(T_k(u_1)-T_k(u_2)) \,|Du_1| \,,
\end{multline}
and
\begin{multline}\label{ec3-u2}
0 = \int_{\{ T_k(u_1)>T_k(u_2) \}} (\z_2,DT_{\e}(T_k(u_1)-T_k(u_2))) \\+ \int_{\{ T_k(u_1)>T_k(u_2) \}} T_{\e}(T_k(u_1)-T_k(u_2)) \,|Du_2| \,.
\end{multline}
Now, we use Step 3 to have:
\begin{equation*}
(\z_1-\z_2, D(T_k(u_1)-T_k(u_2)))\res{\{ T_k(u_1)>T_k(u_2) \}} = 0 \,.
\end{equation*}
Furthermore, when we take the restriction to the set $\{ 0 <T_k(u_1)-T_k(u_2)<\e \}$ for all $\e >0$, we also get that the measure vanishes.
Due to this fact, when we consider together equations \eqref{ec3-u1} and \eqref{ec3-u2}, we obtain
\begin{multline*}
\int_{\{ T_k(u_1)>T_k(u_2) \}} T_{\e}(T_k(u_1)-T_k(u_2)) \,|Du_1|\\=\int_{\{ T_k(u_1)>T_k(u_2) \}} T_{\e}(T_k(u_1)-T_k(u_2)) \,|Du_2| \,.
\end{multline*}
Now, dividing both integrals by $\e$ and using the Dominated Convergence Theorem we can take limits as $\e$ goes to 0 and then we get
\begin{equation*}
\int_{\{ T_k(u_1)>T_k(u_2) \}} |Du_1|=\int_{\{ T_k(u_1)>T_k(u_2) \}} |Du_2| \,.
\end{equation*}
Finally, the Dominated Convergence Theorem also allows us to take limits as $k \to \infty$ and so
we arrive at
\begin{equation}\label{ec-3}
\int_{\{ u_1>u_2 \}} |Du_1|=\int_{\{ u_1>u_2 \}} |Du_2| \,.
\end{equation}
\bigskip
\textbf{STEP 7: Prove ${\bf Du_1 =Du_2 =0}$ in ${\bf \{ u_1 > u_2 \}}$.}
We begin taking the test function $(T_k(u_1)-T_k(u_2))^+$ in problem \eqref{prob-u1} and having in mind Step 4 and Step 5, we get:
\begin{multline}\label{ec-4}
0= \int_{\{ T_k(u_1) > T_k(u_2) \}} |DT_k(u_1)|-\int_{\{ T_k(u_1) > T_k(u_2) \}} |DT_k(u_2)|
\\+ \int_{\{ T_k(u_1) > T_k(u_2) \}} (T_k(u_1)-T_k(u_2))\,|Du_1| \,.
\end{multline}
Now, the first two integrals are convergent as $k \to \infty$ due to the Dominated Convergence Theorem and the last one converges by the Monotone Convergence Theorem. Hence, when $k$ goes to $\infty$ in \eqref{ec-4} we get
\begin{equation*}
0= \int_{\{ u_1>u_2 \}} (|Du_1|-|Du_2|) + \int_{\{ u_1>u_2 \}} (u_1-u_2)\,|Du_1| \,,
\end{equation*}
and since the first integral is finite, the last one is finite too.
On the other hand, we have proved in Step 6 that the first integral vanishes, then the above equality becomes
\begin{equation*}
0 = \int_{\{ u_1 > u_2 \}} (u_1-u_2)\,|Du_1| \,,
\end{equation*}
and we deduce that $|Du_1|\res{\{ u_1 > u_2 \}} =0$ and also $Du_1 =0$ in $\{ u_1 > u_2 \}$.
To prove that $Du_2 =0$ in $\{ u_1 > u_2 \}$ we use \eqref{ec-3} and since we already know that $Du_1 =0$ in $\{ u_1 > u_2 \}$, it becomes
\begin{equation*}
0=\int_{\{ u_1 > u_2 \}} |Du_2| \,.
\end{equation*}
Therefore we have that $Du_2 =0$ in $\{ u_1 > u_2 \}$.
\bigskip
\textbf{STEP 8: ${\bf u_1\le u_2}$ in ${\bf \Omega}$.}
We have seen that $D(u_1-u_2) =0$ in $\{ u_1 > u_2 \}$ and since $D^j (u_1-u_2) =0$, it holds that $D(u_1-u_2)^+ =0$ in $\Omega$. Moreover, we know that $(u_1-u_2)^+=0$ in $\partial \Omega$, therefore we get that $0=(u_1-u_2)^+$ in $\Omega$.
\end{proof}
\begin{Corollary}
Let $\Omega$ be a bounded open subset of $\mathbb{R}^N$ with Lipschitz boundary. Let $f$ be a non--negative function in $L^1(\Omega)$. Then, problem
\begin{equation*}
\left\{\begin{array}{ll}
\displaystyle -\hbox{\rm div\,}\left(\frac{Du}{|Du|}\right)+|Du|=f(x)&\hbox{ in }\,\Omega\,,\\[3mm]
u=0 &\hbox{ on }\,\partial\Omega\,,
\end{array}\right.
\end{equation*}
has a unique solution $u \in BV(\Omega)$.
\end{Corollary}
\section{Better summability}
In Section 3 we have seen that problem \eqref{prob-prin} has a solution for every non--negative datum of $L^1(\Omega)$, and this solution belongs to $BV(\Omega)\subset L^{\frac N{N-1}}(\Omega)$. We expect that the solution satisfies a better summability if the datum belongs to $L^p(\Omega)$, $p>1$. In this regard we recall that, when data $f$ are in the space $L^p(\Omega)$ with $p>N$, it is proved in \cite{MS} that the solution is always bounded. For datum $f \in L^N(\Omega)$, we proved in \cite{LS} that the solution belongs to $L^q(\Omega)$ with $1< q < \infty$.
In this section, we are showing that solutions belong to $L^{\frac{Np}{N-p}}(\Omega)$ if data are in $L^p(\Omega)$ with $1< p<N$. Observe that this result adjust continuously for $p=1$ and $p=N$ with the known facts.
\bigskip
\begin{center}
\begin{tabular}{|c|c|}
\hline
&\\
Data & Solution\\
&\\
\hline
\hline
&\\
$ f\in L^p(\Omega) $ with $p>N$ & $u\in L^\infty(\Omega)$\\
&\\
\hline
&\\
$f\in L^{N}(\Omega)$ & $u\in L^q(\Omega)$ with $1\le q < \infty$\\
&\\
\hline
&\\
$f\in L^p(\Omega)$ with $1<p<N$ & $u\in L^{\frac{Np}{N-p}}(\Omega)$\\
&\\
\hline
&\\
$f\in L^1(\Omega)$ & $u\in L^{\frac{N}{N-1}}(\Omega)$ \\
&\\
\hline
\end{tabular}
\end{center}
\bigskip
The proof of our theorem relies on certain preliminary results. The first one enable us to take a power of our solution $u^q$ as a test function in problem \eqref{prob-prin}.
\begin{Proposition}\label{proposition-1}
If $u\in BV(\Omega)$ is a solution to problem \eqref{prob-prin} satisfying $u^q \in L^{p'}(\Omega)$ for certain $q>1$, then $u^q$ and $u^{q+1} \in BV(\Omega)$. Moreover,
\begin{equation}\label{ec-prop}
\int_\Omega |Du^q|+\int_\Omega u^q\,|Du|=\int_\Omega u^q\,f \,.
\end{equation}
\end{Proposition}
\begin{proof}
Fixed $k>0$, we define the function $G_k(s)= s-T_k(s)$, and we take the test function $G_\delta(T_k(u)^q)$ with $\delta, k>0$ in problem \eqref{prob-prin} obtaining the following equality:
\begin{equation*}
\int_\Omega (\z,DG_\delta(T_k(u)^q)) + \int_\Omega G_\delta(T_k(u)^q)\,|Du| = \int_\Omega G_\delta(T_k(u)^q)f\,dx \,.
\end{equation*}
Since we know that the Radon--Nikod\'ym derivative of $(\z,DG_\delta(T_k(u)^q))$ and $(\z,DT_k(u))$ with respect their respective total variations are the same (Proposition \ref{Prop2.2})
and $(\z,DT_k(u))=|DT_k(u)|$ holds for all $k>0$, we deduce that
\begin{equation*}
(\z,G_\delta(T_k(u)^q))=|DG_\delta((T_k(u))^q)|\,.
\end{equation*}
Then we can write:
\begin{equation}\label{lim-ademas}
\int_\Omega |DG_\delta(T_k(u)^q)| + \int_\Omega G_\delta(T_k(u)^q)\,|Du| = \int_\Omega G_\delta(T_k(u)^q)f\,dx \,.
\end{equation}
Now, we use the inequality $G_\delta(T_k(u)^q) \le u^q$ and H\"older's inequality to get the following bound:
\begin{equation*}
\int_\Omega G_\delta(T_k(u)^q)f\,dx \le \int_\Omega u^q f\,dx \le \|u^q\|_{p'}\|f\|_p < \infty \,.
\end{equation*}
Therefore, each integral in left hand side of \eqref{lim-ademas} is also bounded:
\begin{equation}\label{BV-q}
\int_\Omega |DG_\delta(T_k(u)^q)| \le \|u^q\|_{p'}\|f\|_p < \infty \,,
\end{equation}
and
\begin{equation}\label{BV-q+1}
\int_\Omega G_\delta(T_k(u)^q)\,|Du| \le \|u^q\|_{p'}\|f\|_p < \infty \,.
\end{equation}
We will take advantage of these bounds to take limits in \eqref{lim-ademas}.
Now, we are able to prove $u^q \in BV(\Omega)$.
Using the Chain Rule in \eqref{BV-q} we can write the following inequalities
\begin{equation*}
\int_\Omega {\raise0pt\hbox{{\LARGE $\chi$}}}_{\{u<k\}\cap\{u^q > \delta\}}|Du^q| =\int_\Omega |DG_\delta(T_k(u)^q)| \le \|u^q\|_{p'}\|f\|_p < \infty \,,
\end{equation*}
and, using Monotone Convergence Theorem, we let $\delta \to 0^+$ to get
\begin{equation*}
\int_\Omega {\raise0pt\hbox{{\LARGE $\chi$}}}_{\{u<k\}}|Du^q| \le \|u^q\|_{p'}\|f\|_p < \infty \,.
\end{equation*}
Lastly, we let $k$ goes to $\infty$ and appealing to the Monotone Convergence Theorem once more, it works out that $u^q$ is a bounded variation function.
Let $0<\delta<1$ and keeping in mind \eqref{BV-q+1}, we get the following bound
\begin{multline*}
\int_\Omega {\raise0pt\hbox{{\LARGE $\chi$}}}_{\{u<k\}\cap\{u^{q+1} > \delta\}}|Du^{q+1}|
= (q+1)\int_\Omega u^q {\raise0pt\hbox{{\LARGE $\chi$}}}_{ \{u<k\} \cap\{u^{q+1}>\delta\}} |Du|
\\ \le (q+1)\int_\Omega (G_\delta(T_k(u)^q) + \delta)\,|Du|
\\ \le (q+1)(\|u^q\|_{p'}\|f\|_p +\delta\|u\|_{BV})< \infty \,.
\end{multline*}
Taking limits when $\delta \to 0^+$ and also when $k \to \infty$ we get
\begin{equation}\label{exp-1}
\int_\Omega |Du^{q+1}| \le (q+1)\|u^q\|_{p'}\|f\|_p < \infty \,,
\end{equation}
that is, $u^{q+1}\in BV(\Omega)$.
To conclude, we take limits in \eqref{lim-ademas} firstly when $\delta \to 0^+$ and secondly when $k \to \infty$ and then we obtain
\begin{equation*}
\int_\Omega |Du^q| + \int_\Omega u^q |Du| = \int_\Omega u^q \, f \,.
\end{equation*}
\end{proof}
\begin{Theorem}\label{teo-induccion}
Let $1<p<N$ and let $f \in L^p(\Omega)$ be a non--negative function. Then the solution to problem \eqref{prob-prin} belongs to $BV(\Omega)\cap L^{s}(\Omega)$ for every $1\le s<\frac{Np}{N-p}$.
\end{Theorem}
\begin{proof}
Let $u \in BV(\Omega)$ denote the unique solution to problem \eqref{prob-prin}. For every $j\in \mathbb{N}$, we will prove that $u \in L^{s_j}(\Omega)$ where
\begin{equation*}
s_j= N'\sum\limits_{k=0}^j \left(\displaystyle\frac{N'}{p'}\right)^k \,.
\end{equation*}
It should be noted that $\lim\limits_{j\to\infty}s_j=N'\sum\limits_{k=0}^\infty \left(\displaystyle\frac{N'}{p'}\right)^k = \displaystyle\frac{Np}{N-p}$. Thus, proving $u \in L^{s_j}(\Omega)$ for all $j\in\mathbb{N}$, we are done.
Firstly, we choose $q=\frac{N'}{p'}$ and since $u^q \in L^{p'}(\Omega)$, we may apply Proposition \ref{proposition-1} to conclude that $u^{q+1} \in BV(\Omega)\subseteq L^{N'}(\Omega)$ and therefore $u \in L^{N'\left(\frac{N'}{p'}+1\right)}(\Omega)$, that is, $u\in L^{s_1}(\Omega)$.
Assuming now that $u \in L^{s_j}(\Omega)$, we take
\begin{equation*}
q=\frac{N'}{p'}\sum_{k=0}^j \left(\frac{N'}{p'}\right)^k \,.
\end{equation*}
By hypothesis, we already know that $u \in L^{qp'}(\Omega)$, and using Proposition \ref{proposition-1} we get $u^{q+1} \in BV(\Omega) \subseteq L^{N'}(\Omega)$. Hence, $u \in L^{N'(q+1)}(\Omega)= L^{s_{j+1}}(\Omega)$.
\end{proof}
Now we are ready to prove the main result of this section.
\begin{Theorem}\label{teo-regularidad}
Let $f$ be a non--negative function belonging to $L^p(\Omega)$ with $1<p<N$. Then the unique solution $u$ to problem \eqref{prob-prin} satisfies $u \in BV(\Omega)\cap L^{\frac{Np}{N-p}}(\Omega)$.
\end{Theorem}
\begin{proof}
To show that $u \in L^{\frac{Np}{N-p}}(\Omega)$, we first claim that inequality \eqref{claim0} below holds for every $0<q<\displaystyle\frac{N(p-1)}{N-p}$.
If we choose $0<q<\displaystyle\frac{N(p-1)}{N-p}$, then we have that $qp'<\displaystyle\frac{N(p-1)}{N-p}\displaystyle\frac{p}{p-1}=\displaystyle\frac{Np}{N-p}$. Therefore, applying Theorem \ref{teo-induccion} and Proposition \ref{proposition-1} we arrive at $u^{q+1}\in BV(\Omega)$.
Now, we use Sobolev's inequality and inequality \eqref{exp-1} to get
\begin{multline}\label{exp-2}
\left( \int_\Omega u^{(q+1)N'} \,dx\right)^{\frac{1}{N'}} \le C(p,N) \int_\Omega |Du^{q+1}|
\\ \le C(p,N)(q+1)\|f\|_p \left(\int_\Omega u^{qp'} \,dx\right)^{\frac{1}{p'}} \,.
\end{multline}
Moreover, since $qp' < (q+1)N'$, we can apply H\"older's inequality and we also get
\begin{equation*}
\int_\Omega u^{qp'} \,dx \le \left(\int_\Omega (u^{qp'})^{\frac{(q+1)N'}{qp'}} \,dx\right)^{\frac{qp'}{(q+1)N'}} |\Omega|^{1-\frac{qp'}{(q+1)N'}} \,.
\end{equation*}
Summing up, we have
\begin{equation*}
\left( \int_\Omega u^{(q+1)N'} \,dx\right)^{\frac{1}{N'}}
\le C(p,N)(q+1)\|f\|_p \left(\int_\Omega u^{(q+1)N'} \,dx \right)^{\frac{q}{(q+1)N'}} |\Omega|^{\frac{1}{p'}-\frac{q}{(q+1)N'}} \,,
\end{equation*}
that is, we have proved our claim:
\begin{equation}\label{claim0}
\left( \int_\Omega u^{(q+1)N'} \,dx \right)^{\frac{1}{N'}\left( 1-\frac{q}{q+1} \right)} \le C(p,N)(q+1)\|f\|_p |\Omega|^{\frac{1}{p'}-\frac{q}{(q+1)N'}} \,.
\end{equation}
Now, let $0<q_n<\displaystyle\frac{N(p-1)}{N-p}$ define a non--decreasing sequence convergent to $\frac{N(p-1)}{N-p}$. Hence, for every $n \in \mathbb{N}$ it holds
\begin{equation*}
\left( \int_\Omega u^{(q_n+1)N'} \,dx \right)^{\frac{1}{N'}\frac{1}{q_n +1}} \le C(p,N)(q_n+1)\|f\|_p |\Omega|^{\frac{1}{p'}-\frac{q_n}{(q_n+1)N'}} \,.
\end{equation*}
Thanks to Fatou's lemma, letting $n \to \infty$, we get
\begin{multline*}
\int_\Omega u^{\frac{p\,(N-1)}{N-p}N'} \,dx \le \liminf_{n\to \infty} \left[ C(p,N)(q_n+1)\|f\|_p |\Omega|^{\frac{1}{p'}-\frac{q_n}{(q_n+1)N'}} \right]^{N'(q_n+1)}
\\ \le \left[ C(p,N)\frac{p\,(N-1)}{N-p}\|f\|_p \right]^{\frac{Np}{N-p}} \,.
\end{multline*}
Therefore, $u \in L^{\frac{Np}{N-p}}(\Omega)$ holds.
\end{proof}
\begin{remark}
Going back to Proposition \ref{proposition-1}, it follows from $u^{\frac{N(p-1)}{N-p}} \in L^{p^\prime}(\Omega)$ that $u^{\frac{N(p-1)}{N-p}}$ can be taken as a test function in problem \eqref{prob-prin}, that is,
\[
\int_\Omega |D(u^{\frac{N(p-1)}{N-p}})|+\int_\Omega u^{\frac{N(p-1)}{N-p}}|Du|=\int_\Omega fu^{\frac{N(p-1)}{N-p}}\,.
\]
\end{remark}
\section{Explicit examples}
This section is devoted to show radial examples of solutions in a ball.
These examples allow us to provide evidence that our regularity result is sharp (see Remark \ref{regul} below).
In the sequel, we denote by $B_R(0)$ the open ball centered at $0$ and of radius $R$.
\begin{Example}\label{ej-1}
Let $R>0$, we consider problem
\begin{equation}\label{ejemplo-1}
\left\{\begin{array}{ll}
\displaystyle -\hbox{\rm div\,}\left(\frac{Du}{|Du|}\right)+|Du|=\frac{\lambda}{|x|^q} & \hbox{ in
}B_R(0)\,,\\[3mm]
u=0 & \hbox{ on }\partial B_R(0)\,,
\end{array}\right.
\end{equation}
with $1<q<N$ and $\lambda >0$.
\end{Example}
We know that solution $u$ to problem \eqref{ejemplo-1} must be a non--negative function of bounded variation with no jump part and there also exists a vector field $\z \in \DM(\Omega)$ with $\|\z\|_\infty \le 1$ such that
\begin{equation}\label{cond-distribucion}
-\hbox{\rm div\,} \z +|Du| =\frac{\lambda}{|x|^q} \,\text{ in }\,\dis (\Omega)\,,
\end{equation}
\begin{equation*}
(\z, DT_k(u))=|DT_k(u)| \,\text{ as measures in }\, \Omega \;\; (\mbox{for every } k >0)\,,
\end{equation*}
and
\begin{equation*}
u\big|_{\partial \Omega} = 0\,.
\end{equation*}
We assume that the solution is radial, that is, $u(x)=h(|x|)=h(r)$. Moreover, in order to satisfy the Dirichlet condition, we want that $h(R)=0$ holds. In addition, we also assume $h'(r)\le 0$ for all $0\le r\le R$.
If $h^\prime(r)<0$ in an interval, then the vector field is given by $\z(x)=\frac{h^\prime(|x|)}{|h^\prime(|x|)|}=-\frac{x}{|x|}$ and $\hbox{\rm div\,} \z(x)= -\frac{N-1}{|x|}$.
Therefore, equation \eqref{cond-distribucion} becomes
\begin{equation}\label{dinar}
\frac{N-1}{r} - h'(r) = \frac{\lambda}{r^q} \,.
\end{equation}
Since we are assuming that $h'(r)<0$, then
\begin{equation*}
\frac{N-1}{r}- \frac{\lambda}{r^q} <0\,.
\end{equation*}
Now, we define
\begin{equation*}
\rho_\lambda = \left( \frac{N-1}{\lambda}\right)^{\frac{1}{1-q}} \,.
\end{equation*}
Thus, if $r \le \rho_\lambda$, then $h'(r)<0$ may hold, and if $r>\rho_\lambda$ the solution must satisfy $h'(r)=0$.
We assume $0<\rho_\lambda<R$. Then, when $\rho_\lambda \le r\le R$ the solution to problem \eqref{ejemplo-1} is constant, and since we know that $h(R)=0$ we deduce that $h(r)=0$ for all $\rho_\lambda \le r\le R$.
On account of \eqref{dinar}, if $0\le r < \rho_\lambda$, then solution is given by
\begin{multline*}
h(\rho_\lambda)-h(r)=\int_r^{\rho_\lambda} h'(s)\,ds= \int_r^{\rho_\lambda} \left(\frac{N-1}{s}-\frac{\lambda}{s^q}\right) \,ds
\\= (N-1)\log\left(\frac{\rho_\lambda}{r}\right) +\frac{\lambda}{1-q}(r^{1-q}-\rho_\lambda^{1-q})\,.
\end{multline*}
Therefore,
\begin{equation*}
u(x)=\left\{
\begin{array}{lcc}
(N-1)\log\left(\displaystyle\frac{|x|}{\rho_\lambda}\right) +\displaystyle\frac{\lambda}{1-q}(\rho_\lambda^{1-q}-|x|^{1-q}) & \mbox{if} & 0\le |x| <\rho_\lambda \,, \\
0 & \mbox{if} & \rho_\lambda < |x| \le R \,.
\end{array} \right.
\end{equation*}
The vector field $\z$ must be identified.
When $\rho_\lambda \le r\le R$ we know that the vector field is $\z(x)=-\frac{x}{|x|}$, and when $0\le r < \rho_\lambda$, we assume that the vector field is radial: $\z(x)=x\,\xi (|x|)$. Thus, $\hbox{\rm div\,} \z(x)= N\,\xi(|x|)+|x|\,\xi'(|x|)$, and equation \eqref{cond-distribucion} becomes
\begin{equation*}
-(N\,\xi(r)+r\,\xi'(r)) = \frac{\lambda}{r^q} \,.
\end{equation*}
That is,
\begin{equation*}
-r^N\,\xi(r)=-\int \left(r^N\,\xi(r)\right)' \,dr = \int \lambda \, r^{N-1-q} \,dr = \frac{\lambda}{N-q}\, r^{N-q} +C \,,
\end{equation*}
for some constant $C$ to be determinate.
Then
\begin{equation*}
\xi(r) = -\frac{\lambda}{N-q}\, r^{-q} -C\, r^{-N} \,.
\end{equation*}
Since we need a continuous vector field and we know that $\z(x)=-\frac{x}{\rho_\lambda}$ for $x$ with $|x|=\rho_\lambda$, we get the following equation
\begin{equation*}
\rho_\lambda^{-1} =\frac{\lambda}{N-q}\, \rho_\lambda^{-q} +C\, \rho_\lambda^{-N} \,.
\end{equation*}
Finally, using that $\lambda = (N-1)\,\rho_\lambda^{q-1}$ we deduce
\begin{equation*}
C=\rho_\lambda^{N-1}\frac{1-q}{N-q} \,,
\end{equation*}
and therefore, the vector field is given by
\begin{equation*}
\z(x)=\left\{
\begin{array}{lcc}
-\displaystyle\frac{x}{|x|} & \mbox{if} & 0\le |x| <\rho_\lambda \,, \\
-\displaystyle\frac{x}{N-q}\left((N-1) \displaystyle\frac{\rho_\lambda^{q-1}}{|x|^{q}} +(1-q)\displaystyle\frac{\rho_\lambda^{N-1}}{|x|^N}\right) & \mbox{if} & \rho_\lambda < |x| \le R \,.
\end{array}
\right.
\end{equation*}
\begin{remark}\label{regul}
In our Theorem \ref{teo-regularidad} we have prove that if $f\in L^{\frac Nq}(B_R(0))$, then
$u \in L^{\frac N{q-1}}(B_R(0))$.
Since $\frac{\lambda}{|x|^q} \in L^s(B_R(0))$ for all $s<\frac{N}{q}$, it follows that $u \in L^r(B_R(0))$ for all $r<\frac{N}{q-1}$. This is exactly what it is shown.
\end{remark}
\begin{remark}
In \cite[Proposition 4.4]{LS} it was proved that for any ``small'' datum $f \in W^{-1,\infty}(B_R(0))$, the solution to problem \eqref{prob-prin} is always trivial. Nevertheless, in our examples we always get a positive solution. This is due to the fact that the datum $f(x)=\lambda \,|x|^{-q}$ when $1<q<N$ is not in the space $W^{-1,\infty}(B_R(0))$:
Let $s=N-q$, then function $v(x)=|x|^{-s}-R^{-s} \in W^{1,1}_0(B_R(0))$ since $s<N-1$. However, the product $f(x)v(x)=\lambda|x|^{-N}-f(x)R^{-s}\not \in L^1(B_R(0))$. We conclude that $f \not \in W^{-1,\infty}(B_R(0))$.
\end{remark}
It may be worth comparing our example with that occurring when the datum is $\frac{\lambda}{|x|^q}$, with $0<q<1$. In the same way as in Example \ref{ej-1}, the solution to problem \eqref{ejemplo-1} depends on a value
\begin{equation*}
r_\lambda = \left(\frac{N-q}{\lambda}\right)^{\frac{1}{1-q}} \,.
\end{equation*}
When $0<q<1$, the solution to problem \eqref{ejemplo-1} is given by
\begin{equation*}
u(x)=\left\{
\begin{array}{lcc}
(N-1) \log\left(\displaystyle\frac{r_\lambda}{R}\right)+\displaystyle\frac{\lambda}{1-q}(R^{1-q}-r_\lambda^{1-q}) & \mbox{if} & 0 \le |x| \le r_\lambda \,, \\
(N-1)\log\left(\displaystyle\frac{|x|}{R}\right) +\displaystyle\frac{\lambda}{1-q}(R^{1-q}-|x|^{1-q}) & \mbox{if} & r_\lambda< |x| \le R \,,
\end{array} \right.
\end{equation*}
and the vector field associated is
\begin{equation*}
\z(x)=\left\{
\begin{array}{lcc}
-\displaystyle\frac{\lambda}{N-q}x|x|^{-q} & \mbox{if} & 0\le |x| \le r_\lambda \,, \\
-\displaystyle\frac{x}{|x|} & \mbox{if} & r_\lambda < |x| \le R \,.
\end{array}
\right.
\end{equation*}
It is easy to see that, since $0<q<1$, this solution is always bounded.
|
\section{Introduction}
Recently, the possibility to realize strong photon-photon interactions boosted the study of many-body physics with light~\cite{CarusottoRMP13}.
Indeed, new experimental platforms, such as semiconductor microcavities~\cite{WeisbuchPRL92,DeveaudBOOK} and superconducting circuits~\cite{SchoelkopfNature08,YouNature11}, allow one to realize photonic resonators with relatively large nonlinearities.
This enables the achievement of new highly-interacting regimes which, for decades, were confined to the textbook study of quantum optics~\cite{WallsBOOK}.
In this framework, a new flourishing field is that of reservoir engineering, whose goal is the manipulation of the photon exchanges between a nonlinear resonator and the environment~\cite{PoyatosPRL96,VerstraeteNatPhys09,TanPRA13,LinNature13,ArenzJPB13,AsjadJPB14,RoyPRA15,LeghtasScience15}.
These techniques permit the realization of new models with nontrivial drive and dissipation.
In this context, exactly-solvable models are of particular interest.
The analytic solution allows one to explore the full range of system parameters rather than the limiting regimes of small or high photon densities.
The latter are respectively tackled through numerical techniques and semiclassical approximations.
An example of a solvable model is the driven-dissipative Kerr model, for which Drummond and Walls derived the steady-state solution via the complex $P$-representation~\cite{DrummondJPA80a}.
Beyond the single resonator case, analytic solutions proved to be very useful for an efficient implementation of Gutzwiller mean-field decoupling for arrays of coupled cavities~\cite{DiehlPRL10,TomadinPRA11,LeBoitePRL13,LeBoitePRA14,JinPRL13,JinPRA14,WilsonarXiv16}.
In the present work, we use the complex $P$-representation to provide an exact solution for the steady state of a general class of driven-dissipative nonlinear resonators.
More precisely, we consider a standard driven-dissipative Kerr model (subject to the usual coherent pumping and one-photon dissipation) driven by an additional parametric two-photon pump and subject to two-photon losses.
Recently, these processes have been engineered for superconducting resonators~\cite{LeghtasScience15}.
The growing interest towards such kind of models is motivated by the emergence of nonclassical metastable and steady states in their dynamics, such as mixtures of quasi-coherent states or photonic Schr\"odinger cats, which lead to multi-modal Wigner functions~\cite{VogelPRA89,KerckhoffOE11,MingantiSciRep16}.
The possibility to control and protect such states is promising for the implementation of quantum computation protocols~\cite{GilchristJOB04,OurjoumtsevScience06,MirrahimiNJP14,GotoPRA16,PuriarXiv16}.
The exact solution derived in this work allows one to explore the quantum properties of the steady state beyond the semiclassical approximation, capturing the emergence of dissipative phase transitions~\cite{CarmichaelPRX15,CasteelsarXiv16}.
Furthermore, the exact solution paves the way to a numerically-efficient exploration of resonator lattices through Gutzwiller decoupling.
The paper is organized as follows.
We start in Sec.~\ref{Sec:System} by introducing the model.
Than, in Sec.~\ref{Sec:Solution}, we exploit the formalism of the complex $P$-representation to derive the exact solution for the steady-state of the considered model.
Section~\ref{Sec:Study} is devoted to the study of the steady-state properties.
We compare semiclassical and quantum solutions in Sec.~\ref{subsec:semiclassical}.
In Sec.~\ref{Subsec:PhaseTransitions}, we show the emergence of dissipative phase transitions in the thermodynamic limit of large excitation numbers.
Finally, we present conclusions and perspectives in Sec.~\ref{sec:conclusion}.
\begin{figure}[t!]
\includegraphics[width=0.48\textwidth]{Fig_OurScheme}
\caption{A sketch of the considered class of systems.
The picture represents a photon resonator subject to one-photon losses with rate $\gamma$, and coherently driven by a one-photon pump of amplitude $F$.
The resonator is also subject to a coherent two-photon driving of amplitude $G$ and two-photon losses with rate $\eta$.
The strength of the photon-photon interaction is quantified by $U$.
On the right, we sketch the effects of these physical processes on the Fock (number) states~$\vert n \rangle$.}
\label{fig:syst_ph}
\end{figure}
\section{Nonlinear resonator including two-photon processes}\label{Sec:System}
Let us introduce the general model of a driven-dissipative Kerr nonlinear resonator with two-photon drive and dissipation which is sketched in Fig. \ref{fig:syst_ph}. In the the Hamiltonian, $\omega_c$ is the cavity-mode frequency and $U$ the strength of the Kerr nonlinearity, quantifying the photon-photon interaction.
In the absence of pumping we get ($\hbar=1$)
\begin{equation}
\hat{\mathcal{H}}_0=\omega_c\,\hat{a}^\dagger\hat{a}+\frac{U}{2}\hat{a}^\dagger\ad\hat{a}\aop,
\end{equation}
where $\hat{a}$ and $\hat{a}^\dagger$ are, respectively, the annihilation and creation operator for photons inside the resonator.
A coherent drive with amplitude $F$ and frequency $\omega_p$ can be described by
\begin{equation}\label{Eq:H1ph}
\hat{\mathcal{H}}_{\rm1ph}=F\,e^{- i \omega_p t}\hat{a}^\dagger+F^*\,e^{ i \omega_p t}\hat{a}.
\end{equation}
From now on we will denote this mechanism as one-photon pumping.
Similarly, a parametric process coherently adding photons pairwise is described by
\begin{equation}\label{Eq:H2ph}
\hat{\mathcal{H}}_{\rm2ph}=\frac{G}{2}\,e^{-i \omega_2t}\hat{a}^\dagger\ad+\frac{G^*}{2}\,e^{i \omega_2t}\hat{a}\aop,
\end{equation}
where $G$ is the pump amplitude and $\omega_2$ its frequency.
Such a two-photon pumping mechanism can be obtained by engineering the exchange of photons between the cavity and the environment.
Recently, this has been realized by coupling two superconducting resonators via a Josephson junction~\cite{LeghtasScience15}.
In order to get a time-independent Hamiltonian, we consider $\omega_2=2\omega_p$.
Hence, we use the unitary transformation $\hat{\mathcal{U}}=e^{- i \omega_pt\hat{a}^\dagger\hat{a}}$, which removes the time-dependence from the Hamiltonian.
This allows us to describe the system in the reference frame rotating at the coherent pump frequency $\omega_p$.
The full Hamiltonian, hence, becomes
\begin{align}\label{Eq:HCompleteOur}
\hat{\mathcal{H}}=&-\Delta\hat{a}^\dagger\hat{a}+\frac{U}{2}\hat{a}^\dagger\ad\hat{a}\aop
\nonumber\\
&+F\hat{a}^\dagger+F^*\hat{a}+\frac{G}{2}\hat{a}^\dagger\ad+\frac{G^*}{2}\hat{a}\aop,
\end{align}
where $\Delta=\omega_p-\omega_c$ is the pump-cavity detuning.
For the considered system, photon losses are typically appreciable and can not be neglected~\cite{HarocheBOOK}.
The Markov-Born approximation gives an excellent description of these losses in terms of a Lindblad dissipation super-operator $\mathcal{D}(\hat{\mathcal{C}})$ of the form ~\cite{HarocheBOOK, CarmichaelBOOK}
\begin{equation}
\mathcal{D}(\hat{\mathcal{C}})\,\hat{\rho}=
2\,\hat{\mathcal{C}}\,\hat{\rho}\,\hat{\mathcal{C}}^\dagger
-\hat{\mathcal{C}}^\dagger \hat{\mathcal{C}}\,\hat{\rho}
-\hat{\rho}\,\hat{\mathcal{C}}^\dagger\hat{\mathcal{C}},
\end{equation}
where $\hat{\mathcal{C}}$ is the quantum jump operator corresponding to the specific dissipation process.
Usually, photons are lost individually to the environment and the jump operator is the annihilation operator $\hat{a}$~\cite{HarocheBOOK}. In addition, we also consider two-photon losses, which naturally emerge together with the engineered two-photon pumping~\cite{LeghtasScience15}.
These losses are included through the jump operator $\hat{a}^2$.
The resulting Lindblad master equation describing the evolution of the the system density matrix $\hat{\rho}$ is
\begin{equation}\label{Eq:MasterEquationOur}
i \frac{\partial\hat{\rho}}{\partial t}=
\left[\hat{\mathcal{H}},\hat{\rho}\right]
+ i \frac{\gamma}{2}\,\mathcal{D}(\hat{a})\,\hat{\rho}
+ i \frac{\eta}{2}\,\mathcal{D}(\hat{a}^2)\,\hat{\rho},
\end{equation}
where $\gamma$ and $\eta$ are, respectively, the one- and two-photon dissipation rates and $\hat{\mathcal{H}}$ is the one given in Eq.~\eqref{Eq:HCompleteOur}.
\section{$P$-representation and exact solution for the steady state}\label{Sec:Solution}
The steady-state properties are of central interest in the context of out-of-equilibrium quantum systems. These properties are encoded in the steady-state density matrix, which is the solution of Eq.~\eqref{Eq:MasterEquationOur} for $\partial_t\hat{\rho}=0$.
To this purpose, we consider the $P$-representation of the density matrix, i.e. we decompose $\hat{\rho}$ using the over-complete basis of coherent states $\ket{\alpha}$, such that $\hat{a}\ket{\alpha}=\alpha\ket{\alpha}$.
We use the complex $P$-representation $P(\alpha,\beta)$~\cite{DrummondJPA80b}, which is defined by
\begin{equation}\label{Eq:PRepresentation}
\hat{\rho}=\int_\mathcal{C}\!\!d\alpha \int_{\mathcal{C}'}\!\!d\beta\,
\frac{\ket{\alpha}\bra{\beta^*}}{\braket{\beta^*|\alpha}}\,
P(\alpha,\beta),
\end{equation}
where the closed integration contours $\mathcal{C}$ and $\mathcal{C}'$ must be carefully chosen to encircle all the singularities of the function $P(\alpha,\beta)$.
Once the definition~\eqref{Eq:PRepresentation} is inserted in Eq.~\eqref{Eq:MasterEquationOur}, the action of the annihilation and creation operators on the projector $\ket{\alpha}\!\bra{\beta^*}$ allows one to map the master equation for $\hat{\rho}$ into a complex Fokker-Planck equation for $P(\alpha,\beta)$. Further details on this procedure are presented in appendix~\ref{App:Solution}.
For the case $G=0$, the complex $P$-representation solution for the steady state was derived by Drummond and Walls~\cite{DrummondJPA80a}, and is given by
\begin{equation}\label{Eq:PDrummondWalls}
P_{\rm ss}(\alpha,\beta)\propto e^{2\alpha\beta}
\frac{e^{-2f/\alpha}}{\alpha^{2+2c}}
\frac{e^{-2f^*\!/\beta}}{\beta^{2+2c^*}}.
\end{equation}
In Eq.~\eqref{Eq:PDrummondWalls}, the system parameters are resumed by the dimensionless quantities \mbox{$c=(\Delta+i\gamma/2)/(U-i\eta)$} and \mbox{$f=F/(U-i\eta)$}.
For the general case corresponding to the master equation~\eqref{Eq:MasterEquationOur}, we find
\begin{equation}\label{Eq:PRepOur}
\begin{split}
P_{\rm ss}(\alpha,&\beta)=\frac{
e^{2\alpha\beta}}{\mathcal{N}}
\frac{1}{\left(\alpha^2+g\right)^{1+c}}
\exp\left[-\frac{2f}{\sqrt{g}}\arctan\left(\frac{\sqrt{g}}{\alpha}\right)\right]\\
&\times\,\frac{1}{\left(\beta^2+g^*\right)^{1+c^*}}
\exp\left[-\frac{2f^*}{\sqrt{g^*}}\arctan\left(\frac{\sqrt{g^*}}{\beta}\right)\right].
\end{split}
\end{equation}
All details on the derivation of Eq.~\eqref{Eq:PRepOur} are given in appendix~\ref{App:Solution}.
In Eq.~\eqref{Eq:PRepOur} we introduced the dimensionless parameter \mbox{$g=G/(U-i\eta)$}.
We stress that in the limit $g\to0$ Eq.~\eqref{Eq:PRepOur} reduces to Eq.~\eqref{Eq:PDrummondWalls}, as expected. We note that some particular cases have been considered in~\cite{KryuchkyanOC96,MeaneyEPJQT14,ElliottarXiv16}.
The normalization factor $\mathcal{N}$ in Eq.~\eqref{Eq:PRepOur} ensures that $\Tr{\hat{\rho}}=1$.
By imposing this condition we get
\begin{align}
\mathcal{N}=&
\int_\mathcal{C}\!\!d\alpha \int_{\mathcal{C}'}\!\!d\beta\,\,
e^{2\alpha\beta} \frac{1}{\left(\alpha^2+g\right)^{1+c}}
\frac{1}{\left(\beta^2+g^*\right)^{1+c^*}}
\nonumber\\
\times&\exp\left[-\frac{2f}{\sqrt{g}}\arctan\left(\frac{\sqrt{g}}{\alpha}\right)
-\frac{2f^*}{\sqrt{g^*}}\arctan\left(\frac{\sqrt{g^*}}{\beta}\right)\right].
\end{align}
One can Taylor-expand $e^{2\alpha\beta}$ and swap the resulting sum with the integral.
The two contour integrals over $\alpha$ and $\beta$ thus decouple, leading to
\begin{align}\label{Eq:Normalization}
\mathcal{N}=\sum_{m=0}^{\infty} \frac{2^m}{m!}\, \left|\mathcal{F}_m\left(f,g,c\right)\right|^2,
\end{align}
where we introduced
\begin{equation}\label{Eq:DefinitionFFormal}
\mathcal{F}_m\left(f,g,c\right)=\!\!\int_\mathcal{\!C}
\frac{\alpha^m\, d\alpha}{\left(g+\alpha^2\right)^{1+c}}\,
\exp\left[
-\frac{2f}{\sqrt{g}}\,\arctan\left(\frac{\sqrt{g}}{\alpha}\right)
\right]\!.
\end{equation}
Note that, $\mathcal{F}_m\left[f^*,g^*,c^*\right]=\mathcal{F}_m^*\left[f,g,c\right]$.
Performing the integral in Eq.~\eqref{Eq:DefinitionFFormal} requires an appropriate choice
of the contour $\mathcal{C}$.
In the present case, we used the Pochhammer path (more details are given in appendix~\ref{App:Solution}), which gives
\begin{equation}
\label{Eq:FormulaFcomplete}
\mathcal{F}_m\left(f,g,c\right)=\left(i\sqrt{g}\right)^m
\,_2 F_1\left(-m,-c-i\,f/\sqrt{g};-2c;2\right),
\end{equation}
where $\,_2 F_1$ is the Gauss hypergeometric function~\cite{BaileyBOOK}.
\subsection{Exact results for steady-state quantities}\label{subsec:ssvalues}
The steady-state quantities can be expressed in terms of the $\mathcal{F}_m$ functions~\eqref{Eq:FormulaFcomplete}.
Let us consider the correlation functions
\begin{align}\label{Eq:CorrijFormal}
\corr{i}{j}&=\Tr{\ada{i}{j}\,\hat{\rho}}
\nonumber\\&=
\!\!\int_\mathcal{C}\!\!d\alpha \int_{\mathcal{C}'}\!\!d\beta\,\,
\frac{P(\alpha,\beta)}{\braket{\beta^*|\alpha}}\,\,
\Tr{\ada{i}{j}\ket{\alpha}\bra{\beta^*}}.
\end{align}
Since $\Tr{\ada{i}{j} \ket{\alpha}\bra{\beta^*}}=\alpha^j\beta^i\braket{\beta^*|\alpha}$, we have
\begin{equation}
\label{Eq:CorrijGeneral}
\corr{i}{j}=\frac{1}{\mathcal{N}}\,\sum_{m=0}^{\infty} \frac{2^m}{m!}\, \mathcal{F}_{m+j}\left[f,g,c\right]\, \mathcal{F}_{m+i}^*\left[f,g,c\right].
\end{equation}
Similarly, one can derive the matrix elements of the steady-state density matrix $\hat{\rho}_{\rm ss}$ in the Fock basis:
\begin{align}\label{Eq:RhopqGeneral}
&\braket{p|\hat{\rho}_{\rm ss}|q}=
\!\!\int_\mathcal{C}\!\!d\alpha \int_{\mathcal{C}'}\!\!d\beta\,\,
\frac{P(\alpha,\beta)}{\braket{\beta^*|\alpha}}
\frac{\alpha^p\beta^q}{\sqrt{p!q!}}
\nonumber\\
&\quad=\frac{1}{\mathcal{N}\,\sqrt{p!q!}} \sum_{m=0}^{\infty} \frac{1}{m!}
\mathcal{F}_{m+p}\left[f,g,c\right]\, \mathcal{F}_{m+q}^*\left[f,g,c\right].
\end{align}
An useful tool to visualize the properties of the steady state is the Wigner function~\cite{WignerPR32}.
Given a density matrix $\hat{\rho}$, the corresponding Wigner function $W(z)$ is a real-valued function of the complex variable $z$, defined as~\cite{CahillPR69}
\begin{equation}
W(z)=\frac{2}{\pi}\,\Tr{\hat{D}_z\, e^{i\pi \hat{a}^\dagger\hat{a}}\, \hat{D}_z^\dagger\, \hat{\rho}},
\end{equation}
with $\hat{D}_z=e^{z \hat{a}^\dagger- z^* \hat{a}}$ the displacement operator.
Substituting $\hat{\rho}$ with its $P$-representation, the crucial quantity to evaluate is $\Tr{\hat{D}_z\, e^{i\pi \hat{a}^\dagger\hat{a}}\, \hat{D}_z^\dagger\ket{\alpha}\bra{\beta^*}}$.
Using the identity $\hat{D}_z^\dagger\,\hat{D}_\alpha=e^{(\alpha z^*\!-z\alpha^*)/2} \hat{D}_{\alpha-z}$, after lengthy but straightforward calculations, one gets
\begin{align}
\Tr{\hat{D}_z\, e^{i\pi \hat{a}^\dagger\hat{a}}\, \hat{D}_z^\dagger\ket{\alpha}\bra{\beta^*}}
=\braket{\beta^*|\alpha}\frac{e^{2\alpha z^*} e^{2\beta z}}{e^{2\alpha\beta} e^{2|z|^2}}.
\end{align}
The Wigner function can thus be written as
\begin{equation}
W(z)=\frac{2\,e^{-2|z|^2}}{\pi}
\int_\mathcal{C}\!\!d\alpha \int_{\mathcal{C}'}\!\!d\beta\,\,
\frac{P(\alpha,\beta)}{e^{2\alpha\beta}}\,
e^{2\alpha z^*} e^{2\beta z}.
\end{equation}
This time, the integrals over $\alpha$ and $\beta$ are already independent.
By Taylor expanding the exponentials, we find that
\begin{equation}
\label{Eq:WignerGeneral}
W(z)=\frac{2}{\pi\,\mathcal{N}} \left|
\sum_{m=0}^{\infty} \frac{(2z^*)^m}{m!}\, \mathcal{F}_m\left[f,g,c\right]
\right|^2 e^{-2|z|^2}.
\end{equation}
Therefore, the Wigner function given in Eq.~\eqref{Eq:WignerGeneral} is real and positive over the whole complex plane for \emph{any} choice of the system parameters.
We point out that Eqs.~\eqref{Eq:Normalization}, \eqref{Eq:CorrijGeneral}, \eqref{Eq:RhopqGeneral}, and~\eqref{Eq:WignerGeneral}, together with the definition of $\mathcal{F}_m$ given in Eq.~\eqref{Eq:FormulaFcomplete}, summarizes the exact analytic results of this work.
For sake of completeness, in the case $g=0$, the definition~\eqref{Eq:FormulaFcomplete} can be reduced to \mbox{$\mathcal{F}_m(f,0,c)=(-2f)^m/\Gamma(m-2c)$}~\cite{DrummondJPA80a}.
Although the exact results presented here apply for the general case of complex $F$ and $G$, in what follows, unless differently specified, we will take them as real parameters.
\subsection{Benchmarking in the low-density regime}
\begin{figure}[t]
\includegraphics[width=0.42\textwidth]{Fig_Low_F0}\vspace{0.3cm}
\includegraphics[width=0.42\textwidth]{Fig_Low_FG}
\caption{(Color online)
Mean photon number $\braket{\ada{}{}}$ as a function of the pump-cavity detuning $\Delta$ normalized by the photon-photon interaction strength $U$.
Different curves and data sets correspond to different pump intensities (cf. legend).
Solid lines represent the analytic solution while the points are the numerical results obtained by diagonalization of the Liouvillian superoperator of the master equation on a truncated Fock basis.
Top: results in the absence of one-photon pumping, i.e. $F=0$ [Eq.~\eqref{Eq:CorrijNoF} for $i=j=1$].
Bottom: results in the presence of both one- and two-photon pumping with $F=G$ [Eq.~\eqref{Eq:CorrijGeneral}].
In both panels, vertical dot-dashed red (dashed blue) lines mark the position of odd (even) photonic resonances.
One- and two-photon dissipation rates were set to $\gamma=\eta=0.03U$.}
\label{Fig:Numerics}
\end{figure}
Before exploiting the analytic solution, note that the results summarized in Eqs.~\eqref{Eq:Normalization}, \eqref{Eq:CorrijGeneral}, \eqref{Eq:RhopqGeneral}, and~\eqref{Eq:WignerGeneral}, contain infinite sums of $\mathcal{F}_m$ functions.
In the special cases $G=0$ (studied in~\cite{DrummondJPA80a}) or $F=0$ (cf. appendix~\ref{subsec:closedForms}) such sums can be analytically computed, resulting in combinations of hypergeometric functions.
For the general case of finite one- and two-photon pumping (i.e., $F,G\neq0$), the series can be computed with arbitrary precision (see appendix~\ref{Subsec:Convergence} for further details).
In order to benchmark the analytic solution with numerical approaches, we study it in the low-density regime.
The two panels of Fig.~\ref{Fig:Numerics} show the results obtained in the presence of only two-photon pumping (top) and for both one- and two-photon driving (bottom).
The agreement with numerics is excellent, thus corroborating the validity of the analytic solution.
The parameters have been chosen to clearly visualize the photonic resonances, which are expected when the energy of $n$ pump photons is equal to that of $n$ photons inside the resonator.
Thus, beside the one-photon resonance occurring for $\Delta=0$, the multi-photon resonances arise when $\Delta/U=(n-1)/2$.
For $F=0$ only resonances with an even number of photons appear, while all of them are observed in the presence of a one-photon pumping.
The resonances progressively merge in a continuum by increasing the pump intensities.
In the high-density regime this behavior triggers a dissipative phase transition~\cite{CarmichaelPRX15}, discussed below in Sec.~\ref{Subsec:PhaseTransitions}.
\section{Properties of the steady state}\label{Sec:Study}
\begin{figure}[t]
\includegraphics[width=0.48\textwidth]{Fig_GP}
\caption{(Color online)
Left: mean steady-state photon number $\corr{}{}$ as a function of the dimensionless detuning parameter $\Delta/U$.
The green circles (red crosses) mark the stable (unstable) semiclassical steady-state solutions.
The black line is the analytic solution given by Eq.~\eqref{Eq:CorrijGeneral} ($i=j=1$).
One- and two-photon dissipation rates were set to $\gamma=\eta=0.1U$.
Right: zoom-in on the region in which the almost-degenerate high-density semiclassical solutions get unstable.}
\label{Fig:GP}
\end{figure}
\begin{figure*}[t]
\includegraphics[width=0.85\textwidth]{Fig_Wigners}
\caption{(Color online)
Top: steady-state photon number $\corr{}{}$ as a function of the dimensionless detuning parameter $\Delta/U$ for $F=U$, $G=10U$, $\gamma=\eta=0.1U$.
The results have been obtained through the exact solution~\eqref{Eq:CorrijGeneral} for $i=j=1$.
The vertical grid lines mark the values of $\Delta/U$ for which we evaluated the steady-state Wigner function (cf. bottom panels).
Bottom: steady-state Wigner functions $W(z)$ calculated according to Eq.~\eqref{Eq:WignerGeneral} for the same parameters as in the top panel and for different values of $\Delta/U$ (see frame labels).
The black dots mark the position of the corresponding stable semiclassical solutions.}
\label{Fig:Wigners}
\end{figure*}
The exact analytic solution of the Lindblad equation~\eqref{Eq:MasterEquationOur}, derived and benchmarked in Sec.~\ref{Sec:Solution}, allows us to compute the average steady-state quantities of the considered system in any physical regime, from low- to high-density phases, passing through the nontrivial mesoscopic regime.
In this section, we investigate how the properties of the steady state evolve through these different regimes.
\subsection{Quantum \emph{vs} semiclassical behavior}\label{subsec:semiclassical}
When the resonator has a small population, the solution of the master equation~\eqref{Eq:MasterEquationOur} can be obtained numerically.
For instance, by integrating the master equation on a truncated Fock basis or by diagonalizing the Liouvillian super-operator~\cite{CasteelsPRA16}.
On the other hand, for high photon number the cavity field can be typically approximated by a coherent state $\ket{\alpha}$.
Thus, the master equation reduces to a differential equation for the complex amplitude $\alpha$.
This corresponds to the so-called semiclassical approximation~\cite{CarusottoRMP13}.
In this case, however, all quantum correlations are neglected, which makes our exact analytic solution a precious tool.
The differential equation for the complex amplitude $\alpha$ can be easily derived from $\partial_t\braket{\hat{a}}=\Tr{\hat{a}\,\partial_t \rho}$, by assuming $\hat{\rho}=\ket{\alpha}\!\bra{\alpha}$, namely:
\begin{equation}\label{Eq:GP}
\partial_t\alpha=(i\Delta-\gamma/2)\alpha -iF -iG\alpha^* -(iU+\eta)\alpha^*\alpha^2.
\end{equation}
Note that the latter equation is coupled to the one for the conjugate variable $\alpha^*$.
Solving for the steady state $\partial_t\alpha,\partial_t\alpha^*=0$ one can get, depending on the system parameters, up to five solutions, of which at most three are dynamically stable~\cite{MeaneyEPJQT14,ElliottarXiv16}.
In Fig.~\ref{Fig:GP} we show the semiclassical prediction for the mean photon number according to the semiclassical analysis.
For large and negative detuning, Eq.~\eqref{Eq:GP} predicts a single low-density steady-state solution.
By increasing $\Delta$, the low-density solution gets unstable and two high-density ones emerge.
The corresponding values of $\corr{}{}$ are almost equal, but the phases of their complex amplitudes differ approximatively by $\pi$.
Eventually, a third low-density stable solution appears, coexisting with the two high-density ones until a parameter-dependent threshold is reached (see zoom-in panel in Fig.~\ref{Fig:GP}).
Then, only the low-density stable state is present.
By comparing these results with the exact one given by Eq.~\eqref{Eq:CorrijGeneral} (also plotted in Fig.~\ref{Fig:GP}), we note that the multi-stable behavior does not appear in the analytic solution.
We point out that the quantum solution is unique, while the semiclassical approach gives multiple dynamically stable solutions.
However, the exact and unique quantum solution can display a multimodal mixed-state behavior.
The presence of one (or more) semiclassical solution(s) in the steady state can be visualized by the Wigner function $W(z)$, whose analytic expression is in Eq.~\eqref{Eq:WignerGeneral}.
The case $F=0$ has already been discussed in~\cite{MeaneyEPJQT14}, in particular the evolution of $W(z)$ across the density drop.
We present, in Fig.~\ref{Fig:Wigners}, the results for the general case $F,G\neq0$.
In the multiple-solution regime, even for $F/G\ll1$, the one-photon driving prevents the system from being in a balanced mixture of coherent states, which is the case for $F=0$~\cite{MeaneyEPJQT14,MirrahimiNJP14,LeghtasScience15,MingantiSciRep16,ElliottarXiv16}.
By looking at the bottom panel of Fig.~\ref{Fig:Wigners}, one notes that a bimodal Wigner function only exists nearby the transitions from low- to high-density regimes.
Elsewhere, $W(z)$ always exhibits a single peak.
In the low-density regimes, we recover a squeezed-vacuum steady state as the one observed for $F=0$~\cite{MeaneyEPJQT14,ElliottarXiv16}.
This squeezing of the state can be seen by looking at the elongated elliptic shape of the corresponding Wigner function in the bottom panels of Fig.~\ref{Fig:Wigners}.
\begin{figure}[t]
\includegraphics[width=0.48\textwidth]{Fig_PhaseF}
\caption{(Color online)
Steady-state Wigner functions $W(z)$ calculated according to Eq.~\eqref{Eq:WignerGeneral} for $\Delta=28U$, $G=10U$, $\gamma=\eta=0.1U$ and for different complex values of $F$.
For panel~(e) we took $F=0$.
In the others, $F/U=e^{i\phi}$ and the phase $\phi$ changes as sketched in the bottom-right scheme.}\label{Fig:PhaseF}
\end{figure}
Remarkably, as shown in Fig.~\ref{Fig:PhaseF}, the dominant peak in the multi-modal Wigner function is selected by the relative phase of $F$ and $G$.
For this analysis, we took the same parameters as in Fig.~\ref{Fig:Wigners}, setting the detuning around the threshold value.
In the outer panels we have varied the relative phase $\phi=\arg(F/G)$, changing the relative weight of the Wigner-function peaks.
The central panel~(e) shows the case $F=0$, for which the three peaks have comparable heights.
This property can be a valuable tool for the control of two-photon driven resonators for quantum computation based on quasi-orthogonal coherent states~\cite{GotoPRA16,PuriarXiv16}.
Indeed the relative phase $\phi$ could be experimentally controlled and adjusted at will.
In this direction, it is worth stressing that expression~\eqref{Eq:WignerGeneral} allows to predict precisely the shape of the multi-modal Wigner function even in highly populated regimes, where a numerical approach would be extremely demanding.
\subsection{Emergence of dissipative phase transitions}\label{Subsec:PhaseTransitions}
\begin{figure}[t]
\includegraphics[width=0.45\textwidth]{Fig_BreakdownG0}
\caption{(Color online)
Mean photon number $\braket{\ada{}{}}$ (top panel) and normalized second-order correlation function $g^{(2)}$ (bottom panel) as a function of the pump-cavity detuning $\Delta$ normalized by the photon-photon interaction strength $U$ for a resonator subject only to one-photon coherent driving ($G=0$, $F\neq0$).
Different curves (and colors) correspond to different pumping intensities $F/U$, varied between 0.1 and 300, as indicated beside each curve in the top panel.
One- and two-photon dissipation rates were set to $\gamma=\eta=0.03U$.}
\label{Fig:BreakdownG0}
\end{figure}
In this section, we show how our analytic solution allows to capture the steady-state properties of dissipative phase transitions in the thermodynamic limit.
The latter, in the present context, is defined as the regime in which $\corr{}{}\to+\infty$~\cite{CarmichaelPRX15,CasteelsarXiv16}.
Let us start by considering the case in which the resonator is subject only to a coherent drive (i.e., $G=0$).
In the top panel of Fig.~\ref{Fig:BreakdownG0} we show the evolution of the mean photon density $\corr{}{}$ as a function of the detuning for different driving amplitudes $F$.
For a small drive amplitude $F\lesssim U$, the photon number shows well-resolved multi-photon resonances.
In the intense-pumping regime $F\gg U$, instead, these resonances are replaced by a continuous and monotonous increase of the photon density, up to a sharp transition from a high- to a low-density phase.
Corresponding to the drop in $\corr{}{}$, the normalized second-order correlation function $g^{(2)}$ exhibits a sharp peak, shown in the bottom panel of Fig.~\ref{Fig:BreakdownG0}.
This quantity is defined as \mbox{$g^{(2)}=\corr{2}{2}/\corr{}{}^2$}.
At the transition, photons have a highly super-Poissonian distribution ($g^{(2)}\gg1$).
\begin{figure}[t]
\includegraphics[width=0.45\textwidth]{Fig_BreakdownF0}
\caption{(Color online)
Same as Fig.~\ref{Fig:BreakdownG0}, but in the presence of two-photon driving only, i.e. $F=0$ and $G\neq0$. Different curves (and colors) correspond to different pumping strengths $G/U$, spanning from 0.1 to 300 as labeled beside each curve in the top panel.
One- and two-photon dissipation rates were set to $\gamma=\eta=0.03U$.}
\label{Fig:BreakdownF0}
\end{figure}
A similar analysis can be performed in the presence of two-photon pumping.
The results obtained for $F=0$ and different values of $G/U$ are presented in Fig.~\ref{Fig:BreakdownF0}.
In the top panel, we observe a similar behavior of the photon density as in Fig.~\ref{Fig:BreakdownG0}.
Note that the analytic solution allows us to reach very high values of $\corr{}{}$ (up to $\sim1300$ photons for $G=300U$).
The behavior of the second-order correlation function $g^{(2)}$ dramatically differs from the case $G=0$ considered in Fig.~\ref{Fig:BreakdownG0}.
For $G\geqslant10U$, we find a sub-Poissonian statistics ($g^{(2)}<1$) for small $\Delta$ and a strong peak corresponding to the drop in density.
After the peak, in the low-density phase, $g^{(2)}$ drops but stays considerably larger than one and, furthermore, it keeps growing roughly quadratically.
This high probability of observing photons pairwise is a clear consequence of the two-photon pumping mechanism.
\begin{figure}[t!]
\includegraphics[width=0.42\textwidth]{Fig_PhaseTransitionG0}\vspace{0.5cm}
\includegraphics[width=0.42\textwidth]{Fig_PhaseTransitionF0}
\caption{(Color online)
Top: for the case $G=0$ (coherent driving only), rescaled mean photon density $\chi=\corr{}{}/|f|^{2/3}$ as a function of the dimensionless parameter $\tau={\rm sgn}[\Delta]|c|/|f|^{2/3}$ for different values of the dimensionless coherent drive intensity $|f|$ (see legend).
The smoothest curve corresponds to $|f|=1$, while for increasing values of $|f|$ the curve gets steeper, acquiring a triangular shape.
The points in the inset mark height and position of the peak in $\partial\chi/\partial\tau$ as a function of $|f|$.
The solid lines are power-law fits with exponents $\pm2/3$, performed on the last four points.
Bottom: for the case $F=0$ (two-photon driving only), rescaled mean photon density $\chi=\corr{}{}/|g|$ as a function of $\tau={\rm sgn}[\Delta]|c|/|g|$ for different values of $|g|$ (see legend).
The smoothest curve corresponds to $|g|=1$, while $\chi$ progressively tends to a triangular-shaped curve for increasing $|g|$.
In the inset we show the rapid growth in the derivative $\partial\chi/\partial\tau$ around $\tau=-1$. The smoothest behavior corresponds to $|g|=100$ and the curve progressively acquires a discontinuity by increasing $|g|$ (cf. inset legend).
Overall, dissipations have been set to $\eta=0.1U$ and $\gamma=0.1|\Delta|$.}
\label{Fig:PhaseTransition}
\end{figure}
The abrupt change in the density observed above both for $G=0$ and $F=0$ is the result of a dissipative phase transition.
This kind of phenomenon arises in the nonequilibrium context due to the interplay of nonlinearity, drive and dissipation~\cite{AlsingQO91,KilinJOSAB91,CarmichaelPRX15}.
Hence, a dissipative phase transition similar to the one studied numerically by Carmichael for the Jaynes-Cummings model~\cite{CarmichaelPRX15} is also present in our Kerr system.
Our exact solutions proves it unambiguously and allows us to capture also the critical exponents.
In order to further characterize the transition, we consider a scaling which leads to an universal behavior in the thermodynamic limit.
In the coherent-pumping case $G=0$, starting from the semiclassical equation~\eqref{Eq:GP}, one finds that for a large photon number $\corr{}{}\propto|f|^{2/3}$ [as a reminder, $f=F/(U-i\eta)$].
Hence, we expect an universal behavior of the quantity $\chi=\corr{}{}|f|^{-2/3}$.
In Fig.~\ref{Fig:BreakdownG0}, we saw that the high-to-low density transition is triggered by the detuning $\Delta$.
In a more general description, we expect the phase transition to be controlled by the dimensionless complex detuning $c=(\Delta+i\gamma/2)/(U-i\eta)$.
Hence, in the top panel of Fig.~\ref{Fig:PhaseTransition} we show the behavior of $\chi(\tau)$ for $\tau={\rm sgn}[\Delta]|c||f|^{-2/3}$.
In the thermodynamic limit $|f|\to\infty$, $\chi(\tau)$ shows a discontinuous first-order phase transition.
For finite values of $|f|$, the derivative $\partial\chi/\partial\tau$ is peaked at the transition point.
We find that the height and position of this peak follow the power-law behaviors $\max[\partial\chi/\partial\tau]\propto|f|^{2/3}$ and $|\tau_{\rm max}-\tau_c|\propto|f|^{-2/3}$ [cf. inset of Fig.~\ref{Fig:PhaseTransition}~(top)].
For the selected parameters, we find $\tau_c\sim2.41$.
We now perform the same analysis for the two-photon driven case $F=0$, for which, in the thermodynamic limit, one expects $\corr{}{}\propto|g|$ [with $g=G/(u-i\eta)$].
In the bottom panel of Fig.~\ref{Fig:PhaseTransition} we plot, for different values of $|g|$, the function $\chi(\tau)$ where we defined $\chi=\corr{}{}/|g|$ and $\tau={\rm sgn}[\Delta]|c|/|g|$.
Once again, the behavior becomes universal for $|g|\ggg1$, with a sharp transition at positive detuning.
The critical-exponent analysis of the derivative is compatible with $\max[\partial\chi/\partial\tau]\propto|g|$ and $|\tau_{\rm max}-\tau_c|\propto1/|g|$ for $\tau_c\sim2.62$.
The divergent behavior of the derivative in the thermodynamic limit signals the first-order nature of this phase transition.
The latter has the same character of the one observed for $G=0$ and both occur in the regime for which the semiclassical solution predicts optical multistability.
Remarkably, in the case $F=0$ we can identify another phase transition, taking place for $\tau\simeq-1$.
Although $\chi$ stays continuous in the thermodynamic limit, its derivative, shown in the inset of Fig.~\ref{Fig:PhaseTransition}~(bottom), acquires a discontinuity.
This second-order phase transition has no counterpart in the driven-dissipative Kerr model without the two-photon processes.
It takes place around the semiclassical bifurcation point, i.e. when the Wigner function acquires a bi-modal character.
\section{Conclusions and perspectives}\label{sec:conclusion}
In this work, we have investigated the paradigmatic model of a driven-dissipative nonlinear resonator subject to both one- and two-photon processes.
We have shown that, remarkably, the steady-state of such system can be analytically obtained through the complex $P$-representation of its density matrix.
The exact solution, benchmarked against numerical techniques, stands as a powerful tool to investigate the physics of this general model.
We have discussed the limitations of the semiclassical approach in the high-density regime by comparing its prediction to the analytic results.
In this context, we pointed out the emergence of multi-modal Wigner functions, whose structure can not be fully determined semiclassically.
We have also shown that the multimodal character of $W(z)$ can be controlled by external parameters, such as the relative phase of the one- and two-photon pumps.
Furthermore, the exact solution allowed us to explore the physics of a first-order dissipative phase transition in the regime where the semiclassical approach predicts optical multistability.
Moreover, in the two-photon-driven Kerr model (i.e., for $F=0$) we also revealed a second-order phase transition.
The latter has no counterpart in the driven-dissipative Kerr model with $G=0$.
Both theoretical and experimental perspectives of the present work are numerous.
The one- and two-photon driven-dissipative resonator is already realizable with present techniques~\cite{LeghtasScience15}.
The exact solution allows us to predict how the external experimental parameters affects the steady state. Hence, one can generate and manipulate precisely coherent-like states or superpositions of them, which is of great interest towards quantum computation~\cite{GilchristJOB04,OurjoumtsevScience06,MirrahimiNJP14,GotoPRA16,PuriarXiv16}.
The exact results of this work, combined with mean-field~\cite{LeBoitePRL13,LeBoitePRA14} and renormalization techniques~\cite{SchollwockRMP05,FinazziPRL15}, pave the way to the study of exotic many-body phases of light in networks of nonlinear resonators.
Indeed, the flourishing field of reservoir and coupling engineering in circuit QED makes it possible to envision a plethora of combinations between one- and two-photon driving, dissipation, and hopping mechanisms~\cite{JinPRL13,JinPRA14}.
Moreover, effective two-photon processes can arise in the momentum-space Hamiltonian of systems that do not include two-photon mechanisms in real space.
For example, this is the case for a single-cavity polarization-dependent corss-Kerr model~\cite{ParaisoNatMat10,TakemuraNatPhys14} and for the driven-dissipative Bose-Hubbard model~\cite{GeraceNatPhys09,CarusottoPRL09,CarusottoRMP13,WilsonarXiv16}.
\section*{Acknowledgment}
We acknowledge support from the ERC via the Consolidator Grant “CORPHO”, No.~616233.
|
\section{Introduction}
\label{sec:intro}
In Quantum ChromeDynamics (QCD)
the chiral symmetry is one of the most important symmetry to investigate properties of hadron.
In particular, pion is identified as the pseudo Nambu-Goldstone (NG) boson corresponding to the spontaneously symmetry breaking of the chiral symmetry, which means that dynamics of pion is described by the low energy theorem.
On the other hand, there are the spin-1 mesons at about 1 GeV mass, as shown in FIG.\,\ref{Spectra}.
It is a variable problem to describe the spin-1 mesons and pion in a discussion of symmetries and their breaking.
\begin{figure}[h]
\centering \includegraphics[width=80mm]{spectrum-SU4-v4.eps}
\caption{Spectra of spin-1 mesons \cite{Agashe:2014kda}.}
\label{Spectra}
\end{figure}
Recently, the existence of an SU$(4)$ symmetry in the spin-1 meson sector is suggested by
Refs.\,\cite{Denissenya:2014ywa,Glozman:2015qva,Cohen:2015ekf} via the lattice QCD calculation. This is also shown in unbroken limit of the chiral symmetry in two-flavor case in Ref.\,\cite{Shifman:2016efc}.
In the references, the symmetry is called as an emergent symmetry in QCD.
The SU(4) symmetry includes the chiral symmetry SU(2)$_L\times$SU(2)$_R\times$U(1)$_A$, which
is corresponding to the rotation of
the quark field $\psi$ written as $\psi^T=\braa{u_R , d_R , u_L , d_L}$.
This means that the mesons denoted by $\braa{\rho, a_1, \rho', \omega', b_1, f_1, h_1}$ are members belonging to a multiplet of the SU(4) symmetry,
and that the mass differences of members are caused by the spontaneous chiral symmetry breaking.
This indicates that there exist extended Goldberger-Treiman\,(GT) relations between the mass differences of spin-1 mesons and their couplings to pions which are Nambu-Goldstone bosons associated with the chiral symmetry breaking.
In this work, we construct an effective Lagrangian with an $\mbox{SU(4)}\times\mbox{U(1)}$ hidden local symmetry\,(HLS)\,\cite{Bando:1984ej,Bando:1984pw,Bando:1985rf,Bando:1987br,Bando:1987ym,Harada:2003jx},
which includes
the spin-1 mesons, $\braa{\omega, \rho, a_1, \rho', \omega', b_1, f_1, h_1}$, as gauge fields of the HLS.
The symmetry of the Lagrangian is $\left[ \mbox{SU(2)}_R \times \mbox{SU(2)}_L \times \mbox{U(1)}_A \right]_{\rm chiral} \times \left[ \mbox{SU(4)} \times \mbox{U(1)} \right]_{\rm HLS}$, which is broken to $\mbox{SU(2)}_{\rm isospin}$ symmetry by the chiral condensate.
Then, we show extended GT relations, by which each coupling of the interaction among one pion and two spin-1 mesons is proportional to the mass difference of the relevant spin-1 mesons.
In addition, we can derive relations among the coupling of the spin-1 mesons to one pion thanks to the existence of the SU(4) symmetry.
Furthermore, we show that there is a relation among the mass of $\rho'$ meson, the $\rho'\pi\pi$ coupling and the $\rho'$-photon mixing strength as well as the Kawarabayashi-Suzuki-Riazuddin-Fayyazuddin relations for the $\rho$ meson.
These relations give us predictions for one-pion decay widths of spin-1 mesons and the electromagnetic form factor of the pion, which can be verified by future experiments.
In this paper, we conduct the analyses:
In Sec.\,\ref{sec:construction},
we construct a Lagrangian with the SU(4)$\times$U$(1)$ HLS to introduce the spin-1 mesons.
In Sec.\,\ref{sec:Vector mesons and pion}, we give eigenstates and masses of the spin-1 mesons.
In Sec.\,\ref{sec:GT}, we obtain extended GT relations. We study one-pion decays of the spin-1 mesons in section~\ref{sec:one pion decays} and extended KSRF relations in section~\ref{sec:KSRF}.
In Sec.\,\ref{sec: numerical analysis}, we make a numerical analysis to determine the parameters and give a prediction on the electromagnetic form factor of pion.
The summery and discussions are given in Sec.\,\ref{sec:summary}.
\section{Construction }
\label{sec:construction}
We construct a chiral Lagrangian with the ${{\rm SU(4)}\times {\rm U(1)}}$ hidden local symmetry.
The Lagrangian has
the chiral symmetry~\footnote{
Explicit breaking of $\mbox{U(1)}_A$ by anomaly is added later together with the explicit chiral symmetry breaking from the current quark masses of up and down quarks. }
$\mbox{SU(2)}_{R} \times \mbox{SU(2)}_L \times \mbox{U(1)}_A $
and
the gauge symmetry $G_{\rm local}=\brac{\mbox{SU(4)}\times \mbox{U(1)}}_{\rm HLS}$.
Here
spin-1 mesons are introduced as the gauge fields of $G_{\rm local}$,
which are identified as $\braa{\omega, \rho, a_1, \rho', \omega', b_1, f_1, h_1}$ mesons.
The spontaneously symmetry breaking is represented as
$G_{\rm global} \times G_{\rm local}( =\left[ \mbox{SU(2)}_R \times \mbox{SU(2)}_L \times \mbox{U(1)}_A \right]_{\rm chiral} \times \brac{\mbox{SU(4)}\times \mbox{U(1)}}_{\rm HLS} ) \ra H ( = \mbox{SU(2)}_{V} )$,
where the NG-bosons identified as the pions and the eta meson emerge.
The NG bosons associated with
the coset-space $G_{\rm global}/H$ are introduced
through the 2 by 2 unitary matrix field $U$
as
\begin{equation}
U = \exp \left( i \frac{\eta}{ f_\eta } + i \sum_{a=1}^3 \frac{ \pi^a \sigma^a }{ f_\pi } \right) \ ,
\end{equation}
where $\eta$ and $\pi^a$ ($a=1,2,3$) are the eta meson and the pion fields, and $\sigma^a$ is the Pauli matrix.
For introducing the $\mbox{SU(4)} \times \mbox{U(1)}$ HLS, we embed this $U$ into 4 by 4 matrix field ${\mathcal U}$ as
\bee{
\mc{U} =
\braa{\mt{
0&U^\dagger\\U &0
}}
}
which transforms under the chiral symmetry $\mbox{SU(2)}_L\times\mbox{SU(2)}_R\times\mbox{U(1)}_A$ as
\begin{equation}
{\mathcal U} \ \to \ {\mathcal G} \cdot {\mathcal U} \cdot {\mathcal G}^\dag \ ,
\end{equation}
where $\mc{G}$ is an element of $G_{\rm global}= \mbox{SU(2)}_L\times\mbox{SU(2)}_R\times\mbox{U(1)}_A$ written as
\bee{
\mc{G} =
\braa{\mt{g_R &0\\0&g_L }}
\braa{\mt{g_A&0\\0&g_A^\dagger}}
}
by using $g_{L,R} \in \mbox{SU(2)}_{L,R}$ and $g_{A} \in \mbox{U(1)}_{A}$.
The generators of the chiral symmetry $G_{\rm global}$ are
\bee{
T^A_{\rm global}=&
\brab{
\frac{S^a + X_{(3)}^a}{\sqrt{2}} \ ,~~
\frac{S^a - X_{(3)}^a}{\sqrt{2}} \ ,~~
X_{(3)}^0
}
\nn\\
=&
\brab{
T^a_{R}
\ ,~~
T^a_{L}
\ ,~~
X_{(3)}^0
}
\ ,
\label{generators of chiral}
}
whose
explicit form are given in Appendix~\ref{sec:Generators}.
After the spontaneous symmetry breaking of $G_{\rm global}$,
the generators expressed by $S$ are corresponding to the unbroken ones, while $X$ are broken generators.
This implies that transformations generated by $S$ belong to $H$.
Let us decompose $\mc{U}$ as
\bee{
\mc{U}
=&
\Xi^\dagger (x)
\cdot
\Xi^\dagger_m (x)
\cdot
\bar{\Sigma}
\cdot
\Xi_{m} (x)
\cdot
\Xi (x)
}
by using
\bee{
\bar{\Sigma}
\equiv
\braa{\mt{
0&1_2\\1_2&0
}}
\ .
}
These fields transform under $G_{\rm global} \times G_{\rm local}\times H_{\rm extra}$ as
\bee{
\Xi (x) \ra& \tilde{\mc{G}} (x) \cdot \Xi (x) \cdot \mc{G}^\dagger
\ ,
\nn\\
\Xi_{m} (x) \ra& \tilde{h} (x) \cdot \Xi_{m} (x) \cdot \tilde{\mc{G}}^\dagger (x)
}
where ${\mc{G}} \in G_{\rm global}$, $\tilde{\mc{G}} (x) \in G_{\rm local}$, and
$\tilde{h} (x) $ is an element of an $H_{\rm extra} \left( =\mbox{U(2)} \right)$ extra local symmetry,
whose generator is written as
\bee{
T^A_{\rm extra}
=&
\brab{
S^a , S^0 }
\ .
}
From the above transformation properties,
the covariant derivatives are expressed as
\begin{align}
D_\mu \Xi (x) = &
\partial_\mu \Xi(x) - i V_\mu \Xi(x) + i \Xi(x) {\mathcal V}_\mu
\notag\\
D_\mu \Xi_{m}(x) = & \partial_\mu \Xi_m(x) - i \tilde{V}_\mu \Xi_m(x) + i \Xi_m(x) V_\mu
\ .
\end{align}
where
$V_\mu$ is the HLS gauge field, $V_\mu = V_\mu^A T^A$, for $G_{\rm local}$,
$\mc{V}_\mu$ is the external gauge field written by
\bee{
\mc{V}_\mu = \mc{R}_\mu^a \cdot T_R^a +\mc{L}_\mu^a \cdot T_L^a
+ \sqrt{2} \mc{V}_\mu^0 S^0 +\sqrt{2} \mc{A}_\mu^0 X_{\perp (3)}^0
\ ,
\label{definition of external gauge}
}
and
$\tilde{V}_\mu$ is the gauge field for $H_{\rm extra}$.
We note that we do not introduce the kinetic term for this $\tilde{V}_\mu$, so that it is not a dynamical field in the present analysis.
The field strength of the HLS gauge field is written as
\bee{
V_{\mu\nu} \equiv \pd_\mu V_\nu - \pd_\nu V_\mu -i \brac{ V_\mu , V_\nu} \ .
}
For constructing the Lagrangian, it is convenient to introduce the following
covariantized Maurer-Cartan 1-forms:
\bee{
\hat{\alpha}_{\mu} (x) \equiv&
\frac{1}{i}
\Xi_{m} (x) \cdot
\braa{
D_\mu \Xi (x) \cdot \Xi^\dagger(x)
}
\cdot \Xi_{m}^\dagger (x)
\ ,
\nn\\
\hat{\alpha}_{\mu}^{(m)} (x) \equiv&
\frac{1}{i}
D_\mu \Xi_{m}(x) \cdot \Xi_{m}^\dagger (x)
\ ,
\label{definition of alpha}
}
which
transform under $G_{\rm global}\times G_{\rm local}\times H_{\rm extra}$, the parity $\mc{P}$, and the charge conjugation $\mc{C}$ as
in Table\,\ref{Table:transformation properties}.
\begin{table}
\caption[]{Transformation properties of the Maurer-Cartan 1-forms $\hat{\alpha}_\mu$ and $\hat{\alpha}^{(m)}_\mu$ under $G_{\rm global} \times G_{\rm local}$, $\mc{P}$, and $\mc{C}$ transformations.}
\begin{center}
\begin{tabular}{c|rl}\hline\hline
$G_{\rm global} \times G_{\rm local}$&
$\hat{\alpha}_{\mu} (x)
\ra$&$
\tilde{h} (x) \cdot
\hat{\alpha}_{\mu} (x)
\cdot
\tilde{h}^\dagger (x)
$\\
$\mc{P}$ trans.&
$
\hat{\alpha}_{\mu } (x)
\rax{\mc{P}}{}$&$
\bar{\Sigma} \cdot
\hat{\alpha}_{\mu } (x)
\cdot
\bar{\Sigma}
$\\
$\mc{C}$ trans.&
$
\hat{\alpha}_{\mu} (x)
\rax{\mc{C}}{}$&$
-\bar{\Sigma} \cdot
\braa{\hat{\alpha}_{\mu} (x)}^*
\cdot
\bar{\Sigma}
$\\\hline
$G_{\rm global} \times G_{\rm local}$&
$
\hat{\alpha}_{\mu}^{(m)} (x)
\ra$&$
\tilde{h} (x) \cdot
\hat{\alpha}_{\mu}^{(m)} (x)
\cdot
\tilde{h}^\dagger (x)
$\\
$\mc{P}$ trans.&
$
\hat{\alpha}_{\mu }^{(m)} (x)
\rax{\mc{P}}{}$&$
\bar{\Sigma} \cdot
\hat{\alpha}_{\mu }^{(m)} (x)
\cdot
\bar{\Sigma}
$\\
$\mc{C}$ trans.&
$
\hat{\alpha}_{\mu}^{(m)} (x)
\rax{\mc{C}}{}$&$
-\bar{\Sigma} \cdot
\braa{\hat{\alpha}_{\mu}^{(m)} (x)}^*
\cdot
\bar{\Sigma}
$
\\\hline\hline
\end{tabular}
\end{center}
\label{Table:transformation properties}
\end{table}
Now, the 1-form $\hat{\alpha}_\mu$ is
classified as
\bee{
\hat{\alpha}_{\mu \parallel } (x)
\equiv&
2\tr\brac{\hat{\alpha}_{\mu} \cdot S^a
}
S^a
\ ,
\nn\\
\hat{\alpha}_{\mu \perp (1)} (x)
\equiv&
2\tr\brac{\hat{\alpha}_{\mu} \cdot X_{(1)}^a
}
X_{(1)}^a
\ ,
\nn\\
\hat{\alpha}_{\mu \perp (2)} (x)
\equiv&
2\tr\brac{\hat{\alpha}_{\mu} \cdot X_{(2)}^a
}
X_{(2)}^a
\ ,
\nn\\
\hat{\alpha}_{\mu \perp (3)}(x)
\equiv&
2\tr\brac{\hat{\alpha}_{\mu} \cdot X_{(3)}^a
}
X_{(3)}^a
\ ,
\nn\\
\hat{\alpha}_{\mu \parallel }^{ (I=0)} (x)
\equiv&
2\tr\brac{\hat{\alpha}_{\mu} \cdot S^0
}
S^0
\ ,
\nn\\
\hat{\alpha}_{\mu \perp (1)}^{ (I=0)} (x)
\equiv&
2\tr\brac{\hat{\alpha}_{\mu} \cdot X_{(1)}^0
}
X_{(1)}^0
\ ,
\nn\\
\hat{\alpha}_{\mu \perp (2)}^{ (I=0)} (x)
\equiv&
2\tr\brac{\hat{\alpha}_{\mu} \cdot X_{(2)}^0
}
X_{(2)}^0
\ ,
\nn\\
\hat{\alpha}_{\mu \perp (3)}^{ (I=0)}(x)
\equiv&
2\tr\brac{\hat{\alpha}_{\mu} \cdot X_{(3)}^0
}
X_{(3)}^0
\ ,
}
and similarly for $\alpha_{\mu}^{(m)}$.
Note that the $\tilde{V}_\mu$ gauge field for $H_{\rm extra}$ is included only in $\hat{\alpha}_{\mu\parallel}^{(m)}$ and $\hat{\alpha}_{\mu\parallel}^{(m)(I=0)}$.
Furthermore, from
the definitions \eqref{definition of alpha}
of $\hat{\alpha}_\mu$ and $\hat{\alpha}^{(m)}_\mu$,
we find that the sum
\bee{
\hat{\alpha}_\mu (x)+\hat{\alpha}_\mu^{(m)} (x)
=&
\frac{1}{i}
D_\mu
\brac{\Xi_{m} (x) \cdot \Xi (x) }
\cdot
\brac{\Xi_{m} (x) \cdot \Xi (x) }^\dagger
\nn\\
\in&
\brab{X_{(3)}^A , S^A}
}
is expanded in terms of
$X_{(3)}^A$ and $S^A (=T_{\rm extra}^A)$ only because
the Maurer-Cartan 1-forms of $\Xi_{m} (x) \cdot \Xi (x)$ are constructed by the broken generators corresponding to $G_{\rm global}/H$ and $H_{\rm extra}$.
This relation yield
\bee{
\hat{\alpha}_{\mu \perp (1)}(x)+\hat{\alpha}_{\mu \perp (1)}^{(m)} (x) & = 0 \ ,
\notag\\
\hat{\alpha}_{\mu \perp (2)}(x)+\hat{\alpha}_{\mu \perp (2)}^{(m)} (x) & =0 \ ,
\notag\\
\hat{\alpha}_{\mu \perp (1)}^{(I=0)}(x)+\hat{\alpha}_{\mu \perp (1)}^{(m)(I=0)} (x) & = 0 \ ,
\notag\\
\hat{\alpha}_{\mu \perp (2)}^{(I=0)}(x)+\hat{\alpha}_{\mu \perp (2)}^{(m)(I=0)} (x) & =0 \ .
}
To take account of the effect of current quark masses,
we introduce an external source $\chi$ which transforms under the chiral symmetry $G_{\rm global}$ as
\bee{
\chi \ra& \mc{G}\cdot \chi \cdot \mc{G}^\dagger \ .
}
We assume that its expectation value is given as
\bee{
\brae{\chi}=m_\pi^2 ~\bar{\Sigma}
=\braa{\mt{0 & m_\pi^2 1_2\\ m_\pi^2 1_2&0}}
\ .
}
We redefine the external source as
\bee{
\hat{\chi}
\equiv
\Xi_{m} (x)\cdot
\Xi (x)
\cdot\chi\cdot
\Xi^\dagger (x)
\cdot
\Xi_{m}^\dagger (x)
}
such that it transforms
under $G_{\rm global}\times G_{\rm local}$, $\mc{P}$, and $\mc{C}$ as
\bee{
\hat{\chi} \ra&~
\tilde{h} (x)\cdot
\hat{\chi}\cdot
\tilde{h}^\dagger (x)
\ ,\nn\\
\hat{\chi} \rax{\mc{P}}{}&~
\bar{\Sigma}\cdot
\hat{\chi}\cdot \bar{\Sigma}
\ ,\nn\\
\hat{\chi} \rax{\mc{C}}{}&~
\bar{\Sigma}\cdot
\braa{\hat{\chi}}^*\cdot \bar{\Sigma}
\ ,
}
respectively.
By using a standard order counting manner for the fields:
\bee{
V_\mu
\sim
\tilde{V}_\mu
\sim
\mc{O} (p)
\ ,~~
\alpha_\mu
\sim
\mc{O} (p)
\ ,~~
\hat{\chi}
\sim
\mc{O} (p^2)
\ ,
}
possible operators invariant
under $G_{\rm global} \times G_{\rm local} (\times H_{\rm extra} )$ as well as $\mc{P}$ and $\mc{C}$ at $\mc{O}(p^2)$
which do not include $\hat{\alpha}_{\parallel}^{(m)} (x)$ are $\hat{\alpha}_{\parallel}^{(m)(I=0)} (x)$
are written as
\bee{
\mc{L}_{1}=&
F^2\tr\brac{\braa{\hat{\alpha}_{\mu \perp (3)} (x)+\hat{\alpha}_{\mu \perp (3)}^{(m)} (x)}^2}
\ ,
\nn\\
\mc{L}_{2}=&
F^2\tr\brac{\braa{\hat{\alpha}_{\mu \parallel} (x)}^2}
\ ,
\nn\\
\mc{L}_{3}=&
F^2\tr\brac{\braa{\hat{\alpha}_{\mu \perp (1)}^{(m)} (x)}^2}
\ ,
\nn\\
\mc{L}_{4}=&
F^2\tr\brac{\braa{\hat{\alpha}_{\mu \perp (2)}^{(m)} (x)}^2}
\ ,
\nn\\
\mc{L}_{5}=&
F^2\tr\brac{\braa{\hat{\alpha}_{\mu \perp (3)}^{(m)} (x)}^2}
\ ,
\nn\\
\mc{L}_{6}=&
2F^2\tr\brac{\hat{\alpha}_{\mu \parallel}^{(m)} (x) \cdot \hat{\alpha}^{{(m)} \mu}_{\perp (1)} (x)\cdot \bar{\Sigma}}
\ ,
\nn\\
\mc{L}_{7}=&
-2F^2\tr\brac{\braa{\hat{\alpha}_{\mu \perp (3)} (x)+\hat{\alpha}_{\mu \perp (3)}^{(m)} (x)} \cdot \hat{\alpha}^{{(m)} \mu}_{\perp (3)} (x)}
\ ,
\nn\\
\mc{L}_{8}=&
F^2\tr\brac{\braa{\hat{\alpha}_{\mu \perp (3)}^{(I=0)} (x)+\hat{\alpha}_{\mu \perp (3)}^{(m) (I=0)} (x)}^2}
\ ,
\nn\\
\mc{L}_{9}=&
F^2\tr\brac{\braa{\hat{\alpha}_{\mu \parallel}^{ (I=0)} (x)}^2}
\ ,
\nn\\
\mc{L}_{10}=&
F^2\tr\brac{\braa{\hat{\alpha}_{\mu \perp (1)}^{(m) (I=0)} (x)}^2}
\ ,
\nn\\
\mc{L}_{11}=&
F^2\tr\brac{\braa{\hat{\alpha}_{\mu \perp (2)}^{(m) (I=0)} (x)}^2}
\ ,
\nn\\
\mc{L}_{12}=&
F^2\tr\brac{\braa{\hat{\alpha}_{\mu \perp (3)}^{(m) (I=0)} (x)}^2}
\ ,
\nn\\
\mc{L}_{13}=&
2F^2\tr\brac{\hat{\alpha}_{\mu \parallel}^{ (I=0)} (x) \cdot \hat{\alpha}^{{(m)} (I=0) \mu}_{\perp (1)} (x)\cdot \bar{\Sigma}}
\ ,
\nn\\
\mc{L}_{14}=&
-2F^2\tr \Big[
\braa{\hat{\alpha}_{\mu \perp (3)}^{(I=0)} (x)+\hat{\alpha}_{\mu \perp (3)}^{(m) (I=0)} (x)}
\notag\\
& \qquad\qquad\qquad
\cdot \hat{\alpha}^{{(m)} (I=0) \mu}_{\perp (3)} (x) \Big]
\ ,
\nn\\
\mc{L}_{\chi}=&
F^2\tr\brac{\hat{\chi}\cdot \bar{\Sigma}}
\label{Lagrangian operators1}
\ ,
}
where $F$ is a constant with dimension one. Note that this $F$ is not the pion decay constant, which will be determined later.
In addition,
there are eight operators including $\hat{\alpha}_{\parallel}^{(m)} (x)$ or $\hat{\alpha}_{\parallel}^{{(m)}(I=0)} (x)$:
\bee{
{\mc{L}}'_{1}=&
F^2\tr\brac{\braa{\hat{\alpha}_{\mu \parallel}^{(m)} (x)}^2}
\ ,
\nn\\
{\mc{L}}'_{2}=&
2F^2\tr\brac{\hat{\alpha}_{\mu \parallel}^{(m)} (x) \cdot \hat{\alpha}_{ \parallel}^{\mu} (x)}
\ ,
\nn\\
{\mc{L}}'_{3}=&
2F^2\tr\brac{\hat{\alpha}_{\mu \parallel}^{(m)} (x) \cdot \hat{\alpha}^{ \mu}_{\perp (1)} (x)\cdot \bar{\Sigma}}
\ ,
\nn\\
{\mc{L}}'_{4}=&
2F^2\tr\brac{\hat{\alpha}_{\mu \parallel}^{(m)} (x) \cdot \hat{\alpha}^{{(m)} \mu}_{\perp (1)} (x)\cdot \bar{\Sigma}}
\ ,
\nn\\
{\mc{L}}'_{5}=&
F^2\tr\brac{\braa{\hat{\alpha}_{\mu \parallel}^{(m) (I=0)} (x)}^2}
\ ,
\nn\\
{\mc{L}}'_{6}=&
2F^2\tr\brac{\hat{\alpha}_{\mu \parallel}^{(m) (I=0)} (x) \cdot \hat{\alpha}_{ \parallel}^{(I=0)\mu} (x)}
\ ,
\nn\\
{\mc{L}}'_{7}=&
2F^2\tr\brac{\hat{\alpha}_{\mu \parallel}^{(m) (I=0)} (x) \cdot \hat{\alpha}^{ (I=0) \mu}_{\perp (1)} (x)\cdot \bar{\Sigma}}
\ ,
\nn\\
{\mc{L}}'_{8}=&
2F^2\tr\brac{\hat{\alpha}_{\mu \parallel}^{(m) (I=0)} (x) \cdot \hat{\alpha}^{{(m)} (I=0) \mu}_{\perp (1)} (x)\cdot \bar{\Sigma}}
\label{Lagrangian operators3}
\ .
}
Although the term given by
\bee{
F^2\tr\brac{\hat{\chi}}
}
is also an allowed operator, it contributes only to the vacuum energy.
The Lagrangian with the $\mbox{SU(4)}\times \mbox{U(1)}$ HLS at $\mc{O}(p^2)$ is written as
\bee{
\tilde{\mc{L}}_{\brac{\rm SU(4)\times U(1)}_{\rm HLS}}^{\mc{O} (p^2) }=&
\sum_{n=1}^{14} {a}_{(n)} {\mc{L}}_n
+
\sum_{n=1}^{8} {b}_{(n)} {{\mc{L}}'}_n
\nn\\
&
-
\frac{1}{2g^2}\tr\brac{V_{\mu\nu}V^{\mu\nu}}
\nn\\
&
-
\braa{\frac{1}{g^2_B} -\frac{1}{g^2}}\tr\brac{V_{\mu\nu}}\tr\brac{V^{\mu\nu}}
\nn\\
&
+
a_\chi\mc{L}_{\chi}
\label{Lagrangian full}
}
with arbitrary real coefficients ${a}_{(n)}$, ${b}_{(n)}$, and $a_\chi$.
$g$ and $g_B$ are the gauge coupling corresponding to
$\mbox{SU(4)}_{\rm HLS}$ and $\mbox{U(1)}_{\rm HLS}$, respectively\footnote{The values of the couplings are scaled by $\sqrt{2}$ comparing with usual way as defined in Ref.\,\cite{Harada:2003jx}.}.
In the above Lagrangian (\ref{Lagrangian full}),
$\Xi (x)$ and $\Xi_m (x)$ are parametrized as
\bee{
\Xi (x)\equiv &
e^{ -i p /F_p} \cdot e^{ i s /F_s } \cdot e^{i \pi /F_\pi }
= \Xi \braa{p}^\dagger \cdot\Xi \braa{s} \cdot \Xi \braa{\pi}
\ ,
\nn\\
\Xi_{m} (x)\equiv &e^{ -i \tilde{s} /F_{\tilde{s}}}\cdot e^{ i p /F_p}
=\Xi^\dagger \braa{\tilde{s}} \cdot \Xi \braa{p}
\label{parametrizing}
}
where the $p$, $s$, $\tilde{s}$, and $\pi$ are
\bee{
p (x) =&
p^A_{(1)} (x) \cdot X_{(1)}^A
+ p^A_{(2)} (x) \cdot X_{(2)}^A
+p^A_{(3)} (x)\cdot X_{(3)}^A
\ ,
\nn\\
s (x) =& s^A_s (x) \cdot S^A
\ ,~~~~
\tilde{s} (x) = \tilde{s}^A_s (x) \cdot S^A
\ ,~~~~
\nn\\
\pi (x)=&\pi^A (x)\cdot X^A_{(3)}
\ ,
}
respectively.
$\pi (x)$ is the NG-boson field corresponding to the breaking of the chiral symmetry, which is
identified with the pion.
$F_p$, $F_s$, $F_\pi$, and $F_{\tilde{s}}$ are constants with one mass-dimension, in particular $F_\pi$ is the pion decay constant.
$p(x)$, $s(x)$, and $\tilde{s}(x)$ are also the NG-boson fields which are eaten by the gauge fields.
Since $\tilde{V}_\mu$ included in $\hat{\alpha}_{\mu \parallel}^{(m)}$ is
not a dynamical field, we fix $\tilde{s}(x)=0$ and integrate out the gauge field.
Then $\mc{L}'_n$ become the terms given in Eqs.~\eqref{Lagrangian operators1}, and the Lagrangian is written by
\begin{equation}
{\mc{L}}_{\brac{\rm SU(4)\times U(1)}_{\rm HLS}}^{\mc{O} (p^2) }= {\mathcal L}_V + {\mathcal L}_k + a_\chi{\mathcal L}_{\chi} \ ,
\label{Lagrangian full2}
\end{equation}
where
\begin{align}
{\mathcal L}_V = & \sum_{n=1}^{14} \bar{a}_{(n)} {\mc{L}}_n
\nn\\
{\mathcal L}_k = &
- \frac{1}{2g^2}\tr\brac{V_{\mu\nu}V^{\mu\nu}}
\notag\\
& - \braa{\frac{1}{g^2_B} -\frac{1}{g^2}}\tr\brac{V_{\mu\nu}}\tr\brac{V^{\mu\nu}}
\ ,
\label{Lagrangian terms}
\end{align}
with the coefficients $\bar{a}_{(n)}$ being certain
linear combinations of $a_{(n)}$ and $b_{(n)}$.
To analyze the dynamics of the spin-1 mesons together with the pion in the model,
in the following analysis, we take the unitary gauge
\bee{
p=s=0
}
as well as $\tilde{s} =0$.
As shown in Appendix~\ref{sec:expansion}, the expanded form of the Maurer-Cartan 1-forms are written by using $\pi$ and $V_\mu$.
\section{Eigenstates and masses}
\label{sec:Vector mesons and pion}
In this section,
we obtain the mass eigenstates of the spin-1 mesons and their masses.
By using the generators of $\mbox{SU(4)}\times \mbox{U(1)}$ listed in Appendix~\ref{sec:Generators}, the
HLS gauge field is decomposed as
\bee{
V_{\mu}=&
V_{\mu \parallel}
+ V_{\mu \perp (1)}
+ V_{\mu \perp (2)}
+ V_{\mu \perp (3)}
\nn\\
&
+
V_{\mu \parallel}^{(I=0)}
+
V_{\mu \perp (1)}^{(I=0)}
+
V_{\mu \perp (2)}^{(I=0)}
+
V_{\mu \perp (3)}^{(I=0)}
\ ,
}
where
\bee{
V_{\mu \parallel}
\equiv&
2\tr\brac{V_{\mu} \cdot S^a}S^a
\equiv
V_{\mu \parallel}^a \cdot S^a
\ ,
\nn\\
V_{\mu \parallel}^{(I=0)}
\equiv&
2\tr\brac{V_{\mu} \cdot S^0}S^0
\equiv
V_{\mu \parallel}^0 \cdot S^0
\ ,
\nn\\
V_{\mu \perp (i)}
\equiv&
2\tr\brac{V_{\mu} \cdot X^a_{(i)}}X^a_{(i)}
\equiv
V_{\mu \perp}^a \cdot X^a_{(i)}
\ ,
\nn\\
V_{\mu \perp (i)}^{(I=0)}
\equiv&
2\tr\brac{V_{\mu} \cdot X^0_{(i)}}X^0_{(i)}
\equiv
V_{\mu \perp}^0 \cdot X^0_{(i)}
\ .
}
They are classified as $\braa{\rho, \omega, a_1, \rho', \omega', b_1, f_1, h_1}$
by the properties of transformation under $\mc{P}$ and $\mc{C}$.
Because the fields satisfy
\bee{
V_\mu \rax{\mc{P}}{}&
\bar{\Sigma}\cdot V^{\mu}\cdot \bar{\Sigma}
\nn\\
=&
V^{\mu}_{\parallel}
+
V^\mu_{\perp (1)}
-
V^\mu_{\perp (2)}
-
V^\mu_{\perp (3)}
\nn\\
&+
V_{\parallel}^{\mu (I=0)}
+
V_{\perp (1)}^{\mu (I=0)}
-
V_{\perp (2)}^{\mu (I=0)}
-
V_{\perp (3)}^{\mu (I=0)}
\ ,
\nn\\
V_\mu \rax{\mc{C}}{}&
-
\bar{\Sigma}\cdot \braa{V_{\mu}}^*\cdot \bar{\Sigma}
\nn\\
=&
-
\braa{V_{\mu \parallel}}^*
-
\braa{V_{\mu \perp (1)}}^*
-
\braa{V_{\mu \perp (2)}}^*
+
\braa{V_{\mu \perp (3)}}^*
\nn\\
&
-
\braa{V_{\mu \parallel}^{(I=0)}}^*
-
\braa{V_{\mu \perp (1)}^{(I=0)}}^*
-
\braa{V_{\mu \perp (2)}^{(I=0)}}^*
+
\braa{V_{\mu \perp (3)}^{(I=0)}}^*
\ ,
}
they are identified as the spin-1 mesons:
\bee{
&
\braa{V_{\mu \parallel} \ , V_{\mu \perp {(1)}}}
\Ra \rho
\ ,
\rho'
\ , ~~
\notag\\
&
\braa{V_{\mu \parallel}^{(I=0)} \ , V_{\mu \perp {(1)}}^{(I=0)}}
\Ra \omega
\ ,
\omega'
\ ,
\nn\\
&
V_{\mu \perp {(2)}} \Ra b_1
\ , ~~
V_{\mu \perp {(2)}}^{(I=0)} \Ra h_1
\ , ~~
\notag\\
&
V_{\mu \perp {(3)}} \Ra a_1
\ , ~~
V_{\mu \perp {(3)}}^{(I=0)} \Ra f_1
\label{def of the vector meson}
\ .
}
\begin{widetext}
From ${\mc{L}}_{V}+ a_\chi{\mathcal L}_{\chi} $, the quadratic terms with respect to the fields are given as
\bee{
{\mc{L}}_{V}+ a_\chi{\mathcal L}_{\chi}
=&
\frac{1}{2}\frac{F^2}{F_\pi^2}
\braa{\bar{a}_{(1)}-\frac{\bar{a}_{(7)}^2}{\bar{a}_{(5)} }}\braa{\pd_\mu \pi^a}^2
+
\frac{1}{2}\braa{a_{\chi}\frac{F^2}{F_\pi^2}} m_\pi^2 \braa{\pi^a}^2
+
\frac{1}{2}\frac{F^2}{F_\pi^2}
\braa{\bar{a}_{(8)}-\frac{\bar{a}_{(14)}^2}{\bar{a}_{(12)} }} \braa{\pd_\mu \eta}^2
+
\frac{1}{2}\braa{a_{\chi}\frac{F^2}{F_\pi^2}} m_\pi^2 \braa{\eta}^2
\nn\\
&+
\sqrt{2}\frac{F^2}{F_\pi}
\braa{\bar{a}_{(1)}-\frac{\bar{a}_{(7)}^2}{\bar{a}_{(5)} }}
\braa{ \mc{A}_\mu^a \pd^\mu \pi^a}
+
\sqrt{2}\frac{F^2}{F_\pi}
\braa{\bar{a}_{(8)}-\frac{\bar{a}_{(14)}^2}{\bar{a}_{(12)} }}
\braa{ \mc{A}_\mu^0 \pd^\mu \eta}
\nn\\
&+
\frac{1}{2}
\braa{\bar{a}_{(4)}F^2g^2}
\frac{1}{g^2}\braa{V_{\mu \perp{(2)}}^a }^2
+
\frac{1}{2}
\braa{\bar{a}_{(11)}F^2g^2}
\frac{1}{g^2}\braa{V_{\mu \perp{(2)}}^0}^2
\nn\\
&
+
\frac{1}{2}
\braa{\bar{a}_{(5)}F^2g^2}
\frac{1}{g^2}\braa{V_{\mu \perp{(3)}}^a -\frac{\bar{a}_{(7)}}{\bar{a}_{(5)}} \frac{1}{ F_\pi}\pd_\mu \pi^a}^2
+
\frac{1}{2}
\braa{\bar{a}_{(12)}F^2g^2}
\frac{1}{g^2}\braa{V_{\mu \perp{(3)}}^0 - \frac{\bar{a}_{(14)}}{\bar{a}_{(12)}}\frac{1}{ F_\pi}\pd_\mu \eta}^2
\nn\\
&
+
g^2F^2
\tr\brac{
\braa{\mt{
\frac{1}{g}V_{\mu \parallel} &
\frac{1}{g}\bar{\Sigma}\cdot V_{\mu \perp{(1)}}
}}
\braa{\mt{
\bar{a}_{(2)}&-\bar{a}_{(6)}
\\
-\bar{a}_{(6)}&\bar{a}_{(3)}
}}
\braa{\mt{
\frac{1}{g}V_{\mu \parallel} \\
\frac{1}{g}\bar{\Sigma}\cdot V_{\mu \perp{(1)}}
}}
}
\nn\\
&
+
g^2F^2
\tr\brac{
\braa{\mt{
\frac{1}{g_B}V_{\mu \parallel}^{(I=0)} &
\frac{1}{g}\bar{\Sigma}\cdot V_{\mu \perp{(1)}}^{(I=0)}
}}
\braa{\mt{
\frac{g_B^2}{g^2}\bar{a}_{(9)}&-\frac{g_B}{g}\bar{a}_{(13)}
\\
-\frac{g_B}{g}\bar{a}_{(13)}&\bar{a}_{(10)}
}}
\braa{\mt{
\frac{1}{g_B}V_{\mu \parallel}^{(I=0)} \\
\frac{1}{g}\bar{\Sigma}\cdot V_{\mu \perp{(1)}}^{(I=0)}
}}
}
+
\cdots
\label{L_V and L_chi}
}
where the axial external gauge
field is defined as
$\mc{A}^a_\mu =\frac{1}{2}\braa{\mc{R}_\mu^a - \mc{L}_\mu^a}$.
Note that the field $\eta$ defined by $\eta \equiv 2\tr\brac{\pi \cdot X^0_{(3)}}$ is the linear combination of the lowest eta meson and $\eta'(958)$.
To normalize the kinetic terms of $\pi^a$ and $\eta$, we set
\bee{
F^2\braa{\bar{a}_{(1)}-\frac{\bar{a}_{(7)}^2}{\bar{a}_{(5)} }}=&F_\pi^2
\ ,~~~
F^2\braa{\bar{a}_{(8)}-\frac{\bar{a}_{(14)}^2}{\bar{a}_{(12)} }}=F_\pi^2
\ ,
\label{normalization of pion}
}
together with $a_{\chi}F^2 =F_\pi^2$ which makes the pion mass be $m_\pi$.
The second line in Eq.\,\eqref{L_V and L_chi} implies that the physical pion decay constant is defined as
\bee{
f_\pi \equiv \sqrt{2}F_\pi
\ ,
}
whose value is given in Table\,\ref{Table:Experimental values}.
Furthermore, the mass eigenstates of $a_1$ and $f_1$ are defined by
\bee{
\braa{a_1}_\mu \equiv& \frac{1}{g} \braa{V_{\mu \perp{(3)}}^a - \frac{r_{a_1}}{ f_\pi}\pd_\mu \pi^a}X_{(3)}^a
\ ,
~~
\braa{f_1}_\mu \equiv \frac{1}{g} \braa{V_{\mu \perp{(3)}}^0 - \frac{r_{f_1}}{ f_\pi}\pd_\mu \eta}X_{(3)}^0
\label{definition: a1 f1}
\ ,
}
where
\bee{
r_{a_1}\equiv&
\sqrt{2}\frac{\bar{a}_{(7)}}{\bar{a}_{(5)}}
\ ,~~~~
r_{f_1}\equiv
\sqrt{2}\frac{\bar{a}_{(14)}}{\bar{a}_{(12)}}
\ .
}
$r_{a_1}$ ($r_{f_1}$) expresses the mixing rate between the $a_1$ ($f_1$) meson and the pion $\pi$ ($\eta$).
Their masses are obtained as
\bee{
m_{a_1}^2 =
\bar{a}_{(5)}g^2F^2
\ ,~~~
m_{f_1}^2 =
\bar{a}_{(12)}g^2F^2
\ .
\label{masses af}
}
The mixing of Eq.\,\eqref{definition: a1 f1}
implies that the currents corresponding to the generators $X_{(3)}^a$ and $X_{(3)}^0$ of the \mbox{SU(4)} HLS are coupled to the axial current of the chiral symmetry with the factors of $r_{a_1}$ and $r_{f_1}$, respectively.
\end{widetext}
The physical states and masses for $b_1$ and $h_1$ are defined as
\bee{
\braa{b_1}_\mu \equiv& \frac{1}{g}V_{\mu \perp {(2)}}
~~~~
m_{b_1}^2=\bar{a}_{(4)}g^2F^2
\ , \nn\\
\braa{h_1}_\mu \equiv& \frac{1}{g}V_{\mu \perp {(2)}}^{(I=0)}
~~~~
m_{h_1}^2=\bar{a}_{(11)}g^2F^2
\ ,
\label{masses bh}
}
respectively.
By diagonalizing the mass matrices from Eq.\,\eqref{L_V and L_chi}, the eigenstates for the vector mesons are expressed as
\begin{align}
\rho_\mu &= \rho_\mu^a S^a \equiv \frac{1}{g}\braa{\cos \theta_\rho V_{\mu \parallel} -\sin \theta_\rho \bar{\Sigma}\cdot V_{\mu \perp{(1)}}}
\ , \notag\\
\braa{\rho'}_\mu &= \braa{\rho'}_\mu^a S^a\equiv \frac{1}{g}\braa{\cos \theta_\rho \bar{\Sigma}\cdot V_{\mu \perp{(1)}} + \sin \theta_\rho V_{\mu \parallel}}
\ , \notag\\
\hat{\omega}_\mu &= \omega_\mu S^0 \equiv \frac{1}{g_B}\cos \theta_\omega V_{\mu \parallel}^{(I=0)} -\frac{1}{g}\sin \theta_\omega \bar{\Sigma}\cdot V_{\mu \perp{(1)}}^{(I=0)}
\ , \notag\\
\hat{\braa{\omega'}}_\mu &= \braa{\omega'}_\mu S^0 \equiv \frac{1}{g}\cos \theta_\omega \bar{\Sigma}\cdot V_{\mu \perp{(1)}}^{(I=0)} + \frac{1}{g_B}\sin \theta_\omega V_{\mu \parallel}^{(I=0)}
\ .
\end{align}
The masses of these states are obtained as
\bee{
m_{\rho , \rho'}^2
\equiv&
\frac{1}{2}\braa{\bar{a}_{(2)}+\bar{a}_{(3)}
\mp \sqrt{\braa{\bar{a}_{(2)} - \bar{a}_{(3)}}^2+4\bar{a}_{(6)}^2}}g^2F^2
\ ,\nn\\
m_{\omega , \omega'}^2
\equiv&
\frac{1}{2}\braa{\frac{g_B^2}{g^2}\bar{a}_{(9)}+\bar{a}_{(10)}
\right.
\nn\\&
\left.\mp\sqrt{\braa{\frac{g_B^2}{g^2}\bar{a}_{(9)}-\bar{a}_{(10)}}^2+4\frac{g_B^2}{g^2}\bar{a}_{(13)}^2}}g^2F^2
\ ,
\label{masses rho omega}
}
where
the mixing angles $\theta_\rho$ and $\theta_\omega$ are determined as
\bee{
\tan 2\theta_\rho \equiv&
\frac{2\bar{a}_{(6)} }{\bar{a}_{(2)} - \bar{a}_{(3)}}
\ ,~~~
\tan 2\theta_\omega \equiv
\frac{2g_B g\bar{a}_{(13)} }{g^2_B\bar{a}_{(9)}-g^2\bar{a}_{(10)}}
\ .
}
We can fix the values of ten parameters
from the physical values of eight spin-1 mesons listed in Table~\ref{Table:masses}
using Eqs.~(\ref{masses af}), (\ref{masses bh}) and (\ref{masses rho omega}) together with
two conditions given in Eq.~\eqref{normalization of pion}.
The model still has six free parameters:
\bee{
g
\ ,~~
g_B
\ ,~~
r_{a_1}
\ ,~~
r_{f_1}
\ ,~~
\cos \theta_\rho
\ ,~~
\cos \theta_\omega
\ ,
}
which relates several interactions among the spin-1 mesons.
In this paper, since we do not treat decays of the eta meson,
the parameter $r_{f_1}$ is irrelevant.
So,
we will determine the values of five parameters except for $r_{f_1}$.
\begin{table}
\caption[]{Masses of the relevant spin-1 mesons in PDG~\cite{Agashe:2014kda}}
\begin{center}
\begin{tabular}{c|cc}\hline\hline
Mesons &$I\braa{J^{PC}}$&mass (MeV)
\\\hline
$\rho$&$1\braa{1^{--}}$&$775.26\pm0.25$
\\
$\omega$&$0\braa{1^{--}}$&$782.65\pm0.12$
\\
$\rho'$&$1\braa{1^{--}}$&$1465\pm 25$
\\
$\omega'$&$0\braa{1^{--}}$&$1400$ -- $1450$
\\
$a_1$&$1\braa{1^{++}}$&$1230\pm 40$
\\
$f_1$&$0\braa{1^{++}}$&$1281.9 \pm 0.5$
\\
$b_1$&$1\braa{1^{+-}}$&$1229.5\pm3.2$
\\
$h_1$&$0\braa{1^{+-}}$&$1170\pm20$
\\\hline\hline
\end{tabular}
\end{center}
\label{Table:masses}
\end{table}
\begin{table}
\caption[]{Experimental values from PDG~\cite{Agashe:2014kda}.}
\begin{center}
\begin{tabular}{c|c}\hline\hline
$\Gamma \braa{\rho \ra \pi\pi}
$&$
147.8 \pm 0.9 \,{\rm MeV}
$
\\
$\Gamma \braa{\rho^0 \ra e^+e^-}
$&$
7.04 \pm 0.06 \,{\rm keV}
$
\\
$\Gamma(\omega \to e^+ e^-)$ & $0.60 \pm 0.02\,\mbox{keV}$
\\
$m_{\pi^\pm}
$&$
139.57018 \pm 0.00035 \,{\rm MeV}
$\\
$m_e
$&$
548.57990946 \pm 0.00000022 \,{\rm keV}
$\\
$\alpha\equiv\frac{e^2}{4\pi}
$&$
\frac{1}{137}
$\\
$f_\pi
$&$
92.21 \pm 0.14\,{\rm MeV}
$
\\
$\brae{r^2}^{\pi^\pm}_V
$&$
0.452 \pm 0.011
\,{\rm fm^2}
$
\\\hline\hline
\end{tabular}
\end{center}
\label{Table:Experimental values}
\end{table}
In the following, we summarize the extended GT relations, the
relations among one-pion decays of spin-1 mesons
and the extended KSRF relations in the separated sections.
To obtain some predictions analytically and numerically from them, we use experimental values in Table\,\ref{Table:Experimental values}.
\section{Extended Goldberger-Treiman relation}
\label{sec:GT}
In this section, we investigate an extended Goldberger-Treiman relation for one-pion interactions of two different spin-1 mesons.
First, we give a general discussion for the extended GT relation.
Next, we derive several relations in the HLS model.
Let us start a general
discussion in the case that the final spin-1 meson state has the different parity from the initial state.
By requiring
Lorentz covariance and parity invariance, the
amplitudes of two spin-1 states coupled with the axial current $j_5^\alpha$
are written as\,\footnote
Indices for isospin are omitted. Note that terms including
the antisymmetric tenser $\epsilon^{\mu\nu\alpha\beta }$ are also allowed if
the initial and final states have the same parity.
In the HLS model,
these contributions are obtained from the intrinsic parity odd terms,
which are listed in Appendix\,\ref{sec:Intrinsic parity odd terms}.
}
\bee{
\mc{M}^\alpha
=&
\int d^4x e^{-iqx}\bra{V_\mu (p_2)} j^\alpha_5 (x) \ket{V_\nu (p_1)}
\nn\\
=&
\epsilon^*_\mu (p_2)
\brac{
g_1 (q^2 )
g^{\mu\nu}
i p^\alpha
+
g_2(q^2 )
\frac{q^{\mu}q^{\nu}}{m^2_1 + m_2^2}
i p^\alpha
\right.\nn\\&
~~~~~~~
+
g_3(q^2 )
\braa{
iq^{\mu}g^{\nu \alpha}
+
iq^{\nu}g^{\mu \alpha}
}
\nn\\&
~~~~~~~
+
g_4(q^2 )
\braa{
i q^{\mu}g^{\nu \alpha}
-
i q^{\nu}g^{\mu \alpha}
}
\nn\\&\left.
~~~~~~
+
h_1 (q^2 )
g^{\mu\nu}i q^\alpha
+
h_2 (q^2 )
\frac{q^{\mu}q^{\nu}}{m^2_1 + m_2^2} i q^\alpha
}
\epsilon_\nu (p_1)
\label{axial form factor}
}
where $p= \frac{p_1+p_2}{2}$, $q= p_1-p_2$, and $g_i(q^2)$ ($i=1,2,3,4$) and $h_j(q^2)$ ($j=1,2$) are
independent form factors, which are generally complex functions of $q^2$.
Since the axial vector current is conserved in the chiral limit, we have the Ward-Takahashi identity as
$q_\alpha \mc{M}^\alpha = 0 $,
which leads to
\bee{
&
g_1 (q^2 ) \braa{p\cdot q}
\braa{\epsilon^* \cdot \epsilon}
+
g_2 (q^2 )
\braa{p\cdot q}
\frac{\braa{\epsilon^* \cdot q}
\braa{q \cdot \epsilon}}{m_1^2+m_2^2}
\nn\\&
+
2g_3 (q^2 )
\braa{\epsilon^* \cdot q}
\braa{q \cdot \epsilon}
\nn\\&
+
h_1 (q^2 ) q^2
\braa{\epsilon^* \cdot \epsilon}
+
h_2 (q^2 ) q^2
\frac{\braa{\epsilon^* \cdot q}
\braa{q \cdot \epsilon}}{m_1^2+m_2^2}
=0
\ .
\label{WT identity}
}
The form factors
$h_1 (q^2)$ and $h_2 (q^2)$ include a massless pole of the pion contribution:
\bee{
h_{n} (q^2) = \frac{f_\pi}{q^2} G_{V_1V_2\pi}^{(n)} + \cdots
\label{hn}
}
with $n=1,2$.
In the soft pion limit $q^2 \ra 0$,
the left hand side of Eq.~(\ref{WT identity}) is reduced to
\begin{align}
&
\mbox{(LHS of Eq.~(\ref{WT identity}))} \notag\\
& \quad = \bigg[ g_1 (0) \frac{m_1^2-m_2^2}{2}
+ f_\pi \, G_{V_1V_2 \pi}^{(1)} \bigg] \left( \epsilon^\ast \cdot \epsilon \right) \ ,
\end{align}
where we used
$p\cdot q=\frac{m_1^2 -m_2^2}{2}$. From this
one can obtain
\bee{
G_{V_1 V_2 \pi}^{(1)}
=&
-
\frac{m_1^2- m_2^2}{2f_\pi}
g_1 (0 )
\ .
\label{Extended GT relation from general discussion}
}
This is an extended Goldberger-Treiman relation among a mass difference of two spin-1 mesons, their coupling to one pion and
the axial form factor.
It should be noted that, if
the mass splitting of the initial and final states were large, the soft pion limit would not be reasonable.
The existence of the emergent symmetry in QCD implies that
the mass difference of the spin-1 mesons $\braa{\rho, a_1, \rho', \omega', b_1, f_1, h_1}$ comes from the breaking of the chiral symmetry.
Thus,
the emergent symmetry together with the chiral symmetry ensures low energy theorems
for the members of a multiplet of the symmetry.
Next, we turn to make an analysis based on the present model.
In the $\mc{L}_V$ part of the Lagrangian\,(\ref{Lagrangian full2}), there are no interactions among two HLS gauge fields and one pion field. However, due to the existence of the
$a_1$-$\pi$ and $f_1$-$\eta$ mixings as shown in Eq.\,\eqref{definition: a1 f1}, the HLS gauge field $V_\mu$ include the fields for
the physical pion in addition to the physical
spin-1 mesons.
Then,
the interactions among two spin-1 mesons and one pion are generated from
\bee{
\mc{L}_{\rm int}^{(3)}
=&
-\frac{1}{ig^2}
\tr\brac{\braa{\pd_\mu V_\nu - \pd_\nu V_\mu} \brac{V^\mu, V^\nu}}
}
included in $\mc{L}_k$ of Eq.\,\eqref{Lagrangian terms}.
As a result, all the interactions among two spin-1 mesons and one pion are proportional to the ratio $r_{a_1}/f_\pi$.
The explicit forms of the effective vertices are written as
\bee{
&\Gamma^{\mu\nu} \brac{\braa{V_1}_\mu^a (p_1), \braa{V_2}_\nu^b (p_2) , \pi^c }
\nn\\
=&
g_{V_1 V_2 \pi}\,
\epsilon^{abc}
\braa{p_1^2 P^{\mu\nu}(p_1)-p_2^2 P^{\mu\nu}(p_2)}
\ ,\nn\\
&\Gamma^{\mu\nu} \brac{\braa{V_1}_\mu^a (p_1), \braa{V_2^{(I=0)}}_\nu (p_2) , \pi^b }
\nn\\
=&
g_{V_1 V_2 \pi}\,
\delta^{ab}
\braa{p_1^2 P^{\mu\nu}(p_1)-p_2^2 P^{\mu\nu}(p_2)}
\ ,
\label{effective vertex for VVpi}
}
where the projection operator is defined as
\bee{
P^{\mu\nu} (p) \equiv g^{\mu\nu} -\frac{p^\mu p^\nu}{p^2}
\ ,
}
and $g_{V_1 V_2 \pi}$ expresses the corresponding coupling:
\bee{
g_{\rho a_1 \pi}
=&
g_{\rho' h_1 \pi}
=
\frac{r_{a_1}}{\sqrt{2}f_\pi}\cos \theta_\rho
\ ,\nn\\
g_{\rho' a_1 \pi}
=&
-g_{\rho h_1 \pi}
=
\frac{r_{a_1}}{\sqrt{2}f_\pi}\sin \theta_\rho
\ ,\nn\\
g_{b_1 \omega \pi}
=&
-\frac{r_{a_1}}{\sqrt{2}f_\pi}\sin \theta_\omega
\ ,~~
g_{b_1 \omega' \pi}
=
\frac{r_{a_1}}{\sqrt{2}f_\pi}\cos \theta_\omega
\ .
\label{GT relation}
}
As shown in Appendix\,\ref{sec:expansion}, interactions among three spin-1 mesons including $a_1$
are also obtained from $\mc{L}_{\rm int}^{(3)}$.
The direct coupling of two spin-1 mesons with the axial external gauge field does not exist at the leading order of the present model, and only two diagrams shown in FIG.\,\ref{figure: couple to axial current}
contribute to the coupling to the axial vector current. The pion in Fig.~\ref{figure: couple to axial current}(a) contributes to only $h_1$, and the $a_1$ meson in Fig.~\ref{figure: couple to axial current}(b) contributes to $h_1$, $g_1$ and $g_4$. We summarize their contributions in Table~\ref{Table:axial form factor}. Substituting these contributions into Eq.~(\ref{WT identity}) we can easily verify that the Ward-Takahashi identity is actually satisfied for any $q^2$.
We next
consider the soft-pion limit, $q^2 \to 0$. As expected in the general consideration given above, the pion contribution dominates over the $a_1$ meson contribution in $h_1$. As a result, $h_1(q^2)$ is expressed as in Eq.~(\ref{hn}), where $G_{V_1V_2\pi}^{(1)}$ is listed in the first column of Table~\ref{Table:coupling VVpi}. On the other hand, $g_1(0)$ is determined by taking $q^2 = 0$ limit of the $a_1$ meson contribution, which is listed in the second column of Table~\ref{Table:coupling VVpi}.
Since the coupling $g_{a_1}$ in the second column is given by $g_{a_1}= -\frac{r_{a_1} m_{a_1}^2}{g}$ as shown in Appendix\,\ref{sec:expansion},
we can easily confirm that these actually satisfy the extended GT relation in Eq.~\eqref{Extended GT relation from general discussion}.
\begin{figure*}
(a)
\includegraphics[width=40mm]{VVpi3.eps}
~~(b)
\includegraphics[width=40mm]{VVa1v2.eps}
\caption[]{Diagrams contributing to the amplitude given in Eq.\,\eqref{axial form factor}.}
\label{figure: couple to axial current}
\end{figure*}
\begin{table*}
\caption[]{Axial form factors given in the SU(4) HLS model. The function $D^{a_1} (q^2)$
is defined as $ \displaystyle D^{a_1} (q^2) \equiv \frac{m_{a_1}^2}{m_{a_1}^2-q^2}$.
We also find that the other form factors equal to zero at the $\mc{O}(p^2)$ order: $h_2(q^2)=g_2(q^2)=g_3(q^2)=0$.
}
\begin{center}
\begin{tabular}{c|c|c|c}\hline\hline
&$h_1 (q^2)$&$g_1 (q^2)$&$g_4 (q^2)$
\\\hline
$a_1 \ra \rho$&$ \displaystyle
\braa{m_{a_1}^2-m_{\rho}^2}\braa{\frac{r_{a_1}}{\sqrt{2}f_\pi}\cos \theta_\rho}\frac{ f_\pi}{q^2}
-\frac{g}{\sqrt{2}}\cos\theta_\rho \frac{g_{a_1}}{m_{a_1}^2}
\frac{m_{a_1}^2-m_{\rho}^2}{m_{a_1}^2}
D^{a_1} (q^2)
$&$ \displaystyle
\sqrt{2} g\cos\theta_\rho
\frac{g_{a_1}}{m_{a_1}^2}
D^{a_1} (q^2)
$&$ \displaystyle
\sqrt{2} g\cos\theta_\rho \frac{g_{a_1}}{m_{a_1}^2}
D^{a_1} (q^2)
$
\\
$\rho' \ra a_1$&$ \displaystyle
\braa{m_{\rho'}^2-m_{a_1}^2}\braa{\frac{r_{a_1}}{\sqrt{2}f_\pi}\sin \theta_\rho}\frac{ f_\pi}{q^2}
-\frac{g}{\sqrt{2}}\sin\theta_\rho \frac{g_{a_1}}{m_{a_1}^2}
\frac{m_{\rho'}^2-m_{a_1}^2}{m_{a_1}^2}
D^{a_1} (q^2)
$&$ \displaystyle
\sqrt{2} g\sin\theta_\rho \frac{g_{a_1}}{m_{a_1}^2}
D^{a_1} (q^2)
$&$ \displaystyle
0
$
\\
$h_1 \ra \rho$
&$ \displaystyle
-
\braa{m_{h_1}^2-m_{\rho}^2}\braa{-\frac{r_{a_1}}{\sqrt{2}f_\pi}\sin \theta_\rho}\frac{ f_\pi}{q^2}
-\frac{g}{\sqrt{2}}\sin\theta_\rho \frac{g_{a_1}}{m_{a_1}^2}
\frac{m_{h_1}^2-m_{\rho}^2}{m_{a_1}^2}
D^{a_1} (q^2)
$&$ \displaystyle
\sqrt{2} g\sin\theta_\rho \frac{g_{a_1}}{m_{a_1}^2}
D^{a_1} (q^2)
$&$ \displaystyle
-\frac{ g}{\sqrt{2}}\sin\theta_\rho \frac{g_{a_1}}{m_{a_1}^2}
D^{a_1} (q^2)
$
\\
$\rho' \ra h_1$&$ \displaystyle
\braa{m_{\rho'}^2-m_{h_1}^2}\braa{\frac{r_{a_1}}{\sqrt{2}f_\pi}\cos \theta_\rho}\frac{ f_\pi}{q^2}
-\frac{g}{\sqrt{2}}\cos\theta_\rho \frac{g_{a_1}}{m_{a_1}^2}
\frac{m_{\rho'}^2-m_{h_1}^2}{m_{a_1}^2}
D^{a_1} (q^2)
$&$ \displaystyle
\sqrt{2} g\cos\theta_\rho \frac{g_{a_1}}{m_{a_1}^2}
D^{a_1} (q^2)
$&$ \displaystyle
-\frac{ g}{\sqrt{2}}\cos\theta_\rho \frac{g_{a_1}}{m_{a_1}^2}
D^{a_1} (q^2)
$
\\
$b_1 \ra \omega$&$ \displaystyle
-\braa{m_{b_1}^2-m_{\omega}^2}\braa{-\frac{r_{a_1}}{\sqrt{2}f_\pi}\sin \theta_\omega}\frac{ f_\pi}{q^2}
-\frac{g}{\sqrt{2}}\sin\theta_\omega \frac{g_{a_1}}{m_{a_1}^2}
\frac{m_{b_1}^2-m_{\omega}^2}{m_{a_1}^2}
D^{a_1} (q^2)
$&$ \displaystyle
\sqrt{2} g\sin\theta_\omega \frac{g_{a_1}}{m_{a_1}^2}
D^{a_1} (q^2)
$&$ \displaystyle
-\frac{g}{\sqrt{2}}\sin\theta_\omega \frac{g_{a_1}}{m_{a_1}^2}
D^{a_1} (q^2)
$
\\
$\omega' \ra b_1$&$ \displaystyle
\braa{m_{\omega'}^2-m_{b_1}^2}\braa{\frac{r_{a_1}}{\sqrt{2}f_\pi}\cos \theta_\omega}\frac{ f_\pi}{q^2}
-\frac{g}{\sqrt{2}}\cos\theta_\omega \frac{g_{a_1}}{m_{a_1}^2}
\frac{m_{\omega'}^2-m_{b_1}^2}{m_{a_1}^2}
D^{a_1} (q^2)
$&$ \displaystyle
\sqrt{2} g\cos\theta_\omega \frac{g_{a_1}}{m_{a_1}^2}
D^{a_1} (q^2)
$&$ \displaystyle
-\frac{g}{\sqrt{2}}\cos\theta_\omega \frac{g_{a_1}}{m_{a_1}^2}
D^{a_1} (q^2)
$
\\\hline\hline
\end{tabular}
\end{center}
\label{Table:axial form factor}
\end{table*}
\begin{table}
\caption[]{ One pion and axial couplings in the SU(4) HLS model.
}
\begin{center}
\begin{tabular}{cc|cccccc}\hline\hline
$V_1$&$V_2$&$G_{V_1 V_2 \pi}^{(1)}$&$g_1(0)$
\\\hline
$a_1$&$ \rho$&
$ \displaystyle
\braa{m_{a_1}^2-m_{\rho}^2}\braa{\frac{r_{a_1}}{\sqrt{2}f_\pi}\cos \theta_\rho}
$&
$ \displaystyle
\sqrt{2} g\cos\theta_\rho \frac{g_{a_1}}{m_{a_1}^2}
$
\\
$\rho' $&$ a_1$&$ \displaystyle
\braa{m_{\rho'}^2-m_{a_1}^2}\braa{\frac{r_{a_1}}{\sqrt{2}f_\pi}\sin \theta_\rho}
$&$ \displaystyle
\sqrt{2} g\sin\theta_\rho \frac{g_{a_1}}{m_{a_1}^2}
$
\\
$h_1 $&$ \rho$
&$ \displaystyle
\braa{m_{h_1}^2-m_{\rho}^2}\braa{\frac{r_{a_1}}{\sqrt{2}f_\pi}\sin \theta_\rho}
$&$ \displaystyle
\sqrt{2} g\sin\theta_\rho \frac{g_{a_1}}{m_{a_1}^2}
$
\\
$\rho' $&$ h_1$&$ \displaystyle
\braa{m_{\rho'}^2-m_{h_1}^2}\braa{\frac{r_{a_1}}{\sqrt{2}f_\pi}\cos \theta_\rho}
$&$ \displaystyle
\sqrt{2} g\cos\theta_\rho \frac{g_{a_1}}{m_{a_1}^2}
$
\\
$b_1 $&$ \omega$&$ \displaystyle
\braa{m_{b_1}^2-m_{\omega}^2}\braa{\frac{r_{a_1}}{\sqrt{2}f_\pi}\sin \theta_\omega}
$&$ \displaystyle
\sqrt{2} g\sin\theta_\omega \frac{g_{a_1}}{m_{a_1}^2}
$
\\
$\omega' $&$ b_1$&$ \displaystyle
\braa{m_{\omega'}^2-m_{b_1}^2}\braa{\frac{r_{a_1}}{\sqrt{2}f_\pi}\cos \theta_\omega}
$&$ \displaystyle
\sqrt{2}g\cos\theta_\omega \frac{g_{a_1}}{m_{a_1}^2}
$
\\\hline\hline
\end{tabular}
\end{center}
\label{Table:coupling VVpi}
\end{table}
\section{Relations among one-pion decays of spin-1 mesons}
\label{sec:one pion decays}
In this section we give several relations among
one-pion interactions of two spin-1
mesons.
We would like to stress that all the one-pion decays of spin-1 mesons are expressed by one parameter $r_{a_1}/f_\pi$ reflecting the existence of the
$\mbox{SU(4)}$ symmetry as shown in Eq.\,\eqref{GT relation}.
By using Eq\,\eqref{GT relation},
the one-pion decay widths of spin-1 mesons are easily calculated:
\bee{
\Gamma (V_i \ra V_f \pi)
=&
\frac{1}{8\kappa_{V_i \ra V_f }}\frac{\brad{\vec{p}_{V_i\ra V_f}}}{\pi m_{V_i}^2}
\braa{
g_{V_i V_f \pi}
}^2
\nn\\
&
\times\braa{
m_{V_i}^2
-
m_{V_f}^2
}^2
\braa{3+
\frac{\brad{\vec{p}_{V_i\ra V_f}}^2}{m_{V_f}^2}
}
\ ,
}
where the momentum is given as
\bee{
\brad{\vec{p}_{V_i \ra V_f}}
\equiv
\frac{1}{2}
\sqrt{m_{V_i}^2 - 2\braa{m_{V_f}^2 + m_{\pi}^2}+\frac{\braa{m_{V_f}^2- m_\pi^2}^2}{m_{V_i}^2}}
\ ,
}
and the factor $\kappa$ depends on the isospin of the initial and final states:
\bee{
\kappa_{h_1 \ra \rho} =& \kappa_{\omega' \ra b_1} = 1
\ ,\nn\\
\kappa_{a_1 \ra \rho} =& \kappa_{\rho' \ra a_1} = \frac{3}{2}
\ ,\nn\\
\kappa_{b_1 \ra \omega} =& \kappa_{\rho' \ra h_1} = 3
\ .
}
The unknown parameter $r_{a_1}/f_\pi$ in the coupling $g_{VV\pi}$ is canceled by taking ratios of these decay widths:
\bee{
&
\frac{\Gamma (\rho' \ra h_1 \pi)}{\Gamma (a_1 \ra \rho \pi)}
=
0.16 \pm 0.07
\ ,
\nn\\
&
\frac{\Gamma (h_1 \ra \rho \pi)}{\Gamma (\rho' \ra a_1 \pi)}
=
6.4\pm4.3
\ ,
\nn\\
&
\frac{\Gamma (\rho' \ra a_1 \pi)}{\Gamma (a_1 \ra \rho \pi)}
=
\braa{
0.15\pm0.11
} \tan^2\theta_\rho
\ ,
\nn\\
&
\frac{\Gamma (h_1 \ra \rho \pi)}{\Gamma (a_1 \ra \rho \pi)}
=
\braa{
1.0\pm0.3
} \tan^2\theta_\rho
\ ,
\nn\\
&
\frac{\Gamma (b_1 \ra \omega \pi)}{\Gamma (\omega' \ra b_1 \pi)}
=
\braa{
3.9\pm1.9
} \tan^2\theta_\omega
\ ,
\label{ratio between the widths}
}
where the numerical factors in the RHS are simply evaluated from the corresponding kinematical factors calculated by using the masses listed in Table\,\ref{Table:masses}.
Errors in the RHS are estimated from the errors listed in the table.
Since the first two relations are independent of the parameters,
experimental measurements of these ratios will check the existence of the $\mbox{SU(4)}$ symmetry.
Then, we can determine the mixing angles from the latter three relations.
\section{Extended KSRF relations}
\label{sec:KSRF}
In this section, we
derive the Kawarabayashi-Suzuki-Riazuddin-Fayyazuddin (KSRF) relations
among the $\rho$ meson mass, the $\rho\pi\pi$ coupling and the $\rho$-photon mixing strength, as well as their extension to the $\rho'$ meson.
The interactions among one gauge field and two pion fields are included in
the $\mc{L}_V$ of the Lagrangian\,\eqref{Lagrangian full2}.
Similarly to the one-pion interactions studied in the previous section, due to the existence of the $a_1$-$\pi$ mixing, the three point interaction ${\mathcal L}_{\rm int}^{(3)}$ generates the interactions among a spin-1 meson and two pions.
The resultant
effective vertices among two pions and one vector meson are given by
\bee{
&\Gamma^\mu \brac{\rho_\mu^a(p), \pi^b (p_1) , \pi^c (p_2)}
\nn\\
=&
-i\epsilon^{abc}\brac{g^{(T)}_{\rho \pi\pi}(p^2)P^{\mu\nu} (p) \brac{p_1-p_2}_\nu
+ g^{(L)}_{\rho \pi\pi}(p_1 ,p_2){p}^\mu}
\ ,
\nn\\
&\Gamma^\mu \brac{\braa{\rho'}_\mu^a(p), \pi^b (p_1) , \pi^c (p_2)}
\nn\\
=&
-i\epsilon^{abc}\brac{g^{(T)}_{\rho' \pi\pi}(p^2)P^{\mu\nu} (p) \brac{p_1-p_2}_\nu
+ g^{(L)}_{\rho' \pi\pi}(p_1 ,p_2){p}^\mu
}
}
where $p=p_1+p_2$,
and
\bee{
g^{(T)}_{\rho \pi\pi}(p^2)=&\frac{m_\rho^2 -r_{a_1}^2 p^2}{\sqrt{2}gf_\pi^2}\cos \theta_\rho
\ ,
\nn\\
g^{(T)}_{\rho' \pi\pi}(p^2)=&\frac{m_{\rho'}^2 -r_{a_1}^2 p^2}{\sqrt{2}gf_\pi^2}\sin \theta_\rho
\ ,
\nn\\
g^{(L)}_{\rho \pi\pi}(p_1 ,p_2)=&\frac{m_\rho^2\cos \theta_\rho}{\sqrt{2}gf_\pi^2}
\frac{\braa{p_2}^2-\braa{p_1}^2}{\braa{p_1 + p_2}^2}
\ ,
\nn\\
g^{(L)}_{\rho' \pi\pi}(p_1 ,p_2)=&\frac{m_{\rho'}^2\sin \theta_\rho}{\sqrt{2}gf_\pi^2}
\frac{\braa{p_2}^2-\braa{p_1}^2}{\braa{p_1 + p_2}^2}
\label{rho pi pi coupling}
\ .
}
Since $g_{V\pi\pi}^{(L)}(p_1,p_2)$ ($V=\rho\,,\,\rho'$) vanishes for on-shell pion, only $g_{V\pi\pi}^{(T)}(p^2)$ is relevant for $V\to \pi\pi$ decay and the electromagnetic form factor of pion.
Furthermore, when the vector mesons are on their mass shell, the ratio of two $V\pi\pi$ couplings is related to the mixing angle as
\begin{equation}
\frac{g_{\rho'\pi\pi}^{(T)}(m_{\rho'}^2) }{ g_{\rho\pi\pi}^{(T)}(m_{\rho}^2) } = \frac{ m_{\rho'}^2 }{m_\rho^2} \, \tan\theta_\rho \ .
\label{ratio Vpp}
\end{equation}
We introduce the photon field $A_\mu$ by replacing the external gauge field as
\begin{equation}
{\mathcal V}_\mu = e A_\mu Q \ ,
\end{equation}
where $e$ is the electromagnetic coupling constant, and
\begin{equation}
Q = \sqrt{2}\braa{S^3+\frac{1}{3}S^0}
=\braa{\mt{t^3 + \frac{1}{6}1_2 &0\\0&t^3 + \frac{1}{6}1_2 }}
\ .
\end{equation}
The $\mc{L}_V$ part of the Lagrangian generates the mixing between a vector meson and the photon.
The mixing strengths for $\rho$, $\rho'$, $\omega$ and $\omega'$ mesons are expressed as
\begin{align}
g_{\rho}\equiv&
\frac{\sqrt{2}m_{\rho}^2\cos \theta_\rho}{g}
\ ,~~~~
g_{\rho'}\equiv
\frac{\sqrt{2}m_{\rho'}^2\sin \theta_\rho}{g}
\ ,
\label{coupling grho} \\
g_{\omega}\equiv&
\frac{\sqrt{2}m_{\omega}^2\cos \theta_\omega}{3g_B}
\ ,~~~~
g_{\omega'}\equiv
\frac{\sqrt{2}m_{\omega'}^2\sin \theta_\omega}{3g_B}
\ .
\label{coupling gomega}
\end{align}
Similarly to Eq.\,(\ref{ratio Vpp}), several ratios of two of above quantities are expressed as
\begin{align}
\frac{g_{\rho'} }{ g_\rho } = & \frac{ m_{\rho'}^2 }{m_\rho^2} \, \tan\theta_\rho \ , \notag\\
\frac{g_{\omega} }{ g_\rho } = & \frac{1}{3}\, \frac{ m_{\omega}^2 }{m_\rho^2} \frac{g}{g_B}\frac{\cos\theta_\omega}{\cos\theta_\rho}\ , \notag\\
\frac{g_{\omega'} }{ g_\omega } = & \frac{ m_{\omega'}^2 }{ m_\omega^2} \, \tan\theta_\omega \ ,
\label{ratio Vg}
\end{align}
Now, comparing the expressions in Eq.\,(\ref{coupling grho}) with the two-pion vertices in Eq.\,(\ref{rho pi pi coupling}),
one
can find the KSRF I relation and the extended one for $\rho'$ in the soft momentum limit, $p =0$:
\bee{
g_\rho=2 g_{\rho\pi\pi}^{(T)}(p^2=0) f_\pi^2
\ ,~~~
g_{\rho'}=2 g_{\rho' \pi\pi}^{(T)}(p^2=0) f_\pi^2
\ .
}
On the other hand, for the on-shell vector mesons, they become
\bee{
\frac{2 g_{\rho\pi\pi}^{(T)}(m_\rho^2) f_\pi^2}{g_\rho}
=\frac{2 g_{\rho'\pi\pi}^{(T)}(m_{\rho'}^2) f_\pi^2}{g_{\rho'}}=1-r_{a_1}^2
\label{KSRF I on-shell}
\ .
}
This implies that the deviation for the on-shell $\rho$ from the KSRF I relation is caused by the term including $r_{a_1}$, which is generated from $\mc{L}_{\rm int}^{(3)}$, in $g_{\rho\pi\pi}^{(T)}(p^2)$.
Let us consider the relations among several relevant decay widths.
The decay widths for the $\rho \to \pi\pi$ and $\rho^0 \to e^+e^-$
are calculated as
\bee{
\Gamma \braa{\rho \ra \pi\pi}
=&
\frac{1}{6\pi m_\rho^2}\brac{\frac{m_\rho^2 -4m_\pi^2}{4}}^{\frac{3}{2}}
\brad{g^{(T)}_{\rho \pi\pi}(m_\rho^2)}^2
\ ,
\label{Decay width rho to pipi}
\\
\Gamma \braa{\rho^0 \ra e^+e^-}
=&
\frac{4\pi \alpha^2}{3}
\brad{\frac{g_{\rho}}{m_\rho^2}}^2
\frac{m_\rho^2 + 2m_e^2}{m_\rho^2}\sqrt{m_\rho^2-4m_e^2}
\label{Decay width rho to ee}
\ ,
}
and similarly for $\rho' \to \pi\pi$ and ${\rho '}, \omega, \omega' \to e^+ e^-$.
Combining the relations in Eq.\,(\ref{ratio Vpp}) and (\ref{ratio Vg}), we obtain
\bee{
\frac{\Gamma \braa{\rho' \ra \pi\pi}}
{\Gamma \braa{\rho \ra \pi\pi}}
=&
\braa{
28 \pm 2
}
\tan^2 \theta_\rho
\ ,
\label{rho pipi decay}\\
\frac{\Gamma \braa{{\rho'}^0 \ra e^+e^-}}
{\Gamma \braa{\rho^0 \ra e^+e^-}}
=&
\braa{
1.89\pm 0.03
}
\tan^2 \theta_\rho
\ ,
\label{rhoprime decays} \\
\frac{ \Gamma \left( \omega \to e^+ e^- \right) }{ \Gamma \left( \rho^0 \to e^+ e^- \right) } = & \left(
0.112 \pm 0.000
\right) \, \frac{g^2}{g_B^2}\frac{ \cos^2\theta_\omega }{ \cos^2 \theta_\rho } \ ,
\label{omega ee decay} \\
\frac{ \Gamma \left( \omega' \to e^+ e^- \right) }{ \Gamma \left( \omega \to e^+ e^- \right) } = & \left(
1.01\pm 0.03
\right) \, \tan^2\theta_\omega \ ,
\label{omegap ee decay}
}
where $0.000$ in the third equations implies that the error is smaller than $0.0005$.
Taking the ratio of Eq.~(\ref{rho pipi decay}) and Eq.~(\ref{rhoprime decays}), we obtain the following parameter free relation:
\begin{equation}
\frac{ \Gamma \left( \rho' \to \pi\pi \right) }{ \Gamma \left( \rho \to \pi\pi \right) }
\frac{ \Gamma \left( \rho^0 \to e^+e^- \right) }{ \Gamma \left( \rho^{\prime0} \to e^+e^- \right) } = \left(
15 \pm 1
\right) \ ,
\end{equation}
which is regarded as an experimental check of the existence of $\mbox{SU(4)}$ symmetry for spin-1 mesons.
The relations in Eqs.\,(\ref{rho pipi decay})-(\ref{omegap ee decay}) can be
used to determine the relevant model parameters.
At this moment, we can set up an upper limit for the mixing angle $\tan^2\theta_\rho$ in the following way: The
total decay width of $\rho'$ and the partial decay width of the $\rho \ra \pi\pi$ channel are known as $\Gamma^{\rm total} (\rho') = 400\pm 60\,\mbox{MeV}$ and $\Gamma(\rho \ra \pi\pi) = 147.8\,\mbox{MeV}$, respectively.
Then the upper limit of the LHS of Eq.~(\ref{rho pipi decay}) is estimated as
\begin{equation}
\frac{ \Gamma \left( \rho' \to \pi\pi \right) }{ \Gamma \left( \rho \to \pi\pi \right) } \le
\frac{ \Gamma^{\rm total} \left( \rho' \right) }{ \Gamma \left( \rho \to \pi\pi \right) } \sim 3 \ .
\end{equation}
From this together with Eq.~(\ref{rho pipi decay}),
the limit is obtained as
\bee{
\tan^2\theta_\rho \lesssim 0.1
\label{upper limit of tan2}
\ ,
}
which implies that the mixing between
$V_{\parallel}$ and $V_{\perp(1)}$
is not large.
From this upper limit for the mixing angle, the ratio of $e^+e^-$ decays of $\rho'$ and $\rho$ mesons has an upper limit as
\begin{equation}
\frac{\Gamma \braa{{\rho'}^0 \ra e^+e^-}}
{\Gamma \braa{\rho^0 \ra e^+e^-}} \lesssim 0.2 \ .
\end{equation}
Using the upper limit for $\tan^2\theta_\rho$ in Eqs.\,\eqref{upper limit of tan2}, we obtain the upper limits for the ratios of one-pion decays of spin-1 mesons:
\begin{align}
& \frac{\Gamma (\rho' \ra a_1 \pi)}{\Gamma (a_1 \ra \rho \pi)} \lesssim 0.02 \ ,
\notag\\
& \frac{\Gamma (h_1 \to \rho \pi)}{\Gamma (a_1 \to \rho \pi)} \lesssim 0.1 \ ,
\notag\\
\end{align}
which can be tested in future experiments.
Next, we consider a constraint from Eq.~(\ref{omega ee decay}).
Using the experimental value $\Gamma(\omega \to e^+ e^-) = 0.60 \pm 0.02\,\mbox{keV}$, we obtain
\begin{equation}
\frac{g^2}{g_B^2}\frac{\cos^2\theta_\omega}{\cos^2\theta_\rho} =
0.76 \pm 0.03
\label{ratio cos omega}
\ .
\end{equation}
When the gauge couplings for $\mbox{SU(4)}$ HLS and $\mbox{U(1)}$ HLS are equal to each other, this together with Eq.~(\ref{upper limit of tan2}) gives a constraint as
\begin{equation}
\tan^2 \theta_\omega \lesssim 0.5
\ ,
\label{upper limit theta omega}
\end{equation}
and
\bee{
& \frac{\Gamma (b_1 \ra \omega \pi)}{\Gamma (\omega' \ra b_1 \pi)} \lesssim 2 \ ,
\nn\\
&
\frac{ \Gamma \left( \omega' \to e^+ e^- \right) }{ \Gamma \left( \omega \to e^+ e^- \right) } \lesssim 0.5 \ .
}
\section{Numerical analysis}
\label{sec: numerical analysis}
In this section, we determine the model parameters from the relevant experimental data, and make several phenomenological predictions.
We first construct the electromagnetic form factor of pion.
From the second term in the last line of Eq.\,\eqref{Lagrangian including alpha}, we can read the
direct $\gamma \pi \pi $ coupling as
\bee{
g_{\gamma \pi \pi}\equiv&
1-\frac{m_{\rho}^2\cos^2 \theta_\rho+m_{\rho'}^2\sin^2 \theta_\rho}{g^2f_\pi^2} \ .
\label{direct gamma pi pi}
}
From this and Eqs.\,(\ref{rho pi pi coupling}) and (\ref{coupling grho}), the
pion space-like form factor is given by
\bee{
F^{\pi^\pm}_V (Q^2)
=&
g_{\gamma\pi\pi}
+
\frac{g_\rho g_{\rho \pi\pi}^{(T)} (-Q^2)}{m_\rho^2 +Q^2}
+
\frac{g_{\rho'} g_{\rho' \pi\pi}^{(T)} (-Q^2)}{m_{\rho'}^2 +Q^2}
\label{pion form factor}
}
with $Q^2 = -q^2$, where $q$ is the photon momentum.
This form factor is normalized as $F^{\pi^\pm}_V (Q^2=0) =1$
reflecting the existence of the electromagnetic U(1) symmetry.
From this, the pion charge radius is calculated as
\bee{
\brae{r^2}^{\pi^\pm}_V\equiv&
-6\left.
\frac{\pd F^{\pi^\pm}_V (Q^2)
}{\pd Q^2}
\right|_{Q^2=0}
=
6
\frac{1-r_{a_1}^2}{g^2 f_\pi^2}
\label{pion charge radius}
\ .
}
By using Eqs.\,\eqref{coupling grho},
\eqref{KSRF I on-shell}, and
\eqref{pion charge radius},
the parameters are expressed as
\bee{
\brad{r_{a_1} }=& \sqrt{1- \braa{\frac{2 g_{\rho\pi\pi}^{(T)}(m_\rho^2) f_\pi^2}{g_\rho}}}
\ ,\label{parameter ra1}\\
\brad{g}
=&
\frac{m_\rho}{f_\pi}\sqrt{
\braa{\frac{2g_{\rho\pi\pi}^{(T)}(m_\rho^2)f_\pi^2}{g_\rho}}
/
\braa{\frac{1}{6}\brae{r^2}^{\pi^\pm}_V m_\rho^2}
}
\ ,\label{parameter g}\\
\tan^2 \theta_\rho
=&
\frac{\braa{\frac{1}{6}\brae{r^2}^{\pi^\pm}_V m_\rho^2}}{\braa{\frac{2g_{\rho\pi\pi}^{(T)}(m_\rho^2) f_\pi^2}{g_\rho}}
\braa{\frac{g_\rho}{\sqrt{2}m_\rho f_\pi}}^2
}
-1
\ .
\label{parameters expressed by grhopipi grho r2}
}
We should note that, since the pion charge radius is positive and
\begin{equation}
\frac{2 g_{\rho\pi\pi}^{(T)}(m_\rho^2) f_\pi^2}{g_\rho}=\frac{g^2 f_\pi^2}{6 \brae{r^2}^{\pi^\pm}_V}
\end{equation}
is satisfied, one can find that the couplings $g_{\rho\pi\pi}^{(T)}(m_\rho^2)$ and $g_{\rho}$ have the same sign.
Then, the inside of the square root in the RHS of Eq.~(\ref{parameter g}) is always positive.
Substituting the experimental values listed in Table\,\ref{Table:Experimental values}
into Eqs.~(\ref{parameter ra1})-(\ref{parameters expressed by grhopipi grho r2}), we have
\bee{
\brad{r_{a_1} }=&
0.41 \pm 0.10
\ ,\nn\\
\brad{g}
=&
7.1 \pm 0.5
\ ,\nn\\
\tan^2 \theta_\rho
=&
-0.03 \pm 0.25
\label{values of parameters}
}
where we added
10\% errors expected from higher order corrections~\cite{Harada:2003jx}.
Then the upper limit of $\tan^2\theta_\rho$,
$\tan^2\theta_\rho \lesssim 0.1$ given in Eq.\,\eqref{upper limit of tan2},
is within the errors of above determination.
The electromagnetic form factor obtained from these values is
shown in Fig.\,\ref{vector pion form factor} together with the experimental data.
\begin{figure}[h]
\centering \includegraphics[width=90mm]{form-factor-Q2-v7.eps}
\caption{Electromagnetic form factor of pion. Predicted form factor from the central values of the parameters in Eq.~(\ref{values of parameters}) is shown by the red curve. Shaded area shows the errors of the parameters. Black
dots express the experimental data given in Refs.\,\cite{Amendolia:1986wj,Bebek:1977pe,
Volmer:2000ek,Horn:2006tm,Tadevosyan:2007yd}.}
\label{vector pion form factor}
\end{figure}
This shows that the predicted form factor reasonably reproduce
the experimental data, taking account of their errors.
We would like to note that the both numerators of the $\rho$ and $\rho'$ contributions in Eq.\,\eqref{pion form factor} have the same sign, which is contrasted to the result by a holographic QCD model~\cite{Harada:2010cn}.
Furthermore, the direct $\gamma\pi\pi$ coupling is evaluated through
\bee{
g_{\gamma\pi\pi}
=&
1-
\frac{m_{\rho'}^2}{m_\rho^2}
\frac{\braa{\frac{1}{6}\brae{r^2}^{\pi^\pm}_V m_\rho^2}}{\braa{\frac{2g_{\rho\pi\pi}^{(T)}(m_\rho^2) f_\pi^2}{g_\rho}}
}
+
\braa{\frac{g_\rho}{\sqrt{2}m_\rho f_\pi}}^2
\braa{\frac{m_{\rho'}^2}{m_\rho^2}-1}
\ ,
}
as $g_{\gamma\pi\pi}=
-0.28 \pm 0.13$ by using
the values in Table\,\ref{Table:Experimental values}.
This result means that there is a slight deviation from the vector meson dominance.
At the end of this section, we estimate several decay widths of spin-1 mesons by using the parameter set given in Eqs.\,\eqref{values of parameters} and show them in Table\,\ref{Table:decay widths from numerical analysis}.
\begin{table}
\caption[]{ Predicted values of one-pion decay widths of spin-1 mesons estimated from the parameter set in Eq.\,\eqref{values of parameters}.}
\begin{center}
\begin{tabular}{c|c}\hline\hline
Decay mode & Partial width (MeV)
\\\hline
$\Gamma \braa{a_1 \ra \rho\pi}$ &
$470 \pm 400$
\\
$\Gamma \braa{\rho' \ra a_1\pi}$ & $< 50$
\\
$\Gamma \braa{h_1 \ra \rho\pi}$ & $<60$
\\
$\Gamma \braa{\rho' \ra h_1\pi}$ & $80 \pm 300$
\\\hline\hline
\end{tabular}
\end{center}
\label{Table:decay widths from numerical analysis}
\end{table}
\section{Summary and Discussions}
\label{sec:summary}
We constructed a chiral Lagrangian with an $\mbox{SU(4)}\times\mbox{U(1)}$ hidden local symmetry which includes the spin-1 mesons,
$\braa{\rho, a_1, \rho', \omega', b_1, f_1, h_1}$, together with pion.
We found
that each coupling of the interaction among one pion and two spin-1 mesons is proportional to the mass difference of the relevant spin-1 mesons similarly to the Goldberger-Treiman relation.
In addition, there were
the relations among one-pion decays of spin-1 mesons thanks to the existence of the SU(4) emergent symmetry.
Furthermore, we found a relation among the mass of $\rho'$ meson, the $\rho'\pi\pi$ coupling and the $\rho'$-photon mixing strength as well as the Kawarabayashi-Suzuki-Riazuddin-Fayyazuddin relation for the $\rho$ meson.
We summarize these predictions in Table\,\ref{Table:Predictions}.
\begin{table}
\begin{center}
\caption[]{
Predictions obtained from the $\mbox{SU(4)}\times\mbox{U(1)}$ HLS model.
}
\label{Table:Predictions}
\begin{tabular}{cccccccc}\hline\hline
\multicolumn{3}{c}{Independent of the parameters}
&
\\\hline
$\displaystyle \frac{\Gamma (\rho' \ra h_1 \pi)}{\Gamma (a_1 \ra \rho \pi)}$
&$=$
&$0.16 \pm 0.07$
&
\\
$\displaystyle \frac{\Gamma (h_1 \ra \rho \pi)}{\Gamma (\rho' \ra a_1 \pi)}$
&$=$
&$6.4\pm4.3$
&
\\\hline\hline
\multicolumn{3}{c}{Dependent on $\theta_\rho$ or $\theta_\omega$}
\\\hline
$\displaystyle \frac{\Gamma (\rho' \ra a_1 \pi)}{\Gamma (a_1 \ra \rho \pi)}$
&$=$
&$\braa{0.15\pm0.11}\, \tan^2\theta_\rho$
&$\lesssim$&
$0.015$
\\
$\displaystyle \frac{\Gamma (h_1 \ra \rho \pi)}{\Gamma (a_1 \ra \rho \pi)}$
&$=$
&$\braa{1.0\pm0.3}\, \tan^2\theta_\rho$
&$\lesssim$&
$0.10$
\\
$\displaystyle \frac{\Gamma \braa{\rho' \ra \pi\pi}}{ {\Gamma \braa{\rho \ra \pi\pi}}}$
&$=$&
$\braa{28 \pm 2}\tan^2 \theta_\rho$
&&(Input)
\\
$\displaystyle \frac{\Gamma \braa{{\rho'}^0 \ra e^+e^-}}{\Gamma \braa{\rho^0 \ra e^+e^-}}$
&$=$&
$\braa{1.89\pm 0.03}\tan^2 \theta_\rho$
&$\lesssim$&
$0.2$
\\
$\displaystyle \frac{\Gamma \left( \omega \to e^+ e^- \right) }{ \Gamma \left( \rho^0 \to e^+ e^- \right)} $ &$=$
&$\left(0.11219 \pm 0.00004\right) \,
\displaystyle \frac{ g^2 }{ g_B^2} \frac{ \cos^2\theta_\omega }{ \cos^2 \theta_\rho }$
&&(Input)
\\
$\displaystyle \frac{\Gamma (b_1 \ra \omega \pi)}{\Gamma (\omega' \ra b_1 \pi)}$
&$=$
&$\braa{3.9\pm1.9}\, \tan^2\theta_\omega$
\\
$\displaystyle \frac{\Gamma \left( \omega' \to e^+ e^- \right) }{ \Gamma \left( \omega \to e^+ e^- \right)}$ &$=$&
$\left( 1.01\pm 0.03\right) \, \displaystyle \tan^2\theta_\omega $
\\\hline\hline
\end{tabular}
\end{center}
\end{table}
Using two ratios indicated in Table\,\ref{Table:Predictions} together with the total widths of $\rho'$ and $\omega'$ mesons, we obtained several upper limit for the ratios as shown in the last column of the table.
By testing them in future experiments, we can verify the existence of the emergent symmetry.
There also exist hadronic decays which involve the intrinsic parity odd terms shown
in Appendix\,\ref{sec:Intrinsic parity odd terms}.
Possible decay
modes are expressed by ``$\times$'' in
Table\,\ref{Table:Decay channel},
such as $\rho' \ra \omega \pi$ and $\omega \ra 3\pi$.
We listed the allowed operators in Appendix\,\ref{sec:Intrinsic parity odd terms}.
Similarly to the result obtained in the generalized HLS at ${\mathcal O}(p^2)$~\cite{Bando:1987ym,Bando:1987br},
we also found $\Gamma\braa{a_1 \ra \pi \gamma}=0$ together with $\Gamma\braa{b_1 \ra \pi \gamma}=0$ and $\Gamma\braa{h_1 \ra \pi \gamma}=0$ at the leading order,
as we showed some detail calculations in Appendix\,\ref{sec:to pi gamma}.
As in the case of the generalized HLS, we expect that non-vanishing contributions will be produced
by higher order correction of the derivative expansion\,\cite{Bando:1987ym,Bando:1987br,Nagahiro:2008cv}.
We have to remark that our analysis are done in the chiral broken phase since we used the nonlinear realization of the chiral symmetry.
On the other hand, the existence of the emergent symmetry is proposed
by reducing the Dirac zero mode in the lattice QCD, which corresponds to
remove the dominant contribution of the chiral symmetry breaking
as shown by the Banks-Casher relation.
It will be interesting to clarify the correspondence between our model and the lattice QCD result, which we leave for future works.
\begin{table}
\begin{center}
\caption[]{Two-body pionic decay channels.
``$-$'' implies that the decay mode is prohibited kinematically or by the isospin, $\mc{P}$, and $\mc{C}$ while
``$\times$'' and ``$\sqrt{}$'' mean that such decay channel is allowed through the operators with the intrinsic parity odd and even, respectively.
}
\begin{tabular}{cc|ccccccccc|ccc}\hline\hline
Initial &mass& \multicolumn{10}{|c}{channel}
\\
&(MeV)&
$\pi \pi$&$\rho\pi$&$\omega\pi$&$\rho'\pi$&$\omega'\pi$&$a_1 \pi$&$f_1\pi$&$b_1\pi$&$h_1\pi$&$3\pi$&$\eta \pi\pi$&$4\pi$
\\\hline
$\rho$&$775.26$&$\sqrt{}$
&$-$&$-$&$-$&$-$&$-$&$-$&$-$&$-$
&$-$&$-$&$\sqrt{}$
\\
$\omega$&$782.65$&
$-$&$-$&$-$&$-$&$-$&$-$&$-$&$-$&$-$
&$\times$&$-$&$-$
\\
$\rho'$&$1465$&
$\sqrt{}$
&$-$
&$\times$&$-$&$-$&$\sqrt{}$&$-$&$-$&$\sqrt{}$
&$-$&$\times$&$\sqrt{}$
\\
$\omega'$&$\sim1425$&
$-$&$\times$
&$-$
&$-$&$-$&$-$&$-$&$\sqrt{}$
&$-$&$\times$&$-$
&$-$
\\
$a_1$&$1230$&
$-$&$\sqrt{}$
&$-$&$-$&$-$&$-$&$-$&$-$&$-$
&$\sqrt{}$&$-$
&$-$
\\
$f_1$&$1281.9$&
$-$&$-$
&$-$&$-$&$-$&$-$&$-$&$-$&$-$
&$-$&$\sqrt{}$&$\times$
\\
$b_1$&$1229.5$&
$-$&$-$
&$\sqrt{}$&$-$&$-$&$-$&$-$&$-$&$-$
&$-$&$\sqrt{}$&$\times$
\\
$h_1$&$1170$&
$-$&$\sqrt{}$
&$-$&$-$&$-$&$-$&$-$&$-$&$-$
&$\sqrt{}$&$-$&$-$
\\\hline\hline
\end{tabular}
\end{center}
\label{Table:Decay channel}
\end{table}
\section*{ACKNOWLEDGEMENTS}
We would like to thank Yuichi Motohiro for useful discussions.
The work of MH is supported in part by the JSPS Grant-in-Aid for Scientific Research (c) No.~16K05345.
The work of HN is supported in part by JSPS KAKENHI Grant Number~JP16J03578.
|
\section{Introduction}\label{sec:start}
Quantum graphs were recently introduced as general models to explore the fundamental limits of nonlinear optics and to discover the topological and geometrical properties of structures that maximize the nonlinear optical response\cite{shafe12.01,lytel13.01,lytel13.04,lytel14.01,lytel15.01}. Like other Hamiltonian models, quantum graphs with specific shapes and topologies have nonlinear optical response near the fundamental limits\cite{lytel16.01}.
A quantum graph is a metric graph on which electron dynamics is confined to the edges of the graph. Quantum graphs were first studied as tractable molecular models\cite{pauli36.01,kuhn48.01,ruede53.01,scher53.01,platt53.01} and have been invoked as models of mesoscopic systems\cite{kowal90.01}, optical waveguides \cite{flesi87.01}, quantum wires\cite{ivche98.01,sanch98.01}, excitations in fractals \cite{avish92.01}, and fullerines, graphene, and carbon nanotubes\cite{amovi04.01,leys04.01,kuchm07.01}. Quantum graphs are also exactly solvable models of quantum chaos\cite{kotto97.01,kotto99.02,kotto00.01,blume01.04}. In nonlinear optics, quantum graphs are models of a branched nano-wire structure, or a quasi-linear molecule, such as a donor-acceptor, with side groups.
The solution of quantum graphs for nonlinear optics has used the sum over states perturbation theory developed decades ago\cite{orr71.01}. The process requires determination of the eigenfunctions and energies of the graph for large numbers of states, and careful determination of the degenerate states comprising the spectrum. A Monte Carlo calculation of the response for a simple graph can take hours of computer time, frequent checks that the solutions for each graph are correct through the use of sum rules, and careful delineation of all degeneracies. The value of such simulations to the molecular designer has been investigated, and new design rules have been posited\cite{lytel15.02}. But the method does not scale well to graphs with many edges and large numbers of degenerate states that are often impossible to determine from a direct diagonalization of the Hamiltonian.
Recently, the Dalgarno-Lewis (DL) formulation of perturbation theory\cite{dalga55.01} has been adapted to off-resonance and dispersive nonlinear optics\cite{mossm15.01}. The DL approach requires knowledge of only the ground state of the structure. (For intrinsic nonlinearities\cite{kuzyk00.01}, the ground and first excited state energies are also required.) As such, it offers a huge computational advantage for large-scale calculations, as only a single eigenfunction must be specified. Degeneracies are implicitly incorporated into the DL formulation. Application of the DL formalism to quantum graphs opens up the possibility of a general computational approach for exploring the theory of fundamental limits, as well as for simulating molecular and nanostructures that might achieve larger intrinsic nonlinear optical responses.
Section \ref{sec:QG} reviews the solution of quantum graphs and the use of the sum over states to compute the nonlinearities. Section \ref{sec:DL} introduce the DL formalism and formally adapts it directly to quantum graphs. Section \ref{sec:DLQG} develops the general algorithm for solving quantum graphs using the results of Section \ref{sec:DL} and illustrates it with wire, star, loop, and a composite, seven-edged graph, and displays the accuracy and computational speedup of the method over the sum over states. Section \ref{sec:end} summarizes the benefits of the method and its extension to dispersive nonlinear optics.
\section{Computation with quantum graphs}\label{sec:QG}
Optimization studies using quantum graphs are based upon Monte Carlo computations with tens to hundreds of thousands of possible geometries for a particular topology, such as a star or loop motif\cite{lytel13.02}, or a composite of stars, loops, and wires. For each structure, the Hamiltonian is diagonalized to find the energy spectrum $E_{n}$ and the eigenstates $|n\rangle$, where $n$ is the mode number. The transition moments $r^{i}_{nm}=\langle n|r^{i}|m\rangle$ are computed for the graph using a union operation\cite{lytel13.01}, and the first and second hyperpolarizability tensors $\beta_{ijk}$ and $\gamma_{ijkl}$ are calculated using perturbation theory based upon the sum over states\cite{orr71.01}:
\begin{equation}\label{beta_ijk}
\beta_{ijk}=\left(\frac{e^3}{2}\right)P_{ijk}{\sum_{nm}}'\frac{r_{0n}^{i}\bar{r}_{nm}^{j}r_{m0}^{k}}{E_{n0}E_{m0}}
\end{equation}
and
\begin{eqnarray}\label{gamma_ijkl}
\gamma_{ijkl}&=&\left(\frac{e^4}{6}\right)P_{ijkl}\Biggl({\sum_{nmp}}'\frac{r_{0n}^{i}\bar{r}_{nm}^{j}\bar{r}_{mp}^{k}r_{p0}^{l}}{E_{n0}E_{m0}E_{p0}}\nonumber \\
&-&{\sum_{nm}}\frac{'r_{0n}^{i}r_{n0}^{j}r_{0m}^{k}r_{m0}^{l}}{E_{n0}E_{n0}E_{m0}}\Biggr)
\end{eqnarray}
where $\bar{r}^{i}_{nm}=r^{i}_{nm}-\delta_{nm}r^{i}_{00}$, $E_{n0}=E_{n}-E_{0}$, and the permutation operator P permutes all indices into all possible permutations without regard to order. The indices are $i=x,y$ for a planar graph embedded in two-dimensional space. Both hyperpolarizabilities have been normalized to their maximum values\cite{kuzyk00.01} and are bounded above by unity. The lower bound on $|\beta_{ijk}|$ is zero, while that for $\gamma_{ijkl}$ is $-0.25$. The geometry of the structure largely determines the total contributions of the transition moments from each edge, each of which is determined by the mode overlap of the eigenfunctions with the appropriate position operator and the relative orientation of the edges. Tailoring this overlap and the change in the dipole moment for electron excitation from the ground state to an excited state can lead to a giant enhancement of the nonlinear optical response due to phase disruption of the eigenfunctions $\psi_{n}(s)=\langle s|n\rangle$, where $s$ is measured along the graph\cite{lytel15.02}.
Figure \ref{fig:graph5edge} illustrates the notation for computation with quantum graphs. Each edge has a coordinate $s_{p}$ measured along the edge from zero to $a_{p}$ and an angle $\theta_{p}$ with respect to an external axis, which we choose to be the $x-axis$ for convenience. The edge functions $\phi^{(p)}_{n}(s_{p})$ for edge $p$ and mode $n$ on that edge satisfy $H\phi^{(p)}_{n}=E_{n}\phi^{(p)}_{n}$, where the energies $E_{n}$ are the same on all $p$ edges. The full eigenfunctions $\psi_{n}$ are constructed using a mathematical union operation
\begin{equation}\label{edgeUnion}
\psi_{n}(s)=\bigcup_{p=1}^{E}\phi^{(p)}_{n}(s_{p})
\end{equation}
operating on all $E$ edges\cite{lytel13.01}. The Hilbert space for the solutions to the entire graph is a direct sum of those for each individual edge. The interpretation of the coordinates in Eqn \ref{edgeUnion} is that when the electron is on edge $p$, the eigenfunction $\psi_{n}(s)$ is equal to $\phi_{n}^{(p)}(s_{})$ for any mode $n$.
\begin{figure}\centering
\includegraphics[width=3.4in]{graph5edge.eps}\\
\caption{A five-edge quantum graph, with two 3-star motifs, defining the coordinates and geometry used in the text.}\label{fig:graph5edge}
\end{figure}
The construction of the union of edge functions to make an eigenfunction of the graph requires matching of boundary conditions at the vertices within the graph. For quantum graph unions, the boundary conditions are that the edge functions at each vertex are equal, and that the net probability flux into or out of a vertex is zero, or equivalently
\begin{equation}\label{flux}
\sum_{p}^{E_{v}}\frac{d\phi^{(p)}_{n}(s_{p})}{ds_{p}}\Big|_{v}=0
\end{equation}
where the sum is over the $E_{v}$ edges going into or out of the vertex $v$. There are always enough boundary conditions to compute all of the constants of integration from the solutions of the edge functions.
The transition moments are computed by projecting the coordinate vector $s_{i}$ for a given edge $i$ onto the $x-axis$ to compute $\langle n|x^{i}|m\rangle =\langle n|s_{i}|m\rangle \cos{\theta_{i}}$ and onto the $y-axis$ to compute $\langle n|y^{i}|m\rangle =\langle n|s_{i}|m\rangle \sin{\theta_{i}}$ for each edge. Projected into coordinate space, the transition moments take the form of sums of definite integrals. For example, for $x_{nm}$, we get
\begin{equation}\label{moments}
x_{nm}=\sum_{p=1}^{E}\cos{\theta_{p}}\int_{0}^{a_{p}} ds_{p}s_{p}\phi^{(p)}_{n}(s_{p})\phi^{(p)}_{m}(s_{p})
\end{equation}
where the sum is over the $E$ edges, with a similar expression with $cos\rightarrow sin$ for $y_{nm}$. The solution of a quantum graph yields the spectrum and transition moments required by Eqns \ref{beta_ijk} and \ref{gamma_ijkl} for the hyperpolarizability tensors. Once a graph is specified by its metric variables $a_{p},\theta_{p}$ and the potentials acting on each edge, the dependence of the magnitudes of the hyperpolarizabilities on the geometry and topology of the graph is computable and may be studied using Cartesian or spherical tensor analysis\cite{jerph78.01,lytel13.01}.
Explorations of the topological structures maximizing the hyperpolarizabilities require computation of $\beta_{ijk}$ and $\gamma_{ijkl}$ for each of tens of thousands of graphs. Computation for each graph requires solving the graph for its energies and eigenfunctions, typically for at least $10-50$ modes, depending on the nature of the potential energy. A single Monte Carlo run for a simple three-edged graph might require several hours of execution time on an eight-threaded, quad-core Intel $i7$ CPU clocked at $3.7 GHz$. The computation requires careful handling of the solutions of the graph to ensure that any and all degenerate states are discovered and enumerated. For a typical computation, it is necessary to invoke the Thomas-Reiche-Kuhn (TRK) sum rules\cite{thom25.01,reich25.01,kuhn25.01} to ensure that all such states have been correctly captured and the spectra correctly determined\cite{lytel13.01}. Computation of a quantum graph for nonlinear optics is an elaborate procedure that is time-consuming and often frustrating when graph geometries produce numerous degeneracies. The Dalgarno-Lewis perturbation theory offers a simplification of the Monte Carlo computation and an execution time speedup of over a factor of $50$.
\section{Dalgarno-Lewis perturbation theory}\label{sec:DL}
The Dalgarno-Lewis perturbation method has been reviewed in the literature\cite{mavro91.01,dalga55.01,harri77.01,maize11.01} and in classic quantum mechanics textbooks\cite{schif68.01}. A paper in the present volume describes its application to dispersive nonlinear optics and solves a number of one-dimensional examples. Quantum graphs are quasi-one dimensional objects, with particle motion along the edges of the graph and not transverse to it. But they live in two dimensions and have $x$ and $y$ projections that contribute to the hyperpolarizability tensors. Moreover, their Hilbert space is a direct sum of Hilbert spaces for each of the edges, which adds complexity to the DL formalism and for which no algorithms have yet been developed. This section summaries the DL formalism for one dimension and then extends it to quantum graphs for the first time.
\subsection{DL theory in one dimension}\label{sec:DLone}
The Dalgarno-Lewis method for perturbation theory in one dimension begins by assuming the existence of an operator $F$ whose commutator with the Hamiltonian $H=p^2/2m+V$ takes the form
\begin{equation}\label{defineF}
\left[F,H\right]=\bar{x}
\end{equation}
with $\bar{x}=x-x_{00}$. The right-hand side is essentially the perturbing potential on a molecule in an electric field, $V=-ex$, with the ground state expectation value subtracted and the electric charge e dropped, as it will be incorporated elsewhere later.
Acting on the ground state with both sides of Eqn \ref{commFH} yields
\begin{equation}\label{commFH}
\left[F,H\right]|0\rangle = \bar{x}|0\rangle
\end{equation}
If we operate on the left of both sides of Eqn \ref{commFH} with $\langle n|$ and use $H|n\rangle = E_{n}|n\rangle$, we get
\begin{equation}\label{Fn0}
F_{n0}=-\frac{\bar{x}_{n0}}{E_{n0}}
\end{equation}
which defines matrix elements of the F operator in a way that allows a substitution into the sum over states expressions for $\beta_{xxx}$:
\begin{eqnarray}\label{betaxxx}
\beta_{xxx}&=&3e^3{\sum_{n,m}}'\frac{x_{0n}\bar{x}_{nm}x_{m0}}{E_{n0}E_{m0}} \\
&=&-3e^3{\sum_{n,m}}'F_{0n}\bar{x}_{nm}F_{m0}\nonumber \\
&=& -3e^3\langle 0|F\bar{x}F|0\rangle\nonumber
\end{eqnarray}
where the last line in Eqn \ref{betaxxx} follows by using completeness of the eigenstates. If the DL operator exists, then the first hyperpolarizability is computable through the ground state expectation value of the composite operator $F\bar{x}F$. Evidently, the effect of the perturbation on the excited states and their contributions to $\beta_{xxx}$ are embodied in the $F$ operator.
To obtain a spatial representation for $F$ that we may use in computations, we project Eqn \ref{commFH} into coordinate space as follows. The Hamiltonian, in coordinate space, is given by
\begin{equation}\label{Hamiltonian}
H=-\frac{\hbar^2}{2m}\frac{d^2}{dx^2} + V(x),
\end{equation}
We want to project Eqn \ref{defineF} into position space, where we know $\langle x|0\rangle=\psi_{0}(x)$ is the ground state wavefunction. We will need the projection of F into x-space, which is defined through
\begin{eqnarray}\label{F0}
\langle x|F|0\rangle&=&\int dx'\langle x|F|x'\rangle\langle x'|0\rangle \\
&=&F(x)\psi_{0}(x)
\end{eqnarray}
which defines the function $F(x)$. We are asserting that $F(x)$ exists for our operator. If we can derive an expression for it, then we will have found a representation for the $F$ operator and may use Eqn \ref{betaxxx} to compute the first hyperpolarizability as an expectation value of $F\bar{x}F$ in the \emph{unperturbed} ground state. Note that we do not even need to know the spectrum. The energy scale won't enter into the problem until we divide by $\beta_{max}$ from the theory of fundamental limits. We'll return to this later.
Projecting the left-hand side (LHS) of Eqn \ref{commFH} into x-coordinate space gives
\begin{eqnarray}\label{LHSF}
LHS&=& F(x) \left(- \frac{\hbar^2}{2m} \frac{d^2}{dx^2} + V(x)\right) \psi_{0}(x) \\
&-&\left( - \frac{\hbar^2}{2m} \frac{d^2}{dx^2} + V(x) \right) F(x) \psi_{0}(x)\nonumber \\
&=& \frac{\hbar^2}{2m} \left(\frac{d^2F(x)}{dx^2}\psi_{0}(x) + 2 \frac{dF(x)}{dx} \frac{d\psi_{0}(x)}{dx}\right)\nonumber \\
&=&\frac{\hbar^2}{2m}\frac{1}{\psi_{0}(x)} \frac{d}{dx}\left(\left(\psi_{0}(x)\right)^2\frac{dF(x)}{dx}\right)\nonumber.
\end{eqnarray}
Projecting the right-hand side of Eqn \ref{commFH} into coordinate space yields
\begin{equation}\label{RHSF}
RHS=\bar{x}\psi_{0}(x)
\end{equation}
Equating Eqns \ref{LHSF}, and \ref{RHSF}, we get the differential equation
\begin{equation}\label{diffEQforF}
\frac{d}{dx}\left(\left(\psi_{0}(x)\right)^2\frac{dF(x)}{dx}\right) = \frac{2m}{\hbar^2}\bar{x} \left(\psi_{0}(x)\right)^2.
\end{equation}
The general solution to Eqn \ref{diffEQforF} is obtained with two quadratures:
\begin{eqnarray}\label{Fofx}
F(x)&=&F(0)+\frac{2m}{\hbar^2}\int_{0}^{x}\frac{dx'}{\psi_{0}(x')^{2}}\\
&\times&\left(\int_{0}^{x'}dx''{\bar{x}}''\psi_{0}(x'')^{2}+\frac{\hbar^2}{2m}F'(0)\psi_{0}(0)^{2}\right)\nonumber
\end{eqnarray}
Eqn \ref{Fofx} is the main result. Note that the lower limit of the inner integral could have been chosen to be some value $\delta$, where $F'(\delta)\psi_{0}(\delta)^2=0$, but we choose $\delta=0$ for quantum graphs. Once the ground state wavefunction is known, the function $F(x)$ may be found by analytical or numerical integration, once two boundary conditions are specified. Actually, for bounded problems, such as a quantum well where the wavefunction vanishes at each end, the second term in Eqn \ref{Fofx} vanishes without specifying a boundary condition $F'(0)$. We may always choose $F(0)$ any way we please, because the operator $F$ is determined by a commutator and is therefore only known up to a constant anyway.
To calculate $\gamma$ from the sum over states, we need an additional operator $G$. In analogy with Eqn \ref{commFH}, we intuit that this is
\begin{equation}\label{commGH}
\left[G,H\right]|0\rangle = \Big[\bar{x}F-\langle 0|\bar{x}F|0\rangle\Big] |0\rangle
\end{equation}
which leads to a differential equation for $G(x)$, the projection of $G$ into coordinate space. The solution to this equation may be written down by inspection of Eqn \ref{Fofx} and is
\begin{eqnarray}\label{Gofx}
G(x)&=& G(0) \\
&+&\frac{2m}{\hbar^2}\int_0^x \frac{dx'}{\psi_{0}(x')^{2}}\Bigg( \int_{0}^{x'}dx''\big[{\bar{x}}''F(x'')\nonumber \\
&-&<0|{\bar{x}}''F|0>\big] \psi_{0}(x'')^{2} +\frac{\hbar^2}{2m}G'(0)\psi_{0}(0)^{2}\Bigg).\nonumber
\end{eqnarray}
Eqn \ref{commGH} implies that
\begin{equation}\label{Gn0}
G_{n0}=-\frac{\langle\bar{x}F\rangle_{n0}}{E_{n0}}
\end{equation}
which may be used to write Eqn \ref{gamma_ijkl} in a compact sum of ground state expectation values of operator products of the form $F^px^qG^r$ with integers $p,q,r$. We present an explicit form in the next section that is applicable to quantum graphs.
Note that we have defined the action of the anti-Hermitian operator $F$ on the ground state $F|0\rangle$ to be equal in x-space to $F(x)\psi_{0}(x)$. This implies that the adjoint has a minus sign: $\langle 0|F$ becomes $-F^{*}(x)\psi_{0}^{*}(x)$. This is important because operator products like $FF$ will produce minus signs when they are sandwiched between ground states. The same is true for $G$ if it acts to the left.
\subsection{Extension to quantum graphs}\label{sec:exten}
We can easily extended the one dimensional Dalgarno-Lewis formalism to quasi-one dimensional graphs using the fact that the Hamiltonian rotates as a quadratic form, viz., $H(x,y)\rightarrow H(s)+H(\tau)$ and ignoring the transverse part\cite{shafe11.02}. This implies that all of our one-dimensional work carries through for the $F$ and $G$ functions, except now the coordinates on the right hand side of Eqn \ref{commFH} and \ref{commGH} are either $x$ or $y$, depending on the transition moment one wishes to calculate. With this in mind, we arrive at
\begin{eqnarray}\label{Fi}
F_{i}(s)&=& F_{i}(0)+\frac{2m}{\hbar^2}\int_0^s \frac{ds'}{\psi_{0}(s')^{2}} \\
&\times&\Bigg( \int_{0}^{s'}\bar{r}^{i}(s'')\psi_{0}(s'')^{2} ds''+\frac{\hbar^2}{2m}{F_{i}}'(0)\psi_{0}(0)^{2}\Bigg)\nonumber
\end{eqnarray}
where $r^{1}=x$ and $r^{2}=y$. The coordinate variable $s$ is measured along the graph edges. The wavefunction and DL functions in Eqn \ref{Fi} are unions of functions defined on the graph edges. The interpretation of the integral is that when $s$ is on edge $p$, the values of the edge functions on edge $p$ are to be used, and the coordinate variable spans the range of that edge function. We quantify this description in Section \ref{sec:DLQG}.
For $\gamma$, we use the new set of functions, $G_{ij}$ defined by the commutator $[G_{ij},H]=\bar{r}^{i}F_{j}-<0|\bar{r}^{i}F_{j}|0>$, which translates in s-space to an equation for $G_{ij}(s)$ which is identical to that for $F_{i}$ in Eqn \ref{Fi}, but with an inner integrand given by the product $\bar{r}^{i}F_{j}-<0|\bar{r}^{i}F_{j}|0>$ expressed in s-space. The explicit form is
\begin{eqnarray}\label{Gij}
G_{ij}(s)&=& G_{ij}(0)=\frac{2m}{\hbar^2}\int_0^s \frac{ds'}{\psi_{0}(s')^{2}} \\
&\times&\Bigg( \int_{0}^{s'}\big[\bar{r}^{i}(s'')F_{j}(s'')-<0|\bar{r}^{i}F_{j}|0>\big] \psi_{0}(s'')^{2} ds''\nonumber \\
&+&\frac{\hbar^2}{2m}{G_{ij}(0)}'\psi_{0}(0)^{2}\Bigg).\nonumber
\end{eqnarray}
Note that the $s$ variable appearing in each of Eqns \ref{Fi} and \ref{Gij} roams over the entire graph and that $\psi_{0}(s)$ is the ground state wavefunction for the graph, which is the union of the edge functions.
It is now straightforward to derive simple expressions for the hyperpolarizabilities. First, we note that it is always possible to add a constant to F to make $\langle 0|F|0\rangle=0$ so we will assume that has always been done. Eqn \ref{Fn0} then allows us to write the expression for $\beta_{ijk}$ in Eqn \ref{beta_ijk} without energy denominators as
\begin{equation}\label{beta_ijk_simple}
\beta_{ijk}=-P_{ijk}\left(\frac{e^3}{2}\right)\left(\langle 0|F_{i}r^{j}F_{k}|0\rangle-r_{00}^{j}\langle 0|F_{i}F_{k}|0\rangle\right).
\end{equation}
We note that we can also add a constant to G to make $\langle 0|G|0\rangle=0$ without changing the commutator that defined G. We will always assume we have done this. Given this, we may use Eqn \ref{Gn0} to write down the general expressions for $\gamma_{ijkl}=\gamma^{(1)}_{ijkl}-\gamma^{(2)}_{ijkl}$ as
\begin{eqnarray}\label{gamma_ijkl_simple}
\gamma^{(1)}_{ijkl}&=&P_{ijkl}\left(\frac{e^4}{6}\right)\langle 0|F_{i}\bar{r}^{j}G_{kl}|0\rangle \\
\gamma^{(2)}_{ijkl}&=&P_{ijkl}\left(\frac{e^4}{6}\right)\langle 0|F_{i}F_{j}|0\rangle\langle 0|r^{k}F_{l}|0\rangle\nonumber
\end{eqnarray}
This completes the \emph{formal} specification of the F and G functions, and their use in calculating the hyperpolarizabilities. To apply this formalism to quantum graphs, we must express the DL functions in the language of the direct sum Hilbert space they represent.
\section{General solution for quantum graphs.}\label{sec:DLQG}
For graphs, the meaning of eigenfunctions $\psi_{n}(s)$ and ground state expectation values require sharp definition of the union operation that stitches together the Hilbert spaces on each edge to yield a direct sum Hilbert space with energies equal to those of the spaces for each edge. Continuity and flux conservation, Eqn \ref{flux}, showed how this is accomplished for the edge functions. With this algorithm, Eqn \ref{moments} showed how expectation values of operators acting across a graph are calculated, edge by edge, for transition moments. For $F_{i}$ and $G_{ij}$, we need the union of an $F$ or $G$ function tied to each edge:
\begin{equation}\label{Fip}
F_{i}(s)=\bigcup_{p=1}^{E}F_{i}^{(p)}(s_{p})
\end{equation}
with a similar union operation defined for $G_{ij}$:
\begin{equation}\label{Fip}
G_{ij}(s)=\bigcup_{p=1}^{E}G_{ij}^{(p)}(s_{p}).
\end{equation}
Here and going forward, the index $i$ refers to the Cartesian component $x$ or $y$ of an object, while the index $p$ refers to a mode number.
The matching of the edge $F_{i}^{(p)}$ and $G_{ij}^{(p)}$ functions at vertices is accomplished by demanding continuity of each at the vertex and a net flow into or out of the vertex equal to zero. A ground state expectation value of the form of Eqn \ref{moments} (with m=n=0) but for $xF_{x}$, say, would then be calculated as
\begin{equation}\label{xF00}
\langle 0|xF_{x}|0\rangle=\sum_{p=1}^{E}\cos{\theta_{p}}\int_{0}^{a_{p}} ds_{p}s_{p}F_{x}^{(p)}(s_{p})\phi_{0}^{(p)}(s_{p})\phi_{0}^{(p)}(s_{p})
\end{equation}
with a similar expression for expectation values of other products of operators. The boundary conditions arising from the union operation completely specify all integration constants appearing in the edge DL functions and provide a general solution to the problem of calculating the first and second hyperpolarizabilities from only expectation values of products of DL operators. For clarity, going forward, \emph{we drop the mode subscript $0$ on the edge functions in the ground state by writing $\phi^{p}_{0}(s_{p})\equiv\phi^{p}(s_{p})$}.
The explicit forms of the $F_{i}^{(p)}$ and $G_{ij}^{(p)}$ DL edge functions are
\begin{eqnarray}\label{Fip}
F_{i}^{(p)}(s_{p})&=& F_{i}^{(p)}(0)+\frac{2m}{\hbar^2}\int_0^{s_{p}} \frac{ds_{p}'}{\phi^{(p)}(s_{p}')^2} \\
&\times&\Bigg( \int_{0}^{s_{p}'}\bar{r}^{i}(s_{p}'')\phi^{(p)}(s_{p}'')^2 ds_{p}''+\frac{\hbar^2}{2m}{F'_{i}}^{(p)}(0){\phi^{(p)}}(0)^2\Bigg)\nonumber
\end{eqnarray}
and
\begin{eqnarray}\label{Gijp}
G_{ij}^{(p)}(s_{p})&=& G_{ij}^{(p)}(0)=\frac{2m}{\hbar^2}\int_0^{s_{p}} \frac{ds_{p}'}{\phi^{(p)}(s_{p}')^2} \\
&\times&\Bigg( \int_{0}^{s_{p}'}[\bar{r}^{i}(s_{p}'')F_{j}^{(p)}(s_{p}'')-\langle 0|\bar{r}^{i}F_{j}|0\rangle] \phi^{(p)}(s_{p}'')^2 ds''\nonumber \\
&+&\frac{\hbar^2}{2m}{G'_{ij}}^{(p)}(0){\phi^{(p)}(0)}^2\Bigg).\nonumber
\end{eqnarray}
where the angular brackets signify the expectation value in the ground state DL edge function. The integration constants are selected to satisfy the boundary conditions on the DL functions so their union is a function that is the projection of the DL operator $F$ or $G$ into coordinate space over the entire graph. The boundary conditions are that the $F_{i}^{(p)}(s_{p})$ are equal at the vertex where the edges which they are index meet, and the sum of the derivatives with respect to their edge coordinates, measured into or out of a vertex, of all edges meeting at that vertex vanishes. Similar conditions hold for $G_{ij}^{(p)}$ and its derivatives.
The first integral of the DL equation for $F_{i}^{(p)}(s_{p})$ for any edge p of a graph may be written as
\begin{equation}\label{Fp1}
{F'_{x}}^{(p)}(s_{p})\phi^{(p)}(s_{p})^2={F'_{x}}^{(p)}(0)\phi^{(p)}(0)^2+\frac{2m}{\hbar^2}\int_{0}^{s_{p}}ds'_{p}\bar{x}(s'_{p})\phi^{(p)}(s'_{p})^2
\end{equation}
where $\bar{x}(s'_{p})=x_{0p}+\cos{\theta_{p}}s'_{p}-x_{00}$ for the ith edge. Here, $x_{0p}$ is the offset from the origin for edge p, and $F'=d/ds$. In particular, the following relationship holds at the endpoints of an edge:
\begin{eqnarray}\label{Fp2}
&&{F'_{x}}^{(p)}(a_{p})\phi^{p}(a_{p})^2-{F'_{x}}^{(p)}(0)\phi^{(p)}(0)^2 \\
&=&\frac{2m}{\hbar^2}\int_{0}^{a_{p}}ds'_{p}\bar{x}(s'_{p})\phi^{(p)}(s'_{p})^2\nonumber \\
&\equiv& \frac{2m}{\hbar^2}\langle\bar{x}_{p}\rangle\nonumber
\end{eqnarray}
Note that each DL function for edge $p$ has two integration constants. We show how the boundary conditions determine these constants. Write Eqn \ref{Fip} as
\begin{eqnarray}\label{Fi2}
{F_{x}}^{(p)}(s_{p})&=&{F_{x}}^{(p)}(0)+\frac{2me}{\hbar^2}\int_{0}^{s_{p}}\frac{ds'_{p}}{\phi^{(p)}(s'_{p})^2}\\
&\times&\left(\int_{0}^{s'_{p}}ds_{p}''\bar{x}(s''_{p})\phi^{(p)}(s''_{p})+C_{p}\right)\nonumber
\end{eqnarray}
where the first integration constants are
\begin{equation}\label{Cp}
C_{p}=\frac{\hbar^2}{2m}{F'_{x}}^{(p)}(0)\phi^{p}(0)^2
\end{equation}
and are proportional to the net flux entering edge $p$. Now write Eqn \ref{Fp2} as
\begin{equation}\label{Cout}
{F'_{x}}^{(p)}(a_{p})\phi^{p}(a_{p})^2=\frac{2m}{\hbar^2}\big[C_{p}+\langle\bar{x}_{p}\rangle\big]
\end{equation}
which is proportional to the net flux exiting edge $p$ times the square of the $pth$ edge function evaluated at the exit vertex. Eqns \ref{Cp} and \ref{Cout} reveal that two edges $p1$ and $p2$ have a common vertex, say at $a_{p1}=a_{p2}=0$, then the $C_{p1}$ and $C_{p2}$ for edges $p1$ and $p2$ are related by flux conservation. Eqns \ref{Cp} and \ref{Cout} may thus be used to conserve DL function flux at each vertex. A final set of equations matches the values of the DL functions at each vertex:
\begin{equation}\label{Fbc}
{F_{x}}^{(E_1)}(v;1)={F_{x}}^{(E_2)}(v;2)=\ldots{F_{x}}^{(E_d)}(v;d)
\end{equation}
for every set of $d$ edges that come together at vertex $v$, and the notation $(v;n)$ means evaluate at the position $s_{n}$ for that edge at vertex $v$.
The continuity of $F_{x}^{(p)}$ and conservation of ${F'_{x}}^{(p)}$ resemble the boundary conditions on the edge functions in a quantum graph. These embody the definition of a union of functions operating on the edges.
For the G functions, a similar approach applies, with Eqn \ref{Gijp} replaced by
\begin{eqnarray}\label{Gp1}
{G_{xx}}^{(p)}(s_{p})&=&{G_{xx}}^{(p)}(0)+\frac{2m}{\hbar^2}\int_{0}^{s_{p}}\frac{ds'_{p}}{\phi^{(p)}(s'_{p})^2}\\
&\times&\bigg(\int_{0}^{s'_{p}}ds_{p}''\left[\bar{x}(s''_{p}){F_{x}}^{(p)}(s''_{p})-\langle 0|\bar{x}F_{x}|0\rangle\right]\phi^{(p)}(s''_{p})^2\nonumber \\
&+& D_{p}\bigg)\nonumber
\end{eqnarray}
for the xx component, with suitable modifications made for the yy and off-diagonal components. The constants $D_{p}$ are proportional to the $G_{xx}^{(p)}$ flux entering edge $p$.
\begin{equation}\label{Dp}
D_{p}=\frac{\hbar^2}{2m}{G'_{xx}}^{(p)}(0)\phi^{p}(0)^2
\end{equation}
where this latter derivative may be obtained from an expression similar to Eqn \ref{Fp2}, viz.,
\begin{eqnarray}\label{Gp2}
&& {G'_{xx}}^{(p)}(a_{p})\phi^{p}(a_{p})^2-{G'_{xx}}^{(p)}(0)\phi^{(p)}(0)^2 \\
&=&\frac{2m}{\hbar^2}\int_{0}^{a_{p}}ds'_{p}\big[\bar{x}_{p}(s'_{p})F_{x}^{(p)}(s'_{p})-\langle 0|\bar{x}F_{x}|0\rangle\big]\phi^{(p)}(s'_{p})^2\nonumber \\
&\equiv& \frac{2m}{\hbar^2}\langle\big[\bar{x}_{p}F_{x}^{(p)}-\langle 0|\bar{x}F_{x}|0\rangle\big]\rangle\nonumber
\end{eqnarray}
leading to
\begin{equation}\label{Dout}
{G'_{xx}}^{(p)}(a_{p})\phi^{(p)}(a_{p})^2=\frac{2m}{\hbar^2}\bigg[D_{p}+\langle\big[\bar{x}_{p}F_{x}^{(p)}-\langle 0|\bar{x}F_{x}|0\rangle\big]\rangle\bigg]
\end{equation}
where again the brackets on the last term are the ground state expectation value in edge $p$.
Eqns \ref{Dp} and \ref{Dout} may be used to conserve DL function flux at each vertex. A final set of equations matches the values of the DL functions at each vertex:
\begin{equation}\label{Gbc}
{G_{xx}}^{(E_1)}(v;1)={G_{xx}}^{(E_2)}(v;2)=\ldots{G_{xx}}^{(E_d)}(v;d)
\end{equation}
for every set of $d$ edges that come together at vertex $v$, where again the notation $(v;n)$ means evaluate at the position $s_{n}$ for that edge at vertex $v$.
The method derived in this section is applicable to any quantum graph. It is easy to use and will be illustrated with examples of wires, stars, and loops, and then applied to solve a seven-edge graph with three stars. This graph may have numerous degeneracies, depending on the relative edge lengths, that are challenging to discover\cite{lytel14.01,lytel15.01}. With the DL method, the solutions may be written down by inspection and easily coded, requiring only the ground state of the graph. This is an enormous simplification in solving the graph and leads to reduction in execution time for a large scale Monte-Carlo calculation from several hours to several minutes.
In the following examples, we consider $i=x$ for clarity. A similar set of results applies for the y direction by replacing x with y in the inner integrals' transition moment on each edge. \emph{We suppress the x-index to simplify notation}.
\subsection{Wire graphs.}\label{sec:3graphs}
Consider a three wire graph, as shown in the left panel of Figure \ref{fig:3graphs}. We need to solve for the $F_{x}^{(p)}$ (and $F_{y}^{(p)}$) on each edge so that we may compute matrix elements of the union over edges of $F_{x}=\bigcup_{p}F_{x}^{(p)}$ (and $F_{y}=\bigcup_{p}F_{y}^{(p)}$) in the ground state.
\begin{figure}\centering
\includegraphics[width=1.2in]{3wireBare.eps}\includegraphics[width=1.1in]{3star.eps}\includegraphics[width=0.9in]{loop.eps}\\
\caption{Three wire (left), 3-star (center), and loop (right) graphs used to illustrate how to compute the F functions for the edges.}\label{fig:3graphs}
\end{figure}
Examining the graph and Eqn \ref{Cp}, it is easy to see that $C_{1}=0$ because the edge function $\phi^{(1)}(0)=0$. To calculate $C_{2}$, we note that it is just the flux entering edge 2, which is also the flux exiting edge 1, and by Eqn \ref{Cout}, we get $C_{2}=C_{1}+\langle\bar{x}_{1}\rangle$, so $C_{2}=\langle\bar{x}_{1}\rangle$. Continuing, the flux entering edge three is $C_{3}$ is the flux exiting edge 2, which by Eqn \ref{Cout} is $C_{2}+\langle\bar{x}_{2}\rangle$, so $C_{3}=\langle\bar{x}_{1}\rangle+\langle\bar{x}_{2}\rangle$. Summarizing,
\begin{eqnarray}\label{Ci3wire}
C_{1}&=&0 \\
C_{2}&=&\langle\bar{x}_{1}\rangle\nonumber \\
C_{3}&=&\langle\bar{x}_{1}\rangle+\langle\bar{x}_{2}\rangle\nonumber \\
&=& -\langle\bar{x}_{3}\rangle\nonumber
\end{eqnarray}
where the last equality in Eqn \ref{Ci3wire} follows from the fact that summing over all edges E,
\begin{equation}\label{sumxbar}
\sum_{p=1}^{E}\langle\bar{x}_{p}\rangle=\sum_{p=1}^{E}\langle x_{p}\rangle-x_{00}=0
\end{equation}
by definition of $x_{00}$.
Eqn \ref{Fbc} then requires that
\begin{eqnarray}\label{Fcontinuity}
F_{1}(a_{1})&=&F_{2}(0)\\
F_{2}(a_{2})&=&F_{3}(0)\nonumber
\end{eqnarray}
It is straightforward to arrange these conditions once the Fs are computed from Eqns \ref{Fip}, \ref{Cp}, and \ref{Cout}. Note that this leaves one overall undetermined constant, $F_{1}(0)$. We can always choose this constant such that the unions of the edge DL functions have zero ground state expectation value, $\langle 0|F|0\rangle =0$. The Gs are computed in the same way using \ref{Gijp}, \ref{Dp}, and \ref{Dout}. This allows us to use the simple expressions for the hyperpolarizabilities displayed in Eqns \ref{beta_ijk_simple} and \ref{gamma_ijkl_simple}.
\subsection{Star graphs}\label{sec:star}
Consider next a 3-star graph, as shown in the center panel of Figure \ref{fig:3graphs}. If we write the three edges in the canonical form of Eqn \ref{Fip} and measure distance from the center vertex, we find from Eqn \ref{Cp} and Eqn \ref{Cout} that $C_{1}=-\langle\bar{x}_{1}\rangle$, $C_{2}=-\langle\bar{x}_{2}\rangle$, and $C_{3}=-\langle\bar{x}_{3}\rangle$ which is also equal to $\langle\bar{x}_{1}\rangle+\langle\bar{x}_{2}\rangle$ by Eqn \ref{sumxbar}. All of the $F_{x}^{(p)}$ are equal at the central vertex, and this is automatically satisfied if each of the F constants in Eqn \ref{Fi} are equal. The value of this single constant may be set by demanding $\langle 0|F|0\rangle =0$, as noted above.
Figure \ref{fig:beta_tensors_vs_angles_DL_3star_degen} shows the results of using DL to calculate all of the tensor components of $\beta_{ijk}$ for a 3-star graph with two fixed edges of lengths $0.4$ and $0.2$, angles of $180$ and $90$ degrees relative to the x-axis, and one rotating edge of fixed length $a_{3}=0.6$, and compares them to those calculated by using the sum over states with thirty modes. The DL method works with high precision. Figure \ref{fig:gamma_tensors_vs_angles_DL_3star_degen} shows the calculations for $\gamma_{ijkl}$. Note that the length scale is arbitrary, as these are intrinsic hyperpolarizabilities; the relative lengths determine the response of the structure.
The most remarkable thing about using DL for these sets of star graphs is that states $(4,5)$, $(10,11)$, etc. are degenerate in energy (where $n=0$ is the ground state), and the sum over states calculation requires identification of these states and diagonalization of the degenerate subspaces before it may be used. The DL computations don't know about any of this--they just require the (nodeless) ground state wavefunction. The DL operators, which implicitly sum up all of the contributions in perturbation theory from the excited states, automatically account for the degeneracies, as they followed from completeness of the unperturbed eigenstates.
\begin{figure}\centering
\includegraphics[width=3.4in]{beta_tensors_vs_angles_DL_3star_degen.eps}\\
\caption{Comparison of the Dalgarno-Lewis results for $\beta_{ijk}$ for the 3-star graph with those obtained using the full SOS for all tensor components. In the legend, $xxx\equiv\beta_{xxx}$, etc.}\label{fig:beta_tensors_vs_angles_DL_3star_degen}
\end{figure}
\begin{figure}\centering
\includegraphics[width=3.4in]{gamma_tensors_vs_angles_DL_3star_degen.eps}\\
\caption{Comparison of the Dalgarno-Lewis results for $\gamma_{ijkl}$ for the 3-star graph with those obtained using the full SOS for all tensor components. In the legend, $xxxx\equiv\gamma_{xxxx}$, etc.}\label{fig:gamma_tensors_vs_angles_DL_3star_degen}
\end{figure}
The second remarkable feature of the DL summation technique is illustrated in Figure \ref{fig:speedup}, which shows the massive speedup of DL over the sum over states with large numbers of modes.
\begin{figure}\centering
\includegraphics[width=3in]{speedup.eps}\\
\caption{Speedup of Dalgarno-Lewis computations over the sum over states.}\label{fig:speedup}
\end{figure}
\begin{figure}\centering
\includegraphics[width=3.4in]{FGloop.eps}\\
\caption{Plots of the $F_{i}$ and $G_{ij}$ functions for the closed loop graphs, showing their periodicity, as expected. Numerical values on the y axis have been suppressed for clarity.}\label{fig:FGloop}
\end{figure}
\subsection{Loops}\label{sec:loops}
Figure \ref{fig:3graphs}, right, shows a three-sided, closed loop graph. The computation of $\beta_{ijk}$ and $\gamma_{ijkl}$ for this class of graphs has been extensively discussed\cite{shafe12.01}. The ground state is nondegenerate, and is a constant, with zero energy. The excited state spectrum is doubly-degenerate. The eigenvalues are $k_{n}=2\pi n/L$, where $L$ is the perimeter of the loop.
The computation of the hyperpolarizabilities using DL may be accomplished by using Eqn \ref{Fip} but with a single edge and sequential calculation\cite{lytel12.01}. For the x direction, say, we get
\begin{eqnarray}\label{Fsingle}
F(s)&=&F(0)+\frac{2m}{\hbar^2}\int_{0}^{s}\frac{ds'}{\phi(s')^{2}}\\
&\times&\left(\int_{0}^{s'}ds''\bar{x}(s'')\phi(s'')^{2}+ C\right)\nonumber
\end{eqnarray}
with a similar equation for the y component of F. For a continuous wire graph, including a loop, we may use the sequential edge method of calculation\cite{lytel12.01}, where we measure distance s along the graph from an arbitrary starting point and change the value of $\bar{x}(s)$ in Eqn \ref{Fsingle} when we move from one edge to the next. In this way, we can solve for the F and G functions for this graph by using Eqn \ref{Fsingle} and an equivalent one for G by selecting the correct boundary conditions.
For a loop graph, we demand that $F(L)=F(0)$, in order to make F periodic as the particle circulates the graph. From Eqn \ref{Fsingle}, this requires that the first integration constant C be chosen so that the double integral on the right hand side vanishes at $s=L$. We arrive at
\begin{eqnarray}\label{Fbc}
C&=&-\int_{0}^{L}\frac{ds'}{\phi(s')^{2}}\left(\int_{0}^{s'}ds''\bar{x}(s'')\phi(s'')^{2}\right)\nonumber \\
&\div&\int_{0}^{L}\frac{ds'}{\phi(s')^{2}}
\end{eqnarray}
for the first integration constant. Finally, we note that the slopes are continuous, $F'(L)=F'(0)$, since the integral of $\bar{x}$ over the graph is zero by definition.
Similar remarks hold for the $G$ functions. For example, for $G$ in the x-direction, we find
\begin{eqnarray}\label{Gsingle}
G(s)&=&G(0)+\frac{2m}{\hbar^2}\int_{0}^{s}\frac{ds'}{\phi(s')^{2}}\\
&\times&\left(\int_{0}^{s'}ds''\big[\bar{x}(s'')F(s'')-\langle 0|\bar{x}F|0\rangle\big]\phi(s'')^{2}+ D\right)\nonumber
\end{eqnarray}
where the first integration constant is given by
\begin{eqnarray}\label{Gbc}
D&=&-\int_{0}^{L}\frac{ds'}{\phi(s')^{2}}\\
&\times&\left(\int_{0}^{s'}ds''\big[\bar{x}(s'')F(s'')-\langle 0|xF|0\rangle)\big]\phi(s'')^{2}\right)\nonumber \\
&\div&\int_{0}^{L}\frac{ds'}{\phi(s')^{2}}\nonumber
\end{eqnarray}
Fig \ref{fig:FGloop} shows the two F functions and for G for the loop graph. It is clear that the boundary conditions caused the functions to assume periodic behavior.
\subsection{Multiple star graph}\label{sec:7edge}
\begin{figure}\centering
\includegraphics[width=2.5in]{3prong.eps}\\
\caption{Three prong graph with seven edges.}\label{fig:3prong}
\end{figure}
We can extend the general algorithm to more complicated structures, such as the three prong graph having seven edges (a star at each end, and a star connecting these), shown in the left panel of Figure \ref{fig:3prong}. For convenience, we can put the origin at the first star vertex, where edges $a_{1},a_{2},a_{3}$ are joined. Using the algorithm in Eqn 4, we need to find seven first constants of integration, $C_{i},i=1-7$. These are given by
\begin{eqnarray}\label{Ci3prong}
C_{1}&=&-\langle\bar{x}_{1}\rangle\nonumber \\
C_{2}&=&-\langle\bar{x}_{2}\rangle\nonumber \\
C_{3}&=&\langle\bar{x}_{1}\rangle+\langle\bar{x}_{2}\rangle\nonumber \\
C_{4}&=&-\langle\bar{x}_{4}\rangle\\
C_{5}&=&\langle\bar{x}_{1}\rangle+\langle\bar{x}_{2}\rangle+\langle\bar{x}_{3}\rangle+\langle\bar{x}_{4}\rangle\nonumber \\
C_{6}&=&-\langle\bar{x}_{6}\rangle\nonumber \\
C_{7}&=&-\langle\bar{x}_{7}\rangle\nonumber \\
\end{eqnarray}
Note that we could have used the algorithm to rewrite $C_{6}$ as
\begin{eqnarray}\label{constants_i}
C_{6}&=&\langle\bar{x}_{1}\rangle+\langle\bar{x}_{2}\rangle+\langle\bar{x}_{3}\rangle \\
&+&\langle\bar{x}_{4}\rangle+\langle\bar{x}_{5}\rangle+\langle\bar{x}_{7}\rangle\nonumber
\end{eqnarray}
by Eqn \ref{sumxbar}.
Figure \ref{fig:panel1_quad_3prong} shows the edge functions along the graph and illustrates their continuity and the conservation of flux. The edge lengths $a_1$ to $a_7$ are $0.63, 3.61, 1.36, 1.50, 2.26, 2.70$, and $4.36$, in arbitrary units, and their respective angles, in degrees, are $180, 60,0,60,0,36$, and $0$. Edges 1,2,4,6,7 terminate at an infinite potential wall, and the edge functions vanish there, as may be seen in the Figure.
Figure \ref{fig:panel2_quad_3prong} shows that the SOS and DL results are identical using this formalism. The DL calculations executed ten times faster than using 20 modes in the SOS calculation. Moreover, the solution of the SOS calculation required careful computation of the amplitudes due to degeneracies in the graph. As graphs get larger, degeneracies become very difficult to compute, and the SOS method is impractical. The DL method does not require any information other than the ground state, which is nondegenerate.
\begin{figure}\centering
\includegraphics[width=3.4in]{panel1_quad_3prong.eps}\\
\caption{Edge functions on three prong graph. The x-axis denotes distance from the origin (black dot) in Fig 5. The three vertices of this graph are shown as the intersection points of the edge functions. Continuity of the edge functions and conservation of their derivatives at each vertex are evident in the Figure.}\label{fig:panel1_quad_3prong}
\end{figure}
\begin{figure}\centering
\includegraphics[width=3.4in]{panel2_quad_3prong.eps}\\
\caption{First hyperpolarizability tensors, computed by DL (solid lines) and SOS (open lines) for the seven-edge graph described in the text. Both methods agree to under a few hundreds of a percent. The SOS calculation used 20 modes and executed ten times more slowly than the DL calculation.}\label{fig:panel2_quad_3prong}
\end{figure}
Figures \ref{fig:panel3_quad_3prong} and \ref{fig:panel4_quad_3prong} show the $F_{i}$ function along the x and y directions, respectively. The figures show that the computation has produced edge DL functions that are continuous and have conserved slope at their vertices.
\begin{figure}\centering
\includegraphics[width=3.4in]{panel3_quad_3prong.eps}\\
\caption{$F_{x}^{(p)}$ functions for all graph edges using the same ordering scheme as used in Fig \ref{fig:panel1_quad_3prong} for the edge functions Note that the DL edge functions are continuous at the vertices, and their derivatives are conserved at the vertices.}\label{fig:panel3_quad_3prong}
\end{figure}
\begin{figure}\centering
\includegraphics[width=3.4in]{panel4_quad_3prong.eps}\\
\caption{$F_{y}^{(p)}$ functions for all graph edges using the same ordering scheme as used in Fig \ref{fig:panel1_quad_3prong} for the edge functions. Note that the DL edge functions are continuous at the vertices, and their derivatives are conserved at the vertices.}\label{fig:panel4_quad_3prong}
\end{figure}
\section{Conclusions}\label{sec:end}
The computation of the optical nonlinearities of quasi-one dimensional systems, modeled by quantum graphs, has been adapted from the sum over states (SOS) to the Dalgarno-Lewis (DL) method and resulted in an execution time speedup of anywhere from a factor of $10-50$ for a simple star graph. For large-scale Monte Carlo simulations normally taking hours with the SOS, the computation time is reduced to well under half an hour and often requires only a few minutes. SOS requires solving the quantum graph for all of its eigenfunctions and energies, whereas DL requires only knowledge of the ground state and, for intrinsic nonlinearities, the first two energy levels. A quantum graph with many edges has an exceedingly complex secular equation whose $M$ roots must be found for an SOS over $M$ states. Often, the roots are so closely spaced that numerical solutions are inaccurate, and it is impossible, a priori, to even know which roots are degenerate, let alone what the degree of degeneracy is. Subspaces of eigenfunctions with degenerate states must be diagonalized. The time to sum over states for $\beta$ scales as $M^2$, while for $\gamma$, it scales as $M^3$. The adaptation of the DL method for quantum graphs provides a fast, reliable way for researchers to solve quantum graph models that would otherwise be nearly intractable.
Another benefit of the DL method is that it requires only the ground state of the system. This implies that one can, in principle, attempt to model structures with large response by testing various models of charge distribution in the ground state of a molecule. In a real system, the Hamiltonian determines the exact eigenstates and spectra, and for a given set of boundary conditions, the ground state would be unique. For modeling a system, one could mock up a ground state based upon intuition about how the charge is distributed in a test molecule, and then compute $\beta$ and $\gamma$ using the DL method. The test molecule's actual Hamiltonian might not produce a ground state that exactly resembles the mocked up version, but the result should be meaningful.
The DL method has been solved for the dispersive case and is presented in a paper in this special issue. The extension of the DL method for graphs to include dispersion is identical to the presentation here, except that the DL functions take a different form that depends on the optical frequencies and the linewidths.
\section{Acknowledgements}
SMM and MGK thank the National Science Foundation (ECCS-1128076) for generously supporting this work.
|
\section{Introduction}
Hot B subdwarf (sdB) stars are helium core burning objects that populate
the so-called Extreme Horizontal Branch (EHB). They are expected to have
a mass around 0.47 $M_{\odot}$ and are characterized by a very
thin hydrogen-rich residual envelope containing at most
$\sim 0.02$ $M_{\odot}$. For this reason, they remain hot
and compact throughout all their helium core burning evolution, with
effective temperatures, $T_{\rm eff}$, and surface gravities, $\log g$,
ranging from 22\,000 K to 40\,000 K and from 5.2 to 6.2, respectively
\citep{he09,fo12}.
The presence of pulsations in some sdB stars make them good candidates
for probing their interior with the technique of asteroseismology.
A first group of nonradial sdB pulsators with periods of a few minutes
was theoretically predicted by \citet{ch96} and effectively discovered by
\citet{ki97}.
These pulsators, now referred to as the V361\,Hya stars,
show low-order, low-degree pressure ($p$-)modes that are driven by
a $\kappa$-mechanism induced by the partial ionization of iron-group
elements occurring in the "Z-bump" region and powered-up by radiative
levitation \citep{ch96,ch97}. Long period oscillations of $\sim1-4$ h
were later discovered by \citet{gr03}, forming another group of sdB
pulsators known as the V1093~Her stars. The latter show mid-order gravity
($g$-)modes driven by the same mechanism \citep{fo03}.
Hybrid pulsators that show both $p$- and $g$-mode oscillations simultaneously
have also been reported \citep[e.g., ][]{ss06}. Tight seismic constraints
have indeed been obtained from the measured frequencies using both types of
sdB pulsators, in particular based on high-quality photometric data gathered
from space-borned telescopes \citep[e.g., ][]{ch11b,van10}. However, the
reason behind the apparent variability of some oscillation modes in sdB stars,
already noticed from repeated ground based campaigns \citep[e.g., ][]{ki07},
has remained poorly understood.
The temporal variation of oscillation modes in pulsating sdB stars is
beyond the scope of the standard {\sl linear} nonradial stellar oscillation
theory in which eigenmodes have a stable frequency and amplitude \citep{un89}.
These behaviors must be studied within a {\sl nonlinear} framework to interpret
the modulations. In particular nonlinear resonant mode coupling effects
are expected to affect some oscillation modes, as noted, e.g., in the helium
dominated atmosphere white dwarf variable (DBV) star GD~358 \citep{go98}.
Different types of resonant coupling have been investigated within the
framework of the amplitude equation (AE) formalism since the 1980's,
among them the \threemodes{} resonance \citep{dz82,mo85} and the 2:1
resonance in Cepheid stars \citep{bu86}. The AE formalism was then
extended to nonadiabatic nonradial pulsations in Eulerian and Lagrangian
formulations by \citet{go94} and \citet{van94}, respectively.
A theoretical exploration of specific cases of nonradial
resonances was developed in \citet{bu95,bu97}, including notably the
resonance occurring in a mode triplet that is caused by slow stellar rotation
and which satisfies the relationship $\nu_+ + \nu_- \sim 2\nu_0$ ,
where $\nu_0$ is the frequency of the central $m=0$ component.
However, these theoretical developments based on AEs have since considerably
slowed down, in part due to the lack of clear observational data to rely on.
The launch of instruments for ultra high precision photometry from space has
changed the situation, making it now possible to capture amplitude and/or
frequency modulations occurring on timescales of months or even years
that were difficult to identify from ground-based observatories.
It is however from ground based data that \citet{va11} proposed for the
first time that resonant couplings within triplets could explain the
long-term variations, both in amplitude and frequency, seen in several
oscillation modes monitored in the GW\,Virginis pulsator PG\,0122+200,
through successive campaigns.
The observation of a multitude of pulsating stars, including sdB and white
dwarf stars, by the {\sl Kepler} spacecraft has open up new opportunities
to indentify and characterize the mechanisms that could modulate the
oscillation modes. {\sl Kepler} monitored a 105
deg$^2$ field in the Cygnus-Lyrae region for around four years without
interruption, thus obtaining unprecedented high quality photometric data
for asteroseismology \citep{gi10}. These uninterrupted data are particularly
suited for searching long-term temporal amplitude and frequency
modulations.
In the context of white dwarf pulsators, for instance, \citet[hereafter Z16]{zo16}
found that \dbv{} shows clear signatures of nonlinear effects attributed to
resonant mode couplings. In this star, three rotational multiplets show
various types of behaviors that can be related to different regimes
of the nonlinear resonant mode coupling mechanism. In particular some
amplitude and frequency modulation timescales are found to be consistant
with theoretical expectations. This finding suggests that the variations of
some oscillation modes in sdB stars may also be related to nonlinear
resonance effects. It is in this context that we decided to search clues of
similar nonlinear phenomena involving mode interactions in pulsating sdB stars.
Eighteen sdB pulsators have been monitored with {\sl Kepler}
(see \citealt{os14} and references therein). In this paper,
we focus on one of them, the star KIC~10139564, which was discovered
in quarter\,Q2.1 and then continuously observed from Q5.1 to Q17.2.
A preliminary analysis based on one month of short cadence data originally
showed that KIC~10139564 is a V361-Hya type (rapid, $p$-mode) sdB pulsator
featuring also a low-amplitude $g$-mode oscillation \citep{ka10}.
With extended data, \citet{ba12} detected up to 57 periodicities
including several multiplets attributed to the rotation of the star.
These multiplets are characterized by common frequency spacings, both for
the $p$- and $g$-modes, indicating that KIC~10139564 has a rotation period
of $25.6\pm1.8$\,d. These authors did not find any radial-velocity
variations from their dedicated spectroscopy and derived the atmospheric
parameter values $T_{\rm eff}=31~859$ K and $\log g=5.673$ for this star.
An interesting finding concerning KIC~10139564 is that two of the identified
multiplets may have degrees $\ell$ greater than 2, a possibility further investigated
by \citet{bo13}. The detection of several multiplets in this star
continuously monitored for more than three years makes it a target of choice
for studying eventual nonlinear resonant mode couplings in sdB stars.
In this study, we show that several multiplets in KIC\,10139564 have indeed
amplitude and frequency modulations suggesting nonlinear resonant mode
couplings, which constitutes the first clear-cut case reported for sdB pulsators,
so far.
In Sect.\,2, we present the thorough analysis of the frequency
content of the {\sl Kepler} photometry available on KIC\,10139564, including
our analysis of the frequency and amplitude modulations identified in several
multiplets and linear combination frequencies.
In Sect.\,3, we recall some theoretical background related to nonlinear resonant
mode couplings, focusing mainly on two types of resonances.
The interpretation of the observed modulations which may relate to nonlinear
resonant mode couplings is discussed in Sect.\,4. The summary and conclusion are
then given in Sect.\,5.
\begin{figure}
\includegraphics[width=8.5cm]{lc_a.eps}
\includegraphics[width=8.5cm]{lc_b_o.eps}
\caption{{\sl Top panel}: Condensed representation of the full
{\sl Kepler} light curve of KIC\,10139564 (Amplitude as the
residual relative to the mean brightness intensity of the star vs time in
Barycentric Julian Date) covering from Q5.1 to Q17.2
($\sim 1147.5$ days). {\sl Bottom panel}:
Close-up view showing 0.8 days of the {\sl Kepler} light curve
by slices of 0.08 days. At this scale the oscillations
are clearly apparent.
\label{lc}}
\end{figure}
\begin{figure*}
\includegraphics[width=17cm]{LS}
\caption{Lomb-Scargle Periodogram
(LSP; Amplitude in \% of the mean brightness vs frequency
in $\mu$Hz) of the {\sl Kepler} light curve for KIC\,10139564.
The represented range, up to the Nyquist frequency, covers the
long-period {\sl g}-mode and the short-period {\sl p}-mode frequency
domains. The region between the two dashed vertical lines at 5200 and
6400\,$\mu$Hz is where peaks have the largest amplitudes. However, weaker
peaks outside of this particular region are present and are
made visible by scaling up amplitudes by a factor of 20.
The dashed horizontal line represents the 5.6$\sigma$ detection
threshold (see text). Some well-known {\sl Kepler} instrumental
artefacts are present, but can easily be recognized.
\label{lsp}}
\end{figure*}
\section{The frequency content of KIC~10139564 revisited}
\subsection{The $Kepler$ photometry}
The pulsating sdB star KIC\,10139564 was observed by
{\sl Kepler} in short-cadence (hereafter SC) mode
during quarter Q2.1 and from Q5.1 to Q17.2 (i.e., until the
spacecraft finally lost its second inertia reaction wheel and stopped its
operations). Results based on parts of these data have already been
published in the literature \citep[e.g., ][]{ba12,bo13}.
we obtained the light curves through the $Kepler$ Asteroseismic
Science Consortium (KASC)\footnote{htpp://astro.phys.au.dk/KASC}.
These data were processed through the standard {\sl Kepler}
Science Processing Pipeline \citep{je10}.
For our purposes, we do not further consider the "short"
(one month) light curve of Q2.1 which is well disconnected from the main
campaign and would introduce a large and detrimental gap for our
upcoming analysis. This leaves us with a nearly contiguous 38-month
light curve starting from BJD~2~455~276.5 and ending on BJD~2~456~424
(which spans $\sim$ 1~147.5 days), with a duty circle of $\sim 89\%$.
The full light curve was constructed from each quarter
"corrected" light curves, which most notably include
a correction of the amplitudes taking into account contamination
by nearby objects (this correction estimates that $\sim 83.2\%$ of the
light comes from KIC\,10139564). Each quarter light curve
was individually detrended to correct for residual drifts by performing a
sixth-order polynomial fit. Then, data points that differ significantly
from the local standard deviation of the light curve were removed by
applying a running 3$\sigma$ clipping filter. Note that the latter operation
decreases slightly the overall noise level in Fourier space, but has
no incidence on the measured frequencies.
The fully assembled light curve of KIC\,10139564 is shown in the top panel
of Fig.\,\ref{lc} while the bottom panel expands a 0.8-day portion
of the data. Low-amplitude multi-periodic oscillations dominated by
periodicities of a few minutes are clearly visible. Their presence is
confirmed in the corresponding Lomb-Scargle Periodgram
(LSP, Fig.\,\ref{lsp}; \citealt{sc82}). The LSP shows two distinct regions
with significant signal corresponding to $p$-modes at high frequencies and
$g$-modes at low frequencies. This identifies KIC\,10139564 as a hybrid
pulsating sdB star \citep{ss06} whose oscillations are however largely
dominated by $p$-modes. The formal frequency resolution achieved with these
data is $\sim0.010$ $\mu$Hz.
\subsection{Frequency extraction}
A dedicated software, \felix{} (Frequency Extraction for LIghtcurve
eXploitation) developed by one of us (S.C.), was used to first extract
the frequency content of KIC\,10139564 down to a chosen detection
threshold. The latter was established following the same method as in Z16
(see their Sect.\,2.2), leading also in the present case to a conservative
5.6$\sigma$ criterion (in practice, we searched down to $\sim 5\sigma$ if
a frequency is suspected to be part of a multiplet; see below).
The extraction method is a standard prewhithening and nonlinear least
square fitting technique \citep{de75}, which works with no difficulty in
the present case. The code \felix{} greatly accelerates and eases
the application of this procedure, especially for treating very long time-series
obtained from space with, e.g., CoRoT and {\sl Kepler} \citep{ch10,ch11b}.
\begin{table*} \caption[]{List of frequencies detected in KIC\,10139564
on which we focus our analysis.}
\begin{center}
\begin{tabular}{cccccccccrl}
\hline
\hline
Id. &Frequency &$\sigma_f$ & Period &$\sigma_P$&Amplitude&$\sigma_A$&Phase&$\sigma_\mathrm{Ph}$&S/N &$^\dagger$Comment \\
& ($\mu$Hz) &($\mu$Hz) &(s) &(s) &(\%) &(\%) \\
\hline
&&\\
\multicolumn{3}{l}{Multiplet frequencies:}\\
$f_{39}$ & 315.579243 & 0.000566 & 3168.776214 & 0.005687 & 0.005851 & 0.000596 & 0.2492 & 0.0516 & 9.8 & $T_{2,-1}$\\
$f_{21}$ & 315.820996 & 0.000219 & 3166.350599 & 0.002193 & 0.015155 & 0.000596 & 0.6107 & 0.0199 & 25.4 & $T_{2,0}$\\
$f_{11}$ & 316.066440 & 0.000070 & 3163.891744 & 0.000702 & 0.047276 & 0.000596 & 0.2063 & 0.0064 & 79.3 & $T_{2,+1}$\\
&&\\
$f_{27}$ & 394.027385 & 0.000342 & 2537.894669 & 0.002202 & 0.009667 & 0.000594 & 0.2589 & 0.0312 & 16.3 & $D_{1, 0}$\\
$f_{32}$ & 394.289823 & 0.000397 & 2536.205455 & 0.002555 & 0.008323 & 0.000594 & 0.5123 & 0.0363 & 14.0 & $D_{1, +1}$\\
&&\\
$f_{34}$ & 518.900359 & 0.000437 & 1927.152262 & 0.001624 & 0.007526 & 0.000592 & 0.6648 & 0.0401 & 12.7 & $T_{3, -1}$\\
$f_{28}$ & 519.151796 & 0.000352 & 1926.218898 & 0.001305 & 0.009351 & 0.000592 & 0.9059 & 0.0323 & 15.8 & $T_{3, 0}$\\
$f_{31}$ & 519.402391 & 0.000367 & 1925.289559 & 0.001360 & 0.008964 & 0.000592 & 0.5369 & 0.0337 & 15.2 & $T_{3, +1}$\\
&&\\
$f_{08}$ & 5286.149823 & 0.000053 & 189.173601 & 0.000002 & 0.064784 & 0.000614 & 0.6712 & 0.0047 & 105.4 & $Q_{1,-2}$\\
$f_{10}$ & 5286.561766 & 0.000060 & 189.158861 & 0.000002 & 0.057105 & 0.000614 & 0.4356 & 0.0053 & 92.9 & $Q_{1,-1}$\\
$f_{07}$ & 5286.976232 & 0.000038 & 189.144032 & 0.000001 & 0.088857 & 0.000614 & 0.1202 & 0.0034 & 144.6 & $Q_{1, 0}$\\
$f_{05}$ & 5287.391879 & 0.000019 & 189.129163 & 0.000001 & 0.179339 & 0.000615 & 0.3374 & 0.0017 & 291.8 & $Q_{1,+1}$\\
$f_{06}$ & 5287.805883 & 0.000029 & 189.114355 & 0.000001 & 0.119329 & 0.000615 & 0.7941 & 0.0025 & 194.2 & $Q_{1,+2}$\\
&&\\
$f_{22}$ & 5410.701146 & 0.000234 & 184.818931 & 0.000008 & 0.014871 & 0.000627 & 0.9524 & 0.0203 & 23.7 & $M_{1, 0}$\\
$f_{67}$ & 5411.143448 & 0.000958 & 184.803824 & 0.000033 & 0.003637 & 0.000627 & 0.4591 & 0.0830 & 5.8 & $M_{1, 0}$\\
$f_{13}$ & 5411.597301 & 0.000136 & 184.788325 & 0.000005 & 0.025636 & 0.000627 & 0.6770 & 0.0118 & 40.9 & $M_{1, 0}$\\
$f_{15}$ & 5412.516444 & 0.000185 & 184.756944 & 0.000006 & 0.018812 & 0.000627 & 0.8925 & 0.0160 & 30.0 & $M_{1, 0}$\\
$f_{12}$ & 5413.389096 & 0.000084 & 184.727161 & 0.000003 & 0.041339 & 0.000627 & 0.4037 & 0.0073 & 65.9 & $M_{1, 0}$\\
$f_{19}$ & 5413.814342 & 0.000222 & 184.712651 & 0.000008 & 0.015718 & 0.000627 & 0.7225 & 0.0192 & 25.1 & $M_{1, 0}$\\
&&\\
$f_{01}$ & 5760.167840 & 0.000005 & 173.606052 & $\dots$ & 0.825132 & 0.000761 & 0.0744 & 0.0004 &1084.9 & $T_{1,-1}$\\
$f_{03}$ & 5760.586965 & 0.000008 & 173.593421 & $\dots$ & 0.554646 & 0.000761 & 0.6388 & 0.0005 & 729.3 & $T_{1,0}$\\
$f_{02}$ & 5761.008652 & 0.000007 & 173.580715 & $\dots$ & 0.567034 & 0.000761 & 0.5845 & 0.0005 & 745.5 & $T_{1,+1}$\\
&&\\
\multicolumn{3}{l}{Linear combination frequencies $C_1$:}\\
$f_{23}$ & 6076.234996 & 0.000252 & 164.575597 & 0.000007 & 0.014360 & 0.000650 & 0.7906 & 0.0210 & 22.1 & $f_{11}+f_{01}$ \\
$f_{35}$ & 6076.408232 & 0.000510 & 164.570905 & 0.000014 & 0.007091 & 0.000650 & 0.7821 & 0.0426 & 10.9 & $f_{21}+f_{03}$ \\
$f_{74}$ & 6076.650684 & 0.001120 & 164.564338 & 0.000030 & 0.003225 & 0.000650 & 0.5520 & 0.0937 & 5.0 & $f_{11}+f_{03}$ \\
\hline
\end{tabular}
\end{center}
\label{t1}
\tablefoot{
\tablefoottext{$\dagger$}{The first subscript is the identity of the multiplet and the second one indicates the value of $m$. The $m$-values for the $\ell>2$
multiplet $M_1$ are not provided, as the degree $\ell$ is not known.}
}
\end{table*}
We provide in Table~\ref{t2} (see Appendix) a list of all the extracted
frequencies with their fitted attributes (frequency in $\mu$Hz, period in second,
amplitude in percent of the mean brightness, phase relative to a reference
time $t_0$, and signal-to-noise ratio) and their
respective error estimates ($\sigma_f$, $\sigma_P$, $\sigma_A$, and
$\sigma_\mathrm{Ph}$). For convenience, because in this study we focus on
a particular subset of the observed frequencies, we repeat some of these
information in Table~\ref{t1} for the relevant modes. The "Id." column in both
tables uniquely identify a detected frequency with the number indicating the
rank by order of decreasing amplitude.
We have detected 60 clear independent frequencies that comes out well
above the 5.6$\sigma$ detection threshold (Table~\ref{t2}), of
which 29 frequencies consist of three triplets, one doublet, one
quintuplet and two incomplete multiplets with $\ell>2$ (Table~\ref{t1}).
We also detect another three frequencies that appear as significant but are
linked to other frequencies through linear combinations.
Five additional "forests" of frequencies, each containing many close
peaks in a very narrow frequency range, are detected in the 5400--6400\,$\mu$Hz
region. We also prewhitened 14 frequencies whose amplitudes are above
5.0$\sigma$ but below 5.6$\sigma$ which, we suspect, are real pulsations.
Our well-secured extracted frequencies agree well with the
independent analysis of \citet{bo13}, but we detect a few more low-amplitude
frequencies because the data that we consider here cover about one more year.
We do not investigate further these "forests" of frequencies
(G1--G5, see Table~\ref{t2} in Appendix) that show very complicated
structures. These were
discussed in \citet{ba12}. We point out that our extracted frequencies
may differ in amplitude compared with the work of \citet{bo13} because
some of these frequencies have variable amplitudes.
\begin{figure*}[htbp]
\centering
\includegraphics[width=8.5cm]{econ}
\includegraphics[width=8.5cm]{erand}
\caption{{\sl Left panel}: 2-D~distribution of the frequency and amplitude
deviations between the prewhitened and the injected values for 1\,000
artificial modes of constant amplitude. S/N denotes the signal-to-noise
ratio of the injected signals and the deviations have been normalized by
the 1$\sigma$ error, $\sigma_A$ and $\sigma_f$, derived from the
prewhitenning procedure implemented in the code \felix{}.
The 2-D distribution is also projected into 1-D histograms
(frequency and amplitude) to be compared with the Normal Distribution,
$\mathcal{N}(0,1)$ plotted as a red solid curve.
{\sl Right panel}: Same as above but for 1\,000 modes with random amplitudes.
The injected modes are divided into three groups of S/N ratios in the
ranges $[5,15]$, $(15, 25]$, and $(25, 60]$, respectively (represented by
three different colors and symbols).
\label{afd}}
\end{figure*}
\subsection{Error estimates on frequencies and amplitudes}
Before proceeding further in our analysis, we briefly discuss our
quantitative evaluation of the uncertainties associated with the measured
frequencies and amplitudes given in Table~\ref{t1} and Table~\ref{t2}.
The reliability of these error estimates is particularly important when it
comes to discuss amplitude and frequency variations with time, in particular
to assess if these are significant or not.
With \felix{}, errors are estimated following the formalism proposed by
\citet{mont99}, with the particularity, however, that $\sigma_A$, the error
on the amplitude of a mode, is measured directly in the Lomb-Scargle
periodogram. A window around each frequency is chosen and the median
value of the amplitudes in that frequency range defines $\sigma_A$. The
relations given in \citet{mont99} are then used to compute the other errors,
in particular $\sigma_f$, the error on the measured frequency. In order to
test that this procedure is correct and does not largely under or
overestimate the true errors, we conduct two Monte-Carlo experiments.
We first construct an artificial light curve covering about 200 days (similar
to the time baseline of each light curve pieces considered in the next subsection)
with the same SC-mode sampling provided by {\sl Kepler} in which we add white
random gaussian noise. We further inject in this light curve 1\,000 sinusoidal signals
with the same amplitude (S/N~$\sim160$) but of frequency increasing by steps of
$\sim8.2\,\mu$Hz per signal. {In practice, a random frequency shift of a few tenth
$\mu$Hz is performed on each injected frequency in order to reduce the number
of harmonics and linear combinations.}
The generated time series is then analysed with our code \felix{} that
extracts and measures each signal and evaluates the uncertainties associated
to the measured frequencies and amplitudes ($\sigma_A$ and $\sigma_f$).
Since the true values of these quantities are perfectly known from the
signals we injected, the real distribution of the deviations between measured
(prewhitenned) values and true values can be evaluated. For that purpose,
we define the frequency and amplitude deviations normalized by their
1$\sigma$ errors (as estimated with the code \felix{} from the procedure
described above), $\Delta{}f = (f_{\rm pre} - f_{\rm inj})/\sigma_f$ and
$\Delta{}A = (A_{\rm pre} - A_{\rm inj})/\sigma_A$, where the subscripts indicate
the prewhitened value and the injected one, respectively. A variant of this
test is also performed by again injecting 1\,000 sinusoidal signals, but this
time with random amplitudes (instead of constant ones) chosen in the
S/N $\in (5, 60)$ range. This second test allows us to check also the
reliability of our error estimates as a function of amplitude,
considering that $\sigma_f$ in particular depends on the mode S/N ratio
($\sigma_f$ increases when S/N decreases).
Figure\,\ref{afd} shows the results obtained in both cases. The 2-D
distributions of the frequency and amplitude deviations are well confined
within 3$\sigma$. Moreover, the associated 1-D histograms show that
for both quantities, the measured deviations closely follow the Normal
Distribution, $\mathcal{N}(0,1)$, plotted as a red solid curve. Only a few
data points fall outside the $[-3\sigma, +3\sigma]$ range (within which
99.73\% of the measurements should be for the normal distribution,
$\mathcal{N}(0,1)$). This is the behavior we expect for an accurate
determination of the error estimates, $\sigma_A$ and $\sigma_f$, with
the code \felix{}. Hence, these tests demonstrate that error values
derived in our frequency analysis are robust.
\subsection{Amplitude and frequency modulations}
From now on, we concentrate our discussion on the six multiplets, which
include three triplets $T_1$, $T_2$ and $T_3$, one doublet $D_1$, one
quintuplet $Q_1$ and a likely $\ell=4$ multiplet $M_1$ (see again
Table~\ref{t2}). Interestingly, three of these multiplets
($T_2$, $D_1$, and $T_3$) involve $g$-modes, while the others
($Q_1$, $M_1$, and $T_1$) are $p$-modes. We also examine three linear
combination frequencies ($C_1$). The fine structures of the multiplets
are shown in the left top
panels of Figs.\,\ref{p1}, \ref{g1}, \ref{g3} and \ref{g2}-\ref{s1}.
The average frequency spacing between the components of these well-defined,
nearly symmetric multiplets is $\sim 0.25$\,$\mu$Hz for the $g$-modes and
0.423\,$\mu$Hz for the $p$-modes, thus suggesting that the $g$-modes are
dipoles ($\ell=1$) in a star rotating rigidly with a period of $\sim 26$ days.
\begin{figure*}
\centering
\includegraphics[width=15.5cm]{p}
\caption{Frequency and amplitude modulations in the $T_1$ {\sl p}-mode triplet near
5760\,$\mu$Hz. {\sl Top-left panel} presents the fine structures of the well defined
triplet with near symmetric frequency spacings. The dashed horizontal line in red
represents the 5.6\,$\sigma$ detection threshold.
{\sl Middle-left panel} shows the sliding Lomb-Scargle Periodogram
(sLSP giving the amplitude in \% as a function of frequency in $\mu$Hz and time
in days) of the triplet as a whole.{\sl Bottom-left panel} shows
expanded views around the average frequency (the solid vertical
lines, also in the middle left panel) of each component,
obtained from prewhitening subsets
of the data, thus measuring precisely the frequencies, as a function
of time. {\sl Right panel} provides the measured amplitudes
as a function of time obtained for each subset of data
(see text for details). Note that the errors for each measurement is
smaller than the symbol itself.\label{p1}}
\end{figure*}
\begin{figure}
\includegraphics[width=8cm]{P1_mod_am.eps}
\caption{Frequency modulations after removing the long-term
trend in the $T_1$ triplet by applying a second-order polynomial
fit. The
solid curves represent the best fits of one pure {\sl Sine}
wave to the frequency modulations. The associated formal
errors for the periods and phases are also estimated.
\label{pr1}}
\end{figure}
\begin{figure*}
\begin{center}
\includegraphics[width=15.5cm]{g1}
\end{center}
\caption{Same as Fig.\,\ref{p1} but for the $T_2$ {\sl g}-mode
triplet near 316\,$\mu$Hz.
\label{g1}}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=15.5cm]{g3}
\end{center}
\caption{Same as Fig.\,\ref{p1} but for the $T_3$ {\sl g}-mode
triplet near 519\,$\mu$Hz.
\label{g3}}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=14cm]{3sigma.eps}
\caption{Frequency and amplitude scattering for the three components
forming the {\sl g}-mode triplet $T_3$ around their averaged values.
The red and blue dashed rectangles indicate the 2$\sigma$
and 3$\sigma$ error boxes, respectively. All the data points are within
3$\sigma$.
\label{sg3}}
\end{figure*}
In order to investigate the time variability of these oscillation modes
and their relationships, we used our software \felix{} to
compute sliding Lomb-Scargle Periodograms (sLSP) of the
data set. This method constructs time-frequency diagrams by
filtering in only parts of the data set as a function of time.
We chose a filter window of 200-day width moved along the entire light
curve by time steps of 20 days. This ensures a good compromise, for our
purposes, between the frequency resolution (to resolve close
structures of peaks in each LSP), time resolution, and signal-to-noise.
The sLSP offers an overall view of the amplitude and frequency variations
that may occur for a given mode (see, e.g., the middle left panel of
Fig.\,\ref{p1}). As a complementary (and more precise) technique, we also
extracted the frequencies (through prewhitening and nonlinear least square
fitting) in various parts of the light curve. The
38-month light curve of KIC\,10139564 was divided into 32 time
intervals, each containing 9 months of {\sl Kepler} data (for
the purposes of precision in the measurements) except for the
last 3 intervals at the end of the observations. This second
approach provides a measure of the (averaged) frequencies and
amplitudes at a given time, along with the associated errors (see,
e.g., the right and bottom left panels of Fig.\,\ref{p1}).
\begin{figure*}
\sidecaption
\includegraphics[width=12cm]{g2}
\caption{Same as Fig.\,\ref{p1} but for the $D_1$ {\sl g}-mode
doublet near 394\,$\mu$Hz. The solid curves in right panel show
the best fit of a pure {\sl Sine} wave to the amplitude modulations.
\label{g2}}
\end{figure*}
Figure\,\ref{p1} shows the amplitude and frequency modulations
for the three components forming the triplet $T_1$ near 5760\,$\mu$Hz.
As mentioned already, the top-left panel shows the triplet as revealed
by the full data set with components nearly equally spaced in frequency.
We note, however, that this spacing is not strictly symmetric, with a
difference (or "frequency mismatch") of 0.0026\,$\mu$Hz.
Frequency variations with time are illustrated by the sLSP diagram
in the middle-left panel where the color scale represents the amplitude of
the modes. An expanded view centered on the average frequency of each
component is then provided in the bottom-left panel while the amplitude
behavior with time for each component is shown in the right panel. The
latter two are obtained from prewhitening parts of the light curve as
described above.
From the sLSP diagram, we find that both the
amplitudes and frequencies have varied during the {\sl Kepler}
observations. These variations are more clearly seen in the
bottom-left and right panels. The side components both show
suggestions of a quasi-periodic modulation in frequency and
evolve in antiphase. We also note a long timescale trend as the frequencies
of the two side components gradually approach toward each other, as well as
toward the central component. In order to filter out these trends, we
applied a parabolic fit to each component, leaving the remaining
signature of the quasi-periodic modulation of the frequencies
(see Fig.\,\ref{pr1}). In the process, we find that the two side components
had frequencies about 0.06\,$\mu$Hz closer to each other at the end of the
run compared to the beginning of the observations. Figure\,\ref{pr1} shows
that the data almost cover two cycles of the quasi-periodic frequency variations.
While clearly not strictly sinusoidal, although not very far from it, if we
fit the closest pure sine wave to each curve, we find that all have a very similar
(quasi-)period of $\sim 570$ days.
The variations for the side components (retrograde and prograde modes)
are clearly in antiphase.
For the amplitude variations, we also find suggestions of a quasi-periodic
modulation for the central and prograde components. The retrograde mode
for its part has a more regular amplitude evolution (increase) during the
course of the observations.
Figure\,\ref{g1} illustrates the amplitude and frequency modulations
for the {\sl g}-mode triplet $T_2$ near 316\,$\mu$Hz using the same
presentation as in Fig.\,\ref{p1}. In this case, the triplet shows a
very small (but significant) asymmetry of 0.0036\,$\mu$Hz.
The frequencies appear to be stable over the 38-month {\sl Kepler}
observations. The amplitude is essentially constant for the retrograde
($m=-1$) mode while the other two components display some variations.
The central
one may show a small oscillatory behavior, but more precise measurements
would be needed to really confirm that trend. The prograde ($m=1$) mode has
its amplitude rising continuously throughout the observations, from 400\,ppm
up to about 600\,ppm.
The amplitude and frequency variations of the {\sl g}-mode
triplet $T_3$ near 519\,$\mu$Hz are shown in Fig.\,\ref{g3}.
In this triplet, which is almost perfectly symmetric, the three
components have stable frequencies and amplitudes within the quoted
uncertainties. This stability is further illustrated with Figure\,\ref{sg3}
that shows the scattering of the measured frequencies and amplitudes for these
modes from all the data chunks considered throughout the entire light curve.
Almost all measurements are indeed confined within 2$\sigma$ around their
average values (and all are within 3$\sigma$). It is interesting to note that
the triplet $T_3$ therefore has different characteristics compared to the
two triplets $T_1$ and $T_2$.
Figure\,\ref{g2} shows the amplitude and frequency modulations
for the {\sl g}-mode doublet $D_1$ near 394\,$\mu$Hz. The frequencies
of each component forming the doublet appear to be stable over the
38-month {\sl Kepler} observations, while the amplitudes show very suggestive
indications of quasi-periodic variations. We find that the amplitude
modulations of the two components have very similar periods, about 1100
days, as illustrated by the best-fit sine waves to the data (The red
solid curves in right panel of Fig.\,\ref{g2}). Hence the available {\sl Kepler}
data just cover about one cycle of this variation, but it is remarkable that
almost all the amplitude measurements match very closely the fitted sine
curves. This estimated period is almost twice the period of modulations
occurring in the main triplet $T_1$.
Moreover, we clearly see that the amplitudes of the two
components evolve almost in antiphase during the observing run.
Figure\,\ref{q1} shows the amplitude and frequency modulations
for the {\sl p}-mode quintuplet $Q_1$ near 5287\,$\mu$Hz. In
this complete quintuplet, the $m=\pm 2$ modes and possibly the $m=0$ central
component show significant frequency variations. The other modes,
with $m=\pm 1$, have frequencies which are rather stable (with only marginal
fluctuations) over the entire observation run. In constrast, the amplitudes
for all the modes in the quintuplet vary with patterns that cannot clearly be
connected to periodic modulations, based on the available data. Of course,
quasi periodic modulations with a timescale longer than twice the present
{\sl Kepler} observation cannot be ruled out.
We also note that the frequency variations of the $m=-2$ and $m=+2$ components
and the amplitude modulations of the $m=-2$ and $m=-1$ components are roughly
in antiphase during the observation.
\begin{figure*}
\begin{center}
\includegraphics[width=17.5cm]{q}
\end{center}
\caption{Same as Fig.\,\ref{p1}
but for the $Q_1$ {\sl p}-mode quintuplet near 5287\,$\mu$Hz.
\label{q1}}
\end{figure*}
Figure\,\ref{s1} shows the amplitude and frequency modulations
for the $\ell>2$ {\sl p}-mode multiplet $M_1$ near 5413\,$\mu$Hz.
The $\ell$-value for this group of modes is not clearly assessed yet,
but a plausible interpretation is that it corresponds to
an $\ell=4$ nonuplet \citep{ba12} with three undetected components and
one component barely visible in the LSP of the full data set (see top-left
panel of Fig.\,\ref{s1}) but which is too low in amplitude to be studied in
subsets of the light curve.
Some of the frequencies and amplitudes of the five clearly visible modes in
this multiplet show significant variations during the 38 months of
{\sl Kepler} observation. In particular, the frequencies of the side
components drifted toward each other by $\sim 0.032$\,$\mu$Hz from the
beginning to the end of the run. This trend may be comparable to that
observed in the side components of the $T_1$ triplet (see Fig.\,\ref{p1}).
Moreover, the same phenomenon also occurs for some modes observed in the
long-period-dominated sdB pulsator KIC\,2697388 (M.D. Reed, private
communication).
In addition to the six multiplets discussed above, we identified other
possible multiplets in the data, such as six modes near 5571\,$\mu$Hz (see
Table\,\ref{t2} in Appendix),
in which however most of the components have amplitudes too low to be
well studied in a time-frequency analysis or by
prewhitenning shorter parts of the light curve. We therefore do not
consider them further in this work.
Beyond the multiplets generated by the rotation of the star, we also
focus on the interesting narrow frequency region near 6076\,$\mu$Hz
where three structures ($f_{23}$, $f_{35}$ and $f_{74}$)
show amplitude and frequency modulations as illustrated in Fig.\,\ref{lcf}.
These frequencies are in fact related to components of the
$T_1$ and $T_2$ triplets through linear combinations. We find that
$f_{23}\sim{}f_1+f_{11}$, $f_{35}\sim{}f_3+f_{21}$, and
$f_{74}\sim{}f_3+f_{11}$. Interestingly, we note that the frequency and
amplitude variation pattern of $f_{23}$ is similar to the variations
observed for the mode $f_1$. Similarities also exist between the variations
observed in $f_{35}$ and $f_2$. The peak $f_{74}$, for its part, shows a
rather large frequency variation covering up to $\sim0.1$\,$\mu$Hz
(the scales of these frequency variations are indicated by the green
shadowed region in the top panel of Fig.\,\ref{lcf}).
After this description of the various behaviors encountered, we concentrate,
in the following sections, on plausible theoretical interpretations for
the observed amplitude and frequency modulations.
\begin{figure*}
\begin{center}
\includegraphics[width=17.5cm]{s}
\end{center}
\caption{Same as Fig.\,\ref{p1}
but for the $\ell>2$ {\sl p}-mode multiplet $M_1$ near 5413\,$\mu$Hz.
Note that at least three components are missing in this multiplet
and the red vertical dashed lines indicate the expected position
for two of them.
\label{s1}}
\end{figure*}
\begin{figure*}
\sidecaption
\includegraphics[width=12cm]{lcff}
\caption{Frequency and amplitude modulations of a group of linear
combination frequencies $C_1$ near 6076\,$\mu$Hz. The
red dashed line indicates the 5.6$\sigma$ detection threshold.
The green shadowed areas in top panel represent the scales of
variation of these frequencies (see text for details).
\label{lcf}}
\end{figure*}
\section{Resonant mode coupling and amplitude equations}
In this section, we recall the most natural theoretical background to
understand the behavior of the modes forming the six multiplets induced by
stellar rotation. These are indeed prone to develop nonlinear interactions
through resonant mode couplings, which is the mechanism that we ultimately
support from our present analysis. But before moving forward in discussing
details on the nonlinear resonant coupling mechanism, we first rule out
several other possibilities as the cause of the observed modulations, such
as instrumental deffects, binarity, stochastically driven pulsations, or
stellar "weather".
Instrumental modulations can possibly occur, e.g., on a per quarter basis,
such as a slightly varying contamination from nearby stars that could
affect the amplitude of the modes. Such effects would however affect all
frequencies similarly, which is not what is observed with KIC\,10139564
where the modes show different types of behavior. Another effect related
to the instrument that could induce frequency and amplitude modulations
is the slight shift of the Nyquist frequency associated with the movement
of the {\sl Kepler} spacecraft in the Solar System barycentric reference
frame. Fortunately, the multiplets that we consider (with frequencies
below 5761\,$\mu$Hz), are far away from the Nyquist frequency limit
($\sim 8496$\,$\mu$Hz). Moreover, such well-structured nearly equally spaced
multiplets can obviously not be the mirror reflected frequencies of signals
occurring above the Nyquist limit \citep{ba12}.
The presence of orbiting companions around compact stars could also induce
frequency variations. However, these should occur in all frequencies and be
correlated in phase, such as in the sdB pulsator V391\,Peg \citep{si07}.
The variations that we find in several frequencies of KIC\,10139564 are
clearly not correlated in phase. In addition, radial velocity measurements
from spectroscopy do not show any significant variation, thus ruling out
the presence of a stellar companion \citep[][but for substellar objects,
a higher precision would be needed to exclude this possibility]{ba12}.
Stochastically driven pulsations by envelope convection have long been
observed in the Sun and solar-like stars. It has been claimed in the past
that stochastic oscillations could also occur in some sdB pulsators, based
on the observation that mode amplitudes could vary from season to season
\citep{ki10,re07}. \citet{os14} recently announced that stochastic pulsations
were found in the sdB star KIC\,2991276,
in which the amplitude and phase of the modes vary substantially and
irregularly on a timescale of a month. However, the mechanism responsible
for the oscillations in sdB stars, a well identified $\kappa$-effect
involving iron-group elements \citep{ch97,fo03}, is very different in nature
from the stochastic driving occurring in the convective envelope of
solar-like stars. Such a mechanism would indeed hardly be efficient in sdB stars
that have radiative envelopes, except may be for a very narrow convective
layer generated by the accumulation of iron in the Z-bump region (the latter
being however extremely weak).
Beyond these theoretical considerations, we find in
the case of KIC\,10139564 that several mode behaviors, e.g., the
frequencies in the triplet $T_1$ and the amplitudes in the doublet $D_1$,
show correlations that would be difficult to account for with a stochastic
driving mechanism and that essentially rule out this interpretation.
Changes in the background physical state of the star such as possibly
induced by magnetic cycles could also be invoked for explaining amplitude and
frequency modulations. Magnetic cycles indeed have an impact on the frequencies of
$p$-modes observed in the Sun and lead to small frequency drifts that
correlate in time with tracers of the solar surface activity
(see, e.g., \citealt{sa15} and references therein). However, there is no
clear observational evidence of stellar activity on the surface of
sdB stars which, again, have very stable radiative envelopes and are not
known to be magnetic. Cycles comparable to those observed in solar-like
stars are therefore unlikely to be found in sdB stars. Moreover, such a
mechanism, or more generally a phenomenon modifying the physical state of
the star on a timescale of months could hardly account for the observed
modulations in KIC\,10139564 that show very different modulation behaviors
from mode to mode, while a global change in the star would affect all modes
similarly. Consequently, we also rule out this possibility in the present
case.
\begin{figure}
\centering
\includegraphics[width=7.5cm]{LC1_am_mod}
\includegraphics[width=7.5cm]{LC1_fr}
\includegraphics[width=7.5cm]{LC1_am_rmc}
\caption{Amplitude and frequency variations of the linear
combination frequency $f_{23}=f_1+f_{11}$.
{\sl Top panel}: Measured amplitudes as a function
of time obtained from each data subset (using the same method as for
multiplets).
{\sl Middle panel}: Measured frequencies from each
data subset compared with the frequency sum $f_1 + f_{11}$, both as
a function of time.
{\sl Bottom panel}: Observed amplitudes of $f_{23}$ vs
predicted amplitudes of $R \times$ the product of $f_1$
and $f_{11}$ amplitudes (see Eqn\,6 for the definition of $R$).
In both cases (frequency and amplitude comparisons), the
measurements are found to be within 1$\sigma$.
\label{lc1}}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=7.5cm]{LC2_am_mod}
\includegraphics[width=7.5cm]{LC2_fr}
\includegraphics[width=7.5cm]{LC2_rmc}
\caption{Same as Fig.\,\ref{lc1} but for the linear combination
frequency $f_{35}=f_3+f_{21}$. Note that there are 10 missing measurements
for $f_{21}$, including the first 9 data points, because the
amplitudes were lower than 4$\sigma$. The last data
point is also not shown because of a large associated error, this measurement
being at the end of the data set.
\label{lc2}}
\end{figure}
\subsection{Triplet resonance induced by slow stellar rotation}
We hereafter propose that nonlinear resonant coupling mechanisms could
be a natural explanation for the observed modulations in KIC\,10139564.
Resonant interactions between modes may indeed result in amplitude and
frequency variations occurring on timescales of weeks, months, and
even years.
In the present context, we limit ourselves to the type of resonances
described in \citet{bu95,bu97} involving linear frequency combinations
$\nu_1\,+\nu_2\,\sim\,2\,\nu_0$. More specifically, we focus on a
particular case where dipole ($\ell=1$) modes are split by slow rotation
and form a nearly symmetric triplet (thus following the above relationship
between the frequencies of the components). We also consider the three-mode
couplings of the form $\nu_1\,+\nu_2\,\sim\,\,\nu_0$, which corresponds
to the so-called direct resonances or parametric instabilities
\citep{dz82,wu01b}.
We first recall some basic theoretical background relative to resonances in
mode triplets created by stellar rotation. We emphasize that our
focus on this particular mechanism is obviously motivated by the specific
configuration of the modes observed in KIC\,10139564, most of which are
identified as $\ell=1$ rotationally split triplets. We also point out that
this type of nonlinear resonance has recently been strongly suggested to
explain the modulations of the {\sl g}-modes triplets in the DB white dwarf
KIC\,08626021 (Z16). It is therefore the most natural effect that one could
think of in the present case.
The AEs formalism could also, in principle, be extended to multiplets of
degree $\ell>1$ at the expense of solving a larger set of coupled amplitude
equations \citep{bu95}. However, such development has yet to be done,
which is beyond the scope of our present paper. The latter would be needed
for KIC\,10139564 in order to fully interpret the several multiplets
with $\ell>1$ that show variations.
The behavior for more complex $\ell>1$ multiplets may indeed differ from
the simpler (better documented) $\ell=1$ triplet case, although we expect
some similarities in general.
Details on the theory of nonlinear resonant couplings for three-mode
interactions, such as in $\ell=1$ triplet, can be found in
\citep[][Z16]{bu95,bu97}. We summarize below the most relevant aspects
(for our purposes) of the theory. In particular,
The quantity $\delta\nu$ (which we thereafter call the frequency asymmetry),
measuring the departure from exact resonance (that would occur if, e.g.,
triplets were perfectly symmetric), is in fact essential for driving the
various resonant mode coupling behaviors.
Contributions to the frequency asymmetry in a given triplet generally
involves higher order effects of stellar rotation on the pulsation
frequencies \citep{dz92,jo89}, but could also have additional origins,
such as the presence of a weak magnetic field\footnote{The asymmetry
would be proportional to the strength of magnetic field
$\lvert{}\vec{B}\lvert{}^2$ and the frequency of each component of the
triplet (except the central, $m=0$, one) would be shifted in the
same direction \citep{jo89}.}. We do not however consider further that
possibility since no evidence of significant magnetism exists for
sdB stars \citep{pe12,la12}.
The rotationally split frequencies up to the second order, which should be
the main contribution to the frequency asymmetry, are given by the formula
\begin{equation}
\nu_m-\nu_0 = (1-C_{k\ell})m\Omega + D_{k\ell}\frac{m^2\Omega^2}{\nu_0},
\end{equation}
where the $C_{k\ell}$ coefficient is the well-known first order Ledoux
constant, $D_{k\ell}$ involves a complex integration of the eigenfunctions
of the modes, and $\Omega\,=\,1/P_{\rm rot}$ is the rotation frequency of
the star (expressed in Hertz).
The value of $C_{k\ell}$ is typically $\sim 0.5$ for dipole $g$-modes
when approaching the asymptotic regime, while it is usually very small
($C_{k\ell} \ll 1$) for $p$-modes. The second order coefficient
$D_{k\ell}$ is roughly 4$C_{k\ell}$ for dipole $g$-modes \citep{dz92,go98}
but can vary significantly from one $p$-mode to another \citep{dz92,sa81}.
The rotation period of KIC\,10139564, $P_{\rm rot}$, can be estimated from the
average of the frequency separations between the components of the multiplets
using the first order approximation $\Delta\nu=(1-C_{k\ell})\Omega$.
We find $P_{\rm rot}\sim 26$\,days for KIC\,10139564 (see Sect. 2).
An "observed" frequency asymmetry can also be evaluated directly from the measured
frequency of each triplet component, simply from the relation
\begin{equation}
\delta\nu_o = \nu_- + \nu_+ - 2 \nu_0.
\end{equation}
We note at this stage that $\delta\nu_o$ may actually differ from asymmetries expected
from linear developments (such as discussed above) because nonlinear
effects can modify the frequencies of the modes.
The numerical solutions of the AEs for mode interactions in triplets
mainly reveal three distinct regimes of resonances (see, e.g.,
\citealp[][and in Z16]{bu97}). The first state is the "frequency lock"
regime where all the components in the triplet have constant frequencies
and amplitudes and the asymmetry tend to be zero (triplets become perfectly
symmetric). The opposite configuration is the nonresonant regime
where the triplet configuration is like predicted by the linear theory of
stellar oscillations. Between the two, there is an intermediate regime in
which all the modes in the triplet show modulated frequencies and
amplitudes which can be periodic, irregular, or even chaotic.
In order of magnitude, the occurrence of these three regimes is roughly
linked to a parameter $D$ defined as
\citep[see][]{go98}
\begin{equation}
D \equiv \frac{2\pi\delta\nu}{\kappa_0},
\end{equation}
where $\kappa_0$ is the linear growth rate of the $m=0$ mode in the triplet
(a nonadiabatic quantity). However, the ranges of values for this parameter
which define the different regimes depend somewhat on the values of the
nonlinear coefficients in the real star. $D$ is also a quantitative indicator
that measures how far the triplet modes are away from the exact resonance
center ($D=0$). We nonetheless summarize some of the properties encountered
in previous studies as a function of $D$ :
\begin{itemize}
\item In the frequency lock regime ($\delta\nu \rightarrow 0$),
the $D$-parameter roughly corresponds to values in the range $\sim 0-1$
according to the AEs formalism. However, in the case of the white dwarf
star GD358, \citet{go98} found that $D$ could be up to 20 and still
correspond to a frequency locked situation. These ranges, therefore,
are somewhat dependent on the specific properties of the mode being
considered, in particular on the scale of their linear growth rate,
$\kappa_0$.\\
\item The intermediate regime occurs when the triplet components move away
from the resonance center ($\delta\nu\ne0$). In this situation, if periodic
variations indeed affect the considered modes, these can be
expected to have a modulation timescale of
\begin{equation}
P_{\rm mod} \sim \frac{1}{\delta\nu}\simeq\frac{2\pi}{\kappa_0}\frac{1}{D},
\label{eqnpmod}
\end{equation}
i.e., roughly the timescale derived from the inverse of the linear (i.e.,
unperturbed) frequency asymmetry of the triplet dominated by the second
order effect of stellar rotation (following Eqn.\,2). This timescale is
also connected to the inverse of the growth rate of the oscillation mode
through the $D$ parameter \citep{go98}.\\
\item The modes recover a configuration of steady pulsations with the
nonresonant regime when the involved frequencies are such that the modes
are now far from the resonance condition ($D >> 1$). In this regime, the
nonlinear interaction between modes is very weak and nonlinear frequency
shifts become very small. Consequently, the mode frequencies are close to
the linear ones.
\end{itemize}
We finally point out that in addition to the above mentioned three main regimes,
a narrow hysteresis (transitory) regime exists between the frequency lock
and intermediate regimes in which the frequencies can be locked (i.e., constant),
while the amplitudes still have a modulated behavior \citep{bu97}.
\subsection{Three mode resonance of the type \threemodes{}}
In this section, we recall some properties of nonlinear interactions
between oscillations modes not within triplets but whose frequencies are
close to a resonance condition such that \threemodes{}.
Frequencies with such a relationship could also result from
simple linear combination frequencies, i.e., exact sum or difference of
frequencies (where the "child" frequency is not a true eigenmode),
which may be related to nonlinearities in the mixing process affecting the depth
of a convective zone in the outer layer of a pulsating star \citep{wu01a}, or
to nonlinearities in the flux response induced by the surface geometrical and
temperature distorsions triggered by the propagating waves \citep{br95}.
A useful quantity, $R$, connecting the observed amplitude of the combination
frequency and the amplitudes of its "parent" modes, has been defined as
\citep[][]{va00,wu01a}
\begin{equation}
R = \frac{A_0}{A_1\cdot{}A_2},
\end{equation}
where the $A_0$, $A_1$, and $A_2$ are the amplitudes of the frequencies
$\nu_0$, $\nu_1$, and $\nu_2$, respectively.
This ratio $R$ is typically less than 10 for simple linear combinations
related, e.g., to nonlinearities in the flux response. Consequently,
in pulsating sdB stars, the "child" frequency resulting from this effect
usually has a very low amplitude compared to its "parent"
frequencies. In the large amplitude and brightest known
pulsating sdB star, Balloon\,090100001, where such linear combination frequencies
have been unambiguously observed, the amplitude ratios are 3.9, 3.7, 3.0 and 5.5
for the linear combination frequencies of four {\sl p}-modes and one
{\sl g}-mode $f_1+f_2$, $f_1+f_3$, $f_1+f_4$ and $f_1-f_B$ in
$B$-band photometry \citep{ba08}, respectively. In the
present work, we however find that the identified linear combination
frequencies $C_1$ have amplitude ratios in the $10-100$ range, i.e.,
one order of magnitude larger than typical linear combination frequencies
observed so far. {One possible interpretation for the high amplitude ratios
is that the frequency sum/difference is near the resonance condition of
\threemodes{} and its amplitude is boosted significantly by the resonance
\citep[e.g.,][]{dz82,br14}}
The AEs formalism treating the \threemodes{} type of resonance, including
the parametric instability and the direct resonance (see below),
is similar to the case of a triplet resonance
\citep[e.g., see the amplitude equations in][Z16]{bu97}. According to \citet{dz82},
the three-mode interactions can be described by the following coupled system
\begin{subequations}
\begin{align}
\frac{d\vec{A_0}}{dt} & = \kappa_0\vec{A_0} + i\frac{q}{2\nu_0 I_0}\vec{A_1A_2}\exp(-i\delta\nu{}t), \label{seq61} \\
\frac{\vec{dA_1}}{dt} & = \kappa_1\vec{A_1} + i\frac{q}{2\nu_1 I_1}\vec{A_0A_2^*}\exp(-i\delta\nu{}t), \label{seq62} \\
\frac{d\vec{A_2}}{dt} & = \kappa_2\vec{A_2} + i\frac{q}{2\nu_2 I_2}\vec{A_0A_1^*}\exp(-i\delta\nu{}t). \label{seq63}
\end{align}
\end{subequations}
Where $\vec{A_j^*}$ is the complex conjugate of the amplitude $\vec{A_j}$
($\vec{A_j}=A_je^{i\phi_j}$),
$I_{j}$ is the mode inertia, and $\kappa_{j}$ is the linear growth rate
for the three involved frequencies. The quantity $q$ is a nonlinear
coupling coefficient, and $\delta{}\nu$ is the frequency mismatch relative to
pure resonance defined by the relationship $\delta{}\nu=\nu_0-\nu_1-\nu_2$.
The nonlinear equations (\ref{seq61}), (\ref{seq62}) and (\ref{seq63}) cannot be
solved by analytical methods, but solutions for the equilibrium state (all time derivatives
set to zero) can be obtained. In particular, the equilibrium solution leads to an amplitude
ratio
\begin{equation}
R = \frac{q}{2\nu_0\kappa_0I_0} \equiv \frac{A_0}{A_1\cdot{}A_2}.\label{nonratio}
\end{equation}
{The stability of the equilibrium-state depends on the growth (damping) rates and the
frequency mismatch \citep[e.g.,][]{dz82}.}
In a three-mode direct resonance, the child mode is damped and has a frequency
very close to the sum of frequencies of its two parent modes which are linearly driven
(unstable). The child mode amplitude is very sensitive to its mode inertia, linear
growth rate, and to the nonlinear coupling coefficient (see Eqn.\,\ref{nonratio}).
This near resonance mode can grow up to a very large amplitude if the
quantity $q/\kappa_0{}I_0$ is sufficiently large.
The coupling coefficient $q$ follows from a complex integration
of the coupled mode eigenvectors and its explicit form can be found in
\citet{dz82}. It may be possible, in principle, to calculate this
coefficient provided that the mode eigenfunctions are known. However,
this would require that a precise seismic solution is found for KIC\,10139564,
which still has to be obtained.
To test the theory of a three-mode direct resonance, we would also need to
know the linear damping (growth) rate and the inertia of the damped
mode, eventually corrected by the effect of slow stellar rotation
\citep[e.g, see][]{ca82}. {However, the situation could be simplified
in the case of unstable equilibrium state where the amplitude and
frequency of the child mode should exactly follow those of its parent modes,
even if the growth rates and coupling coefficients are unknown. Fortunately,
the equilibrium state of three-mode direct resonances seems always unstable
because the growth (damping) rates cannot satisfy the stability criteria
of the Hurwitz theorem \citep{dz82}. Therefore, each frequency and amplitude
measurement could be used as one independent test of these particular
nonlinear couplings. Furthermore, this also provides a method to
separate the child mode from their parent modes according to the amplitude
and frequency relationships \citep[e.g.,][]{br14}.}
The parametric instability is another form of three-mode resonant coupling
that could destabilize a pair of stable daughter modes from an
overstable (driven) parent mode \citep{wu01b}.
In this mechanism, the overstable parent mode gains energy through the
driving engine (a $\kappa$-mechanism in our case) and the two independent
damped child modes dissipate energy. This configuration would lead
the system to reach limit cycles under certain conditions
\citep[e.g., if $\delta\nu<\kappa_1$ or $\kappa_2$, ][]{wu01b,mo85}.
During such limit cycles, the amplitude of the parent mode first
increases slowly on a timescale of $\kappa_0^{-1}$, then decreases
rapidly on a timescale of $\kappa_{1,2}^{-1}$. A the same time, the amplitude
of the daughter modes follow the opposite behavior.
In sdB stars, we point out that the linear growth rate of the parent
mode $\kappa_0$ would be usually far smaller than the damping rate of
the daughter modes $\kappa_{1,2}$.
We further mention that the nonlinear interactions between the parent/child modes
would also affect their periods as a result of phase variations.
The nonlinear frequency shift could be of the order of a few $\mu$Hz
in some extreme conditions \citep{wu01b,mo85}. We point out that a
parametric instability can also occur in multiplets. In such
circumstances, different $m$ components that forms the multiplet may share
some common damped daughter modes. Having common daughter modes
involved in different parametric resonances, i.e., involving
different parent overstable components, will obviously induce more
complex dynamic modulations than simple periodic variations that could be
expected from pure three-mode only interactions.
We indeed point out that both the triplet resonance that was
explored by \citet{bu95,bu97} and the three-mode
\threemodes{} resonances that were investigated by \citet{dz82},
\citet{mo85} and \citet{wu01b} are treated as isolated systems,
i.e., assuming only interactions between the three involved modes and
ignoring the possible influence of other modes. Modes with the
highest amplitudes are more likely to efficently couple with different
resonances, e.g., in a multiplet resonance and in a \threemodes{}
resonance.
\section{Connections with mode behaviors seen in KIC\;10139564}
In light of the theoretical background summarized in the last section,
we tentatively interpret some of the behaviors described in Section 2 for
the frequencies listed in Table\,1. These indeed show striking similarities
with patterns expected for nonlinear resonant mode interactions that occur
in various regimes.
\subsection{Multiplets in the intermediate regime}
The first connection is for the modes belonging to multiplets that
show quasi-periodic amplitude and frequency modulations. In particular, the
$p$-mode triplet near 5760\,$\mu$Hz ($T_1$) shows indication that it could be evolving
within the so-called intermediate regime of a triplet resonance. We recall
(see Section 2; Fig.\ref{pr1}) that the frequencies of the two side components in
this triplet, besides showing a long term drift, vary quasi-periodically in
antiphase with a timescale of $\sim550$ days. The central component
of $T_1$, for its part, has a frequency modulation which also vary, possibly with
a slightly longer period of $\sim600$ days.
For comparison purposes, the modulation timescale is expected to be related to
the inverse of the linear (i.e., unperturbed) frequency asymmetry in the
triplet (see Eqn.\ref{eqnpmod}), which therefore should be $\delta\nu\sim 0.02$
$\mu$Hz. Assuming that this frequency asymmetry originaly comes from the second
order effect of slow rotation, and given the average rotation frequency of
the star ($\sim 0.42$ $\mu$Hz, corresponding to $\sim 26$ days), the $D_{k\ell}$
coefficient in Eqn.~1 can be estimated to $\sim 200$ for that mode. This value is
plausible because the $D_{k\ell}$ coefficient is found to vary over a large range
for dipole $p$-modes \citep{dz92,sa81}. However, to compute $D_{kl}$
in this specific case and compare with this value, a precise seismic solution
for KIC\,1013956 has to be worked out, but is not available yet.
It has to be noted that the frequency asymmetry measured from the averaged
frequencies given in Table\,1 is only 0.0026\,$\mu$Hz that is one order of
magnitude lower than the value derived from the modulation frequency
($\sim0.02$\,$\mu$Hz). We note, however, since the frequencies are varying with time,
that the maximum extent of the observed frequency asymmetry is $\sim0.02$\,$\mu$Hz
when considering the 33 measurements independently (see Fig\,\ref{p1}).
We point out that these observed values (0.0026\,$\mu$Hz on average and
$\sim0.02$\,$\mu$Hz for the maximum asymmetry) are similar to those observed
in the main triplet of \dbv{}, which is also in the intermediate regime (Z16).
Nonlinear resonant interactions are bound to perturb the linear
frequencies of the modes, forcing them in some cases to shift toward the
exact resonance (obtained when the system is locked). It is therefore not
surprising to observe a frequency asymmetry that can be significantly smaller
than the theoretical shift expected in the linear theory context.
In terms of amplitude modulations, the situation is bit less clear as only the
prograde component in $T_1$ shows a quasi-periodic modulation, with a
timescale of $\sim800$\,days, while, for the other two components, particularly
for the retrograde mode, their amplitude variations appear somewhat
irregular.
In addition to the frequency variations of $T_1$ discussed above,
we note that the three components that form this triplet feature a
regular drift toward each other which, if nothing change, would lead
them to merge into one frequency on a timescale of $\sim10$ years.
Such a merging is of course not conceivable and what we observe is more
likely a small fraction of a variation cycle occurring on a timescale
much longer than the duration of the {\sl Kepler} observations.
This suggests that the triplet resonance is probably not the only
mechanism that affects the stability of $T_1$. This added complexity
may also explain the more erratic behavior of the amplitude variations
in this triplet. Quite notably, we indeed find that all the components
of $T_1$ can be linked to other frequencies forming linear
combinations satisfying the conditions for a three-mode
resonance \threemodes{}. This will be further discussed in Sect.\,4.4.
The quintuplet $Q_1$ also shows components with amplitude
and frequency variations (see Fig.\,\ref{q1}) that may be
associated to the intermediate regime.
In this case, however, we cannot estimate timescales for the
modulations which appear to have a more complex behavior
than the modulations detected in the $T_1$ triplet or, if
we compare to other cases, in the triplets of the pulsating
DB star KIC\,08626021 (Z16).
The averaged frequency mismatch, $\delta\nu_o$, for $Q_1$ is
about 0.0018\,$\mu$Hz. This could either be the result of
the nonlinear coupling mechanism locking the modulated
components close to the exact resonance, even if they are
in the intermediate regime (see the case of $T_1$, as well
as the triplets in the DBV star KIC\,08626021), or it could
indicate that the modulation timescale for $Q_1$ is
$\sim 17.6$ yrs (the inverse to $\delta\nu_o$,
{if their amplitudes have a periodic behavior}).
As there has been no theoretical exploration of the nonlinear
five-mode interaction yet, the connection of $Q_1$ with the
intermediate regime is based on the assumption that nonlinear
five-mode interactions has also mainly three distinct regimes.
The coupled amplitude equations for the five-mode resonance
involve more terms in each AE and the numerical solutions
are more difficult to search for.
Another case may be connected with the intermediate regime:
the multiplet $M_1$ which shows amplitude and frequency
variations (see Fig.\,\ref{s1}). But, again, we cannot
determine any timescale for the complex modulations
occurring in this multiplet. In that case, there is also a
slow trend leading frequencies, particularly for the most side
components, to seemingly converge. This trend is very similar
to the slow variation observed in the $T_1$ triplet. It could
possibly be a fraction of a variation cycle with a much
longer timescale than the duration of the observations, but
more observations would be needed to confirm this hypothesis.
This multiplet $M_1$ should be the siege of even more complex
resonant coupling interactions than the quintuplet $Q_1$, since
there are six detected components, with at least three
components missing.
\subsection{Triplets in the transitory regime}
Another type of behavior encountered in our data can be linked to
the narrow transitory hysteresis regime which is between the frequency
lock and intermediate regimes. This state is characterized by stable
frequencies but varying amplitudes. This is notably observed in the
$g$-mode triplet $T_2$ (see Fig.\,\ref{g1}). For this triplet,
the observed frequency mismatch is about 0.0036\,$\mu$Hz,
i.e., very similar to the value measured for the $T_1$ triplet
(see left-top panel of Fig.\,\ref{p1} and \ref{g1}).
We also point out that $T_2$ may couple with the $p$-mode
triplet $T_1$ through a three-mode resonance \threemodes{},
as discussed in Sect.4.4.
The incomplete triplet $D_1$ may also be associated to this transitory
regime as it shows quasi-periodic amplitude modulations and
stable frequencies. Due to the missing component, we cannot measure
the frequency mismatch for this doublet. We note that the AEs
for the triplet resonance indicate that the modes cannot be stable,
i.e, there is no fixed-point solution, if one of the visible modes
forming the incomplete triplet is the central ($m=0$) component \citep{bu95}.
Thus, at this stage, we may just fail to detect either the
third component of the triplet whose amplitude may be lower than
the detection threshold (meaning that the triplet is indeed in the
narrow transitory regime), or the nonlinear modulation of the frequencies,
which may be smaller in amplitude than our current precision (meaning a
doublet in the intermediate regime, as predicted by the AEs).
\subsection{A triplet in the frequency lock regime}
The last case occurring in a different regime is the $g$-mode triplet
$T_3$, which shows stable amplitudes and frequencies (see Fig.\,\ref{g3}
and \ref{sg3}). This suggests that $T_3$ is in the configuration of the
frequency lock regime where the triplet approaches the resonance
center, i.e., $\delta\nu\rightarrow 0$ and both the frequencies and
amplitudes are constant. Indeed, we find that the observed frequency
asymmetry, $\delta\nu_o$, is 0.0008\,$\mu$Hz (Fig.\,\ref{g3}) for
$T_3$, which is less than the measured error 0.0011\,$\mu$Hz. The triplet
$T_3$ is therefore exactly (within measurement errors) at the resonance
center, contrary to $T_1$ and also $T_2$ which has constant frequencies
but a small non-zero frequency mismatch.
In summary, the various behaviors encountered in the multiplets detected
in KIC\,10139564 seem to cover all the different regimes expected in a
context of resonant mode coupling in multiplets. This mechanism is
therefore quite likely responsible, at least in part, for the observed
phenomena. In the following section, we discuss another type of
resonance, the \threemodes{} nonlinear interaction.
\subsection{Three-mode resonance}
In this section, we discuss the variations of a group of
frequencies $C_1$, including $f_{23}$, $f_{35}$ and $f_{74}$,
that are involved in a relationship \threemodes{}.
We find that the variations of these frequencies have
strong correlations with the variations of the components
in the triplet $T_1$ (see Fig.\,\ref{p1} and \ref{lcf}).
The large frequency variations first suggest that the $C_1$
frequencies correspond to three-mode resonances rather than
simple linear combination frequencies.
The result of prewhitening the frequencies $f_{23}$ and $f_{35}$
(using the same method as for the multiplets) is shown in
Fig.\,\ref{lc1} and \ref{lc2}, respectively.
Most of the amplitude and frequency measurements for $f_{23}$ and $f_{35}$
are exactly following the variation of amplitude and frequency of
the sums $f_1+f_{11}$ and $f_3+f_{21}$ within 1$\sigma$, respectively
(see in particular the middle and bottom panels of Fig.\,\ref{lc1} and
\ref{lc2}).
The amplitude ratio $R$ is 37 and 85 for $f_{23}\sim f_1+f_{11}$
and $f_{35}\sim f_3+f_{21}$, respectively. These values are significantly
higher than those observed for normal linear combination frequencies
in sdB stars (see the example given in Sect.\,3.2). There is also a possible
true linear combination frequency in KIC\,10139564 with the frequency
$f_{79}\sim f_1$\,-\,$f_4$ (see Table~2) which indeed has a very low
amplitude (signal-to-noise ratio of 5.1) and an amplitude ratio $R$ less
than one. Thus, we propose that there should be real pulsation modes near
the position of the linear combination frequencies $f_1+f_{11}$ and
$f_3+f_{21}$ and these modes had their amplitudes boosted through a
resonance. In the \threemodes{} resonance, the child mode indeed
follows the behavior of its parent modes (see, again, Fig.\,\ref{lc1}
and \ref{lc2} and examples provided by \citealt{br14}).
We note that another frequency, $f_{74}$, is also in the
region near 6076\,$\mu$Hz, but has an amplitude too low to be monitored
over time using the prewhitening technique on subsets of the data.
However, Fig.\,\ref{lcf} still clearly shows that this frequency is
varying smoothly during the observation, from $\sim6076.58\,\mu$Hz
(the first half part of the run) to $\sim6076.68\,\mu$Hz (the second
half part). We speculate that there is possibly a real mode, with a
frequency around $6076.58-6076.69\,\mu$Hz, which first interacts with
the frequency sum $f_2+f_{39}\sim6076.59\,\mu$Hz, then with the
frequency sum $f_3+f_{11}\sim6076.66\,\mu$Hz, because the influence
from the latter modes become stronger than the former ones during the last
half of the observation run, due to the amplitude of $f_3$ increasing
significantly in the second half of the {\sl Kepler} time series.
This, again, suggests that the $C_1$ frequencies are really part of
\threemodes{} resonances instead of being simple linear combination
frequencies.
All of the involved frequencies, $f_{1, 2, 3, 11, 21, 39}$,
are the components of the triplets $T_1$ and $T_2$. They are expected
to be overstable (driven) modes, thus meaning that they are involved in
a three-mode direct resonance and not a parametric resonance which
involves one overstable parent mode and two damped unstable daughter modes.
At this stage, it becomes natural to interpret the complex variations
observed in the components of $T_1$ and $T_2$ to be linked with the fact
that these modes are simultaneously involved in two different types of
resonances, i.e., a triplet resonance and \threemodes{} direct resonances.
In this situation, the triplet resonance may be the dominating nonlinear
interaction occurring in the triplet, while the nonlinear coupling with
the modes outside the triplet could strongly perturb the periodic
amplitude and frequency modulations expected if the triplets were pure
isolated systems. This shows that nonlinear mode interactions in real
stars are certainly more complex configurations than those treated by
current simplified theoretical approaches.
Moreover, since the $T_2$ triplet is in the transitory regime,
with frequencies locked by the nonlinear coupling within the triplet,
the nonlinear interactions outside this triplet are therefore unable
to destroy this frequency locking, resulting in no long-term frequency
variation as can be seen in the $T_1$ triplet.
Interestingly, the resonant mode coupling theory predicts that a limit
cycle (steady equilibrium state) may not be reached in the case of
three-mode direct resonance (which is likely at work here as discussed
above). The evolution of this long-term frequency variation in $T_1$,
whether the mode frequencies will further converge or eventually
diverge, and the evolution of $T_2$, whether frequencies will remain
constant or the locked regime will eventually be broken, will need further
observation either from ground or by future space instruments currently in
preparation, e.g., TESS and PLATO \citep{ra14,ri14}.
\subsection{The $D$-parameter and further insight on the modulations}
The value of the $D$-parameter, that defines how far the modes are from the
resonance center, is usually connected to the kind of regime a multiplet
is in when undergoing resonant mode interactions.
This $D$-value is in particular sensitive to the linear growth rate of
the oscillation modes (see again Section 3).
For the frequency lock regime that is observed in KIC\,10139564 with
the $T_3$ triplet, $D$ is near or exactly zero, as predicted by the
AEs \citep{bu95,go98}.
The $D$-value for the other triplets $T_1$ ($\delta\nu_o \sim 0.0026$\,$\mu$Hz)
and $T_2$ ($\delta\nu_o \sim 0.0036$\,$\mu$Hz) may reflects more their linear
growth rates since, with very similar frequency mismatches, the two triplets are
found to be in different regimes. The growth rate values are indeed
substantially different between $p$- and $g$-modes \citep{ch99,fo03}.
Assuming the growth rate for $p$-modes is of the order of 10$^{-6}$\,s$^{-1}$
\citep{ch99}, the corresponding $D$ value for the $T_1$ triplet would be far less
than 1, indicating that $T_1$ should be in the frequency lock regime
\citep{bu97,go98}, but we find it to be in the intermediate regime.
We note however that this estimate for the value of $D$ is based on the measured
frequency mismatch which may not be representative of the unperturbed frequency
asymmetry that enters in the definition of $D$. The latter is likely
much larger (see Section 4.1), leading to a somewhat larger $D$-value more
in line with the observed regime of the resonance for $T_1$.
The $D$-value for the $g$-mode triplet $T_2$, for its part, could be much
larger than that of $T_1$, considering the much smaller growth rates of
the $\ell=1$ $g$-modes \citep{fo03}. Extended ranges for $D$ were also
found in \dbv{} (Z16), where the $D$-values for the triplets are at least
two orders of magnitude larger than those suggested in \citet[]{go98}.
This suggests that the nonlinear behaviors not only depend on the magnitude of
$D$, but also on the specific coupling coefficients for each specific mode
\citep{bu95}.
Further quantitative comparisons between the observed modulations and
the theoretical framework would require to solve the amplitude equations
for the specific case of KIC\,10139564. This would require to calculate
the coupling coefficients in the AEs, which, in principle, could be
extracted from the observed amplitude and frequency modulations
\citep{bu95}. With these known coupling coefficients, one could then
determine the ranges of $D$-values related to each different regime
of the nonlinear resonance. A measurement of the growth rates of the
oscillation modes would then possibly follow with the determination of this
parameter, which may lead for the first time to an independent estimation of
the linear nonadiabatic growth rates of the modes and a direct test of
nonadiabatic pulsation calculations in sdB stars.
\begin{balance}
\section{Summary and conclusion}
While studying the high-quality and long-duration photometric data provided
by the {\sl Kepler} spacecraft on the pulsating sdB star KIC\,10139564,
we have identified different patterns in the frequency and amplitude
modulations of the oscillation modes belonging to several rotationally
split multiplets or linear combination frequencies.
These modulations show signatures that can be associated to nonlinear
resonant mode coupling mechanisms that could occur between the
multiplet components themselves and with other modes under certain
conditions, i.e., satisfying a \threemodes{} resonance relationship.
This is the first time that such signatures are quite clearly identified
in pulsating hot B subdwarf stars, and the second case reported so far for
a compact pulsator monitored with {\sl Kepler} photometry (see Z16).
We first reanalysed the 38-month of {\sl Kepler} photometry obtained
for KIC\,10139564, leading to the detection of 60 independent
frequencies above a secured detection threshold (5.6$\sigma$; see
Table\,\ref{t2}).
Among these, 29 frequencies consist of three triplets, one doublet,
one quintuplet and two incomplete multiplets with $\ell>2$
(see Table\,\ref{t1}). Another three detected
frequencies are linked to other frequencies through linear combinations.
Five additional groups of frequencies are found in the region
between 5400 and 6400\,$\mu$Hz, which have very complicated
structures.
Finaly, we also find 14 independent frequencies and two
frequencies satisfying linear combination relationships that could
be real as their amplitudes are between 5 and 5.6$\sigma$ above the
noise. In general, our well secured frequencies are in good agreement
with the former analysis from \citet{ba12}. In this paper, we particularly
concentrated our study on six multiplets and three linear combination
frequencies observed near 6076\,$\mu$Hz.
We found different types of mode behaviors occurring in the
above mentioned frequencies. A "short timescale" quasi-periodic amplitude
and frequency modulations along with a slow trend of the frequencies to
convergence toward each other occur in the dominant $p$-mode triplet near
5760\,$\mu$Hz ($T_1$). The $\sim570$-day quasi-periodic frequency
modulation evolve in antiphase between the two side components in this
triplet. Modulated frequencies and amplitudes are also found in
a quintuplet near 5287\,$\mu$Hz ($Q_1$) and a ($\ell>2$) multiplet near
5412\,$\mu$Hz ($M_1$), but the modulations do not show a clear periodicity.
One triplet near 316\,$\mu$Hz ($T_2$) has a distinct
behavior from the above mentioned multiplets, as it shows stable frequencies
but varying amplitudes. A similar phenomenon occurs in a doublet
near 394\,$\mu$Hz ($D_1$) which shows constant frequencies and
an $\sim 1100$ days periodic amplitude modulations. Another triplet at
518\,$\mu$Hz ($T_3$) completely differs from all the above multiplets,
with constant amplitudes and frequencies throughout the whole observation run.
In addition, we also discovered amplitude and frequency variations in three
frequencies near 6076\,$\mu$Hz ($C_1$) that are linked to other independent
frequencies through linear combinations.
After ruling out various possible causes for the modulations,
we showed that these mode behaviors could be related to the
different types of nonlinear resonances that should occur
according to the amplitude equation formalism. In particular,
nonlinear resonant couplings within a multiplet can lead to three
main regimes, all of which are possibly occurring in KIC\,10139564.
The multiplets $T_1$, $Q_1$ and $M_1$ can be associated
with the intermediate regime of the resonance where
the involved modes have modulated amplitudes and frequencies.
The triplet $T_2$ and doublet $D_1$ have a different behavior that
could be associated to a narrow transitory regime in which the
frequencies of the modes can be locked (constant) while the amplitudes
experience modulations. The behavior of the triplet $T_3$ is the unique
case found in this star that can be associated to the frequency
lock regime of the resonance, where both amplitudes and
frequencies are stable. In addition, the large amplitude ratios
between the $C_1$ frequencies and their main parent modes,
together with the large variation of amplitude and frequency
observed for these peaks, suggest that $C_1$ correspond to three-mode
direct resonances. We indeed found that the frequencies of $C_1$
exactly follow the evolution of their main parent modes. Moreover,
as the parent modes of $C_1$ are also the components of $T_1$ and
$T_2$, we suggest part of the complexity of the mode behaviors could be
related to these cross interactions between the various modes. In particular,
the slow variations occurring in $T_1$ may be related to the \threemodes{}
resonance superimposed to the triplet resonance occurring between the
components.
We emphasize that the observed frequency modulations likely induced by
nonlinear mode interactions could challenge future attempts
to measure the evolutionary effects on the oscillation mode periods
in pulsating sdB stars. Compared to the resonant variations taking place
on timescales of years, the rate of period change of the pulsations due
to stellar evolution in sdB stars is much longer, typically occurring on
a timescale of $\sim 10^6$\,yr \citep{ch02}.
Nonlinear modulations of the frequencies can potentially jeopardize any
attempt to measure reliably such rates, unless they can be corrected
beforehand.
These nonlinear modulations could also complicate the detection of
exoplanets or stellar companions around sdB stars using the
technique of measuring phase changes in the pulsations \citep{si07}.
It should be possible however to distinguish between the two effects,
considering that nonlinear couplings may induce different behaviors
on different modes, while external causes such as an orbiting body
should affect all modes similarly.
Finally, we note that our analysis suggests that resonances
occurring in real stars, in which modes could be involved in two
or more types of different couplings, lead to more complicated
patterns than those predicted by current theoretical frameworks which
treat the modes only as isolated systems within one type of resonance
and ignore the nonlinear interactions that could occur simultaneously
outside of the system.
This should motivate further theoretical work to develop nonlinear
stellar pulsation theory for more precise predictions of the
mode behaviors in pulsating stars in general.
\begin{acknowledgements}
Funding for the {\sl Kepler} mission is provided
by NASA's Science Mission Directorate. We greatfully acknowledge the
{\sl Kepler} Science Team and all those who have contributed to making
the {\sl Kepler} mission possible. WKZ acknowledges the financial support
from the China Scholarship Council.
This work was supported in part by the
Programme National de Physique Stellaire (PNPS, CNRS/INSU, France) and
the Centre National d'Etudes Spatiales (CNES, France).
\end{acknowledgements}
|
\section{Introduction}
Gender is one of the most important individual-level attributes known to structure patterns of social connectivity in human social networks \cite{ridgeway1999, ridgeway2006}. In Sociology, scholars have documented the organization of interaction according to gender lines across private, formal, informal, and public spheres of association \cite{ridgeway2004, mcpherson2001, munch1997, mcpherson1986, ibarra1997, hagan1998, smith1993}. Given this, it is not surprising to find that beginning with some of the earliest attempts to introduce a network perspective into the social sciences \cite[e.g.]{moreno1953, bott1955}, researchers have found consistent differences in the relative structure, composition, and relational quality of men and women's social networks \cite{fischer1982, marsden1987, coates1987, moore1990, dunbar1995, mcpherson2006}.
For instance, women have been found to include more kin as contacts with whom they discuss ``important matters'' or get social and emotional support than men \cite{fischer1983, marsden1987, coates1987, moore1990, dunbar1995}; to be more likely to generate a larger number of ``strong ties'' (contacts with whom the feel emotionally close) in free recall of friends and acquaintances \cite{fischer1982, marsden1987}; to be less likely to include co-workers in their core discussion networks \cite{fischer1983, marsden1987, moore1990}; and to be more likely to interact with members of their personal in private or informal settings rather than in public settings or formal organizations \cite{mcpherson1986, coates1987, marsden1987, mcpherson1986}.
In addition, the observation that friendship choices in social networks display a strong pattern of ``gender-segregation'' (also referred to as ``gender homophily'' or ``assortative mixing by gender'') is one of the most well-replicated and long-standing empirical generalizations in the social and behavioral sciences \cite{maccoby1988, benenson1990, smith1993, mcpherson2001, mehta2009}. The basic observation here is that same-gender preferences in friendship choices emerges very early in the life-course (at about 3 to 4 years of age), strengthens during adolescence and remains fairly strong through young adulthood, middle and old age \cite{mehta2009}. The other key finding in this literature is that women tend to display significantly stronger same-gender preference in friendships especially for close friends \cite{mehta2009}, although this pattern waxes and wanes depending on the stage in the life-course \cite{fischer1983, palchykov2012}.
Taken together, research on gender differences in the structure and composition of social networks along with work examining gender differences in communication propensity, partner preferences, and relationship maintenance style have the potential to generate diffusion and wide-randing effects in the macro-level structure of human social networks. In this paper we contribute to the literature on the origins of compositional differences in men and women's social networks. We do this both by replicating long-standing findings in the literature (e.g. stronger same-gender preferences for women in communication networks) and by extending the gender-differences paradigm to a new set of social network metrics (e.g. distributional tendencies across alters and gender categories in communication and proximity networks) and behavioral indicators (e.g. predictability of spatial mobility patterns).
\section{Theoretical Approaches to Studying Gender Differences in Social Networks}
Social network theorists have proposed two basic approaches to understanding the origins of gender differences in the structure and composition of social networks: \emph{dispositional} and \emph{structural} \cite{moore1990, smith1993}. Dispositional explanations for gender differences in social network structure and composition point to enduring, habitual behavioral tendencies (e.g. towards reciprocation, transitivity, frequency of communication, and so on) that make women's relationship building and maintenance style distinctive from that of men \cite{fischer1983}. Dispositional theorists point to the origins of these enduring behavioral tendencies in a combination of cultural, socialization, and personality-level influences. Structural theorists on the other hand explain differences in the structure and composition of men and women's networks by pointing to the different positions occupied by men and women in the social system, especially when these different positions entail differential access and opportunity to generate social connections of a given type (e.g. weak ties versus strong ties).
A third approach, the \emph{life cycle} perspective, attempts to mediate between dispositional and structural perspectives by noting that both factors may be at play but that whether one or the other is most prevalent depends on an individual's age \cite{kalmijn2002}. In this respect, life-cycle theorists point to the possibility that gender differences in personal network structure and composition may be due more to the stage of the life-cycle in which persons find themselves. From this perspective, family formation, entry into (or exit from) the workforce, the advent of children, and the unequal distribution of family responsibilities generates tendencies for women to draw on kin ties and strong ties for social and emotional support \cite{fischer1983, munch1997, palchykov2012}.
Previous work attempting to isolate dispositional differences in men and women's approach to relationship formation and maintenance has had trouble separating these from structural, and life-cycle influences. This line of research has suffered from two major drawbacks: First, there has been an undue focus on static qualities of personal networks obtained from free-recall of personal contacts obtained via the name generator method \cite{stehle2013}. Second, average gender differences are computed for groups who are either at different stages in the life cycle or are located different structural positions, thus combining information from persons in both distinct positions in the structure and at different stages in the life-cycle.
Reliance on personal networks constructed from free-recall of personal contacts elicited by name generator prompts have been subject to criticisms on various methodological grounds \cite{brewer2000, marsden2003, marin2004}. For our purposes, the key limitation of this approach when it comes to drawing inferences related to the dispositional origins gender differences in social networks is that we cannot separate gender differences connected to cognitive and emotional biases in recall and naming of contacts from dispositional differences connected to communication \emph{behavior}. For instance, the fact that women tend to name more kin and ``close" contacts than men, may be due to the fact that both kin and emotionally close contact are more cognitively and affectively salient for women or that they women are in fact more likely to sustain ties that require a high level of behavioral engagement. Self report of the number of ties to which the persons feels subjectively close, however, may not tell us much about dispositional differences in behaviors relevance to network formation and relationship maintenance. In what follows we deal with this issue by focusing our attention on \emph{behavioral} differences in the social networking behavior of men and women obtained via non-obtrusive data collection strategies tied to a remote social-sensing platform keyed to automatically capture social interactions between individuals \cite{striegel2013}.
Reliance on groups that are heterogeneous on both life-cycle stage and structural location on the other hand, leads to problems in isolating behavioral differences relevant to the structure and composition in men and women's social networks that come from differences in opportunity structure, gender inequalities in the domestic division of labor, childcare, and workforce participation \cite{munch1997, smith1993, moore1990}. For instance, the over-representation of kin and the stronger same-gender preference patterns in women's networks may be due to dispositional factors (actual preference for social contact with kin and same-gender alters) or to the fact that women continue to be more likely to do the bulk of household related and childcare related activities \cite{fuwa2004}, in settings where they are more likely to come into contact with other women and draw on close relatives for help and support \cite{munch1997, smith1993}.
In what follows, we isolate dispositional from structural and life course factors by focusing our analysis on a population that is largely homogeneous in terms of life-course stage (e.g. unmarried, childless, young adults between the ages of 17-19 making the transition to college), located in a setting that provides most persons with a similar opportunity structure for network formation (a self-contained, majority-residential campus, located in a mid-sized town). This is an ideal time in the life-course to study dispositional differences in network-related behaviours as the life-course events associated with gender-based structural differences in the composition and structure of social networks such as marriage, labor-force entry and childbearing \cite{fischer1983, moore1990, munch1997}, have yet to take effect.
\section{Previous work on Gender Differences in Social Networks Using Unobtrusive Data Collection Strategies}
While not extensive, an emerging literature using unobtrusive data collection strategies points to possibility of important dispositional differences in the relational behavior of men and women especially when it comes to revealed preferences and usage patterns of remote telecommunication media, such as phone calls or text messaging.
\subsection{Mobile Communication Networks}
Several studies on gender differences in relational and network behaviors relies on mobile communication network (MCNs) data constructed from large data sets (usually containing millions of users) of mobile phone call data records. For instance, a study of ``temporal motifs'' finds that all-female subgraphs of size three are over-represented in these types of networks \cite{kovanen2013}, suggesting that women are (a) more likely to initiate communication events and direct them at other women (``out-star'' temporal motif), (b) more likely to be the recipient of communication events from other women (``in-star'' temporal motif), (c) induce their female contacts to generate subsequent communication events of their own post-contact (``causal chain'' temporal motif), and (d) be the recipient of a communication event from another woman after directing a communication event to a seemingly unrelated contact (``non-causal chain'' temporal motif).
Additional MCN-based studies have been able to replicate the classic observation of strong same-gender preference patterns for each user's strongest dyadic connections in the context of mobile phone communication data \cite{palchykov2012}. Taking a life-course perspective, this work shows that same-gender preference for the strongest link is pervasive and strong at all ages, but that (a) it is stronger for women, and (b) the strength of the gender difference depends on life course stage \cite{palchykov2012}. Finally, other MCN-based work shows that women are more likely to engage in cross-generation communicative interactions than men, tend to call more frequently and spend more time on the phone, and display a stronger preference for same-gender communication \cite{stoica2010, dong2014}. This work replicates stylized facts regarding women's propensity towards greater sociality and kin connectivity that had been uncovered in previous social network research using free-recall name-generators \cite{fischer1983, marsden1987, moore1990}.
\subsection{Online and Social Media Networks}
Research dealing with gender differences in the context of online social networks sites (SNSs) replicates the long standing finding of stronger assortative mixing preferences for women. This same research, however shows that cross-gender connections are more common for users with a larger circle of friends \cite{volkovich2014}. While not sharing all of the features of social interactions in natural settings, recent work reveals strong gender-specific effects in relational behavior in SNSs dedicated to online gaming \cite{szell2013}. This work finds that women tend to be more risk averse, while demonstrating stronger same-gender preferences (assortative mixing) than men in this setting. Related work finds that these type of settings elicit preferential behavior directed towards women, including a larger volume of incoming communication \cite{griffiths2004}. Using data from Facebook profiles for a population very similar to the one under consideration here (the Fall of 2005 incoming Freshman cohort at a ``diverse private college in the Northeast U.S."), Harvard researchers found that women tended to have larger, more active, and more diverse Facebook networks than men \cite{lewis2008}.
\subsection{Spatial Proximity Networks}
Recent work has begun to tackle issues of gender differences in network related behaviors by constructing spatial proximity networks (SPNs) from wearable sensing badges \cite{paradiso2010}. These sensors, worn continuously by study participants, are designed to capture each individual location signature patterns, allowing researchers to study proximity behaviors. This work shows that generalizations obtained from communication network analysis (e.g. mobile phones and social media), such as the stronger preference for same-gender contact displayed by women, may not be completely generalizable to spatial proximity networks.
For instance, in a study of contact proximity among high school students, it was found that while female students seemed to display little same-gender preference (in fact spending most of their time around other-gender students) in proximity behaviors, male students were much more likely to spend the majority of their time in close proximity to other male students. These differences were partially attenuated once the marginal distributions of students of each gender were accounted \cite{fournet2014}. A recent study of of primary school children (aged 6 - 12) using wearable sensor methodology finds similar results. While same-gender preference in spatial proximity behaviors is found for boys and girls of of all ages, boys tend to display a stronger pattern of same gender spatial proximity preference and this effect becomes larger as boys get older \cite{stehle2013}. In all, emerging results on patterns of interactive gender-segregation in spatial proximity networks using unobtrusive methods contrasts sharply with the recurrent pattern of stronger same-gender preference for women found in the work communication-based networks considered above \cite{stoica2010, palchykov2012}.
\section{Our Work}
Social sensing has emerged as a new paradigm of crowdsourcing the data collection tasks to average individuals~\cite{wang2015social}. Our work leverages high quality information from a social sensing platform to provide a wide-ranging exploration of dispositional differences in men and women's relationship building and relationship maintenance style that is relatively (but not completely) unconfounded by life-course and structural influences. We move beyond previous work by studying patterns of communication among young adults making the transition to college in a natural setting in a context in which they are in the process of ``turning over'' their previous (high school) network and forming new relationships \cite{oswald2003, kane2011}.
This unique dataset \cite{meng2014} allow us to consider a wider range of network-relevant communication behaviors and network formation tendencies than previous work on the subject. In particular we seek to move beyond the emphasis on gender differences in static characteristics of social networks such as assortative mixing and reciprocity \cite{rivera2010, szell2013}. We focus on gender-specific difference in managing and maintaining interactions via three communication channels (phone call, text messaging, and via spatial proximity). We further examine how individuals sustain or modify their behaviors in these three types of communication channels in their turn. Gender differences on each criterion behavior are assessed via the construction of several null models, in which we compare observed gender patterns to what we would expect by chance \cite{kovanen2013,yang2014}.
In keeping with the argument outlined above, we focus on dispositional differences that can be tied most directly to behavior (and not issues of the cognitive or emotional salience of personal contacts). In particular we focus on broad behavioral patterns that may be tied to dispositional differences between men and women. Some of these patterns have been addressed in previous work (e.g. same-gender preference in communication and proximity networks), while others are analyzed for the first time here (e.g. gender differences in distributional tendencies and spatial mobility diversity).
\subsection{Gender Differences in Social Networks}
We address the question of the extent to which we can observe a pattern of same-gender preference across communication (voice call and text messaging) and spatial proximity networks. We also address the question of whether there are gender differences in the strength and direction of the same-gender preference effect across communication and proximity networks. Our review of previous studies suggests that we should find strong patterns of gender based homophily across these networks. However, given the different pattern of results obtained in the communication network and spatial proximity studies \cite{fournet2014, stehle2013}, we anticipate finding stronger same-gender preference patterns for women in communication networks, and stronger same-gender preference patterns for men in spatial proximity networks. In addition, we address the question of whether there exist gender differences in the \emph{coupling} of communication and proximity. To address this issue, we examine the extent to which we can observe: (a) increases in dyadic proximity behaviors following connectivity in the communication networks, and (b) the extent to which the strength of this effect differs for men and women.
\subsection{Gender Differences in Sociability, Popularity and Reciprocity}
Previous work on MCNs (mobile communication networks) and SNSs (social networks sites) has shown strong differences between men and women in both their tendencies to direct and receive communications from others, with women being more likely to engage in and receive communications than men. We ascertain whether we can observe similar differences in these two behavioral outomes in our two directed (voice call and text messaging) networks. Reciprocity is a key construct in social network theory \cite{hallinan1979}, and an important behavioral marker in social interactions \cite{gouldner1960}. Here we address the question of whether there exist gender differences in the tendency to reciprocate communicative interactions.
\subsection{Tendencies Towards Equitable Distribution of Communications}
While previous work has examined the general tendencies of individuals to distribute their communications unequally across contacts \cite{miritello2013}, the question of whether men and women differ on this crucial behavioral trait has not been examined before. This is largely because of the unavailability of data containing fine-grained records of communication frequencies across contacts \cite{eagle2010}. We introduce a quantitative metric of an individual's ability to distribute their communications equitably (or inequitably) across their contacts and ascertain whether men and women are different along this metric. We extend this work by examining the extent to which men and women differ not only in their distribution across contacts, but also in their distribution tendencies across \emph{types of contacts} (classified by gender). In addition, we also extend this framework to examine gender differences in distributive tendencies regarding the \emph{time} spend with other persons in the spatial proximity network.
\subsection{Gender Differences in Location Diversity}
Different relationships are tied to different spatial locations \cite{doreian2012}. In addition, research in the emerging field of human mobility studies reveals surprising levels of predictability in human mobility behavior \cite{song2010}. If the structure of men and women's social networks differ, this may be tied to dispositional differences in the way that men and women distribute their time across spatial locations \cite{mehta2009}. However, this is a question that has yet to be addressed in social network and mobility studies. We have at our disposal detailed records the traces of students in 54 locations on campus, which allows us to explore gender-specific difference patterns of mobility (Figure~\ref{fig_mobility_flow}). We examine three questions: Are men and women different in the relative \emph{diversity} (e.g. predictability or lack thereof) of their mobility patterns? Are same-gender dyads more similar in mobility patterns than cross-gender dyads? Do mobility patterns change after persons become connected in the communication network?
\subsection{Paper Organization}
The rest of the paper is organized as follows, section~\ref{sec_data} we introduce and describe our main data source. We follow with a detailed description of preliminaries in section~\ref{sec_prelim}. In section~\ref{sec_analysis} we report the key dispositional gender differences that emerge in our analysis. We conclude by outlining the implications of our results for future work on gender differences in social networks in section~\ref{sec_discussion}.
\section{Data}
\label{sec_data}
The ND mobility data \cite{meng2014} is drawn from the University of Notre Dame's NetSense smartphone study \cite{striegel2013}. This study was launched in August of 2011 with the goal of monitoring the communication and relationship formation behavior of 200 freshmen entering the university in the Fall of 2011. The primary data collection strategy of social interactions was via unobtrusive social sensing. Study participants were equipped with smartphones modified to track, the location, calls and texts made and received (but not the content of their communications) over a two-year period using a monitoring app employed on each phone. This app logged and then transmitted to a secure database a Call Data Record (CDR) for each communication event.
Each CDR contains the phone numbers of the sender and receiver along with a timestamp indicating when the event occurred. There are two broad categories of communication types among these students: digital communications (text messages and phone calls) and face-to-face interactions (proximity observed via Bluetooth). Additionally we also have location traces of each student. There are in total 54 locations labelled in this dataset; among them 18 locations are for study usage (i.e., classrooms and library), 27 of them are residence halls, and rest of them are church, football stadium, dining hall and etc. All of data described above are collected between August of 2011 and December of 2012. The data collection spans across three semesters: 1) 2011-08 to 2011-12; 2) 2012-01 to 2012-05; 3) 2012-08 to 2012-12. We consider the following four categories of data.
\begin{figure*}[ht]
\centering
\centerline{\includegraphics[width=3in]{mobility_flow.jpg}}
\caption{Mobility Flows on Campus. The color ranges from blue to white. White color indicates that the density of flux is high, while blue means that the flux is rare.}
\label{fig_mobility_flow}
\end{figure*}
\subsection{Voice Call Network}
We built the voice call network from the CDR log. In the voice call networks vertices are individual students and an edge exists between individual $u$ and an individual $v$ if there is CDR showing that $u$ initiated a voice call with $v$. The voice call network is both \emph{directed} and \emph{weighted}.
\subsection{Text Message Network}
In the same way, we built the text message network from the CDR log. In the text message network vertices are individual students and an edge exists between individual $u$ and an individual $v$ if there is CDR showing that $u$ sent a text message to individual $v$. Like the voice call network, the text message network is both \emph{directed} and \emph{weighted}.
\subsection{Spatial Proximity Network}
In the proximity network, vertices are individual students and an edge exists between individual $u$ and an individual $v$ if bluetooth logs show them to be proximate to one another at that time. The weight of the $(u,v)$ edge is given by the total duration that they have been proximate as the weight of the link $(u,v)$. Given the inherent non-directionality of the proximity relation, this network is \emph{undirected} and \emph{weighted}.\footnote{We only consider proximity network in three semesters, due to the fact that there are rare proximity interactions when students are on vacations. Thus our validations and experiments are conducted on semester 1 (2011-08 to 2011-12), semester 2 (2012-01 to 2012-05), and semester 3 (2012-08 to 2012-12) proximity networks.}
\subsection{Location Traces}
The GPS device in each cellphone also records the location traces of each student between August of 2011 and December of 2012. If an individual student is at the location $L$, then the monitor software will send this information with the current timestamp to a server. Thus we have the location trace for each student, which allows us to analyze their mobility behaviors (Figure~\ref{fig_mobility_flow}). We divided campus location into 54 distinct spatial regions. Among these, 18 are dedicate for schooling and study activities (e.g., classrooms and library), 27 are (same-gender only) residence halls, and the rest consist of such function and activity specialized locations as the Basilica (Church), the football stadium, the dining halls (two major ones ``North'' and ``South''), and so on. In this way we are able construct the location visiting signature $T_{L} (u)$ for each individual $u$. We can then compute the similarity of location visiting patterns between any pair of students. Additionally a student $u$'s location signature $T_{L} (u)$ allows us to calculate the extent to which he or she spends time at a small number of locations or spreads his or her time evenly across locations (see Equations~\ref{eq:location} and~\ref{eq:locdiv} below).
\section{Network Metrics}
\label{sec_prelim}
One main goal of this study is to determine the existence of dispositional differences between men and women in communication behavior, such as the distribution of communicative behaviors across alters (among others). In this section we present and formalize several behavioral metrics aimed at capturing gender differences in relationship formation and relationship maintenance behavior across our four different sources of relational data, namely phone call, text messaging, face-to-face interaction, and mobility traces.
\subsection{Topological Diversity}
We use Shannon's entropy \cite{eagle2010,shannon1949} to capture the diversity of an individual node $s$' local structure. For a node $s$ we can extract its ego network (as shown in Figure~\ref{fig_diversity}), and the diversity of $s$ is calculated as below:
\begin{equation}
H_{t}(s) = -\sum_{j=1}^{k}p_{sj}log(p_{sj})
\label{eq:diversity}
\end{equation}
where $k$ is the number of $s$ neighbors and $p_{sj}$ is the fraction of interactions that happen between $s$ and $j$,
\begin{equation}
p_{sj} = \frac{w_{sj}}{\sum_{j=1}^{k} w_{sj}}
\end{equation}
where $w_{sj}$ is the weight of edge between $s$ and $j$. For the communication networks (voice calls and text messages), the Shannon entropy gives us an ego-level metric of the extent to which any one individual concentrates their communications among a few contacts or distributes them equally across contacts. For the proximity network, this measure can be interpreted as the likelihood that an individual distributes her ``face time'' equally among all of her neighbors, or spends most of her time in spatial proximity to small proportion of her total number of possible contacts.
\begin{figure*}[h]
\begin{center}
\centerline{\includegraphics[width = 3.2in]{diversity_example.jpg}}
\caption{Ego Network diversity for a hypothetical individual. If an individual's communications are distributed evenly across different alters (brown), then the diversity value is high; conversely, if the bulk of communications go to a small number of alters (blue), the diversity value is small.}
\label{fig_diversity}
\end{center}
\end{figure*}
\subsection{Gender Diversity}
As we have seen, the study of gender segregation of friendship choice is a classic line of analysis in the literature. We extend this approach to study the extent to which men and women exhibit a preference to distribute their communications (in the case of voice calls and text messages) or ``face time'' (in the case of proximity) to same-gender or different-gender partners. To that end, we develop a \emph{gender diversity metric} (GDM) as follows.
We can classify individuals and their neighbors as a man (m) and a woman (w). We then use the Shannon entropy to measure the gender diversity of every individual. This is given by:
\begin{equation}
H_{g}(s) = -p_{w} log (p_{w}) - p_{m} log (p_{m})
\label{eq:gendiv}
\end{equation}
where $p_{w}$ is the proportion of $s$' total interactions with women, and and $p_{m}$ is the proportion of $s$'s total interactions with men. For the proximity network, this measure can be interpreted as the likelihood that an individual distributes his ``face time'' equally among members of the same gender, or spends most of his time in spatial proximity to same-gender peers.
\subsection{Location Signature}
In this data, we have the traces of each student in 54 locations. For a student $s$, we denote her/his cumulative staying time in a location $L_{i}$ as $T(s, L_{i})$. In this way, for a student $s$, we can construct a vector $T_{L} (s)$,
\begin{equation}
T_{L} (s) = \{ T(s, L_{1}), T(s, L_{2}), ..., T(s, L_{k}) \}
\label{eq:location}
\end{equation}
where $k$ is the total number of locations. The vector $T_{L} (s)$ can be considered a signature of student $s$'s location visiting behaviors. The location similarity between two individuals $u$ and $v$ is computed by the Pearson Correlation Coefficient between the respective location vectors $T_{L} (u)$ and $T_{L} (v)$ (see Figure~\ref{fig_location}).
\subsection{Location Diversity}
We calculate the location diversity $H_{L}$, for each individual $s$ using that individual's location signature:
\begin{equation}
H_{L} (s) = -\sum_{i=1}^{k} p(s, L_{i}) log(p(s, L_{i}))
\label{eq:locdiv}
\end{equation}
where $k$ is the number of locations and $p(s, L_{i})$ is the fraction of time (hours) $s$ stays at location $L_{i}$. The location diversity $H_{L}$ captures students' preference in location visiting (see Figure~\ref{fig_location}).
\begin{figure*}[ht]
\begin{center}
\centerline{\includegraphics[width = 3.2in]{location_example.jpg}}
\caption{Location Visiting Signature Example. The ND mobility data contains traces of the time (in minutes) each individual spends in 54 campus locations. We calculate the cumulative time spent at location $L_{i}$ for each individual $u$. The resulting $1 \times 54$ vector of location durations can thus be considered that individual's spatial signature on campus.}
\label{fig_location}
\end{center}
\end{figure*}
\section{Gender Difference Analysis}
\label{sec_analysis}
\subsection{Significance Validation}
We begin by ascertaining whether we observe assortative mixing by gender in each of the communication networks (voice call and text messaging). A straightforward way to measure and validate the existence of same-gender preference in communication partners is to compare the numbers of links between four possible ``mixtures'' of (directed) links (MM, WW, MW, WM) in each network to a null model where the gender classification of node is randomly assigned, while keeping the underlying (observed) network structure intact. Positive and negative Z-scores denote the number of standard deviations by which a directed link of each type is over or under represented. We compare real data to 10,000 randomized cases where the gender classification of each node is randomly shuffled. The statistical significance of gender mixture combination is measured by the {Z} score:
\begin{equation}
Z = \frac{N_{obs} - \mu(N_{rand})}{\sigma(N_{rand})}
\label{eqn_gender_z}
\end{equation}
where $N_{obs}$ is the number of links by type (e.g. MM, WW, MW, WM) in the observed networks, $\mu(N_{rand})$, and $\sigma(N_{rand})$ are the average of links and the standard deviation in the gender-reshuffled networks respectively. The statistically significant Z-score values are displayed in the boldface type in the Tables.
\begin{table}[h]
\caption{Gender-Specific Number of Calls and Text Messages Validation ($Z$-scores) ($H_{0}$ rejected at $2$ sigmas). Homophily Validation.}
\resizebox{4in}{!}{
\begin{tabular*}{\textwidth}{c @{\extracolsep{\fill}} cccccccc}
\hline \hline
&\multicolumn3c{Voice Call Network} & \multicolumn3c{Text Message Nework}\\
& Unweighted & Weighted & Weighted & Unweighted & Weighted & Weighted\\
&&(N. Events)&(Duration)&&(N. Events)&(Text Length) \\
\hline
MM & 0.759 & 0.334 & -0.648 & -0.346 & -1.106 & -1.089 \\
WW & {\bf 2.164} & -0.548 & -0.431 & {\bf 2.112} & 0.331 & 0.164 \\
MW & {\bf -3.083} & -0.822 & 0.511 & {\bf -2.138} & 0.702 & 0.919 \\
WM & {\bf -2.468} & 1.022 & {\bf 1.953} & -1.649 & 0.735 & 0.735 \\
\hline
\label{tab_homo_validation}
\end{tabular*}
}
\end{table}
\subsection{Gender Differences in Assortative Mixing}
\subsubsection{Voice Call and Text Messaging Networks}
The gender homophily analysis results for the voice call and text-messaging networks are shown in Table~\ref{tab_homo_validation}. First, it is clear that in both the voice call and text messaging networks women display statistically significant same-gender preference (both are above $2$ sigmas) but men do not ($Z$ = $0.759$ and $-0.346$ for voice call and text message networks respectively). This is consistent with previous work showing stronger same-gender preference in friendships for women in MCNs \cite{stoica2010, palchykov2012}. In both cases, cross-gender links are significantly under-represented (WM links is not significantly under-represented in the message network however, $Z$ = $-1.649$).
In addition, when considering the overall volume of communication (the weighted representation of each network) there seems to be an over-representation of weighted cross-gender links although the z-score for this edge type is not significant ($Z$=$0.735$). Note, however, that same-gender preferences pattern are largely attenuated (dropping to non-significance) when we consider the weighted representation of each network irrespective of whether the weight represents the number of events (phone calls, messages), call duration, or text message length. This suggests that studies that rely on binary representations of friendship choices may largely overestimate the existence of gender-segregation in social networks.
\begin{table}[h]
\caption{Gender-specific proximity validation (Z-score). We examine the significance of three types of proximity interactions, MM, WW and MW/WM (proximity network is undirected, thus MW and WM are equivalent) ($H_{0}$ accepted).}
\resizebox{4in}{!}{
\begin{tabular*}{\textwidth}{c @{\extracolsep{\fill}} lcccccc}
\hline \hline
&\multicolumn2c{Semester 1} & \multicolumn2c{Semester 2} & \multicolumn2c{Semester 3} \\
& All Pairs & Connected & All Pairs & Connected & All Pairs & Connected \\
\hline
MM & {\bf 3.645} & {\bf 3.723} & {\bf 4.195} & {\bf 3.906} & {\bf 1.974} & {\bf 2.044} \\
WW& {\bf 2.701} & {-0.072} & {\bf 2.525} & -0.009 & {\bf 2.638} & {0.583} \\
MW/WM & {\bf -7.907} & {\bf -4.444} & {\bf -8.745} & {\bf -4.466} & {\bf -5.616} & {\bf -3.652} \\
\hline
\end{tabular*}
}
\label{tab_proximity_homo}
\end{table}
\subsubsection{Proximity Network}
Table~\ref{tab_proximity_homo} shows the corresponding gender homophily analysis results for the proximity network. We partition the proximity log into three temporal snapshots. The first one is from Aug of 2011 to Dec of 2011, corresponding to the first semester that students spent on campus (Semester 1); the second snapshot goes from Jan of 2012 to May of 2012 (Semester 2); and the third snapshot covers the period from Aug of 2012 to Dec of 2012 (Semester 3). We do not consider proximity data from other time periods, since these are logged during summer and winter vacations when the majority of students are off campus. Additionally when calculating the cumulative proximity time between two students (how long they have been proximate to each other in one semester), we exclude proximity times logged between 9:00 PM and 9:00 AM, in order to remove large (spurious) proximity values generated by students who spend time sleeping in the same dorm. For each proximity metric, we present separate sets of results for all of the participants in the study (All Pairs) and for those participants who have an edge in either the voice call or text messaging networks (Connected).
The results of our analysis show that in contrast to the communication-based networks, men exhibit a stronger levels of assortative mixing by gender than women. The all-male (MM) dyad is overrepresented in the proximity network in all three time periods ($Z > 1.96$), whether we consider all pairs or only respondents who are directly connected via a communication link. Women on the other hand, exhibit a tendency to be spatially proximate to other women (WW dyad) when we consider all pairs, but this tendency disappears when we restrict our analysis to those individuals with whom they are connected. This suggests that while both men and women experience high levels of spatially based gender assortativity, only men seem to couple direct social connectivity with spatial contiguity. Women on the other hand seem to spend equal amounts of time being spatially proximate to their social contacts regardless of their gender. Finally, mixed gender dyads are significantly underepresented in all three periods for both all pairs and connected dyads ($Z < -2$).
\begin{table}[h]
\caption{Gender-Specific Number of Calls and Text Messages Validation ($Z$-scores) ($H_{0}$ rejected at $2$ sigmas): Action Validation.}
\centering
\resizebox{4in}{!}{
\begin{tabular*}{\textwidth}{c @{\extracolsep{\fill}} cccccccc}
\hline \hline
&\multicolumn3c{Voice Call Network} & \multicolumn3c{Text Message Nework}\\
& Unweighted & Weighted & Weighted & Unweighted & Weighted & Weighted \\
& &(N. Events)&(Duration)& &(N. Events)&(Text Length) \\
\hline
From M & -0.629 & -0.187 & -0.428 & {\bf -1.971} & -0.920 & -0.730\\
From W & 0.629 & 0.187 & 0.428 & {\bf 1.971} & 0.920 & 0.730\\
\hline
To M & -0.544 & 1.159 & 0.082 & -1.152 & -0.897 & -0.862 \\
To W & 0.544 & -1.159 & -0.082 & 1.152 & 0.897 & 0.862 \\
\hline
\label{tab_action_validation}
\end{tabular*}
}
\end{table}
\subsection{Gender Differences in Sociability (Outgoing Communications)}
A long standing finding in the MCNs (mobile communication networks) and SNSs (social network sites) literature is that women tend to initiate and produce a larger volume of interactions than men \cite{lewis2008, miritello2013}. As shown in the first two rows of the bottom panel of Table~\ref{tab_action_validation}, we find mixed support for this hypothesis. In the phone network men and women are equally likely to be the initiator of communications. In the message network, we observe that women are indeed more likely to initiate communications ($Z$ = $1.971$). This pattern of results suggests that dispositional gender differences in network-related behaviors may depend on the communication channel that is being considered. Below we consider other results that are consistent with this expectation.
\subsection{Gender Differences Popularity (Incoming Communications)}
Previous work in online environments (i.e., social network sites) has shown that women are more likely to be the target of communication attempts \cite{griffiths2004}. As shown by the last two rows of Table~\ref{tab_action_validation} we find little support for this hypothesis in our data ($|Z| < 2$). Men and women are equally likely to be the recipient of communications in both the voice call and text message networks. This result remains the same whether we consider the binary or weighted representation of each network.
\begin{table}[h]
\caption{Gender-specific Reciprocity of Call and Text Message Validation ($Z$-scores).}
\centering
\resizebox{4in}{!}{
\begin{tabular*}{\textwidth}{c @{\extracolsep{\fill}} ccc}
\hline
& {Voice Call Network} & {Text Message Network} \\ \hline \hline
{MM}& {\bf 2.019} & {\bf -1.969} \\
{WW}& -0.312 & -1.005 \\
{MW}& -0.501 & {\bf 2.541} \\
{WM}& -1.515 & 1.171
\\
\hline
\label{tab_reciprocity}
\end{tabular*}
}
\end{table}
\subsection{Differences in Reciprocity by Gender Mix}
We consider the question of whether directed links are more likely to be reciprocated depending on whether these links are same-gender versus cross-gender links and whether the participants in the interaction are men or women. To that end, we follow the same procedure used above to examine homophily patterns. First, we compute the observed rate of reciprocation along the four types of links classified by the gender mix. Second, we compare the observed frequencies with those obtained from a null model where the gender classification of each node is randomly assigned but the network structure remains intact. In each round of simulation, we obtain counts of reciprocated relationships for the four link types in the randomized network. The simulation is repeated for 10,000 times. Finally we use equation~\ref{eqn_gender_z} to compute a reciprocation z-score for each gender mixture.
Table~\ref{tab_reciprocity} shows the results. We find that relations between men (MM dyads) are significantly more likely to be reciprocated in the voice call network than we would expect by chance. However, we find the \emph{opposite} pattern of results in the text message network; here men are less likely to reciprocate interactions initiated by other men. In addition, we find significant differences in reciprocation behavior for cross-gender pairs depending on the gender of the initiator. Women are more likely to reciprocate a text message coming from a man than we would expect by chance ($Z = 2.541$), but men are equally likely to reciprocate a text message regardless of the sender's gender ($Z = 1.171$). Women exhibit no tendencies for higher levels of within-gender reciprocity than we would expect by chance on either the voice call or text messaging network ($|Z| < 2$).
\subsection{Gender Differences in Topological and Gender Diversity}
\subsubsection{Voice Call and Text Messaging Networks}
We define the topological diversity as an ego-level metric capturing an individual's tendency to concentrate communication on a few contacts or distribute them evenly across neighbors (see equation~\ref{eq:diversity}). We define the gender diversity as ego-level metric capturing the tendency to concentrate communications either men or women contacts or to spread communications evenly across gender lines (see equation~\ref{eq:gendiv}). Both measures range from 0 to 1, with a higher value indicating a greater propensity to distribute communications equitably across neighbors or gender categories. A key question is whether men and women differ in both of these behavioral propensities. As before we generate a null model by keeping the distribution of link weights the same but reshuffling the gender labels and calculating the topological and gender diversity for each individual at each iteration. We compute Z-scores using the procedure outlined above.
\begin{table}[h]
\caption{Gender-Specific Diversities Validation (Z-scores). $H_{0}$ is rejected at 2 sigmas.}
\resizebox{4in}{!}{
\begin{tabular*}{\textwidth}{c @{\extracolsep{\fill}} lcccc}
\hline \hline
& \multicolumn2c{Voice Call Network} & \multicolumn2c{Text Message Network} \\
& $H_{g}$ & $H_{t}$ & $H_{g}$ & $H_{t}$ \\ \hline
Gender Effect (Women) & {\bf 2.148} & {\bf 2.041} & {\bf 2.913} & {\bf -2.006} \\ \hline
\label{tab_z_score_topological}
\end{tabular*}
}
\end{table}
Table~\ref{tab_z_score_topological} shows the results. We find that women differ systematically from men in their tendency to concentrate communications on a few contacts but that the direction of this difference depends on the communication channel (voice versus text). In the voice call network, women are more likely than men to distribute their communications evenly across contacts \emph{and} gender categories. In the text messaging network, on the other hand, women are more likely than men to concentrate their social interactions on a few persons, but are \emph{less} likely to concentrate their communications on members of only one gender. This suggests that women tend to be more likely to structure their communication messaging networks in terms of a core/periphery regime, but that the \emph{gender} of the preferred contact is not necessarily predictable. Men on the other hand are more likely to partition their voice call network in terms of a core-periphery regime, \emph{and} they tend to select a same-gender alter as a preferred voice call contact.
\begin{table}[h]
\caption{Gender Diversity and Topological Diversity in Proximity Network (Z-score). We observe that women have significantly higher gender diversity and topological diversity than men. This observation persists in three semesters. ($H_{0}$ rejected at 2 sigmas)}
\resizebox{4in}{!}{
\begin{tabular*}{\textwidth}{c @{\extracolsep{\fill}} ccccccc}
\hline \hline
&\multicolumn2c{Semester 1} & \multicolumn2c{Semester 2} & \multicolumn2c{Semester 3} \\
& $H_{g}$ & $H_{t}$ & $H_{g}$ & $H_{t}$ & $H_{g}$ & $H_{t}$ \\
\hline
Gender Effect (Women)& {\bf 1.972} & {\bf 1.960} & {\bf 2.067} & {\bf 2.077} & {\bf 1.960} & {\bf 1.966} \\
\hline
\end{tabular*}
}
\label{tab_proximity_diversity}
\end{table}
\subsubsection{Proximity Network}
Do men and women exhibit similar patterns of topological and gender diversity when it comes to spatial proximity as they do for voice calls and text messages? Table~\ref{tab_proximity_diversity} shows the results. We find that women have higher topological and higher gender diversity than men throughout all three periods. This suggests that, in contrast to men, who seem to segment their face-to-face interactions so that they end up spending a disproportionate amount of time spatial proximity to a select group of other men, women distribute their ``face time'' equally across all of their possible contacts regardless of their gender category.
\subsection{Gender Differences in Communication and Proximity Coupling}
It is unlikely that the gender-based dynamics in the communication (voice call and text message) networks and the proximity networks are independent. Instead, social network theory suggests that there are fundamental dependencies between the two, such that propinquity leads to connectivity via communicative interactions, and communicative interactions lead to propinquity by increasing the motivation and opportunities to engage in face to face interaction \cite{feld1981, carley1991, doreian2012}. However, our results suggest that there should exist systematic differences between men and women in the extent to which social connectivity affects subsequent patterns of spatial proximity and face-to-face interaction. In essence, men seem to couple social interaction and physical proximity in a stronger way than women (see the results summarized in Table~\ref{tab_proximity_diversity}). We evaluate this hypothesis by ascertaining whether we can observe significant changes in the proximity network once we know that two individuals are connected in the communication network and whether this effect is the same for men and women.
\subsubsection{Procedure}
We proceed as follows. First, we calculate the daily average proximity ($t$) for each pair of individuals, $u$ and $v$, before they are connected in the communication network, we then calculate the same quantity ($t'$) \emph{after} we observe those individuals interact via voice call or text messaging. We then compute the difference between these two quantities ($\Delta t$ = $t'-t$). This gives us an estimate of the effect of connectivity on time spent in social proximity. We then use the same null model generating procedure (see Equation~\ref{eqn_gender_z}) described above to determine the existence of significant differences in proximity differences between pre and post-connectivity proximities for same-gender (MM and WW) and mixed gender (MW/WM) dyads.
\begin{figure}[ht]
\centering
\includegraphics[width=2.5in]{beforeafter_proximity_new.jpg}
\caption{The impact of friendship on mobility behaviors. All-men dyads (MM) display higher levels of spatial proximity after connectivity in the Voice Call and Text Messaging networks. This effect is not observed in the all-women (WW) and the mixed-gender (WM/MW) dyads}.
\label{fig_beforeafter_proximity}
\end{figure}
\subsubsection{Results}
The results are summarized in Figure~\ref{fig_beforeafter_proximity}. We find that, consistent with expectations, spatial proximity increases after observing connectivity in the communication network only for the all-men (MM) dyad. After being connected in the communication network, members of the average all-male dyad are observed to spend about 3.5 additional hours near one another, which is a substantively significant effect. For all-women (WW) and the mixed gender (MW/WM) dyad there is no statistically appreciable effect of connectivity on subsequent patterns of spatial proximity. This supports the hypothesis that men tend to couple social and physical connectivity, such that they end up spending more time in gender-segregated groups than women do.
\begin{table}[h]
\caption{Gender-specific difference in location diversity (Z-score). We observe that women have significantly higher location diversity than men. This observation persists in three semesters. ($H_{0}$ rejected at 2 sigmas)}
\resizebox{4in}{!}{
\begin{tabular*}{\textwidth}{c @{\extracolsep{\fill}} lcccc}
\hline \hline
&Semester 1 & Semester 2 & Semester 3 \\
\hline
Men& {\bf -2.183} & {\bf -1.958} & {\bf -2.250} \\
Women& {0.052} & {0.031} & {-0.057} \\
\hline
\end{tabular*}
}
\label{tab_location_diversity}
\end{table}
\subsection{Gender Differences in Mobility Behaviors}
\label{sec_location}
\subsubsection{Location Diversity}
In this section we address the question of whether we can observe any gender differences in location diversity patterns. Table~\ref{tab_location_diversity} shows the results. We find that on the whole women display higher levels of location diversity than men. This mirrors in terms of spatial mobility behavior the way that women distribute their communication and face-to-face interaction time across alters: in both of these respects women are less ``predictable'' than men. Note that this result has the implication that men not only spend more time interacting in gender-segregated groups with a relatively smaller number range of individuals, but that they may do so in a smaller set of geographic locations than women, including being more likely to use the same-gender segregated spatial settings repeatedly. In this respect, there seems to be an empirical connection between the male pattern of gender-segregated spatial interaction and their relatively more predictable patterns of spatial mobility.
\begin{figure*}[ht]
\centering
\centerline{\includegraphics[width=4in]{mobility_gender_S3.jpg}}
\caption{Location Stay Density of Women and Men. Green indicates men, and pink indicates women. The higher brightness indicates larger concentration density.}
\label{fig_gender_density}
\end{figure*}
Figure~\ref{fig_gender_density} provides a visual representation of this effect, overlaid against an actual representation of the campus spatial layout. The figure provides clear evidence of the relative concentration of spatial mobility for men and the relative dispersion of mobility for women. On the whole men are more likely to be found frequenting the same locations on campus for longer periods of time. Women on the other hand disperse their activities across a wider range of locations, leading them to spend relatively smaller amounts of time at any one location.
\begin{table}[h]
\caption{Gender-specific location visiting validation (Z-score). We examine the significance of three types of location visiting similarity, MM, WW and MW/WM.}
\resizebox{4in}{!}{
\begin{tabular*}{\textwidth}{c @{\extracolsep{\fill}} lccccc}
\hline \hline
&\multicolumn2c{Semester 1} & \multicolumn2c{Semester 2} & \multicolumn2c{Semester 3} \\
Types of Pairs\\
& All Pairs & Connected & All Pairs & Connected & All Pairs & Connected \\
\hline
MM & {\bf 4.136} & {\bf 4.718} & {\bf 3.913} & {\bf 4.108} & {\bf 3.077} & {\bf 3.450} \\
WW& {0.548} & {1.701} & {-1.208} & 0.590 & {-1.731} & {-0.097} \\
MW/WM & {\bf -10.229} & {\bf -8.730} & {\bf -9.096} & {\bf -5.825} & {\bf -6.054} & {\bf -5.345} \\
\hline
\end{tabular*}
}
\label{tab_location_homo}
\end{table}
\subsubsection{Similarity in Location Signature by Dyadic Gender Mix}
Are same-gender dyads more likely to share mobility habits than cross gender dyads? To answer this question, we develop a measure of dyadic similarity in location signatures by computing the Pearson correlation coefficient $r_{uv}$ between the location signature vectors of each individual at each time period. We then use the null model validation method described above to ascertain whether this correlation is higher than one would expect by chance across the same gender (MM, WW) and across different genders (MW/WM) dyads. We generate validation results both for all pair of individuals (All Pairs) and for individuals who are connected in the Voice Mail or the Text Message communication networks (Connected). Consistent with the results reported above, we expect to find higher levels of location signature similarities among all-men dyads (MM) than in the other gender-based dyadic combinations.
The results are shown on Table~\ref{tab_location_homo}. In support of our male similarity hypothesis, we find that members of all-men dyads tend to display higher levels of location signature similarity than we would expect by chance. This location signature similarity pattern emerges in the first semester, and can be found in all subsequent time periods. For the all-women dyad, location signature similarities are close to what we would expect under a random-pairing regime. For cross-gender dyads, on the other hand, we find a smaller correlation between location signature pattern than predicted under a random pairing regime.
This pattern of results is the same irrespective of whether we compare all pairs of individuals or only those directly connected in the communication network. This suggests that the location signature correlation is independent of social connectivity (it is not the result of a direct social influence process) but may reflect common underlying dispositional attributes correlated with both the odds of individuals forming a social connection and their underlying mobility habits. These results reinforce the impression that men tend to have more predictable and have less diverse mobility habits than women. Men also tend to couple mobility and proximity behaviors in a strong way. In addition, the fact that men and women tend to have such distinct location signature patterns, means that any cross-gender pair drawn at random from the population is likely to have negatively correlated location signature patterns, whether they share a direct social link or not.
\section{Discussion}
\label{sec_discussion}
In this paper we have examined gender differences in dispositional factors relevant for the formation and maintenance of social networks. We attempted to isolate dispositional differences between men and women by studying a population in a relatively closed context at the same stage of the life course (the transition to college) before the onset of life-transitions that tend to exacerbate structurally induced (non-dispositional) differences in the social connectivity behavior between men and women (e.g. women's greater reliance on kin; men's greater likelihood to connect to co-workers). We also extended previous work by looking at patterns of behavior (e.g. distributional and proximity tendencies) where gender differences have yet to be documented systematically.
Building on previous work that has looked at gender differences in social networks with unobtrusively collected data, we have examined gender differences across both communication (cell phone and text messaging) and mobility (spatial proximity) networks. Our results are consistent with (and thus validate in a natural setting) previous work done on MCNs and SNSs \cite{lewis2008, kovanen2013, stoica2010, dong2014, palchykov2012}. They are also consistent with an emerging line of research that suggests that gender differences in behavioral dispositions associated with mobility and spatial proximity behaviors differ in significant ways than those associated with communication behavior \cite{fournet2014, stehle2013}. The key implication of our study therefore is that when studying gender differences in social networks, researchers should be wary of extending conclusions reached when studying one realm of social behavior (e.g. communication) into an unrelated realm (spatial proximity patterns).
For instance, our analysis of same-gender preference in both (voice call and SMS) of our communication networks turned up the now well-documented pattern of stronger same-gender preference for women, but the gender differences were more pronounced in the binary representation of the network than when call volume was taken into account. Most importantly, however, we found the \emph{opposite} same-gender preference pattern in the proximity network, with men displaying stronger behavioral patterns of same-gender proximity than women. These patterns emerged early on (after the first semester) and remained over a long temporal scale (one year). Gender differences in spatial proximity behavior also extended to the \emph{distribution} of face-to-face interaction across contacts. In all, men were more likely to unevenly distribute their face time in favor of same-gender alters, and were more likely to concentrate their face-to-face time on a single contact. This contrasts with the results obtained in the communication (SMS) network, which showed that women were more likely to concentrate their communication behavior on a few contacts than men.
Related to the tendency of men to concentrate their face-to-face interaction on same gender contacts, we found that there is a stronger linkage between communication and subsequent time spent in close proximity for men but not for women, suggesting that men tend to couple their face-to-face time to with their previous communicative interactions in a stronger way than women. The tendency of men to form gender-segregated ``bands'' that spend a lot of time in close proximity affect the predictability of their mobility behavior: On the whole, we found that men's mobility behavior (as given by the time spent across a set of demarcated locations) is more predictable than that of women because it tends to concentrate itself in a smaller number of locations. In accordance with this result our analysis reveals that there is a stronger level of concordance in spatial mobility habits in all-male dyads than there is in mixed-gender or all-women dyads.
In all the results reported in this paper paint a picture of substantial differences in the mobility and spatial proximity behavior of men and women. These differences stand in stark contrast to the relatively small gender-differences found when looking at purely communicative interactions, supporting the view that communication and exchange networks should not be analyzed in isolation from their embedding in a spatial context \cite{doreian2012}.
The greater likelihood of men's proximity and spatial behavior, and the greater-similarity of all-male dyads in mobility habits is consistent with previous sociological work. These studies note that men's friendship same and cross-gender selection and behavior is driven by more by dispositional factors than those of women, who seem be more likely to respond to the opportunity structure of contacts \cite{kalmijn2002}. Our results validate a long line of ethnographic observational research showing the stronger effect of same-gender peer groups constructed around such activities as games and sports on the ability of men to construct coherent masculine identities especially during key transition points in the life course \cite{messner1990, maccoby1988, smith1993}. Our results are also consistent with social psychological perspectives that point to the strong affinity of males, beginning at young ages, to interact in relatively larger groups in specific settings outside the household while strongly avoiding cross-gender interactions in the same settings \cite{ridgeway1999, mehta2009}.
Our finding of greater spatial homophily among men is in line with dispositional accounts of gender differences in friendship patterns at the intersection of the social and biological sciences (e.g. evolutionary psychological frameworks). This line of work is beginning to move beyond traditional scholarly understandings of women as predisposed to be more ``social'' than men \cite{cross1997}, and towards a more comprehensive account that sees men and women as predisposed to distinct forms of sociality and intimacy \cite{baumeister1997, benenson2009}. These researchers point to the ubiquity and functional role (e.g. mutual protection, resource hoarding, etc.) of stronger dispositions towards homosociality among men across cultures and historical settings \cite{benenson2014}. This is evidenced by the greater likelihood of observing ``coalitions'' of unrelated males in close physical proximity especially in public spaces \cite{rodseth2000, vigil2007}; the greater investment (e.g. reported enjoyment and interaction preferences) displayed by males (starting early in life) same-gender peer relationships centered in specific spatial contexts, such as neighborhoods and classrooms \cite{benenson1998}; the greater likelihood that same-gender relationships among men of being embedded in a larger group context \cite{benenson2003}; women's greater preference for interaction in dyads rather than groups \cite{benenson1997}; the greater dispositional capacity of males to resolve conflict with same-gender peers in comparison to females \cite{benenson2014}; and greater likelihood of same-gender best-friendships among males to endure longer and to survive in the face of relationship conflict and strain \cite{benenson2009} .
The basic empirical implication of these perspectives is that we should expect to find predictable \emph{asymmetries} in the way that men and women structure the temporal and spatial organization of their same-gender peer relations \cite{vigil2007}. While men prefer socialize in relatively large, function-specific groups that interact in public settings, women prefer to interact across emotionally intimate dyads socializing in private settings. The asymmetry hypothesis reconciles the discrepant findings of stronger same-gender preference and greater unevenness in the distribution of communications tendencies in the SMS network for women and the stronger same-gender preference and greater unevenness distribution of spatial interactions in the proximity network for men. In all, our results provide striking support for the proposition that there is an intimate connection between social and communicative behavior and mobility and proximity patterns but that the empirical linkage between these is governed by cultural, personality, and possibly biological factors (glossed under the dispositional view) that differ systematically between men and women.
\bibliographystyle{nws}
|
\section{Introduction}
The particle-in-cell (PIC) method has been a standard tool in the simulation of kinetic plasmas for over 50 years \cite{birdsall1969clouds,buneman1963computer,dawson1962one,langdon1970theory,morse1970multidimensional}, and continues to be popular today - see e.g.\ \cite{birdsall2004plasma,chen2013energy,fubiani2015developpment, garrigues2015characterization,philippov2014ab,taitano2013development,tenbarge2014collisionless}, among many other references. Its numerous desirable features explain in large part its success - it is conceptually simple, readily parallelizable, relatively robust, and can incorporate a wide variety of physical phenomena. Even so, PIC simulations of the complex, three-dimensional systems that arise in modern plasma physics applications still require many hours on a massively parallel machine \cite{exum2013ppps, fiuza2013ion, fonseca2013exploiting, fubiani2015developpment, wang2015modern}.
There are several reasons for these enormous run times, but prominent among them is the statistical error introduced by the particle representation. Good statistical resolution requires many particles \textit{per cell}, and in two or three dimensions, the number of cells may be very large in order to properly resolve the features of the system under study. The combination of these two requirements forces the total number of particles to be overwhelmingly large, making for a computation that requires both huge CPU and memory resources.
The result of this interaction between the grid and particles requirements is an algorithm with two of the worst features of both particle and continuum methods: One is not only saddled with the unavoidably slow convergence of a particle scheme, but also complexity that grows exponentially with dimension, like a continuum method. While the reduction in dimension - from six to three - makes this a worthwhile sacrifice in many contexts, further acceleration of PIC simulations is clearly of great practical interest.
In this paper, we propose the use of so-called `sparse grids' in concert with PIC to obtain a scheme whose complexity is nearly independent of dimension. The idea of a sparse grid is not new, having been studied extensively in the applied mathematics community \cite{bungartz2004sparse,garcke2001data,griebel1998adaptive,griebel1990combination}. It has also seen some use in continuum codes for plasma applications \cite{ali2015fault,guo2016sparse,heene2013load,hinojosa2015towards,kowitz2013sparse,kowitz2012combination,pfluger2014exahd}.
While most of the sparse grid literature deals with a hierarchical-basis representation \cite{bungartz2004sparse}, here we will make use of an alternative characterization called the `combination technique' \cite{griebel1990combination}.
Specifically, a function - e.g. charge density or current - is approximated on a variety of different grids, each of which has a different resolution in each coordinate direction. By combining these approximations intelligently, one can achieve accuracy close to that of a well-resolved regular grid, but at dramatically reduced cost.
Crucially for PIC, the combination technique grids have very large cells relative to a comparable regular grid. This improves statistical resolution by increasing the number of particles \textit{per cell} without increasing the overall particle number. The use of sparse grids can thus accelerate not only the computation of the electromagnetic fields, but also the particle operations, which typically dominate the computation and storage requirements. The key to achieve this end is to reinterpret the particle shape function typically used in PIC as an approximation to a delta function rather than a physical charge density.
The remainder of the article is structured as follows. In section 2, we present the necessary background on the combination technique and particle-in-cell methods in general. In section 3, we present our reinterpretation of PIC methodology and the use of sparse grids that it enables. In section 4, we proceed to results of numerical experiments in 2- and 3-D. In section 5, we discuss our results, prospects for future work, and conclusions.
\section{Prerequisites}
In this section, we establish the necessary background on the combination technique for sparse grids and PIC schemes.
\subsection{Sparse grid combination technique}
While the combination technique can be applied to problems of any dimensionality greater than one, we will discuss the key ideas in the context of two dimensional simulations for the simplicity of the presentation. We refer the interested reader to previous work \cite{griebel1990combination,griebel1992combination,hegland2007combination,reisinger2012analysis} for a more complete treatment. Furthermore, even though the sparse grids are not limited to approximation schemes based on multi-linear bases \cite{achatz2003higher,bungartz2004higher,bungartz2004sparse}, we restrict for simplicity our presentation here to multilinear bases, and leave the possibility of PIC schemes with higher-order bases on sparse grids to future work.
Consider, then, a function $u(x,y)$ on the unit square $[0,1] \times [0, 1]$ that we wish to approximate on a rectangular grid with cell width $h_x$ and height $h_y$. Two simple circumstances of this type are (a) $u$ is the solution of an elliptic PDE that we solve with bilinear finite elements, and (b) $u$ is known at the grid points and we wish to approximate it on the entire domain via bilinear interpolation. For these and many other approximation schemes based on a piecewise bilinear representation of a smooth function, the error between the exact function $u$ and its approximation $\mathfrak{u}$ at a particular point can be written has
\begin{equation} \label{uerr}
u(x,y) - \mathfrak{u}(x,y) = C_1(h_x) h_x^2 + C_2(h_y) h_y^2 + C_3(h_x,h_y) h_x^2 h_y^2.
\end{equation}
The $C_i$ above are functions with a uniform upper bound. Importantly, by allowing the $C_i$ to be arbitrary functions of their arguments, they can be chosen so that (\ref{uerr}) is not simply the leading terms in an asymptotic expansion, but rather the \textit{exact} error \cite{griebel1990combination}.
In the absence of additional information about $u$, one typically chooses $h_x = h_y = h$ and finds an error of $O(h^2)$. In the best case scenario, the computational complexity of the scheme scales linearly in the number of grid points, giving a complexity that scales like $O(h^{-2})$. Thus, the complexity $\kappa$ is related to the error $\varepsilon$ by
\begin{equation}
\kappa \sim \varepsilon^{-1}.
\end{equation}
In arbitrary dimension $d$, the best case computational complexity scales as $O(h^{-d})$, meaning the above generalizes to $\kappa \sim \varepsilon^{-d/2}$.
The combination technique improves on this circumstance by using cancellation across different grids. Suppose the desired resolution is $h_n = 2^{-n}$ for some positive integer $n$. Let $h_x^i = 2^{-i}$, $h_y^j = 2^{-j}$ and $\mathfrak{u}_{i,j}$ be the approximation of $u$ on the corresponding grid. Then, consider the quantity $\mathfrak{u}_n$ defined by
\begin{equation} \label{combo}
\mathfrak{u}_n = \sum_{i+j=n+1} \mathfrak{u}_{i,j} - \sum_{i+j=n} \mathfrak{u}_{i,j}.
\end{equation}
In each of the sums, $i$ and $j$ are strictly positive integers. This combination is depicted graphically in fig.\ \ref{combdiag}.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.4\textwidth]{SGdiagram1.png}
\caption{A graphical depiction of the combination of grids used in (\ref{combo}). The green `+' signs represent grids that give a positive contribution, while red `$-$' signs are subtracted. Cancellation arises from pairing neighboring grids along vertical and horizontal axes.}
\label{combdiag}
\end{center}
\end{figure}
By considering figure 1 and the error formula (\ref{uerr}), we see that a great deal of cancellation occurs in computing the error corresponding to $\mathfrak{u}_n$. In particular, for any particular $i$ between $1$ and $n-1$, a grid with horizontal spacing $h_x^i$ appears exactly once in each of the two sums in (\ref{combo}). For those two grids, the term $C_1(h_x) h_x^2$ that appears in (\ref{uerr}) exactly cancels, because it is independent of $h_y$. The only contribution from the $O(h_x^2)$ term thus comes from the grid with $h_x = 2^{-n}$. An analogous argument holds in the $y$-direction. We thus find
\begin{equation}
u - \mathfrak{u}_n = C_1(h_n) h_n^2 + C_2 (h_n) h_n^2 + h_n^2 \left\{\frac{1}{4} \sum_{i+j=n+1} C_3(h_x^i,h_y^j) - \sum_{i+j=n} C_3(h_x^i,h_y^j) \right\},
\end{equation}
where we have used the fact that $h_x^i h_y^j = h_n/2$ when $i+j = n+1$ and $h_x^i h_y^j = h_n$ when $i+j = n$. The expression in braces contains $2n-1$ terms which are all uniformly bounded by constants, so we find that
\begin{equation}
|u - \mathfrak{u}_n| = O(n h_n^2) = O(h_n^2 |\log h_n|).
\end{equation}
That is, $\mathfrak{u}_n$ approximates $u$ \textit{nearly as well} as an approximate solution using $h_x = h_y = 2^{-n}$. Even better, each grid used in the combination technique has $O(h_n^{-1})$ grid points, and there are $O(n)$ grids. Thus, for schemes that scale with the number of grid nodes, one has $\kappa = O(h_n^{-1} | \log h_n |)$. Using the fact that the logarithm grows slower than any polynomial, we thus find a new relationship between complexity $\kappa$ and error $\varepsilon$:
\begin{equation} \label{sparsecomplexity}
\kappa \sim \varepsilon^{-1/2} |\log \varepsilon|^{2},
\end{equation}
At least asymptotically, one can thus achieve the same accuracy considerably faster with the combination technique than with a single regular grid.
This idea extends naturally to higher dimensions. In three dimensions, for example, if we let $h_z^k = 2^{-k}$, then
\begin{equation}
\mathfrak{u}_n = \sum_{i+j+k = n+2} \mathfrak{u}_{i,j,k} - 2 \sum_{i+j+k=n+1} \mathfrak{u}_{i,j,k} + \sum_{i+j+k = n} \mathfrak{u}_{i,j,k}
\end{equation}
gives a sparse approximation of $u$ - see \cite{griebel1990combination} for more detail. In general, a $d$-dimensional function is represented with error $O(h_n^2 |\log h_n|^{d-1})$ and complexity $O(h_n^{-1} |\log h_n|^{d-1})$. The scaling of complexity with error shown in (\ref{sparsecomplexity}) thus generalizes to
\begin{equation}
\kappa \sim \varepsilon^{-1/2} |\log \varepsilon|^{2(d-1)}.
\end{equation}
For very high dimensional functions, this is an enormous savings over the $\kappa \sim \varepsilon^{-d/2}$ scaling for regular grids.
Of course, this technique is not without its limitations. Firstly, it requires a structured grid. An unstructured triangular mesh, for instance, has no natural sense of $h_x$ and $h_y$ that can be refined independently, making it unclear how these ideas might be applied. That being said, it is not necessary that the domain be rectangular. It is only necessary that the domain can be mapped to a rectangle (or a sequence of connected rectangles \cite{bungartz2004sparse}) via some coordinate transformation.
A second drawback is the increased regularity demanded of the function $u$. In particular, for a regular grid, the leading order constants in the error - $C_1$ and $C_2$ - are proportional to the second derivatives $u_{xx}$ and $u_{yy}$. In contrast, the leading term in the sparse grid error - related to $C_3$ - is proportional to the fourth-order mixed derivative $u_{xxyy}$. As a result, the constant in front of the error scaling will often be larger for a sparse grid than a regular grid, and the predicted scaling may not even apply if $u$ is not sufficiently smooth. Local adaptive refinement techniques have successfully mitigated this difficulty in some scenarios \cite{griebel1998adaptive}.
\subsection{Particle-in-cell}
For simplicity, we restrict our presentation to a relatively basic physical situation in which PIC applies, namely the electrostatic Vlasov equation with fixed magnetic field for the electrons, and ions assumed to form a uniform, immobile, neutralizing background:
\begin{equation} \label{ESVlasov}
\begin{split}
&\partial_t f + \mathbf{v} \cdot \nabla_x f + q (\mathbf{E} + \mathbf{v} \times \mathbf{B} )\cdot \nabla_v f = 0, \\
&\mathbf{E} = -\nabla \varphi, \qquad -\nabla^2 \varphi = q \int f \, d\mathbf{v} + \rho_i.
\end{split}
\end{equation}
In the above, $f(\mathbf{x},\mathbf{v},t)$ is the electron phase-space distribution, $q$ the electron charge, and $\rho_i$ the constant charge density of the ions.
Traditionally, a PIC scheme for (\ref{ESVlasov}) is comprised of four features:
\begin{enumerate}[i.]
\item Represent $f$ as a collection of simulated ``particles" $(\mathbf{x}_p, \mathbf{v}_p)$, which are evolved via some finite difference scheme for
\begin{equation} \label{ParPush}
\dot{\mathbf{x}}_p = \mathbf{v}_p, \qquad \dot{\mathbf{v}}_p = q (\left. \mathbf{E} + \mathbf{v} \times \mathbf{B} )\right|_{\mathbf{x} = \mathbf{x}_p}
\end{equation}
\item Assign to each particle a charge density $S(\mathbf{x} - \mathbf{x}_p)$, and approximate the overall electron charge density at grid points $\mathbf{x}_k$ via
\begin{equation} \label{PICparadigm}
\rho_e(\mathbf{x}_k) \approx \varrho_e (\mathbf{x}_k) \coloneqq \sum_p S(\mathbf{x}_k - \mathbf{x}_p).
\end{equation}
\item Use some grid-based Poisson solver to compute $\varphi$ by solving $-\nabla^2 \varphi = \rho_e + \rho_i$, then evaluate $\mathbf{E}$ by numerical differentiation.
\item Evaluate $\mathbf{E}$ at the particle positions $\mathbf{x}_p$ by interpolation. Use this to repeat step i.
\end{enumerate}
One loop through each of these four phases represents a single time step. We will denote the total number of simulated particles by $N_p$. For simplicity, we will assume the spatial domain is rectangular and the use of a regular Cartesian grid with cell width $h_x$, $h_y$, and $h_z$ in the coordinate directions.
Let us consider the simplest practical PIC scheme, which uses a leapfrog or Boris scheme for the particle push and a second-order finite element scheme for the field solve. Typically, the grid spacing is approximately equal in each coordinate direction - call it $h$. The overall error $\varepsilon$ of the scheme has the following scalings \cite{birdsall2004plasma}:
\begin{equation} \label{stderrscalings}
\underbrace{\varepsilon \sim \Delta t^2}_{\substack{\text{Time-stepping}\\\text{error}}}, \qquad \underbrace{\varepsilon \sim h^2}_{\substack{\text{Grid}\\\text{error}}}, \qquad \underbrace{\varepsilon \sim (N_p h^{d_x})^{-1/2}}_{\substack{\text{Particle sampling}\\\text{error}}}.
\end{equation}
Observe that an ideal particle method has a sampling error that scales like $N_p^{-1/2}$, but PIC's sampling error scales not with the total particle number, but the number of particles \textit{per cell}. A small grid error requires $h^2 \ll 1$, but small sampling error requires not only $N_p \gg 1$ but also the more onerous condition $N_p \gg h^{-d_x}$.
It is also evident that $N_p$ must grow rapidly as $d_x$ increases. This point is further emphasized by looking at the complexity of the particle based operations - the particle push and interpolation to/from the grid. The complexity $\kappa$ scales with $N_p / \Delta t$. A simple calculation leads from (\ref{stderrscalings}) to
\begin{equation}
\kappa \sim \varepsilon^{-(2.5 + d_x/2)}.
\end{equation}
Evidently, a 2-D PIC scheme is more expensive than a 1-D scheme by a factor of $1/\sqrt{\varepsilon}$, while the complexity of a 3-D scheme is increased by $1/\varepsilon$ compared to 1-D.
\subsubsection{The Shape Function}
This work is concerned with the use of sparse grid ideas, which necessitate a reinterpretation of the shape function $S$ - which maps the particle properties onto the grid - and the errors associated with that process. To help clarify our novel conception of $S$, we first describe here the usual interpretation.
While a variety of choices exist, we will assume for simplicity that $S$ is built from the so-called \textit{tent function} $\tau$, defined by
\begin{equation}
\tau(x) = \left\{ \begin{array}{ccr} 1 - |x| & : & |x| \leq 1 \\ 0 & : & |x| > 1 \end{array} \right.
\end{equation}
$S$ is then given by
\begin{equation}
S(\mathbf{x}) = \frac{Q}{N_p} \prod_{i=1}^{d_x} \frac{\tau \left(x_i / h_i \right)}{h_i},
\end{equation}
where $h_i$ is the cell width in the $i^\textrm{th}$ coordinate direction. The scaling factor $Q$ is interpreted as the total (electron) charge in the system, and this choice of $S$ fixes that value independent of $N_p$. Physically, if the total number of \textit{physical} particles in the system is $\mathcal{N} = Q / q$, each simulated particle is interpreted as a \textit{macroparticle} with charge $q \mathcal{N} / N_p$ that represents \textit{many} physical particles.
With these standard concepts set, we are now ready to revisit the PIC method in the context of sparse grids and the combination technique.
\section{Merging PIC with Sparse Grids}
In this section, we show how PIC may be used in concert with the combination technique to obtain a scheme with complexity that depends only logarithmically on dimension. We proceed first by reinterpreting the usual PIC shape function $S$. We then carry out a formal error and complexity analysis. We conclude the section by summarizing the new algorithm.
\subsection{Reinterpreting the shape function}
Define $\bar{f} = f / \mathcal{N}$. Note that $\bar{f}$ is non-negative and its phase space integral is unity. It may thus be interpreted as a probability density. Moreover, by definition,
\begin{equation}
\rho_e(\mathbf{x}) = q \int f \, d\mathbf{v} = q\mathcal{N} \int \bar{f}(\tilde{\mathbf{x}},\mathbf{v}) \delta(\mathbf{x} - \tilde{\mathbf{x}}) \, d\mathbf{v} d\tilde{\mathbf{x}} = Q \mathbb{E}_{\bar{f}(\tilde{\mathbf{x}},\mathbf{v})} [ \delta(\mathbf{x} - \tilde{\mathbf{x}}) ],
\end{equation}
where we have introduced notation for the expected value over the probability density $\bar{f}(\tilde{\mathbf{x}},\mathbf{v})$.
Of course, the mean of a Dirac delta function is an entirely formal concept and impossible to compute from particle data. With the introduction of a spatial grid, however, it is natural to introduce an approximate delta function - we choose
\begin{equation}
\delta(\mathbf{x}) \approxeq \mathcal{S} (\mathbf{x}) \coloneqq \prod_{i=1}^{d_x} \frac{\tau \left(x_i / h_i \right)}{h_i}.
\end{equation}
We note that the error in approximating the delta function by $\mathcal{S}$ can be written as
\begin{equation} \label{Serrorexpansion}
\rho_e(\mathbf{x}) - Q \mathbb{E}_{\bar{f}(\tilde{\mathbf{x}},\mathbf{v})}[\mathcal{S}(\mathbf{x} - \tilde{\mathbf{x}})] = C_1(h_x) h_x^2 + C_2(h_y) h_y^2 + C_3(h_x,h_y) h_x^2 h_y^2
\end{equation}
in two dimensions, while a directly analogous expression holds in 3-D. This can be seen by simple Taylor expansion of $\bar{f}$ in position space. As before, the $C_i$ are functions with a uniform upper bound and can be chosen to make this expression a true equality. Already, this bears a promising resemblance to (\ref{uerr}). Moreover, we note that any of the B-splines commonly used as shape functions for PIC will have an error of this form - indeed, any non-negative $\mathcal{S}$ can do no better than second order in this sense. The error given in \eqref{Serrorexpansion} may be interpreted as the grid error of the PIC method, as will be more apparent from our error analysis in Section \ref{sec:error}
Of course, in PIC, we cannot evaluate $\mathbb{E}_{\bar{f}(\tilde{\mathbf{x}},\mathbf{v})} [\mathcal{S}]$ exactly. It is approximated by a sum over the particles, which in our probablistic interpretation are regarded as samples from the probablity density $\bar{f}$. That is, we approximate $\rho_e$ at each grid point $\mathbf{x}_k$ by
\begin{equation} \label{reinterp_approx}
\rho_{e} (\mathbf{x}_k) \approx Q\mathbb{E}_{\bar{f}(\tilde{\mathbf{x}},\mathbf{v})} [\mathcal{S}(\mathbf{x}_k - \tilde{\mathbf{x}})] \approx \frac{Q}{N_p} \sum_{p} \mathcal{S} (\mathbf{x}_k - \mathbf{x}_p) = \varrho_e(\mathbf{x}_k).
\end{equation}
For our $\mathcal{S}$, this approximation of $\rho_e$ is \textit{identical} to the standard viewpoint of (\ref{PICparadigm}). The difference between both sides of \eqref{reinterp_approx} may be interpreted as the statistical or particle error of the PIC method, as our analysis will clearly show in the next section.
\subsection{Formal error analysis}\label{sec:error}
Let us now assume that the initial particle states have been chosen by independent sampling from $\bar{f}(t=0)$ - a so-called \textit{noisy start}.
When $N_p \gg 1$, each particle has a small influence on the bulk field used to push the particles, meaning individual particle states remain approximately independent. We may thus regard our approximation of $\rho_e(\mathbf{x}_k)$ as a random variable with mean $Q \mathbb{E}_{\bar{f}(\tilde{\mathbf{x}},\mathbf{v})} [\mathcal{S}(\mathbf{x}_k - \tilde{\mathbf{x}})]$ and variance given by $Q^2 \textrm{Var}_{\bar{f}(\tilde{\mathbf{x}},\mathbf{v})} [\mathcal{S}(\mathbf{x}_k - \tilde{\mathbf{x}})] / N_p$.
Taylor expanding $\bar{f}$ and evaluating simple integrals gives an expression for $\textrm{Var}_{\bar{f}(\tilde{\mathbf{x}},\mathbf{v})} [\mathcal{S}(\mathbf{x}_k - \tilde{\mathbf{x}})]$. Namely, we have
\begin{equation}
Q^2\textrm{Var}_{\bar{f}(\tilde{\mathbf{x}},\mathbf{v})} [\mathcal{S}(\mathbf{x}_k - \tilde{\mathbf{x}})] \approx \frac{4/9}{h_x h_y} Q\rho_e(\mathbf{x}_k)
\end{equation}
to leading order. With this, we find that
\begin{equation}
\rho_e(\mathbf{x}_k) - \varrho_e(\mathbf{x}_k) = C_1(h_x) h_x^2 + C_2(h_y) h_y^2 + C_3(h_x,h_y) h_x^2 h_y^2 + \xi_k,
\end{equation}
where $\xi_k$ is a random variable with
\begin{equation} \label{varprops}
\mathbb{E}[\xi_k] = 0, \qquad \mathrm{Var}[\xi_k] \approx \frac{4Q \rho_e(\mathbf{x}_k)}{9} \frac{1}{h_x h_y N_p}.
\end{equation}
As usual, a directly analogous expression holds in three dimensions.
From here, we can directly see the grid and particle error scalings that were quoted in (\ref{stderrscalings}). The introduction of $\mathcal{S}$ to approximate the Dirac delta function has introduced an $O(h^2)$ error, while approximation of an integral by a sum over $N_p$ particles has led to a random error of size $O(1/\sqrt{h_x h_y N_p})$. Moreover, when $\varrho_e$ is extended to the entire computational domain by bilinear interpolation, exactly the same expression for the error holds because the interpolation error has the same form as the existing grid error.
We are now in a position to leverage the sparse grid ideas outlined in section 2.1. For notational simplicity, we once again assume the computational domain is $[0,1] \times [0,1]$, but all results are easily generalizable to an arbitrary rectangle. Fix the particle number $N_p$ and let $\varrho_{i,j}$ denote the approximation of $\rho_e$ computed using $h_x = 2^{-i}$ and $h_y = 2^{-j}$ in (\ref{reinterp_approx}) and extended to the entire domain via bilinear interpolation. Then
\begin{equation}
\varrho_n = \sum_{i+j = n+1} \varrho_{i,j} - \sum_{i+j=n} \varrho_{i,j}
\end{equation}
takes advantage of many of the same cancellations described in section 2.1. In particular,
\begin{equation}
\begin{split}
\rho_e - \varrho_n &= h_n^2 \left\{C_1(h_n) + C_2(h_n) + \frac{1}{4} \sum_{i+j=n+1} C_3(h_x^i,h_y^j) - \sum_{i+j=n} C_3(h_x^i,h_y^j) \right\} \\
&+ \sum_{i+j = n+1} \xi_{i,j} - \sum_{i+j=n} \xi_{i,j}
\end{split}
\end{equation}
As in standard versions of the combination technique, the term in braces is $O(n) = O(|\log h_n|)$. Moreover, the mean-square size of the random error is
\begin{equation}
\mathbb{E} \left[ \left( \sum_{i+j = n+1} \xi_{i,j} - \sum_{i+j=n} \xi_{i,j} \right)^2 \right] = O \left(n^2 (N_p h_n)^{-1} \right)
\end{equation}
since the left side contains $(2n-1)^2$ terms of the form
\begin{equation}
\begin{split}
\mathbb{E} \left[ \xi_{i,j} \xi_{i',j'} \right] &\leq \left( \textrm{Var}[\xi_{i,j}] \textrm{Var}[\xi_{i',j'}] \right)^{1/2} \\
&= O\left( (N_p h_x^i h_y^j)^{-1/2} \right) \cdot O \left( (N_p h_x^{i'} h_y^{j'})^{-1/2} \right) \\
&= O \left( (N_p h_n)^{-1} \right)
\end{split}
\end{equation}
by the Schwarz inequality and the definitions of $h_x^i$ and $h_y^j$. Thus, we find that the root-mean-square particle sampling error is now $O((N_p h_n)^{-1/2} |\log h_n |)$.
Thus we find the overall error in the approximation of $\rho_e$ has the scaling
\begin{equation}
|\rho_e - \varrho_n| = \underbrace{O(h_n^2 |\log h_n|)}_{\textrm{Grid error}} + \underbrace{O\left( \frac{|\log h_n|}{\sqrt{h_n N_p}}\right)}_{\textrm{Particle error}}
\end{equation}
for 2-D sparse grids, and the only difference in 3-D is an extra power of $|\log h_n|$ in each term. This gives rise to a sparse version of the standard PIC scalings in (\ref{stderrscalings}):
\begin{equation}
\underbrace{\varepsilon \sim \Delta t^2}_{\substack{\text{Time-stepping}\\\text{error}}}, \qquad \underbrace{\varepsilon \sim h_n^2 |\log h_n|^{d_x - 1}}_{\text{Grid error}}, \qquad \underbrace{\varepsilon \sim |\log h_n|^{d_x-1} (N_p h_n)^{-1/2}}_{\text{Particle sampling error}}.
\end{equation}
If we again assume that particle steps dominate the complexity of the PIC scheme, then we find that the complexity scales as
\begin{equation}
\kappa \sim \frac{N_p}{\Delta t} \sim \varepsilon^{-3} | \log \varepsilon|^{3(d_x - 1)}
\end{equation}
We see that, in contrast to standard PIC on regular grids, the complexity is \textit{nearly} independent of dimension, in that the complexity only increases by logarithmic powers of the desired error as dimension increases.
The improvement over standard PIC can be intuitively understood in the following way. The figure of merit for the statistical error in a PIC scheme is the number of particles \textit{per cell} - $N_p / N_c$, where $N_c$ denotes the number of cells. Since $N_c$ scales inversely with the cell volume, on a regular grid, $N_c \sim h^{-3}$ in 3-D. However, on a 3-D sparse grid, $N_c \sim (4 h)^{-1}$. We thus achieve \textit{many} more particles per cell, even with the total particle number fixed, by using a sparse grid.
\subsection{Algorithm outline}
Using PIC in concert with the combination technique functions very much like standard PIC as outlined in section 2.2. The analogous outline - in 2-D - is:
\begin{enumerate}[i.]
\item Push particles exactly as in standard PIC.
\item Assign to each particle a sequence of shape functions $\mathcal{S}_{i,j}(\mathbf{x} - \mathbf{x}_p) = \tau(2^i (x-x_p)) \tau(2^j (y-y_p)) / 2^{i+j}$, and approximate the overall electron charge density via
\begin{equation} \label{sparsePICparadigm}
\rho_e \approx \varrho_e \coloneqq \sum_{i+j = n+1} \varrho_{i,j} - \sum_{i+j=n} \varrho_{i,j},
\end{equation}
where $\rho_{i,j}$ is defined at grid points $\mathbf{x}_{k,\ell} = (k 2^{-i}, \ell 2^{-j})$ by
\begin{equation}
\varrho_{i,j}(\mathbf{x}_{k,\ell}) = \frac{Q}{N_p} \sum_{p} \mathcal{S}_{i,j} (\mathbf{x}_p - \mathbf{x}_{k,\ell})
\end{equation}
and extended to the entire domain using bilinear interpolation.
\item Use some grid-based Poisson solver to compute $\varphi_{i,j}$ by solving $-\nabla^2 \varphi_{i,j} = \varrho_{i,j} + \rho_i$, then evaluate $\mathbf{E}_{i,j}$ by numerical differentiation of $\varphi_{i,j}$.
\item Evaluate $\mathbf{E}$ at the particle positions $\mathbf{x}_p$ via
\begin{equation}
\mathbf{E}(\mathbf{x}_p) = \sum_{i+j=n+1} \mathbf{E}_{i,j}(\mathbf{x}_p) - \sum_{i+j=n} \mathbf{E}_{i,j}(\mathbf{x}_p).
\end{equation}
Use this to repeat step i.
\end{enumerate}
Step ii takes advantage of sparse grids to achieve an elevated number of particles per cell in our representation of the electron density. Moreover, steps iii and iv also accelerate the ``field solve" in the standard manner of sparse grids as described in section 2.1.
\section{Numerical Tests}
We work in dimensionless variables in which distance is measured in multiples of the Debye length $\lambda_D = \sqrt{\epsilon_0 T/q \rho_e}$, time in multiples of the inverse plasma frequency $\omega_p^{-1} = \sqrt{\epsilon_0 m_e/q \rho_e}$, and the electron mass $m_e$ and charge $q$ are each normalized to one. In this initial study, we restrict our attention to a periodic box with side-lengths $L$.
We perform three types of numerical test:
\begin{itemize}
\item Linear Landau damping in 3-D
\item Nonlinear Landau damping in 3-D
\item Diocotron instability in 2-D
\end{itemize}
The purpose of the linear Landau damping test is to confirm the correct damping rate. For the nonlinear Landau damping tests, we compare the accuracy and timing of the sparse PIC method to those of a standard, explicit PIC method. The diocotron instability provides an example with a magnetic field and a chance to see sparse grids attempt to handle fine scale structure.
Throughout this section, we will refer to the number of simulated particles \textit{per cell}, which we denote by $P_c$. For runs with standard PIC, the definition is standard:
\begin{equation}
P_c = \frac{N_p}{N_c} = \frac{N_p h_x h_y}{L^2}
\end{equation}
in 2-D, and analogously in 3-D. For the sparse-PIC runs, however, we count the total number of cells in \textit{all} the grids used in the combination technique. Here, for an overall resolution of $2^n$ cells in each direction, one finds
\begin{equation}
P_c = \frac{N_p}{n2^{n+1} + (n-1)2^n} = \frac{N_p 2^{-n}}{3n - 1}
\end{equation}
in 2-D, and similarly in 3-D. For brevity, we also refer to a sparse grid whose maximum resolution in each coordinate is $2^n$ as being ``$2^n \times 2^n$" in 2-D and similarly in 3-D.
We will measure accuracy in density $\rho_e$ and the electric field $\mathbf{E}$. We do this by comparing to a highly resolved reference solution computed with standard PIC. The reported error in a given quantity $\psi$ will be given by
\begin{equation}
\mathcal{E}(\psi) = \frac{\left\lVert \psi - \psi_{\textrm{ref}} \right\rVert_{L^2}}{\left\lVert \psi_{\textrm{ref}} \right\rVert_{L^2}} = \sqrt{\frac{\int \left| \psi - \psi_{\textrm{ref}} \right|^2 \, d\mathbf{x}}{\int \left| \psi \right|^2 \, d\mathbf{x}}},
\end{equation}
where $\psi_{\textrm{ref}}$ is the relevant reference solution, and integrals are taken over the computational domain. We compute these integrals with the trapezoidal rule. Since all the quantities are periodic and since we rely on an equispaced grid in all direction, quadratures with the trapezoidal rule give us high order accuracy \cite{trefethen2014exponentially}.
When comparing the efficiency of sparse and regular PIC schemes, we measure both total computation time (in seconds) and total memory usage (in gigabytes). The memory usage metric is of particular importance for large-scale applications to massively parallel architectures, where computations are increasingly memory bound. In typical simulations, the dominant memory load is in storing the states (position and velocity) of the particles. This is the usage we report, assuming double precision arithmetic.
Finally, we note that energy conservation is an important consideration in PIC schemes of the type considered here. We do not present detailed results, but do note that in all tests performed, the energy conservation properties were identical between regular- and sparse-PIC.
\subsection{Notes on implementation and parallelization}
Before discussing detailed computational results, it is useful to discuss the implementation of the algorithm used. We use the FFT to solve Poisson's equation for $\varphi$ and to differentiate it to find $\mathbf{E}$. Additionally, to accommodate the enormous numbers of particles necessary, reference solutions are computed on the NYU HPC cluster `Mercer'. However, all other computations are performed on a quad-core personal workstation except where specifically noted. Regular PIC computations are performed entirely in serial, while sparse-PIC computations are `lightly parallelized' in the following sense.
In a shared memory context, the sparse-PIC algorithm is even more parallelizable than standard PIC. To see this, consider the three operations that are performed on the particles in a PIC scheme:
\begin{itemize}
\item Push particles
\item Map grid data to particles
\item Map particle data to grid
\end{itemize}
The first two operations are easily parallelizable since, if each thread is assigned to an individual particle, no two threads ever attempt to write to the same memory address. However, as noted in \cite{buyukkecceci2013portable,stantchev2008fast} the mapping of particle data onto the grid is prone to memory collisions when threads are assigned particle-wise, as multiple particles may interact with a single grid cell. This can be overcome through use of atomic operations, but a loss of performance results.
In contrast, in the sparse-PIC method described here, particle data is interpolated onto multiple \textit{distinct} grids. If each thread is assigned to a single grid used in the combination technique, then these computations can be performed simultaneously with no risk of memory collisions. Our sparse-PIC implementation takes advantage of this added parallelism relative to standard PIC when computing the particle-to-grid map, but computations are otherwise performed in serial for fair comparison to standard PIC.
\subsection{Linear Landau damping}\label{sec:linearlandau}
In the limit of long wavelength, the damping rate $\gamma$ of a warm plasma wave is given in our dimensionless variables by \cite{krall1973principles}
\begin{equation}
\gamma \approx \sqrt{\frac{\pi}{2}} \frac{\omega^2}{2k^3} \exp \left\{ -\frac{\omega^2}{2k^2} \right\}
\end{equation}
when the velocity distribution is Maxwellian. Here, the frequency $\omega$ and wavenumber $k$ are related by the Bohm-Gross dispersion relation
\begin{equation}
\omega^2 = 1 + 3 k^2.
\end{equation}
We introduce the shorthand
\begin{equation}
g(x; \alpha, \beta) = 1 + \alpha \cos \frac{2 \beta \pi x}{L}
\end{equation}
and choose the initial distribution
\begin{equation}
f(t=0) = \frac{1}{(2\pi)^{3/2}} \exp \left\{ -\frac{v_x^2 + v_y^2 + v_z^2}{2} \right\} g\left(x; 0.05,1\right) g(y; 0.05, 1) g(z; 0.05,1),
\end{equation}
$L = 22$, $\Delta t = 1/20$ and final time $T = 25$ for these tests.
\begin{figure}[h]
\begin{center}
\includegraphics[width=.6\textwidth]{LinLandauPE_correct.png}
\caption{Potential energy as a function of time for the 3-D linear Landau damping test case described in section \ref{sec:linearlandau}. Sparse (solid red) and regular (dashed blue) PIC simulations both agree well with predicted damping rate (black).}
\label{PElandau}
\end{center}
\end{figure}
The results are shown in figure \ref{PElandau}. Both the regular and sparse grid solutions use a $64 \times 64 \times 64$ grid resolution with $500$ particles per cell. Both solutions agree well with each other and with the predicted damping rate. The sparse solution is obtained with total particle number $N_p = 3.968 \times 10^6$, while the regular grid solution uses $N_p = 1.31 \times 10^8$ and requires approximately eight times the computation time of the sparse grid solution in our implementation.
\subsection{3-D Nonlinear Landau Damping}
For our nonlinear damping tests, we choose $L = 160$. The initial distribution is given by
\begin{equation}
f(t=0) = \frac{1}{(2\pi)^{3/2}} e^{-|v|^2/2} g\left(x; \, 0.2,4\right) g\left(y; \, 0.15, 3\right) g\left(z; \, 0.2, 4 \right).
\end{equation}
We use $\Delta t = 1/20$, and the reference solution is taken to be an ensemble average of 32 independent computations on a $128 \times 128 \times 128$ grid with $N_p = 1.5 \times 10^9$. We perform standard- and sparse-PIC runs using overall grid resolutions of $16$, $32$, $64$, $128$, and with $P_c = 25, 50, 100, 400, 800$. In addition, we run sparse tests at grid resolution $256$. All sparse-PIC runs are performed on a personal workstation, but the memory requirements of regular-PIC necessitate running the most computationally demanding cases on a cluster.
In all cases, we use $\Delta t = 1/20$ and we measure accuracy at the final time $T = 2.7$. We begin by plotting the density at the final time for our reference solution, as well as sample sparse- and regular-PIC solutions in figure.
\begin{figure}[h]
\begin{center}
\includegraphics[width=1.1\textwidth]{ThreeDcomparison.png}
\caption{Comparison of density $\rho_e$ - normalized so that mean value is unity - between reference solution computed using regular PIC (left), sparse solution on $128 \times 128 \times 128$ grid with $P_c = 800$ (center), and regular-PIC solution on $128 \times 128 \times 128$ grid with $P_c = 800$ (right). Approximate computation times are shown, revealing the dramatic speedup enabled by sparse grids. Interestingly, even at the same $P_c$ value, sparse-PIC experiences visibly smaller statistical errors.}
\label{3dcomparison}
\end{center}
\end{figure}
Next, we plot computation time and memory usage as functions of accuracy $\mathcal{E}(\rho_e)$ and $\mathcal{E}(\mathbf{E})$. Results are shown in figure \ref{3deff}. The advantages of the sparse approach are immediately evident in $\rho_e$, where for a given accuracy sparse-PIC is more than an order of magnitude faster and less memory intensive. When measuring accuracy in the electric field, sparse-PIC is competitive in computation time while still generating reduction in memory usage.
\begin{figure}[h]
\begin{center}
\includegraphics[width=.75\textwidth]{ThreeDAccFigs.png}
\caption{Computation time and memory usage for 3-D sparse (blue square) and full (red x) PIC runs. Run times can be comparable, but sparse PIC consistently uses less memory, frequently by an order of magnitude or more. This reduction in memory usage allows more accurate solutions for given computing resources. }
\label{3deff}
\end{center}
\end{figure}
In addition, we plot $\mathcal{E}(\rho_e)$ and $\mathcal{E}(\mathbf{E})$ as functions of grid resolution in figure \ref{3dconv}. As before, the sparse solution is superior when measuring $\rho_e$, but the electric field appears to be more sensitive to statistical error when sparse grids are used.
\begin{figure}[h]
\begin{center}
\includegraphics[width=.7\textwidth]{ThreeDconvFigs.png}
\caption{Accuracy in density $\rho_e$ and electric field $\mathbf{E}$ as functions of grid resolution. }
\label{3dconv}
\end{center}
\end{figure}
\subsection{Diocotron Instability}
We consider here a variation on the diocotron instability, which is an instability that leads to the formation of vortices in electron plasmas with hollow density profiles confined by a uniform magnetic field \cite{aydemir1994unified,davidson2001physics,driscoll1990experiments}. The variation resides in the fact that in our case the electrons are immersed in a uniform, immobile and neutralizing background ion population, as described in \eqref{ESVlasov}, and we impose periodic boundary conditions. While this setup may seem somewhat artificial from a plasma physics point of view, it is often used to study vortex dynamics in neutral fluids \cite{melander1987axisymmetrization}, in which the diocotron instability is also observed \cite{dritschel1986nonlinear}, and may be viewed as an illustration of the Kelvin-Helmholtz instability \cite{driscoll1990experiments}. We expect fine scale structures to form as a result of this process, and are interested in the way our sparse grid scheme handles these structures.
To solve \eqref{ESVlasov} numerically, we choose magnetic field strength $B = 15$ oriented along the $z$-axis, $L = 22$ and
\begin{equation}
f(t=0) = \frac{C}{2\pi} e^{-|v|^2/2} \exp \left\{ -\frac{(r-L/2)^2}{2 (0.03L)^2} \right\}, \quad r = \sqrt{(x - L/2)^2 + (y-L/2)^2}
\end{equation}
where $C$ is a normalization constant chosen to enforce $\int f dx dv = 1$. As before, $\Delta t = 1/20$, while final time $T = 35$.
We perform tests at $P_c = 20, 40, 80$. For regular grids, we test grid resolutions of $64, 128, 256, 512$, while for sparse grids we additionally test $1024, 4096, 8192, 16384$. Sample plots of $\rho_e$ are shown using both methods in figure \ref{2dcomp}.
\begin{figure}[h]
\begin{center}
\includegraphics[width=.8\textwidth]{comparisonFig2D.png}
\caption{Two-dimensional test case of the diocotron instability at $t=0$ (left), $20$ (middle) and $35$ (right) using both regular PIC (top) and sparse-PIC (bottom). Plots show density $\rho_e$. Regular PIC used a $256 \times 256$ grid and $P_c = 40$, requiring $246$ seconds. Sparse-PIC used a $1024 \times 1024$ grid and $P_c = 40$, requiring $225$ seconds. }
\label{2dcomp}
\end{center}
\end{figure}
We plot computation time and memory load as functions of $\mathcal{E}(\rho_e)$ in figure \ref{2deff}. Evidently, regular-PIC is more efficient than sparse-PIC in 2-D for the parameters tested. One can clearly see the improved asymptotic scaling of the computation time for sparse-PIC, but the so-called ``cross-over" point, at which sparse-PIC becomes more efficient than regular PIC occurs at larger scales than we are able to test here.
\begin{figure}[h]
\begin{center}
\includegraphics[width=.8\textwidth]{effrho2D.png}
\caption{Computation time (left) and memory usage(right) as functions of accuracy for the 2-D diocotron example.}
\label{2deff}
\end{center}
\end{figure}
It appears much of the error in the sparse-PIC runs arises from the tendency of sparse grids to ``smear" out the fine-scale structure developed by the solution - see the right-most portions of figure \ref{2dcomp}. Higher-order sparse grid schemes and a better understanding of the interactions between sparse grids and the field solve remain promising options to rectify this shortcoming.
\section{Conclusions}
In this initial study, we have presented the use of the sparse grid combination technique in concert with PIC methodology. Although the method is still early in the development stage, initial results presented here demonstrate the method's potential to accelerate large scale kinetic plasma simulations. This is particularly true in three dimensions, for which the benefits of the sparse approach are immediately visible, even with our relatively simple implementation. In the current state of the scheme, however, it is less clear whether sparse grids provide an advantage for 2-D simulations, except perhaps at very large grid sizes.
More work is needed to understand the interplay between statistical error and the field-solve part of the algorithm in the sparse grid context, as evidenced by figure \ref{3dconv}. A possible alternative to the approach presented here is to compute the density $\rho_e$ on a sparse grid, but then to project it onto a regular grid for the purpose of the field-solve. This is likely to reduce errors in the electric field due to the sparse grid representation. A drawback of this approach is that it would slow down the field-solve portion of the scheme to speeds commensurate with standard PIC, but in contexts where particle operations dominate the computation, this would be a small price to pay.
There exist many other directions for future development. Higher order interpolation schemes hold the potential to make sparse approaches even more attractive, even for two-dimensional simulations, by improving the ability to resolve fine structure. Non-periodic boundaries, complex geometries, electromagnetic and collisional simulations are also subjects of ongoing work, with results to be reported at a later date.
\section*{Acknowledgements}
The authors are particularly grateful to Bokai Yan, who is responsible for first introducing them to sparse grids. Enlightening conversations were also had with Harold Weitzner and Russel Caflisch. The authors were supported by the U.S. Department of Energy, Office of Science, Fusion Energy Sciences under Award No. DE-FG02-86ER53223. The second author emphasizes that the keys ideas for this article were contributed by the first author.
\vspace{4em}
\bibliographystyle{plain}
|
\section{Introduction}
This paper can be viewed from different perspectives, and indeed it has a two-fold purpose. On the one hand, it yields (in its four appendices) a study of unitary eigenfunction transforms of a novel type, generalizing the transmission-reflection and bound state picture associated with one-dimensional nonrelativistic Schr\"odinger operators
\begin{equation}\label{Schr}
- \frac{d^2}{dr^2}+V(r).
\end{equation}
Here we are thinking of real-valued potentials $V(r)$, $r\in {\mathbb R}$, that are smooth and decay rapidly for $r\to\pm\infty$. As is well known, for such potentials the operator~\eqref{Schr} can have finitely many bound states with negative eigenvalues, whereas the scattering can be encoded in the so-called Jost solutions, which have eigenvalue $k^2>0$. (We view $r$ and~$k$ as dimensionless variables.) Specifically, the Jost solutions $J_{\pm}(r,k)$ satisfy the time-independent Schr\"odinger equation
\begin{equation}\label{Jost}
(-\partial_r^2+V(r))J_{\pm}(r,k)=k^2J_{\pm}(r,k),\ \ \ k>0,
\end{equation}
and are characterized by the asymptotic behavior
\begin{equation}
J_+(r,k)\sim
\left\{
\begin{array}{ll}
a(k)e^{-irk}+b(k)e^{irk}, & r\to\infty, \\
e^{-irk}, & r\to -\infty,
\end{array}
\right.
\end{equation}
\begin{equation}
J_-(r,k)\sim \left\{
\begin{array}{ll}
e^{irk}, & r\to\infty, \\
a(k)e^{irk}-b(-k)e^{-irk}, & r\to -\infty.
\end{array}
\right.
\end{equation}
It is then customary to define the transmission and reflection coefficients by
\begin{equation}
t(k)=1/a(k),\ \ \ r(k)=b(k)/a(k).
\end{equation}
The generalized Jost functions we study in the appendices share the asymptotic behavior just recalled, but their transmission and reflection coefficients have a (purely imaginary) period in the spectral variable~$k$. For certain special choices they converge to ordinary Jost functions as this period goes to $i\infty$.
On the other hand, the paper is concerned with special cases of the eigenfunctions arising for two particles of opposite charge in the relativistic hyperbolic Calogero-Moser system, and with the associated Hilbert space theory. We intend to return to the Hilbert space theory for the general case in a companion paper. In the present one, however, we already lay the groundwork for the general case by introducing and discussing the pertinent eigenfunctions, both in this section and in more detail in Section~2. In Section~2 we are also better placed to put these Hilbert space aspects in a wider context and summarize previous literature in the area.
We proceed by recalling the two types of interaction for the hyperbolic Calogero-Moser two-particle system in its center-of-mass frame.
For the nonrelativistic Calogero-Moser case they are given by the potentials
\begin{equation}\label{Vra}
V_s(r)=\lambda(\lambda-1)/\sinh^2(r),\ \ V_o(r)=-\lambda(\lambda-1)/\cosh^2(r).\ \ \
\end{equation}
(In the physics literature, these are often referred to as P\"oschl-Teller potentials, see for example~\cite{F74}.)
Restricting attention to $\lambda> 1$, this corresponds to a description of a particle pair with the same/opposite charge having a repulsive/attractive interaction, the charge interpretation being borrowed from the electromagnetic force.
The relativistic generalizations of the repulsive and attractive nonrelativistic Hamiltonians associated with~\eqref{Vra} can be taken to be of the form
\begin{equation}\label{Hs}
H_s(\rho,\tau;r)=\left(\frac{\sinh(r+i\tau)}{\sinh(r)}\right)^{1/2}\exp(i\rho \partial_r)\left(\frac{\sinh(r-i\tau)}{\sinh(r)}\right)^{1/2}+(r\to -r),
\end{equation}
\begin{equation}\label{Ho}
H_o(\rho,\tau;r)=\left(\frac{\cosh(r+i\tau)}{\cosh(r)}\right)^{1/2}\exp(i\rho \partial_r)\left(\frac{\cosh(r-i\tau)}{\cosh(r)}\right)^{1/2}+(r\to -r).
\end{equation}
Here we have $\rho,\tau>0$, and the nonrelativistic differential operators given by~\eqref{Schr} and~\eqref{Vra} arise from these analytic difference operators in the limit $\rho \to 0$ with $\tau=\rho\lambda$. (For a comprehensive survey of the relativistic Calogero-Moser $N$-particle systems together with their nonrelativistic and Toda limits we refer to the lecture notes~\cite{R94}.)
The repulsive nonrelativistic Hamiltonian can be diagonalized by using the conical (or Mehler) function specialization of the hypergeometric function. The attractive potential arises from the repulsive one by the two analytic continuations $r\to r\pm i\pi/2$. The attractive nonrelativistic Hamiltonian can then be diagonalized by employing a suitable linear combination of the two correspondingly continued conical functions, taking also bound states into account, which do not arise for the repulsive case. (More details on these diagonalizations from disparate viewpoints can be found for instance in~\cite{F74}, Problems 38 and 39, \cite{T62}, Sections 4.18 and 4.19, and~\cite{K84}.)
The repulsive Hamiltonian~$H_{s}$ can be diagonalized by a relativistic generalization of the conical function. A detailed study of this relativistic conical function and its various limits can be found in~\cite{R11}. We use this function as a starting point to arrive at the eigenfunctions of the attractive counterpart~$H_{o}$ that are the key to the Hilbert space reinterpretation of~$H_o$. As we shall show, they are once more obtained by suitable linear combinations of the $r\to r\pm i\pi/2$ continuations of the relativistic conical function.
For the special coupling constants $\tau=(N+1)\pi$, $N\in{\mathbb N}$, the associated attractive eigenfunction transforms have certain periodicity features that motivated the general framework that is set out in the appendices. The latter yield a largely self-contained account, which can be read independently of the main text.
We stress, however, that the assumptions we make in Appendix~A and~B would seem far-fetched without their concrete realizations coming from the main text, as they are not satisfied for the eigenfunctions of the nonrelativistic operators~\eqref{Schr} with a nontrivial $V(r)$. Since one might question whether they can be realized at all, we have included the simplest explicit examples in the appendices, so that one need not delve into the main text to see that the assumptions are not vacuous.
The Hilbert space theory associated with the opposite-charge Hamiltonian~$H_{o}$ for the case of vanishing reflection has been worked out before in Section~4 of~\cite{R00}. (This case corresponds to couplings of the form~$\tau=(N+1)\rho$, $N\in{\mathbb N}$.) The reflectionless case was also handled via a slightly more general framework, cf.~Section~2 in~\cite{R00}. The latter setting is subsumed by our account in the appendices of the present paper, in the sense that the assumptions we make in Appendix~A and~B are satisfied by the eigenfunctions studied in~\cite{R00}.
More specifically, in the appendices we start from wave functions
\begin{equation}\label{Psipm}
\Psi(r,k)=\Psi^+(r,k)e^{irk} +\Psi^-(r,k)e^{-irk},
\end{equation}
which are counterparts of the above function~$J_-(r,k)/a(k)$, and impose various requirements on the plane wave coefficients $\Psi^{\pm}(r,k)$ of a function-theoretic and asymptotic nature. (The case $\Psi^-(r,k)=0$ yields reflectionless transforms.) The assumptions ensure that we can push through a comprehensive Hilbert space analysis for the associated transforms, without mentioning any Hamiltonians until the last Appendix~D, which is devoted to time-dependent scattering theory.
With the functional analysis relegated to the appendices, the main text demonstrates that the assumptions are satisfied by the pertinent eigenfunctions of the attractive Hamiltonian~$H_{o}$ when the coupling constant~$\tau$ is of the form~$(N+1)\pi$, $N\in{\mathbb N}$. (The reflectionless $\tau$-choices~$(N+1)\rho$, $N\in{\mathbb N}$, are briefly reviewed as well.) Therefore, when viewed as an analytic difference operator---as opposed to a Hilbert space operator---the Hamiltonian reduces to the free one
\begin{equation}
H_f=\exp(i\rho \partial_r)+\exp(-i\rho \partial_r).
\end{equation}
As it stands, the latter can of course be diagonalized by the Fourier transform with kernel $\exp(irk)$, yielding multiplication by $2\cosh(\rho k)$. The crux is, however, that the $H_o$-eigenfunctions for general $\tau$ and $\rho$, specialized to $\tau=(N+1)\pi$, do not reduce to a plane wave at all. Instead, they are of the form~\eqref{Psipm}, with nontrivial coefficients $\Psi^{\pm}(r,k)$ that are $i\rho$-periodic in~$r$, and they give rise to reflection and transmission coefficients $r(k)$ and $t(k)$ that, together with the scattering function $u(k)$ for the repulsive case, satisfy the Yang-Baxter equations.
The latter equations actually hold for the arbitrary coupling scattering amplitudes, cf.~\eqref{t}--\eqref{YB2}. This state of affairs lends further credence to the long-standing conjecture that the many-particle generalizations of the Hamiltonians~\eqref{Hs} and~\eqref{Ho} lead to a factorized $S$-matrix. For the classical versions of the systems this solitonic scattering has been proved in~\cite{R93}, alongside a detailed account of the connection to several solitonic field theories.
For the special case $\tau=\pi/2$, a prime example of the latter is the sine-Gordon classical field theory. For this choice of $\tau$, the scattering amplitudes $t(k)$, $r(k)$ and $u(k)$ coincide with those for quantum sine-Gordon opposite-charge and equal-charge fermions~\cite{ZZ79,S92}. This state of affairs is in agreement with earlier work from a related viewpoint~\cite{R01}, to which we intend to return in the companion paper dealing with general $\tau$-values.
Among the special cases considered in this paper, of particular interest is the case $N=0$, yielding $\tau=\pi$. Indeed, for this choice the repulsive eigenfunction (essentially given by the relativistic conical function) reduces to the plane wave sum~$\sin(rk)$, so that $u(k)=1$. By contrast, the scattering coefficients $t(k)$ and $r(k)$ are nontrivial. (They amount to specializations of~$T_+(k)$~\eqref{tnu} and~$R_+(k)$~\eqref{rnu}.) Physically speaking, this parameter choice yields a system in which particles of the same charge do not interact, whereas oppositely charged particles have an attractive interaction. In particular, an oppositely-charged pair can form a bound state, provided $\rho$ is suitably restricted.
In order to tie in the main text with the appendices, we need to handle the residue sums that are spawned by contour shifts arising for the pertinent transforms and their adjoints. For isometry of the transforms, it is necessary that these sums vanish, whereas eventual bound states show up via nonzero residue sums associated with the adjoint transforms. At the end of the main text, we also present an in-depth study of isometry breakdown for sufficiently small $\rho$, a phenomenon without a nonrelativistic counterpart.
Having presented a bird's eye view of the aims and contents of the paper, we proceed with a more detailed sketch of its results and organization.
As already mentioned, our starting point is the relativistic conical function. More specifically, we work with a renormalized version~${\cal R}_{\rm ren}(a_+,a_-,b;x,y)$. (We recall one of the many integral representations for this function below, cf.~\eqref{Rrdef}.) From a quantum-mechanical perspective, all of its five variables have dimension [position]. More specifically, the parameters~$a_+$ and $a_-$ can be viewed as the interaction length and Compton wave length $\hbar/mc$, while the parameter~$b$ and~variable~$y$ play the role of coupling constant and spectral variable. The unorthodox choice of the latter (inasmuch as it is customary to choose a momentum variable as spectral variable) is inspired by the symmetry of this function under the interchange of $x$ and~$y$ (self-duality). It is written entirely in terms of the hyperbolic gamma function~$G(a_+,a_-;z)$ from~\cite{R97}, which is invariant under swapping $a_+$ and $a_-$, cf.~\eqref{modinv} below. This `modular invariance' is inherited by~${\cal R}_{\rm ren}(a_+,a_-,b;x,y)$. The renormalized version has the advantage of being meromorphic in~$b$, $x$ and~$y$, with no poles in~$b$ that do not depend on~$x$ and~$y$. In more detail, its poles can only be located at
\begin{equation}\label{cRpoles}
\pm z=2ia-ib+ika_++ila_-,\ \ \ z=x,y,\ \ \ k,l\in {\mathbb N},
\end{equation}
where
\begin{equation}
a\equiv (a_++a_-)/2.
\end{equation}
Thanks to these symmetry properties, this function satisfies not only the A$\Delta$E (analytic difference equation)
\begin{equation}\label{Rade}
A(a_+,a_-,b;x){\cal R}_{\rm ren}(a_+,a_-,b;x,y)=2\cosh(\pi y/a_+){\cal R}_{\rm ren}(a_+,a_-,b;x,y),
\end{equation}
where the A$\Delta$O (analytic difference operator) is defined by
\begin{equation}\label{defA}
A(a_+,a_-,b;z)\equiv V(a_+,b;z)\exp(-ia_-\partial_z) +(z\to -z),
\end{equation}
with
\begin{equation}\label{defV}
V(a_+,b;z)\equiv \frac{\sinh(\pi (z-ib)/a_+)}{\sinh(\pi z/a_+)},
\end{equation}
but also three more A$\Delta$Es obtained by swapping~$a_+$ and~$a_-$, and/or~$x$ and~$y$ in~\eqref{Rade}.
Each of the resulting four A$\Delta$Os has two more avatars, obtained by similarity transformations involving the generalized Harish-Chandra function
\begin{equation}\label{defc}
c(b;z)\equiv G(z+ia-ib)/G(z+ia),
\end{equation}
and weight function
\begin{equation}\label{defw}
w(b;z)\equiv 1/c(b;z)c(b;-z).
\end{equation}
Here and often below, we suppress the dependence on~$a_+$ and~$a_-$ when no ambiguity can arise. (For example, the $c$- and~$w$-functions just introduced are invariant under swapping $a_+$ and~$a_-$.)
From now on, we also use the abbreviations
\begin{equation}
c_{\delta}(z)\equiv \cosh(\pi z/a_{\delta}),\ \ s_{\delta}(z)\equiv \sinh(\pi z/a_{\delta}),\ \ e_{\delta}(z)\equiv \exp(\pi z/a_{\delta}),\ \ \delta=+,-.
\end{equation}
For our present purposes, we only need two among these eight additional operators. The most important one is the Hamiltonian
\begin{equation}\label{HA}
H(a_+,a_-,b;x) \equiv w(b;x)^{1/2}A(a_+,a_-,b;x) w(b;x)^{-1/2},\ \ \ (b,x)\in(0,2a)\times(0,\infty),
\end{equation}
but we also have occasion to use
\begin{equation}\label{cA}
{\cal A}(a_+,a_-,b;y) \equiv c(b;y)^{-1}A(a_+,a_-,b;y) c(b;y).
\end{equation}
These operators can also be written as
\begin{equation}\label{Hb}
H(b;x)= V(a_+,b;-x)^{1/2}\exp(ia_-\partial_x) V(a_+,b;x)^{1/2}+(x\to -x),
\end{equation}
and
\begin{equation}\label{cAb}
{\cal A}(b;y)=\exp(-ia_-\partial_y) + V(a_+,b;-y)\exp(ia_-\partial_y) V(a_+,b;y).
\end{equation}
From these formulas one can read off that they are formally self-adjoint, cf.~\eqref{defV}.
The $w$-function~\eqref{defw} is positive for $(b,x)\in(0,2a)\times(0,\infty)$, and throughout the paper we take positive square roots of positive functions. This positivity assertion and the alternative formulas for~$H(b;x)$ and~${\cal A}(b;y)$ readily follow from key features of the hyperbolic gamma function. For completeness, we briefly digress to summarize the salient properties.
The hyperbolic gamma function can be defined as the meromorphic solution to one of the first order A$\Delta$Es
\begin{equation}\label{Gades}
\frac{G(z+ia_{\delta}/2)}{G(z-ia_{\delta}/2)}=2c_{-\delta}(z),\ \ \delta=+,-,\ \ a_+,a_->0,
\end{equation}
which is uniquely determined by
the normalization~$G(0)=1$ and a minimality property; the second A$\Delta$E is then satisfied as well. (This property amounts to requiring absence of zeros and poles in a certain $|{\rm Im}\, z|$-strip, and `optimal' asymptotics for $|{\rm Re}\, z|\to\infty$, cf.~\cite{R97}.) In the strip $|{\rm Im}\, z|<a$ it has the integral representation
\begin{equation}\label{Gint}
G(a_+,a_-;z)=\exp\left( i
\int_0^\infty\frac{dy}{y}\left(\frac{\sin 2yz}{2\sinh(a_{+}y)\sinh(a_{-}y)} - \frac{z}{a_{+}a_{-} y}\right)\right),
\end{equation}
from which one reads off absence of zeros and poles in this strip and the properties
\begin{equation}\label{modinv}
G(a_-,a_+;z) = G(a_+,a_-;z),\ \ \ ({\rm modular\ invariance}),
\end{equation}
\begin{equation}\label{refl}
G(-z) = 1/G(z),\ \ \ ({\rm reflection\ equation}),
\end{equation}
\begin{equation}\label{Gcon}
\overline{G(a_+,a_-;z)}=G(a_+,a_-;-\overline{z}).
\end{equation}
The hyperbolic gamma function has its poles at
\begin{equation}\label{Gpo}
-ia -ika_+-il a_-,\ \ \ \ k,l\in{\mathbb N},\ \ \ (G{\rm -poles}),
\end{equation}
and its zeros at
\begin{equation}\label{Gze}
ia +ika_++il a_-, \ \ \ \ k,l\in{\mathbb N},\ \ \ \ (G{\rm -zeros}).
\end{equation}
The pole at $-ia$ is simple, and so is the zero at $ia$.
Finally, we list the asymptotic behavior for ${\rm Re}\, (z)\to\pm \infty$:
\begin{equation}\label{Gas}
G(a_+,a_-;z) = \exp \big(\mp i\left(\chi+ \pi z^2/2a_+a_-\right)\big)\big(1 + O(\exp(-r |{\rm Re}\, (z)|))\big).
\end{equation}
Here, we have
\begin{equation}\label{chi}
\chi \equiv \frac{\pi}{24}\left(\frac{a_+}{a_-} + \frac{a_-}{a_+}\right),
\end{equation}
and the decay rate can be any positive number satisfying
\begin{equation}
r <2\pi \min(1/a_+,1/a_-).
\end{equation}
The A$\Delta$O~\eqref{HA} can be viewed as the defining Hamiltonian for the same-charge center-of-mass two-particle system. The opposite-charge defining Hamiltonian is given by
\begin{equation}\label{tHtA}
\tilde{H}(b;x)= \tilde{w}(b;x)^{1/2}\tilde{A}(b;x) \tilde{w}(b;x)^{-1/2},\ \ \ (b,x)\in(-a_+/2,a_+/2+a_-)\times {\mathbb R},
\end{equation}
where
\begin{equation}\label{tA}
\tilde{A}(b;x)\equiv \tilde{V}(b;x)\exp(-ia_-\partial_x) +(x\to -x),
\end{equation}
\begin{equation}\label{tV}
\tilde{V}(b;x)\equiv \frac{\cosh(\pi (x-ib)/a_+)}{\cosh(\pi x/a_+)},
\end{equation}
and the attractive weight function is given by
\begin{equation}\label{tilw}
\tilde{w}(b;x)\equiv \prod_{\sigma=+,-}\frac{G(\sigma x+ia_-/2)}{G(\sigma x+ia_-/2-ib)}.
\end{equation}
Hence the attractive Hamiltonian can also be written as
\begin{equation}\label{tHb}
\tilde{H}(b;x) = \tilde{V}(b;-x)^{1/2}\exp(ia_-\partial_x) \tilde{V}(b;x)^{1/2}+(x\to -x).
\end{equation}
Clearly, the attractive coefficient~$\tilde{V}(b;x)$ arises from the repulsive coefficient~$V(a_+,b;x)$
by either one of the analytic continuations $x\to x\pm i a_+/2$. Hence the same is true for $\tilde{A}$ and~$A$; for~$\tilde{H}$ and~$H$ this is also the case, as follows from the alternative representations~\eqref{Hb} and~\eqref{tHb}.
Even so, the weight function $\tilde{w}(b;x)$ does not arise in this way from~$w(b;x)$. Indeed, the two functions $w(b;x\pm i a_+/2)$ are not positive for $(b,x)\in(-a_+/2,a_+/2+a_-)\times {\mathbb R}$, whereas $\tilde{w}(b;x)$ does have this positivity feature. (As before, this readily follows from~\eqref{Gades}--\eqref{Gcon}.) On the other hand, we clearly have
\begin{equation}\label{deftw}
\tilde{w}(b;x)=1/\tilde{c}(b;x)\tilde{c}(b;-x),
\end{equation}
where
\begin{equation}\label{deftc}
\tilde{c}(b;x)\equiv G(x+ia_-/2-ib)/G(x+ia_-/2)=c(b;x-ia_+/2).
\end{equation}
In words, the attractive $c$-function does arise from the repulsive $c$-function~\eqref{defc} by one of the analytic continuations at issue. We also point out that the conjugation relation~\eqref{Gcon} entails that the $\tilde{c}$-function has the same conjugation property
\begin{equation}
\overline{\tilde{c}(b;x)}=\tilde{c}(b;-x),\ \ \ (b,x)\in {\mathbb R}\times{\mathbb R},
\end{equation}
as the $c$-function; by contrast to the latter, however, it is regular at the origin (for generic~$b$), whereas~$c(b;x)$ has a simple pole.
Because the scale parameter $a_+$ and variable~$x$ are singled out, modular invariance and self-duality are not preserved for the attractive regime. Therefore, it is no longer crucial to insist on the symmetric parametrization of the repulsive regime. Even so, for reasons of notational economy we continue to employ it for most of the main text.
In the appendices, however, it is more convenient to work with dimensionless variables~$r$ and~$k$ such that the plane wave~$\exp(i\pi xy/a_+a_-)$ becomes~$\exp(irk)$. Specifically,
we take
\begin{equation}\label{xryk}
x/a_-\to r/\rho,\ \ \ y/a_-\to k/\kappa,
\end{equation}
with
\begin{equation}\label{rhoka}
\rho\kappa =\pi a_-/a_+.
\end{equation}
The coupling parameter~$b$ can be traded for a dimensionless parameter~$\tau$ given by
\begin{equation}\label{tau}
\tau \equiv \pi b/a_+.
\end{equation}
The Hamitonians~$H_s$~\eqref{Hs} and~$H_o$~\eqref{Ho} then arise from~$H$ and~$\tilde{H}$ by using these dimensionless quantities together with the choice $\kappa=1$.
In Section~2 we begin by summarizing pertinent results on the repulsive eigenfunctions. Their abundance of symmetries leaves no doubt concerning their uniqueness, and indeed there already exists a fully satisfactory Hilbert space theory for $H(b;x)$~\cite{R11,R03III}.
For the attractive case, however, there is an enormous ambiguity, and it is no longer obvious how to proceed. Once we have put this problem in perspective, we are prepared to survey the previous literature that has a bearing on the key problem of promoting \emph{formally} self-adjoint A$\Delta$Os to self-adjoint Hilbert space operators, using an explicit unitary eigenfunction transform that yields a concrete realization of the spectral theorem~\cite{RS72}.
After this terse literature overview (which can be found above~\eqref{Rpmades}), we detail how our attractive eigenfunctions result from the repulsive ones by taking a suitable linear combination of the two analytically continued functions~${\cal R}_{\rm ren}(a_+,a_-,b;x\pm ia_+/2,y)$. We exemplify this for the simplest case $b=a_+$. This yields the elementary function~\eqref{psizero}, all of whose function-theoretic and asymptotic features can be directly read off.
In view of the ambiguity in obtaining the opposite-charge eigenfunctions, our choice may seem unmotivated at face value. It is, however, singled out by its very special and desirable features. We have not attempted to show that this renders the choice unique, but we have little doubt that this is true.
To lend credence to this conviction, we obtain already in this paper the dominant $|{\rm Re}\, x|\to\infty$ asymptotics of the joint eigenfunction~$\psi(b;x,y)$~\eqref{defpsi}, cf.~Proposition~2.1. Specifically, together with the scattering function~$u(b;y)$~\eqref{u} of the repulsive case, the resulting attractive transmission and reflection coefficients~$t(b;y)$~\eqref{t} and~$r(b;y)$~\eqref{r} satisfy the Yang-Baxter equations~\eqref{sjk}--\eqref{YB2}. General-$b$ Hilbert space aspects, however, are beyond the scope of this paper.
In the remaining sections we focus on the special $b$-values $(N+1)a_+$. Section~3 contains a detailed account of function-theoretic and asymptotic properties of the corresponding eigenfunctions~$\psi_N(x,y)$~\eqref{psiN}. Their elementary character is inherited from that of the relativistic conical functions~$R_N(x,y)$~\eqref{RN}. The latter were studied in~\cite{R99}. We need various results from this source, which we summarize in the equations~\eqref{RNexp}--\eqref{SNasm}. In the remainder of the section we obtain the properties of~$\psi_N(x,y)$ that served as a template for the axiomatic account in the appendices.
In Section~4 we then apply the Hilbert space results from the appendices to the eigenfunction transforms associated with~$\psi_N(x,y)$, and to the adjoints of the transforms. First, we study the residue sums arising for the transforms, showing that they vanish for the parameter interval~$a_-\in((N+1/2)a_+,\infty)$ in Theorem~4.1. With isometry proved, it follows from Appendix~D that the transforms can be viewed as the incoming wave operators from time-dependent scattering theory. Then we study the adjoint transforms for parameters in this isometry interval. Here a distinction arises: For the subinterval~$a_-\in[(N+1)a_+,\infty)$ we prove in Theorem~4.2 that the associated residue sums vanish, so that the transforms are unitary. By contrast, for~$a_-\in((N+1/2)a_+,(N+1)a_+)$, Theorem~4.4 reveals the presence of a bound state~$\psi_N(x)$, given by~\eqref{psiNx}. As we detail at the end of Section~4, it arises as (a multiple of) the residue of $\psi_N(x,y)$ at its only $y$-pole $i(N+1)a_+-ia_-$ in the strip ${\rm Im}\, y\in(0,a_-)$.
After these `constructive' results in Section~4, the final Section~5 furnishes some `destructive' results pertaining to the remaining $a_-$-interval $(0,(N+1/2)a_+]$. In particular, we prove that for $a_-\in(Na_+,(N+1/2)a_+)$ isometry of the transform breaks down in a way that we make quite explicit. For $N>0$ it is not an easy matter to handle any positive $a_-$, and we shall not do so. Indeed, as $a_-$ decreases, there is a cascade of ever larger and more baroque deviations from isometry, much like for the reflectionless cases handled in~\cite{R00}. For $N=0$ we do work out the details, which are summarized in Proposition~5.1.
As already mentioned, the appendices contain an axiomatic account of the functional-analytic aspects, with the assumptions inspired by the features of the eigenfunctions $\psi_N(x,y)$~\eqref{psiN}. Appendix~A deals with a general transform~${\cal F}$ from `momentum space' to `position space'. The former space is described in terms of two-component square-integrable functions of a (dimensionless) variable $k\in(0,\infty)$, so as to lead to a $2\times 2$ $S$-matrix $S(k)$~\eqref{Sm}. The assumptions are sufficiently restrictive to lead to a picture of the transform as being isometric away from a subspace encoded in residue terms, cf.~Theorem~A.1. In the axiomatic setting, however, it is far from clear whether this subspace is necessarily finite-dimensional, as is the case for the transforms associated with $\psi_N(x,y)$.
The adjoint transform~${\cal F}^*$ is analyzed in Appendix~B, with additional assumptions on the $k$-dependence of the plane wave coefficients enabling us to arrive at the counterpart Theorem~B.1 of Theorem~A.1. Simple explicit examples are included in this appendix via Propositions~B.3--B.5, with Proposition B.4 presenting example transforms that go beyond the main text.
A pivotal technical ingredient in the proofs of Theorems~A.1 and B.1 is relegated to Appendix~C, namely, the asymptotic analysis of the boundary terms arising from the rectangular integration contours at issue in the proofs.
The paper is concluded with Appendix~D, in which it is shown that when the transforms are unitary, then they may be viewed as the incoming wave operators for a large class of self-adjoint Hilbert space dynamics. More specifically, the transform~${\cal F}$ from Appendix~A plays this role for a class containing the (Hilbert space version of the) opposite-charge Hamiltonian~$\tilde{H}((N+1)a_+;x)$ given by~\eqref{HN}, whereas for its adjoint~${\cal F}^*$ the class contains a Hamiltonian that arises from the dual A$\Delta$O ${\cal S}((N+1)a_+;y)$ given by~\eqref{cSN}. It should be stressed that in the axiomatic framework of Appendix~D the self-adjoint dynamics at issue are \emph{defined} via the transforms. Likewise, since we rely on the results of the appendices to associate self-adjoint operators to the A$\Delta$Os \eqref{HN} and~\eqref{cSN}, the feasibility of doing so hinges on the \emph{isometry} of their eigenfunction transforms.
Below Theorem~D.1 we also discuss the issue of parity and time reversal symmetry. Both symmetries are present for the transforms arising from the main text, but we leave the question open whether time reversal symmetry holds in the general setting of the appendices.
\section{Repulsive and attractive eigenfunctions for general coupling}
As is detailed in~\cite{R11}, the relativistic conical function has a great many distinct representations. For completenes, we quote one of them, namely (cf.~Eq.~(1.3) and Eq.~(1.6) in~\cite{R11}),
\begin{equation}\label{Rrdef}
{\cal R}_{\rm ren}(b;x,y)=\frac{G(ia-ib)}{\sqrt{a_+a_-}} \int_{{\mathbb R}}dz \frac{G(z+ (x-y)/2-ib/2)G(z- (x-y)/2-ib/2)}{G(z+ (x+y)/2+ib/2)G(z- (x+y)/2+ib/2)},
\end{equation}
where we take $(b,x,y)\in (0,2a)\times {\mathbb R}^2$ to begin with. (Note that the $G$-asymptotics~\eqref{Gas} entails the convergence of the integral.) This is the only known representation that is both manifestly self-dual and modular invariant. Moreover, evenness in~$x$ and~$y$ follows by using the reflection equation~\eqref{refl}, and real-valuedness results from the conjugation relation~\eqref{Gcon}.
The function
\begin{equation}\label{FR}
{\mathrm F}(b;x,y) \equiv w(b;x)^{1/2}w(b;y)^{1/2}{\cal R}_{\rm ren}(b;x,y),\ \ \ (b,x,y)\in(0,2a)\times (0,\infty)^2,
\end{equation}
is also self-dual and modular invariant, so it is a joint eigenfunction of four independent A$\Delta$Os, namely~$H(a_+,a_-,b;x), H(a_+,a_-,b;y), H(a_-,a_+,b;x), H(a_-,a_+,b;y)$. In fact, it has yet another symmetry, which is not manifest:
It satisfies
\begin{equation}\label{bsym}
{\mathrm F}(2a-b;x,y) ={\mathrm F}(b;x,y).
\end{equation}
This can be understood from the four Hamiltonians having this symmetry property. By contrast, the A$\Delta$O $A(a_+,a_-,b;z)$~\eqref{defA} and the three factors of~${\mathrm F}(b;x,y)$ are not invariant under $b\to 2a-b$.
It follows from previous results~\cite{R11,R03III} that the transform
\begin{equation}\label{cFs}
({\cal F}_s(b) f)(x)\equiv \frac{1}{\sqrt{2a_+a_-}} \int_0^{\infty}dy\, {\mathrm F}(b;x,y)f(y),\ \ \ b\in(0,2a),
\end{equation}
yields a unitary transformation from~$L^2((0,\infty),dy)$ onto~$L^2((0,\infty),dx)$. It satisfies
\begin{equation}\label{Fssym}
{\cal F}_s(b)^*={\cal F}_s(b),\ \ {\cal F}_s(2a-b)={\cal F}_s(b),
\end{equation}
with the self-adjointness following from real-valuedness and self-duality of its kernel, and the $b$-symmetry from~\eqref{bsym}. For $b$ equal to~$a_+$ or~$a_-$ the transform amounts to the sine transform, and its $b=0$ and $b=2a$ limits amount to the cosine transform.
For~$b\in(0,2a)$ the even weight function has a second order zero at the origin (cf.~\eqref{Gze}), so its square root continues from~$(0,\infty)$ to an odd function. Therefore, the transform can also be viewed as a unitary involution on the odd subspace of $L^2({\mathbb R})$. More generally, the eigenfunctions for $N$ particles with the same charge are antisymmetric under permutations~\cite{HR14}: Equal-charge particles obey fermionic statistics.
By contrast, a fermion and an antifermion can be distinguished by their charge, so we should aim for eigenfunctions of~$\tilde{H}(b;x)$ that have nontrivial even and odd reductions. This opposite-charge Hamiltonian, however, has no `modular partner', so we can no longer insist on invariance under interchange of $a_+$ and~$a_-$, a requirement that renders the above same-charge eigenfunction~${\mathrm F}(b;x,y)$ essentially unique. Now any $\tilde{H}(b;x)$-eigenfunction remains an eigenfunction upon multiplication by a function $M(x,y)$ that is meromorphic and $ia_-$-periodic in~$x$ and has an arbitrary $y$-dependence. (This is the ambiguity alluded to in the Introduction.) A priori, however, there is no reason to expect that in this infinite-dimensional eigenfunction space there exist any special ones with the requisite orthogonality and completeness properties.
More generally, to date there exists no general Hilbert space theory for A$\Delta$Os. There is, however, a growing supply of explicit (mostly one-variable) A$\Delta$Os that admit a reinterpretation as bona fide self-adjoint Hilbert space operators. This hinges on the existence of special eigenfunctions that give rise to the unitary transform featuring in the spectral theorem for self-adjoint operators. We continue by surveying the literature in this field, which is quite scant by comparison to the vast literature dealing with Hilbert space aspects of (linear) differential operators and discrete difference operators.
By far the simplest cases are analytic difference operators that admit
orthogonal polynomials as eigenfunctions.
The earliest example is given by the Askey-Wilson polynomials~\cite{AW85}. From the perspective of Calogero-Moser type systems, these can be viewed as the special eigenfunctions of the $BC_1$ relativistic trigonometric Calogero-Moser system that give rise to a reinterpretation of the Askey-Wilson 4-parameter A$\Delta$O as a self-adjoint Hilbert space operator. Likewise, the multi-variable orthogonal polynomials introduced by Macdonald~\cite{M95} yield the sought-for joint eigenfunctions of the commuting A$\Delta$Os arising for the $A_{N-1}$ (i.~e., $N$-particle) relativistic trigonometric Calogero-Moser system, and the Koornwinder polynomials those for the $BC_N$ case~\cite{K92, D95}.
A large class of one-variable A$\Delta$Os yielding reflectionless unitary eigenfunction transforms has been introduced in~\cite{R05}. (The reflectionless A$\Delta$Os studied in~\cite{R00} form a tiny subclass.) Their eigenfunctions are closely connected to soliton solutions of various nonlocal evolution equations. The unitarity proof for the associated Hilbert space transforms makes essential use of previous results on the connection between the classical hyperbolic relativistic Calogero-Moser $N$-particle systems and the KP and 2D Toda hierarchies.
Further unitary eigenfunction transforms for 4-parameter A$\Delta$Os of Askey-Wilson type have been constructed in~\cite{KS01, R03III}, and for the elliptic 8-parameter A$\Delta$Os introduced by van Diejen~\cite{D94} in~\cite{R15}. Other relevant papers are~\cite{R03, R04, FT15}.
In all of these cases, the eigenfunctions have some very special properties that play a pivotal role for pushing through the associated Hilbert space theory.
In spite of this store of examples, a general theory has not emerged yet.
Returning to the problem at hand, it is clear from~\eqref{Rade}--\eqref{defV} and~\eqref{tA}--\eqref{tV} that we have
\begin{equation}\label{Rpmades}
\tilde{A}(b;x){\cal R}_{\rm ren}(b;x\pm ia_{+}/2,y)=2c_+(y) {\cal R}_{\rm ren}(b;x\pm ia_+/2,y).
\end{equation}
By virtue of~\eqref{tHtA}, this implies that we get two independent $\tilde{H}(b;x)$-eigenfunctions,
\begin{equation}
\tilde{H}(b;x)\tilde{w}(b;x)^{1/2}{\cal R}_{\rm ren}(b;x\pm ia_{+}/2,y)=2c_+(y)\tilde{w}(b;x)^{1/2} {\cal R}_{\rm ren}(b;x\pm ia_+/2,y).
\end{equation}
However, these functions remain eigenfunctions when they are multiplied by functions that are $ia_-$-periodic in~$x$ and that have an arbitrary $y$-dependence, so it is at this point that we need further constraints to reduce the ambiguity.
To this end, consider the auxiliary function
\begin{equation}\label{cZ}
{\cal Z}(b;x, y)\equiv {\cal R}_{\rm ren}(b;x,y)/c(b;-y).
\end{equation}
In view of~\eqref{cA} it satisfies
\begin{equation}
{\cal A}(b;-y){\cal Z}(b;x,y)=2c_+(x){\cal Z}(b;x,y),
\end{equation}
which entails
\begin{equation}
{\cal A}(b;-y){\cal Z}(b;x\pm ia_+/2,y)=\pm 2is_+(x){\cal Z}(b;x\pm ia_+/2,y).
\end{equation}
We now introduce the formally self-adjoint A$\Delta$O
\begin{equation}\label{cS}
{\cal S}(b;y)\equiv V(a_+,b;y)\exp(-ia_-\partial_y)V(a_+,b;-y)-\exp(ia_-\partial_y),
\end{equation}
and observe that in view of~\eqref{cAb} we have similarities
\begin{equation}\label{cScA}
\frac{e_-(\pm (ib-y)/2)}{s_-(ib-y)}{\cal A}(b;-y)\frac{s_-(ib-y)}{e_-(\pm (ib-y)/2)}=\pm i{\cal S}(b;y).
\end{equation}
Therefore, the functions
\begin{equation}\label{psipm}
\psi_{\pm}(b;x,y)\equiv \pm \tilde{w}(b;x)^{1/2}
\frac{e_-(\pm (ib-y)/2)}{2s_-(ib-y)}{\cal Z}(b;x\pm ia_+/2,y),
\end{equation}
satisfy not only the same $\tilde{H}$-A$\Delta$E,
\begin{equation}\label{tHade}
\tilde{H}(b;x)\psi_{\pm}(b;x,y) =2c_+(y)\psi_{\pm}(b;x,y),
\end{equation}
but also the same ${\cal S}$-A$\Delta$E,
\begin{equation}\label{cSade}
{\cal S}(b;y)\psi_{\pm}(b;x,y) =2s_+(x)\psi_{\pm}(b;x,y).
\end{equation}
As a consequence, the function
\begin{multline}\label{defpsi}
\psi(b;x,y)\equiv \psi_+(b;x,y)+\psi_-(b;x,y)
\\
=\frac{\tilde{w}(b;x)^{1/2}}{2s_-(ib-y)c(b;-y)}\big( e_-((ib-y)/2){\cal R}_{\rm ren}(x+ia_+/2,y)
\\
-e_-((y-ib)/2){\cal R}_{\rm ren}(x-ia_+/2,y)\big),
\end{multline}
satisfies both A$\Delta$Es, too. (We recall $c(b;-y)$ is given by~\eqref{defc}, and $\tilde{w}(b;x)$ by~\eqref{tilw}.)
As will transpire, the function~$\psi(b;x,y)$ we just defined is the sought-for attractive eigenfunction. We shall substantiate this for the special cases $b=(N+1)a_+$, $N\in{\mathbb N}$, in the present paper, whereas the Hilbert space theory for the general-$b$ case will be dealt with in a companion paper. The special cases $b=(N+1)a_-$, $N\in{\mathbb N}$, have been treated before in~Section~4 of~\cite{R00}, but in view of the above-mentioned ambiguity it is not immediate that the functions occurring there are basically the same as the functions $\psi((N+1)a_-;x,y)$. At the end of Section~3 we shall show that this holds true.
To be sure, at this point it is far from clear that even for the simplest cases $b=a_+$ and $b=a_-$ the eigenfunction $\psi(b;x,y)$ has all of the desired properties. We continue by working out the details for these $b$-values, since this involves little effort and the resulting formulas are illuminating.
As announced above, for~$b=a_+$ we are dealing with an equal-charge (reduced) 2-particle system that is `free', in the sense that no scattering occurs. Specifically, we have (cf.~Eq.~(4.7) in~\cite{R11})
\begin{equation}\label{N0}
{\cal R}_{\rm ren}(a_+;x,y)=\frac{\sin(\pi xy/a_+a_-)}{2s_-(x)s_-(y)}.
\end{equation}
Also, \eqref{defc} and~\eqref{Gades} imply
\begin{equation}\label{cwsp}
c(a_+;z)=1/2is_-(z),\ \ \ w(a_+;z)=4s_-(z)^2,
\end{equation}
so that \eqref{FR} yields
\begin{equation}
{\mathrm F}(a_+;x,y)=2\sin(\pi xy /a_+a_-).
\end{equation}
This entails that ${\cal F}_s(a_+)$~\eqref{cFs} amounts to the sine transform, as mentioned before.
On the other hand, from~\eqref{deftc} and~\eqref{deftw} we have
\begin{equation}
\tilde{c}(a_+;x)=1/2is_-(x-ia_+/2),\ \ \tilde{w}(a_+;x)=4s_-(x+ia_+/2)s_-(x-ia_+/2).
\end{equation}
Also, \eqref{cZ} gives
\begin{equation}\label{cZsp}
{\cal Z}(a_+;x,y)=-i\sin(\pi xy /a_+a_-)/s_-(x),
\end{equation}
so from~\eqref{defpsi} we obtain
\begin{multline}\label{psizero}
\psi(a_+;x,y) = \frac{(s_-(x+ia_+/2)s_-(x-ia_+/2))^{1/2}}{2s_-(ia_+-y)}\sum_{\tau=+,-}\tau
\frac{e_-(\tau (y-ia_+)/2)}{s_-(x-i\tau a_+/2)}
\\
\times \big(e_-(\tau y/2) \exp(i\pi xy/a_+a_-)-e_-(-\tau y/2) \exp(-i\pi xy/a_+a_-).
\end{multline}
With the substitutions~\eqref{xryk} and~\eqref{rhoka} in force, this yields the wave function~$\Psi(r,k)$ featuring in Prop.~B.3, cf.~\eqref{cFp0}.
Turning to the special case $b=a_-$, we can again make use of the formulas~\eqref{N0}, \eqref{cwsp} and~\eqref{cZsp}, but now with $a_+$ and~$a_-$ swapped. Then, \eqref{deftc} and~\eqref{deftw} yield the quite different outcome
\begin{equation}
\tilde{c}(a_-;x)=1/2c_+(x),\ \ \ \tilde{w}(a_-;x)=4c_+(x)^2.
\end{equation}
Using~\eqref{defpsi}, we now obtain the attractive wave function
\begin{equation}\label{psifree}
\psi(a_-;x,y)=\exp(i\pi xy/a_+a_-),
\end{equation}
which is manifestly `free', just as its repulsive counterpart.
Before specializing to the $b$-values~$(N+1)a_+$ for~$N>0$, it is expedient to derive already in this paper the general-$b$ asymptotic behavior of~$\psi(b;x,y)$ as $|{\rm Re}\, x|\to\infty$. Indeed, this will illuminate how we arrived at the above $y$-dependence.
To begin with, the $G$-function asymptotics~\eqref{Gas}
implies that the $c$-function~\eqref{defc} satisfies
\begin{equation}\label{cas}
c(b;z)\sim \phi(b)^{\pm 1}\exp(\mp \pi bz/a_+a_-),\ \ \ {\rm Re}\, z\to\pm\infty,
\end{equation}
where we have introduced the constant
\begin{equation}\label{phib}
\phi(b)\equiv \exp(i\pi b(b-2a)/2a_+a_-).
\end{equation}
Introducing next
\begin{equation}\label{u}
u(b;z)\equiv -c(b;z)/c(b;-z)=-\prod_{\delta=+,-}G(z+i\delta (a-b))/G(z+ i\delta a),
\end{equation}
we deduce
\begin{equation}\label{uas}
u(b;z)\sim -\phi(b)^{\pm 2},\ \ \ \ {\rm Re}\, z\to \pm\infty.
\end{equation}
Also, the reflection equation~\eqref{refl} and the complex conjugation relation~\eqref{Gcon}
entail
\begin{equation}
u(b;-z)u(b;z)=1,\ \ \ |u(b;z)|=1,\ \ b,z\in{\mathbb R}.
\end{equation}
The function~$u(b;y)$ encodes the scattering associated with the A$\Delta$O ${\cal A} (a_{+},a_-,b;x)$ and its modular partner, reinterpreted as commuting self-adjoint operators on the Hilbert space $L^2((0,\infty),dx)$. (To be sure, this reinterpretation requires $b\in[0,2a]$, cf.~\cite{R11}.)
The assertion just made hinges on the asymptotic behavior
of the joint eigenfunction~${\mathrm E}(b;x,y)$ of these A$\Delta$Os, defined by
\begin{equation}\label{ER}
{\mathrm E}(b;x,y)\equiv \frac{\phi(b)}{c(b;x)c(b;y)}{\cal R}_{\rm ren}(b;x,y).
\end{equation}
By contrast to ${\cal R}_{\rm ren}(b;x,y)$ and~${\mathrm F}(b;x,y)$, this function is not even, but satisfies
\begin{equation}\label{Erefl}
{\mathrm E}(b;-x,y)=-u(b;x){\mathrm E}(b;x,y).
\end{equation}
It follows from~Theorem~1.2 in~\cite{R03II} that the ${\mathrm E}$-function has asymptotics
\begin{equation}\label{EasR}
{\mathrm E}(b;x,y)\sim e^{i\pi xy/a_+a_-}- u(b;-y)e^{-i\pi xy/a_+a_-},\ \ (b,y)\in{\mathbb R}\times(0,\infty), \ \ {\rm Re}\, x\to \infty.
\end{equation}
(The specialization of the `$BC_1$'-functions from~\cite{R03II,R03III} to the `$A_1$'-functions of this paper is detailed in Section~2 of~\cite{R11}.) Using \eqref{Erefl} and~\eqref{uas}, this yields
\begin{equation}\label{EasL}
{\mathrm E}(b;x,y)\sim \phi(b)^2\big( e^{-i\pi xy/a_+a_-}- u(b;-y) e^{i\pi xy/a_+a_-}\big),\ \ (b,y)\in{\mathbb R}\times(0,\infty), \ {\rm Re}\, x\to -\infty.
\end{equation}
To complete our preparation for the last result of this section, we define transmission and reflection coefficients by
\begin{equation}\label{t}
t(b;y)\equiv \frac{s_-(y)}{s_-(ib-y)}u(b;y),
\end{equation}
\begin{equation}\label{r}
r(b;y)\equiv \frac{s_-(ib)}{s_-(ib-y)}u(b;y).
\end{equation}
This entails that when we set
\begin{equation}\label{sjk}
s_{jk}\equiv s(y_j-y_k),\ \ \ s=u,t,r,\ \ \ 1\le j<k\le 3,
\end{equation}
then the well-known $(u,t,r)$-Yang-Baxter equations given by
\begin{equation}\label{YB1}
r_{12}t_{13}u_{23}=t_{23}u_{13}r_{12}+r_{23}r_{13}t_{12},
\end{equation}
and
\begin{equation}\label{YB2}
u_{12}r_{13}u_{23}=t_{23}r_{13}t_{12}+r_{23}u_{13}r_{12},
\end{equation}
are easily verified. We recall that these equations encode the consistent factorization of the multi-particle $S$-matrix into a product of two-particle scattering amplitudes~\cite{YB90}.
\begin{proposition}
The dominant large-$|{\rm Re}\, x|$ asymptotic behavior of $\psi(b;x,y)$ is given by
\begin{equation}\label{psias}
\psi(b;x,y)\sim
\left\{
\begin{array}{ll}
t(b;y)\exp(i\pi xy/a_+a_-), & {\rm Re}\, x\to\infty, \\
\exp(i\pi xy/a_+a_-)-r(b;y)\exp(-i\pi xy/a_+a_-) , & {\rm Re}\, x\to-\infty ,
\end{array}
\right.
\end{equation}
where $b\in(-a_+/2,a_-+a_+/2)$ and $y>0$. Furthermore, we have an identity
\begin{equation}\label{psirev}
\psi(b;x,y)=t(b;y)\psi(b;-x,-y)-r(b;y) \psi(b;x,-y).
\end{equation}
\end{proposition}
\begin{proof}
Using the evenness of ${\cal R}_{\rm ren}(b;x,y)$ in~$x$ and~$y$, the identity~\eqref{psirev} follows from~\eqref{defpsi} by a straightforward calculation. (It encodes `time reversal invariance', cf.~the discussion below Theorem~D.1.)
In order to prove~\eqref{psias}, we first note that from \eqref{cZ} we have
\begin{equation}
{\cal Z}(b;x,y)=-\phi(b)^{-1}c(b;x)u(b;y){\mathrm E}(b;x,y).
\end{equation}
Combining this relation with~\eqref{cas}, \eqref{EasR} and~\eqref{EasL}, we obtain
\begin{equation}
{\cal Z}(b;x,y)\sim
\left\{
\begin{array}{ll}
e^{-\pi bx/a_+a_-}\big(e^{-i\pi xy/a_+a_-}-u(b;y)e^{i\pi xy/a_+a_-} \big), & {\rm Re}\, x\to\infty, \\
e^{\pi bx/a_+a_-}\big(e^{i\pi xy/a_+a_-}-u(b;y)e^{-i\pi xy/a_+a_-} \big), & {\rm Re}\, x\to-\infty .
\end{array}
\right.
\end{equation}
Now from \eqref{deftw}, \eqref{deftc} and \eqref{cas} we get
\begin{equation}
\tilde{w}(b;x)^{1/2}\sim \exp(\pm \pi bx/a_+a_-),\ \ \ {\rm Re}\, x\to\pm \infty,
\end{equation}
where we need $b\in(-a_+/2,a_-+a_+/2)$ to ensure $\tilde{w}(b;0)>0$. From this we deduce
\begin{multline}
\tilde{w}(b;x)^{1/2}{\cal Z}(b;x\pm ia_+/2,y)\sim \\
e_-(\mp i b/2)\big(e_-(\pm y/2)e^{-i\pi xy/a_+a_-}-e_-(\mp y/2)u(b;y)e^{i\pi xy/a_+a_-} \big),\ {\rm Re}\, x\to\infty,
\end{multline}
\begin{multline}
\tilde{w}(b;x)^{1/2}{\cal Z}(b;x\pm ia_+/2,y)\sim \\
e_-(\pm i b/2) \big(e_-(\mp y/2)e^{i\pi xy/a_+a_-}-e_-(\pm y/2)u(b;y)e^{-i\pi xy/a_+a_-} \big) ,\ {\rm Re}\, x\to-\infty.
\end{multline}
Hence \eqref{psipm} entails
\begin{multline}
2\psi_{\pm}(b;x,y)s_-(ib-y)\sim \\
\pm\big(e^{-i\pi xy/a_+a_-}-e_-(\mp y)u(b;y)e^{i\pi xy/a_+a_-} \big),\ \ \ {\rm Re}\, x\to\infty,
\end{multline}
\begin{multline}
2\psi_{\pm}(b;x,y)s_-(ib-y)\sim \\
\pm e_-(\pm ib)\big(e_-(\mp y)e^{i\pi xy/a_+a_-}-u(b;y)e^{-i\pi xy/a_+a_-} \big) , \ \ \ {\rm Re}\, x\to-\infty .
\end{multline}
Recalling \eqref{defpsi}, we now obtain~\eqref{psias}.
\end{proof}
\section{A close-up of the special eigenfunctions $\psi_N(x,y)$}
In this section we specialize the coupling constant~$b$ to the sequence of values $(N+1)a_+$, $N\in{\mathbb N}$, and study the opposite-charge eigenfunctions
\begin{equation}\label{psiN}
\psi_N(x,y)\equiv \psi((N+1)a_+;x,y),\ \ \ N\in{\mathbb N}.
\end{equation}
They are obtained from the same-charge eigenfunctions
\begin{equation}\label{RN}
R_N(x,y)\equiv {\cal R}_{\rm ren}((N+1)a_+;x,y),\ \ \ N\in{\mathbb N},
\end{equation}
via~\eqref{defpsi}. By contrast to the case of a generic coupling $b\in{\mathbb R}$, the $c$-function and $\tilde{w}$-function featuring in these formulas are elementary periodic functions, given by
\begin{equation}\label{crec}
1/c((N+1)a_+;-y)=\prod_{j=0}^N (-2i)s_-(y+ija_+),
\end{equation}
\begin{equation}\label{twN}
\tilde{w}((N+1)a_+;x)=\prod_{j=0}^N 4s_-(x+i(j+1/2)a_+)s_-(x-i(j+1/2)a_+).
\end{equation}
(This easily follows from the $G$-A$\Delta$Es~\eqref{Gades}.)
The crux of the special $b$-values is that the functions \eqref{RN} are elementary functions, too. Specifically, from Eqs.~(4.8) and (4.5) in~\cite{R11} we have
\begin{equation}\label{RNexp}
R_N(x,y)=(-i)^{N+1}(K_N(x,y)-K_N(x,-y))\Big/\prod_{j=-N}^N 4s_-(x+ija_+)s_-(y+ija_+),
\end{equation}
with (cf.~Eqs.~(4.10)--(4.12) in~\cite{R11})
\begin{equation}\label{KN}
K_N(x,y) \equiv \exp(i\pi xy/a_+a_-)\Sigma_N(x,y),
\end{equation}
\begin{equation}\label{SigN}
\Sigma_N(x,y)\equiv\sum_{k,l=0}^Nc_{kl}^{(N)}(e_-(ia_+))e_-((N-2k)x+(N-2l)y)).
\end{equation}
The coefficients in the sum are specified in Section~II of~\cite{R99}, where the functions $K_N(x,y)$ and $R_N(x,y)$ have been studied in great detail. (See also~\cite{DK00} for a related account.)
In the sequel we need substantial information concerning the above quantities, which we proceed to collect.
First, the coefficients $c_{kl}^{(N)}(q)$ are Laurent polynomials in~$q$ with integer coefficients. In particular, we have~$c^{(0)}_{00}=1$ (so that $R_0(x,y)$ reduces to~\eqref{N0}), and
\begin{equation}\label{ckl1}
c^{(1)}_{00}=c^{(1)}_{11}=q,\ \ c^{(1)}_{01}=c^{(1)}_{10}=-q^{-1},\ \ \ q=e_-(ia_+).
\end{equation}
Second, the coefficients have symmetries
\begin{equation}\label{csym}
c^{(N)}_{kl}=c^{(N)}_{lk}=c^{(N)}_{N-k,N-l}=(-)^N\overline{c^{(N)}_{k,N-l}},\ \ \ k,l=0,\ldots,N.
\end{equation}
Clearly, these are equivalent to the following features of $K_N$ and~$\Sigma_N$:
\begin{equation}\label{Sigsd}
K_N(x,y)=K_N(y,x),\ \ \ \Sigma_N(x,y)=\Sigma_N(y,x),
\end{equation}
\begin{equation}\label{Sigrefl}
K_N(x,y)=K_N(-x,-y),\ \ \ \Sigma_N(x,y)=\Sigma_N(-x,-y),
\end{equation}
\begin{equation}\label{Sigcon}
K_N(x,y)=(-)^N\overline{K_N(x,-y)},\ \ \Sigma_N(x,y)=(-)^N\overline{\Sigma_N(x,-y)},\ \ \ x,y\in{\mathbb R}.
\end{equation}
Third, the function $K_N(x,y)$ satisfies the A$\Delta$E
\begin{equation}\label{KNade}
s_-(x+iNa_{+})K_N(x-ia_+,y)+s_-(x-iNa_{+})K_N(x+ia_+,y)=2s_-(x)c_-(y)K_N(x,y).
\end{equation}
(This is a similarity transformed version of the A$\Delta$E~\eqref{Rade} with~$b=(N+1)a_+$.)
Fourth, we need the following explicit evaluations:
\begin{equation}\label{Kspec}
K_N(x,\pm iNa_+)=K_N(\pm iNa_+,y)=\prod_{j=N+1}^{2N}2s_-(ija_+),
\end{equation}
\begin{equation}\label{KBy}
K_N(\pm i(N-k)a_+,y)=i^NB^{(N)}_k(c_-(y)),\ \ \ k=0,\ldots,N,
\end{equation}
\begin{equation}\label{KBx}
K_N(x,\pm i(N-k)a_+)=i^NB^{(N)}_k(c_-(x)),\ \ \ k=0,\ldots,N.
\end{equation}
Here, $B^{(N)}_k(u)$ is a polynomial of degree~$k$ and parity~$(-)^k$ with real coefficients, and the restrictions
\begin{equation}\label{arestr}
ja_+\notin a_-{\mathbb N},\ \ \ j=1,\ldots, 2N,
\end{equation}
must be imposed for the degree-$k$ property to hold true.
In particular, from~\eqref{ckl1} we easily get
\begin{equation}\label{Bsp}
K_1(\pm ia_+,y)=2s_-(2ia_+),\ \ K_1(0,y)=4s_-(ia_+)c_-(y),
\end{equation}
and the restrictions $a_+\neq la_-/2$, $l\in{\mathbb N}$, ensure that the coefficients do not vanish.
The above features can all be gleaned from~Theorem~II.1 in~\cite{R99}. Finally, we also need the summation identity
\begin{equation}\label{SNasp}
\sum_{l=0}^Nc_{0l}^{(N)}e_-((N-2l)y)=\prod_{j=1}^N2s_-(y+ija_+),
\end{equation}
which results upon combining Eqs.~(2.19)--(2.21) with Eq.~(2.55) in~\cite{R99}. From this we obtain a second identity
\begin{equation}\label{SNasm}
\sum_{l=0}^Nc_{Nl}^{(N)}e_-((N-2l)y)=(-)^N\prod_{j=1}^N2s_-(y-ija_+),
\end{equation}
by using $c^{(N)}_{Nl}=c^{(N)}_{0,N-l}$ and relabeling.
We are now in the position to work out an explicit expression for $\psi_N(x,y)$ by using the above building blocks. From this and the summation identities~\eqref{SNasp} and~\eqref{SNasm} we can then determine the asymptotic behavior of~$\psi_N(x,y)$. First, however, we do so for
\begin{equation}\label{EN}
{\mathrm E}_N(x,y)\equiv {\mathrm E} ((N+1)a_+;x,y),
\end{equation}
since this yields a simple template for the application of the above formulas.
From~\eqref{phib} we calculate
\begin{equation}\label{phiN}
\phi((N+1)a_+)=(-i)^{N+1}e_-(i(N+1)Na_+/2),
\end{equation}
and then~\eqref{ER} and~\eqref{crec} yield
\begin{equation}\label{ENexp}
{\mathrm E}_N(x,y)=e_-(i(N+1)Na_+/2)
(K_N(x,y)-K_N(x,-y))\Big/\prod_{j=1}^N 4s_-(x+ija_+)s_-(y+ija_+).
\end{equation}
Combining \eqref{KN} and \eqref{SigN} with~\eqref{SNasp}, it is now easy to verify
\begin{equation}\label{ENas}
{\mathrm E}_N(x,y)= e^{i\pi xy/a_+a_-}-u_N(-y)e^{-i\pi xy/a_+a_-}+O(e_-(-2 x)),\ \ \ {\rm Re}\, x\to\infty,
\end{equation}
where
\begin{equation}\label{uN}
u_N(z)\equiv u((N+1)a_+;z)=\prod_{j=1}^N \frac{s_-(ija_++z)}{s_-(ija_+-z)},
\end{equation}
and where the implied constant can be chosen uniformly for~${\rm Im}\, x$ varying over ${\mathbb R}$-compacts and~$y$ varying over compact subsets of ${\mathbb C}$ with the $y$-poles removed.
This agrees with the specialization of~\eqref{EasR}, and by using~\eqref{SNasm} we readily obtain a more precise version of~\eqref{EasL} as well.
Turning to $\psi_N(x,y)$, we begin by defining a weight function~$w_N(x)$ that plays the role of $w(r)$ in Appendix~A. It is given by (recall~\eqref{twN})
\begin{equation}\label{wN}
w_N(x)\equiv 1/\tilde{w}((N+1)a_+;x)=1\Big/ \prod_{j=0}^N 4s_-(x+i(j+1/2)a_+)s_-(x-i(j+1/2)a_+).
\end{equation}
Next, we introduce a function $v_N(y)$ that corresponds to the function $v(k)$ from Appendix~B. To ease the notation, we define its reciprocal:
\begin{equation}\label{vyrec}
1/v_N(y)\equiv 2s_-(y-i(N+1)a_+)c(((N+1)a_+;-y)\prod_{j=-N}^N2s_-(y+ija_+).
\end{equation}
Using~\eqref{crec} and simplifying, this yields
\begin{equation}\label{vhy}
v_N(y)=1\Big/\prod_{j=1}^{N+1}2is_-(y-ija_+).
\end{equation}
From~\eqref{defpsi} we now obtain, using the above formulas,
\begin{multline}\label{psiNexp}
\psi_N(x,y)=(-)^Ni^{N+1}w_N(x)^{1/2}v_N(y) \\
\times\sum_{\delta=+,-}2\delta s_-(x-i\delta (N+1/2)a_+)e_-(\delta(i(N+1)a_+-y)/2)\\
\times \Big( \exp(i\pi xy/a_+a_-)e_-(-\delta y/2)\Sigma_N(x+i\delta a_+/2,y) -(y\to -y)\Big).
\end{multline}
As a check, note this coincides with \eqref{psizero} for $N=0$.
We proceed to derive properties of~$\psi_N(x,y)$ from these explicit formulas. A first conclusion is that we have
\begin{equation}\label{psiNcon}
\overline{\psi_N(x,y)}=\psi_N(x,-y),\ \ \ (x,y)\in{\mathbb R}^2.
\end{equation}
Indeed, this readily follows by combining \eqref{wN}--\eqref{psiNexp} with the conjugation relation~\eqref{Sigcon} for $\Sigma_N$.
Next, we obtain the large-$|{\rm Re}\, x|$ asymptotics of $\psi_N(x,y)$. Just as for $E_N(x,y)$, the identities \eqref{SNasp} and \eqref{SNasm} are the key ingredients. A little more effort is needed due to the extra factors, but it is still straightforward to deduce
\begin{equation}\label{psiNasx}
\psi_N(x,y)=
\left\{
\begin{array}{ll}
t_N(y)e^{i\pi xy/a_+a_-}+O(e_-(-2x)), & {\rm Re}\, x\to\infty, \\
e^{i\pi xy/a_+a_-}-r_N(y)e^{-i\pi xy/a_+a_-} +O(e_-(2x)) , & {\rm Re}\, x\to-\infty ,
\end{array}
\right.
\end{equation}
with
\begin{equation}\label{tNrN}
t_N(y)\equiv\frac{s_-(y)}{s_-(i(N+1)a_+-y)}u_N(y),\ \ \ r_N(y)\equiv\frac{s_-(i(N+1)a_+)}{s_-(i(N+1)a_+-y)}u_N(y),
\end{equation}
where the implied constants can be chosen uniformly for~${\rm Im}\, x$ varying over ${\mathbb R}$-compacts and~$y$ varying over ${\mathbb C}$-compacts with the $y$-poles removed. This renders the dominant general-$b$ asymptotics given by~\eqref{t}--\eqref{psias} more precise.
We can also use the dual versions of~\eqref{SNasp}--\eqref{SNasm} (i.~e., $y\to x$ and $c_{0l}/c_{Nl}\to c_{l0}/c_{lN}$) to obtain the large-$|{\rm Re}\, y|$ asymptotics of $\psi_N(x,y)$. This again needs a bit of work, the result being
\begin{equation}\label{psiNasy}
\psi_N(x,y)=
\left\{
\begin{array}{ll}
w_N(x)^{1/2}C_N(x)\exp(i\pi xy/a_+a_-)+O(e_-(-y) ), & {\rm Re}\, y\to\infty, \\
w_N(x)^{1/2}\overline{C_N(x)}\exp(i\pi xy/a_+a_-)+O(e_-(y) ), & {\rm Re}\, y\to-\infty ,
\end{array}
\right.
\end{equation}
where
\begin{equation}\label{CN}
C_N(x)\equiv (-)^{N+1}e_-(i(N+1)^2a_+/2)
\prod_{j=0}^N2s_-(x+i(j+1/2)a_+),
\end{equation}
and the implied constants are uniform for~${\rm Im}\, y$ in ${\mathbb R}$-compacts and~$x$ varying over ${\mathbb C}$-compacts with the $x$-singularities removed. Thus we obtain a reflectionless asymptotics of the form
\begin{equation}
\psi_N(x,y)\sim U_N(x)^{\pm 1/2} \exp(i\pi xy/a_+a_-),\ \ \ {\rm Re}\, y\to\pm\infty,
\end{equation}
with
\begin{equation}\label{UN}
U_N(x)\equiv e_-(i(N+1)^2a_+)
\prod_{j=0}^N\frac{s_-(x+i(j+1/2)a_+)}{s_-(x-i(j+1/2)a_+)}.
\end{equation}
We stress that there are no restrictions on $a_+$ and $a_-$ in~\eqref{psiNasx}--\eqref{UN}, save for our standing positivity assumption. This reveals a major flaw of time-independent (as opposed to time-dependent) scattering theory in the present setting: The unitary asymptotics exhibited by these formulas does not show any anomaly. However, as we shall see later on, we must \emph{restrict} the scale parameters for (a suitable multiple of) $\psi_N(x,y)$ to yield the integral kernel of a \emph{unitary} eigenfunction transform.
In order to clarify the unitarity issue (`orthogonality and completeness' in quantum-mechanical parlance), we rely on the results in the appendices. To make contact with their setup, we need to verify the various assumptions made there. To this end we rewrite~\eqref{psiNexp} as
\begin{equation}\label{psiNalt}
\psi_N(x,y)=w_N(x)^{1/2}v_N(y)\sum_{\tau=+,-}\ell_N^{\tau}(x,y)\exp(i\tau \pi xy/a_+a_-),
\end{equation}
so that we obtain entire coefficients
\begin{multline}\label{ellNm}
\ell_N^{\tau}(x,y)=(-)^Ni^{N+1} \tau \sum_{\delta=+,-}2\delta s_-(x-i\delta (N+1/2)a_+)
\\
\times e_-(\delta(i(N+1)a_+-y )/2)e_-(-\delta\tau y/2)
\Sigma_N(x+i\delta a_+/2,\tau y).
\end{multline}
From this we read off the (anti)periodicity features
\begin{equation}\label{perx}
\ell_N^{\tau}(x+ia_-,y)=(-)^{N+1}\ell_N^{\tau}(x,y),\ \ \ \tau=+,-,
\end{equation}
\begin{equation}\label{pery}
\ell_N^{\tau}(x,y+ia_-)=\tau(-)^{N+1}\ell_N^{\tau}(x,y),\ \ \ \tau=+,-.
\end{equation}
Also, from~\eqref{wN} it is clear that~$w_N(x)$ is $ia_-$-periodic, while~\eqref{vhy} shows that we have
\begin{equation}\label{vNper}
v_N(y+ia_-)=(-)^{N+1}v_N(y).
\end{equation}
When we now make the substitutions~\eqref{xryk}, then it follows from this that the assumptions in Appendix~A and~B concerning periodicity/antiperiodicity in the variables~$r$ and~$k$ are satisfied. Likewise, \eqref{Psiconj} is clear from~\eqref{psiNcon}, and the assumptions regarding asymptotic behavior in~$r$ and~$k$ are satisfied, too.
Finally, we show that the critical evenness assumptions hold true.
\begin{proposition}
With the substitutions~\eqref{xryk} in $\ell^{\pm}_N(x,y)$~\eqref{ellNm} and $v_N(y)$~\eqref{vhy}, the evenness assumptions \eqref{Lass} and~\eqref{Mass} hold true.
\end{proposition}
\begin{proof}
Staying with the variables~$x$ and~$y$,
\eqref{Lass} amounts to
invariance of
\begin{equation}\label{LNtt}
L^{\tau,\tau'}_N(y,x,x')\equiv \ell_N^{\tau}(x,y)\ell_N^{\tau'}(x',-y)+\ell_N^{-\tau}(-x,y)\ell_N^{-\tau'}(-x',-y),
\end{equation}
under taking $y,x,x'\to -y,-x,-x'$. To check this symmetry property, we suppress the subscripts in~\eqref{ellNm} and use the notation
\begin{equation}
\eta_N\equiv i(N+1/2)a_+,\ \ \ \nu=i(N+1)a_+/2.
\end{equation}
Then, \eqref{ellNm} entails
\begin{multline}
\ell^{\tau}(x,y)\ell^{\tau'}(x',-y)\sim \tau\tau'\sum_{\delta,\delta'}\delta\de' s(x-\delta\eta_N)s(x'-\delta'\eta_N)
\Sigma(x+\delta \eta_0,\tau y)
\Sigma(x'+\delta' \eta_0,-\tau' y)
\\
\times e((\delta'\tau'-\delta\tau)y/2)
e((\delta+\delta')\nu)e((\delta'-\delta)y/2).
\end{multline}
From this we deduce by using~\eqref{Sigrefl},
\begin{multline}\label{Lttex}
L^{\tau,\tau'}_N(y,x,x')\sim
\sum_{\delta,\delta'}\delta\de' s(x-\delta\eta_N)s(x'-\delta'\eta_N)
\Sigma(x+\delta \eta_0,\tau y)
\Sigma(x'+\delta' \eta_0,-\tau' y)
\\
\times e((\delta'\tau'-\delta\tau)y/2)
\big[e((\delta+\delta')\nu)e((\delta'-\delta)y/2)+(\delta,\delta'\to -\delta,-\delta')\big]
\\
=\sum_{\delta}s(x-\delta\eta_N)s(x'-\delta\eta_N)
\Sigma(x+\delta \eta_0,\tau y)
\Sigma(x'+\delta \eta_0,-\tau' y)
e(\delta(\tau'-\tau)y/2)\cdot 2c(2\nu)
\\
-\sum_{\delta}s(x-\delta\eta_N)s(x'+\delta\eta_N)
\Sigma(x+\delta \eta_0,\tau y)
\Sigma(x'-\delta \eta_0,-\tau' y)
e(-\delta(\tau'+\tau)y/2)\cdot 2c(y).
\end{multline}
Invoking~\eqref{Sigrefl} once more, we now see that both sums are invariant under taking $y,x,x'$ to~$-y,-x,-x'$.
Having verified~\eqref{Lass}, the weaker assumptions~\eqref{Mass} follow, too.
\end{proof}
As promised below~\eqref{defpsi}, we conclude this section by deriving explicit expressions for the attractive eigenfunctions~$\psi((N+1)a_-;x,y)$ with $N>0$, and comparing them to Section~4 in~\cite{R00}. (Recall we already calculated~$\psi(a_-;x,y)$, cf.~\eqref{psifree}.) To this end we can make use of~\eqref{crec} and~\eqref{RNexp}--\eqref{SigN} with~$a_+$ and~$a_-$ swapped, whereas~\eqref{twN} is replaced by
\begin{equation}\label{twNrep}
\tilde{w}((N+1)a_-;x)=2c_+(x)\prod_{j=-N}^N 2c_+(x-ija_-).
\end{equation}
Employing these formulas in combination with~\eqref{defpsi}, we obtain
\begin{multline}\label{psiNr=0}
\psi((N+1)a_-;x,y)=\exp(i\pi xy/a_+a_-)\\
\times
\frac{\sum_{k,l=0}^N(-)^kc_{kl}^{(N)}(e_+(ia_-))e_+((N-2k)x+(N-2l)y))}{\prod_{j=1}^N2s_+(y-ija_-)[4c_+(x-ija_-)c_+(x+ija_-)]^{1/2} }.
\end{multline}
We point out that these functions are not only joint eigenfunctions of $\tilde{H}((N+1)a_-;x)$ and~${\cal S}((N+1)a_-;y)$ (cf.~\eqref{tHade} and~\eqref{cSade}), but also of the two `free' A$\Delta$Os
\begin{equation}\label{2free}
\exp(-ia_+\partial_x)+\exp(ia_+\partial_x),\ \ \ \exp(-ia_+\partial_y)+\exp(ia_+\partial_y).
\end{equation}
(Indeed, the functions multiplying the plane wave are $ia_+$-periodic in~$x$ and~$y$.) By contrast, for other $b$-values no additional independent A$\Delta$Os appear to exist for which $\psi(b;x,y)$ are eigenfunctions.
The connection with the functions $F_a(\nu,\beta;x,p)$ given by Eq.~(1.32) of~\cite{R00} can be made by setting
\begin{equation}\label{RIMS}
a_+=\pi/\nu,\ \ a_-=\hbar\beta,\ \ y=\beta p/\nu.
\end{equation}
Then the dimensionless parameter $a=\hbar\beta\nu$ used there equals $\pi a_-/a_+$, and the coefficient matrices are related by
\begin{equation}
(-)^mc_{mn}^{(N)}(e_+(ia_-))=i^Nc^{(a)}_{mn},\ \ \ m,n=0,\ldots,N,
\end{equation}
cf.~Eq.~(1.39) in~\cite{R00}. With these reparametrizations, we wind up with the relation
\begin{equation}\label{psiFa}
F_a(\nu,\beta;x,p)=(-i)^N\psi((N+1)a_-;x,y)\prod_{j=1}^N[s_+(y-ija_-)/s_+(y+ija_-)]^{1/2}.
\end{equation}
We mention that in~\cite{DK00} these reflectionless eigenfunctions were tied in with basic hypergeometric series. It is an interesting question whether this link can still be made for the eigenfunctions~$\psi((N+1)a_+;x,y)$.
\section{The transforms associated with $\psi_N(x,y)$}
In this section we focus on the Hilbert space aspects of the opposite-charge A$\Delta$Os
\begin{equation}\label{HN}
\tilde{H}_N(x)\equiv \tilde{H}((N+1)a_+;x)=\exp(-ia_-\partial_x) +\exp(ia_-\partial_x),
\end{equation}
\begin{equation}\label{cSN}
{\cal S}_N(y)\equiv {\cal S}((N+1)a_+;y)=\exp(-ia_-\partial_y) -\exp(ia_-\partial_y),
\end{equation}
(cf.~\eqref{tHb} and~\eqref{cS}), and their joint eigenfunctions~$\psi_N(x,y)$~\eqref{psiNexp}. To this end we invoke the results in Appendix~A and~B, with $\psi_N(x,y)$ (reparametrized by~\eqref{xryk}) in the role of $\Psi(r,k)$. We are entitled to do so, since we have shown in the previous section that the assumptions made there are satisfied for generic scale parameters $a_+,a_-$. We have not isolated the exceptional parameters yet, and now proceed by studying this issue.
To start with, it easily follows from \eqref{psiNexp} that the transform resulting from~\eqref{cF}--\eqref{Fm} is bounded whenever $w_N(x)$ and $v_N(y)$ have no pole at the origin. Indeed, the integral kernels are then given by a finite sum of terms of the form
\begin{equation}\label{cFform}
B(r)\exp(i\tau rk)\hat{B}(k),\ \ \ \tau \in \{ +,-\},
\end{equation}
where $B(r)$ and $\hat{B}(k)$ are bounded functions on~${\mathbb R}$ and~$[0,\infty)$, resp. Clearly, each such term gives rise to a product of three bounded operators.
The only eventual constraint on $a_+$ and $a_-$ encountered thus far comes from the condition `no poles at the origin' just mentioned. We are using the adjective `eventual', since we do not know whether it can ever happen that the wave function~$\psi(b;x,y)$ itself has a pole for~$x=0$ or~$y=0$. We can illustrate this with~$\psi_0(x,y)$~\eqref{psizero}: Its factors are singular at the origin if and only if $a_+\in a_-{\mathbb N}^*$, but in fact we easily calculate
\begin{equation}\label{psi0free}
\psi_0(x,y)=(-)^{l-1}\exp(i\pi xy/a_+a_-),\ \ \ a_+=la_-,\ \ \ l\in{\mathbb N}^*.
\end{equation}
More generally, whenever the parameter~$b$ is simultaneously a multiple of~$a_+$ and of~$a_-$ (so that the factors of~$\psi_N(x,y)$ have poles at the origin), the function $\psi_N(x,y)$ is actually a (constant) phase multiple of the plane wave~$\exp(i\pi xy/a_+a_-)$. This can be concluded from the features of the functions $F_a(\nu,\beta;x,p)$~\eqref{psiFa} expounded at the end of~Section~4 in~\cite{R00}.
Fortunately our ignorance about this absence of real poles is of little consequence, as we now explain. First, it is clear from its definition~\eqref{wN} that $w_N(x)$ has no real poles for
\begin{equation}\label{unitary}
a_->(N+1/2)a_+=b-a_+/2.
\end{equation}
For the remainder of this section, we restrict attention to this parameter interval.
As we shall show in the next one, for $a_-\le (N+1/2)a_+$ the transform associated with~$\psi_N(x,y)$ is not isometric, save for a discrete parameter set, hence not acceptable from a quantum-mechanical viewpoint.
Second, from~\eqref{vhy} we read off that~$v_N(y)$ has no real poles for~$a_->(N+1)a_+$. For~$a_-=(N+1)a_+$, however, $\psi_N(x,y)$ reduces to the plane wave~\eqref{psifree}, so the corresponding transform amounts to the Fourier transform. For the remaining interval~$a_-\in ((N+1/2)a_+,(N+1)a_+)$ we do not get any real poles in~$v_N(y)$, so the upshot is that there is no real pole problem for the parameters~\eqref{unitary} at issue in this section.
In Appendix~A and~B, however, we have additional assumptions concerning poles, which amount to the poles of the weight functions~$w_N(x)$ and~$v_N(y)v_N(-y)$ being simple. Clearly, in the interval~\eqref{unitary} $w_N(x)$ has double poles in the critical strip~${\rm Im}\, x\in (0,a_-)$ whenever the poles at $x=i(j+1/2)a_+$, $j=0,\ldots,N$, collide with the poles at $x=ia_--i(k+1/2)a_+$, $k=0,\ldots,N$. However, this is of no consequence for the first major result of this section.
\begin{theorem}
With the parameter restriction~\eqref{unitary} in effect, the transform~${\cal F}_N(\rho,\kappa)$ given by~\eqref{cF}--\eqref{Fm} with kernel
\begin{equation}
\Psi(r,k) =\psi_N(a_- r/\rho,a_- k/\kappa ),\ \ \ \rho\kappa =\pi a_-/a_+,
\end{equation}
is an isometry.
\end{theorem}
\begin{proof}
The requirement~\eqref{unitary} amounts to
\begin{equation}\label{rkint}
\rho\kappa\in ((N+1/2)\pi,\infty).
\end{equation}
There is a finite set~${\cal E}_N$ of exceptional $\rho\kappa$-values in this interval for which at least one pole of (cf.~\eqref{wN})
\begin{equation}\label{wr}
w(r)=1\Big/ \prod_{j=0}^N 4\sinh\Big(\frac{\pi }{\rho}\Big( r+i(j+1/2)\frac{\pi}{\kappa}\Big)\Big)
\sinh\Big(\frac{\pi }{\rho}\Big( r-i(j+1/2)\frac{\pi}{\kappa}\Big)\Big),
\end{equation}
is not simple. But the bounded multiplication operators featuring in the transform kernel (cf.~\eqref{cFform}) are clearly strongly continuous in $\rho$ on the whole interval~\eqref{rkint}, so the same is true for the transform~${\cal F}_N(\rho,\kappa)$. Thus we need only show
isometry for the non-exceptional subintervals to conclude isometry for all of~\eqref{rkint}.
To prove isometry for
\begin{equation}
\rho\kappa \in ((N+1/2)\pi,\infty)\setminus {\cal E}_N,
\end{equation}
we may invoke Theorem~A.1, since all of its assumptions are satisfied, with $L$ equal to $N+1$. In view of its Corollary~A.2, it suffices to show that each of the~$N+1$ terms in the sum on the rhs of~\eqref{Rdk} vanishes.
To this end, consider the dependence on the indices~$\nu$ and~$\nu'$ of one of the terms. For it to vanish, it is sufficient that the bracketed expression on the second line of~\eqref{Rdk} is proportional to~$\nu\nu'$. Indeed, assuming it is, we can invoke the identity~\eqref{id1} to infer that the term vanishes.
Now in the case at hand, we have
\begin{equation}
\mu^{\tau}(\rho x/a_-,\kappa y/a_-)=
v_N(y)\ell^{\tau}_N(x,y)\exp(i\tau \pi xy/a_+a_-),\ \ \ \tau=+,-,
\end{equation}
where~$\ell^{\tau}_N(x,y)$ is given by~\eqref{ellNm}. Recalling~\eqref{KN}, we see that the only dependence on~$\tau$ comes from the factor
\begin{equation}
\tau K_N(x+i\delta a_+/2,\tau y).
\end{equation}
Moreover, for the present case we may choose as pole locations
\begin{equation}\label{rj}
r_j\equiv i\pi (j+1/2)/\kappa\Leftrightarrow x_j\equiv i(j+1/2)a_+,\ \ \ j=0,\ldots,N.
\end{equation}
Thus we only encounter the values
\begin{equation}\label{Kval}
K_N(\pm ima_+,\tau y),\ \ \ m=0,\ldots,N+1,
\end{equation}
with the case $m=N+1$ corresponding to~$x_N$ and $\delta=+$. In the latter case, however, the factor~$s_-(x-i\delta(N+1/2)a_+)$ in~\eqref{ellNm} vanishes.
With the parameter restrictions~\eqref{arestr} in force, we can invoke~\eqref{KBy} to conclude that the values~\eqref{Kval} with~$m<N+1$ are even in~$y$. By continuity this is still true for the excluded parameters. Hence these values do not depend on~$\tau$.
The upshot is that the two $\mu^{\tau}$-products in~\eqref{Rdk} solely depend on~$\nu$ and~$\nu'$ via a factor $\nu\nu'$, so that all terms in the sum vanish.
\end{proof}
The proof of the above theorem only made use of Appendix~A. To obtain the next theorem, we need the results of Appendix~B.
\begin{theorem}
Assume
\begin{equation}\label{parunit}
a_-\in[(N+1)a_+,\infty).
\end{equation}
Then the transform~${\cal F}_N(\rho,\kappa)$ of Theorem~4.1 is unitary.
\end{theorem}
\begin{proof}
By Theorem~4.1 the transform is isometric, and it equals the Fourier transform at the interval endpoint, cf.~\eqref{psifree}. Letting next
\begin{equation}
a_-\in ((N+1)a_+,\infty) \Leftrightarrow \rho\kappa \in ((N+1)\pi,\infty),
\end{equation}
there is a finite set~$\hat{{\cal E}}_N$ of exceptional $\rho\kappa$-values in this interval for which at least one pole of (cf.~\eqref{vhy})
\begin{equation}\label{hatw}
\hat{w}(k)=\Big( \prod_{j=1}^{N+1} 4\sinh\big(\frac{\pi k}{\kappa}+ij\frac{\pi^2}{\rho\kappa}\big)
\sinh\big(\frac{\pi k}{\kappa}-ij\frac{\pi^2}{\rho\kappa}\big) \Big)^{-1},
\end{equation}
is not simple. By continuity, we need only prove unitarity for
\begin{equation}
\rho\kappa \in ((N+1)\pi,\infty)\setminus \hat{{\cal E}}_N.
\end{equation}
For such parameters all assumptions of Theorem~B.1 are satisfied, with~$\hat{L}=N+1$.
As a consequence, it suffices to prove that each of the $N+1$ terms on the rhs of~\eqref{defR} vanishes. We choose as pole locations
\begin{equation}\label{kj}
k_j\equiv i\pi j/\rho\Leftrightarrow y_j\equiv ija_+,\ \ \ j=1,\ldots,N+1.
\end{equation}
Now from~\eqref{ellNm} we get
\begin{multline}\label{lamN}
\lambda^{\tau}_N(x,y)\equiv \exp(i\tau\pi xy/a_+a_-)\ell_N^{\tau}(x,y) =(-)^Ni^{N+1} \tau \sum_{\delta=+,-}2\delta s_-(x-i\delta (N+1/2)a_+)
\\
\times e_-(\delta(i(N+1)a_+-y )/2)
K_N(x+i\delta a_+/2,\tau y).
\end{multline}
The dependence on $\tau$ and $\tau'$ of the two $\lambda$-products in~\eqref{defR} is therefore given by an overall factor $\tau\tau'$ and by the arguments of the two $K_N$-functions.
But just as in the previous proof, it follows from~\eqref{KBx} that the values
\begin{equation}\label{Kvalres}
K_N(x,\pm \tau ija_+),\ \ \ j=1,\ldots,N,
\end{equation}
do not depend on~$\tau$.
In view of the identity~\eqref{id1} with $A,A'\to iA,iA'$, we then deduce that the terms in the sum on the rhs of~\eqref{defR} vanish for $j=1,\ldots,N$.
Next, we set
\begin{multline}\label{Lam}
\Lambda_N^{\tau,\tau'}(y,x,x')\equiv \lambda_N^{\tau}(x, y )\lambda_N^{\tau'}(x',-y )+\lambda_N^{-\tau}(- x,y )\lambda_N^{-\tau'}(-x',- y )
\\
=L_N^{\tau,\tau'}(y,x,x')\exp(i\pi (\tau x-\tau' x')y/a_+a_-),
\end{multline}
and consider~$\Lambda_N^{\tau,\tau'}(y_{N+1},x,x')$. From~\eqref{lamN} we get
\begin{multline}
\Lambda_N^{\tau,\tau'}(y_{N+1},x,x')=(-)^{N+1}8\tau\tau' c_-(i(N+1)a_+)\sum_{\delta,\delta'=+,-}\delta\de' s_-(x-i\delta (N+1/2)a_+)\\
\times s_-(x'-i\delta' (N+1/2)a_+)
K_N(x+i\delta a_+/2,\tau y_{N+1})K_N(x'+i\delta' a_+/2,-\tau' y_{N+1}),
\end{multline}
where we have used \eqref{Sigrefl} to rewrite the second $\lambda$-product. (Note that this agrees with~\eqref{Lttex}.) Invoking~\eqref{lamN} once more, this can be rewritten as
\begin{equation}\label{Lamalt}
\Lambda_N^{\tau,\tau'}(y_{N+1},x,x')=-2c_-(i(N+1)a_+)\lambda_N^{\tau}(x,i(N+1)a_+)\lambda_N^{-\tau'}(x',i(N+1)a_+).
\end{equation}
We continue by evaluating~$\lambda_N^{\tau}(x,i(N+1)a_+)$ explicitly. First, from~\eqref{lamN} we have
\begin{equation}\label{lamid}
\lambda_N^{\tau}(x,i(N+1)a_+)= (-i)^{N+1}\tau p_N,
\end{equation}
where we have introduced
\begin{equation}\label{pN}
p_N\equiv 2\sum_{\nu=+,-}\nu s_-(x+i\nu (N+1/2)a_+)K_N(x-i\nu a_+/2,i\tau (N+1)a_+).
\end{equation}
We are suppressing the manifest $\tau$- and $x$-dependence of the entire functions on the rhs for the new quantity~$p_N$, since it does not depend on these variables. We shall prove this claim shortly. Taking it for granted, we may take $x=i(N+1/2)a_+ $ and invoke~\eqref{Kspec} to obtain the product formula
\begin{equation}\label{pNexp}
p_N=\prod_{j=N+1}^{2N+1}2s_-(ija_+).
\end{equation}
Hence we have
\begin{equation}
\lambda_N^{\tau}(x,i(N+1)a_+)=\tau \prod_{j=N+1}^{2N+1}2\sin(\pi ja_+/a_-),
\end{equation}
so that
\begin{equation}\label{Lamspec}
\Lambda_N^{\tau,\tau'}(y_{N+1},x,x')=2\tau\tau'\cos(\pi(N+1)a_+/a_-)\prod_{j=N+1}^{2N+1}4\sin(\pi ja_+/a_-)^2.
\end{equation}
Thus the~$j= N+1$ term in the sum on the rhs of~\eqref{defR} vanishes, too, and the theorem follows.
It remains to prove our constancy claim. To this end, consider the $x$-dependence of the sum $S(x)$ on the rhs of~\eqref{pN}. It is clear from the definitions~\eqref{KN} and~\eqref{SigN} that $S(x)$ has period~$2ia_-$. Next, we have
\begin{multline}
S(x+ia_+/2)-S(x-ia_+/2)=\\
s_-(x+i(N+1)a_+)K_N(x,\tau y_{N+1} )
-s_-(x-iNa_+)K_N(x+i a_+,\tau y_{N+1})\\
-s_-(x+iNa_+)K_N(x-i a_+,\tau y_{N+1})+s_-(x-i(N+1)a_+)K_N(x,\tau y_{N+1} ) ,
\end{multline}
so from the A$\Delta$E~\eqref{KNade} we deduce that the rhs equals
\begin{equation}
\big( s_-(x+i(N+1)a_+)+s_-(x-i(N+1)a_+)-2s_-(x)c_-(y_{N+1})\big) K_N(x,\tau y_{N+1} )=0.
\end{equation}
Therefore $S(x)$ has period~$ia_+$, too. Taking $a_+/a_-$ irrational, it follows that $S(x)$ is constant. By virtue of real-analyticity in $x,a_+,a_-$, it then follows that $S(x)$ can only depend on $a_+$ and~$a_-$, and so our claim is proved.
\end{proof}
For parameters satisfying~\eqref{parunit}, this theorem enables us to promote the A$\Delta$Os~\eqref{HN} and~\eqref{cSN} to self-adjoint Hilbert space operators. Indeed, for this parameter range all assumptions of Appendix~D are satisfied, so that we may associate to $\tilde{H}_N(x)$ the operator~$M_{CM}$ (cf.~\eqref{muCM} and~\eqref{HCM}), and to ${\cal S}_N(y)$ the operator~$\hat{D}_{CM}$ (cf.~\eqref{dCM} and~\eqref{HhCM}). From this the following result is nearly immediate.
\begin{corollary}
Assume $a_+,a_-$ satisfy~\eqref{parunit}. Then the self-adjoint operator $M_{CM}$ associated to the A$\Delta$O $\tilde{H}_N(x)$ has absolutely continuous spectrum $[2,\infty)$ with multiplicity two, whereas the self-adjoint operator~$\hat{D}_{CM}$ associated to the A$\Delta$O ${\cal S}_N(y)$ has absolutely continuous spectrum $(-\infty,\infty)$ with multiplicity one.
\end{corollary}
\begin{proof}
Indeed, the unitary transform and its adjoint intertwine the respective operators with multiplication operators for which these spectral features are plain.
\end{proof}
Even though the action of the above operators on the functions in their domains is just the free A$\Delta$O action, they are quite different from the Hilbert space versions of the A$\Delta$Os defined via the Fourier transform~${\cal F}_0$, as evidenced by the nontrivial $S$-operators at hand.
A further comment on this distinction may be in order: It arises from the vastly different definition domains. In particular, viewing~$M_{CM}$ as a self-adjoint operator on $L^2({\mathbb R},dx)$ by undoing the reparametrization~\eqref{xryk}, its definition domain consists of functions that have an analytic continuation to the strip $|{\rm Im}\, x|<a_-$, but the locations of their square-root branch points for~$x\in i(-a_-,a_-)$ (which can be read off from the pertinent eigenfunction transform) entail that a pairwise intersection of domains for different parameters consists of the zero function. This domain behavior is radically different from that for a pair of distinct Hilbert space versions~$\hat{D}_1,\hat{D}_2$ of the same free (i.~e., constant-coefficient) differential operator~$D$. Indeed, typically their Hilbert space domains are encoded in distinct boundary conditions, so that the domain intersection is still a dense subspace, but neither a core for~$\hat{D}_1$ nor for~$\hat{D}_2$.
We proceed with the last theorem of this section.
\begin{theorem}
Letting
\begin{equation}\label{parbs}
a_-\in((N+1/2)a_+,(N+1)a_+),
\end{equation}
the transform~${\cal F}_N(\rho,\kappa)$ of Theorem~4.1
satisfies
\begin{equation}\label{cFN}
{\cal F}_N(\rho,\kappa)^*{\cal F}_N(\rho,\kappa)={\bf 1}_{ \hat{{\cal H} }},\ \ \ {\cal F}_N(\rho,\kappa){\cal F}_N(\rho,\kappa)^*={\bf 1}_{ {\cal H} }-\Psi_N\otimes \Psi_N/(\Psi_N,\Psi_N)_1.
\end{equation}
Here, we have
\begin{equation}\label{PsiN}
\Psi_N(r)\equiv 2\cosh(\kappa r) w_N(a_- r/\rho)^{1/2},
\end{equation}
with~$w_N(x)$ defined by~\eqref{wN}, and the inner product is given by
\begin{equation}\label{ipN}
(\Psi_N,\Psi_N)_1=(-)^{N+1}\frac{\pi\prod_{j=1}^N\sin (j\pi^2/\rho\kappa)}{\kappa\prod_{j=N+1}^{2N+1}\sin (j\pi^2/\rho\kappa)},\ \ \ \ \ \ \rho\kappa\in((N+1/2)\pi,(N+1)\pi).
\end{equation}
\end{theorem}
\begin{proof}
Proceeding as in the proof of Theorem~4.2, we choose again as simple pole locations in the strip ${\rm Im}\, k\in (0,\kappa)$ the numbers
\begin{equation}\label{kjres}
k_j\equiv ij\pi /\rho\Leftrightarrow y_j\equiv ija_+,\ \ \ j=1,\ldots,N,
\end{equation}
but now we need to choose
\begin{equation}\label{kspec}
k_{N+1}\equiv i (N+1)\pi /\rho-i\kappa\Leftrightarrow y_{N+1}\equiv i(N+1)a_+-ia_-,
\end{equation}
since we have $(N+1)\pi/\rho\in (\kappa,2\kappa)$. Thus the residue terms with $j=1,\ldots,N$ vanish as before, but we need to reconsider the $j=N+1$ contribution.
Now from \eqref{lamN} and~\eqref{pery} we obtain
\begin{equation}\label{lamper}
\lambda^{\tau}_N(x,y-ia_-)=\tau (-)^{N+1}e_+(\tau x)\lambda^{\tau}_N(x,y),
\end{equation}
so \eqref{Lam} entails
\begin{equation}
\Lambda_N^{\tau,\tau'}(y-ia_-,x,x') = \tau\tau'e_+(\tau x-\tau'x')\Lambda_N^{\tau,\tau'}(y,x,x').
\end{equation}
On account of~\eqref{Lamspec} we then get
\begin{equation}\label{Lambs}
\Lambda_N^{\tau,\tau'}(i(N+1)a_+-ia_-,x,x')= 2e_+(\tau x-\tau' x')\cos(\pi(N+1)a_+/a_-)\prod_{j=N+1}^{2N+1}4\sin(\pi ja_+/a_-)^2.
\end{equation}
The upshot is that the residue sum~\eqref{defR} reduces to
\begin{multline}\label{defRN}
R(r,r')=2i\hat{w}_{N+1}\cos((N+1)\pi^2/\rho\kappa)\prod_{j=N+1}^{2N+1}4\sin( j\pi^2/\rho\kappa )^2 \\
\times\sum_{\tau,\tau'=+,-}\frac{ \exp(\kappa(\tau r-\tau' r')}{1-\tau\tau'\exp(\kappa (\tau' r'-\tau r))} .
\end{multline}
Next, from \eqref{hatw} we calculate
\begin{equation}
\hat{w}_{N+1}=(-)^N\frac{\kappa}{i\pi}\frac{\sin((N+1)\pi^2/\rho\kappa)}{\prod_{j=1}^{2N+2}2\sin(j\pi^2/\rho\kappa)},
\end{equation}
and via the identity resulting from~\eqref{AAp} and~\eqref{S0} we obtain the sum in~\eqref{defRN}. Simplifying, we get the final result
\begin{equation}
R(r,r')=(-)^N\frac{4\kappa}{\pi}\cosh(\kappa r)\cosh(\kappa r')\frac{\prod_{j=N+1}^{2N+1}\sin (j\pi^2/\rho\kappa)}{\prod_{j=1}^N \sin(j\pi^2/\rho\kappa)}.
\end{equation}
From this we arrive at~\eqref{cFN} and~\eqref{ipN} by the same reasoning as for the special case $N=0$, cf.~\eqref{Psi0}--\eqref{inpr0}.
\end{proof}
To conclude this section, we derive further information concerning the bound state $\Psi_N(r)$ given by~\eqref{PsiN}. It is expedient to do so in its guise (cf.~\eqref{xryk})
\begin{multline}\label{psiNx}
\psi_N(x)\equiv \Psi_N(\rho x/a_-)= 2c_+(x)w_N(x)^{1/2}
\\
=2c_+(x)\Big/\prod_{j=0}^{N}\big[4s_-(x-i(j+1/2)a_+)s_-(x+i(j+1/2)a_+)\big]^{1/2}.
\end{multline}
Let us consider the poles of $\psi_N(x,y)$ in the strip ${\rm Im}\, y\in(0,a_-)$. Recalling~\eqref{vhy} and~\eqref{psiNexp}, we see that poles of $\psi_N(x,y)$ can only occur at the locations $y\equiv ija_+$ (mod $ia_-$), with $j=1,\ldots ,N+1$.
Requiring first $a_-\in ((N+1)a_+,\infty)$, it follows that the only eventual pole locations in the critical strip are the numbers
\begin{equation}\label{yj}
y_j := ija_+,\ \ \ j=1,\ldots,N+1.
\end{equation}
Now $\psi_N(x,y)$ has factors $K_N(x\pm ia_+/2,y)-K_N(x\pm ia_+/2,-y)$ (cf.~\eqref{psiNexp}), and on account of~\eqref{KBx}, these factors vanish at $y_j$ for $j=1,\ldots,N$. Therefore $\psi_N(x,y)$ is regular at $y_1,\ldots,y_N$.
Next, we recall that we have
\begin{equation}
\psi_N(x,y)=w_N(x)^{1/2}v_N(y)\sum_{\tau=+,-}\lambda^{\tau}_N(x,y).
\end{equation}
In view of the identity~\eqref{lamid}, this entails that $\psi_N(x,y)$ is regular at $y_{N+1}$ as well.
The upshot is that $\psi_N(x,y)$ is holomorphic in the critical strip ${\rm Im}\, y\in(0,a_-)$ for the $a_-$-interval featuring in Theorem~4.2. (Recall $\psi_N(x,y)$ is equal to the plane wave~\eqref{psifree} at the endpoint.)
For the $a_-$-interval~\eqref{parbs} of Theorem~4.4, however, the pole locations $y_1,\ldots,y_N$ are still in the critical strip, whereas~$y_{N+1}$ is not. Obviously, $\psi_N(x,y)$ is still regular at $y_1,\ldots,y_N$. But now we also have a pole of $v_N(y)$ in the strip at $y_{N+1}-ia_-$. We can determine its residue by combining~\eqref{lamper} and~\eqref{lamid}. Indeed, from these identities we deduce
\begin{equation}
\lambda_N^{\tau}(x,y_{N+1}-ia_-)=i^{N+1}p_Ne_+(\tau x),\ \ \tau=+,-.
\end{equation}
Therefore, the residue is proportional to the bound state~\eqref{psiNx}. To be specific, from~\eqref{vhy} we calculate
\begin{equation}
{\rm Res}\ v_N(y)\Big|_{y=y_{N+1}-ia_-}=i^{N+1}a_-/\pi \prod_{j=1}^N 2s_-(ija_+),
\end{equation}
and together with~\eqref{pNexp} this yields
\begin{equation}
{\rm Res}\ \psi_N(x,y)\Big|_{y=y_{N+1}-ia_-}=
(-)^{N+1} \frac{ia_-\prod_{j=N+1}^{2N+1}\sin (ja_+/a_- )}{\pi\prod_{j=1}^N \sin(ja_+/a_- )}\psi_N(x).
\end{equation}
The eigenvalue of the A$\Delta$O $\tilde{H}_N(x)$ \eqref{HN} on $\psi_N(x)$~\eqref{psiNx} is therefore given by
\begin{equation}
2c_+(i(N+1)a_+-ia_-)=2(-)^{N+1}\cos(\pi a_-/a_+).
\end{equation}
(This can be easily checked directly.) Of course, we define the operator on $L^2({\mathbb R},dx)$ associated to $\tilde{H}_N(x)$ to have the same eigenvalue, and then the last result of this section easily follows.
\begin{corollary}
Assuming \eqref{parbs}, the self-adjoint operator associated to $\tilde{H}_N(x)$ has a nondegenerate positive eigenvalue
\begin{equation}
E_N\equiv 2(-)^{N+1}\cos(\pi a_-/a_+)\in (0,2),\ \ \ a_-\in((N+1/2)a_+,(N+1)a_+),
\end{equation}
below its absolutely continuous spectrum $[2,\infty)$ with multiplicity~2. The corresponding eigenfunction~$\psi_N(x)$~\eqref{psiNx} has norm given by~\eqref{ipN}.
\end{corollary}
In view of Theorem~D.1, Theorem~4.2 and Theorem~4.4, the self-adjoint Hamiltonian on~$L^2({\mathbb R},dx)$ associated to the A$\Delta$O $\tilde{H}_N(x)$ yields a well-defined time-dependent scattering theory for $a_->(N+1/2)a_+$, with its $S$-operator encoded in the transmission and reflection coefficients $t_N(y)$ and~$r_N(y)$ given by~\eqref{tNrN} and~\eqref{uN}. Note that the `Jost function' $\psi_N(x,y)/t_N(y)$ converges to a multiple of the bound state~\eqref{psiNx} as~$y$ converges to $i(N+1)a_+-ia_-$.
\section{Isometry breakdown}
As we have seen in the proof of Theorem 4.3, the reason for the transform~${\cal F}$ not being unitary is the presence of a nonzero residue term. It emerges when $a_-$ decreases beyond the critical value $(N+1)a_+$, at which the transform reduces to the Fourier transform~${\cal F}_0$. Since the adjoint ${\cal F}^*$ of~${\cal F}$ is not isometric for $a_-\in((N+1/2)a_+,(N+1)a_+)$, we can no longer use~${\cal F}^*$ to associate a self-adjoint operator to ${\cal S}_N(y)$~\eqref{cSN}. Even so, since ${\cal F}$ is still an isometry in this $a_-$-interval, ${\cal F}^*$ is still a partial isometry.
As we shall make clear in this section, for generic $a_-\in (0,(N+1/2)a_+)$ isometry of the eigenfunction transform ${\cal F}$ breaks down, so that we cannot use it any longer to associate a self-adjoint operator to $\tilde{H}_N(x)$~\eqref{HN}. This is due to nonzero residue terms that emerge when the numbers (cf.~\eqref{rj})
\begin{equation}
x_j:=i(j+1/2)a_+,\ \ \ j=0,\ldots,N,
\end{equation}
move out of the strip ${\rm Im}\, x\in(0,a_-)$ as $a_-$ decreases. More precisely, the nonzero terms are spawned by poles $x_j-in_ja_-$, with $n_j>0$ chosen such that these locations are in the critical strip. (Except for the $N=0$ case, we do not consider the $a_-$-values such that the origin is among the locations $x_j-ila_-$, $l\in{\mathbb N}^*$.)
We proceed to analyze the state of affairs for the $N=0$ case in complete detail. (On first reading, the reader may wish to skip to Proposition~5.1, in which the results are summarized.) We first recall that the transform ${\cal F}_0(\rho,\kappa)$ of Theorem~4.1 equals the transform~${\cal F}_+(\phi_0)$ of Proposition~B.3, and that we have special cases
\begin{equation}
{\cal F}_0(\rho,\kappa)=(-)^{l-1}{\cal F}_0,\ \ \ \ \ a_-=a_+/l\Leftrightarrow \rho\kappa=\pi/l,\ \ \ l\in{\mathbb N}^*,
\end{equation}
cf.~\eqref{psi0free}. Thus the transform is not only isometric for $a_+<2a_-$, but also for~$a_+$ equal to an arbitrary multiple of~$a_-$.
Consider now the remaining $\rho\kappa$-intervals
\begin{equation}\label{Ipm}
I_n^-\equiv \Big(\frac{\pi}{2n+2},\frac{\pi }{2n+1}\Big),\ \ I_n^+\equiv \Big(\frac{\pi }{2n+1},\frac{\pi }{2n}\Big),\ \ \ \ n\in{\mathbb N}^*.
\end{equation}
We have (cf.~\eqref{Psi0})
\begin{equation}
w_0(r)= 1\Big/4\sinh \Big(\frac{\pi}{\rho}\big( r+i\frac{\pi}{2\kappa}\big)\Big)\sinh \Big(\frac{\pi}{\rho}\big( r-i\frac{\pi}{2\kappa}\big)\Big),
\end{equation}
so we get a simple $w_0(r)$-pole in the critical strip ${\rm Im}\, r\in(0,\rho)$ at the location
\begin{equation}
r_1^{(n)}:= i\pi/2\kappa -in\rho\ne i\rho/2,
\end{equation}
with residue
\begin{equation}
w_1=\frac{-i\rho}{4\pi \sin(\pi^2/\rho\kappa)}.
\end{equation}
Also, from~\eqref{cFp0} we read off the summands $\mu^{\pm}(r,k)$, and then a straightforward calculation yields
\begin{equation}\label{mu0res}
\mu^{\tau}(\delta r_1^{(n)},k)=\tau (-)^{n+1}\exp(\delta\tau n\rho k)\sinh\big(i\pi^2/\rho\kappa\big)\frac{\exp\big(\delta(\pi k/2\kappa-i\pi^2/2\rho\kappa)\big) }{\sinh(\pi k/\kappa-i\pi^2/\rho\kappa)}.
\end{equation}
Therefore, the residue sum~\eqref{Rdk} becomes
\begin{multline}
\hat{R}_{\delta,\delta'}(k,k')= \frac{\rho\sin(\pi^2/\rho\kappa)\hat{s}_n(k,k')}{4\pi\sinh\big(\pi k/\kappa+i\pi^2/\rho\kappa\big)\sinh\big(\pi k'/\kappa-i\pi^2/\rho\kappa\big) }
\\
\times \Big(\exp\Big(\delta\Big(\frac{\pi k}{2\kappa}+\frac{i\pi^2}{2\rho\kappa}\Big)-\delta'\Big(\frac{\pi k'}{2\kappa}-\frac{i\pi^2}{2\rho\kappa}\Big)\Big) +(\delta,\delta'\to-\delta,-\delta')\Big),
\end{multline}
where
\begin{eqnarray}
\hat{s}_n(k,k') & \equiv & \sum_{\nu,\nu'=+,-}\frac{\nu\nu'\exp(n\rho(\nu k-\nu' k'))}{1-\exp(\rho(\nu' k'-\nu k))}
\nonumber \\
& = & \frac{\sinh((n+1/2)\rho(k-k'))}{\sinh( \rho(k-k')/2)}-
\frac{\sinh((n+1/2)\rho(k+k'))}{\sinh( \rho(k+k')/2)}.
\end{eqnarray}
Now we have a recurrence
\begin{eqnarray}
(\hat{s}_n-\hat{s}_{n-1})(k,k') & = & 2\cosh(n\rho (k-k'))-2\cosh(n\rho (k+k'))
\nonumber \\
& = & -4\sinh(n\rho k)\sinh(n\rho k'),
\end{eqnarray}
whence we easily deduce
\begin{equation}
\hat{s}_n(k,k')=-4\sum_{j=1}^n \sinh(j\rho k)\sinh(j\rho k').
\end{equation}
Introducing
\begin{equation}
\chi_{\delta}^{(\alpha,j)}(k)\equiv \frac{\sinh(j\rho k)}
{\sinh(\pi k/\kappa+i\pi^2/\rho\kappa)}\exp\Big(\alpha\delta\Big(\frac{\pi k}{2\kappa}+\frac{i\pi^2}{2\rho\kappa}\Big)\Big),\ \ \ \alpha,\delta=+,-,
\end{equation}
we therefore obtain
\begin{equation}\label{Rchi}
\hat{R}_{\delta,\delta'}(k,k')=- \frac{\rho\sin(\pi^2/\rho\kappa)}{\pi}\sum_{\alpha=+,-}\sum_{j=1}^n
\chi_{\delta}^{(\alpha,j)}(k)\overline{ \chi_{\delta'}^{(-\alpha,j)}(k')}.
\end{equation}
Finally, we use the parity operator~$\hat{{\cal P}}$~\eqref{pari} to first substitute
\begin{equation}
\chi^{(-,j)}=-\hat{{\cal P}}\chi^{(+,j)},
\end{equation}
and then we rewrite~\eqref{Rchi} in terms of the even and odd functions
\begin{eqnarray}\label{chie}
\chi^{(e,j)}(k) & \equiv & \chi^{(+,j)}(k)+\hat{{\cal P}}\chi^{(+,j)}(k)
\nonumber \\
& = & \frac{\sinh(j\rho k)}
{\cosh(\pi k/2\kappa+i\pi^2/2\rho\kappa)} \left(\begin{array}{c}
1 \\
-1
\end{array}\right),\ \ \ \ \ k>0,
\end{eqnarray}
\begin{eqnarray}\label{chio}
\chi^{(o,j)}(k) & \equiv & \chi^{(+,j)}(k)-\hat{{\cal P}}\chi^{(+,j)}(k)
\nonumber \\
& = & \frac{\sinh(j\rho k)}
{\sinh(\pi k/2\kappa+i\pi^2/2\rho\kappa)} \left(\begin{array}{c}
1 \\
1
\end{array}\right),\ \ \ \ \ k>0.
\end{eqnarray}
As a result, we obtain the manifestly self-adjoint rank-$(2n)$ kernel
\begin{equation}\label{hatRfin}
\hat{R}_{\delta,\delta'}(k,k')= \frac{\rho\sin(\pi^2/\rho\kappa)}{2\pi} \sum_{j=1}^n\Big(
\chi_{\delta}^{(e,j)}(k)\overline{ \chi_{\delta'}^{(e,j)}(k')}-\chi_{\delta}^{(o,j)}(k)\overline{ \chi_{\delta'}^{(o,j)}(k')}\Big),
\end{equation}
with $\delta,\delta'=+,-$, and $ k,k'>0$.
We continue by calculating the residue sum~$R(r,r')$~\eqref{defR}. The dual weight function is given by
\begin{equation}\label{wod}
\hat{w}_0(k)=1\Big/4\sinh \Big(\frac{\pi}{\kappa}\big( k+i\frac{\pi}{\rho}\big)\Big)\sinh \Big(\frac{\pi}{\kappa}\big( k-i\frac{\pi}{\rho}\big)\Big),
\end{equation}
and we consider its simple poles in the critical strip ${\rm Im}\, k\in(0,\kappa)$ of the form
\begin{equation}
k_1^{(m)}=i\pi/\rho -im\kappa \ne i\kappa/2,\ \ \ m\in{\mathbb N}^*,
\end{equation}
with residue
\begin{equation}\label{hatres}
\hat{w}_1=\frac{-i\kappa}{4\pi \sin(2\pi^2/\rho\kappa)}.
\end{equation}
More specifically, we choose $\rho\kappa$ in the interval (cf.~\eqref{Ipm})
\begin{equation}\label{Im}
I_m\equiv \Big(\frac{\pi}{m+1},\frac{\pi}{m}\Big)
=\left\{\begin{array}{cc}
I_{m/2}^+, & m\ \mathrm{even}, \\
I_{(m-1)/2}^-, & m\ \mathrm{odd},
\end{array}\right.
\end{equation}
with~$\rho\kappa\ne \pi/(m+1/2)$ to avoid the double pole location. Using~\eqref{Psi0} we calculate
\begin{equation}
\lambda^+(r,k_1^{(m)})=2(-)^me^{m\kappa r}\sin(\pi^2/\rho\kappa),
\end{equation}
\begin{equation}
\lambda^+(r,-k_1^{(m)})=2(-)^me^{-m\kappa r}\big(\sin(2\pi^2/\rho\kappa)-e^{2\pi r/\rho}\sin(\pi^2/\rho\kappa)\big),
\end{equation}
\begin{equation}
\lambda^-(r,k_1^{(m)})=-2\sin(\pi^2/\rho\kappa)e^{-m\kappa r},
\end{equation}
\begin{equation}
\lambda^-(r,-k_1^{(m)})=-2\sin(\pi^2/\rho\kappa)e^{m\kappa r}e^{-2\pi r/\rho},
\end{equation}
and then the definition~\eqref{Lamdef} yields
\begin{equation}\label{Lameq}
\Lambda^{\tau,\tau}(k_1^{(m)},r,r')=4 \sin(2\pi^2/\rho\kappa)\sin(\pi^2/\rho\kappa)\exp(\tau m\kappa( r-r')),
\end{equation}
\begin{equation}\label{Lamdi}
\Lambda^{\tau,-\tau}(k_1^{(m)},r,r')= (-)^{m+1}4\sin(2\pi^2/\rho\kappa)\sin(\pi^2/\rho\kappa)\exp (\tau m\kappa( r+r')).
\end{equation}
(Note this checks with~\eqref{Lamspec} for $m=0$ and with~\eqref{Lambs} for~$m=1$.)
The residue sum~$R(r,r')$~\eqref{defR} can therefore be written
\begin{equation}
R(r,r')=\frac{\kappa\sin(\pi^2/\rho\kappa)}{\pi}w_0(r)^{1/2}w_0(r')^{1/2}s_m(r,r'),
\end{equation}
where
\begin{eqnarray}
s_m(r,r') & \equiv & \sum_{\tau=+,-}\Big( \frac{\exp(\tau m\kappa(r-r'))}{1-\exp(\tau\kappa(r'-r))}+(-)^{m+1}\frac{\exp(\tau m\kappa(r+r'))}{1-\exp(-\tau\kappa(r'+r))}
\nonumber \\
& = & \frac{\sinh((m+1/2)\kappa(r-r'))}{\sinh( \kappa(r-r')/2)}+(-)^{m+1}
\frac{\cosh((m+1/2)\kappa(r+r'))}{\cosh( \kappa(r+r')/2)}.
\end{eqnarray}
This sum obeys the recurrence
\begin{eqnarray}
(s_m-s_{m-1})(r,r') & = & 2\cosh(m\kappa (r-r'))+(-)^{m+1}2\cosh(m\kappa (r+r'))
\nonumber \\
& = & \left\{\begin{array}{cc}
-4\sinh(m\kappa r)\sinh(m\kappa r'), & m\ \mathrm{even}, \\
4\cosh(m\kappa r)\cosh(m\kappa r') , & m\ \mathrm{odd},
\end{array}\right.
\end{eqnarray}
whose unique solution reads
\begin{equation}
s_m(r,r')=\sum_{j=0}^{m-1} (-)^jh_j(r)h_j(r'),
\end{equation}
with
\begin{equation}\label{hj}
h_j(r)=\left\{\begin{array}{cc}
2\sinh(j\kappa r), & j\ \mathrm{odd}, \\
2\cosh(j\kappa r) , & j\ \mathrm{even}.
\end{array}\right.
\end{equation}
As a consequence, we obtain the manifestly self-adjoint rank-$m$ kernel
\begin{equation}\label{Rfin}
R(r,r')=\frac{\kappa\sin(\pi^2/\rho\kappa)}{\pi}\sum_{j=0}^{m-1} (-)^j\Psi^{(j)}(r)\Psi^{(j)}(r'),\ \ \ \rho\kappa\in I_m\setminus \{ \pi/(m+1/2)\},
\end{equation}
where
\begin{equation}\label{Psij}
\Psi^{(j)}(r)\equiv h_j(r)w_0(r)^{1/2}.
\end{equation}
We now summarize the above $N=0$ results.
\begin{proposition}
Letting~$\rho\kappa\le\pi/2$, the transform ${\cal F}_0(\rho,\kappa)={\cal F}_+(\phi_0)$ of Proposition~B.3 has the following properties. For~$\rho\kappa$ in the intervals~$I_n^-$ and~$I_n^+$ defined by~\eqref{Ipm}, it satisfies
\begin{equation}\label{stFF}
{\cal F}_0(\rho,\kappa)^*{\cal F}_0(\rho,\kappa)={\bf 1}_{\hat{ {\cal H}} }+
\frac{\rho\sin(\pi^2/\rho\kappa)}{2\pi} \sum_{j=1}^n\Big(
\chi^{(e,j)}\otimes \overline{ \chi^{(e,j)}}-\chi^{(o,j)}\otimes \overline{ \chi^{(o,j)}}\Big),
\end{equation}
with $\chi^{(e,j)}$/$\chi^{(o,j)}$ given by~\eqref{chie}/\eqref{chio}. For~$\rho\kappa$ in the interval~$I_m$~\eqref{Im}, it satisfies
\begin{equation}\label{FstF}
{\cal F}_0(\rho,\kappa){\cal F}_0(\rho,\kappa)^*={\bf 1}_{ {\cal H} }+\frac{\kappa\sin(\pi^2/\rho\kappa)}{\pi}\sum_{j=0}^{m-1} (-)^j\Psi^{(j)}\otimes \Psi^{(j)},
\end{equation}
with $\Psi^{(j)}$ defined by~\eqref{Psij} and~\eqref{hj}. Furthermore, at the interval endpoints we have
\begin{equation}\label{cFmfree}
{\cal F}_0(\rho,\kappa)=(-)^{m-1}{\cal F}_0,\ \ \rho\kappa =\pi/m,\ \ \ \ m=2,3,\ldots,
\end{equation}
where ${\cal F}_0$ is the Fourier transform.
\end{proposition}
\begin{proof}
We obtain~\eqref{stFF} from Theorem~A.1 and~\eqref{hatRfin}. Likewise, \eqref{FstF} follows from Theorem~B.1 and~\eqref{Rfin}, except at the midpoint of~$I_m$, at which~$\hat{w}_0(k)$~\eqref{wod} has a double pole. However, the pole on the rhs of~\eqref{hatres} is cancelled by a zero coming from~\eqref{Lameq}--\eqref{Lamdi}, so that $R(r,r')$ is regular for~$\rho\kappa=\pi/(m+1/2)$. By continuity, therefore, \eqref{FstF} holds on all of~$I_m$. Finally, \eqref{cFmfree} follows from~\eqref{psi0free}.
\end{proof}
For $N>0$ we need to keep track of more than one pole moving out of the ${\rm Im}\, r$- and ${\rm Im}\, k$-strips as $a_-$ decreases. We shall only detail the case where the `highest' $r$-pole has moved out, so as to explicitly reveal the isometry breakdown for the interval
\begin{equation}\label{isovio}
a_-\in (Na_+,(N+1/2)a_+)\Leftrightarrow \rho\kappa \in (N\pi, (N+1/2)\pi).
\end{equation}
Thus we can still work with the pole locations~$r_j$~\eqref{rj} for $j=0,\ldots,N-1$, but now we need
\begin{equation}
r_N^{(1)}:=i\pi(N+1/2)/\kappa-i\rho\Leftrightarrow x_N^{(1)}:=i(N+1/2)a_+-ia_-.
\end{equation}
From~\eqref{KN} and~\eqref{SigN} we deduce
\begin{equation}
K_N(\delta (Na_+-ia_-),y)=(-)^Ne_+(\delta y)K_N(\delta iNa_+,y)=(-)^Ne_+(\delta y)\prod_{j=N+1}^{2N} 2s_-(ija_+),
\end{equation}
so via~\eqref{lamN} we obtain
\begin{equation}
\mu^{\tau}_N(\delta x_N^{(1)},y)=\tau v_N(y)e_+(\delta\tau y)e_-(\delta(y-i(N+1)a_+)/2)\prod_{j=N+1}^{2N+1} 2is_-(ija_+).
\end{equation}
Recalling~\eqref{vhy}, this entails
\begin{multline}
\mu^{\tau}_N(\delta a_- r_N^{(1)}/\rho,a_-k/\kappa)=\tau \exp(\delta\tau \rho k)\exp\big(\delta\big(\pi k/2\kappa-i(N+1)
\pi^2/2\rho\kappa\big)\big)
\\
\times \frac{\prod_{j=N+1}^{2N+1}\sinh\big(ij \pi^2/\rho\kappa\big)}{\prod_{j=1}^{N+1}\sinh\big(\pi k/\kappa-ij \pi^2/\rho\kappa\big)}.
\end{multline}
As should be the case, this reduces to~\eqref{mu0res} for $N=0$ and~$n=1$. Using the residue (cf.~\eqref{wr})
\begin{equation}
w_N=\rho\Big(2\pi\prod_{j=1}^{2N+1}2\sinh(ij\pi^2/\rho\kappa)\Big)^{-1},
\end{equation}
and following the same steps as for $N=0$, this readily yields the arbitrary-$N$ and $n=1$ counterpart of~\eqref{stFF}. Specifically, we obtain
\begin{equation}\label{stFFN}
{\cal F}_N(\rho,\kappa)^*{\cal F}_N(\rho,\kappa)={\bf 1}_{\hat{ {\cal H}} }+(-)^N
\frac{\rho\prod_{j=N+1}^{2N+1}\sin(j\pi^2/\rho\kappa)}{2\pi\prod_{j=1}^{N}\sin(j\pi^2/\rho\kappa)} \Big(
\chi_N^{(e,1)}\otimes \overline{ \chi_N^{(e,1)}}-\chi_N^{(o,1)}\otimes \overline{ \chi_N^{(o,1)}}\Big),
\end{equation}
with even and odd functions
\begin{eqnarray}\label{chieN}
\chi_N^{(e,1)}(k)
& \equiv & \frac{1}{\prod_{j=1}^N2\sinh(\pi k/\kappa+ij\pi^2/\rho\kappa)}
\nonumber \\
& \times & \frac{\sinh(\rho k)}
{\cosh(\pi k/2\kappa+i(N+1)\pi^2/2\rho\kappa)} \left(\begin{array}{c}
1 \\
-1
\end{array}\right),
\end{eqnarray}
\begin{eqnarray}\label{chioN}
\chi_N^{(o,1)}(k)
& \equiv & \frac{1}{\prod_{j=1}^N2\sinh(\pi k/\kappa+ij\pi^2/\rho\kappa)}
\nonumber \\
& \times & \frac{\sinh(\rho k)}
{\sinh(\pi k/2\kappa+i(N+1)\pi^2/2\rho\kappa)} \left(\begin{array}{c}
1 \\
1
\end{array}\right).
\end{eqnarray}
An inspection of the proof of Theorem~4.3 reveals that the reasoning can be applied to the interval~\eqref{isovio} as well. This readily yields
\begin{equation}
{\cal F}_N(\rho,\kappa){\cal F}_N(\rho,\kappa)^*={\bf 1}_{ {\cal H} }+(-)^N\frac{\kappa\prod_{j=N+1}^{2N+1}\sin (j\pi^2/\rho\kappa)}{\pi\prod_{j=1}^N\sin (j\pi^2/\rho\kappa)} \Psi_N\otimes \Psi_N,
\end{equation}
with $\Psi_N$ given by \eqref{PsiN} and~\eqref{wN}. This concludes our account of isometry violation for the interval~\eqref{isovio}.
The reader who has followed us to this point will realize that all ingredients are in place to handle the intervals arising when $a_-$ is further decreased, but we shall not pursue this.
\begin{appendix}
\section{The transform ${\cal F}$}\label{AppA}
Our general transform ${\cal F}$ involves two Hilbert spaces
\begin{equation}\label{cH1}
{\cal H}\equiv L^2({\mathbb R}, dr),
\end{equation}
and
\begin{equation}\label{cH2}
\hat{{\cal H}}\equiv L^2((0,\infty), dk)\otimes {\mathbb C}^2,
\end{equation}
with inner products
\begin{equation}
(d,e)_1\equiv \int_{-\infty}^{\infty} dr\,\overline{d(r)}e(r),\ \ \ d,e\in{\cal H},
\end{equation}
\begin{equation}
(f,g)_2\equiv \sum_{\delta=+,-} \int_{0}^{\infty} dk\, \overline{f_{\delta}(k)}g_{\delta}(k),\ \ \ f,g\in\hat{{\cal H}}.
\end{equation}
(Physically speaking, these spaces can be thought of as the reduced position and momentum space of a particle pair; the variables~$r$ and~$k$ are dimensionless and related to center-of-mass position~$x=x_1-x_2$ and momentum~$p=p_1-p_2$ by $r=\nu x$ and $k=p/2\hbar\nu $.) The transform is at first defined on the dense $\hat{{\cal H}}$-subspace~$\hat{{\cal C}}$ of ${\mathbb C}^2$-valued smooth functions $f=(f_+,f_-)$ with compact support in~$(0,\infty)$, which are assumed to be mapped into~${\cal H}$:
\begin{equation}\label{cF}
{\cal F}\, :\, \hat{{\cal C}}\equiv C_0^\infty((0,\infty))^2\subset \hat{{\cal H}} \to {\cal H}.
\end{equation}
Specifically, we have
\begin{equation}\label{cFdef}
({\cal F} f)(r)=\frac{1}{\sqrt{2\pi}}\int_{0}^{\infty} dk \sum_{\delta=+,-} F_{\delta}(r,k)f_{\delta}(k).
\end{equation}
The two transform kernels are defined in terms of one function $\Psi(r,k)$:
\begin{equation}\label{Fp}
F_+(r,k)=\Psi(r,k),
\end{equation}
\begin{equation}\label{Fm}
F_-(r,k)=-\Psi(-r,k).
\end{equation}
For the case of no interaction we have
\begin{equation}\label{free}
\Psi(r,k)=\exp(irk),
\end{equation}
and then we denote the corresponding transform by~${\cal F}_0$. Thus, ${\cal F}_0$ amounts to the Fourier transform, with~$\hat{f}\in L^2({\mathbb R},dk)$ corresponding to~$f\in\hat{{\cal H}}$ via
\begin{equation}\label{ident}
\hat{f}(k)=
\left\{
\begin{array}{ll}
f_+(k), & k>0, \\
-f_-(-k), & k<0.
\end{array}
\right.
\end{equation}
(The phase of $F_-$ is a matter of convention, the minus sign being chosen for later purposes.)
We proceed with an initial list of assumptions on the function $\Psi(r,k)$. (This list will be supplemented in Appendix~B.) To begin with, we assume $\Psi(r,k)$ is a smooth function on ${\mathbb R}^2$ that satisfies
\begin{equation}\label{Psiconj}
\overline{\Psi(r,k)}=\Psi(r,-k),\ \ \ (r,k)\in{\mathbb R}^2,
\end{equation}
and is of the form
\begin{equation}\label{Psiform}
\Psi(r,k)= w(r)^{1/2}\sum_{\tau=+,-} m^{\tau}(r,k)\exp(i\tau rk).
\end{equation}
The weight function $w(r)$ is a positive even function and the positive square root is taken in~\eqref{Psiform}. It extends to a meromorphic function that has period $i\rho$ with~$\rho>0$. Its only singularities in the period strip ${\rm Im}\, r\in (0,\rho)$ are finitely many simple poles. If $r_0$ is one of these poles, it follows from evenness and $i\rho$-periodicity that $i\rho -r_0$ is another such pole whose residue has opposite sign. Thus we can pair off the poles, obtaining $2L$ distinct poles $r_1,\ldots,r_{2L}$ related by
\begin{equation}\label{poles}
r_{j+L}=i\rho -r_j,\ \ \ j=1,\ldots,L,
\end{equation}
with residues
\begin{equation}\label{wres}
w_j\equiv {\rm Res}\,( w(r))|_{r=r_j},\ \ \ j=1,\ldots,2L,
\end{equation}
satisfying
\begin{equation}\label{resid}
w_{j+L}=-w_j,\ \ \ j=1,\ldots,L.
\end{equation}
The two coefficients $m^{\pm}(r,k)$ are smooth and satisfy
\begin{equation}\label{mconj}
\overline{m^{\tau}(r,k)}=m^{\tau}(r,-k),\ \ \ \tau=+,-,\ \ \ (r,k)\in{\mathbb R}^2,
\end{equation}
in accord with~\eqref{Psiconj}. As functions of~$r$, they extend to entire functions that are either both $i\rho$-periodic or both $i\rho$-antiperiodic.
Next, we assume that the asymptotic behavior of $\Psi(r,k)$ for ${\rm Re}\, r\to \pm\infty$ is given by
\begin{equation}\label{mpas}
w(r)^{1/2}m^+(r,k)=
\left\{
\begin{array}{ll}
T(k)+O(\exp(-\gamma r)), & \re r \to \infty, \\
1+O(\exp(\gamma r)), & \re r \to -\infty,
\end{array}
\right.
\end{equation}
\begin{equation}\label{mnas}
w(r)^{1/2}m^-(r,k)=
\left\{
\begin{array}{ll}
O(\exp(-\gamma r)), & \re r \to \infty, \\
- R(k)+O(\exp(\gamma r)), & \re r \to -\infty,
\end{array}
\right.
\end{equation}
and that we also have
\begin{equation}\label{asextra}
\partial_r(w(r)^{1/2}m^\tau (r,k))=O(\exp(\mp\gamma r)),\ \ \tau=+,-,\ \ {\rm Re}\, r\to \pm \infty.
\end{equation}
Here, we have $\gamma>0$ and the implied constants can be chosen uniformly for ${\rm Im}\, r$ and~$k$ in ${\mathbb R}$-compacts. (To leave no doubt regarding the meaning of the uniformity assumption, let us spell it out for the bound~\eqref{asextra}: For given compact subsets $K_1,K_2$ of~${\mathbb R}$, it says that there exist positive constants~$C$ and~$R$ such that the modulus of the lhs is bounded above by $C\exp(-\gamma |{\rm Re}\, r|)$ for all $k\in K_1$ and all~$r\in{\mathbb C}$ with ${\rm Im}\, r\in K_2$ and $|{\rm Re}\, r| \ge R$.)
The transmission and reflection coefficients $T(k)$ and $R(k)$ are assumed to be smooth functions on ${\mathbb R}$ satisfying
\begin{equation}\label{trconj}
\overline{T(k)}=T(-k),\ \ \ \overline{R(k)}=R(-k),\ \ \ k\in{\mathbb R},
\end{equation}
\begin{equation}\label{unit1}
T(-k)T(k)+R(-k)R(k)=1,
\end{equation}
\begin{equation}\label{unit2}
T(-k)R(k)+R(-k)T(k)=0.
\end{equation}
As a consequence, the matrix multiplication operator on $\hat{{\cal H}}$ (`$S$-matrix') given by
\begin{equation}\label{Sm}
S(k)\equiv
\left(\begin{array}{cc}
T(k) & R(k) \\
R(k) & T(k)
\end{array}\right),\ \ \ k>0,
\end{equation}
is a unitary operator. We also point out that our initial assumption that ${\cal F}$ maps $\hat{{\cal C}}$ into ${\cal H}$ readily follows from the asymptotic behavior we have just assumed.
We need two more assumptions on the coefficients. To this end we introduce
\begin{equation}\label{Mtsdef}
M_{\alpha}^{\tau}(r,k,k')\equiv m^+(-r,-k)m^{\tau}(-\alpha r,k')+m^-(r,-k)m^{-\tau}(\alpha r,k'),\ \ \ \tau,\alpha=+,-.
\end{equation}
Then we assume
\begin{equation}\label{Mass}
M_{\alpha}^{\alpha}(r,k,k)=M_{\alpha}^{\alpha}(-r,k,k),\ \ \alpha=+,-,\ \ \ (r,k)\in{\mathbb R}^2.
\end{equation}
By contrast to previous ones, these assumptions may seem unintuitive. For now, we point out that
\eqref{mpas} and~\eqref{mnas} imply
\begin{equation}\label{M1as}
w(r)M_{+}^{+}(r,k,k')
=
\left\{
\begin{array}{ll}
1+O(\exp(-\gamma r)), & \re r \to \infty, \\
T(-k)T(k')+R(-k)R(k')+O(\exp(\gamma r)), & \re r \to -\infty,
\end{array}
\right.
\end{equation}
\begin{equation}\label{M2as}
w(r)M_{-}^{-}(r,k,k')
=
\left\{
\begin{array}{ll}
O(\exp(-\gamma r)), & \re r \to \infty, \\
-T(-k)R(k')-R(-k)T(k')+O(\exp(\gamma r)), & \re r \to -\infty.
\end{array}
\right.
\end{equation}
Consequently, the evenness assumptions~\eqref{Mass} may be viewed as generalizations of the unitarity assumptions~\eqref{unit1} and~\eqref{unit2}.
Before stating a theorem that only involves the above assumptions on the various coefficients determining $\Psi(r,k)$, we add two simple choices that satisfy these assumptions. Besides $\rho$, they involve two positive parameters~$\kappa$ and~$\varphi$. The parametrization of the $k$-dependence and the choice of numerical factors anticipate the further assumptions made in Appendix~B.
Specifically, we set
\begin{equation}\label{wsp}
w(r)\equiv 1/4\sinh(\pi r/\rho+i\varphi)\sinh(\pi r/\rho-i\varphi),
\end{equation}
\begin{equation}\label{mnu}
m_{\sigma}^{\tau}(r,k)\equiv \frac{\ell_{\sigma}^{\tau}(r,k)}{2i\sinh (\pi k/\kappa-2i\varphi)},\ \ \ \tau,\sigma=+,-,
\end{equation}
\begin{equation}\label{ellnum}
\ell_{\sigma}^-(r,k)\equiv 2i\sigma\sinh(2i\varphi)\exp(-\pi r/\rho),\ \ \ \sigma=+,-,
\end{equation}
\begin{equation}\label{ellnup}
\ell_{\sigma}^+(r,k)\equiv 2i\big(\exp(-\pi r/\rho)\sinh(\pi k/\kappa-2i\varphi) -
\exp(\pi r/\rho) \sinh(\pi k/\kappa)\big),\ \ \sigma=+,-,
\end{equation}
and denote the associated special transforms by ${\cal F}_{\sigma}(\varphi)$, $\sigma=+,-$. Note first that all of these functions are $i\pi$-periodic in $\varphi$, so we may and will restrict $\varphi$ to the period interval $(0,\pi]$. For the special choices $\varphi=\pi$ and $\varphi=\pi/2$, the weight function $w(r)$ does not satisfy the simple pole restriction, but one readily verifies
\begin{equation}\label{phiexc}
w(r)^{1/2}\sum_{\tau=+,-} m_{\pm}^{\tau}(r,k)\exp(i\tau rk)=\left\{
\begin{array}{ll}
-\exp(irk), & \varphi =\pi, \\
\exp(irk), & \varphi =\pi/2.
\end{array}\right.
\end{equation}
Thus we have ${\cal F}_{\pm}(\pi)=-{\cal F}_0$ and ${\cal F}_{\pm}(\pi/2)={\cal F}_0$, cf.~\eqref{free}.
Choosing next
\begin{equation}\label{phiint}
\varphi \in (0,\pi/2)\cup (\pi/2,\pi),
\end{equation}
it is routine to check that all assumptions are satisfied, with $L=1$, $\gamma=2\pi/\rho$, and
\begin{equation}\label{tnu}
T_{\pm}(k)=\frac{\sinh(\pi k/\kappa)}{\sinh(2i\varphi-\pi k/\kappa )},
\end{equation}
\begin{equation}\label{rnu}
R_{\pm}(k)=\frac{\pm\sinh(2i\varphi )}{\sinh(2i\varphi-\pi k/\kappa )}.
\end{equation}
We are now prepared for the following theorem.
\begin{theorem}
Letting $f,g\in\hat{{\cal C}}$, we have
\begin{multline}\label{cFres}
({\cal F} f,{\cal F} g)_1= (f,g)_2+ i\sum_{j=1}^{L}w_j
\sum_{\delta,\delta'=+,-}\delta\de'\int_{0}^{\infty} dk\, \overline{f_{\delta}(k)}\int_{0}^{\infty} dk'\, g_{\delta'}(k')
\\
\times \sum_{\nu,\nu'=+,-}\frac{\exp(i r_j(\nu k-\nu' k'))}{1-\exp(\rho (\nu' k'-\nu k))} M_{\delta \delta'}^{\delta\de'\nu\nu'}(\nu r_j,k,k').
\end{multline}
\end{theorem}
\begin{proof}
To start with, we stress that the~$4L$ residue integrals on the rhs are absolutely convergent due to our evenness assumptions~\eqref{Mass}. Indeed, fixing $j,\delta,\delta'$, the denominator has zeros on the integration region only when~$k=k'$ and $\nu =\nu'$. Now the two terms in the sum with $\nu'=\nu$ involve $M_{+}^{+}(\nu r_j,k,k')$ for the two cases $\delta=\delta'$, and $M_{-}^{-}(\nu r_j,k,k')$ for $\delta=-\delta'$. Hence \eqref{Mass} ensures the cancellation of the poles arising for $k=k'$.
Proceeding with the proof, we use Fubini's theorem to write
\begin{equation}\label{Lamlim}
({\cal F} f,{\cal F} g)_1= \lim_{\Lambda \to\infty}\sum_{\delta,\delta'=+,-}\delta\de'\int_{0}^{\infty} dk\, \overline{f_{\delta}(k)}\int_{0}^{\infty} dk'\, g_{\delta'}(k')I_{\delta \delta'}(\Lambda ,k,k'),
\end{equation}
with
\begin{equation}
I_{\sigma}(\Lambda ,k,k')\equiv \frac{1}{2\pi}\int_{-\Lambda }^\Lambda dr\, \overline{\Psi( r,k)}\Psi(\sigma r,k'),\ \ \ \sigma=+,-,
\end{equation}
cf.~\eqref{cFdef}--\eqref{Fm}. Using~\eqref{Psiconj} and~\eqref{Psiform}, we obtain
\begin{equation}\label{Idd}
I_{\sigma}(\Lambda ,k,k')=\frac{1}{2\pi}\sum_{\tau,\tau'=+,-} \int_{-\Lambda }^\Lambda dr\, w(r)J^{\tau,\tau'}_{\sigma}(r,k,k'),
\end{equation}
where
\begin{equation}\label{Jdef}
J^{\tau,\tau'}_{\sigma}(r,k,k')\equiv m^{\tau} ( r,-k)m^{\tau'} (\sigma r,k')\exp(ir(\sigma\tau'k'-\tau k)).
\end{equation}
Fixing $\tau$ and $\tau'$, the integrand in~\eqref{Idd} picks up an $r$-independent multiplier when~$r$ is shifted by $i\rho$. Indeed, $w(r)$ is $i\rho$-periodic, while
\begin{equation}
J^{\tau,\tau'}_{\sigma}(r+i\rho,k,k')=\exp(\rho(\tau k-\sigma\tau' k'))J^{\tau,\tau'}_{\sigma}(r,k,k').
\end{equation}
To exploit this, we use Cauchy's theorem to get
\begin{multline}\label{Jint}
\big(1-\exp(\rho(\tau k-\sigma\tau' k'))\big) \int_{-\Lambda }^\Lambda dr\, w(r)J^{\tau,\tau'}_{\sigma}(r,k,k')
\\
=2\pi i\sum_{j=1}^{2L} w_j J^{\tau,\tau'}_{\sigma}(r_j,k,k')+B^{\tau,\tau'}_{\sigma}(\Lambda ,k,k'),
\end{multline}
where
\begin{equation}\label{BttLam}
B^{\tau,\tau'}_{\sigma}(\Lambda ,k,k')\equiv
-\left( \int_\Lambda ^{\Lambda +i\rho}+\int_{-\Lambda +i\rho}^{-\Lambda } \right) dr\, w(r)J^{\tau,\tau'}_{\sigma}(r,k,k').
\end{equation}
Here, we have chosen $\Lambda $ sufficiently large so that the rectangular contour with corners~$-\Lambda,\Lambda,\Lambda+i\rho,-\Lambda+i\rho$, encloses all of the $w$-poles $r_1,\ldots,r_{2L}$.
Using \eqref{poles}--\eqref{resid}, we can write the residue sum as
\begin{multline}\label{ressum}
\sum_{j=1}^{2L} w_j J^{\tau,\tau'}_{\sigma}(r_j,k,k')
= \sum_{j=1}^{L} w_j \Big(
\exp(i r_j(\sigma\tau' k'-\tau k))
m^{\tau}( r_j,-k)m^{\tau'}(\sigma r_j,k')
\\
-\exp(\rho(\tau k-\sigma\tau' k'))\exp(-i r_j(\sigma\tau' k'-\tau k))
m^{\tau}(- r_j,-k)m^{\tau'}(-\sigma r_j,k')\Big).
\end{multline}
As a consequence, we obtain
\begin{multline}
\sum_{\tau,\tau'=+,-}\sum_{j=1}^{2L}w_j J^{\tau,\tau'}_{\sigma}(r_j,k,k')\big/[1-\exp(\rho(\tau k-\sigma\tau' k'))]
\\
=\sum_{j=1}^L\sum_{\nu,\nu'=+,-}\frac{w_j}{1-\exp(\rho(\nu k-\nu' k'))}\Big[ \exp(ir_j(\nu' k'-\nu k))m^{\nu}( r_j,-k)m^{\sigma\nu'}(\sigma r_j,k')
\\
-\exp(\rho(\nu k-\nu' k'))\exp(-ir_j(\nu' k'-\nu k))m^{\nu}(- r_j,-k)m^{\sigma\nu'}(-\sigma r_j,k')\Big],
\end{multline}
where we have changed variables $\tau,\tau'\to \nu,\sigma\nu'$. When we now take $\nu,\nu'\to -\nu,-\nu'$ in the first sum, we get
\begin{multline}\label{crit1}
\sum_{j=1}^Lw_j\sum_{\nu,\nu'=+,-}\frac{\exp(ir_j(\nu k-\nu' k'))}{1-\exp(\rho(\nu' k'-\nu k))}
\\
\times\big[ m^{-\nu}( r_j,-k)m^{-\sigma\nu'}(\sigma r_j,k')+m^{\nu}(- r_j,-k)m^{\sigma\nu'}(-\sigma r_j,k')\big].
\end{multline}
Combining this with~\eqref{Lamlim}--\eqref{Jint}, the $4L$ residue integrals on the rhs of~\eqref{cFres} result upon rewriting the second line of~\eqref{crit1} as $M_{\sigma}^{\sigma\nu\nu'}(\nu r_j,k,k')$, cf.~\eqref{Mtsdef}. (The latter equality is not immediate, but it can be readily verified case by case.)
We proceed to rewrite~$B^{\tau,\tau'}_{\sigma}(\Lambda ,k,k')$. Taking $r\to -r+i\rho$ in the second integral, we use evenness and $i\rho$-periodicity of~$w(r)$ to obtain
\begin{equation}
B^{\tau,\tau'}_{\sigma}(\Lambda ,k,k')=\int_\Lambda ^{\Lambda +i\rho} dr\, w(r)\big( J^{\tau,\tau'}_{\sigma}(-r+i\rho,k,k')
-J^{\tau,\tau'}_{\sigma}(r,k,k') \big).
\end{equation}
Shifting $r$ over $i\rho/2$ and using~\eqref{Jdef}, this entails
\begin{multline}\label{boundint}
\frac{B^{\tau,\tau'}_{\sigma}(\Lambda ,k,k')}{1-\exp(\rho(\tau k-\sigma\tau' k'))}=\frac{1}{2\sinh(\rho(\tau k-\sigma\tau' k')/2)}\int_{\Lambda -i\rho/2}^{\Lambda +i\rho/2} dr\, w(r+i\rho/2)
\\
\times\sum_{\epsilon=+,-}\epsilon \exp(i\epsilon r(\sigma\tau' k'-\tau k))m^{\tau}(\epsilon(r+i\rho/2),-k)m^{\tau'}(\sigma\epsilon(r+i\rho/2),k').
\end{multline}
We now study the $\Lambda \to\infty$ limit of
\begin{equation}\label{start}
\delta\de'\int_{0}^{\infty} dk\, \overline{f_{\delta}(k)}\int_{0}^{\infty} dk'\, g_{\delta'}(k')\sum_{\tau,\tau'=+,-} \frac{B^{\tau,\tau'}_{\delta\de'}(\Lambda ,k,k')}{1-\exp(\rho(\tau k-\delta\de'\tau' k'))},
\end{equation}
for the four cases~$\delta,\delta'=+,-$.
First let $\delta=\delta'$. Then for the terms in~\eqref{start} with $\tau=-\tau'$, the denominator does not vanish on the integration region. Therefore, using~\eqref{boundint} and the asymptotics assumptions~\eqref{mpas}--\eqref{mnas}, it readily follows that they vanish for $\Lambda \to\infty$. (For the dominant asymptotics the $r$-integration is elementary, and one need only invoke the Riemann-Lebesgue lemma and dominated convergence, while the subdominant terms vanish by dominated convergence.)
Consider next the terms with $\tau=\tau'$. From~\eqref{boundint} and~\eqref{Mtsdef} we infer
\begin{multline}\label{firstcase}
\sum_{\tau=+,-}\frac{B^{\tau,\tau}_{+}(\Lambda ,k,k')}{1-\exp(\rho\tau (k-k'))}
=\frac{1}{2\sinh(\rho( k'- k)/2)}\int_{\Lambda -i\rho/2}^{\Lambda +i\rho/2} dr\, w(r+i\rho/2)
\\
\times\sum_{\alpha=+,-}\alpha\exp(i\alpha r(k-k'))M_{+}^{+}(\alpha(r+i\rho/2),k,k').
\end{multline}
Now we change variables
\begin{equation}\label{stt}
t=\rho k/2,\ \ t'=\rho k'/2,\ \ s=2r/\rho,
\end{equation}
and set
\begin{equation}
G_{\alpha}(s,t,t')\equiv w(\rho(s+i)/2))M_{+}^{+}(\alpha \rho (s+i)/2, 2t/\rho,2t'/\rho),\ \ \ \alpha=+,-.
\end{equation}
This implies that the rhs of \eqref{firstcase} can be rewritten as
\begin{equation}
\frac{\rho}{4\sinh(t'-t)}\int_{2\Lambda /\rho -i}^{2\Lambda /\rho +i}ds\, \sum_{\alpha=+,-}\alpha\exp(i\alpha s(t-t'))G_{\alpha}(s,t,t').
\end{equation}
In view of our evenness assumptions~\eqref{Mass}, the functions $G_{\pm}(s,t,t')$ satisfy~\eqref{Geq}. Moreover, \eqref{M1as} implies that $G_{\pm}(s,t,t')$ also satisfy~\eqref{Gpmas}--\eqref{rhojpart}, with~$\eta=\gamma\rho/2$, and
\begin{equation}
A_1(t,t')=1,\ \ \ \ \ A_2(t,t')=T(-2t/\rho)T(2t'/\rho)+R(-2t/\rho)R(2t'/\rho).
\end{equation}
Finally, our unitarity assumption \eqref{unit1} entails that \eqref{Aeq} is obeyed, with~$A(t)=1$.
Putting
\begin{equation}
\phi(t,t')\equiv \frac{1}{2\pi\rho}\,\overline{f_{\delta}(2t/\rho)}g_{\delta}(2t'/\rho),
\end{equation}
it follows that all assumptions of Lemma~C.1 are fulfilled.
As a result, we deduce
\begin{equation}\label{deplim}
\lim_{\Lambda \to\infty}\int_{(0,\infty)^2} dk\, dk'\,\frac{\overline{f_{\delta}(k)} g_{\delta}(k')}{2\pi}\sum_{\tau,\tau'=+,-} \frac{B^{\tau,\tau'}_{+}(\Lambda ,k,k')}{1-\exp(\rho(\tau k-\tau'k'))}=\int_{0}^{\infty} dk\, \overline{f_{\delta}(k)}g_{\delta}(k).
\end{equation}
It remains to show that for the two remaining cases $\delta'=-\delta$ the pertinent limits vanish.
As before, this is easily checked when $\tau'=\tau$.
Consider now the case $\tau'=-\tau$. As the analog of~\eqref{firstcase}, we then get
\begin{multline}\label{thirdcase}
\sum_{\tau=+,-}\frac{B^{\tau,-\tau}_{-}(\Lambda ,k,k')}{1-\exp(\rho\tau (k-k'))}
=\frac{1}{2\sinh(\rho( k'- k)/2)}\int_{\Lambda -i\rho/2}^{\Lambda +i\rho/2} dr\, w(r+i\rho/2)
\\
\times\sum_{\alpha=+,-}\alpha\exp(i\alpha r(k-k'))M_{-}^{-}(\alpha(r+i\rho/2),k,k').
\end{multline}
Using again the variable change~\eqref{stt}, but now putting
\begin{equation}
G_{\alpha}(s,t,t')\equiv w(\rho(s+i)/2))M_{-}^{-}(\alpha \rho (s+i)/2, 2t/\rho,2t'/\rho),\ \ \ \alpha=+,-,
\end{equation}
\begin{equation}
\phi(t,t')\equiv \frac{1}{2\pi\rho}\,\overline{f_{\delta}(2t/\rho)}g_{-\delta}(2t'/\rho),
\end{equation}
it follows once more that all assumptions of Lemma~C.1 are fulfilled, now with
\begin{equation}
A_1(t,t')=0,\ \ \ A_2(t,t')=-T(-2t/\rho)R (2t'/\rho)-R(-2t/\rho)T(2t'/\rho),
\end{equation}
and $A(t)=0$ (cf.~\eqref{Mass}, \eqref{M2as}, and \eqref{unit2}).
Thus the cases $\delta'=-\delta$ yield limit zero, so that the proof is complete.
\end{proof}
To apply the theorem in the main text, it is expedient to rewrite the residue sums in an alternative form, which is detailed in the following corollary.
\begin{corollary}
Setting
\begin{multline}\label{Rdk}
\hat{R}_{\delta,\delta'}(k,k')\equiv i\delta\de'\sum_{j=1}^{L} w_j
\sum_{\nu,\nu'=+,-}\frac{ 1}{1-\exp(\rho (\nu' k'-\nu k))}
\\
\times
\big[ \mu^{-\delta\nu}(\delta r_j,-k)\mu^{-\delta'\nu'}(\delta' r_j,k')+\mu^{\delta\nu}(-\delta r_j,-k)\mu^{\delta'\nu'}(-\delta' r_j,k')\big],
\end{multline}
where
\begin{equation}\label{mudef}
\mu^{\tau}(r,k)\equiv \exp(i\tau rk)m^{\tau}(r,k),\ \ \ \tau=+,-,
\end{equation}
the transform ${\cal F}$ is isometric if and only if the residue sums $\hat{R}_{\delta,\delta'}(k,k')$, $\delta,\delta'=+,-$, vanish.
\end{corollary}
\begin{proof}
This readily follows from~\eqref{cFres} by invoking~\eqref{crit1} with $\sigma=\delta\de'$.
\end{proof}
For the special cases~\eqref{wsp}--\eqref{ellnup}, we have
\begin{equation}
r_1=i\rho\varphi/\pi\in i(0,\rho),\ \ \ w_1=\frac{\rho}{4\pi\sinh(2i\varphi)},\ \ \ \varphi \in (0,\pi/2)\cup (\pi/2,\pi),
\end{equation}
\begin{equation}
\ell_{\sigma}^-(\pm r_1,k)=2i\sigma\sinh(2i\varphi)\exp(\mp i\varphi),
\end{equation}
\begin{equation}
\ell_{\sigma}^+(\pm r_1,k)=-2i\sinh(2i\varphi)\exp(\pm (\pi k/\kappa-i\varphi)).
\end{equation}
From this we obtain the following ratios:
\begin{multline}\label{ratios}
M_{\alpha}^{\alpha}(-r_1,k,k')/M_{\alpha}^{\alpha}(r_1,k,k')=e^{\pi(k'-k)/\kappa},\ \
M_{\alpha}^{-\alpha}(r_1,k,k')/M_{\alpha}^{\alpha}(r_1,k,k')=-\sigma e^{\pi k'/\kappa},
\\
M_{\alpha}^{-\alpha}(-r_1,k,k')/M_{\alpha}^{\alpha}(r_1,k,k')=-\sigma e^{-\pi k/\kappa},\ \ \ \alpha=+,-.
\end{multline}
Hence the residue sum in~\eqref{cFres} is proportional to
\begin{multline}
\frac{1}{1-\exp(\rho(k'-k))}+\frac{\exp((2ir_1+\pi/\kappa)(k'-k))}{1-\exp(\rho(-k'+k))}
\\
-\sigma\frac{\exp((2ir_1+\pi/\kappa)k')}{1-\exp(\rho(-k'-k))}
-\sigma\frac{\exp(-(2ir_1+\pi/\kappa)k)}{1-\exp(\rho(k'+k))}.
\end{multline}
For this to vanish we can choose either
\begin{equation}\label{phi1}
\sigma=+,\ \ \ 2ir_1+\pi/\kappa=0,\ \ \ \rho\kappa>\pi/2,
\end{equation}
or
\begin{equation}\label{phi2}
\sigma=+,-,\ \ \ 2ir_1+\pi/\kappa=-\rho,\ \ \ \rho\kappa>\pi,
\end{equation}
with the inequalities due to the requirement that~$r_1$ belong to $i(0,\rho)$.
Indeed, for these choices the asserted vanishing comes down to the identities
\begin{equation}\label{id1}
\frac{1}{1-A'/A}+\frac{1}{1-A/A'}-\frac{1}{1-1/A'A}-\frac{1}{1-A'A}=0,
\end{equation}
\begin{equation}\label{id2}
\frac{1}{1-A'/A}+\frac{A}{A'}\,\frac{1}{1-A/A'} -\frac{1}{A'}\,\frac{\sigma}{1-1/A'A} -A\,\frac{\sigma}{1-A'A}=0,\ \ \sigma=+,-,
\end{equation}
which are easily checked.
Thus, setting
\begin{equation}\label{rk0}
\phi_0\equiv\frac{\pi^2}{2\rho\kappa},\ \ \ \rho\kappa>\pi/2,
\end{equation}
\begin{equation}\label{rke}
\phi_e\equiv\frac{\pi^2}{2\rho\kappa}+\frac{\pi}{2},\ \ \ \rho\kappa>\pi,
\end{equation}
we deduce that the three transforms ${\cal F}_+(\phi_0)$, ${\cal F}_+(\phi_e)$ and ${\cal F}_-(\phi_e)$ are isometric.
In case the restrictions on $\rho\kappa$ are not satisfied, we need to subtract a suitable multiple of $i\rho$ from $i\rho\phi_j/\pi$, $j=0,e$, to obtain a pole location $r_1$ with ${\rm Im}\, r\in (0,\rho)$. Generically, this yields a nonvanishing residue sum, hence isometry breakdown.
The transform ${\cal F}_+(\phi_0)$ amounts to the $N=0$ transform in the main text. The `extra' transforms~${\cal F}_{\pm}(\phi_e)$ go to show that our assumptions allow realizations beyond the main text. In Section~5 we shall elaborate on the issue of isometry violation. In particular, Proposition~5.1 encodes the salient features of~${\cal F}_+(\phi_0)$ for $\rho\kappa\le\pi/2$.
\section{The transform ${\cal F}^*$ }\label{AppB}
In this appendix we retain the assumptions on ${\cal F}$ made in Appendix~A. As we have pointed out, they imply in particular that ${\cal F}$ maps~$\hat{{\cal C}}$ into~${\cal H}$. However, they do not imply that~${\cal F}$ is a bounded operator. (To be sure, boundedness is plain when the residue sum vanishes.) In fact, it might not even follow from the assumptions made thus far that~${\cal F}$ has an adjoint that is densely defined.
From the additional assumptions in this appendix it shall follow that~${\cal F}^*$ is indeed densely defined. In particular, we shall see that we have
\begin{equation}\label{cTst}
{\cal F}^*\, :\, {\cal C}\equiv C_0^\infty({\mathbb R})\subset {\cal H} \to \hat{{\cal H}}.
\end{equation}
Of course, it follows without further assumptions that for~$h\in {\cal C}$ we obtain
\begin{equation}
(h,{\cal F} f)_1=\frac{1}{\sqrt{2\pi}}\sum_{\delta=+,-}\int_{-\infty}^{\infty} dr\, \overline{h(r)}
\int_{0}^{\infty} dk \, F_{\delta}(r,k)f_{\delta}(k),\ \ \ f\in \hat{{\cal C}},
\end{equation}
cf.~\eqref{cFdef}. From this and~\eqref{Fp}--\eqref{Psiconj} we easily deduce
\begin{equation}\label{cTstdef}
({\cal F}^*h)_{\delta}(k)=\frac{\delta}{\sqrt{2\pi}} \int_{-\infty}^{\infty} dr\, \Psi(\delta r,-k)h(r),\ \ \ \delta=+,-,\ \ \ k>0,
\end{equation}
whenever both integrals yield functions of~$k$ that are square-integrable on~$(0,\infty)$. But we can only ensure this property by making more assumptions.
These extra assumptions concern the $k$-dependence of $m^{\pm}(r,k)$. They will allow us to study ${\cal F}^*$ in much the same way as ${\cal F}$ itself. We would like to stress, however, that it is at this point that we part company with the nonrelativistic framework, inasmuch as the Jost functions~\eqref{Jost} do not have the periodicity in the spectral variable~$k$ we are about to require.
First, we assume that the coefficients $m^{+}(r,k)$ and $m^{-}(r,k)$ are $i\kappa$-periodic and $i\kappa$-antiperiodic functions of~$k$, resp., with $\kappa>0$. Second, they are of the form
\begin{equation}\label{mtform}
m^{\tau}(r,k)= v(k)\ell^{\tau}(r,k),\ \ \ \tau=+,-.
\end{equation}
Here, $\ell^{\pm}(r,k)$ are assumed to be entire in~$k$, whereas $v(k)$ is meromorphic and either $i\kappa$-periodic or $i\kappa$-antiperiodic. Third, $v(k)$ satisfies
\begin{equation}\label{vass}
\overline{v(k)}=v(-k),\ \ \ k\in{\mathbb R}.
\end{equation}
Note that these assumptions are satisfied for the special cases given by~\eqref{mnu}--\eqref{ellnup}, with $v(k)$ being $i\kappa$-antiperiodic.
Introducing the dual weight function
\begin{equation}\label{hwdef}
\hat{w}(k)\equiv |v(k)|^2=v(-k)v(k),\ \ \ k\in{\mathbb R},
\end{equation}
it follows from the assumptions just made that it extends to a meromorphic $i\kappa$-periodic function. We assume that its
only singularities in the period strip ${\rm Im}\, k\in (0,\kappa)$ are finitely many simple poles. Just as for $w(r)$, it then follows that we can pair off the poles, obtaining $2\hat{L}$ distinct poles $k_1,\ldots,k_{2\hat{L}}$ related by
\begin{equation}\label{whpoles}
k_{j+\hat{L}}=i\kappa -k_j,\ \ \ j=1,\ldots,\hat{L},
\end{equation}
with residues
\begin{equation}\label{whres}
\hat{w}_j\equiv {\rm Res}\,( \hat{w}(k))|_{k=k_j},\ \ \ j=1,\ldots,2\hat{L},
\end{equation}
satisfying
\begin{equation}\label{whresid}
\hat{w}_{j+\hat{L}}=-\hat{w}_j,\ \ \ j=1,\ldots,\hat{L}.
\end{equation}
Clearly, these assumptions are satisfied for~\eqref{mnu}--\eqref{ellnup}, with~$\hat{L}=1$, and
\begin{equation}\label{hw}
\hat{w}(k)\equiv \frac{1}{4\sinh (\pi k/\kappa -2i\varphi)\sinh (\pi k/\kappa +2i\varphi)},
\end{equation}
provided $\varphi$ is restricted by~\eqref{phiint} and not equal to $ \pi/4$ or $3\pi/4$. (The latter cases yield double poles at $k= i\kappa/2$.)
Next, we assume that the $|{\rm Re}\, k|\to \infty$ asymptotics of the coefficients is given by
\begin{equation}\label{mpask}
m^+(r,k)=
\left\{
\begin{array}{ll}
C(r)+O(\exp(-\hat{\gamma} k)), & {\rm Re}\, k\to\infty, \\
\overline{C(r)}+O(\exp(\hat{\gamma} k)), & {\rm Re}\, k\to -\infty,
\end{array}
\right.
\end{equation}
\begin{equation}\label{mmask}
m^-(r,k)= O(\exp(\mp \hat{\gamma} k)), \ \ \ {\rm Re}\, k\to\pm \infty,
\end{equation}
\begin{equation}\label{mexk}
\partial_km^{\tau}(r,k)=O(\exp(\mp \hat{\gamma} k)), \ \ \ \tau=+,-,\ \ \ {\rm Re}\, k\to\pm \infty,
\end{equation}
with $\hat{\gamma}>0$ and the implied constants uniform for ${\rm Im}\, k$ and~$r$ in ${\mathbb R}$-compacts. Moreover, we assume that the function $C(r)$ extends from the real line to an entire $i\rho$-periodic or $i\rho$-antiperiodic function satisfying
\begin{equation}\label{cw}
|C(r)|^2=1/w(r),\ \ \ r\in{\mathbb R}.
\end{equation}
Introducing
\begin{equation}\label{Ur}
U(r)^{1/2}\equiv C(r)w(r)^{1/2},\ \ \ r\in{\mathbb R},
\end{equation}
it follows from the above assumptions that $U(r)$ extends to a meromorphic $i\rho$-periodic function that is a phase for real~$r$, and that we have a reflectionless asymptotics
\begin{equation}\label{Psikas}
\Psi(r,k)=
\left\{
\begin{array}{ll}
U(r)^{1/2}\exp(irk)+O(\exp(-\hat{\gamma}k)), & {\rm Re}\, k\to\infty, \\
U(r)^{-1/2}\exp(irk) +O(\exp(\hat{\gamma} k)), & {\rm Re}\, k\to -\infty,
\end{array}
\right.
\end{equation}
uniformly for ${\rm Im}\, k$ and $r$ in ${\mathbb R}$-compacts.
From this asymptotic behavior it easily follows that ${\cal F}^*$ is defined on~${\cal C}$ and given by~\eqref{cTstdef}, as announced. Actually, it may well follow from our assumptions that ${\cal F}$ must be bounded. We have not tried to show this, however, since we do not need this property for the general analysis undertaken in this appendix. (It is not hard to see that the example transforms ${\cal F}_{\pm}(\varphi)$ and the transforms arising in the main text are bounded, cf. the paragraph containing~\eqref{cFform}.)
Our final assumptions generalize the evenness assumptions~\eqref{Mass}. Introducing
\begin{equation}\label{Lttp}
L^{\tau,\tau'}(k,r,r')\equiv \ell^{\tau}(r,k)\ell^{\tau'}(r',-k)+\ell^{-\tau}(-r,k)\ell^{-\tau'}(-r',-k),\ \ \ \tau,\tau'=+,-,
\end{equation}
it follows from the above that $L^{\tau,\tau}(k,r,r')$ is $i\kappa$-periodic in~$k$, whereas $L^{\tau,-\tau}(k,r,r')$ is $i\kappa$-antiperiodic. Also,
it is plain that we have symmetries
\begin{equation}\label{keven}
L^{\tau,\tau'}(k,r,r')=L^{-\tau,-\tau'}(k,-r,-r'),\ \ \ L^{\tau,\tau'}(k,r,r)=L^{\tau',\tau}(-k,r,r).
\end{equation}
But the evenness properties
\begin{equation}\label{Lass}
L^{\tau,\tau'}(k,r,r')=L^{\tau,\tau'}(-k,-r,-r'),\ \ \ \tau,\tau'=+,-,
\end{equation}
amount to further assumptions on the coefficients, strengthening~\eqref{Mass}. Indeed, for $\tau'=\tau=+$ and $r'=r$, \eqref{Lass} amounts to~\eqref{Mass} with $\alpha=+$, while for $\tau'=-\tau=-$ and $r'=-r$, \eqref{Lass} reduces to~\eqref{Mass} with $\alpha=-$.
It is easy to check that the assumptions on the $|{\rm Re}\, k|\to\infty$ asymptotics are satisfied for the example cases \eqref{mnu}--\eqref{ellnup}, with~$\hat{\gamma}=\pi/\kappa$, and
\begin{equation}
C(r)= -2\exp (i\varphi)\sinh (\pi r/\rho +i\varphi),\ \ U(r)= \exp (2i\varphi)\frac{\sinh (\pi r/\rho +i\varphi)}{\sinh (\pi r/\rho -i\varphi)}.
\end{equation}
With due effort, the assumptions~\eqref{Lass} can also be checked for these special cases.
\begin{theorem}
With the above assumptions in effect, let $f,g\in {\cal C}$. Then we have
\begin{multline}\label{cFstres}
({\cal F}^* f,{\cal F}^* g)_2= (f,g)_1+i\sum_{j=1}^{\hat{L}} \hat{w}_j
\int_{-\infty}^{\infty} dr\, \overline{f(r)}\int_{-\infty}^{\infty} dr'\, g(r')w(r)^{1/2}w(r')^{1/2}
\\
\times \sum_{\tau,\tau'=+,-}\frac{\exp(i k_j(\tau r- \tau'r'))}{1-\tau\tau'\exp(\kappa (\tau' r' - \tau r))}L^{\tau,\tau'}( k_j,r,r').
\end{multline}
\end{theorem}
\begin{proof}
The proof proceeds along the same lines as the proof of Theorem~A.1. We begin by noting that we have an identity
\begin{equation}\label{idextra}
L^{\tau,\tau}(k,r,r)=L^{-\tau,-\tau}(k,r,r),
\end{equation}
as readily follows from~\eqref{Lttp} and~\eqref{Lass}. Thus the residue integrals involving~$\tau'=\tau$ are absolutely convergent. (For~$\tau'=-\tau$ absolute convergence is immediate, since then the denominator has no zeros.)
Next, we use \eqref{cTst} and Fubini's theorem to write
\begin{equation}\label{Lalim}
({\cal F}^* f,{\cal F}^* g)_2= \lim_{\Lambda \to\infty} \int_{-\infty}^{\infty} dr\, \overline{f(r)}\int_{-\infty}^{\infty} dr'\, g(r')I(\Lambda ,r,r'),
\end{equation}
with
\begin{equation}
I (\Lambda ,r,r')\equiv \frac{1}{2\pi}\int_{0}^\Lambda dk\, \big(\Psi(r,k) \Psi( r',-k)+\Psi(-r,k) \Psi(- r',-k)\big),
\end{equation}
cf.~\eqref{Fp}--\eqref{Psiconj}. Using~\eqref{Psiform}, \eqref{mtform}--\eqref{hwdef}, and our evenness assumptions~\eqref{Lass}, we now deduce
\begin{equation}\label{IhJ}
I (\Lambda ,r,r')=\frac{1}{4\pi}w(r)^{1/2}w(r')^{1/2}\sum_{\tau,\tau'=+,-} \int_{-\Lambda }^\Lambda dk\, \hat{w}(k)J^{\tau,\tau'} (k,r,r'),
\end{equation}
where
\begin{equation}\label{Jddef}
J^{\tau,\tau'} (k,r,r')\equiv L^{\tau,\tau'}(k,r,r') \exp(ik (\tau r-\tau' r')).
\end{equation}
Fixing $\tau$ and $\tau'$, the integrand in~\eqref{IhJ} picks up a $k$-independent multiplier when~$k$ is shifted by $i\kappa$. Indeed, this is clear for the plane wave factor, whereas $\hat{w}(k)L^{\tau,\tau'}(k,r,r') $ is $i\kappa$-periodic/$i\kappa$-antiperiodic for $\tau\tau'=+$/$\tau\tau'=-$. Using Cauchy's theorem, we therefore get
\begin{multline}\label{hJint}
\big(1-\tau\tau'\exp(\kappa(\tau' r'-\tau r))\big)\int_{-\Lambda }^\Lambda dk\, \hat{w}(k)J^{\tau,\tau'}(k,r,r')
\\
= 2\pi i\sum_{j=1}^{2\hat{L}} \hat{w}_j J^{\tau,\tau'}(k_j,r,r')+B^{\tau,\tau'}(\Lambda ,r,r'),
\end{multline}
where
\begin{equation}\label{BdttLam}
B^{\tau,\tau'} (\Lambda ,r,r')\equiv
-\left( \int_\Lambda ^{\Lambda +i\kappa}+\int_{-\Lambda +i\kappa}^{-\Lambda } \right) dk\, \hat{w}(k)J^{\tau,\tau'}(k,r,r').
\end{equation}
As before, we choose $\Lambda $ sufficiently large so that the contour encloses all of the $\hat{w}$-poles.
We proceed to rewrite the residue sum as
\begin{multline}\label{ressum2}
\sum_{j=1}^{2\hat{L}} \hat{w}_j J^{\tau,\tau'}(k_j,r,r')
= \sum_{j=1}^{\hat{L}} \hat{w}_j \Big(
\exp(i k_j(\tau r-\tau' r'))
L^{\tau,\tau'}( k_j,r,r')
\\
-\tau\tau'\exp(\kappa(\tau' r'-\tau r))\exp(-i k_j(\tau r-\tau' r'))
L^{\tau,\tau'}(- k_j,r,r')
\Big),
\end{multline}
where we used \eqref{whpoles}--\eqref{whresid}. Hence we deduce
\begin{multline}
\sum_{\tau,\tau'=+,-}\sum_{j=1}^{2\hat{L}}\hat{w}_j J^{\tau,\tau'} (k_j,r,r')\big/[1-\tau\tau'\exp(\kappa(\tau' r'-\tau r))]
\\
=\sum_{j=1}^{\hat{L}}\sum_{\tau,\tau'=+,-}\frac{ \hat{w}_j}{1-\tau\tau'\exp(\kappa(\tau' r'- \tau r))}
\big[\exp(ik_j(\tau r-\tau' r'))L^{\tau,\tau'}( k_j,r,r')
\\
-\tau\tau'\exp(\kappa(\tau' r'- \tau r))\exp(-ik_j(\tau r-\tau' r'))L^{-\tau,-\tau'}( k_j,r,r')\big],
\end{multline}
where we have used an identity equivalent to~\eqref{Lass}, viz.,
\begin{equation}\label{Lasseq}
L^{\tau,\tau'}(-k,r,r')=L^{-\tau,-\tau'}(k,r,r').
\end{equation}
In the second sum we now take $\tau,\tau'\to -\tau,-\tau'$, thus arriving at
\begin{equation}\label{crit2}
2\sum_{j=1}^{\hat{L}} \sum_{\tau,\tau'=+,-}\frac{ \hat{w}_j }{1-\tau\tau'\exp(\kappa(\tau' r'- \tau r))}
\exp(ik_j(\tau r-\tau' r'))L^{\tau,\tau'}( k_j,r,r').
\end{equation}
Combining this with~\eqref{Lalim}--\eqref{hJint}, we readily obtain the residue integrals on the rhs of~\eqref{cFstres}.
Next, we change variables to rewrite \eqref{BdttLam} as
\begin{equation}
B^{\tau,\tau'} (\Lambda ,r,r')=
\int_{\Lambda-i\kappa/2} ^{\Lambda +i\kappa/2} dk\, \hat{w}(k+i\kappa/2)\big(J^{\tau,\tau'}(-k+i\kappa/2,r,r')-J^{\tau,\tau'}(k+i\kappa/2,r,r')\big).
\end{equation}
From this we infer, using \eqref{Jddef},
\begin{multline}\label{bndint}
\frac{B^{\tau,\tau'} (\Lambda ,r,r')}{1-\tau\tau'\exp(\kappa(\tau' r'-\tau r))} =\frac{\exp(\kappa(\tau' r'-\tau r)/2)}{1-\tau\tau'\exp(\kappa(\tau' r'-\tau r))} \int_{\Lambda -i\kappa/2}^{\Lambda +i\kappa/2} dk\, \hat{w}(k+i\kappa/2)
\\
\times\sum_{\sigma=+,-}\sigma \exp(i\sigma k(\tau' r'-\tau r))L^{\tau,\tau'}(-\sigma k+i\kappa/2,r,r') .
\end{multline}
We now study the $\Lambda\to\infty$ limit of the four terms
\begin{equation}
T^{\tau,\tau'}(\Lambda)\equiv \frac{1}{4\pi}\int_{-\infty}^{\infty} dr\, \overline{f(r)}\int_{-\infty}^{\infty} dr'\, g(r')w(r)^{1/2}w(r')^{1/2}\frac{B^{\tau,\tau'} (\Lambda ,r,r')}{1-\tau\tau'\exp(\kappa(\tau' r'-\tau r))},
\end{equation}
contributing to the rhs of~\eqref{Lalim}. For the cases $\tau'=-\tau$, it follows from the uniform asymptotics~\eqref{mpask}--\eqref{mmask} that we have
\begin{equation}
\lim_{\Lambda\to\infty}T^{\tau,-\tau}(\Lambda)=0,\ \ \ \tau=+,-.
\end{equation}
(Note that in \eqref{bndint} we may replace $\hat{w}(k+i\kappa/2)$ by $\hat{w}(-\sigma k+i\kappa/2)$, since $\hat{w}$ is even and $i\kappa$-periodic.) For the cases $\tau'=\tau$, the rhs of~\eqref{bndint} can be rewritten as
\begin{multline}\label{brhs}
\frac{1}{2\sinh(\kappa (r'-r)/2)} \int_{\Lambda -i\kappa/2}^{\Lambda +i\kappa/2} dk\, \sum_{\alpha=+,-}\alpha
\exp(i\alpha k( r- r'))
\\
\times\hat{w}(\alpha k+i\kappa/2)L^{+,+}(\alpha k+i\kappa/2,r,r'),
\end{multline}
where we used \eqref{Lasseq} with $\tau'=\tau=-$ and $i\kappa$-periodicity.
We now change variables
\begin{equation}\label{stt2}
t=\kappa r/2,\ \ t'=\kappa r'/2,\ \ s=2k/\kappa,
\end{equation}
and set
\begin{equation}
G_{\alpha}(s,t,t')\equiv \hat{w}(\alpha \kappa (s+i)/2)L^{+,+}(\alpha \kappa (s+i)/2, 2t/\kappa,2t'/\kappa),\ \ \ \alpha=+,-.
\end{equation}
Then \eqref{brhs} can be rewritten as
\begin{equation}
\frac{\kappa}{4\sinh(t'-t)}\int_{2\Lambda /\kappa -i}^{2\Lambda /\kappa +i}ds\, \sum_{\alpha=+,-}\alpha\exp(i\alpha s(t-t'))G_{\alpha}(s,t,t').
\end{equation}
Thanks to~\eqref{keven}, the functions $G_{\pm}(s,t,t')$ satisfy~\eqref{Geq}. Moreover, our assumptions \eqref{mpask}--\eqref{mexk} about the ${\rm Re}\, k\to\pm \infty$ asymptotics of the coefficients $m^{\tau}(r,k)$ ensure that $G_{\pm}(s,t,t')$ also satisfy~\eqref{Gpmas}--\eqref{rhojpart}, with $\eta=\hat{\gamma}\kappa/2$, and
\begin{equation}
A_1(t,t')=C(2t/\kappa)\overline{C(2t'/\kappa)},\ \ \ \ \ A_2(t,t')= \overline{A_1(t,t')}.
\end{equation}
Putting
\begin{equation}
\phi(t,t')\equiv \frac{1}{4\pi\kappa}\,\overline{f(2t/\kappa)}g(2t'/\kappa)w(2t/\kappa)^{1/2}w(2t'/\kappa)^{1/2},
\end{equation}
it follows that the assumptions of Lemma~C.1 are obeyed, with
\begin{equation}
A(t)=|C(2t/\kappa)|^2.
\end{equation}
In view of~\eqref{cw}, we therefore get
\begin{equation}
\lim_{\Lambda\to\infty}T^{\tau,\tau}(\Lambda)=(f,g)_1/2,\ \ \ \tau=+,-,
\end{equation}
and hence the theorem follows.
\end{proof}
Just as for the residue sum in Theorem~A.1, in the main text it is more expedient to use an alternative form for the residue sum in Theorem~B.1, as specified by the following corollary.
\begin{corollary}
Setting
\begin{equation}\label{defR}
R(r,r')\equiv iw(r)^{1/2}w(r')^{1/2}\sum_{j=1}^{\hat{L}} \hat{w}_j
\sum_{\tau,\tau'=+,-}\frac{ \Lambda^{\tau,\tau'}(k_j,r,r')}{1-\tau\tau'\exp(\kappa (\tau' r'-\tau r))},
\end{equation}
where
\begin{equation}\label{Lamdef}
\Lambda^{\tau,\tau'}(k,r,r')\equiv \lambda^{\tau}(r, k)\lambda^{\tau'}(r',- k)+\lambda^{-\tau}(- r,k)\lambda^{-\tau'}(-r',- k),
\end{equation}
\begin{equation}\label{lamdef}
\lambda^{\tau}(r,k)\equiv \exp(i\tau rk)\ell^{\tau}(r,k),
\end{equation}
the operator ${\cal F}^*$ is isometric if and only if the residue sum~$R(r,r')$ vanishes.
\end{corollary}
\begin{proof}
This easily follows from~\eqref{cFstres} and~\eqref{Lttp}.
\end{proof}
For the examples~\eqref{wsp}--\eqref{ellnup}, we should distinguish two cases for the simple pole locations~$k_1$ and $i\kappa-ik_1$ in the period strip ${\rm Im}\, k\in(0,\kappa)$, viz.,
\begin{equation}\label{kap1}
k_1=2i\kappa\varphi/\pi,\ \ \ \varphi\in(0,\pi/2), \ \ \ \varphi\ne \pi/4,
\end{equation}
and
\begin{equation}\label{kap2}
k_1=2i\kappa\varphi/\pi -i\kappa,\ \ \ \varphi\in(\pi/2,\pi),\ \ \ \varphi\ne 3\pi/4.
\end{equation}
Also, we obtain
\begin{equation}
\ell_{\sigma}^-(r,\pm k_1)=2i\sigma\sinh(2i\varphi)\exp(-\pi r/\rho),
\end{equation}
\begin{equation}
\ell_{\sigma}^+(r, k_1)=\mp 2i\sinh(2i\varphi)\exp(\pi r/\rho),
\end{equation}
\begin{equation}
\ell_{\sigma}^+(r, -k_1)=\mp 2i\sinh(2i\varphi)\big(2\cosh(2i\varphi)\exp(-\pi r/\rho)-\exp(\pi r/\rho)\big),
\end{equation}
so that \eqref{Lttp} yields
\begin{equation}\label{Lp}
L^{\tau,\tau}( k_1,r,r')=-8\sinh(2i\varphi)^2\cosh(2i\varphi)\exp(\tau \pi (r-r')/\rho),\ \ \tau=+,-,
\end{equation}
\begin{equation}\label{Lm}
L^{\tau,-\tau}( k_1,r,r')=\pm 8\sigma\sinh(2i\varphi)^2\cosh(2i\varphi)\exp(\tau \pi (r+r')/\rho),\ \ \sigma,\tau=+,-,
\end{equation}
with the upper/lower sign referring to~\eqref{kap1}/\eqref{kap2}. From this we calculate ratios
\begin{equation}
L^{-,-}/L^{+,+} =e^{2\pi(r'-r)/\rho},\ \
L^{+,-}/L^{+,+} =\mp \sigma e^{2\pi r'/\rho},
\ \
L^{-,+} /L^{+,+} =\mp \sigma e^{-2\pi r/\rho}.
\end{equation}
With~\eqref{kap1} in force, the residue sum in~\eqref{cFstres} is therefore proportional to
\begin{multline}
\frac{1}{1-\exp(\kappa(r'-r))}+\frac{\exp((2ik_1+2\pi/\rho)(r'-r))}{1-\exp(\kappa(-r'+r))}
\\
-\sigma\frac{\exp((2ik_1+2\pi/\rho)r')}{1+\exp(\kappa(-r'-r))}
-\sigma\frac{\exp(-(2ik_1+2\pi/\rho)r)}{1+\exp(\kappa(r'+r))}.
\end{multline}
For this to vanish we can choose
\begin{equation}
\sigma =+,\ \ \ ik_1+\pi/\rho=0\Leftrightarrow \varphi =\pi^2/2\rho\kappa=\phi_0,
\end{equation}
cf.~\eqref{rk0}. (Indeed, vanishing boils down to~\eqref{id1} with $A,A'\to iA, iA'$.) As a consequence, the transform ${\cal F}_+(\phi_0)$ is unitary for $\phi_0=\pi^2/2\rho\kappa\in(0,\pi/2)$, $\phi_0\ne \pi/4$. Using a continuity argument, it follows that ${\cal F}_+(\phi_0)$ is unitary for the double-pole case $\phi_0=\pi/4$, too.
Next, we consider the case~\eqref{kap2}. Then~\eqref{hw} yields the residue
\begin{equation}\label{hw1}
\hat{w}_1=\frac{\kappa}{4\pi\sinh(4i\varphi)}.
\end{equation}
From~\eqref{Lp}--\eqref{Lm} we now deduce
\begin{equation}
R(r,r')={\cal N}(\varphi)w(r)^{1/2}w(r')^{1/2}S_{\sigma}(\varphi;r,r'),
\end{equation}
where we have introduced
\begin{equation}
{\cal N}(\varphi)\equiv \frac{\kappa\sin(2\varphi)}{\pi}\in (-\infty,0),\ \ \ \varphi\in(\pi/2,\pi),
\end{equation}
and
\begin{equation}\label{Sdef}
S_{\sigma}(\varphi;r,r')\equiv \sum_{\tau=+,-}\Big(\frac{\exp(\tau (ik_1+\pi/\rho)( r- r'))}{1-\exp(\tau\kappa (r' - r))}
+\sigma
\frac{\exp(\tau (ik_1+\pi/\rho)( r+ r'))}{1+\exp(-\tau\kappa (r' + r))} \Big).
\end{equation}
We proceed to study the two $\varphi$-choices $\phi_0$, $\phi_e\in(\pi/2,\pi)$, for which we already know that ${\cal F}_+(\phi_0)$ and ${\cal F}_{\pm}(\phi_e)$ are isometric. (Recall $\rho\kappa\in(\pi/2,\pi)$ in the first case and $\rho\kappa\in (\pi,\infty)$ in the second one, cf.~\eqref{rk0}--\eqref{rke}.) For the case
$\varphi=\phi_0$ we obtain, putting
\begin{equation}\label{AAp}
A\equiv \exp(\kappa r),\ \ \ A'\equiv \exp(\kappa r'),
\end{equation}
and setting $ik_1+\pi/\rho=\kappa$ in~\eqref{Sdef},
\begin{multline}\label{S0}
S_+(\phi_0;r,r')=\frac{A}{A'}\,\frac{1}{1-A'/A}+\frac{A'}{A}\,\frac{1}{1-A/A'} +AA'\,\frac{1}{1+1/A'A}+\frac{1}{A'A}\,\frac{1}{1+A'A}
\\
=\frac{A'}{A}+\frac{A}{A'}+A'A+\frac{1}{A'A}=4\cosh(\kappa r)\cosh(\kappa r').
\end{multline}
Introducing the function (cf.~\eqref{wsp})
\begin{equation}\label{Psi0}
\Psi_0(r)\equiv 2\cosh(\kappa r) w_0(r)^{1/2}, \ \ \ w_0(r)\equiv 1\big/4\sinh \Big(\frac{\pi}{\rho}\big( r+i\frac{\pi}{2\kappa}\big)\Big)\sinh \Big(\frac{\pi}{\rho}\big( r-i\frac{\pi}{2\kappa}\big)\Big),
\end{equation}
we get
\begin{equation}
\Psi_0(\cdot)\in {\cal H}, \ \ \rho\kappa \in(\pi/2,\pi).
\end{equation}
Thus we can rewrite~\eqref{cFstres} as
\begin{equation}\label{F0st}
({\cal F}_+(\phi_0)^* f,{\cal F}_+(\phi_0)^* g)_2= (f,g)_1+\frac{\kappa \sin (\pi^2/\rho\kappa)}{\pi}(f,\Psi_0)_1(\Psi_0,g)_1.
\end{equation}
Now ${\cal F}_+(\phi_0)$ is isometric, so ${\cal F}_+(\phi_0)^*$ has a continuous extension from~${\cal C}$ to a partial isometry such that
\begin{equation}
{\cal F}_+(\phi_0)^*{\cal F}_+(\phi_0)={\bf 1}_{\hat{{\cal H}} },\ \ \ {\cal F}_+(\phi_0){\cal F}_+(\phi_0)^*={\bf 1}_{ {\cal H} }-P,
\end{equation}
where $P$ is the projection on the orthocomplement of ${\cal F}_+(\phi_0)(\hat{{\cal H}} )$.
In particular, this implies \eqref{F0st} holds true for all $f,g\in{\cal H}$, which yields
\begin{equation}
{\cal F}_+(\phi_0){\cal F}_+(\phi_0)^*={\bf 1}_{ {\cal H} }+\frac{\kappa \sin (\pi^2/\rho\kappa)}{\pi}\Psi_0\otimes \Psi_0.
\end{equation}
Hence we have~${\cal F}_+(\phi_0)^*\Psi_0=0$ and
\begin{equation}\label{inpr0}
(\Psi_0,\Psi_0)_1=-\frac{\pi}{\kappa \sin (\pi^2/\rho\kappa)},\ \ \rho\kappa \in(\pi/2,\pi).
\end{equation}
By continuity these formulas are also valid for the double-pole case $\phi_0=3\pi/4$.
Let us now summarize our findings regarding~${\cal F}_+(\phi_0)$.
\begin{proposition}
The transform~${\cal F}_+(\phi_0) $ defined by~\eqref{cF}--\eqref{Fm} with
\begin{multline}\label{cFp0}
\Psi(r,k)\equiv \frac{w_0(r)^{1/2}}{\sinh\Big( \frac{\pi}{\kappa}\big( k-\frac{i\pi}{\rho}\big)\Big)}
\Big( e^{irk}\Big[ e^{-\pi r/\rho}\sinh\Big( \frac{\pi}{\kappa}\big( k-\frac{i\pi}{\rho}\big)\Big) -
e^{\pi r/\rho}\sinh\big( \frac{\pi}{\kappa} k \big)\Big]
\\
+e^{-irk}e^{-\pi r/\rho}\sinh\big( \frac{i\pi^2}{\rho\kappa} \big)\Big),\ \ \ \rho\kappa\in(\pi/2,\infty),
\end{multline}
is unitary for $\rho\kappa\in [\pi,\infty)$. For $\rho\kappa\in (\pi/2,\pi)$ it is isometric and satisfies
\begin{equation}
{\cal F}_+(\phi_0){\cal F}_+(\phi_0)^*={\bf 1}_{ {\cal H} }-\Psi_0\otimes \Psi_0/(\Psi_0,\Psi_0)_1.
\end{equation}
Here, $w_0(r)$ and $\Psi_0(r)$ are defined by~\eqref{Psi0} and the inner product is given by~\eqref{inpr0}.
\end{proposition}
As mentioned below~\eqref{psizero}, the function~$\Psi(r,k)$ in this proposition is a reparametrized version of the $N=0$ function $\psi(a_+;x,y)$ of the main text. We also point out that from~\eqref{cFp0} one can read off that for $\rho\kappa\in[\pi,\infty)$ the function $\Psi(r,k)$ is analytic in the strip ${\rm Im}\, k \in [0,\kappa)$. By contrast, for $\rho\kappa\in(\pi/2,\pi)$ it has a pole in this strip at~$k=i\pi/\rho-i\kappa$, and the bound state $\Psi_0(r)$ is proportional to the residue at this pole. Finally, we repeat that Proposition~5.1 yields a complete picture of the transform~${\cal F}_+(\phi_0) $ for~$\rho\kappa\le \pi/2$.
Next, we study the choice $\varphi=\phi_e$. Then we need to set $ik_1+\pi/\rho=0$ in~\eqref{Sdef}, yielding
\begin{equation}
S_{\sigma}(\phi_e;r,r')= \frac{1}{1-A'/A}+ \frac{1}{1-A/A'} +\frac{\sigma}{1+1/A'A}+ \frac{\sigma}{1+A'A}.
\end{equation}
For $\sigma=-$ this vanishes, implying that the transform~${\cal F}_-(\phi_e)$ is unitary.
By contrast, we have
\begin{equation}
S_{+}(\phi_e;r,r')=2.
\end{equation}
Defining (cf.~\eqref{wsp})
\begin{equation}\label{we}
w_e(r)\equiv 1\Big/4\cosh \Big(\frac{\pi}{\rho}\big( r+i\frac{\pi}{2\kappa}\big)\Big)\cosh \Big(\frac{\pi}{\rho}\big( r-i\frac{\pi}{2\kappa}\big)\Big),
\end{equation}
we get
\begin{equation}\label{Psie}
\Psi_e(r)\equiv w_e(r)^{1/2} \in {\cal H}, \ \ \rho\kappa \in(\pi,\infty).
\end{equation}
Thus we can rewrite~\eqref{cFstres} as
\begin{equation}
({\cal F}_+(\phi_e)^* f,{\cal F}_+(\phi_e)^* g)_2= (f,g)_1-\frac{2\kappa \sin (\pi^2/\rho\kappa)}{\pi}(f,\Psi_e)_1(\Psi_e,g)_1.
\end{equation}
As before, we deduce from this that we have ${\cal F}_+(\phi_e)^*\Psi_e=0$ and
\begin{equation}\label{inpre}
(\Psi_e,\Psi_e)_1=\frac{\pi}{2\kappa \sin (\pi^2/\rho\kappa)},\ \ \ \rho\kappa\in(\pi,\infty).
\end{equation}
Once more, by continuity this is also true for $\phi_e=3\pi/4$.
We proceed to summarize these results.
\begin{proposition}
Consider the transforms~${\cal F}_{\sigma}(\phi_e) $, $\sigma=+,-$, defined by~\eqref{cF}--\eqref{Fm} with
\begin{multline}\label{cFe}
\Psi(r,k)\equiv \frac{w_e(r)^{1/2}}{\sinh\Big( \frac{\pi}{\kappa}\big( k-\frac{i\pi}{\rho}\big)\Big)}
\Big( e^{irk}\Big[ e^{-\pi r/\rho}\sinh\Big( \frac{\pi}{\kappa}\big( k-\frac{i\pi}{\rho}\big)\Big) +
e^{\pi r/\rho}\sinh\big( \frac{\pi}{\kappa} k \big)\Big]
\\
+\sigma e^{-irk}e^{-\pi r/\rho}\sinh\big( \frac{i\pi^2}{\rho\kappa} \big)\Big), \ \ \ \rho\kappa\in(\pi,\infty),
\end{multline}
where $w_e(r)$ is given by \eqref{we}. The transform ${\cal F}_{-}(\phi_e) $
is unitary, whereas ${\cal F}_{+}(\phi_e) $ is isometric and satisfies
\begin{equation}
{\cal F}_+(\phi_e){\cal F}_+(\phi_e)^*={\bf 1}_{ {\cal H} }-\Psi_e\otimes \Psi_e/(\Psi_e,\Psi_e)_1.
\end{equation}
Here, $\Psi_e(r)$ is defined by~\eqref{Psie} and the inner product is given by~\eqref{inpre}.
\end{proposition}
From \eqref{cFe} one sees that for $\sigma=-$ the function $\Psi(r,k)$ is analytic in the strip ${\rm Im}\, k\in[0,\kappa)$. For $\sigma=+$ it has a pole in this strip at $k=i\pi/\rho$, and the bound state $\Psi_e(r)$ is proportional to the residue of~$\Psi(r,k)$ at this pole.
To conclude this appendix, we add one more explicit example satisfying all of the assumptions. We begin by noting that when we have an additional assumption
\begin{equation}\label{addass}
m^-(r,k)=0,
\end{equation}
then we obtain a reflectionless transform, and it also follows that we have
\begin{equation}\label{Mr0}
M^-_{\alpha}(r,k,k')=0,\ \ \ \ \ M^+_{\alpha}(r,k,k')=m^+(-r,-k)m^+(-\alpha r,k'),\ \ \alpha=+,-,
\end{equation}
\begin{equation}\label{Lr0}
L^{\tau,-\tau}(k,r,r')=0,\ \ L^{\tau,\tau}(k,r,r')=\ell^+(\tau r,k)\ell^+(\tau r',-k),\ \ \tau=+,-,
\end{equation}
cf.~\eqref{Mtsdef} and \eqref{Lttp}. Therefore, the critical evenness assumptions~\eqref{Mass} and~\eqref{Lass} reduce to the sole assumption
\begin{equation}\label{L0ass}
\ell^+(r,k)\ell^+(r',-k)=\ell^+(-r,-k)\ell^+(-r',k).
\end{equation}
Moreover, in view of \eqref{ressum}, the condition for the transform to be isometric reduces to
\begin{equation}\label{iso0}
0=\sum_{j=1}^Lw_j
\big[
\mu^{+}(\delta r_j,-k)\mu^{+}(\delta' r_j,k')
-\exp(\rho(\delta k-\delta' k'))
\mu^{+}(-\delta r_j,-k)\mu^{+}(-\delta' r_j,k')\big],
\end{equation}
where we have used the notation~\eqref{mudef}.
We are now prepared for our last example, which yields a reflectionless transform.
\begin{proposition}
The transform~${\cal F}_a $ defined by~\eqref{cF}--\eqref{Fm} with
\begin{equation}\label{cFa}
\Psi(r,k)\equiv w_a(r)^{1/2} e^{irk}\Big[ e^{-\pi r/\rho}+e^{\pi r/\rho}\sinh\Big( \frac{\pi}{\kappa}\big( k+\frac{i\pi}{\rho}\big)\Big)\Big/\sinh\Big( \frac{\pi}{\kappa}\big( k-\frac{i\pi}{\rho}\big)\Big) \Big],
\end{equation}
\begin{equation}\label{wa}
w_a(r)\equiv
1\Big/4\cosh \Big(\frac{\pi}{\rho}\big( r+i\frac{\pi}{\kappa}\big)\Big)\cosh \Big(\frac{\pi}{\rho}\big( r-i\frac{\pi}{\kappa}\big)\Big),
\end{equation}
is isometric for $\rho\kappa\in (2\pi,\infty)$, and satisfies
\begin{equation}
{\cal F}_a{\cal F}_a^*={\bf 1}_{ {\cal H} }-\Psi_a\otimes \Psi_a/(\Psi_a,\Psi_a)_1,\ \ \ \Psi_a(r)\equiv w_a(r)^{1/2},
\end{equation}
where
\begin{equation}
(\Psi_a,\Psi_a)_1=\frac{\pi}{\kappa \sin (2\pi^2/\rho\kappa)},\ \ \ \rho\kappa\in(2\pi,\infty).
\end{equation}
\end{proposition}
\begin{proof}
Here we have
\begin{equation}\label{vhell}
v(k)=1\Big/2i \sinh\Big( \frac{\pi}{\kappa}\big( k-\frac{i\pi}{\rho}\big)\Big), \ \ell^+(r,k)=
2ie^{\pi r/\rho}\sinh\Big( \frac{\pi}{\kappa}\big( k+\frac{i\pi}{\rho}\big)\Big)+(\rho\to -\rho).
\end{equation}
From this the evenness assumption \eqref{L0ass} is readily verified, and all other assumptions are clearly satisfied as well, with
\begin{equation}
T_a(k)=\sinh\Big( \frac{\pi}{\kappa}\big( k+\frac{i\pi}{\rho}\big)\Big)\Big/\sinh\Big( \frac{\pi}{\kappa}\big( k-\frac{i\pi}{\rho}\big)\Big),\ \ \ \ \ R_a(k)=0,
\end{equation}
\begin{equation}
C_a(r)=2\exp(i\pi^2/\rho\kappa)\cosh(\pi r/\rho+i\pi^2/\rho\kappa).
\end{equation}
Consequently, Theorems~A.1 and~B.1 apply, so it remains to study the residue sums.
For the residue sum on the rhs of~\eqref{iso0} we have $L=1$ and we can take
\begin{equation}\label{r1}
r_1=i\pi/\kappa +i\rho/2\in i[0,\rho),\ \ \ \rho\kappa\in(2\pi,\infty).
\end{equation}
From~\eqref{vhell} we then obtain
\begin{equation}\label{ellp}
\ell^+(\nu r_1,k)=-2\nu \sinh(2i\pi^2/\rho\kappa) \exp(\nu\pi k/\kappa),
\end{equation}
which entails
\begin{equation}
\mu^+(\nu r_1,k)=-2\nu \sinh(2i\pi^2/\rho\kappa)v(k) \exp(-\nu \rho k/2).
\end{equation}
From this we see that \eqref{iso0} holds true, so~${\cal F}_a$ is isometric.
Turning to the residue sum in Theorem~B.1, we have $\hat{L}=1$ and
\begin{equation}
k_1=i\pi/\rho\in i[0,\kappa),\ \ \ \hat{w}_1=-i\kappa/4\pi\sin(2\pi^2/\rho\kappa),\ \ \rho\kappa\in(2\pi,\infty).
\end{equation}
From \eqref{vhell} we obtain
\begin{equation}
\ell^+(r,\nu k_1)=-2\nu \sin(2\pi^2/\rho\kappa)\exp(\nu\pi r/\rho),
\end{equation}
so \eqref{Lr0} yields
\begin{equation}
L^{\tau,\tau}( k_1,r,r')=-4\sin(2\pi^2/\rho\kappa)^2\exp(\tau\pi(r-r')/\rho).
\end{equation}
As a result, we have
\begin{eqnarray}
\sum_{\tau=+,-}\frac{\exp(i\tau k_1( r- r'))}{1-\exp(\tau\kappa (r' - r))}L^{\tau,\tau}( k_1,r,r') & = & -4\sin(2\pi^2/\rho\kappa)^2\sum_{\tau=+,-}\frac{1}{1-\exp(\tau\kappa (r' - r))}
\nonumber \\
& = & -4\sin(2\pi^2/\rho\kappa)^2.
\end{eqnarray}
Therefore, \eqref{cFstres} becomes
\begin{equation}
({\cal F}_a^* f,{\cal F}_a^* g)_2= (f,g)_1-\frac{\kappa\sin(2\pi^2/\rho\kappa)}{\pi}(f,\Psi_a)_1(\Psi_a,g)_1,
\end{equation}
whence the proposition follows.
\end{proof}
Once more, the bound state $\Psi_a(r)$ is proportional to the residue of $\Psi(r,k)$ at the pole $k=i\pi/\rho$ in the strip ${\rm Im}\, k\in [0,\kappa)$.
Admittedly, the transform~${\cal F}_a$ may seem to come out of the blue. We have included it, because the function $\Psi(r,k)$~\eqref{cFa} corresponds to the function $\Psi(2a_-;x,y)$ in the main text, cf.~\eqref{psiNr=0}. In Section~4 of~\cite{R00} the reflectionless transforms associated with $\Psi((N+1)a_-;x,y)$, $N\in{\mathbb N}$, were already analyzed, and $\Psi(r,k)$ arises from the function given by Eq.~(4.29) in~\cite{R00}; as such, it satisfies the eigenvalue equation
\begin{multline}\label{aeigen}
\left( \frac{\cosh(\pi r/\rho -i\pi^2/\rho\kappa)}{\cosh(\pi r/\rho)}\exp(i\pi \partial_r/\kappa)+(i\to -i)\right)w_a(r)^{-1/2}\Psi(r,k)
\\
=2\cosh(\pi k/\kappa)w_a(r)^{-1/2}\Psi(r,k).
\end{multline}
(To tie this in with $H_o$~\eqref{Ho}, first take $\rho \leftrightarrow \pi/\kappa$, then put $\kappa =1$. We have deviated from the reparametrization \eqref{xryk} by swapping~$a_+$ and~$a_-$, so that the $r$- and $k$-periodicity assumptions apply to the reflectionless wave function~\eqref{cFa}.)
In fact, all of the assumptions (including~\eqref{addass}) are obeyed by the arbitrary-$N$ attractive eigenfunctions from Section~4 in~\cite{R00}. In particular, the evenness assumption~\eqref{L0ass} is obeyed due to parity features of the latter.
\section{A boundary lemma}
The following lemma yields a template for handling the $\Lambda \to\infty$ limits of the boundary terms arising in the analysis of the transform ${\cal F}$ and its adjoint. It is adapted from the proof of Theorem~2.1 in Appendix~A of~\cite{R00}.
In order to handle ${\cal F}$ and~${\cal F}^*$ at once, we start from two ${\mathbb C}$-valued functions~$G_{\pm}(s,t,t')$ that are defined on~$\{{\rm Re}\, s>0\}\times {\mathbb R}^2$ and have the following features. They are analytic in~$s$ and smooth in~$t,t'$. Their dominant asymptotics as ${\rm Re}\, s\to\infty$ is given by smooth functions~$A_j(t,t')$, $j=1,2$, in the sense that
\begin{equation}\label{Gpmas}
G_+(s,t,t')=A_1(t,t')+\rho_1(s,t,t'),\ \ \ G_-(s,t,t')=A_2(t,t')+\rho_2(s,t,t'),
\end{equation}
where
\begin{equation}\label{rhoj}
\rho_j(s,t,t')=O(\exp(-\eta s)),\ \ {\rm Re}\, s\to\infty,\ \ \eta>0,
\end{equation}
\begin{equation}\label{rhojpart}
\partial_3\rho_j(s,t,t')=O(\exp(-\eta s)),\ \ {\rm Re}\, s\to\infty,\ \ \eta>0,
\end{equation}
with implied constants that are uniform for ${\rm Im}\, s,t,t'$ in compact subsets of~${\mathbb R}$. Finally, we assume
\begin{equation}\label{Geq}
G_+(s,t,t)=G_-(s,t,t),\ \ \ {\rm Re}\, s>0,\ \ t\in{\mathbb R},
\end{equation}
and
\begin{equation}\label{Aeq}
A_j(t,t)=A(t),\ \ j=1,2,\ \ \ t\in{\mathbb R}.
\end{equation}
We are now prepared for our boundary lemma.
\begin{lemma}
Letting $\phi(t,t')\in C_0^\infty ({\mathbb R}^2)$, define
\begin{equation}\label{IR}
I_R\equiv \int_{{\mathbb R}^2}dtdt'\, \phi(t,t') \frac{B_R(t,t')}{\sinh(t'-t)},
\end{equation}
where
\begin{equation}\label{BR}
B_R(t,t')\equiv \int_{R-i}^{R+i}ds\, \sum_{\alpha=+,-}\alpha G_{\alpha}(s,t,t')\exp(i\alpha s(t-t')) ,\ \ \ R>0.
\end{equation}
Then we have
\begin{equation}
\lim_{R\to\infty} I_R=4\pi \int_{-\infty}^{\infty} \phi(t,t)A(t)dt.
\end{equation}
\end{lemma}
\begin{proof}
We begin by noting that by~\eqref{Geq} the integrand in~\eqref{BR} is a smooth function of $(t,t')\in{\mathbb R}^2$ that vanishes for $t=t'$, so that $B_R(t,t')$ has the same properties. Therefore the integrand in~\eqref{IR} belongs to $C_0^\infty({\mathbb R}^2)$. Hence $I_R$ is well defined.
Next, we write
\begin{equation}\label{Btel}
B_R(t,t')= \sum_{j=1}^4
\int_{R-i}^{R+i}ds\, b_j(s,t,t'),
\end{equation}
\begin{equation}\label{b1}
b_1\equiv i\sin
(s(t-t'))A_+(t,t'),\ \ \
b_2\equiv i\sin (s(t-t'))\rho_{+}(s,t,t'),
\end{equation}
\b
b_3\equiv \cos
(s(t-t'))A_-(t,t'),\ \ \
b_4\equiv \cos (s(t-t'))\rho_{-}(s,t,t'),
\end{equation}
where
\begin{equation}\label{Apm}
A_{\pm}(t,t') \equiv A_1(t,t')\pm A_2(t,t'),
\end{equation}
\begin{equation}\label{rhopm}
\rho_{\pm}(s,t,t') \equiv \rho_1(s,t,t')\pm \rho_2(s,t,t').
\end{equation}
Each of the terms in the sum on the rhs of \eqref{Btel}
is a smooth function of $t$ and $t'$ that vanishes
for $t=t'$. Thus the integrals
\begin{equation}\label{defIj}
I_j(R) \equiv \int_{{\mathbb R}^2} dt dt'\,
\frac{\phi(t,t')}{\sinh(t'-t)} \int_{R-i}^{R+i}ds\, b_j(s,t,t'),\ \
j=1,\ldots,4,
\end{equation}
are well defined and it suffices to prove
\begin{equation}\label{I1}
\lim_{R\to\infty}I_1(R)=4\pi \int_{-\infty}^{\infty} \phi(t,t)A(t)dt,
\end{equation}
\begin{equation}\label{Ij}
\lim_{R\to\infty}I_j(R)=0,\ \ \ j=2,3,4.
\end{equation}
In order to prove (\ref{I1}), we use (\ref{defIj}) and
(\ref{b1}) to calculate
\begin{equation}
I_1(R)= \int_{{\mathbb R}^2} dt dt' \,\phi(t,t')\frac{2\sin
R(t-t')}{t-t'}A_+(t,t').
\end{equation}
Invoking the tempered distribution limit
\begin{equation}
\lim_{R\to\infty}\frac{\sin Rx}{x}=\pi \delta(x),
\end{equation}
and using our assumption~\eqref{Aeq}, we now deduce (\ref{I1}).
We continue by studying the integral $I_2(R)$. It can be written
\begin{equation}
i\int_{{\mathbb R}^2} dt dt' \,\phi(t,t') \frac{t-t'}{\sinh
(t'-t)}\int_{R-i}^{R+i}ds\,\frac{\sin
(s(t-t'))}{t-t'}\rho_{+}(s,t,t').
\end{equation}
The integrand of the $s$-integral can be estimated by using
\begin{equation}
\left| \frac{\sin
s(t-t')}{t-t'}\right|=\frac{1}{2}\left|
\int_{-s}^sdxe^{ix(t-t')}\right| \le |s|e^{|t-t'||{\rm Im}\,
s|},
\end{equation}
\begin{equation}
\rho_{+}(s,t,t')=O(\exp (-\eta s)),\ \ \ {\rm Re}\, s\to \infty,
\end{equation}
where the latter bound is uniform for ${\rm Im}\, s, t,t'$ in ${\mathbb R}$-compacts, cf.~\eqref{rhoj} and \eqref{rhopm}. Hence we easily deduce \eqref{Ij} for
$j=2$.
Consider next $I_3(R)$. This integral equals
\begin{equation}
2\int_{{\mathbb R}^2} dt dt' \,\phi(t,t')\cos
R(t-t') \frac{A_-(t,t') }{t-t'}.
\end{equation}
Thus its integrand equals $\cos R(t-t')$ times a
function in $C_0^{\infty}({\mathbb R}^2)$. Its $R\to\infty$ limit
then vanishes by virtue of the Riemann-Lebesgue lemma.
Finally, we take $j=4$ in (\ref{defIj}) and write
\begin{equation}\label{Ifin}
I_4(R)=\int_{{\mathbb R}^2} dt dt' \,\phi(t,t') \frac{t-t'}{\sinh
(t'-t)}
\int_{R-i}^{R+i}ds\,\cos
(s(t-t'))\frac{\rho_{-}(s,t,t')}{t-t'}.
\end{equation}
Now due to our assumptions~\eqref{Geq} and~\eqref{Aeq}, the function $\rho_{-}(s,t,t')$ vanishes for $t=t'$. For $(t,t')$ belonging to the support of $\phi$ we therefore have
\begin{equation}\label{rhomin}
\left|\frac{\rho_{-}(s,t,t')}{t-t'}\right| =\left|
\frac{1}{t-t'}\int_t^{t'}du\,\partial_3\rho_{-}(s,t,u)\right|\le
\max_{(t,t',\theta)\in {\rm supp}( \phi)\times [0,1]}
|\partial_3\rho_{-}(s,t,t+\theta(t'-t))|.
\end{equation}
Invoking our assumption~\eqref{rhojpart}, we infer that the rhs of
(\ref{rhomin}) is $O(\exp (-\eta{\rm Re}\, s))$ for ${\rm Re}\, s\to
\infty$, uniformly for ${\rm Im}\, s,t,t'$ in ${\mathbb R}$-compacts. Clearly,
this entails that (\ref{Ifin}) has limit 0 for $R\to
\infty$, completing the proof of the lemma.
\end{proof}
\section{Time-dependent scattering theory}
In the previous appendices we have not introduced any dynamics, yet we have freely referred to well-known objects from scattering theory, including reflection/transmission coefficients and $S$-matrix. In this appendix we shall explain the relation to time-dependent scattering theory in the general setting of Appendices~A and~B.
First, we recall that this setting solely involves a number of features of the function $\Psi(r,k)$ in terms of which the transform ${\cal F}$ is defined, cf.~\eqref{cF}--\eqref{Fm}. These features were chosen such that they form sufficient hypotheses for Theorems~A.1 and B.1 to hold true. We have already shown that there are nontrivial concrete realizations of these assumptions, yielding the transforms~${\cal F}_{\pm}(\varphi)$, cf.~\eqref{wsp}--\eqref{ellnup}. These assumptions are in force throughout this appendix.
To connect the general transform~${\cal F}$ to time-dependent scattering theory, however, an additional assumption is critical. This assumption is that ${\cal F}$ is an isometry, or equivalently, that the residue sums~\eqref{Rdk} vanish. Admittedly, at face value this seems an assumption of a quite inaccessible nature. The special cases worked out below Corollary~A.2 show that it is not vacuous, but they only involve the simplest case~$L=1$. Indeed, without the further examples coming from the main text, it would be far from obvious that there exist nontrivial arbitrary-$L$ functions $\Psi(r,k)$ satisfying all of the pertinent assumptions.
At any rate, in this appendix we are not concerned with explicit realizations of our assumptions, this being the focus of the main text. More is true: We need not even restrict attention to the special dynamics that pertains to the relativistic hyperbolic Calogero-Moser systems.
Specifically, we consider a vast class of dynamics to which the latter belongs. We start from multiplication operators on $\hat{{\cal H}}$ of the diagonal form
\begin{equation}\label{hatM}
(\hat{M}f)_{\delta}(k)=\mu(k)f_{\delta}(k),\ \ \ \delta=+,-,\ \ \ k>0.
\end{equation}
Here, $\mu(k)$ is any real-valued, smooth, even function on~${\mathbb R}$ whose (odd) derivative
satisfies
\begin{equation}\label{muder}
\mu'(k)>0,\ \ \ k>0.
\end{equation}
Thus $\mu(k)$ is strictly increasing on~$(0,\infty)$, but not necessarily unbounded. (For example, the function $1-\exp(-k^2)$ satisfies the assumptions.) On its maximal multiplication domain ${\cal D}(\hat{M})\subset \hat{{\cal H}}$, the operator~$\hat{M}$ is self-adjoint, and the subspace $\hat{{\cal C}}=C_0^{\infty}((0,\infty))^2\subset {\cal D}(\hat{M})$ is a core (domain of essential self-adjointness).
We can now define an operator $M$ on the subspace
\begin{equation}\label{DMdef}
{\cal D}(M)\equiv {\cal F}({\cal D}(\hat{M}))\subset {\cal H},
\end{equation}
by setting
\begin{equation}\label{Mdef}
M{\cal F} f\equiv {\cal F} \hat{M} f,\ \ f\in {\cal D}(\hat{M}).
\end{equation}
Since ${\cal F}$ is isometric (by asumption), this yields a self-adjoint operator~$M$ on the Hilbert space
\begin{equation}\label{ranF}
{\cal H}^r\equiv {\cal F}(\hat{{\cal H}}).
\end{equation}
Furthermore, ${\cal F}(\hat{{\cal C}})$ is a core for~$M$. In case ${\cal H}^r$ is a proper subspace of ${\cal H}$ (so that ${\cal F}$ is not unitary), we define $M$ to be equal to an arbitrary self-adjoint operator on the orthogonal complement of~${\cal H}^r$; this choice plays no role in the scattering theory of this appendix, but it will be made definite in the main text for the dynamics at issue there, cf.~the last paragraph of Section~4.
With the dynamics $M$ thus defined as a self-adjoint operator on ${\cal H}$, we now consider the associated `interacting' unitary time evolution
\begin{equation}
{\cal U}(t)\equiv \exp(-itM),\ \ t\in{\mathbb R},
\end{equation}
in relation to a `free' evolution defined by using the Fourier transform~${\cal F}_0$, cf.~\eqref{free}. Specifically, with~${\cal F}_0$ in the role of~${\cal F}$, we obtain a self-adjoint operator $M_0$ on~${\cal H}$ with associated time evolution
\begin{equation}\label{U0t}
{\cal U}_0(t)\equiv \exp(-itM_0),\ \ t\in{\mathbb R}.
\end{equation}
This evolution can be compared to ${\cal U}(t)$ in the usual sense of time-dependent scattering theory. We recall that this amounts to studying the (strong) limits of the unitary family ${\cal U}(-t){\cal U}_0(t)$, yielding the isometric wave operators
\begin{equation}\label{Wpm}
{\cal W}_{\pm}\equiv \lim_{t\to \pm \infty}{\cal U}(-t){\cal U}_0(t),
\end{equation}
in case the limits exist~\cite{RS79}.
The crux is now that our assumptions already suffice to prove that these limits do exist. Moreover, the transform ${\cal F}$ is substantially equivalent to the incoming wave operator ${\cal W}_-$, in the sense that it is equal to ${\cal W}_-{\cal F}_0$.
To avoid possible confusion, we should stress that this equality does not give rise to an isometry proof for ${\cal F}$. Indeed, we need to \emph{assume} that~${\cal F}$ is isometric to begin with, so as to obtain a unitary evolution~${\cal U}(t)$, cf.~\eqref{DMdef}--\eqref{Mdef}.
Before stating the pertinent theorem, we specify the function~$\mu(k)$ that corresponds to the defining relativistic Calogero-Moser dynamics in the main text: It reads
\begin{equation}\label{muCM}
\mu_{CM}(k)=2\cosh(\rho k).
\end{equation}
The action of the associated operator $M_{CM}$ on the core ${\cal F}(\hat{{\cal C}})$ is that of the analytic difference operator
\begin{equation}\label{HCM}
H_{CM}=\exp(i\rho \partial_r)+\exp(-i\rho \partial_r).
\end{equation}
Indeed, $H_{CM}$ yields the eigenvalues~\eqref{muCM} when acting on~$\Psi(r,k)$. To be more specific, in the main text the $w(r)$-poles are on the imaginary axis and the coefficients~$w(r)^{1/2}m^{\pm}(r,k)$ multiplying the plane waves~$\exp(\pm irk)$ are $i\rho$-periodic for $|{\rm Re}\, r|>0$. (Note that the assumptions~\eqref{mpas}--\eqref{mnas} entail that this also holds true in the present axiomatic setting when we let $|{\rm Re}\, r|>|{\rm Re}\, r_j|$, $j=1,\ldots,2L$, so as to avoid the branch points.)
\begin{theorem}
Assuming the transform~${\cal F}$ is isometric, define dynamics by~\eqref{hatM}--\eqref{U0t}. Then the strong limits \eqref{Wpm} exist. They are explicitly given by
\begin{equation}\label{Wm}
{\cal W}_-={\cal F} {\cal F}_0^*,
\end{equation}
and
\begin{equation}\label{Wp}
{\cal W}_+={\cal W}_-{\cal F}_0S(\cdot)^*{\cal F}_0^*,
\end{equation}
where $S(k)$ is the unitary matrix multiplication operator~\eqref{Sm} on~$\hat{{\cal H}}$.
\end{theorem}
\begin{proof}
We first prove that~${\cal W}_-$ exists and is given by~\eqref{Wm}. Since~${\cal F}$ is isometric and the time evolutions are unitary, we need only show that the ${\cal H}$-norm
\begin{equation}\label{gm}
\| \big( e^{itM}e^{-itM_0}{\cal F}_0-{\cal F}\big)f\|_1=\| ({\cal F}_0-{\cal F})e^{-it\hat{M}}f\|_1,
\end{equation}
with $f=(f_+,f_-)$ an arbitrary function in the dense subspace~$\hat{{\cal C}}$ of~$\hat{{\cal H}}$, vanishes for~$t\to-\infty$. To this end, consider
\begin{eqnarray}
\big( ({\cal F}_0-{\cal F})e^{-it\hat{M}}f\big)(r) &=&\frac{1}{\sqrt{2\pi}}\int_{0}^{\infty} dk\, \sum_{\delta=+,-}\delta \big(e^{i\delta rk}-\Psi(\delta r,k)\big)e^{-it\mu(k)}f_{\delta}(k)
\nonumber \\
& =: & \frac{1}{\sqrt{2\pi}}\sum_{\delta=+,-}\delta \psi_{\delta}(r).
\end{eqnarray}
Taking $r\to -r$ in $\psi_-(r)$, we see that it suffices to prove that the ${\cal H}$-norm of the function
\begin{equation}
g(t;r)\equiv \int_{0}^{\infty} dk\, \big(e^{i rk}-\Psi( r,k)\big)e^{-it\mu(k)}f(k),\ \ \ f\in C_0^\infty((0,\infty)),
\end{equation}
vanishes for $t\to -\infty$. Recalling~\eqref{Psiform}, we obtain
\begin{equation}
g(t;r)=\sum_{\tau=+,-}\tau g^{\tau}(t;r),
\end{equation}
where
\begin{equation}
g^+(t;r)\equiv \int_{0}^{\infty} dk\, e^{i rk}(1-w(r)^{1/2}m^+(r,k))e^{-it\mu(k)}f(k),
\end{equation}
\begin{equation}
g^-(t;r)\equiv \int_{0}^{\infty} dk\, e^{-irk}w(r)^{1/2}m^-(r,k)e^{-it\mu(k)}f(k).
\end{equation}
From this we see that we need only change variables $k\to x\equiv \mu(k)$ and invoke the Riemann-Lebesgue lemma to obtain convergence to zero of $g^{\pm}(t;r)$ for $t\to -\infty$ and fixed~$r$. (Note that our assumption~\eqref{muder} ensures that this change of variables is well defined.)
As a consequence, it remains to supplement this pointwise convergence with $L^1({\mathbb R})$ dominating functions for the functions~$|g^{\pm}(t;r)|^2$. To do so, we split up the integration over~${\mathbb R}$ into intervals $[-R,R]$, $(-\infty,-R]$, and $[R,\infty)$, where~$R$ is chosen large enough so that we may invoke the asymptotics \eqref{mpas}--\eqref{mnas}. On the first interval the functions~$|g^{\pm}(t;r)|^2$ are clearly bounded uniformly in~$t$, so this contribution vanishes for $t\to -\infty$.
Next, we bound~$|g^+(t;r)|^2$ on the two tail intervals. For the left interval we use~\eqref{mpas} to obtain an exponentially decreasing dominating function. On the right interval we get two contributions from~\eqref{mpas}, the second one again yielding an exponentially decreasing dominating function. Thus we are left with obtaining a suitable bound for the function
\begin{equation}\label{gright}
\int_{0}^{\infty} dk\, e^{i rk}(1-T(k))e^{-it\mu(k)}f(k),\ \ r\in[R,\infty),
\end{equation}
with $t\le -1$, say. To this end, we write the exponentials as
\begin{equation}\label{exp1}
(r-t\mu'(k))^{-1}(-i\partial_k)\exp(irk-it\mu(k)),
\end{equation}
noting that by our assumption~\eqref{muder} the denominator $r-t\mu'(k)$ is bounded away from zero on the compact support of~$f(k)$ for~$t\le -1$ and $r\ge R$.
Integrating by parts and estimating in the obvious way, this yields an $O(1/r)$-majorization that is uniform for~$t\le -1$. Thus the modulus squared of the function~\eqref{gright} is bounded above by $C/r^2$ for all $t\le -1$, so that by the dominated convergence theorem its $L^2([R,\infty))$-norm vanishes for~$t\to -\infty$.
It remains to bound~$|g^-(t;r)|^2$ on the tail intervals. On the right one we get an exponentially decreasing dominating function from~\eqref{mnas}. On the left, the second term in~\eqref{mnas} yields again an exponentially decreasing dominating function, so it remains to bound
\begin{equation}\label{gleft}
\int_{0}^{\infty} dk\, e^{-i rk}R(k)e^{-it\mu(k)}f(k),\ \ r\in(-\infty,-R],
\end{equation}
uniformly for $t\le -1$. Writing
\begin{equation}\label{exp2}
(r+t\mu'(k))^{-1}(i\partial_k)\exp(-irk-it\mu(k)),
\end{equation}
and integrating by parts, we readily obtain a uniform $O(1/r)$-majorization that suits our purpose. As a result, we have now proved existence of~${\cal W}_-$ and its explicit form~\eqref{Wm}.
Next, we show that~${\cal W}_+$ exists as well, and that ${\cal W}_+{\cal F}_0$ equals ${\cal F}
S(\cdot)^*$. (In view of~\eqref{Wm}, this amounts to~\eqref{Wp}.) Proceeding along the same lines as before, we study
\begin{equation}
\| \big( e^{itM}e^{-itM_0}{\cal F}_0-{\cal F} S(\cdot)^*\big)f\|_1=\| ({\cal F}_0-{\cal F} S(\cdot)^*)e^{-it\hat{M}}f\|_1,\ \ \ f\in\hat{{\cal C}},
\end{equation}
for $t\to\infty$. We have
\begin{equation}
\big( ({\cal F}_0-{\cal F} S(\cdot)^*)\exp(-it\hat{M})f\big)(r)= \frac{1}{\sqrt{2\pi}}\sum_{\delta=+,-}\delta \phi_{\delta}(r),
\end{equation}
where
\begin{equation}
\phi_{\delta}(r)\equiv \int_{0}^{\infty} dk\, \big(e^{i\delta rk}-\Psi(\delta r,k)T(-k)+\Psi(-\delta r,k)R(-k)\big)e^{-it\mu(k)}f_{\delta}(k).
\end{equation}
Taking $r\to -r$ in $\phi_-(r)$, we deduce that we need only show that the $L^2({\mathbb R})$-norm of the function
\begin{equation}
h(t;r)\equiv \int_{0}^{\infty} dk\, \big(e^{i rk}-\Psi( r,k)T(-k)+\Psi(- r,k)R(-k)\big)e^{-it\mu(k)}f(k),\ \ \ f\in C_0^\infty((0,\infty)),
\end{equation}
vanishes for $t\to \infty$. Now from~\eqref{Psiform} we get
\begin{equation}
h(t;r)=\sum_{\tau=+,-}\tau h^{\tau}(t;r),
\end{equation}
where
\begin{equation}
h^+(t;r)\equiv \int_{0}^{\infty} dk\, e^{i rk}\big(1-w(r)^{1/2}m^+(r,k)T(-k)+w(r)^{1/2}m^-(-r,k)R(-k)\big)e^{-it\mu(k)}f(k),
\end{equation}
\begin{equation}
h^-(t;r)\equiv \int_{0}^{\infty} dk\, e^{-irk}\big(w(r)^{1/2}m^-(r,k)T(-k)-w(r)^{1/2}m^+(-r,k)R(-k)\big)e^{-it\mu(k)}f(k).
\end{equation}
As before, the Riemann-Lebesgue lemma implies that the functions~$h^{\pm}(t;r)$ vanish for $t\to \infty$ and fixed~$r$, so it remains to supply dominating functions for $t\ge 1$, say. This can be done by adapting the above reasoning for the incoming wave operator. More in detail, for~$h^+(t;r)$ we can invoke the unitarity relation~\eqref{unit1} when $r\ge R$ and use~\eqref{exp1} for integration by parts when $r\le -R$, whereas~$h^-(t;r)$ can be handled by using~\eqref{unit2} for $r\le -R$ and~\eqref{exp2} for $r\ge R$. Thus our proof is now complete.
\end{proof}
In the language of time-independent scattering theory, this theorem reveals that the integral kernel of ${\cal F}$ is the incoming wave function
\begin{equation}\label{incom}
\Psi^{in}(r,k)=
\left(\begin{array}{c}
\Psi(r,k) \\
-\Psi(-r,k)
\end{array}\right),\ \ \ k>0,
\end{equation}
cf.~\eqref{cFdef}--\eqref{Fm}, whose relation to the outgoing one
\begin{equation}\label{outgo}
\Psi^{out}(r,k)=\left(\begin{array}{c}
\Phi(r,k) \\
-\Phi(-r,k)
\end{array}\right),\ \ \ k>0,
\end{equation}
is given by
\begin{equation}\label{Phi}
\Phi(r,k)=T(-k)\Psi(r,k)-R(-k)\Psi(-r,k).
\end{equation}
Equivalently, we have
\begin{equation}
\Psi^{in}(r,k)=S(k)\Psi^{out}(r,k),
\end{equation}
with the $S$-matrix $S(k)$ given by~\eqref{Sm}.
It follows from our assumptions that each of the above dynamics defined via~${\cal F}$ is invariant under the usual parity operator
\begin{equation}
({\cal P} f )(r)\equiv f(-r),\ \ \ f\in{\cal H}.
\end{equation}
Indeed, we readily calculate that the operator
\begin{equation}
\hat{{\cal P}}\equiv {\cal F}^* {\cal P} {\cal F},
\end{equation}
is given by
\begin{equation}\label{pari}
(\hat{{\cal P}}g)(k)=
\left(\begin{array}{cc}
0 & -1 \\
-1 & 0
\end{array}\right)g(k),\ \ \ g\in\hat{{\cal H}}.
\end{equation}
This implies that the range~${\cal H}^r$ of ${\cal F}$ is left invariant by~${\cal P}$, and since $\hat{M}$ commutes with~$\hat{{\cal P}}$, we also have $[{\cal P},\exp(itM)]=0$. (To be quite precise, this holds when we define $M$ on the orthocomplement~$({\cal H}^r)^{\perp}$ of~${\cal H}^r$ in such a way that it also commutes with~${\cal P}$ on~$({\cal H}^r)^{\perp}$.)
By contrast, the state of affairs for the customary time reversal operator
\begin{equation}\label{cT}
({\cal T} f)(r)\equiv \overline{f(r)},\ \ f\in{\cal H},
\end{equation}
is not clear in the present axiomatic context (as opposed to the main text, as we shall see shortly). We proceed to elaborate on this. First, we can easily calculate
\begin{equation}
\hat{{\cal T}}_0 \equiv {\cal F}_0^* {\cal T} {\cal F}_0,
\end{equation}
yielding
\begin{equation}
(\hat{{\cal T}}_0g)(k) =\left(\begin{array}{cc}
0 & -1 \\
-1 & 0
\end{array}\right)\overline{g(k)},\ \ \ g\in\hat{{\cal H}}.
\end{equation}
Thus $\hat{{\cal T}}_0$ commutes with all of the dynamics~$\hat{M}$ given by~\eqref{hatM}. Clearly, this entails
\begin{equation}\label{cTM0}
{\cal T}\exp(-itM_0)=\exp(itM_0){\cal T},\ \ \ t\in{\mathbb R}.
\end{equation}
The difficulty is now that we do not know whether our assumptions imply~\eqref{cTM0} with $M_0\to M$.
To explain what is involved, let us assume that this is indeed the case.
Then, the definition~\eqref{Wpm} of the wave operators entails
\begin{equation}
{\cal T} {\cal W}_-={\cal W}_+{\cal T}.
\end{equation}
Using~\eqref{Wm} and~\eqref{Wp}, we readily deduce
\begin{equation}\label{intid}
{\cal F}={\cal T}{\cal F} S(\cdot)^*\hat{{\cal T}}_0.
\end{equation}
When we now compare the kernels of the transforms in~\eqref{intid}, then
we obtain
\begin{equation}
\Psi(r,k)=T(k)\overline{\Psi(-r,k)}-R(k)\overline{\Psi(r,k)}.
\end{equation}
In view of our standing assumption~\eqref{Psiconj}, this amounts to the identity
\begin{equation}\label{cTsym}
\Psi(r,k)=T(k)\Psi(-r,-k)-R(k)\Psi(r,-k).
\end{equation}
On account of~\eqref{incom}--\eqref{Phi}, this identity can also be rewritten as the relation
\begin{equation}\label{outin}
\Psi^{out}(r,k)=\left(\begin{array}{cc}
0 & -1 \\
-1 & 0
\end{array}\right)\overline{\Psi^{in}(r,k)}.
\end{equation}
Clearly, this argument can be reversed: Assuming~\eqref{cTsym} holds true, we deduce
\begin{equation}\label{cTM}
{\cal T}\exp(-itM)=\exp(itM){\cal T},\ \ \ t\in{\mathbb R}.
\end{equation}
For the transforms coming from the main text, the time reversal identity~\eqref{cTsym} is indeed valid, cf.~\eqref{psirev}. Moreover, it can also be verified for the two transform kernels~\eqref{cFe} in Proposition~B.4. But we do not know whether~\eqref{cTsym} is a necessary consequence of the assumptions we made in Theorem~D.1.
From now on, we add to the assumptions of Theorem~D.1 the extra assumption that~${\cal F}$ is unitary (equivalently, that~${\cal H}^r$ equals~${\cal H}$, cf.~\eqref{ranF}). As we shall show next, this implies that the function~$U(r)$ arising from the large-$|k|$ asymptotics of the coefficients (cf.~\eqref{Psikas}) may be viewed as the $S$-matrix of a large class of `dual' dynamics. This class arises by starting from real-valued, smooth, odd functions~$d(r)$, which satisfy
\begin{equation}\label{derpos}
d'(r)>0,\ \ \ r\in{\mathbb R}.
\end{equation}
Thus $d(r)$ is strictly increasing, but need not be unbounded. (For example, the function $\tanh r$ satisfies the assumptions.)
Any such function gives rise to a self-adjoint multiplication operator~$D$ on its natural domain~${\cal D}(D)\subset {\cal H}$. Since ${\cal F}$ is assumed to be unitary, we can now define a self-adjoint operator $\hat{D}$ on~$\hat{{\cal H}}$
by
\begin{equation}\label{Dhdef}
\hat{D}{\cal F}^* f\equiv {\cal F}^* D f,\ \ f\in {\cal D}(D).
\end{equation}
We continue to compare the associated `interacting' unitary time evolution
\begin{equation}
\hat{{\cal U}}(t)\equiv \exp(-it\hat{D}),\ \ t\in{\mathbb R},
\end{equation}
to the `free' evolution defined by using~${\cal F}_0^*$. Thus, replacing~${\cal F}^*$ by~${\cal F}_0^*$ in~\eqref{Dhdef}, we obtain a self-adjoint operator $\hat{D}_0$ on~$\hat{{\cal H}}$, which yields a time evolution
\begin{equation}
\hat{{\cal U}}_0(t)\equiv \exp(-it\hat{D}_0),\ \ t\in{\mathbb R}.
\end{equation}
As before, our goal is to show that the dual wave operators,
\begin{equation}\label{Whpm}
\hat{{\cal W}}_{\pm}\equiv \lim_{t\to \pm \infty}\hat{{\cal U}}(-t)\hat{{\cal U}}_0(t),
\end{equation}
exist, and to clarify their relation to~${\cal F}^*$ and the unitary operator
\begin{equation}\label{Udef}
(Uf)(r)\equiv U(r)f(r),\ \ \ f\in{\cal H}.
\end{equation}
Before doing so, we detail the function~$d(r)$ that arises from the dual relativistic Calogero-Moser dynamics. It is given by
\begin{equation}\label{dCM}
d_{CM}(r)=2\sinh(\kappa r),
\end{equation}
and the corresponding operator $\hat{D}_{CM}$ acts on the core ${\cal F}^*({\cal C})$ as the analytic difference operator
\begin{equation}\label{HhCM}
\hat{H}_{CM}=\exp(-i\kappa \partial_k)-\exp(i\kappa \partial_k).
\end{equation}
To see that this gives rise to the `eigenvalues'~\eqref{dCM}, recall that we have assumed (above~\eqref{mtform}) that~$m^+(r,k)$ is $i\kappa$-periodic in~$k$, whereas~$m^-(r,k)$ is assumed to be $i\kappa$-antiperiodic.
\begin{theorem}
The strong limits \eqref{Whpm} exist and are given by
\begin{equation}\label{Whm}
\hat{{\cal W}}_{\pm}={\cal F}^* U(\cdot)^{\mp 1/2} {\cal F}_0,
\end{equation}
where $U(r)$ is the unitary multiplication operator on~${\cal H}$ given by the square of~\eqref{Ur}.
\end{theorem}
\begin{proof}
Our proof is patterned after the proof of Theorem~D.1.
To show that~$\hat{{\cal W}}_-$ exists and is given by~\eqref{Whm}, we start from the $\hat{{\cal H}}$-norm
\begin{equation}\label{gmd}
\| \big( e^{it\hat{D}}e^{-it\hat{D}_0}{\cal F}_0^*-{\cal F}^*U(\cdot)^{1/2}\big)f\|_2=\| ({\cal F}_0^*-{\cal F}^*U(\cdot)^{1/2})e^{-itD}f\|_2,\ \ f\in{\cal C}.
\end{equation}
In order to prove it vanishes for~$t\to-\infty$, we consider
\begin{equation}\label{difde}
\big( {\cal F}_0^*-{\cal F}^*U(\cdot)^{1/2})e^{-itd(\cdot)} f\big)_{\delta}(k) =\frac{\delta}{\sqrt{2\pi}}\int_{{\mathbb R}} dr\, \big(e^{-i\delta rk}-\Psi(\delta r,-k)U(r)^{1/2}\big)e^{-itd(r)}f(r).
\end{equation}
By virtue of the Riemann-Lebesgue lemma, this function vanishes for $t\to -\infty$ and~$k>0$ fixed, so we need only exhibit a suitable dominating function in~$L^1((0,\infty),dk)$.
For $\delta=+$ we should look at the functions
\begin{equation}\label{gpp}
\hat{g}_+^+(t;k)\equiv \int_{{\mathbb R}} dr\, e^{-i rk}(1-w(r)^{1/2}m^+(r,-k)U(r)^{1/2})e^{-itd(r)}f(r),
\end{equation}
\begin{equation}\label{gpm}
\hat{g}_+^-(t;k)\equiv \int_{{\mathbb R}} dr\, e^{irk}w(r)^{1/2}m^-(r,-k)U(r)^{1/2}e^{-itd(r)}f(r).
\end{equation}
To supply dominating functions for~$|\hat{g}_+^{\pm}(t;k)|^2$, we split integration over~$(0,\infty)$ into intervals $(0,R]$ and $[R,\infty)$, where~$R$ is chosen large enough for the asymptotics \eqref{mpask}--\eqref{mmask} to be valid. Since the functions~$|\hat{g}_+^{\pm}(t;k)|^2$ are bounded uniformly in~$t$ on~$(0,R]$, the contribution of this interval vanishes for $t\to -\infty$.
To bound~$|\hat{g}_+^+(t;k)|^2$ on $[R,\infty)$, we need only invoke~\eqref{mpask} and~\eqref{Ur}. Indeed, from this we deduce that the dominant contribution cancels, so we are left with an exponentially decreasing dominating function. For~$|\hat{g}_+^-(t;k)|^2$ the existence of such a function is immediate from~\eqref{mmask}, so it now follows that the norm of the function~\eqref{difde} with~$\delta=+$ vanishes for~$t\to -\infty$.
For the choice~$\delta=-$, we should majorize the functions
\begin{equation}\label{gmp}
\hat{g}_-^+(t;k)\equiv \int_{{\mathbb R}} dr\, e^{i rk}(1-w(r)^{1/2}m^+(-r,-k)U(r)^{1/2})e^{-itd(r)}f(r),
\end{equation}
\begin{equation}\label{gmm}
\hat{g}_-^-(t;k)\equiv \int_{{\mathbb R}} dr\, e^{-irk}w(r)^{1/2}m^-(-r,-k)U(r)^{1/2}e^{-itd(r)}f(r).
\end{equation}
As before, boundednes for $k\in (0,R]$ is plain, and
just as for~\eqref{gpm}, we get an exponentially decreasing dominating function for~\eqref{gmm} on~$[R,\infty)$ right away from~\eqref{mmask}.
Invoking once again~\eqref{mpask} and~\eqref{Ur}, we deduce that it remains to bound the function
\begin{equation}\label{ghright}
\int_{{\mathbb R}} dr\, e^{i rk}\big(1-w(r)\overline{C(-r)}C(r)\big)e^{-itd(r)}f(r),\ \ k\in[R,\infty),
\end{equation}
for $t\le -1$. Writing
\begin{equation}\label{exp3}
(k-td'(r))^{-1}(-i\partial_r)\exp(irk-itd(r)),
\end{equation}
we observe that by our assumption~\eqref{derpos} the denominator is bounded away from zero on the compact support of~$f(r)$ for~$t\le -1$ and $k\ge R$.
It easily follows that when we integrate by parts we can obtain an $O(1/k)$-bound that is uniform for~$t\le -1$. Thus the $L^2([R,\infty),dk)$-norm of~\eqref{ghright} vanishes for~$t\to -\infty$.
The upshot is that we have completed the proof that $\hat{W}_-$ exists and is given by~\eqref{Whm}. The proof for~$\hat{W}_+$ only involves some obvious changes, so we omit it.
\end{proof}
\end{appendix}
\vspace{5mm}
\noindent
{\Large\bf Acknowledgments}
\vspace{8mm}
\noindent
We would like to thank the referees for their comments, which helped us to improve the exposition.
\vspace{4mm}
\bibliographystyle{amsalpha}
|
\section{Introduction}
We shall work in an $n-$dimensional vector space $\mathbb R^n$ with standard orthonormal basis $e_1,...,e_n$ and a scalar product $\langle \cdot,\cdot\rangle$. The standard Euclidean length is denoted by $|\cdot|$.
A set $K$ in $\mathbb R^n$ is said to be convex if together with every pair of points it contains the interval connecting them. Compact convex sets with non-empty interior are called convex bodies.
The standard Lebesgue measure of a set $A$ in $\mathbb R^n$ shall be denoted by $|A|$ or, sometimes, $|A|_n$. When the standard Lebesgue measure on a subspace of dimension $k$ is considered, it shall be denoted by $|\cdot|_k$. We shall denote the unit ball centered at the origin in $\mathbb R^n$ by $B_2^n$, and the unit sphere by ${{\mathbb S}^{n-1}}.$
Given a convex body $K$ in $\mathbb R^n$, its Gauss map $\nu_K:\partial K\rightarrow {{\mathbb S}^{n-1}}$ is the map that corresponds to every $y\in \partial K$ the set of normal vectors at $y$ with respect to $K.$ The surface area measure of $K$ is the measure on the unit sphere defined as the push forward to the sphere of the $(n-1)-$dimensional Hausdorff measure on $\partial K$ via the map $\nu_K$. It is denoted by $\sigma_K.$
Minkowski's existence theorem guarantees that every barycentered measure on ${{\mathbb S}^{n-1}}$ which is not supported on any great subsphere is a surface area measure for some convex body; moreover, a convex body is determined by its surface area measure uniquely up to a shift.
For $p\in\mathbb R$, the $L_p$ surface area measure of a convex body with the support function $h_K$ is the measure on the sphere given by $d\sigma_{p,K}(u)=h_K^{1-p}(u) d\sigma_{K}(u)$. It was introduced by Lutwak. The normalized $L_p$ surface area is given by $d\bar{\sigma}_{p,K}(u)=\frac{1}{|K|}d\sigma_{p,K}(u)$. An extension of Minkowski's Theorem, called \emph{$L_p-$Minkowski problem} is open in general. It asks which conditions should be required in order for a measure on the sphere to be an $L_p-$surface area measure, as well as whether $L_p-$surface area measure determines a convex body uniquely. Lutwak, Yang, Zhang have solved the normalized $L_p$-Minkowski problem with even data for the case $p\leq 0$, and showed the uniqueness of the solution when $p<0$. B\"or\"oczky, Lutwak, Yang, Zhang \cite{BLYZ}, \cite{BLYZ-1}, \cite{BLYZ-2} have studied the case $p=0$ and have, in particular, obtained the uniqueness in the case of symmetric convex bodies on the plane. Stancu \cite{St}, \cite{St1} has treated this problem for polytopes on the plane. Huang, Liu, Xu \cite{HLX} have established uniqueness in $\mathbb R^3$ in the case when the $L_p$ surface area is constant. The $L_p$-Minkowski problem is one of the main questions in the rapidly developing Brunn-Minkowski-Firey theory (see more in Ludwig \cite{Lud}, Lutwak \cite{Lut1}, \cite{Lut2}, Lutwak, Yang, Zhang \cite{LYZ-1}, \cite{LYZ-2}, \cite{LYZ-3}, Lutwak, Oliker \cite{LO}, Meyer, Werner \cite{MW}, Ryabogin, Zvavitch \cite{RZ}, Zhu \cite{zhu}, \cite{zhu1}, and the references therein).
In this manuscript, we prove an analogue of Minkowski's theorem in a different setting. Let $\mu$ be an absolutely continuous measure on $\mathbb R^n$. We study the surface area measure of convex bodies with respect to $\mu$.
\begin{definition}
Let $K$ be a convex body and $\nu_K$ be its Gauss map. Let $\mu$ be a measure on $\mathbb R^n$ with density $g(x)$ continuous on its support. Define $\sigma_{\mu,K}$ on ${{\mathbb S}^{n-1}}$, a surface area measure of $K$ with respect to $\mu$, as follows: for every Borel set $\Omega\subset{{\mathbb S}^{n-1}}$, let
$$\sigma_{\mu,K}(\Omega)=\int_{\nu_K^{-1}(\Omega)} g(x) dH_{n-1}(x),$$
where $H_{n-1}$ stands for the $(n-1)-$dimensional Hausdorff measure on $\partial K$, and $\nu_K^{-1}(\Omega)$ stands for the full pre-image of $\Omega$ under $\nu_K.$
\end{definition}
When $\mu$ is the standard Lebesgue measure, the measure $\sigma_{\mu, K}$ coincides with $\sigma_K,$ the classical surface area measure.
Let $p\in (0,+\infty)$. We say that a function $f:\mathbb R^n\rightarrow [0,\infty]$ is $p-$concave if $f^p(x)$ is a concave function on its support. That is, for every $x,y\in supp(f)$ and for every $\lambda\in [0,1]$ we have
$$f^p(\lambda x+(1-\lambda) y)\geq \lambda f^p(x)+(1-\lambda) f^p(y).$$
Let $r\in (-\infty,+\infty)$. We say that a function $f:\mathbb R^n\rightarrow [0,\infty]$ is $r-$ homogenous if for every $a>0$ and for every $x\in\mathbb R^n$ we have $f(ax)=a^{r} f(x)$.
We shall consider the class of measures on $\mathbb R^n$ with densities that have a positive degree of homogeneity and a positive degree of concavity. In fact, all such densities are $p-$concave and $\frac{1}{p}$-homogenous for the same $p\geq 0$ (see the Proposition \ref{p-concave-proposition} from the Appendix). This class of measures was considered by E. Milman and L. Rotem \cite{MilRot}, where they studied their isoperimetric properties. We remark that such measures are necessarily supported on convex cones. An example of a density function with said properties is $f(x)=1_{\{\langle x,\theta\rangle>0\}}|\langle x, \theta\rangle|^{\frac{1}{p}}$, where $\theta$ is a vector.
We prove an extension of Minkowski's existence theorem to the class of surface area measures with respect to measures with positive degree of concavity and positive degree of homogeneity.
\begin{theorem}\label{mink-existence-intro}
Let $\mu$ on $\mathbb R^n$ be a measure and $g(x)$ be its even $r-homogenous$ density for some $r\geq 0$, and the restriction of $g$ to some half space is $p-concave$ for a $p\geq 0$. Let $\varphi(u)$ be an arbitrary even measure on ${{\mathbb S}^{n-1}}$, not supported on any great subsphere, such that $supp(\varphi)\subset int(supp(g))\cap {{\mathbb S}^{n-1}}$. Then there exists a symmetric convex body $K$ in $\mathbb R^n$ such that
$$d\sigma_{K,\mu}(u)=d\varphi(u).$$
Moreover, such convex body is determined uniquely up to a set of $\mu-$measure zero.
\end{theorem}
In Theorem \ref{mink-existence-intro}, and throughout the paper, uniqueness up to $\mu-$measure zero means that for every pair of $K$ and $L$, symmetric convex bodies with $\sigma_{K,\mu}=\sigma_{L,\mu}$, the measure of their symmetric difference $\mu(K\Delta L)=0.$
We apply Theorem \ref{mink-existence-intro} to extend the study of volume comparison and unique determination of convex bodies related to projections.
Given a unit vector $u\in{{\mathbb S}^{n-1}}$, we consider an $(n-1)-$dimensional hyperplane orthogonal to it:
$$u^{\perp}=\{x\in\mathbb R^n\,:\, \langle x,u\rangle=0\}.$$
An orthogonal projection of a convex body $K$ to a subspace $u^{\perp}$ shall be denoted by $K|u^{\perp}$; that is,
$$K|u^{\perp}=\{x\in u^{\perp}\,:\, \exists t\in \mathbb R\, s.t.\, x+tu\in K\}.$$
Let K be an origin symmetric convex body in $\mathbb R^n$ with curvature function $f_K$. The projection body $\Pi K$ of $K$ is defined as the origin symmetric convex body in $\mathbb R^n$ whose support function in every direction is equal to the volume of the hyperplane projection of K in this direction.
The Shephard problem (see Shephard \cite{shep}) is the following question: \emph{given symmetric convex bodies $K$ and $L$ such that for every $u\in {{\mathbb S}^{n-1}}$
$$|K|u^{\perp}|_{n-1}\leq |L| u^{\perp}|_{n-1},$$
does it follow that $|K|_n\leq |L|_n$?}
The problem was solved independently by Petty \cite{P} and Schneider \cite{Sc1}. They showed that the answer is affirmative if $n\leq 2$ and negative if $n\geq 3$. More precisely, the answer to Shephard's problem is affirmative if and only if $L$ is a projection body. As for general symmetric convex bodies, Ball \cite{B} proved that if the volumes of projections of $K$ are less than or equal to the volumes of projections of $L$ in every direction, then $|K|\leq \sqrt{n} |L|$, for every dimension $n$. Goodey and Zhang \cite{GZ} obtained a generalization of the Shephard problem for lower dimensional projections. A Fourier analytic approach to Shephard's problem was presented by Koldobsky, Ryabogin and Zvavitch \cite{KRZ}. Ryabogin and Zvavitch \cite{RZ} solved the generalization of Shephard's problem for Firey projections.
The Busemann-Petty problem is in a sense dual to the Shephard problem. It asks whether symmetric convex bodies with larger central hyperplane sections necessarily have greater volume. The Busemann-Petty problem has been solved affirmatively for $n\leq 4$ and negatively for $n\geq 5$ (see Gardner, Koldobsky, Schlumprecht \cite{GKS} and Zhang \cite{zhang}). The answer to Busemann-Petty problem is affirmative if and only if the body with larger sections is an intersection body (see Lutwak \cite{Lu} for the definition and properties of intersection bodies, and Koldobsky \cite{K6} for Fourier analytic approach to intersection bodies). Zvavitch solved an isomorphic version of Busemann-Petty problem for Gaussian measures \cite{zvav1}, and completely generalized the solution of Busemann-Petty problem to arbitrary measures with positive density \cite{zvav}. Koldobsky \cite{K8}, and further Koldobsky and Zvavitch \cite{KZ} obtained estimates for the isomorphic version of Busemann-Petty problem for arbitrary measures; a discrete analog of those estimates was very recently obtained by Alexander, Zvavitch, Henk \cite{AHZ}.
We refer the reader to the books by Koldobsky \cite{Kold} and Koldobsky, Yaskin \cite{KY} for a deep, yet accessible study of the Fourier-analytic approach to the Busemann-Petty and Shephard problems, as well as a general introduction to Fourier analysis in Convex geometry.
Aleksandrov in \cite{A2} proved that any symmetric convex body in $\mathbb R^n$ is determined uniquely by the $(n-1)-$dimensional volumes of its projections. See Zhang \cite{zhang} for the discrete version of that statement under natural assumptions. In Section 5 we generalize Aleksandrov's theorem to measures with positive degree of concavity and positive degree of homogeneity.
First, we find a natural analogue of the Lebesgue measure of projection of a convex body to other measures.
\begin{definition}\label{proj_gen}
Let $\mu$ be a measure on $\mathbb R^n$ with density $g$ continuous on its support, and let $K$ be a convex body. Consider a unit vector $\theta\in {{\mathbb S}^{n-1}}$. Define the following function on the cylinder ${{\mathbb S}^{n-1}}\times [0,1]$:
\begin{equation}\label{mu-proj-t}
p_{\mu, K}(\theta,t):=\frac{n}{2}\int_{{{\mathbb S}^{n-1}}} |\langle \theta, u\rangle| d\sigma_{\mu, tK}(u).
\end{equation}
We also consider $\mu-projection$ function on the unit sphere:
\begin{equation}\label{mu-proj}
P_{\mu, K}(\theta):=\int_0^1 p_{\mu, K}(\theta,t) dt.
\end{equation}
\end{definition}
In the particular case of Lebesgue measure $\lambda$ we have
$$P_{\lambda, K}(\theta)=|K|\theta^{\perp}|_{n-1}.$$
The Definition \ref{proj_gen} is natural since it is a generalization of Cauchy's projection formula (see below (\ref{Cauchy})). For even $g$, the notion of $p_{\mu,K}(\theta,t)$ can be understood geometrically as the projected weight of the boundary of $tK$, $t\in[0,1]$. More specifically, we define a measure $\mu_{tK}$ on $\theta^{\perp}$ to be the marginal measure of $1_{\partial (tK)}(x)g(x)dx$. In other words, for a measurable set $\Omega\subset \theta^{\perp}$, let $\mu_{tK}(\Omega)=\int_{\Omega} g(\pi_{tK}^{-1}(w)) dw$, where $\pi_{tK}^{-1}(w)$ is the full pre-image of $w$ under the projection of $tK$ onto $\theta^{\perp}$. Then
$$p_{\mu,K}(\theta, t)=\mu_{tK}(tK|\theta^{\perp})=\mu_{tK}(K|\theta^{\perp}),$$
where the last equality holds since $tK\subset K$. Hence,
$$P_{\mu,K}(\theta)=\int_0^1\mu_{tK}(K|\theta^{\perp})dt.$$
We prove the following result.
\begin{theorem}\label{main}
Fix $n\geq 1$; let $\mu$ on $\mathbb R^n$ be a measure and $g(x)$ be its even $r-homogenous$ density for some $r\geq 0$, and the restriction of $g$ to some half space is $p-concave$ for a $p\geq 0$.
Let $K$ and $L$ be symmetric convex bodies, and let $L$ additionally be a projection body. Assume that for every $\theta\in {{\mathbb S}^{n-1}}$ we have
$$P_{\mu, K}(\theta)\leq P_{\mu, L}(\theta).$$
Then $\mu(K)\leq \mu(L)$.
\end{theorem}
To compliment Theorem \ref{main} we prove the following.
\begin{theorem}\label{main-2}
Fix $n\geq 1$; let $\mu$ on $\mathbb R^n$ be a measure and $g(x)$ be its even $r-homogenous$ density for some $r\geq 0$, and the restriction of $g$ to some half space is $p-concave$ for a $p\geq 0$. Assume further that the closure of the support of $\mu$ is the whole space.
Let $L$ be a symmetric convex body which is not a projection body. Then there exists a symmetric convex body $K$ such that for every $\theta\in {{\mathbb S}^{n-1}}$ we have
$$P_{\mu, K}(\theta)\leq P_{\mu, L}(\theta),$$
but $\mu(K)> \mu(L)$.
\end{theorem}
We remark that in the case of Lebesgue measure Theorems \ref{main} and \ref{main-2} are generalizations of the well-known solution to the classical Shephard problem (see Koldobsky \cite{Kold}, Chapter 8).
This paper is organized as follows. In Section 2 we present the preliminaries on the subject. In Section 3 we introduce and study the notion of mixed measure and prove an analogue of Minkowski's first inequality for measures. In Section 4 we prove Theorem \ref{mink-existence-intro}. In Section 5 we prove two types of uniqueness results: one is the extension of Aleksandrov's theorem, and the other is related to the uniqueness of the solution of certain PDE in the class of support functions. In Section 6 we prove Theorems \ref{main} and \ref{main-2}. In Section 7 we discuss stability and separation results for Theorem \ref{main}, and their corollaries.
\textbf{Acknowledgement.} The author would like to thank Alex Koldobsky, Artem Zvavitch, Liran Rotem and Ben Jaye for very fruitful discussions and encouragement.
\section{Preliminaries}
\subsection{Brunn-Minkowski theory.}
Below we present classical concepts and results of Convex geometry and Brunn-Minkowski theory. We refer the reader to books by Ball \cite{Ball-book}, Milman, Schechtman \cite{MS}, Schneider \cite{book4} for a detailed introduction to the subject.
Standard Minkowski's addition for sets $A, B\subset \mathbb R^n$ is defined as
$$A+B:=\{a+b\,:\,a\in A, b\in B\}.$$
Scalar multiplication for $\alpha\in \mathbb R$ and a set $A\subset \mathbb R^n$ is defined as
$$\alpha A:=\{\alpha a\,:\, a\in A\}.$$
For Borel sets $A$, $B$ in $\mathbb R^n$ and for arbitrary $\lambda\in [0,1]$, Brunn-Minkowski inequality states that
$$|\lambda A+(1-\lambda) B|^{\frac{1}{n}}\geq \lambda |A|^{\frac{1}{n}}+(1-\lambda) |B|^{\frac{1}{n}}.$$
See Gardner \cite{gardner} for an exhaustive survey on the subject. We remark that for convex bodies the equality in the Brunn-Minkowski inequality is attained if and only if the sets $A$ and $B$ are closed, convex dilates of each other.
First mixed volume of convex bodies $K$ and $L$ in $\mathbb R^n$ is defined as follows:
$$V_1(K,L):=\frac{1}{n}\liminf_{\epsilon\rightarrow 0} \frac{|K+\epsilon L|-|K|}{\epsilon}.$$
Note that for any convex body $K$ one has
\begin{equation}\label{1mink-volume}
V_1(K,K)=|K|.
\end{equation}
Brunn-Minkowski inequality implies Minkowski's first inequality:
\begin{equation}\label{1mink}
V_1(K,L)\geq |K|^{\frac{n-1}{n}} |L|^{\frac{1}{n}}.
\end{equation}
There is equality in Minkowski's first inequality if and only if $K$ and $L$ are closed convex dilates of each other (see Schneider \cite{book4} for more details).
A particular case of mixed volume, is the surface area of a convex set $K$ in $\mathbb R^n$:
$$|\partial K|^+:=nV_1(K,B_2^n)=\liminf_{\epsilon\rightarrow 0} \frac{|K+\epsilon B_2^n|-|K|}{\epsilon}.$$
Therefore, (\ref{1mink}) implies classical isoperimetric inequality:
$$\frac{|\partial K|^+}{|K|^{\frac{n-1}{n}}}\geq \frac{|\partial B_2^n|^+}{|B_2^n|^{\frac{n-1}{n}}}.$$
Next, we shall discuss Brunn-Minkowski inequality for $p-$concave measures (see Gardner \cite{gardner} for more details). For $p\in\mathbb R$ and for $a,b\geq 0$, $\lambda\in[0,1]$ we define a $p-$average as follows:
\begin{equation}\label{p-av}
M_p(a,b,\lambda)=\left(\lambda a^p+(1-\lambda)b^p\right)^{\frac{1}{p}}.
\end{equation}
In the special cases $p=0,$ $p=+\infty$ and $p=-\infty$ we have
$$M_0(a,b,\lambda)=a^{\lambda} b^{1-\lambda},$$
$$M_{-\infty}(a,b,\lambda)=\min(a,b),$$
$$M_{+\infty}(a,b,\lambda)=\max(a,b).$$
We say that a function $g:\mathbb R^n\rightarrow \mathbb R^+$ is $p-$concave if for every $x,y\in \mathbb R^n$ such that $g(x)g(y)>0,$ and for every $\lambda \in [0,1]$ one has
$$g(\lambda x+(1-\lambda)y)\geq M_p(g(x),g(y),\lambda).$$
We remark that $0-$concave functions are also called log-concave.
The following generalized Brunn-Minkowski inequality is well known (see e.g. Borell \cite{bor}, Gardner \cite{gardner}). Let $p\in [-\frac{1}{n},+\infty]$, and let $\mu$ be a measure on $\mathbb R^n$ with $p-$concave density $g.$ Let $$q=\frac{p}{np+1}.$$
Then the measure $\mu$ is $q-$concave on $\mathbb R^n$. That is, for every pair of Borel sets $A$ and $B$ and for every $\lambda\in[0,1]$ one has
\begin{equation}\label{genbm}
\mu(\lambda A+(1-\lambda) B)\geq M_q(\mu(A),\mu(B),\lambda).
\end{equation}
\subsection{The surface area measure, its properties and applications.}
Support hyperplane of a convex body $K$ at a point $y\in \partial K$ is a hyperplane which contains $y$ and does not contain any of the interior points of $K$. By convexity, such hyperplane exists at every point $y\in \partial K$, and is unique almost everywhere with respect to the $(n-1)$-dimensional Hausdorff measure on $\partial K$. The vector orthogonal to a support hyperplane at $y\in\partial K$ is called \emph{normal vector} at $y$; if such vector is unique it shall be denoted $n_y.$ The Gauss map $\nu_K:\partial K\rightarrow {{\mathbb S}^{n-1}}$ corresponds $y\in \partial K$ to the set of its normal vectors.
The push forward of the $(n-1)-$dimensional Hausdorff measure on $\partial K$ under the Gauss map $\nu_K$ to the sphere is called surface area measure of $K$ and is denoted by $\sigma_K.$
In particular, $|\partial K|^+$ (the surface area of $K$) can be found as
$$|\partial K|^+=\int_{{{\mathbb S}^{n-1}}} d\sigma_K(u).$$
A class of strictly convex bodies whose support function is twice continuously differentiable we shall denote by $C^{2,+}$ (strict convexity means that the interior of every interval connecting a pair of points in the body is fully contained in the interior of the body). For such bodies, the Gauss map is a bijection, and the surface area measure $\sigma_K$ has a continuous density $f_K(u)$, which is called curvature function of $K.$
One can see via approximation by polytopes, that
$$\int_{{{\mathbb S}^{n-1}}} u d\sigma_K(u)=0.$$
Conversely, the following Minkowski's existence Theorem holds (see e. g. Schneider \cite{book4} or Koldobsky \cite{Kold}).
\begin{theorem}[Minkowski]\label{Minkowski}
Let $\mu$ be a measure on the sphere, not supported on any subspace, and such that
$$\int_{{{\mathbb S}^{n-1}}} u d\mu(u)=0.$$
Then there exists a unique convex body $K$ so that $d\sigma_K(u)=d\mu(u)$ for all $u\in{{\mathbb S}^{n-1}}.$
\end{theorem}
We refer the reader to Schneider \cite{book4} for an accessible proof of Minkowski's existence theorem, and to Pogorelov \cite{Pog} for a detailed survey on the differential geometric approach to Minkowski's existence theorem, its strengthening and related results.
The support function $h_K$ of a convex body $K$, containing the origin, is defined on $\mathbb R^n$ via
$$h_K(x)=\max_{y\in K} \langle x,y\rangle.$$
Geometrically, for a unit vector $\theta$, the value of $h_K(\theta)$ represents distance to the support hyperplane of $K$ in the direction $\theta.$ Due to the fact that $h_K$ is 1-homogenous, one has
\begin{equation}\label{gradprop}
\langle \nabla h_K(u),u\rangle=h_K(u),
\end{equation}
for every $u\in {{\mathbb S}^{n-1}}$, provided that $\nabla h_K(u)$ is well-defined. In this case, $\nabla h_K(n_y)=y$ for all $y\in\partial K.$
We state a formula for a volume of a convex body $K$ with surface area measure $\sigma_K$:
\begin{equation}\label{volume}
|K|=\frac{1}{n}\int_{{{\mathbb S}^{n-1}}} h_K(u)d\sigma_K(u).
\end{equation}
The validity of this formula can be seen in the case when $K$ is a polytope and the general case follows by approximation. Moreover, for arbitrary convex bodies $K$ and $L$ one has the following:
\begin{equation}\label{mixed_volume}
V_1(K,L)=\frac{1}{n}\int_{{{\mathbb S}^{n-1}}} h_L(u)d\sigma_K(u).
\end{equation}
Another formula involving surface area measure is the so called Cauchy projection formula:
\begin{equation}\label{Cauchy}
|K|\theta^{\perp}|_{n-1}=\frac{1}{2}\int_{{{\mathbb S}^{n-1}}} |\langle u, \theta \rangle| d\sigma_K(u),
\end{equation}
where $\theta$ is an arbitrary unit vector, and $K$ is a convex symmetric body. The validity of (\ref{Cauchy}), once again, can be seen for polytopes and it follows by approximation for arbitrary convex bodies. See Koldobsky \cite{Kold} for more details about (\ref{volume}), (\ref{mixed_volume}) and (\ref{Cauchy}).
\subsection{Fourier transform on ${{\mathbb S}^{n-1}}$ and its applications to Convex geometry.}
Fourier transform in Convexity plays a very important role. See books by Koldobsky \cite{Kold}, Koldobsky, Yaskin \cite{KY}, and a survey by Koldobsky, Ryabogin, Zvavitch \cite{KRZ-2} for a detailed introduction to the subject.
The Schwartz class $\bf{S}$ is the space of complex valued rapidly decreasing infinitely differentiable functions on $\mathbb R^n$. Every locally integrable real valued function $f$ on $\mathbb R^n$ with power growth at infinity represents a distribution acting by integration:
$$\langle f,\varphi\rangle=\int_{\mathbb R^n}f(x)\varphi(x)dx,$$
for $\varphi\in\bf{S}$.
The Fourier transform $\widehat{f}$ of a distribution $f$ is defined by
$$\langle\widehat{f},\widehat{\varphi}\rangle= (2\pi)^n\langle f,\varphi\rangle,$$
for every test function $\varphi\in\bf{S}$.
Let $\mu$ be a finite Borel measure on the unit sphere ${{\mathbb S}^{n-1}}$. Let $\mu_e$ be a $-(n+1)$-homogenous extension of $\mu$ to $\mathbb R^n.$ $\mu_e$ is called the extended measure of $\mu$ if for every $\varphi\in\bf{S},$
$$\langle \mu_e, \varphi \rangle=\frac{1}{2}\int_{{{\mathbb S}^{n-1}}} \langle r^{-2},\varphi(ru)\rangle d\mu(u).$$
The following geometric representation of Fourier transform on the sphere was proved by Koldobsky, Ryabogin, Zvavitch \cite{KRZ} (see also Koldobsky \cite{Kold}):
\begin{equation}\label{Fourier_scalar}
\widehat{\mu_e}(\theta)=-\frac{\pi}{2}\int_{{{\mathbb S}^{n-1}}} |\langle u, \theta\rangle|d\mu(u),
\end{equation}
for every $\theta\in {{\mathbb S}^{n-1}}.$
Note that (\ref{Cauchy}) and (\ref{Fourier_scalar}) impy that
\begin{equation}\label{Fourier_curvature}
d\widehat{\sigma_K}(\theta)=-\pi|K|\theta^{\perp}|d\theta,
\end{equation}
where $\sigma_K$ is the surface area measure of a symmetric convex body $K$, extended to $\mathbb R^n$ with degree of homogeneity $-(n+1)$.
The following Parseval-type identity was proved by Koldobsky, Ryabogin, Zvavitch \cite{KRZ} (see also Koldobsky \cite{K98}, \cite{Kold}): for symmetric convex bodies $K, L$, so that the support function of $K$ is infinitely smooth,
\begin{equation}\label{parseval}
\int_{{{\mathbb S}^{n-1}}}\widehat{h_K}(\theta)\widehat{f_L}(\theta)=(2\pi)^n\int_{{{\mathbb S}^{n-1}}}h_K(\theta)f_L(\theta),
\end{equation}
where the Fourier transform of $h_K$ is considered with respect to its $1-$homogenous extension, and the Fourier transform of $f_L$ is considered with respect to its $-(n+1)-$homogenous extension.
By Minkowski's existence Theorem, for every symmetric convex body $L$ and for every even density $g$, not supported on a great subsphere, there exists a symmetric convex body $\tilde{L}$ such that
$$\sigma_{\mu,L}=\sigma_{\tilde{L}}.$$
Therefore, for all infinitely smooth symmetric convex bodies $K, L$ in $\mathbb R^n$, and for every even density $g$ continuous on its support, one has
\begin{equation}\label{parseval_measures}
\int_{{{\mathbb S}^{n-1}}}\widehat{h_K}(\theta)d\widehat{\sigma_{\mu,L}}(\theta)=(2\pi)^n\int_{{{\mathbb S}^{n-1}}}h_K(\theta)d\sigma_{\mu,L}(\theta),
\end{equation}
where the Fourier transform of $h_K$ is considered with respect to its $1-$homogenous extension, and the Fourier transform of $\sigma_{\mu,L}$ is considered with respect to its $-(n+1)-$homogenous extension.
Another observation is that (\ref{Fourier_scalar}) implies:
\begin{equation}\label{Fourier_scalar_flghl}
d\widehat{\sigma_{\mu,L}}(\theta)=-\frac{\pi}{2}\int_{{{\mathbb S}^{n-1}}} |\langle u, \theta\rangle| d\sigma_{\mu,L}(u),
\end{equation}
where the Fourier transform of $\sigma_{\mu,L}$ is considered with respect to its $-(n+1)-$homogenous extension.
In particular, considering $tL$ in place of $L$ we get
\begin{equation}\label{Fourier_p_t}
\widehat{\sigma_{\mu,tL}}(\theta)=-\frac{\pi}{n}p_{\mu,L}(\theta,t),
\end{equation}
and
\begin{equation}\label{Fourier_p}
\widehat{\int_0^1 \sigma_{\mu,tL}(\theta)dt}=-\frac{\pi}{n}P_{\mu,L}(\theta).
\end{equation}
\begin{remark}
The degree of homogeneity with which a function on the sphere is extended to $\mathbb R^n$ impacts radically its Fourier transform, and, in particular, the restriction of its Fourier transform back to the unit sphere (see more in Goodey, Yaskin, Yaskina \cite{GYY}.) We would like to emphasize the fact that the homogeneity properties of the measure $\mu$ on $\mathbb R^n$ are completely irrelevant to the study of Fourier transforms of $h_K$ and $\sigma_{\mu,K}$. In fact, we always extend $h_K$ and $\sigma_{\mu,K}$ in the most convenient way, after having already translated all the information about the underlying measure $\mu$ onto the sphere. The proof of Theorem \ref{main}, much like the classical Shephard's problem (see \cite{KRZ}), consists of gluing together Fourier transform and Brunn-Minkowski theory; the part which involves Fourier transform works for arbitrary measures, while the Brunn-Minkowski part is what reinforces the assumptions of concavity and homogeneity on the density of $\mu.$
\end{remark}
\subsection{Projection bodies}
Let K be an origin symmetric convex body in $\mathbb R^n$ with curvature function $f_K$. The projection body $\Pi K$ of $K$ is defined as the origin symmetric convex body in $\mathbb R^n$ whose support function in every direction is equal to the volume of the orthogonal projection of K in this direction. We extend $h_{\Pi K}$ to a homogeneous function of degree $1$ on $\mathbb R^n$. By
(\ref{Fourier_curvature}),
$$
h_{\Pi K}(\theta)=-\frac{1}{\pi} \widehat{f_K}(\theta).
$$
The curvature function of a convex body is non-negative. Therefore, $\widehat{h_{\Pi K}}\leq 0$. On the other hand, by Minkowski's existence theorem, an origin symmetric convex body $K$ in $\mathbb R^n$ is the projection body of some origin symmetric convex body if and only if there exists a measure $\mu$ on ${{\mathbb S}^{n-1}}$ so that
$$\widehat{h_K}=-\mu_e.$$
The condition that $L$ is a projection body is equivalent to $L$ being a centered zonoid (see Gardner \cite{G}). Zonoids are characterized as polar bodies of unit balls of finite dimensional sections of $L_1$.
Every origin symmetric convex body on the plane is a projection body (see Herz \cite{He}, Ferguson \cite{Fe}, Lindenstrauss \cite{Li}). It was proved by Koldobsky \cite{K1} that $p-$balls in $\mathbb R^n$ for $n\geq 3$ and $p\in [1,2]$ are not projection bodies.
\section{Mixed measures and related results}
\subsection{Mixed measures}
As an analogue of the classical mixed volume consider the following notion.
\begin{definition}\label{mixed measure}
Given sets $K$ and $L$, we define their \textbf{mixed $\mu-$measure} as follows.
$$\mu_1(K,L)=\liminf_{\epsilon\rightarrow 0} \frac{\mu(K+\epsilon L)-\mu(K)}{\epsilon}.$$
\end{definition}
We observe that in the absence of homogeneity of $\mu$, the mixed measure $\mu_1(K,L)$ is not homogenous in $K$. However, it is necessarily homogenous in $L$:
$$\mu_1(K, sL)=s\mu_1(K,L).$$
If, additionally, the measure $\mu$ is $\alpha-$homogenous, i.e.
$$\mu(tA)=t^{\alpha}\mu(A)$$
for all $t\in \mathbb R^+$ and Borel sets $A,$ then
$$\mu_1(tK, L)=t^{\alpha-1}\mu_1(K,L).$$
\begin{definition}\label{mixed measure-volume}
We also introduce the following analogue of mixed volume:
$$V_{\mu,1}(K,L)=\int_0^1 \mu_1(tK,L) dt.$$
\end{definition}
Note that in the case of the Lebesgue measure $\lambda$ we have
$$V_{\lambda,1}(K,L)=V_1(K,L).$$
Definition \ref{mixed measure} implies that for $t\in (0,\infty)$,
\begin{equation}\label{measure-derivative}
\mu_1(tK,K)=\mu(tK)'_t;
\end{equation}
this derivative exists by monotonicity. Therefore,
\begin{equation}\label{measure-derivative-integral}
V_{\mu,1}(K,K)=\int_0^1 \mu_1(tK,K) dt=\int_0^1 \mu(tK)' dt=\mu(tK)|_0^1=\mu(K).
\end{equation}
Recall that we use the notation $\sigma_{\mu,K}$ for a surface area measure of a convex body $K$ with respect to a measure $\mu$ on $\mathbb R^n$. That is, for a Borel set $A\subset {{\mathbb S}^{n-1}},$
$$\sigma_{\mu,K}(A)=\int_{\nu_K^{-1}(A)} g(x) dH_{n-1}(x),$$
where $dH_{n-1}(x)$ stands for the $(n-1)-$dimensional Hausdorff measure on $\partial K$. Following the idea from the appendix of \cite{L}, we prove the following representation for $\mu_1(K,L)$.
\begin{lemma}\label{formula-mixed-measure}
Given convex bodies $K$ and $L$ containing the origin, and a measure $\mu$ with continuous density $g$ on $\mathbb R^n$, we have
$$\mu_1(K,L)=\int_{{{\mathbb S}^{n-1}}} h_L(u) d\sigma_{\mu,K}(u).$$
Here $h_K$ and $h_L$ are support functions of $K$ and $L$ and $\sigma_{\mu,K}$ is the surface area measure of $K$.
\end{lemma}
The proof is outlined in the Appendix (see Lemma \ref{jacapp}).
In order to provide some intuition about $\sigma_{\mu,K}$, we describe it explicitly in a couple of partial cases.
\begin{proposition}\label{sam_smooth}
If a body $K$ is $C^2-$smooth and strictly convex then its surface area measure has representation
$$d\sigma_{\mu,K}(u)=f_K(u)g(\nabla h_K(u))du.$$
\end{proposition}
\begin{proposition}\label{sam_pol}
The surface area measure of a convex polytope $P$ with respect to a measure $\mu$ has representation
$$d\sigma_{\mu,P}(u)=\sum_{i=1}^N \delta_{u_i} \mu_{n-1}(F_i)du,$$
where $u_i$, $i=1,...,N$ are the normals to the faces of the polytope, $F_i$ are the corresponding faces, and $\mu_{n-1}(F_i)$ stands for $\int_{F_i} g(x)dx$.
\end{proposition}
See the Appendix for the proofs of Propositions \ref{sam_smooth} (Proposition \ref{1ref}) and \ref{sam_pol} (Proposition \ref{2ref}).
We remark that Lemma \ref{formula-mixed-measure}, Proposition \ref{sam_smooth}, along with (\ref{Fourier_p_t}) and (\ref{parseval}) imply for all symmetric convex infinitely smooth bodies $K$ and $L$:
$$\mu_1(tK,L)=(2\pi)^{-n}\int_{{{\mathbb S}^{n-1}}} \widehat{h_L}(u) d\widehat{\sigma_{\mu,tK}}(u) du=$$
\begin{equation}\label{mu1-dual}
-\frac{\pi}{n}(2\pi)^{-n}\int_{{{\mathbb S}^{n-1}}} \widehat{h_L}(u) p_{\mu,K}(t,u) du.
\end{equation}
As an immediate corollary of Lemma \ref{formula-mixed-measure} and (\ref{measure-derivative-integral}) we derive the following expression of the measure of a $C^{2,+}$ convex body (see also \cite{CLM}).
\begin{lemma}\label{formula} Let $\mu$ be a measure with continuous density $g$. Let $K$ be a $C^{2,+}$ convex body with support function $h_K$ and curvature function $f_K$. Then
\begin{equation}\label{volume formula2}
\mu(K)=\int_{\mathbb S^{n-1}} h_K(u)f_K(u)\int_0^{1} t^{n-1} g\left(t\nabla h_K(u)\right)dt du.
\end{equation}
\end{lemma}
We outline that if the density of a measure $\mu$ on $\mathbb R^n$ is $r-$homogenous, then
\begin{equation}\label{outl}
\mu(K)=\int_0^1 \mu_1(tK,K)dt= \mu_1(K,K)\int_0^1 t^{n+r-1}dt=\frac{1}{n+r}\mu_1(K,K).
\end{equation}
In view of (\ref{outl}), Lemma \ref{formula-mixed-measure} and Proposition \ref{sam_pol} imply the following.
\begin{proposition}\label{poly-vol}
Let $\mu$ be a measure with $r-$homogenous density $g(x)$ on $\mathbb R^n$, and consider a polytope with $N$ faces:
$$P=\{x\in\mathbb R^n:\,\,\langle x,u_i\rangle\leq \alpha_i\},$$
where $u_i\in{{\mathbb S}^{n-1}}$ and $\alpha_i>0,$ $i=1,...,N.$ Let $F_i$ be faces of $P$ orthogonal to $u_i$, $i=1,...,N.$ Then
$$\mu(P)=\frac{1}{n+r}\sum_{i=1}^N \alpha_i \mu_{n-1}(F_i),$$
where $\mu_{n-1}(F_i)$ stands for $\int_{F_i} g(x)dx$.
\end{proposition}
\subsection{Minkowski's first inequality generalized}
The main result of this subsection is the following theorem.
\begin{theorem}\label{mink_gen}
Let $\mu$ on $\mathbb R^n$ be a measure. Assume that $\mu$ is $F(t)-$concave, i.e. there exists a differentiable invertible function $F:\mathbb R^+\rightarrow \mathbb R$ such that for every $\lambda\in [0,1]$ and for every pair of Borel sets $K$ and $L$ in a certain class, we have
\begin{equation}\label{condbm}
\mu(\lambda K+(1-\lambda) L)\geq F^{-1}\left(\lambda F(\mu(K))+(1-\lambda)F(\mu(L))\right).
\end{equation}
Then the following holds:
\begin{equation}\label{condbm-inf}
\mu_1(K,L)\geq \mu_1(K,K)+\frac{F(\mu(L))-F(\mu(K))}{F'(\mu(K))},
\end{equation}
for all $K,$ $L$ in that class.
\end{theorem}
\begin{proof} We write
$$\mu(K+\epsilon L)=\mu\left((1-\epsilon)\frac{K}{1-\epsilon}+\epsilon L\right)\geq$$
$$F^{-1}\left((1-\epsilon) F\left(\mu(\frac{K}{1-\epsilon})\right)+\epsilon F\left(\mu(L)\right)\right)=:G_{K,L,\mu,F}(\epsilon).$$
Note that $G_{K,L,\mu,F}(0)=\mu(K)$. Therefore,
$$\mu_1(K,L)\geq G_{K,L,\mu,F}'(0).$$
We note that
$$\mu\left(\frac{K}{1-\epsilon}\right)'|_{\epsilon=0}=\mu_1(K,K).$$
Using the above along with standard rules of differentiation, such as
$$(F^{-1}(a))'=\frac{1}{F'(F^{-1}(a))},$$
we get the statement of the Theorem.
\end{proof}
A standard argument implies that the equality cases of the inequality (\ref{condbm-inf}) coincide with equality cases of (\ref{condbm}). We shall formulate a few corollaries of Theorem \ref{mink_gen} in some special cases.
\begin{corollary}\label{p-concave}
Let $p\geq 0$. Let $g:\mathbb R^n\rightarrow \mathbb R^+$ be a p-concave density of measure $\mu$, continuous on its support. Let $q=\frac{1}{n+\frac{1}{p}}$. Then for every pair of Borel sets $K$ and $L$ we have
$$\mu_1(K,L)\geq \mu_1(K,K)+\frac{\mu(L)^q-\mu(K)^q}{q\mu(K)^{q-1}}.$$
\end{corollary}
The corollary \ref{p-concave} follows from Theorem \ref{mink_gen} via considering $F(t)=t^q$. We also obtain the following nicer-looking corollary for measures with $p-$concave and $\frac{1}{p}-$homogenous densities. It was originally proved by E. Milman and L. Rotem \cite{MilRot}.
\begin{corollary}[E. Milman, L. Rotem]\label{p-concave-homo}
Let $p\geq 0$. Let $g:\mathbb R^n\rightarrow \mathbb R^+$ be a $p$-concave $\frac{1}{p}-$homogenous density of measure $\mu$. Let $q=\frac{1}{n+\frac{1}{p}}$. Then for every pair of Borel sets $K$ and $L$ we have
\begin{equation}\label{mink-gen-1}
\mu_1(K,L)\geq \frac{1}{q}\mu(K)^{1-q}\mu(L)^q,
\end{equation}
and
\begin{equation}\label{mink-gen-2}
V_{\mu,1}(K,L)\geq \mu(K)^{1-q}\mu(L)^q.
\end{equation}
\end{corollary}
\begin{proof} Note that if $g$ is $\frac{1}{p}-$homogenous then $\mu$ is an $(n+\frac{1}{p})=\frac{1}{q}-$homogenous measure. Therefore,
\begin{equation}\label{homo}
V_{\mu,1}(K,L)=\int_0^1 \mu_1(tK,L)dt=\mu_1(K,L)\int_0^1
t^{\frac{1}{q}-1} dt=q\mu_1(K,L),
\end{equation}
and in particular
\begin{equation}\label{homo-K}
\mu(K)=q\mu_1(K,K)
\end{equation}
Corollary \ref{p-concave} together with (\ref{homo-K}) implies (\ref{mink-gen-1}). Also, (\ref{mink-gen-1}) together with (\ref{homo}) implies (\ref{mink-gen-2}).
\end{proof}
Recall that a measure $\mu$ is called log-concave if for all Borel sets $K$ and $L$,
$$\mu(\lambda K + (1-\lambda)L)\geq \mu(K)^{\lambda}\mu(L)^{1-\lambda}.$$
Applying Theorem \ref{mink_gen} with $F(t)=\log t$ (as $\log t$ is an increasing function), we get the following corollary.
\begin{corollary}\label{log-concave}
Let measure $\mu$ be log-concave. Then for every pair of Borel sets $K$ and $L$ we have
$$\mu_1(K,L)\geq \mu_1(K,K)+\mu(K)\log\frac{\mu(L)}{\mu(K)}.$$
\end{corollary}
In particular, the following isoperimetric-type result follows from Theorem \ref{mink_gen}.
\begin{proposition}\label{isoper}
Let a measure $\mu$ be log-concave. Then for every pair of Borel sets $K$ and $L$ such that $\mu(K)=\mu(L)$, one has
$$\mu_1(K,L)\geq \mu_1(K,K).$$
\end{proposition}
For example, if $\gamma$ is the standard Gaussian measure $\gamma$ (that is, the measure with density $\frac{1}{\sqrt{2\pi}^n} e^{-\frac{|x|^2}{2}}$), and $K$ is a convex set containing the origin, then the expression
$$\int_{\partial K} \langle y, \nu_L(y)\rangle e^{-\frac{|y|^2}{2}} d\sigma(y)$$
is minimized when $L=K,$ where $L$ is such convex region that $\gamma(K)=\gamma(L)$, and $\nu_L$ is it Gauss map.
Another strengthening of Corollary \ref{log-concave} in the case of the standard Gaussian measure is possible to obtain using Ehrhard's inequality (see Ehrhard \cite{E}, Borell \cite{bor-erch}). Recall the notation
$$\psi(a)=\frac{1}{\sqrt{2\pi}}\int_{-\infty}^a e^{-\frac{t^2}{2}} dt.$$
It was shown by Ehrhard (for convex sets), and further extended by Borell, that for every pair of Borel sets $K$ and $L$ and for every $\lambda\in [0,1]$ we have
$$\psi^{-1}\left(\gamma(\lambda K+(1-\lambda) L)\right)\geq \lambda \psi^{-1}(\gamma(K))+(1-\lambda)\psi^{-1}(\gamma(L)).$$
Hence the next Corollary follows.
\begin{corollary}\label{gaussian}
For the standard Gaussian measure $\gamma$ and for every pair of convex sets $K$ and $L$ we have
$$\gamma_1(K,L)\geq \gamma_1(K,K)+e^{-\frac{\psi^{-1}(\gamma(K))^2}{2}}\left(\psi^{-1}(\gamma(L))-\psi^{-1}(\gamma(K))\right).$$
\end{corollary}
To obtain this corollary we use the fact that $\psi$ is an increasing function and the relation
$$\psi^{-1}(a)'=e^{\frac{\psi^{-1}(a)^2}{2}}.$$
\section{Extension of the Minkowski's existence theorem.}
This section is dedicated to proving an extension of Minkowski's existence theorem. We use ideas from the proof of the classical Minkowski's existence theorem (see Schneider \cite{book4}).
First, we state a definition.
\begin{definition}
For a measure $\mu$ on $\mathbb R^n$, we say that a convex body $K$ in $\mathbb R^n$ with particular properties is \textbf{$\mu-$unique} if every pair of convex bodies with said properties coincides up to a set of $\mu$-measure zero.
\end{definition}
\begin{theorem}\label{mink-existence}
Let $\mu$ on $\mathbb R^n$ be a measure and $g(x)$ be its even $r-homogenous$, continuous on its support density for some $r\geq 0$, such that a restriction of $g$ on some half space is $p-concave$ for $p\geq 0$. Let $\varphi$ be an arbitrary even measure on ${{\mathbb S}^{n-1}}$, not supported on any great subsphere, such that $supp(\varphi)\subset int(supp(g))\cap {{\mathbb S}^{n-1}}$. Then there exists a $\mu-$unique convex body $K$ in $\mathbb R^n$ such that
$$d\sigma_{K,\mu}(u)=d\varphi(u).$$
\end{theorem}
The existence part of Theorem \ref{mink-existence} follows by approximation from the lemma below. We remark, that for an $(n-1)-$dimensional surface $F$, the notation $\mu_{n-1}(F)$ stands for
$$\mu_{n-1}(F)=\int_F g(x)dx,$$
where $g(x)$ is the density of $\mu,$ and $dx$ is the area element on $F.$
\begin{lemma}\label{mink-existence-polytopes}
Let $\mu$ on $\mathbb R^n$ be a measure and $g(x)$ be its even $r-homogenous$ continuous on its support density for some $r>-n$. Let $N\geq 2n$ be an even integer. Let $u_1,...,u_N$ be unit vectors spanning the $\mathbb R^n$, $u_i\in int(supp(g))$, such that $u_i=-u_{\frac{N}{2}+i}$. Let $f_1,...,f_N$ be arbitrary positive numbers such that $f_i=f_{\frac{N}{2}+i}$.
Then there exist positive $\alpha_1,...,\alpha_N$ such that the convex polytope
$$P=\cap _{i=1}^N \{|\langle x,u_i\rangle|\leq \alpha_i\}$$
with faces $F(u_1),...,F(u_N)$ satisfies
$$\mu_{n-1}(F(u_i))=f_i.$$
Moreover, if restriction of $g$ on a half space is $p-concave$ for $p\geq 0$ then such polytope $P$ is $\mu-$unique.
\end{lemma}
\begin{proof} For a vector $A=(\alpha_1,...,\alpha_N)\in\mathbb R^N$ we shall consider a polytope
$$P(A)=\cap_{i=1}^N\{x\in\mathbb R^n\,:\,|\langle x,u_i\rangle| \leq \alpha_i\}.$$
Consider a set $M\subset \mathbb R^N$ defined as follows:
$$M:=\{A\in \mathbb R^N\,:\, \mu(P(A))\geq 1\}.$$
Note that $M\subset \{A:\,\, \alpha_i\geq 0\,\,\forall i=1,...,N\}$. It is nonempty since the measure is unbounded. As the set $M$ is closed, and $f_i>0,$ the linear functional
$$\varphi(A)=\frac{1}{n+r}\sum_{i=1}^N f_i \alpha_i$$
attains its minimum on $M.$ Let $A^*=(\alpha_1^*,...,\alpha_N^*)$ be the minimizing point, $P^*=P(A^*)$, and let $F_i^*$ stand for the facet of $P^*$ orthogonal to $u_i$. Denote the value of the minimum $\varphi(A^*)=m^{n+r-1}$.
We show that $m P^*$ is the polytope which solves the problem. Indeed, consider hyperplanes
$$H_1=\{A\in \mathbb R^N\,:\, \frac{1}{n+r}\sum_{i=1}^N f_i \alpha_i=m^{n+r-1}\},$$
$$H_2=\{A\in \mathbb R^N\,:\, \frac{1}{n+r}\sum_{i=1}^N \mu_{n-1}(F_i^*) \alpha_i=1\}.$$
Note that all $\alpha_i^*>0$. Thus, by Proposition \ref{poly-vol},
$$\mu(P^*)=\frac{1}{n+r}\sum_{i=1}^N \alpha_i^*\mu_{n-1}(F_i^*).$$
On the other hand, the linear functional $\varphi$ attains its minimum on the boundary of $M$, and hence
\begin{equation}\label{mu^*}
\mu(P^*)=1.
\end{equation}
We conclude that $A^*\in H_1\cap H_2.$
Observe that $H_1\cap int(M)=\varnothing$, as otherwise $A^*$ would not be the minimum. Consider a vector $A\in H_1$ different from $A^*$. For any $\lambda\in [0,1]$, the vector $\lambda A^*+(1-\lambda) A\in H_1$, and hence
$$\mu(P(\lambda A^*+(1-\lambda) A))\leq 1.$$
Note also that
$$\lambda P(A^*)+(1-\lambda) P(A)\subset P(\lambda A^*+(1-\lambda) A),$$
and thus
\begin{equation}\label{muleq}
\mu(\lambda P^*+(1-\lambda) P(A))\leq 1.
\end{equation}
Therefore, by homogeneity of $\mu$, (\ref{mu^*}) and (\ref{muleq}),
$$\mu_1(P^*,P(A))=\liminf_{\epsilon\rightarrow 0}\frac{\mu(P^*+\epsilon P(A))-\mu(P^*)}{\epsilon}=$$
$$\liminf_{\epsilon\rightarrow 0}\frac{(1+\epsilon)^{n+r}\mu(\frac{1}{1+\epsilon}P^*+\frac{\epsilon}{1+\epsilon} P(A))-1}{\epsilon}\leq \liminf_{\epsilon\rightarrow 0}\frac{(1+\epsilon)^{n+r}-1}{\epsilon}=n+r.$$
On the other hand, if $\alpha_i>0$, by Proposition \ref{sam_pol} and Lemma \ref{formula-mixed-measure} we have
$$\mu_1(P^*,P(A))=\sum_{i=1}^N \alpha_i \mu_{n-1}(F^*_i),$$
and hence
\begin{equation}\label{muleq-1}
\frac{1}{n+r}\sum_{i=1}^N \alpha_i \mu_{n-1}(F_i^*)\leq 1.
\end{equation}
Therefore, there exists an open subset of $H_1$,
$$U:=H_1\cap \{A\in\mathbb R^N\,:\,\alpha_i>0\},$$
which is fully contained in the half space
$$H_2^-=\{A\in \mathbb R^N\,:\, \frac{1}{n+r}\sum_{i=1}^N \mu_{n-1}(F_i^*) \alpha_i\leq1\},$$
and, in addition, the interior of $U$ contains $A^*\in H_1\cap H_2$. This implies that $H_1=H_2.$
Therefore,
$$\mu_{n-1}(F_i^*)m^{n+r-1}=f_i.$$
Using homogeneity of $g$ once again, we conclude that the polytope
$$m P^*=\cap_{i=1}^N \{x\in \mathbb R^n\,:\, \langle x,u_i\rangle \leq \beta_i\},$$
with $\beta_i=m \alpha_i^*$, satisfies the conclusion of the Lemma.
The uniqueness part follows in the same manner as in subsection \ref{un} for all convex bodies, therefore we skip the argument here.
\end{proof}
We remark that no concavity was necessary to prove the existence part for polytopes; however, it is used in the proof for uniqueness, and it is used in the approximation argument below.
\subsection{Proof of the uniqueness part of Theorem \ref{mink-existence}.}\label{un}
\begin{proof} Let $\tilde{\mu}$ be measure with density $\tilde{g}(u)=g(u)1_{\{\langle u,v\rangle>0\}}$, for some unit vector $v$, such that $\tilde{g}$ is $p-$concave and $\frac{1}{p}-$homogenous on its support for some $p\geq 0$ (assumptions of the Theorem along with Proposition \ref{p-concave-proposition} of the appendix allow us to select such vector). Fix $q=\frac{p}{np+1}$. Assume that there exist two symmetric convex bodies $K$ and $L$ such that
\begin{equation}\label{condition1111}
d\sigma_{\mu,K}(u)=d\sigma_{\mu,L}(u)
\end{equation}
for all $u\in{{\mathbb S}^{n-1}}$. Observe that
$$\mu_1(K,L)=\int_{{{\mathbb S}^{n-1}}} h_K(u) d\sigma_{\mu,L}(u)=$$
$$\int_{{{\mathbb S}^{n-1}}} h_K(u) d\sigma_{\mu,K}(u)=\mu_1(K,K)=\frac{1}{q} \mu(K).$$
By symmetry of $K$ and $L$, it implies that
$$\tilde{\mu}_1(K,L)=\frac{1}{q} \tilde{\mu}(K).$$
By Corollary \ref{p-concave-homo},
\begin{equation}\label{mink-torefer}
\frac{1}{q}\tilde{\mu}(K)=\tilde{\mu}_1(K,L)\geq \frac{1}{q}\tilde{\mu}(K)^{1-q}\tilde{\mu}(L)^q,
\end{equation}
and hence $\tilde{\mu}(K)\geq\tilde{\mu}(L)$. Analogously, by considering $\tilde{\mu}_1(L,K)$, we get that $\tilde{\mu}(K)\leq \tilde{\mu}(L)$. Hence $\tilde{\mu}(K)=\tilde{\mu}(L)$, and hence there is equality in (\ref{mink-torefer}). Milman and Rotem (\cite{MilRot} Corollary 2.17) proved, using the results from Dubuc \cite{Dub}, that in this case $K$ and $L$ have to coincide up to a dilation and a shift on the support of $\tilde{\mu}$. As we assume that $K$ and $L$ are symmetric, we get that $K=aL$ for some $a>0$ almost everywhere with respect to $\tilde{\mu}$. But as $g$ is $\frac{1}{p}$-homogenous, we have
$$
d\sigma_{\mu,K}(u)=d\sigma_{\mu,aL}(u)=a^{n+\frac{1}{p}-1}d\sigma_{\mu,L}(u),
$$
and hence by (\ref{condition1111}), $a=1$. Which means that $K=L$ $\mu$-almost everywhere.
\end{proof}
\subsection{Proof of the existence part of Theorem \ref{mink-existence}.}
\begin{proof}
We shall use Lemma \ref{mink-existence-polytopes} and argue by approximation. Let $d\varphi(u)$ be an even measure on ${{\mathbb S}^{n-1}}$. For a positive integer $k$, consider a symmetric partition of ${{\mathbb S}^{n-1}}\cap supp(\varphi)$ into disjoint sets $A_i$, $i=1,...,2N$ with spherically convex closures of diameters at most $\frac{1}{k}$ (recall that a subset of the sphere is called spherically convex if the geodesic interval connecting any pair of points in the set is fully contained in this set). Consider the vector
$$c_i=\frac{1}{\varphi(A_i)}\int_{A_i} ud\varphi(u).$$
Note that $c_i\neq 0$. Select $u_i\in{{\mathbb S}^{n-1}}$ and $f_i\in\mathbb R^+$ to be such that $c_i=f_i u_i.$ Note that $u_i\in int(A_i)$. Therefore, for every $u\in A_i$, $|u-u_i|\leq \frac{1}{k}$, and hence
\begin{equation}\label{f_i_est}
1-\frac{1}{k}\leq f_i\leq 1.
\end{equation}
According to Lemma \ref{mink-existence-polytopes}, there exists a polytope
$$P_k=\{x\in\mathbb R^n:\, |\langle x,u_i\rangle|\leq \alpha_i\}$$
with faces $F_{P_K}$, such that
$$\mu_{n-1}(F_{P_K}(u_i))=\int_{A_i} \varphi(u)du.$$
Consider a measure $\varphi_k$ on ${{\mathbb S}^{n-1}}$ such that for every Borel set $\Omega\subset {{\mathbb S}^{n-1}},$
$$\varphi_k(\Omega)=\sum_{u_i\in\Omega}\mu_{n-1}(F_{P_K}(u_i)).$$
Consider a bounded Lipschitz function $a(u)$ on ${{\mathbb S}^{n-1}}.$ Observe that
$$\left|\int_{{{\mathbb S}^{n-1}}} a(u)d\varphi(u)-\int_{{{\mathbb S}^{n-1}}} a(u)d\varphi_k(u)\right|\leq \sum \int_{A_i}|a(u)-a(u_i) f_i| d\varphi(u).$$
Observe as well, that by (\ref{f_i_est}),
$$|a(u)-f_i a(u_i)|\leq |a(u_i)-f_i a(u_i)|+|a(u)-a(u_i)|\leq$$
$$\frac{1}{k}||a||_{Lip}+||a||_{\infty}|1-f_i|\leq \frac{1}{k}(||a||_{Lip}+||a||_{\infty})\rightarrow_{k\rightarrow \infty} 0.$$
Thus $\varphi_k\rightarrow \varphi$ weakly, as $k$ tends to infinity.
It remains to show that all the polytopes $P_k$ are bounded on the support of $\mu$: then, by Blaschke selection theorem (see \cite{book4}, Theorem 1.8.6), applied on the support of $\mu$, there exists a subsequence of $\{P_k\}$ which converges to some convex body $P$ in Hausdorff metric. Then $\sigma_{\mu,P_k}\rightarrow \sigma_{\mu,P}$ weakly (see Proposition \ref{app} from the appendix), and hence, by the uniqueness of the weak limit, we have $d\sigma_{\mu,P}(u)=d\varphi(u)$.
To show the boundedness, observe first that $\mu^+(\partial P_k)=\int_{{{\mathbb S}^{n-1}}} \varphi(u)du=:\tilde{C}_{\varphi}$, where $\mu^+(\partial P_k)$ stands for $\mu_1(P_k,B_2^n)$.
Let $\tilde{g}$ be the restriction of $g$ to a half space where it is $p-concave.$ By Corollary \ref{p-concave-homo},
$$\tilde{\mu}(P_k)\leq \left(q\tilde{\mu}(B_2^n)^{-q}\tilde{\mu}^+(\partial P_k)\right)^{\frac{1}{1-q}},$$
and hence, by symmetry of $P_k,$
\begin{equation}\label{a1}
\mu(P_k)\leq \left(q\mu(B_2^n)^{-q}\mu^+(\partial P_k)\right)^{\frac{1}{1-q}}\leq C_{\mu,\varphi}.
\end{equation}
Here $q=\frac{p}{np+1}$, and $C_{\mu,\varphi}$ depends only on the measures $\mu$ and $\varphi.$ On the other hand, for any $x\in P_k$ we have
$$h_{P_k}(u)\geq \langle u,x\rangle^+=|x|\langle u,v\rangle^+,$$
where $v\in {{\mathbb S}^{n-1}}$ is such that $x=|x|v,$ and $\langle u,x\rangle^+$ stands for the positive part of $\langle u,x\rangle$. We note that for $k$ large enough,
$$\int_{{{\mathbb S}^{n-1}}} \langle u,v\rangle^+ d\varphi_k(u)\geq \frac{1}{2}\int_{{{\mathbb S}^{n-1}}} \langle u,v\rangle^+ d\varphi(u)=:C_{\varphi}>0,$$
where $C_{\varphi}>0$ is a positive constant depending on $\varphi$ only. Therefore,
\begin{equation}\label{a2}
\mu(P_k)=\frac{1}{n+r}\int_{{{\mathbb S}^{n-1}}} h_{P_K}(u) d\varphi_k(u) \geq |x| C_{\varphi}.
\end{equation}
By (\ref{a1}) and (\ref{a2}), $|x|\leq \frac{C_{\mu,\varphi}}{C_{\varphi}}$. As $x$ was an arbitrary point from $P_k$, we conclude that the sequence $\{P_k\}$ is indeed uniformly bounded.
\end{proof}
\section{Applications to the questions about uniqueness.}\label{logmink}
\subsection{An extension of Aleksandrov's theorem.}
\begin{theorem}\label{aleks}
Let $\mu$ be a measure with density with positive degree of concavity and positive degree of homogeneity. Let $K$ and $L$ be symmetric convex bodies such that in every direction $\theta$, $P_{\mu, K}(\theta)=P_{\mu,L}(\theta)$. Then $K=L$ $\mu-$almost everywhere.
\end{theorem}
\begin{proof} Given $g(x)$ on $\mathbb R^n,$ the density of $\mu,$ let $\tilde{\mu}$ on $\mathbb R^n$ be the measure with density $\tilde{g}(x)=\frac{g(x)+g(-x)}{2}$. Recall that by (\ref{Fourier_p_t}),
$$
d\widehat{\sigma_{\tilde{\mu},K}}(\theta)=-C(\mu)\frac{\pi}{n}P_{\tilde{\mu},K}(\theta)
$$
and
$$
d\widehat{\sigma_{\tilde{\mu},L}}(\theta)=-C(\mu)\frac{\pi}{n}P_{\tilde{\mu},L}(\theta),
$$
where $C(\mu)$ depends only on the dimension and the degree of homogeneity of $\mu$, and the Fourier transform is considered with respect to $-(n+1)-$homogenous extensions of $\sigma_{\tilde{\mu},K}$ and $\sigma_{\tilde{\mu},L}$.
Note that $P_{\mu, K}(\theta)=P_{\mu,L}(\theta)$ implies $P_{\tilde{\mu}, K}(\theta)=P_{\tilde{\mu},L}(\theta)$ for every $\theta$. By Fourier inversion formula, we get that $\sigma_{\tilde{\mu},K}=\sigma_{\tilde{\mu},L}$ everywhere on the sphere. By Theorem \ref{mink-existence-intro} we conclude that $K$ and $L$ coincide up to a set of $\mu$-measure zero.
\end{proof}
\subsection{Uniqueness of solutions for certain PDE's in the class of support functions.}
\begin{proposition}\label{logbm}
Let $K$ and $L$ be two symmetric $C^{2,+}$ convex bodies in $\mathbb R^n$ with support functions $h_K$ and $h_L$ and curvature functions $f_K$ and $f_L$ such that
$$\frac{\partial h_K(u)}{\partial x_1} f_K(u)=\frac{\partial h_L(u)}{\partial x_1} f_L(u)$$
for every $u\in{{\mathbb S}^{n-1}}.$ Then $K=L.$
\end{proposition}
\begin{proof} Let $g:\mathbb R^n\rightarrow \mathbb R^+$ be given via
$$g(x)=|x_1|.$$
Then, for every $x\in \mathbb R^n,$
$$
g(\nabla h_K)=\left|\frac{\partial h_K(u)}{\partial x_1}\right|.
$$
By the symmetry, the Proposition \ref{sam_smooth} and the condition of the Corollary,
\begin{equation}\label{conevolappl}
\sigma_{\mu,K}=f_K(u) g(\nabla h_K(u))=f_L(u) g(\nabla h_L(u))=\sigma_{\mu,L}
\end{equation}
for every $u\in{{\mathbb S}^{n-1}}.$
Observe that the restriction of $g$ onto $\{x\in \mathbb R^n:\,x_1>0\}$ is $1-$homogenous and $1-$concave. Therefore, it satisfies the condition of theorem \ref{mink-existence}, and thus, by (\ref{conevolappl}), $K=L$ $\mu$-almost everywhere. In this case it means that $K=L$ coincide almost everywhere with respect to Lebesgue measure, and as they are also convex bodies, it means that $K=L$.
\end{proof}
We remark that the curvature function $f_K$ can be written in the Aleksandrov's form as $det(\delta_{ij} h+h_{ij})$, where $h$ is the support function of $K$, $h_{ij}$ are derivatives of it taken with respect to an orthonormal frame on ${{\mathbb S}^{n-1}}$, and $\delta_{ij}$ is the usual Kroneker symbol. Therefore, Proposition \ref{logbm} implies that a PDE
$$\frac{\partial h}{\partial x_1} det(\delta_{ij} h+h_{ij})=F$$
has a unique solution in the class of even support functions of convex bodies. The existence of such solution for even continuous function $F$ which is not supported on any great subsphere can be derived from Theorem \ref{mink-existence-intro}.
\begin{remark}
Observe that
$$\frac{\partial (h_K(u)f_K(u))}{\partial x_1}=\frac{\partial h_K(u)}{\partial x_1} f_K(u)+\frac{\partial f_K(u)}{\partial x_1} h_K(u).$$
Hence, by Proposition \ref{logbm}, the following pair of conditions guarantee equality of smooth symmetric sets $K$ and $L$:
\begin{enumerate}
\item $h_K (u)f_K(u)=h_K (u)f_K(u)$ at every $u\in{{\mathbb S}^{n-1}}$;\\
\item $\frac{\partial f_K(u)}{\partial x_1} h_K(u)=\frac{\partial f_L(u)}{\partial x_1} h_L(u)$ at every $u\in{{\mathbb S}^{n-1}}$.
\end{enumerate}
\end{remark}
\begin{remark}\label{logbm-remark}
Instead of requiring the condition of Proposition \ref{logbm} it is in fact enough to require that there exists a vector $v$ such that for every $u\in{{\mathbb S}^{n-1}},$
$$f_K(u)\langle\nabla h_K(u),v\rangle=f_L(u)\langle\nabla h_L(u),v\rangle.$$
In this case we still conclude that $K=L.$
\end{remark}
\begin{remark}
The Log-Minkowski problem (see e. g. B\"or\"oczky, Lutwak, Yang, Zhang \cite{BLYZ}, \cite{BLYZ-1}, \cite{BLYZ-2}, Lutwak, Yang, Zhang \cite{LYZ}, Lutwak, Oliker \cite{LO}, Stancu \cite{St}, Huang, Liu, Xu \cite{HLX}) asks whether a symmetric convex body $K$ is uniquely defined by its cone volume measure $\frac{1}{n}h_K(u) f_K(u)$, where $u\in{{\mathbb S}^{n-1}}$.
Suppose that symmetric convex bodies $K$ and $L$ satisfy
\begin{equation}\label{cv}
h_K(u)f_K(u)=h_L(u)f_L(u),
\end{equation}
for every $u\in {{\mathbb S}^{n-1}}$. Consider a vector field
$$a(u)=\nabla h_K(u) f_K(u)-\nabla h_L(u) f_L(u).$$
Note that by (\ref{gradprop}), (\ref{cv}) is equivalent to the fact that $a(u)$ is a tangent field, that is $a(u)\perp u$.
In view of Corollary \ref{logbm}, unique determination of a smooth convex body would follow if one could show that in fact $a(u)$ has to be identically zero. Moreover, in view of the previous remark it would suffice to show that there exists a vector $v\in {{\mathbb S}^{n-1}}$ such that $\langle a(u), v\rangle=0$ for all $u\in{{\mathbb S}^{n-1}}.$
\end{remark}
\section{Extensions of the solution to Shephard's problem.}
We shall follow the scheme of the proof for the classical Shephard problem (see Koldobsky \cite{Kold}), which suggests glueing together harmonic-analytic results with the Brunn-Minkowski theory.
\subsection{General preparatory lemmas.}
To prove Theorem \ref{main}, we first need the following Lemma.
\begin{lemma}\label{shephard_general}
Let $\mu$ be a measure with density $g$ continuous on its support, and let $K, L$ be symmetric convex bodies. Assume additionally that $L$ is a projection body. Assume that for a given $t\in[0,1]$ and for every $\theta\in{{\mathbb S}^{n-1}}$ we have
$$p_{\mu, K}(\theta,t)\leq p_{\mu, L}(\theta,t).$$
Then
$$\mu_1(tK,L)\leq \mu_1(tL,L).$$
\end{lemma}
\begin{proof} Without loss of generality we may assume that $K$ and $L$ are infinitely smooth strictly convex bodies; the general case then follows via standard approximation argument (see, e.g., Koldobsky \cite{Kold} Section 8).
Consider a symmetrization of $\mu$. Let $\tilde{\mu}$ be the measure with density
$$\tilde{g}(x)=\frac{g(x)+g(-x)}{2}.$$
Since $K$ and $L$ are symmetric, we have for all $\theta\in{{\mathbb S}^{n-1}}$ and $t\in[0,1]$:
$$p_{\tilde{\mu}, K}(\theta,t)=p_{\mu, K}(\theta,t);$$
$$p_{\tilde{\mu}, L}(\theta,t)=p_{\mu, L}(\theta,t),$$
and hence
\begin{equation}\label{111111}
p_{\tilde{\mu}, K}(\theta,t)\leq p_{\tilde{\mu}, L}(\theta,t).
\end{equation}
Assume for a moment that $K$ and $L$ are strictly convex and infinitely smooth. By (\ref{Fourier_p_t}),
$$\widehat{\sigma_{\mu,tL}}(\theta)=-\frac{\pi}{n}p_{\mu,L}(\theta,t).$$
Hence, by Proposition \ref{sam_smooth},
$$p_{\tilde{\mu}, K}(\theta,t)=-\frac{n}{\pi}\widehat{f_{tK}\tilde{g}(\nabla h_{tK})}(\theta).$$
By (\ref{111111}), we get
$$\widehat{f_{tK}\tilde{g}(\nabla h_{tK})}(\theta)\geq \widehat{f_{tL}\tilde{g}(\nabla h_{tL})}(\theta),$$
for every $\theta\in{{\mathbb S}^{n-1}}$ and for every $t\in [0,1].$ As $L$ is a projection body, we have $\widehat{h_L}(\theta)\leq 0$. Thus
\begin{equation}\label{toint}
\widehat{h_L}(\theta)\widehat{f_{tK}\tilde{g}(\nabla h_{tK})}(\theta)\leq \widehat{h_L}(\theta)\widehat{f_{tL}\tilde{g}(\nabla h_{tL})}(\theta),
\end{equation}
for every $\theta\in{{\mathbb S}^{n-1}}$ and for every $t\in [0,1].$ Integrating (\ref{toint}) over the unit sphere, and applying Parseval's identity (\ref{parseval}) on both sides of the inequality, we get
\begin{equation}\label{fin}
\int_{{{\mathbb S}^{n-1}}}h_L(\theta)f_{tK}(\theta)\tilde{g}(\nabla h_{tK}(\theta))d\theta\leq \int_{{{\mathbb S}^{n-1}}}h_L(\theta)f_{tL}(\theta)\tilde{g}(\nabla h_{tL}(\theta))d\theta.
\end{equation}
Lemma \ref{formula-mixed-measure} applied along with (\ref{fin}) implies that
$$\tilde{\mu}_1(tK,L)\leq \tilde{\mu}_1(tL,L).$$
Using symmetry of $K$ and $L$ once again, we note that
$$\tilde{\mu}_1(tK,L)=\mu_1(tK,L);$$
$$\tilde{\mu}_1(tL,L)=\mu_1(tL,L),$$
and the lemma follows.
\end{proof}
Via the same scheme as above, invoking Lemma \ref{formula} along with the fact that $V_{\mu,1}(L,L)=\mu(L)$, we get the following
\begin{lemma}\label{shephard_general-integrated}
Let $\mu$ be a measure with density $g$ continuous on its support, and let $K, L$ be symmetric convex bodies. Assume additionally that $L$ is a projection body. Assume that for every $\theta\in{{\mathbb S}^{n-1}}$ we have
$$P_{\mu, K}(\theta)\leq P_{\mu, L}(\theta).$$
Then
$$V_{\mu,1}(K,L)\leq \mu(L).$$
\end{lemma}
\subsection{Proof of the Theorem \ref{main}.}
\begin{proof}
As is shown in Proposition \ref{p-concave-proposition} of the Appendix, if a non-negative function has a positive degree of homogeneity and a positive degree of concavity, then there exists $p\geq 0$ such that $g$ is $p-$concave and $\frac{1}{p}-$homogenous. Additionally, such function is necessarily supported on a convex cone.
The assumptions of the Theorem allow us to apply Lemma \ref{shephard_general-integrated} and obtain:
\begin{equation}\label{part1}
V_{1,\mu}(K,L)\leq \mu(L).
\end{equation}
On the other hand, we apply part (\ref{mink-gen-2}) of Corollary \ref{p-concave-homo} and write
$$\mu(L)\geq V_{1,\mu}(K,L)\geq \mu(K)^{1-q}\mu(L)^q,$$
where $q=\frac{p}{np+1}$. Hence $\mu(L)\geq \mu(K)$.
\end{proof}
\begin{remark}
Theorem \ref{main} does not hold for all measures. Indeed, consider measure $\mu$ with density $1_{B_2^n}$ and convex bodies $L=r B_2^n$, $K=R B_2^n$ such that $r\leq 1\leq R$ and $R\geq r^{-\frac{1}{n-1}}$. Then $P_{\mu,K}(\theta)\leq P_{\mu, L}(\theta)$ for all $\theta\in{{\mathbb S}^{n-1}}$ but $\mu(K)\geq\mu(L)$. However, requiring the inequality $p_{\mu,K}(\theta,t)\leq p_{\mu, L}(\theta, t)$ for all $\theta\in{{\mathbb S}^{n-1}}$ and for all $t\in[0,1]$ may suffice to conclude that $\mu(K)\leq\mu(L)$ for a wide class of measures with some basic concavity properties.
\end{remark}
\subsection{A general statement}
Finally, we present a measure comparison-type result for a more general class of measures. It may prove useful for considering this problem in greater generality.
\begin{proposition}
Let $\mu$ be a measure on $\mathbb R^n$ with density continuous on its support. Suppose that $\mu$ is $F(t)-$concave for some invertible $C^1$ function $F:\mathbb R^+\rightarrow \mathbb R.$ Let $K$ and $L$ be convex symmetric bodies, and let $L$ in addition be a projection body. Assume that for every $\theta\in {{\mathbb S}^{n-1}}$ and for every $t\in [0,1]$ we have
$$p_{\mu,L}(\theta,t)\geq p_{\mu,K}(\theta,t).$$
Then
\begin{enumerate}[label = (\roman*)\,]
\item $\displaystyle \mu(L)\geq \mu(K)+\int_0^1 \frac{F(\mu(tL))-F(\mu(tK))}{tF'(\mu(tK))} dt;$\\
\item $\displaystyle \mu(L)\geq \mu(K)+\int_0^1 \left[\mu(tL)-\mu(tK)+\frac{F(\mu(tL))-F(\mu(tK))}{F'(\mu(tK))}\right]dt$.\end{enumerate}
\end{proposition}
\begin{proof} By Lemma \ref{shephard_general}, we get that $\mu_1(tK,L)\leq \mu_1(tL,L)$ for every $t\in [0,1]$, and therefore
\begin{equation}\label{1}
\mu_1(tK,tL)=t\mu_1(tK,L)\geq t\mu_1(tL,L)=\mu_1(tL,tL).
\end{equation}
Applying (\ref{1}) along with Theorem \ref{mink_gen} we get
\begin{equation}\label{eq}
t\mu_1(tL,L)\geq t\mu_1(tK,K)+\frac{F(\mu(tL))-F(\mu(tK))}{F'(\mu(tK))}.
\end{equation}
After dividing both sides by $t$ and integrating we get
\begin{equation}\label{rfr}
\int_0^1\mu_1(tL,L)dt\geq \int_0^1\mu_1(tK,K)dt+ \int_0^1\frac{F(\mu(tL))-F(\mu(tK))}{tF'(\mu(tK))}dt,
\end{equation}
hence (i) follows from (\ref{measure-derivative-integral}) and (\ref{rfr}).
Next, we integrate by parts:
\begin{equation}\label{11111}
\int_0^1 t\mu_1(tL,L)dt=\mu(L)-\int_0^1 \mu(tL)dt.
\end{equation}
Thus (\ref{eq}) and (\ref{11111}) imply (ii).
\end{proof}
\subsection{Proof of Theorem \ref{main-2}.}
\begin{proof}
Without loss of generality we may assume that the boundary of $L$ is infinitely smooth (see the approximation argument in Koldobsky \cite{Kold}, Section 8). Inasmuch as $L$ is not a projection body we have that $\widehat{h_L}$ is positive on an open set $\Omega\subset {{\mathbb S}^{n-1}}$; recall as well that, per our assumptions, the curvature function $f_L$ is positive everywhere on the sphere, and $L$ is symmetric. Let $v:{{\mathbb S}^{n-1}}\rightarrow \mathbb R$ be a non-negative infinitely smooth even function supported on $\Omega.$ Let $\tilde{g}(x)$ be the restriction of $g(x)$ on the half space where is has positive homogeneity, and let $\tilde{\mu}$ be the measure with density $\tilde{g}$. Since we assume that $g$ is supported on the whole space, $\tilde{g}$ is fully supported on a half space.
Define a symmetric convex body $K$ via the relation
\begin{equation}\label{equation}
d\sigma_{\mu,K}(u)=d\sigma_{\mu,L}(u)-\epsilon \widehat{v}(u)
\end{equation}
for every $u\in{{\mathbb S}^{n-1}}$. Here $\epsilon>0$ is chosen small enough so that the right hand side of (\ref{equation}) stays non-negative. Theorem \ref{mink-existence} guarantees that such convex body exists. Applying Fourier transform to $-(n+1)$-homogenous extensions of both sides of (\ref{equation}), we get
$$
-\frac{\pi}{n q}P_{\tilde{\mu},K}(\theta)=-\frac{\pi}{n q}P_{\tilde{\mu},L}(\theta)-\epsilon v(\theta),
$$
and hence, by symmetry of $K$ and $L$,
\begin{equation}\label{p_comp}
-P_{\mu,K}(\theta)=-P_{\mu,L}(\theta)-\frac{n q}{\pi}\epsilon v(\theta).
\end{equation}
Recall that
$$V_{\mu,1}(K,L)=\int_0^1\int_{{{\mathbb S}^{n-1}}} h_L(u) f_{tK}(u) g(\nabla h_{tK}(u)) du dt,$$
and that $P_{\mu,K}(\theta)$ is the Fourier transform of the $-(n+1)$-homogenous extension of
$$-\frac{\pi}{n}\int_0^1f_{tK}(u)g(\nabla h_{tK}(u))dt.$$
Note that $\widehat{h_L}(u) v(u)$ is positive for all $u\in\Omega$. Therefore, by Parseval's type formula (\ref{parseval}),
$$V_{\mu,1}(K,L)=V_{\mu,1}(K,L)=-(2\pi)^{-n}\frac{\pi}{n}\int_{{{\mathbb S}^{n-1}}} \widehat{h_L}(u) P_{\mu,K}(u) du=$$
$$-(2\pi)^{-n}\frac{\pi}{n}\int_{{{\mathbb S}^{n-1}}} \widehat{h_L}(u) P_{\mu,L}(u) du-(2\pi)^{-n}q\epsilon \int_{\Omega} \widehat{h_L}(u) v(u) du<$$
$$-(2\pi)^{-n}\frac{\pi}{n}\int_{{{\mathbb S}^{n-1}}} \widehat{h_L}(u) P_{\mu,L}(u) du=\mu(L).$$
Using the above along with Corollary \ref{p-concave-homo} we get that
$$\mu(L)> V_{\mu,1}(K,L)\geq \mu(K)^{1-q}\mu(L)^q,$$
and hence $\mu(L)> \mu(K)$. On the other hand, (\ref{p_comp}) implies that $P_{\mu,L}(\theta)\leq P_{\mu,K}(\theta)$ for every $\theta\in{{\mathbb S}^{n-1}}$.
\end{proof}
\section{Stability and separation for Shephard's problem extension.}
\subsection{Separation result for Theorem \ref{main}.}
\begin{theorem}\label{stability}
Fix $n\geq 1$, $p\in [0,\infty)$ and consider a measure $\mu$ on $\mathbb R^n$ whose density $g:\mathbb R^n\rightarrow \mathbb R^+$ is $p-$concave and $\frac{1}{p}$-homogenous function. Set $q=\frac{p}{np+1}$.
Let $K$ and $L$ be symmetric convex bodies, and let $L$ additionally be a projection body. Fix $\epsilon>0.$ Assume that for every $\theta\in {{\mathbb S}^{n-1}}$ we have
$$P_{\mu, K}(\theta)\leq P_{\mu, L}(\theta)-\epsilon.$$
Then
$$\mu(K)^{1-q}\leq \mu(L)^{1-q}-C(\mu)\epsilon,$$
where $C(\mu)$ is a constant which only depends on the measure $\mu.$
\end{theorem}
We formulate the following notable corollary of Theorem \ref{stability}.
\begin{corollary}\label{hyperplane}
Fix $n\geq 1$, $p\in [0,\infty)$ and consider a measure $\mu$ on $\mathbb R^n$ whose density $g:\mathbb R^n\rightarrow \mathbb R^+$ is $p-$concave and $\frac{1}{p}$-homogenous function. Set $q=\frac{p}{np+1}$.
Let $L$ be a strictly convex symmetric projection body. Then
$$\mu(L)^{1-q}\geq C(\mu)\min_{\theta\in{{\mathbb S}^{n-1}}}P_{\mu,L}(\theta),$$
where $C(\mu)$ is a constant which only depends on the measure $\mu.$
\end{corollary}
Corollary \ref{hyperplane} is an analogue of a hyperplane inequality for Lebesgue measure of projections (see Gadrner \cite{gardner}, or Koldobsky \cite{K7}).
\begin{proof}[Proof of Theorem \ref{stability}.]
Let $\tilde{\mu}$ be, as before, the symmetrization of $\mu$, i.e. the measure with the density $g(x)=\frac{g(x)+g(-x)}{2}$.
Assume without loss of generality that $K$ and $L$ are infinitely smooth. The assumptions $\widehat{h_L}\leq 0$ and
$$P_{\mu, K}(\theta)\leq P_{\mu, L}(\theta)-\epsilon,$$
lead to the following chain of inequalities:
$$V_{\tilde{\mu},1}(K,L)=-(2\pi)^{-n}\frac{\pi}{n}\int_{{{\mathbb S}^{n-1}}} \widehat{h_L}(u)P_{\tilde{\mu}, K}(u)du\leq$$
$$-(2\pi)^{-n}\frac{\pi}{n}\int_{{{\mathbb S}^{n-1}}} \widehat{h_L}(u)P_{\tilde{\mu}, L}(u)du+\epsilon (2\pi)^{-n}\frac{\pi}{n}\int_{{{\mathbb S}^{n-1}}}\widehat{h_L}(u)du=$$
$$\tilde{\mu}(L)+\epsilon (2\pi)^{-n}\frac{\pi}{n} \int_{{{\mathbb S}^{n-1}}}\widehat{h_L}(u)du.$$
By Corollary \ref{p-concave-homo}, we have
\begin{equation}\label{food}
\tilde{\mu}(L)+\epsilon (2\pi)^{-n}\frac{\pi}{n} \int_{{{\mathbb S}^{n-1}}}\widehat{h_L}(u)du\geq \tilde{\mu}(K)^{1-q}\tilde{\mu}(L)^q.
\end{equation}
Let $S={{\mathbb S}^{n-1}}\cap supp(g)$. By Theorem \ref{mink-existence} there exists a symmetric convex body $Q$ (depending on the measure $\mu$) with
$$d\sigma_{\tilde{\mu}, Q}=\frac{1}{q}\widehat{1_S},$$
and therefore satisfying
$$P_{\tilde{\mu},Q}(\theta)=1_S(\theta).$$
We then estimate
$$(2\pi)^{-n}\frac{\pi}{n}\int_{{{\mathbb S}^{n-1}}}\widehat{h_L}(u)du\leq (2\pi)^{-n}\frac{\pi}{n}\int_{S}\widehat{h_L}(u)du=(2\pi)^{-n}\frac{\pi}{n}\int_{{{\mathbb S}^{n-1}}}\widehat{h_L}(u)P_{\tilde{\mu},Q}(\theta)du=$$
\begin{equation}\label{food1}
-V_{\tilde{\mu},1}(Q, L)\leq -\tilde{\mu}(Q)^{1-q}\tilde{\mu}(L)^q.
\end{equation}
Letting $C(\mu)=\tilde{\mu}(Q)^{1-q}$, by (\ref{food}) and (\ref{food1}), we get
$$\tilde{\mu}(L)-\epsilon C(\mu)\tilde{\mu}(L)^q\geq \tilde{\mu}(K)^{1-q}\tilde{\mu}(L)^q,$$
which implies the statement of the Theorem for $\tilde{\mu}$ in place of $\mu$, and hence the Theorem follows for $\mu$ as well.
\end{proof}
\subsection{Stability for Theorem \ref{main}.}
Finally, we prove the stability result.
\begin{theorem}\label{separation}
Fix $n\geq 1$, $p\in [0,\infty)$ and consider a measure $\mu$ on $\mathbb R^n$ whose density $g:\mathbb R^n\rightarrow \mathbb R^+$ is $p-$concave and $\frac{1}{p}$-homogenous function. Set $q=\frac{p}{np+1}$.
Let $K$ and $L$ be symmetric convex bodies, and let $L$ additionally be a projection body. Let $\epsilon>0.$ Assume that for every $\theta\in {{\mathbb S}^{n-1}}$ we have
$$P_{\mu, K}(\theta)\leq P_{\mu, L}(\theta)+\epsilon.$$
Then $\mu(K)^{1-q}\leq \mu(L)^{1-q}+C(\mu,L)\epsilon$, where $C(\mu,L)$ is a constant which depends on the measure $\mu$ and the body $L.$
\end{theorem}
\begin{proof} Suppose that
$$P_{\mu,K}(\theta)\leq P_{\mu,L}(\theta)+\epsilon.$$
Assume without loss of generality that $K$ and $L$ are infinitely smooth. Then, similarly to the proof of Theorem \ref{stability}, we have
$$\mu(L)-\epsilon (2\pi)^{-n}\frac{\pi}{n} \int_{{{\mathbb S}^{n-1}}}\widehat{h_L}(u)du\geq \mu(K)^{1-q}\mu(L)^q.$$
For the unit ball $B_2^n$ we have
$$(2\pi)^{-n}\frac{\pi}{n}\int_{{{\mathbb S}^{n-1}}}\widehat{h_L}(u)du=-\nu_{n-1}^{-1} V_1(B_2^n,L).$$
Let $R(L)$ be the smallest positive number such that $L\subset R(L) B_2^n$. Note that
$$V_1(B_2^n,L)=\lim_{\epsilon\rightarrow 0} \frac{|B_2^n+\epsilon L|-|B_2^n|}{n\epsilon}\leq \nu_n \frac{(1+\epsilon R(L))^n-1}{n\epsilon}=\nu_n R(L).$$
Letting $C(L,\mu)=\frac{\nu_n}{\nu_{n-1}} R(L) \mu(L)^{-q}$, we get the statement of the Theorem.
\end{proof}
\pagebreak
|
\section{Introduction}
This long introduction summarizes the main contributions of the paper and the intuitions behind many of the definitions, results and proofs. We start with the main object of this paper: introducing the Brascamp-Lieb inequalities, and surveying known structural results and our new algorithmic results. We then describe how to instantiate the operator scaling algorithm of~\cite{GGOW} directly in the context of BL-inequalities, in a way that does not require previous familiarity with it. The properties of this algorithm lead to further structural results, as well as highlight it as a provably efficient instance of the general {\em alternate minimization} \footnote{sometimes also called alternating minimization.} heuristic. Finally, we discuss several known and potential applications of Brascamp-Lieb inequalities in computer science and optimization, which further motivate this and future studies.
\subsection{The Brascamp-Lieb inequalities: basic notions}
Many important inequalities, including H\"older's inequality, Loomis-Whitney inequality, Young's convolution inequality, hypercontractivity inequalities and many others are all special cases of the extremely general Brascamp-Lieb inequalities, introduced by these authors in~\cite{BL76,Lieb90}. Yet many others, including Prekopa-Lindler inequality, versions of Brunn-Minkowski inequality and others are special cases of Barthe's reverse form of Brascamp-Lieb~\cite{Barthe2}. We discuss below only the original form, and only for Euclidean spaces. The notation we use is taken mostly from~\cite{BCCT}, which is an excellent source both for background and motivation on this topic, as well as the state-of-art on the basic questions in this field. As is common, we often use BL as abbreviation for Brascamp-Lieb.
A {\em Brascamp-Lieb datum} is specified by two $m$-tuples: one of linear transformations
\linebreak $\mathbf{B} = (B_1, B_2, \dots, B_m)$, with $B_j:{\mathbb{R}}^n \rightarrow {\mathbb{R}}^{n_j}$
(along with $(n, n_1,\ldots,n_m)$, the vector of {\em dimensions}), and another of
non-negative reals $\mathbf{p}=(p_1, p_2,\dots , p_m)$ (which is a vector of {\em exponents}).
This combined information is denoted by $(\mathbf{B,p})$.
A {\em Brascamp-Lieb inequality}\footnote{While this is the common name of the inequality below found in the literature, we note that only a restricted form appears in the original~\cite{BL76} paper, and this general form appeared only in the later paper of Lieb~\cite{Lieb90}.} with the datum above states that for {\em every} tuple of $m$
non-negative functions, $f=(f_1, f_2, \dots, f_m)$, which are integrable according to the Lebesgue
measure, we have the following inequality.
$$ \int_{x\in {\mathbb{R}}^n} \prod_{j=1}^m (f_j(B_j x))^{p_j} dx \, \leq \,
C \,\prod_{j=1}^m \left(\int_{x_j \in {\mathbb{R}}^{n_j}} f_j(x_j) dx_j \right)^{p_j} $$
where $C \in (0, \infty ]$ is independent of the functions $f$. Of course, we really get an interesting
inequality if $C$ is finite. In that case, we call the datum $(\mathbf{B,p})$ {\em feasible}, and we
denote smallest $C$ for which this inequality holds by $\textnormal{BL}(\mathbf{B,p})$, called the {\em \textnormal{BL}-constant}.
Many familiar special cases of this inequality are listed in the introduction of~\cite{BCCT}; we describe only a very simple one here, which hopefully makes concrete and intuitive the complicated formal expression above and its constituents. A special case of the Loomis-Whitney inequality asserts that {\em the volume of every measurable set $S$ in the unit cube is at most the square root of of the product of the areas of its projections on the three coordinate planes}. In this theorem, the linear transformations $B_j:{\mathbb{R}}^3 \rightarrow {\mathbb{R}}^2$ are defined by the simple projections $B_1(x,y,z) = (y,z)$, $B_2(x,y,z) = (x,z)$ and $B_3(x,y,z) = (x,y)$, the functions $f_j$ are the indicator functions of these three projections of the set $S$, and the exponent vector is $p=(\frac12,\frac12,\frac12)$. The corresponding BL-constant in this case happens to be 1 (this case is an important situation which will be important later). The reader may be familiar with an entropy\footnote{Appropriately defined for these continuous variables.} version of the inequality above, namely that $H(X,Y,Z) \leq \frac12 (H(X,Y)+H(Y,Z)+H(X,Z))$ for every random variable $(X,Y,Z)$. Indeed, BL-inequalities may be viewed in general as entropy subadditivity inequalities, as developed in~\cite{Carlen09, Liu16, LiuCCV17}.
There has been extensive work on precisely understanding (and computing) when a given datum
$(\mathbf{B,p})$ is feasible, and if it is, determining the
BL-constant $\textnormal{BL}(\mathbf{B,p})$. Clean characterizations exist for both questions, which clarify that they are
both decidable (have a finite algorithm). Let us overview these characterizations and then turn to the computational complexity
of the existing algorithms and our new ones. The first, regarding feasibility, is from the paper~\cite{BCCT},
and the second, on the optimal constant, is from~\cite{Lieb90}. Before doing so, we make two general comments, one about the ``reverse BL-inequalities'' and the second regarding how we measure input size when analyzing algorithms.
\paragraph{\textit{Reverse BL inequalities}}
This informal comment simply clarifies that all results in this paper stated for BL-inequalities hold for their reverse form.
In~\cite{Barthe2}, Barthe introduced a reverse form of the Brascamp-Lieb inequalities, sometimes called RBL inequalities, which turn out to generalize some known inequalities not captured by the the original BL inequalities. These RBL inequalities take the {\em same} data as standard BL inequalities, and have an optimal RBL constant associated with each. All questions raised above for the BL inequalities, like feasibility, computation of the constant and its continuity properties are relevant in this reverse setting. However, they are actually equivalent to the original ones for a simple reason: Barthe~\cite{Barthe2} proved that for any feasible datum the optimal BL constant and the optimal RBL constant multiply to $1$, and when the BL datum is infeasible the RBL constant is $0$. In short, these two optimal constants determine each other. Thus all structural results above translate to the reverse setting, and so do all our algorithmic results.
\paragraph{\textit{Input size conventions}}
Before we start, let us formalize the input conventions to all algorithms, and the parameters we
use to measure their complexity. The input to all algorithms will be a BL datum $(\mathbf{B,p})$, and the ``size"
of each part will be measured differently. The entries of the matrices $B_j$ will be rational numbers,
given in binary. We will let $b=b(\mathbf{B})$ denote their total binary length (note that in particular $b \geq nm$, and so lower bounds the ``combinatorial size'' of the input).
The vector $\mathbf{p}$ will be given as a sequence of rationals with a common denominator, namely
$p_j = c_j/d$ with $c_j, d$ integers. We use this convention for two reasons. First, our algorithms will use
this integer representation, and their complexity will depend on $d=d(p)$ (while in many cases $d$ is only
polynomial in the dimensions $n,m$ of the problem, it can definitely be as large as exponential in other cases).
Second, in the context of {\em quiver representations}, which is very relevant to this study as we shall
later see\footnote{With hindsight, our reduction from BL data to operator scaling data may be viewed as a special case of the reduction of Derksen and Makam~\cite{DM15} of general quivers to the Left-Right quiver. More on this in Section \ref{sec:bl-to-cap}.}, it is natural to use these integer ``weights'' $c_j$ and $d$.
Summarizing, the two size parameters for a BL datum $(\mathbf{B,p})$ will be $b$ and $d$ as above.
\subsection{The Brascamp-Lieb inequalities: known and new results}
\paragraph{\textit{Testing feasibility} (\textit{and more})}
The following theorem of Bennett et al~\cite{BCCT,BCCT2} precisely characterizes when a given BL datum is feasible. This work will provide a different proof of this important theorem (see Corollary~\ref{cor:BCCT}).
\begin{theorem}[\cite{BCCT}]\label{feasibility}
The datum $(\mathbf{B,p})$ is feasible if and only if the following conditions hold:
\begin{enumerate}
\item $n= \sum_j p_j n_j.$
\item $\dim (V) \leq \sum_j p_j \dim(B_jV)$ \,\, holds for all subspaces $V$ of ${\mathbb{R}}^n$.
\end{enumerate}
\end{theorem}
Note that these are simply linear conditions on the exponent vector $\mathbf{p}$, albeit defined
by infinitely many subspaces $V$. However, as the coefficients are integers in $[n] = \{1,2,\ldots,n\}$, there are at
most $n^m$ different inequalities, and so the linear maps $\mathbf{B}$ define a polytope $P_\mathbf{B}$ (sometimes called the {\em BL polytope}) in ${\mathbb{R}}^n$, such that
$(\mathbf{B,p})$ is feasible iff $\mathbf{p} \in P_\mathbf{B}$.
The BL polytopes $P_\mathbf{B}$ have received a lot of attention. For example, their vertices were characterized for the ``rank-1 case'' (namely, when all dimensions $n_j=1$) by Barthe \cite{Barthe2}, and this was extended to other cases by Valdimarsson~\cite{Vald10}. In the same paper, Valdimarsson considers the question of generating the inequalities defining $P_\mathbf{B}$, and gives a finite algorithm to do so (needless to say, after they are given, feasibility testing becomes a linear programming problem). No explicit upper bound on the complexity of this algorithm is given, but it is at least exponential in $m$. The same holds for the algorithm in~\cite{christ2013communication} for generating the inequalities defining $P_\mathbf{B}$.
We give a polynomial time algorithm for the feasibility problem, and much more. Our algorithm also
gives a ``separation oracle'' (namely finds a violated inequality when infeasible).
Recall again that the exponents $\mathbf{p}$ in the BL datum $(\mathbf{B,p})$ are given by
$p_j=c_j/d$ where $c_j, d$ are integers.
\begin{theorem}[Corollary \ref{cor:bl_feasibility} rephrased]\label{feasibility-alg}
There is an algorithm that on input $(\mathbf{B,p})$ of binary length $b$ and common denominator $d$ runs in
time $\textnormal{poly}(b,d)$
and provides the following information:
\begin{description}[leftmargin= 1in]
\item [ Membership oracle:] Tests if $\mathbf{p} \in P_\mathbf{B}.$
\item [ Separation oracle:] When $\mathbf{p} \not\in P_\mathbf{B}$, provides a violated inequality, namely
a basis for a vector space $V$ in ${\mathbb{R}}^n$ such that $\dim (V) > \sum_j p_j \dim(B_jV).$
\end{description}
\end{theorem}
We believe that the ability to efficiently optimize over such a wide family of polytopes with
exponentially many facets should be useful for different optimization problems via appropriate
reductions to this setting. We will give some simple examples of such reductions in Section~\ref{sec:mat_int}.
One concrete challenge is e.g. trying to find such a reduction which
embeds the Edmonds polytope~\cite{Edm65} of perfect matchings in a general (non-bipartite)
graph into some BL polytope.
It is an interesting open problem if one can improve the algorithm to depend
polynomially on $\log d$ instead of $d$. This
would allow for optimization over the BL polytopes in polynomial time via the ellipsoid algorithm. We believe even the current separation oracle for BL polytopes with polynomial dependence on $d$ should allow for approximate optimization over BL polytopes but we haven't been able to prove this yet.
\paragraph{\textit{Characterizing the BL constant}}
The following theorem of Lieb~\cite{Lieb90} characterizes the BL constant $\textnormal{BL}(\mathbf{B,p})$ in all cases
where the BL datum is feasible.
The heart of the proof establishes that the optimal constant in any BL inequality is attained (or approached) by plugging in density functions of appropriate centered Gaussians. For such densities, the BL constant has a nice expression via the following identity:
$$
\int_{\mathbb{R}^n} \textnormal{exp} \left( - \pi x^T A x\right) dx = \textnormal{Det}(A)^{-1/2}
$$
\begin{theorem}[\cite{Lieb90}]\label{BLconstant}
Assume $(\mathbf{B,p})$ is feasible. Then
$$\textnormal{BL}(\mathbf{B,p}) =
\left[ \sup \frac{\prod_j (\det X_j)^{p_j}}{\det \left(\sum_j p_j B_j^{\dagger} X_j B_j \right)} \right]^{1/2} $$
where the supremum is taken over all choices of positive definite matrices $X_j$ in dimension $n_j$, and
$B_j^{\dagger}$ denotes the adjoint map corresponding to $B_j$.
\end{theorem}
Thus the BL constant for given BL datum is a solution to an optimization problem.
However, as defined, it is not even convex. We are not aware of any algorithms to compute the
BL constant in general. Moreover, to the best of our knowledge no explicit bounds on the BL
constant (when finite) in terms of the BL datum were known. We resolve both issues: provide an
explicit bound and give a polynomial time algorithm for computing the BL constant to any accuracy.
\begin{theorem}[Corollary \ref{cor:bl_upperbound_rational} rephrased]\label{thmintro:BL_ub}
For any feasible \textnormal{BL} datum $(\mathbf{B,p})$ that has binary length $b$ and common denominator $d$, it holds that $\textnormal{BL}(\mathbf{B,p}) \leq \exp(O(b \log(bd)))$.
\end{theorem}
\begin{theorem}[Theorem \ref{thm:computing_bl} rephrased]\label{constant-alg}
There is an algorithm that on input $(\mathbf{B,p})$ of binary length $b$ and common denominator $d$, and an accuracy parameter
$\epsilon >0$, runs in time $\textnormal{poly}(b, d, 1/\epsilon)$ and computes a factor
$(1+\epsilon)$-approximation of $\textnormal{BL}(\mathbf{B,p}).$ Furthermore, the algorithm outputs a scaling $\mathbf{B'}$ which is almost geometric i.e. $\textnormal{BL}(\mathbf{B'},\mathbf{p}) \le 1+\epsilon$.
\end{theorem}
\subsection{Efficient computation of the BL-constant via scaling}
The algorithm underlying the proof of Theorem \ref{constant-alg} will be shown (via a reduction) to be a special case
of the operator scaling algorithm of Gurvits~\cite{gurvits2004}, which the current authors analyzed
and proved to be polynomial time in~\cite{GGOW}.
In this subsection, we explicitly describe the algorithm in this special case of BL-inequalities only, without referring to~\cite{GGOW}.
It will be instructive to see how
notion of scaling, which naturally exists in the theory of BL inequalities, can be used algorithmically. Moreover, it will become clear how the algorithm implies (known and new) structural consequences to the BL theory.
This description will also help to motivate our reduction in the technical chapters which follow (which will provide the proof for its run-time).
Let us return to the BL constant, and to an important family of BL data called {\em geometric}
\footnote{The analogous term in the operator scaling setting will be called {\em doubly stochastic}.}
defined below. It was introduced by Ball~\cite{Ball} and extended by Barthe~\cite{Barthe2}.
\begin{definition}\label{def:normalization}
A BL datum $(\mathbf{B,p})$ is called {\em geometric} if it satisfies the following conditions
(with $I_k$ denoting the $k\times k$ identity matrix).
\begin{description}[leftmargin= 1in]
\item [ Projection:] For every $j\in [m]$, $B_j$ is a projection, namely $B_j B_j^{\dagger} = I_{n_j}$.
\item [ Isotropy:] $\sum_j p_j B_j^{\dagger} B_j = I_n$
\end{description}
\end{definition}
Ball proved that for the special case of geometric BL datum with $n_j=1$, the constant is always one. Barthe extended this to the general case\footnote{The analogous
theorem for operator scaling is that the capacity of doubly stochastic operators is always $1$. Another family of inequalities where the
constant is always $1$ are the ``discrete'' BL inequalities of \cite{CDKSY15}.}.
\begin{theorem}[\cite{Ball, Barthe2}]\label{geometric}
For every geometric BL datum $(\mathbf{B,p})$ we have $\textnormal{BL}(\mathbf{B,p})=1$.
\end{theorem}
A simple and natural action of the general linear groups on the Euclidean spaces involved (which simply performs a basis change in each space) yields an equivalence relation on BL data\footnote{This group action naturally calls for a study of BL from an invariant theory viewpoint, which indeed exists in a much more general context that will be relevant to us in several ways, namely that of {\em quiver representations}. An extensive survey is ~\cite{DW07}.}. Specifically, we say that $(\mathbf{B,p})$ and $(\mathbf{B',p'})$ are {\em equivalent} if there exist matrices $C\in GL_n({\mathbb{R}})$, $C_j \in GL_{n_j}({\mathbb{R}})$ (which are called {\em intertwining transformations}\footnote{They will be called {\em scaling matrices} in the operator scaling setting.} in~\cite{BCCT}) such that $B_j' = C_j^{-1}B_jC$ and $p_j' = p_j$ for all $j$. It is easy to compute the effect of such action on the BL constant.
\begin{proposition}[\cite{BCCT}]\label{basis-change}
Assume $(\mathbf{B,p})$ and $(\mathbf{B',p'})$ are equivalent via $C,C_j$ as above. Then
$$ \textnormal{BL}(\mathbf{B',p'}) = \frac{\prod_j (\det(C_j))^{p_j}}{\det (C)} \textnormal{BL}(\mathbf{B,p}).$$
where $\det$ is the determinant polynomial on matrices of the appropriate dimension.
\end{proposition}
Equivalence and this simple, efficiently computable formula suggests a natural path to computing BL constants. Given BL datum $(\mathbf{B,p})$, compute an equivalent {\em geometric} datum $(\mathbf{B',p'})$ (if one exists), and the intertwining transformations relating the two. The main questions are, is there such a geometric equivalent datum, and how to compute these transformations. The theory of (quantum) operator scaling~\cite{gurvits2004,GGOW} suggests that a simple, greedy procedure will work. This algorithm will underly (most of) the statements in Theorem~\ref{feasibility-alg} and Theorem~\ref{constant-alg}. As mentioned, we will explain the connection and reduction to operator scaling in the next sections, and here describe informally how it is applied to the BL setting.
\paragraph{\textit{The scaling algorithm}}
The key idea of {\em greedy, iterative} scaling, which goes back to Sinkhorn \cite{Sink} in the (classical) matrix scaling, is that when trying to achieve a pair of conditions as in the definition of geometric BL datum, we should try to satisfy them one at a time! Say we are given $(\mathbf{B,p})$ and assume for example that the isotropy condition is not met. A non-triviality condition (otherwise $(\mathbf{B,p})$ is not feasible) implies that $C \triangleq \sum_j p_j B_j^{\dagger} B_j $ is invertible, and so we can set $B_j \leftarrow B_j C^{-1/2}$ to get a new datum which satisfies isotropy.
So, we ``only'' have to fix the projection property; lets do it. Again, non-triviality implies that each of
$C_j \triangleq B_j B_j^{\dagger}$
is invertible, and we can now set $B_j \leftarrow C_j^{-1/2} B_j $, satisfying the projection property. Of course, we may now have ruined isotropy. No problem: lets fix it again in the same way, and repeat alternately fixing the unsatisfied property. The magic is that this sequence of {\em normalization} steps converges, and moreover, converges in polynomial time, to a geometric datum, whenever the original datum $(\mathbf{B,p})$ is feasible!\footnote{Note that the sequence might only converge to a geometric datum and it is possible that no element of the sequence is geometric.} Let us describe the BL algorithm more precisely and then state its properties formally.
\begin{Algorithm}[H]
\textbf{Input}: A BL datum $(\mathbf{B,p})$ \\ \\
Set $\mathbf{B^0}=\mathbf{B}$.\\
Repeat for $i=1$ to $t$ normalization steps.
\begin{itemize}
\item if $i$ is odd, normalize $\mathbf{B^i}$ from $\mathbf{B^{i-1}}$ to satisfy isotropy.
\item if $i$ is even, normalize $\mathbf{B^i}$ from $\mathbf{B^{i-1}}$ to satisfy projection.
\end{itemize}
\textbf{Output}: $\mathbf{B^t}$.
\caption{The BL scaling algorithm}
\label{The BL scaling algorithm}
\end{Algorithm}
The BL scaling algorithm defines a dynamics in the space of BL data which are restricted to stay in one equivalence class. If $(\mathbf{B,p})$ is the original datum which is not already geometric, it defines a sequence $\mathbf{B^0}=\mathbf{B}, \mathbf{B^1}, \mathbf{B^2},\dots$ which alternately satisfies the isotropy or projection properties (note that the exponent vector $p$ is used, but not changed in this process). We now state the main three properties which underlie the analysis.
\begin{theorem}\label{analysis}
For every $(\mathbf{B,p})$ the following holds:
\begin{enumerate
\item \emph{Upper bound:} If $(\mathbf{B,p})$ is feasible, then $\textnormal{BL}(\mathbf{B,p}) \leq \exp(O(b \log(bd)))$.
\item \emph{Lower bound:} If the datum $(\mathbf{B,p})$ is either isotropy-normalized or projection-normalized, then $\textnormal{BL}(\mathbf{B,p}) \geq 1$.
\item \emph{Progress per step:} Let $(\mathbf{B',p})$ denote the result of applying either isotropy-normalization or projection-normalization to $(\mathbf{B,p})$. Then $($as long as $\textnormal{BL}(\mathbf{B,p}) \geq 1+\epsilon)$, we have $$\textnormal{BL}(\mathbf{B',p}) \leq (1-\textnormal{poly} \left( \epsilon/nd \right)) \textnormal{BL}(\mathbf{B,p}).$$
\end{enumerate}
\end{theorem}
We will not need to prove this theorem, as it will follow via our reduction\footnote{In the case when the exponent vector $\mathbf{p}$ is irrational, the reduction to operators does not exist and so in this case, we provide an analysis of the number of iterations of Algorithm \ref{The BL scaling algorithm} in Section \ref{sec:irrational}.} from the analogous theorem in the more general operator scaling setting\footnote{Where statements made here regarding the BL constants will be replaced by statements about a related notion called the {\em capacity} of an operator.}. However, let us say a few words here about what goes into each item above using BL language. Property (3) on progress per step is the simplest: it follows from a robust version of the AM-GM inequality. Property (2) follows from plugging in densities of spherical Gaussians (it was proved in \cite{Vald11}, also see Section \ref{sec:properties}).
Property (1) is the hardest, in that it requires degree bounds on the generating polynomial invariants in the null-cone of the quiver associated to the group action above. We will briefly explain this mouthful in Section \ref{sec:ub_bl}, and refer the reader to~\cite{GGOW} for details. In Section \ref{sec:ub_bl}, we will also give an improved and simplified analysis of the main theorem from \cite{GGOW} and derive the desired upper bounds on the BL constant via our main reduction. All in all, it is worth stressing again that beyond this non-trivial use of algebra, on the analytic side the most general BL inequalities follow from the most basic one, AM-GM.
This analysis above implies that we can choose the number $t$ of steps in the BL scaling algorithm
to be $\textnormal{poly}(b,d, 1/\epsilon)$ so that $\textnormal{BL}(\mathbf{B^{i},p}) \leq 1+\epsilon$ for some $1 \le i \le t$
(whenever the initial datum $(\mathbf{B^0,p})$ is feasible). Testing if the lower bound is met
($\textnormal{BL}(\mathbf{B^{i},p}) \leq 1+\epsilon$) decides feasibility\footnote{It is not immediately clear how to
test $\textnormal{BL}(\mathbf{B^{i},p}) \leq 1+\epsilon$. However this can be done by checking closeness to the
geometric position in an appropriate metric, similar to the distance to double stochasticity for
operators that will be defined later.}, thus proving the first item in Theorem~\ref{feasibility-alg}.
For proving the second item
we will employ a very different, beautiful combinatorial algorithm for operator scaling (only the decision version of it), that was
developed after ours by Ivanyos, Qiao and Subrahmanyam~\cite{IQS15b}. Their algorithm relies
on even tighter degree bounds from invariant theory which were proved very recently~\cite{DM15},
and has several advantages over the algorithm in~\cite{GGOW}; in particular it can find violated
dimension inequalities. Going back to Theorem~\ref{constant-alg} on approximating the BL
constant, note that it suffices to multiply together the (respective) normalizing matrices used in all
steps of the BL scaling algorithm, using the formula in Proposition~\ref{basis-change}.
We now return to the scaling algorithm above, and discuss connections and applications of its structure and properties in the next two bullets.
\paragraph{\textit{The BL constant, scaling and alternate minimization}}
The problem of computing the BL constant, as it is formulated here, is a non-convex problem.
We presented an efficient algorithm for this optimization problem which may be viewed as an instance of a general heuristic called {\em alternate minimization}. Let us describe and discuss this method.
In general, alternate minimization
is a very general technique introduced by~\cite{CT84}, which was devised to solve problems of the following
kind. Given some universe $U$, a (distance) function $d : U \times U \rightarrow {\mathbb{R}}_+$ and two subsets
$X, Y \subseteq U$, find
$\displaystyle ({\mathbf x}^*, {\mathbf y}^*) = \textnormal{arg}\min_{\substack{{\mathbf x} \in X, {\mathbf y} \in Y}} d({\mathbf x}, {\mathbf y})$. We note that
such problems are in general NP-hard, even over convex sets $X,Y$ and natural distances $d$.
Nevertheless, alternate minimization is widely used to solve special cases of such problems in practice.\footnote{Note that such problems can be defined with more than two components, and the alternate minimization approach below can be extended to such problems, but we stick here with two.}
An alternate minimization algorithm for the problem above starts with a arbitrary point $(x_0,y_0) \in X \times Y$ and generates a sequence of points $(x_i,y_i)$ as follows. In even steps $i$, $x_{i+1}=x_i$, and $y_{i+1}$ is chosen to be the value of $y$ which minimizes $d(x_i,y)$ over the second coordinate alone (an optimization problem that is often far simpler). In odd steps the roles of $x$ and $y$ are reversed. It is easy to cast the scaling algorithm above as a process of this nature. It is also easy to see that it enjoys some nice properties in general situations. For example, if both $X$ and $Y$ are strictly convex, and $d$ is a metric, then this process converges to the optimum. Of course, the speed of convergence is the main question. We list below some other, possibly familiar instances of alternate minimization for which such convergence is not known.
The famous Lemke-Howson algorithm for finding a Nash equilibrium for 2-player
games~\cite{Lemke-Howson} is of that nature. Recall that in that algorithm one starts with an arbitrary strategy for both players, and then one alternates fixing the strategy of one, and finding the best response for the other via linear programming.
Another such example is the work of Zafeiriou and Petrou~\cite{ZP10}
on computing non-negative tensor factorization. As above, the analysis of such alternating
minimization algorithms just provide convergence of the procedure, without proving {\em how fast}
it converges to the minimum, with the hope that this heuristic will lead to quick convergence on instances arising in practice.
Of course, there are examples, e.g. the alternate minimization algorithms for matrix completion in~\cite{JNS13, hardt14} which prove rapid convergence under certain conditions on the inputs.
In this work, as well as in~\cite{GGOW}, we prove that our alternate minimization algorithm
not only converges, but also converges to the infimum in polynomial time. Our proof is based on the introduction
of a potential function which measures how much progress we make in every step of the alternating
minimization algorithm. In our case the potential function {\em is} the very BL constant we are trying to optimize, which is a special case of the so-called {\em capacity} we use as potential in our operator scaling algorithm~\cite{GGOW}. As it happens, the analysis of convergence relies on a combination of algebraic tools, in particular from (the commutative) invariant theory and (the non-commutative) theory of skew fields, as explained in that paper. We hope that such potential functions as well as the methods to analyze their progress, which we call {\em capacity methods}, will be of use in the analyses of other problems which use alternate
minimization, especially when the result of optimizing each of the solution components in individual optimization steps may be viewed as the action of a group. We are currently exploring this direction for alternate minimization with more than two components, and these lead to interesting algebraic questions related to questions that arise in invariant theory as well as in Geometric Complexity Theory of Mulmuley and Sohoni~\cite{GCT1, GCT5}.
\paragraph{\textit{Bounds and continuity of the BL constant, and non-linear BL inequalities}}
Another important property of the BL scaling algorithm above, evident from its form, is that it is smooth! Namely, the orbits under the algorithm of two sufficiently close BL data will remain close to each other for the duration of the algorithm. This implies that the BL constants of both will be close as well, and moreover this can be quantified! This simple observation gives a much stronger result than existing qualitative results on the continuity of BL constants which has recently received significant attention. This is yet another example highlighting the usefulness of algorithmic results in mathematics. We briefly overview the motivation and known results before stating ours.
How smooth (or regular) the BL constant is (in terms of the BL data) is of course a natural question, especially as the expression in Lieb's theorem~\ref{BLconstant} is not even convex. Further motivation to study this question arises from a variety of non-linear variants of the BL inequalities, in which the maps $B_j$ are non-linear, but at least smooth enough so as to be approximable by linear ones in a small ball.
Some such variants, needed in specific applications, were proved directly. The papers~\cite{BB10,BBFL15} ask this question in full generality, and demonstrate the importance of such smoothness conditions of the BL constant for general non-linear extensions of BL inequalities. These papers also
exposit the many diverse applications of such generalized BL inequalities
in analysis, number theory and other areas (possibly the most impressive recent ones are the applications of ``Kakeya-type'' BL inequalities to number theory in~\cite{BoDe15,BDG15}, resolving long standing open problems including the ``Vinogradov mean-value conjecture'').
Local boundedness, a condition weaker than continuity, is established in~\cite{BBFL15}.
\begin{theorem}[\cite{BBFL15}]\label{locally-bounded}
If $(\mathbf{B,p})$ is feasible, then there are $\delta > 0$ and $C < \infty$ such that for all $(\mathbf{B',p})$
such that $\|\mathbf{B}'-\mathbf{B}\|_2 \leq \delta$ we have $\textnormal{BL}(\mathbf{B',p}) \leq C$.
\end{theorem}
This was followed recently by the paper~\cite{BBCF16}, which proves continuity of the BL constant.
\begin{theorem}[\cite{BBCF16}]\label{continuous}
The \textnormal{BL} constant $\textnormal{BL}(\mathbf{B,p})$ is continuous in $\mathbf{B}$. Namely, for every feasible $(\mathbf{B,p})$ and $\epsilon>0$, there exists a $\delta >0$ such that
if $\|\mathbf{B}'-\mathbf{B}\|_2 \leq \delta$, then we have that $|\textnormal{BL}(\mathbf{B',p})-\textnormal{BL}(\mathbf{B,p})| \leq \epsilon$.
\end{theorem}
In both theorems above, no quantitative bounds are given. This is not surprising, as the proofs of both use compactness.
On the other hand, the availability of a smooth algorithm, together with the quantitative analysis of its continuity provided in~\cite{GGOW}, imply via our reduction the following quantitative bound on the continuity\footnote{While Theorem \ref{stability} only implies continuity of $\textnormal{BL}(\mathbf{B,p})$ for rational $\mathbf{B}$ and $\mathbf{p}$, we describe in Section \ref{sec:cont_bl} how our proof can be extended to the case of irrational $\mathbf{B}$ and speculate how this might also extend to irrational $\mathbf{p}$.} of the BL constant (giving a multiplicative approximation).
\begin{theorem}[Theorem \ref{BL_continuity} rephrased]\label{stability}
For every parameters $b, d, \epsilon >0$ there exists a
$$
\delta = \delta(b,d,\epsilon) \leq \textnormal{exp} \left( -\frac{\textnormal{poly}(b,d)}{\epsilon^3} \right)
$$
such that for every \textnormal{BL} data $(\mathbf{B,p})$ of size $b,d$ and every \textnormal{BL} data $(\mathbf{B',p})$ such that $\|\mathbf{B}'-\mathbf{B}\|_2 \leq \delta$, we have that if $\textnormal{BL}(\mathbf{B,p})$ is finite, then $\textnormal{BL}(\mathbf{B',p})$ is finite. Furthermore,
$$
(1-\epsilon) \textnormal{BL}(\mathbf{B,p}) \leq \textnormal{BL}(\mathbf{B',p}) \leq (1+\epsilon)\textnormal{BL}(\mathbf{B,p})
$$
\end{theorem}
We hope that this explicit bound will be useful to making non-linear BL inequalities quantitative, and to their applications.
\subsection{Applications of Brascamp-Lieb in Computer Science, Optimization and Beyond}
Besides the applications to other areas of mathematics as we mentioned above, the structural theory of
BL inequalities (sometimes with the connection to operator scaling) have encountered several applications in computer science and optimization. We give below a (partial) survey of these applications.
Many diverse applications exist even for the special case of rank-$1$ BL setting (when all the $n_i$'s are $1$), which has been studied extensively. In this case, the BL polytopes are exactly the {\em basis polytopes} \cite{Barthe2}. More specifically, if the maps $B_i$ are given by $B_i(x) = \langle v_i,x \rangle$, then the BL polytope is the convex hull of the following set:
$$
\{1_I\,:\, I \subset [m], |I|=n, (v_i)_{i \in I} \: \text{forms a basis for $\mathbb{R}^n$}\}
$$
Hence Theorem \ref{feasibility-alg} for this special case is well known. Furthermore, computing the BL constant in this special case can be formulated as a convex problem and hence Theorem \ref{constant-alg} in this special case can be obtained by the ellipsoid algorithm \cite{gur-sam1, HardtM13} (\cite{gur-sam1} in fact solves a more general problem). We list some of the applications of this setting.
\begin{itemize}
\item In \cite{HardtM13}, Hardt and Moitra provide an application of scaling the rank-$1$ BL datum to a geometric position (which they call isotropic) to the problem of {\em robust subspace recovery}.
\item Forster \cite{Forster02} used the existence of a scaling to geometric position for rank-$1$ BL datum (sometimes called Barthe's theorem) to give the first nontrivial lower bounds for the {\em sign rank} of an explicit matrix (in particular the Hadamard matrix), and thereby prove the first nontrivial {\em unbounded error communication complexity} lower bounds for the inner product function.
\item In~\cite{DSW_LCC14}, Dvir et al. use Barthe's theorem as part of the first super-quadratic lower bounds on the size of 3-query {\em LCCs (Locally Correctable Codes)} over the Real numbers.
\item Recently, Nikolov and Singh \cite{NikolovS16} use a optimization problem similar to the rank-$1$ BL constant to give a polynomial time algorithm for approximating the maximum value of the determinant of a submatrix of a psd matrix under partition constraints.
\end{itemize}
The class of BL polytopes is a rich class of polytopes. As mentioned above, the rank-$1$ case gives rise to basis polytopes. The rank-$2$ case (when all $n_i$'s are $1$ or $2$) has been completely characterized by Valdimarsson \cite{Vald10}. An important special case is given by the {\em linear matroid intersection} polytopes which we describe next. Suppose $(v_1,\ldots,v_m)$ and $(w_1,\ldots,w_m)$ are two collection of vectors in $\mathbb{R}^n$. Then the matroid intersection polytope given by these collections is the convex hull of the following set:
$$
\{1_I \,:\, I \subset [m], |I|=n, (v_i)_{i \in I}, (w_i)_{i \in I} \: \text{both form a basis for $\mathbb{R}^n$}\}
$$
Let us define the maps $B_i : \mathbb{R}^{2n} \rightarrow \mathbb{R}^2$ in the following way:
$$
B_i(x,y) = (\langle v_i,x\rangle, \langle w_i, y\rangle)
$$
Then the BL polytope $P_{\mathbf{B}}$ is exactly the corresponding matroid intersection polytope. In particular, the bipartite matching polytopes are a special case of rank-$2$ BL polytopes.
We will give proofs of these statements in Section~\ref{sec:mat_int}, with the hope of demonstrating the simplicity by which such exponential size linear programs of this form capture (simple) combinatorial optimization problems.
As mentioned before, it remains a very interesting challenge to see if the general matching polytopes (or in general the linear matroid matching polytopes) are a special case of BL polytopes, for which it may still suffice to stay in the rank-2 case. We note a peculiar discrepancy between operator scaling and BL-inequalities when encoding the matroid intersection problem. In operator scaling, this problem was one of the first special cases considered in Gurvits' paper~\cite{gurvits2004}, and he showed that one could encode it by an operator scaling problem with matrices of rank $1$. Here, when encoded as a BL polytope, we use rank-$2$ projectors, and believe it is impossible to do in rank $1$. Even though BL is a special case of operator scaling (as our main reduction shows) the relative power of these two formulations with respect to rank is not clear.
Characterizing the vertices of BL-polytopes for projectors of rank $3$ and higher remains a very interesting question, as is the case of rank-$2$ matrices in the formulation of operator scaling.
BL-inequalities (and operator scaling) have recently found applications in computational and
combinatorial geometry in the paper of Dvir et al.~\cite{DGOS16}. They generalize
the rank lower bounds for design matrices given in~\cite{BDWY12, DSW14} to the setting of
block matrices. This result allows them to obtain sharp bounds on Sylvester-Gallai type
theorems for arrangements of subspaces, to obtain new incidence bounds for high-dimensional
line/curve arrangements, as well as to prove structural rigidity results in the projective setting.
Beyond {\em applying} BL-inequalities, we note that in recent years computer scientists started getting interested
in {\em proving} them efficiently. Efficiency here here should be taken to mean in the Sum-of-Squares (SoS)
framework, a hierarchy of semi-definite programs that is perhaps the most powerful general algorithmic technique
for a variety of optimization problems. The paper~\cite{BKS14}, generalizing from many particular examples on
the power of SoS algorithms, suggests a general framework for rounding solutions in this system, in which a
crucial element is efficient SoS proofs of inequalities (which in all known applications are special cases of
BL-inequalities). This viewpoint has lead to many new recent SoS algorithms~\cite{BKS15, MSS16, HSSS16},
and naturally also lead to the question of efficiently proving (in this sense) all BL-inequalities. This was taken up by
Lei and Sheng~\cite{LS16}, who have (in a very general setting) SoS proofs of degree
$\textnormal{poly} (d)$ for BL-inequalities that have denominator $d$ in its exponent $\mathbf{p}$-vector. This yields an algorithm for
testing feasibility that is {\em exponential} in $d$. Our algorithms depend polynomially in $d$, and raise the
question of relating operator scaling and SoS algorithms, and whether such a relation can help improve the
bounds above. Needless to say, relating the power of these two optimization methods is motivated for many
other reasons!
\paragraph{\textit{Dictionary}}
We give here a basic dictionary describing translation between central notions in the theory of Brascamp-Lieb inequalities and completely positive operators. This will be useful in reading the technical sections below.
The basic notions concerning completely positive operators will be defined in Sections \ref{sec:square} and \ref{sec:rect}.
\begin{center}
\begin{tabular}{| l | l |}
\hline
\textbf{Brascamp-Lieb inequalities} & \textbf{Completely positive operators} \\ \hline
BL datum & Kraus operators of completely positive operators \\ \hline
Intertwining transformations & Scaling operations \\ \hline
Geometric BL datum & Doubly stochastic operator \\ \hline
BL constant & Capacity \\ \hline
\end{tabular}
\end{center}
\paragraph{\textit{Organization of the paper}}
All but the last two sections are devoted to providing background and then proving the main technical results of this paper.
Preliminaries and results from~\cite{gurvits2004,GGOW} about (square) completely positive operators, their scaling, capacity and more will be reviewed in Section \ref{sec:square}. In Section \ref{sec:rect}, we will extend these results to {\em rectangular} completely positive operators via a simple reduction. In Section \ref{sec:bl-to-cap}, we describe our main reduction, from BL data to (rectangular) completely positive operators, and prove Theorem \ref{feasibility-alg}. We also show how this reduction yields, as a corollary, a simple proof of the characterization (Theorem \ref{feasibility}) of \cite{BCCT}. In Section \ref{sec:ub_bl}, we describe an upper bound on the Brascamp-Lieb constants and prove Theorem \ref{thmintro:BL_ub}. In Section \ref{sec:compute_bl}, we show how to approximate Brascamp-Lieb constants by using the main reduction above and our operator scaling algorithm (proving Theorem \ref{constant-alg}). In Section \ref{sec:cont_bl}, we use the smoothness of the scaling algorithm to give explicit bounds for the continuity of Brascamp-Lieb constants and prove Theorem \ref{stability}. In Section~\ref{sec:mat_int} (which is independent from all others) we exemplify the power of exponentially large linear programs arising as BL-polytopes to capture certain simple combinatorial optimization problems. Section~\ref{sec:properties} contains some additional properties of the Brascamp-Lieb constant, strengthening the results in \cite{Vald11}. Section~\ref{sec:irrational} contains an analysis of the number of iterations required in Algorithm \ref{The BL scaling algorithm} to get close to geometric position. Finally, we conclude with some open problems in Section \ref{sec:open}.
\section{Square completely positive operators}\label{sec:square}
\input{square.tex}
\section{Rectangular operators}\label{sec:rect}
\input{rectangular.tex}
\section{Reduction from Brascamp-Lieb to completely positive operators}\label{sec:bl-to-cap}
\input{main_reduction.tex}
\section{Upper bound on the Brascamp-Lieb constant}\label{sec:ub_bl}
\input{cap-lower-bound.tex}
\section{Computing the Brascamp-Lieb constant}\label{sec:compute_bl}
\input{computing-bl-const.tex}
\section{Continuity of the Brascamp-Lieb constant}\label{sec:cont_bl}
\input{continuity.tex}
\section{Interesting special cases of BL polytopes} \label{sec:mat_int}
\input{matroid.tex}
\section{Additional properties of the Brascamp-Lieb constant}\label{sec:properties}
\input{properties.tex}
\section{Iteration analysis for Algorithm \ref{The BL scaling algorithm}}\label{sec:irrational}
\input{irrational.tex}
\section{Conclusion and open questions}\label{sec:open}
\input{conclusion.tex}
\section*{Acknowledgments}
We would like to thank Zeev Dvir for pointing us to the paper \cite{BCCT}. We would
also like to thank Jonathan Bennett, Larry Guth and anonymous reviewers for helpful
comments.
We thank Rohit Gurjar for suggesting that BL polytopes can encode the linear
matroid intersection polytopes, and we would like to thank Nisheeth Vishnoi and
Damian Straszak for pointing out to us a mistake in our previous proof of
Lemma~\ref{lem:BL-eq-mat} and for providing us a fix for it.
\bibliographystyle{alpha}
\subsection{Connection to quiver semi-invariants}\label{sec:quiver}
In this short subsection, we provide an invariant theoretic view of our reduction\footnote{We remark that other connections between BL inequalities and quiver representations are mentioned in \cite{BCCT}.}. It is far from being self-contained and we refer the reader to the survey \cite{DW2000} about semi-invariants of quiver representations. In this theory, there is a general way of generating semi-invariants for general quivers by a reduction to the Kronecker quiver (left-right action) (\cite{DW2000, DZ2001, SVdB2001}, also see Section $5$ in \cite{DM15}).
As it happens, the data to both BL inequalities and to completely positive operators fits the descriptions of the ``star quiver''and the ``Kronecker quiver'' respectively, and with hindsight, the reduction described in Construction~\ref{cons:main_reduction} (when composed with the reduction from rectangular to square operators) can be viewed as a special case of the reduction above. More specifically, semi-invariants for the star quiver with dimensions $(n, n_1, n_2, \ldots, n_m)$ and weight vector $(d, -c_1,-c_2,\ldots,-c_m)$ can be generated by reducing to a Kronecker quiver of appropriate dimensions (see Figure \ref{fig:star_to_Kronecker}). It is worth emphasizing here the intriguing connection (that this paper exploits) between analysis (BL inequalities) and algebra (semi-invariants of the star quiver): a BL datum $(B_1,\ldots,B_m, c_1/d,\ldots,c_m/d)$ is not feasible iff $(B_1,\ldots,B_m)$ is the common zero of the semi-invariants of the star quiver with dimensions $(n, n_1, n_2, \ldots, n_m)$ and weight vectors \linebreak $\{ (l\cdot d, -l \cdot c_1,-l \cdot c_2,\ldots,-l \cdot c_m): l \in \mathbb{N}\}$\footnote{One can even define such a connection for the finiteness of a specific BL constant: $\textnormal{BL}(\mathbf{B,p}) = \infty$ iff $\mathbf{B}$ is in the null cone of the subgroup of $\text{GL}_n({\mathbb{C}}) \times \text{GL}_{n_1}({\mathbb{C}}) \times \cdots \text{GL}_{n_m}({\mathbb{C}})$ satisfying $\textnormal{Det}(C) = \prod_{i=1}^m \textnormal{Det}(C_i)^{p_i}$.}.
\begin{figure}[H]
\centering
\begin{tikzpicture}
\tikzset{vertex/.style = {shape=circle,draw,minimum size=1.5em}}
\tikzset{edge/.style = {->,> = latex'}}
\node[vertex] (cr) at (0,0) [label={[label distance=0.01cm]45:$(n,d)$}] {cr};
\node[vertex] (p_1) at (2.5,0) [label={[label distance=0.01cm]0:$(n_1,-c_1)$}] {$p_1$};
\node[vertex] (p_2) at (0,-2.5) [label={[label distance=0.01cm]-90:$(n_2,-c_2)$}] {$p_2$};
\node[vertex] (p_3) at (-2.5,0) [label={[label distance=0.01cm]180:$(n_3,-c_3)$}] {$p_3$};
\node[vertex] (p_4) at (0,2.5) [label={[label distance=0.01cm]90:$(n_4,-c_4)$}] {$p_4$};
\draw[edge] (cr) -- node[below] {$B_1$} ++ (p_1);
\draw[edge] (cr) -- node[left] {$B_2$} ++ (p_2);
\draw[edge] (cr) -- node[above] {$B_3$} ++ (p_3);
\draw[edge] (cr) -- node[left] {$B_4$} ++ (p_4);
\node[vertex] (l) at (6,0) [label={[label distance=0.01cm]-95:$(dn,1)$}] {l};
\node[vertex] (r) at (10,0) [label={[label distance=0.01cm]-85:$(dn,-1)$}] {r};
\draw[edge][yshift=0.5cm] (l) to [bend right=40] node[sloped,midway,above] {$A_{m'-1}$} (r);
\draw[edge][yshift=0.5cm] (l) to [bend left=40] node[sloped,midway,above] {$A_2$} (r);
\draw[edge][yshift=0.5cm] (l) to [bend right=80] node[sloped,midway,above] {$A_{m'}$} (r);
\draw[edge][yshift=0.5cm] (l) to [bend left=80] node[sloped,midway,above] {$A_1$} (r);
\node[fill,circle,scale=0.1] at (8,0) {};
\node[fill,circle,scale=0.1] at (8,0.2) {};
\node[fill,circle,scale=0.1] at (8,-0.4) [label={[label distance=1cm]-90:$m'=d \cdot \left( \sum_{i=1}^4 c_i\right)$ multiple edges}] {};
\end{tikzpicture}
\caption{Reduction from star quiver to the Kronecker quiver. Nodes are labelled with $(d,w)$: $d$ is the dimension of the vector space at this node and $w$ is the weight corresponding to this node. Edges are labelled with maps from the vector space at the head of the edge to the vector space at the tail of the edge.} \label{fig:star_to_Kronecker}
\end{figure}
|
\section{Introduction}\label{intro}
Nuclear activity and star formation in galaxies are observed to often coexist across all redshifts \citep{farrah03, alexander05, shi09, hernan09, hatzimi10, mainieri11, lamassa13,harris16,alag16}, up to extremely high active galactic nuclei (AGN) and starburst luminosities \citep[e.g.][]{gen98,carilli01,omont01,farrah02,efst14,magdis14,rosen15}. Moreover, there is a tight correlation between the stellar velocity dispersion in the bulge and the mass of the supermassive black hole (BH) residing in the centre for nearby quiescent galaxies (e.g. \citealt{magorrian98,ferrarese00,tremaine02,haering04}). These observations suggest that there exists a deep connection between stellar and black hole mass assembly events in galaxies.
The nature of, and connection between, star formation and AGN activity in galaxies is unclear. For starbursts, the majority of star-forming systems lie on a `main sequence' whose mean star formation rate (SFR) rises with redshift, from about 10\,M$_\odot \mathrm{yr}^{-1}$ at $z=0.5$ to roughly 100\,M$_\odot \mathrm{yr}^{-1}$ at $z=2$ \citep[as in e.g.][]{elb11,rodig11,schreiber15}. The origin of this main sequence is however still debated, as is the trigger for star formation as a function of a galaxy's position relative to the main sequence. For AGN, it is difficult to complete a robust census of AGN from surveys, since they are obscured by dust and gas for a significant fraction of their lives, with this fraction possibly dependent upon both luminosity and redshift \citep[e.g.][]{mart05}. It is also unclear if star formation events and AGN activity can directly affect one another. A direct relation is motivated by models for galaxy assembly to improve consistency between predictions and observations, most often via `quenching' of star formation by an AGN \citep[e.g.][]{bower06, croton06, booth09, fabian12}. However, observational studies of quenching remain inconclusive. Indirect manifestations of feedback, such as large molecular gas outflows \citep[e.g.][]{feruglio10,spo13} and powerful AGN-driven winds \citep[e.g.][]{perna15}, are found routinely (see also e.g. \citealt{farrah09,bridge13,spo13}), but studies that claim a causal relation are rarer \citep{farrah12,page12}, and sometimes controversial \citep[e.g.][]{harrison14}.
As a result, the scaling relations between star formation and AGN activity across their respective duty cycles remain uncertain. Some authors find a scaling between SFR and AGN luminosity \citep[e.g][]{ima11,Young2014,delv15}, while others do not \citep[e.g.][]{shao10,mullaney12,harrison14,ma15,stanley15}. Recently, \citet{harris16} have shown that a correlation between SFR and AGN luminosity may only exist over certain AGN luminosity, SFR and redshift ranges.
An insightful way to study the connections between star formation and AGN activity in the context of galaxy assembly events is to examine their scaling relations in populations that signpost specific regions of the AGN luminosity and SFR parameter space. One such population are optically luminous type 1 quasars that host luminous, off-main sequence starbursts. This population is straightforward to find in optical surveys, corresponds to a specific phase in the AGN duty cycle, and may signpost the extremes of the starburst duty cycle. As such, they may illustrate how the processes that convert free baryons to stellar and BH mass may change at the most extreme luminosities. They are also an excellent population in which to search for the initial stages of AGN feedback.
A related population within which it would appear intuitive to search for evidence of AGN feedback is that of the Broad Absorption Line (BAL) quasars \citep{lynds67,turnshek84}. The BAL quasars possess broad P-Cygni-like absorption features in their ultraviolet spectra that are blue-shifted with respect to the nominal wavelength of the emission line. These features may stem from outflows from the quasars and may further be associated with high mass-loss rates \citep{dekool02, chartas03}. These outflows could arise in two ways; the outflows could be a random process, present for only a random fraction of the quasar lifetime and/or over certain viewing angles \citep[e.g.][]{elvis00}. The observation of BALs in only 10-15 per cent of quasars \citep[e.g.][]{hewett03} then arises from a suitable combination of these two factors. Second, BALs may signpost outflows that occur only at a particular point in a quasar's lifetime, most likely `young' objects recently (re)fuelled by mergers that also trigger star formation in their hosts. It is this second scenario that would mean that BAL quasars may signpost quasar-mode feedback. This may manifest itself via BAL quasars having different far-infrared (FIR) properties, on average, to those of ordinary quasars.
In this paper, we undertake such a study, exploring the relationship between star formation and AGN activity in optically luminous quasars over $0.5<z<4$ that host extremely high SFRs, from 40$-$4000\,M$_\odot$yr$^{-1}$, placing these hosts approximately one dex above the star formation main sequence. To do so, we start with quasars in the Sloan Digital Sky Survey (SDSS) that also lie within extragalactic survey fields covered by the Spectral and Photometric Imaging Receiver \citep[SPIRE;][]{griffin10} instrument onboard \textit{Herschel} \citep{pilbratt10}. We then restrict ourselves to those quasars that are individually detected by \textit{Herschel}. We infer SFRs from the \textit{Herschel} data, and compare these to the properties of the AGN, as measured from the SDSS catalogue data. We also examine the properties of BAL quasars to see if their properties show evidence for AGN feedback.
The paper is structured as follows. The SDSS and {\it Herschel} data, as well as the matched quasar catalogue, are described in Sec. \ref{sec:data}. Further, Sec. \ref{sec:analysis} describes the spectral energy distribution (SED) fitting, the SFR derivation and uncertainties and discusses the way in which the current study complements previous works in terms of samples and methodology. Sec. \ref{sec:results_full} presents the SFRs in the full quasar sample and in the BAL quasar sub-sample separately, as a function of the quasars' intrinsic properties, namely accretion luminosity, BH mass and Eddington ratio. Lastly, Sec. \ref{sec:discussion} discusses our results and places them in a greater context. Throughout this work, we assume $H_0 = 72 \hspace{0.1cm} \textnormal{km} \hspace{0.1cm} \textnormal{s}^{-1} \hspace{0.1cm} \textnormal{Mpc}^{-1}$, $\Omega_\Lambda = 0.7$ and $\Omega_\textnormal{M}=0.3$.
\begin{table*}
\begin{center}
\begin{tabular}{l C{2cm} crrrrrrr}
\hline
&&& \multicolumn{3}{c}{SDSS quasars} && \multicolumn{3}{c}{{\it Herschel}/SDSS quasars} \\
\cline{4-6}
\cline{8-10}
\noalign{\vskip0.025cm}
Field & Overlapping Region (deg$^2$) && Total & In DR7 & In DR10 && Total & In DR7 & In DR10\\
\hline
\vspace{-0.25cm}
&&&&&&&&\\
HerMES & 60 && 1,293 & 586 & 739 && 83 & 43 & 44\\
HerS & 79 && 4,524 & 2,154 & 2,779 && 212 & 150 & 90\\
HeLMS & 210 && 8,827 & 2,488 & 6,796 && 218 & 92 & 144\\
\hline
\end{tabular}
\caption{Quasar detection statistics by survey. The overlapping region is the area of each \textit{Herschel} field that is also covered by SDSS. The SDSS detections are those quasars that lie in their respective \textit{Herschel} fields. The joint detections columns represent those SDSS quasars with SPIRE detections above 3$\sigma$ at 250$\micron$. We further describe these objects by the data release in which they are found: DR7 or DR10. The sum of the DR7 and DR10 columns is always higher than the columns showing the totals, as the latter only consider unique entries. The SDSS detections columns represent those SDSS quasars located in each respective \textit{Herschel} field, while the joint detections are the quasars with 250\,$\micron$ detections at $>$3$\sigma$.}
\label{tab:jointDetections}
\end{center}
\end{table*}
\begin{figure}
\begin{center}
\vspace{-.2cm}
\includegraphics[width=8cm]{Mi_z_withHisto.pdf}
\vspace{-2.5cm}
\caption{Absolute \textit{i}-band magnitude versus redshift. The red circles represent those objects included in SDSS DR7, while the blue squares are those objects in DR10.
The filled and open shapes differentiate between the BAL and non-BAL samples, respectively (see Section \ref{sec:BAL}). To characterise these BAL and non-BAL subsets we use the objects that lie within the black box, with its limits defined by 1.5\,$\le$\,\textit{z}\,$\le $\,3.3 and $-28.5$\,$\le$\,$M_i$\,$\le$\,$-24.3$, in which most of the BAL quasars lie. The redshift histograms are shown at the top, with the filled red, dashed blue and solid black showing the DR7, DR10 and full sample of objects, respectively. The green spikes represent the redshift distribution of the BAL quasars.}
\label{fig:MI_z}
\end{center}
\end{figure}
\section{Data}\label{sec:data}
\subsection{SDSS Data}\label{sec:sdss}
We assemble our parent quasar sample from a combination of the SDSS quasar catalogues from DR7 \citep{schneider10} and DR10 \citep{paris14}. The location within the SDSS colour-space provides the basis for the selection of most quasar candidates, with \cite{richards02} describing the original SDSS selection criteria. For an object with an ultraviolet excess (typical for quasars) to be selected for spectroscopic follow-up, it should be brighter than $i \sim$19.1 ($i \sim$ 20.2) if its estimated redshift based on its colours is $z \la 3$ ($z \ga 3$). However, the quasar candidate selection criteria applied up through DR7 were biasing the selection against objects with $z \ga 2.2$. The Baryon Oscillation Spectroscopic Survey (BOSS) quasars included in DR10, on the other hand, were sought for in the redshift range between 2.15 and 3.5, down to fainter magnitudes ($g<22.0$ or $r<21.85$), in order to reach a higher quasar surface density \citep{ross12}. We include both DR7 and DR10 quasars to get a larger and more uniform redshift coverage. Had we opted for DR7 (DR10) alone, we would have missed the bulk of the $z > 2$ ($z<2$) quasars. The resulting combined sample, though not complete in redshift, covers a more representative space in absolute $i$-band magnitude, $M_\textit{i}$, and redshift than would either of the two data releases individually.
As a function of redshift, DR7 is more than 90 per cent complete up to redshifts of about 2.2 \citep{richards02}. Thus, DR10 was designed to target objects above this redshift; that is, quasars at $2.15 \le z \le 3.5$, over which the CORE selection is homogeneous and uniform \citep{ross12, white12}. Within this range, DR10 has a given maximum completeness from single-epoch data of seventy per cent at $z=2.2$ \citep[][fig. 14]{ross12}. As such, some of the highest redshift BAL quasars might be missing from our sample.
The combined use of DR7 and DR10, without any restrictions, potentially affects the overall completeness and uniformity of the sample, since the different apparent magnitude cuts in the two releases could make the combined sample incomplete in terms of redshift. K-S tests on the distribution of the various colours and $M_\textit{i}$ however yield $p\,$-\,values (all above 0.4) that cannot reject the null hypothesis that the DR7 and DR10 sub-samples come from the same colour and $M_\textit{i}$ parent distributions. Moreover, \cite{ma15}, who also study the FIR properties of a combined sample of quasars from DR7 and DR10, similarly reach the conclusion that any incompleteness resulting from the non-uniform selection of targets in the two releases does not affect the validity of such a study.
\subsection{Herschel Data}\label{sec:sample}
We look for \textit{Herschel} counterparts to our SDSS quasar sample in the {\it Herschel} Multi-tiered Extragalactic Survey \citep[HerMES;][]{oliver12} and the {\it Herschel} Stripe 82 Survey \citep[HerS;][]{viero14}. HerS consists of 79 deg$^2$ of contiguous SPIRE imaging of the SDSS ``Stripe 82''. The HerMES fields considered here are the HerMES Large Mode Survey (HeLMS) field, the widest HerMES tier spanning 270 deg$^2$ observed with SPIRE alone, as well as the northern Level 5 and 6 fields of Bootes, EGS, ELAIS-N1, Lockman, FLS and \textit{XMM}-LSS, selected for their overlap with the SDSS footprint. The total area we select from is thus $\sim$ 350 deg$^2$.
The point source catalogues from which we extract the SPIRE fluxes in the smaller HerMES fields are presented in \cite{wang14}. The HeLMS catalogue is part of the HerMES DR4 and is also briefly discussed in \cite{wang15}. The HeLMS SPIRE maps are also part of the HerMES DR3. Here we give a summary of the point source catalogue creation, as well as its completeness, accuracy and reliability. For a full description, see \cite{clarke16}\footnote{Doctoral thesis, available through Sussex Research Online: \href{http://sro.sussex.ac.uk/61474/}{http://sro.sussex.ac.uk/61474/}}. The HeLMS point source catalogue was created using the HerMES pipeline, a combination of the \textsc{starfinder} and \textsc{desphot} algorithms. \textsc{Starfinder} \citep{diolaiti00} is an iterative source-finding algorithm initially developed for crowded stellar fields. As such, \textsc{starfinder} can de-blend confused {\it Herschel} sources to find positions, but it requires the background to be mostly free from source flux. In {\it Herschel} SPIRE maps, however, the sky is comprised almost entirely of flux from sources. As such, \textsc{starfinder} does not accurately estimate fluxes. Thus, \textsc{desphot} \citep[De-blended SPIRE Photometry algorithm;][]{roseboom10,roseboom12} is used instead, with the \textsc{starfinder} objects as positional priors.
Both the HeLMS and HerS fields are contaminated by Galactic cirrus. This is in contrast to the other HerMES fields on which the source extraction techniques were developed, which are relatively free from cirrus; thus, both the maps and the source extraction techniques have to be modified. FIR emission from cirrus boosts fluxes of sources behind it such that a naive source extraction will find more and brighter sources in a cirrus-contaminated region. Cirrus has structure on large scales, as opposed to the small-scale fluctuations in the map from extragalactic sources, so emission from cirrus can be removed using a high-pass filter. For consistency, we adopt the filtering technique applied for HerS, as defined in \cite{viero14}, for the creation of the HeLMS catalogue.
The confusion noise and background have been calculated from the entire map rather than on individual tiles, into which the large map has been segmented at the time of source extraction \citep[for details see][]{viero14}. The total error on each source is given as $\sigma_{\rm{T}} = \sqrt{\sigma_{\rm{conf}}^2 + \sigma_{\rm{inst}}^2}$, with $\sigma_{\rm{inst}}^2$ as the instrumental noise. The confusion noise, $\sigma_{\rm{conf}}^2$, has been calculated on the residual map by fitting a line to the flux as a function of coverage, where the confusion noise is the y-intercept. Any source with $\chi^2 > 10$ is removed from the catalogue, leaving a total of 92,256 sources at 250\,\micron.
To test the positional and flux accuracy, we inject sources of known fluxes at random positions in the map. The maps are then run through the source extraction pipeline again. We find the positional accuracy to be within 5 arcsec in both right ascension and declination. The flux accuracy is already above 90 per cent at 30\,mJy and reaches 98 per cent at 40\,mJy.
Finally, to match the catalogues in the three SPIRE bands and assess completeness, we follow \cite{smith12}. We find a maximum completeness of 93 per cent, attained at $\sim$60\,mJy. The remaining sources are missed by the \textsc{starfinder} algorithm, but are not associated with any particular region of the map \citep[see also][]{wang14}.
\subsection{SDSS-\textbf{\textit{Herschel}} matching}\label{sec:matching}
\cite{wang14} have shown through simulations, in the fashion mentioned above, that the positional accuracy of sources in the DR2 HerMES point source catalogues also peaks at 5 arcsec. We therefore look for SDSS DR7 and DR10 quasars with $>3\sigma$ detections at 250\,$\micron$ applied across all the fields, using a 5 arcsec matching radius between the SDSS and {\it Herschel} catalogues. The results of this match are provided in Table \ref{tab:jointDetections}. Note that we rely on the individually detected objects only and do not stack the SPIRE fluxes of the non-detected SDSS quasars. Our final catalogue contains 83, 212 and 218 unique SDSS quasars (i.e. they only appear once in the catalogue even if they belong to both DR7 and DR10) in the HerMES fields, HerS and HeLMS, respectively. In what follows, we call these 513 quasars the {\it Herschel}/SDSS quasar sample.
The {\it Herschel}/SDSS quasars span the redshift range $0.1627 \le z \le 4.679$, and the bulk lie in the $M_\textit{i}$ \citep[as defined in][]{schneider10} range between -22.5 and -29. The $M_\textit{i}-z$ plane of the sample is shown in Fig. \ref{fig:MI_z}, where the DR7 and DR10 objects are marked in red circles and blue squares, respectively. The filled symbols denote BAL quasars and will be discussed separately. The histograms of the quasars belonging to the DR7 and DR10 releases, as well as that of the final sample, are shown in the top figure as filled red bars, dashed blue and solid black lines, respectively. The green spikes show the redshift distribution of the BAL quasars.
The 350 and 500\,$\micron$ fluxes are extracted on the 250\,$\micron$ positions, but we apply no cuts in signal-to-noise for either the 350 or the 500\,$\micron$ fluxes. This results in 505 detections at 350\,$\micron$ and 473 detections at 500\,$\micron$. The SPIRE flux distributions of the {\it Herschel}/SDSS quasar sample are shown in Fig. \ref{fig:flux250}.
\begin{figure}
\begin{center}
\vspace*{-2cm}
\includegraphics[width=8cm]{spireflux_dist_new.pdf}
\vspace*{-1.5cm}
\caption{SPIRE flux distributions of the {\itshape Herschel}/SDSS quasar sample. In the solid blue, filled green and dashed red histograms, we show the 250, 350 and 500\,$\micron$ fluxes, respectively.}
\label{fig:flux250}
\end{center}
\end{figure}
The \textit{Herschel} detection rates at 250\,$\micron$ of SDSS quasars in our fields average at $\sim$5 per cent, as seen when comparing the 3rd and 6th columns of Table \ref{tab:jointDetections}, with a constant detection fraction up to $M_i \sim$ -27 and an increase by a factor of $\sim$1.5 towards brighter absolute magnitudes. This number is smaller than the detection rates mentioned in previous works (e.g. \citealt{hatzimi10}, who find a 30 per cent detection rate in Lockman and FLS at 5$\sigma$ but for both type 1 and type 2 AGN combined). The lower detection rates within our sample are largely due to the fact that HerS and HeLMS, where 85 per cent our quasars are located, are shallower than are the HerMES fields. Our detection rates are instead more in line with the 8 per cent detection rate presented in \cite{bonfield11}, using a simple positional cross-match in the H-ATLAS science demonstration field.
\begin{figure*}
\begin{center}
\includegraphics[width=8cm]{nonbal_sed.pdf}
\hspace{0.5cm}
\includegraphics[width=8cm]{bal_sed.pdf}
\caption{Example SED fits. The left panel shows the fit for one of our general population quasars, while the right panel shows a fit for one of our BAL quasars. The photometric data are shown as blue triangles. The total fits are shown in solid black, while the two components of the fit are shown in the long dashed red (AGN) and the short dashed green (starburst) lines.}
\label{fig:sed}
\end{center}
\end{figure*}
\subsection{Ancillary data}\label{AncillaryData}
Of our 513 quasars, 473 are also detected by the Wide-Field Infrared Survey Explorer \citep[WISE;][]{wright10}. We also assemble near-infrared detections from three surveys: the Two Micron All Sky Survey \citep[2MASS;][]{skrutskie06}, the UKIRT Infrared Deep Sky Survey \cite[UKIDSS;][]{lawrence07}, and the VISTA Hemisphere Survey \citep[VHS;][]{mcmahon13} Data Release 3. There are 196 objects in the HerS catalogue with a measurement in one or more of the UKIDSS bands and 178 within HeLMS. There are also 64 objects (16, 14 and 34 in HerMES, HerS and HeLMS, respectively) with VHS photometry. Finally, 29 objects have a measurement in at least one of the 2MASS bands. While HerS and HeLMS also have 2MASS data available, we choose to use the UKIDSS data, as it is deeper. As a result, for more than 90 per cent of our objects, there is a minimum of 11 and a maximum of 16 photometric data points available from the optical to the FIR.
\section{Analysis}\label{sec:analysis}
\subsection{Starburst and AGN luminosities}\label{sec:sedfit}
To infer starburst and AGN luminosities, we use the multi-component SED fitting technique described in \citealt{hatzimi08,hatzimi09}. In the most general case, the method fits fluxes and their errors with three separate models: stellar, AGN, and starburst. Our sample however consists entirely of luminous type 1 quasars with extremely high SFRs. This allows us to make two key simplifications to the general method. First, it is almost certain that the stellar emission will be outshone by the AGN and/or the starburst at all wavelengths, so we do not include a stellar template in the fit. Second, since our sources are type 1 quasars, it is reasonable to conclude that the optical to mid-infrared (MIR) coverage will be dominated by the AGN \citep{hatzimi05}, and that the FIR coverage will be dominated by the starburst \citep[e.g][]{hatzimi10,ma15,harris16}. The infrared emission reprocessed by the AGN torus strongly depends on the dust geometry. In particular, the ratio between the outer and inner radius of the torus plays a major role in determining the AGN SED at longer wavelengths (in the models used in this paper the maximum value of this ratio is 150). As a result, the peak of the infrared emission of the AGN torus models lies at $10-30$\,$\micron$ (fig. 5 of \citealt{feltre12}) and the maximum width of the infrared bump (as defined in \citealt{feltre12}) for these models is $\sim 50$\,$\micron$. We therefore do not expect any significant contribution from the AGN emission at wavelengths larger than $\sim 50$\,$\micron$. Complementary, observationally motivated discussion on this point can be found in \S$5.3$ of \citealt{harris16}. We note the caveat though that, with our data, which do not cover the range between 22 and 250\,\micron, we cannot completely rule out significant FIR emission from the AGN. We thus use the shorter wavelength data to constrain the properties of the AGN, and, having done so, then use the SPIRE data to constrain the properties of the starburst. Two example SED fits are given in Fig. \ref{fig:sed}; one for a classical quasar and one for a BAL quasar.
The AGN component of the fit includes both the direct emission and the reprocessed emission, that is, optical/ultraviolet light that is absorbed by the dust in the torus and reradiated in the infrared. We call the luminosity of the best-fitting AGN model the accretion luminosity, $L_\textnormal{acc}$. We have assumed the smooth, continuous distribution of dust within the torus described in \cite{fritz06} and \cite{feltre12}. This set of models has been shown to accurately reproduce the SEDs of both obscured and unobscured AGN \citep[e.g.][]{fritz06, hatzimi08, hatzimi09, gruppioni16}. We constrained the fit to a subset of 216 torus templates (out of the $\sim$3700 models in the full grid), selecting a representative set of parameters defined in \cite{hatzimi08}. This is done to minimise degeneracies among model parameters that we do not seek to constrain, without affecting the quality of the fits for the luminosities, given that there is a large overlap in SED shapes when varying the model parameters \citep[for a description of the effect of each of the parameters on the SED, see][]{fritz06}.
The starburst component is modelled with a library of templates of six starburst galaxies: Arp220 (merger), M82 (disk/amorphous), M83 (SAB(s)c), NGC1482 (E/int), NGC4102 (Sb), and NGC7714 (Sc). The starburst templates span systems with low (e.g. M82 \& M83, \citealt{nikola12,foyle13}) to high (e.g. Arp 220, \citealt{farrah03,rang11}) extinctions and a broad range in mid- to far-IR SED shapes, star-formation histories, and Hubble types. Thus, they plausibly span at least a substantial fraction of the properties of star-forming galaxies at $z>0.5$ and have been used to model starbursts at all redshifts before \citep[see e.g.][]{feltre13}. The luminosity of the starburst component, $L_\textnormal{SB}$, is that of the best-fitting template scaled to the observed data points, integrated between 8 and 1000\,$\micron$. Finally, $L_\textnormal{SB}$ is then converted into an SFR using the original (i.e. assuming a Salpeter IMF) relation derived by \cite{kennicutt98}
\begin{equation}\label{eq:kenn}
\frac{\textnormal{SFR}}{\textnormal{M}_\odot \hspace{0.1cm} \textnormal{yr}^{-1}} = 4.5 \times 10^{-44} \frac{L_\textnormal{SB}}{\textnormal{erg} \hspace{0.1cm} \textnormal{s}^{-1}}
\end{equation}
\begin{figure}
\begin{center}
\vspace*{-2cm}
\includegraphics[width=7.5cm]{lacc_comp.pdf}\\
\vspace*{-3cm}
\includegraphics[width=7.5cm]{lsb_comp.pdf}\\
\vspace*{-1.5cm}
\caption{A comparison between the luminosities extracted from the best-fitting SEDs and from the median of their respective PDFs for $L_\textnormal{acc}$ (top) and $L_\textnormal{SB}$ (bottom). The agreement is excellent in both cases.}
\label{fig:lcomp}
\end{center}
\end{figure}
The $L_\textnormal{acc}$ and $L_\textnormal{SB}$ values derived from this approach have been shown to be robust \citep[see][]{hatzimi08}. To estimate the uncertainties on $L_\textnormal{acc}$ and $L_\textnormal{SB}$ specific to our set of templates, we calculate the median of the probability distribution function (PDF) of the values obtained using all combinations of models in the fit. Fig. \ref{fig:lcomp} compares the best fit $L_\textnormal{acc}$ (top) and $L_\textnormal{SB}$ (bottom) with the median of their respective PDFs. The root-mean-square deviation of the two sets of distributions are 7 per cent and 15 per cent, respectively. In other words, the best-fitting $L_\textnormal{acc}$ ($L_\textnormal{SB}$) differs, on average by 7 (15) per cent from the median value of the PDF.
\subsection{Black hole masses \& Eddington ratios}\label{sec:agn_prop}
We estimate black hole masses ($M_\textnormal{BH}$) from the FWHM of the C\begin{scriptsize}IV\end{scriptsize}, Mg\begin{scriptsize}II\end{scriptsize} or H$\beta$ lines (e.g. \citealt{wandel99, kaspi00, mclure02}), using:
\begin{equation}\label{eq:Mbh_form}
\frac{M_\textnormal{BH}}{\textnormal{M}_{\odot}} = \alpha \left(\frac{\lambda L_{\lambda}}{10^{37} \textnormal{W}} \right)^{\beta} \left( \frac{\textnormal{FWHM}}{\textnormal{km} \hspace{0.05cm} \textnormal{s}^{-1}} \right)^2
\end{equation}
\noindent with $\alpha$\,\,=\,\,4.6, 5.2 and 8.2, $\beta$\,\,=\,\,0.53, 0.62 and 0.50, and $\lambda$\,\,=\,\,1350 \AA, 3000 \AA \, and 5100 \AA \, for C\begin{scriptsize}IV\end{scriptsize}, Mg\begin{scriptsize}II\end{scriptsize} and H$\beta$, respectively, as described in \cite{vester06} and \cite{shen11}, the latter of which offers a recalibration of the Mg\begin{scriptsize}II\end{scriptsize} formula initially described by \cite{mclure04}. We then estimate Eddington ratio as $\lambda_{Edd} = L_\mathrm{acc}/{L_\mathrm{Edd}}$, where $L_\mathrm{Edd} $ is the Eddington luminosity, calculated as
\begin{equation}\label{eqnLed}
\frac{L_\mathrm{Edd}}{\rm{erg}\,\rm{s}^{-1}} = \frac{4 \pi G m_{p} c} {\sigma_{T}} {\it M}_{\rm{BH}} = 1.26 \times 10^{38}\frac{{\it M}_{\rm{BH}}}{\rm{M}_{\odot}}
\end{equation}
\noindent in which $\sigma_{T}$ is the Thompson scattering cross-section for the electron, $m_{p}$ is the mass of the proton, $c$ is the speed of light and $G$ the gravitational constant.
The uncertainties on the measurements of the FWHMs of C\begin{scriptsize}IV\end{scriptsize}, Mg\begin{scriptsize}II\end{scriptsize}, and H$\beta$ are of the order of 10 per cent \citep{shen11}. The same authors discuss in detail the systematic errors on $M_\textnormal{BH}$ introduced by the use of the different emission lines. They show the offset between the estimates using Mg\begin{scriptsize}II\end{scriptsize} and C\begin{scriptsize}{IV}\end{scriptsize} to be negligible. For those objects in the redshift range such that both the C\begin{scriptsize}{IV}\end{scriptsize} and Mg\begin{scriptsize}{II}\end{scriptsize} lines are present in the spectra, we use the Mg\begin{scriptsize}{II}\end{scriptsize} line as the BH mass tracer. There are no relevant emission line data available to compute the BH masses for eleven of our objects; we retain these objects in the sample but discard them from the relevant sections.
\subsection{Complementarity with other studies}\label{sec:comple}
This paper complements and builds upon other works on the star-forming properties of quasar host galaxies. We here describe how it relates to the two most closely associated works, those of \citealt{ma15} and \citealt{harris16}, hereafter MY15 and H16, respectively.
In many ways, the present work is an expansion of the work in MY15, who also study the star-forming properties of {\it Herschel}-selected SDSS quasars, using a sample of 354 quasars in all the HerMES DR2 catalogues, regardless of the depth of the fields, as well as the HerS and H-ATLAS \citep{eales10} catalogues. Our sample, on the other hand, includes the HeLMS data, never previously been used in a study of this type, as well as HerS and the wider (and shallower) northern HerMES fields listed in Sec. \ref{sec:sample}. Moreover, while MY15 match SDSS positions to the positions in the HerMES, HerS and H-ATLAS public catalogues using a matching radius of 3 arcsec, we adopted a less conservative matching radius of 5 arcsec based on the simulations in \cite{wang14}. From a visual inspection of all the sources, we find the resulting matches to be highly reliable. As a result of all the above, our {\it Herschel}/SDSS sample has 187 objects (about a third of its size) in common with the sample in MY15.
Our approach to extracting SFRs also differs to that used by MY15. MY15 considered the coldest dust component alone, by fitting a single-temperature modified black body to the SPIRE data. For testing purposes, a two-component SED fitting including an AGN and a starburst component was performed on a handful of objects with available PACS photometry. Conversely, in this work we constrain the AGN properties by focusing on the AGN-dominated part of the SED, using data in the near-IR (2MASS, UKIDSS or VHS photometry) and MIR (WISE), and by using a set of AGN models that have proved to accurately reproduce the optical-to-MIR quasar SED. Therefore, the key difference in the derived SFRs is that, in this work, the SFRs are measured on the full range 8-1000\,$\micron$, including the contribution to the MIR.
Turning to H16, our paper complements and extends this work. In H16, the authors also use {\it Herschel} data to study the star-forming properties of SDSS quasars. However, H16 focus on quasars in a narrower, systematically higher redshift range, $2<z<3$, and have a sample which is mostly individually {\itshape un}detected by {\it Herschel}. Hence their analysis relies on stacked observations rather than individual detections. At the same redshift, therefore, we study quasars with much higher SFRs on average.
\section{Results}\label{sec:results_full}
\subsection{Star formation rates}\label{sec:sfrs_full}
The distribution of the SFRs in our sample is shown in Fig. \ref{fig:sfr_full}. The SFRs are extremely high, with a mean of $1000\substack{+1400 \\ -600}$\,M$_\odot \hspace{0.1cm} \textnormal{yr}^{-1}$. This aligns well with those rates estimated for other individually FIR-detected populations, such as submillimetre galaxies (e.g. \citealt{chapman10,wardlow11,barger14}) and infrared-bright radio galaxies (e.g. \citealt{drouart14}), over a similar redshift range. The mean SFR is a factor $\sim$2 higher than that in MY15\footnote{MY15 quote 50 per cent of their objects exhibiting SFRs above 300\,M$_\odot \hspace{0.1cm} \textnormal{yr}^{-1}$, derived using Chabrier IMF. The value would be $\sim$510\,M$_\odot \hspace{0.1cm} \textnormal{yr}^{-1}$ assuming a Salpeter IMF.}. The difference is likely due to the fact that MY15 only considered SFRs measured from the coldest components, while our SFRs also include emission in the MIR.
\begin{figure}
\begin{center}
\includegraphics[width=7.5cm]{sfr-dist.pdf}
\caption{The SFR distribution for the {\it Herschel}/SDSS quasar sample. Our sample represents some of the most extreme starburst systems, as evidenced by the high inferred SFRs, which places them above the main sequence, for their respective redshifts, by over a dex on average.}
\label{fig:sfr_full}
\end{center}
\end{figure}
Next, we examine the behaviour of the SFR as a function of $L_\textnormal{acc}$ (as defined in Sec. \ref{sec:sedfit}), $M_\textnormal{BH}$, and $\lambda_\textnormal{Edd}$. To account for the fact that this is a flux-limited sample, we split the sample into three redshift bins, $z<1.0$, $1.0 \le z < 2.0$ and $z \ge 2.0$. Each redshift bin is further split into $L_\textnormal{acc}$, $M_\textnormal{BH}$, or $\lambda_\textnormal{Edd}$ bins, and the average SFR is computed in each of these bins.
We find the SFR to increase by a factor of $\sim$8 from the low ($z<1.0$) to the high ($z \ge 2.0$) redshift bins for quasars of the same $L_\textnormal{acc}$. In a given redshift interval, however, the SFR is consistent with being constant for increasing $L_\textnormal{acc}$ for all redshift bins. This is shown in the top panel of Fig. \ref{fig:SFR_trends}, where the average points are overplotted on the individual objects.
The typical uncertainties on $L_\textnormal{acc}$ and $L_\textnormal{SB}$, discussed in Sec. \ref{sec:sedfit}, are small relative to the size of the bins and have no effect on the results. The green line shows the relation described in H16 and will be discussed in Sec. \ref{sec:discussion}. The bins in Fig. \ref{fig:SFR_trends}, however, are large, such that the apparent lack of a trend might be a result of degeneracies between redshift and luminosity. We tried a finer binning, with bins half the size of the ones presented here in terms of both redshift and luminosity, and found consistent results. We also found consistent results by examining SDSS DR7 and DR10 separately. The (lack of) trends are, therefore, not a bias as far as we can tell.
\begin{figure}
\begin{center}
\includegraphics[width=9cm]{sfr-lacc.pdf} \\
\vspace*{.15cm}
\includegraphics[width=9cm]{sfr-mbh.pdf} \\
\vspace*{0.15cm}
\includegraphics[width=9cm]{sfr-edd.pdf}
\caption{SFR versus $L_\textnormal{acc}$ (top panel), $M_\textnormal{BH}$ (middle panel) and $\lambda_\textnormal{Edd}$ (bottom panel), colour-coded by redshift. The three redshift bins are $z<1.0$, $1.0 \le z < 2.0$ and $z \ge 2.0$. SFRs increase with redshift, but do not correlate with any of the intrinsic quasar properties within the same redshift range. The green lines in the top and middle panels show the relations described in \protect\cite{harris16} for quasars at $2 < z< 3$ with a factor $2-3$ lower SFRs than in our sample. The filled squares and associated error bars show the mean value in each bin. The diamonds and triangles show the results for the non-BAL and BAL subsamples, respectively.}
\label{fig:SFR_trends}
\end{center}
\end{figure}
The middle panel of Fig. \ref{fig:SFR_trends} plots SFR against $M_\textnormal{BH}$. We again see that the SFR is constant with increasing $M_\textnormal{BH}$ inside a given redshift bin. Here there is no hint of a rise in SFR; while the uncertainties on the SFR and $M_\textnormal{BH}$ are significant, the data are entirely consistent with a flat relation. Finally, in the bottom panel of Fig. \ref{fig:SFR_trends}, we plot SFR against the Eddington ratios of the quasars. We again see no relation.
\subsection{Broad absorption line quasars}\label{sec:BAL}
A quasar is classified as BAL quasar if the equivalent width of at least one absorption line (usually C\begin{scriptsize}{IV}\end{scriptsize}) exceeds 2000\,$\textnormal{km} \hspace{0.1cm} \textnormal{s}^{-1}$ \citep{weymann91}. BAL quasars can be further split into three sub-populations. The most numerous of these is that of the high-ionisation BAL (HiBAL) quasars. HiBAL quasars display absorption in Ly$\alpha \hspace{0.025cm} \lambda$1216\textup{\AA}, N\begin{scriptsize}{V}\end{scriptsize}$\hspace{0.025cm} \lambda$1240\textup{\AA}, Si\begin{scriptsize}{IV}\end{scriptsize}$\hspace{0.025cm} \lambda$1394\textup{\AA} and C\begin{scriptsize}{IV}\end{scriptsize}$\hspace{0.025cm} \lambda$1549\textup{\AA}. In addition to the absorption features found in HiBAL quasars' spectra, low-ionisation BAL (LoBAL) quasars also display absorption features in such low-ionisation ions as Al\begin{scriptsize}{III}\end{scriptsize}$\hspace{0.025cm} \lambda$1857\textup{\AA} and Mg\begin{scriptsize}{II}\end{scriptsize}$\hspace{0.025cm} \lambda$2799\textup{\AA}. The third class, FeLoBAL quasars, shares the same properties as LoBAL quasars with additional absorption features from excited levels of iron.
\begin{table}
\begin{center}
\begin{tabular}{L{1.15cm} C{2cm} C{2.05cm}}
\hline
Field & SDSS \hspace{3cm} BAL quasars & \textit{Herschel}/SDSS BAL quasars \\
\hline
\noalign{\vskip0.1cm}
HerMES & 92 & 7 \\
HerS & 344 & 20 \\
HeLMS & 666 & 32 \\
\hline
\end{tabular}
\end{center}
\caption{BAL detection statistics by \textit{Herschel} field. The average {\it Herschel} detection rate of BAL quasars is about 5 per cent, the same as for non-BAL quasars, as shown in Table \ref{tab:jointDetections}.}
\label{tab:bal_stats}
\end{table}
We select BAL quasars using the SDSS \texttt{BAL\char`_FLAG}, which is set to 1 when C\begin{scriptsize}{IV}\end{scriptsize} has an absorption feature, 2 when Mg\begin{scriptsize}{II}\end{scriptsize} has one and 3 when both C\begin{scriptsize}{IV}\end{scriptsize} and Mg\begin{scriptsize}{II}\end{scriptsize} show absorption. Table \ref{tab:bal_stats} shows the number of BAL quasars in each of the three \textit{Herschel} surveys with at least a $3\sigma$ detection at 250\,$\micron$. The {\it Herschel} detection rates of BAL quasars are around 5 per cent, identical to those of non-BAL quasars. Of the 59 BAL quasars detected in the three {\it Herschel} fields, 57 are HiBAL quasars and two are LoBAL quasars. As HiBAL and LoBAL quasars may be physically different \citep[see e.g.][]{streb10}, we restrict our analysis to the 57 HiBAL quasars.
We create two sub-samples from the sources in the rectangle in Fig. \ref{fig:MI_z}. The lower redshift cut ($z$=1.5) corresponds approximately to the wavelength at which the C\begin{scriptsize}{IV}\end{scriptsize} line shifts out of the SDSS bandpass, eliminating two of the BAL quasars, while the upper redshift cut ($z$=3.3) is set where the quasar density drops significantly, and only leaves out one BAL quasar from the full set. The corresponding bright and faint absolute magnitude cuts are -28.5 and -24.3, respectively, leaving another three BAL quasars out. A total of 51 BAL and 208 non-BAL quasars reside in the defined area.
Looking at Fig. \ref{fig:SFR_trends}, there is no preferred location for the BAL quasars in any of the parameter spaces considered (the lack of BAL quasars with low SFR is due to the lower redshift cut). To further test if the behaviour of these BAL quasars is identical to that of the non-BAL quasars, we perform K-S tests to compare the $L_\textnormal{acc}$, $\lambda_\textnormal{Edd}$ and SFR distributions of the two subsets. The BAL quasars have indistinguishable $L_\textnormal{acc}$ and $\lambda_\textnormal{Edd}$ distribution from the non-BAL quasars, as indicated by the $p$\,-\,values from the K-S test (0.47 and 0.46, respectively). Lastly, their SFR distributions are also indistinguishable ($p$\,-\,value = 0.46). We thus find no evidence that the accretion rate, accretion efficiency or star forming properties of HiBAL quasars are different from those of classical quasars.
\section{Discussion \& Conclusions}\label{sec:discussion}
The relationship between SFRs and AGN properties gives insights into the co-evolution of these activities. While we do not have stellar masses for our sample, their extremely high SFRs coupled with reasonable assumptions for quasar host stellar masses at this epoch mean they almost certainly lie above the `main sequence' for star formation at their respective epochs. In the following, we consider the relationship between SFRs and both redshift and AGN properties. Finally, we discuss the implications for BAL quasars.
\subsection{Evolution with redshift}\label{sec:disczevo}
We find that the SFRs of our sample increase by a factor of up to $\sim$\,8 from $z<1$ to $z \ge 2$. We do not see any evidence for a rise in SFRs with redshift beyond $z\sim2$.
This is consistent with observations of the global comoving SFR density, which increases by a factor of at least ten over $0<z<1$ and peaks around $z=2$ \citep[e.g.][]{reddy08, magn09, karim11, wuyts11, bethermin13, burgarella13, wang13, chen16}. Thus, in terms of evolution over $0\lesssim z \lesssim 3$, extremely luminous star formation events in quasar hosts appear to evolve, on average, in a manner consistent with the global star formation history as derived from the general galaxy population.
Our results are also consistent with those of MY15, who find that the fraction of IR-luminous quasars peaks at $z\sim2$, and with H16, who find that SFRs in quasars do not change appreciably as a function of redshift over $2<z<3$ (but see also e.g. \citealt{netzer16}). They are also consistent with several previous studies of the evolution with redshift of star formation in AGN hosts \citep[e.g.][]{serjeant09,serjeant10,shao10,bonfield11,mullaney12}. These previous studies mostly sample different parts of the $z$-SFR-$L_\textnormal{acc}$ parameter space for quasars to our study. For example, the sample in \citealt{mullaney12} spans a similar redshift range but has lower $L_\textnormal{acc}$ values and SFRs by $\gtrsim 2$ orders of magnitude, while the sample in H16 has comparable $L_\textnormal{acc}$ values, but lies at higher redshifts with a factor $\sim 3$ lower SFRs. The apparent consistency of our results with theirs on the evolution of SFRs with redshift in quasar hosts is thus remarkable. This consistency suggests that the physical processes that lead to star formation in AGN hosts, such as mergers or secular evolution, do so in a way that gives a similar evolution with redshift for star formation events spanning tens to thousands of $M_{\odot}$ per year and accretion luminosities $\gtrsim10^{10}$\,L$_\odot$. We caution however that this does not account for selection effects between studies, and only applies to time-scales significantly longer than the AGN and starburst duty cycles.
\subsection{Star formation and accretion luminosity}\label{sec:discsfagn}
We first consider the relation between SFR and $L_\textnormal{acc}$. The top panel of Fig. \ref{fig:SFR_trends} is consistent with an approximately flat $\textnormal{SFR} - L_\textnormal{acc}$ relation, in all redshift bins. This is in agreement with some previous studies \citep[e.g][]{priddey03, shao10, dicken12, mullaney12, harrison12, rosario13, ma15, baner15, stanley15}, but not others \citep[e.g.][]{netzer09, hatzimi10, ima11, rafferty11, mull12a, chen13, Young2014, delv15, xu15, bian16}.
Several explanations for an observed lack of a relation between SFR and $L_\textnormal{acc}$ have been proposed. They fall, broadly, into four categories. First, the strength of the relation depends on redshift, with an observable relation only emerging at $z\gtrsim2$ \citep{hatzimi10, rovilos12, delv15, harris16}. This could arise due to, for example, a higher free gas fraction at higher redshifts. Second, AGN luminosities can vary substantially on time-scales of days to months. If AGN luminosities are measured using methods that are particularly sensitive to such variations, such as X-ray observations \citep{barr86,ulr97,grupe01,chitnis09}, then underlying trends could be masked \citep{gabor13, volon15}. Third, it is possible that at very high SFRs and/or $L_\textnormal{acc}$ values, we are sampling an intrinsically different quasar population, one in which e.g. a different trigger for star formation leads to a weaker or absent correlation between SFR and $L_\textnormal{acc}$. Fourth, at high SFRs or high $L_\textnormal{acc}$, the properties of the starburst become decoupled from those of the AGN, plausibly because internal self-regulation processes become the dominant effect in regulating luminosities on time-scales comparable to or shorter than the AGN or starburst duty cycles.
The first possibility is, {\itshape on its own}, not plausible for our sample, as our sample spans a wide range in redshift and we do not see a significant correlation in any redshift bin. The second seems unlikely, since our sample is large and has AGN luminosities measured from optical, rather than X-ray data. We note however that there exists a conceptually similar possibility - that the duty cycles of starburst and AGN activity are sufficiently mismatched in duration that no correlation is observed between their luminosities. The available constraints on the duty cycle lengths of starburst and AGN phases are however not accurate enough for us to comment on this possibility. The third possibility also seems unlikely, since we found no differences between the {\it Herschel}/SDSS sample and the general SDSS quasar population in Section \ref{sec:sdss}, though we cannot formally exclude it.
The fourth scenario, however, is plausible, since our sample harbours the highest SFRs seen in quasars, where self-regulation may be expected to be seen, if it occurs. Moreover, it is consistent with observations of systems with comparable SFRs at $z>1$, i.e. submillimetre galaxies (SMGs, \citealt{barger14}), which exhibit a flattening in their SFRs at extremely high SFR values. An interesting corollary comes from the work of MY15, who find that star-forming regions in IR-luminous quasars do not increase in size at very high SFRs, but instead may increase in SFR {\itshape density}. An increase in SFR density could mean that starburst self-regulation becomes more effective, via e.g. approaching the Eddington limit for star formation, where winds from supernovae can expel free gas and thus stall star formation \citep[e.g.][]{thom05,murray05,dia12}. We thus propose that the flat relation between SFR and $L_\textnormal{acc}$ in our sample arises because self-regulation by the starburst becomes the dominant factor controlling the SFRs. We cannot, however, rule out a contribution from redshift-driven effects, or from mismatches in the lengths of duty cycles of starburst and AGN episodes.
The top panel of Fig. \ref{fig:SFR_trends} also shows no downturn in SFRs at the highest $L_\textnormal{acc}$ values. This is, in principle, inconsistent with the idea that luminous AGN can quench star formation in their host galaxies (see also e.g. \citealt{vil16}). An alternative interpretation is however possible; {\itshape assuming that AGN feedback occurs}, this instead suggests that $L_\textnormal{acc}$ (as derived from rest-frame UV observations) is not a good proxy for the strength of AGN feedback. This is consistent with AGN feedback being a brief phase in the AGN duty cycle, signposted by properties other than AGN luminosity \citep[e.g.][]{farrah12}.
\subsection{Star formation, black hole masses, and Eddington ratios}\label{sec:discsfagn2}
The middle panel of Fig. \ref{fig:SFR_trends} is consistent with there being no relation between SFR and $M_\textnormal{BH}$. This lack of a correlation is, in some senses, surprising. For example, it has been suggested that SFRs should scale with $M_\textnormal{BH}$ in quasars since SFRs should scale with the total available free gas mass or its large-scale distribution \citep{cen15}, and both the free gas mass and $M_\textnormal{BH}$ should scale with the total baryon density. Thus, on $\gtrsim100$\,Myr time-scales, SFRs and $M_\textnormal{BH}$ in quasars may be expected to correlate since a larger free baryon reservoir would favor both higher current SFRs and higher $L_\textnormal{acc}$ {\itshape in the past}. We, however, do not see such a correlation.
A plausible explanation comes from considering our results in context with those of H16, who found that SFRs in quasars at $2<z<3$ do correlate with $M_\textnormal{BH}$ (and $L_\textnormal{acc}$), but only up to an SFR of $\sim600$\,M$_{\odot}$yr$^{-1}$. Above this SFR, they find no correlation. Our SFRs are mostly higher than those in H16 and are measured individually, rather than stacked as in H16, so it is interesting to compare their results to ours. In the top and middle panels of Fig. \ref{fig:SFR_trends}, we plot the relations that H16 derive between SFR and $L_\textnormal{acc}$ and $M_\textnormal{BH}$, respectively. For the SFR-$L_\textnormal{acc}$ plot our results are, given the sizes of the error bars on the average points in each redshift bin, consistent with the H16 relation. For the SFR-$M_\textnormal{BH}$ plot the consistency is weaker, but fig. 17 of H16 shows, plausibly, a flattening in the SFR-$M_\textnormal{BH}$ relation at about $600$\,M$_{\odot}$yr$^{-1}$ as well. Overall therefore, we argue that our results and those in H16 are both consistent with the idea that SFRs in quasars correlate with $L_\textnormal{acc}$ and $M_\textnormal{BH}$, but only up to an SFR when internal `self-regulation' by the starburst becomes important.
The bottom panel of Fig. \ref{fig:SFR_trends} is consistent with there being no relation between SFR and $\lambda_\textnormal{Edd}$, at any redshift. This result is also in line with those in H16, who also find no evidence for an $\textnormal{SFR} - \lambda_\textnormal{Edd}$ relation, at any SFR (modulo possible enhanced SFRs at low $\lambda_\textnormal{Edd}$ values. Overall, this is consistent with there being no relation between SFRs in quasars and how efficiently the black hole is accreting, at {\itshape any} SFR. It is also consistent with the suggestion that $\lambda_\textnormal{Edd}$ values in quasars do not depend strongly on redshift, at least at $z\gtrsim0.5$ (e.g. \citealt{shan13,capl15}, see also \citealt{fine06,bluck11,luss12}).
\subsection{Star formation in broad absorption-line quasars }\label{sec:dischibal}
We find that there is no difference in the SFR, $L_\textnormal{acc}$, $M_\textnormal{BH}$, and $\lambda_\textnormal{Edd}$ distributions between the HiBAL and non-BAL quasars in our sample; all are consistent with being drawn from the same parent population. A minor caveat to this result is that our sample may be incomplete for HiBALs at very high redshifts (Sec. \ref{sec:sdss}), an effect that we cannot account for using available data; we think it unlikely that this incompleteness could be a significant factor leading to the lack of differences between BAL and non-BAL quasars that we observe.
This result aligns with previous studies, which find no differences between HiBAL and classical quasars, at any redshift \citep{priddey03,priddey07,gallagher07,pu15,harris16}. In particular, it is consistent with the study of \citealt{cao12}, who also use {\it Herschel}-SPIRE data to compare the properties of BAL versus non-BAL quasars, from the H-ATLAS survey, and find no differences between the two populations. The two studies, however, sample different regimes in SFR; our study examines quasars with typical SFRs of $\sim1000$\,M$_{\odot}$yr$^{-1}$, whereas the typical SFRs of the quasars in \citealt{cao12} are $\sim240$\,M$_{\odot}$yr$^{-1}$. The combined results suggest that BALs are not seen preferentially in certain $L_\textnormal{acc}$ or $M_\textnormal{BH}$ regimes, and not over specific ranges in SFR, even for those quasars harbouring the most luminous star formation events seen in any quasar at any redshift.
It has been argued that BAL quasars may be a promising quasar population within which to look for evidence of AGN feedback, since the BAL winds provide a natural mechanism to couple momentum from accretion-disk winds to ISM gas. The lack of any differences between the HiBALs and the non-BALs in our sample however argues that HiBAL quasars are, as a population, not sites for AGN feedback, unless the AGN feedback phase is much shorter than the lifetime of the BAL winds. Instead, our results are consistent with HiBAL quasars being normal quasars observed along a particular line of sight, with the outflows in HiBAL quasars not having any measurable effect on the star formation in their hosts (see also e.g. \citealt{vio16}).
\section*{ACKNOWLEDGEMENTS}
Funding for the creation and distribution of the SDSS Archive has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Aeronautics and Space Administration, the National Science Foundation, the U.S. Department of Energy, the Japanese Monbukagakusho and the Max Planck Society. The SDSS website is http://www.sdss.org/. The results described in this paper are based on observations obtained with {\it Herschel}, an ESA space observatory with science instruments provided by European-led Principal Investigator consortia and with important participation
from NASA. SPIRE has been developed by a consortium of institutes led by Cardiff Univ. (UK) and including: Univ. Lethbridge (Canada); NAOC (China); CEA, LAM (France); IFSI, Univ. Padua (Italy); IAC (Spain); Stockholm Observatory (Sweden); Imperial College London, RAL, UCL-MSSL, UKATC, Univ. Sussex (UK); and Caltech, JPL, NHSC, Univ. Colorado (USA). This development has been supported by national funding agencies: CSA (Canada); NAOC (China); CEA, CNES, CNRS (France); ASI (Italy); MCINN (Spain); SNSB (Sweden);
STFC, UKSA (UK); and NASA (USA). This research has made use of data from the HerMES project (http://hermes.sussex.ac.uk/). HerMES is a {\it Herschel} Key Programme utilizing Guaranteed Time from the SPIRE instrument team, ESAC scientists, and a mission scientist. HerMES is described in \cite{oliver12}. The HerMES data presented in this paper are available through the {\it Herschel} Database in Marseille (http://hedam.lam.fr/HerMES/). AF acknowledges support from the ERC via an Advanced Grant under grant agreement no. 321323 – NEOGAL. This work makes extensive use of TOPCAT \citep{taylor05}.
|
\section{Introduction} \label{sec:intro}
The anomalous magnetic moment of the muon $a_\mu \equiv (g-2)_\mu/2$ is among the most precisely known quantities in the Standard Model (SM), and therefore, provides us with a sensitive probe of new physics beyond the SM (BSM)~\cite{Czarnecki:2001pv, Jegerlehner:2009ry}.
There is a long-standing $3.6\, \sigma$ discrepancy between the SM prediction~\cite{Hagiwara:2011af, Aoyama:2012wk, Kurz:2016bau} and the measured value of $a_\mu$~\cite{Agashe:2014kda}:
\begin{equation} \label{eq:gm2}
\Delta a_\mu \ \equiv \ a_\mu^\text{exp} - a_\mu^\text{SM} \ \simeq \ (288 \pm 80) \times 10^{-11}\, .
\end{equation}
The uncertainties in the experimental measurement, which come from the E821 experiment at BNL~\cite{Bennett:2006fi}, can be reduced by about a factor of four in the upcoming Muon $g-2$ experiment at Fermilab~\cite{Grange:2015fou}. If comparable progress can be made in reducing the uncertainties of the SM prediction~\cite{Blum:2013xva, Blum:2015gfa, Blum:2015you, Chakraborty:2016mwy}, we will have a definite answer to the question whether or not $\Delta a_\mu$ is evidence for BSM physics.
Thus from a theoretical point of view, it is worthwhile investigating simple BSM scenarios which can account for the $(g-2)_\mu$ anomaly, should this endure, and at the same time, have complementary tests in other ongoing and near future experiments.
With this motivation, we discuss here a simple $Z^\prime$ interpretation of the $(g-2)_\mu$ anomaly.
A sufficiently muonphilic $Z^\prime$ can address the $(g-2)_\mu$ discrepancy~\cite{Foot:1994vd, Gninenko:2001hx, Murakami:2001cs, Baek:2001kca, Ma:2001md, Pospelov:2008zw, Heeck:2011wj, Davoudiasl:2012ig, Carone:2013uh, Harigaya:2013twa, Altmannshofer:2014cfa, Tomar:2014rya, Altmannshofer:2014pba, Lee:2014tba, Allanach:2015gkd, Heeck:2016xkh, Patra:2016shz}; however, in order to avoid stringent bounds from the charged lepton sector, while being consistent with a sizable contribution to $(g-2)_\mu$, the $Z^\prime$ coupling must violate lepton universality.\footnote{There are other experimental hints of lepton flavor violation or the breakdown of lepton flavor universality in processes involving muons and taus, e.g. in $B^+\to K^+\ell^+ \ell^-$ decays at the LHCb~\cite{Aaij:2014ora}, in $B \to D^{(*)} \tau \nu$ decays at BaBar~\cite{Lees:2012xj}, Belle~\cite{Huschle:2015rga, Abdesselam:2016cgx} and LHCb~\cite{Aaij:2015yra}, and in the $h\to \mu\tau$ decay at both CMS~\cite{Khachatryan:2015kon} and ATLAS~\cite{
Aad:2015gha} (which however seems to have disappeared in the early run-II LHC data~\cite{CMS:2016qvi, Aad:2016blu}). See e.g. Refs.~\cite{Boucenna:2016wpr, Buttazzo:2016kid, Altmannshofer:2016oaq, Nandi:2016wlp, Bauer:2015knc, Das:2016vkr, Wang:2016rvz, Tobe:2016qhz} for the most recent attempts to explain some of these anomalies. In this work we concentrate
on $(g-2)_\mu$ and only comment on $h\to \mu\tau$.}
For instance, a sizable $Z^\prime$ coupling to electrons is strongly constrained over a large range of $Z^\prime$ masses from $e^+e^- \to e^+e^-$ measurements at LEP~\cite{Schael:2013ita}, electroweak precision tests~\cite{Hook:2010tw, Curtin:2014cca}, $e^+e^-\to \gamma \ell^+\ell^-$ (with $\ell=e,\mu$) at BaBar~\cite{Lees:2014xha}, $\pi^0\to \gamma \ell^+\ell^-$ at NA48/2~\cite{Batley:2015lha}, the $g-2$ of the electron~\cite{Pospelov:2008zw}, and neutrino-neutrino scattering in supernova cores~\cite{Manohar:1987ec, Dicus:1988jh}. Similarly, a
sizable flavor-diagonal $Z'$ coupling to muons is strongly constrained from neutrino trident production $\nu_\mu N\to \nu_\mu N \mu^+\mu^-$~\cite{Altmannshofer:2014pba} using the CCFR data~\cite{Mishra:1991bv}.
In addition, charged lepton flavor-violating (LFV) processes, such as $\mu\to e\gamma$, $\mu\to 3e$, $\tau\to \mu\gamma$, $\tau\to 3e$, $\tau\to ee\mu$, $\tau\to e\mu\mu$, constrain all the lepton-flavor-diagonal couplings of the $Z^\prime$, as well as the flavor off-diagonal couplings to electrons and muons~\cite{Farzan:2015hkd, Cheung:2016exp, Yue:2016mqm, Kim:2016bdu}. There also exist stringent LHC constraints from di-lepton resonance searches: $pp\to Z' \to ee,\mu\mu$~\cite{CMS:2015nhc, Aaboud:2016cth}, $\tau\tau$~\cite{CMS:2016zxk} and $e\mu$~\cite{atlas:emu, CMS:2016dfe}. All these constraints require the flavor-diagonal $Z'$ couplings, as well as the flavor off-diagonal couplings involving electrons to be very small, or equivalently, push the $Z^\prime$ mass scale to above multi-TeV range~\cite{Langacker:2008yv}.
We propose a simplified leptophilic $Z^\prime$ scenario with {\it only} a flavor off-diagonal coupling to the muon and tau sector [see Eq.~\eqref{lagZp} below], which trivially satisfies all the above-mentioned constraints, and moreover, can be justified from symmetry arguments, as discussed below.
In such a scenario, we find that the most relevant constraints come from leptonic $\tau$ decays in low-energy precision experiments, and to some extent, from the leptonic decays of the SM $W$ boson at the LHC. In particular, we show that the $(g-2)_\mu$ anomaly can be accounted for only with $m_{Z^\prime}>m_\tau-m_\mu$ and by allowing a larger $Z^\prime$ coupling to the right-handed charged-leptons than to the left-handed ones, whereas the lighter $Z'$ scenario (with $m_{Z'}<m_\tau-m_\mu$) is ruled out completely from searches for $\tau\to \mu$+invisible decays.
We emphasize that the entire allowed range can likely be tested in future low-energy precision measurements of lepton flavor universality in $\tau$ decays at Belle 2, as well as in the leptonic decay of the $W$ boson at the LHC. A striking four-lepton collider signature consisting of like-sign di-muons and like-sign di-taus can be probed at the high luminosity phase of the LHC (HL-LHC) as well as at a future electron-positron collider running at the $Z$ pole. We also point out an interesting possibility for the detection of our flavor-violating $Z^\prime$ scenario by the scattering of ultra-high energy neutrinos off lower-energy neutrinos, which leads to characteristic spectral absorption features that might be observable in large volume neutrino telescopes like IceCube and KM3NeT.
The rest of the paper is organized as follows: in Section~\ref{sec:model}, we present our phenomenological model Lagrangian, which can be justified in a concrete BSM scenario. In Section~\ref{sec:gm2}, we show how the $(g-2)_\mu$ anomaly can be resolved in our LFV $Z'$ scenario. Section~\ref{sec:lfv} discusses the lepton flavor universality violating tau decays for $Z'$ masses larger than the tau mass. Section~\ref{sec:2body} discusses the two-body tau decays for a light $Z'$. In Section~\ref{sec:lhc}, we derive the LHC constraints on our model from leptonic $W$ decays. Section~\ref{sec:lep} derives the LEP constraints from $Z$-decay measurements. Section~\ref{sec:4lepton} presents a sensitivity study for the new collider signature of this model. Section~\ref{sec:icecube} discusses some observational prospects of the $Z'$ effects in neutrino telescopes. Our conclusions are given in Section~\ref{sec:concl}.
\section{A Simplified Model} \label{sec:model}
Our simplified model Lagrangian for the $Z'$ coupling exclusively to the muon and tau sector of the SM is given by
\begin{eqnarray}
\label{lagZp}
\mathcal L_{Z^\prime} & \ = \ & g_L^\prime \big( \bar \mu \gamma^\alpha P_L \tau + \bar \nu_\mu \gamma^\alpha P_L \nu_\tau \big) Z^\prime_\alpha \nonumber \\
&& \qquad +\, g_R^\prime \big(\bar \mu \gamma^\alpha P_R \tau\big) Z^\prime_\alpha + {\rm H.c.}\, ,
\end{eqnarray}
where $P_{L,R}=(1\mp\gamma^5)/2$ are the chirality projection operators.
Due to $SU(2)_L$ invariance, the couplings of the left-handed neutrinos and charged leptons are identical, whereas we do not introduce right-handed neutrinos in order to keep the model minimal.
The left-handed and right-handed couplings $g_L^\prime$ and $g_R^\prime$ could in principle contain $C\!P$ violating phases. We will take into account the complex nature of these couplings in all the equations below; in our numerical analysis however, we will take them to be real for simplicity. We allow different LFV couplings of the $Z'$ to left- and right-handed charged leptons, which will be crucial for the $(g-2)_\mu$ explanation.
We assume the $Z'$ can acquire mass from the spontaneous breaking of some extra $U(1)'$ symmetry, under which it is charged. The details of the mechanism that generates the $Z^\prime$ mass are irrelevant for our phenomenological purposes, and we treat $m_{Z'}$ as a free parameter in the following.
Since $U(1)_Y$ is the only flavor-blind $U(1)$ symmetry that is anomaly-free with the SM field content,
the advantage of the extra $U(1)'$ is that the associated $Z'$ can couple differently to different SM fermion families.
As mentioned above, most of the existing experimental constraints involve first generation fermions, which may be regarded as more `fundamental' in the sense that these comprise ordinary matter around us. Thus, we assume that the couplings of the $Z^\prime$ to the first generation fermions are vanishingly small or non-existent~\cite{Kile:2014jea}, so that all these stringent experimental constraints are readily avoided.\footnote{This can be realized, for instance, in concrete models with a gauged $U(1)_{L_\mu-L_\tau}$ symmetry~\cite{Baek:2001kca, Ma:2001md, Harigaya:2013twa, Altmannshofer:2014cfa, Patra:2016shz}, which is in fact the only anomaly-free $U(1)$ group with nonzero charge assignments to SM neutrinos that can lead to
an experimentally viable light $Z'$ without requiring the addition of any exotic fermions.
Another possibility is a $U(1)$ group charged under only muon or tau number, but this requires new chiral fermions charged under both $SU(2)_L$ and $U(1)_Y$, as well as under the new $U(1)_\mu$ or $U(1)_\tau$ group.}
If the $Z'$ does not couple universally to quarks, there will be no Glashow-Iliopoulos-Maiani (GIM) suppression of the flavor changing neutral current (FCNC) processes in the quark sector, and the current experimental bounds on neutral meson mixing, such as $K-\bar{K},\, D_0-\bar{D}_0, B_d-\bar{B}_d,\, B_s-\bar{B}_s$~\cite{Isidori:2014rba, Altmannshofer:2014cfa}, as well as FCNC decays of the top, bottom and strange quarks~\cite{Altmannshofer:2014cfa, Fuyuto:2015gmk} will force the $Z'$ couplings to be rather small.
Therefore, we will assume that the $Z'$ in our case is leptophilic, and more specifically, couples only to second and third generation leptons.
The phenomenological Lagrangian in Eq.~\eqref{lagZp} can then be justified by imposing an exact discrete symmetry under which~\cite{Foot:1994vd}
\begin{align}
L_\mu \ & \leftrightarrow \ L_\tau \, , \qquad \mu_R \ \leftrightarrow \ \tau_R \, , \nonumber \\
B^\alpha \ & \leftrightarrow \ B^\alpha \, , \qquad Z'^\alpha \ \leftrightarrow \ -Z'^\alpha \, ,
\label{discrete}
\end{align}
where $L_\ell \equiv (\nu_\ell, \, \ell)_L$ and $\ell_R$ are respectively the usual $SU(2)_L$ lepton doublets and singlets in the SM in the gauge eigenstate basis and $B^\alpha$ is the $U(1)_Y$ gauge field.\footnote{The discrete charge assignment in Eq.~\eqref{discrete} would require an extended Higgs sector to give masses to all the charged leptons~\cite{Foot:1994vd}, but this does not affect the $Z'$ phenomenology discussed here.}
Since the $B^\alpha$ gauge field, and hence, the photon and $Z$ fields are even under the discrete symmetry, we can forbid kinetic $Z-Z'$ mixing and $\gamma - Z^\prime$ mixing to all orders, thus removing a few more stringent experimental constraints, e.g.~from neutrino-electron scattering~\cite{Laha:2013xua} and beam dump experiments~\cite{Alekhin:2015byh}.
\section{Muon Anomalous Magnetic Moment} \label{sec:gm2}
The flavor-violating $Z^\prime$ coupling in Eq.~\eqref{lagZp} gives rise to a new contribution to $(g-2)_\mu$, as shown in Fig.~\ref{fig:1}, and is given by the general expression~\cite{Leveille:1977rc}\footnote{A typo in Ref.~\cite{Leveille:1977rc} is corrected in the second line of Eq.~\eqref{gm2}.}
\begin{align}
a_\mu \ & = \ \frac{m_\mu^2}{4\pi^2}\int_0^1 dx\Bigg[C_V^2\bigg\{(x-x^2)\left(x+\frac{2m_\tau}{m_\mu}-2\right)\nonumber \\
& \quad -\frac{x^2}{2m_{Z'}^2}(m_\tau-m_\mu)^2\left(x-\frac{m_\tau}{m_\mu}-1\right)\bigg\}\nonumber \\
& \quad +C_A^2\bigg \{m_\tau\to -m_\tau\bigg \}\Bigg]\nonumber \\
& \times \Big[m_\mu^2 x^2+m_{Z'}^2(1-x)+x(m_\tau^2-m_\mu^2)\Big]^{-1} \, ,
\label{gm2}
\end{align}
where $C_V=|g'_R+g_L'|/2$ and $C_A=|g'_R-g_L'|/2$ in the notation of Eq.~\eqref{lagZp}. For $m_{Z^\prime} \gg m_\tau$, this reduces to
\begin{equation}
a_\mu \ \simeq \ \frac{1}{12\pi^2} \frac{m_\mu^2}{m_{Z^\prime}^2} \left[ 3\: \text{Re}(g_L^\prime g_R^{\prime *}) \frac{m_\tau}{m_\mu} - |g_L^\prime|^2 - |g_R^\prime|^2 \right] \, ,
\label{eq:g-2}
\end{equation}
Note that in the presence of both left-handed and right-handed couplings, the contributions of the flavor changing $Z^\prime$ are enhanced by a factor $m_\tau / m_\mu$. This is in contrast to contributions from flavor-blind new physics, that do not enjoy such an enhancement. Moreover, a purely left-handed or right-handed coupling would lead to a negative contribution to $a_\mu$, thus making the $\Delta a_\mu$ discrepancy worse than in the SM.
\begin{figure}[t]
\centering
\includegraphics[width=4cm]{feyn1n.pdf}
\caption{Feynman diagram for the $Z'$ contribution to the anomalous magnetic moment of the muon in our model.}
\label{fig:1}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.46\textwidth]{couplings.png}
\caption{The $g_L^\prime$ vs. $g^\prime_R$ plane for $m_{Z^\prime} = 100$~GeV. The green band is preferred at $2\, \sigma$ by the $(g-2)_\mu$ anomaly, whereas the gray region is disfavored at $>5 \, \sigma$ (see Section~\ref{sec:gm2}). The red region is excluded by lepton flavor universality in tau decays (see Section~\ref{sec:lfv}). The dashed red contours show values of constant lepton flavor universality violation in tau decays. The black dashed curve shows the 95\% CL LHC exclusion from searches for leptonic $W$ decays (see Section~\ref{sec:lhc}) and the purple dashed curve shows the 95\% CL LEP exclusion from $Z$ coupling measurements (see Section~\ref{sec:lep}). The orange dotted curve shows the expected $3\sigma$ sensitivity to the process $pp \to \mu^\pm \mu^\pm \tau^\mp \tau^\mp$ at the high-luminosity LHC (see Section~\ref{sec:4lepton}).}
\label{fig:2}
\end{figure}
\begin{figure*}[t!]
\centering
\includegraphics[width=0.46\textwidth]{gminus2_a.png}
\includegraphics[width=0.46\textwidth]{gminus2_b.png}
\caption{Slices of $m_{Z^\prime}$ vs. $g^\prime_R$ parameter space. The left-handed coupling is set to $g_L^\prime = g_R^\prime$ in the left panel and $g_L^\prime = g_R^\prime / 10$ in the right panel. The green band is the $2\, \sigma$-preferred range by the $(g-2)_\mu$ anomaly, while the gray region is disfavored at $> 5 \, \sigma$ (see Section~\ref{sec:gm2}). The red region is excluded at $2\, \sigma$ by lepton flavor universality in tau decays (see Section~\ref{sec:lfv}). The blue region is excluded at 95\% CL by searches for the two-body decay $\tau \to \mu Z^\prime$ (see Section~\ref{sec:2body}). The black dashed curve shows the 95\% CL LHC exclusion from searches for leptonic $W$ decays (see Section~\ref{sec:lhc}) and the purple dashed curve shows the 95\% CL LEP exclusion from $Z$ coupling measurements (see Section~\ref{sec:lep}). The orange and blue dotted lines show the expected $3\sigma$ sensitivity in searches for the $\mu^\pm \mu^\pm \tau^\mp \tau^\mp$ final state at the high-luminosity LHC and
at a future electron-positron collider running at the $Z$ pole (see Section~\ref{sec:4lepton}).}
\label{fig:3}
\end{figure*}
In Figs.~\ref{fig:2} and~\ref{fig:3} we show regions of parameter space that allow to address the $(g-2)_\mu$ discrepancy. The plot in Fig.~\ref{fig:2} shows the $g_L^\prime$ vs. $g_R^\prime$ plane for a fixed $Z^\prime$ mass $m_{Z^\prime} = 100$~GeV; the plots in Fig.~\ref{fig:3} show the $m_{Z^\prime}$ vs. $g_R^\prime$ plane for two choices of $g_L^\prime$, namely, $g_L^\prime = g_R^\prime$ (left) and $g_L^\prime = g_R^\prime /10$ (right). The green bands correspond to the $2\,\sigma$ preferred region from Eq.~(\ref{eq:gm2}). In the gray regions, the discrepancy is larger than $5\,\sigma$ which we consider to be excluded.
Note that both left-handed and right-handed couplings are required to explain the anomaly.
Pure left-handed or pure right-handed couplings of the $Z^\prime$ necessarily enlarge the discrepancy in $(g-2)_\mu$, as seen from Eq.~\eqref{eq:g-2}, and hence, are not entertained here. Other constraints shown in Figs.~\ref{fig:2} and \ref{fig:3} are explained below.
\section{Lepton Flavor Universality Violation in Tau Decays}\label{sec:lfv}
Constraints on our flavor violating $Z^\prime$ scenario can be derived from leptonic tau decays.
In the SM, the leptonic decays of the tau, $\tau^- \to \mu^- \nu_\tau \bar\nu_\mu$ and $\tau^- \to e^- \nu_\tau \bar\nu_e$, are mediated by the tree-level exchange of a $W$ boson.
Integrating out the $W$, we arrive at the following effective Hamiltonian describing the decays:
\begin{eqnarray}
\mathcal H_\text{SM} & \ = \ & \frac{g_2^2}{2 m_W^2} (\bar \nu_\tau \gamma_\alpha P_L \tau)\sum_{\ell=e,\mu}(\bar \ell \gamma^\alpha P_L \nu_\ell)
\,,
\end{eqnarray}
where
$g_2 = e / \sin\theta_W \simeq 0.65$ is the $SU(2)_L$ gauge coupling.
Due to lepton flavor universality of the weak interactions, the ratio of the branching ratios of the leptonic tau decays is close to unity. In the SM, the ratio can be predicted with extremely high accuracy~\cite{Pich:2013lsa}:
\begin{equation} \label{eq:tauSM}
R_{\mu e}^\text{SM} = \frac{\text{BR}(\tau \to \mu \nu_\tau \bar\nu_\mu)_\text{SM}}{\text{BR}(\tau \to e \nu_\tau \bar\nu_e)_\text{SM}} = 0.972559 \pm 0.000005 \,,
\end{equation}
where the deviation from unity is almost entirely due to phase space effects.
On the experimental side, the most precise measurement of this ratio comes from BaBar~\cite{Aubert:2009qj}. The PDG average~\cite{Agashe:2014kda} also includes less precise determinations from CLEO~\cite{Anastassov:1996tc} and ARGUS~\cite{Albrecht:1991rh}:
\begin{equation} \label{eq:tauexp}
R_{\mu e}^\text{PDG} \ = \ 0.979 \pm 0.004 \,.
\end{equation}
We observe a slight tension with the SM prediction at the level of $1.6\, \sigma$.
Combining Eqs.~(\ref{eq:tauSM}) and~(\ref{eq:tauexp}) we find
\begin{equation}\label{eq:taubound}
\frac{R_{\mu e}}{R_{\mu e}^\text{SM}} - 1 \ = \ 0.0066 \pm 0.0041 \,.
\end{equation}
The tree level exchange of the considered flavor violating $Z^\prime$ cannot affect the $\tau \to e \nu_\tau \bar\nu_e$ decay. However, it does give additional contributions to the $\tau \to \mu \nu_\tau \bar\nu_\mu$ decay and induces the new tau decay mode $\tau \to \mu \nu_\mu \bar\nu_\tau$, as shown in Fig.~\ref{fig:4}.
The decay $\tau \to \mu \nu_\mu \bar\nu_\tau$ is absent in the SM, but has exactly the same experimental signature as $\tau \to \mu \nu_\tau \bar\nu_\mu$. In the following we will therefore consider the sum of the two decay modes that we denote with $\tau \to \mu \nu \bar\nu$. As long as $m_{Z^\prime} \gg m_\tau$ the treatment of the $Z^\prime$ effect in terms of an effective Hamiltonian is valid and we find
\begin{eqnarray}
\mathcal H_{Z^\prime} & \ = \ & -\frac{|g_L^\prime|^2}{m_{Z^\prime}^2} (\bar \mu \gamma^\alpha P_L \tau) (\bar \nu_\tau \gamma^\alpha P_L \nu_\mu) \nonumber \\
&& -\, \frac{g_R^\prime g_L^{\prime *}}{m_{Z^\prime}^2} (\bar \mu \gamma^\alpha P_R \tau) (\bar \nu_\tau \gamma_\alpha P_L \nu_\mu) \nonumber \\
&& -\, \frac{(g_L^{\prime})^2}{m_{Z^\prime}^2} (\bar \mu \gamma^\alpha P_L \tau) (\bar \nu_\mu \gamma_\alpha P_L \nu_\tau) \nonumber \\
&& -\, \frac{g_R^\prime g_L^\prime}{m_{Z^\prime}^2} (\bar \mu \gamma^\alpha P_R \tau) (\bar \nu_\mu \gamma_\alpha P_L \nu_\tau) \,.
\label{HamZp}
\end{eqnarray}
The Hamiltonian in Eq.~\eqref{HamZp} leads to the following correction to the lepton flavor universality ratio $R_{\mu e}$:
\begin{equation}
\frac{R_{\mu e}}{R_{\mu e}^\text{SM}} \ = \ 1 + \frac{|g_L^\prime|^2}{g_2^2} \frac{4 m_W^2}{m_{Z^\prime}^2} + \left(\frac{|g_L^\prime g_R^\prime|^2}{g_2^4} + \frac{|g_L^\prime|^4}{g_2^4} \right) \frac{8 m_W^4}{m_{Z^\prime}^4} \,.
\label{eq:heavyZp}
\end{equation}
Note that our model can only increase the ratio $R_{\mu e}$ compared to the SM prediction.
Thus, the result in Eq.~(\ref{eq:taubound}) gives strong constraints on the $Z^\prime$ parameter space.
If we neglect the term that contains the right-handed $Z^\prime$ coupling, we find the following approximate constraint at the $2\,\sigma$ level
\begin{equation}
\frac{m_{Z^\prime}}{|g_L^\prime|} \ \gtrsim \ 2~\text{TeV} ~.
\label{eq:mZpbound}
\end{equation}
Note that in the absence of $g_L^\prime$, the $Z^\prime$ does not couple to neutrinos and the constraint from lepton flavor universality in tau decays can be avoided.
\begin{figure}[t]
\centering
\includegraphics[width=5cm]{feyn2n.pdf}
\caption{Feynman diagram for the $Z'$ contribution to the lepton flavor universality violating tau decay.}
\label{fig:4}
\end{figure}
The constraint~(\ref{eq:taubound}) is shown in Fig.~\ref{fig:2} in the $g_L^\prime$--$g^\prime_R$ plane in red, for a fixed $Z^\prime$ mass of $m_{Z^\prime} = 100$~GeV. Large values of $g_L^\prime$ are strongly constrained, leaving a compact region of $g_L^\prime$-$g_R^\prime$ parameter space, where the $(g-2)_\mu$ anomaly can be explained.
The dashed red lines show values of constant lepton flavor universality violation, i.e. $R_{\mu e}/R_{\mu e}^\text{SM} -1 = 1\%, 0.3\%, 0.1\%, 0.03\%$.
Probing lepton flavor universality violation in tau decays down to a level of $0.1\%$ would allow us to conclusively test the entire remaining parameter space relevant for our explanation of the $(g-2)_\mu$ anomaly. This should be possible to achieve at Belle~2~\cite{Aushev:2010bq} with 50 ab$^{-1}$ luminosity, assuming that systematic uncertainties can be kept under control.
For $Z^\prime$ masses of the order of the tau mass, the momentum transfer along the $Z^\prime$ propagator in Fig.~\ref{fig:4} has to be taken into account. In this case we find
\begin{eqnarray}
\frac{R_{\mu e}}{R_{\mu e}^\text{SM}} & \ = \ & 1 + \frac{|g_L^\prime|^2}{g_2^2} \frac{4 m_W^2}{m_{Z^\prime}^2}f\left(\frac{m_\tau^2}{m_{Z^\prime}^2}\right)+ \nonumber \\
&& + \left(\frac{|g_L^\prime g_R^\prime|^2}{g_2^4} + \frac{|g_L^\prime|^4}{g_2^4} \right) \frac{2 m_W^4}{m_{Z^\prime}^4} g\left(\frac{m_\tau^2}{m_{Z^\prime}^2}\right) ,
\label{eq:genZp}
\end{eqnarray}
with the functions
\begin{eqnarray}
f(z) &=& \frac{2}{z^4} \left[ \frac{5}{6} z^3 + 2z^2 -2z - (1-z)^2(2+z)\log(1-z) \right] \,, \nonumber \\
g(z) &=& \frac{2}{z^4} \left[ - z^3 -3z^2 +6z + 6(1-z)\log(1-z) \right] \,. \nonumber
\end{eqnarray}
In the limit $m_{Z^\prime} \gg m_\tau$, we have $\lim\limits_{z\to 0}f(z) =1$, $\lim\limits_{z\to 0}g(z)= 1$ and then Eq.~\eqref{eq:genZp} reduces to Eq.~(\ref{eq:heavyZp}).
Note that in the above expression we still neglected the muon mass.
Once the $Z^\prime$ mass comes close to the tau mass, such that $m_\tau - m_{Z^\prime} \sim m_\mu$ also the muon can no longer be treated massless.
In our numerical analysis we take into account the muon mass.
In Fig.~\ref{fig:3} we show in red the regions in the $m_{Z^\prime}$--$g^\prime_R$ plane that are excluded by the current experimental measurement of $R_{\mu e}$, for two choices of $g_L^\prime$. As expected, the constraint is strongest for large values of $g_L^\prime$. For a heavy $Z^\prime$ the constraint agrees well with the approximate bound in Eq.~(\ref{eq:mZpbound}).
We observe that in the case $g_L^\prime = g_R^\prime$, the tau decays exclude entirely our explanation of the $(g-2)_\mu$ anomaly for any $Z^\prime$ mass larger than $m_\tau$.
\begin{figure}[t]
\centering
\includegraphics[width=0.46\textwidth]{Emu_spectrum.png}
\caption{Differential muon energy spectrum in the decay $\tau \to \mu \nu \bar\nu$ in the tau rest frame in the presence of a light $Z'$ with the indicated masses.}
\label{fig:5}
\end{figure}
A light $Z^\prime$ of the order of the tau mass not only affects the overall rate of the $\tau \to \mu \nu_\tau \bar\nu_\mu$ decay, but also modifies the muon energy spectrum. In Fig.~\ref{fig:5} we show the muon energy spectrum in the tau rest frame for various choices of the $Z^\prime$ mass. We set $g_L^\prime = g_R^\prime / 10$ and choose $g_R^\prime$ such that the $Z^\prime$ leads to a 10\% increase of the $\tau \to \mu \nu_\tau \bar\nu_\mu$ decay rate, i.e. $R_{\mu e} = 1.1 \times R_{\mu e}^\text{SM}$.
For a $Z'$ mass close to the tau mass, the $Z^\prime$ exchange leads to muons that tend to be slightly softer compared to that from the SM.
We caution the reader that a possible impact of the modified muon spectrum is not taken into account when deriving the bound in Fig.~\ref{fig:3}. A careful analysis of the experimental acceptances and efficiencies would be required to ascertain the robustness of the bound shown in the parameter region $ m_\tau - m_\mu < m_{Z^\prime} \lesssim$ few~GeV (shown by the dotted red curve), which is beyond the scope of this work.
A more detailed study of the $Z'$ effect might include a Michel parameter analysis for tau decays~\cite{Pich:2013lsa, Tobe:2016qhz}. Moreover, a study of the tau polarization via its decays may be useful to probe differences from the SM, induced by the $Z'$ effect.
\section{Two-body Tau Decays} \label{sec:2body}
If the $Z^\prime$ mass is smaller than the difference of tau and muon mass, $m_{Z^\prime} < m_\tau - m_\mu$ the two body decay $\tau \to \mu Z^\prime$ opens up kinematically. This is illustrated for the two cases $m_{Z^\prime} = 1.5$~GeV and $m_{Z^\prime} = 1$~GeV by the peaks in the muon energy spectrum shown in Fig.~\ref{fig:5}. In this region of parameter space the only available decay mode of the $Z^\prime$ is into neutrinos. Direct searches for the decay $\tau \to \mu +$missing energy can then be used to constrain the $Z^\prime$ parameter space.
The $\tau \to \mu Z^\prime$ decay width reads
\begin{align}
\Gamma_{\mu Z'} & = \frac{m_\tau^3}{32\pi m_{Z^\prime}^2} \Bigg[ \left( |g_L^\prime|^2 + |g_R^\prime|^2 \right) \bigg\{ \left( 1 + \frac{2m_{Z^\prime}^2}{m_\tau^2} \right)\left( 1 - \frac{m_{Z^\prime}^2}{m_\tau^2} \right) \nonumber \\
& - \frac{m_\mu^2}{m_\tau^2} \left( 2 - \frac{m_{Z^\prime}^2}{m_\tau^2} -\frac{m_\mu^2}{m_\tau^2}\right) \bigg\} - 12\, \text{Re}(g_L^\prime g_R^{\prime *}) \frac{m_\mu}{m_\tau} \frac{m_{Z^\prime}^2}{m_\tau^2} \Bigg] \nonumber \\
& \times \sqrt{\left( 1 - \frac{m_{Z^\prime}^2}{m_\tau^2} \right)^2 - \frac{m_\mu^2}{m_\tau^2} \left( 2 + \frac{2m_{Z^\prime}^2}{m_\tau^2} -\frac{m_\mu^2}{m_\tau^2}\right) }\,.
\label{eq:muzp}
\end{align}
Neglecting terms that are suppressed by the muon mass, this can be simplified to
\begin{eqnarray}
\Gamma_{\mu Z^\prime} & \ \simeq \ & \frac{m_\tau^3}{32\pi m_{Z^\prime}^2} \left( |g_L^\prime|^2 + |g_R^\prime|^2 \right) \nonumber \\
&& \times \left( 1 + \frac{2m_{Z^\prime}^2}{m_\tau^2} \right)\left( 1 - \frac{m_{Z^\prime}^2}{m_\tau^2} \right)^2 \,.
\end{eqnarray}
In our numerical analysis we keep muon mass effects and use Eq.~\eqref{eq:muzp}.
Searches for the two body decay $\tau \to \mu \phi$ by ARGUS~\cite{Albrecht:1995ht}, where $\phi$ is an unobservable particle, directly apply to our case; they give bounds on the corresponding branching ratio for masses up to $1.6$~GeV.
The region of $Z^\prime$ parameter space that is excluded by this search is shown in Fig.~\ref{fig:3} in blue.
The bound from $\tau \to \mu Z^\prime$ is remarkably strong, and as a result, our explanation of the $(g-2)_\mu$ is entirely excluded for $m_{Z^\prime} \lesssim m_\tau - m_\mu$ by orders of magnitude, independent of the relative size of $g_L^\prime$ and $g_R^\prime$.
\section{LHC Constraints} \label{sec:lhc}
The direct LHC constraints on $Z'$ from simple resonance searches like $pp\to Z'\to \ell^+\ell^-$ and $pp\to Z'\to jj$ are not applicable in our case, since the $Z'$ does not couple to quarks at the tree level. Moreover, the flavor-violating $Z'$ searches at the LHC have only focused on the $e\mu$ channel so far~\cite{atlas:emu, CMS:2016dfe}.
Nevertheless, we can derive LHC constraints on the $\mu\tau$ coupling from the leptonic decays of the $W$ boson, since $pp\to W\to \mu \nu$ will also get a contribution from $pp\to W\to \tau \nu$, followed by the $Z'$-mediated decay of tau, as shown in Fig.~\ref{fig:4}. This will lead to an isolated muon and three neutrinos in the final state, where the neutrinos will be registered in the LHC detectors simply as missing energy, without any information on their number or flavor content. So we can use the constraints derived from this channel in our case, as long as the missing energy criterion $E_T^{\rm miss}>25$ GeV used in the corresponding $\mu\nu$ search at $\sqrt s=13$ TeV LHC~\cite{Aad:2016naf} is satisfied.
To check this, we implemented our model Lagrangian~\eqref{lagZp} into MadGraph5~\cite{Alwall:2014hca} for event generation with CT14NNLO PDFs~\cite{Dulat:2015mca}, used PYTHIA 6.4~\cite{Sjostrand:2006za} for showering and hadronization, and DELPHES 3~\cite{deFavereau:2013fsa} for a fast detector simulation. We find that most of our $W \to \mu 3\nu$ signal events pass the event selection cuts of Ref.~\cite{Aad:2016naf} for a wide range of $Z'$ masses of interest. Here we have used the narrow-width approximation and have written down the $Z'$-induced cross section as
\begin{eqnarray}
&& \sigma(pp\to W\to \tau\nu_\tau \to \mu\nu_\tau\nu_\mu\nu_\tau) \ = \ \sigma_{\rm SM}(pp\to W\to \tau\nu_\tau)\nonumber \\
&& \qquad \qquad \qquad \qquad \qquad \qquad \quad \times\: {\rm BR}(\tau \to \mu\nu_\tau\nu_\mu\nu_\tau)\, .
\end{eqnarray}
For $m_{Z'}> m_\tau-m_\mu$, we use the following expression for the width of the 3-body decay $\tau^\pm \to \mu^\pm Z'^*\to \mu^\pm \nu\bar{\nu}$:
\begin{align}
\Gamma_{\mu\nu\bar{\nu}} \ = \ |g_L'|^2(|g_L'|^2+|g_R'|^2)\frac{m_\tau^5}{768 \pi^3 m_{Z'}^4} a\left(\frac{m_\mu^2}{m_\tau^2}\right) \, ,
\label{3body}
\end{align}
where $a(x)=1-8x+8x^3-x^4-12x^2\log x$ and we included both channels $\tau \to \mu \nu_\mu \bar{\nu}_\tau$ and $\tau \to \mu \nu_\tau \bar{\nu}_\mu$. For $m_{Z'}< m_\tau-m_\mu$, we use the 2-body decay formula as in Eq.~\eqref{eq:muzp}.
Comparing the measured value of the $pp\to W\to \mu \nu$ cross section $\sigma_{\rm exp} =20.64\pm 0.70$~nb~\cite{Aad:2016naf} with the SM NNLO prediction~\cite{Catani:2009sm} $\sigma_{\rm SM} =20.08\pm 0.66$~nb obtained using CT14NNLO PDFs~\cite{Dulat:2015mca}, we derive 95\% CL upper limits on the $Z'$ couplings, as shown in Figs.~\ref{fig:2} and \ref{fig:3} by the black dashed curves. We find that the LHC constraints are weaker than the low-energy constraints directly derived from $\tau$ decay. Future run-II LHC data, as well as the high-luminosity phase~\cite{Rossi:2011zc} and/or a future 100 TeV collider~\cite{Golling:2016gvc} will perhaps be able to probe a large portion of the allowed parameter space in Figure~\ref{fig:3}, if the systematics and the SM theory uncertainty could be improved.
\section{LEP Constraints} \label{sec:lep}
While our $Z'$ does not couple to electrons and quarks at tree level, it can contribute, however, to the processes $e^+e^-\to Z\to \mu^+\mu^-, \tau^+\tau^-$, and $\nu \bar\nu$ via one-loop diagrams involving $Z'$, as shown in Fig.~\ref{fig:lep}.
Measurements of the SM $Z$ couplings~\cite{ALEPH:2005ab} to muons, taus and neutrinos can therefore be used to set constraints on the $Z'$ parameter space.
We find the following modifications of the $Z$ couplings due to the $Z'$ loop:
\begin{eqnarray}
\frac{g_{L\tau}}{g_{Le}} \ \simeq \ \frac{g_{L\mu}}{g_{Le}} & \ \simeq \ & 1 + \frac{|g_L^\prime|^2}{16\pi^2} ~\mathcal{K}(m_Z^2/m_{Z^\prime}^2) \, , \\
\frac{g_{R\tau}}{g_{Re}} \ \simeq \ \frac{g_{R\mu}}{g_{Re}} & \ \simeq \ & 1 + \frac{|g_R^\prime|^2}{16\pi^2} ~\mathcal{K}(m_Z^2/m_{Z^\prime}^2) \, , \\
\frac{g_{L\nu}}{g_{Re}- g_{Le}} & \ \simeq \ & 1 + \frac{2}{3}\frac{|g_L^\prime|^2}{16\pi^2} ~\mathcal{K}(m_Z^2/m_{Z^\prime}^2) \, ,
\end{eqnarray}
with the loop function $\mathcal{K}$~\cite{Haisch:2011up}
\begin{eqnarray}
\mathcal{K}(x) &=& - \frac{4+7x}{2x} + \frac{2+3x}{x}\log x \nonumber \\
&-& \frac{2(1+x)^2}{x^2}\big[\log x \log(1+x) + \text{Li}_2(-x) \big] \, ,
\label{loop}
\end{eqnarray}
where $\text{Li}_2(x) = - \int_{0}^x dt \log(1-t)/t$ is the di-logarithm.
In the above expressions, we use the electron couplings $g_{Le}$ and $g_{Re}$ as convenient normalization, as they are not affected by $Z'$ loops.
\begin{figure}[t]
\centering
\includegraphics[width=5cm]{feyn4n.pdf}
\caption{Example Feynman diagram for the one-loop correction to the $Z$-decay vertex due to LFV $Z'$ interactions.}
\label{fig:lep}
\end{figure}
The combined experimental results for the $Z$ couplings from LEP and SLD read~\cite{ALEPH:2005ab}
\begin{eqnarray}
g_{L\nu} & \ = \ & +0.5003 \pm 0.0012 \,, \\
g_{Le} & \ = \ & -0.26963 \pm 0.00030 \,, \\
g_{L\mu} & \ = \ & -0.2689 \pm 0.0011 \,, \\
g_{L\tau} & \ = \ & -0.26930 \pm 0.00058 \,, \\
g_{Re} & \ = \ & +0.23148 \pm 0.00029 \,, \\
g_{R\mu} & \ = \ & +0.2323 \pm 0.0013 \,, \\
g_{R\tau} & \ = \ & +0.23274 \pm 0.00062 \,,
\end{eqnarray}
with the error correlation matrix
\begin{equation}
\rho =
{\small \begin{pmatrix}
1 & -0.52 & 0.12 & 0.22 & 0.37 & -0.06 & -0.17 \\
-0.52 & 1 & -0.11 & -0.07 & 0.29 & -0.06 & 0.04 \\
0.12 & -0.11 & 1 & 0.07 & -0.07 & 0.90 & -0.04 \\
0.22 & -0.07 & 0.07 & 1 & 0.01 & -0.03 & 0.44 \\
0.37 & 0.29 & -0.07 & 0.01 & 1 & -0.09 & -0.03 \\
-0.06 & -0.06 & 0.90 & -0.03 & -0.09 & 1 & 0.04 \\
-0.17 & 0.04 & -0.04 & 0.44 & -0.03 & 0.04 & 1
\end{pmatrix}.}
\end{equation}
To derive bounds on the $Z'$ couplings and mass we perform a simple $\chi^2$ fit, setting the electron couplings $g_{Le}$ and $g_{Re}$ to the measured values. The resulting constraint is shown in the plots of Figs.~\ref{fig:2} and~\ref{fig:3} as dashed purple curves. Above the curves $\Delta \chi^2 > 4$, corresponding to a $95\%$ CL exclusion. The constraint vanishes around $m_{Z'}=25$~GeV, where the loop function~\eqref{loop} has a zero crossing.
We observe that the LEP constraint is generically weaker than the constraint obtained from the tree-level leptonic tau decays. Similarly, the constraints obtained from the modifications to the $W$ and $Z$ total widths due to the $Z'$ effects~\cite{Laha:2013xua} are weaker than those in the whole parameter space of interest, and therefore, are not shown in Fig.~\ref{fig:3}.
\section{Future Collider Signatures}\label{sec:4lepton}
\begin{figure}[t]
\centering
\includegraphics[width=5cm]{feyn5n.pdf}
\caption{A striking collider signature of our LFV $Z'$ scenario. This is applicable to both lepton and hadron colliders (depending on whether the initial state fermion $f$ is a SM charged lepton or quark). There exists a similar diagram with an intermediate muon, which is not shown here, but included in our calculation.}
\label{fig:col}
\end{figure}
The leptophilic $Z'$ scenarios have characteristic multi-lepton signatures at both lepton and hadron colliders~\cite{delAguila:2014soa, Elahi:2015vzh}. A particularly interesting signal in our LFV $Z'$ scenario is the 4-lepton final state with two same-sign muons and taus at the LHC, i.e.
\begin{align}
pp \ \to \ \mu^\pm \tau^\mp Z'^{(*)} \ \to \ \mu^\pm \mu^\pm \tau^\mp \tau^\mp \, ,
\label{lhc}
\end{align}
as shown in Fig.~\ref{fig:col}. This signal is very clean and effectively background-free. Through this process, one might also be able to determine the $Z^\prime$ mass for $m_{Z'}>m_\tau+m_\mu$, when the $Z'$ in Fig.~\ref{fig:col} goes on-shell and one of the $\mu\tau$ pairs will have an invariant mass at $m_{Z'}$. Although the tau reconstruction poses some practical challenges, Eq.~\eqref{lhc} could provide a `smoking gun' signal for our $Z'$ scenario at the LHC.
We simulate the process \eqref{lhc} to estimate the sensitivity reach at the $\sqrt s=14$ TeV LHC. The parton level events are generated using MadGraph5~\cite{Alwall:2014hca}, which are then fed to PYTHIA 6.4~\cite{Sjostrand:2006za} for showering and hadronization, and DELPHES 3~\cite{deFavereau:2013fsa} for a fast detector simulation. We impose the basic trigger cuts following a previous analysis for light $Z'$ searches in the $pp\to Z\to 4\mu$ channel~\cite{Elahi:2015vzh}:
\begin{itemize}
\item[(i)] the leading lepton must satisfy the transverse momentum cut $p_T>20$ GeV, while the sub-leading leptons are required to satisfy a milder cut $p_T>15$ GeV;
\item[(ii)] all the four leptons must satisfy the pseudo-rapidity $|\eta|<2.7$ and the isolation cut $\Delta R>0.1$.
\end{itemize}
These values are set to be as inclusive as possible for an optimistic analysis.
Since we are interested in the final states with same-sign muon pairs, we select the hadronic decay mode of the taus. In the SM, each tau decays hadronically with a probability of $\sim$
65\%, producing a tau-jet mostly containing neutral and charged pions. In our case with a pair of taus in the final state, 42\% of the events will contain two tau-jets. The hadronic tau decays have low charged track multiplicity (one or three prongs) and a relevant fraction of the electromagnetic energy deposition due to photons coming from the decay of neutral pions. Moreover, when the momentum of the tau is large compared to its mass, the tau-jets will be highly collimated and produce localized energy deposit in the electromagnetic and hadronic calorimeters. These characteristics can be exploited to enhance the identification of hadronic tau decays~\cite{Bagliesi:2007qx}. We have assumed an optimistic value of 70\% for the tau-tagging efficiency in our analysis. Since the SM background is negligible for the same-sign di-lepton pairs $\mu^\pm \mu^\pm \tau^\mp \tau^\mp$, we can simply estimate the signal sensitivity as ${\cal N}=S/\sqrt{S+B}\simeq \sqrt{{\cal L}\sigma_{\rm signal}}$, where ${\cal L}$ is the
integrated luminosity and $\sigma_{\rm signal}$ is the signal cross section times efficiency, as obtained from our detector simulation.
Our results for the $3\, \sigma$ sensitivity reach (corresponding to ${\cal N}>3$) in the high-luminosity phase of the LHC with 3 ab$^{-1}$ integrated luminosity are shown in Figs.~\ref{fig:2} and \ref{fig:3} by the orange curves. The LHC sensitivity gets weaker for a very light $Z^\prime$ with mass $m_{Z'}<m_\tau+m_\mu$, since both the intermediate tau as well as $Z'$ in Fig.~\ref{fig:col} are off-shell in this case.
The small bump around $Z$-mass is because the $Z$ also becomes off-shell for $m_{Z'}>m_Z-m_\mu$.
Overall, we find that for $m_{Z'}>m_\tau+m_\mu$, the HL-LHC has good sensitivity to large parts of the $(g-2)_\mu$-favored region. The LHC sensitivity again becomes weaker for a very heavy $m_{Z'}>2$ TeV or so, simply due to the kinematic suppression. A future $\sqrt s=100$ TeV $pp$ collider could extend our LFV $Z'$ sensitivity reach to the multi-TeV range.
The collider sensitivity can be further improved for $m_{Z'}<m_Z$ by considering a lepton collider operating at the $Z$-pole, i.e. with $\sqrt s=m_Z$. As an example, we consider a next generation $e^+e^-$ $Z$ factory such as the FCC-ee, and simulate the process (cf. Fig.~\ref{fig:col})
\begin{align}
e^+e^- \ \to \ Z \ \to \ \mu^\pm \tau^\mp Z'^{(*)} \ \to \ \mu^\pm \mu^\pm \tau^\mp \tau^\mp
\label{ilc}
\end{align}
for $m_{Z'}<m_Z$ using the procedure outlined above. Our results are shown in Fig.~\ref{fig:3} by the blue curves for the maximum achievable integrated luminosity of 2.6 ab$^{-1}$ at FCC-ee~\cite{d'Enterria:2016yqx}. We find that the sensitivity can be improved by a factor of 2-3, thus covering almost the entire $(g-2)_\mu$-favored region for $m_{Z'}<m_Z$.
Our $Z'$ scenario can in principle also affect the SM Higgs decays. First of all, the $h\to \mu^+\mu^-$ decay will receive a one-loop correction due to the LFV $Z'$ interactions. Although it is enhanced by the tau Yukawa coupling,
due to the loop suppression factor, and given that BR$(h\to \mu\mu)=2.19\times 10^{-4}$ in the SM~\cite{Heinemeyer:2013tqa}, the deviation due to the $Z'$-loop correction is extremely difficult to be observed at the HL-LHC or even at a dedicated Higgs factory.
The $Z'$ interactions could also induce a LFV decay of $h\to \tau^+\tau^-\to \mu^\pm \tau^\mp Z'$, where the $Z'$ goes undetected for a sufficiently small $m_{Z'}$. However, for the allowed range of masses and couplings in Fig.~\ref{fig:3}, this effect is again small and easily compatible with the LHC searches for $h \to \mu\tau$ that imply BR$(h\to \mu\tau) \lesssim 1.5\%$~\cite{CMS:2016qvi,Aad:2016blu}.
\section{Signal at Neutrino Telescopes} \label{sec:icecube}
\begin{figure}[t]
\centering
\includegraphics[width=4cm]{feyn3an.pdf}
\includegraphics[width=4cm]{feyn3bn.pdf}
\caption{$Z'$ contributions to the charged-current neutrino-nucleon interactions. Similar diagrams for incident neutrinos of muon flavor, as well as for antineutrinos and also for neutral-current interactions, are not shown here.}
\label{fig:6}
\end{figure}
In this section, we briefly discuss a complementary way to test our LFV $Z'$ hypothesis using ultra-high energy (UHE) neutrinos in large volume neutrino telescopes like IceCube and KM3NeT. First, we note that the $Z'$ interactions in our model induce new channels for the neutrino-nucleon interactions, as illustrated in Fig.~\ref{fig:6} for the charged-current (CC) process. For $m_{Z'}> m_\tau+m_\mu$, they could potentially give rise to a novel signature with simultaneous muon and tau events coming from the $Z'\to \mu\tau$ decay. However, it turns out that in presence of both left and right-handed $Z'$ couplings to charged leptons, as required for the $(g-2)_\mu$ explanation, there is a destructive interference between the two diagrams in Fig.~\ref{fig:6}, which leads to a cross section too small to be ever observed. Moreover, the stringent limits on the $Z'$ couplings from tau decays (cf. Figs.~\ref{fig:2} and \ref{fig:3}) necessarily imply that even if we disregard the $(g-2)_\mu$ favored
region by taking $g_R'\gg g_L'$, the total cross section for the processes in Fig.~\ref{fig:6} is still small, as compared to that of the SM CC interaction. To give an example, for a benchmark point with $m_{Z'}=2$ GeV, $g_R'=0.02$ and $g_L'=0.0004$ which satisfies the $\tau\to \mu\nu\bar{\nu}$ constraint, we find the total cross section for the processes shown in Fig.~\ref{fig:6} (including the antineutrino initial states) for an incoming neutrino energy $E_\nu= 1$ PeV to be $1.54\times 10^{-38}~{\rm cm}^2$, as compared to the corresponding SM CC cross section of $7.3\times 10^{-34}~{\rm cm}^2$, both calculated using the CT14NNLO PDFs~\cite{Dulat:2015mca}. It is difficult to measure such a small cross section at IceCube even with large statistics, since it will be overshadowed by various uncertainties in the incoming neutrino flux, flavor composition, and parton distribution functions (see e.g.~\cite{Chen:2013dza, Chen:2014gxa, Vincent:2016nut}).
A better possibility to detect a light $Z'$ at IceCube might be through its effect on neutrino-neutrino scattering due to on-shell $Z'$ production. In fact, the resonant absorption of UHE neutrinos by the cosmic neutrino background (C$\nu$B)~\cite{Weiler:1982qy, Weiler:1983xx, Weiler:1997sh, Roulet:1992pz, Yoshida:1996ie, Fargion:1997ft} in the presence of a light mediator has been invoked~\cite{Ibe:2014pja, Araki:2014ona, Araki:2015mya, Kamada:2015era, DiFranzo:2015qea} to explain the apparent energy gap in the IceCube neutrino data~\cite{Aartsen:2013jdh, Aartsen:2014gkd, Aartsen:2015zva} just below the PeV deposited energy bin. However, this scenario works only for an MeV-scale $Z'$, which is unfortunately ruled out in our model due to the $\tau \to \mu Z'$ constraint. For a higher $Z'$ mass, the incoming neutrino energy required to observe a resonance feature at the IceCube will be shifted upwards:
\begin{align}
E_{\nu}^{\rm res} \ = \ \frac{m_{Z'}^2}{2m_\nu(1+z)} \, ,
\label{res}
\end{align}
where $z$ is the redshift parameter at which the scattering occurs (typically taken to be the source redshift),\footnote{The redshift factor $(1+z)$ in Eq.~\eqref{res} is due to the fact that the energy $E_{\nu_i}^s$ of the cosmic neutrino $\nu_i$ at the source position $z$ is $(1+z)$ times the energy $E_{\nu}$ measured at IceCube in an expanding Universe.} $m_\nu$ is the mass of the target C$\nu$B, which is assumed to be larger than the effective temperature of the thermal distribution of the C$\nu$B, $T_{\nu}=1.7\times 10^{-4}(1+z)$ eV.\footnote{If the lightest neutrino is nearly massless, $m_\nu$ in Eq.~\eqref{res} should be replaced with the thermally averaged momentum $\langle p_\nu \rangle = \frac{7\pi^4 T_\nu}{180\zeta(3)}\approx 3.15 T_\nu$.}
The total cross section for $\nu_i\bar{\nu}_j\to Z'\to f\bar{f'}$, where $\{i,j\}=\{\mu,\tau\}$ and $\{f,f'\} = \{\nu_\mu,\nu_\tau\}$ or $\{\mu,\tau\}$ (with $i\neq j \,, f\neq f'$), is given by
\begin{align}
\sigma(s) \ = \ \frac{1}{6\pi}|g_L'|^2(2|g_L'|^2+|g_R'|^2)\frac{s}{(s-m_{Z'}^2)^2+m_{Z'}^2\Gamma_{Z'}^2} \, ,
\label{cross}
\end{align}
where $s$ is the squared center of mass energy and $\Gamma_{Z'}$ is the total width of the $Z'$. Here we have ignored the $t$-channel contribution for the $\nu\bar{\nu}$ final state, as it is highly suppressed relative to the $s$-channel resonance. Also we have assumed $s \gg (m_\tau+m_\mu)^2$. For $m_{Z'}>m_\tau+m_\mu$, there are two decay modes of $Z'\to \nu_{\mu(\tau)}\bar{\nu}_{\tau(\mu)},\: \mu^\pm \tau^\mp$, with the corresponding decay widths given by
\begin{align}
\Gamma_{\nu_\mu\bar{\nu}_\tau} & = \frac{|g_L'|^2 m_{Z'}}{24\pi}\, ,\\
\Gamma_{\mu^-\tau^+} & = \frac{m_{Z'}}{24\pi} \beta \tilde\beta \left[ \tilde \beta^2 \left(3 -\beta^2 \right) C_V^2 + \beta^2 \left(3 -\tilde\beta^2 \right) C_A^2 \right] \, ,
\end{align}
where $\beta=\sqrt{1-\frac{(m_\tau+m_\mu)^2}{m_{Z'}^2}}$, $\tilde\beta=\sqrt{1-\frac{(m_\tau-m_\mu)^2}{m_{Z'}^2}}$ and $C_{V,A}$ are defined below Eq.~\eqref{gm2}. The total decay width of $Z'$ is then given by $\Gamma_{Z'}=2\: (\Gamma_{\nu_\mu\bar{\nu}_\tau}+\Gamma_{\mu^-\tau^+})$, taking into account two possibilities for each decay mode.
\begin{figure}[t]
\centering
\includegraphics[width=8cm]{nuscat.pdf}
\caption{Cross section for $\nu_i\bar{\nu}_j\to Z'\to f\bar{f'}$ as a function of the energy of one of the initial state neutrinos. For the second neutrino $\nu_j$, we consider two cases: C$\nu$B (red solid curve) and supernova neutrinos with MeV energy (blue solid and dashed curves). The numbers above the peaks show the $Z'$ mass. For comparison, we also show the SM neutrino-nucleon CC and $\bar{\nu}_e e$ cross sections.}
\label{fig:7}
\end{figure}
The cross section~\eqref{cross} is plotted in Fig.~\ref{fig:7} as a function of the energy of the incoming UHE neutrino $\nu_i$ for three different cases, depending on the energy of the other neutrino $\nu_j$. First, we consider the C$\nu$B for which the effective temperature $T_\nu$ is smaller than at least two of the light neutrino masses, so $s=2m_\nu E_\nu$. From Eq.~\eqref{res}, it is clear that for $m_{Z'}$ above the tau mass, the resonance will occur at very high energies well beyond the energy scale currently being probed at the IceCube.
For an illustration, we choose a benchmark point from Fig.~\ref{fig:3} (right panel) satisfying all the constraints: $m_{Z'}=1.8$ GeV, $g_R'=0.01$, and $g_L'=g_R'/10$ and take the light neutrino mass $m_\nu=\sqrt{\Delta m_{\rm atm}}\simeq 0.05$ eV and a typical source redshift $z=0.2$. For this benchmark, we find the resonance energy to be at $2.7\times 10^{10}$ GeV,\footnote{For comparison, the SM $Z$ resonance occurs at $6.9\times 10^{13}$ GeV for $z=0.2$.} as shown by the red solid curve in Fig.~\ref{fig:7}. The other two light neutrino mass eigenstates will induce similar peaks at different energies, depending on their mass hierarchy.
For comparison, we also show the SM neutrino-nucleon CC and $\bar{\nu}_e e$ cross sections, with the latter having the Glashow resonance~\cite{Glashow:1960zz} at 6.3 PeV. In spite of the resonance enhancement, the $\nu\bar{\nu}$ cross section turns out to be much smaller than the SM $\nu N$ cross section.
In order to check the condition under which the UHE neutrinos $\nu_i$ will likely have at least one interaction with the C$\nu$B during their entire journey from the source to Earth, we calculate their mean free path (MFP), given by
\begin{align}
\lambda(E_\nu, z) & \ = \left[\int\frac{d^3p}{(2\pi)^3}\frac{1}{\exp{\{{\mathbf{p}/T_\nu(1+z)}\}}+1}\sigma(E_{\nu_i}^s,\mathbf{p}) \right]^{-1} \nonumber \\
& \ = \ \frac{1}{n_\nu \sigma(E_\nu)} \, ,
\end{align}
where $n_\nu = \frac{3}{4\pi^2}\zeta(3)T_\nu^3\simeq 56(1+z)^3~{\rm cm}^{-3}$ is the number density of the C$\nu$B (for each flavor) and $\sigma$ is given by Eq.~\eqref{cross}. The MFP will be the minimum at the resonance energy which corresponds to the maximum cross section. The survival rate of the high-energy neutrino $\nu_i$ travelling from the source at $z$ to Earth (at $z=0$) is then given by
\begin{align}
P(E_\nu,z) \ = \ \exp{\left[-\int_0^z dz'\frac{1}{\lambda(E_\nu,z')}\frac{dL}{dz'}\right]} \, ,
\end{align}
where $dL/dz=c/(H_0\sqrt{\Omega_m(1+z)^3+\Omega_\Lambda})$, $c$ is the speed of light in vacuum, and the present best-fit values of the cosmological parameters in a $\Lambda$CDM Universe are $H_0=100h~{\rm km \, s}^{-1}{\rm Mpc}^{-1}$ with $h=0.678$, the matter energy density $\Omega_m=0.308$ and the dark energy density $\Omega_\Lambda=0.692$~\cite{Ade:2015xua}. Thus, if the traveling distance of the UHE neutrinos is larger than the MFP, they will be attenuated by the $C\nu$B and their survival rate will have a `dip' at the resonance energy. This will lead to a characteristic absorption feature in the UHE neutrino energy spectrum. For the benchmark discussed above, we find $\lambda(E_\nu^{\rm res})\simeq 6$ kpc, which means that all extragalactic sources like gamma-ray bursts and active galactic nuclei (with typical distances of Mpc or larger) or even far-away galactic sources like supernova remnants could in principle show an absorption feature in their neutrino spectrum due to the presence of a light $Z'$.
The resonance energy could be lowered significantly if we consider interactions of the high-energy neutrinos with other relativistic neutrinos naturally available, e.g. supernovae neutrinos (after they have oscillated into muon and tau flavors) which have a typical energy $E'$ in the MeV range~\cite{Formaggio:2013kya}. In this case, the center-of-mass energy of the system is $s=4E_\nu E'$, and the resonance condition~\eqref{res} gets modified to $E_\nu^{\rm res}=m_{Z'}^2/4E'(1+z)$, independent of the light neutrino mass, thus lowering the resonance energy down to the TeV scale. This is illustrated in Fig.~\ref{fig:7} for two choices of $m_{Z'}=1.8$ GeV (blue solid curve) and 10 GeV (blue dashed curve). Below the TeV scale, it will be difficult to observe the resonance feature, since it will be swamped by the atmospheric neutrino background. The neutrino number density at the supernova core surface is much larger, e.g. $\gtrsim 10^{34}~{\rm cm}^{-3}$ for SN1987A~\cite{Hirata:1987hu}. Hence, the MFP can be
much smaller, thus allowing for the possibility of observing the absorption feature from both galactic and extragalactic sources, provided the incoming high-energy neutrinos encounter a supernova core collapse en route to Earth. The likelihood of such an arrangement somewhat depends on the origin of the high-energy neutrino source, and cannot be excluded at the moment.
The LFV interactions could also alter the ratio of astrophysical neutrino flavors at detection on Earth from the standard expectation of $(\nu_e:\nu_\mu:\nu_\tau)=(1:1:1)$. The detailed predictions for the event rate and the track-to-shower ratio will depend on many parameters, including the source neutrino flux normalization and spectral index, redshift, as well as the PDF uncertainties, but in spite of all these uncertainties, the anomalous features could plausibly be measured~\cite{Barenboim:2004di, D'Olivo:2005uh, Lunardini:2013iwa, Ioka:2014kca, Ng:2014pca, Blum:2014ewa} by IceCube or next generation neutrino telescopes like IceCube-Gen2, thereby opening a new era of `cosmic neutrino spectroscopy'.
\section{Conclusion} \label{sec:concl}
We have discussed a simple new physics interpretation of the long-standing anomaly in the muon anomalous magnetic moment in terms of a purely flavor off-diagonal $Z'$ coupling only to the muon and tau sector of the SM. We have discussed the relevant constraints from lepton flavor universality violating tau decays for $m_{Z'}>m_\tau$ and from $\tau\to \mu ~+$ invisibles decay for $m_{Z'}<m_\tau$, as well as the latest LHC constraints from $W\to \mu\nu$ searches.
We find that for a $Z'$ lighter than the tau, the low-energy tau decay constraints rule out the entire $(g-2)_\mu$ allowed region by many orders of magnitude.
However, a heavier $Z'$ solution to the $(g-2)_\mu$ puzzle is still allowed, provided the $Z'$ coupling to the charged leptons has both left- and right-handed components, and the right-handed component is larger than the left-handed one. The deviations from lepton flavor universality in the tau decays predicted in this model can be probed at Belle 2, while a large part of the $(g-2)_\mu$ allowed region can be accessed at future colliders such as the high-luminosity LHC and/or an $e^+e^-$ $Z$-factory such as FCC-ee. The on-shell production of $Z'$ in high-energy neutrino interactions with either cosmic neutrino background or with other natural neutrino sources such as supernova neutrinos could lead to characteristic absorption features in the neutrino spectrum, which might be measured in neutrino telescopes.
\section*{Acknowledgments}
W.A. acknowledges discussions with Stefania Gori and financial support by the University of Cincinnati.
The work of C.-Y.C. is
supported by NSERC, Canada. Research at the Perimeter Institute is supported in part
by the Government of Canada through NSERC and by the Province of Ontario through
MEDT. The work of B.D. is supported by the DFG grant No. RO 2516/5-1. B.D. also acknowledges partial support from the
TUM University Foundation Fellowship, the DFG cluster of excellence
``Origin and Structure of the Universe", and the Munich Institute for Astro- and Particle Physics (MIAPP) during various stages of this work. The work of A.S. is supported in part by the US DOE Contract No. DE-SC 0012704. B.D. and A.S. thank the organizers of WHEPP XIV at IIT Kanpur for the hospitality during an earlier phase of this work.
|
\section{introduction}
The idea is inspired by the quantum trajectory description of
decoherence\cite{trajectory}. The trajectory was first proposed by Bohm when
he made a suggested interpretation of the quantum theory for hidden
variables\cite{Bohm}. The theory is known as the de Broglie-Bohm(BB)
interpretation of quantum mechanics. In the theory, all particles have
well-defined trajectories. The motions of the particles are governed by the
wave functions that satisfy the Schrodinger equation. Therefore the BB quantum
theory of motion is a suitable tool with which to study coherence and
decoherence\cite{BBdecoherence,BBdecoherence1}. The one reason of the
decoherence is the open quantum system interacts with the environment.
Unfortunately, it is very difficult to deal with decoherence problem using
this quantum trajectory approach because the environment usually involves
large number of degrees of the freedom. To overcome this drawback, we assume
the environment to be the Markovian environment and describe the whole system
by a Markovian master equation. This equation introduces two contributions:
the time-evolution of the coherent state and the quenching factor leading to
decoherence. The quenching factor accounts for physical properties of the
environment and its interaction with the coherent system. Combining the
trajectory theory with reduced density matrix theory yields a new trajectory
called reduced quantum trajectory\cite{trajectory}. The advantage of this
reduced quantum trajectory is that the environment effects are described by a
time-dependent damping factor when these trajectories are applied to the study
of an open quantum system. The reduced quantum trajectory then describes in
detail the evolution of the coherent state. These provide insight in
understanding decoherence.
Recently, the discovery of the high critical temperature of $MgB_{2}$ has
inspired a widely interest in the charged two-condensate
superconductors\cite{two-gap1,two-gap2,TG-E}. The two charged condensates in
the superconductor are tightly bound fermion pairs, or some other charged
bosonic fields such as electronic or protonic Cooper pairs in metallic
hydrogen under certain condition\cite{liquidm}. The charged two-condensate
wave functions correspond to the order parameters of the two different parts
of the Fermi surface. They are coupled because of their electromagnetic
interaction. The system is described by the Ginzberg-Landau model with two
flavors of Cooper pairs\cite{TGBB,TGBB1,TGBB2}. In\cite{TGBB}, the authors
show the charged -condensate Ginzberg-Landau model can be mapped onto a
version of the nonlinear $O\left( 3\right) $ $\sigma$-model and found this
system possesses a hidden $O\left( 3\right) $ symmetry. There is a stable
knot solution in the superconductor. This provides us with a new way to
investigate the coherent quantum system.
The topology and geometry play an important role in physics and mathematics
and a great deal of works have been done in the topology and
geometry\cite{Debrus,Eguchi,Morandi,Nakahara,Nash,Schwarz}. Especially, the
vorticity of the vortex in condensate meter and topology of the physical
system have been studied by applying the $\phi-$mapping topological current
theory\cite{Duan,London e,Shi,Shi1}. In this paper, we present the relation
between the current of coherent state and the supercurrent of the two-gap
condensate system. The paper is organized as follows: in part II, the $\phi
-$mapping topological current theory in reduced density trajectory is given.
The current of the coherent state is also presented. In part III, the new
expression of the current is derived. We find this current is similar to the
supercurrent of the charged two-condensate system. In part IV, the symmetry
and the topological properties of the current of the coherent state are
studied based on Faddeev's $O\left( 3\right) $ nonlinear $\sigma$-model.
Finally, we make a conclusion.
\section{$\phi-$ mapping topological current theory in reduced density
trajectory and the current of the coherent state}
We give a brief review of the reduced quantum trajectory approach as presented
in\cite{trajectory}. We start with the calculation of the reduced density
matrix. The total density matrix of the system is given b
\begin{equation}
\widehat{\rho}=|\psi_{t}\rangle\times\langle_{t}\psi|, \label{1
\end{equation}
where the subscript $t$ denotes the time-dependence of the wave function. We
take the environment degrees of freedom to be $\mathbf{r}_{i}$ $\left(
i=1,\cdots,N\right) $. The system's reduced density matrix is then given by
tracing the total density matrix $\widehat{\rho}$ over the environment degrees
of freedom, resulting in
\begin{equation}
\widetilde{\rho_{t}}\left( r,r^{\prime}\right) =\int\langle\mathbf{r,r
_{1}\mathbf{,\cdots r}_{N}|\psi_{t}\rangle\times\langle_{t}\psi|\mathbf{r
^{\prime}\mathbf{,r}_{1}^{\prime}\mathbf{,\cdots r}_{N}^{\prime}\rangle
d\mathbf{r}_{1}\cdots d\mathbf{r}_{N}.
\end{equation}
Next the system reduced quantum density current can be derived as follows
\begin{equation}
\widetilde{\mathbf{J}_{t}}=\frac{\hbar}{m}\operatorname{Im}\left[ \nabla
_{r}\widetilde{\rho_{t}}\left( \mathbf{r,r}^{\prime}\right) \right]
_{r=r^{\prime}}, \label{J
\end{equation}
where $\widetilde{\mathbf{J}_{t}}$ satisfies the continuity equation, which is
given a
\begin{equation}
\partial_{t}\widetilde{\rho_{t}}+\nabla\widetilde{\mathbf{J}_{t}}=0. \label{P
\end{equation}
Where $\widetilde{\rho_{t}}$ is the diagonal element of the reduced density
matrix, which provides the measured intensity. We now define the Bohmian-like
velocity using (\ref{J}) and (\ref{P}
\begin{equation}
\mathbf{V}=\frac{\widetilde{\mathbf{J}_{t}}}{\widetilde{\rho_{t}}}. \label{V
\end{equation}
Therefore, we can define a new trajectory associated with the reduced density
matri
\begin{equation}
\mathbf{V}=\frac{\hbar}{m}\frac{\operatorname{Im}\left[ \nabla_{r
\widetilde{\rho_{t}}\left( r,r^{\prime}\right) \right] }{\operatorname{Re
\left[ \widetilde{\rho_{t}}\left( r,r^{\prime}\right) \right]
}_{r=r^{\prime}}. \label{V0
\end{equation}
which is called reduced quantum trajectory. The disadvantage of this
definition of velocity is it is difficult to give the detailed information at
$\widetilde{\rho_{t}}=0$, or at the zero points of the wave functions. These
zero points are the singularity of the velocity. Next, we will illustrate the
exact expression of the velocity field and its topology at zero point of wave
functions based on $\phi-$mapping topological current theory. To do this, we
must consider the BB quantum mechanics ansatz of the wave functio
\begin{equation}
\left\langle r\right\vert \psi_{t}\rangle=R_{t}\left( r\right)
e^{iS_{t}\left( r\right) /\hbar},
\end{equation}
from the topological viewpoint, the wave function $\left\langle r\right\vert
\psi_{t}\rangle$ is the section of the complex linear bundle, i.e. a section
of 2-dimensional real vector bundle. We can then write this ansatz a
\begin{equation}
\left\langle r\right\vert \psi_{t}\rangle=\phi^{1}+i\phi^{2}.
\end{equation}
Defining the unit vector of this ansatz yield
\begin{equation}
n^{1}=\frac{\phi^{1}}{\left\Vert \left\langle r\right\vert \psi_{t
\rangle\right\Vert }\text{ \ }n^{2}=\frac{\phi^{2}}{\left\Vert \left\langle
r\right\vert \psi_{t}\rangle\right\Vert }. \label{unit
\end{equation}
It is obvious that the unit vector satisfies the conditio
\begin{equation}
n^{a}n^{a}=1\text{ \ \ }a=1,2.
\end{equation}
Using this unit vector and (\ref{V0}), we write the velocity a
\begin{equation}
\mathbf{V}_{i}=\frac{\hbar}{m}\epsilon_{ab}n^{a}\partial_{i}n^{b}.
\end{equation}
In traditional quantum mechanics, the curl of the velocity vanishes at zero
points of the wave functions. However, the curl of the velocity must be
modified along trajectories because $\boldsymbol{\nabla}\times\mathbf{V}$ need
not vanish at nodal points of the wave function\cite{Shi}. The curl of the
velocity i
\begin{equation}
\nabla\times\mathbf{V}=\frac{\hbar}{m}\left( \epsilon^{ijk}\epsilon
_{ab}\partial_{j}n^{a}\partial_{k}n^{b}\right) \mathbf{e}_{i}. \label{17
\end{equation}
Using Eqs.(\ref{unit}), the curl of the velocity can further be written a
\begin{equation}
\nabla\times\mathbf{V}=\frac{\hbar}{m}\mathbf{e}_{i}\epsilon^{ijk
\epsilon_{ab}\frac{\partial}{\partial\phi^{c}}\frac{\partial}{\partial\phi
^{a}}\left( \ln\left\Vert \phi\right\Vert \right) \partial_{j}\phi
^{c}\partial_{k}\phi^{b}. \label{18
\end{equation}
Defining the vector Jacobian of $\mathbf{\phi}$ b
\begin{equation}
\mathbf{e}_{i}\epsilon^{ijk}\partial_{j}\phi^{c}\partial_{k}\phi^{b
=\epsilon^{cb}\mathbf{D}\left( \frac{\phi}{x}\right) , \label{19
\end{equation}
and using the well-known result from the Green's function theory in $\phi
-$space, we find tha
\begin{equation}
\frac{\partial}{\partial\phi^{a}}\frac{\partial}{\partial\phi^{a}
\ln\left\Vert \phi\right\Vert =2\pi\delta^{2}\left( \phi\right) . \label{20
\end{equation}
Finally, the curl of the velocity i
\begin{equation}
\nabla\times\mathbf{V}=\frac{\hbar}{m}2\pi\delta^{2}\left( \phi\right)
\mathbf{D}\left( \frac{\phi}{x}\right) .
\end{equation}
where $\mathbf{D}\left( \frac{\phi}{x}\right) $ is the vector Jacobian of
$\mathbf{\phi}$ and satisfies $\epsilon^{ijk}\partial_{j}\phi^{c}\partial
_{k}\phi^{b}=\epsilon^{cb}D^{i}\left( \frac{\phi}{x}\right) $. From this, we
learn that the trajectory is at the zero point of the wave function. We
consider, in general, a vector field $\mathbf{\phi}$ on the smooth manifold
$\Sigma$; a zero point $p$ is a singular point of $\mathbf{\phi}$ if
$\mathbf{\phi}_{p}=0$. Consider a closed curve $\gamma\in\Sigma$ encircling
but never touching $p$. In completing one turn along $\gamma$, the vector
field $\mathbf{\phi}$ will turn around itself a certain number of times. By
appropriately assigning signs to the direction of the turn, the algebraic sum
of turns is called index of the curve. It is well known the sum of all the
indices of a chosen vector field $\mathbf{\phi}$ on a compact differentiable
manifold $\Sigma$ equals the Euler-Poincare characteristic of $\Sigma$ that
describes the topological properties of singular points. In application here,
all nodal points form the zero-line of wave function and the zero-line of wave
function is just the locations of trajectories in de Broglie-Bohm quantum
mechanics. The zero points can be denoted by $z_{l}^{i},$ where $l$ represent
the $\ell$ isolated zero points on $\Sigma.$ We assume that $u=(u_{1},u_{2})$
are the coordinates, so that $\delta^{2}\left( \phi\right) $ can be expanded
at the zero poin
\begin{equation}
\delta^{2}\left( \phi\right)
{\displaystyle\sum\limits_{l=1}^{\ell}}
C_{l}\delta^{2}\left( x^{i}-z_{l}^{i}\right) , \label{25
\end{equation}
where $C_{l}$ are positive coefficients. The winding number of the $lth$
trajectory i
\begin{align}
W\left( \phi,z_{i}\right) & =C_{l}\int_{\Sigma}\delta^{2}\left(
x^{i}-z_{l}^{i}\right) D\left( \frac{\phi}{x}\right) d^{2}x\label{26}\\
& =C_{l}D\left( \frac{\phi}{u}\right) _{z_{l}}.\nonumber
\end{align}
Here, $D\left( \frac{\phi}{u}\right) $ i
\begin{equation}
D\left( \frac{\phi}{u}\right) =\frac{1}{2}\epsilon^{jk}\epsilon_{ab
\frac{\partial}{\partial u^{j}}\phi^{a}\frac{\partial}{\partial u^{k}}\phi
^{b}. \label{27
\end{equation}
If we le
\begin{equation}
\left\vert W_{l}\right\vert =\left\vert W\left( \phi,z_{l}\right)
\right\vert =\beta_{l}, \label{30
\end{equation}
where $\beta_{l}$ is Hopf index of $\phi-$mapping on $\Sigma$, with the
interpretation that the function $\mathbf{\phi}$ covers the corresponding
region in $\phi-$space $\beta_{l}$ times when a point covers the neighborhood
of the zero point $z_{l}^{i}$ once. Furthermore, $\delta^{2}\left(
\phi\right) $ can be expressed a
\begin{equation}
\delta^{2}\left( \phi\right)
{\displaystyle\sum\limits_{l=1}^{\ell}}
\frac{\beta_{l}}{\left\vert D\left( \frac{\phi}{u}\right) \right\vert
_{z_{l}}}\delta^{2}\left( x^{i}-z_{l}^{i}\right) . \label{31
\end{equation}
Let us defin
\begin{equation}
\eta_{l}=signD\left( \frac{\phi}{u}\right) _{z_{l}}=\frac{D\left(
\frac{\phi}{u}\right) }{\left\vert D\left( \frac{\phi}{u}\right)
\right\vert }_{z_{l}}=\pm1, \label{32
\end{equation}
which is called the Brouwer degree of the map $x\rightarrow\phi\left(
x\right) $. Finally, the vorticity of the velocity at the zero points on
$\Sigma$\ i
\begin{equation}
\Gamma=\int_{\Sigma}\left( \nabla\times\mathbf{V}\right) \cdot
d\mathbf{S}=\frac{h}{m
{\displaystyle\sum\limits_{l}}
\beta_{l}\eta_{l}=\frac{h}{m}W, \label{W
\end{equation}
where $W$ is the winding number of the zero points of the trajectories on
$\Sigma$. The zero points on the plane can be seen as the topological
solutions of the equation $\delta^{2}\left( \phi\right) $ and can be written
a
\begin{align}
\phi^{1}\left( x^{\mu}\right) & =0,\nonumber\\
\phi^{2}\left( x^{\mu}\right) & =0, \label{E
\end{align}
where $\mu=1,2,3..$
Considering a quantum system in the double-slit experiment, the system is
described by the coherent state of a particle and the state of the
environment. The coherent state of a particle i
\begin{equation}
|\Psi_{t}\rangle=c_{1}|\psi_{1,t}\rangle+c_{2}|\psi_{2,t}\rangle,
\end{equation}
where the coefficients $c_{a}$ satisfies the conditio
\begin{equation}
\left\vert c_{1}\right\vert ^{2}+\left\vert c_{2}\right\vert ^{2
=1.\label{inerproduct
\end{equation}
We assume the environment states are subject to the elastic system-environment
scattering conditions\cite{trajectory}, then only the environment state will
evolve with time. The environment state associated with each partial wave is
denoted by $|H_{\alpha}\rangle$. The initial state of the environment states
can be given b
\begin{equation}
|H_{1}\rangle=|H_{2}\rangle=|H_{0}\rangle.
\end{equation}
Using BB quantum mechanics anzatz, the coherent state can be described without
considering the interaction between coherence states and the environment
\begin{equation}
\Psi_{t}\left( r\right) =\langle r|\Psi_{t}\rangle.
\end{equation}
The density matrix associated with coherent state is
\begin{equation}
\rho_{t}\left( \mathbf{r,r}^{\prime}\right) =\Psi_{t}\left( \mathbf{r,r
^{\prime}\right) \left[ \Psi_{t}\left( \mathbf{r,r}^{\prime}\right)
\right] ^{\ast}.\label{density
\end{equation}
The diagonal element of this density matrix is the measured intensity. We
write it as
\begin{equation}
\rho_{t}=\left\vert c_{1}\right\vert ^{2}\left\vert \psi_{1,t}\right\vert
^{2}+\left\vert c_{2}\right\vert ^{2}\left\vert \psi_{2,t}\right\vert
^{2}+2\left\vert c_{1}\right\vert \left\vert c_{2}\right\vert \left\vert
\psi_{1,t}\right\vert _{t}\left\vert \psi_{2,t}\right\vert \cos\delta
_{t},\label{D1
\end{equation}
where $\delta_{t}$ is the time-dependent phase shift between the
partial\ waves. Similarly the partial wave function $\psi_{i,t}$ can be
written a
\begin{equation}
\psi_{i,t}=\langle r|\psi_{i,t}\rangle.
\end{equation}
In addition to writing the measured intensity for $\Psi_{t}\left( r\right)
$, we define the measured intensity of the partial wave function $\rho
_{t}^{\left( i\right) }$ as
\begin{equation}
\rho_{t}^{\left( i\right) }=\psi_{i,t}^{\ast}\psi_{i,t}\text{ \ }i=1,2,.
\end{equation}
The partial wave function can also be expressed a
\begin{equation}
\psi_{i,t}=\phi_{i,t}^{1}+i\phi_{i,t}^{2}.
\end{equation}
Recalling (\ref{unit}), the unit vector $\mathbf{n}_{\left( i\right) }$ of
the partial wave function $\psi_{i,t}$ is defined b
\begin{equation}
n_{\left( i\right) }^{1}=\frac{\phi_{i,t}^{1}}{\left\Vert \psi
_{i,t}\right\Vert },n_{\left( i\right) }^{2}=\frac{\phi_{i,t}^{2
}{\left\Vert \psi_{i,t}\right\Vert }.\label{unit1
\end{equation}
The general initial coherent states get entangled with the environment states
when the environment is considered. The initial entangled state i
\begin{equation}
|\Psi\rangle=|\Psi_{0}\rangle\otimes|H_{0}\rangle,
\end{equation}
where $|\Psi_{0}\rangle$ is the wave function $|\Psi_{t}\rangle$ at time
$t=0$. The time-dependence of the entangled state i
\begin{equation}
|\Psi_{t}\rangle=c_{1}|\psi_{1,t}\rangle\otimes|H_{1,t}\rangle+c_{2
|\psi_{2,t}\rangle\otimes|H_{2,t}\rangle, \label{density1
\end{equation}
where $|H_{i,t}\rangle$ is the time-dependent environment. Then we obtain the
measured intensity of the entangled state by tracing the full density matrix
over the environment stat
\begin{equation}
\widetilde{\rho_{t}}
{\displaystyle\sum\limits_{a=1}^{2}}
\langle H_{a,t}\left\vert \widehat{\rho}\right\vert H_{a,t}\rangle.
\label{reduced density
\end{equation}
Substitute (\ref{density1}) and (\ref{1}) into (\ref{reduced density}), one
obtains the measured intensity by tracing the total density matrix over the
environmental degrees of freedo
\begin{equation}
\widetilde{\rho}_{t}=\left( 1+\left\vert a_{t}\right\vert ^{2}\right)
{\displaystyle\sum\limits_{i=1}^{2}}
\left\vert c_{i}\right\vert ^{2}\psi_{i,t}^{\ast}\psi_{i,t}+2a_{t}c_{1
c_{2}^{\ast}\psi_{1,t}\psi_{2,t}^{\ast}+c.c.. \label{density2
\end{equation}
This equation means the interaction between the coherence state and the
environment is the reason of the decoherence. The coefficient $a_{t}=\langle
H_{2,t}|H_{1,t}\rangle$ is called the damping factor and indicates the degree
of coherence. The cross terms $c_{1}c_{2}^{\ast}\psi_{1,t}\psi_{2,t}^{\ast}$
and its conjugate complex in (\ref{density2}) disappear, when $a_{t}=0,$ the
coherent state suffers a total loss of coherence. If one introduces the
coherence time $\tau,$ then this damping factor can be written as
$a_{t}=e^{-t/\tau}$. By using (\ref{V0}) and (\ref{density2}), the current is
given b
\begin{align}
\mathbf{J} & \mathbf{=}\widetilde{\rho_{t}}\mathbf{V}=\frac{i\left(
1+\left\vert a_{t}\right\vert ^{2}\right) \hbar}{2m
{\displaystyle\sum\limits_{i=1}^{2}}
\left\vert c_{i}\right\vert ^{2}\left[ \left( \psi_{i,t}^{\ast}\nabla
\psi_{i,t}-\psi_{i,t}\nabla\psi_{i,t}^{\ast}\right) \right] \label{29}\\
& +\frac{i\hbar}{m}\left\vert a_{t}\right\vert c_{1}c_{2}^{\ast}\left[
\left( \psi_{2,t}^{\ast}\nabla\psi_{1,t}-\psi_{1,t}\nabla\psi_{2,t}^{\ast
}\right) \right] +C.C..\nonumber
\end{align}
\section{\bigskip the current as a supercurrent in two-condensate system}
From equation (\ref{29}), the current is seen to be expressed as a sum of two
contributions: the first term, which does not include the cross term of the
partial wave functions, will be denoted by $\mathbf{J}_{1}$
\begin{equation}
\mathbf{J}_{1}=\frac{i\left( 1+\left\vert a_{t}\right\vert ^{2}\right)
\hbar}{2m
{\displaystyle\sum\limits_{i=1}^{2}}
\left\vert c_{i}\right\vert ^{2}\left[ \left( \psi_{i,t}^{\ast}\nabla
\psi_{i,t}-\psi_{i,t}\nabla\psi_{i,t}^{\ast}\right) \right] ,
\end{equation}
and the second term, which includes the cross term which indicates the
coherent effects, will be written by $\mathbf{J}_{2}
\begin{align}
\mathbf{J}_{2} & =\frac{i\hbar}{m}\left\vert a_{t}\right\vert c_{1
c_{2}^{\ast}\left[ \left( \psi_{2,t}^{\ast}\nabla\psi_{1,t}-\psi_{1,t
\nabla\psi_{2,t}^{\ast}\right) \right] \nonumber\\
& +\frac{i\hbar}{m}\left\vert a_{t}\right\vert c_{1}^{\ast}c_{2}\left[
\left( \psi_{1,t}^{\ast}\nabla\psi_{2,t}-\psi_{2,t}\nabla\psi_{1,t}^{\ast
}\right) \right] .
\end{align}
In terms of the partial measured intensity of the partial wave function
$\rho_{t}^{\left( i\right) }$, $\mathbf{J}_{1}$ is
\begin{equation}
\mathbf{J}_{1}=\frac{i\left( 1+\left\vert a_{t}\right\vert ^{2}\right)
\hbar}{2m}\left[ \left\vert c_{1}\right\vert ^{2}\left( \psi_{1,t}^{\ast
}\psi_{1,t}\right) \frac{\left( \psi_{1,t}^{\ast}\nabla\psi_{1,t}-\psi
_{1,t}\nabla\psi_{1,t}^{\ast}\right) }{\left( \psi_{1,t}^{\ast}\psi
_{1,t}\right) }+\left\vert c_{2}\right\vert ^{2}\left( \psi_{2,t}^{\ast
\psi_{2,t}\right) \frac{\left( \psi_{2,t}^{\ast}\nabla\psi_{2,t}-\psi
_{2,t}\nabla\psi_{2,t}^{\ast}\right) }{\left( \psi_{2,t}^{\ast}\psi
_{2,t}\right) }\right] .\label{V1
\end{equation}
In a similar manner, $\mathbf{J}_{2\text{ }}$ is also rewritten a
\begin{align}
\mathbf{J}_{2} & =\frac{i\hbar}{m}\left\vert a_{t}\right\vert \left[ \left(
c_{1}c_{2}^{\ast}\psi_{1,t}\psi_{2,t}^{\ast}\frac{\psi_{1,t}^{\ast}\nabla
\psi_{1,t}}{\psi_{1,t}^{\ast}\psi_{1,t}}-c_{1}^{\ast}c_{2}\psi_{1,t}^{\ast
}\psi_{2,t}\frac{\psi_{1,t}\nabla\psi_{1,t}^{\ast}}{\psi_{1,t}^{\ast
\psi_{1,t}}\right) \right] \nonumber\\
& +\frac{i\hbar}{m}\left\vert a_{t}\right\vert \left[ \left( c_{1}^{\ast
}c_{2}\psi_{2,t}\psi_{1,t}^{\ast}\frac{\psi_{2,t}^{\ast}\nabla\psi_{2,t}
{\psi_{2,t}^{\ast}\psi_{2,t}}-c_{1}c_{2}^{\ast}\psi_{2,t}^{\ast}\psi
_{1,t}\frac{\psi_{2,t}\nabla\psi_{2,t}^{\ast}}{\psi_{2,t}^{\ast}\psi_{2,t
}\right) \right] .
\end{align}
Let us define the complex variable $\Lambda=c_{1}c_{2}^{\ast}\psi_{1,t
\psi_{2,t}^{\ast};$ then $\Lambda^{\ast}=c_{1}^{\ast}c_{2}\psi_{1,t}^{\ast
}\psi_{2,t}$, the current $\mathbf{J}_{2}$ can be rewritten as
\begin{align}
\mathbf{J}_{2} & =\frac{i\hbar}{m}\left\vert a_{t}\right\vert \left[ \left(
\Lambda\frac{\psi_{1,t}^{\ast}\nabla\psi_{1,t}}{\psi_{1,t}^{\ast}\psi_{1,t
}-\Lambda^{\ast}\frac{\psi_{1,t}\nabla\psi_{1,t}^{\ast}}{\psi_{1,t}^{\ast
\psi_{1,t}}\right) \right] \nonumber\\
& +\frac{i\hbar}{m}\left\vert a_{t}\right\vert \left[ \left( \Lambda^{\ast
}\frac{\psi_{2,t}^{\ast}\nabla\psi_{2,t}}{\psi_{2,t}^{\ast}\psi_{2,t}
-\Lambda\frac{\psi_{2,t}\nabla\psi_{2,t}^{\ast}}{\psi_{2,t}^{\ast}\psi_{2,t
}\right) \right] .\label{velocity
\end{align}
It is convenient to write $\Lambda=\Lambda_{1}+i\Lambda_{2}$ and
$\Lambda^{\ast}=\Lambda_{1}-i\Lambda_{2}$, where $\Lambda_{1}$ and
$\Lambda_{2}$ are real numbers. Substituting $\Lambda_{1}$ and $\Lambda_{2}$
into (\ref{velocity}), one obtain
\begin{align}
\mathbf{J}_{2} & =\frac{i\hbar}{m}\left\vert a_{t}\right\vert \left[
\Lambda_{1}\left( \frac{\psi_{1,t}^{\ast}\nabla\psi_{1,t}}{\psi_{1,t}^{\ast
}\psi_{1,t}}-\frac{\psi_{1,t}\nabla\psi_{1,t}^{\ast}}{\psi_{1,t}^{\ast
\psi_{1,t}}\right) +i\Lambda_{2}\left( \frac{\psi_{1,t}^{\ast}\nabla
\psi_{1,t}}{\psi_{1,t}^{\ast}\psi_{1,t}}+\frac{\psi_{1,t}\nabla\psi
_{1,t}^{\ast}}{\psi_{1,t}^{\ast}\psi_{1,t}}\right) \right] \nonumber\\
& +\frac{i\hbar}{m}\left\vert a_{t}\right\vert \left[ \Lambda_{1}\left(
\frac{\psi_{2,t}^{\ast}\nabla\psi_{2,t}}{\psi_{2,t}^{\ast}\psi_{2,t}
-\frac{\psi_{2,t}\nabla\psi_{2,t}^{\ast}}{\psi_{2,t}^{\ast}\psi_{2,t}}\right)
-i\Lambda_{2}\left( \frac{\psi_{2,t}^{\ast}\nabla\psi_{2,t}}{\psi_{2,t
^{\ast}\psi_{2,t}}+\frac{\psi_{2,t}\nabla\psi_{2,t}^{\ast}}{\psi_{2,t}^{\ast
}\psi_{2,t}}\right) \right] .
\end{align}
In term of the relation
\begin{equation}
\nabla\ln\left( \psi_{i,t}^{\ast}\psi_{i,t}\right) =\left( \frac{\psi
_{i,t}^{\ast}\nabla\psi_{i,t}}{\psi_{i,t}^{\ast}\psi_{i,t}}+\frac{\psi
_{i,t}\nabla\psi_{i,t}^{\ast}}{\psi_{i,t}^{\ast}\psi_{i,t}}\right) \text{
\ }i=1,2,
\end{equation}
finally, $\mathbf{J}_{2\text{ }}$can be expressed b
\begin{align}
\mathbf{J}_{2} & =\frac{i\hbar}{m}\left\vert a_{t}\right\vert \Lambda
_{1}\left[ \left( \frac{\psi_{1,t}^{\ast}\nabla\psi_{1,t}}{\psi_{1,t}^{\ast
}\psi_{1,t}}-\frac{\psi_{1,t}\nabla\psi_{1,t}^{\ast}}{\psi_{1,t}^{\ast
\psi_{1,t}}\right) +\left( \frac{\psi_{2,t}^{\ast}\nabla\psi_{2,t}
{\psi_{2,t}^{\ast}\psi_{2,t}}-\frac{\psi_{2,t}\nabla\psi_{2,t}^{\ast}
{\psi_{2,t}^{\ast}\psi_{2,t}}\right) \right] \nonumber\\
& +\frac{\hbar}{m}\left\vert a_{t}\right\vert \Lambda_{2}\mathbf{\nabla
}\left[ \ln\left( \frac{\psi_{1,t}^{\ast}\psi_{1,t}}{\psi_{2,t}^{\ast
\psi_{2,t}}\right) \right] .\label{D2
\end{align}
This formula shows there is a topological reason leading to the decoherence.
The new parameter $\Lambda_{1}$ can be used to indicate the coherent degree.
This parameter also can be called damping factor, but it is very different
from the parameter $a_{t}$. The parameter $a_{t}$ relates to the degrees of
the freedom of the environment. But from (\ref{D1}), the parameter
$\Lambda_{1}$ relates to the phase shift of the partial wave functions. The
parameter $\Lambda_{1}$ is indispensable to give the exact expression
(\ref{D2}), which is essential for giving the topological structure of the
current. Then the parameter $\Lambda_{1}$ is important to the topological
structure of the current, but the parameter $a_{t}$ has nothing to do with the
topological structure. In addition, we find $\hbar\mathbf{\nabla}\left[
\ln\left( \frac{\psi_{1,t}^{\ast}\psi_{1,t}}{\psi_{2,t}^{\ast}\psi_{2,t
}\right) \right] $ is a vector, then a new $U\left( 1\right) $ gauge
potential is defined by
\begin{equation}
\mathbf{A}=\hbar\mathbf{\nabla}\left[ \ln\left( \frac{\psi_{1,t}^{\ast
\psi_{1,t}}{\psi_{2,t}^{\ast}\psi_{2,t}}\right) \right] .
\end{equation}
Therefore, the current $\mathbf{J}_{2}$ is
\begin{align}
\mathbf{J}_{2} & =\frac{i\hbar}{m}\left\vert a_{t}\right\vert \Lambda
_{1}\left[ \left( \frac{\psi_{1,t}^{\ast}\nabla\psi_{1,t}}{\psi_{1,t}^{\ast
}\psi_{1,t}}-\frac{\psi_{1,t}\nabla\psi_{1,t}^{\ast}}{\psi_{1,t}^{\ast
\psi_{1,t}}\right) +\left( \frac{\psi_{2,t}^{\ast}\nabla\psi_{2,t}
{\psi_{2,t}^{\ast}\psi_{2,t}}-\frac{\psi_{2,t}\nabla\psi_{2,t}^{\ast}
{\psi_{2,t}^{\ast}\psi_{2,t}}\right) \right] \nonumber\\
& +\frac{1}{m}\left\vert a_{t}\right\vert \Lambda_{2}\mathbf{A.
\end{align}
We assume the system is in the coherence, that is to say, the damping factor
$\left\vert a_{t}\right\vert =1.$ The current using the measured intensity of
the partial wave function $\rho_{t}^{\left( i\right) }$ i
\begin{align}
\mathbf{J} & =\frac{i\hbar}{m}\left( \left\vert c_{1}\right\vert ^{2
\rho_{t}^{\left( 1\right) }+\Lambda_{1}\right) \left[ \frac{\left(
\psi_{1,t}^{\ast}\nabla\psi_{1,t}-\psi_{1,t}\nabla\psi_{1,t}^{\ast}\right)
}{\psi_{1,t}^{\ast}\psi_{1,t}}\right] +\frac{i\hbar}{m}\left( \left\vert
c_{2}\right\vert ^{2}\rho_{t}^{\left( 2\right) }+\Lambda_{1}\right) \left[
\frac{\left( \psi_{2,t}^{\ast}\nabla\psi_{2,t}-\psi_{2,t}\nabla\psi
_{2,t}^{\ast}\right) }{\psi_{2,t}^{\ast}\psi_{2,t}}\right] \nonumber\\
& +\frac{1}{m}\Lambda_{2}\mathbf{A
\end{align}
In order to study the current in detail, we define new charges $q_{1
=\frac{\left( \left\vert c_{1}\right\vert ^{2}\rho_{t}^{\left( 1\right)
}+\Lambda_{1}\right) }{\rho_{t}^{\left( 1\right) }}$ and $q_{2
=\frac{\left( \left\vert c_{2}\right\vert ^{2}\rho_{t}^{\left( 2\right)
}+\Lambda_{1}\right) }{\rho_{t}^{\left( 2\right) }}.$ Then the new
$U\left( 1\right) $ gauge potential is given b
\[
\widetilde{\mathbf{A}}=\frac{\Lambda_{2}}{4\left( q_{1}^{2}+q_{2}^{2}\right)
\left( \left\vert \psi_{1,t}\right\vert ^{2}+\left\vert \psi_{2,t}\right\vert
^{2}\right) }\mathbf{A.
\]
Based on the new gauge potential $\widetilde{\mathbf{A}}\mathbf{,}$ The
current of the coherent system can be expressed by
\begin{align}
\mathbf{J} & =\frac{i\hbar q_{1}}{m}\left( \psi_{1,t}^{\ast}\nabla\psi
_{1,t}-\psi_{1,t}\nabla\psi_{1,t}^{\ast}\right) +\frac{i\hbar q_{2}
{m}\left( \psi_{2,t}^{\ast}\nabla\psi_{2,t}-\psi_{2,t}\nabla\psi_{2,t}^{\ast
}\right) \nonumber\\
& +\frac{4\left( q_{1}^{2}+q_{2}^{2}\right) }{m}\left( \left\vert
\psi_{1,t}\right\vert ^{2}+\left\vert \psi_{2,t}\right\vert ^{2}\right)
\widetilde{\mathbf{A}}.\label{totalcurrent
\end{align}
However, we find the total current can be deduced from the following free
energ
\begin{equation}
F=\left[ \frac{1}{2m}\left\vert \left( \hbar\partial_{k}+i\frac{2q_{1}
{c}\widetilde{\mathbf{A}}_{k}\right) \psi_{1,t}\right\vert ^{2}+\frac{1
{2m}\left\vert \left( \hbar\partial_{k}+i\frac{2q_{2}}{c}\widetilde
{\mathbf{A}}_{k}\right) \psi_{2,t}\right\vert ^{2}+V\left( \left\vert
\psi_{a,t}\right\vert ^{2}\right) +\frac{\widetilde{\mathbf{B}}^{2}}{8\pi
}\right] ,\label{free energy
\end{equation}
where $\widetilde{\mathbf{B}}=\mathbf{\nabla\times}\widetilde{\mathbf{A}}$ is
$U(1)$ gauge field. The potential $V\left( \left\vert \psi_{a,t}\right\vert
^{2}\right) $ i
\begin{equation}
V\left( \left\vert \psi_{a,t}\right\vert ^{2}\right) =-b_{a}\left\vert
\psi_{a,t}\right\vert ^{2}+\frac{c_{a}}{2}\left\vert \psi_{a,t}\right\vert
^{4}.\text{ \ }a=1,2.
\end{equation}
It is well known this free energy is called Ginzberg-Landau free energy, which
is used to describe the charged two-condensate Bose system \cite{TGBB}. The
total current (\ref{totalcurrent}) of quantum coherent system is similar to
the supercurrent of the charged two-condensate Bose system. In two-condensate
superconductor, the charged two-condensate wave functions, or charged order
parameters, can carry the electronic charges. The interaction of charged order
parameters is mediated by the electromagnetic potential $\mathbf{A}_{e}$. In
this description, we find the coherent system interacting with the environment
is similar to the two-condensate superconductor. The partial wave functions
can be seen as the charged order parameters. The partial wave functions are
weakly-coupled because they carry the charges $q_{1}$ and $q_{2}$, which is
different from the electronic charge. The interaction of the partial wave
functions is mediated by the new $U\left( 1\right) $ gauge potential
$\widetilde{\mathbf{A}}$, not the electromagnetic potential.
\section{the symmetry of the current and its topology}
In this section, we try to study the free energy, symmetry and the topological
properties of the current of the coherent state. Let us define the partial
wave function a
\begin{equation}
\psi_{a,t}=\sqrt{2m}\rho\xi_{a}\text{ \ }a=1,2,
\end{equation}
where the complex variable $\xi_{a}=\left\vert \xi_{a}\right\vert e^{i\theta
$. The modular $\rho$ i
\begin{equation}
\rho=\frac{1}{2}\left( \frac{\left\vert \psi_{1,t}\right\vert ^{2}}{m
+\frac{\left\vert \psi_{2,t}\right\vert ^{2}}{m}\right) .
\end{equation}
By using these new variables, the Ginzberg-Landau-like free energy of the
coherent state is given as
\begin{align}
F & =\hbar^{2}\left( \partial\rho\right) ^{2}+\hbar^{2}\rho^{2}\left\vert
\left( \partial_{k}+i\frac{2q_{1}}{\hbar c}\widetilde{\mathbf{A}}\right)
\xi_{1}\right\vert ^{2}+\hbar^{2}\rho^{2}\left\vert \left( \partial
_{k}+i\frac{2q_{2}}{\hbar c}\widetilde{\mathbf{A}}\right) \xi_{2}\right\vert
\nonumber\\
& +V\left( \left\vert \psi_{a,t}\right\vert ^{2}\right) +\frac
{\widetilde{\mathbf{B}}^{2}}{8\pi}.\label{FE
\end{align}
It can be rewritten b
\begin{align}
F & =\hbar^{2}\left( \partial\rho\right) ^{2}+\hbar^{2}\rho^{2}\left(
\left\vert \partial\xi_{1}\right\vert ^{2}+\left\vert \partial\xi
_{2}\right\vert ^{2}\right) +V\left( \left\vert \psi_{a,t}\right\vert
^{2}\right) +\frac{\widetilde{\mathbf{B}}^{2}}{8\pi}\nonumber\\
& +\hbar^{2}\rho^{2}\left[ i\frac{2q_{1}}{\hbar c}\left( \widetilde
{\mathbf{A}}\xi_{1}\partial\xi_{1}^{\ast}-\widetilde{\mathbf{A}}\xi_{1}^{\ast
}\partial\xi_{1}\right) +\frac{4q_{1}^{2}}{\hbar^{2}c^{2}}\left\vert \xi
_{1}\right\vert ^{2}\widetilde{\mathbf{A}}\right] \nonumber\\
& +\hbar^{2}\rho^{2}\left[ i\frac{2q_{2}}{\hbar c}\left( \widetilde
{\mathbf{A}}\xi_{2}\partial\xi_{2}^{\ast}-\widetilde{\mathbf{A}}\xi_{2}^{\ast
}\partial\xi_{2}\right) +\frac{4q_{2}^{2}}{\hbar^{2}c^{2}}\left\vert \xi
_{2}\right\vert ^{2}\widetilde{\mathbf{A}}\right] .
\end{align}
The supercurrent of the free energy can be derived a
\begin{equation}
\mathbf{J=}i\hbar^{2}\rho^{2}\left[ \frac{2q_{1}}{\hbar c}\left( \xi
_{1}\partial\xi_{1}^{\ast}-\xi_{1}^{\ast}\partial\xi_{1}\right) +\frac
{2q_{2}}{\hbar c}\left( \xi_{2}\partial\xi_{2}^{\ast}-\xi_{2}^{\ast
\partial\xi_{2}\right) \right] +\hbar^{2}\rho^{2}\left( \frac{4q_{1}^{2
}{\hbar^{2}c^{2}}\left\vert \xi_{1}\right\vert ^{2}+\frac{4q_{2}^{2}
{\hbar^{2}c^{2}}\left\vert \xi_{2}\right\vert ^{2}\right) \widetilde
{\mathbf{A}}.
\end{equation}
Let $\Delta=\left( \frac{4q_{1}^{2}}{\hbar^{2}c^{2}}\left\vert \xi
_{1}\right\vert ^{2}+\frac{4q_{2}^{2}}{\hbar^{2}c^{2}}\left\vert \xi
_{2}\right\vert ^{2}\right) $, the new supercurrent $\widetilde{\mathbf{J}}$
is given a
\begin{align}
\widetilde{\mathbf{J}} & =\frac{\mathbf{J}}{\hbar^{2}\rho^{2}\Delta
}\nonumber\\
& =i\left[ \frac{2q_{1}}{\hbar c\Delta}\left( \xi_{1}\partial\xi_{1}^{\ast
}-\xi_{1}^{\ast}\partial\xi_{1}\right) +\frac{2q_{2}}{\hbar c\Delta}\left(
\xi_{2}\partial\xi_{2}^{\ast}-\xi_{2}^{\ast}\partial\xi_{2}\right) \right]
+\widetilde{\mathbf{A}}.
\end{align}
To find the symmetry of the coherent system, a new complex variable
$\widetilde{\xi_{a}}$ is defined by
\begin{equation}
\widetilde{\xi_{a}}=\sqrt{\frac{2q_{a}}{\hbar\Delta Qc}}\xi_{a},
\end{equation}
where the real number $Q$ guarantee the new partial wave functions satisf
\begin{equation}
\left\vert \widetilde{\xi_{1}}\right\vert ^{2}+\left\vert \widetilde{\xi_{2
}\right\vert ^{2}=1.
\end{equation}
In terms of the new complex variable, the supercurrent $\widetilde{\mathbf{J
}$ i
\begin{equation}
\widetilde{\mathbf{J}}\mathbf{=}iQ\left[ \left( \widetilde{\xi}_{1
\partial\widetilde{\xi}_{1}^{\ast}-\widetilde{\xi}_{1}^{\ast}\partial
\widetilde{\xi}_{1}\right) +\left( \widetilde{\xi}_{2}\partial\widetilde
{\xi}_{2}^{\ast}-\widetilde{\xi}_{2}^{\ast}\partial\widetilde{\xi}_{2}\right)
\right] +\widetilde{\mathbf{A}}.
\end{equation}
Next we define a gauge invariant unit vector $\widetilde{\mathbf{n}}
\begin{equation}
\widetilde{\mathbf{n}}\mathbf{=}\left( \overline{\widetilde{\xi
},\mathbf{\sigma}\widetilde{\xi}\right) ,
\end{equation}
where $\overline{\widetilde{\xi}}=\left( \widetilde{\xi}_{1}^{\ast
},\widetilde{\xi}_{2}^{\ast}\right) $ and $\sigma$ are the Pauli matrices. It
is obvious the unit vector satisfie
\[
\widetilde{\mathbf{n}}\bullet\widetilde{\mathbf{n}}=1.
\]
Then a new vector $\mathbf{C}$ can be defined b
\begin{equation}
\mathbf{C=}Q\frac{\mathbf{j}}{2}+\widetilde{\mathbf{A}},
\end{equation}
where $\mathbf{j=}i\left[ \left( \widetilde{\xi}_{1}\partial\widetilde{\xi
}_{1}^{\ast}-\widetilde{\xi}_{1}^{\ast}\partial\widetilde{\xi}_{1}\right)
+\left( \widetilde{\xi}_{2}\partial\widetilde{\xi}_{2}^{\ast}-\widetilde{\xi
}_{2}^{\ast}\partial\widetilde{\xi}_{2}\right) \right] .$ We add and
subtract from (\ref{FE}) a term $\frac{1}{4}\hbar^{2}\rho^{2}Q^{2}\Delta
^{2}\mathbf{j}^{2}$, the two charged free energy of the coherent state can be
expressed with these new variable
\begin{align}
F & =\hbar^{2}\left( \partial\rho\right) ^{2}+\frac{\hbar^{2}\rho^{2
Q^{2}\Delta^{2}}{4}\left( \partial\widetilde{\mathbf{n}}\right) ^{2
+\frac{1}{8\pi}\left[ \left( \partial_{i}C_{j}-\partial_{j}C_{i}\right)
-\frac{Q}{4}\widetilde{\mathbf{n}}\cdot\mathbf{\partial}_{i}\widetilde
{\mathbf{n}}\times\mathbf{\partial}_{j}\widetilde{\mathbf{n}}\right]
+\hbar^{2}\rho^{2}\Delta\mathbf{C}^{2}+V\nonumber\\
& +\hbar^{2}\rho^{2}\left[ \left( 1-\frac{2q_{1}}{\hbar c}\right)
\left\vert \partial\xi_{1}\right\vert ^{2}+\left( 1-\frac{2q_{2}}{\hbar
c}\right) \left\vert \partial\xi_{2}\right\vert ^{2}\right] .
\end{align}
Considering the London limit, we have $\partial\rho=0$, the free energy is
given b
\begin{align}
F & =\frac{\hbar^{2}\rho^{2}Q^{2}\Delta^{2}}{4}\left( \partial
\widetilde{\mathbf{n}}\right) ^{2}+\frac{1}{8\pi}\left[ \left( \partial
_{i}C_{j}-\partial_{j}C_{i}\right) -\frac{Q}{4}\widetilde{\mathbf{n}
\cdot\mathbf{\partial}_{i}\widetilde{\mathbf{n}}\times\mathbf{\partial
_{j}\widetilde{\mathbf{n}}\right] +\hbar^{2}\rho^{2}\Delta\mathbf{C
^{2}+V\nonumber\\
& +\hbar^{2}\rho^{2}\left[ \left( 1-\frac{2q_{1}}{\hbar c}\right)
\left\vert \partial\xi_{1}\right\vert ^{2}+\left( 1-\frac{2q_{2}}{\hbar
c}\right) \left\vert \partial\xi_{2}\right\vert ^{2}\right] .
\end{align}
Finally, we find there is a stable knotted solution in coherent system, which
is described by the Skyme-Faddeev-Niemi actio
\begin{equation}
F_{0}=\frac{\rho^{2}\hbar^{2}Q^{2}\Delta^{2}}{4}\left( \partial
\widetilde{\mathbf{n}}\right) ^{2}+\frac{Q}{32\pi}\widetilde{\mathbf{n}
\cdot\mathbf{\partial}_{i}\widetilde{\mathbf{n}}\times\mathbf{\partial
_{j}\widetilde{\mathbf{n}}.
\end{equation}
The knotted solution displays a $O\left( 3\right) $ symmetry in the free
energy. The knotted solution is just the nontrivial map
\begin{equation}
\widetilde{\mathbf{n}}:S^{3}\rightarrow S^{2}.
\end{equation}
The boundary condition of this knotted solution i
\begin{equation}
\widetilde{\mathbf{n}}\left( x\right) \rightarrow\widetilde{\mathbf{n}
_{0}\text{ \ \ }\mathbf{x\rightarrow\infty
\end{equation}
where $\mathbf{n}_{0}$ is the constant vector in spatial direction. The
knotted solution has an important relation to the current of the coherent
state. It is convenient to write the current a
\begin{equation}
\mathbf{J}=\mathbf{J}_{1}+\mathbf{J}_{2},
\end{equation}
where
\begin{equation}
\mathbf{J}_{1}=\frac{\hbar q_{1}\rho_{t}^{\left( 1\right) }}{im}\left[
\frac{\left( \psi_{1,t}^{\ast}\nabla\psi_{1,t}-\psi_{1,t}\nabla\psi
_{1,t}^{\ast}\right) }{\psi_{1,t}^{\ast}\psi_{1,t}}\right] +\frac{4q_{1
^{2}}{m}\left\vert \psi_{1,t}\right\vert ^{2}\widetilde{\mathbf{A}
\end{equation}
an
\begin{equation}
\mathbf{J}_{2}=\frac{\hbar q_{2}\rho_{t}^{\left( 2\right) }}{im}\left[
\frac{\left( \psi_{2,t}^{\ast}\nabla\psi_{2,t}-\psi_{2,t}\nabla\psi
_{2,t}^{\ast}\right) }{\psi_{2,t}^{\ast}\psi_{2,t}}\right] +\frac{4q_{2
^{2}}{m}\left\vert \psi_{2,t}\right\vert ^{2}\widetilde{\mathbf{A}}.
\end{equation}
Recalling the unit vector $\mathbf{n}_{\left( i\right) }$, these components
can be derived a
\begin{equation}
\mathbf{J}_{1}=\frac{\hbar q_{1}\rho_{t}^{\left( 1\right) }}{m}\epsilon
_{ab}n_{(1)}^{a}\partial_{i}n_{(1)}^{b}+\frac{4q_{1}^{2}}{m}\left\vert
\psi_{1,t}\right\vert ^{2}\widetilde{\mathbf{A}
\end{equation}
an
\begin{equation}
\mathbf{J}_{2}=\frac{\hbar q_{2}\rho_{t}^{\left( 2\right) }}{m}\epsilon
_{ab}n_{(2)}^{a}\partial_{i}n_{(2)}^{b}+\frac{4q_{2}^{2}}{m}\left\vert
\psi_{2,t}\right\vert ^{2}\widetilde{\mathbf{A}}.
\end{equation}
By making use of $\phi-mapping$ topological current theory, the vorticity of
the current is given a
\begin{equation}
\Gamma=\int_{\Sigma_{i}}\left( \nabla\times\mathbf{J}\right) \cdot
d\mathbf{S}=\int_{\Sigma_{i}}\left( \nabla\times\mathbf{J}_{1}\right) \cdot
d\mathbf{S+}\int_{\Sigma_{i}}\left( \nabla\times\mathbf{J}_{2}\right) \cdot
d\mathbf{S.
\end{equation}
Then the curl of the currents $\mathbf{J}_{1}$ and $\mathbf{J}_{2}$ are
calculated as
\begin{equation}
\left( \nabla\times\mathbf{J}_{1}\right) \mathbf{=}\frac{\hbar q_{1}\rho
_{t}^{\left( 1\right) }}{m
{\displaystyle\sum\limits_{l=1}^{\ell}}
\beta_{l}^{\left( 1\right) }\eta_{l}^{\left( 1\right) }\delta_{\left(
1\right) }^{2}\left( x^{i}-z_{l}^{i}\right) \frac{dx_{\left( 1\right)
}^{i}}{ds}+\frac{4q_{1}^{2}}{m}\left\vert \psi_{1,t}\right\vert ^{2
\mathbf{\nabla}\times\widetilde{\mathbf{A}
\end{equation}
an
\begin{equation}
\left( \nabla\times\mathbf{J}_{2}\right) \mathbf{=}\frac{\hbar q_{2}\rho
_{t}^{\left( 2\right) }}{m
{\displaystyle\sum\limits_{l=1}^{\ell}}
\beta_{l}^{\left( 2\right) }\eta_{l}^{\left( 2\right) }\delta_{\left(
2\right) }^{2}\left( x^{i}-z_{l}^{i}\right) \frac{dx_{\left( 2\right)
}^{i}}{ds}+\frac{4q_{2}^{2}}{m}\left\vert \psi_{2,t}\right\vert ^{2
\mathbf{\nabla\times}\widetilde{\mathbf{A}}.
\end{equation}
Furthermore, the vorticity of the current i
\begin{equation}
\Gamma=\frac{\hbar q_{1}\rho_{t}^{\left( 1\right) }}{m}W_{1}+\frac{\hbar
q_{2}\rho_{t}^{\left( 2\right) }}{m}W_{2}+\frac{4\left( q_{1}^{2}\rho
_{t}^{\left( 1\right) }+q_{1}^{2}\rho_{t}^{\left( 2\right) }\right)
{m}\int_{\Sigma_{i}}\left( \mathbf{\nabla\times}\widetilde{\mathbf{A
}\right) \cdot d\mathbf{S}.
\end{equation}
It is well known the property of a supercurrent is the magnetic flux passing
through any area bounded by such a current is quantized. The quantization of
the flux in the superconductor i
\[
\int_{\Sigma_{i}}\left( \mathbf{\nabla\times A}_{E}\right) \cdot
d\mathbf{S=}\frac{h}{2e}\widetilde{W},
\]
where $e$ is the electronic charge. Similarly, we give the flux quantization
of this new $U\left( 1\right) $ gauge potential $\widetilde{\mathbf{A}}
\begin{equation}
\int_{\Sigma_{i}}\left( \mathbf{\nabla\times}\widetilde{\mathbf{A}}\right)
\cdot d\mathbf{S}=\frac{h}{q_{1}+q_{2}}\widetilde{W}.
\end{equation}
Finally, the vorticity of the current i
\begin{equation}
\Gamma=\frac{\hbar q_{1}\rho_{t}^{\left( 1\right) }}{m}W_{1}+\frac{\hbar
q_{2}\rho_{t}^{\left( 2\right) }}{m}W_{2}+\frac{4\left( q_{1}^{2}\rho
_{t}^{\left( 1\right) }+q_{1}^{2}\rho_{t}^{\left( 2\right) }\right)
{m}\frac{\hbar}{q_{1}+q_{2}}\widetilde{W}.
\end{equation}
\section{conclusion}
In this paper, the relation between the coherent quantum system and the
charged two-condensate system is investigated. The new expression of the
current of the coherent state is given based on reduced density trajectory and
$\phi-mapping$ topological current theory. A topological reason leading to the
decoherence is found. By defining a new $U\left( 1\right) $ gauge potential
$\widetilde{\mathbf{A}}$ and new charges $q_{1}$ and $q_{2},$ we find that the
coherent system can be described by the Ginzberg-Landau-like model with two
charged Cooper pairs. The corresponding relation between coherent system and
two-gap superconductor is shown as follows: the partial wave functions of the
coherence correspond to the charged two-condensate wave functions; the charges
$q_{1}$ and $q_{2}$ correspond to the electronic charges; the new $U\left(
1\right) $ gauge potential $\widetilde{\mathbf{A}}$ corresponds to the
electromagnetic potential $\mathbf{A}e$. Finally, the hidden $O\left(
3\right) $ symmetry of the coherent state is found using Faddeev's $O\left(
3\right) $ nonlinear $\sigma$-model and the topological properties of the
knot solution are studied based on $\phi-$mapping topological current theory.
\begin{acknowledgments}
This work is supported by the Fundamental Research Funds for the Central Universities.
\end{acknowledgments}
|
\section{Introduction}
\label{sec:intro}
Binary neutron star (BNS) mergers are primary sources of gravitational
waves (GWs) and are associated with a variety of electromagnetic (EM)
emissions.
GW observations of BNS are eagerly expected in the upcoming
LIGO-Virgo runs, after the first binary black hole (BBH) GW detections
GW150914~\cite{Abbott:2016blz} and GW151226~\cite{Abbott:2016nmj}.
Such GWs will allow us to place constraints on
the nature of matter at densities above nuclear density,
e.g.~\cite{DelPozzo:2013ala}, and to identify the origin of EM
emissions like kilo/macronovae and short gamma-ray burst (SGRB),
e.g.~\cite{Rosswog:2015nja,Fernandez:2015use}.
Kilo/macronovae events are transient emissions in the optical or
near-infrared band observed in e.g.~\cite{Tanvir:2013pia}. They are believed to be
triggered by the radioactive decay of r-process nuclei
in the neutron-rich material ejected during a BNS merger.
SGRBs models are instead based on highly relativistic outflows, powered
e.g.~by the merger remnant accretion disk~\cite{Paczynski:1986px,Eichler:1989ve}.
Their joint observation with GWs might be challenging due to the short duration of
the burst and to the fact that they are highly collimated
emissions~\cite{Soderberg:2006bn}.
Additionally, the interaction of mildly or sub- relativistic outflows with
the surrounding material generates synchrotron radiation known as radio
flares~\cite{Nakar:2011cw}. This emission can persist from months to
years after merger, depending on the composition, which makes radio
flares a particularly attractive EM counterpart to detect.
Understanding the dependency of the GW and EM emissions on the
source's parameters is of key importance for GW astronomy and
multimessenger astrophysics. The BNS parameter space is composed by the
component masses (and spins), and by a choice of equation of state
(EOS) describing the NS matter. The ranges of the mass and spins
parameters are rather uncertain.
The expected NS mass range is $\sim 0.9 - 3 M_\odot$. The lower bound is
inferred from the formation scenario (gravitational-collapse) and from
current observations, although those measurements have typically large uncertainties,
see e.g.~\cite{Rawls:2011jw,Ozel:2012ax}. The upper bound is inferred
from stability argument (maximum theoretical mass), from precise
measurements of $\sim 2M_\odot$ NSs in double NS
systems~\cite{Demorest:2010bx,Antoniadis:2013pzd}, and from models of
SGRBs, which suggest a maximum mass of
$2.2M_\odot$, e.g.~\cite{Lawrence:2015oka}. There exist observations
of larger NS masses but with large uncertainties, so they cannot give
strong constraints on the NS maximum mass \cite{Lattimer:2012nd}.
Also, the maximum mass of a NS is determined by the particular
EOS. Most tabulated EOS compatible with astrophysical constraints
support NSs with maximum masses in a range of $\sim 2-3M_\odot$.
The above considerations suggest that the BNS mass-ratio
\begin{equation}
q := M^A/M^B\geq1 \ ,
\end{equation}
where $M^{A,B}$ are the individual gravitational masses of the NSs (in
isolation), is most likely constrained to $q\lesssim 2-3$.
Observations suggest that BNS systems consist of equal mass NSs with
masses of around $\sim 1.35 M_\odot$ and that the mass-ratio is close
to one e.g.~\cite{Swiggum:2015yras,Lattimer:2012nd}. However, only
approximately a dozen BNS systems are known so far and only $6$ of these
systems have well determined masses and will merge within a Hubble time.
The lack of very unequal mass configurations might only be a selection
effect. For example Ref.~\cite{Martinez:2015mya} discovered a compact
binary system with a mass ratio of $q \approx 1.3$, suggesting that
BNS with larger mass ratios exist. Population synthesis models for binaries
formed ``\emph{in situ}'' predict a wider range of masses and
mass ratios up to $q\approx 1.9$~\cite{Dominik:2012kk,Dietrich:2015pxa}.
Also NS spins are constrained by theoretical arguments and observations,
e.g. \cite{Burgay:2003jj,Lynch:2011aa,Miller:2014aaa,Dietrich:2015pxa}.
However, we do not consider here the NS rotation
and we remind to the above references and to future work \cite{Dietrich:2016prep2}.
Parameter space investigation of BNS are challenging due to the
unknown EOS, and the need of simulating each masses (and spins)
configuration with different EOS.
Most numerical relativity studies of BNS systems have focused on equal
masses and irrotational configurations. The first simulations of unequal
mass systems have been presented in
\cite{Shibata:2003ga,Rezzolla:2010fd,Hotokezaka:2012ze} using
polytropic and piecewise polytropic EOS.
Mass ejecta in $q\neq1$ BNS simulations have been studied in
e.g.~\cite{Bauswein:2013yna,Hotokezaka:2012ze,Dietrich:2015iva}.
Unequal mass simulations with microphysical EOS and neutrino cooling
have been presented in
e.g.~\cite{Bernuzzi:2015opx,Lehner:2016wjg,Lehner:2016lxy}
and with radiation-hydrodynamics in~\cite{Sekiguchi:2016bjd}.
Previous works were restricted to mass ratios $q\leq 1.35$.
Overall, the main results are that i) asymmetric mergers produce more
massive ejecta with smaller electron fractions than the corresponding equal mass setups
ii) unequal mass systems are likely to produce kilonovae and
iii) the remnant disk mass increases for an increasing mass ratio.
In \cite{Dietrich:2015pxa} we reported an upgrade of the SGRID code
able to generate generic initial data for BNS simulations together
with few preliminary evolutions. Among other results, we showed the
possibility of generating ``large mass-ratio'' configurations with
$q\sim2$. The test evolution of a $q=2$ BNS showed interesting features,
including large mass ejection and mass transfer from one star to the
other during the last revolutions.
In this work we study the effect of the binary's mass-ratio $q$ on the
GWs and on the characteristics of possible EM emission
associated to dynamical mass ejecta, in particular macronovae and radio flares.
We present a new set of (3+1)D numerical relativity simulations
of the merger and postmerger phase, and focus on a previously
inaccessible region of the binary parameter space spanning
$q\in[1,1.75]$ for different masses and equations of state, and a
setup with $q=2$.
The article is structured as follows:
In Sec.~\ref{sec:simeth}, we give a short description of the numerical methods
and describe important quantities used to analyze our simulations.
Section~\ref{sec:config} summarizes our configurations and the investigated part
of the BNS parameter space. Section~\ref{sec:dynamics} deals with the dynamics of the simulation,
where in particular we focus on the mass-transfer, the ejecta, and the final remnant.
The GW signal is investigated in Sec.~\ref{sec:GW} with respect to spectrograms,
the sky location and the emitted GW energy per mode.
Sec.~\ref{sec:EM} focuses on EM counterparts and
we conclude in Sec.~\ref{sec:conclusion}.
In Appendix~\ref{sec:accuracy} we test the accuracy of our simulations with respect to conserved quantities,
convergence, the constraints.
Throughout this work we use geometric units, setting $c=G=M_\odot=1$,
though we will sometimes include $M_\odot$ explicitly or quote values
in CGS units for better understanding. Spatial indices are denoted by
Latin letters running from 1 to 3 and Greek letters are used for
spacetime indices running from 0 to 3.
\section{Simulation methods}
\label{sec:simeth}
\subsection{Initial configurations}
Our initial configurations are constructed with the SGRID
code~\cite{Tichy:2006qn,Tichy:2009yr,Tichy:2009zr}.
SGRID uses pseudospectral methods to accurately compute spatial derivatives.
To obtain stationary configurations that are appropriate
as initial data, we use the conformal thin sandwich equations~\cite{Wilson:1995uh,Wilson:1996ty,York:1998hy}
together with assumptions of the constant rotational velocity approach~\cite{Tichy:2011gw,Tichy:2012rp},
which allows to construct generic NS binaries in hydrodynamical equilibrium.
The computational domain is divided into six patches (Fig.~1 of \cite{Dietrich:2015pxa}).
The domain reaches spatial infinity and thus allows to impose exact boundary conditions at spatial infinity.
Recent changes presented in~\cite{Dietrich:2015pxa} allow us to construct
configurations in almost all corners of the BNS parameter space
including high mass ratios, spinning configurations, low and high eccentricity inspirals,
as well as more realistic EOSs and highly compact stars.
In this work we do not study the influence of the eccentricity and spin, but
focus on the high-mass ratio configuration and the influence of the EOS.
We employ $n_A=28,n_B=28,n_\varphi=8,n_{\rm Cart} = 24$ points for the spectral grid,
cf.~\cite{Dietrich:2015pxa} for more details. In Tab.~\ref{tab:ID_and_grid_irr} we report
the initial parameters, in particular, the initial ADM-mass and angular momentum of the system as well as
the initial GW frequency.
\subsection{Evolutions}
Dynamical simulations are performed with the BAM
code~\cite{Brugmann:2008zz,Thierfelder:2011yi,Dietrich:2015iva}.
We use the Z4c scheme~\cite{Bernuzzi:2009ex,Hilditch:2012fp} and
employ the 1+log and gamma-driver conditions for the evolution of the lapse
and shift~\cite{Bona:1994a,Alcubierre:2002kk,vanMeter:2006vi}.
The equations of general-relativistic hydrodynamics (GRHD) are solved in
conservative form by defining Eulerian conservative variables from the
rest-mass density $\rho$, pressure $p$, internal energy $\epsilon$, and
3-velocity, $v^i$. The system is closed by an EOS.
We model the EOS with piecewise polytropic fits of~\cite{Read:2008iy},
see Tab.~\ref{tab:EOS} below.
Thermal effects are included during the simulation by an additive pressure
contribution given by
$p_{\rm th} = (\Gamma_{\rm th}-1)\rho\epsilon$~\cite{Shibata:2005ss,Bauswein:2010dn}.
As in our previous work we set $\Gamma_{\rm th}=1.75$.
The evolution algorithm is based on the method-of-lines with explicit
4th order Runge-Kutta time integrators. Finite difference stencils of 4th order
are employed for the spatial derivatives of the metric.
GRHD is solved by means of a high-resolution-shock-capturing
method~\cite{Thierfelder:2011yi} based on primitive reconstruction
and the Local-Lax-Friedrichs (LLF) central scheme for the numerical
fluxes. Primitive reconstruction is performed with the 5th order WENOZ
scheme of~\cite{Borges20083191}. We do not employ the higher-order algorithm
presented recently in~\cite{Bernuzzi:2016pie}, because the new method was
not tested at the time these simulations were performed.
The numerical domain is made of a hierarchy of cell-centered nested
Cartesian grids. The hierarchy consists of $L$ levels of refinement labeled
by~$l = 0,...,L-1$. A refinement level $l$ has one or more
Cartesian grids with constant grid spacing $h_l$ and $n$ points per
direction. The refinement factor is two such that $h_l = h_0/2^l$. The grids
are properly nested, i.e., the coordinate extent of any grid at
level~$l$, $l > 0$, is completely covered by the grids at level~$l-1$.
Some of the mesh refinement levels $l>l^{\rm mv}$ can be dynamically moved and adapted
during the time evolution according to the technique of~``moving boxes''.
The BAM grid setups considered in this work consist of 7 refinement levels.
We distinguish between two different grid setups.
For the first setup, we substitute
the outermost level ($l=0$) by a multipatch cubed-sphere
grid~\cite{Ronchi:1996,Thornburg:2004dv,Pollney:2009yz,Hilditch:2012fp}
on which we do not solve the GRHD equations.
The spheres allow us to apply constraint preserving boundary conditions~\cite{Ruiz:2010qj}.
This `shell' setup is denoted in Tab.~\ref{tab:ID_and_grid_irr} with the ending `s'.
The second grid setup does not make use of the multipatch cubed-spheres, so
all the refinement levels are Cartesian boxes. While this has the disadvantage that only
Sommerfeld boundary conditions are used, it increases the computational speed
and the matter fields are evolved on a larger region. We denote this `box'
setup by the ending `b' in Tab.~\ref{tab:ID_and_grid_irr}.
The Berger-Oliger algorithm is employed for the time stepping~\cite{Berger:1984zza}
on the inner levels. As in~\cite{Dietrich:2014wja,Dietrich:2015iva,Dietrich:2015pxa}
we make use of an additional refluxing algorithm to enforce mass conservation
across mesh refinement boundaries based on~\cite{East:2011aa,Berger:1989}. Restriction and prolongation between the
refinement levels is performed with an average scheme and 2nd order essentially non-oscillatory
scheme, respectively.
\subsection{Simulation analysis}
\label{sec:diagno}
We briefly want to summarize important quantities for our simulation analysis.
For this purpose we follow the description of~\cite{Dietrich:2015iva} including
also quantities to monitor the mass transfer and the GW spectrogram. \\
\paragraph*{Mass transfer:}
Because of the high mass ratios considered in this work,
one might expect mass transport between the two NSs during the inspiral prior to merger.
In our last work~\cite{Dietrich:2015iva} we estimated the amount of material
flowing from one star to the other by measuring how
much mass is inside the Cartesian refinement boxes.
Now, we estimate the mass transfer by computing the change of baryonic mass
inside a coordinate sphere around the center of the NS.
The center of the NS is defined as the minimum of the lapse.
We perform the integration for different coordinate radii $r_{c,{\rm min}}=8.0$
up to $r_{c,{\rm max}}=13.0$ in steps of $\Delta r_c=0.5$.
While a radius chosen too small does not cover all of the tidally deformed
star, if the radius is chosen too large the mass measurement will be effected by the other star a
significant time before the actual merger.
We will focus on the radii $r_{c1}= r_{c2}=10.5$.\\
\paragraph*{Merger remnant:}
In agreement with the literature, e.g.~\cite{Baumgarte:1999cq},
we define the merger remnant as a hypermassive neutron star
(HMNS) if its rest-mass is larger than the maximum rest-mass of
a stable uniformly rotating star with the same EOS;
or as a supramassive neutron star (SMNS)
if its rest-mass is smaller than the maximum rest-mass
of a stable uniformly rotating star, but above the
mass of a stable TOV-star.
In case its mass is also below the maximum supported mass
of a TOV-star, we call it simply a massive neutron star (MNS).
Notice that these definitions \textit{cannot} be used strictly and should be seen as a
qualitative description since they refer to equilibrium configurations
assuming barotropic EOS and axisymmetry.
For a HMNS the merger remnant collapses on a dynamical timescale to a BH.
Typically the lifetime $\tau$ is the time from the moment of merger
to the time an apparent horizon forms.
The final BH is characterized by its horizon
mass $M_{\rm BH}$ and dimensionless spin $j_{\rm BH}$.
The accretion disk around the BH has a mass of
\begin{equation}
M_\text{disk} = \int_{r>r_{\rm AH}} \text{d}^3 x \ q^{(D)} \ ,
\end{equation}
with $ q^{(D)} = \sqrt{\gamma} D$ where $D$ is the fluid's rest frame baryonic mass, $\gamma$ the
determinant of the 3-metric, and the domain of
integration excludes the spherical region inside the apparent horizon.\\
\paragraph*{Mass ejecta:}
As in~\cite{Dietrich:2015iva} we label material as ejecta when
\begin{equation}
\label{eq:unbound}
u_t<-1 \ \ \text{and} \ \ \bar{v}_r = v^i x_i >0 \ ,
\end{equation}
where $u_t = -W (\alpha - \beta_i v^i) $ is
the first lower component of the fluid 4-velocity
with the lapse $\alpha$, the shift $\beta^i$, and the Lorentz factor $W$.
and $x^i = (x,y,z)$.
Equation~\eqref{eq:unbound} assumes that the fluid elements follow geodesics
and requires that the orbit is unbound and has an outward pointing velocity,
cf.~also~\cite{East:2012ww}. Other ways of estimating the ejecta mass
can be found in e.g.~\cite{Kastaun:2014fna,Radice:2016dwd}.
The total ejecta mass is given by
\begin{equation}
M_\text{ej} =
\int_{\mathcal{U}} \text{d}^3 x \ q^{(D)} \ ,
\end{equation}
where the integral is computed on the region,
\begin{equation}
\mathcal{U}=\{ x^i=(x,y,z)\, : \, u_t<-1 \ \ \text{and} \ \ \bar{v}_r >0 \} \ ,
\end{equation}
on which material is unbound according to Eq.~\eqref{eq:unbound}.
The kinetic energy of the ejecta can be approximated as the difference
between the total energy $E_\text{ejecta}$ (excluding gravitational
potential energy), the rest-mass, and the total internal energy
$U_\text{ejecta}$~\cite{Hotokezaka:2012ze,Dietrich:2015iva},
\begin{equation}
T_\text{ej} = E_\text{ej} - ( M_\text{ej} + U_\text{ej} ) =
\int_{\mathcal{U}} \text{d}^3 x q^{(D)} (e-1-\epsilon) , \label{eq:Tejecta}
\end{equation}
with $e=\alpha u^t h - p/(\rho \alpha u^t)$.
Additionally, we compute the $D$-weighted integral of $v^2=v_i v^i$ inside
the orbital plane and in the $x$-$z$-plane,
\begin{eqnarray}
\mean{\bar{v}}_{\rho}&=&\sqrt{\frac{\int_{\mathcal{U}_{z=0}} \text{d}^3 x q^{(D)} v^2}
{\int_{\mathcal{U}_{z=0}} \text{d}^3 x q^{(D)} }} , \label{eq:meanbarvrho} \\
\mean{\bar{v}}_{z}&=&\sqrt{\frac{\int_{\mathcal{U}_{y=0}} \text{d}^3 x q^{(D)} v^2}
{\int_{\mathcal{U}_{y=0}} \text{d}^3 x q^{(D)} }} , \label{eq:meanbarvz}
\end{eqnarray}
and the $D$-weighted integrals
\begin{equation}
\mean{v}_{\rho} = \sqrt{ \left( \frac{\int_{\mathcal{U}_{z=0}} \text{d}^3 x q^{(D)} v^{x}}
{\int_{\mathcal{U}_{z=0}} \text{d}^3 x q^{(D)} } \right)^2 +
\left( \frac{\int_{\mathcal{U}_{z=0}} \text{d}^3 x q^{(D)} v^{y}}
{\int_{\mathcal{U}_{z=0}} \text{d}^3 x q^{(D)} }\right)^2 } \label{eq:barvrho}
\end{equation}
and
\begin{eqnarray}
\mean{|v|}_{\rho} & = & \sqrt{ \frac{\int_{\mathcal{U}_{z=0}} \text{d}^3 x q^{(D)} (v^{x})^2}
{\int_{\mathcal{U}_{z=0}} \text{d}^3 x q^{(D)} } +
\frac{\int_{\mathcal{U}_{z=0}} \text{d}^3 x q^{(D)} (v^{y})^2}
{\int_{\mathcal{U}_{z=0}} \text{d}^3 x q^{(D)} } }, \label{eq:meanabsvrho} \\
\mean{|v|}_{z} & = & \sqrt{\frac{\int_{\mathcal{U}_{y=0}} \text{d}^3 x q^{(D)} (v^{z})^2}
{\int_{\mathcal{U}_{y=0}} \text{d}^3 x q^{(D)} }}. \label{eq:meanabsvz}
\end{eqnarray}
Our velocity measurements can be interpreted in the following way:
$\mean{\bar{v}}_{\rho,z}$ is the mean velocity of the ejected material inside the orbital, but not
necessarily in the direction of the cylindrical radius, and perpendicular to it,
$\mean{|v|}_{\rho,z}$ gives the average velocities, and
$\mean{v}_{\rho}$ gives an estimate of the ``kick'' velocity produced by the
ejecta.
Furthermore we consider the entropy ``indicator'',
\begin{equation}
\hat{S} = \frac{p}{K_i \rho^{\Gamma_i}}, \label{eq:entropy}
\end{equation}
where $\Gamma_i$ and $K_i$ are locally determined by the density $\rho$
and the EOS.
In cases where the additional thermal
contribution to the pressure $p_{th}$ is small $\hat{S}\sim 1$,
while in presence of shock heating $\hat{S} \gg 1$.\\
\section{BNS Configurations}
\label{sec:config}
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{fig01.pdf}
\caption{
The four different EOSs employed in this work.
The EOSs are modeled by piecewise-polytropic fits to the tabulated EOSs.
The fits are taken from~\cite{Read:2008iy}, see also Tab.~\ref{tab:EOS}.
Markers correspond to spherical individual stars in our
configurations, and dashed black lines the total masses $M$ of our
configurations for which we investigate all four EOSs. In the
figure, we do not include the total mass of the setup MS1b-094194,
since for those masses only the EOS MS1b was employed.}
\label{fig:EOS}
\end{figure}
Our configurations span the mass ratios $q=1.0,1.25,1.5,1.75$ and
total binary masses of $M=2.5,2.75,2.9167M_\odot$. For each pair
$(q,M)$ we simulate the four EOS ALF2, H4, MS1b, SLy EOS, cf.~Tab.~\ref{tab:EOS}.
Additionally, we consider a $q=2.06$ configuration with the stiff EOS
MS1b, simulated at a single resolution in ~\cite{Dietrich:2015pxa}.
The mass-radius relations of each of the considered EOS are shown in
Fig.~\ref{fig:EOS}, where we also include as markers the isolated star
configurations, and as black dashed lines the total masses $M=M^A+M^B$ of the
systems.
The properties of the initial configuration are listed in
Tab.~\ref{tab:ID_and_grid_irr}. Among various quantities, we report
the tidal polarizability coefficient
\begin{equation}
\kappa^T_2 = 2 \left( \frac{q^4}{(1+q)^5} \frac{k_2^A}{C_A^5} +
\frac{q}{(1+q)^5} \frac{k_2^B}{C_B^5}
\right) \ ,
\label{eq:kappa}
\end{equation}
which describes at leading order the NSs tidal interactions, and depends on
the EOS via the quadrupolar dimensionless Love number $k_2$ of isolated
star configurations, e.g.~\cite{Damour:2009vw}. In
Eq.~\eqref{eq:kappa}, $\mathcal{C}^{A}$ is the compactness of star A
defined as the ratio of the gravitational mass in isolation $M^A$ with
the star's proper radius. It has been shown in
\cite{Bernuzzi:2014kca,Bernuzzi:2015rla} that $\kappa_2^T$ is
the relevant parameter encoding all the EOS information to
characterize the BNS dynamics during both the merger and the
postmerger phases.
In addition, the GW of BNS is almost entirely determined by the
mass ratio $q$ and the $\kappa^T_2$, because the binary total mass
scales trivially in absence of tides and its dependency in the tidal
waveform is hidden in the $\kappa^T_2$. Although other tidal
polarizability parameters, corresponding to higher-than-quadrupole
interactions,
do play a role in the detailed modeling of the
GW~\cite{Bernuzzi:2012ci,Bernuzzi:2014owa,Hinderer:2016eia}, the leading
order $\kappa^T_2$ encodes the main effect and it is the only tidal
parameter measurable in GW searches.
The coverage of the $q$-$\kappa^T_2$ parameter space by
our simulations is shown in Fig.~\ref{fig:parspace}.
All our configurations are simulated with at least two resolutions to
control numerical artifacts and to have (at least rough) estimate of
error bars for \textit{all} the quantities. The specific grid setups
can be found in Tab.~\ref{tab:ID_and_grid_irr}.
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{fig02.pdf}
\caption{BNS parameter space in terms of $q$ and $\kappa^T_2$.
Different colors represent different EOSs, while different markers stand for
different total masses.}
\label{fig:parspace}
\end{figure}
\begin{table}[t]
\centering
\caption{\label{tab:EOS} Properties of the equations of state (EOSs)
used in this work. Our EOSs use a crust with
$\kappa_\text{crust}=\kappa_0=8.94989\times10^{-2}$ and
$\Gamma_\text{crust}=1.35692$. The divisions for the
individual parts are at $\rho_{1}={\rho}_\text{crust} \times 10^{-4}$,
$\rho_{2}=8.11322\times 10^{-4}$ and $\rho_{3} =
1.61880\times 10^{-3}$. The columns
refer to: the name of the EOS, the maximum density in the crust,
the three polytropic exponents $\Gamma$
for the individual pieces, and
the maximum supported gravitational mass $M^\text{max}$, maximum
baryonic mass $M_\text{b}^\text{max}$, and maximum dimensionless compactness
$\mathcal{C}^\text{max}$, respectively, of an isolated
nonrotating star. }
\begin{tabular}{l|cccc|ccc}
\hline EOS & ${\rho}_\text{crust}$& $\Gamma_1$ & $\Gamma_2$ & $\Gamma_3$ & $M^\text{max}$ &
$M_\text{b}^\text{max}$ & $\mathcal{C}^\text{max}$ \\ \hline
SLy & 2.36701 & 3.005 & 2.988 & 2.851 & 2.06 & 2.46 & 0.31 \\
ALF2 & 3.15280 & 4.070 & 2.411 & 1.890 & 1.99 & 2.32 & 0.26 \\
H4 & 1.43709 & 2.909 & 2.246 & 2.144 & 2.03 & 2.33 & 0.26 \\
MS1b & 1.83977 & 3.456 & 3.011 & 1.425 & 2.76 & 3.35 & 0.31 \\ \hline \hline
\end{tabular}
\end{table}
\begin{table*}[t]
\centering
\caption{Configurations and grid setups.
The first column defines the configuration name.
Next 11 columns describe the physical properties: EOS,
gravitational mass of the individual stars $M^{A,B}$,
baryonic mass of the individual stars $M_{b}^{A,B}$,
stars' compactnesses $\mathcal{C}^{A,B}$,
the tidal polarizability coefficient $\kappa^T_2$,
the initial dimensionless GW frequency $M \omega_{22}^0$,
the initial data ADM-Mass $M_{\rm ADM}$ and ADM-angular momentum $J_{\rm ADM}$.
Next 8 columns describe the grid configuration:
finest grid spacing $h_{L-1}$, radial resolution inside the
shells $h_r$, number of points $n$ $(n^{mv})$ in the fix (moving) levels,
radial point number $n_r$ and azimuthal number of points
$n_\theta$ in the shells, inradius $r_1$ up to which GRHD
equations are solved , and the outer boundary $r_b$.
Notice that we divide most configurations in 3 different grid setups R1, R2, R3
(compare the specific simulation name).
}
\begin{Tiny}
\begin{tabular}{c|l|c|ccc|ccc|cccc|cccccccc}
$(q,M)$ & Name & EOS & $M^A$ & $M_b^A$ & $\mathcal{C}^A$ & $M^B$ & $M_b^B$ & $\mathcal{C}^B$ &
$\kappa^T_2$ & $M \omega_{22}^0$ &$ M_{ADM}$ & $J_{ADM}$ &
$h_{L-1}$ & $h_{r}$ & $n$ & $n^{mv}$ & $n_r$ & $n_\theta$ & $r_1$ & $r_b$ \\
\hline
\hline
\multirow{8}{*}{\rotatebox[origin=c]{90}{\textbf{(1.00,2.75)}}}
& ALF2-137137-R1s & ALF2 & 1.375 & 1.518 & 0.164 & 1.375 & 1.518 & 0.164 & 125 & 0.036 & 2.728 & 8.120 & 0.25 & 8.000 & 128 & 64 & 128 & 64 & 572 & 1564 \\
& ALF2-137137-R2s & ALF2 & 1.375 & 1.518 & 0.164 & 1.375 & 1.518 & 0.164 & 125 & 0.036 & 2.728 & 8.120 & 0.16 & 5.333 & 192 & 96 & 196 & 96 & 552 & 1555 \\
& H4-137137-R1s & H4 & 1.375 & 1.499 & 0.150 & 1.375 & 1.499 & 0.150 & 188 & 0.035 & 2.728 & 8.093 & 0.25 & 8.000 & 128 & 72 & 128 & 64 & 572 & 1564 \\
& H4-137137-R2s & H4 & 1.375 & 1.499 & 0.150 & 1.375 & 1.499 & 0.150 & 188 & 0.035 & 2.728 & 8.093 & 0.16 & 5.333 & 192 & 108 & 196 & 96 & 552 & 1555 \\
& MS1b-137137-R1s & MS1b & 1.375 & 1.497 & 0.145 & 1.375 & 1.497 & 0.145 & 262 & 0.035 & 2.729 & 8.158 & 0.25 & 8.000 & 128 & 72 & 128 & 64 & 572 & 1564 \\
& MS1b-137137-R2s & MS1b & 1.375 & 1.497 & 0.145 & 1.375 & 1.497 & 0.145 & 262 & 0.035 & 2.729 & 8.158 & 0.16 & 5.333 & 192 & 108 & 196 & 96 & 552 & 1555 \\
& SLy-137137-R1s & SLy & 1.375 & 1.526 & 0.177 & 1.375 & 1.526 & 0.177 & 65 & 0.036 & 2.728 & 8.006 & 0.23 & 7.392 & 128 & 64 & 128 & 64 & 529 & 1445 \\
& SLy-137137-R2s & SLy & 1.375 & 1.526 & 0.177 & 1.375 & 1.526 & 0.177 & 65 & 0.036 & 2.728 & 8.006 & 0.15 & 4.928 & 192 & 96 & 196 & 96 & 510 & 1437 \\
\hline
\hline
\multirow{8}{*}{\rotatebox[origin=c]{90}{\textbf{(1.25,2.75)}}}
& ALF2-122153-R1s & ALF2 & 1.527 & 1.707 & 0.182 & 1.222 & 1.334 & 0.147 & 127 & 0.036 & 2.728 & 7.956 & 0.25 & 8.000 & 128 & 64 & 128 & 64 & 572 & 1564 \\
& ALF2-122153-R2s & ALF2 & 1.527 & 1.707 & 0.182 & 1.222 & 1.334 & 0.147 & 127 & 0.036 & 2.728 & 7.956 & 0.16 & 5.333 & 192 & 96 & 196 & 96 & 552 & 1555 \\
& H4-122153-R1s & H4 & 1.527 & 1.683 & 0.167 & 1.222 & 1.318 & 0.133 & 193 & 0.035 & 2.729 & 8.025 & 0.25 & 8.000 & 128 & 72 & 128 & 64 & 572 & 1564 \\
& H4-122153-R2s & H4 & 1.527 & 1.683 & 0.167 & 1.222 & 1.318 & 0.133 & 193 & 0.035 & 2.729 & 8.025 & 0.16 & 5.333 & 192 & 108 & 196 & 96 & 552 & 1555 \\
& MS1b-122153-R1s & MS1b & 1.527 & 1.680 & 0.159 & 1.222 & 1.318 & 0.130 & 267 & 0.035 & 2.729 & 8.032 & 0.25 & 8.000 & 128 & 72 & 128 & 64 & 572 & 1564 \\
& MS1b-122153-R2s & MS1b & 1.527 & 1.680 & 0.159 & 1.222 & 1.318 & 0.130 & 267 & 0.035 & 2.729 & 8.032 & 0.16 & 5.333 & 192 & 108 & 196 & 96 & 552 & 1555 \\
& SLy-122153-R1s & SLy & 1.527 & 1.719 & 0.198 & 1.222 & 1.338 & 0.157 & 69 & 0.036 & 2.728 & 7.934 & 0.23 & 7.392 & 128 & 64 & 128 & 64 & 529 & 1445 \\
& SLy-122153-R2s & SLy & 1.527 & 1.719 & 0.198 & 1.222 & 1.338 & 0.157 & 69 & 0.036 & 2.728 & 7.934 & 0.15 & 4.928 & 192 & 96 & 192 & 96 & 510 & 1437 \\
\hline
\hline
\multirow{8}{*}{\rotatebox[origin=c]{90}{\textbf{(1.50,2.5)}}}
& ALF2-100150-R1s & ALF2 & 1.500 & 1.672 & 0.178 & 1.000 & 1.073 & 0.123 & 223 & 0.031 & 2.482 & 6.637 & 0.25 & 8.000 & 128 & 64 & 128 & 64 & 572 & 1564 \\
& ALF2-100150-R2s & ALF2 & 1.500 & 1.672 & 0.178 & 1.000 & 1.073 & 0.123 & 223 & 0.031 & 2.482 & 6.637 & 0.16 & 5.333 & 192 & 96 & 196 & 96 & 552 & 1555 \\
& H4-100150-R1s & H4 & 1.500 & 1.649 & 0.164 & 1.000 & 1.063 & 0.110 & 363 & 0.030 & 2.283 & 6.664 & 0.25 & 8.000 & 128 & 72 & 128 & 64 & 572 & 1564 \\
& H4-100150-R2s & H4 & 1.500 & 1.649 & 0.164 & 1.000 & 1.063 & 0.110 & 363 & 0.030 & 2.283 & 6.664 & 0.16 & 5.333 & 192 & 108 & 196 & 96 & 552 & 1555 \\
& MS1b-100150-R1s & MS1b & 1.500 & 1.647 & 0.156 & 1.000 & 1.063 & 0.109 & 460 & 0.030 & 2.483 & 6.657 & 0.25 & 8.000 & 128 & 72 & 128 & 64 & 572 & 1564 \\
& MS1b-100150-R2s & MS1b & 1.500 & 1.647 & 0.156 & 1.000 & 1.063 & 0.109 & 460 & 0.030 & 2.483 & 6.657 & 0.16 & 5.333 & 192 & 108 & 196 & 96 & 552 & 1555 \\
& SLy-100150-R1b & SLy & 1.500 & 1.672 & 0.193 & 1.000 & 1.074 & 0.129 & 138 & 0.031 & 2.482 & 6.587 & 0.23 & 7.392 & 160 & 64 & - & - & 1190 & 1190 \\
& SLy-100150-R2b & SLy & 1.500 & 1.672 & 0.193 & 1.000 & 1.074 & 0.129 & 138 & 0.031 & 2.482 & 6.587 & 0.15 & 4.928 & 240 & 96 & - & - & 1188 & 1188 \\
\hline
\multirow{8}{*}{\rotatebox[origin=c]{90}{\textbf{(1.50,2.75)}}}
& ALF2-110165-R1b & ALF2 & 1.650 & 1.862 & 0.197 & 1.100 & 1.190 & 0.134 & 133 & 0.036 & 2.729 & 7.686 & 0.25 & 8.000 & 160 & 64 & - & - &1288 & 1288 \\
& ALF2-110165-R2b & ALF2 & 1.650 & 1.862 & 0.197 & 1.100 & 1.190 & 0.134 & 133 & 0.036 & 2.729 & 7.686 & 0.16 & 5.333 & 240 & 96 & - & - &1285 & 1285 \\
& H4-110165-R1s & H4 & 1.650 & 1.835 & 0.182 & 1.100 & 1.177 & 0.121 & 209 & 0.035 & 2.729 & 7.820 & 0.25 & 8.000 & 128 & 72 & 128 & 64 & 572 & 1564 \\
& H4-110165-R2s & H4 & 1.650 & 1.835 & 0.182 & 1.100 & 1.177 & 0.121 & 209 & 0.035 & 2.729 & 7.820 & 0.16 & 5.333 & 192 & 108 & 196 & 96 & 552 & 1555 \\
& MS1b-110165-R1s & MS1b & 1.650 & 1.830 & 0.171 & 1.100 & 1.177 & 0.118 & 282 & 0.035 & 2.729 & 7.799 & 0.25 & 8.000 & 128 & 72 & 128 & 64 & 572 & 1564 \\
& MS1b-110165-R2s & MS1b & 1.650 & 1.830 & 0.171 & 1.100 & 1.177 & 0.118 & 282 & 0.035 & 2.729 & 7.799 & 0.16 & 5.333 & 192 & 108 & 196 & 96 & 552 & 1555 \\
& SLy-110165-R1b & SLy & 1.650 & 1.878 & 0.215 & 1.098 & 1.190 & 0.142 & 78 & 0.036 & 2.727 & 7.700 & 0.23 & 7.392 & 160 & 64 & - & - & 1190 & 1190 \\
& SLy-110165-R2b & SLy & 1.650 & 1.878 & 0.215 & 1.098 & 1.190 & 0.142 & 78 & 0.036 & 2.727 & 7.700 & 0.15 & 4.928 & 240 & 96 & - & - & 1188 & 1188 \\
\hline
\multirow{6}{*}{\rotatebox[origin=c]{90}{\textbf{(1.50,2.92)}}}
& ALF2-117175-R1b & ALF2 & 1.750 & 1.992 & 0.210 & 0.167 & 0.127 & 0.141 & 96 & 0.035 & 2.895 & 8.792 & 0.25 & 8.000 & 160 & 64 & - & - & 1288 & 1288 \\
& ALF2-117175-R2b & ALF2 & 1.750 & 1.992 & 0.210 & 0.167 & 0.127 & 0.141 & 96 & 0.035 & 2.895 & 8.792 & 0.16 & 5.333 & 240 & 96 & - & - & 1285 & 1285 \\
& H4-117175-R1s & H4 & 1.750 & 1.961 & 0.195 & 1.167 & 1.253 & 0.128 & 146 & 0.038 & 2.894 & 8.612 & 0.25 & 8.000 & 128 & 64 & 128 & 64 & 572 & 1564\\
& H4-117175-R2s & H4 & 1.750 & 1.961 & 0.195 & 1.167 & 1.253 & 0.128 & 146 & 0.038 & 2.894 & 8.612 & 0.16 & 5.333 & 192 & 108 & 196 & 96 & 552 & 1555 \\
& MS1b-117175-R1s & MS1b & 1.750 & 1.954 & 0.180 & 1.167 & 1.253 & 0.125 & 206 & 0.038 & 2.894 & 8.612 & 0.25 & 8.000 & 128 & 72 & 128 & 64 & 572 & 1564 \\
& MS1b-117175-R2s & MS1b & 1.750 & 1.954 & 0.180 & 1.167 & 1.253 & 0.125 & 206 & 0.038 & 2.894 & 8.612 & 0.16 & 5.333 & 192 & 108 & 196 & 96 & 552 & 1555 \\
& SLy-117175-R1b & SLy & 1.750 & 2.012 & 0.230 & 1.167 & 1.272 & 0.150 & 53 & 0.033 & 2.893 & 8.850 & 0.23 & 7.392 & 160 & 64 & - & - & 1190 & 1190 \\
& SLy-117175-R2b & SLy & 1.750 & 2.012 & 0.230 & 1.167 & 1.272 & 0.150 & 53 & 0.033 & 2.893 & 8.850 & 0.15 & 4.928 & 240 & 96 & - & - & 1188 & 1188 \\
\hline
\hline
\multirow{6}{*}{\rotatebox[origin=c]{90}{\textbf{(1.75,2.75)}}}
& ALF2-100175-R1 & ALF2 & 1.750 & 1.992 & 0.210 & 1.000 & 1.074 & 0.123 & 140 & 0.032 & 2.731 & 7.674 & 0.25 & 8.000 & 160 & 64 & - & - & 1288 & 1288 \\
& ALF2-100175-R2 & ALF2 & 1.750 & 1.992 & 0.210 & 1.000 & 1.074 & 0.123 & 140 & 0.032 & 2.731 & 7.674 & 0.16 & 5.333 & 240 & 96 & - & - & 1285 & 1285 \\
& H4-100175-R1 & H4 & 1.750 & 1.961 & 0.195 & 1.000 & 1.063 & 0.110 & 230 & 0.035 & 2.730 & 7.531 & 0.25 & 8.000 & 128 & 64 & 128 & 64 & 572 & 1564 \\
& H4-100175-R2 & H4 & 1.750 & 1.961 & 0.195 & 1.000 & 1.063 & 0.110 & 230 & 0.035 & 2.730 & 7.531 & 0.16 & 5.333 & 192 & 108 & 196 & 96 & 552 & 1555 \\
& MS1b-100175-R1 & MS1b & 1.750 & 1.954 & 0.180 & 1.000 & 1.063 & 0.109 & 300 & 0.035 & 2.730 & 7.531 & 0.25 & 8.000 & 128 & 72 & 128 & 64 & 572 & 1564\\
& MS1b-100175-R2 & MS1b & 1.750 & 1.954 & 0.180 & 1.000 & 1.063 & 0.109 & 300 & 0.035 & 2.730 & 7.531 & 0.16 & 5.333 & 192 & 108 & 196 & 96 & 552 & 1555 \\
& SLy-100175-R1b & SLy & 1.750 & 2.012 & 0.230 & 1.000 & 1.075 & 0.129 & 88 & 0.031 & 2.735 & 7.756 & 0.23 & 7.392 & 160 & 64 & - & - & 1190 & 1190 \\
& SLy-100175-R2b & SLy & 1.750 & 2.012 & 0.230 & 1.000 & 1.075 & 0.129 & 88 & 0.031 & 2.735 & 7.756 & 0.15 & 4.928 & 240 & 96 & - & - & 1188 & 1188 \\
\hline
\multirow{3}{*}{\rotatebox[origin=c]{90}{\textbf{(2.06,\ }}}
\multirow{3}{*}{\rotatebox[origin=c]{90}{\textbf{\ 2.88)}}}
& MS1b-094194-R1 & MS1b & 1.944 & 2.200 & 0.199 & 0.944 & 1.000 & 0.103 & 250 & 0.036 & 2.868 & 7.850 & 0.250 & 8.000 & 128 & 72 & 144 & 64 & 572 & 1692 \\
& MS1b-094194-R2 & MS1b & 1.944 & 2.200 & 0.199 & 0.944 & 1.000 & 0.103 & 250 & 0.036 & 2.868 & 7.850 & 0.167 & 5.333 & 192 & 108 & 216 & 96 & 552 & 1683 \\
& MS1b-094194-R3 & MS1b & 1.944 & 2.200 & 0.199 & 0.944 & 1.000
& 0.103 & 253 & 0.036 & 2.868 & 7.850 & 0.125 & 4.000 & 256 & 144
& 288 & 128 & 542 & 1678 \\
\hline
\hline
\end{tabular}
\end{Tiny}
\label{tab:ID_and_grid_irr}
\end{table*}
\section{Dynamics}
\label{sec:dynamics}
\subsection{Mass transfer}
\begin{figure}[t]
\includegraphics[width=0.45\textwidth]{fig03a.pdf}
\includegraphics[width=0.45\textwidth]{fig03b.pdf}
\caption{Mass-transfer between the two neutron stars.
Top: The panels show he time evolution of the baryonic mass and the
``mass defect'' $|\Delta M| = |M(t)-M(t=5000M_\odot)|$ inside coordinate
spheres of radius $10.5M_\odot$ around the two neutron stars as a function of the simulation time $t$.
While the mass is increasing for the primary star (solid lines) the secondary star ``looses'' mass (dashed lines).
For comparison we rescale everything to the mass at $t=5000M_\odot$, with
this approach only influences at the late stage of the inspiral are taken
into account. Note that the dashed line can be seen as an upper bound for the mass
transfer, since mass is also lost due to ejecta.
Bottom: mass transfer as a function of the shifted time $\tilde{t}$, such that
the moment of merger happens at the same time $\tilde{t}^{\rm mrg}$ for all models.
The gray vertical dashed lines refer to the times shown in Fig.~\ref{fig:2d_rho_MS1b-094194}}
\label{fig:mass_transfer_MS1b-094194}
\label{fig:mass_transfer_MS1b-094194_rescaled}
\end{figure}
\begin{figure*}[t]
\includegraphics[width=\textwidth]{fig04.png}
\caption{Density and entropy on the orbital plane for MS1b-094194 in
the final merger phase.
Left plots:
Density profile inside the orbital plane.
The presented times are: (2562.3,2648.9,2735.5,2822.0)$M$ for the
upper left, upper right, lower left, lower right panel respectively.
Those times correspond to the times marked in Fig.~\ref{fig:mass_transfer_MS1b-094194_rescaled}
as vertical dashed lines.
We color the density from blue to red and the unbound density from brown to green.
During the inspiral mass is ejected from the tidal tail of the less massive star due to torque.
The upper right panel corresponds to the moment of merger.
Right plots: Entropy indicator $\hat{S}$ inside the orbital plane for
MS1b-094194. The
times are the same as in the right panels. We also include white contour
lines for densities $\rho=10^{-6},10^{-5},10^{-4}$ and black lines for the
ejecta material with $\rho_u =
10^{-12},10^{-11},10^{-10},10^{-9},10^{-8},10^{-7},10^{-6},10^{-5}.$
During the simulation we observe that the entropy between the two neutron
stars is small, i.e.~no shock heating produced ejecta can be observed
for this setup at the chosen times.}
\label{fig:2d_rho_MS1b-094194}
\label{fig:2d_entropy_MS1b-094194}
\end{figure*}
\paragraph*{MS1b-094194:}
We reported in~\cite{Dietrich:2015pxa} for MS1b-094194-R1s that a rest mass
up to $\approx10^{-2}M_\odot$ was transfered between the stars during the
last orbits before the actual merger.
However, our previous investigations suffered from two facts,
which we already mentioned in~\cite{Dietrich:2015pxa}:
(i) we used a relatively low resolution (quite similar to the R1 setup);
(ii) the infrastructure to compute baryonic masses inside coordinate spheres around the
NSs was not implemented and the mass transfer was estimated by investigating mass changes
across Cartesian refinement boxes.
Here, we can investigate the mass transfer in more
detail using higher resolutions. We find that the mass transfer is not robust for
varying resolution, and the amount of mass decreases at higher resolutions.
We compute the baryonic mass around the NSs
inside coordinate spheres of $r_{c1}=r_{c2}=10.5M_\odot$. While for the
more massive star a mass increase is directly related to the accretion of
material, a mass loss of the less massive star is also produced by
material ejected prior to merger.
Figure~\ref{fig:mass_transfer_MS1b-094194} summarizes our findings. The
upper panel shows the baryonic mass inside $r_{c1/2}$ rescaled to the
value at $t\approx5000M_\odot\approx1731M$. We decided not to rescale the
mass with respect to the initial data to study only mass changes within a
few revolutions prior to merger to allow an easier investigation. The
middle panel shows the mass difference in solar masses with respect
to the mass at $t\approx 1731M$.
The solid lines indicate the mass gain of the
more massive star, while the dashes lines show the mass loss of the secondary NS.
For increasing resolution the transfered mass decreases, while
for R1s we observe a mass difference of $\sim 10^{-2}$, the
transfered mass decreases up to a factor of 10 for R3s.
Due to different numerical dissipation for different resolutions
the moment of merger, i.e.~the peak in the GW amplitude, differs.
For a more simplified and straight forward comparison of the different configurations,
we compensate this effect by applying a time rescaling according to $\tilde{t}=\eta t$,
where the factor $\eta$ is the quotient $t^{\rm mrg}_{R3}/t^{\rm mrg}_{RX}$ with $X={1,2}$.
$t^{\rm mrg}$ denotes the moment of merger, see also \cite{Hotokezaka:2015xka}.
Thus the merger happens for all simulations at the same time
$\tilde{t}^{\rm mrg}$
and we can easily compare the transfered mass for different resolutions.
To give an impression about the density profile during the simulation,
we show the rest mass density inside the orbital plane
at the times
$\tilde{t}=2562.3M (7400M_\odot),
2648.9M (7650M_\odot)$,
$2735.5M (7900M_\odot),
2822.0M (8150M_\odot)$
in Fig.~\ref{fig:2d_rho_MS1b-094194}.
Those times are also marked inside Fig.~\ref{fig:mass_transfer_MS1b-094194_rescaled} (bottom panel)
as vertical dashed lines. The bottom panel of Fig.~\ref{fig:mass_transfer_MS1b-094194_rescaled}
shows the mass gain of the more massive star over the time $\tilde{t}$.
In total, we estimate the transfered mass to be smaller than $5 \times 10^{-3}M_\odot$.
Furthermore, while the lower resolution simulation suggested that the mass transfer sets in several
revolutions before the merger, mass is only transfered $\sim 2$ revolutions before merger for the R3 setup.
In addition to the amount of transfered mass it is also interesting whether
the mass transfer could heat up the more massive NS.
To investigate this, we are showing the entropy indicator $\hat{S}$, cf.~Eq.\eqref{eq:entropy},
together with contour density lines (white dashed)
in Fig.~\ref{fig:2d_entropy_MS1b-094194}.
According to the entropy indicator $\hat{S}$ no shocks are produced at
the surface of the more massive star.
Consequently this region of the NSs does not become significantly hotter
than other parts. \\
\paragraph*{Other configurations:}
Our discussion about the mass transfer is based on MS1b-094194,
but similar results can be obtained for other configurations.
Most notably, we can verify that with increasing the resolution from R1 to R2
the transfered mass is decreasing. Our simulations are also in agreement with the naive understanding
that for stiff EOSs and higher mass ratios the transfered mass is increasing.
Motivated by this observation,
we conclude that the mass transfer estimated from the setup MS1b-094194
can be seen as an upper bound for astrophysical realistic systems.
Because of this finding together with the fact that no shocks are produced due to mass
accretion to the massive neutron star, we do not expect that for astrophysical realistic systems
the energy released by the mass-transfer might lead to EM counterparts which
could be observed before the actual merger of the BNS system.
\subsection{Ejecta}
The mass ejection in our simulations is caused by two effects.
i) Part of the unbound mass is expelled either during the (partial) tidal
disruption or from the tidal tail
of the companion by a centrifugal effect,
see e.g.~Fig.~\ref{fig:2d_rho_MS1b-094194}. These ejecta are emitted
already during the last orbits in an essentially adiabatic way (small
entropy value). As shown in simulations that includes microphysics
e.g. \cite{Sekiguchi:2015dma,Lehner:2016lxy}, the composition is rich in neutrons and
has a small electron fraction.
ii) Part of the unbound mass is expelled when the two NS cores
collide. These
shock-triggered ejecta are characterized by large entropy values and
a higher electron fraction.
The two components are clearly distinguishable by plotting the entropy
indicator $\hat{S}$, Eq.~\eqref{eq:entropy}. Mechanism i) is dominant
for configurations with large $q$.
In addition to the amount of ejecta, we also compute the kinetic energy, as well as the
average velocity, cf.~Sec.~\ref{sec:diagno}.
The results are summarized in Tab.~\ref{tab:ejecta}. In the following
we will present two exemplary cases before discussing the general influence of
the mass ratio and total mass. We often assign an uncertainty as given by the
difference between different resolutions. It is however important to
notice that this does not necessarily corresponds to the total
uncertainty since systematic errors are significant
\cite{Dietrich:2015iva}. Most notably, when the fluid expands it is
possible that the density falls below the artificial atmosphere
value producing mass losses in ejecta. Additionally material can also
leave the grid, see Appendix~\ref{sec:accuracy}.
These are the reasons why the estimated ejecta quantities are evaluated at
a time, where most of the ejected material is rather close to the center
of the system, at a distance of $\approx 150$--$250M_\odot$. While the
ejected mass does not depend much on this fact, the momenta and velocities
of the ejecta are typically overestimated due to fact that the gravitational
potential of merger remnant is neglected in Eqs.~\eqref{eq:Tejecta}--\eqref{eq:meanabsvz}.
A simple Newtonian estimate shows that the ejecta velocities and the
linear momenta are therefore overestimated by up to
$20$--$30\%$.
\\
\paragraph*{MS1b-094194:}
Figure~\ref{fig:2d_rho_MS1b-094194} (left panels) presents the 2D density profile of
MS1b-095194 ($\rho$ from blue to red) and the unbound density ($\rho_u$
from brown to green). The largest amount of ejecta comes from
the tidal tail of the less massive NS, i.e. mechanism i) described above.
We also present snapshots of
the entropy indicator $\hat{S}$ in Fig.~\ref{fig:2d_entropy_MS1b-094194} (right panels).
The density is marked as white thin, dashed contour lines and the
ejecta by black contours. From the plots it is clear that no shock
heating happens inside the tidal tail.
The low entropy value suggests that for such large mass ratios
and stiff EOS, the ejected material will be neutron rich as found in e.g.~\cite{Lehner:2016lxy}.\\
\begin{figure}[htpb]
\includegraphics[width=0.42\textwidth]{fig05.png}
\caption{Entropy indicator $\hat{S}$ for SLy-137137 in the early
postmerger phase.
The panels represent
snapshots at $t=2708M, t=2714, 2721M$ from top to bottom in the $x$-$z$-plane.
The last panel shows the entropy inside the orbital plane at $t=2721M$.
We also include white contour lines for densities $\rho=10^{-6},10^{-5},10^{-4}$ and black lines
for the ejecta material with $\rho_u =
10^{-12},10^{-11},10^{-10},10^{-9},10^{-8},10^{-7},10^{-6},10^{-5}$.}
\label{fig:2d_entropy_SLy137137}
\end{figure}
\paragraph*{SLy-137137:}
As an opposite scenario, we present the results for SLy-137137, an
equal mass NSs described by a soft
EOS. Figure~\ref{fig:2d_entropy_SLy137137} shows $\hat{S}$ for
SLy-137137-R1 inside the $x$-$z$-plane for $t=2708M, 2714M, 2721 M$
and inside the orbital plane ($x$-$y$-plane) at $t=2721M$.
The colors and contour lines are identical to the ones used in
Fig.~\ref{fig:2d_entropy_MS1b-094194}. The first three panels show an
ejection of material, colored by black contour lines in an angle $\gtrsim
45^\circ$ to the orbital plane. The ejection is triggered from high entropic
regions near the surface of the NS (white dashed lines).
Also, we find that after the collision of the cores the central region of the
HMNS in the orbital plane is low entropic; this suggests that no hot
core forms during the merger, but that instead the hot material is
confined in streams of matter expanding from the NS-NS interface and
at the HMNS surface (see also \cite{Bernuzzi:2015opx}).
Type ii) ejecta are triggered from these regions. \\
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{fig06.pdf}
\caption{Ejected properties as a function of the mass ratio for masses $M=2.75$ and $M=2.888$.
We show the difference between two adjacent resolutions as an error bar in our plot.
Top: Ejecta mass, where we see that for an increasing mass ratio stiffer EOS produce significantly more ejecta,
while a similar effect can not be observed for softer EOS.
Upper middle: Kinetic energy of the ejecta, where for an increasing mass ratio ejecta have significantly more kinetic energy, in particular for stiff EOSs.
Lower middle: $\mean{|v_\rho|}$, Eq.~\eqref{eq:meanabsvrho}, as a function of q.
Bottom: $\mean{|v_z|}$, Eq.~\eqref{eq:meanabsvz}, as a function of q.
While the velocity inside the orbital plane seems to be almost constant independent of the mass ratio,
the velocity orthogonal to the orbital plane decreases for higher mass ratios.}
\label{fig:M_ej(q)}
\end{figure}
\paragraph*{Effect of the mass-ratio:}
Figure~\ref{fig:M_ej(q)} (top panel) shows the dependence of the mass of the
ejected material for different mass ratios. We mainly focus on
simulations with a total mass of $M=2.75M_\odot$, but also include the
MS1b-094194, which has a total mass of $M=2.888M_\odot$ to support our argumentation.
Different EOSs are represented by different colors and
the uncertainty is given as an error bar based on the
difference between different resolutions solely (cf. error discussion above).
We find that for stiff EOSs more mass is ejected for larger mass
ratios. This is in-line with previous studies \cite{Dietrich:2015iva,Hotokezaka:2012ze},
although the mass ratios considered here are significantly larger.
The trend in $q$ is almost linear regarding MS1b (red circles), H4
EOS (green circles), ALF2 (orange circles).
A rather similar dependence on the mass ratio was already found for BHNS scenarios,
e.g.~\cite{Foucart:2012nc,Kawaguchi:2016ana}, for
the ejecta and the disk mass, but up to our knowledge, never presented for BNS configurations.
For soft EOSs (SLy) no strong correlation between the mass-ratio and
the ejected mass is visible, all the configurations produce the same
amount of ejecta between
$1$--$2\times 10^{-2} M_\odot$.
Figure~\ref{fig:M_ej(q)} (middle panel) shows the dependence of the
kinetic energy as a function of the mass ratio for the $M=2.75
M_\odot$ and $M=2.888M_\odot$ configurations. As for the amount of
ejected material it is obvious that for stiff EOSs a higher mass ratio
leads to ejecta with a higher kinetic energy.
Again the relation between $T_{\rm ej}$ and $q$ is almost linear.
Figure~\ref{fig:M_ej(q)} (bottom panel) shows the estimated velocities of the ejecta inside the orbital plane and orthogonal
to the orbital plane. We find that the velocity of the ejecta inside
the orbital plane are $\mean{|v_\rho|}\sim 0.17 \sim 50000{\rm km/s}$
for all configurations independent of the EOS and the mass ratio.
However, we stress again that the ejecta velocities are typically overestimated with our current approach.
The velocity orthogonal to the orbital plane is in general smaller than $\mean{|v_\rho|}$. Furthermore,
we find that, independent of the EOS and above $q=1.25$, $\mean{|v_z|}$ decreases for an increasing mass ratio.
This suggests that for large mass ratio configurations more ejection happens
inside the orbital plane than in other directions. The effect is
expected since ejecta due to shock heating play a subdominant role in
those cases. Furthermore this leads to a more oblate and less isotropic
shape of the ejecta.\\
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{fig07.pdf}
\caption{Ejected mass as a function of the total mass of the binary system for a mass ratio
of $q=1.5$. We show the difference between two adjacent resolutions as an error bar.
The influence of the total mass is smaller than the effect of the mass ratio, however,
less ejecta are produced for higher total masses in most cases.}
\label{fig:M_ej(q=1.5)}
\end{figure}
\paragraph*{Effect of the total mass:}
Finally, we study the influence of the total mass on the amount of ejected material.
For this purpose we compare all our configurations with a mass ratio of $q=1.5$.
We have evolved these configurations for total masses of $M=2.5M_\odot$, $M=2.75M_\odot$, and $M\approx2.92M_\odot$.
The data show that the influence of the total mass is smaller than the influence
of the mass ratio, see Fig.~\ref{fig:M_ej(q=1.5)}. However, for
increasing total mass, less massive ejecta are produced in agreement
with previous studies, e.g.~\cite{Hotokezaka:2012ze}.
The reason for this is that larger mass systems form more compact and more massive merger remnants
and the material is more bound by the larger gravitational potential.
The dependence of the kinetic energy on the total mass can be read off
from Tab.~\ref{tab:ejecta}. Similarly to the ejecta mass, the kinetic energy is decreasing for
higher total masses, although the uncertainties are even larger than for the mass of the ejecta.
For the velocities of the ejecta no strict statement about the influence of the total mass can be made
according to our simulations because of the large uncertainties.
\begin{table*}[t]
\centering
\caption{
Ejecta properties. The columns refer to:
the name of the configuration, the mass of the ejecta, the kinetic energy of the ejecta,
the $D$-weighted integral $v_\rho$, Eq.~\eqref{eq:barvrho},
the average velocity of the ejecta inside the orbital plane $\mean{|v|}_\rho$
and perpendicular to it $\mean{|v|}_z$, cf.\ Eqs.~\eqref{eq:meanabsvrho} and~\eqref{eq:meanabsvz},
and the average of $v^2$ of fluid elements inside the orbital plane
$\mean{\bar{v}}^\rho$ and perpendicular to it $\mean{\bar{v}}^z$, Eqs.~\eqref{eq:meanbarvrho}
and~\eqref{eq:meanbarvz}.
We present results for R2 resolved runs and give R1 results inside brackets for an error estimate.
(With the exception of MS1b-094194, where we give results for R3 and R2.) }
\begin{tabular}{l|cccccccc}
Name & $M_{\rm ej} \ [10^{-2}M_\odot] $ & $T_{\rm ej}\ [10^{-4}]$ &
$v_\rho $ & $\mean{|v|}_\rho$ & $\mean{|v|}_z$ & $\mean{\bar{v}}^\rho$ & $\mean{\bar{v}}^z$ \\
\hline
\hline
ALF2-137137 & 0.34 (0.20) & 0.76 (0.22) & 0.066 (0.058) & 0.17 (0.12) & 0.10 (0.11) & 0.17 (0.12) & 0.22 (0.15) \\
H4-137137 & 0.34 (0.06) & 0.89 (0.10) & 0.039 (0.080) & 0.19 (0.13) & 0.10 (0.14) & 0.19 (0.13) & 0.23 (0.22) \\
MS1b-137137 & 0.23 (0.55) & 0.22 (0.56) & 0.030 (0.032) & 0.13 (0.11) & 0.06 (0.08) & 0.13 (0.11) & 0.14 (0.14) \\
SLy-137137 & 1.6 (1.3) & 2.7 (2.8) & 0.060 (0.042) & 0.17 (0.19) & 0.10 (0.12) & 0.17 (0.19) & 0.17 (0.20) \\
\hline
\hline
ALF2-122153 & 0.75 (0.97) & 2.2 (2.1) & 0.077 (0.089) & 0.17 (0.09) & 0.12 (0.10) & 0.17 (0.09) & 0.23 (0.17) \\
H4-122153 & 0.66 (0.88) & 1.7 (1.7) & 0.032 (0.002) & 0.18 (0.15) & 0.11 (0.11) & 0.18 (0.16) & 0.22 (0.28) \\
MS1b-122153 & 0.48 (0.58) & 0.92 (0.77) & 0.038 (0.104) & 0.15 (0.14) & 0.11 (0.07) & 0.16 (0.14) & 0.20 (0.21) \\
SLy-122153 & 1.8 (1.4) & 4.7 (3.5) & 0.014 (0.076) & 0.16 (0.16) & 0.11 (0.10) & 0.16 (0.16) & 0.22 (0.21) \\
\hline
\hline
ALF2-100150 & 2.1 (1.8) & 2.7 (1.9) & 0.095 (0.092) & 0.15 (0.14) & 0.07 (0.06) & 0.15 (0.14) & 0.15 (0.15) \\
H4-100150 & 2.7 (2.6) & 4.5 (3.0) & 0.130 (0.084) & 0.17 (0.14) & 0.03 (0.04) & 0.17 (0.15) & 0.16 (0.15) \\
MS1b-100150 & 3.2 (2.8) & 4.4 (2.8) & 0.124 (0.078) & 0.16 (0.14) & 0.03 (0.04) & 0.16 (0.14) & 0.17 (0.13) \\
SLy-100150 & 1.8 (1.5) & 5.1 (1.8) & 0.095 (0.023) & 0.19 (0.14) & 0.12 (0.06) & 0.19 (0.14) & 0.23 (0.14) \\
\hline
\hline
ALF2-110165 & 2.4 (1.5) & 4.2 (2.1) & 0.101 (0.088) & 0.17 (0.15) & 0.07 (0.08) & 0.17 (0.15) & 0.18 (0.16) \\
H4-110165 & 1.7 (2.0) & 2.7 (2.9) & 0.123 (0.105) & 0.17 (0.16) & 0.04 (0.04) & 0.17 (0.16) & 0.18 (0.17) \\
MS1b-110165 & 2.6 (2.5) & 4.1 (3.5) & 0.126 (0.101) & 0.17 (0.16) & 0.04 (0.05) & 0.17 (0.16) & 0.16 (0.17) \\
SLy-110165 & 1.6 (1.0) & 4.3 (2.5) & 0.064 (0.054) & 0.19 (0.17) & 0.11 (0.12) & 0.19 (0.18) & 0.21 (0.24) \\
\hline
\hline
ALF2-117175 & 2.5 (1.8) & 6.0 (4.9) & 0.100 (0.088) & 0.19 (0.16) & 0.06 (0.09) & 0.18 (0.18) & 0.20 (0.20) \\
H4-117175 & 1.4 (1.6) & 2.6 (2.6) & 0.125 (0.105) & 0.18 (0.17) & 0.05 (0.06) & 0.19 (0.17) & 0.19 (0.17) \\
MS1b-117175 & 2.4 (2.5) & 4.3 (3.9) & 0.108 (0.099) & 0.18 (0.16) & 0.05 (0.08) & 0.18 (0.16) & 0.17 (0.19) \\
SLy-117175 & 0.65 (0.60) & 3.1 (1.6) & 0.118 (0.063) & 0.25 (0.14) & 0.11 (0.07) & 0.38 (0.15) & 0.24 (0.20) \\
\hline
\hline
ALF2-100175 & 3.6 (5.0) & 7.1 (9.5) & 0.118 (0.113) & 0.18 (0.18) & 0.03 (0.05) & 0.19 (0.19) & 0.21 (0.17) \\
H4-100175 & 4.0 (4.2) & 7.0 (6.2) & 0.131 (0.106) & 0.17 (0.16) & 0.02 (0.03) & 0.18 (0.17) & 0.29 (0.18) \\
MS1b-100175 & 4.9 (5.5) & 8.5 (9.4) & 0.134 (0.115) & 0.17 (0.17) & 0.03 (0.03) & 0.18 (0.17) & 0.19 (0.18) \\
SLy-100175 & 2.4 (2.8) & 5.0 (4.8) & 0.096 (0.044) & 0.19 (0.16) & 0.03 (0.05) & 0.21 (0.16) & 0.21 (0.22) \\
\hline
\hline
MS1b-094194 & 6.5 (6.8) & 12 (13) & 0.130 (0.126) & 0.18 (0.18) & 0.02 (0.02) & 0.18 (0.18) & 0.17 (0.21) \\
\hline
\hline
\end{tabular}
\label{tab:ejecta}
\end{table*}
\subsection{Merger Remnant}
\label{sec:merger_remnant}
\begin{table*}[t]
\centering
\caption{Properties of the merger remnant.
The columns represent:
(i) the classification of the merger remnant;
(ii) the lifetime in case a HMNS has formed and collapsed during our
simulation (given in multiples of $100 M_\odot$)
or a description as `p.c.' for a prompt collapse
of the remnant;
(iii) the final mass of the BH $M_{BH}$;
(iv) the dimensionless spin of the final BH $j_{BH}$;
(v) the mass of the disk surrounding the BH $M_{disk}$ measured
$250M_\odot$ after BH formation.
We mark runs where the apparent horizon finder did not work reliably
with a * and runs where no BH
has formed for one resolution but for the other with **.}
\begin{tabular}{l|ccccc}
Name & remnant & $\tau \ [100 M_\odot] $& $M_{BH}[M_\odot]$ & $j_{BH}$ & $M_{disk} [M_\odot]$ \\
\hline
\hline
ALF2-137137 & HMNS$\rightarrow$BH & 26 (24) & 2.49 (2.50) & 0.65 (0.65) & 0.20 (0.17) \\
H4-137137 & HMNS$\rightarrow$BH & 42 (39) & 2.44 (*) & 0.59 (*) & 0.23 (0.11) \\
MS1b-137137 & MNS & - & - & - & - \\
SLy-137137 & HMNS$\rightarrow$BH & 77 (36) & 2.42 (2.43) & 0.60 (0.61) & 0.23 (0.21) \\
\hline
\hline
ALF2-122153 & HMNS$\rightarrow$BH & 23 (**) & 2.48 (**) & 0.64 (**) & 0.23 (**) \\
H4-122153 & HMNS & - & - & - & - \\
MS1b-122153 & MNS & - & - & - & - \\
SLy-122153 & HMNS$\rightarrow$BH & 33 (29) & 2.44 (2.45) & 0.59 (0.60) & 0.20 (0.20) \\
\hline
\hline
ALF2-100150 & HMNS & - & - & - & - \\
H4-100150 & HMNS & - & - & - & - \\
MS1b-100150 & MNS & - & - & - & - \\
SLy-100150 & HMNS & - & - & - & - \\
\hline
ALF2-110165 & HMNS$\rightarrow$BH & 21 (42) & 2.48 (2.40) & 0.61 (0.57) & 0.24 (0.30) \\
H4-110165 & HMNS & - & - & - & - \\
MS1b-110165 & MNS & - & - & - & - \\
SLy-110165 & HMNS$\rightarrow$BH & 58 (**) & 2.45 (**) & 0.61 (**) & 0.26 (**) \\
\hline
ALF2-117175 & BH & p.c. & 2.70 (2.67) & 0.71 (0.70) & 0.20 (0.21) \\
H4-117175 & HMNS$\rightarrow$BH & 14 (17) & 2.65 (2.62) & 0.68 (0.65) & 0.26 (0.24) \\
MS1b-117175 & SMNS & - & - & - & - \\
SLy-117175 & BH & p.c. & 2.75 (2.73) & 0.73 (0.72) & 0.14 (0.15) \\
\hline
\hline
ALF2-100175 & HMNS$\rightarrow$BH & 25 (26) & 2.48 (2.42) & 0.61 (0.57) & 0.26 (0.29) \\
H4-100175 & HMNS & - & - & - & - \\
MS1b-100175 & MNS & - & - & - & - \\
SLy-100175 & BH & p.c. & * (*) & * (*) & 0.18 (0.21) \\
\hline
\hline
MS1b-094194 & SMNS & - & - & - & - \\
\hline
\hline
\end{tabular}
\label{tab:remnant}
\end{table*}
Tab.~\ref{tab:remnant} summarizes the merger outcome (second column)
and the properties of the merger remnant.
In most cases ($15$ configurations) the merger remnant is a HMNS, which is temporary stabilized from collapse by
thermal pressure and centrifugal support. The HMNS will collapse
within a dynamical timescale to a BH; this happens within the
simulated time for $9$ configurations.
$3$ configurations undergo a prompt collapse and for $7$ cases the centrifugal support is sufficient
to support a MNS or SMNS.
In the following we discuss the lifetime of the HMNSs and the
properties of the final BH and disk structure.\\
\paragraph*{Lifetime of the merger remnant:}
The collapse time of the merger remnant is very sensitive to numerical
error and grid resolution, e.g.~\cite{Bernuzzi:2013rza}.
Precise numbers are difficult to obtain, and uncertainties can be of the
other of several milliseconds.
From Tab.~\ref{tab:remnant} one observes that a larger total mass of
the system leads to an earlier collapse, as expected by the fact that a larger
rest-mass star is closer to the collapse threshold.
The influence of the EOS is less clear than the effect of the total mass,
however, as outlined in~\cite{Dietrich:2015iva} one can expect that systems with a softer EOSs
collapse earlier than systems with a stiff EOS.
The EOSs SLy, ALF2, H4 EOS support approximately the same maximum mass regarding
single spherical stars, cf.~Fig.~\ref{fig:EOS} and Tab.~\ref{tab:EOS}.
But, the stiffer the EOS, the longer is the lifetime of the HMNS.
Binaries with stiffer EOS are in general less bound at merger than soft EOSs
binaries. This fact has been quantified in \cite{Bernuzzi:2014kca},
that computed universal relations of the binary's binding energy and
angular momentum at formation of the merger remnant as functions of
the parameter $\kappa^T_2$, Eq. \eqref{eq:kappa}. A consequence of
those relations is that stiffer EOS lead to binaries with larger centrifugal support.
In addition to the centrifugal support, the pressure support in the central regions
is larger for stiffer EOS, and the merger remnant is even further stabilized.
Focusing on the effect of the mass-ratio on the lifetime of the
remnant, we observe for several cases that larger $q$ give slightly larger lifetimes.
Regarding, e.g., setups employing the H4 EOS we find that the equal mass setup
collapses to a BH, but the HMNSs formed during the merger of most of the
unequal mass systems
survive until the end of the simulation.
However, for a more conclusive statement we suggest that higher
resolution are needed since for several setups the differences in the lifetime of the
merger remnant caused by different mass ratios
lie within the error bar obtained from different resolutions.
For larger $q$ the merger remnant is typically more non-axisymmetric than
for equal-mass configurations, and thus has a larger angular momentum support.
Also, looking again at Eq. \eqref{eq:kappa}, for fixed EOS ($k_2$) and fixed
masses, larger $q$ correspond to larger $\kappa^T_2$.
Using the relations of \cite{Bernuzzi:2014kca}, this implies indeed
that remnants from binaries with larger $q$ have larger angular
momentum at formation.
\\
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{fig08.pdf}
\caption{Rest mass of the disk surrounding the final BH for the ALF2 configurations with a mass of $M=2.75$ and mass ratios
$q=1.0;1.25,1.5$. Clearly visible is that for an increasing mass-ratio the disk is more massive.
However, ALF2-100175 and ALF-110165 give almost the same disk masses. We have not included
ALF2-100175 in this figure since the simulation after the collapse to the BH is much shorter than
for the other configurations.}
\label{fig:Mdisk}
\end{figure}
\paragraph*{Final black hole and disk:}
For $12$ configurations a black hole has formed during our simulations, and we expect that if
we evolved our configurations for longer the remaining $6$ HMNS would have formed a BH as well.
In cases where a BH forms we also report the BH mass $M_{BH}$, the dimensionless spin $j_{BH}$
of the BH and the mass of the accretion disk $M_{disk}$.
The mass of the BH depends trivially on the mass of the binary systems,
i.e.~more massive systems produce more massive BHs.
Regarding $M_{BH}/M$, we find that for systems with $M=2.75 M_\odot$
at a time $t = t^{\rm mrg} + 250 M_\odot$
$M_{BH}$ is $\sim 88\%$- up to $91\%$ of the total mass.
For systems with a total mass of $M\approx2.92$ at $t=t^{\rm mrg} + 250 M_\odot$
the BH mass is $89\%$--$94\%$ of the total mass.
Interestingly for the irreducible mass $M_{\rm irr}=\sqrt{A_{\rm BH}/(16 \pi)}$ (with $A_{\rm BH}$ being the horizon area)
we find that $M_{\rm irr}/M$ is almost independent of the EOS, mass, and mass ratio and takes a value around
$\sim0.85$ for all setups.
As the final BH mass, also the final dimensionless spin of the BH $j_{BH}$
is larger for more massive systems, see Tab.~\ref{tab:remnant} and lies around $\sim 0.7$
for our systems with a mass of $2.92M_\odot$.
On the other hand, the imprint of the mass-ratio on the
mass and spin of the BH is less clear. We find that the
black hole mass is almost independent of the mass-ratio and that
the dimensionless spin is slightly decreasing for an increasing mass ratio,
cf.~in particular the systems employing the ALF2 EOS.
The decrease in the dimensionless spin is caused by the fact that
outer region and the formed disk has larger angular momentum.
Among the different EOS considered and for fixed mass ratio and
total mass, the ALF2 EOS produces in most cases a faster rotating
black hole than the other configurations.
We find in our simulations that, for configurations undergoing prompt
collapse, the system has no sufficient time to redistribute angular
momentum to outer regions. Thus, the formed disk is less massive.
In other configurations that form a HMNS, larger $q$ give more
massive disks \cite{Shibata:2006nm,Rezzolla:2010fd}.
In Fig.~\ref{fig:Mdisk} we present the disk mass as a function of time
after BH formation for the ALF2 configurations up to $q=1.50$ with
$M=2.75$.
The disk mass increases up $30\%$ comparing $q=1$ and $q=1.75$.
We find that in general disks produced by larger mass ratios
have smaller maximum densities, but a larger radii.
For SLy no monotonic trend is present in our simulations.
\section{Gravitational Waves}
\label{sec:GW}
\begin{figure*}[t]
\includegraphics[width=1\textwidth]{fig09.pdf}
\caption{Real part and amplitude of the (2,2)-multipole $h_{lm}(u)$
of the GWs for different
configurations and the highest resolution.
Different colors correspond to different EOS:
ALF2 is orange, H4 green, MS1b red, and SLy blue. The lower
resolution runs are shown as a black dotted line to understand
the influence of the resolution in our configurations. }
\label{fig:GW_h22}
\end{figure*}
GWs are calculated extracting the curvature invariant
$\Psi_4$ on coordinate spheres and computing projections on the
spin-weighted spherical harmonics for
spin $-2$, ${}^{-2}Y_{lm}$, see e.g.~\cite{Brugmann:2008zz}. The metric multipoles $r
h_{\ell m}$ are reconstructed from the curvature multipoles using the frequency domain
integration of~\cite{Reisswig:2010di}. The initial circular GW frequency
is used as a cutting frequency. The GW strain is then
given by
\begin{equation}
h(t,\theta,\phi) = \sum_{l=2}^{l_\text{max}} \sum_{m=-l}^{l} r\,
h_{lm}(t) {}^{-2}Y_{lm}(\theta,\phi) \ ,
\end{equation}
where we include individual modes up to $l_\text{max}=4$.
All waveforms are plotted against the retarded time,
\begin{equation}
u=t-r_*=t-r_{\rm extr}-2M\ln\left(r_{\rm extr}/2M-1\right),
\end{equation}
where the extraction radius is set to $r_{\rm extr}\sim 1000\,M_\odot$.
As explained in~\cite{Bernuzzi:2016pie} for such an extraction radius the error compared to
extrapolated waveforms is $\lesssim0.5\%$ for the amplitude and
below $0.1$~rad for the phase of the GW.
The GW energy radiated during the simulations is calculated as
\begin{align}
\mathcal{E}_{\rm rad} &= \sum_{l,m}^{l_{\rm max}} E_{l,m} = \label{eq:GW_Erad}
\dfrac{1}{16\pi}\sum_{l,m}^{l_{\rm max}} \int_{0}^t
dt'\left|r\,\dot{h}_{lm}(t')\right|^2 \ ,
\end{align}
with $l_{\rm max}=8$. We also define the $m$-mode contributions as
\begin{equation}
E_m := \sum_{l=m}^{l_{\rm max}} E_{lm} \ .
\end{equation}
The \textit{total} GW energy emitted during the binary history is given by
\begin{equation}
E^{\rm tot}_{\rm rad} = M - M_{\rm ADM}(t=0) + \mathcal{E}_{\rm rad}
\ .
\end{equation}
In addition to the waveforms, we compute spectra and
spectrograms by applying the Fourier transform $\mathcal{F}$ (in fact,
the discrete FFT) and compute the power spectral density (PSD)
of the GWs as in \cite{Bernuzzi:2013rza},
\begin{eqnarray}
\tilde{h}_{lm}(f) & = & |\mathcal{F}[h_{lm}(t)]|, \label{eq:htilde} \\
\tilde{h}(f; \theta,\phi) & = & |\mathcal{F}[h(t; \theta,\phi)]| \ . \label{eq:htilde}
\end{eqnarray}
The spectrogram allows to identify to which part
of the wave [and so of the dynamics] the spectrum peaks correspond to.
They are computed with chunks of width $\Delta t = 400M_\odot$.
Note also that $\tilde{h}$ depends on the inclination $\theta$ and
azimuthal angle $\phi$,
i.e.~the location of the detector with respect to our source.
For most of our analysis we pick $\theta=\pi/4$ and $\phi=0$, i.e.,
the line of sight to the detector is $45^\circ$ above the orbital plane.
But, we will also consider one BNS configuration for which we investigate the influence of
$\theta$ on $\tilde{h}$.
Figure~\ref{fig:GW_h22} shows the real part of the dominant (2,2)
multipole for our configurations, where
the solid colored lines represent the highest resolved runs (R2) and the dashed black
lines the lower resolution (R1). The signal is composed by the
well-known chirp corresponding the inspiral-to-merger transition,
which formally ends at the wave amplitude's peak, followed by the
post-merger emission which corresponds to the HMNS/MNS-phase.
In some simulations the signal also contains the collapse to the BH and
the quasi-normal ringing.
Table~\ref{tab:GWs} reports the number of orbits from the start of the
simulation to the end of the chirp, $u:=u_\text{mrg}$, defined as the
time of the
$h_{22}$ amplitude peak. We also report there the GW frequency at
$u_\text{mrg}$ as $f_\text{mrg}$.
We find that the dimensionless frequency at merger
$M \omega_\text{mrg}$ depends on the EOS and the mass ratio. While stiffer EOSs merge
with a lower frequency, softer EOSs merge at higher frequencies.
Furthermore, higher mass ratios lead to smaller merger frequencies.
This behavior is understood in terms of the leading order tidal
coupling constant $\kappa^T_2$, and can be encoded with high precision
in the quasi-universal relations proposed in \cite{Bernuzzi:2014kca}.
In the following, we discuss spectra, spectrograms, and GW energy focusing
on the post-merger phase. Similar analysis have been reported in
e.g.~\cite{Bernuzzi:2013rza,Clark:2015zxa,Foucart:2015gaa,Rezzolla:2016nxn}.
GW from unequal-masses BNS, in particular, have been computed recently
in e.g. \cite{Hotokezaka:2012ze,Takami:2014tva,Bernuzzi:2015rla,Lehner:2016lxy}.
\begin{table*}[t]
\centering
\begin{small}
\caption{GW quantities. The columns refer to:
the name of the configurations, the number of orbits until merger from the beginning of the simulation estimated as
$N_\text{orb}=\Phi^\text{mrg}/4\pi$ with $\Phi^\text{mrg}$ being the accumulated phase,
the dimensionless merger frequency $M \omega_{\rm mrg}$,
the merger frequency in kHz $f_{\rm mrg}$,
the dominant postmerger frequencies for
extracted from the (2,1),(2,2),(3,3)-mode, and a possible secondary peak in the (2,2)-mode.
Also these frequencies are stated in kHz. In cases where no secondary peak is found, we mark this
simulations with $-$. We abbreviate the prompt collapse of some configurations with p.c..
Results in brackets refer to R1 resolved runs expect for
MS1b-094194, where we show results for R2 in brackets.}
\begin{tabular}{l|cc|cccccc}
Name & $N_\text{orb}$ & $M\omega_{\rm mrg}$ & $f_{\rm mrg}$ & $f_1$ & $f_{2}$ & $f_3$ & $f_{s}$ \\
\hline
\hline
ALF2-137137 & 11.7 (11.2) & 0.144 (0.142) & 1.72 (1.70) &
1.55 (1.46) & 2.80 (2.77) & 4.30 (4.06) & 1.70 (1.70)\\
H4-137137 & 10.9 (10.7) & 0.133 (0.127) & 1.59 (1.52) &
1.27 (1.38) & 2.50 (2.58) & 3.74 (3.84) & 1.61 (1.74) \\
MS1b-137137 & 10.9 (10.7) & 0.121 (0.118) & 1.45 (1.41)&
1.06 (1.06) & 2.14 (2.06) & 3.10 (3.07) & 1.58 (1.57) \\
SLy-137137 & 11.4 (11.2) & 0.167 (0.163) & 2.00 (1.95) &
1.83 (1.76)& 3.66 (3.47) & 5.39 (5.16) & 2.80 (2.56) \\
\hline
\hline
ALF2-122153 & 10.9 (10.4) & 0.133 (0.131) & 1.59 (1.57) &
1.42 (1.44) & 2.72 (2.68) & 4.11 (4.13) & 2.41 (2.43) \\
H4-122153 & 11.1 (10.9) & 0.114 (0.115) & 1.36 (1.38) &
1.28 (1.24) & 2.42 (2.38) & 3.78 (3.70) & 1.75 ( - ) \\
MS1b-122153 & 10.6 (10.2) & 0.110 (1.08) & 1.32 (1.29) &
1.07 (1.08) & 2.02 (2.16) & 3.12 (3.24) & 1.82 (1.80) \\
SLy-122153 & 11.3 (11.3) & 0.143 (0.144) & 1.71 (1.72) &
1.65 (1.66) & 3.30 (3.36) & 5.09 (5.04) & 2.91 (3.00) \\
\hline
\hline
ALF2-100150 & 13.8 (12.9) & 0.106 (0.103) & 1.37 (1.33) &
1.19 (1.20) & 2.44 (2.38) & 3.69 (3.60) & - \\
H4-100150 & 13.4 (12.8) & 0.086 (0.084) & 1.11 (1.09) &
1.15 (1.05) & 2.20 (2.10) & 3.25 (3.05) & - \\
MS1b-100150 & 13.4 (12.8) & 0.084 (0.084) & 1.11 (1.11) &
1.00 (0.95) & 1.95 (1.95) & 2.90 (2.95) & - \\
SLy-100150 & 14.3 (13.6) & 0.114 (0.111) & 1.51 (1.47) &
1.46 (1.43) & 2.96 (2.87) & 4.40 (4.23) & 2.59 (2.54) \\
\hline
ALF2-110165 & 10.2 (9.8) & 0.119 (0.118) & 1.40 (1.39) &
1.45 (1.32) & 2.74 (2.74) & 4.17 (4.06) & 1.89 (1.95) \\
H4-110165 & 11.6 (10.9) & 0.101 (0.098) & 1.19 (1.15) &
1.28 (1.24) & 2.57 (2.43) & 3.83 (3.54) & - \\
MS1b-110165 & 10.4 (10.0) & 0.098 (0.096) & 1.15 (1.13) &
0.99 (1.06) & 1.94 (1.98) & 3.00 (3.02) & - \\
SLy-110165 & 11.6 (11.3) & 0.133 (0.130) & 1.56 (1.53) &
1.76 (1.62) & 3.51 (3.51) & 5.25 (5.04) & 2.31 (-) \\
\hline
ALF2-117175 & 12.4 (12.0) & 0.131 (0.127) & 1.45 (1.41)&
p.c. & p.c. & p.c. & p.c. \\
H4-117175 & 9.9 (9.6) & 0.110 (0.109) & 1.22 (1.21) &
1.38 (1.36) & 2.82 (2.80) & 4.21 (4.05) & 1.95 (1.90) \\
MS1b-117175 & 9.3 (9.0) & 0.109 (0.110) & 1.21 (1.22) &
1.11 (1.04) & 2.09 (1.99) & 3.20 (3.10) & - \\
SLy-117175 & 14.6 (14.2) & 0.148 (0.137) & 1.64 (1.52)&
p.c. & p.c. & p.c. & p.c. \\
\hline
\hline
ALF2-100175 & 14.0 (13.4) & 0.106 (0.109) & 1.27 (1.31)&
p.c. & p.c. & p.c. & p.c. \\
H4-100175 & 11.0 (10.6) & 0.089 (0.090) & 1.07 (1.07) &
1.25 (1.21) & 2.44 (2.50) & 3.77 (3.68) & - \\
MS1b-100175 & 10.4 (10.0) & 0.088 (0.087) & 1.05 (1.04) &
0.96 (0.93) & 1.98 (2.05) & 3.03 (2.99) & - \\
SLy-100175 & 16.6 (15.9) & 0.122 (0.105) & 1.35 (1.25) &
p.c. & p.c. & p.c. & p.c. \\
\hline
\hline
MS1b-094194 & 10.7 (10.5) & 0.087 (0.088) & 0.96 (0.97) &
0.92 (1.01) & 2.14 (2.00) & 3.20 (3.23) & - \\
\hline
\hline
\end{tabular}
\label{tab:GWs}
\end{small}
\end{table*}
\subsection{Spectrograms}
\label{sec:GW_spectra}
\begin{figure*}[t]
\includegraphics[width=1\textwidth]{fig10.png}
\caption{Spectra $\tilde{h}(\theta=\pi/4,\phi=0)$ for all configurations.
Clearly visible is the chirp signal of the inspiral-merger phase
followed by the post-merger signal (if no prompt collapse).
In most of the cases, the spectrogram highlights that several
frequencies, other than the $f_2$, are excited in the postmerger.
Note also that configurations like ALF2-137137, H4-137137, and
SLy-122153, in which a HMNS forms and collapse within dynamical
times, clearly shows that the postmerger spectrum is not
discrete, but rather continuous, and similar to the
inspiral-merger chirp.}
\label{fig:GW_h22_spectra}
\end{figure*}
We present the spectrograms for our configurations in
Fig.~\ref{fig:GW_h22_spectra}.
The color bar goes from red to blue and is given in arbitrary units, since
we are only interested in the frequencies and the relative strength.
Additional information are given in Table~\ref{tab:GWs}.
During the inspiral-merger one observes the typical chirp signal, with
frequency and amplitude increasing monotonically over time.
The end of the chirp, $u=u_\text{mrg}$, is marked by an amplitude
peak followed by a sharp amplitude's cut-off.
The (2,2)-mode is by far the most dominant mode emitting more than 99\%
of the total energy. Thus, higher modes can not be seen when
$\tilde{h}$ is considered, but can be studied looking at
$\tilde{h}_{lm}$ (see below).
The postmerger spectra are mainly characterized by a dominant emission
frequency $f_2$, related to the (2,2)-mode. A prominent secondary peak
in the (2,2) channel is also visible at frequency $f_s<f_2$ for several configurations (see also
Fig.~\ref{fig:hspectra} below). Interpretation of the latter have
been proposed in e.g.~\cite{Takami:2014tva,Rezzolla:2016nxn,Bauswein:2015yca,Clark:2015zxa}
and references therein. Our spectrograms indicate the peak originates
during the very early postmerger phase, right after the GW amplitude peak
that marks the end of the chirping signal (inspiral-merger), see
e.g.~the panel of MS1b-137137 for $u\sim3000M$. During this short
period the waveforms are characterized, in addition to the oscillation
at frequency $f_2$, by an amplitude modulation corresponding to the
fluid's mass axisymmetric mode.
Notably, for larger mass ratios the spectra becomes more complicated
as shown in Fig.~\ref{fig:hspectra}. At fixed EOS and $M$, the peak of
the spectra for larger $q$ has less power and more peaks appear at frequencies
$f_\text{mrg}<f<f_2$. The secondary peak, in particular, at $f_s$ has
maximum power for $q=1$ and progressively disappears for larger values
$q>1$. Correspondingly, we checked that, for large $q$, the fluid's
mass axisymmetric oscillations diminish (see also the wave's amplitude
in Fig.~\ref{fig:GW_h22}).
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{fig11.pdf}
\caption{Effect of mass-ratio on the GW spectra.
Top: $\tilde{h}_{22}$ for MS1b setups for mass ratios $q=1.00;1.25;1.50;1.75$.
Bottom: $\tilde{h}_{22}$ for H4 setups for mass ratios $q=1.00;1.25;1.50;1.75$.
We mark the $f_2$-frequencies as diamonds, the $f_{\rm mrg}$ as triangles and
the secondary peak frequencies $f_s$ as circles.}
\label{fig:hspectra}
\end{figure}
By considering the multipolar waveforms $\tilde{h}_{lm}$, and not only the $(2,2)$ mode,
we observe other modes excited during the postmerger.
These mode frequencies are named $f_1,f_3,...,f_m$ and roughly
correspond to the azimuthal $m=1,2,3,...$ modes of
oscillations of the fluid\footnote{To avoid confusion, we explicitly mention the difference between
our $f_1$ and $f_3$ frequency and the one defined in~\cite{Takami:2014tva}.
However, we decided to stick to the notation used in~\cite{Dietrich:2015iva,Radice:2016gym}, since this
notation is better suited once more than just the dominant $(2,2)$-mode is considered.}.
These frequencies are more robustly extracted in the ``late''
postmerger phase, i.e.~few milliseconds after the waveform amplitude
peak. For a clear interpretation we extract the $f_1$ frequency from the $(2,1)$ and the
$f_3$-frequency from the $(3,3)$ mode, but they are present in all the
GW multipoles unless killed for symmetry reason in the projection
integrals.
As a further illustration of the multipoles effect, we present
$\tilde{h}_{lm}$ for H4-137137-R2 in Fig.~\ref{fig:GW_h22_spectra_H137137}. We have rescaled the
individual contributions for better visibility (we multiplied $\tilde{h}_{21}$
and $\tilde{h}_{33}$ by $10^{2.5}$ and $\tilde{h}_{44}$ by $10^{3.5}$).
In this case it is clear that the (2,2) mode is dominated by the
$f_2$ frequency, the (2,1)-mode by $f_2$ and $f_1$, the (3,3)-mode by $f_2$ and $f_3$, and
for the (4,4)-mode we find peaks at the $f_2$,$f_3$,$f_4$ frequency.
As expected the frequency of the (4,4)-mode during the inspiral-merger is
approximately twice the frequency of the (2,2)-mode.
This harmonicity of the frequencies is also present in the post-merger phase.
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{fig12.png}
\caption{Spectra of individual GW modes for H4-137137-R2 setup. We
present $\tilde{h}_{lm}$ for
$(lm)=(22),(21),(33),(44)$. The amplitudes are rescaled for better
visibility, i.e.\
$\tilde{h}_{21}$ and $\tilde{h}_{33}$ are rescaled by a factor of $10^{2.5}$ and $\tilde{h}_{44}$ by $10^{3.5}$.}
\label{fig:GW_h22_spectra_H137137}
\end{figure}
Let us specify the spectrogram features for each of the different
postmerger scenarios.
\paragraph*{MNS/SMNS waveforms:}
For the configurations employing the MS1b EOS, the merger remnant is a
stable MNS or a SMNS, see lines for the total mass of the systems in Fig.~\ref{fig:EOS}.
The spectrogram is dominated by the frequency $f_2$, but one can clearly
see contributions from other multipoles in the $m=1,3$ and even the
$m=4$ channels. The emission of
energy and angular momentum decreases over time, as
also visible in Fig.~\ref{fig:GW_h22} by a decreasing amplitude.
Characteristic timescales for the GW emission have been identified in
\cite{Bernuzzi:2015opx}.
\paragraph*{HMNS waveforms:}
In cases of a HMNS and, in particular for configurations undergoing
gravitational collapse within dynamical times, the postmerger signal
is shorter and peaks at specific frequencies $f_1,f_2,f_3$ are more
difficult to extract than for MNS/SMNS.
Considering the $f_2$-frequency one clearly observes a ``postmerger-chirp'', i.e.~that the
frequency increases over time up to the formation of the BH, cf.~Fig.~\ref{fig:hspectra}
for $q=1$ H4\footnote{Note that the
wiggles for $f>2500$~Hz are just due to the Fourier transform of the finite-length signal.}.
But contrary to the inspiral the postmerger-chirp is characterized by a decreasing amplitude.
The feature is physically expected from the increase of rotational velocity and compactness of the star over time
\cite{Bernuzzi:2015rla}. We observe it in all our configurations.
We stress that this indicates that spectra are actually \textit{continuous}, and
they can be modeled with discrete frequencies only for cases in which
the remnant lifetime is sufficiently long so that most of the GW
energy is radiated at frequencies close to $f_2$. This timescale is
$\gtrsim20$~ms~\cite{Bernuzzi:2015opx}.
\paragraph*{Prompt Collapse waveforms:}
In cases of a prompt collapse to a BH, the spectra/spectrograms have
cut-off after the chirp with no other signal.
Since no additional refinement levels are added once the BH forms,
the resolution around the puncture is lower than in our BH simulations~\cite{Brugmann:2008zz}.
Thus although the quasi-normal ringing is visible, we can only resolve between five to eight
local maxima of $r\Psi_{4,22}$ after the merger. This is not sufficient to extract
accurately the quasi-normal modes of the newly formed BH.
\subsection{Source sky location}
\begin{figure}[t]
\includegraphics[width=0.48\textwidth]{fig13.png}
\caption{Angular dependence of the GW spectra for H4-110165-R2. We
present $\tilde{h}$ for:
$(\theta=0,\phi=0)$ (top panel),
$(\theta=\pi/4,\phi=0)$ (middle panel), and $(\theta=\pi/2,\phi=0)$ (bottom panel).}
\label{fig:spectra_detector}
\end{figure}
We discuss the influence of the source's sky location on the measured GW spectrum
for the model H4-110165. We pick as fiducial angles
$(\theta=0,\phi=0)$, $(\theta=\pi/4,\phi=0)$, $(\theta=\pi/2,\phi=0)$ and
present the results in Fig.~\ref{fig:spectra_detector}.
The inspiral-merger signal is strongest for $\theta=0$, which can be explained by the
fact that the (2,2)-mode contribution to $h$ is largest for this angle
since ${}^{-2}Y_{22}(\theta)$ has its maximum
for $\theta=0$. The post-merger signal is dominated by the $f_2$-mode,
and no $f_1$ frequency could be detected because
${}^{-2}Y_{21}(0)={}^{-2}Y_{2-1}^{-2}(0)=0$.
Increasing the inclination $\theta$, the contribution from the
$f_2$-frequency decreases, and the detectability of the
$f_1$-frequency increases. Note however that is unlikely the $m=1$
modes will be detected in GW observations \cite{Radice:2016gym}.
\subsection{GW Energy}
\label{sec:GW_energy}
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{fig14.pdf}
\caption{Total GW energy emitted for different EOSs and mass ratios.
For the total energy we also take into account contributions before the beginning of the
simulation ($M-M_{\rm ADM}(t=0)$). Shown as dashed lines are
estimates for the released energy during BBH mergers (see text for details).
The left panel shows the energies at the moment of merger, i.e.~peak of the GW amplitude,
right panels represent data $20$ ms after merger.
We have removed those setups for which the initial linear momentum, Tab.~\ref{tab:initPandecc},
was large and an artificial drift of the center of mass was present.
}
\label{fig:Erad_modes_q_total}
\end{figure}
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{fig15.pdf}
\caption{GW energy emitted by different modes (up to
$l=m=4$) and rescaled by the amount of energy emitted by the (2,2)-mode.
Black lines correspond to $l=2$ contributions, red lines the
$l=3$ contributions, and $l=4$ is represented by blue lines. Different
dashing corresponds to different $m$. We present the configurations
MS1b-137137 (upper left panel), MS1b-122153 (upper right panel),
MS1b-110165 (lower left panel), MS1b-100175 (lower right panel).}
\label{fig:Erad_modes_rescaled}
\end{figure}
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{fig16.pdf}
\caption{GW energy emitted by different modes $E_m= \sum_{l=m}^{8} E_{lm}$.
We show $E_m$ for $m=1;3;4$ rescaled by $E_2$.
Left panels show the energies at the moment of merger, i.e.~peak of the GW amplitude,
right panels represent data $20$ ms after merger.
We have removed those setups for which the initial linear momentum, Tab.~\ref{tab:initPandecc},
was large and an artificial drift of the center of mass was present.}
\label{fig:Erad_modes_q}
\end{figure}
Finally, let us discuss the influence of the mass ratio on the emitted GW energy.
The emitted energy is calculated according to Eq.~\eqref{eq:GW_Erad} and
we compute the contributions for all modes individually.
To the end of the simulation $\sim1$--$5\times 10^{-2} M_\odot$ is
radiated. The largest amount of energy is emitted by the (2,2)-mode,
which emits 2-3 orders of magnitude more energy than other modes.
The total radiated energy $E^{\rm tot}_{\rm rad}$ is
shown in Fig.~\ref{fig:Erad_modes_q_total} at the moment of merger,
i.e.~$u_{\rm mrg}$, (left panel), and $\sim 4000M_\odot$,
i.e.~$20$ms after the merger (right panel).
The thin dashed lines represent the emitted energy for a non-spinning BBH systems.
To estimate the emitted energy at merger, we use the effective-one-body model
of~\cite{Damour:2012ky}. The final total energy is instead computed by
the fitting formula of~\cite{Barausse:2012qz}.
As for BBHs, the BNS emitted energy decreases for an increasing mass ratio.
Since NSs merge at larger separations and are less relativistic than corresponding BBH systems,
the emitted energy at merger is smaller than the corresponding emitted energy for
BBHs~\cite{Bernuzzi:2015opx}.
In the postmerger phase the influence of the mass ratio becomes even more prominent than
during merger. More energy is released for equal mass systems than for
unequal mass setups. We also find an imprint of the EOS,
where in general NSs employing a soft EOS emit
more energy than for a stiff EOS.
Both observations can be explained in terms of the quasi-universal relations:
systems with larger $\kappa^T_2$ produce a more bound merger remant and release
more energy~\cite{Bernuzzi:2014kca}.
We present in Fig.~\ref{fig:Erad_modes_rescaled} the
rescaled energies $E_{lm}/E_{22}$ for simulations with mass $M=2.75 M_\odot$ and the EOS MS1b.
For equal mass configurations the second dominant mode during the inspiral-merger is the (4,4) mode followed by the
(3,2)-mode, see the upper left panel of Fig.~\ref{fig:Erad_modes_rescaled}.
All other modes do not produce a significant contribution to the total
emitted energy during this phase and contribute only $\sim0.1\%$ to the total energy up to the merger.
For unequal mass configurations the important subdominant modes are (3,3), (4,4), (2,1), (3,2) in
descending order\footnote{With the exception of ALF2-100175, SLy-117175,Sly-100175, for which the
(2,1) is particularly large because of center of mass drift due to
the initial residual linear momentum, see
Tab.~\ref{tab:initPandecc}.}.
This is in qualitative accord to the post-Newtonian theory:
the amplitudes of (3,3) and (2,1) modes are nonzero for unequal mass
configurations and they are proportional to the mass ratio at leading order,
e.g.~\cite{Blanchet:2013haa}.
In our simulations we find that for a mass ratio of $q=1.75$ around
$1\%$ of the emitted energy at merger comes from the (3,3) mode.
In the post-merger phase several different modes are excited.
The (2,2)-mode is still dominant. At the end of our simulations the
(3,3) mode is the second strongest followed by (2,1), (4,4), (3,2), (2,0), where
the exact ordering depends on the mass ratio and EOS.
We find that also for $q=1$ the (2,1) and (3,3) mode are non-zero,
cf.~\cite{Paschalidis:2015mla,Radice:2016gym}.
Computing the luminosity $dE_{lm}(t)/dt$ (not shown in the plot) we
find that while $ {\rm d} E_{22} / {\rm d}t$ decreases
over time due to the very efficient emission in this channel
\cite{Bernuzzi:2015opx}, other modes actually increase up to the
collapse time, in particular $ {\rm d} E_{33} / {\rm d}t$ and $ {\rm d}
E_{21} / {\rm d}t$, e.g.~\cite{Radice:2016gym,Lehner:2016wjg}.
In Fig.~\ref{fig:Erad_modes_q} we present the energy released for $E_m$ with $m=1,3,4$ divided by $E_2$.
We see that at the merger (left panels) for an increasing mass ratio more energy is emitted for $m=1,3$
with respect to the total energy. For $m=4$ the energy mode is approximately constant and contributed $\sim0.2\%$
to the total emitted energy.
The clear imprint of the mass-ratio for $m=1,3$ is lost after the merger (right panels).
However, a small trend towards more energy release for unequal mass
ratios for larger mass ratios
is still present for $m=3$.
In addition to the imprint of the mass ratio we find
that during the postmerger phase the amount
of energy emitted in the subdominant modes is much higher compared to the inspiral
independent of the mass ratio. In general up to $\sim 3\%$
of the total released energy can be emitted in the subdominant modes.
\section{Electromagnetic counterparts}
\label{sec:EM}
\begin{table*}[t]
\centering
\caption{Electromagnetic Counterparts.
The columns refer to: the name of the configuration,
the time in which the peak in the near infrared occurs $t_{\rm peak}$,
the corresponding peak luminosity $L_{\rm peak}$,
the temperature at this time $T_{\rm peak}$,
the time of peak in the radio band $t_{\rm peak}^{\rm rad}$,
and the corresponding radio fluence.
As in other tables, we present results for R2 and R3 resolved simulations and
results for the second highest resolution for all configurations
are given in brackets.}
\begin{tabular}{l|ccccc}
Name & $t_{\rm peak}$ & $L_{\rm peak}$ & $T_{\rm peak}$ & $t^{\rm rad}_{\rm peak}$ & $F^{\nu {\rm rad}}_{\rm peak}$ \\
& [days] & [$10^{40}{\rm erg/ s}$] & [$10^3$ K] & [years] & [mJy] \\
\hline
\hline
ALF2-137137 & 2.0 (1.8) & 2.6 (1.9) & 2.5 (2.7) & 6.4 (6.1) & 0.041 (0.007) \\
H4-137137 & 1.9 (0.9) & 2.8 (1.4) & 2.5 (3.3) & 5.9 (3.5) & 0.058 (0.005) \\
MS1b-137137 & 2.0 (3.1) & 1.9 (2.5) & 2.7 (2.4) & 7.3 (10.6) & 0.006 (0.013) \\
SLy-137137 & 4.5 (3.7) & 4.5 (4.6) & 1.9 (2.0) & 10.0 (7.9) & 0.143 (0.203) \\
\hline
\hline
ALF2-122153 & 2.9 (4.2) & 3.7 (2.9) & 2.2 (2.2) & 8.0 (17.6) & 0.139 (0.046) \\
H4-122153 & 2.7 (3.4) & 3.5 (3.5) & 2.2 (2.1) & 7.3 (9.6) & 0.105 (0.074) \\
MS1b-122153 & 2.5 (3.0) & 2.9 (2.8) & 2.3 (2.3) & 7.2 (9.1) & 0.046 (0.026) \\
SLy-122153 & 4.7 (4.2) & 4.7 (4.3) & 1.9 (2.0) & 12.3 (11.3) & 0.237 (0.173) \\
\hline
\hline
ALF2-100150 & 5.4 (5.4) & 4.5 (4.0) & 1.9 (1.9) & 12.6 (13.6) & 0.100 (0.056) \\
H4-100150 & 6.1 (6.5) & 5.1 (4.6) & 1.8 (1.8) & 14.0 (15.7) & 0.187 (0.090) \\
MS1b-100150 & 6.9 (6.9) & 5.1 (4.5) & 1.8 (1.8) & 15.9 (17.2) & 0.152 (0.072) \\
SLy-100150 & 4.5 (4.9) & 5.2 (3.8) & 1.9 (2.0) & 9.9 (13.3) & 0.359 (0.055) \\
\hline
\hline
ALF2-110165 & 5.6 (4.6) & 5.0 (4.1) & 1.8 (2.0) & 12.8 (11.4) & 0.190 (0.083) \\
H4-110165 & 4.8 (5.4) & 4.3 (4.5) & 1.9 (1.9) & 11.8 (12.9) & 0.111 (0.109) \\
MS1b-110165 & 6.1 (6.0) & 5.0 (4.8) & 1.8 (1.8) & 14.1 (14.1) & 0.161 (0.131) \\
SLy-110165 & 4.3 (3.3) & 4.9 (4.0) & 1.9 (2.1) & 9.8 (8.3) & 0.288 (0.159) \\
\hline
\hline
ALF2-117175 & 5.5 (4.9) & 5.3 (4.6) & 1.8 (1.9) & 12.8 (13.0) & 0.317 (0.233) \\
H4-117175 & 4.2 (4.7) & 4.2 (4.3) & 2.0 (1.9) & 10.3 (11.5) & 0.129 (0.108) \\
MS1b-117175 & 5.6 (5.8) & 5.0 (5.0) & 1.8 (1.8) & 12.8 (13.0) & 0.195 (0.170) \\
SLy-117175 & 2.4 (3.0) & 4.1 (2.8) & 2.2 (2.3) & 6.0 (11.5) & 0.340 (0.055) \\
\hline
\hline
ALF2-100175 & 6.9 (8.0) & 5.9 (6.6) & 1.7 (1.6) & 14.7 (16.0) & 0.340 (0.460) \\
H4-100175 & 7.4 (7.7) & 5.9 (5.7) & 1.7 (1.7) & 16.1 (16.8) & 0.293 (0.234) \\
MS1b-100175 & 8.2 (8.7) & 6.3 (6.5) & 1.6 (1.6) & 17.1 (18.0) & 0.361 (0.386) \\
SLy-100175 & 5.5 (6.3) & 5.2 (5.0) & 1.8 (1.8) & 12.5 (15.6) & 0.254 (0.177) \\
\hline
\hline
MS1b-094194 & 9.4 (9.5) & 7.0 (7.2) & 1.5 (1.5) & 18.8 (18.5) & 0.489 (0.581) \\
\hline
\hline
\end{tabular}
\label{tab:EM}
\end{table*}
Since we do not simulate the evolution of the electron fraction and of
the internal composition of the fluid, we rely on simplified models
for estimating the EM luminosity, fluxes and light curves.
We use in particular the analytical model of \cite{Grossman:2013lqa} which
assumes that the diffusion time is less than the dynamical times
for computing the peak luminosity and
temperature. Light-curves are computed with the model described in \cite{Kawaguchi:2016ana},
originally introduced for the merger of BHNS systems.
Radio flares peak fluxes are instead computed following \cite{Nakar:2011cw}.
Simulations including microphysics have been presented
e.g.~in~\cite{Sekiguchi:2016bjd,Lehner:2016lxy,Palenzuela:2015dqa,Radice:2016dwd,Just:2014fka,Wanajo:2014wha,Radice:2016dwd}. Although more detailed
in term of simulated physics, none of these work have explored the effect of large $q$ as we do here.
\subsection{Macronovae}
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{fig17.pdf}
\caption{Peak time $t_{\rm peak}$ (top panel), peak luminosity $L_{\rm peak}$ (middle panel),
and peak temperature $T_{\rm peak}$ (bottom panel) of marcronovae produced
by the BNS mergers considered in this article as a function of the mass ratio.
As before we consider configurations with $M=2.75M_\odot$ and $M=2.888M_\odot$.}
\label{fig:Lpeak(q)}
\end{figure}
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{fig18.pdf}
\caption{Top panel: Predicted lightcurve (absolute bolometric luminosity) for the
predicted macronova produced by the merger of the configurations employing the MS1b EOS.
The lightcurve is produced with the publicly available code of~\cite{Kawaguchi:2016ana}.
Middle and bottom panel: absolute magnitudes for the ugrizJHK-band assuming that the macronova
is produced 200{\rm Mpc}\ away for MS1b-094194 (middle panel) and MS1b-137137 (bottom panel).}
\label{fig:lightcurve}
\end{figure}
Following~\cite{Grossman:2013lqa}, we can estimate the
time $t_{\rm peak}$ at which the peak in the near-infrared occurs,
the bolometric luminosity at this time $L_{\rm peak}$, and the corresponding temperature
$T_{\rm peak}$:
\begin{align}
t_{\rm peak} & = 4.9 \ {\rm d} \nonumber \\
& \times \left( \frac{M_{ej}}{10^{-2} M_\odot} \right)^{\frac{1}{2}}
\left( \frac{\kappa}{10 {\rm cm^2 g^{-1}} } \right)^{\frac{1}{2}}
\left( \frac{v_{\rm ej}}{0.1} \right)^{-\frac{1}{2}} , \label{eq:tpeak} \\
L_{\rm peak}& = 2.5 \ 10^{40} {\rm erg \, s^{-1}} \nonumber \\
& \times \left( \frac{M_{ej}}{10^{-2} M_\odot} \right)^{1-\frac{\alpha}{2}}
\left( \frac{\kappa}{10 {\rm cm^2 g^{-1}} } \right)^{-\frac{\alpha}{2}}
\left( \frac{v_{\rm ej}}{0.1} \right)^{\frac{\alpha}{2}} , \label{eq:Lpeak} \\
T_{\rm peak} & = 2200 {\rm K} \ \nonumber \\
& \times \left( \frac{M_{ej}}{10^{-2} M_\odot} \right)^{-\frac{\alpha}{8}}
\left( \frac{\kappa}{10 {\rm cm^2 g^{-1}} } \right)^{-\frac{\alpha+2}{8}}
\left( \frac{v_{\rm ej}}{0.1} \right)^{\frac{\alpha-2}{8}} . \label{eq:Kpeak}
\end{align}
Ref.~\cite{Grossman:2013lqa} assumes that the energy release due to
the radioactive decay is proportional to $\sim t^{-\alpha}$ with
$\alpha=1.3$. Furthermore, we set the average opacity to $\kappa =
10~{\rm cm^2 g^{-1}}$ as in~\cite{Grossman:2013lqa}.
We summarize the time of the peak, the corresponding luminosity and
temperature for our configurations in Tab.~\ref{tab:EM}.
Results are also shown in Fig.~\ref{fig:Lpeak(q)}.
We find that for an increasing mass ratio the peak time increases from
$\sim2$ days up to $\sim 10$ days. This effect is larger for stiffer EOSs.
Also the peak luminosity increases with an increasing mass
ratio. Except for the SLy setup the peak luminosity scales almost
linear to the mass ratio. Finally (lower panel of
Fig.~\ref{fig:Lpeak(q)}) we see that the peak temperature decreases
from $\sim2500$K to $\sim1500$K when the mass ratio increases, except
for SLy for which the temperature for all mass ratios is around $\sim1700$K.
Light curves are computed following \cite{Kawaguchi:2016ana}.
The model assumes homologous expansion of the ejecta, a gray opacity,
diffusion approximation for the radiation transfer and that the
photons diffuse only from the latitudinal edge.
It was originally developed for the EM radiation
produced during the merger of a BHNS system. The main difference
between the ejecta of BNS and BHNS mergers is that during BHNS mergers
it is unlikely that shocks form. However, we have shown in
Fig.~\ref{fig:2d_entropy_MS1b-094194} that the entropy of the ejected
material for MS1b-094194 is relatively small. Similar statements hold
for other configurations employing stiff EOSs. For this reason we
expect that the model of~\cite{Kawaguchi:2016ana} is
appropriate for the MS1b as well as other stiff configurations and
systems without shocks.
We use the publicly available program of~\cite{Kawaguchi_web}.
Input parameters are taken from Tab.~\ref{tab:ejecta}. The latitude opening angles are estimated
by evaluation of Eq.~(19) and Eq.~(20) of~\cite{Dietrich:2015iva}.
Furthermore, we use as longitudinal opening angle $\pi$(rad),
$\kappa = 10 {\rm cm^2 g^{-1}}$ for the opacity,
$\dot{\epsilon} = 1.58 \times 10^{10}$erg/g/s for the heating rate coefficient,
$\alpha = 1.2$ for the heating rate power and
$\epsilon_{th}=0.5$ for the thermalization
efficiency as in~\cite{Kawaguchi:2016ana,Kawaguchi_web}.
Figure~\ref{fig:lightcurve} presents our results.
We find that the considered configurations will have
a luminosity between $10^{39}-10^{42}$erg/s (upper panel).
Since the luminosity strongly correlates to the mass of the ejecta,
we observe that for an increasing mass ratio the luminosity
increases for more than one order of magnitude for our configurations.
Because of the increasing luminosity also the observed magnitude of the macronova
at a hypothetical distance of $200{\rm Mpc}$ is larger for MS1b-094194 (middle and bottom panel).
But also for MS1b-137137 counterparts~\footnote{Notice that the model of~\cite{Kawaguchi:2016ana}
was tested for ejecta masses above $0.01M_\odot$ and that higher uncertainties and errors
might be present for low mass ejecta.}
are observable for the first days after merger with 8-m
class telescopes assuming that magnitudes between 26 and 27 are
detectable. Contrary, for high mass ratio configurations like
MS1b-094194 macronovae can be observed for several days up to
weeks after the merger.
Thus, the mass ratio has a clear observational imprint on the EM
counterparts.
\subsection{Radio Flares}
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{fig19.pdf}
\caption{Peak time in the radio band $t_{\rm peak}^{\rm rad}$ (top panel) and corresponding
radio fluence $F^{\nu {\rm rad}}_{\rm peak}$ (bottom panel),
as a function of the mass ratio. We consider BNS configurations
with $M=2.75M_\odot$ and $M=2.888M_\odot$.}
\label{fig:Rad2(q)}
\end{figure}
In order to estimate the radio emission, we use the model of~\cite{Nakar:2011cw}
in which the strongest signal is expected at a time
\begin{align}
t_{\rm peak}^{\rm rad} & = 1392 \ {\rm d} \nonumber \\
& \times \left( \frac{T_{\rm ej} }{10^{49}{\rm erg}} \right)^{\frac{1}{3}}
\left( \frac{n_0}{\rm cm^{-3}}\right)^{-\frac{1}{3}}
\left( \frac{v_{\rm ej}}{0.1} \right)^{-\frac{5}{3}}
\end{align}
with a radio fluence of
\begin{align}
{F^\nu}^{\rm rad}_{\rm peak} & = 0.3 \ {\rm mJy} \times
\left( \frac{T_{\rm ej} }{10^{49}{\rm erg}} \right)
\left( \frac{n_0}{\rm cm^{-3}}\right)^{\frac{p+1}{4}}
\left( \frac{\epsilon_B}{\rm 0.1}\right)^{\frac{p+1}{4}} \nonumber \\
& \times \left( \frac{\epsilon_e}{\rm 0.1}\right)^{p-1}
\left( \frac{v_{\rm ej}}{1} \right)^{\frac{5p-7}{2}}
\left( \frac{D}{10^{27} {\rm cm}} \right)^{-2} \nonumber \\
& \times \left( \frac{\nu_{\rm obs}}{1.4 {\rm GHz}} \right)^{-\frac{p-1}{2}}
\end{align}
for an observation frequency $\nu_{\rm obs}$, which is expected to be higher than the
self-absorption and synchrotron peak frequency at a distance $D$.
The parameters $\epsilon_B$ and $\epsilon_e$ determine how efficiently the
energy of the blast wave is transfered to the magnetic field and to
electrons. We set both parameters to $\epsilon_B=\epsilon_e=0.1$ in our analysis as in~\cite{Nakar:2011cw}.
The variable $n_0$ denotes the surrounding particle density and is set to $0.1 {\rm
cm^{-3}}$. Following~\cite{Nakar:2011cw} we additionally set $p=2.3$ and
$\nu_{\rm obs}=1.4 {\rm GHz}$.
We report our results for $t_{\rm peak}^{\rm rad}$ and ${F^\nu}^{\rm
rad}_{\rm peak}$ in Tab.~\ref{tab:EM} and in Fig.~\ref{fig:Rad2(q)}.
As for $t_{\rm peak}$, $t_{\rm peak}^{\rm rad}$ increases with an
increasing mass ratio. The peak time $t_{\rm peak}^{\rm rad}$ is in
the range of a few years up to two decades for very high mass
ratios. The corresponding radio fluence is characterized by large
uncertainties, however, for stiff EOS an almost linear growth for an
increasing mass ratio is observed.
\section{Summary}
\label{sec:conclusion}
In this article we studied the effect of the mass-ratio by a large set
of new numerical relativity simulations spanning, for the first time,
up to values $q\sim2$. Our findings are summarized in what follows. \\
\paragraph*{Mass transfer:}
A resolution study of simulations with $q=2$ showed that mass transfer
during the last orbits is very dependent on the grid
resolution. In particular, by increasing resolution the amount of
transfered mass decreases. We conclude that no significant mass
transfer happens during the merger of BNS, even in cases with large
mass-ratio. \\
\paragraph*{Mass ejection:}
Mass ejection in large $q$ systems is primary due to a centrifugal effect
and originates from the companion's tidal tail (or partial disruption)
during the late inspiral and merger. Ejecta components due to
shock-driven ejecta are only dominant for configurations with a soft
EOS and rather independent of the mass ratio.
We showed, for the first time in the context of BNS, that the
dependence of the ejecta mass and kinetic energy is essentially linear
on the mass-ratio $q$ for stiff EOSs, see Fig.~\ref{fig:M_ej(q)}.
Also the velocity of the ejecta depend significantly on the mass
ratio. In particular, the component perpendicular to the orbital plane
decreases for increasing $q$, since torque becomes the dominant ejecta mechanism.
For large $q$ the ejecta are almost entirely about the orbital plane.
Overall, these ejecta properties for large $q$ resamble those of black
hole - neutron star binaries and lead to characteristic features of
electromagnetic counterparts (see below).
Finally, the total mass of the configuration plays a minor role and is
less important than the EOS or the mass-ratio. \\
\paragraph*{Merger remnant:}
The lifetime of the merger remnant depends strongly on the EOS,
in most cases softer EOSs lead to an earlier collapse. The mass-ratio
is a secondary effect, but larger $q$ lead to delay collapse.
We also showed that in most cases for which a black hole forms the rest
mass of the accretion disk increases. In cases of a prompt collapse
no massive accretion disk forms. \\
\paragraph*{Gravitational Waves:}
Varying the mass ratio leads to quantitative changes to the GW
frequency, and to qualitative changes of the postmerger spectra.
The GW merger frequency is in general largest for equal masses and
decreases for increasing $q$. For MS1b $f_{\rm mrg}$ decreases from
1.45~kHz to 0.09~kHz when $q$ goes from $1$ to $1.75$ respectively.
No significant effect are instead on the postmerger frequency $f_2$.
We believe the latter is due to the limited accuracy the peak
frequency can be extracted, ultimately related to the broad character
of the spectra peak.
However, we find that for unequal masses the characteristic
secondary peaks $f_s$ in the spectrum tend to disappear for large $q$,
and they are actually absent for high mass ratio systems,
see Fig.~\ref{fig:hspectra}.
Our spectrograms highlight the rich structure of the multipolar GW
waveform. Modes with azimuthal number $m=1,3$, in particular, become
progressively more relevant for larger $q$, although it is unlikely
their effect will be observed in next LIGO/Virgo observations,
e.g. \cite{Radice:2016gym}.
Furthermore, for configurations producing a MNS merger remnant, the
spectrograms show that the postmerger frequencies increase over time
in a chirp-like fashion as the merger remnant
becomes more compact. This implies that the postmerger spectrum is
in fact continuous and not discrete as anticipated in
\cite{Bernuzzi:2015rla}.
The total emitted GW energy is a decreasing function of $q$ during
both the inspiral and the post-merger phase.
This qualitative behavior is already known from binary black hole systems, but
here we extend the result to BNS. This is nontrivial since for BNS
tidal interactions play an important role during inspiral-merger and
the postmerger phase has different physics from black hole binaries.
We find that neutron stars with softer EOS emit more energy.
In addition to the total energy, also the mode hierarchy changes by
varying $q$. We find that the energy in the $m=3$ (and $m=1$) emission channel
increase for larger $q$ and contribute up to $1\%$ to the total emitted
energy up to merger. \\
\paragraph*{Electromagnetic counterparts:}
A GW detection of BNSs will trigger observations to capture
follow-up electromagnetic emissions. We used simplified models to
estimate the luminosity,
peak time, and light curves of macronovae counterparts and the peak
time and fluence of radio flares.
We showed that the peak luminosity, peak time, and persistency of these
counterparts are strongly dependent on the mass ratio $q$. Unequal mass
BNS systems are more luminous in the EM, than equal mass systems
because of more massive ejecta. The larger the mass ratio, the more delayed
is the luminosity peak. Also our estimated macronova lightcurves are
more persistent for larger mass ratios; they could be detected up
to a few weeks, see Fig. \ref{fig:lightcurve}.
Similarly to black hole neutron star mergers and differently from
$q\sim1$ BNS, the dynamical ejecta of large $q$ BNS are confined to
the equatorial plane and will not obscure optical emissions from the
disk wind~\cite{Kasen:2014toa}. Thus, the latter might be detectable
for face-on binaries \footnote{We thank D.Radice for pointing this out.}.
\\
Numerical uncertainties have been carefully evaluated on multiple quantities,
see Appendix~\ref{sec:accuracy}, and we are confident on the presented
results. However, our simulations do not account for microphysics as
done in other works
e.g.~\cite{Just:2014fka,Bernuzzi:2015opx,Palenzuela:2015dqa,Lehner:2016wjg,Lehner:2016lxy,Sekiguchi:2016bjd,Radice:2016dwd}.
The simplified assumptions in our work have been necessary to simulated a large number of
BNS configurations, and in order to better control the numerical
errors. We believe our results hold at least at a qualitative level;
dedicated simulations of selected configurations that include a more
sophisticated treatment of microphysics should be performed in the future
to validate our predictions.
|
\section{Introduction}
Radiative captures are reactions during which two nuclei $b$ and $c$ fuse to form a nucleus $a$ by emiting a $\gamma$:
\beq
b+c\rightarrow a+\gamma.
\eeqn{e1}
These reactions, also noted $c(b,\gamma)a$, take place in many astrophysical sites.
For example, as discussed by Mossa during this conference \cite{Mos16}, $\rm d(p,\gamma)^3He$ is one of the reactions of the pp chain which takes place in the Sun and has happened during the Big-Bang nucleosynthesis.
The $s$ and $r$ processes that take place during supernova-explosions consist of sequences of neutron radiative captures $(n,\gamma)$ \cite{BBFH,Ata16}.
To constrain astrophysical models, the corresponding cross sections need to be measured at the relevant energies.
These energies being rather low (of the order of a few tens of keV in stars), the direct measurement of the radiative-capture cross sections can be quite difficult, either because of the Coulomb barrier between the colliding nuclei or because they involve neutrons.
The former hinders the capture by repelling the colliding nuclei from each other, and the seconds are difficult to handle experimentally.
Hence the interest in indirect techniques to infer these cross sections \cite{BBR86,BR96,Mer16}.
Coulomb breakup is one of them \cite{BBR86,BR96}.
In that reaction, the projectile dissociates into lighter fragments through its interaction with a heavy (high $Z$) target $T$:
\beq
a+T\rightarrow b+c+T.
\eeqn{e2}
Assuming the dissociation to be due to the sole Coulomb interaction, the reaction can be described as an exchange of virtual photons between the projectile and the target.
It can thus be seen as the time-reversed reaction of the radiative capture of the fragments, which should enable us to deduce easily the radiative-capture cross section from breakup measurements \cite{BBR86,BR96}.
Using accurate reaction models, various studies have shown that higher-order effects and other reaction artefacts play a significant role in Coulomb breakup, which hinder the simple extraction of radiative-capture cross sections from breakup measurements \cite{EBS05,CB05}.
The case of \ex{15}C is of particular interest to analyse this indirect method as both its Coulomb breakup~\cite{Nak09} and the radiative capture $\rm ^{14}C(n,\gamma)^{15}C$~\cite{Rei08} have been measured accurately.
Summers and Nunes have recently proposed an innovating analysis of the Coulomb-breakup measurement \cite{SN08}.
They have confirmed the significant influence of dynamical effects observed in previous analyses \cite{EBS05,CB05} and, accordingly, the need of an accurate reaction model to study properly such reactions.
Since breakup reactions are mostly peripheral, in the sense that they probe mostly the tail of the projectile wave function \cite{CN07}, they have suggested to use the comparison between their calculations and the measurements to deduce the Asymptotic Normalisation Coefficient (ANC) of the \ex{15}C bound state from experimental data \cite{SN08}.
They then suggest to rely on this inferred ANC to compute the cross section for the $\rm ^{14}C(n,\gamma)^{15}C$ radiative capture.
Their idea leads to inferred radiative-capture cross sections in good agreement with the direct measurements \cite{SN08,Esb09}.
In the present work we analyse the influence of the description of the \ex{14}C-n continuum upon breakup calculations.
In \Sec{C15}, we present the model of \ex{15}C we use in this study, emphasising on the \ex{14}C-n interaction used in the continuum.
We then discuss our results obtained for the Coulomb breakup of \ex{15}C on Pb at $68A$~MeV (\Sec{Cbu}) and present our analysis of the extraction of the cross section for the radiative capture $\rm ^{14}C(n,\gamma)^{15}C$ following the prescription of Summers and Nunes in \Sec{ng}.
We show also how this method can be improved by selecting the data at low energy in the \ex{14}C-n continuum.
We end by the conclusions and perspectives of this work.
\section{Model of \ex{15}C}\label{C15}
We consider a simple two-body model of \ex{15}C, viz. a neutron loosely bound to an inert \ex{14}C core.
The \ex{14}C-n interaction is simulated by a phenomenological Woods-Saxon potential, whose depth is adjusted to reproduce the low-energy spectrum of \ex{15}C.
For simplicity, the spins of the fragments are ignored in this study.
Hence the $1/2^+$ ground state is described as the $1s$ state of the potential while the $5/2^+$ excited bound state is reproduced by the $0d$ state.
To confirm the sensitivity of the breakup calculations to the ANC of the bound states, we consider Woods-Saxon form factors with two different geometries: one with a regular diffuseness ($a=0.6$~fm) leading to a regular ANC, and one with a large diffuseness ($a=1.5$~fm), which leads to a much larger ANC, i.e.\ with a density of probability significantly shifted towards large $^{14}$C-n distances [see \fig{f1} (left)].
In this way, the $^{14}$C-n potential is constrained in the $s$ and $d$ waves, but not in the $p$ wave.
Since the dominant transition, in both the Coulomb breakup of $^{15}$C and the radiative capture $\rm ^{14}C(n,\gamma)^{15}C$, is an E1 transition to (resp. from) the $p$ continuum from (resp. to) the $s$ ground state, the description of the $p$ wave is the one that matters, if the description of the continuum plays any role in those reactions.
To test this hypothesis, we use four different potentials in the $p$ wave in order to vary the phaseshift as much as possible [see \fig{f1} (right)].
The first one is the regular $^{14}$C-n potential used to describe $^{15}$C ground state ($a_p=0.6$~fm, solid line).
The second exhibit the same geometry but has its depth adjusted to set the forbidden $0p$ bound state at the one-neutron separation energy of $^{14}$C ($E_{0p}=-8$~MeV, dashed line).
To obtain a significant change, we also use the very diffuse potential mentioned earlier ($a_p=1.5$~fm, dotted line).
Finally, we also test what happens when the $^{14}$C and the neutron do not interact in the $p$ wave ($V_p=0$, dash-dotted line).
\begin{figure}
\center
\includegraphics[width=7.3cm]{C15foel.ps}
\includegraphics[width=7.3cm]{C15pdep.ps}
\caption{Left: wave functions describing the $^{15}$C $s$ ground state using a regular diffuseness ($a=0.6$~fm, red) and a very diffuse potential ($a=1.6$~fm, blue).
Right: $p$-wave phaseshifts in the $^{14}$C-n continuum obtained with four different potentials (see text for details).}\label{f1}
\end{figure}
\section{Coulomb breakup of \ex{15}C on Pb at $68A$~MeV}\label{Cbu}
Combining the two descriptions of $^{15}$C in its ground state and the four ones of its $p$ continuum, we obtain eight models of the projectile, which we use to compute the Coulomb breakup of \ex{15}C on Pb at $68A$~MeV, corresponding to the RIKEN experiment \cite{Nak09}.
The calculations are performed with the Dynamical Eikonal Approximation, which provides excellent agreement with experiments on one-nucleon halo nuclei \cite{BCG05,GBC06}.
\fig{f3} shows the corresponding $^{15}$C breakup cross sections as a function of the \ex{14}C-n relative energy $E$ after dissociation.
As Summers and Nunes \cite{SN08}, we observe a significant influence of the ground-state ANC.
The cross sections computed with the diffuse state [blue (top) curves] are much larger than those obtained with the regular diffuseness [red (bottom) curves].
However, the ANC is not the only parameter that affects the calculations.
As already observed in \Ref{CN06}, the description of the $p$ continuum significantly influences the value of the breakup cross section.
Interestingly, the order of the curves is the same for both descriptions of the ground state.
The calculations performed setting the $0p$ forbidden bound state at the $^{14}$C one-neutron separation energy (dashed lines) are about 15\% below the calculations performed using the regular $s$-wave potential in the $p$ wave (solid lines).
The latter is very close to the results obtained with no \ex{14}C-n interaction in the continuum, as expected from the fact that both descriptions lead to very similar phaseshifts (see \fig{f1} right).
Using the very diffuse potential in the $p$ wave (dotted lines) leads to a significant change of the shape of the breakup cross section compared to the other calculations.
This is to be related to the weird behaviour of the $p$ phaseshift observed for that potential in \fig{f1}.
These results put at stake the idea of Summers and Nunes.
The significant influence of the \ex{14}C-n continuum observed in \fig{f3} shows that the ANC extracted by fitting breakup calculations to experimental data is spoiled by the description of the continuum.
\begin{figure}
\center
\includegraphics[width=11cm]{C15fullsdE.ps}
\caption{Breakup cross section of \ex{15}C on Pb at $68A$~MeV as a function of the \ex{14}C-n relative energy $E$ after dissociation. The calculations are performed with the \ex{14}C-n potentials described in Sec.~2. Data are from \Ref{Nak09}.}\label{f3}
\end{figure}
\section{Inferring the cross section for the radiative capture $\rm ^{14}C(n,\gamma)^{15}C$}\label{ng}
Following the results presented in the previous section, is the method proposed in \Ref{SN08} to extract the radiative capture cross section from breakup reliable?
To answer this question, we follow the method suggested by Summers and Nunes \cite{SN08} and compare the prediction for the radiative-capture cross section obtained by the fit of our eight theoretical breakup cross sections to the RIKEN data.
This fit consist simply in multiplying the theoretical cross section by a scaling factor and adjust that factor to obtain the best $\chi^2$ with the data.
That factor is then used in the direct radiative capture calculation.
\begin{figure}
\center
\includegraphics[width=7.5cm]{C15scaledsdE.ps}
\includegraphics[width=7.5cm]{C14ngscaled.ps}
\caption{Left: theoretical breakup cross sections scaled to the data of \Ref{Nak09}.
Right: radiative-capture cross sections obtained with our eight descriptions of \ex{15}C scaled with the factors obtained from the fit of the breakup data. Experimental radiative-capture data are from \Ref{Rei08}.}\label{f4}
\end{figure}
The results of this fit are displayed in \fig{f4} (left).
We observe a good agreement between the fitted theoretical cross sections and the data, but for the cases in which the (very unrealistic) diffuse potential is used in the \ex{14}C-n $p$ wave (dotted blue and red lines).
In these two cases no scaling parameter could enable us to reproduce the data with such a narrow peak.
Originally this method is based on the idea that the scaling factor would correct the theoretical ANC to its actual value, and that this value can then be used to predict reliably the radiative-capture cross section.
However, the influence of the continuum description on the calculations tells us that no actual ANC can be extracted in such a way.
Nevertheless, we naively apply this method to our calculations.
The corresponding radiative-capture cross sections are depicted in \fig{f4} (right) as a function of the \ex{14}C-n energy.
They are found in fairly good agreement with the direct data of Reifart \etal \cite{Rei08}.
We observe a larger spread of our theoretical predictions than that obtained by Summers and Nunes \cite{SN08}, and an average value slightly larger than the direct data.
Nevertheless, this good result indicates that most of the dependence of our calculations to the projectile description, in both its bound and scattering states, can be captured in this scaling constant, and that this dependence is very similar for both breakup and radiative-capture processes.
This is mostly due to the fact that both reactions are sensitive to the same physical inputs from the \ex{15}C structure, namely the ANC of its bound state and the phaseshift in its continuum.
Thanks to this, the idea of Summers and Nunes provides reliable radiative-capture cross sections, even when varying the projectile description in the continuum.
However, according to what has been said before, the scaling factor extracted from the $\chi^2$ fit of the the breakup calculation to the experimental data cannot be directly related to the ground-state ANC.
Apart from the sensitivity of the breakup calculations to the projectile continuum, \fig{f4} (left) teaches us that all calculations, even the very unrealistic ones, exhibit the same behaviour at low energy, e.g.\ before the cross sections reach a maximum at about 0.5~MeV.
This energy range is also the range of astrophysical interest: the radiative-capture cross sections are needed at low energy, not at a few MeV in the continuum.
Hence the idea to apply Summers and Nunes' method not on the whole range of the breakup data, but to restrict it at low energy.
The result of such a fit is displayed in \fig{f5}.
The left panel shows that all breakup cross sections, including the very exotic ones, fall on top of each other when fitted below 0.5~MeV.
Using the fitting parameter obtained in such a way, we get radiative-capture cross sections in much better agreement with the direct data, even with unrealistic \ex{14}C-n potentials [see \fig{f5} (right)].
This interesting result confirms the validity of the method proposed by Summers and Nunes and that it enables to capture most of the nuclear-structure inputs that matter in the radiative capture
In addition, it shows that it works better if applied to low \ex{14}C-n continuum energies, which actually matter in astrophysical processes.
\begin{figure}
\center
\includegraphics[width=7.5cm]{C15scaledlowEsdE.ps}
\includegraphics[width=7.5cm]{C14ngscaledlowE.ps}
\caption{Left: theoretical breakup cross sections scaled to the data of \Ref{Nak09} restricted in the low-energy range ($E<0.5$n MeV).
Right: radiative-capture cross sections obtained that low-energy fit.}\label{f5}
\end{figure}
\section{Conclusions and perspectives}
Coulomb breakup has been proposed as an indirect technique to infer radiative-capture cross section at astrophysical energies \cite{BBR86,BR96}.
This idea is based on the hypothesis that Coulomb breakup, corresponding to an exchange of virtual photons between the projectile and the target, can be seen as the time-reversed reaction of the radiative capture.
Unfortunately, the direct extraction of the latter cross section from the former's can only be done if the Coulomb breakup takes place at first order, which we know is not the case \cite{EBS05,CB05}.
To circumvent this issue, Summers and Nunes have proposed a method based on the fact that breakup reactions are mostly peripheral \cite{CN07}.
In this method, an ANC for the projectile bound state is extracted from the fit of accurate breakup calculations to breakup data \cite{SN08,Esb09}.
In this work, we have studied the sensitivity of this method to the description of the projectile continuum, which has been ignored in Summers and Nunes' analysis.
From accurate calculations of the breakup of \ex{15}C on Pb at $68A$~MeV using different descriptions of the projectile, we have shown that the sensitivity to the \ex{14}C-n continuum cannot be overlooked and that the scaling factor extracted from the fit suggested by Summers and Nunes contains information not only about the ANC of the projectile bound state, but also about its continuum.
Nevertheless, the method works fine.
We understand this by the fact that the structure information absorbed in this fitting procedure is important for both processes.
We have observed that this fit does not have much sense for the extreme descriptions of the \ex{14}C-n continuum, as they lead to significant distortions in the breakup cross sections compared to regular potentials.
To account for this, we suggest to restrict the fit suggested by Summers and Nunes to low \ex{14}C-n energies, i.e. those that are of significance for astrophysics.
The radiative-capture cross sections obtained in such a way are in excellent agreement with the direct data.
This variant of Summers and Nunes' idea hence enables to extract reliably radiative-capture cross section from Coulomb breakup data without having to worry about the description of the two-body projectile in both its bound state and its continuum.
This, in a sense, revives the original idea of Baur, Bertulani and Rebel \cite{BBR86}.
It would be interesting to see if this variant can be improved by selecting breakup data at forward angles, where the process is fully dominated by the Coulomb interaction.
Another interesting perspective is to see whether this method can be applied to charged cases, like for \ex{3}He($\alpha$,$\gamma$)\ex{7}Be or \ex{7}Be(p,$\gamma$)\ex{8}B.
\acknowledgments
This work is part of the Belgian Research Initiative on eXotic
nuclei (BriX), program P7/12 on inter-university attraction
poles of the Belgian Federal Science Policy Office.
It was supported in part by the Research Credit No. 19526092 of the Belgian Funds for Scientific Research F.R.S.-FNRS.
|
\section{Introduction}
Over the past few years---starting with ref.~\cite{lam}---a paradigm
has been developed in which the moduli of the entries of the lepton mixing
(Pontecorvo--Maki--Nakagawa--Sakata or PMNS) matrix $U$
are fixed by some discrete,
finite,
non-Abelian group $G$.
In this paradigm,
the (unspecified) family theory of the leptons
is symmetric under $G$.
That group breaks,
through some unspecified mechanism,
into two different subgroups $G_\ell \subset G$ and $G_\nu \subset G$,
which only intersect at the identity element of $G$.
The mass matrices $M_\ell$ and $M_\nu$
of the charged leptons and of the neutrinos,
respectively,
are separately invariant under $G_\ell$ and $G_\nu$,
respectively.
As a consequence,
the unitary matrices that diagonalize $M_\ell$ and $M_\nu$,
named $U_\ell$ and $U_\nu$,
respectively,
also diagonalize the matrices of the restrictions to $G_\ell$ and $G_\nu$,
respectively,
of a three-dimensional representation of $G$.
The diagonalization of $M_\ell$ and of $M_\nu$ is thus replaced
by the diagonalization of the matrices representing
the two subgroups $G_\ell$ and $G_\nu$ in some representation of $G$.
The numerical values of the moduli of the matrix elements
of $U = U_\ell^\dagger U_\nu$ are traced back in this way
to a three-dimensional representation of a finite group $G$.
The irreducible representations of finite groups are finite in number
and may be systematically studied.
That study may be carried out exclusively through
theoretical means~\cite{hagedorn},
but is nowadays greatly facilitated by the free availability
of the computer software {\tt GAP}~\cite{GAP},
which manipulates groups and their representations,
and of the complete library {\tt SmallGroups}
of all the discrete groups of order less than 2\,000~\cite{SG}.
The paradigm mentioned in the previous paragraph
has firstly been developed under the assumption
that the neutrinos are Majorana fields.
In that case,
the group $G_\nu$ must be (isomorphic to)
the Klein group $\mathbbm{Z}_2 \times \mathbbm{Z}_2$,\footnote{We have assumed
in this statement that $G_\nu$ is a subgroup of $SU(3)$.
If $G_\nu$ is a subgroup of $U(3)$ but not of $SU(3)$,
then it should be of the form $\mathbbm{Z}_2 \times \mathbbm{Z}_2 \times \mathbbm{Z}_2$.}
because the neutrino Majorana masses must remain real and positive
under a rephasing of the neutrino fields
and therefore this rephasing can at most be a change of sign.
Systematic searches using {\tt GAP}
were produced under this assumption~\cite{lam2,lim}
and a thorough classification of the PMNS matrices
achievable in this way has been derived~\cite{fonseca}.
The paradigm has been extended
to the cases of Dirac neutrinos~\cite{hagedorn} and of quarks~\cite{quarks};
then,
$G_\nu$ may be a general $\mathbbm{Z}_n$ group with $n > 2$.
An extensive theoretical as well as {\tt SmallGroups} investigation
of those cases has been presented in ref.~\cite{chinese}.
The papers mentioned in the previous paragraph aimed at
fixing the whole matrix $\left| U \right|^2$,
defined as
$\left( \left| U \right|^2 \right)_{ij} \equiv \left| U_{ij} \right|^2$,
through group theory.
In this paper,
following ref.~\cite{lavouraludl},
we have the more modest aim of only fixing either one row
or one column of $\left| U \right|^2$.
This allows the prediction of two out of
the four parameters
of $\left| U \right|^2$.\footnote{Those four parameters
are usually taken to be three mixing (`Euler') angles and one phase,
but they may alternatively be chosen to be
four of the entries of $\left| U \right|^2$~\cite{Branco}.}
This happens when either $G_\ell$ or $G_\nu$,
respectively,
is represented by twice-degenerate matrices,
\textit{i.e.}\ by $3 \times 3$ matrices which have two equal eigenvalues
while the third eigenvalue is different.
For instance,
if the matrices representing $G_\ell$ have two of their eigenvalues equal,
then only one of columns of their diagonalizing matrix $U_\ell$ is well-defined,
\textit{i.e.}\ defined but for an arbitrary overall phase.
Choosing that column to be the third one,
the third row of $U = U_\ell^\dagger U_\nu$ is well-defined
while the first and second rows may be mixed and cannot,
therefore,
be predicted.
In ref.~\cite{lavouraludl} this job of predicting either one row
or one column of $\left| U \right|^2$ was undertaken
under the assumption that the neutrinos are Majorana fields;
it was found that most rows/columns thus found have some zero entries
and are therefore of no practical interest,
since the phenomenology indicates the absence of zeros from the PMNS matrix.
In this paper we shall assume instead that
neutrinos are Dirac fields;\footnote{We thus treat the lepton sector
in exactly the same way as the quark sector,
just as was done in ref.~\cite{chinese}.
We drop any attempt to explain the smallness of the neutrino masses
through the usual see-saw mechanism.
Other versions to that mechanism,
like for instance a see-saw mechanism for the vacuum expectation values
of Higgs doublets~\cite{radovcic}, or a see-saw mechanism
with extra vector-like neutrinos~\cite{valle}, may possibly be employed.}
this allows for a larger variety of groups $G_\ell $ and $G_\nu$---namely,
they will either be cyclic groups of order larger than 2
or possibly groups $\mathbbm{Z}_m \times \mathbbm{Z}_n$---and consequently
to a much larger variety of predictions
for rows/columns of $\left| U \right|^2$.
In section~\ref{theory} we expose the theory behind our group search.
In section~\ref{searches}
we explain how we have used {\tt GAP} to perform the search.
Section~\ref{results} is devoted to the presentation of the rows/columns
that resulted from the search.
Section~\ref{comparison} compares our presumptive rows/columns
to the actual phenomenological values of $\left| U \right|^2$,
checking which rows/columns are realistic.
In section~\ref{conclusions} we make a short summary
of our work.
Appendix~\ref{groups} is devoted to some definitions in group theory
and may be skipped by an uninterested reader.
We clarify that this paper reports on a pure computational search
made by using {\tt GAP}.
We have made neither any attempt at studying analytically
a particular group or set of groups,
nor at explaining analytically
the results found through our computational search.
For comparison,
ref.~\cite{talbert} gives
other papers relying strongly on the power of the {\tt GAP} software.
\section{Theory} \label{theory}
We work in the context of the three-generation Standard Model
with the addition of three right-handed neutrinos.
The neutrinos are assumed to be standard Dirac fields;
they have no Majorana mass terms.
The charged-lepton mass matrix $M_\ell$
and the neutrino mass matrix $M_\nu$ are defined through the mass terms
\begin{equation}
\mathcal{L}_\mathrm{lepton\ mass} =
- \overline{\ell_L} M_\ell \ell_R - \overline{\nu_L} M_\nu \nu_R
+ \mathrm{H.c.}
\end{equation}
The matrices $H_\ell \equiv M_\ell M_\ell^\dagger$
and $H_\nu \equiv M_\nu M_\nu^\dagger$ are diagonalized as
\begin{subequations}
\begin{eqnarray}
U_\ell^\dagger H_\ell U_\ell &=&
\mbox{diag} \left( m_e^2,\ m_\mu^2,\ m_\tau^2 \right),
\\
U_\nu^\dagger H_\nu U_\nu &=&
\mbox{diag} \left( m_1^2,\ m_2^2,\ m_3^2 \right),
\end{eqnarray}
\end{subequations}
where $U_\ell$ and $U_\nu$ are $3 \times 3$ unitary matrices.
The PMNS matrix is $U = U_\ell^\dagger U_\nu$.
We assume that the matrices $H_\ell$ and $H_\nu$ are invariant
under the action of two invertible matrices $T_\ell$ and $T_\nu$,
respectively.
This invariance is defined through
\begin{subequations}
\begin{eqnarray}
T_\ell H_\ell T_\ell^{-1} &=& H_\ell, \label{biuytp}
\\
T_\nu H_\nu T_\nu^{-1} &=& H_\nu. \label{buiyo}
\end{eqnarray}
\end{subequations}
Equation~(\ref{biuytp}) states that $T_\ell$ and $H_\ell$ commute,
therefore they are simultaneously diagonalizable.
Similarly,
equation~(\ref{buiyo}) implies that
$T_\nu$ and $H_\nu$ are simultaneously diagonalizable.
Therefore,
$U_\ell$ diagonalizes $T_\ell$ and $U_\nu$ diagonalizes $T_\nu$:
\begin{subequations}
\label{buisp}
\begin{eqnarray}
U_\ell^\dagger T_\ell U_\ell &=& \hat T_\ell
\ \, \equiv\ \, \mathrm{diag} \left( l_1,\ l_2,\ l_3 \right), \label{bxihp}
\\
U_\nu^\dagger T_\nu U_\nu &=& \hat T_\nu
\ \, \equiv \ \, \mathrm{diag} \left( n_1,\ n_2,\ n_3 \right).
\end{eqnarray}
\end{subequations}
We now make the crucial assumption that
$T_\ell$ and $T_\nu$ together generate a
(three-dimensional representation of a)
group which is \emph{finite}.
This assumption allows one to restrict the PMNS matrix,
which in general belongs to the continuous set of unitary matrices,
to a discrete set of matrices defined by the theory of finite groups,
thereby generating some predictive power.
This predictive power is further enhanced if we assume that
the group generated by $T_\ell$ and $T_\nu$ is \emph{small},
\textit{i.e.}\ that its order is smaller than some arbitrary number.
In our practical search
we have assumed that the order of $G$ is smaller than 2\,000,
since this is the present reach of the {\tt SmallGroups} library.
\subsection{Main search}
In our main search we have assumed that two,
and only two,
of the three eigenvalues $l_{1,2,3}$ of $T_\ell$ are equal,
while the eigenvalues $n_{1,2,3}$ of $T_\nu$ are all distinct.
Suppose for instance that $l_1 = l_2 \neq l_3$.
Then,
the third column of $U_\ell$,
which according to equation~\eqref{bxihp}
is the normalized eigenvector of $T_\ell$ corresponding to the eigenvalue $l_3$,
is well-defined---only its overall phase is arbitrary---but the first
two columns of $U_\ell$ are not,
because they are eigenvectors corresponding to the same eigenvalue $l_1 = l_2$
and may therefore be arbitrarily mixed between themselves.
As a consequence,
the third row of $U = U_\ell^\dagger U_\nu$ will be fixed
except for its phase,
while the first two rows will remain arbitrary
(they will only be restricted to being orthogonal to the third row
and to each other).
Our assumption thus allows us to `predict' the moduli of the matrix elements
of the third row of $U$.
In the same way,
if $\ell_1 = \ell_3 \neq \ell_2$,
then the second row of $U$ is predicted;
if $\ell_2 = \ell_3 \neq \ell_1$,
then the first row of $U$ is predicted.
In practice,
we compute those moduli in the following way.
Let $p$ and $q$ be two integers,
then,
from equations~\eqref{buisp},\footnote{The use of traces
like those in equation~\eqref{uiyhp}
has been first advocated in ref.~\cite{Branco}.}
\begin{eqnarray}
\mathrm{tr} \left( T_\ell^p T_\nu^q \right) &=&
\mathrm{tr} \left( U_\ell \hat T_\ell^p U_\ell^\dagger
U_\nu \hat T_\nu^q U_\nu^\dagger \right)
=
\mathrm{tr} \left( U^\dagger \hat T_\ell^p U \hat T_\nu^q \right)
=
\sum_{i=1}^3 \sum_{j=1}^3 l_i^p n_j^q \left| U_{ij} \right|^2
\nonumber \\ &=&
l_1^p n_1^q \left( 1 - \left| U_{21} \right|^2 - \left| U_{31} \right|^2 \right)
+ l_1^p n_2^q \left( 1 - \left| U_{22} \right|^2 - \left| U_{32} \right|^2 \right)
\nonumber \\ & &
+ l_1^p n_3^q \left( \left| U_{21} \right|^2 + \left| U_{31} \right|^2
+ \left| U_{22} \right|^2 + \left| U_{32} \right|^2 - 1 \right)
+ l_2^p n_1^q \left| U_{21} \right|^2
\nonumber \\ & &
+ l_2^p n_2^q \left| U_{22} \right|^2
+ l_2^p n_3^q \left( 1 - \left| U_{21} \right|^2 - \left| U_{22} \right|^2 \right)
+ l_3^p n_1^q \left| U_{31} \right|^2
\nonumber \\ & &
+ l_3^p n_2^q \left| U_{32} \right|^2
+ l_3^p n_3^q \left( 1 - \left| U_{31} \right|^2 - \left| U_{32} \right|^2 \right)
\nonumber \\ &=&
\left( l_2^p - l_1^p \right) \left( n_1^q - n_3^q \right)
\left| U_{21} \right|^2
+ \left( l_2^p - l_1^p \right) \left( n_2^q - n_3^q \right)
\left| U_{22} \right|^2
\nonumber \\ & &
+ \left( l_3^p - l_1^p \right) \left( n_1^q - n_3^q \right)
\left| U_{31} \right|^2
+ \left( l_3^p - l_1^p \right) \left( n_2^q - n_3^q \right)
\left| U_{32} \right|^2
\nonumber \\ & &
+ l_1^p n_1^q
+ l_1^p n_2^q
- l_1^p n_3^q
+ l_2^p n_3^q
+ l_3^p n_3^q.
\label{uiyhp}
\end{eqnarray}
Next,
using our extra assumption that $l_1 = l_2 \neq l_3$,
\begin{eqnarray}
\mathrm{tr} \left( T_\ell^p T_\nu^q \right) &=&
\left( l_3^p - l_1^p \right) \left( n_1^q - n_3^q \right)
\left| U_{31} \right|^2
+ \left( l_3^p - l_1^p \right) \left( n_2^q - n_3^q \right)
\left| U_{32} \right|^2
\nonumber \\ & &
+ l_1^p n_1^q
+ l_1^p n_2^q
+ l_3^p n_3^q.
\label{vyurp}
\end{eqnarray}
Writing both $\mathrm{tr} \left( T_\ell T_\nu \right)$
and $\mathrm{tr} \left( T_\ell T_\nu^2 \right)$ as in equation~\eqref{vyurp}
(\textit{i.e.}\ with $p= 1$ and $q = 1$ and $2$, respectively),
one obtains two equations for $\left| U_{31} \right|^2$
and $\left| U_{32} \right|^2$.
They are solved to yield
\begin{subequations}
\label{sdihp}
\begin{eqnarray}
\left| U_{31} \right|^2 &=& \frac{\mathrm{tr} \left( T_\ell T_\nu^2 \right)
+ l_1 \chi - \left( n_2 + n_3 \right) \mathrm{tr} \left( T_\ell T_\nu \right)
- l_1 n_1^2 + l_3 n_2 n_3}
{\left( l_3 - l_1 \right) \left( n_1 - n_2 \right)
\left( n_1 - n_3 \right)},
\label{uigdp} \\
\left| U_{32} \right|^2 &=& \frac{\mathrm{tr} \left( T_\ell T_\nu^2 \right)
+ l_1 \chi - \left( n_1 + n_3 \right) \mathrm{tr} \left( T_\ell T_\nu \right)
- l_1 n_2^2 + l_3 n_1 n_3}
{\left( l_3 - l_1 \right) \left( n_2 - n_1 \right)
\left( n_2 - n_3 \right)},
\label{bihpi} \\
\left| U_{33} \right|^2 &=& \frac{\mathrm{tr} \left( T_\ell T_\nu^2 \right)
+ l_1 \chi - \left( n_1 + n_2 \right) \mathrm{tr} \left( T_\ell T_\nu \right)
- l_1 n_3^2 + l_3 n_1 n_2}
{\left( l_3 - l_1 \right) \left( n_3 - n_1 \right)
\left( n_3 - n_2 \right)},
\label{buigf}
\end{eqnarray}
\end{subequations}
where $\chi \equiv n_1 n_2 + n_1 n_3 + n_2 n_3$.
(We have used $\left| U_{33} \right|^2 = 1 - \left| U_{31} \right|^2
- \left| U_{32} \right|^2$ to derive equation~\eqref{buigf}
from equations~\eqref{uigdp} and~\eqref{bihpi}.)
Equations~\eqref{sdihp} allow us to compute
the third row of $\left| U \right|^2$
from the matrices $T_\ell$ and $T_\nu$ and from their eigenvalues,
without having to explicitly diagonalize those matrices.
We emphasize that our use of traces $\mbox{tr} \left( T_\ell^p T_\nu^q \right)$
constitutes an important technical progress
over other searches using {\tt GAP},
because {\tt GAP} frequently gives the matrices
of a group representation in non-unitary form;
by using the traces one may directly use those matrices in that form,
without having firstly to unitarize them
and then to compute their eigenvectors;
one thus saves a considerable amount of computer time.
\subsection{Secondary search}
We have also considered the possibility that
there are two matrices $T_{\nu 1}$ and $T_{\nu 2}$
which commute both with each other and with the Hermitian matrix $H_\nu$;
therefore,
they are all diagonalized by the same unitary matrix $U_\nu$:
\begin{subequations}
\begin{eqnarray}
U_\nu^\dagger T_{\nu 1} U_\nu &=& \mbox{diag} \left( n_1,\ n_2,\ n_3 \right),
\\
U_\nu^\dagger T_{\nu 2} U_\nu &=&
\mbox{diag} \left( \bar n_1,\ \bar n_2,\ \bar n_3 \right).
\end{eqnarray}
\end{subequations}
We have considered the situation in which
both $T_{\nu 1}$ and $T_{\nu 2}$ have two identical eigenvalues,
but together they act as if all their eigenvalues are different:
$n_1 = n_2 \neq n_3$ and $\bar n_1 = \bar n_3 \neq \bar n_2$.
We identify this property by computing the invariant quantity
\begin{eqnarray}
q &\equiv& 3 \left[ \mathrm{tr} \left( T_{\nu 1} T_{\nu 2} \right) \right]^2
- 2 \left( \mathrm{tr}\, T_{\nu 1} \right) \left( \mathrm{tr}\, T_{\nu 2} \right)
\mathrm{tr} \left( T_{\nu 1} T_{\nu 2} \right)
\nonumber \\ & &
+ \left( \mathrm{tr}\, T_{\nu 1} \right)^2 \mathrm{tr} \left( T_{\nu 2}^2 \right)
+ \mathrm{tr} \left( T_{\nu 1}^2 \right) \left( \mathrm{tr}\, T_{\nu 2} \right)^2
- 3\, \mathrm{tr} \left( T_{\nu 1}^2 \right) \mathrm{tr} \left( T_{\nu 2}^2 \right).
\label{huigp}
\end{eqnarray}
It is easy to check that
$q = - \left( n_1 - n_3 \right)^2
\left( \bar n_1 - \bar n_2 \right)^2 \neq 0$
if $n_1 = n_2$ and $\bar n_1 = \bar n_3$,
while
$q = 0$ if $n_1 = n_2$ and $\bar n_1 = \bar n_2$.
So,
we select $q \neq 0$.
Now,
with
\begin{subequations}
\begin{eqnarray}
U_\ell^\dagger T_\ell U_\ell &=& \mbox{diag} \left( l_1,\ l_1,\ l_3 \right),
\\
U_\nu^\dagger T_{\nu 1} U_\nu &=& \mbox{diag} \left( n_1,\ n_1,\ n_3 \right),
\\
U_\nu^\dagger T_{\nu 2} U_\nu &=& \mbox{diag} \left( \bar n_1,\ \bar n_2,\ \bar n_1
\right),
\end{eqnarray}
\end{subequations}
and $U = U_\ell^\dagger U_\nu$ as before,
one easily finds that
\begin{subequations}
\label{huoiy}
\begin{eqnarray}
\left| U_{33} \right|^2 &=& \frac{\mathrm{tr} \left( T_\ell T_{\nu 1} \right)
- l_1 n_1 - l_1 n_3 - l_3 n_1}{\left( l_1 - l_3 \right)
\left( n_1 - n_3 \right)},
\\
\left| U_{32} \right|^2 &=& \frac{\mathrm{tr} \left( T_\ell T_{\nu 2} \right)
- l_1 \bar n_1 - l_1 \bar n_2 - l_3 \bar n_1}{\left( l_1 - l_3 \right)
\left( \bar n_1 - \bar n_2 \right)}.
\end{eqnarray}
\end{subequations}
So,
in this case one can once again derive the entries
in the third row of $\left| U \right|^2$ by using invariant traces.
\section{The search} \label{searches}
We have scanned all the finite groups with order less than $2\,000$
by making use of the library {\tt SmallGroups}.
We have identified which of them
have three-dimensional faithful irreducible representations (`irreps') and,
moreover,
are \emph{not}\/ the direct product of a group $\mathbbm{Z}_n$ with $n \ge 2$
by some other group.\footnote{The cyclic group $\mathbbm{Z}_n$
is formed by the $n$-roots of unity
under the standard multiplication of complex numbers.
It is of course an Abelian group,
because the multiplication of complex numbers is commutative.}
We only need to scan for \emph{faithful}\/ representations;
if a representation is unfaithful
(\textit{i.e.}\ if it represents several elements of the group $G$
by the same matrix),
then it is the faithful representation of a subgroup $G^\prime \subset G$
and we will find it when we scan for the faithful representations of $G^\prime$.
Moreover,
we only need to scan for \emph{irreducible}\/ representations:
if a representation is reducible,
then all its matrices may simultaneously be rotated to a basis
in which they are all block-diagonal;
hence,
$T_\ell$ and $T_\nu$ will be block-diagonal in some basis,
and therefore $U_\ell$, $U_\nu$,
and the PMNS matrix will also be block-diagonal;
but this is in contradiction with the phenomenology,
which indicates that $U$ has no zero matrix elements.
Finally,
we shed groups $G$ of the form $G = G^\prime \times \mathbbm{Z}_n$
with $n \ge 2$,
because a faithful representation of $G^\prime \times \mathbbm{Z}_n$
is necessarily also a faithful representation of the smaller group $G^\prime$
and we will find it when we study $G^\prime$.
Since the number of groups of order 1\,536 is much too large
for all of them to be scanned within a reasonable time,
we have used the conjecture in ref.~\cite{chinese}
that both nilpotent groups\footnote{One can check
whether the group with {\tt SmallGroups} identifier $\left[ m, n \right]$
($m$ and $n$ are integers;
$m$ is the order of the group)
is nilpotent by first typing the {\tt GAP} command {\tt G:=SmallGroup([m,n])}
and then the {\tt GAP} command {\tt IsNilpotentGroup(G)}.
The latter command produces the answer {\tt True}
if the group $\left[ m, n \right]$ is nilpotent.}
and groups with a normal Sylow 3-subgroup\footnote{The Sylow 3-subgroups
of a group $G$ may be found by typing the {\tt GAP} command
{\tt SylowSubgroup(G,3)}.
The {\tt GAP} command {\tt IsNormal(G,U)} returns {\tt True}
if $U$ is a normal subgroup of $G$.}
never have three-dimensional faithful irreps.\footnote{In appendix~\ref{groups}
we attempt to explain in simple terms
what are nilpotent groups and groups
with a normal Sylow 3-subgroup~\cite{grouptheory}.}
Since 99.97\% of the groups of order 1\,536 are in one of those two categories,
this conjecture has allowed us
to outright disconsider most of the groups of that order.
Tables~\ref{table1},
\ref{table2},
and~\ref{table3} present all the groups $G$ that we have found
\begin{table}[ht!]
\begin{tabular}{|l|}
\hline
\begin{minipage}[t]{0.96\columnwidth}
{\bf [12, 3]}, [21, 1], {\bf [24, 12]}, [27, 3], [27, 4], {\bf [36, 3]},
[39, 1], {\bf [48, 3]}, {\bf [48, 30]}, {\bf [54, 8]}, [57, 1],
{\bf [60, 5]}, [63, 1], {\bf [75, 2]}, [81, 6], {\bf [81, 7]},
{\bf [81, 8]}, {\bf [81, 9]}, {\bf [81, 10]}, [81, 14],
{\bf [84, 11]}, [93, 1], {\bf [96, 64]}, {\bf [96, 65]}, {\bf [108, 3]},
{\bf [108, 11]}, {\bf [108, 15]}, {\bf [108, 19]}, {\bf [108, 21]},
{\bf [108, 22]}, [111, 1], [117, 1], [129, 1], {\bf [144, 3]},
[147, 1], {\bf [147, 5]}, {\bf [150, 5]}, {\bf [156, 14]}, {\bf [162, 10]},
{\bf [162, 12]}, {\bf [162, 14]}, {\bf [162, 44]}, {\bf [168, 42]},
[171, 1], [183, 1], [189, 1],
[189, 4], [189, 5], [189, 7], [189, 8], {\bf [192, 3]}, {\bf [192, 182]},
{\bf [192, 186]}, [201, 1], {\bf [216, 17]}, {\bf [216, 25]},
{\bf [216, 88]}, {\bf [216, 95]}, [219, 1], {\bf [225, 3]}, {\bf [228, 11]},
[237, 1], {\bf [243, 16]}, {\bf [243, 19]}, {\bf [243, 20]}, [243, 24],
{\bf [243, 25]}, {\bf [243, 26]}, {\bf [243, 27]}, [243, 50],
{\bf [243, 55]}, {\bf [252, 11]}, [273, 3], [273, 4], [279, 1], [291, 1],
{\bf [294, 7]}, {\bf [300, 13]}, {\bf [300, 43]}, [309, 1], {\bf [324, 3]},
{\bf [324, 13]}, {\bf [324, 15]}, {\bf [324, 17]}, {\bf [324, 43]},
{\bf [324, 45]}, {\bf [324, 49]},
{\bf [324, 50]}, {\bf [324, 51]}, {\bf [324, 60]},
{\bf [324, 102]}, {\bf [324, 111]}, {\bf [324, 128]},
[327, 1], [333, 1], {\bf [336, 57]}, [351, 1], [351, 4], [351, 5], [351, 7],
[351, 8], {\bf [363, 2]}, {\bf [372, 11]}, [381, 1], {\bf [384, 568]},
{\bf [384, 571]}, {\bf [384, 581]},
[387, 1], [399, 3], [399, 4], [417, 1],
{\bf [432, 3]}, {\bf [432, 33]}, {\bf [432, 57]},
{\bf [432, 100]},{\bf [432, 102]}, {\bf [432, 103]},
{\bf [432, 239]}, {\bf [432, 260]}, {\bf [432, 273]},
[441, 1], {\bf [441, 7]}, {\bf [444, 14]}, [453, 1],
{\bf [468, 14]}, [471, 1], {\bf [486, 26]},
{\bf [486, 28]}, {\bf [486, 61]},
{\bf [486, 125]}, {\bf [486, 164]}, [489, 1], [507, 1], {\bf [507, 5]},
[513, 1], [513, 5], [513, 6], [513, 8], [513, 9], {\bf [516, 11]},
{\bf [525, 5]}, [543, 1], [549, 1], [567, 1], [567, 4], [567, 5], [567, 7],
{\bf [567, 12]}, {\bf [567, 13]},
{\bf [567, 14]}, {\bf [567, 23]}, [567, 36], {\bf [576, 3]},
[579, 1], {\bf [588, 11]}, {\bf [588, 16]}, {\bf [588, 60]},
[597, 1], {\bf [600, 45]}, {\bf [600, 179]}, [603, 1],
{\bf [624, 60]}, [633, 1], {\bf [648, 19]}, {\bf [648, 21]},
{\bf [648, 23]}, {\bf [648, 244]}, {\bf [648, 259]}, {\bf [648, 260]},
{\bf [648, 266]}, {\bf [648, 352]}, {\bf [648, 531]}, {\bf [648, 532]},
{\bf [648, 533]}, {\bf [648, 551]}, {\bf [648, 563]}, [651, 3], [651, 4],
[657, 1], [669, 1], {\bf [675, 5]}, {\bf [675, 9]},
{\bf [675, 11]}, {\bf [675, 12]}, {\bf [684, 11]}, [687, 1], [711, 1],
[723, 1], {\bf [726, 5]},
{\bf [729, 62]},{\bf [729, 63]}, {\bf [729, 64]}, {\bf [729, 80]},
{\bf [729, 86]}, [729, 94],
{\bf [729, 95]}, {\bf [729, 96]}, {\bf [729, 97]}, {\bf [729, 98]},
{\bf [729, 284]}, [729, 393],
{\bf [729, 397]}, {\bf [732, 14]}, [741, 3], [741, 4], {\bf [756, 11]},
{\bf [756, 113]}, {\bf [756, 114]}, {\bf [756, 116]}, {\bf [756, 117]}
\end{minipage}\tabularnewline
\hline
\end{tabular}
\caption{The {\tt SmallGroups} identifiers of the groups $G$
with three-dimensional irreducible representations.
Part 1: groups with $\mbox{order}(G) < 768$.
The identifiers in boldface denote the groups which have three-dimensional
irreducible representations in which some of the matrices
have twice-degenerate eigenvalues;
only those groups are relevant for this paper.}
\label{table1}
\end{table}
\begin{table}[ht!]
\begin{tabular}{|l|}
\hline
\begin{minipage}[t]{0.96\columnwidth}
{\bf [768, 1083477]}, {\bf [768, 1085333]}, {\bf [768, 1085335]},
{\bf [768, 1085351]},
[777, 3], [777, 4], {\bf [804, 11]}, [813, 1], [819, 3], [819, 4],
[831, 1], [837, 1],
[837, 4], [837, 5], [837, 7], [837, 8], [849, 1], {\bf [864, 69]},
{\bf [864, 194]}, {\bf [864, 675]}, {\bf [864, 701]}, {\bf [864, 703]},
{\bf [864, 737]}, {\bf [867, 2]}, [873, 1], {\bf [876, 14]}, {\bf [900, 66]},
[903, 5], [903, 6], {\bf [912, 57]}, [921, 1], [927, 1], [939, 1],
{\bf [948, 11]}, {\bf [972, 3]}, {\bf [972, 29]}, {\bf [972, 31]},
{\bf [972, 64]}, {\bf [972, 117]}, {\bf [972, 121]}, {\bf [972, 122]},
{\bf [972, 123]}, {\bf [972, 147]}, {\bf [972, 152]}, {\bf [972, 153]},
{\bf [972, 170]}, {\bf [972, 309]}, {\bf [972, 348]}, {\bf [972, 411]},
{\bf [972, 520]}, {\bf [972, 550]}, {\bf [975, 5]}, [981, 1], [993, 1],
[999, 1], [999, 5], [999, 6], [999, 8], [999, 9], {\bf [1008, 57]},
[1011, 1], {\bf [1014, 7]}, [1029, 6], {\bf [1029, 9]}, [1047, 1],
[1053, 16], [1053, 25], [1053, 26], [1053, 2], {\bf [1053, 29]},
{\bf [1053, 32]}, {\bf [1053, 35]}, {\bf [1053, 37]}, [1053, 47],
{\bf [1080, 260]}, [1083, 1], {\bf [1083, 5]}, {\bf [1089, 3]},
{\bf [1092, 68]}, {\bf [1092, 69]}, [1101, 1], {\bf [1116, 11]}, [1119, 1],
[1137, 1], [1143, 1], [1161, 6], [1161, 9],
[1161, 10], [1161, 11], [1161, 12], {\bf [1164, 14]}, {\bf [1176, 57]},
{\bf [1176, 243]}, [1191, 1], [1197, 3], [1197, 4],
{\bf [1200, 183]}, {\bf [1200, 384]}, {\bf [1200, 682]}, [1209, 3],
[1209, 4], [1227, 1], {\bf [1236, 11]}, [1251, 1], [1263, 1], [1281, 3],
[1281, 4], {\bf [1296, 3]}, {\bf [1296, 35]}, {\bf [1296, 37]},
{\bf [1296, 39]}, {\bf [1296, 220]}, {\bf [1296, 222]}, {\bf [1296, 226]},
{\bf [1296, 227]}, {\bf [1296, 228]}, {\bf [1296, 237]}, {\bf [1296, 647]},
{\bf [1296, 688]}, {\bf [1296, 689]}, {\bf [1296, 699]}, {\bf [1296, 1239]},
{\bf [1296, 1499]}, {\bf [1296, 1995]}, {\bf [1296, 2113]},
{\bf [1296, 2203]}, [1299, 1], {\bf [1308, 14]}, [1317, 1], [1323, 1],
[1323, 4], [1323, 5], [1323, 7], [1323, 8], {\bf [1323, 14]}, {\bf [1323, 40]},
{\bf [1323, 42]}, {\bf [1323, 43]}, {\bf [1332, 14]}, {\bf [1344, 393]},
{\bf [1350, 46]}, [1359, 1], [1371, 1], [1389, 1], {\bf [1404, 14]},
{\bf [1404, 137]}, {\bf [1404, 138]}, {\bf [1404, 140]}, {\bf [1404, 141]},
[1407, 3], [1407, 4], [1413, 1], {\bf [1425, 5]}, [1443, 3], [1443, 4],
{\bf [1452, 11]}, {\bf [1452, 34]}, {\bf [1458, 615]}, {\bf [1458, 618]},
{\bf [1458, 659]}, {\bf [1458, 663]}, {\bf [1458, 666]}, {\bf [1458, 1095]},
{\bf [1458, 1354]}, {\bf [1458, 1371]}, [1461, 1], [1467, 1],
{\bf [1488, 57]}, [1497, 1], [1521, 1], {\bf [1521, 7]}, {\bf [1524, 11]},
[1533, 3], [1533, 4]
\end{minipage}\tabularnewline
\hline
\end{tabular}
\caption{The {\tt SmallGroups} identifiers of the groups
with three-dimensional irreducible representations.
Part 2: groups $G$ with $768 \le \mbox{order}(G) < 1536$.
The identifiers in boldface stand for groups with three-dimensional
irreducible representations in which some of the matrices
have twice-degenerate eigenvalues.}
\label{table2}
\end{table}
\begin{table}[ht!]
\begin{tabular}{|l|}
\hline
\begin{minipage}[t]{0.96\columnwidth}
{\bf [1536, 408544632]}, {\bf [1536, 408544641]}, {\bf [1536, 408544678]},
{\bf [1536, 408544687]}, [1539, 16], [1539, 25], [1539, 26], [1539, 27],
{\bf [1539, 29]}, {\bf [1539, 32]}, {\bf [1539, 35]}, {\bf [1539, 37]},
[1539, 47], {\bf [1548, 11]}, [1569, 1], {\bf [1575, 7]}, {\bf [1587, 2]},
{\bf [1596, 55]}, {\bf [1596, 56]}, [1623, 1], [1629, 1], [1641, 1],
[1647, 6], [1647, 9], [1647, 10], [1647, 11], [1647, 12], [1659, 3],
[1659, 4], {\bf [1668, 11]}, [1677, 3], [1677, 4], [1701, 68],
{\bf [1701, 102]}, {\bf [1701, 112]}, {\bf [1701, 115]}, [1701, 126],
[1701, 127], [1701, 128], {\bf [1701, 130]}, {\bf [1701, 131]},
{\bf [1701, 135]}, {\bf [1701, 138]}, [1701, 240], {\bf [1701, 261]},
[1713, 1], {\bf [1728, 3]}, {\bf [1728, 185]}, {\bf [1728, 953]},
{\bf [1728, 1286]}, {\bf [1728, 1290]}, {\bf [1728, 1291]},
{\bf [1728, 2785]}, {\bf [1728, 2847]}, {\bf [1728, 2855]},
{\bf [1728, 2929]}, [1731, 1], {\bf [1734, 5]}, [1737, 1],
{\bf [1764, 11]}, {\bf [1764, 91]}, [1767, 3], [1767, 4],
{\bf [1776, 60]}, [1791, 1], [1803, 1], [1809, 6], [1809, 9], [1809, 10],
[1809, 11], [1809, 12], {\bf [1812, 11]}, [1821, 1],
[1839, 1], [1857, 1], {\bf [1872, 60]}, {\bf [1875, 16]},
{\bf [1884, 14]}, [1893, 1], [1899, 1], [1911, 3], [1911, 4],
{\bf [1911, 14]}, [1929, 1], {\bf [1944, 35]},
{\bf [1944, 37]}, {\bf [1944, 70]}, {\bf [1944, 707]},
{\bf [1944, 746]}, {\bf [1944, 832]}, {\bf [1944, 833]}, {\bf [1944, 849]},
{\bf [1944, 1123]}, {\bf [1944, 2293]}, {\bf [1944, 2294]},
{\bf [1944, 2333]}, {\bf [1944, 2363]}, {\bf [1944, 2415]},
{\bf [1944, 3448]}, [1953, 3], [1953, 4], {\bf [1956, 11]},
[1971, 6], [1971, 9], [1971, 10], [1971, 11], [1971, 12], [1983, 1]
\end{minipage}\tabularnewline
\hline
\end{tabular}
\caption{The {\tt SmallGroups} identifiers of the groups
with three-dimensional irreducible representations.
Part 3: groups with $1536 \le \mbox{order}(G) < 2000$.
The identifiers in boldface denote groups with three-dimensional
irreducible representations with matrices
having twice-degenerate eigenvalues.}
\label{table3}
\end{table}
not to be of type $G = \mathbbm{Z}_n \times G^\prime$
and to possess at least one faithful three-dimensional irrep.
Tables~\ref{table1} and~\ref{table2}
reproduce equation~(42) of ref.~\cite{kingludl},
while table~\ref{table3} is new.
We have explicitly constructed all the three-dimensional irreps
of the groups in tables~\ref{table1},
\ref{table2},
and~\ref{table3}.
From each of those irreps
we have discarded matrices proportional to the unit matrix.
We have divided the remaining matrices (of each irrep of each group)
into two sets:
the ones that have non-degenerate eigenvalues (set~$N$)
and the ones that have twice-degenerate eigenvalues (set~$T$).
We recall that a matrix $M$ has degenerate eigenvalues when
\begin{eqnarray}
4 \left[ I_2 \left( M \right) \right]^3
- \left( \mathrm{tr}\, M \right)^2 \left[ I_2 \left( M \right) \right]^2
+ 27 \left( \det{M} \right)^2
& & \nonumber \\
+ 4 \left( \mathrm{tr}\, M \right)^3 \det{M}
- 18 \left( \mathrm{tr}\, M \right) \left[ I_2 \left( M \right) \right]
\det{M} &=& 0,
\end{eqnarray}
where $I_2 \left( M \right) \equiv M_{11} M_{22} + M_{11} M_{33} + M_{22} M_{33}
- M_{12} M_{21} - M_{13} M_{31} - M_{23} M_{32}$ is the second-order invariant of $M$.
We have found that there are many groups for which set $T$ is empty;
we have discarded those groups.
The remaining groups---the ones that
have at least one three-dimensional faithful irrep
with a non-empty set $T$---are marked boldface in tables~\ref{table1},
\ref{table2},
and~\ref{table3}.
Those were the sole relevant groups and irreps for the remainder of our search.
We have then explicitly computed the eigenvalues of all the matrices
in both sets $N$ and $T$.
In our main search,
we have considered all possible pairs of
one matrix $T_\ell$ from set $T$ and one matrix $T_\nu$ from set $N$.
For each of those pairs we have computed the three $\left| U_{3j} \right|^2
(j = 1, 2, 3)$ by using equations~\eqref{sdihp};
in those equations,
$l_1$ is the degenerate eigenvalue of $T_\ell$,
$l_3$ is the non-degenerate eigenvalue of $T_\ell$,
and $n_{1,2,3}$ are the three eigenvalues of $T_\nu$.
We have discarded the set of the three $\left| U_{3j} \right|^2$
whenever any one of them turned out to vanish;
we have only collected the sets for which all three numbers
$\left| U_{31} \right|^2$,
$\left| U_{32} \right|^2$,
and $\left| U_{33} \right|^2$ were non-zero.
For our secondary search we have picked all possible couples
of two matrices $T_{\nu 1}$ and $T_{\nu 2}$ from the set $T$
and selected those couples that commute and that moreover
have a non-zero quantity $q$ defined in equation~\eqref{huigp}.
We have then picked one third matrix $T_\ell$,
and have computed the three $\left| U_{3j} \right|^2$
by using equations~\eqref{huoiy} together with
$\left| U_{31} \right|^2 = 1 - \left| U_{32} \right|^2 - \left| U_{33} \right|^2$.
Once again,
we have discarded all the sets of three $\left| U_{3j} \right|^2$
in which any one of them happened to vanish.
Since all the finite-dimensional representations of finite groups are unitary,
all the representations that we have dealt with are in principle equivalent
to representations through unitary matrices.
{\tt GAP} usually gives the representations in non-unitary form,
but we never have had to bring the representations to unitary form,
because all our computations were performed
in terms of basis-invariant quantities.
At the end of our search we have used {\tt GAP}
to find out the form of the group generated by the matrices $T_\ell$ and $T_\nu$
(in the main search) alone.
Let $\left\langle T_\ell,\ T_\nu \right\rangle$ denote that group.
It coincides in most cases with the initial group $G$,
but sometimes it is just a subgroup of it.
\section{Results} \label{results}
The searches described in the previous section
have produced a total of sixty sets
of three non-zero numbers $\left| U_{31} \right|^2$,
$\left| U_{32} \right|^2$,
and $\left| U_{33} \right|^2$;\footnote{Actually,
all the sets except one have been produced by the main search.
All but one of the sets produced by the secondary search merely reproduce sets
that had already been obtained in the main search.}
in each set,
$\left| U_{31} \right|^2 + \left| U_{32} \right|^2
+ \left| U_{33} \right|^2 = 1$.
From now on,
we let $V_j \equiv \left| U_{3j} \right|^2\ (j = 1, 2, 3)$
denote the three numbers in each set
and we assume that they have been ordered as $V_1 \le V_2 \le V_3$.
We christen each such set $\left\{ V_1,\ V_2,\ V_3 \right\}$ a `structure'.
We have plotted the sixty structures that we have found
as sixty---blue, green, and red---points in figure~\ref{figura}.
\begin{figure}[h!]
\begin{center}
\includegraphics[scale=0.47]{figurerows}
\end{center}
\vspace*{-3mm}
\caption{A depiction of the sixty structures
$\left\{ V_1,\ V_2,\ V_3 \right\}$ that have been produced by our searches.
The horizontal line gives $V_1$ and the vertical line gives $V_2$,
while $V_3 = 1 - V_1 - V_2$.
We have discarded structures
in which any of the three $V_j\ (j = 1, 2, 3)$ is zero.
The structures are ordered as $V_1 \le V_2 \le V_3$.
All the points are within the triangle bounded by the vertical axis
(equivalent to $V_1 = 0$) and by the green line.
The latter is composed of two segments with equations
$V_1 = V_2$ and $V_1 + 2 V_2 = 1\, \Leftrightarrow\, V_2 = V_3$.
The two segments meet at the point $V_1 = V_2 = V_3 = 1/3$.
The blue curve unites points with similar features (see text).}
\vspace*{3mm}
\label{figura}
\end{figure}
We note that all the sixty structures that we have found would also
have been found if we had restricted our search to subgroups of $SU(3)$.
Namely, for all the structures,
there is always at least one case in
which both matrices $T_\ell$ and $T_\nu$ have determinant 1.
We do not know why this happens but,
like Ludl has pointed out~\cite{ludl2},
it is possible that every finite subgroup of $U(3)$
is equivalent in physical terms to some finite subgroup of $SU(3)$,
because they can only produce the same Lagrangians.
We divide all the structures into three types.
Each of these types is described in one of the following subsections.
\subsection{Structures on the blue curve}
The first type encompasses a total of 44 structures.
They are described by the analytical formula
\begin{equation}
\label{suigp}
V_2 = \frac{1}{2} \left( 1 - V_1 - \sqrt{2 V_1 - 3 V_1^2} \right).
\end{equation}
This is depicted as a blue curve in figure~\ref{figura}.
The relevance of that curve had already been noticed
in refs.~\cite{lim,grimuslavoura}.
All but two of the 44 structures on the blue curve may be written
\begin{subequations}
\label{yutpe}
\begin{eqnarray}
V_1 &=& \frac{1}{3}
\left( 1 - \cos{\frac{2 \pi k}{3 n}} \right),
\\
V_2 &=& \frac{1}{3}
\left[1 - \cos{\frac{2 \pi \left( k - n \right)}{3 n}} \right],
\\
V_3 &=& \frac{1}{3}
\left[1 - \cos{\frac{2 \pi \left( k + n \right)}{3 n}} \right],
\end{eqnarray}
\end{subequations}
for positive integers $2 \le n \le 17$
and $k < n/2$ given in tables~\ref{tablenk}
\renewcommand{\arraystretch}{1.1}
\begin{table}[ht]
\centering
\begin{tabular}{|c||c|c|c|c|c|}
\hline
$n$ & 17 & 16 & 14 & 13 & 11
\\ \hline
$k$ & 1, 2, 3, 4, 5, 6, 7, 8 & 1, 3, 5, 7 & 1, 3, 5
& 1, 2, 3, 4, 5, 6 & 1, 2, 3, 4, 5
\\ \hline
\end{tabular}
\caption{The values of $n$ and $k$ to be used in equations~\eqref{yutpe}.}
\label{tablenk}
\end{table}
\renewcommand{\arraystretch}{1.0}
and~\ref{tablenk2}.
\vspace*{-5mm}
\renewcommand{\arraystretch}{1.1}
\begin{table}[ht]
\centering
\begin{tabular}{|c||c|c|c|c|c|c|c|c|c|}
\hline
$n$ & 10 & 9 & 8 & 7 & 6 & 5 & 4 & 3 & 2
\\ \hline
$k$ & 1, 3 & 1, 2, 4 & 1, 3 & 1, 2, 3 & 1 & 1, 2 & 1 & 1 & 1
\\ \hline
\end{tabular}
\caption{Continuation of table~\ref{tablenk}.}
\label{tablenk2}
\end{table}
\renewcommand{\arraystretch}{1.0}
\vspace*{5mm}
There are two more structures that are also described by equation~\eqref{suigp}
but cannot be described through equations~\eqref{yutpe}
for any small integers $n$ and $k$.
They are
\begin{equation}
\label{ghoty}
\left\{ \frac{1}{ 6+\csc \frac{3\pi}{14} },\
\frac{1}{ 6-\sec \frac{\pi}{7} },\
\frac{1}{ 6-\csc \frac{\pi}{14} }
\right\}
\approx \left\{ 0.131512,\ 0.204495,\ 0.663993 \right\}
\end{equation}
and
\begin{equation}
\label{bhigo}
\left\{ \frac{3 - \sqrt{5}}{8},\ 1/4,\ \frac{3 + \sqrt{5}}{8} \right\}
\approx \left\{ 0.0954915,\ 0.25,\ 0.654508 \right\}.
\end{equation}
The structures given by equations~\eqref{yutpe}
are produced by groups $\left\langle T_\ell,\ T_\nu \right\rangle$
which are either $\Delta (6 n^2)$ or $D \left( 9n, 1, 1; 2, 1, 1 \right)$,
where $n$ is a positive integer~\cite{su3}.
We recall that those are subgroups of $SU(3)$.
The group $\Delta (6 n^2)$~\cite{6n2}
has $6 n^2$ elements and is generated by
\begin{equation}
E = \left( \begin{array}{ccc} 0 & 1 & 0 \\ 0 & 0 & 1 \\ 1 & 0 & 0
\end{array} \right),
\quad
B = \left( \begin{array}{ccc} -1 & 0 & 0 \\ 0 & 0 & -1 \\ 0 & -1 & 0
\end{array} \right),
\quad \mbox{and} \quad
\left( \begin{array}{ccc} 1 & 0 & 0 \\ 0 & \eta & 0 \\ 0 & 0 & \eta^{-1}
\end{array} \right),
\end{equation}
where $\eta = \exp{\left( 2 i \pi / n \right)}$.
The group $D \left( 9n, 1, 1; 2, 1, 1 \right)$
has $162 n^2$ elements and is generated by $E$,
$B$, and
$\mathrm{diag} \left( \epsilon, \epsilon, \epsilon^{-2} \right)$,
where $\epsilon = \exp{\left[ 2 i \pi /
\! \left( 9 n \right) \right]}$~\cite{ludl1}.
In tables~\ref{Delta}
\renewcommand{\arraystretch}{1.3}
\begin{table}[h]
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline
$\Delta \left( 6 \times 2^2 \right)$ & $\Delta \left( 6 \times 3^2 \right)$ &
$\Delta \left( 6 \times 4^2 \right)$ & $\Delta \left( 6 \times 5^2 \right)$ &
$\Delta \left( 6 \times 6^2 \right)$
\\*[0.0pt]
[24, 12] & [54, 8] & [96, 64] & [150, 5] & [216, 95]
\\ \hline
$\Delta \left( 6 \times 7^2 \right)$ & $\Delta \left( 6 \times 8^2 \right)$ &
$\Delta \left( 6 \times 9^2 \right)$ & $\Delta \left( 6 \times 10^2 \right)$ &
$\Delta \left( 6 \times 11^2 \right)$
\\*[0.0pt]
[294, 7] & [384, 568] & [486, 61] & [600, 179] & [726, 5]
\\ \hline
$\Delta \left( 6 \times 12^2 \right)$ & $\Delta \left( 6 \times 13^2 \right)$ &
$\Delta \left( 6 \times 14^2 \right)$ & $\Delta \left( 6 \times 15^2 \right)$ &
$\Delta \left( 6 \times 16^2 \right)$
\\*[0.0pt]
[864, 701] & [1014, 7] & [1176, 243] & [1350, 46] & [1536, 408544632]
\\ \hline
$\Delta \left( 6 \times 17^2 \right)$ & $\Delta \left( 6 \times 18^2 \right)$ &
& &
\\*[0.0pt]
[1734, 5] & [1944, 849] & & &
\\ \hline
\end{tabular}
\caption{The {\tt SmallGroups} identifiers
of the groups $\Delta (6 n^2)$ with order smaller than 2\,000.}
\label{Delta}
\end{table}
\renewcommand{\arraystretch}{1.0}
and~\ref{D}
\renewcommand{\arraystretch}{1.1}
\begin{table}[h]
\centering
\begin{tabular}{|c|c|c|}
\hline
$n = 1$ & $n = 2$ & $n = 3$
\\*[0,0pt]
[162, 14] & [648, 259] & [1458, 659]
\\ \hline
\end{tabular}
\caption{The {\tt SmallGroups} identifiers of the groups
$D \left( 9 n, 1, 1; 2, 1, 1 \right) \equiv D^{(1)}_{9n, 3n}$~\cite{ludl1}
with order smaller than 2\,000.}
\label{D}
\end{table}
\renewcommand{\arraystretch}{1.0}
we give the {\tt SmallGroups} identifiers of these groups,
which have structures
\begin{equation}
\left( \mathbbm{Z}_n \times \mathbbm{Z}_n \right) \rtimes S_3
\quad \mbox{and} \quad
\left( \mathbbm{Z}_{9n} \times \mathbbm{Z}_{3n} \right) \rtimes S_3,
\end{equation}
respectively,
where $S_3$ is the permutation group of three objects.
The structure~\eqref{ghoty} is produced in the main search
by the group $\left\langle T_\ell,\ T_\nu \right\rangle
= [168, 42] = \Sigma \left( 168 \right)$,
which is an `exceptional' finite subgroup of $SU(3)$.
The structure~\eqref{bhigo}
is produced by another exceptional subgroup of $SU(3)$,
the group
$[60, 5] = \Sigma \left( 60 \right) \cong A_5$
(the symmetry group of the icosahedron)
in the secondary search~\cite{varzielas}.
That structure is the only one produced in the secondary search
that was not also a result of the main search;
it had already been found in refs.~\cite{varzielas,lavouraludl}.
It is convenient to number all the structures belonging to the blue curve
of figure~\ref{figura}
according to increasing values of $V_1$.
We thus construct the following list of the 44 structures:
\begin{subequations}
\label{u8}
\begin{eqnarray}
\mbox{structure 1}\ (n = 17,\ k = 1): & &
\left\{ 0.0025265,\ 0.463262,\ 0.534212 \right\};
\nonumber \\
\mbox{structure 2}\ (n = 16,\ k = 1): & &
\left\{ 0.00285171,\ 0.460894,\ 0.536254 \right\};
\nonumber \\
\mbox{structure 3}\ (n = 14,\ k = 1): & &
\left\{ 0.00372306,\ 0.455114,\ 0.541163 \right\};
\nonumber \\
\mbox{structure 4}\ (n = 13,\ k = 1): & &
\left\{ 0.00431658,\ 0.451535,\ 0.544148 \right\};
\nonumber \\
\mbox{structure 5}\ (n = 11,\ k = 1): & &
\left\{ 0.00602377,\ 0.442356,\ 0.55162 \right\};
\nonumber \\
\mbox{structure 6}\ (n = 10,\ k = 1): & &
\left\{ 0.00728413,\ 0.436339,\ 0.556377 \right\};
\nonumber \\
\mbox{structure 7}\ (n = 9,\ k = 1): & &
\left\{ 0.00898504,\ 0.428934,\ 0.562081 \right\};
\nonumber \\
\mbox{structure 8}\ (n = 17,\ k = 2): & &
\left\{ 0.0100677,\ 0.424554,\ 0.565378 \right\};
\nonumber \\
\mbox{structure 9}\ (n = 8,\ k = 1): & &
\left\{ 0.0113581,\ 0.419606,\ 0.569036 \right\};
\nonumber \\
\mbox{structure 10}\ (n = 7,\ k = 1): & &
\left\{ 0.0148091,\ 0.407507,\ 0.577684 \right\};
\nonumber \\
\mbox{structure 11}\ (n = 13,\ k = 2): & &
\left\{ 0.0171545,\ 0.400009,\ 0.582837 \right\};
\nonumber \\
\mbox{structure 12}\ (n = 6,\ k = 1): & &
\left\{ 0.0201025,\ 0.391216,\ 0.588681 \right\};
\nonumber \\
\mbox{structure 13}\ (n = 17,\ k = 3): & &
\left\{ 0.0225093,\ 0.384464,\ 0.593027 \right\};
\nonumber \\
\mbox{structure 14}\ (n = 11,\ k = 2): & &
\left\{ 0.0238774,\ 0.380772,\ 0.595351 \right\};
\nonumber \\
\mbox{structure 15}\ (n = 16,\ k = 3): & &
\left\{ 0.0253735,\ 0.376842,\ 0.597784 \right\};
\nonumber \\
\mbox{structure 16}\ (n = 5,\ k = 1): & &
\left\{ 0.0288182,\ 0.368176,\ 0.603006 \right\};
\nonumber \\
\mbox{structure 17}\ (n = 14,\ k = 3): & &
\left\{ 0.0330104,\ 0.358243,\ 0.608746 \right\};
\nonumber \\
\mbox{structure 18}\ (n = 9,\ k = 2): & &
\left\{ 0.0354558,\ 0.352715,\ 0.611829 \right\};
\nonumber \\
\mbox{structure 19}\ (n = 13,\ k = 3): & &
\left\{ 0.0381813,\ 0.346755,\ 0.615063 \right\};
\nonumber \\
\mbox{structure 20}\ (n = 17,\ k = 4): & &
\left\{ 0.0396626,\ 0.343598,\ 0.616739 \right\};
\nonumber \\
\mbox{structure 21}\ (n = 4,\ k = 1): & &
\left\{ 0.0446582,\ 0.333333,\ 0.622008 \right\};
\nonumber \\
\mbox{structure 22}\ (n = 11,\ k = 3): & &
\left\{ 0.0529155,\ 0.317473,\ 0.629612 \right\};
\nonumber \\
\mbox{structure 23}\ (n = 7,\ k = 2): & &
\left\{ 0.0579204,\ 0.308423,\ 0.633656 \right\};
\nonumber \\
\mbox{structure 24}\ (n = 17,\ k = 5): & &
\left\{ 0.0612677,\ 0.302577,\ 0.636155 \right\};
\nonumber \\
\mbox{structure 25}\ (n = 10,\ k = 3): & &
\left\{ 0.063661,\ 0.298491,\ 0.637848 \right\};
\nonumber \\
\mbox{structure 26}\ (n = 13,\ k = 4): & &
\left\{ 0.0668524,\ 0.293154,\ 0.639993 \right\}; \hspace*{7mm}
\nonumber \\
\mbox{structure 27}\ (n = 16,\ k = 5): & &
\left\{ 0.0688822,\ 0.289825,\ 0.641293 \right\};
\nonumber \\
\mbox{structure 28}\ (n = 3,\ k = 1): & &
\left\{ 0.0779852,\ 0.275451,\ 0.646564 \right\}; \hspace*{7mm}
\nonumber \\
\mbox{structure 29}\ (n = 17,\ k = 6): & &
\left\{ 0.086997,\ 0.262022,\ 0.650981 \right\};
\nonumber \\
\mbox{structure 30}\ (n = 14,\ k = 5): & &
\left\{ 0.0889827,\ 0.25916,\ 0.651858 \right\};
\nonumber \\
\mbox{structure 31}\ (n = 11,\ k = 4): & &
\left\{ 0.0920887,\ 0.254747,\ 0.653164 \right\};
\nonumber \\
\mbox{structure 32}: & & \left\{ 0.0954915,\ 0.25,\ 0.654508 \right\};
\nonumber \\
\mbox{structure 33}\ (n = 8,\ k = 3): & &
\left\{ 0.0976311,\ 0.24706,\ 0.655309 \right\};
\nonumber \\
\mbox{structure 34}\ (n = 13,\ k = 5): & &
\left\{ 0.102425,\ 0.240594,\ 0.656981 \right\};
\nonumber \\
\mbox{structure 35}\ (n = 5,\ k = 2): & &
\left\{ 0.11029,\ 0.230328,\ 0.659383 \right\};
\nonumber \\
\mbox{structure 36}\ (n = 17,\ k = 7): & &
\left\{ 0.116461,\ 0.222548,\ 0.660991 \right\};
\nonumber \\
\mbox{structure 37}\ (n = 7,\ k = 3): & &
\left\{ 0.125503,\ 0.211553,\ 0.662944 \right\};
\nonumber \\
\mbox{structure 38}\ (n = 16,\ k = 7): & &
\left\{ 0.130413,\ 0.205772,\ 0.663815 \right\};
\nonumber \\
\mbox{structure 39}: & & \left\{ 0.131512,\ 0.204495,\ 0.663993 \right\};
\nonumber \\
\mbox{structure 40}\ (n = 9,\ k = 4): & &
\left\{ 0.13428,\ 0.201307,\ 0.664413 \right\};
\nonumber \\
\mbox{structure 41}\ (n = 11,\ k = 5): & &
\left\{ 0.139981,\ 0.194862,\ 0.665157 \right\};
\nonumber \\
\mbox{structure 42}\ (n = 13,\ k = 6): & &
\left\{ 0.143978,\ 0.190436,\ 0.665586 \right\};
\nonumber \\
\mbox{structure 43}\ (n = 17,\ k = 8): & &
\left\{ 0.149212,\ 0.184754,\ 0.666034 \right\};
\nonumber \\
\mbox{structure 44}\ (n = 2,\ k = 1): & &
\left\{ 1/6,\ 1/6,\ 2/3 \right\}.
\nonumber
\end{eqnarray}
\end{subequations}
\subsection{Structures on the green line}
The second type of structures features two equal $V_j$,
\textit{i.e.}\ either $V_1 = V_2$ or $V_2 = V_3$.
These structures straddle the green line in figure~\ref{figura}.
One of them is of course structure~44,
which corresponds to the point where the green line meets the blue curve.
There are six more such structures on the green line.
They are
\begin{subequations}
\begin{eqnarray}
\mbox{structure 45}\
\left( V_1 = \frac{3 - \sqrt{5}}{12} \right): & &
\left\{ 0.063661,\ 0.063661,\ 0.872678 \right\};
\nonumber \\
\mbox{structure 46}\
\left( V_1 = \frac{5 - \sqrt{5}}{20} \right): & &
\left\{ 0.138197,\ 0.138197,\ 0.723607 \right\};
\nonumber \\
\mbox{structure 47}\
\left( V_1 = \frac{1}{4} \right): & &
\left\{ 0.25,\ 0.25,\ 0.5 \right\};
\nonumber \\
\mbox{structure 48}\
\left( V_1 = \frac{1}{3} \right): & &
\left\{ 0.333333,\ 0.333333,\ 0.333333 \right\};
\nonumber \\
\mbox{structure 49}\
\left( V_1 = \frac{5 - \sqrt{5}}{10} \right): & &
\left\{ 0.276393,\ 0.361803,\ 0.361803 \right\};
\nonumber \\
\mbox{structure 50}\
\left( V_1 = \frac{3 - \sqrt{5}}{6} \right): & &
\left\{ 0.127322,\ 0.436339,\ 0.436339 \right\}.
\nonumber
\end{eqnarray}
\end{subequations}
The structures 45, 46, 49, and 50 all originate
in $\left\langle T_\ell,\ T_\nu \right\rangle = A_5$.
Structure~47 originates in the permutation group
$S_4 \cong \Delta (6 \times 2^2)$.
Structure~48 originates in the alternating group
$A_4 \cong \Delta (3 \times 2^2)$,
with {\tt SmallGroups} identifier [12, 3].
Both structures~46 and~49 justify the \textit{Ansatz}~\cite{goldenratio}
\begin{equation}
\cot{\theta_{12}} = \varphi \equiv \frac{1 + \sqrt{5}}{2},
\end{equation}
relating the lepton mixing angle $\theta_{12}$ to the `golden ratio' $\varphi$.
Indeed,
$2 V_1 =
1 \! \left/ \left( 1 + \varphi^2 \right) \right.$ in structure~46
and $ V_1 =
1 \! \left/ \left( 1 + \varphi^2 \right) \right.$ in structure~49.
\subsection{Isolated structures}
Besides the points on the blue curve and the points on the green line,
there are ten isolated points marked red in figure~\ref{figura}.
The corresponding structures are
\begin{eqnarray}
& & \mbox{structure~51}:\
\left\{0.0347854,\ 0.166667,\ 0.798548 \right\}
\nonumber \\*[1mm]
&=& \left\{ \frac{5 - \sqrt{21}}{12},\ \frac{1}{6},\ \frac{5 + \sqrt{21}}{12}
\right\};
\nonumber \\*[2mm]
& & \mbox{structure~52}:\
\left\{ 0.0389375,\ 0.306554,\ 0.654508 \right\}
\nonumber \\*[1mm]
&=& \left\{ \frac{5 + \sqrt{3} - \sqrt{5} - \sqrt{15}}{16},\
\frac{5 - \sqrt{3} - \sqrt{5} + \sqrt{15}}{16},\
\frac{3 + \sqrt{5}}{8}
\right\};
\nonumber \\*[2mm]
& & \mbox{structure~53}:\
\left\{ 0.0442811,\ 0.25,\ 0.705719 \right\} =
\left\{ \frac{3 - \sqrt{7}}{8},\ \frac{1}{4},\ \frac{3 + \sqrt{7}}{8}
\right\};
\nonumber \\*[2mm]
& & \mbox{structure~54}:\
\left\{ 0.0625591,\ 0.138197,\ 0.799244 \right\}
\nonumber \\*[1mm]
&=& \left\{ \frac{15 + \sqrt{5} - \sqrt{150 + 30 \sqrt{5}}}{40},\
\frac{5 - \sqrt{5}}{20},\
\frac{15 + \sqrt{5} + \sqrt{150 + 30 \sqrt{5}}}{40} \right\};
\nonumber \\*[2mm]
& & \mbox{structure~55}:\
\left\{ 0.08592426701,\ 0.201689718788,\ 0.712386014201 \right\}
\nonumber \\*[1mm]
&=& \left\{ \frac{2}{3 \left( 2+\csc \frac{\pi}{18} \right)},\
\frac{2}{3 \left( 2+\sec \frac{2 \pi}{9} \right)},\
\frac{2}{3 \left( 2-\sec \frac{\pi}{9} \right)}
\right\};
\nonumber \\*[2mm]
& & \mbox{structure~56}:\
\left\{ 0.0914501,\ 0.361803,\ 0.546747 \right\}
\nonumber \\*[1mm]
&=& \left\{ \frac{15 - \sqrt{5} - \sqrt{150 - 30 \sqrt{5}}}{40},\
\frac{5 + \sqrt{5}}{20},\
\frac{15 - \sqrt{5} + \sqrt{150 - 30 \sqrt{5}}}{40} \right\};
\nonumber \\*[2mm]
& & \mbox{structure~57}:\
\left\{ 0.0954915,\ 0.101940,\ 0.802569 \right\}
\nonumber \\*[1mm]
&=& \left\{ \frac{3 - \sqrt{5}}{8},
\frac{5 - \sqrt{3} + \sqrt{5} - \sqrt{15}}{16},\
\frac{5 + \sqrt{3} + \sqrt{5} + \sqrt{15}}{16}
\right\};
\nonumber \\*[2mm]
& & \mbox{structure~58}:\
\left\{ 0.105662,\ 0.394338,\ 0.5 \right\} =
\left\{ \frac{3 - \sqrt{3}}{12},\ \frac{3 + \sqrt{3}}{12},\ \frac{1}{2}
\right\};
\nonumber \\*[2mm]
& & \mbox{structure~59}:\
\left\{ 0.158494,\ 0.25,\ 0.591506 \right\} =
\left\{ \frac{3 - \sqrt{3}}{8},\ \frac{1}{4},\ \frac{3 + \sqrt{3}}{8}
\right\};
\nonumber \\*[2mm]
& & \mbox{structure~60}:\
\left\{ 1/6,\ 1/3,\ 1/2 \right\}.
\nonumber
\end{eqnarray}
All structures~51--60 but structures~55 and~58
are either rows or columns of matrices found in ref.~\cite{fonseca};
we have taken their analytic expressions from that paper.
Structures~51--60 originate in the following
groups $\left\langle T_\ell,\ T_\nu \right\rangle$:
\begin{itemize}
\item The exceptional $SU(3)$ subgroup $\Sigma \left( 168 \right)$,
which has {\tt SmallGroups} identifier [168, 42],
for structures~51 and~53;
\item The exceptional $SU(3)$ subgroup
$\Sigma \left( 360 \times 3 \right)$,
which has {\tt SmallGroups} identifier [1080, 260],
for structures~52, 54, 56, and~57;
\item any of three groups
of order 648---with {\tt SmallGroups} identifiers [648, 531],
[648, 532],
and [648, 533]---for structures~55,
58,
and~60.
Those three groups have similar structure
$\left\{ \left[ \left( \mathbbm{Z}_3 \times \mathbbm{Z}_3 \right) \rtimes \mathbbm{Z}_3 \right]
\rtimes Q_8 \right\} \rtimes \mathbbm{Z}_3$,
where $Q_8$ is the quaternion group,
a subgroup of $SU(2)$.
The group [648, 532] is the $SU(3)$ exceptional subgroup
$\Sigma \left( 216 \times 3 \right)$.
\item for structure~59,
$\left\langle T_\ell,\ T_\nu \right\rangle$ may be either
the exceptional $SU(3)$ subgroup $\Sigma \left( 36 \times 3 \right)$,
with {\tt SmallGroups} identifier [108, 15],
or some other groups with analogous structures
and orders which are multiple of 108,
like [216, 25], [324, 111], [432, 57], and so on.
\end{itemize}
Table~\ref{exceptional} gives the {\tt SmallGroups} identifiers
of all six exceptional subgroups of $SU(3)$.
\renewcommand{\arraystretch}{1.1}
\begin{table}[h]
\centering
\begin{tabular}{|c|c|c|c|c|c|}
\hline
$\Sigma \left( 60 \right)$ & $\Sigma \left( 168 \right)$ &
$\Sigma \left( 36 \times 3 \right)$ & $\Sigma \left( 72 \times 3 \right)$ &
$\Sigma \left( 216 \times 3 \right)$ & $\Sigma \left( 360 \times 3 \right)$
\\*[0.0pt]
[60, 5] & [168, 42] & [108, 15] & [216, 88] & [648, 532] & [1080, 260]
\\ \hline
\end{tabular}
\caption{The {\tt SmallGroups} identifiers
of the exceptional finite subgroups of $SU(3)$.}
\label{exceptional}
\end{table}
\renewcommand{\arraystretch}{1.0}
\section{Comparison with the data} \label{comparison}
The matrix $\left| U \right|^2$ is parameterized as
\begin{equation}
\left| U \right|^2 = \left( \begin{array}{ccc}
c_{12}^2 c_{13}^2 & s_{12}^2 c_{13}^2 & s_{13}^2 \\
s_{12}^2 c_{23}^2 + c_{12}^2 s_{23}^2 s_{13}^2 + Y &
c_{12}^2 c_{23}^2 + s_{12}^2 s_{23}^2 s_{13}^2 - Y & s_{23}^2 c_{13}^2 \\
s_{12}^2 s_{23}^2 + c_{12}^2 c_{23}^2 s_{13}^2 - Y &
c_{12}^2 s_{23}^2 + s_{12}^2 c_{23}^2 s_{13}^2 + Y & c_{23}^2 c_{13}^2
\end{array} \right),
\end{equation}
where $c_{ij} \equiv \cos{\theta_{ij}}$ and $s_{ij} \equiv \sin{\theta_{ij}}$
for $(ij) = (12), (23), (13)$.
The quantity $Y \equiv 2 c_{12} s_{12} c_{23} s_{23} s_{13} \cos{\delta}$.
There are in the literature three global phenomenological fits
to the parameters $\theta_{12}$,
$\theta_{23}$,
$\theta_{13}$,
and $\delta$.
The specific bounds on each parameter depend on which of refs.~\cite{tortola},
\cite{fogli}, or~\cite{schwetz} one uses
and also on whether a `normal' or `inverted' ordering
is assumed for the neutrino masses.
For definiteness,
we shall use the values of ref.~\cite{tortola} for an inverted ordering.
They are
\begin{equation}
s_{12}^2 \in \left[ 0.278,\ 0.375 \right], \quad
s_{23}^2 \in \left[ 0.403,\ 0.640 \right], \quad
s_{13}^2 \in \left[ 0.0183,\ 0.0297 \right]
\end{equation}
at $3 \sigma$ level,
\begin{equation}
s_{12}^2 \in \left[ 0.292,\ 0.357 \right], \quad
s_{23}^2 \in \left[ 0.432,\ 0.621 \right], \quad
s_{13}^2 \in \left[ 0.0202,\ 0.0278 \right]
\end{equation}
at $2 \sigma$ level,
and
\begin{equation}
\begin{array}{lcl}
s_{12}^2 \in \left[ 0.307,\ 0.339 \right],
& &
s_{23}^2 \in \left[ 0.530,\ 0.598 \right],
\\*[1mm]
s_{13}^2 \in \left[ 0.0221,\ 0.0259 \right],
& &
\delta \in \left[ 1.16\, \pi,\ 1.82\, \pi \right]
\end{array}
\end{equation}
at $1 \sigma$ level.
Note that $\cos{\delta}$ is free at both $3 \sigma$ and $2 \sigma$ levels.
Up to now,
we have considered that the structures correspond to predictions
for the third row of $\left| U \right|^2$.
This is because $T_\ell$ and $H_\ell$ are simultaneously diagonalizable
and we have assumed that the eigenvalues of $T_\ell$ obey $l_1 = l_2 \neq l_3$.
However,
\begin{itemize}
\item we might instead have assumed either $l_1 = l_3 \neq l_2$
or $l_2 = l_3 \neq l_1$,
and then we would have predicted,
through exactly the same mathematics,
either the second or the first row,
respectively,
of $\left| U \right|^2$;
\item equations~\eqref{sdihp} are invariant under simultaneous permutations
of
\begin{equation}
\left( \left| U_{31} \right|^2,\ \left| U_{32} \right|^2,\
\left| U_{33} \right|^2 \right)
\quad \mbox{and} \quad \left( n_1,\ n_2,\ n_3 \right),
\end{equation}
hence,
by altering the ordering of the eigenvalues of $T_\nu$
one may alter the ordering of the matrix elements
in the row of $\left| U \right|^2$ that one is predicting;
\item $\left( H_\ell,\ T_\ell \right)$ and $\left( H_\nu,\ T_\nu \right)$
may be interchanged in the reasoning of section~\ref{theory},
and then we would be predicting a \emph{column}\/ instead of a \emph{row}\/
of $\left| U \right|^2$.
\end{itemize}
Thus,
the structures in the previous
section must be interpreted as predictions for either \emph{any row}\/
or \emph{any column}\/ of $\left| U \right|^2$;
moreover,
for each such column or row of $\left| U \right|^2$
the three numbers of the structure may be taken in any order.
We may now state the results of the confrontation of our structures
with the data:
\begin{description}
\item If one tries to fit the first row of $\left| U \right|^2$
through one of our structures,
one finds that none of them is able to do it.
Indeed,
though structures~12--16 have $V_1$ adequate
to fit $\left| U_{13} \right|^2$,
their $V_2$ is much too large
to agree with the upper bound on $s_{12}^2$.
\item If one tries to fit the third column of $\left| U \right|^2$
through one of our structures,
one finds that structures~12 and~16 do it at $3\sigma$ level
and structures~13--15 manage it at $2\sigma$ level.
\item Many structures on the blue curve of figure~\ref{figura}
fit the first column of $\left| U \right|^2$:
structures~29--44 do it at $1\sigma$ level,
structures~26--28 at $2\sigma$ level,
and structures~24 and~25 at $3\sigma$ level.
The structures on the green line in figure~1
(with the exception of structure~44)
and the isolated structures in that figure
cannot fit the first column of $\left| U \right|^2$,
even though some of the corresponding points appear close to the blue
line in figure~\ref{figura}.
Notice that structure~32 incorporates
one of the `golden ratio' \textit{Ans\"atze}\/
for the mixing angle $\theta_{12}$,
namely $\cos{\theta_{12}} = \varphi / 2\, \Leftrightarrow\,
\cos^2{\theta_{12}}
= \left. \left( 3 + \sqrt{5} \right) \right/ \! 8$~\cite{rodejohann};
it was already pointed out in ref.~\cite{varzielas} that that particular
structure gives an excellent fit to the first column of $\left| U \right|^2$.
\item The second column of $\left| U \right|^2$
is fitted at $2 \sigma$ level by structure~48
and at $3 \sigma$ level by structures~49 and~60.
Structure~48 corresponds
to the well-known `trimaximal mixing'~\cite{trimaximal}
or `TM$_2$'~\cite{TM2}
\textit{Ansatz}\/ for that column of $\left| U \right|^2$.
Structure~49 corresponds to one of the two~\cite{adulpravitchai}
`golden ratio' \textit{Ans\"atze}\/ for $\left| U_{12} \right|^2$.
\item The second row of $\left| U \right|^2$ may be fitted
at $1\sigma$ level by structure~56,
at $2\sigma$ level by structures~47, 50, and~58--60,
and at $3\sigma$ level by structure~21.
\item The third row of $\left| U \right|^2$ may be fitted
at $1\sigma$ level by structure~50,
at $2\sigma$ level by structures~47, 56, 58, and~60,
and at $3\sigma$ level by structure~49.
The adequateness of structure~47 had already been pointed out
in ref.~\cite{lavouraludl}.
\end{description}
It should be stressed that the fits in this section
are rather sensitive to the precise phenomenlogical bounds that one utilizes
for the three $s_{ij}^2$ and for $\cos{\delta}$.
We have used here the bounds in ref.~\cite{tortola}
with an inverted neutrino mass ordering but,
if we had instead used the bounds
in either ref.~\cite{fogli} or ref.~\cite{schwetz},
or the bounds for a normal ordering,
then our results would have differed somewhat.
\section{Summary} \label{conclusions}
In this paper we have assumed that neutrinos are Dirac particles,
\textit{i.e.}\ that the lepton sector is similar to the quark sector.
We have assumed that mixing in the lepton sector originates
from a discrete horizonal symmetry group $G$
that breaks into two distinct subgroups $G_\ell$ and $G_\nu$
under which the mass matrices of the charged leptons and of the neutrinos
are separately invariant.
We have considered the special situation in which one of the subgroups
$G_{\ell,\nu}$ is generated by a matrix with two equal eigenvalues
while the other subgroup
is generated by a matrix with non-degenerate eigenvalues.
Under these assumptions,
by making a {\tt GAP/SmallGroups} search
of all possible groups $G$ of order smaller than 2\,000,
we have found 60 possible structures for either a row or a column
of the lepton mixing matrix.
Several of those structures constitute realistic
predictions for either any of the columnsor for the second row
or for the third row of the PMNS matrix.
\begin{appendix}
\section{Nilpotent groups and groups with a normal Sylow 3-subgroup}
\label{groups}
\setcounter{equation}{0}
\renewcommand{\theequation}{A\arabic{equation}}
\subsection{Nilpotent groups}
The \emph{commutator}\/ $\left[ f, g \right]$
of two group elements $f$ and $g$ is the group element $f^{-1} g^{-1 } f g$.
Clearly,
if $f$ commutes with $g$,
\textit{i.e.}\ if $f g = g f$,
then $\left[ f, g \right]$ is the identity element $e$ of the group;
the converse is also true:
if $\left[ f, g \right] = e$ then $f g = g f$.
Let $G$ be a finite group.
Let $F$ be one of its subgroups.
Then,
we define $\left[ F, G \right]$ as the subgroup of $G$
generated by all the elements $\left[ f, g \right]$ of $G$
which are the commutators of some $f \in F$ and some $g \in G$.
(Notice that the commutators do not in general close
under the group multiplication;
therefore,
one must use them to \emph{generate}\/ a subgroup of $G$.)
We next define the \emph{descending central series}\/ of a group $G$.
This is the series $G, G_1, G_2, G_3, \ldots$ of subgroups of $G$
defined through the procedure $G_1 = \left[ G, G \right]$,
$G_2 = \left[ G_1, G \right]$,
$G_3 = \left[ G_2, G \right]$,
and so on.
We finally define a \emph{nilpotent}\/ group:
it is a group whose descending central series ends up
in the trivial group---the one consisting solely of the identity element.
Let us give an example of a nilpotent group:
the $D_4$ group generated by the two matrices
\begin{equation}
\label{dd44}
A = \left( \begin{array}{cc} 0 & 1 \\ 1 & 0 \end{array} \right)
\quad \mbox{and} \quad
B = \left( \begin{array}{cc} 1 & 0 \\ 0 & -1 \end{array} \right).
\end{equation}
Clearly,
\begin{equation}
\label{uhiy}
\left[ A, A \right] = \left[ B, B \right] =
\left( \begin{array}{cc} 1 & 0 \\ 0 & 1 \end{array} \right)
\quad \mbox{and} \quad
\left[ A, B \right] =
\left( \begin{array}{cc} -1 & 0 \\ 0 & -1 \end{array} \right).
\end{equation}
Therefore,
$\left[ D_4, D_4 \right] = \mathbbm{Z}_2$
is formed by the two matrices in~\eqref{uhiy}.
Since those two matrices commute with all the matrices of $D_4$
(indeed, they commute with any $2 \times 2$ matrix),
$\left[ \left[ D_4, D_4 \right], D_4 \right]$
is just the unit $2 \times 2$ matrix,
\textit{i.e}\ it is the trivial group.
Therefore,
$D_4$ is nilpotent.
We next give an example of a non-nilpotent group:
the $D_3$ group generated by the two matrices
\begin{equation}
\label{ss33}
A = \left( \begin{array}{cc} 0 & 1 \\ 1 & 0 \end{array} \right)
\quad \mbox{and} \quad
C = \left( \begin{array}{cc} \omega & 0 \\ 0 & \omega^2 \end{array} \right),
\end{equation}
where $\omega = \exp{\left( 2 i \pi / 3 \right)}$.
Clearly,
\begin{equation}
\label{upiy}
\left[ A, A \right] = \left[ C, C \right] =
\left( \begin{array}{cc} 1 & 0 \\ 0 & 1 \end{array} \right)
\quad \mbox{and} \quad
\left[ A, C \right] =
\left( \begin{array}{cc} \omega^2 & 0 \\ 0 & \omega \end{array} \right) = C^{-1}.
\end{equation}
Therefore,
$\left[ D_3, D_3 \right] = \mathbbm{Z}_3$
is the group generated by $C^{-1}$.
We compute $\left[ C^{-1}, A \right] = C^{-1}$
to conclude that $\left[ \left[ D_3, D_3 \right], D_3 \right]$
is again $\mathbbm{Z}_3$.
Therefore,
the descending central series of $D_3$ is $D_3$,
$\mathbbm{Z}_3$,
$\mathbbm{Z}_3$,
$\mathbbm{Z}_3, \ldots$;
this central series never ends up in the trivial group,
hence $D_3$ is not nilpotent.
\subsection{Groups with a normal Sylow 3-subgroup}
Let $G$ be a finite group.
Let $F$ be one of its subgroups.
Let $g \in G$ be an element of $G$.
Then,
the \emph{left coset}\/ of $F$ with respect to $g$
is the set of all elements $h \in G$ that may be written $h = g f$
for some $f \in F$.
Similarly,
the \emph{right coset}\/ of $F$ with respect to $g$
is defined to be the set of all the $h^\prime \in G$
that may be written $h^\prime = f g$ for some $f \in F$.
The subgroup $F$ of $G$ is said to be \emph{normal}\/
if its left coset with respect to any $g \in G$
coincides with the right coset with respect to $g$.
Take for instance the group $D_3$ generated by the matrices in~\eqref{ss33}.
It has a $\mathbbm{Z}_3$ subgroup formed by $C$,
$C^2$,
and the $2 \times 2$ unit matrix.
It is easily seen that $A C = C^2 A$ and $A C^2 = C A$.
Therefore,
the left and right cosets of $\mathbbm{Z}_3$ relative to $A$ are identical;
hence,
$\mathbbm{Z}_3$ is a normal subgroup of $D_3$.
The \emph{order}\/ of an element $g$ of a finite group $G$
is the smallest positive integer $o$
such that $g^o = e$ is the identity of $G$.
Let $p$ be a prime number,
then a $p$-group is a group where \emph{all}\/ the elements have order $o$
which is a power of $p$
(different elements may have different orders,
but all the orders are powers of $p$).
Thus,
a 3-group is a group where all the elements either have order one,
or order three,
or order nine,
\textit{etc.}
An obvious example is the well-known group $\Delta(27)$,
generated by
\begin{equation}
\left( \begin{array}{ccc} 0 & 1 & 0 \\ 0 & 0 & 1 \\ 1 & 0 & 0
\end{array} \right)
\quad \mbox{and} \quad
\left( \begin{array}{ccc} 1 & 0 & 0 \\ 0 & \omega & 0 \\ 0 & 0 & \omega^2
\end{array} \right);
\end{equation}
all 27 elements of $\Delta(27)$ except the identity have order three.
Similarly,
the group $D_4$ generated by the matrices in~\eqref{dd44} is a 2-group---all
its elements either have order one, or order two, or order four.
In general,
\emph{a finite group is a $p$-group if and only if
the number of elements of the group is a power of $p$};
thus,
the 3-groups are the groups with $3^n$ elements,
for some integer $n$.
A subgroup $S$ of a group $G$
is called a \emph{Sylow $p$-subgroup}\/ if it is a $p$-group
and if there is no larger $p$-subgroup of $G$ that contains $S$
as a proper subgroup.
Thus,
a normal Sylow 3-subgroup $S$ of a finite group $G$
is a normal subgroup of $G$ with $3^n$ elements
such that there is no subgroup of $G$ with $3^m$ elements,
$m > n$,
that contains $S$.
For instance,
the $\mathbbm{Z}_3$ subgroup of $D_3$
generated by the matrix $C^{-1}$ in~\eqref{upiy} is a Sylow 3-subgroup of $D_3$;
indeed,
the group $D_3$ has six elements and therefore any three-element subgroup of it
is necessarily a Sylow 3-subgroup.
Since we already know that $\mathbbm{Z}_3$ is a normal subgroup of $D_3$,
we conclude that $\mathbbm{Z}_3$ is a normal Sylow 3-subgroup of $D_3$.
The conjecture mentioned in section~\ref{searches} then informs us that
$D_3$ does \emph{not}\/ have a faithful three-dimensional irrep;
this is indeed true.
In the same way,
a normal Sylow 3-subgroup of a group $G$ of order $1536 = 3 \times 2^9$
is just a normal $\mathbbm{Z}_3$ subgroup of $G$.
A different example is the group $A_4$ generated by
\begin{equation}
\label{uoiyp}
D = \left( \begin{array}{ccc} 1 & 0 & 0 \\ 0 & -1 & 0 \\ 0 & 0 & -1
\end{array} \right)
\quad \mbox{and} \quad
E = \left( \begin{array}{ccc} 0 & 1 & 0 \\ 0 & 0 & 1 \\ 1 & 0 & 0
\end{array} \right).
\end{equation}
The group $A_4$ has $12 = 3 \times 2^2$ elements
and therefore its $\mathbbm{Z}_3$ subgroup formed by $E$,
$E^2$,
and the unit $3 \times 3$ matrix is automatically a Sylow 3-subgroup.
However,
it is not a normal subgroup,
because the set $\left\{ D E, D E^2 \right\}$
does not coincide with the set $\left\{ E D, E^2 D \right\}$.
Thus,
$A_4$ does not have a normal Sylow-3 subgroup
and it is allowed by the conjecture of section~\ref{searches}
to have a faithful three-dimensional irrep;
this is indeed the representation generated by the matrices in~\eqref{uoiyp}.
\end{appendix}
\paragraph{Acknowledgements:}
We gratefully acknowledge the collaboration of Pa\-trick Otto Ludl
in the beginning of this work;
we also thank him for valuable discussions all along.
The work of D.J.\ was supported by the Lithuanian Academy of Sciences
through project DaFi2016.
The work of L.L.\ was supported by
the Portuguese \textit{Funda\c{c}\~ao para a Ci\^encia e a
Tecnologia} through the projects CERN/FIS-NUC/0010/2015
and UID/FIS/00777/2013,
which are partially funded by POCTI (FEDER),
COMPETE,
QREN,
and the European Union.
|
\section{Introduction}
One of the main interests of relativistic heavy ion collision experiments is to explore the phase structure of the QCD matters. The QCD phase structure can be displayed in the two dimensional QCD phase diagram, in which the temperature $T$ is plotted as a function of the baryon chemical potential $\mu_\text{B}$. Lattice QCD calculations show that the transition from Quark-Gluon Plasma (QGP) to hadronic phase is a crossover transition~\cite{crossover} when $\mu_\text{B}=0$, and the QCD based model
calculations predict the transition at large $\mu_\text{B}$ is of the first order~\cite{firstorder}. Thus, the crossover region and first order region should be connected by a so called QCD critical point~\cite{QCP_Prediction,science,location}. Due to sign problem in Lattice QCD calculations at finite baryon density, there are large uncertainties to determine the location of critical point by theoretical and/or QCD based model calculations~\cite{location,qcp_Rajiv}. The fluctuation of conserved quantities, such as net-baryon number, and its proxy observable net-proton number, served as the observable sensitive to the correlation length of nuclear matter~\cite{ejiri2006hadronic,qcp_signal,Neg_Kurtosis,Asakawa}, have been extensively studied experimentally~\cite{2010_NetP_PRL,netcharge_PRL,STAR_BES_PRL} and theoretically~\cite{HRG_Karsch,PBM_netpdis,Lattice,2015_JianDeng_fluctuation,2014_Bengt_flu,Asakawa_formula,BFriman_EPJC,2015_Swagato_evolution,2015_Vovchenko,baseline_PRC,huichao,HRG_Nahrgang,kenji_morita,freezeout}.
\begin{figure}[!htb]
\hspace{-0.5cm}
\includegraphics[width=2.4in]{KVnetp}
\vspace{-0.5cm}
\caption{Energy dependence of cumulants ratios ($\kappa \sigma^{2}$) of net-proton distributions from RHIC beam energy scan measured by STAR experiment. The dashed shadow band represents the results from UrQMD calculations. Figure is obtained from~\cite{luo2015exploring}.}
\label{fig:kvsd}
\end{figure}
In recent experimental search for the critical point by the STAR experiment at RHIC, the forth order net-proton fluctuations ($\kappa \sigma^{2}$) displayed a non-monotonic behavior in the energy dependence~\cite{2014_Luo_CPOD,luo2015exploring}. One of the most striking observation is the large deviation above unity of the net-proton $\kappa \sigma^{2}$ in the most central Au+Au collisions at $\sqrt{s_{\text{NN}}} = 7.7\,\text{GeV}$, which is consistent with receiving strong positive critical contribution predicted by theoretical calculations. This observation cannot be described by the UrQMD and AMPT model~\cite{kurtosis_nahrgang,luo2014baseline,urqmd,Jan_kurtosis,UrQMDMoments}, both of which are transport model without mean field and QCD phase transition. This motivates us to investigate whether the large increase in the forth order net-proton fluctuations ( $\kappa \sigma^{2}$) are caused by non-critical contributions, such as mean field potentials, which may play an important role at low collision energies and high baryon density region. Theoretical model estimated that the size of the correlation length near the critical point is about 2$\sim$3 fm~\cite{2to3fm} in high energy heavy-ion collisions. One should note that due to the collisions and freeze-out dynamics, even the system indeed passes through the critical region, the critical behavior may not survive in the final observables. It is therefore important to discuss non-critical behavior in order to understand the experimental observations. On the other hand, it is also important to understand the possible experimental effects such as the efficiency correction and associated systematic errors~\cite{volker_eff}, which are still under studied and not discussed in this paper.
To study the effects of the mean field potential and softening of equation of state (EoS), we utilized a hadronic transport (JAM) model~\cite{nara1999relativistic} to simulate the net-proton (baryon) fluctuations in Au+Au collisions at $\sqrt{s_{\text{NN}}} = 5\,\text{GeV}$, which is dedicated center of mass energy covered by the future fixed target heavy-ion collisions program in the Compress Baryonic Matter (CBM) experiment~\cite{CBM_experiment} at GSI, Germany and J-PARC experiment~\cite{JPARC} at JAEA/KEK, Japan. The mean field potential and the effects of softening of EoS have been added in Monte Carlo simulation. This studies will allow us to understand the contribution from these two effects to the fluctuations of net-proton (baryon) in heavy-ion collisions. It will provide a baseline and reveal the non-critical background contribution for the baryon number fluctuations when searching for the QCD critical point in heavy-ion collisions.
The paper is organized as follows. We provide the definition of the cumulant observables in the section II. In the section III, we briefly discuss the JAM model with the mean field potential and the attractive scattering orbit, which is used to simulate the effects of the softening of EoS. Then, the results and discussion will be given in the section IV. Finally, we will summarize our studies in the section V.
\section{Observables}
Fluctuations of net-proton (baryon) can be characterized by their cumulants of the event-by-event multiplicity distribution. It can be computed as
\begin{align}
C_1 &= \left<N\right> \\
C_2 &= \left<(\delta N)^2\right> \\
C_3 &= \left<(\delta N)^3\right> \\
C_4 &= \left<(\delta N)^4\right> - 3\left<(\delta N)^2\right>^2
\end{align}
Where $N$ is net-proton (baryon) number and the $\left<N\right>$ is average over events, $\delta N = N-\left<N\right>$. In the STAR experiments, the fluctuations of net-proton number is usually used as a proxy of net-baryon number fluctuations, for the reason that neutral baryons, like neutrons are invisible to the detector. With the definition of cumulants, we can also define mean ($M$), variance ($\sigma^{2}$), skewness ($S$) and kurtosis ($\kappa$) as:
\begin{eqnarray}
M = C_{1}, \sigma^2 = C_{2}, S=\frac{C_{3}}{(C_{2})^\frac{3}{2}}, \kappa = \frac{C_{4}}{(C_{2})^2}
\end{eqnarray}
In addition, the moments product $\kappa\sigma^2$ and $S\sigma$ can be expressed in terms of the ratios of cumulants:
\begin{eqnarray}
\label{eq6} \kappa\sigma^2 = \frac{C_{4}}{C_{2}}, S\sigma = \frac{C_{3}}{C_{2}}, \sigma^{2}/M=\frac{C_{2}}{C_{1}}
\end{eqnarray}
The ratios of cumulants are intensive variables, which is independent on system volume. The statistical errors of those cumulants and cumulants ratios are evaluated
by the Delta theorem~\cite{Delta_theory,Unified_Errors}. In general, the statistical uncertainties strongly depend on the shape of the distributions, especially the width of the distributions. For e.g.,
the statistical errors of various order cumulants ($C_{n}$) can be approximated as $error(C_{n}) \propto \sigma^{n}/(\sqrt{N} \epsilon^{n})$, where $\sigma$ is the measured width of the distribution, $N$ represents the number of events and $\epsilon$ is
the particle detection efficiency.
\section{JAM Model}
The JAM (Jet AA Microscopic Transportation Model)~\cite{nara1999relativistic} is a hadronic transport model providing a useful tool to study the heavy ion collision from 100A MeV to RHIC energies. The $hh$ inelastic collision part is modeled by hadronic resonance productions at low energies and above this resonance region, the string excitation is adapted. The description of interacting $N$-body system of JAM comes from Relativistic QMD (RQMD) approach ~\cite{isse2005mean, maruyama1995relativistic, maruyama1996relativistic}, which regarded as a hadronic transport model. A simplified version of RQMD (RQMD/S) that focuses in effective using of CPU was adapted in JAM. The Hamiltonian of this $N$-body system has been derived to be
\begin{equation}
H \approx \sum_{i=1}^N u_i(p_i^2 - m_i^2 -2m_i V_i)
\end{equation}
where $u_i=1/2p_i^0$, $p_i^0 = \sqrt{\boldsymbol{p}_i^2 + m_i^2 +2m_i V_i}$. In the mean field case, a Skyrme type density dependent and Lorentzian-type momentum dependent potential was introduced. The total potential energy of the system has the form
\begin{align}
V &=
\int \mathrm{d}\boldsymbol{r}\,
\left[
\alpha\frac{\rho^2(\boldsymbol{r})}{2\rho_0}
+
\beta \frac{\rho^{\gamma+1}(\boldsymbol{r})}{(1+\gamma)\rho_0^\gamma}
\right]
\nonumber \\
&+
\sum_{k=1,2}{\frac{C_k}{2\rho_0}}
\int \mathrm{d}\boldsymbol{r}\,\mathrm{d}\boldsymbol{p}\,\mathrm{d}\boldsymbol{p}'
\frac{f(\boldsymbol{r},\boldsymbol{p})f(\boldsymbol{r},\boldsymbol{p}')}{1+[(\boldsymbol{p}-\boldsymbol{p}')/\mu_k]^2}
\end{align}
where $\rho(\boldsymbol{r})$ is the baryon density distribution. $\alpha$, $\beta$, $\gamma$, $\rho_0$, ${C_k}$, $\mu_k$ are empirical parameters same as reference \cite{nara2015does}.
The effect of the softening of EoS has been investigated in the paper \cite{nara2016directed}. Accompanying with the occurrence of EoS softening, the system usually experienced a first order phase transition\cite{rischke1995phase}. In the JAM model, extra pressure of the system in addition to the free streaming is obtained by two-body scattering given by the virial theorem \cite{nara2016directed}. The attractive orbit in two-body scattering is introduced in the JAM model to simulate the effect of the softening of EoS. By doing this, the pressure of the system can be reduced and the negative slope of rapidity dependence of directed flow are successfully reproduced.
In this work, data of Au+Au collisions at $\sqrt{s_{\text{NN}}} = 5 \,\text{GeV}$ were generated with JAM model in the three modes, which are cascades, mean field potential and attractive scattering orbit, respectively.
\begin{figure}[htb]
\includegraphics[height=2.8in]{fdndy}
\includegraphics[height=3in]{Edn}
\caption{Rapidity distribution (top) of net-proton (baryon), and event-by-event net-proton (baryon) distributions (bottom) from JAM model simulation for three different modes.}
\label{fig:fdndy}
\end{figure}
\section{Results}
The calculation results of net-proton (baryon) cumulants would be modified by including the mean field potential in cascade model. This additional mean field potential might have significant effects on the fluctuations of particle multiplicity at low collision energies, and the rapidity distribution of particles in final state. Since the event-by-event net-proton (baryon) cumulants are calculated within certain rapidity window or transverse momentum cuts in order to archive the condition of grand canonical ensemble. It is reasonable to postulate that the calculation results may depend on the types of potentials that applied to the Monte Carlo simulation.
\begin{figure*}[htb]
\hspace{-0.7cm}
\includegraphics[height=3. in]{Pcum2y}
\hspace{1.2cm}
\includegraphics[height=3.in]{Bcum2y}
\caption{Rapidity dependence for the cumulants ($C_1 \sim C_4$) of net-proton (left) and net-baryon (right) distributions in the most central Au+Au collisions from JAM model with cascade, mean field potential and attractive scattering orbit, respectively.
The dashed horizontal lines are with the value of zero. }
\label{fig:cul2y}
\end{figure*}
\begin{figure*}[htb]
\includegraphics[height=2.6in]{r21sd2y}
\hspace{-1.cm}
\includegraphics[height=2.6in]{r31kv2y}
\caption{Rapidity dependence for the Cumulants ratios of net proton and net baryon multiplicity distributions in Au+Au collisions at $\sqrt{s_{\text{NN}}} = 5\,\text{GeV}$ GeV from JAM model computed in the three different modes. In the left shows $\sigma^{2}/M$ ($C_2/C_1$) and $C_3/C_1$. The figure in the right shows $S\sigma$ ($C_3/C_2$) and $\kappa\sigma^2$($C_4/C_2$) . The dashed horizontal lines are with the value of unity.} \label{fig:ratios}
\end{figure*}
To avoid the effects of auto-correlation, the collision centralities in the simulation are determined by the multiplicities of charged pion and kaon within pseudo-rapidity $|\eta|<1$ (refmult3) in Au+Au collisions, which is also applied for the data analysis of the net-proton fluctuation in the STAR experiment~\cite{STAR_BES_PRL}. In our study, we will focus on discussing the results from the 0-5\% most central Au+Au collisions, which is defined by the top 5\% fraction of the refmult3 multiplicity distribution for the minimum-biased event ensemble. The volume fluctuations are suppressed by applying the so called centrality bin width correction~\cite{technique}. In the top of the figure~\ref{fig:fdndy}, we show the rapidity distributions ($\text{d}N/\text{d}y$) for the net-proton (baryon) of the most central (0-5\%) Au+Au collisions from JAM cascade, mean field potential and attractive scattering orbit, respectively. The ratios between cascade and the other two cases are displayed in lower panels, respectively. It shows that both the mean field potentials and attractive scattering orbit have strong impacts on the $\text{d}N/\text{d}y$ distributions for net-proton (baryon). Due to the nucleon repulsive potentials, the results from mean field yield less stopping thus wider $\text{d}N/\text{d}y$ distributions than cascade. It is observed that the magnitude of the $\text{d}N/\text{d}y$ from mean field is smaller than the cascade around mid-rapidity ($|y|<0.5$). On the other hand, the attractive orbit scattering results in significant enhancement with respect to the cascade at mid-rapidity and a narrower $\text{d}N/\text{d}y$ distributions. The enhancement of the $\text{d}N/\text{d}y$ at mid-rapidity for attractive orbit can be attributed to the reduction of the pressure of the system and thus stronger nucleon stopping. We also have studied the event-by-event net-proton (baryon) distributions at mid-rapidity for central Au+Au collisions with JAM model in the three modes. The net-proton (baryon) distributions show similar shape for cascade and mean field cases, while the net-proton (baryon) distributions calculated from attractive scattering orbit case show larger mean values comparing with the distributions from cascade and mean field data.
\begin{figure*}[htb]
\hspace{-0.5cm}
\includegraphics[height=4.3in]{ProPtDep}
\caption{Rapidity dependence for the cumulants ratios of net-proton distributions in Au+Au collisions at $\sqrt{s_{NN}}=5$ GeV from JAM model computed in the three different modes and various transverse momentum ranges. From top to bottom are $\sigma^{2}/M$ ($C_2/C_1$), $C_3/C_1$, $S\sigma$ ($C_3/C_2$) and $\kappa\sigma^2$($C_4/C_2$), respectively. The dashed horizontal lines are with the value of unity.}
\label{fig:pt}
\end{figure*}
Figure~\ref{fig:cul2y} displays rapidity dependence for the cumulants (up to forth order) of net-proton and net-baryon distributions from JAM model computed with three different modes. The rapidity coverage $\Delta y$ of particles are centered at zero and the rapidity cut in the analysis is $|y|< \Delta y/2$. It is found that the net-proton (baryon) cumulants for the three modes show similar trends. Since cumulants are sensitive to number of particles in analyze, the net-baryon cumulants show larger values than that of net-protons. With increasing the rapidity acceptance $\Delta y$, cumulants are rising linearly, for the reason that they are proportional to the volume of the system. The further increasing on $\Delta y$ will lead to significant suppressions due to the effects of baryon number conservation.
Figures~\ref{fig:ratios} shows cumulant ratios of net-proton (baryon) distributions in Au+Au collisions at $\sqrt{s_{\text{NN}}} = 5 \,\text{GeV}$ from JAM model. Those ratios of cumulants of net-proton (baryon) distributions are constructed to eliminate the volume dependence and can be used to compare with the theoretical calculations. When increasing the rapidity acceptance ($\Delta y$), the net-proton (baryon) cumulant ratios will decrease, reach a minimum and then increase, which is the typical effects of baryon number conservation~\cite{bzdak2013baryon}. For different net-proton (baryon) cumulant ratios, the position of the minimum are different. It indicates the mean field potential and softening of EoS will not lead to large increase above unity for the net-proton (baryon) cumulants ratios. Instead, due to the baryon number conservation, large suppression for the fluctuations of net-proton (baryon) are observed. The rapidity dependence for the cumulants ratios calculated from the three modes are with the similar trend. It suggests that the observed similar trends obtained by JAM model without implementing critical physics are dominated by the effects of baryon number conservation. On the other hand, one observes that the net-baryon cumulant ratios show larger suppression with respect to unity than the net-proton and the higher order cumulant ratios also show larger suppression than the lower order. Since the mean field potential implemented in the JAM model is momentum dependent, it is important to study the momentum dependence for the cumulants of net-proton distributions. In Fig. \ref{fig:pt}, for different transverse momentum range, we plot the cumulant ratios of net-proton distributions as a function of rapidity coverage, which are calculated with the three different modes in the JAM model. The results computed from different modes are with the similar trends. When the $p_{T}$ coverage is enlarged, the cumulant ratios are suppressed with respected to unity, the Poisson expectations. When the $p_{T}$ range is small, the fluctuations are dominated by Poisson statistics and the cumulant ratios are close to the unity. We also noticed the recent study for baryon number fluctuations due to mean field effects. This was done with a Relativistic Mean Field (RMF) approach~\cite{Kenji_vector}. Both scaler field ($\sigma$), vector field ($\omega$) as well as the liquid-gas phase transition are included in the RMF calculations for a static nuclear medium. As a result, it is found that the $\kappa \sigma^{2}$ is suppressed by the vector field ($\omega$) while an enhancement, due to the liquid-gas phase transition, was identified at a low temperature $T$=21 MeV and large baryon chemical potential $\mu_{B}$= 906 MeV. On the other hand, here we employed a full dynamic approach with a momentum dependent scaler mean field. Relative to the cascade mode, we do not observe any further suppression from the scaler mean field. In the future, it would be interesting to include the vector field in our dynamic calculations for baryon number fluctuations.
\section{Summary}
With the JAM model, we studied the rapidity and transverse momentum dependence for the cumulants and cumulants ratios of net-proton (baryon) distributions in 0-5\% most central Au+Au collisions at $\sqrt{s_{\text{NN}}} = 5 \,\text{GeV}$. The model simulation is performed with three different modes, which are cascade, mean field and attractive scattering orbit, respectively. The results from three different modes show similar suppression trends, when enlarging the rapidity and/or transverse momentum coverage. Those suppressions in net-proton(baryon) cumulant ratios are expected for the conservation of baryon numbers in strong interactions. Therefore, it indicates that the significant increasing of the $\kappa \sigma^{2}$ of net-proton distribution above unity observed in Au+Au collisions at $\sqrt{s_{NN}}=7.7$ GeV measured by the STAR experiment, cannot be explained by the JAM model with effects of the momentum dependent scaler mean field potential or softening of EoS. This study can provide us important information about the non-critical contribution to the fluctuations of net-proton (baryon) and baseline for the QCD critical point search in the large net-baryon density region.
\section*{Acknowledgments}
The work was supported in part by the MoST of China 973-Project No.2015CB856901, NSFC under grant No. 11575069, 11221504.
|
\section{Introduction}
The observational evidence for an accelerated cosmic expansion was first given by
Type Ia Supernovae (SNIa) \citep{Riess1998, Perlmutter1999}. Since then, measurements of the
cosmic microwave background (CMB) anisotropies
\citep[e.g.][]{Jaffe2001, Pryke2002, Spergel2007, Planck2015} and of
Baryon Acoustic Oscillations (BAOs) \citep[e.g.][]{Eisenstein2005}, in
combination with independent Hubble parameter measurements
\citep[e.g.][]{Freedman2012}, have provided ample evidence of the
presence of a dark energy (DE) component in the Universe.
To the present day, the main geometrical tracer of the cosmic
acceleration has been SNIa at redshifts $z \lesssim 1.5$
\citep[e.g.][]{Suzuki2012, Betoule2014}. It is of great
importance to use alternative geometrical probes at higher redshifts
in order to verify the SNIa results and to obtain
more stringent constrains in the cosmological parameters solution
space \citep{Plionis2011}, with the final aim of discriminating among the various theoretical
alternatives that attempt to explain the accelerated expansion of the
Universe \citep[cf.][]{Suyu2012}.
The $L(\mathrm{H}x) - \sigma$\ relation between the velocity dispersion ($\sigma$) and
Balmer-line luminosity ($L[\mathrm{H}x]$, usually H$\beta$) of H\,\textsc{ii}\ galaxies has
already proven its potential as a cosmological tracer
\citep[e.g.][and references therein]{Melnick2000, Siegel2005, Plionis2011,
Chavez2012, Chavez2014, Terlevich2015}. It has been shown that the
$L(\mathrm{H}\beta) - \sigma$\ relation can be used in the local Universe to constrain the
value of $H_0$ \citep{Chavez2012}. At high-$z$ it can set constraints
on the parameters of the DE Equation of State (EoS) \citep{Terlevich2015}.
H\,\textsc{ii}\ Galaxies are a promising tracer for the parameters of
the DE EoS precisely because they can be observed, using the current available
infrared instrumentation, up to $z \sim 3.5$ \citep[cf.][]{Terlevich2015}. Even when their scatter
on the Hubble diagram is about a factor of two
larger than in the case of SNIa, this disadvantage is
compensated by the fact that H\,\textsc{ii}\ galaxies are observed to much larger
redshifts than SNIa where the degeneracies for different
DE models are substantially reduced \citep[cf.][]{Plionis2011}.
In addition, because the $L(\mathrm{H}\beta) - \sigma$\ relation systematic uncertainty sources\citep{Chavez2012, Chavez2014}
are not the same as those of SNIa, H\,\textsc{ii}\ galaxies constitute
an important complement to SNIa in the local
Universe, contributing to a better understanding of the systematic
errors of both empirical methods.
In this paper we perform an HII/BAO/CMB joint likelihood analysis and
compare the resulting cosmological constraints with those of
a BAO/CMB and a SNIa/BAO/CMB joint likelihood analysis (for the
latter we use the \emph{Union 2.1} SNIa compilation \citep{Suzuki2012}).
Furthermore, we present extensive Monte-Carlo simulations,
tailored to the specific uncertainties of the H\,\textsc{ii}\ galaxies
$L(\mathrm{H}\beta) - \sigma$\ relation and currently available instrumentation, to
demonstrate its potential possibilities as a cosmological tracer
to $z\lesssim 3.5$, to
probe a region where the Hubble function is very sensitive to the
variations of cosmological parameters \citep{Melnick2000,
Plionis2011}.
The paper is organised as follows: in section 2 we succinctly describe the data used
and associated systematic uncertainties; cosmological constraints
that can be obtained from the data are explored in section 3; in
section 4 we discuss the Monte-Carlo simulations, in section
5 we discuss the planned data acquisition in order to
obtain better constraints on the cosmological parameters.
Finally in section 6 we present our
conclusions.
\section{H II galaxies data}
Our current sample consists of a low-$z$ subsample
of 107 H\,\textsc{ii}\ galaxies ($0.01 \leq z \leq 0.16$) extensively analysed in
\citet{Chavez2014} and 24 Giant Extragalactic H\,\textsc{ii}\ Regions (GEHR) at
$ z \leq 0.01$ described in \citet{Chavez2012}.
The sample also includes a high-$z$ subsample composed by 6 star-forming
galaxies, selected from \citet{Hoyos2005, Erb2006b, Erb2006} and
\citet{Matsuda2011}, that we observed \citep{Terlevich2015} using
X-SHOOTER \citep{Vernet2011} at the Very Large Telescope in Paranal. The data of
19 objects taken from
\citet{Erb2006, Maseda2014} and \citet{Masters2014} complete the
sample. Altogether, the redshift
range covered by the high-$z$ subsample is
$0.64 \leq z \leq 2.33$.
It has been demonstrated \citep[cf.][]{Terlevich1981, Melnick1988,
Terlevich2003, Plionis2011, Chavez2012, Chavez2014} that the
$L(\mathrm{H}\beta) - \sigma$\ relation for H\,\textsc{ii}\ galaxies and GEHR can be used to measure
distances via the determination of their Balmer emission line
luminosity,
$L(\mathrm{H}\beta)$, and the velocity dispersion ($\sigma$) of the young starforming
cluster from
measurements of the line width. The relevant relation can be expressed as:
\begin{equation}
\log L(\mathrm{H}\beta) = (5.05 \pm 0.097) \log \sigma (\mathrm{H}\beta) + (33.11 \pm 0.145) .
\label{eq:LS}
\end{equation}
Distance moduli are then obtained from:
\begin{equation}
\mu^{o} = 2.5 \log L(\mathrm{H}\beta)_{\sigma} - 2.5 \log f(\mathrm{H}\beta) - 100.195
\label{eq:mu1}
\end{equation}
where $L(\mathrm{H}\beta)_{\sigma}$ is the luminosity estimated
from the $L(\mathrm{H}\beta) - \sigma$\ relation as in eq. (\ref{eq:LS}) and $f(\mathrm{H}\beta)$ is the measured flux in the H$\beta$ line. The uncertainty on
the distance moduli, $\sigma_{\mu^{o}}$, is propagated from the
uncertainties in $\sigma_i$ and $f_i$ and the slope and intercept of
the distance estimator in eq. (\ref{eq:LS}).
\subsection{Systematic Errors}
\begin{table}
\center{
\caption {Systematic error budget on the distance moduli, $\mu$. The typical uncertainty
contribution of each source of systematic error is given. }
\tabcolsep 15pt
\begin{tabular}[h] { l l l }
\hline
\hline
Source & Error \\
\hline
Size of the Burst & 0.175 \\
Age of the burst & 0.05 \\
Abundances & 0.05 \\
Extinction & 0.175 \\
\hline
Total & 0.257 \\
\hline
\label{tab:SEB}
\end{tabular}
}
\end{table}
\subsubsection{Size of the burst}
The scatter found in the $L(\mathrm{H}\beta) - \sigma$\ relation for
H\,\textsc{ii}\ galaxies suggests a dependence on a second parameter
\citep[cf.][]{Terlevich1981, Melnick1987}. Indeed
\citet{Chavez2014}, using SDSS DR7 effective Petrosian radii,
corrected for seeing, for a sample of local H\,\textsc{ii}\ galaxies, found
the size of the starforming region to be this second parameter.
For the high-$z$ samples, unfortunately, we do not have any
size measurements, so using it as a second parameter in the
correlation is impossible. The error induced by not using the size of
the burst as a second parameter appears in the uncertainties in the
slope and zero point of the $L(\mathrm{H}\beta) - \sigma$\ relation, i.e.\ our
uncertainty values already incorporate this effect. In Table
\ref{tab:SEB} we show the typical contribution of the size of the
burst to the uncertainty on the distance moduli.
\subsubsection{Age of the burst}
\citet{Melnick2000} have demonstrated that H\,\textsc{ii}\ galaxies with equivalent width of H$\beta$, $W(\mathrm{H}\beta)
< 25\ \mathrm{\AA}$, do follow an $L(\mathrm{H}\beta) - \sigma$\ relation with a similar slope
but different intercept than those with larger $W(\mathrm{H}\beta)$,
i.e.\ older starbursts follow a parallel less luminous
$L(\mathrm{H}\beta) - \sigma$\ relation. For the study presented here, the starburst age is a
controlled parameter in the sense that we have selected our sample to
be composed of very young objects \citep[$\lesssim 5$ Myr, for
instantaneous burst models cf.][]{Leitherer1999} by putting a high
lower limit to the value of $W(\mathrm{H}\beta) >
50\ \mathrm{\AA}$. Therefore only the youngest bursts were considered
and in this way the effects of the age of the burst as a systematic
error on the $L(\mathrm{H}\beta) - \sigma$\ relation has been minimised; this selection also minimises
the contamination by an older underlying stellar component.
We have demonstrated \citep{Chavez2014} that using the $W(\mathrm{H}\beta)$ as a
second parameter in the $L(\mathrm{H}\beta) - \sigma$\ correlation reduces only slightly the
scatter because of the small dynamic range of the age of our sample
objects. We chose \citep{Terlevich2015}
not to use the $W(\mathrm{H}\beta)$ as a parameter to
`correct' the $L(\mathrm{H}\beta) - \sigma$\ relation. Therefore, the small effect of the age
of the burst on the correlation manifests itself in the uncertainties of
the slope and zero point of the $L(\mathrm{H}\beta) - \sigma$\ relation that we are
adopting. In Table \ref{tab:SEB} we show the typical contribution of
the age of the burst to the uncertainty on the distance moduli.
\subsubsection{Abundances}
The oxygen abundance of H\,\textsc{ii}\ galaxies was considered in the past
\citep[eg.][]{Melnick1987, Melnick2000, Siegel2005} as a second
parameter for the $L(\mathrm{H}\beta) - \sigma$\ relation.
We have explored again this issue in
\citet{Chavez2014} for our local sample and
concluded that the effect albeit present is very small.
We chose \citep{Terlevich2015} not to use the oxygen
abundance as a parameter to `correct' the $L(\mathrm{H}\beta) - \sigma$\ relation, and thus
the small effect of the metallicity of the burst on the correlation is
already part of the uncertainties of the slope and zero point of the
$L(\mathrm{H}\beta) - \sigma$\ relation that we are adopting. The typical contribution of the abundances to the uncertainty on the
distance moduli is shown in Table \ref{tab:SEB}.
\subsubsection{Extinction}
The internal extinction correction was performed on the low-$z$
subsample following the procedure described in \citet{Chavez2014} and
using the extinction coefficients derived from SDSS DR7 spectra. For
the high-$z$ subsample
we used the extinction coefficients given in the literature
\citep{Erb2006b, Erb2006, Matsuda2011, Maseda2014,
Masters2014}. Typical contribution of the
extinction to the distance modulus uncertainty is also shown in
Table \ref{tab:SEB}.
\subsubsection{Malmquist bias}
The Malmquist bias is a selection effect in flux limited samples.
Due to the preferential detection of the most luminous objects
as a function of distance and limiting flux, at any distance
there are always more
faint objects being randomly scattered-in of the flux-limited
sample than bright
objects being randomly scattered-out of the sample. Therefore the
source mean
absolute magnitude at some large distance will be systematically
fainter than what expected due to the flux limit of the catalogue at
that distance.
The Malmquist bias for our flux limited low-$z$ calibrating sample was calculated
following the procedure given by \citet{Giraud1987}. In the first place, using the Luminosity Function for H\,\textsc{ii}\ galaxies \citep{Chavez2014} we estimated the expected value of the
luminosity at any redshift as:
\begin{equation}
\langle L \rangle = \frac{\int_{L_i}^{L_s}{L^{\alpha} L dL}}{\int_{L_i}^{L_s}{L^{\alpha} dL}} ,
\end{equation}
where $L_i = 10^{39.7}$ is the lower limit of the luminosity function,
$L_s = 10^{42.5}$ is the upper limit and $\alpha = -1.5$ is the slope
\citep{Chavez2014}.
Subsequently, at each $z$ we calculate the luminosity expected when
we change the lower limit of the Luminosity Function to the value
given by the flux limit at that redshift:
\begin{equation}
\langle L(z) \rangle = \frac{\int_{L_l(z)}^{L_s}{L^{\alpha} L dL}}{\int_{L_l(z)}^{L_s}{L^{\alpha} dL}} ,
\end{equation}
where the value of $L_l(z)$ can be calculated from:
\begin{equation}
\log L_l(z) = \log f_l + 2\log(d_L[z, \mathbf{p}]) + 50.08 ,
\end{equation}
where $\log f_l = -14.3$ is the flux limit of our low-$z$ sample and $d_L$ is the luminosity distance as function of $z$ and a set of
cosmological parameters $\mathbf{p}$.
Finally the bias is a function of the difference of the unbiased and biased expected values of the luminosity and can be obtained as:
\begin{equation}
b( \log L_{\mu}) = \frac{\sigma_0^2}{\sigma_{L_0}^2 + \sigma_0^2} (\log \langle L \rangle - \log \langle L(z) \rangle)
\end{equation}
where $\sigma_0$ is the
dispersion of residuals of the $L(\mathrm{H}\beta) - \sigma$\ relation and $\sigma_{L_0}$ is
the dispersion of the distribution of luminosities in the
sample. From the above equation the bias for a certain distance modulus can be obtained as $b(\mu) = 2.5 b( \log L_{\mu})$.
The typical value of the Malmquist bias found for our low-$z$
calibrating sample is $b(\mu) = 0.03$, extremely small compared to
the other uncertainties.
\subsection{Gravitational Lensing Effects}
Details of the expected effect of
gravitational lensing on the distance modulus of high-$z$ standard
candles \citep[e.g.][and references therein]{Holz1998, Holz2005,
Brouzakis2008} were given in \citet{Plionis2011}. The basic
assumption used in developing a
correction procedure for this effect is that the magnification distribution resembles a
lognormal with zero mean (the mean flux of each source
over all possible different paths is conserved, since lensing does not
affect photon numbers),
a mode shifted towards the de-magnified regime and a long tail towards
high magnification. This sort of distribution has been found in
analyses based on Monte-Carlo procedures and ray-tracing techniques \citep[cf.][]{Holz2005}.
Therefore most high-$z$ sources will be demagnified (will appear artificially fainter), inducing an apparently
enhanced accelerated expansion, while a few will be highly magnified. The effect is obviously
stronger for higher redshift sources since the lower the redshift the smaller the optical
depth of lensing.
It is important to note that the effect of gravitational lensing is
not only to increase the distance modulus uncertainty, which is
proportional to the redshift, but
also to induce a systematic shift of the mode of the
distance modulus distribution to de-magnified (fainter) values.
These effects appear to be independent of the underlying cosmology and
the details of the density profile of cosmic structures \citep[eg.][]{Wang2002}.
A procedure, first suggested by
\citet{Holz2005}, to correct statistically for such an
effect was explained in detail in \citet{Plionis2011}. The reader is referred to that work.
We apply this procedure to our analysis of the Hubble
expansion cosmological probe, using either H\,\textsc{ii}\ galaxies or SNIa, but
find minimal effects on the resulting cosmological parameter constraints.
\section{Cosmological Constraints}
A variety of observational probes have been developed through the years in order to
provide constraints on the cosmological parameters, which in turn
determine the specifics of the evolution of the Universe.
These probes may be divided in two general classes; {\em geometrical}
and {\em dynamical} and both use the redshift dependence of the comoving
distance to a source:
\begin{equation}\label{eq:Dc}
d_{C}(z)= \int_{0}^{z}{\frac{c dz^{'}}{H(z^{'})}}\;,
\end{equation}
where the Hubble function $H(z) [\equiv H_0 E(z)]$ is derived from the
first Friedman equation and $E(z)$ is given in the matter-dominated
era for a flat Universe with matter and DE, by:
\begin{equation}\label{eq:Ez}
E^2(z)= \left[ \Omega_{m,0}(1+z)^3 + \Omega_{w,0} (1+z)^{3y}
\exp\left( \frac{-3 w_a z}{z+1}\right) \right]
\end{equation}
with $y=(1 + w_0 + w_a)$.
The parameters $w_0$ and $w_a$ refer to the DE equation of
state, the general form of which is:
\begin{equation}
p_{w} = w(z) \rho_{w} \;,
\end{equation}
with $p_{w}$ the pressure and $\rho_{w}$ the density of the
postulated DE fluid.
Different DE models have been proposed and many are
parametrized using a Taylor expansion around the present epoch:
\begin{equation}
w(a) = w_0 + w_a(1-a)\Longrightarrow w(z)=w_0+w_a \frac{z}{1 + z}\;,
\end{equation}
\citep[CPL model;][]{Chevallier2001, Linder2003, Peebles2003,
Dicus2004, Wang2006}.
The cosmological constant is just a special case of DE, given
for $(w_0,w_a)=(-1,0)$, while the so called {\em quintessence} (QDE) models
are such that $w_a=0$ but $w_0$ can take values $\neq -1$.
Therefore, assuming a flat Universe ($\Omega_{m}+\Omega_{w}=1$),
a negligible radiation density
parameter and the generic CPL DE EoS parametrisation,
the most general set of cosmological parameters that is
necessary to be constrained in order to define the actual cosmological
model, is given by $\mathbf{p} = \{\Omega_{m,0}, w_0, w_a\}$. Note
that we do not include as a parameter the Hubble constant because, as
it will become clear further below, the dependence on $H_0$ is
factored out.
In what follows, we will consider two parametrisation of the DE EoS, assuming a flat Universe, i.e.,
\begin{enumerate}
\item QDE model with $\mathbf{p} = \{\Omega_{m,0}, w_0, 0\}$, and
\item CPL model with $\mathbf{p} = \{\Omega_{m,0}, w_0, w_a\}$
\end{enumerate}
The {\em geometrical} probes, which are independent of the underline
gravity theory,
are used to probe the Hubble function
through the redshift dependence of the luminosity, $d_{L}(z)$, or the angular
diameter, $d_{A}(z)$, distance.
These methods utilize extragalactic sources for which their luminosity
is either known a priori (e.g. standard candles) or it can be estimated by
using a distance-independent observational parameter. Alternatively, they can use cosmic
phenomena for which their metric size is known (e.g. standard rulers).
Then the cosmic
expansion history is traced via the luminosity distance $d_{L}(z)$, in
the first case, or the angular diameter distance $d_{A}(z)$, in the
second case. To date such observations probe
the integral of the Hubble expansion rate $H(z)$ either up to
redshifts of order $z\simeq 1.5$ (e.g., SNIa, BAO, clusters), or at
the redshift of recombination, $z_{rec}\sim 1100$ (CMB fluctuations).
{\em Dynamical} probes, on the other hand, map the expansion history
based on measures of the growth rate of cosmological perturbations
and therefore depend on the theory of gravity \citep[cf.][and
references therein]{Bertschinger2006, Nesseris2008,
Basilakos2013}. Such methods are also confined to
relatively low redshifts, up to $z\simeq 1$.
It is therefore clear
that the redshift range $1.5\lesssim z\lesssim 1000$ is not directly
probed to date by
any of the above cosmological tests, and as discussed in
\citet{Plionis2011} the redshift range $1.5\lesssim z\lesssim 3.5$ is
of crucial importance to constrain the DE EoS,
since different DE models manifest their largest
deviations in this redshift range.
Therefore the fact that H\,\textsc{ii}\ galaxies can be observed relatively
easily at such redshifts make them ideal and indispensable tools for
cosmological studies.
Below we present the basics of the two {\em geometrical} probes that are
extensively used to constrain the DE EoS parameters.
\subsection{Standard Candle Probes}
As discussed previously, for standard candle probes we need to use the luminosity
distance of the sources tracing the Hubble expansion, given by
$d_{L}=(1+z) d_C$.
For convenience, which will be understood below, we define a further
parameter, independent of the Hubble constant, by:
\begin{equation}
D_L(z,{\bf p})= (1+z) \int_{0}^{z}{\frac{dz^{'}}{E(z^{'},{\bf p})}}\;.
\end{equation}
i.e., $d_L=c D_L/H_0$.
Using the luminosity distance, as calculated
from a set of cosmological parameters, $\mathbf{p}$, and
the redshift, $z$, we can obtain the
`theoretical' distance modulus of a source as:
\begin{equation}
\mu_{th} = 5 \log d_L (\mathbf{p}, z) +25= 5 \log D_L (\mathbf{p}, z) + \mu_{0},
\label{eq:mu2}
\end{equation}
where $\mu_{0}=42.384-5\log h$.
Therefore, to restrict a given set of cosmological parameters, we
define the usual $\chi^2$ minimisation function as:
\begin{equation}
\label{eq:mu22}
\chi_{sc}^{2}({\bf p})=
\sum\limits_{i=1}^{N}\frac{[\mu_{\rm obs}(z_i)-\mu_{\rm th}(z_{i},{\bf p})]^2}
{\sigma_{\mu,i}^2}\;,
\end{equation}
where $N$ is the total number of sources used, the suffix $sc$
indicates the standard candle probe and
$\mu_{\rm obs}(z_{i})$ and $\sigma^{2}_{\mu,i}$ are the distance
moduli and the corresponding uncertainties at the observed redshift $z_{i}$.
Inserting the second equality of eq.(\ref{eq:mu2}) into eq.(\ref{eq:mu22})
we find after some simple algebra that
\begin{equation}\label{eq:expand-xi2}
\chi_{sc}^2({\bf p})=A({\bf p})-2B({\bf p})\mu_0+C\mu_0^2\;,
\end{equation}
where
\begin{eqnarray}\label{eq:xi2-A-B-C}
A({\bf p}) &=& \sum\limits_{i=1}^{N}\frac{[\mu_{\rm obs}(z_{i})-
5{\rm log}D_{L}(z_{i},{\bf p})]^2}{\sigma_{\mu,i}^2}\;,\nonumber \\
B({\bf p}) &=&\sum\limits_{i=1}^{N}\frac{\mu_{\rm obs}(z_{i})-5{\rm log}D_{L}(z_{i},{\bf p})}
{\sigma_{\mu,i}^2}\;,\nonumber \\
C &=& \sum\limits_{i=1}^{N} \frac{1}{\sigma_{\mu,i}^2}\;. \nonumber
\end{eqnarray}
Obviously for $\mu_{0}=B/C$ eq.(\ref{eq:expand-xi2})
has a minimum at
\begin{equation}
{\tilde \chi}^{2}({\bf p})=A({\bf p})-\frac{B^2({\bf p})}{C}.
\end{equation}
Therefore, instead of using
$\chi^{2}$ we now minimise ${\tilde \chi}^{2}$
which is independent of $\mu_{0}$ and thus of the value of the Hubble constant.
For more details concerning the above treatment
the reader is referred to \citet{Nesseris2005}.
\subsection{Standard Ruler Probes}
The first standard ruler probe is provided by
the first peak of the CMB temperature perturbation spectrum, appearing
at $l_1^{TT}$, which refers to the angular scale of the sound horizon at the last
scattering surface, $\theta_{1}^{TT}\sim 1/l_1^{TT}$. Then by
calculating its comoving scale, $r_s(z_{rec})$, we can derive its angular
diameter distance by:
\begin{equation}
d_A(z_{rec}, {\bf p})=\frac{r_s(z_{rec}, {\bf p})}{\theta_1^{TT}}=\frac{d_C(z_{rec},
{\bf p})}{1+z_{rec}}\;.
\end{equation}
Since the above equation is model dependent, through the CMB physics
determination of $r_s$, a model independent parameter has been
defined, the so-called {\em shift parameter} \citep{Bond1997,
Nesseris2007}, which is the ratio of the position of the first peak
to that of a reference model, and for spatially flat models it is
given by:
\begin{equation}\label{shiftparameter}
R({\bf p})=\sqrt{\Omega_{m,0}}\int_{0}^{z_{rec}} \frac{dz}{E(z, {\bf p})}\,.
\end{equation}
The observationally measured shift parameter, according to the recent
{\em Planck} data \citep{Shafer2014} is $R=1.7499\pm 0.0088$ at the
redshift of decoupling (viz. at the last scattering surface, $z_{rec}=1091.41$).
At this point we would like to remind the reader that
when dealing with the CMB
shift parameter we need to include also the radiation density
term in the $H(z)$ function since at recombination it amounts to $\sim
23\%$ of the matter density ($\Omega_{r,rec}\simeq 0.23 \Omega_{m,rec}$)
and therefore cannot be ignored. The final minimisation function is:
\begin{equation}
\label{eq:cmb}
\chi^{2}_{\rm CMB}({\bf p})= \frac{[R({\bf p})-1.7499]^2}{0.0088^2}\;.
\end{equation}
The second standard ruler probe that we use
is the Baryonic Acoustic Oscillation (BAO) scale, a feature produced in the last
scattering surface by the competition between the pressure of the
coupled baryon-photon fluid and gravity. The resulting sound
waves leave an overdensity
signature at a certain length scale of the matter distribution. This
length scale is related to the comoving distance that a sound
wave can travel until recombination and in practice it manifests
itself as a feature in the correlation function of galaxies on large
scales ($\sim 100 \;h^{-1}$ Mpc).
In recent years, measurements of the BAO have proven extremely useful as
a ``standard ruler''.
The BAOs were clearly identified, for the first time,
as an excess in the clustering pattern of the SDSS luminous red galaxies
\citep{Eisenstein2005}, and of the 2dFGRS galaxies \citep{Cole2005}.
Since then a large number of dedicated surveys have been used to
measure BAOs, among which the WiggleZ Dark Energy Survey
\citep{Blake2011}, the 6dFGS \citep{Beutler2011} and
the SDSS Baryon Oscillation Spectroscopic Survey (BOSS) of SDSS-III
\citep{Eisenstein2011,Anderson2014,Aubourg2015}.
In the current paper we utilise the results
of \citet[][see their Table 3]{Blake2011}
which are given in terms of the acoustic parameter $A(z)$,
first introduced by \citet{Eisenstein2005}:
\begin{equation}\label{defBAOA}
A({z_i,\bf p})=\frac{\sqrt{\Omega_{m,0}}}{[z_i^{2} E(z_{i}, {\bf p})]^{1/3}}
\left[\int_{0}^{z_i} \frac{dz^{'}}{E(z^{'}, {\bf p})} \right]^{2/3}
\end{equation}
with $z_i$ the redshift at which the signature of the acoustic oscillations has been
measured. The corresponding minimisation function is given by
\begin{equation}
\label{eq:BAO}
\chi^{2}_{\rm BAO}({\bf p})= \sum_{i=1}^{6}\frac{[A(z_i,{\bf p})-A_{obs,i}]^2}{\sigma_i^2}\;.
\end{equation}
where $A_{obs,i}$ are the observed $A_i$ values at six different redshifts, $z_i$,
provided in Blake et al. (2011).
\subsection{Joint Analysis of Different Probes}
In order to place tight constraints on the corresponding
parameter space of the DE EoS, the cosmological
probes described previously must be
combined through a joint likelihood analysis, given by the product
of the individual likelihoods according to:
\begin{equation}
{\cal L}_{\rm tot}({\bf p})= \prod_{i=1}^{n} {\cal L}_i({\bf p})
\label{eq:overalllikelihood}
\end{equation}
where $n$ is the total number of cosmological probes
used\footnote{Likelihoods are normalised to their maximum
values.}. This
translates to an addition for the corresponding joint total $\chi^2_{\rm tot}$
function:
\begin{equation}
\chi^{2}_{\rm tot}({\bf p})=\sum_{i=1}^{n}\chi^{2}_i({\bf p})
\;.
\label{eq:overalllikelihoo}
\end{equation}
In our current analysis we sample the cosmological
parameter space with the following resolution:
$\delta\Omega_{m,0}=0.001$, $\delta w_0=0.003$ and
$\delta w_a=0.016$. Also, the reported uncertainties for each
unknown parameter of the vector ${\bf p}$ are estimated after marginalising one
parameter over the other, such that $\Delta \chi^{2}(\le 2\sigma)$.
Note however that in order to appreciate the possible degeneracy among
the different fitted parameters one must inspect the two-dimensional
likelihood contours.
\begin{figure}
\mbox{\epsfxsize=8.2cm
\epsffile{fig/FIG1a-1.eps}}
\caption{Likelihood contours for $\Delta
\chi^2=\chi^{2}_{\rm tot}-\chi^{2}_{\rm tot,min}$ equal to 2.32 and 6.18
corresponding to the 1$\sigma$ and 2$\sigma$
confidence levels in the $(\Omega_{m,0},w)$ plane.
Results based on the H\,\textsc{ii}\ galaxies are shown in black, on the
CMB shift parameter (green) and on BAO (blue) while the joint contours
are shown in red.}
\label{fig:01}
\end{figure}
\begin{table*}
\caption{Cosmological parameters from the joint analysis of different
combinations of probes and for both parameterisations of the DE EoS.}
\label{tab:02}
\tabcolsep 12pt
\begin{tabular}{cccccc} \hline
Probes & $\Omega_{m,0}$ & $w_0$ & $w_a$ & $\chi^2_{min}$ & df
\\ \hline
& \multicolumn{5}{c}{QDE parameterisation} \\
BAO/CMB & 0.274$\pm0.0145$&$-1.109\pm0.082$& 0 & 1.036 & 6 \\
HII/BAO/CMB & 0.278$\pm0.0143$&$-1.088\pm0.080$& 0 & 213.85 & 162 \\
SNIa/BAO/CMB & 0.287$\pm0.0130$&$-1.034\pm0.056$& 0 & 563.68 & 586 \\
& \multicolumn{5}{c}{CPL parameterisation} \\
BAO/CMB & 0.278& $-1.052\pm0.083$&$-0.112\pm0.35$ & 1.087 & 6 \\
HII/BAO/CMB & 0.278& $-0.992\pm0.084$&$-0.368\pm0.38$ & 213.72 & 162 \\
SNIa/BAO/CMB & 0.278& $-0.983\pm0.057$&$-0.304\pm0.28$& 563.90 &
586 \\ \hline
\end{tabular}
\end{table*}
\begin{figure*}
\mbox{\epsfxsize=17.2cm
\epsffile{fig/FIG4a-2.eps}}
\caption{Comparison of the joint likelihood contours of the
HII/CMB/BAO (black contours) and of the SNIa/CMB/BAO
(red contours) probes.
{\it Left Panel:} QDE dark energy equation of state parametrisation.
{\it Right Panel:} CPL dark energy equation of state parametrisation
using $\Omega_{m,0}=0.278$ as a prior.}
\label{fig:02}
\end{figure*}
\subsection{Results of the Joint Analysis}
As discussed earlier our present sample of H\,\textsc{ii}\ galaxies is dominated by the very-low
redshift regime ($z<0.15$) as it contains only a small number of high-$z$
sources; therefore the cosmological constraints that can be imposed
are very weak \citep[see][]{Terlevich2015}. Nevertheless by joining the
H\,\textsc{ii}\ galaxy analysis with other cosmological probes we can further test the
effectiveness of using H\,\textsc{ii}\ galaxies as alternative tracers of the
Hubble expansion.
To this end we will present and compare our results of the joint analysis
but using as standard candles separately our H\,\textsc{ii}\
galaxies and the SNIa.
In Figure \ref{fig:01} we present the 1$\sigma$ and 2$\sigma$ likelihood
contours in the $(\Omega_{m,0}, w)$ plane for the following
probes: H\,\textsc{ii}\ galaxy Hubble relation (black contours),
CMB shift parameter (green) and BAO (blue), whereas with red we
present the result of the joint analysis. The solution provided by the
H\,\textsc{ii}\ galaxy Hubble relation probe has been shown to be consistent with that
of the SNIa, albeit leaving mostly unconstrained the QDE free
parameters \citep{Terlevich2015}. However,
the joint analysis reduces dramatically the solution space, providing quite
stringent constraints on the two QDE parameters. Even with the current
very broad H\,\textsc{ii}\-galaxy likelihood contours, the joint HII/BAO/CMB
analysis increases the Figure of Merit (FoM) by 13\% with respect to that of the BAO/CMB
joint analysis alone.
In order to compare the performance of the H\,\textsc{ii}\ galaxies (as they stand
today in our sample of only 25 high-$z$ sources) with that of the {\em Union2.1}
SNIa, we display in Figure \ref{fig:02} the joint likelihood contours for
HII/BAO/CMB$_{\rm shift}$ (black contours),
and SNIa/BAO/CMB$_{\rm shift}$ (red contours) probes for both DE EoS parameterisations.
Note that in the case of the CPL analysis we impose an {\em a priori}
value for the cosmological matter density parameter,
$\Omega_{m,0}=0.278$, and allow the two DE EoS parameters,
$w_0$ and $w_a$ to vary.
A first observation is that both joint analyses, based either on H\,\textsc{ii}\
galaxies or SNIa, provide consistent results for both DE equation of
state parameterisations,
although (as expected) the SNIa rate better since
the SNIa sample is much larger and their median redshift is
significantly higher than that of our preliminary H\,\textsc{ii}\ galaxy sample.
For the QDE case, the broad
H\,\textsc{ii}\ galaxy likelihood contours and the corresponding extensive
parameter degeneracy is reduced significantly with the joint
HII/BAO/CMB analysis, while
the degeneracy appears to disappear with the SNIa/BAO/CMB analysis.
As expected for the more
demanding CPL parametrisation the degeneracy between $w_0$ and $w_a$
is present in both sets of joint analyses.
However, what is particularly interesting is that
for the CPL model the two joint analyses
provide the same minimum, as can be seen also in Table \ref{tab:02},
where we list the resulting cosmological parameters and their
uncertainties for the different combinations of cosmological
probes.
It is very encouraging that even with the current H\,\textsc{ii}\ galaxy pilot sample, the
combined analysis of the H\,\textsc{ii}\ data with BAOs
and the CMB shift parameter provides constraints on the cosmological
parameters which are in agreement with those of the joint
SNIa/BAOs/${\rm CMB}_{\rm shift}$.
We plan to considerably increase the current sample of high-$z$ H\,\textsc{ii}\
galaxies (see next section) which together with other future cosmological data, based
for example on {\em Euclid}, will improve significantly the relevant
constraints (especially on $w_{a}$) and thus the validity of
a running EoS parameter, namely $w(z)$,
will be effectively tested.
\section{Monte-Carlo Simulations}
In order to predict the effectiveness of using high-$z$ H\,\textsc{ii}\ galaxies to
constrain the DE EoS, we have performed an
extensive series of Monte-Carlo simulations with which we assess our
ability to recover the input parameters of an {\em a priori} selected cosmological
model, in our case that of the concordance cosmology ($\Omega_{m, 0},
w_0, w_a)=(0.28, -1, 0)$.
We distribute different numbers of mock high-$z$ H\,\textsc{ii}\ galaxies
in redshift according to the observational constraints of
the adequate, for our purpose, instruments and telescopes (in this case
the VLT-KMOS spectrograph at ESO\footnote{As we prepared this work, we
have also procured some 25 high-z H\,\textsc{ii}\ galaxies data with MOSFIRE
at Keck. A paper is in preparation.}). The range of the available near IR bands
for this instrument are shown in Table \ref{tab:03}, as well as the corresponding
redshift ranges within which either the H$\alpha$ or [OIII] emission lines can
be observed. There are practically 4 independent redshift ranges that
can be sampled
centered at $\langle z \rangle \simeq 0.8, 1.4, 2.3$ and 3.3, and
these are the redshift ranges where we will distribute
our mock high-$z$ H\,\textsc{ii}\ galaxies\footnote{Note that other studies that
present simulations of the constraints provided by future
high-$z$ tracers of the Hubble expansion do not always take into
account the limited redshift intervals that can be observationally probed
\citep[cf.][]{Scovacricchi2016}.}. Since the IR bands window function are
clearly not top-hat, we model the distribution of redshifts, within
each $z$-window, by a Gaussian with mean and standard deviation given in
Table \ref{tab:03}.
\begin{table}
\caption{\bf The KMOS FWHM sensitivities and redshift windows}
\label{tab:03}
\tabcolsep 2pt
\begin{tabular}{ccccc} \hline
Band & $\lambda$/nm & H$\alpha$ $z$-window & [OIII] $z$-window & Exp.time (sec) \\
& & & & $S/N\simeq 25$ \\
J & 1175$\pm 40$ & 0.79$\pm 0.026$ & 1.35$\pm 0.046$ & 1800 \\
H & 1635$\pm 65$ & 1.49$\pm 0.060$ & 2.26$\pm 0.090$ & 1500 \\
K & 2145$\pm 65$ & 2.26$\pm 0.070$ & 3.28$\pm 0.100$ & 2100 \\ \hline
\end{tabular}
The width of the wavelength coverage includes only the region
with sensitivity higher than 50\% of the band peak sensitivity.
\end{table}
The Monte-Carlo simulation procedure that we follow entails assigning
to each mock H\,\textsc{ii}\ galaxy the ideal distance modulus for the
selected cosmology and an uncertainty which is determined by the
expected distribution of luminosity and flux errors that enter in
the relation (\ref{eq:mu1}). We then transform these errors in a
distance modulus error distribution
and use this distribution to assign randomly errors to each high-$z$ mock H\,\textsc{ii}\
galaxy. The mean distance modulus uncertainty is thus derived from
propagating the mean velocity dispersion and flux errors via
eq.(\ref{eq:mu1}), i.e.:
\begin{equation}
\sigma_\mu= 2.5 \left( \log\sigma^2 \sigma_a^2 +
a^2\sigma^2_{\log\sigma}+\sigma_b^2 +\frac{\sigma_f^2}{\ln(10)^2
f^2}\right)^{1/2}
\end{equation}
where $a$ and $b$ are the slope and intercept of the
$L_{H\beta}-\sigma$ relation, $\sigma_a$ and $\sigma_b$ are the
corresponding uncertainties of the fit, while $f$ and $\sigma_f$ are
the $H\beta$ line flux and its uncertainty.
Assuming a flux uncertainty of $\lesssim 10\%$ \citep[as indeed we
find for the three $z\gtrsim 1.5$ H\,\textsc{ii}\ galaxies we observed with X-SHOOTER; see][]{Terlevich2015} and the uncertainties of our
$L_{H\beta}-\sigma$ relation, we obtain a mean
$\langle \sigma_\mu\rangle \simeq 0.6$ mag, slightly lower than the
measured values of
our low-$z$ sample ($\langle \sigma_\mu\rangle \simeq 0.7$ mag).
The available high-$z$ H\,\textsc{ii}\ galaxy data from the literature (as well as our own data) indicate a large
dispersion of the distance modulus uncertainty and therefore, for the
purpose of our simulations, we will assume a Gaussian uncertainty
distribution with mean $\langle \sigma_\mu\rangle \simeq 0.6$ mag
and a standard deviation of $\sigma_{\sigma}\simeq 0.24$. Obviously,
the outcome of the simulations are sensitive to the error distribution
and the results presented here are intended as indicative of the potential
of our approach.
\subsection{Results of simulations}
In order to test the effectiveness of our procedure, as a starting
point, we assign to each of the 156 H\,\textsc{ii}\ galaxies and GEHR of our high-quality
velocity dispersions observational sample \citep{Chavez2014, Terlevich2015}
the ideal distance modulus and the actual observed
uncertainty. We then perform our usual
$\chi^{2}$ minimisation procedure and derive the cosmological
constraints, shown in Figure \ref{fig:03} as greyscale contours.
We also overplot the corresponding true
observational constraints of the same H\,\textsc{ii}\ galaxy sample, which
are statistically consistent with the {\em ideal} case (more so for
the QDE parametrization).
If for the {\em ideal} distance modulus case we assign to each source
the model observational uncertainties, discussed previously, we obtain
similar constraints as in the true uncertainties case but with
slightly higher FoM, by a factor of $\lesssim 2$.
\begin{figure}
\mbox{\epsfxsize=8.9cm
\epsffile{fig/FIGSIM1-3.eps}}
\caption{Likelihood contours
corresponding to the 1$\sigma$ and 2$\sigma$
confidence levels for our H\,\textsc{ii}\
galaxy sample but using the ideal concordance cosmology distance
moduli (grey-scale contours). In red we show the corresponding true
constraints of our current sample. {\em Left Panel:} QDE parametrisation. {\em
Right Panel:} CPL parametrisation with $\Omega_{m,0}=0.278$.}
\label{fig:03}
\end{figure}
For our tests we will consistently estimate the increase of the
current FoM, based on the 156 H\,\textsc{ii}\ galaxies and GEHR of our sample using the
{\em ideal} distance moduli
with that provided when we add
different numbers of high-$z$ H\,\textsc{ii}\ galaxies, distributed in the
redshift ranges shown in Table \ref{tab:03}. This exercise will be presented
for both the QDE and CPL parameterisations of the DE EoS.
Note that the distribution of numbers of the mock H\,\textsc{ii}\ galaxies at the different
redshift ranges could also affect the results in the sense that
different cosmological models show the largest deviations from the
concordance model at different redshifts \citep[eg., Fig.1 of][]{Plionis2011}.
After a trial and error procedure we found that an optimal
distribution of the fractions of the total number of high-$z$ H\,\textsc{ii}\
galaxies in the 4 available redshift ranges, shown in Table 3, is 0.2,
0.2, 0.3 and 0.3 (from the lowest to the highest redshift
range). However, the case of equal fraction among the different redshifts
provide similar results.
We performed 100 Monte-Carlo realisations for each selected number of
mock high-$z$ H\,\textsc{ii}\ galaxies, and the aggregate results are presented in
Figure 4, in the form of the ratio between the simulation FoM and that
of our current sample of H\,\textsc{ii}\ galaxies as a function of the number of
mock high-$z$ H\,\textsc{ii}\ galaxies. Thus what is shown is the factor by which
the FoM increases with respect to its current value. This factor
increases linearly with $\rm N_{\rm HII}$ providing the following
rough analytic expressions:
$$F_{\rm QDE}\simeq 0.015 N_{\rm HII}+ 1.72 \;\;\;\;{\rm and} \;\;\;\; F_{\rm CPL}\simeq
0.004 N_{\rm HII}+ 1.51$$
which means that for the very realistic near future expectations of observations of $\sim 500$
high-$z$ H\,\textsc{ii}\ galaxies, we predict a $\sim$ ten-fold increase of the current
FoM for the QDE parametrisation and $\sim$ four-fold increase of the
corresponding FoM for the CPL parametrisation, within the limits of
the parameters shown in Figure \ref{fig:04}.
As an example, we present in Figure \ref{fig:05} the results of one simulation of 500 high-$z$ mock
H\,\textsc{ii}\ galaxies both for the QDE and CPL parameterisations of the DE
EoS (grey-scale contours), which can be compared with
the constraints of our current sample (but using for consistency the
ideal distance moduli).
\section{Feasibility of the project and Future Work}
The realisation of this project relies on two main prerequisites; finding an
adequate number of high-$z$ H\,\textsc{ii}\ galaxy targets and being able to
observe them using a reasonable amount of observing time.
To this end, we compiled a sample of objects searching the literature for high-$z$
H\,\textsc{ii}\ galaxy candidates that we define as compact emission line systems
with either $W$(H$\alpha)>200$ \AA\ and $W$[OIII]$\lambda 5007 > 200$ \AA\ or
with $W$(H$\beta)>50$ \AA\ and FWHM$<150$ \AA\ and with $z>1.2$. We have found
up to now more than 500 candidates in about 20 high galactic latitude fields
(Gonz\'alez-Mor\'an et al.~in preparation).
To estimate the feasibility of our project we calculated the time it
could take to observe the whole sample. For this estimate we have
assumed the use of IR spectroscopic facilities with resolution R
larger than 4000 in 10m class telescopes and with multiplexing
capability. These facilities are at present only two, KMOS at the VLT
and MOSFIRE at Keck. We have used the KMOS Exposure Time Calculator to
estimate the time needed to obtain a S/N 25 or larger in either
H$\alpha$ or [OIII]$\lambda$5007 for the faintest objects in our list
and combine this estimate with their surface density at $z\sim 2.3$.
The typical exposure times are about 3 hours per field. Each search
field is typically populated by 25 objects with about 8 to 15
simultaneously inside either the KMOS or MOSFIRE field of view. Thus
the number of objects that can be observed in a 10 hours night ranges
from 24 to 45,
therefore about 15 observing nights would be needed to observe 500 H\,\textsc{ii}\ galaxies.
This estimate is shown in the upper scale of Figure \ref{fig:04}.
\begin{figure}
\mbox{\epsfxsize=8.2cm
\epsffile{fig/FoM_final1.eps}}
\caption{The factor by which the FoM of the QDE and CPL EoS constraints
increases with respect to its current value (based on the observed 25
high-$z$ HII galaxies) as a function of the number of
mock high-$z$ H\,\textsc{ii}\ galaxies. The FoM has been estimated within the limits of the parameters
shown in Figure \ref{fig:03}. The red and blue points correspond to the
QDE and CPL parameterisations of the DE EoS,
respectively. The solid black lines are
the linear fits to the corresponding coloured curves. The scale at the
top gives the number of 10m class telescope nights needed in order
to observe 500, 1000 and 1500 objects as 15, 30 and 45 nights respectively.}
\label{fig:04}
\end{figure}
\begin{figure}
\mbox{\epsfxsize=8.9cm
\epsffile{fig/FIGSIM2-3.eps}}
\caption{Likelihood contours
corresponding to the 1$\sigma$ and 2$\sigma$
confidence levels for our H\,\textsc{ii}\
galaxy sample but adding 500 high-$z$ mock H\,\textsc{ii}\ galaxies (grey-scale
contours). In red we show the corresponding current
constraints (ie., without the high-$z$ mock H\,\textsc{ii}\ galaxies). We consistently
use the {\em ideal} distance moduli of the concordance cosmology. {\em Left Panel:} QDE parametrisation. {\em
Right Panel:} CPL parametrisation using $\Omega_{m,0}=0.278$.}
\label{fig:05}
\end{figure}
\section{Conclusions}
We have used the Hubble relation of H\,\textsc{ii}\ galaxies
in a joint likelihood
analysis with the BAO and CMB cosmological probes
with the aim of testing the consistency of the derived
cosmological constraints with those of the joint
SNIa/BAO/CMB analysis. This results in two important conclusions:
\begin{itemize}
\item The FoM of the QDE EoS constraints, provided by the joint
HII/BAO/CMB analysis, was found to be larger by 13 percent than
those provided by the BAO/CMB joint analysis, even with the very small
sample of only 25 high-z HII galaxies.
\item Both the QDE and CPL EoS constraints of the HII/BAO/CMB and of
the SNIa/BAO/CMB joint analyses are in excellent consistency with
each other, although (as expected) the SNIa probe still provides a
significantly larger FoM.
\end{itemize}
We have also performed Monte-Carlo simulations tailored to the specific
uncertainties of the $L(\mathrm{H}\beta) - \sigma$\ relation and to the technical
instrumental requirements of KMOS/VLT (and instruments like it). They
address the important question of what is the expected increase of the
FoM as a function of the number of high-z H\,\textsc{ii}\ galaxies in the
redshift windows accessible. Our previous
simulations \citep[cf.][]{Plionis2011} did not take into account
the specific error budget of our $L(\mathrm{H}\beta) - \sigma$\ relation, or the
characteristics of the instruments available and of the accessible
redshifts.
We would like to add that cosmological analyses, like the one
presented in this work, demands a
thorough understanding of the interplay between observational random
and systematic errors and biases, for which mock catalogues are an
essential tool.
\section*{Acknowledgements}
We are thankful to an anonymous referee for careful and constructive comments on the manuscript.
RC, RT, ET and MP are grateful to
the Mexican research council (CONACYT) for supporting this research
under studentship 224117 and grants 263561, CB-2005-01-49847, CB-2007-01-84746 and
CB-2008-103365-F.
SB acknowledges support by the Research
Center for Astronomy of the Academy of Athens in the context of the
program {\it ``Tracing the Cosmic Acceleration''}. MP acknowledges the
hospitality of the KAVLI Institute for Cosmology in Cambridge, where
this work was completed.
|
\section{Introduction}
Self-consistent mean-field models based on relativistic energy density functionals (EDFs)
with density-dependent strength parameters \cite{dd}, have been successfully applied
to studies of a broad variety of nuclear phenomena such as radii, masses,
collective modes, fission and shape coexistence (see e.g.
\cite{Ring96,rprc}). Remarkable results have been obtained both in the
relativistic mean-field (RMF) framework \cite{Ring96,Niksic_2002}, and
more recently using the relativistic Hartree-Fock (RHF) scheme
\cite{par_PKO2,ebr11,gu}, even though RHF applications have mostly been
restricted to spherical nuclei. One of the basic advantages of using functionals
with manifest covariance is the natural inclusion of the nucleon spin degree
of freedom, and the resulting nuclear spin-orbit potential which emerges automatically
with the empirical strength in a covariant formulation \cite{FS.00}.
Non-relativistic EDFs such as, for instance, Skyrme \cite{Vautherin72} or the
Gogny \cite{berg} functionals must, of course, also include a spin-orbit term.
In this case, however, the strength of the phenomenological spin-orbit term has
to be adjusted to data on the energy spacing between spin-orbit partner states.
This approach has been extensively applied and refined over the last four
decades \cite{ben,SR.07,EKR.11}, and it provides an effective description of
spin-orbit effects in nuclei.
The omission of an explicit treatment of exchange terms in the RMF
approach may have an impact on the description of the isovector
channel, in particular for the energy gap between spin-orbit partner
states when the ratio between neutrons and protons becomes very large.
The modification of spin-orbit splittings predicted by RMF-based
models differs from that obtained with non-relativistic, e.g. Skyrme
models \cite{Ring96,shar}. Empirical constraints can be obtained by
studying the changes in neutron spin-orbit splittings when the number
of protons change and vice versa, as in the recent study of
spectroscopic properties of $^{35}$Si and $^{37}$S \cite{burg}. This
task is, however, not straightforward because single-particles
energies and occupations factors are not direct observables
\cite{duguet}.
On the theoretical side, the difference between the isospin dependence
of RMF and non-relativistic spin-orbit interactions can be analyzed by
performing a non-relativistic reduction of the Dirac equation. Such a
study was reported, for instance, by Sulaksono et al. \cite{burn} with
the goal to compare in a global way the magnitude of spin-orbit terms
in these two approaches. In this work we focus on the isospin
dependence of the spin-orbit effect using relativistic EDFs with
density-dependent strength parameters, and evaluate the effect of
explicit treatment of exchange terms in relativistic structure models.
\section{Spin-orbit term in relativistic effective interactions}
\subsection{\label{RMF}The RMF case}
Most SCMF models based on the relativistic mean-field approximation
have been formulated using the finite-range meson-exchange
representation, in which the nucleus is described as a system of Dirac
nucleons coupled to mesons fields through an effective Lagrangian. The
isoscalar scalar $\sigma$-meson, the isoscalar vector $\omega$-meson,
and the isovector vector $\rho$-meson build the minimal set of meson
fields that, together with the electromagnetic field, is necessary for
a description of bulk and single-particle nuclear properties. The
corresponding Lagrangian density reads
\begin{eqnarray}
\mathcal{L} =&& \bar \Psi (i \gamma^\mu \partial_\mu - m)\Psi \nonumber \\
&&+ \frac{1}{2}(\partial_\mu \sigma \partial^\mu \sigma - m_\sigma^2 \sigma^2) - \frac{1}{2}(\frac{1}{2}\Omega^{\mu\nu}\Omega_{\mu\nu} - m_\omega^2\omega_\mu \omega^\mu) - \frac{1}{2}(\frac{1}{2} \boldsymbol{R}_{\mu\nu} \cdot \boldsymbol{R}^{\mu\nu} - m_\rho^2 \boldsymbol{\rho}_\mu \cdot \boldsymbol{\rho}^\mu) \nonumber \\
&&-g_\sigma(\rho_B (\vec r))\bar \Psi \sigma \Psi -g_\omega(\rho_B (\vec r))\bar \Psi \gamma_\mu \omega^\mu \Psi -g_\rho(\rho_B (\vec r))\bar \Psi \gamma_\mu \boldsymbol{\rho}^\mu \cdot \boldsymbol{\tau} \Psi,
\end{eqnarray}
where, for simplicity, we omit the Coulomb term which is not relevant for the present discussion.
$\Psi$ denotes the nucleon spinor, $m$ is the nucleon bare mass, $m_\sigma$, $m_\omega$ and $m_\rho$ denote the meson masses.
$\Omega^{\mu\nu}\equiv \partial^\mu\omega^\nu-\partial^\nu\omega^\mu$ and $\boldsymbol{R}^{\mu\nu}\equiv \partial^\mu\boldsymbol{\rho}^\nu-\partial^\nu\boldsymbol{\rho}^\mu$ are the $\omega$ and $\rho$ meson field tensors. Boldface symbols denote vectors and tensors in
isospin space. The meson-nucleon couplings are assumed to be functions of
the nucleon density (time-like component of the nucleon 4-current) $\rho_B(\vec r)$, and this density dependence
in principle encodes all in-medium many-body correlations.
\newline
In the self-consistent RMF framework the dynamics of independent nucleons is determined by
local scalar and vector self-energies. For simplicity spherical nuclei are considered and time-reversal symmetry is assumed
(pairwise occupied states with Kramers degeneracy), which
ensures that the only non-vanishing components of the 4-vector fields are the time-like
ones and thus there is no net contribution from nucleon currents. Because of charge conservation
only the 3rd component of the vectors in isospin space gives a non-vanishing contribution.
The single-nucleon equation of motion is then the Dirac equation:
\begin{equation}
\label{Dirac}
\left [ {\vec \alpha} \cdot {\vec p} + V + \beta (m+S) \right ] \psi_i = E_i \psi_i
\end{equation}
where $\vec \alpha = \gamma_0\vec\gamma$, $\beta = \gamma_0$,
$\gamma_0$ and $\vec\gamma$ are the Dirac matrices in Dirac representation,
and $\psi_i$ denotes the self-consistent solution for the i-th Dirac state of
energy E$_i$:
\begin{equation}
\left(
\begin{array}[c]{c}
\phi_i\\
\chi_i
\end{array}
\right)
\end{equation}
with $\phi_i$ and $\chi_i$ denoting the large and small component, respectively. The scalar and time-like vector self-energies read:
\begin{equation}
\label{Self-enS}
S(\vec r) = g_\sigma(\rho_B(\vec r)) \sigma(\vec r)
\end{equation}
\begin{eqnarray}
V(\vec r) = && g_\omega(\rho_B(\vec r)) \omega(\vec r) + g_\rho(\rho_B(\vec r)) \tau^3 \rho(\vec r) +
\frac{dg_\sigma}{d\rho_B}\sum_i\bar{\psi}_i(\vec r)\sigma(\vec r)\psi_i(\vec r) \\
&& + \frac{dg_\omega}{d\rho_B}\sum_i\bar{\psi}_i(\vec r) \gamma_0\omega(\vec r)\psi_i(\vec r)+
\frac{dg_\rho}{d\rho_B}\sum_i\bar{\psi}_i(\vec r)\gamma_0\rho(\vec
r)\tau^3\psi_i(\vec r)
\label{Self-enV}
\end{eqnarray}
The explicit dependence of the coupling functions on the baryon
density $\rho_B$ produces rearrangement contributions to the vector
nucleon self-energy. The rearrangement terms result from the variation
of the couplings with respect to the baryon density.
In applications to nuclear matter and finite nuclei, relativistic
models are used in the {\em no-sea} approximation: the Dirac sea of
states with negative energies does not contribute to the densities and
currents. In the nuclear ground state A nucleons occupy the lowest
single-nucleon orbitals, determined self-consistently by the iterative
solution of the Dirac equation (\ref{Dirac}). Expressing the
single-nucleon energy as $E_i = m + \varepsilon_i$, where $m$ is the
nucleon mass, and rewriting the Dirac equation as a system of two
equations for $\phi_i$ and $\chi_i$, then, noticing that for bound
states $\varepsilon_i << m$, the equation for the upper component $\phi_i$ of the Dirac spinor reduces to the
Schr\" odinger-like form \cite{bib82,rei89,lala}
\begin{equation}
\label{Sch}
\left [ {\vec p} {1 \over 2 {\cal M}(r)} {\vec p} + U(r) + V_{so}(r) \right ] \phi_i = \varepsilon_i \phi_i
\end{equation}
for a nucleon with effective mass
\begin{equation}
{\cal M}({r})\equiv m+\frac{1}{2}\left(S({r})-V({r})\right) \;,
\label{eq:meff}
\end{equation}
in the potential
$U(r) \equiv V(r) + S(r)$. The resulting additional spin-orbit potential \cite{bib82,rei89,lala}
\begin{equation}
\label{eq:vso}
V_{so} = \frac{1}{2r {\cal M}^2 (r)} \frac{\mathrm{d}}{\mathrm{d}r}(V-S)~\vec l \cdot \vec s
\end{equation}
plays a crucial role in reproducing the empirical nuclear magic numbers.
The non-relativistic limit corresponds to an $\frac{1}{2 {\cal M} (\vec r)}$ expansion. In the lowest order the isoscalar density $\rho_s(\vec r)$ can be approximated by the non-relativistic nucleon density $\rho_B(\vec r)$.
At the energy scale characteristic for nuclear binding, meson exchange
($\sigma$, $\omega$, $\rho$, $\ldots$) is just a
convenient representation of the effective nuclear interaction. The exchange
of heavy mesons is associated with short-distance dynamics that cannot be
resolved at low energies, and therefore in each channel meson exchange can be
replaced by the corresponding local four-point (contact) interactions between
nucleons. The relation between the two representations: finite-range
(meson exchange) and zero-range (point-coupling), is straightforward in nuclear
matter because of constant nucleon scalar and vector densities.
The Klein-Gordon equations of the
meson-exchange model with meson masses $m_{\phi}$ and
density-dependent couplings $g_{\phi}(\rho)$, are replaced by the corresponding
point-coupling interaction terms with strength parameters
$g_{\phi}^{2}/m_{\phi}^{2}$. In finite nuclei, however, because of
the radial dependence of the densities, the expansion of the
meson propagator in terms of $1/m_{\phi}^{2}$ leads to a series of
gradient terms \cite{NVLR.08}. For the purpose of our discussion it suffices to
consider only the lowest order, in which the
self-energies $S(\vec r)$~(\ref{Self-enS}) and $V(\vec r)$~(\ref{Self-enV}) read
\begin{eqnarray}
S(\vec r) &=& - \frac{g_\sigma^2(\rho_B(\vec r))}{m_\sigma^2}\rho_B(\vec r) \label{eq:se0} \\
V(\vec r) &=& \frac{g_\omega^2}{m_\omega^2}\rho_B(\vec r) + \tau^3 \frac{g_\rho^2}{m_\rho^2}\rho_\tau(\vec r) \nonumber \\
&&- \frac{g_\sigma g_\sigma'}{m_\sigma^2} \rho_B^2(\vec r) + \frac{g_\omega g_\omega'}{m_\omega^2} \rho_B^2(\vec r) + \frac{g_\rho g_\rho'}{m_\rho^2} \rho_\tau^2(\vec r) \;,
\label{eq:se}
\end{eqnarray}
where $\rho_B$ is the nucleon density and $\rho_\tau$=$\rho^{(n)}_B-\rho^{(p)}_B$ is the isovector nucleon density. Introducing the notation:
\begin{eqnarray}
\label{eq:def_alpha}
\alpha_i &\equiv& \frac{g_i^2}{m_i^2} \\
\alpha_i' &\equiv& \frac{\mathrm{d} \alpha_i}{\mathrm{d} \rho_B} = \frac{2g_i g_i'}{m_i^2} \\
\alpha_i'' &\equiv& \frac{\mathrm{d}^2 \alpha_i}{\mathrm{d} \rho_B^2} = 2\frac{(g_i')^2 + g_i g_i''}{m_i^2}
\end{eqnarray}
for $i = \{\sigma,\omega,\rho\}$, and explicitly writing the neutron and proton contributions ($q = \{n,p\}$) with
$\rho_B^{(q-q')} \equiv \rho_B^{(q)} - \rho_B^{(q')}$,
from Eqs (\ref{eq:vso}), (\ref{eq:se0}) and (\ref{eq:se}) one derives:
\begin{eqnarray}
\label{eq:Vso_W1W2} &&V_{so}^{(q)} = \frac{\alpha_\sigma +
\alpha_\omega + \alpha_\rho + 2 \alpha_\omega' \rho_B + 2 \alpha_\rho'
\rho_B^{(q-q')} + \frac{- \alpha_\sigma'' +
\alpha_\omega''}{2}\rho_B^2 + \frac{\alpha_\rho''}{2}
(\rho_B^{(q-q')})^2}{2r\left\{ m - \frac{1}{2}\left[ (\alpha_\sigma +
\alpha_\omega)\rho_B + \alpha_\rho \rho_B^{(q-q')} +
\frac{-\alpha_\sigma' + \alpha_\omega'}{2}\rho_B^2 +
\frac{\alpha_\rho'}{2}(\rho_B^{(q-q')})^2
\right]\right\}^2}\frac{\mathrm{d}\rho_B^{(q)}}{\mathrm{d}r} \vec l
\cdot \vec s \nonumber \\ &&+ \frac{\alpha_\sigma + \alpha_\omega -
\alpha_\rho + 2 \alpha_\omega' \rho_B + \frac{- \alpha_\sigma'' +
\alpha_\omega''}{2}\rho_B^2 + \frac{\alpha_\rho''}{2}
(\rho_B^{(q-q')})^2}{2r\left\{ m - \frac{1}{2}\left[ (\alpha_\sigma +
\alpha_\omega)\rho_B + \alpha_\rho \rho_B^{(q-q')} +
\frac{-\alpha_\sigma' + \alpha_\omega'}{2}\rho_B^2 +
\frac{\alpha_\rho'}{2}(\rho_B^{(q-q')})^2
\right]\right\}^2}\frac{\mathrm{d}\rho_B^{(q'\ne q)}}{\mathrm{d}r}
\vec l \cdot \vec s \nonumber \\
\end{eqnarray}
This expression can be rewritten as
\begin{equation}
\label{W1andW2}
V_{so}^{(q)} = \left[ W_1 \frac{\mathrm{d}\rho_B^{(q)}}{\mathrm{d}r} + W_2 \frac{\mathrm{d}\rho_B^{(q'\ne q)}}{\mathrm{d}r} \right]\vec l \cdot \vec s
\end{equation}
and the relevant ratio that determines the isospin dependence of the spin-orbit potential reads
\begin{equation}
\label{eq:W1W2}
\frac{W_1}{W_2}^{(q)}(\alpha_\sigma,\alpha_\omega,\alpha_\rho) \equiv
\frac{A^q(\alpha_\sigma,\alpha_\omega,\alpha_\rho)+B^q(\alpha_\rho,\rho_B^{(q-q')})}{A^q(\alpha_\sigma,\alpha_\omega,\alpha_\rho)-B^q(\alpha_\rho,0)}\;,
\end{equation}
\vspace{0.5cm}
with
\begin{equation}
A^q(\alpha_\sigma,\alpha_\omega,\alpha_\rho)\equiv
\alpha_\sigma + \alpha_\omega + 2 \alpha_\omega'
\rho_B + \frac{- \alpha_\sigma'' +
\alpha_\omega''}{2}\rho_B^2 + \frac{\alpha_\rho''}{2}
(\rho_B^{(q-q')})^2
\end{equation}
and
\begin{equation}
B^q(\alpha_\rho,\rho_B^{(q-q')})\equiv
\alpha_\rho+ 2 \alpha_\rho' \rho_B^{(q-q')}
\label{eq:isov}
\end{equation}
Equation (\ref{eq:W1W2}) shows that the ratio W$_1$/W$_2$ differs from unity
because of the isovector contribution (\ref{eq:isov}), and is larger than one
for B$^q> 0$.
\subsection{\label{RHF}The RHF case}
In the relativistic Hartree-Fock case, in which exchange terms are treated explicitly, because of non-locality it is not possible to derive a
simple analytic expression for the non-relativistic spin-orbit potential. For a direct comparison with the RMF case, one can first consider the point-coupling approximation to the meson-exchange RHF Lagrangian, and further perform a Fierz transformation to obtain a corresponding RMF Lagrangian \cite{liang}. The interacting part of the RHF Lagrangian reads:
\begin{equation}
\label{eq:Lsimp}
\mathcal{L}_{int} = -g_\sigma(\rho_B) \bar{\Psi} \sigma \Psi - g_\omega(\rho_B) \bar{\Psi} \gamma_\mu \omega^\mu \Psi - g_\rho(\rho_B) \bar{\Psi} \gamma_\mu \boldsymbol{\rho}^\mu \cdot \boldsymbol{\tau}\Psi
\end{equation}
In the lowest-order point-coupling approximation \cite{ddpc} the mesons fields can be expressed
\begin{eqnarray}
\label{eq:champPC}
\sigma &=& - \frac{g_\sigma(\rho_B(\vec r))}{m_\sigma^2}\bar \Psi \Psi \nonumber \\
\omega^\mu &=& \frac{g_\omega(\rho_B(\vec r))}{m_\omega^2}\bar \Psi \gamma^\mu \Psi \nonumber \\
\boldsymbol{\rho}^\mu &=& \frac{g_\rho(\rho_B(\vec r))}{m_\rho^2}\bar \Psi \gamma^\mu \boldsymbol{\tau} \Psi \;,
\end{eqnarray}
and for the equivalent Lagrangian in the point-coupling approximation
\begin{equation}
\label{eq:lagdensity}
\mathcal{L}_{int}^{PC} = -\frac{1}{2}\alpha_\sigma(\bar \Psi \Psi)(\bar \Psi \Psi) -\frac{1}{2}\alpha_\omega(\bar \Psi \gamma_\mu \Psi)(\bar \Psi \gamma^\mu \Psi) -\frac{1}{2}\alpha_\rho(\bar \Psi \gamma_\mu \boldsymbol{\tau} \Psi)\cdot (\bar \Psi \gamma^\mu \boldsymbol{\tau} \Psi)
\end{equation}
one obtains the ground-state expectation value
\begin{equation}
\left\langle\mathcal{L}_{int}^{PC} \right\rangle = -\frac{1}{2} \alpha_\sigma\rho_s^2 -\frac{1}{2} \alpha_\omega\rho_B^2 -\frac{1}{2} \alpha_\rho\rho_{\tau}^2 +\frac{1}{2} \alpha_\sigma\rho_{s,exch}^2 +\frac{1}{2} \alpha_\omega\rho_{v,exch}^2 +\frac{1}{2} \alpha_\rho\rho_{\tau,exch}^2 \;.
\end{equation}
Using the Fierz transformation, the couplings $\tilde \alpha_i$ of the corresponding RMF Lagrangian are expressed in terms of
those of $\mathcal{L}_{int}^{(PC)}$ \cite{liang}:
\begin{eqnarray}
\tilde \alpha_S &=& \frac{7}{8} \alpha_\sigma + \frac{1}{2} \alpha_\omega + \frac{3}{2}\alpha_\rho \\
\tilde \alpha_V &=& \frac{1}{8} \alpha_\sigma + \frac{5}{4} \alpha_\omega + \frac{3}{4}\alpha_\rho \\
\tilde \alpha_{tV} &=& \frac{1}{8} \alpha_\sigma + \frac{1}{4} \alpha_\omega + \frac{3}{4}\alpha_\rho \\
\tilde \alpha_{tS} &=& -\frac{1}{8} \alpha_\sigma + \frac{1}{2} \alpha_\omega - \frac{1}{2}\alpha_\rho \;.
\end{eqnarray}
The resulting Fierz Lagrangian:
\begin{equation}
\left\langle\mathcal{L}_{int}^{(Fierz)}\right\rangle = -\frac{1}{2} \tilde \alpha_S\rho_s^2
-\frac{1}{2} \tilde \alpha_V\rho_B^2 -\frac{1}{2} \tilde
\alpha_{tV}\rho_{\tau}^2 -\frac{1}{2} \tilde \alpha_{tS}\rho_{tS}^2
\label{lag}
\end{equation}
is then equivalent to the RMF Lagrangian of the previous section but,
in addition, an isovector scalar term appears because of the Fierz
transformation. Additional terms in the pseudoscalar and pseudovector
channels do not contribute to the self-consistent ground-state
solution. Using this expression in (\ref{eq:W1W2}) yields the ratio
$\frac{W_1}{W_2}^{(q)}$ for the Fierz Lagrangian:
\begin{equation}
\label{eq:W1W2fierz0}
\frac{W_1}{W_2}^{(q)} \equiv \frac{W_1}{W_2}^{(q)}(\tilde \alpha_S,\tilde \alpha_V,\tilde
\alpha_{tV}+\tilde \alpha_{tS})
\end{equation}
with the explicit functional dependence
\begin{eqnarray}
\label{eq:W1W2fierz}
\frac{W_1}{W_2}^{(q)} = \frac{\tilde \alpha_S + \tilde \alpha_V + \tilde \alpha_{tV} + \tilde \alpha_{tS} + 2 \tilde \alpha_V' \rho_B + 2(\tilde \alpha_{tV}' + \tilde \alpha_{tS}')\rho_B^{(q-q')}}{\tilde \alpha_S + \tilde \alpha_V - \tilde \alpha_{tV} - \tilde \alpha_{tS} + 2 \tilde \alpha_V' \rho_B + \frac{-\tilde \alpha_S'' + \tilde \alpha_V''}{2}\rho_B^2 + \frac{\tilde \alpha_{tV}'' + \tilde \alpha_{tS}''}{2}(\rho_B^{(q-q')})^2}\nonumber \\
+ \frac{\frac{-\tilde \alpha_S'' + \tilde \alpha_V''}{2}\rho_B^2 + \frac{\tilde \alpha_{tV}'' + \tilde \alpha_{tS}''}{2}(\rho_B^{(q-q')})^2}{\tilde \alpha_S + \tilde \alpha_V - \tilde \alpha_{tV} - \tilde \alpha_{tS} + 2 \tilde \alpha_V' \rho_B + \frac{-\tilde \alpha_S'' + \tilde \alpha_V''}{2}\rho_B^2 + \frac{\tilde \alpha_{tV}'' + \tilde \alpha_{tS}''}{2}(\rho_B^{(q-q')})^2} \;.
\end{eqnarray}
The structure of equation (\ref{eq:W1W2fierz}) is similar to that of
Eq. (\ref{eq:W1W2}) but, in addition to the isovector-vector, it
contains also an isovector-scalar contribution but the strength
parameter $\tilde \alpha_{tS}$ of this channel is not independent. In
the meson-exchange representation this channel corresponds to the
exchange of a $\delta$-meson. The isovector-scalar meson $\delta$ can
be, of course, explicitly included in the model Lagrangian but, as it
has been often argued in the literature, it is difficult to determine
its coupling strength from available data on finite nuclei. In the RMF
meson-exchange model DD-ME$\delta$ developed and tested in
Ref~\cite{RM.11}, for instance, the isovector effective mass $m^*_p -
m^*_n$ derived from relativistic Brueckner theory was used to
determine the coupling strength of the $\delta$-meson and its density
dependence. It was noted, however, that the explicit inclusion of the
isovector-scalar meson does not improve the accuracy of calculated
properties of finite nuclei such as masses and radii.
\section{Results and discussion}
Conventional non-relativistic Hartree-Fock mean-field calculations
based on the Skyrme or Gogny force use a spin-orbit potential without
explicit isospin dependence, and with a constant strength parameter.
The explicit treatment of the exchange term constraints the ratio of
the resulting constants in the expression of Eq.~(\ref{W1andW2}) to
$W_1/W_2 = 2$ \cite{ben,shar}. In some cases this choice is too
restrictive, but it can be relaxed if the effective interaction
is interpreted as resulting from an energy density functional in the
sense of Kohn-Sham density functional theory \cite{SR.07,EKR.11}. In
the relativistic mean-field approximation (cf. Sec. \ref{RMF}) a weak
isospin dependence of the effective spin-orbit potential arises
because of the $\rho$-meson contribution (in meson-exchange models) or
the isovector-vector term of the interaction Lagrangian (in
point-coupling models). Exchange terms are not computed explicitly
and, because of the way the spin-orbit potential Eq.~(\ref{W1andW2})
emerges in the non-relativistic reduction of the single-nucleon Dirac
equation, the ratio $W_1/W_2$ in Eq. (\ref{eq:W1W2}) explicitly
depends on proton and neutron densities.
Figure \ref{fig:1} displays the radial dependence of the proton and
neutron ratio ${W_1}/{W_2}$ of parameters of the spin-orbit potential
Eq.~(\ref{eq:W1W2}), for the self-consistent ground states of
$^{16}$O, $^{34}$Si and $^{208}$Pb, calculated with two of the most
successful RMF effective interactions DD-ME2 \cite{LNVR.05}
(meson-exchange) and DD-PC1 \cite{ddpc} (point-coupling). It should be
noted that, in contrast to the value of the ratio $W_1/W_2 = 2$ used in
standard non-relativistic Hartree-Fock calculations, in the RMF
case the ratio ${W_1}/{W_2}$ is close to one. The absolute deviation
from unity can be attributed to the contribution of the $\rho$-meson
exchange, that is, the explicit contribution of the isovector-vector
channel: in the absence of the isovector degree of freedom
in the interaction Lagrangian, Eq. (\ref{eq:W1W2}) gives $W_1=W_2$.
The isovector contribution is, of course, also responsible for the
difference between the effective proton and neutron single-particle
potentials, while the radial (density) profiles depend on the shell
structure of occupied orbitals in the self-consistent solution for a
particular nucleus. In this respect, especially interesting is the
case of $^{34}$Si, for which a possible central depletion of the
proton density distribution has been analysed using a variety of
theoretical approaches \cite{gra,yao}, and experimental constraints on
the strength of the two-body spin-orbit interaction have been reported
\cite{burg,muts}. For the effective interaction DD-ME$\delta$ that
explicitly includes contributions from both $\rho$ and $\delta$ meson
exchange in the direct term, the isovector channel of the spin-orbit potential
is enhanced when compared to DD-ME2, although in both models the total isovector part of
the spin-orbit potential is an order of magnitude weaker than the isoscalar contribution \cite{RM.11}.
\begin{figure}[]
{\includegraphics[width=0.8\textwidth]{RMF.pdf}
}
\caption{Radial dependence of the proton (solid) and neutron (dashed) ratio $(\frac{W_1}{W_2})$ of parameters of the spin-orbit potential
Eq.~(\ref{eq:W1W2}), for the ground states of $^{16}$O, $^{34}$Si and $^{208}$Pb, calculated with the RMF effective interactions DD-ME2 (left) and DD-PC1 (right).}
\label{fig:1}
\end{figure}
To illustrate the effect of the exchange terms in the RHF
approximation on the single-nucleon spin-orbit potential the radial dependence of the ratio ${W_1}/{W_2}$
for protons and neutrons for the same nuclei are plotted in Fig.
\ref{fig:2} using
Eq.~(\ref{eq:W1W2fierz}). This corresponds to using the RMF Lagrangian
(Eq.~(\ref{lag})), obtained by performing the Fierz transformation on
the interaction terms of the point-coupling RHF Lagrangian
(Eq.~(\ref{eq:lagdensity})). The effective RHF interaction is PKO2
\cite{long} which includes the $\sigma$, $\omega$, and $\rho$ meson
exchange, but not the pion or the $\delta$-meson. The most important
result is that in this case the overall value of the ratio
${W_1}/{W_2}$ is around 1.8. This is significantly larger than in the
simple RMF approach based on the Hartree approximation, and much
closer to the value 2 which characterises standard non-relativistic HF
calculations based on Skyrme forces. The difference with respect to
the latter is due to the fact that there is already an isovector
dependence of the effective spin-orbit potential for the Lagrangian
PKO2 which arises because of the $\rho$-meson exchange contribution,
and also due to the non-relativistic reduction of the single-nucleon
Dirac equation to the Schr\" odinger-like form Eq.~(\ref{Sch}) that
explicitly includes the effective spin-orbit potential.
\begin{figure}[]
{\includegraphics[width=0.8\textwidth]{RHFFierz.pdf}
}
\caption{Same as in the caption to Fig. \ref{fig:1}, but the ratio of the parameters is given by Eq.~(\ref{eq:W1W2fierz}) for the case of the relativistic Hartree-Fock effective interaction PKO2.}
\label{fig:2}
\end{figure}
An effect that has been attributed to the isospin dependence of the
effective spin-orbit potential is the change (kink) of charge isotope
shifts across the $N=126$ shell gap \cite{Ang.04,Coc.13}. The charge
isotope shift is the difference between the charge radius $< r^2_{ch}
>$ of a given isotope with respect to the reference nucleus. The best
known example is the kink in the isotope shifts of even Pb nuclei and,
more recently, a similar effect has also been observed in Polonium
isotopes \cite{Coc.13}. Numerous calculations over the last twenty
years have shown that all relativistic mean-field effective
interactions, both at the RMF level (without or with inclusion of the
isovector scalar $\delta$ meson) and in the RHF approach, reproduce
the empirical kink in the isotope shifts of even Pb isotopes
\cite{Ring96,shar,RM.11,RF.95,par_PKO2}. This was explained by a
relatively weak isospin dependence of the corresponding spin-orbit
potentials. Conventional Skyrme HF parameterizations with $W_1/W_2 =
2$ were unable to reproduce the kink and, therefore, in
Ref.~\cite{RF.95} the Skyrme framework was extended with an additional
degree of freedom in the spin-orbit channel which allows to modify the
value of the ratio $W_1/W_2$. This simple modification of the Skyrme
functional, in which the relative weights of the neutron and proton
contributions to the spin-orbit potential can be freely adjusted,
produces values for the isotope shifts of Pb in reasonable agreement
with data.
\begin{figure}[]
{\includegraphics[width=0.6\textwidth]{Pbnew.png}
}
\caption{Isotope shifts for even-A Pb nuclei with respect to the reference nucleus $^{208}$Pb. Experimental values \cite{Ang.04,Coc.13} are shown in
comparison with theoretical results obtained in the RMF calculation with the effective interaction DD-PC1, using the relativistic
Hartree-Fock effective interaction PKO2, and with the RMF model obtained by performing the Fierz transformation of
the point-coupling approximation of PKO2.}
\label{fig:3}
\end{figure}
In Figure \ref{fig:3} we plot the experimental isotope shifts for
even-A Pb nuclei with respect to the reference nucleus $^{208}$Pb, in
comparison with results obtained in the RMF calculation with the
effective interaction DD-PC1, using the relativistic Hartree-Fock
effective interaction PKO2, and with the RMF model obtained by
performing the Fierz transformation of the equivalent point-coupling
approximation of PKO2. In all three cases the theoretical values
reproduce the empirical kink at $N=126$ and, in particular, the kink
is most pronounced in the RMF calculation with the Fierz-transformed
effective interaction PKO2, even though ${W_1}/{W_2} \approx 1.8$ for
this model. This result is consistent with a more recent
interpretation of the change of charge isotope shifts across the
$N=126$ shell gap \cite{GSR.13}, in which the kink is attributed to
the occupation of the $ 1 i_{11/2}$ neutron orbital and the resulting
overlap between neutron and proton orbitals with the same principal
quantum number, $n=1$. It was noted that effective forces for which
the $ 1 i_{11/2}$ neutron orbital has a significant occupation above
$N=126$, display an increase in the isotope shift of the $n=1$ proton
states. This is because when neutrons gradually occupy the $ 1
i_{11/2}$ orbital, proton states expand to larger radii to maximally
overlap with the additional neutrons \cite{GSR.13}. In the present
calculation, both for DD-PC1 and PKO2, the neutron orbitals $ 1
i_{11/2}$ and $ 2 g_{9/2}$ are almost degenerate above $N=126$, and
this leads to significant occupation of $ 1 i_{11/2}$ and the
resulting sudden increase in the isotope shifts. In fact, the
quasi-degeneracy of $ 2 g_{9/2}$ and $ 1 i_{11/2}$ corresponds to an
approximate realization of pseudospin symmetry of single-nucleon
states with $(n, l, j=l+1/2)$ and $(n-1, l+2, j=l+3/2)$. When the
Fierz transformation is performed on the point-coupling approximation
of PKO2, the equivalent RMF Lagrangian leads to the lowering of the
orbital $ 1 i_{11/2}$ below $ 2 g_{9/2}$. Although it cannot directly
be compared to data \cite{duguet}, this discrepancy with the
experimental spectra of $^{209}$Pb and $^{211}$Pb is probably caused
by the fact that the parameters of the equivalent Lagrangian are not
fine-tuned after performing the point-coupling approximation.
Nevertheless, it leads to the pronounced kink shown in
Fig.~\ref{fig:3}. In fact we note that the best agreement with the
empirical kink is obtained with those interactions for which
pseudospin symmetry is realized in the single-neutron spectra (here
DD-PC1 and PKO2, but also other relativistic interactions). If this
symmetry is broken by further lowering $ 1 i_{11/2}$, below $ 2
g_{9/2}$, the kink in the isotope shifts becomes too strong compared
to data (cf. Fig.~\ref{fig:3}). The presence of the kink in the
isotope shifts and the relativistic models that reproduce the data
thus provide evidence for the occurrence of pseudospin symmetry in
neutron-rich Pb nuclei.
In conclusion, we have analyzed the isospin dependence of the
effective spin-orbit potential in standard relativistic meson-exchange
or point-coupling (contact) effective interactions, when used in the
mean-field (Hartree) or Hartree-Fock approximations. By performing a
non-relativistic reduction of the single-nucleon Dirac equation to a
Schr\" odinger-like form that explicitly exhibits the spin-orbit
potential, the corresponding isospin dependence can be directly
compared to that of the non-relativistic Hartree-Fock models based on
effective Skyrme forces. This isospin dependence can be characterised
by the ratio ${W_1}/{W_2}$ of the two parameters in the expression for
the effective spin-orbit potential Eq.~(\ref{W1andW2}). In
conventional non-relativistic Hartree-Fock mean-field calculations
based on the Skyrme force $W_1/W_2 = 2$, whereas in standard RMF
models this ratio is close to 1. The deviation from 1 arises because
of the explicit isovector contribution to the spin-orbit potential.
It should be noted that, because of the medium-dependence of the
effective coupling parameters, either through an explicit density
dependence or higher-order self-interaction terms, the ratio
${W_1}/{W_2}$ is density dependent in the relativistic approach. In
the case of relativistic Hartree-Fock models, to evaluate the effect
of exchange terms we have performed a Fierz transformation of the
point-coupling RHF Lagrangian and derived an equivalent RMF Lagrangian
that, in addition to the isovector-vector contribution of the original
RHF Lagrangian ($\rho$-meson exchange), contains also an
isovector-scalar term. As a result, the ratio ${W_1}/{W_2} \approx
1.8$ is much closer to the value that characterises standard Skyrme
Hartree-Fock models. This result is important in view of recent
experimental efforts to explore the isospin dependence of spin-orbit
forces in nuclei. It shows that, when comparing with results obtained
using conventional Skyrme HF models, Fock terms should also be treated
explicitly in relativistic mean-field models or, if one wants to
preserve the advantage of local density functionals, the
isovector-scalar channel has to be taken into account in addition to
the usual isovector-vector contribution. Of course, this channel has
been considered before in relativistic structure models, however the
standard data (masses, radii) could not be used to discern between the
two isovector channels. Information on the isospin dependence of the
energy spacings between spin-orbit partner states could thus be used
to determine the isovector-scalar channel contribution. We have also
shown that the reproduction of the empirical kink in the isotope
shifts of even Pb nuclei by relativistic effective interactions points
to the occurrence of pseudospin symmetry in the single-neutron spectra
in these nuclei.
\section*{Acknowledgement}
The authors thank T. Nik\v{s}i\'{c} and O. Sorlin for fruitful discussions. This work has been supported in part by the QuantiXLie Centre of Excellence.
|
\section{Introduction}\label{sec:introduction}
\begin{center}
{\em Causality versus localizabity.}
\end{center}
Causality refers to the physical principle according to which information propagates at a bounded speed. Localizability refers to the principle that all must emerge constructively from underlying local mechanisms, that govern the interactions of close by systems.
Classically there may not be much difference between the two. Consider Cellular Automata (CA), for instance, i.e. a grid of cells, each of which may take one in a finite number of possible states. Causality in this context states that the next state of a cell must be a function of its current state and that of its neighbours. But then this update function readily provides us with the underlying local mechanism demanded by localizability. Things get much more involved if causality is relaxed to its topological characterization \cite{Hedlund}, and if the grid is relaxed to time-varying graphs \cite{ArrighiCGD,ArrighiRC}, a.k.a. for Causal Graph Dynamics (CGD). Still, causality is shown to imply localizability.\\
In the reversible setting causality is no different, but localizability is more stringent, because the local mechanism must itself be reversible. Still it was shown that Reversible CA decompose as a finite-depth circuit of reversible, local gates \cite{KariBlock,KariCircuit,Durand-LoseBlock}. The same holds true for Reversible CGD \cite{ArrighiBRCGD} in spite of the dynamicity of the neighbourhood relation. In the probabilistic setting, however, the implication fails \cite{Henson,ArrighiPCA}.
It may therefore come as a surprise that, in the quantum setting, unitarity plus causality implies localizability. This was show successively for two systems \cite{Beckman}, three systems \cite{SchumacherWestmoreland}, a line of systems \cite{SchumacherWerner,ArrighiLATA,ArrighiIJUC} and eventually for an arbitrary fixed graph of systems \cite{ArrighiJCSS} --- encompassing Quantum CA.
In this paper we prove that, in the context of unitary evolutions of quantum superpositions of graphs, causality implies localizability.
\begin{center}
{\em Quantum superpositons of graphs.}
\end{center}
Picture yourself a graph having quantum systems lying at each node. Suppose that the whole thing evolves in discrete time steps, according to a global unitary operator. But in such a way as to respect the graph: in one time step, information propagates from one node to another only if they are close by in the graph. This can all be defined and studied, these are unitary causal operators \cite{ArrighiJCSS}. But now, suppose that the graph itself is subject to the evolution, and gets driven to be in a quantum superposition of graphs---in accordance to the superposition principle. What does it mean to be causal in this strange context? When two nodes are now connected and disconnected, in a superposition, can they signal?
In this paper we propose and formalize a notion of causality in the context of quantum superpositions of time-varying graphs. The notion is well-behaved. In the quantum-but-fixed-topology regime, it specializes down to the more usual notion of causality used for Quantum Cellular Automata or in Algebraic Quantum Field Theory. In the classical-but-dynamical-topology regime, it specializes down to that used for Causal Graph Dynamics. It admits both a Shr\"odinger form, and a dual Heisenberg form, which remain equivalent. We do the same applies to the notion localizability. Even the basic operations of tensor product and partial trace demand slight generalizations in order to address this context.
\begin{center}
{\em Motivations and result.}
\end{center}
Say that a Shr\"odinger cat is in a superposition of having fallen dead and being standing, alive. The cat's position changes the mass distribution in space, and so the curvature of space must also be in a superposition. What mathematical formalism can we use to describe this situation? Can we at least build a simple, discrete model that accounts for it? These sort of questioning have been at the heart of the research in Quantum Gravity.
Quantum Graphity \cite{QuantumGraphity1,QuantumGraphity2}, for instance, considers a complete graph whose edges each carry a qubit that says whether the edge is active or not. The whole thing evolves in continuous-time, according to a Hamiltonian which is a sum of nearest-neighbours---in the sense of being connected by an active edge. Similarly Causal Dynamical Triangulations \cite{LollCDT} considers simplicial complexes evolving according to a path-integral formalism. So is the case of
Loop Quantum Gravity \cite{RovelliLQG} in general, this time over a particular set of labelled graphs called spin foams. Much of the current research effort is dedicated towards understanding how an almost-flat space would emerge at large scale. We do not tackle this issue.
What we provide is a discrete-time formalism for these quantum superpositions of time-varying graphs. Moreover, we provide a structure theorem which decomposes the global unitary operator $U$, into a finite-depth circuit of local unitary gates $\mu$ and $K$. This circuit can be described by a formula:
$$U\ket{\psi} =(\prod \mu_u)(\prod K_u)\ket{\psi},$$
whose meaning is illustrated in Fig. \ref{fig:K} and made clear in the text. This makes the model constructive and parametrizable. Moreover only the $K_u$'s depends on which $U$ we decide to implement that way, and these commute: $[K_u,K_v]=0$. Hence the first product, which is the relevant one, can be spectrally decomposed.
\begin{figure}
\includegraphics[scale=0.7]{cgd_easy_rev_loc.pdf}
\caption{\label{fig:K} $K_u$ is a local unitary gate: it acts locally upon vertex $u$ and its neighbours, and may create superpositions of graphs. Consider global, unitary causal operator, which takes quantum superpositions of entire graphs into quantum superpositions of entire graphs---but in such a way that information propagates from one node to another only if they are neighbours in the graph. In this paper we show that unitary causal operators always decompose as a finite-depth circuit of such local unitary gates.}
\end{figure}
\begin{center}
{\em Plan.}
\end{center}
In Sec. \ref{sec:defs} we formalize our state space: superpositions of graphs. In Sec. \ref{sec:tens} we adapt the notion of tensor and partial trace to this state space. In Sec. \ref{sec:caus} and \ref{sec:loc} we propose and formalize the notions causality and locality in this context, and prove several propositions of general interest. In Sec. \ref{sec:mainth} we prove our main result. Sec. \ref{sec:conclusion} provides a summary and perspectives.
\section{Mathematical preliminaries}
\label{sec:defs}
The graphs we consider are the usual, finite, undirected, bounded-degree graphs, but with three added twists:
\begin{itemize}
\item[$\bullet$] Edges are between ports of vertices, rather than vertices themselves, so that each vertex can distinguish its different neighbours, via the port that connects to it.
\item[$\bullet$] The vertices are given labels taken in an alphabet, so that they may carry an internal state just like the cells of a Cellular Automaton.
\item[$\bullet$] The labelling function is partial, so that we may express our partial knowledge about part of a graph.
\end{itemize}
\begin{Def}[Graph]\label{def:graphs}
Consider $V$ a countable set of vertices and $\pi$ a finite set of ports.
A {\em graph} $G$ is given by
\begin{itemize}
\item[$\bullet$] A finite subset $V(G)$ of $V$, whose elements are called {\em vertices}.
\item[$\bullet$] A set $E(G)$ of non-intersecting two-elements subsets of $V(G)\times\pi$, whose elements are called {\em edges}. In other words an edge $e$ is of the form $\{u \!:\! a, v \!:\! b\}$, and $\forall e,e'\in E(G), e\cap e'\neq \varnothing \Rightarrow e=e'$.
\end{itemize}
The {\em set of all graphs} with vertices in $V$ and ports in $\pi$ is written ${\cal G}_{V,\pi}$. To ease notations, we write $v \in G$ for $v \in V(G)$. We write $\varnothing$ for the empty graph.
\end{Def}
\begin{Def}[Labelled graph]
A {\em labelled graph} is a triple $(G,\sigma)$, also denoted simply $G$ when unambiguous, where $G$ is a graph, and $\sigma$ is a partial function from $V(G)$ to a finite, or countable set $\Sigma$. The {\em set of all labelled graphs} with vertices in $V$, ports in $\pi$, and states in $\Sigma$ is written ${\cal G}_{V,\Sigma,\pi}$, or simply ${\cal G}$.
\end{Def}
We need a notion of {\em union} of graphs, and for this purpose we need a notion of {\em consistency} between the operands of the union, so as at to make sure that both graphs ``do not disagree''.
\begin{Def} [Consistent]
Two graphs $G$ and $H$ in ${\cal G}$ are {\em consistent} if and only if:
\begin{itemize}
\item over the set $W=(V(G)\cap V(H))$ the partial functions $\sigma_{G}$ and $\sigma_{H}$ agree when they are both defined, meaning that: $\forall u\in W$,
$$\big[u\in\textrm{dom}(\sigma_G)\wedge u\in\textrm{dom}(\sigma_H)\Rightarrow \sigma_G(u)=\sigma_H(u)\big]$$
where $\textrm{dom}(\sigma)$ stands for the domain of $\sigma$.
\item the edges adjacent to vertices of $W$ in $G$ and $H$ agree, meaning that: $\forall u\!:\! i\in (W\!:\!\pi)$, $\forall v\!:\! j\in V(G)\!:\!\pi$, $\forall v'\!:\! j' \in V(H)\!:\!\pi$:
\begin{align*}
\big[\{u\!:\! i, v\!:\! j\}\in E(G) \wedge \{u\!:\! i,& v'\!:\! j'\}\in E(H)\\ &\Rightarrow (v=v'\wedge j=j')\big]
\end{align*}
\end{itemize}
\end{Def}
\begin{Def} [Union]
Consider two consistent graphs $G$ and $H$, we define the graph
$G\cup H$ to be the graph:
\begin{itemize}
\item whose set of vertices $V(G\cup H)$ is $V(G)\cup V(H)$.
\item whose partial function $\sigma_{G\cup H}$ has domain $\textrm{dom}(\sigma_{G})\cup\textrm{dom}(\sigma_{H})$ and coincides with $\sigma_{G}$ (resp. $\sigma_H$) over $\textrm{dom}(\sigma_G)$ (resp. $\textrm{dom}(\sigma_H)$).
\end{itemize}
\end{Def}
We will also need ways of taking subgraphs that induced by the neighbours of a vertex.
\begin{Def}[Disks]\label{def:disks}
Consider $G\in {\cal G}$, and $S\subset V$. We write $D(G)^r_S$ for the {\em radius $r$ neighbours of $S$ in $G$}, i.e. all those vertices which can be reached in $r$ steps, or less, following edges of $G$, starting from a vertex in $S$. This includes $S$. The {\em border vertices of $S$}, on the other hand, are the radius $1$ neighbours of $S$ in $G$ which do not lie in $S$.\\
Now consider $D\subset V$, and $\overline{D}=V\setminus D$. We write $G_D$ for the subgraph induced by $D$ and its border vertices, all labellings included except those of the border vertices. We write $\overline{G}_D$ for the subgraph induced by $\overline{D}$, all labellings included.\\
In the special case where $D$ is $D(G)^r_v$ the set of radius $r$ neighbours of vertex $v$ in $G$, we simply write $G^r_v$ instead of $G_{D(G)^r_v}$, and refer to this as a {\em disk}. Similarly we write $\overline{G}^r_v$ instead of $\overline{G}_{D(G)^r_v}$ and refer to this as the {\em complement of a disk}.
\end{Def}
\begin{Rk} Notice that for all $G$ and $D\subset V$, $G=G_D\cup\overline{G}_D$. Indeed, the decomposition does not miss any out edge between $D$ and $\overline{D}$, as these belong to $G_D$.
\end{Rk}
Having defined the set of labelled graphs, which is infinite but countable, we can readily use it as the canonical basis for a Hilbert space of quantum superpositions of graphs.
\begin{Def}[Superpositions of configurations]~\label{def:HCfbis}\\
We define ${\cal H}_{\cal G}$ be the Hilbert space of labelled graphs, as follows: to each labelled graph $G$ is associated a unit vector $\ket{G}$, such that the family $\pa{\ket{G}}_{G\in{\cal G}}$ is the canonical orthonormal basis of ${\cal H}_{\cal G}$. A \emph{state vector} is a unit vector $\ket{\psi}$ in ${\cal H}_{\cal G}$. A \emph{state} is a trace-one positive operator $\rho$ over ${\cal H}_{\cal G}$. To ease notations, we write $\mathcal{H}$ instead of ${\cal H}_{\cal G}$.
\end{Def}
\couic{
\begin{Def}[Isomorphism]
An isomorphism is specified by a bijection $R$ from $V$ to $V$ and acts on a graph $G$ of ${\cal G}$ as follow:
\begin{itemize}
\item $V(R(G))=R(V(G))$
\item $\{u:k,v:l\}\in E(G) \Leftrightarrow \{R(u):k,R(v):l\}\in E(R(G))$
\end{itemize}
The definition naturally extends to labelled graphs. It is linearly extended to any state of $\mathcal{H}$, thus defining a unitary operator.
\end{Def}
\begin{Def}[Dynamics]~\label{def:dyn}\\
A linear operator $U:\mathcal{H}\longrightarrow\mathcal{H}$ is said to be a \emph{dynamics} if and only if for any isomorphism $R$, we have $[U,R]=0$.
\end{Def}
}
\begin{Def}[Vertex preserving]~\label{def:vp}\\
A linear operator $U:\mathcal{H}\longrightarrow\mathcal{H}$ is said to be vertex preserving if and only if for any graph $G\in {\cal G}$, we have that if $U\ket{G}=\sum_i \alpha_i\ket{G(i)}$, then for all $i$, $V(G(i)) = V(G)$.
\end{Def}
\section{Generalized tensors and traces}\label{sec:tens}
In quantum theory, the tensor product is the basic operation used to mathematically represent the joint system of two systems next to one another. Here the systems are lie at vertices of a graph, and so we need to say how they connect to each other. Moreover, the graph itself may be in a superposition, it forms part of the state space. We need to adapt tensor products to this context.
\begin{Def}[Generalized Tensors]
We write $\ket{G}\otimes\ket{H}=\ket{G}\ket{H}$ for $\ket{G\cup H}$. The tensor product definition is then bilinearly extended to pairwise consistent superpositions of state vectors in ${\cal H}$.
\end{Def}
{\em Comment}. Usually a tensor product takes in states $\ket{\psi}^A$ and $\ket{\phi}^B$ from two non-overlapping systems $A$ and $B$, to produce $\ket{\psi}^A\ket{\phi}^B$. However, consider a common subsystem $C$ and input states from overlapping systems $AC$ and $BC$. If we demand that they have that the particular form $\ket{\psi}^A\ket{\vartheta}^C$ and $\ket{\phi}^B\ket{\vartheta}^C$, then we can naturally extend the tensor product to produce $\ket{\psi}^A\ket{\phi}^B\ket{\vartheta}^C$. This is what the above definition does: the consistency requirement amounts to imposing agreement upon the common subsystem.
\noindent Now, if $\rho$ captures the state of an entire system, then $\rho^r_v$ stands for the state of the neighbours of $v$:
\begin{Def}[Generalized partial trace]~\label{def:Tr}\\
Consider $\rho=\ket{G}\bra{H}$ over $\mathcal{H}$, with $G,H \in {\cal G}$.
Let $D=D(G)^r_v\cup D(H)^r_v$. We define its partial trace $\rho^r_v$ to be
$$\ket{G_D}\bra{H_D}\braket{\overline{H}_D}{\overline{G}_D}$$
The definition is linearly extended to any state over $\mathcal{H}$.
\end{Def}
\begin{Rk} From the previous remark, we get that $\textrm{Tr}(\rho)=\textrm{Tr}(\rho^r_v)$.
\end{Rk}
{\em Comment}. On the one hand, the above definition is a straightforward extension of the usual $\textrm{Tr}_B(\ket{ij}^{AB}\bra{kl})=\ket{i}^A\bra{k}\braket{j}{l}$ formula. On the other hand, the definition intends to let the system $A$ be the disk of radius $r$ centered on $v$---but this a priori is an unclear notion for quantum superpositions of graphs. This issue is solved by addressing the basic case, first, and then extending to superpositions.
\section{Causality}\label{sec:caus}
A fundamental symmetry of physics is causality, meaning that information propagates at a bounded speed. In discrete space and time, this means that in order to know the next state and connectivity of a vertex $v$, we only need to know that of its neighbours at the previous time step:
\begin{Def}[Causality]~\label{def:causality}\\
A linear operator $U:\mathcal{H}\longrightarrow\mathcal{H}$ is said to be \emph{causal} if and only if for all $m\geq 0$ there exists $n\geq 0$ such that for any state $\rho$ over $\mathcal{H}$, and for any $v\in V$, we have
\begin{align}
(U\rho U^\dagger)^m_v=(U\rho^n_v U^\dagger)^m_v. \label{eq:causality}
\end{align}
\end{Def}
{\em Comment}. The above is a direct translation of causality, as expressed with our generalized partial trace. Notice how, in the case of basic states, this definition of causality specializes into the usual notion of causality as in classical Causal Graph Dynamics \cite{ArrighiCayley,ArrighiRC}. Notice also how, in the case of fixed graphs whose nodes are labelled by quantum states, this definition of causality again specializes into that used for Quantum Cellular Automata \cite{ArrighiUCAUSAL,ArrighiJCSS}. What happens in the grey zone of quantum superpositions of graphs may seem a little wilder, but in the end it is just the linear extension of these notions.
The following proposition will turn out useful.
\begin{Pro}[Tensorial extension]\label{pro:tensext}
Given a causal unitary operator $U$ over ${\cal H}$,
\begin{align*}
U'':{\cal H}\otimes{\cal H}&\longrightarrow{\cal H}\otimes{\cal H}\\
U''&=U\otimes I
\end{align*}
is causal. Here, ${\cal H} \otimes{\cal H} \equiv {\cal H}_{{\cal G}^2} \equiv {\cal H}_{{\cal G} \uplus {\cal G}} $.
\end{Pro}
\noindent \textbf{Proof.}
\begin{align*}
(U''\rho' U''^\dagger)^m_v=&(U''(\sum_i\rho(i)\otimes\tau(i)) U''^\dagger)^m_v\\
=&\sum_i (U\rho(i)U^\dagger)^m_v\otimes\tau(i)^m_v\\
=&\sum_i (U\rho(i)^n_v U^\dagger)^m_v\otimes(\tau(i)^n_v)^m_v\\
=&(U''\sum_i(\rho(i)\otimes\tau(i))^n_v) U''^\dagger)^m_v\\
=&(U''\rho'^n_v U''^\dagger)^m_v
\end{align*}
~\hfill$\Box$
\section{Locality}\label{sec:loc}
Causal operators change the entire graph in one go. The word causal there
refers to the fact that information does not propagate too fast. Local operations,
on the other hand, act just in one bounded region of the graph, leaving the rest
unchanged:
\begin{Def}[Localization]\label{localizationdef}
A linear operator $A:\mathcal{H}\longrightarrow\mathcal{H}$ is said to be $r$-\emph{localized} upon $v\in V$ if and only if for any $G,H \in {\cal G}$, we have that
\begin{align}
\bra{H}A\ket{G}&=\bra{H_D}A\ket{G_D}\braket{\overline{H}_D}{\overline{G}_D} \label{eq:locality}
\end{align}
with $D=D(G)^r_v\cup D(H)^r_v$.
\end{Def}
In other words, $A$ only changes $v$ and its neighbours, and only requires knowledge of $v$ and its neighbours to do that. Thought of as a measurement, $A$ is only sensitive to changes in $v$ and its neighbours:
\begin{Pro}[Dual localization] \label{prop:dualloc}
Let $A$ be an $r$-local linear operator, upon a vertex $v$. This is equivalent to saying that for any two states $\rho, \rho'$ over $\mathcal{H}$, we have that $\rho^r_v=(\rho')^r_v$ entails that $\textrm{Tr}(A\rho)=\textrm{Tr}(A\rho')$.
\end{Pro}
$[\Rightarrow]$. Suppose that $A$ is $r$-localized upon a vertex $v\in V$. For any $\rho=\ket{G}\bra{H}$ let $D=D(G)^r_v\cup D(H)^r_v$. We have:
\begin{align*}
\textrm{Tr}\pa{A \rho}&=\bra{H}A\ket{G}\\
&= \bra{H_D}A\ket{G_D}\braket{\overline{H}_D}{\overline{G}_D}\\
&= \textrm{Tr}\pa{A\ket{G_D}\bra{H_D}\braket{\overline{H}_D}{\overline{G}_D}}\\
&= \textrm{Tr}\pa{A\rho^r_v}
\end{align*}
Assuming $\rho^r_v={\rho'}^r_v$ yields $\textrm{Tr}\pa{A \rho}=\textrm{Tr}\pa{A\rho^r_v}=\textrm{Tr}\pa{A\rho'^r_v}=\textrm{Tr}\pa{A \rho'}$.\\
$[\Leftarrow]$. Let us assume that $A$ is dually $r$-local around $v$. For any $\rho=\ket{G}\bra{H}$ Let $D=D(G)^r_v\cup D(H)^r_v$. We have:
\begin{align*}
\bra{H}A\ket{G}&=\textrm{Tr}\pa{A\rho}\\
&=\textrm{Tr}\pa{A\rho^r_v}\;\textrm{using dual localization.}\\
&=\textrm{Tr}\pa{A\ket{G_D}\bra{H_D}\braket{\overline{H}_D}{\overline{G}_D}}\\
&=\bra{H_D}A\ket{G_D}\braket{\overline{H}_D}{\overline{G}_D}
\end{align*}
Now that we have a notion of locality, we can rephrase that of causality in the Heisenberg picture, which is the more traditional one in algebraic quantum field theory for instance.
\begin{Pro}[Dual causality]~\\ \label{prop:dual}
Let $U$ be a causal linear operator. This is equivalent to saying that for every operator $A$ $m$-localized upon vertex $v$, then $U^\dagger AU$ is $n$-localized upon vertex $v$.
\end{Pro}
\textbf{Proof.}
$[\Rightarrow]$. Suppose causality and let $A$ be an operator $m$-localized upon vertex $v$. Let $n$ be that from Def. \ref{def:causality}. For every pair of states $\rho$ and $\rho'$ such that
$\rho^n_v={\rho'}^n_v$, we have $\pa{U\rho U^\dagger}^m_v=\pa{U\rho' U^\dagger}^m_v$ and hence $\textrm{Tr}\pa{AU\rho U^\dagger}=\textrm{Tr}\pa{AU\rho' U^\dagger}$. We thus
get $\textrm{Tr}\pa{U^\dagger AU\rho}=\textrm{Tr}\pa{U^\dagger AU\rho'}$. Since
this equality holds for every $\rho$ and $\rho'$ such that
$\rho^n_v={\rho'}^n_v$, we have that $U^\dagger AU$ is $n$-localized.\\
$[\Leftarrow]$. Suppose dual causality and $\rho^n_v={\rho'}^n_v$.
Then, for every operator $B$ $n$-localized upon $v$,
$\textrm{Tr}\pa{B\rho}=\textrm{Tr}\pa{B\rho'}$, and so for every operator $A$ $m$-localized upon vertex $v$, we get:
$\textrm{Tr}\pa{AU\rho U^\dagger}=\textrm{Tr}\pa{U^\dagger
AU\rho}=\textrm{Tr}\pa{U^\dagger AU\rho'}
=\textrm{Tr}\pa{AU\rho'U^\dagger}.$
This entails $\pa{U\rho U^\dagger}^m_v=\pa{U\rho' U^\dagger}^m_v$.\hfill $\Box$\\
\section{Representation theorem}\label{sec:mainth}
The goal of this section is to achieve a representation of causal operators as a bounded-depth circuit of local unitary operators. The idea of this construction is to proceed by local updates. We will construct local operators $K_u$ updating only the neighbourhood of a vertex $u$. All these $K_u$ will be local unitary operators and will commute with each other. To do so, we generalize the construction presented in \cite{ArrighiBRCGD}: the local update $K_u$ consists in applying the causal operator $U$, ``putting aside'' vertex $u$ from the graph, and applying the inverse operator $U^\dagger$.
First, we construct an appropriate state space, allowing us to `mark' vertices in order to ``put them aside'' as computed.
\begin{Def}[Marked graphs and space]
Given a set of graphs ${\cal G}={\cal G}_{V,\Sigma,\pi}$, consider the set of graphs ${\cal G}_{V,\Sigma',\pi'}$ with $\Sigma'=\Sigma\times\{0,1\}$ and $\pi'=\pi\times\{0,1\}$. We define the set of marked graphs ${\cal G}'$, to be the subset of ${\cal G}_{V,\Sigma',\pi'}$ such that for all graph $G\in {\cal G}'$, for all vertex $u\in G$ with $\sigma_G(u)=(x,a)$ and $\{u:(i,b),v:(j,c)\}\in G$, we have $a=c$. We denote by ${\cal H}'$ the state space whose basis vectors are the marked graphs ${\cal G}'$. Given a graph $G\in \mathcal{G}$, it is naturally identified with the same graph in ${\cal G}'$ with all marks set to $0$.
\end{Def}
The following definition introduces our marking mechanism.
\begin{Def}[Mark operator]
Given a set of labels $\Sigma'=\Sigma\times\{0,1\}$ and a set of ports $\pi'=\pi\times\{0,1\}$, we define the marking operation $\mu(.)$ over labels and ports as toggling the bit in the second component:
\begin{itemize}
\item $\forall (x,a)\in \Sigma', \mu(x,a)=(x,1-a)$
\item $\forall (i,a)\in \pi', \mu(i,a)=(i,1-a)$
\end{itemize}
Then, we define the mark operation $\mu_u$ over marked graphs, as attempting to mark the label of vertex $u$ and its opposite ports, if this will not create conflicts between ports. More formally, given a graph $G$ in ${\cal G}'$, we define the mark operation, $\mu_u:{\cal G}'\rightarrow {\cal G}'$ as follows:
\begin{itemize}
\item if $\exists v,w\in G ,i,j\in \pi'$ such that $\{u \!:\! i , v \!:\! j\}\in G$ and $\{ v\!:\! \mu(j) ,w \!:\! k\}\in G$ then $\mu_u G=G$
\item else
\begin{itemize}
\item[$\bullet$] $\sigma_{\mu_u G}(u)=\mu(\sigma_{G}(u))$
\item[$\bullet$] For all $i,j\in\pi'$, $\{u\!:\! \mu(i),u\!:\! \mu(j)\} \in \mu_u G$ if and only if $\{u\!:\! i,u\!:\! j\} \in G$.
\item[$\bullet$] For all $v\in G$ with $v\neq u$ and $i,j\in\pi'$, $\{u\!:\! i,v\!:\! \mu(j)\} \in \mu_u G$ if and only if $\{u\!:\! i,v\!:\! j\} \in G$.
\end{itemize}
with the rest of the graph $G$ left unchanged.
\end{itemize}
Finally,
$\mu_u$ is linearly extended to become a unitary operator over ${\cal H}'$. Moreover each $\mu_u$ is $1$-localized and commutes with $\mu_v$ for all $v$.
\end{Def}
\noindent\textbf{Soundness.} As $\mu_u$ specifies a bijection over the set of graphs ${\cal G}'$, its linear extension to ${\cal H}'$ is unitary. Moreover, $\mu_u$ only changes the label of vertex $u$ and the ports of its adjacent edges, but it does so conditionally upon the edges of the neighbours, which makes it $1$-localized. Finally, when $u\neq v$, then $\mu_u$ and $\mu_v$ either both act independently upon disjoint labels and ports, or they are both the identity---hence they commute.
It turns out that any causal operator admits an extension that is compatible with these marks.
\begin{Pro}[Marked extension]
Given a vertex-preserving causal unitary operator $U$ over ${\cal H}$, there exists a vertex-preserving causal unitary operator $U'$ over ${\cal H}'$ such that:
\begin{align*}
\forall G\in {\cal G}_{V,\Sigma\times\{0\},\pi\times\{0\}},&\quad U'\ket{G}=U\ket{G}\\
\forall G\in {\cal G}_{V,\Sigma\times\{1\},\pi\times\{1\}},&\quad U'\ket{G}=\ket{G}
\end{align*}
\end{Pro}
\noindent \textbf{Proof.} Consider
\begin{align*}
U'':{\cal H}\otimes{\cal H} &\longrightarrow {\cal H}\otimes{\cal H}\\
U''&=U\otimes I
\end{align*}
instead---which is causal by Proposition \ref{pro:tensext}. Notice that ${\cal H}\otimes{\cal H}\equiv {\cal H}_{{\cal G}^2}$. We now consider
\begin{align*}
\varphi:{\cal G}'&\longrightarrow{\cal G}^2\\
G&\mapsto (\overline{G}_M,\mu_M G_M)
\end{align*}
with $M$ the marked vertices of $G$. The function $\varphi$ is injective since if $\varphi (G')=(G,H)$ then we can recover $G'$ as $G\cup (\mu_M H)$ with $M=V(H)\backslash V(G)$. Let $S=\varphi(\cal{G}')$ be the image of $\varphi$.
Notice that, if $(X,H) \in S$ and $Y\in {\cal G}$ is such that $V(X)=V(Y)$, then $(Y,H)\in S$. Indeed, if $(X,H)$ has antecedent $X'=X \cup (\mu_M H)$ then the union $Y'=Y \cup (\mu_M H)$ is well-defined as $\mu_M H$ does not specify any internal state or connectivity in $\pi\times\{0\}$ over $V(H) \cap V(X)=V(H) \cap V(Y)$. Moreover, $\varphi(Y')=(Y,H)$ because since $V(X)=V(Y)$, $\overline{Y'}_M = Y$ and $\mu_M Y'_M=\mu_M X'_M=H$.
The subspace ${\cal H}_S$ is stable under $U''$. Indeed, consider $(G,H)\in S$
\begin{align*}
U''(\ket{G}\otimes \ket{H}) &= (U \ket{G})\otimes \ket{H} \\
&= \sum \alpha_i\ket{G(i)}\otimes\ket{H}
\end{align*}
and notice that for all $i$, $(G(i),H)\in S$ because $V(G(i))= V(G)$.
Next, take $U'$ to be the restriction of $U''$ to $S$:
\begin{align*}
U':{\cal H}'&\longrightarrow {\cal H}'\\
\ket{G'}&\mapsto (\varphi^{\dagger} \circ U''\circ \varphi) \ket{G'}
\end{align*}
where $\varphi$ is linearly extended to be a unitary operator from ${\cal H}'$ to ${\cal H}_S$.
We have the two requested properties. Indeed, $\forall G\in {\cal G}_{V,\Sigma\times\{0\},\pi\times\{0\}}$,
\begin{align*}
U'\ket{G}&=\varphi^\dagger(U\otimes I)(\ket{G}\otimes\ket{\varnothing})\\
&=\varphi^\dagger(U\ket{G}\otimes\ket{\varnothing})\\
&=U\ket{G}
\end{align*}
and $\forall G\in {\cal G}_{V,\Sigma\times\{1\},\pi\times\{1\}}$,
\begin{align*}
U'\ket{G}&=\varphi^\dagger(U\otimes I)(\ket{\varnothing}\otimes\ket{\mu_{V(G)}\,G})\\
&=\varphi^\dagger(\ket{\varnothing}\otimes\ket{\mu_{V(G)}\,G})\\
&=\ket{G}
\end{align*}
Causality and vertex preservation are inherited from $U''$.
~\hfill$\Box$
The following theorem is our main contribution:
\begin{Th}[Structure theorem]~\\ \label{th:locrep}
Let $U$ be a vertex-preserving unitary causal operator over space ${\cal H}$. Then, in ${\cal H}'$ there exists $(K_u)$ such that for all $\ket{\psi}$:
$$U\ket{\psi}=(\prod_{u\in V} \mu_u)\,(\prod_{u\in V} K_u)\,\ket{\psi}$$
where $(K_u)$ is a collection of commuting unitary $n$-localized operators.
\end{Th}
{\bf Proof.} Let us consider a causal operator $U$ over ${\cal H}$.
We define $K_u$ as $U'^\dagger \mu_u U'$, where $U'$ is a marked extension of $U$.
Using the causality of $U'$ and the dual causality property, we have that $K_u$ is a $n$-localized operator. Moreover, it is easy to see that for two distinct vertices $u$ and $v$:
\begin{align*}
K_u K_v &=U'^\dagger \mu_u U'U'^\dagger \mu_v U'&\\
&=U'^\dagger \mu_u \mu_v U'&\\
&=U'^\dagger \mu_v \mu_u U'& \textrm{using commutativity of $\mu_\bullet$}\\
&=U'^\dagger \mu_v U'U'^\dagger \mu_u U'&\\
&=K_v K_u
\end{align*}
Hence, $(K_u)$ is a collection of localized commutating operators, thus the product $(\prod_u K_u)$ is well defined.\\
Now let us unfold the product $(\prod_u K_u)$ and apply it to $\ket{G}$, with $V(G)=\{u_1, \ldots, u_k\}$:
\begin{align*}
K_{u_1} K_{u_2} \cdots K_{u_k} \ket{G} & =U'^\dagger \mu_{u_1} U' U'^\dagger \mu_{u_2} U'\cdots U'^\dagger \mu_{u_k} U'\ket{G}\\
&=U'^\dagger \mu_{u_1} \mu_{u_2} \cdots \mu_{u_k} U' \ket{G}\\
&=U'^\dagger \mu_{u_1} \mu_{u_2} \cdots \mu_{u_k} U\ket{G}\\&\ \ \ \ \ \ \textrm{by construction of $U'$}\\
&= \mu_{u_1} \mu_{u_2} \cdots \mu_{u_k} U\ket{G}
\end{align*}
Hence, applying $(\prod_u \mu_u)(\prod_u K_u)$ results in applying $(\prod_u \mu_u)^2 U$ which is just $U$.
The general implications of this Theorem are discussed in Secs \ref{sec:introduction} and \ref{sec:conclusion}. Mathematically speaking, the following two corollaries immediately follow.
\begin{Cor}[Inverse of unitary causal is causal]~\label{cor:inverse}
Let $U$ be a vertex-preserving unitary causal operator over space ${\cal H}$. Then, $U^\dagger$ is also causal.
\end{Cor}
{\bf Proof outline.} The hypotheses imply localizability. The obtained circuit can then be reversed so as to implement $U^\dagger$. It follows that $U^\dagger$ is localizable, and hence causal.
\begin{Cor}[Unitary 1-causal is causal]~\label{cor:1causality}
An operator $U$ over $\mathcal{H}$ is said to be 1-\emph{causal} if and only if it is vertex preserving and there exists $n\geq 0$ such that for any state $\rho$ over $\mathcal{H}$, and for any $v\in V$, we have
\begin{align}
(U\rho U^\dagger)^1_v=(U\rho^n_v U^\dagger)^1_v. \label{eq:1causality}
\end{align}
Let $U$ be a vertex-preserving unitary $1$-causal operator over space ${\cal H}$. Then, $U$ is causal.
\end{Cor}
{\bf Proof outline.} By inspection of the proofs above, it suffices to be $1$-causal in order to be localizable. But then localizability implies causality.
Notice that we could have followed the same reasoning for $0$-causality if the mark operator could have been made $0$-local, but avoiding port conflicts forces it to be $1$-local. Ultimately the demand to avoid port conflict is to ensure that the graphs remain bounded degree at all times. This in turn is just a choice: with unbounded degree graphs \cite{QuantumGraphity1,QuantumGraphity2} this question would not arise, and $0$-causality would imply causality. But then these graphs would be much harder to interpret geometrically, as dual to pseudo-manifolds \cite{ArrighiSURFACES} or spin networks.
\section{Summary and future work}\label{sec:conclusion}
We took as our state space the Hilbert space of quantum superpositions of finite, undirected, bounded-degree, labelled graphs. We adapted the notions of tensor product, partial trace, causality and locality to this context, mainly through the idea that they should coincide with their traditional counterparts whenever the graph is not in a superposition---and extending from there by linearity. We recovered a number of reassuring results, such as the duality between causality phrased in the Shr\"odinger picture (i.e. $(U\rho U^\dagger)^m_v=(U\rho^n_v U^\dagger)^m_v$) and that phrased in the Heisenberg picture (i.e. if $A$ is localized, so is $U^\dagger A U$), and a similar duality for localizabity.\\
Mainly, we showed that any vertex-preserving, global unitary causal operator $U$, decomposes into a finite-depth circuit of local unitary gates $\mu$ and $K$:
$$U\ket{\psi}=(\prod \mu_u)(\prod K_u)\ket{\psi}$$
thereby making the model constructive and parametrizable. Here $\mu$ is just a fixed, marking operator: only the $K_u$'s depend upon $U$. These commute: $[K_u,K_v]=0$. Hence the first product, which is the relevant one, can be spectrally decomposed.
Here are two items for future work:
\begin{itemize}
\item The structure theorem was proven under the assumption that the global evolution is vertex-preserving. We wish to understand whether this condition can be relaxed. For instance, even in the model as it stands, we could attach to each vertex some `reservoir' of vertices, structured as an infinite binary tree. The quantum causal graph dynamics would then be able to `create' vertices by pulling them out of the reservoir, and to `destroy' them by pushing them back in. We plan to investigate this question in the near future.
\item Our framework is canonical, in the sense that it relies upon a distinguished discrete-time evolution. It cannot, therefore, be manifestly covariant in the sense of general relativity---but the perhaps covariance of some instances could be proven, e.g. in the style of \cite{arrighi2014discrete}.
\end{itemize}
\section{Quantum gravity landscape}
\noindent {\em Digital Physics}, of which a prominent actor is \cite{tHooftCA}, seeks to recover modern theoretical physics concepts as emergent from a Reversible Cellular Automata. The above-presented Quantum Causal Graph Dynamics clearly arises as a two-fold extension of Reversible Cellular Automata --- an extension to dynamical graphs on the one hand, and to quantum theory on the other hand. In this sense it shares common origins with the digital physics program, but it also departs from it: both quantum dynamics and geometrodynamics \cite{WheelerGeo} are seen as fundamental features that one needs to put in.
{\em Quantum Graphity} \cite{QuantumGraphity1,QuantumGraphity2} discretizes `pre-geometrical space' as a complete graph, with qubits on each edge telling whether its end vertices are neighbours or not. These qubits are then made to evolve according to a nearest-neighbours Hamiltonian. Quantum Graphity thus does not place space and time on a equal footing, as one is discrete and the other continuous. It follows that, strictly speaking, after any finite period of time, information has propagated everywhere: causality is approximate, in the sense of a Lieb-Robinson bound \cite{EisertSupersonic}. Quantum Causal Graph Dynamics can be seen as a theory of discrete-time quantum graphity.
{\em Quantum Causal Histories} \cite{FotiniQCH}, on the other hand, does look at global unitary evolutions between two spacelike surfaces---but the fact that it decomposes into local unitary gates is directly assumed in this approach. The local unitary gates are also located at each vertex, and act upon the quantum information circulating upon the edges---but have no influence upon the graph itself. The causal set is given, and in no superposition.
{\em Emergence of almost-flat space.} In quantum graphity, the underlying graph is complete, which makes its geometrical interpretation very difficult. Some bounded-degree graphs, on the other hand, can be understood as pseudo-manifolds \cite{Lickorish,GrasselliGems,ArrighiSURFACES}.
Thus a more long-term aim is to retake this inspiring program of emergence of almost-flat space \cite{QuantumGraphity1,QuantumGraphity2} but in this discrete-space discrete-time bounded-degree-graphs formalism. Alternatively and interestingly, almost-flatness can also be seen as an emergent property of certain probability distribution over graphs \cite{TrugenbergerEmergent1,TrugenbergerEmergent2}, e.g. via clustering \cite{KrioukovEmergent}.
{\em Loop Quantum Gravity} \cite{RovelliLQG}, one of the main contenders for a theory of Quantum Gravity, also provides means of computing the transition amplitude of one space-like graph evolving into another. Yet its relationship with the above-presented Quantum Causal Graph Dymanics is unclear to us at this stage. Indeed, in Loop Quantum Gravity, the transition amplitude between the two spin networks, that make up the past and future boundaries of a spacetime region, is provided in a path-integral form by summing over the possible spin foams that could relate them. It follows that: the `time taken for the transition to happen' is also summed over; that the evolution is not guaranteed to be unitary; and that far-away vertices of the spin networks are allowed to signal to some extent. Actually, spin networks are not directly interpretable as discretized space-like surfaces in Loop Quantum Gravity: only coherent superpositions of them correspond to piecewise-linear manifolds in a one-to-one manner. These are some of the key differences that stand in the way of bridging this gap, which we believe would be a fruitful program.
{\em Causal Dynamical Triangulations} \cite{LollCDT} lets some of these difficulites vanish. It lifts the ambiguity of spin networks and directly works with glued equilateral tetrahedras. It works out the transition amplitude between two discretized space-like space surfaces that are separated by a given proper time. It does so in a path-integral form, by summing over the possible successive surfaces that could relate them --- but these successive surfaces are themselves related by local 'moves'. Still it is unclear whether this induces a unitary evolution operator over discretized space-like space surfaces. Moreover here again and far-away parts of the triangulation are allowed to signal to some extent.
Summarizing, quantum causal graph dynamics provide a mathematical framework for discrete-time quantum gravity models that exhibit both strict causality and strict unitary. To the best of our knowledge, to this day none of the main quantum gravity models gathers all of these features at once. Several of them are not so far. Gathering the reminding features would make the model fall within the scope of the structure theorem of this paper, and therefore decompose into local unitary scattering matrices $K$. This would pave the way towards quantum simulating Quantum Gravity.
\section*{Acknowledgements} This work has been funded by the ANR-12-BS02-007-01 TARMAC grant, the ANR-10-JCJC-0208 CausaQ grant, the John Templeton Foundation grant ID 15619, the STICAmSud project 16STIC05 FoQCoSS. The authors acknowledge enlightening discussions with Marios Christodoulou, Gilles Dowek and Simon Perdrix.
|
\section{Introduction}\label{sec:intro}
Massive multiple-input multiple-output (MIMO) has emerged as one of the disruptive technologies for future fifth generation cellular networks, due to its potential benefits such as improvement of several orders of magnitude in spectral and energy efficiencies \cite{Boccardi2014}. The key idea behind massive MIMO is to deploy a very large number of antennas at the base station (BS) to serve many tens of users simultaneously. With such a deployment, massive MIMO reaps almost all the advantages offered by conventional MIMO, but on a much larger scale \cite{Yang2015}. A substantial implementation burden incurred by deploying hundreds of antennas is the channel estimation for a large number of channels. In order to ease this burden, the time division duplex (TDD) mode is adopted in massive MIMO networks such that the channels between the BS and the users are estimated via the uplink \cite{Lu2014}. Due to the assumption of channel reciprocity in the TDD mode, the estimated channel knowledge can be directly utilized for the downlink transmission \cite{Rusek2013}.
Pilot contamination is a practical problem that any massive MIMO network designer can face, which occurs when non-orthogonal pilot sequences are used across the whole network. In massive MIMO networks, the number of users is in general larger than the number of available pilot sequences. This indicates that the same pilot sequence needs to be assigned to two or more users, resulting in pilot contamination. Therefore, pilot contamination is identified as one of the main performance limiting factors in massive MIMO \cite{Marzetta2010,Larsson2014,Ashikhmin2012}. Several methods have been proposed to address the pilot contamination problem in massive MIMO networks,
such as protocol based methods \cite{Fernandes2013}, precoding based methods \cite{Jose2011}, angle-of-arrival based methods \cite{Yin2013}, and blind methods \cite{Muller2014}. While the aforementioned studies stand on their own merits, very little attention has been paid to the pilot sequence design, which decreases the negative effect of pilot contamination on the network performance. One example is \cite{Shen2015}, which designed pilot sequences and power allocation scheme for a single-cell massive MIMO network. For a multi-cell massive MIMO network, unfortunately, the design of load-achieving pilot sequences has not been explored in the literature. Despite its practical significance, such design is not trivial as multiple cells impose additional constraints that need to be satisfied, e.g., the per-cell quality of service requirements of users.
In this paper, we propose a novel pilot sequence design for an $L$-cell massive MIMO network. In each cell, an $N_{t}$-antenna BS communicates with $K$ single-antenna users in the TDD mode. In our design, the pilot sequences are generated for all the users in the network as per the rules of the generalized Welch bound equality (GWBE) sequence design \cite{Waldron2003}. The rationale behind choosing GWBE lies in its ability of achieving user capacity \cite{Ulukus2001} in code-division-multiple-access (CDMA) systems \cite{Cotae2006}. Notably, the GWBE design for a single cell CDMA system can not be directly utilized in multi-cell massive MIMO networks. Moreover, our design guarantees low correlation between different pilot sequences and thus reduces the detrimental impact of pilot contamination. The primary contributions of this paper are summarized as follows:
\begin{enumerate}
\item
We derive a new expression for the user load of the multi-cell massive MIMO network. The user load is defined as the number of users that can be simultaneously served, with their signal-to-interference-plus-noise ratio (SINR) requirements being satisfied. We then determine the load region of the network, under which the newly derived user load is achieved.
\item
We propose an easy-to-implement algorithm which produces load-achieving pilot sequences satisfying the SINR requirements at all the users within the network. Based on the algorithm, we also determine the power allocation for the downlink data transmission.
\item
We present numerical results to provide useful insights into the advantages of our proposed GWBE design over the existing pilot sequence designs. We show that our design achieves a larger load region and supports a greater range of SINR requirement than the existing designs. We further show that our design satisfies the SINR requirement with finite $N_{t}$, due to the larger load region it achieves, while the existing designs do not, even with infinite $N_{t}$.
\end{enumerate}
\section{Multi-Cell Massive MIMO Networks}\label{sec:system}
In this work we consider an $L$-cell TDD massive MIMO network. In each cell, an $N_{t}$-antenna BS communicates with $K$ single-antenna users. We denote $K_{tot}$ as the total number of users in the network, where $K_{tot}=KL$. In this network we consider both large-scale and small-scale propagation effects. Specifically, we denote $\sqrt{\beta_{i_{j}l}}h_{i_{j}l_{n}}$ as the propagation factor between the $j$th user in the $i$th cell and the $n$th BS antenna in the $l$th cell, where $i\in\left\{1,\dotsc,L\right\}$, $j\in\left\{1,\dotsc,K\right\}$, $l\in\left\{1,\dotsc,L\right\}$, and $n\in\left\{1,\dotsc,N_{t}\right\}$. Here, $\beta_{i_{j}l}$ characterizes the large-scale path loss effect from the $j$th user in the $i$th cell to the BS in the $l$th cell and $h_{i_{j}l_{n}}$ characterizes the small-scale multipath fading effect from the $j$th user in the $i$th cell to the $n$th BS antenna in the $l$th cell. In particular, we assume that $h_{i_{j}l_{n}}$ follows an independent and identically distributed (i.i.d) Rayleigh fading model, i.e., $h_{i_{j}l_{n}}\sim\mathcal{CN}(0,1)$. We also assume a block fading channel model, where the channel remains constant during the coherence time interval of $T$ but changes independently every interval. We further assume channel reciprocity between the uplink and the downlink, which is typical in TDD systems \cite{Jose2011}. Under this assumption, the propagation factor estimated via the uplink can be utilized for the downlink transmission.
\subsection{Channel Estimation via the Uplink}
We first focus on the channel estimation process in the uplink. In this process, the propagation factors in each cell are estimated by the BS using the pilot sequences sent by $K$ same-cell users. All pilot sequences are assumed to have unit energy and the length of $\tau$. We assume perfect synchronization between the uplink pilot sequences, which is regarded as the worst case scenario for pilot contamination \cite{Yin2013}. Noticeably, synchronization errors result in decoration between the pilot sequences. The pilot sequence vector received at the $\textrm{BS}$ in the $l$th cell during the uplink training phase, denoted by a $\tau M \times 1$ vector, $\mathbf{s}_{l}$, is given by
\begin{align}\label{rec_pilot}
\mathbf{s}_{l}=\sum_{i=1}^{L}\sum_{j=1}^{K}\eta_{i_{j}l}{\mathbf{Q}_{i_{j}}}\mathbf{h}_{i_{j}l}+\mathbf{n}_{l},
\end{align}
where $\eta_{i_{j}l} = \sqrt{p_{i_{j}}\beta_{i_{j}l}}$, $\mathbf{Q}_{i_{j}} = {\mathbf{q}_{i_{j}}}\otimes \mathbf{I}_{N_{t}}$ is a $\tau N_{t} \times N_{t}$ matrix, $\mathbf{q}_{i_{j}}$ is the $\tau\times1$ pilot sequence assigned to the $j$th user in the $i$th cell, $\otimes$ denotes the Kronecker product, $\mathbf{I}_{N_{t}}$ denotes the $N_{t}\times{}N_{t}$ identity matrix, $p_{i_{j}}$ is the pilot power for the $j$th user in the $i$th cell, $\mathbf{h}_{i_{j}l}=[h_{i_{j}l_{1}},h_{i_{j}l_{2}},\dotsc,h_{i_{j}l_{n}}]^{T}$ is an $N_{t}\times1$ uplink channel vector from the $j$th user in the $i$th cell to the BS in the $l$th cell, and $\mathbf{n}_{l}$ is the $\tau{}N_{t}\times1$ additive white Gaussian noise (AWGN) at the BS in the $l$th cell.
We assume that the least square (LS) channel estimation method is adopted. It has been demonstrated that mean square error of an LS channel estimator remains nearly constant as $N_{t}$ increases \cite{Khansefid2015}, which makes it attractive for massive MIMO networks. Accordingly, the uplink channel from the $k$th user in the $l$th cell to the BS in the $l$th cell is obtained by utilizing the property of the pilot sequence matrix, given by $\mathbf{Q}_{l_{k}}^{T}\mathbf{Q}_{l_{k}}=\mathbf{I}_{N_{t}}$. Based on \eqref{rec_pilot} and assuming that the uplink power control is enabled with $\eta_{l_{k}l}=1$, we obtain the LS channel estimate as
\begin{align}\label{channel_estimate_2}
\mathbf{\hat{g}}_{l_{k}l}=\mathbf{Q}_{l_{k}}^{T}\mathbf{s}_{l}= \mathbf{h}_{l_{k}l}+\sum_{i,j \neq l,k}\eta_{i_{j}l}\rho_{i_{j}l_{k}}\mathbf{h}_{i_{j}l}+{\mathbf{Q}_{l_{k}}^T}\mathbf{n}_{l},
\end{align}
where $\mathbf{Q}_{l_{k}}^{T}$ denotes the transpose of $\mathbf{Q}_{l_{k}}$, ${\sum}_{i,j\neq{}l,k}=\sum_{i=1}^{L}\sum_{j=1}^{K}$ with $(i,j)\neq(l,k)$, and $\rho_{i_{j}l_{k}}$ is the correlation coefficient between pilot sequences, defined as $\rho_{i_{j}l_{k}}=\mathbf{q}_{l_{k}}^{T}\mathbf{q}_{i_{j}}$, $k \in \left\{1,2,\dotsc, K\right\}$. We note that the value of $\rho_{i_{j}l_{k}}$ varies from $+1$ to $-1$, where $+1$ and $-1$ indicate a perfect positive correlation and a perfect negative correlation between pilot sequences, respectively, while 0 indicates no correlation (or equivalently, orthogonal pilot sequences). It is evident from \eqref{channel_estimate_2} that the use of non-orthogonal pilot sequences for different users in the network, i.e., $\rho_{i_{j}l_{k}}\neq0$, contaminates the uplink channel estimate. This effect is referred to as pilot contamination, which significantly deteriorates the performance of massive MIMO networks.
\subsection{Data Transmission via the Downlink}
We now concentrate on the downlink data transmission. In this transmission, a data symbol $x_{l_{k}}$ is sent to the $k$th user in the $l$th cell from the same-cell BS. The transmit power of $x_{l_{k}}$ at the BS is given by $\mathbb{E}\left[x_{l_{k}}^H{x}_{l_{k}}\right] = P_{l_{k}}$, where $\mathbb{E}[\cdot]$ denotes expectation. We assume that the data symbols are uncorrelated zero mean symbols and linearly precoded by a precoding vector $\mathbf{a}$. Thus, the precoded downlink transmission received by the $k$th user in the $l$th cell is given by
\begin{align}\label{rec_initial}
\hat{y}_{l_{k}}=
\sum_{m=1}^{L}\sum_{n=1}^{K}\sqrt{\beta_{l_{k}m}}\mathbf{h}_{l_{k}m}^{H}\left(\mathbf{a}_{m_{n}}x_{m_{n}}\right)+w_{l_{k}},
\end{align}
where $w_{l_{k}}$ is the AWGN at the $k$th user in the $l$th cell. Assuming that only the statistical information of the channel is available at the user \cite{Jose2011,Shen2015}, we rewrite $\hat{y}_{l_{k}}$ in \eqref{rec_initial} as
\begin{align}\label{received_siga}
\hat{y}_{l_{k}}=\sqrt{\beta_{l_{k}l}}\mathbb{E}\left[\mathbf{h}_{l_{k}l}^H\mathbf{a}_{l_{k}}\right]x_{l_{k}}+u_{l_{k}},
\end{align}
where $u_{l_{k}}=\sqrt{\beta_{l_{k}l}}\left({\mathbf{h}_{l_{k}l}^H\mathbf{a}_{l_{k}}}
-\mathbb{E}\left[{\mathbf{h}_{l_{k}l}^H\mathbf{a}_{l_{k}}}\right]\right)x_{l_{k}}+\sum_{m,n \neq l,k}\sqrt{\beta_{l_{k}m}}\mathbf{h}_{l_{k}m}^{H}\left(\mathbf{a}_{m_{n}}x_{m_{n}}\right)+w_{l_{k}}$.
We clarify that the term $u_{l_{k}}$ can be treated as the effective noise and is uncorrelated with $\sqrt{\beta_{l_{k}l}}\mathbb{E}\left[{\mathbf{h}_{l_{k}l}}\right]x_{l_{k}}$. With the aid of \eqref{received_siga}, we determine the SINR at the users, evaluate the user load, and design the pilot sequences for the multi-cell massive MIMO network in Section \ref{sec:design}.
\section{User Load Analysis and Pilot Sequence Design}\label{sec:design}
In this section, we first derive a new expression for the user load in the multi-cell massive MIMO network. We then determine the load region of the network, under which the derived user load is achieved. We further propose an easy-to-implement algorithm to design pilot sequences that satisfy the SINR requirements at users and achieve the user load.
\subsection{Analysis of User Load}\label{sec:user_capacity_analysis}
Throughout this paper, the user load is defined as the number of users that can be simultaneously served via the downlink in the massive MIMO network such that SINR requirements of all the users are satisfied. Here, we preserve a practical assumption that the number of users in each cell is higher than the length of the pilot sequence, i.e., $K>\tau$. This is due to the fact that the massive MIMO BS typically serves a huge number of users using a limited number of pilot sequences. Under this assumption, the network performance suffers from both inter-cell pilot contamination and intra-cell pilot contamination, which is treated as a worst-case scenario for a pilot contaminated massive MIMO network.
\subsubsection{Signal-to-Interference-Plus-Noise Ratio at Users}
We commence our analysis by formulating the achievable SINR for the $k$th user in the $l$th cell, denoted by $\phi_{l_{k},N_{t}}$. Based on \eqref{received_siga}, we express $\phi_{l_{k},N_{t}}$ as
\begin{align}\label{long_exp}
\phi_{l_{k},N_{t}}=\frac{\left(\mathbb{E}\left[{\mathbf{h}_{l_{k}l}^{H}\mathbf{a}_{l_{k}}}\right]\right)^2\beta_{l_{k}l}P_{l_{k}}}
{\textrm{var}\left[{\mathbf{h}_{l_{k}l}^{H}\mathbf{a}_{l_{k}}}\right]\beta_{l_{k}l}P_{l_{k}} + \overline\phi_{l_{k},N_{t}} + \sigma_{w}^2},
\end{align}
where
\begin{align}\label{long_exp_phi_bar}
\hspace{-0.1cm}\overline\phi_{l_{k},N_{t}}= \sum_{m,n{}\neq{}l,k}\mathbb{E}\left[|{\mathbf{h}_{l_{k}m}^{H}\mathbf{a}_{m_{n}}}|^{2}\right]\beta_{l_{k}m}P_{m_{n}},
\end{align}
$\textrm{var}\left[\cdot\right]$ denotes the variance operation, and $\sigma_{w}^2$ is the variance of $w_{l_{k}}$.
We note that the achievable SINR given by \eqref{long_exp} is a generalized expression which is valid for any precoder. We now determine the achievable SINR with the maximum-ratio transmission
(MRT) precoder. Using \eqref{channel_estimate_2} and the channel hardening property of massive MIMO, the MRT precoding vector for the $k$th user in the $l$th cell is given by
\begin{align}\label{MRT_precoder}
\mathbf{a}_{l_{k}}=\frac{\mathbf{\hat{g}}_{l_{k}l}}{\|\mathbf{\hat{g}}_{l_{k}l}\|}=
\frac{\mathbf{\hat{g}}_{l_{k}l}}{\sqrt{N_{t}
\left(\mathbf{\hat{g}}_{l_{k}l}^{H}\mathbf{\hat{g}}_{l_{k}l}/N_{t}\right)}}
=\frac{\mathbf{\hat{g}}_{l_{k}l}}{\sqrt{N_{t}\delta_{l_{k}}}},
\end{align}
where $\|\cdot\|$ denotes the $l_2$ norm and $\delta_{l_{k}}=\sum_{i=1}^{L}\sum_{j=1}^{K}\eta_{i_{j}l}^2\rho_{i_{j}l_{k}}^{2}+\sigma_{n_{l}}^{2}$. Using the MRT precoding indicated by \eqref{MRT_precoder} together with the LS channel estimation, we present a simplified expression for $\phi_{l_{k},N_{t}}$ in the following Lemma.
\newtheorem{lemma}{Lemma}
\begin{lemma}\label{lemma_red}
If the MRT precoding is used with the LS channel estimation, the achievable SINR is simplified as
\begin{align}\label{SINR}
\phi_{l_{k},N_{t}}=\frac{\beta_{l_{k}l}P_{l_{k}}}{\delta_{l_{k}}\left[\sum\limits_{m,n\neq{}l,k} \frac{\rho_{l_{k}m_{n}}^{2}\eta_{l_{k}m}^2\beta_{l_{k}m}P_{m_{n}}}{\delta_{m_{n}}}+\frac{1}{N_{t}}\left(\overline{P}_{lk}\right)
\right]},
\end{align}
\end{lemma}
where $\overline{P}_{lk}=\sum_{m=1}^{L}\sum_{n=1}^{K}\beta_{l_{k}m}P_{m_{n}} + \sigma_{w}^2$.
We next present an asymptotic expression for the achievable SINR given in \eqref{SINR} when $N_{t}\rightarrow\infty$. We note that $N_{t}\rightarrow\infty$ is a valid and widely-adopted assumption in massive MIMO networks. Under this assumption, the asymptotic expression for $\phi_{l_{k},N_{t}}$, denoted by $\phi_{l_{k},\infty}$, is derived as
\begin{align}\label{SINR_infa}
\phi_{l_{k},\infty}=\frac{\beta_{l_{k}l}P_{l_{k}}}{\delta_{l_{k}}\left(\sum\limits_{m=1}^{L}
\sum\limits_{n=1}^{K}\frac{\rho_{l_{k}m_{n}}^{2}\eta_{l_{k}m}^2\beta_{l_{k}m}P_{m_{n}}}{\delta_{m_{n}}}\right)-\beta_{l_{k}l}P_{l_{k}}}.
\end{align}
The asymptotic expression given by \eqref{SINR_infa} reveals that the pilot contamination is a performance limiting factor in massive MIMO networks, since $\rho_{l_{k}m_{n}}$ still exists and deteriorates the performance even when $N_{t}\rightarrow\infty$.
\subsubsection{User Load of the Network}
We now analyze the user load. We first simplify $\phi_{l_{k},\infty}$ given by \eqref{SINR_infa} using uplink power control assumption and re-express it as
\begin{align}\label{SINR_infb}
\phi_{l_{k},\infty} \geq \overline{\phi}_{l_{k},\infty}=\frac{P_{l_{k}}}{\delta_{l_{k}}\textrm{tr}\left(\mathbf{q}_{l_{k}}^{T} \mathbf{Q}\mathbf{D}\mathbf{A}\mathbf{Q}^{T}\mathbf{q}_{l_{k}}\right)-P_{l_{k}}},
\end{align}
where $\textrm{tr}\left(\cdot\right)$ denotes the trace operation and $\mathbf{Q}$, $\mathbf{D}$, and $\mathbf{A}$ are block matrices given by $\mathbf{Q}=\left[\mathbf{Q}_{1},\dotsc, \mathbf{Q}_{l},\dotsc,\mathbf{Q}_{L}\right]$, $\mathbf{D}=\text{diag}\left[\mathbf{D}_{1},\dotsc, \mathbf{D}_{l},\dotsc,\mathbf{D}_{L}\right]$, and $\mathbf{A}=\text{diag}\left[\mathbf{A}_{1},\dotsc,\mathbf{A}_{l},\dotsc,\mathbf{A}_{L}\right]$, respectively. Here, $\mathbf{Q}_{l}$ is the pilot sequence matrix for the $K$ users in the $l$th cell, given by $\mathbf{Q}_{l}=\left[\mathbf{q}_{l_{1}},\mathbf{q}_{l_{2}},\dotsc,\mathbf{q}_{l_{K}}\right]$, $\mathbf{D}_{l}$ is a diagonal matrix consisting of the transmit power at the BS in the $l$th cell for the $K$ same-cell users, given by $\mathbf{D}_{l}=\textrm{diag}\left\{P_{l_{1}}, P_{l_{2}}, \dotsc, P_{l_{K}}\right\}$, and $\mathbf{A}_{l}$ is a diagonal matrix consisting the inverse of parameter $\delta_{l_{k}}$ for all the $K$ users in the $l$th cell, given by $\mathbf{A}_{l}=\textrm{diag}\left\{1/\delta_{l_{1}}, 1/\delta_{l_{2}}, \dotsc, 1/\delta_{l_{K}}\right\}$, where $\textrm{diag}\{\cdot\}$ denotes a diagonal matrix with indicated elements along the diagonal.
Based on \eqref{SINR_infb}, we find that
\begin{align}\label{interme-appendix6}
\hspace{-0.2cm}\sum_{i=1}^{L}\sum_{j=1}^{K}\left(\frac{1+\overline{\phi}_{i_{j},\infty}}{\overline{\phi}_{i_{j},\infty}}\right)
=\textrm{tr}\left(\mathbf{D}^{-1}\mathbf{A}^{-1}\mathbf{Q}^{T}\mathbf{Q} \mathbf{D}\mathbf{A}\mathbf{Q}^{T}\mathbf{Q}\right).
\end{align}
By defining $\mathbf{R}_{S}\triangleq\mathbf{Q}^{T}\mathbf{Q}$ and $\mathbf{Z}\triangleq\mathbf{D}\mathbf{A}$ in \eqref{interme-appendix6}, we obtain
\begin{align}\label{interme-appendix7}
\sum_{i=1}^{L}\sum_{j=1}^{K}\left(\frac{1+\overline{\phi}_{i_{j},\infty}}{\overline{\phi}_{i_{j},\infty}}\right) &=\textrm{tr}\left(\mathbf{Z}^{-1}\mathbf{R}_{S}\mathbf{Z}\mathbf{R}_{S}\right),\notag\\
&\geq{}K_{tot}+\underbrace{\sum_{p=1}^{L}\sum_{q=1}^{K}\sum_{r=1}^{L}\sum_{s=1}^{K}}_{p>r,q>s}2\rho_{p_{q}r_{s}}^2,\notag\\
&\geq\textrm{tr}\left(\mathbf{R}_{S}\mathbf{R}_{S}\right).
\end{align}
We note that $\mathbf{R}_{S}$ in \eqref{interme-appendix7} is a symmetric matrix. By performing the eigen-decomposition of $\mathbf{R}_{S}$, we simplify \eqref{interme-appendix7} as
\begin{align}\label{interme-appendix8}
\sum_{i=1}^{L}\sum_{j=1}^{K}\left(\frac{1+\overline{\phi}_{i_{j},\infty}}{\overline{\phi}_{i_{j},\infty}}\right)\geq \sum_{i=1}^{K_{tot}}\lambda_{i}^{2}=\frac{1}{\tau}K_{tot}^{2},
\end{align}
where $\lambda_{i}$ is the $i$th eigenvalue of $\mathbf{R}_{S}$.
We denote $\gamma_{i_{j}}$ as the SINR requirement for the $j$th user in the $i$th cell. As such, the achievable SINR with infinite $N_{t}$ needs to be higher than or equal to $\gamma_{i_{j}}$, i.e., $\phi_{i_{j},\infty} \geq \overline{\phi}_{i_{j},\infty}\geq\gamma_{i_{j}}$. This indicates that $\sum_{i=1}^{L}\sum_{j=1}^{K}\left(\frac{1+\gamma_{i_{j}}}{\gamma_{i_{j}}}\right)
\geq\sum_{i=1}^{L}\sum_{j=1}^{K}\left(\frac{1+\overline{\phi}_{i_{j},\infty}}{\overline{\phi}_{i_{j},\infty}}\right)$. Using this inequality, \eqref{interme-appendix8} can be rewritten as
\begin{align}\label{interme-appendix10}
K_{tot}\leq\sqrt{\tau\sum_{i=1}^{L}\sum_{j=1}^{K}\left(\frac{1+\gamma_{i_{j}}}{\gamma_{i_{j}}}\right)}.
\end{align}
We clarify that \eqref{interme-appendix10} gives an upper bound on the user load of a multi-cell massive MIMO network.
\subsubsection{Load Region of the Network}
We next determine the load region of the network under which the previously derived user load can be achieved. By applying the Cauchy-Schwarz inequality, it is proven that the user load indicated by \eqref{interme-appendix10} is always achieved when the following condition holds:
\begin{align}\label{BW_all}
\sum_{i=1}^{L}\sum_{j=1}^{K}\left(\frac{\gamma_{i_{j}}}{1+\gamma_{i_{j}}}\right)\leq\tau.
\end{align}
We refer to the bound given by \eqref{BW_all} as the load region of the network. Under the load region, the user load given by \eqref{interme-appendix10} is achieved. Assuming that the load region is equally divided among the $L$ cells in the network, the upper bound on the load region for the $i$th cell is given by
\begin{align}\label{BW}
\sum_{j=1}^{K}\left(\frac{\gamma_{i_{j}}}{1+\gamma_{i_{j}}}\right)\leq\frac{\tau}{L}.
\end{align}
\subsection{Design of Pilot Sequences}\label{sec:pilot_sequence_design}
In this subsection, we propose an easy-to-implement algorithm to design the load-achieving pilot sequences for the multi-cell massive MIMO network. Here, we define the load-achieving pilot sequences as the sequences that satisfy the SINR requirements for all the users in the network and achieve the user load given by \eqref{interme-appendix10}.
In order to design the load-achieving pilot sequences, we define two $1\times{}K$ vectors $\mathbf{z}$ and $\mathbf{x}$, given by $\mathbf{z}=\left[\gamma_{l_{1}}/\left(1+\gamma_{l_{1}}\right),\gamma_{l_{2}}/\left(1+\gamma_{l_{2}}\right),\dotsc,\gamma_{l_{K}}/ \left(1+\gamma_{l_{K}}\right) \right]$ and $\mathbf{x}=\left[x_{1},x_{2},\cdots,x_{\tau},0,\cdots,0 \right]$, where $\gamma_{l_{1}}\geq\gamma_{l_{2}}\geq\dotsc\geq\gamma_{l_{K}}$ and $x_1\geq x_2\geq\dotsc\geq x_{\tau}$. We highlight that the values for $\gamma_{l_{k}}$ need to be chosen to satisfy \eqref{BW} with equality. We next present three preliminaries based on $\mathbf{x}$ and $\mathbf{z}$, as follows:
\newtheorem{preliminary}{Preliminary}
\begin{preliminary}\label{prem1}
Given vectors $\mathbf{z}$ and $\mathbf{x}$, $\mathbf{x}$ majorizes $\mathbf{z}$, i.e., $\mathbf{x}\succ\mathbf{z}$, if $\sum_{n=1}^{m}x_{n}\geq\sum_{n=1}^{m}z_{n}$, where $m\in\left\{1,\cdots,K-1\right\}$ and $\sum_{n=1}^{m}x_{n}=\sum_{n=1}^{m}z_{n}$, where $m=K$.
\end{preliminary}
\begin{preliminary}\label{prem2}
Given a vector $\mathbf{z}$, a vector $\mathbf{x}$ can be found for the value of $m=\tau$ such that $\mathbf{x}\succ\mathbf{z}$, if the vector $\mathbf{x}$ is given by $x_{i}= \sum_{n=1}^{K}\left(z_{n}/\tau\right)$, where $i\in\left\{1,\cdots,\tau\right\}$ and $x_{j}=0$, where $j\in\left\{\tau+1,\cdots,K\right\}$.
\end{preliminary}
\begin{preliminary}\label{prem3}
If $\mathbf{x}\succ\mathbf{z}$, $\mathbf{z}$ is obtainable by applying at most $K-1$ T--transform operation on $\mathbf{x}$, i.e., $\mathbf{z}=\mathbf{T}_{K-1}\mathbf{T}_{K-2}\cdots\mathbf{T}_{1}\mathbf{x}$, and there exists a matrix $\mathbf{U}=\mathbf{U}_{1}\mathbf{U}_{2}\cdots\mathbf{U}_{K-1}$, where $\mathbf{U}_{i}$ is a unitary matrix generated from $\mathbf{T}_{i}$ at each step of the T-transform \cite{Viswanath1999}.
\end{preliminary}
We now present the step-by-step procedure of the load-achieving pilot sequences design for the multi-cell massive MIMO network. Specifically, this procedure is detailed in \textbf{Algorithm~\ref{algo1}}. This algorithm uses a $1\times K_{tot}$ vector $\mathbf{\Gamma}=\left[\pmb{\gamma}_{1},\cdots,\pmb{\gamma}_{l},\cdots,\pmb{\gamma}_{L}\right]$ and the length of the pilot sequence $\tau$ as the input, where $\pmb{\gamma}_{l}=\left[\gamma_{l_1},\gamma_{l_2},\cdots,\gamma_{l_K}\right]$ is the $1\times{}K$ vector containing the minimum SINR requirements for the $K$ users in the $l$th cell. We clarify that the values of $\gamma_{l_k}$ need to be chosen such that \eqref{BW} is satisfied with equality, which in turn guarantees that \eqref{interme-appendix10} is achieved. The algorithm returns the pilot sequence matrix for the network, $\mathbf{Q}$, as the output.
\begin{algorithm}
\caption{Load--achieving pilot sequence design} \label{algo1}
\begin{algorithmic}[1]
\Procedure{Pilot Design}{$\mathbf{\Gamma},\tau$}\Comment{$\mathbf{\Gamma} = \left[\pmb{\gamma}_{1},\cdots,\pmb{\gamma}_{L}\right]$}
\For{$l\gets 1, L$}
\State $\pmb{\gamma} \gets \pmb{\gamma}_{l}$
\Comment{where $\gamma_{ij}\leq \nicefrac{1}{L-1}$}
\State $\textrm{sum} \gets 0$
\State $\mathbf{x}_{l} \gets \mathbf{0}_{1\times K}$ \Comment{$\mathbf{x}_{l}$ is a $1 \times K$ zero vector}
\State $\mathbf{z}_{l} \gets \mathbf{0}_{1\times K}$ \Comment{$\mathbf{z}_{l}$ is a $1 \times K$ zero vector}
\For{$k\gets 1, K$}
\State $\mathbf{z}_{l}(k)\gets \left(\nicefrac{\gamma_{k}}{1+\gamma_{k}}\right)$
\State $\textrm{sum} \gets \textrm{sum} + \mathbf{z}_{l}(k)$
\EndFor
\State $B_{l} \gets \frac{\textrm{sum}}{\tau}$
\State $\mathbf{x}_{l}\left(1,\cdots,\tau\right) \gets B_{l}$ \Comment{\parbox[t]{.40\linewidth}{First $\tau$ elements of $\mathbf{x}_{l}$ are set to $B_{l}$}}
\State $\mathbf{U}_{l} \gets \Call{T-transform}{\mathbf{z}_{l},\mathbf{x}_{l}}$
\State $\mathbf{V}_{l} \gets \mathbf{U}_{l}\left(\tau,:\right)$ \Comment{\parbox[t]{.48\linewidth}{$\mathbf{V}_{l}$ retains first $\tau$ rows of $\mathbf{U}_{l}$}}
\State $\mathbf{Z}_{l} \gets \text{diag}\{\mathbf{z}_{l}\}$ \Comment{$\mathbf{z}_{l}$ is a diagonal matrix}
\State $\mathbf{Q}_{l} \gets \text{normc}\left(B_{l}^{\frac{1}{2}}\mathbf{V}_{l} \mathbf{Z}_{l}^{-\frac{1}{2}}\right)$
\EndFor
\State $\mathbf{Q} \gets \left[\mathbf{Q}_{1},\cdots,\mathbf{Q}_{L}\right]$ \Comment{\parbox[t]{.45\linewidth}{$\mathbf{Q}$ is the desired pilot sequence matrix}}
\State \textbf{return} $\mathbf{Q}$
\EndProcedure
\end{algorithmic}
\end{algorithm}
\textbf{Algorithm~\ref{algo1}} considers the SINR requirements in one cell at each time. First, the algorithm obtains the SINR requirements for all the $K$ users in the $l$th cell. Second, the algorithm calculates the effective bandwidth for all the $K$ users in the $l$th cell and sets the values in $\mathbf{z}_{l}$. Third, the algorithm finds $\mathbf{x}_{l}$ from $\mathbf{z}_{l}$ using \emph{Preliminary \ref{prem2}}. A $K\times K$ matrix $\mathbf{U}_{l}$ is found by applying the T-transform to $\mathbf{x}_{l}$ and $\mathbf{z}_{l}$, as described in \emph{Preliminary \ref{prem3}}. The $\tau \times K$ matrix $\mathbf{V}_{l}$ only retains the first $\tau$ rows of the vector $\mathbf{U}_{l}$. Finally, the pilot sequence matrix for all the $K$ users in the $l$th cell is found by normalizing the columns of the matrix given by $B_{l}^{\frac{1}{2}}\mathbf{V}_{l}\mathbf{Z}_{l}^{-\frac{1}{2}}$, i.e., $\text{normc}\left(B_{l}^{\frac{1}{2}}\mathbf{V}_{l} \mathbf{Z}_{l}^{-\frac{1}{2}}\right)$, such that each column has unit energy.
The process is repeated for all the $L$ cells until $\mathbf{Q}$ is obtained for the whole network. The finally obtained pilot sequences are known as the GWBE pilot sequences, which are capable of achieving the user load. Notably, \textbf{Algorithm~\ref{algo1}} can be implemented among $L$ BSs in a distributed manner. In the distributed implementation, the BS in the $i$th cell can design $\mathbf{Q}_{i}$ for all the $K$ users in the $i$th cell, without requiring any feedback from other BSs.
We note that \textbf{Algorithm~\ref{algo1}} designs pilot sequences for the case that \eqref{BW} is satisfied with equality. We now focus on the case where \eqref{BW} is not satisfied with equality. Since $\gamma_{i_{j}}/\left(1+\gamma_{i_{j}}\right)$ is a monotonically increasing function, there exists some value, $\hat{\gamma}_{i_{j}}\geq\gamma_{i_{j}}$, such that the equality in \eqref{BW} holds, i.e., $\sum_{j=1}^{K}\left(\frac{\hat\gamma_{i_{j}}}{1+\hat\gamma_{i_{j}}}\right)=\frac{\tau}{L}$. The value of $\hat{\gamma}_{i_{j}}$ is then used in our proposed GWBE design. Accordingly, the downlink transmit power for the $j$th user in the $i$th cell is set as $P_{i_{j}}=\delta_{i_{j}}\hat\gamma_{i_{j}}/\left(1+\hat\gamma_{i_{j}}\right)$.
We clarify that the use of $\hat{\gamma}_{i_{j}}$ also guarantees that the SINR requirements for all the users in the network are met and the user load of the network is achieved.
\section{Numerical Results}\label{sec:Numerical}
In this section we present numerical results that demonstrate the advantage of our proposed pilot sequence design over the current designs. Specifically, we compare the performance of the proposed GWBE design with the performance of two well-known pilot sequence designs, namely, the Welch bound equality (WBE) design and the finite orthogonal set (FOS) design. In the WBE design \cite{Sarwate1999}, the correlation coefficient between different pilot sequences is fixed, which is given by $\rho_{i_{j}l_{k}}=\sqrt{\left(K-\tau\right)/\left(\left(K-1\right)\tau\right)}$, where $\left(i,j\right)\neq\left(l,k \right)$. Therefore, the value of $\delta_{i_{j}}$ is the same for all the users in the network, i.e., $\delta_{i_{j}}=\delta$. In the FOS design \cite{Hoydis2011}, the correlation coefficient between different pilot sequences is always zero. Hence, only the users with the same pilot sequence are considered in the comparison presented in this section. We note that the load regions of the WBE design and the FOS design in a single-cell massive MIMO network have already been found in \cite{Shen2015}. Here, we extend the results in \cite{Shen2015} to a multi-cell network, which facilitate our comparison. Throughout this section, we consider that $K=4$ users in each cell and the length of the pilot sequence is $\tau=3$.
\begin{figure}[!t]
\centering
\includegraphics[height=2.5in,width=3.2in]{area}
\caption{The upper surface boundary of the load regions versus the SINR requirements for the proposed GWBE design and the existing WBE and FOS designs.}
\label{capacity_region}
\end{figure}
We first compare the load region of the proposed GWBE design with those of the WBE and FOS designs. In this comparison we consider a multi-cell massive MIMO network with $L=3$ cells such that $K_{tot}=12$. The SINR requirements for the users in 3 cells are set to $\pmb{\gamma}_{1}=\pmb{\gamma}_{2}=\pmb{\gamma}_{3}=\left[\gamma_{1_{1}},\gamma_{1_{2}},\gamma_{1_{3}},0.1\right]$. For our GWBE design, the values of $\gamma_{i_{j}}$ need to be
carefully chosen such that the majorization condition given by \emph{Preliminary \ref{prem1}} always holds, i.e., $\gamma_{i_{j}}\leq{1}/\left(L-1\right)$. For the FOS design, it is assumed that one pilot sequence is simultaneously used by 6 users in the network. Fig.~\ref{capacity_region} depicts the upper surface boundary of the load regions for three designs. We find that the load region for the proposed GWBE design is $20.9\%$ and $73.5\%$ larger than the WBE and FOS designs, respectively. Importantly, a larger load region indicates that a group of users with higher SINR requirements can be simultaneously served in a pilot contaminated massive MIMO network. This is due to the fact that a larger load region offers more freedom for the users to choose their required SINR level in the network without derogating from the limitation imposed by pilot contamination.
\begin{figure}[!t]
\centering
\includegraphics[height=2.5in,width=3.2in]{num_cells}
\caption{The maximum permitted SINR versus the number of cells for the proposed GWBE design and the existing WBE and FOS designs.}
\label{num_cells}
\end{figure}
We now examine the effect of increasing the number of cells on the maximum permitted SINR in the $i$th cell, denoted by $\gamma_{i}^{\textrm{MAX}}$. Here, $\gamma_{i}^{\text{MAX}}$ is defined as the maximum value of the SINR requirements among the $K$ users in the $i$th cell, i.e. $\gamma_{i}^{\textrm{MAX}}=\max_{1\leq{j}\leq{K}}\gamma_{i_{j}}$.
In this examination, the SINR requirements for the users in the $L$ cells are given by $\pmb{\gamma}_{i}=\left[\gamma,\gamma,\gamma/2,\gamma/2\right]$, $\forall i\in\left\{1,\dots,L\right\}$. As such, we have $\gamma_{i}^{\textrm{MAX}}=\gamma$. Fig.~\ref{num_cells} depicts the maximum permitted SINR when the number of cells increases.
We first observe that the maximum permitted SINR decreases when $L$ increases. For example, when $L$ increases from 2 to 6, the maximum permitted SINR of our GWBE design decreases from 0.84 to 0.19. This is due to the fact that increasing $L$ reduces the load region, which in turn restricts the maximum permitted SINR. Second, we observe that the proposed GWBE design achieves a higher maximum permitted SINR than the WBE and FOS designs. For example, when $L=2$, the maximum permitted SINR of our GWBE design is 0.84, while those of the WBE and FOS designs are 0.57 and 0.46, respectively. Crucially, a higher maximum permitted SINR indicates that the network can support higher SINR requirements.
\begin{figure}[!t]
\centering
\includegraphics[height=2.5in,width=3.2in]{m_antenna_SINR31}
\caption{The achievable SINR versus the number of antennas for the first user in the third cell for the proposed GWBE design and the existing WBE and FOS designs.}
\label{achievable SINR}
\end{figure}
Finally, we compare the achievable SINR with a finite number of antennas of the proposed GWBE design with those of the WBE and FOS designs. In this comparison we consider $L=3$, $\sigma_{w}^{2}=p_{l_{k}}=1$, $\beta_{l_{k}m}=1$, where $l=m$, and $\beta_{l_{k}m}=0.9$, where $l\neq m$. The SINR requirements for the users in 3 cells are set to $\pmb{\gamma}_{1}=\left[0.45, 0.38, 0.25, 0.19\right]$, $\pmb{\gamma}_{2}=\left[0.43, 0.38, 0.28, 0.20\right]$, and $\pmb{\gamma}_{3}=\left[0.47, 0.43, 0.28, 0.13\right]$. We clarify that these SINR requirements remain within the load region for the proposed GWBE design but lie outside the load region for the WBE and FOS pilot designs. This is not surprising since the load region of our GWBE design is larger than those of the WBE and FOS designs, as depicted in Fig.~\ref{capacity_region}. As such, the SINR requirements considered in this comparison demonstrate how a diverse range of SINR requirements can be satisfied by our GWBE design.
As stated in Section~\ref{sec:pilot_sequence_design}, the load-achieving pilot design procedure of our design requires that the value of $\hat{\gamma}_{l_{k}}>\gamma_{l_{k}}$ needs to be chosen such that \eqref{BW_all} is satisfied with equality. As such, we choose $\hat{\pmb{\gamma}}$ as $\pmb{\hat{\gamma}}_{1}=\left[0.48, 0.40, 0.27, 0.21\right]$, $\pmb{\hat{\gamma}}_{2}=\left[0.45, 0.40, 0.30, 0.22\right]$, and
$\pmb{\hat{\gamma}}_{3}=\left[0.49, 0.45, 0.30, 0.15\right]$.
Fig.~\ref{achievable SINR} depicts the achievable SINR for the first user in the third cell using the three designs. A key observation from Fig.~\ref{achievable SINR} is that only the proposed GWBE design satisfies the SINR requirement for the user when the number of antennas exceeds some threshold, i.e., $\phi_{3_{1}}>\gamma_{3_{1}}$ when $N_{t}\geq 93$. In contrast, the WBE and FOS designs do not satisfy the SINR requirement for the user, no matter how large $N_{t}$ is. Notably, the achievable SINR for the WBE and FOS designs using an infinite number of antennas is still below the SINR requirement.
\section{Conclusions}\label{sec:con}
We proposed a novel GWBE design that generates low correlated pilot sequences to address the pilot contamination problem in multi-cell massive MIMO networks. We first derived a new expression for the user load, determined the load region of the network, and then designed an algorithm to produce load-achieving pilot sequences satisfying the SINR requirements for the users. Using numerical results we demonstrated the advantage of our proposed GWBE design over the existing WBE and FOS designs. In particular, we showed that our GWBE design achieves a higher load region and supports a greater range of SINR requirements than the existing designs. We further showed that our GWBE design can satisfy a higher pre-determined SINR requirement with a finite number of antennas, while the other two designs cannot satisfy the same requirements even with an infinite number of antennas.
|
\section{Introduction}
The prime number theorem in arithmetic progressions says that
\begin{equation}
\label{PNTAP}
\sum_{n\leq N}\Lambda(qn+b)= 1_{(q,b)=1}\frac{q}{\varphi(q)}N+o(N),
\end{equation}
where $\Lambda$ is the von Mangoldt function and $\varphi$ the Euler totient function, while $b$ and $q$ are two fixed integers. The Siegel-Walfisz theorem affirms that the asymptotic \eqref{PNTAP} holds uniformly in $b$ and $q$ in the regime where $q=q(N)$ is a function of $N$ satisfying $q=\log^{O(1)}N$ and $b$ varies as $N\log^{O(1)}N$.
Using the notation
$$
\Lambda_q(n)=\frac{q}{\varphi(q)}1_{(n,q)=1}
$$
for $n\in\mathbb{Z}$ or $n\in\mathbb{Z}/q\mathbb{Z}$ and multiplicativity, we can rewrite
equation \eqref{PNTAP} as
\begin{equation}
\label{PNTAP2}
\sum_{n\leq N}\Lambda(qn+b)= N(\prod_{p\mid q}\Lambda_p(b)+o(1)).
\end{equation}
A theorem of
Green and Tao \cite{GT2}, which we now state and which relies on two conjectures later resolved by Green, Tao and Ziegler \cite{MN,GI}, may be seen as a higher-dimensional analogue of
\eqref{PNTAP2}.
\begin{trm}
\label{GT}
Let $L$ be a constant and
$\Psi=(\psi_1,\ldots,\psi_{t}):\mathbb{Z}^d\rightarrow \mathbb{Z}^t$ be a system of affine-linear forms of which no two are affinely related.
Suppose that the coefficients of the
linear part $\dot{\Psi}$ are bounded by $L$, while the constant
coefficients $\psi_i(0)$ satisfy $\abs{\psi_i(0)}\leq LN$.
Let $K\subset [-N,N]^d$ be a convex body. Then
\begin{equation*}
\sum_{n\in\mathbb{Z}^d\cap K}\prod_{i=1}^{t}\Lambda(\psi_i(n))=
\beta_{\infty}\prod_p\beta_p+o_{d,t,L}(N^d),
\end{equation*}
where
$$\beta_{\infty}=\mathrm{Vol}(K\cap\Psi^{-1}(\mathbb{R}_+^t))$$
and
$$\beta_p=\mathbb{E}_{a\in (\mathbb{Z}/p\mathbb{Z})^d}\prod_{i=1}^t\Lambda_p(\psi_i(a)).$$
\end{trm}
Above we have used the notation
$$\mathbb{E}_{a\in A}f(a)=\frac{1}{\abs{A}}\sum_{a\in A}f(a)$$
to denote the averaging operator. We also agree that the letter $p$ is reserved
for primes, thus $\prod_p$ implicitly means $\prod_{p\in\mathcal{P}}$.
The factors $\beta_p$ are known as \emph{local factors}.
An interesting extension was obtained by the same authors, together with Ford and Konyagin \cite{GTFK}. They showed that Theorem \ref{GT} was still valid when the constant coefficients satisfied the less restrictive condition $\abs{\psi_i(0)}\leq N\log^C N$. This relaxed condition recently allowed Tao and Ziegler \cite[Theorem 1.3]{TZ} to obtain an improvement of the error term $o(N^d)$ to $o(\mathrm{Vol}(K))$ in the case where $K=[N]\times [M]^{d-1}$ with $M\gg N\log^{-O(1)}N$ and $\psi_i(n)=n_1+P_i(n_2,\ldots,n_{d})$ for some affine-linear forms $P_1,\ldots,P_t$ whose linear coefficients are bounded.
Here we prove a further extension, allowing unbounded linear coefficients. The impetus for the work came from a discussion on Tao's blog \cite{blog}. Before stating our theorem,
we collect several definitions pertaining to systems of
affine-linear forms.
\begin{dfn}
\label{dfn:systems}
A system $\Psi=(\psi_1,\ldots,\psi_{t}):\mathbb{Z}^d\rightarrow \mathbb{Z}^t$ with $d,t\geq 2$ of affine-linear forms
has \emph{finite complexity} if
no form is in the affine-linear span of another one.
It is called \emph{admissible}
if it is of finite complexity and if $\beta_p\neq 0$ for all $p$.
In particular, if $\Psi$ is admissible,
no prime $p$ divides all coefficients of any form $\psi_i$.
A system $\Psi'=(\psi_{i_1},\ldots,\psi_{i_s})$ for some sequence
$1\leq i_1<\ldots<i_s\leq t$ is called a \emph{subsystem} of $\Psi$.
The system $\Psi$ is called \emph{bounded} when all its linear coefficients are bounded (in terms of the asymptotic parameter $N$) and unbounded otherwise. Its \emph{size at scale $(N,B)$}
is defined as
$$
\nor{\Psi}_{N,B}=\frac{1}{\log^BN}\left(\sum_{i\in[t]}\abs{\frac{\psi_i(0)}{N}}+\sum_{i\in[t],j\in[d]}\abs{\dot{\psi}_i(e_j)}\right)
$$
where $(e_1,\ldots,e_d)$ is the canonical basis of the lattice $\mathbb{Z}^d$ and $\dot{\psi}_i$ denotes the linear part of $\psi_i$.
Furthermore, a prime $p$ is called \emph{exceptional} for $\Psi$ (and we write $p\in P_{\Psi}$) if there exist $i\neq j$ such that $\psi_i$ and $\psi_j$ are affinely related modulo $p$. In particular, if a form $\psi_i$ is a nonzero constant modulo $p$, then $p$ is exceptional.
\end{dfn}
We highlight that our definition of exceptional prime is different (less restrictive) than that of Green and Tao \cite[Theorem D.3]{GT2} and that the size of a system differs from their definition by the initial factor $\log^B N$.
We now check that $\prod_p\beta_p$ is still convergent in the setting where the coefficients are unbounded.
Thus the next lemma plays the role of \cite[Lemma 1.3]{GT2}.
\begin{lm}
\label{lm:convBetap}
Let $\Psi$ be a system of affine-linear forms.
Then if $p$ is not exceptional,
$$
\beta_p=1+O_{d,t}(p^{-2}).
$$
In particular, if $\Psi$ is admissible, the product $\prod_p\beta_p$ is convergent and nonzero.
\end{lm}
\begin{proof}
If two forms $\psi_i$ and $\psi_j$ are not affinely related modulo $p$, then the probability
as $a$ ranges over $(\mathbb{Z}/p\mathbb{Z})^d$ that they vanish simultaneously at $a$ is $p^{-2}$,
by elementary linear algebra (see for instance \cite[Proposition C.5]{bienv}). Inclusion-exclusion then yields $\beta_p=1+O(p^{-2})$ for $p$ unexceptional.
Now, if $\Psi$ is admissible, only finitely many primes are exceptional. Indeed, if $\psi_i$ and $\psi_j$ are affinely related modulo $p$ then
all the $2\times 2$ minors of the matrix $(\dot{\psi}_k(e_\ell))_{k\in\{i,j\},\ell\in[d]}$ are divisible by $p$ although
at least one these minors has to be nonzero because
they are not affinely related as forms over $\mathbb{Z}$. Moreover,
the hypothesis of admissibility implies that $\beta_p\neq 0$ for every prime $p$, so that the product is convergent and nonzero.
\end{proof}
We now state our main theorem.
\begin{trm}
\label{mytrm}
Let $d,t$ be positive integers and $A,B,L$ be positive constants.
Assume $\Psi=(\psi_1,\ldots,\psi_{t}):\mathbb{Z}^d\rightarrow \mathbb{Z}^t$ is an admissible system.
Suppose that $\nor{\Psi}_{N,B}\leq L$ and that $K\subset [-N,N]^d$ is a convex body satisfying $\mathrm{Vol} (K)\gg N^d\log^{-A}N$ and $\Psi(K)\subset\mathbb{R}_+^t$.
Then
\begin{equation}
\label{myequation}
\sum_{n\in\mathbb{Z}^d\cap K}\prod_{i=1}^{t}\Lambda(\psi_i(n))=
\mathrm{Vol}(K)\prod_p\beta_p(1+o_{d,t,A,B,L}(1)).
\end{equation}
\end{trm}
Unlike Theorem \ref{GT}, this theorem
is still meaningful when the convex body $K\subset [-N,N]^d$
satisfies $\mathrm{Vol}(K)=o(N^d)$ and applies even when the linear coefficients are of size polylogarithmic in $N$.
In some special cases, Theorem \ref{mytrm} follows easily
from the work of Green and Tao. In fact, Theorem \ref{GT} is a consequence of an asymptotic for the unbounded system $W\Psi+b$ where
$\Psi$ is a bounded system, $W=\prod_{p\leq w}p=\log^{1+o(1)}N$
and $b=(b_1,\ldots,b_t)\in[W]^t$ is a $t$-tuple of integers coprime to $W$.
More generally, an unbounded system $q\Psi+b$
with $q=\log^{O(1)}N$
and $\Psi$ bounded
is tractable, via an asymptotic for the system $\widetilde{W}\Psi+c$
where $\widetilde{W}=Wq$.\footnote{This remark was already exploited by the author in a previous paper \cite[Section 2.3]{bienv}.} By decomposing into residue classes, this
extends to systems $\Psi$ such that
for each $j$, the coefficients $\dot{\psi}_i(e_j)$ are bounded multiples of a common coefficient $q_j$. We show an example,
corresponding to the count of $k$-term progressions of primes
whose common difference is a multiple of $q$. We have
$$
\sum_{\substack{1\leq n,d\\ n+(k-1)qd\leq N}}\prod_{i=0}^{k-1}\Lambda(n+iqd)=
\sum_{a\in [q]}\sum_{\substack{1\leq n,d\\ n+(k-1)d\leq \frac{N-a}{q}}}
\prod_{i=0}^{k-1}\Lambda(q(n+id)+a)
$$
and are thus left with a system of the form $q\Psi+b$ with $\Psi$
bounded.
We now provide less immediate examples where Theorem \ref{mytrm}
applies.
\begin{ex}
What is the proportion of arithmetic progressions
$n+d\mathbb{N}$ whose $q_1$th,$\ldots$,$q_k$th terms are all primes?
Assume that $q_i=\lfloor \log^i N\rfloor$.
The answer is given by
$$
\sum_{1\leq n,d\leq N}\prod_{i=1}^k\Lambda(n+q_id).
$$
For this system, the factors $\beta_p$ can be easily expressed,
using the notation $h(p)$ for the number of classes modulo
$p$ occupied by $q_1,\ldots,q_k$, as
$$
\beta_p=\left(\frac{p}{p-1}\right)^k\frac{(p-1)(1+p-h(p))}{p^2}.
$$
\end{ex}
\begin{ex}
We can also count $k$-terms arithmetic progressions of primes up to $N$ whose common difference is $q=\lfloor \log N\rfloor$ times a prime.
This time the sum to consider is
$$
\sum_{1\leq n\leq n+(k-1)qd\leq N}\Lambda(d)\prod_{i=0}^{k-1}\Lambda(n+iqd).
$$
To simplify the expression of the local factors, assume
$\prod_{p\leq k}p\mid q$. Then
$$\beta_p=
\left(\frac{p}{p-1}\right)^{k+1}\frac{1}{p^2}
\left\lbrace\begin{array}{cc}
(p-1)^2 & \text{ if } p\mid q\\
(p-1)(p-k) & \text{ if } p\nmid q.
\end{array}\right.
$$
\end{ex}
\begin{ex}
We provide the asymptotic count of solutions to linear equations in the shifted squarefree primes, that is, primes $p$ for
which $p-1$ is squarefree. As it is not a direct application,
we give the details in the final section.
\end{ex}
In view of the Siegel-Walfisz theorem \eqref{PNTAP2}, one may
hope to write
$$\sum_{n\in\mathbb{Z}^d\cap K}\prod_{i=1}^{t}\Lambda(\psi_i(n))=
\mathrm{Vol}(K)\left(\prod_p\beta_p+o_{d,t,A,B,L}(1)\right)$$
instead of \eqref{myequation}, but unfortunately our method does
not yield this. Such an estimate is genuinely stronger given that $\prod_p\beta_p$ might well tend to infinity with $N$ (and possibly the linear coefficients as well). This weaker bound is ultimately due to the ineffectiveness of the Gowers norm estimate
\cite[Theorem 7.2]{GT2}.
To prove Theorem \ref{mytrm}, we first get rid of the convex body by decomposing into reasonably small boxes,
so that the theorem simply needs to be proven on boxes. In this context, the variables all have the same range and
are independent of each other, which makes it possible, after
the introduction of the $W$-trick, to prove a suitable von Neumann theorem.\footnote{The paper of Green and Tao also proceeds via a destruction of the convex body $K$ in Appendix C, but their method ceases to bear fruit as
soon as $\mathrm{Vol}(K)=o((\diam K)^d)$.} The latter eliminates all but one form, say $\psi_1$, thanks to a pseudorandom majorant. The next step is to equalize all coefficients of the chosen form $\psi_1$, by decomposing the ranges of
averaging into congruence classes. We are then left requiring a Gowers norm estimate which was proven by Green and Tao,
conditionally on two conjectures later fully resolved by
Green, Tao and Ziegler \cite{MN,GI}.
We assume a certain amount of familiarity with the original arguments of Green and Tao \cite{GT2}. Wherever only minor changes are needed to accommodate the unbounded coefficients,
we will
simply describe what must be changed in the existing arguments.
\textbf{Asymptotic notation.} The main asymptotic parameter throughout the paper is $N$, and all the standard asymptotic notation $o,O,\ll$ refers to the limit
as $N\rightarrow\infty$. There is an exception: whenever we discuss local factors, the limit is when $p\rightarrow\infty$, see
for instance Lemma \ref{lm:convBetap}. Several statements require, sometimes implicitly, that $N$ be large enough, which we always assume. Many other asymptotic parameters are defined in terms of $N$, such as $M,X,Z,Y$ so that the limit
as $N\rightarrow\infty$ is the same as, say, the limit
as $M\rightarrow\infty$. Indices may be added to symbols such as $O$, in which case they indicate upon which parameters the implied constant depends.
\section{First reductions}
As in \cite[Theorem 4.1]{GT2}, we show that we can assume
that the
affine forms not only satisfy $\psi_i(n)\geq 0$ but actually $\psi_i(n)\geq N^{9/10}$.
We remark that the set $$\{n\in K\cap\mathbb{Z}^d : \exists i\in[t]\quad \psi_i(n)\leq N^{9/10}\}$$
contains only $O(N^{d-1/10})=o(\mathrm{Vol}(K)N^{-1/20})$ elements, so that we can replace $K$ by the convex body $K\cap\bigcap_{i=1}^t\psi_i^{-1}([N^{9/10},+\infty))$ in Theorem \ref{mytrm}.
Moreover, prime powers are so sparse that we can restrict $\Lambda$ to primes, replacing
$\Lambda$ by $\Lambda'=1_\mathcal{P}\log$.
This leads to the first reduction.
\begin{prop}
Let $\Psi=(\psi_1,\ldots,\psi_{t}):\mathbb{Z}^d\rightarrow \mathbb{Z}^t$ be an admissible system.
Suppose that $\nor{\Psi}_{N,B}\leq L$ and that $K\subset [-N,N]^d$ is a convex body satisfying $\mathrm{Vol} (K)\gg N^d\log^{-A}N$ and $\Psi(K)\subset([N^{9/10},+\infty))^t$. Then
\begin{equation*}
\sum_{n\in\mathbb{Z}^d\cap K}\prod_{i=1}^{t}\Lambda'(\psi_i(n))=
\beta_{\infty}\prod_p\beta_p\left(1+o_{d,t,A,B,L}(1)\right).
\end{equation*}
\end{prop}
We perform one more elementary reduction, namely we reduce to normal form; for the definition of this notion, we refer to \cite[Definition 4.2]{GT2}.
The existence of a normal form extension $\Psi'$ for a system $\Psi$ follows from \cite[Lemma 4.4]{GT2}.
We inspect its proof to check that $\nor{\Psi}_{N,B}\leq L$
implies $\nor{\Psi'}_{N,B}\leq L'$ for some constant $L'$ depending only on $L,B,d,t$.
With respect to the paper \cite{GT2}, the change is
that the vector $f_{d+1},\ldots,f_{d'}$ introduced there will now be of size $O_L(\log^C N)$ for some constant $C=O_{B,d,t}(1)$. This is because these vectors are obtained by Cramer's formula, i.e. by computing determinants of matrices of bounded dimensions whose coefficients are $O_L(\log^B N)$. We claim this procedure reduces Theorem \ref{mytrm} to the following proposition, corresponding to \cite[Theorem 4.5]{GT2}.
\begin{prop}
\label{reduction}
Let $d,t$ be positive integers and $A,B,L$ be positive constants.
Let $\Psi=(\psi_1,\ldots,\psi_{t}):\mathbb{Z}^d\rightarrow \mathbb{Z}^t$ be an admissible system in normal form.
Suppose that $\nor{\Psi}_{N,B}\leq L$ and that $K\subset [-N,N]^d$ is a convex body satisfying $\mathrm{Vol} (K)\gg N^d\log^{-A}N$ and $\Psi(K)\subset([N^{9/10},+\infty))^t$. Then
\begin{equation*}
\sum_{n\in\mathbb{Z}^d\cap K}\prod_{i=1}^{t}\Lambda(\psi_i(n))=
\beta_{\infty}\prod_p\beta_p\left(1+o_{d,t,A,B,L}(1)\right).
\end{equation*}
\end{prop}
We now check that the arguments following \cite[Theorem 4.5]{GT2}
yield the deduction of Theorem \ref{mytrm}
from Theorem \ref{reduction}. Let $\Psi$ be an admissible system and $K$ a convex body satisfying the hypotheses of Theorem \ref{mytrm}.
Let $f_{d+1},\ldots,f_{d'}$ be the vectors and $\Psi'$ the system in normal form produced by the procedure given above.
Suppose all $f_i$ satisfy $f_i=O(\log^C N)$.
Letting $M=N\log^{-C}N$, we define an auxiliary convex body $K'$ by
$$K'=\{(n,m_{d+1},\ldots,m_{d'})\in \mathbb{R}^d\times [-M,M]^{d'-d}\mid
n+\sum_{i=d+1}^{d'}m_if_i\in K\}$$
which is included in $[-N',N']^{d'}$ where $N'=O(N)$. Moreover,
we still have $$\mathrm{Vol}(K')=\mathrm{Vol}(K)(2M)^{d'-d}\gg N'^d\log^{-D}N'$$
for some constant $D$.
Finally, the local factors are left unchanged by this operation. Thus if the system $\Psi$ is admissible, so is $\Psi'$, so that
Proposition \ref{reduction} can be applied to it, which concludes
the proof of the reduction.
\section{Reduction to the case of a box}
We show that the main theorem follows from the following very particular case.
\begin{prop}
\label{Boxes}
Let $d,t$ be positive integers and $A,B,L$ be positive constants.
Let $\Psi=(\psi_1,\ldots,\psi_{t}):\mathbb{Z}^d\rightarrow \mathbb{Z}^t$ be an admissible system in normal form.
Suppose that $\nor{\Psi}_{M,B}\leq L$ and that $\Psi([M]^d)\subset([M^{9/10},+\infty))^t$.
Then
\begin{equation*}
\sum_{n\in[M]^d}\prod_{i=1}^{t}\Lambda(\psi_i(n))=
M^d\prod_p\beta_p\left(1+o_{d,t,A,B,L}(1)\right).
\end{equation*}
\end{prop}
\begin{proof}[Proof that Proposition \ref{Boxes} implies Proposition \ref{reduction}.]
Let $K\subset [-N,N]^d$ be a convex body satisfying $\mathrm{Vol} (K)\gg N^d\log^{-A}N$. Let $$K'=\{x\in K\mid d(x,\partial K)\geq N\log^{-A-1}N\}$$ and $$K''=\{x\in \mathbb{R}^d\mid d(x,K)\leq N\log^{-A-1}N\}.$$
These are two convex bodies.
The arguments from elementary convex geometry displayed in \cite[Appendix A]{GT2} allow one to infer that
$$
\mathrm{Vol}(K')=\mathrm{Vol}(K)+O(N^d\log^{-A-1}N)=\mathrm{Vol}(K)(1+o(1))
$$
and the same for $K''$.
Now let $M=N\log^{-A-1}N/\sqrt{d}$ and consider the grid $(M\mathbb{Z})^d$. Let $\mathcal{B}=\{c+[M]^d\mid c\in J\}$ be the collection of boxes defined by this grid that are included in $K$
and $\mathcal{B'}=\{c+[M]^d\mid c\in J'\}$ be the collection of boxes defined by this grid that meet $K$. Note that \begin{equation}
\label{encadrement}
K'\subset \bigcup_{B\in \mathcal{B}}B\subset K\subset\bigcup_{B\in \mathcal{B'}}B\subset K'' .
\end{equation}
The first inclusion is because if a box $B$ from the grid meets $K'$, then it is included in $K$.
Now let $\Psi=(\psi_1,\ldots,\psi_{t}):\mathbb{Z}^d\rightarrow \mathbb{Z}^t$ be a system of affine-linear forms of finite complexity.
Suppose that $\nor{\Psi}_{M,B}\leq L$ and that $\Psi([M]^d)\subset([M^{9/10},+\infty))^t$.
Then
$$
\sum_{B\in\mathcal{B}}\sum_{n\in B\cap\mathbb{Z}^d}\prod_{i\in[t]}\Lambda(\psi_i(n))\leq \sum_{n\in K\cap \mathbb{Z}^d}\prod_{i\in[t]}\Lambda(\psi_i(n))\leq \sum_{B\in\mathcal{B'}}\sum_{n\in B\cap\mathbb{Z}^d}\prod_{i\in[t]}\Lambda(\psi_i(n))
$$
Now if $B=c+[M]^d$ with $c\in\mathbb{Z}^d$, letting $\Psi_c=\Psi+\dot{\Psi}(c)$, we can write
$$
\sum_{n\in B\cap\mathbb{Z}^d}\prod_{i\in[t]}\Lambda(\psi_i(n))=
\sum_{n\in[M]^d}\prod_{i\in[t]}\Lambda(\psi_{c,i}(n)).
$$
We check that the system $\Psi_c$ satisfies $\nor{\Psi_c}_{M,C}=O(1)$ for some constant $C$. Indeed, the linear coefficients are unchanged, so still of size $O(\log^B N)=O(\log^B M)$, and the constant coefficients are of size
$O(N\log^B N)=O(M\log^{A+B+1}M)$. Thus $C=A+B+1$ is good enough.
Moreover $\Psi_c([M]^d)\subset([N^{9/10},+\infty))^t$.
Thus we can apply Proposition \ref{Boxes}. We note that the $\beta_p$ it produces for a box $B=c+[M]^d$ is in fact independent of $c$, because the translation invariance of $\mathbb{Z}/p\mathbb{Z}$ allows one to write
$$
\mathbb{E}_{a\in(\mathbb{Z}/p\mathbb{Z})^d}\prod_{i\in[t]}\Lambda_p(\psi_i(a)+\dot{\psi_i}(c))=\mathbb{E}_{a\in(\mathbb{Z}/p\mathbb{Z})^d}\prod_{i\in[t]}\Lambda_p(\psi_i(a+c))=
\mathbb{E}_{a\in(\mathbb{Z}/p\mathbb{Z})^d}\prod_{i\in[t]}\Lambda_p(\psi_i(a)).
$$
Finally,
$$
\abs{\mathcal{B}}M^d\prod_p\beta_p(1+o(1))\leq \sum_{n\in K\cap \mathbb{Z}^d}\prod_{i\in[t]}\Lambda(\psi_i(n))\leq\abs{\mathcal{B'}}M^d
\prod_p\beta_p(1+o(1))
$$
Because of the inclusions \eqref{encadrement}, we see that
$$
\mathrm{Vol}(K)(1+o(1))=\mathrm{Vol}(K')\leq \abs{\mathcal{B}}M^d\leq \abs{\mathcal{B'}}M^d\leq \mathrm{Vol}(K'')=\mathrm{Vol}(K)(1+o(1)).
$$
This completes the proof of Proposition \ref{reduction}.
\end{proof}
\section{The $W$-trick}
We perform the $W$-trick in the same spirit as \cite{GT2}, so as to eliminate biases modulo small primes.
We introduce
$$
w=w(N)=\log\log N
$$
and $$
W=\prod_{p\leq w}p=\log^{1+o(1)}N.
$$
However, a simple one-dimensional example shows that one has to
adapt it to our situation where coefficients are unbounded.
Indeed, consider the system made of one form in one variable, namely
$n\mapsto qn+b$ with $q$ of size roughly $\log N$. The $W$-trick consists in writing
$$
\sum_{n\leq N}\Lambda(qn+b)=\sum_{a\in[W],(qa+b,W)=1}\frac{W}{\varphi(W)}\sum_{n\leq N/W}\frac{\varphi(W)}{W}\Lambda(Wqn+qa+b).
$$
But imposing that $(qa+b,W)=1$ does not ensure that the inner sum is $N/W(1+o(1))$, because $qa+b$ could well have a common factor greater than $w$ with $q$: when the coefficients are bounded, their factors are all less than $w$ for large enough $N$ but this is not the case any more in our setting. Moreover the
relevant average is not $W/\varphi(W)$ but $Wq/\varphi(Wq)$ which may be different if $q$ has prime factors larger than $w$.\footnote{Nevertheless, it is easy to check using Mertens' theorem that if $w=\log\log N$ and $q\leq \log^BN$ that
$$Wq/\varphi(Wq)=(1+o_B(1))W/\varphi(W).$$
}
This suggests that the coefficients of the system have to be taken into account to determine a suitable parameter $\widetilde{W}$ instead of $W$.
We fix once and for all an admissible system
$\Psi_1=(\psi_1,\ldots,\psi_{t}):\mathbb{Z}^{d_1}\rightarrow \mathbb{Z}^{t_1}$
in normal form
satisfying $\nor{\Psi_1}_{M,B}\leq L$ for some constants $B,L>0$.
Let
\begin{equation}
\label{minors}
Q=\prod_{\substack{i\in[t_1],j\in[d_1]\\\dot{\psi}_i(e_j)\neq 0}}\dot{\psi}_i(e_j)\times
\prod_{\substack{1\leq i<k\leq t_1\\mathbb{1}\leq j<\ell\leq d_1\\
\dot{\psi}_i(e_j)\dot{\psi}_k(e_\ell)
-\dot{\psi}_i(e_\ell)\dot{\psi}_k(e_j)\neq 0
}}(\dot{\psi}_i(e_j)\dot{\psi}_k(e_\ell)
-\dot{\psi}_i(e_\ell)\dot{\psi}_k(e_j))
\end{equation}
be the product of
the nonzero minors of size 1 and 2 in the matrix $(\dot{\psi}_i(e_j))_{i,j}$; thus
$Q=O_L(\log^{O_{d,t,B}(1)}N)$.
Moreover, if a prime $p$ is exceptional for $\Psi_1$, it must divide $Q$ (see the proof of Lemma \ref{lm:convBetap}).
We now introduce $\widetilde{W}=WQ=O(\log^{O(1)}N)$. The hypothesis that $\psi_i(n)>N^{9/10}$ means that if $\psi_i(n)$ is to be
a prime number, it has to be coprime to $\widetilde{W}$.
Writing $$\Lambda'_{\widetilde{W},b}(n)=\frac{\varphi(\widetilde{W})}{\widetilde{W}}\Lambda'(\widetilde{W} n+b)$$
we get
\begin{equation}
\label{WQtrick}
\sum_{n\in[M]^{d_1}}\prod_{i=1}^{t_1}\Lambda'(\psi_i(n))=
\sum_{\substack{a\in [\tilde{W}]^{d_1}\\
\forall i\in[t_1],(\psi_i(a),\widetilde{W})=1}}\left(\frac{\widetilde{W}}{\varphi(\widetilde{W})}\right)^{t_1}
\sum_{n\in [M/\widetilde{W}]^{d_1}}\prod_{i=1}^{t_1}
\Lambda'_{\widetilde{W},b_i(a)}
(\tilde{\psi}_i(n))+O(\log^{O(1)}N)
\end{equation}
where $\tilde{\psi}_i$ differs from $\psi_i$ only in the constant coefficient (by a multiple of $\widetilde{W}$) and $b_i(a)\in[\widetilde{W}]$ is the reduction mod $\widetilde{W}$ of $\psi_i(a)$.
Using the notion of subsystem introduced in Definition \ref{dfn:systems}, we reduce Proposition \ref{Boxes} to the following one.
\begin{prop}
\label{BoxesWtricked}
Let $\Psi_0=(\psi_1^0,\ldots,\psi_{t_0}^0):\mathbb{Z}^{d_0}\rightarrow \mathbb{Z}^{t_0}$ be a subsystem of $\Psi_1$.
Suppose that $\Psi_0([M]^{d_0})\subset([M^{8/10},+\infty))^{t_0}$ and that $b_i\in[\widetilde{W}]$ is coprime to $\widetilde{W}$ for any $i\in[t_0]$.
Then
\begin{equation}
\sum_{n\in [M/\widetilde{W}]^{d_0}}\prod_{i\in[t_0]}(\Lambda'_{\widetilde{W},b_i}(\psi_i^0(n))-1)=o((M/\widetilde{W})^{d_0})
\end{equation}
\end{prop}
We show how the reduction works, adapting the argument following
\cite[Proposition 5.1]{GT2}.
Applying successively the decomposition \eqref{WQtrick},
the trivial identity $x=x-1+1$, and Proposition
\ref{BoxesWtricked} to systems $\tilde{\Psi}$ where $\Psi$ is a subsystem of $\Psi_1$, we have
\begin{align*}
\sum_{n\in[M]^{d_1}}\prod_{i=1}^{t_1}\Lambda(\psi_i(n)) &=
\sum_{\substack{a\in [\widetilde{W}]^{d_1}\\
\forall i\in[t_1],(\psi_i(a),\widetilde{W})=1}}\left(\frac{\widetilde{W}}{\varphi(\widetilde{W})}\right)^{t_1}\left(\frac{M}{\widetilde{W}}\right)^{d_1}
(1+o(1))+O(\log^{O(1)}N)\\
&=M^{d_1}(1+o(1))\mathbb{E}_{a\in[\widetilde{W}]^{d_1}}\prod_{i\in[t_1]}\Lambda_{\widetilde{W}}(\psi_i(a))+O(\log^{O(1)} N).
\end{align*}
By the Chinese remainder theorem, and the fact that
$\Lambda_{p^k}(b)=\Lambda_p(b)$, we have
$$\mathbb{E}_{a\in[\widetilde{W}]^{d_1}}\prod_{i\in[t_1]}\Lambda_{\widetilde{W}}(\psi_i(a))
=\prod_{p\mid WQ}\mathbb{E}_{a\in\Zp{d_1}}\prod_{i\in[t_1]}\Lambda_p(\psi_i(a))
=\prod_{p\mid WQ}\beta_p.
$$
Moreover, if a prime $p$ does not divide $Q$, it is not
exceptional for $\Psi_1$. Then Lemma \ref{lm:convBetap}
implies
that
$\beta_p=1+O(p^{-2})$, so that
$$\prod_{p\nmid WQ}\beta_p=\prod_{p>w}(1+O(p^{-2}))=1+O(w^{-1}).$$
This concludes the reduction.
\section{Reduction to a Gowers norm estimate}
Write $X=M/\widetilde{W}$ and fix a system $\Psi_0$ and a tuple $b_1,\ldots,b_{t_0}$ satisfying the conditions of Proposition \ref{BoxesWtricked}.
If $t_0=1$, this proposition follows directly from the one-dimensional Siegel-Walfisz theorem \eqref{PNTAP}, so we suppose
$t_0\geq 2$.
Let $Q_0$ be the product of $2\times 2$ minors for the system $\Psi_0$
as defined by equation \eqref{minors}. In particular, $Q_0\mid Q$.
We have to prove
$$
\sum_{n\in[X]^d}\prod_{i\in[t_0]}F_i(\psi_i(n))=o(X^d)
$$
for $F_i=\Lambda'_{\widetilde{W},b_i}-1$.
At this point Green and Tao move to some cyclic group $\mathbb{Z}/N'\mathbb{Z}$ with $N'=O(X)$, but we cannot do this here without any wrap around, because of the large (unbounded) coefficients.
\subsection{A pseudorandom majorant}
Recall that we have fixed an admissible system
$$
\Psi_0=(\psi_i^0)_{i\in[t_0]} : \mathbb{Z}^{d_0}\longrightarrow\mathbb{Z}^{t_0}.
$$
We introduce the notion of a \emph{derived system}. This captures
the important properties of the systems that arise from repeated applications of the Cauchy-Schwarz inequality.
\begin{dfn}
A system $\Psi : \mathbb{Z}^d\rightarrow\mathbb{Z}^t$ of affine-linear forms is said to be \emph{derived} from $\Psi_0$ if
the following conditions are all satisfied:
\begin{itemize}
\item $d\leq 2d_0$;
\item $t\leq 2^{d_0}t_0$;
\item $\nor{\Psi}_{N,B}\ll \nor{\Psi_0}_{N,B}$;
\item any exceptional prime for $\Psi$ divides $Q_0$.
\end{itemize}
\end{dfn}
We now define pseudorandomness, based on the so-called linear forms condition; it is a fairly standard notion (see \cite[Section 6]{GT2}) but our definition
differs slightly to allow unbounded coefficients.
\begin{dfn}
We say that a function $\nu_Z : [Z]\rightarrow \mathbb{R}_+$ satisfies the
$\Psi_0$-\emph{linear forms condition}
if for any system $\Psi$
derived from $\Psi_0$
we have
$$
\mathbb{E}_{n\in [Z]^d}\prod_{i\in[t]}\nu(\psi_i(n))=1+o(1).
$$
We also say that $\nu$ is a $\Psi_0$-\emph{pseudorandom measure}.
\end{dfn}
The next proposition is about the existence of a pseudorandom
majorant for a $\widetilde{W}$-tricked von Mangoldt function.
\begin{prop}
\label{existencePRM}
For any integers $b_1,\ldots,b_{t_0}$ in $[\widetilde{W}]$ coprime to $\widetilde{W}$, for $Z\gg N\log^{-O(1)}N$, there exists a $\Psi_0$-pseudorandom measure $\nu$ on $[Z]$
such that
\begin{equation}
\label{Majorant}
1+\Lambda'_{\widetilde{W},b_1}+\cdots+\Lambda'_{\widetilde{W},b_{t_0}}\ll \nu
\end{equation}
on $[Z^{3/5},Z]$.
\end{prop}
The construction of the majorant, identical to \cite{GT2}, shall be explained in Section \ref{pseudorandom}.
\subsection{Generalised von Neumann theorem}
In spite of the impossibility to move to a cyclic group, we attempt to prove an analogue of
\cite[Proposition 7.1]{GT2}. Compared with the setting in a cyclic group, we cannot assume some linear coefficients to be 1, and
the range $[X]$ is not translation invariant. Recall that $\Psi_0
:\mathbb{Z}^{d_0}\rightarrow\mathbb{Z}^{t_0}$ is a fixed system of affine-linear
forms in $s$-normal form; thus without loss of generality,
write its first form as
$$
\psi_1(n_1,\ldots,n_{s+1},y)=q_1n_1+\cdots+q_{s+1}n_{s+1}+\psi_1(0,y)
$$
with $q_i\neq 0$ for all $i$ and $\prod_{j\in[s+1]}\dot{\psi_i}(e_j)=0$ for all $i>1$. We have dropped the exponent $0$ in the names of the form of the system $\Psi_0$ and shall always do so in the sequel.
We now state our variant of the von Neumann theorem.
\begin{trm}
\label{GVNT}
Let $f_1,\ldots,f_{t_0} :\mathbb{Z}\rightarrow \mathbb{R}$ be functions and $\nu$ be a $\Psi_0$-pseudorandom measure such that $\abs{f_i}\leq \nu$ for all $i$.
Then
$$
\abs{\mathbb{E}_{n\in [X]^d}\prod_{i\in[t_0]}f_i(\psi_i(n))}\leq \abs{\mathbb{E}_{y\in[X]^{d-s-1}}\mathbb{E}_{n^{(0)},n^{(1)}\in [X]^{s+1}}\prod_{\omega\in\{0,1\}^{s+1}}f_1(\sum_{i=1}^{s+1} q_in_i^{(\omega_i)}+\psi_1(0,y))}^{1/2^{s+1}}+o(1)
$$
\end{trm}
We adapt the proof of Proposition 7.1'' in \cite{GT2} and, for brevity,
use the notation from that proof without redefining it.
With this notation, the left-hand side equals
\begin{equation}
\label{LHStobound}
\mathbb{E}_{y\in[X]^{d-s-1},x\in[X]^{s+1}}\prod_{B\subseteq[s+1]}F_{B,y}(x_B).
\end{equation}
We observe that
$$
F_{[s+1],y}(x_{[s+1]})=f_1(\sum_{i\in[s+1]}q_ix_i+\psi(0,y)).
$$
We have the bounds $$\abs{F_{B,y}}\leq \nu_{B,y}.$$
The introduction of the functions $F_{B,y},\nu_{B,y} : [X]^B\rightarrow \mathbb{N}$ hides the arithmetic nature of the setting,
blurring away the difference between cyclic group and
intervals of integers.
Thus we can use \cite[Corollary B.4]{GT2},
with $A=[s+1]$ and
$X_\alpha=[X]$ for all $\alpha \in [s+1]$.
Hence we bound
\eqref{LHStobound} by
\begin{equation}
\label{afterCORB4}
\mathbb{E}_{y\in [X]^{d-s-1}}\nor{F_{[s+1],y}}_{\square (\nu_{[s+1],y})}
\prod_{B\subsetneq [s+1]}\nor{\nu_{B,y}}_{\square (\nu_{B,y})}^{2^{\abs{B}-(s+1)}},
\end{equation}
where we recall that
$$
\nor{F}_{\square (\nu_{B,y})}^{2^{\abs{B}}}=\mathbb{E}_{x^{(0)},x^{(1)}\in [X]^B}\prod_{\omega\in \{0,1\}^B}F(x^{(\omega)})
\prod_{C\subsetneq B}\nu_C((x_i^{(\omega_i)})_{i\in C},y).
$$
By Hölder's inequality, it suffices to show that
\begin{equation}
\label{norm}
\mathbb{E}_{y\in [X]^{d-s-1}}\nor{F_{[s+1],y}}_{\square (\nu_{[s+1],y})}^{2^{s+1}}= \mathbb{E}_{y\in [X]^{d-s-1}}\mathbb{E}_{n^{(0)},n^{(1)}\in [X]^{s+1}}\prod_{\omega\in\{0,1\}^{s+1}}f_1(\sum q_in_i^{(\omega_i)}+\psi_1(0,y))+o(1)
\end{equation}
and that
$$
\mathbb{E}_{y\in [X]^{d-s-1}}\nor{\nu_{B,y}}_{\square (\nu_{B,y})}^{2^{\abs{B}}}=1+o(1)
$$
for all non empty $B\subseteq [s+1]$.
To prove the latter, expand the left-hand side as
\begin{equation}
\label{averageNu}
\mathbb{E}_{y\in [X]^{d-s-1}}\mathbb{E}_{n^{(0)},n^{(1)}\in [X]^B}\prod_{C\subseteq B}\prod_{i: \Omega(i)=C}\prod_{\omega\in\{0,1\}^C}\nu(\psi_i((n_j^{(\omega_j)})_{j\in\Omega(i)},y)).
\end{equation}
which is an expression involving the average of $\nu$ on a
system $$\Psi=(\psi_{i,\omega})_{i\in[t_0],\omega\in\{0,1\}^{\Omega(i)}} : \mathbb{Z}^d\rightarrow\mathbb{Z}^t$$
of linear forms and it easy to check that $d\leq 2d_0$ and $t\leq 2^{d_0}t_0$. It is also obvious that $\nor{\Psi}_{M,B}\ll\nor{\Psi_0}_{M,B}$. Let the prime $p$ be
exceptional for $\Psi$ and let us check that it divides $Q_0$; this would mean that $\Psi$ is derived from $\Psi_0$ and
thus we could apply the linear forms condition. So let $\psi_{i,\omega}\neq \psi_{k,\alpha}$ be two forms that are affinely related modulo $p$. Then if $i\neq k$, we conclude that $\psi_i$ and $\psi_k$ are related and thus the prime is exceptional for $\Psi_0$, which implies that it divides $Q_0$. Otherwise $i=k$ and thus $\omega\neq \alpha$, in other
words there exists $j\in[d_0]$ such that $\dot{\psi_i}(e_j)\neq 0$ and $\omega_j\neq \alpha_j$. Thus $p$ must divide $\dot{\psi}_i(e_j)$
and hence also $Q_0$.
By applying the linear forms condition, the term \eqref{averageNu} is $1+o(1)$.
Let us look at \eqref{norm}.
At this point, Green and Tao use the translation invariance of $\mathbb{Z}/N'\mathbb{Z}$ to perform a change of variable which is not possible here, but we make do without it.
As the system is in normal form and $t\geq 2$, the form $\psi_1$ does not genuinely use all the variables. Indeed, the form $\psi_2$ must also have its set of $s+1$ variables that it is the only one to use fully, in particular $\psi_1$ does not use them all. Let us thus assume that $\psi_1$ only uses $x_1,\ldots,x_{d-k}$ with $k\geq 1$, which enables us, by a slight abuse
of notation, to regard $\psi_1$ as a map from $\mathbb{Z}^{d-k}$ to $\mathbb{Z}$. Upon expanding the norm, the left-hand side of \eqref{norm} becomes
\begin{equation*}
\label{expansion}
\mathbb{E}_{x^{(0)},x^{(1)}\in [X]^{s+1},y\in [X]^{d-k-s-1}}\prod_{\omega\in\{0,1\}^{s+1}}f_1(\sum_{i=1}^{s+1} q_ix_i^{(\omega_i)}+\psi_1(0,y))\mathbb{E}_{z\in [X]^{k}}\prod_{\omega\in\{0,1\}^{s+1}}\prod_{C\subsetneq [s+1]}\nu_{C,(y,z)}(x_C^{(\omega_C)})
\end{equation*}
where $(y,z)$ is the vector in $\mathbb{Z}^{d-s-1}$ obtained by concatenating $y$ and $z$.
We want to replace the inner expectation over $z$, which is a function of $(x^{(0)},x^{(1)},y)$ of average 1, by 1.
To do that, by Cauchy-Schwarz, it is enough to prove
$$\mathbb{E}_{x^{(0)},x^{(1)}\in [X]^{s+1},y\in [X]^{d-k-s-1}}\prod_{\omega\in\{0,1\}^{s+1}}\nu(\sum_{i=1}^{s+1} q_ix_i^{(\omega_i)}+\psi_1(0,y))=1+o(1)=O(1)$$
which follows directly from the linear forms condition, and
$$
\mathbb{E}_{x^{(0)},x^{(1)}\in [X]^{s+1},y\in [X]^{d-k-s-1}}\prod_{\omega\in\{0,1\}^{s+1}}\nu(\sum_{i=1}^{s+1} q_ix_i^{(\omega_i)}+\psi_1(0,y))\abs{\mathbb{E}_zW(x,y,z)-1}^2=o(1),
$$
where $W(x,y,z)=\prod_{\epsilon\in\{0,1\}^{s+1}}\prod_{C\subsetneq [s+1]}\nu_{C,(y,z)}(x_C^{(\epsilon_C)})$.
This amounts to
$$
\mathbb{E}_{x^{(0)},x^{(1)}\in [X]^{s+1},y\in [X]^{d-k-s-1}}\prod_{\omega\in\{0,1\}^{s+1}}\nu(\sum_{i=1}^{s+1} q_ix_i^{(\omega_i)}+\psi_1(0,y))(\mathbb{E}_zW(x,y,z))^j=1+o(1)
$$
for $j=0,1,2$.
Let us inspect the left-hand side in the most intricate case,
namely $j=2$. Upon expanding the square, we get an expectation over $x^{(0)},x^{(1)},y,z^{(0)},z^{(1)}$, thus the system is in at most $2d_0$ variables. There are $2^{s+1}$ forms arising from $\psi_1$ and at most $2^{s+2}(t-1)$ other forms, which means together at most $2^{d_0}t_0$ forms. Now the reasoning we used
to analyse the average \eqref{averageNu} also applies here and
yields that the system is derived from $\Psi_0$.
Thus the linear forms condition applies and equation \eqref{norm} is proven, hence also Theorem \ref{GVNT}.
\subsection{A Gowers-norm estimate}
Together with the existence of a pseudorandom majorant provided by Proposition \ref{Majorant},
Theorem \ref{GVNT} reduces Proposition \ref{BoxesWtricked} to the following.
\begin{prop}
Let $b\in[\widetilde{W}]$ be coprime to $\widetilde{W}$. Let $B>0$ and $d\in \mathbb{N}$ be constants. Suppose $q_1,\ldots,q_{d}$ are divisors of $Q$ satisfying $q_i=O(\log^BN)$ while $c=O(N\log^B N)$.
Then we have
\begin{equation}
\label{eq:pseudoGowers}
\mathbb{E}_{x^{(0)},x^{(1)}\in [X]^{d}}\prod_{\omega\in\{0,1\}^{d}}(\Lambda'_{\widetilde{W} ,b}(\sum_{i=1}^d q_ix_i^{(\omega_i)}+c)-1)=o(1).
\end{equation}
\end{prop}
The progress compared to Proposition \ref{BoxesWtricked} is that each variable $x_i^{(\epsilon)}$
for $i\in[d]$ and $\epsilon\in\{0,1\}$ is affected throughout the system
by one and the same coefficient $q_i$. We now attempt to transform
the system so that all variables have the same coefficient $Q'$;
the price we pay is that the variables will not have the same ranges any more.
To this effect, we introduce
$$
Q_i=\prod_{j\neq i}q_j
$$
and variables $n_i^{(\omega_i)},m_i^{(\omega_i)}$
such that $x_i^{(\omega_i)}=Q_in_i^{(\omega_i)}+m_i^{(\omega_i)}$.
Then the left-hand side of equation \eqref{eq:pseudoGowers}
decomposes as
$$
\mathbb{E}_{m_i^{(\omega_i)}\in [q_i]}\mathbb{E}_{n_i^{(\omega_i)}\in [X/q_i]}
\prod_{\omega\in\{0,1\}^d} (\Lambda'_{\widetilde{W},b}(\sum_{i=1}^d Q'n_i^{(\omega_i)}+q_im_i^{(\omega_i)}+c)-1)+o(1),
$$
where $Q'=q_iQ_i$ for any $i$.
We recognise the function
$$
n\mapsto F_a(n)=\frac{\varphi(\widetilde{W})}{\widetilde{W}}\Lambda'(Q'\widetilde{W}{}n+a)=\Lambda'_{Q'\widetilde{W},a},
$$
where the equality holds because $\prod_{p\mid Q'}p$ divides $Q$ and hence $\widetilde{W}$.
The parameters $a$ occurring are
$$
a_\omega=\widetilde{W}(\sum_{i=1}^d q_im_i^{(\omega_i)}+c)+b
$$
and given that $(b,\widetilde{W})=1$, we also have $(a_\omega,\widetilde{W})=1$ and finally $(a_\omega,\widetilde{W}{}Q')=1$.
We remark that for any tuple $a\in[\widetilde{W}{}Q']^{2^d}$ of integers coprime
to $\widetilde{W}{}Q'$, we can create a common $\Xi$-pseudorandom majorant for the functions $1+F_{a_\omega}$ where $\Xi=(\xi_\omega)_{\omega\in\{0,1\}^d}$ is defined by
$$
\xi_{\omega}=(n_1^{(0)},\ldots,n_d^{(0)},n_1^{(1)},\ldots,n_d^{(1)})
\longmapsto \sum_{i=1}^dn_i^{(\omega_i)}.
$$
In fact we
can rewrite Proposition 4.2 with $\widetilde{W}{}Q'$ instead of $\widetilde{W}$, because $Q'$ still satisfies $Q'=O(\log^{O(1)}N)$,
a bound which
allows us to control the effect of exceptional primes in Proposition \ref{GPY}.
Thus we are left to prove that
$$
\mathbb{E}_{n_i^{(\omega_i)}\in [X_i]}\prod_{\omega\in\{0,1\}^d} (F_{a_\omega}(\sum_{i=1}^d n_i^{(\omega_i)})-1)=o(1)
$$
where each $X_i$ satisfies $N\log^{-C}N\ll X_i\leq N$.
Letting $Z=\max_i X_i$ and $K=\prod_i[X_i]$, we have $K\subset [Z]^d$ and $\mathrm{Vol}(K)\gg Z^d\log^{-C'}Z$. Thus we can apply the same
reasoning as in Section 2 where we approximated such a convex body
by a set of small boxes of equal
sides,\footnote{The reader might object that we then used the positivity of the function to average, which is not available here, but we
can just as well use the majorant and the linear forms condition
to bound the contribution of the few boxes included in $K''$ but not in $K'$.}
and it suffices to prove that
\begin{equation}
\label{almostGowers}
\mathbb{E}_{n^{(0)},n^{(1)}\in[Y]^d}\prod_{\omega\in\{0,1\}^d}(F_{a_\omega}(\sum_{i=1}^dn_i^{(\omega_i)})-1)=o(1)
\end{equation}
for some $Y\gg N\log^{-D}N$.
Now that the linear forms have bounded coefficients (namely 0 and 1), there is no more objection to the use of Green-Tao's generalised von Neumann theorem \cite[Proposition 7.1]{GT2},
as long as the functions $1+\Lambda'_{Q'\widetilde{W},a_\omega}$ are dominated
by a pseudorandom measure, in the sense of Green-Tao \cite[Definition 6.2]{GT2}. Green and Tao proved the existence
of such a majorant, except that they had $W$ instead of $Q'\widetilde{W}$,
but this makes no difference as $Q'\widetilde{W}$ is $w$-smooth and divisible by $W$. See Section 6 for a review of the construction of
a majorant and a closer scrutiny of the role of the
linear coefficients in the linear forms condition.
Thus equation \eqref{almostGowers} follows from the claim
\begin{equation}
\label{eq:claim}
\nor{F_a-1}_{U^k([Y'])}=
\nor{\Lambda'_{Q'\widetilde{W},a}-1}_{U^k([Y])}=o(1)
\end{equation}
for any $a\in [Q'\widetilde{W}]$ coprime to $Q'\widetilde{W}$.
This is almost \cite[Proposition 7.2]{GT2}.
Compared with that proposition, we have $W'=Q'\widetilde{W}=O(\log^{O(1)}N)$ instead of $W$.
One can inspect attentively the remainder of the argument of Green and Tao to notice that the required properties of $W$ are
\begin{itemize}
\item that it be divisible by all primes $p\leq w$ for some function $w=w(N)$ tending to infinity;
\item that it be $O(\log^{O(1)}N)$; this is crucial when applying the Möbius-Nilsequence theorem \cite{MN}, which comes with a saving
of size an arbitrary power of $\log N$.
\end{itemize}
These properties are equally satisfied by $Q'\widetilde{W}$. Thus the claim \eqref{eq:claim} holds. To complete the proof of Theorem \ref{mytrm}, there remains only to prove Proposition \ref{existencePRM}, which we shall do in the next section.
\section{The linear forms condition}
\label{pseudorandom}
We recall the notation from \cite{GT2}
$$
\Lambda_{\chi,R}(n)=\log R\left(\sum_{d\mid n}\mu(d)\chi\left(\frac{\log d}{\log R}\right)\right)^2
$$
for $R=N^{\gamma}$ a small power of $N$, a smooth function $\chi$ supported on $[-1,1]$
satisfying $\chi(0)=1$ and $\int{\chi'^2}=1$. The function
$\Lambda_{\chi,R}$ is positive and
if $n$ is a prime
larger than $R$, then $\Lambda_{\chi,R}(n)=\log R$, so that
$\Lambda'\leq \gamma^{-1}\Lambda_{\chi,R}$ on $[N^{\gamma},N]$.
We need to extend the range of application of \cite[Theorem D.3]{GT2}. There it is stated only for forms whose constant coefficients are bounded, although it was then applied to
other natural systems such as $\Phi=W\Psi+c$, as the extension
was straightforward. In \cite[Appendix A]{GTFK}, the constant coefficients
are allowed to be as large $N^{1.01}$.
We claim that the estimate can be pushed further.
\begin{prop}
\label{GPY}
Let $L,B$ be positive constants and
$\Psi=(\psi_1,\ldots,\psi_t)$ an admissible system of affine-linear forms satisfying $\nor{\Psi}_{Z,B}\leq L$.
Let $P_\Psi$ be the set of exceptional primes and
$X=\sum_{p\in P_\Psi}p^{-1/2}$.
Then
$$
\sum_{n\in [Z]^d}\prod_{i\in[t]}\Lambda_{\chi,R}(\psi_i(n))=Z^d\prod_p\beta_p(1+o(1)).
$$
\end{prop}
We now carefully analyse what needs to be changed in the proof of Theorem D.3 of \cite{GT2} when the linear coefficients
are of size up to $\log^B N$. As remarked in \cite[Appendix A]{GTFK}, the first place where the bound on the coefficients
is used is page 1833, where it is said that $\alpha(p,B)=O(1/p)$
for $p$ large enough. In fact, as we have assumed from the outset
that no form of the system $\psi_i$ is divisible by any prime $p$,
this is always the case. This also means that $\beta_p=1+O(1/p)$
for all $p$. The next moment where
Green and Tao invoke the size of the coefficients is
to get the bound $\beta_p=1+O(p^{-2})$;
but in fact it is valid as soon as $p$ is not exceptional, no matter the size of the coefficients.
As seen in the proof of Lemma \ref{lm:convBetap},
the set $P_\Psi$ of exceptional primes is finite but
its cardinality may increase to infinity with $Z$.
What we have seen implies that the asymptotic formula
$$
\sum_{n\in [Z]^d}\prod_{i\in[t]}\Lambda_{\chi,R}(\psi_i(n))=Z^d\prod_p\beta_p(1+ e^{O(X)}\log^{-1/20}R)
$$
is still valid, where $X=\sum_{p\in P_\Psi}p^{-1/2}$. It remains to bound $X$.
If $p\in P_{\Psi}$, as already seen in Lemma \ref{lm:convBetap}, $p$ divides
the parameter $Q=\log^{O(1)}N$ introduced in equation \eqref{minors}.
Letting $\omega(Q)=O(\log Q)=O(\log\log N)$ be the number of
its prime factors, we have
$$
X=\sum_{p\in P_\Psi}p^{-1/2}
\leq \sum_{p\leq \omega(Q)}p^{-1/2}\leq \sum_{n\ll \log\log N}n^{-1/2}\ll \sqrt{\log\log N}.
$$
Thus $e^{O(X)}\ll\log^{1/30} N$ while $\log^{-1/20} R\ll \log^{-1/20} N$, which gives
$e^{O(X)}\log^{-1/20} R=o(1)$. This completes the verification of
Proposition \ref{GPY}.
From Proposition
\ref{GPY}, the proof of Proposition \ref{existencePRM}, that is, the construction of a majorant satisfying the adequate
linear forms condition, runs just as in the paper of Green and Tao.
We provide it here. Let
$b_1,\ldots,b_{t_0}$ be integers in $[\widetilde{W}]$ coprime to $\widetilde{W}$.
Let $Z$ be an asymptotic parameter satisfying $Z\gg N\log^{-A}N$ for some constant $A>0$. Then, writing
$$
\nu(n)=\frac{1}{t_0+1}\left(1+\frac{\varphi(\widetilde{W})}{\widetilde{W}}\sum_{i\in[t_0]}\Lambda_{\chi,R}(\widetilde{W}{}n+b_i)\right),
$$
we have
the bound \eqref{Majorant} for $n\in [Z^{3/5},Z]$
if $\gamma <3/5$.
To show that $\nu$ is a $\Psi_0$-pseudorandom measure, it is enough to
check that for any system $\Psi : \mathbb{Z}^d\rightarrow\mathbb{Z}^t$ derived from $\Psi_0$, any $s\leq t$ and any sequence $1\leq j_1<\cdots<j_s\leq t$, we have
\begin{equation}
\label{unpluso}
\left(\frac{\varphi(\widetilde{W})}{\widetilde{W}}\right)^s\mathbb{E}_{n\in [Z^d]}\prod_{i\in [s]}\Lambda_{\chi,R}(\widetilde{W}\psi_{j_i}(n)+b_{j_i})=1+o(1).
\end{equation}
To this aim, observe that the hypotheses ensure that the system
$\Phi=(\widetilde{W}\psi_{j_i}+b_{j_i})_{i\in[s]}$ is
is admissible.
Moreover, because $\widetilde{W}=O(\log^{O(1)}N)$,
the bound $\nor{\Phi_{Z,B}}=O(1)$ holds for some $B=O(1)$.
So we can use Proposition
\ref{GPY}. For $p\mid\widetilde{W}$, the local factor is simply $(p/(p-1))^s$
while if $p\nmid \widetilde{W}$, the prime $p$ is not exceptional for $\Psi_0$ and hence for $\Phi$, which implies $\beta_p=1+O(p^{-2})$ by Lemma \ref{lm:convBetap}.
Thus
$$
\prod_p\beta_p=\left(\frac{\widetilde{W}}{\varphi(\widetilde{W})}\right)^s\prod_{p>w}(1+O(p^{-2}))=\left(\frac{\widetilde{W}}{\varphi(\widetilde{W})}\right)^s(1+O(w^{-1}).
$$
This compensates exactly for the factor $(\varphi(\widetilde{W})/\widetilde{W})^s$ and
finishes the proof of equation \eqref{unpluso}, hence also of Proposition \ref{existencePRM}, and finally of Theorem \ref{mytrm}.
In the next section, we give a nice application of our higher-dimensional Siegel-Walfisz theorem.
\section{Application to the primes $p$ such that $p-1$ is squarefree}
The set of primes $p$ such that $p-1$ is squarefree is a well-known dense subset of the primes of density $\sum_a\frac{\mu(a)}{\varphi(a^2)}=\prod_p(1-1/p(p-1))$; this is a theorem of Mirsky \cite{Mirsky}. As any dense subset of the primes, it contains arbitrarily long arithmetic progressions, by the Green-Tao theorem \cite{GT1}. However, no asymptotic was available
so far for the count of
$k$-term progressions in this set, nor in fact in any dense subset of the primes (except residue classes). As a consequence
of Theorem \ref{mytrm},
we now prove such an
asymptotic; in fact, we obtain an asymptotic for the number of solutions in this set of primes to any
finite complexity system of equations.
For convenience, let
$F$ be the von Mangoldt function
restricted to the squarefree shifted primes, that is $F(n)=\Lambda(n+1)\mu^2(n)$. Also we denote by $\mathbb{N}$ the set of positive integers.
\begin{trm}
\label{appliSQF}
Let $\Psi : \mathbb{Z}^d\rightarrow\mathbb{Z}^t$ be a system of affine-linear forms
of finite complexity and $K\subset [-N,N]^d$ a convex body. Suppose that the linear coefficients are $O(1)$, the constants ones are $O(N)$ and that $\Psi(K)\subset \mathbb{N}^t$.
Then there exists a constant $C(\Psi)$ (possibly equal to 0) such
that
\begin{equation}
\label{eq:appliSQF}
\sum_{n\in K\cap\mathbb{Z}^d}\prod_{i\in[t]}F(\psi_i(n))=C(\Psi)\mathrm{Vol}(K)+o(N^d).
\end{equation}
\end{trm}
The constant $C(\Psi)$ will appear explicitly in the proof, but
its expression is unpleasant, so we do not give it here.
Throughout the proof of this theorem, we will need the notation
$$
\alpha_\Psi(k_1,\ldots,k_t)=\mathbb{E}_{a\in (\mathbb{Z}/m\mathbb{Z})^d}\prod_{i\in[t]}1_{k_i\mid\psi_i(a)}
$$
where $m=\text{lcm}(k_1,\ldots,k_t)$.
Elementary convex geometry reveals that this is the density of points of the lattice $\{n\in\mathbb{Z}^d : \forall i\in[t]\quad k_i\mid\psi_i(n)\}$ per unit volume, in the sense that
for any convex body $K\subset [-B,B]^d$, we have
that
\begin{equation}
\label{volPacking}
\sum_{n\in K\cap\mathbb{Z}^d}\prod_{i\in[t]}1_{d_i\mid\psi_i(n)}
=\mathrm{Vol}(K)\alpha_\Psi(d_1,\ldots,d_t)+O(B^{d-1}\text{lcm}(d_1,\ldots,d_t)).
\end{equation}
This follows from simple volume packing arguments (see \cite[Appendix A]{GT2}, \cite[Appendix C]{polyprog}).
We now prove Theorem \ref{appliSQF}.
\begin{proof}
We insert the formula $\mu^2(n)=\sum_{a^2\mid n}\mu(a)$
in the left-hand side of \eqref{eq:appliSQF}. Thus
\begin{equation}
\label{eq:ais}
\sum_{n\in K\cap\mathbb{Z}^d}\prod_{i\in[t]}F(\psi_i(n))
=\sum_{(a_1,\ldots,a_t)\in\mathbb{N}^t}\prod_{i\in[t]}\mu(a_i)\sum_{\substack{n\in K\cap\mathbb{Z}^d\\ \forall i\in[t]\, a_i^2\mid\psi_i(n)}}
\prod_{i\in[t]}\Lambda(\psi_i(n)+1).
\end{equation}
Now for any $a=(a_1,\ldots,a_t)\in\mathbb{N}^t$, we introduce the set
$$
L_a=\{n\in\mathbb{Z}^d : \forall i\in[t]\quad a_i^2\mid\psi_i(n)\}.
$$
Fix an $a$ for which
$L_a\neq \emptyset$ and let $n_0\in L_a$.
Then
$$
L_{a}=n_0+\bigcap_{i=1}^t\ker g_i
$$
where $g_i:\mathbb{Z}^d\rightarrow\prod_{i\in[t]}\mathbb{Z}/a_i^2\mathbb{Z}$ is the affine-linear map obtained by applying $\dot{\psi_i}$ and then reducing modulo $a_i^2$. So $L_a$ is an affine sublattice of full rank: indeed, its direction contains $\{\prod_ia_i^2e_1,\ldots,\prod_ia_i^2e_d\}$.
Incidentally, this means that we can suppose that $n_0$ satisfies
$n_0\cdot e_i \in[\prod_ia_i^2]$, in particular
$\nor{n_0}\leq d\log^{2dC}$.
As a lattice of full rank, the direction $\overrightarrow{L_a}$
of $L_a$ has a $\mathbb{Z}$-basis: there exist $f_1,\ldots,f_d$ such that
$$
L_{a}=\{n_0+\sum_{i=1}^dm_if_i\mid (m_1,\ldots,m_d)\in\mathbb{Z}^d\}.
$$
Because of a theorem of Mahler, we can assume that $\nor{f_i}\leq i\lambda_i$ for $i=1,\ldots,d$, where
$\lambda_1\leq \cdots\leq \lambda_d$ are the successive minima of the lattice $\overrightarrow{L_a}$ with respect to the Euclidean unit ball.
Let $R^a$ be the affine transformation of $\mathbb{R}^d$ defined by $R^a(0)=n_0$ and $\dot{R^a}(e_i)=f_i$ for each $i\in[d]$.
Note that $L_a\cap K=R^a(\mathbb{Z}^d\cap K_a)$ where $K_a$ is also a convex body.
For the notions of geometry of numbers alluded to here, see for instance the notes of Green \cite{geonumbers}.
Now if one of the $a_i$ is larger than $\log^C N$, then
$K_a$ is small. Indeed, the set
of $n\in K\cap\mathbb{Z}^d$ such that there exists $i\in [t]$ and $a_i>\log^C N$ satisfying $a_i^2\mid \psi_i(n)$ has $O(N^d\log^{-C}N)$ elements.
This follows from equation \eqref{volPacking} combined with the bound
$\alpha_{\psi_i}(a_i^2)\ll a_i^{-2}$, obtained by multiplicativity, linear algebra (for instance \cite[Corollary C.4]{bienv}) and the fact that the coefficients of $\psi_i$ are bounded, and finally $\sum_{a>x}a^{-2}\ll x^{-1}$.
Bounding the contribution to the left-hand side of \eqref{eq:ais} of this exceptional set of $n\in K\cap\mathbb{Z}^d$ using $F\ll\log$, and supposing that $C\geq 2t$, we obtain
\begin{align*}
\sum_{n\in K\cap\mathbb{Z}^d}\prod_{i\in[t]}F(\psi_i(n))
&=\sum_{\substack{n\in K\cap \mathbb{Z}^d\\ \forall i\in[t]\,\forall a>\log^C N\, a^2\nmid \psi_i(n)}}\prod_{i\in[t]}F(\psi_i(n))+
O(N^d\log^{-C/2}N)\\
&=\sum_{1\leq a_1,\ldots,a_t\leq \log^C N}\prod_{i\in[t]}\mu(a_i)\sum_{n\in K\cap L_a}
\prod_{i\in[t]}\Lambda(\psi_i(n)+1)+
O(N^d\log^{-C/2}N).
\end{align*}
For each $i\in[t]$, the map $\psi_i^a : L_a\rightarrow\mathbb{Z}$ defined by
$$
\psi_i^a(n)=\frac{\psi_i(n)}{a_i^2}
$$
is an affine map. Then introduce $\phi_i^a=\psi_i^a\circ R^a$.
This defines a system $\Phi^a : \mathbb{Z}^d\rightarrow\mathbb{Z}^t$ of
affine-linear forms which is again of finite complexity. Thus the
inner sum in the left-hand side of equation \eqref{eq:ais} may be written as
\begin{equation}
\label{innersum}
\sum_{n\in K\cap L_a}
\prod_i\Lambda(\psi_i(n)+1)=
\sum_{m\in K_a\cap\mathbb{Z}^d}\prod_i\Lambda(a_i^2\phi_i^a(m)+1).
\end{equation}
We now apply Theorem \ref{mytrm} to the inner sum.
One can check that the linear coefficients of $\Phi_a$ have size
$O(\log^{O(1)}N)$. To do this, it is enough to examine the size of
the basis vectors $f_j$ of the lattice $\overrightarrow{L_a}$. Indeed,
$$
a_i^2\abs{\dot{\phi_i^a}(e_j)}=\abs{\dot{\psi_i}(f_j)}\leq \nor{\dot{\psi_i}}\nor{f_j}\ll\nor{f_j}.
$$
Moreover, the constant coefficients are $O(N)$.
As observed,
if $n_0\in L_a$, the lattice
$$
\{n_0+\sum_{i\in[d]}k_ia_i^2e_i\mid k\in\mathbb{Z}^d\}
$$
is a sublattice of $L_a$ and its determinant is
$\prod_i a_i^2\leq \log^{2dC}N$. Hence using Minkowski's second theorem, one finds that
$$
\prod_{i\in[d]} \nor{f_i}\leq d!\prod_{i\in[d]}\lambda_i\ll_d \abs{\det L_a}\leq \log^{2dC}N.
$$
Similarly, we obtain the bound
$$
\mathrm{Vol}(K_{a})=\mathrm{Vol}(K)\det (R^a)^{-1}\geq\mathrm{Vol}(K)\log^{-2dC}N.
$$
Now Theorem \ref{mytrm} tells us that
the right-hand side of \eqref{innersum} is $\mathrm{Vol}(K_a)\prod_p\beta_p(1+o(1))$ as
soon as none of the local factors $\beta_p(a)$ corresponding to the system of the forms
$a_i^2\phi_i^a+1$ vanishes.
Note that if any $\beta_p(a)$ is 0, then for all $m$ there exists $i\in[t]$ such that $p$ divides $a_i^2\phi_i^a(m)+1$. In this case,
the expression $a_i^2\phi_i^a(m)+1$ cannot be prime unless it is equal to $p$, which easily implies that
$$
\sum_{m\in K_a\cap\mathbb{Z}^d}\prod_i\Lambda(a_i^2\phi_i^a(m)+1)
=O(N^{d-1}\log^{O(1)}N).
$$
Moreover, equation \eqref{volPacking} reveals that
$$
\mathrm{Vol}(K_a)=\abs{K_a\cap\mathbb{Z}^d}+O(N^{d-1})=\abs{K\cap L_a}+O(N^{d-1})
=\mathrm{Vol}(K)\alpha_\Psi(a_1^2,\ldots,a_t^2)+O(N^{d-1}\log^{O(1)}N).
$$
Thus the left-hand side of equation \eqref{eq:appliSQF} equals
$$
\mathrm{Vol}(K)(1+o(1))\sum_{1\leq a_1,\ldots,a_t\leq \log^C N}\alpha_\Psi(a_1^2,\ldots,a_t^2)\prod_p\beta_p(a)\prod_{i\in[t]}\mu(a_i)+O(N^d\log^{-C/2}N).
$$
We claim the sum over $a$ is absolutely convergent.
To see this, first observe that
the exceptional primes for the system of the forms $a_i^2\phi_i^a+1$ are divisors of
$a_i^2$ or exceptional primes for the system $\Phi^a$;
in any case, they are divisors\footnote{See equation \eqref{minors} and the remark following it.} of a parameter $Q(a)=O(\prod_ia_i^{O(1)})$.
For all other primes, we have $\beta_p=1+O(p^{-2})$ by Lemma \ref{lm:convBetap}, so that
$$
\prod_p\beta_p(a)\ll\prod_{p\mid Q(a)}\beta_p(a)\leq \left(\frac{Q(a)}{\varphi(Q(a))}\right)^t\ll (\log\log Q(a))^t\ll
(\log\log\prod_ia_i)^t.
$$
Then note that the sum
$$
\sum_{a_1,\ldots,a_t}(\log\log\prod_ia_i)^t\alpha_\Psi(a_1^2,\ldots,a_t^2)
$$
is convergent. Indeed, we have the bound
$$
\alpha_{\Psi}(a_1^2,\ldots,a_t^2)=\prod_p\alpha_\Psi(p^{2v_p(a_1)},\ldots,p^{2v_p(a_t)})
\ll
\prod_pp^{-2v_p(\max_i a_i)}=\text{lcm}(a_1,\ldots,a_t)^{-2},
$$
where the inequality holds because the forms
$\psi_i$ have bounded linear coefficients and
if $p\nmid\psi_i$, then $\alpha_{\psi_i}(p^k)\leq p^{-k}$ (this is linear algebra and Hensel's lemma, see \cite[Corollary C.4]{bienv}). The convergence then follows from a trivial bound for the number of $t$-tuples $a$ of prescribed least common multiple $k$, namely $\tau(k)^t$.
This convergence result implies that
$$\sum_{1\leq a_1,\ldots,a_t\leq \log^C N}\alpha_\Psi(a_1^2,\ldots,a_t^2)\prod_p\beta_p(a)\prod_{i\in[t]}\mu(a_i)
=C(\Psi)+o(1),
$$
where
$$
C(\Psi)=\sum_{(a_1,\ldots,a_t)\in\mathbb{N}^t}\alpha_\Psi(a_1^2,\ldots,a_t^2)\prod_p\beta_p(a)\prod_{i\in[t]}\mu(a_i).
$$ This concludes the proof.
\end{proof}
|
\section{Introduction}
In a \emph{network creation game} (NCG), we are given $n$ players (identified as the nodes of a graph), which attempt to settle an undirected interconnection network. This is realized by letting each player connect herself \emph{directly} to a subset of players, through the activation of the corresponding set of incident links. These links can then be freely used by everyone, and so a player remains connected to non-adjacent players \emph{indirectly}, i.e., by following a \emph{shortest path} in the currently active network. In such a decentralized process, a player has then to strategically balance the sum of two costs: the \emph{building cost}, which is given by the sum of the costs she incurs in activating her links, and the \emph{routing cost}, which is a function of the length of the shortest paths towards the other players.
Due to their generality, it is evident that NCGs can model very different practical situations. Just to mention an example, NCGs are fit to model the decentralized construction of \emph{communication} networks, in which the constituting components (e.g., routers and links) are activated and maintained by different owners, as in the Internet.
In the very first formulation of the game \cite{JW96}, the building and routing costs are defined as follows. Concerning the building cost, each activated link $(i,j)$ has a cost $c_{ij}$ (respectively, $c_{ji}$) for player $i$ (respectively, $j$), and the formation of a link requires the consent of both players involved (since each of the parties pays the corresponding activating cost), while link severance can be done unilaterally. On the other hand, the routing cost is given by a function additively depending on the distances to all the other players.
Later on, Fabrikant \emph{et al.} \cite{FLM03} developed a simplified version of this model, which is also the most popular one in the field of \emph{Algorithmic Game Theory} (AGT), namely that in which the activation of each link has a fixed cost $\alpha > 0$, and this is incurred by the activating player only and without the consent of the adjacent player, while the routing cost is simply the sum of the distances to all the other players (in graph terminology, this is known as the \emph{status} of a node). Besides this simplification, the merit of such a paper was that of emphasizing how the social utility for a (very large) system as a whole is affected by the selfish behavior of the players, which was instead downplayed by the economists, that were more focused on system stability issues. This new perspective inspired then a sequel of papers in the AGT community, as detailed in the following.
\paragraph{Previous work on NCGs}
More formally, the form of a NCG as provided in \cite{FLM03}, which we call \textsc{SumNCG}\xspace, is as follows: we are given a set of $n$ players, say $V$, where the
strategy space of player $u \in V$ is the power set $2^{V \setminus\{u\}}$. Given
a combination of strategies $\sigma=(\sigma_u)_{u \in V}$, let $G(\sigma)$ denote the underlying undirected
graph whose node set is $V$, and whose edge set is
$E(\sigma)=\{(u,v): u \in V \wedge v \in \sigma_u\}$.
Then, the \emph{cost} incurred by player $u$ in $\sigma$ is
\begin{equation}
\label{sum} C_u(\sigma) = \alpha \cdot |\sigma_u| + \sum_{v \in V}
d_{G(\sigma)}(u,v)
\end{equation}
\noindent where $d_{G(\sigma)}(u,v)$ is the distance between $u$
and $v$ in $G(\sigma)$.
When a player takes an action (i.e., activates a subset of incident edges), she aims to keep this cost as low as possible. Under the assumption of a complete knowledge of $G(\sigma)$, we therefore have that a player $u$ is fully aware that after switching from strategy $\sigma_u$ to strategy $\sigma'_u$, the network will change to $G(\sigma_{-u}, \sigma'_u)$.
Thus, a \emph{Nash Equilibrium}\footnote{In this paper, we only
focus on \emph{pure}-strategy Nash equilibria.} (NE) for the game is a strategy profile $\bar{\sigma}$ such that for every player $u$ and every strategy profile $\sigma_u$, we have that $C_u(\bar{\sigma}) \leq C_u(\bar{\sigma}_{-u}, \sigma_u)$. If we characterize the space of NE in terms of the \emph{Price of Anarchy}
(PoA), i.e., the ratio between the social cost of the costlier NE to the optimal
(centralized) social cost, then it has been shown this is constant for all values of $\alpha$ except for $n^{1-\varepsilon} \leq \alpha < 65 \, n$, for any $\varepsilon \geq 1/\log n$ (see \cite{MMM13,MS10}), while an upper bound of $2^{O(\sqrt{\log n})}$ is known for the remaining values of $\alpha$ \cite{DHM12}. Moreover, very recently, in \cite{GHLS13} it was proven that for all constant non-integral $\alpha \geq 2$, the PoA is bounded by $1+ o(1)$.
A first natural variant of \textsc{SumNCG}\xspace was introduced in \cite{DHM12}, where the \emph{eccentricity} of a player rather than her status was considered as a measure of centrality in the network, and then the player cost function was redefined
as follows:
\begin{equation}
\label{max} C_u(\sigma) = \alpha \cdot |\sigma_u| + \max \{d_{G(\sigma)}(u,v):v \in V\}.
\end{equation}
\noindent This variant, named \textsc{MaxNCG}\xspace, received further
attention in \cite{MS10}, where the authors improved the PoA of
the game on the whole range of values of $\alpha$, obtaining in this case that the PoA is constant for all values of $\alpha$ except for $129 > \alpha = \omega(1/\sqrt{n})$, while for the remaining values of $\alpha$, it is at most $2^{O(\sqrt{\log n})}$.
Besides these two basic models, many variations on the theme have been defined.
We mention those obtained by limiting the modification a player can do on her current strategy
(see \cite{ADH10,L12,MS12}), or by budgeting either the
number of edges a player can activate or her eccentricity (see \cite{LPR08,EFM11,BGP12}), or finally by constraining the set of available links to a host graph (see \cite{BGLP12,DHM09}).
Generally speaking, in all the above models the obtained upper bounds on the PoA are asymptotically worse than those we get in the two basic models.
\paragraph{Criticisms to the standard model}
Observe that while the general assumption that players have a \emph{common and complete} information about the ongoing network is feasible for small-size instances of the game, this becomes unrealistic for large-size networks. This is rather problematic, since the asymptotic analysis which guides the AGT literature requires instead a growing size of the input. Moreover, quite paradoxically, the full-knowledge assumption is not simplifying at all: it makes computationally unfeasible for a player to select a best-response strategy, or even to check whether she is actually in a NE!
Very recently, the same observation leads Ballester Pla \emph{et al.} to consider in \cite{CPV09} a more compelling scenario for the related class of \emph{network} (or \emph{graphical}) \emph{games}. In a graphical game, players are embedded in a network, and the cost of a player depends on her action and on those of her neighbors, and thus is correlated to the (knowledge of the) entire network. In \cite{CPV09} the authors assume instead that players have a complete knowledge of the network structure up to a given radius $k$, and use this information to make up a belief about the rest of the network. For this model, they provide a closed formula to compute a \emph{Bayes-Nash equilibrium} for the game, and show an interesting relationship with a scenario in which players have a \emph{bounded rationality} (i.e., they take a step by only exploring a subset of the strategy space).
Another work which constrains the available strategies of the players according to a concept of locality is \cite{H13}. In this work the author studies the non-coordinated process of \emph{matching formation} where each player can (potentially) match with any other player having distance at most $k$ in a graph which depends on the current state of the game.
\paragraph{Our new NCG model}
In this paper we concentrate on \textsc{MaxNCG}\xspace and \textsc{SumNCG}\xspace, in this order of presentation and importance, but we deviate from the standard full-knowledge model, and we explore the theoretical implications on the two games induced by the assumption that players have a partial view of the network. More precisely, we consider the players to have only knowledge of their \emph{$k$-neighborhood} (as in \cite{CPV09}), i.e., each player knows $k$ and the entire network up to the nodes at distance at most $k$ from herself. Furthermore, we also assume that players do not even know the size $n$ of the network (in distributed computing terminology, the system is \emph{uniform}).\footnote{According to the spirit of the game, we assume that in our model the players initially sit on a connected network.}
Despite of this partial knowledge of the network structure, the players keep on using the entire network, and so their cost function is still given by Eqs. \eqref{max} and \eqref{sum}, respectively. However, such a cost must now be revised as implicitly incurred by the players --- as a consequence of using the network --- rather than being explicitly known.
On the other hand, consistently with the model, the strategy space of a player is now restricted to selecting a subset of nodes in her $k$-neighborhood.
So, it is in the best interest of a player to reduce an unknown \emph{global} cost, but with the limitation of only knowing and modifying a \emph{local} portion of the network.
This ambitious task must be modeled through a coherent definition of the players' rational behavior, as we explain in the following.
Actually, a player has a partial (defective) view of the network, and thus before taking a step, she has to evaluate whether such a choice is convenient in \emph{every} realizable network which is compatible with her current view. More formally, let $\sigma_u$ be the strategy played by player $u$, and define $\Sigma|_{\sigma_u}$ to be the set of strategy profiles $\sigma = (\sigma_{-u},\sigma_u)$ of the players such that the network $G(\sigma)$ is realizable according to player $u$'s view. Let
\begin{equation}
\label{eq:delta}
\Delta(\sigma_u,\sigma_u') = \max_{\sigma \in \Sigma|_{\sigma_u}} \{C_u( \sigma_{-u},\sigma'_u)-C_u(\sigma)\}
\end{equation}
\noindent
\hide{let ${\cal G}$ denote the set of realizable networks according to $u$'s view, and let $C_u(\sigma,G)$ denote what her cost would be if the actual network was $G \in \cal G$. Then, let
\begin{equation}
\label{eq:delta}
\Delta(\sigma_u,\sigma_u') = \max_{G \in \cal G} \{C_u((\sigma_{-u},\sigma'_u),G)-C_u(\sigma,G)\}
\end{equation}}
\noindent
denote the worst possible cost difference player $u$ would have in switching from $\sigma_u$ to $\sigma'_u$. For our model, we use a suitable equilibrium concept (weaker than NE), that we call \emph{Local Knowledge Equilibrium} (LKE), and which is defined as a strategy profile $\bar{\sigma}$ such that for every player $u$ and every strategy profile $\sigma_u$, we have that $\Delta(\bar{\sigma}_u,\sigma_u) \geq 0$.
As the set of realizable networks $G(\sigma)$ can be infinite, it might appear that a player is not even able to determine if a strategy is convenient.
However, in Section~\ref{sect:conv_br} we will show that, in contrast with the intuition, this is not the case. In particular, for \textsc{MaxNCG}\xspace we will show that the worst-case scenario for a player is the one in which the network coincides with her view.
Therefore, the player only needs to take into account her view when evaluating a new strategy, ignoring in some sense the portion of the network she cannot see.
This also means that, in our model, a player behaves \emph{exactly} as she would do in the full-knowledge game played on the graph induced by her $k$-neighborhood.
This crucial analogy (which, in contrast, does not hold so tightly for \textsc{SumNCG}\xspace, as we discuss later) also explains why we put first the local-knowledge version of \textsc{MaxNCG}\xspace in our study: it allows us to fairly compare our game with the traditional \textsc{MaxNCG}\xspace, since we are only concerned on how the bounded view will impact on (the dynamics of) the game, given that the behavior of the players will remain the same.
Besides that, we point out another remarkable property of our model:
differently from the standard model, we have that the computational hardness of establishing an improving strategy is now depending not on the size of the entire network, but only on the size of her $k$-neighborhood.
Therefore, although in principle this latter one could be $\Theta(n)$ already for small values of $k$, we believe that in practice the situation may be quite different, as the size of the known network is expected to be constant (or at least very small) compared to $n$.
Regarding \textsc{SumNCG}\xspace, it is instead easy to see that if a player increases the distance of any
vertex $x$ at distance exactly $k$ in her view, then her cost might increase. Indeed, as the rest of the network is unknown to the player, it might be the case that the distance of a large number of non-viewable nodes will increase as well. Nevertheless, for every other strategy, we show that the worst-case network coincides with the player's view.
In conclusion, the above discussion shows that the player can choose and evaluate her strategies as in the classical \textsc{SumNCG}\xspace, but with the exception of the strategies that increase the distances of the nodes at distance $k$, which are forbidden. Thus, there is a dyscrasia between the full- and the local-knowledge version of the game: in this latter one, a player will actually have a more conservative behavior, since an improving strategy w.r.t. her partial view (i.e., ignoring what is outside her view) would not necessarily be an improving strategy w.r.t. the entire network (as it was the case for \textsc{MaxNCG}\xspace)!
\paragraph{Our results}
Having in mind this solution concept and this asymmetric relationship of the two local-knowledge games w.r.t. to their classical counterpart, we characterize the corresponding space of equilibria w.r.t. the social optimum through the study of upper and lower bounds to the PoA. We remark that, as the set of LKEs is broader than the set of NEs, the PoA in our model can only be worse than the PoA in the full-knowledge model.
First, we consider \textsc{MaxNCG}\xspace, and we give three lower bounds to the PoA which are based on different constructions each holding in different ranges of $k$ and $\alpha$. One of these is based on a dense graph, while the other two will have an high social cost due to their large diameter. In this latter case, the difficulty of the construction relies on the fact that, when $\alpha$ is small, we need to guarantee that no player can decrease her cost by buying new edges. We deal with this issue by carefully exploiting the defective views of the players, i.e., we provide a non-trivial construction where every player is not aware that buying a small number of edges would reduce her cost.
We also provide an upper bound to the PoA by considering both the density and the diameter of an equilibrium graph. In order to prove this bound, we take inspiration from techniques successfully used, for example, in \cite{ADH10,DHM12}, which allow to (lower) bound the number of nodes within a certain distance from a player. However, these techniques cannot be directly applied to our model since they require additional work to cope with the concept of locality.
The bounds to the PoA that arise from the various combinations of these results are discussed in detail in Section~\ref{sec:conclusions_max}, and they are essentially tight for many ranges of $\alpha$ and $k$. Here we just outline some prominent implications of our results.
For example, for constant values of $k$ (regardless of $\alpha$) we are able to exhibit stable graphs having diameter $\Omega(n)$. This immediately implies that the PoA is $\Omega(\frac{n}{1+\alpha})$, which is fairly bad.
However, one might expect the PoA to decrease for large values of $\alpha$. This is not the case, as we can show a tight lower bound of $\Omega(n^\frac{1}{\Theta(k)})$. This is in sharp contrast with the classical full-knowledge version of the game where the PoA is constant as soon as $\alpha \ge 129$.
On the other hand, when $k$ increases, the PoA decreases, although this happens quite slowly. Indeed, even when $k=O(2^{\sqrt{\log n}})$ and $\alpha=O(\log n)$ the PoA is still $\Omega(n^{1-\epsilon})$ for every $\epsilon>0$.
On the bright side, as soon as $k=\Omega(n^\epsilon)$ for any $\epsilon > 0$, we have that, in every LKE, each player has a complete knowledge of the network, and so the PoA coincides with the PoA of the full-knowledge game, hence it is mostly constant.
Then, we consider the sum version of the game, and we show that some of the lower bound schemes used for \textsc{MaxNCG}\xspace can be extended to \textsc{SumNCG}\xspace as well. In particular, a strong lower bound of $\Omega\big(\frac{n}{k}\big)$ to the PoA holds if $k \le c' \cdot \sqrt[3]{\alpha}$ and $\alpha \le n$, for a suitable constant $c'$. Observe that the latter lower bound is at least $\Omega(n^{\frac{2}{3}})$. Moreover, we show that for $\alpha \le n$, the set of LKEs coincides with the set of NEs as soon as $k \ge c \cdot \sqrt{\alpha}$, for a suitable constant $c$. Thus, in this region the PoA is constant, as a consequence of the corresponding result for the full-knowledge version of the game. Unfortunately, we cannot exhibit non-trivial upper bounds for the remaining values of the $(\alpha,k)$-space, and we leave this as a future goal of our research.
\paragraph{Paper organization}
The paper is organized as follows: in Section~\ref{sect:conv_br} we characterize the player's behavior, and we provide some remarks on the complexity of computing a best-response strategy and on the convergence issues of the iterated version of the game. In Sections~\ref{sect:PoA} and ~\ref{sec:sum} we focus on the main results of this paper, namely the study of the PoA for \textsc{MaxNCG}\xspace and \textsc{SumNCG}\xspace, respectively. Then, in Section~\ref{sec:experiments} we provide an extensive set of experiments, which for a twofold reason we restricted to \textsc{MaxNCG}\xspace: on one hand, we have for it a more exhaustive theoretical characterization of the PoA space to compare with, and on the other hand, as we will explain in more detail in the section, for \textsc{MaxNCG}\xspace is computationally feasible to find a best-response strategy of a player for reasonably large values of $n$ and $k$. Finally, Section~\ref{sec:conclusions} concludes the paper and provides few directions of future research.
\section{Preliminary remarks}
\label{sect:conv_br}
We start by showing that, despite the defective knowledge of the network, a player is able to evaluate whether a strategy is convenient in a worst-case scenario. In particular, when a player $u$ changes her strategy from $\sigma_u$ to $\sigma'_u$, she needs to evaluate $\Delta(\sigma_u, \sigma_u')$ by figuring out a realizable network $G(\sigma)$ maximizing \eqref{eq:delta}. In the following propositions we will characterize this worst-case network for both \textsc{MaxNCG}\xspace and \textsc{SumNCG}\xspace.
Let $H$ be the view of $u$ in $G(\sigma)$, i.e. the subgraph of $G(\sigma)$ induced by the $k$-neighborhood of $u$, and let $G \in \mathcal{G}$ be a generic realizable network w.r.t.\ $H$.
Let $G' = G \setminus ( \{u\} \times \sigma_u) \cup ( \{u\} \times \sigma'_u)$ be the network $G$ after the strategy change.
In a similar manner, let $H'=H \setminus ( \{u\} \times \sigma_u) \cup ( \{u\} \times \sigma'_u)$ be the old view of $u$ modified according to the strategy change. Notice that $H'$ might not coincide with the view of $u$ in $G'$.
\begin{proposition}
\label{prop:maxncg_best_response}
In \textsc{MaxNCG}\xspace, the worst-case network maximizing \eqref{eq:delta} coincides with $H$.
\end{proposition}
\begin{proof}
Consider a generic network $G$. In switching from $\sigma_u$ to $\sigma_u'$ the player $u$ is paying an additional cost of:
\begin{equation}
\label{eq:cost_difference_max}
\alpha(|\sigma_u'|-|\sigma_u|) + \max_{v} d_{G'}(u,v) - \max_{v} d_{G}(u,v).
\end{equation}
Let $y$ be the vertex maximizing $\max_{v} d_{G'}(u,v)$. If $d_{G}(u,y)<k$ then $y$ belongs to both $H$ and thus to $H'$ as well, therefore the formula \eqref{eq:cost_difference_max} can be upper-bounded by $\alpha(|\sigma_u'|-|\sigma_u|) + \max_{v \in V(H')} d_{H'}(u,v) - \max_{v \in V(G)} d_{G}(u,v)$, which is attained when $G=H$ (hence $G'=H'$) since any graph in $\mathcal{G}$ is a supergraph of $H$.
Otherwise, let $x$ be the unique vertex in a shortest path $\pi$ from $u$ to $y$ in $G$ such that $d_G(u,x)=k$. Notice that the subpath of $\pi$ between $x$ and $y$ also lies in $G'$, hence $d_{G'}(x,y) \le d_G(x,y)$. We can rewrite \eqref{eq:cost_difference_max} as follows: $\alpha(|\sigma_u'|-|\sigma_u|) + d_{G'}(u,y) - \max_{v} d_{G}(u,v) \le \alpha(|\sigma_u'|-|\sigma_u|) + d_{G'}(u,x) + d_{G'}(x,y) - d_{G}(u,y) \le \alpha(|\sigma_u'|-|\sigma_u|) + d_{H'}(u,x) + d_{G}(x,y) - d_{G}(u,x) - d_G(x,y) = \alpha(|\sigma_u'|-|\sigma_u|) + d_{H'}(u,x) - k \le \alpha(|\sigma_u'|-|\sigma_u|) + \max_{v \in V(H')} d_{H'}(u,v) - \max_{v \in V(H)} d_H(u,v)$.
\end{proof}
Notice that, according to the above discussion, the players do not even need to know the value of $k$ in order to play the game.
Regarding \textsc{SumNCG}\xspace, let us define as $F$ the set a vertices at distance exactly $k$ from $u$ in $H$.
\begin{proposition}
\label{prop:sumncg_best_response}
In \textsc{SumNCG}\xspace, every strategy that in\-creas\-es the distance of some vertex of $F$ in $H'$ is not an improving strategy for $u$. For every other strategy of $u$, the worst-case network maximizing \eqref{eq:delta} coincides with $H$.
\end{proposition}
\begin{proof}
Consider a generic network $G$. When $u$ switches from $\sigma_u$ to $\sigma_u'$ she pays an additional cost of:
\begin{equation}
\label{eq:cost_difference_sum}
\alpha(|\sigma_u'|-|\sigma_u|) + \sum_v d_{G'}(u,v) - \sum_v d_G(u,v).
\end{equation}
We first notice that if there exists a vertex $y$ such that $d_G(u,y)=d_H(u,y)=k$ and $d_{H'}(u,y) > k$ then $u$ is not improving in the worst-case scenario. Indeed we can make \eqref{eq:cost_difference_sum} positive by letting $G$ be equal to the graph $H$ where a large number $\eta$ of nodes has been appended to $y$, as \eqref{eq:cost_difference_sum} becomes at least
$\alpha(|\sigma_u'|-|\sigma_u|) + \sum_{v \in H} d_{G'}(u,v) - \sum_{v \in H} d_G(u,v) + \eta$.
Therefore, we restrict to strategies $\sigma_u'$ where if $d_H(u,y)=k$ then $d_{H'}(u,y) \le k$.
Call $P$ the set of vertices $x$ such that $d_G(u,x) < k$, the formula \eqref{eq:cost_difference_sum} becomes:
$ \alpha(|\sigma_u'|-|\sigma_u|) + \sum_{v \in P} \big( d_{G'}(u,v) - d_G(u,v) \big) +
\sum_{v \not\in P} \big( d_{G'}(u,v) - d_G(u,v) \big) \le \alpha(|\sigma_u'|-|\sigma_u|) + \sum_{v \in P} \! \big( d_{H'}(u,v) - d_H(u,v) \big) + \sum_{v \not\in P} \! \big( d_{G'}(u,v) \linebreak - d_G(u,v) \big) \le \alpha(|\sigma_u'|-|\sigma_u|) + \sum_{v \in P} \big( d_{H'}(u,v) - d_H(u,v) \big)$, since for every $v \not\in P$ we have $d_{G'}(u,v) \le d_G(u,v)$.
This upper bound to \eqref{eq:cost_difference_sum} is attained when $G=H$.
\end{proof}
We now provide some remarks on the complexity of computing a best-response strategy and on convergence issues.
We start by noticing that the \mbox{\sf NP}-hardness reductions which are known in the full-knowledge model for finding a best response in \textsc{SumNCG}\xspace and \textsc{MaxNCG}\xspace can be extended to our games for every $k \ge 2$ and $k\geq 1$, respectively.
Indeed, the full-knowledge versions of \textsc{MaxNCG}\xspace and \textsc{SumNCG}\xspace are \mbox{\sf NP}-hard for $\alpha=2/n$ (see \cite{MS10}) and for every $1<\alpha<2$ (see \cite{FLM03}), respectively. Both reductions are from the {\sc Minimum Dominating Set} problem, that is the problem of finding a minimum cardinality subset $U$ of vertices of an undirected graph $G$ such that every vertex $v \in V(G)$ has a neighbor in $U$ or is in $U$ itself. In both reductions, any best response of a new single player that joins the network $G$ is buying all the edges towards the vertices of a minimum dominating set of $G$. Since in the full-knowledge versions of \textsc{MaxNCG}\xspace and \textsc{SumNCG}\xspace, the best response of a player is independent from the strategy she is actually playing, both games remain \mbox{\sf NP}-hard even if the new player is initially buying all the edges towards all the other players. Turning back to our games, this is equivalent to say that the new player always sees the entire network. Moreover, for $k\geq 2$, the new player is aware to have the full-knowledge of the network as she knows $k$, and does not see vertices at distance 2 from her. Therefore, by Proposition \ref{prop:maxncg_best_response}, we have that \textsc{MaxNCG}\xspace is \mbox{\sf NP}-hard for every $k\geq 1$ and $\alpha=2/n$. Furthermore, by Proposition \ref{prop:sumncg_best_response}, we have that \textsc{SumNCG}\xspace is \mbox{\sf NP}-hard for every $k\geq 2$ and every $1<\alpha<2$.
Concerning the convergence issue, let us consider the iterated version of the game.\footnote{We assume that the players other than being \emph{myopic} are also \emph{oblivious}, namely at each time they only argue about the current view, without taking care of previous views.} A natural question is whether a better- or best-response dynamics always converges to an equilibrium state.
Unfortunately, a negative answer to this question follows from the divergence results presented in \cite{KL13} on the full-knowledge model for \textsc{SumNCG}\xspace and \textsc{MaxNCG}\xspace, since they are based on an instance having (small) constant diameter. This immediately implies the existence of a cycling best-response dynamics for both games as soon as $k \ge c$ for a constant $c$.
\section{Results for \protect\textsc{MaxNCG}\xspace}
\label{sect:PoA}
For the sake of exposition, but also for all the other reasons we discussed in the introduction, we first analyze \textsc{MaxNCG}\xspace, and then we consider \textsc{SumNCG}\xspace. Moreover, for technical convenience, we will assume $\alpha > 1$ although our constructions can also be extended to the case $\alpha \le 1$. We recall that in the full-knowledge version of the game, for $\alpha > 1$ the spanning star is the social optimum and has a cost of $\Theta(\alpha n + n)$.
\subsection{Lower bounds for \textsc{MaxNCG}\xspace}
\label{sec:lower_bounds_max_m3}
We present three lower bounds to the PoA based on three graphs with high social cost which are in equilibrium for different ranges of $\alpha$ and $k$.
The first is a cycle. We have the following.
\begin{lemma}
If $k\ge1$ and $\alpha \ge k-1$ then $\mbox{PoA}=\Omega(\frac{n}{1+\alpha})$.
\end{lemma}
\begin{proof}
Consider a cycle on $n \ge 2k+2$ vertices where each player owns exactly one edge.
The view of each player $u$ is a path of length $2k$ with $u$ as the center vertex. In order to decrease her eccentricity $u$ has to buy at least one edge. This will decrease the usage cost of $u$ by at most $k-1$ and increase the building cost of $u$ by at least $\alpha$. Then, $\mbox{PoA} = \Omega\left( \frac{\alpha n + n^2}{\alpha n+ n} \right) = \Omega\left( \frac{n}{1+\alpha} \right)$.
\end{proof}
The next lower bound is based on a dense graph of large girth.
\begin{lemma}
\label{lemma:lb_poa_max_dense}
For each $2 \le k = o(\log n)$ and $\alpha \ge 1$ the PoA is $\Omega(n^{\frac{1}{2k-2}})$.
\end{lemma}
\begin{proof}
For each even integer $g \ge 6$ and prime power $q$ there exists a $q$-regular graph of girth at least $g$ with $n$ vertices and $\Omega(n^{1+\frac{1}{g-4}})$ edges \cite{LUW95}.
We choose $g=2k+2$ and construct such a graph.
The view of each player $u$ is a tree of height $k$ with $q(q-1)^{i-1}$ vertices on level $i$. Moreover, the player $u$ owns at most $q$ edges.
In order to reduce her usage cost by $i$, player $u$ must buy at least $q(q-1)^i-q$ additional edges. If we choose $q \ge 3$ then the increase in the building cost will exceed the decrease in the usage cost. Hence, we have that the PoA is at least $\Omega(\frac{\alpha n^{1+\frac{1}{g-4}}}{ (1+\alpha) n})= \Omega(n^{\frac{1}{2k-2}})$.
It can be shown that the previous construction holds for $k = o(\log n)$.
\end{proof}
The last lower bound is based on a sparse graph with large diameter.
The construction is non-trivial and it is a generalization of the graph shown in \cite{ADH10}. Although the precise definition is critical, we now give some intuition on how the graph is built.
Roughly speaking, the original graph resembles a $2$-dimensional square grid that was rotated by $45^\circ$ and had the vertices on the opposite sides identified in order to form a toroidal shape.
This graph has several useful properties: it is vertex-transitive and the diameter is about the length of a ``side'' of the grid. Moreover, if the value of $k$ is small, each player $u$ is not aware of the toroidal shape as she only sees a ``square'' subgraph. This subgraph has $4$ vertices at distance $k$ from $u$ whose pairwise distance are $2k$. This fact can be used to show that, actually, this graph is stable for small values of $\alpha$ and $k$, e.g. $\alpha=k=1$. Unfortunately, this is no longer true for larger values of $k$ since, for example, the addition of $4$ edges suffices to reduce the eccentricity of $\Omega(k)$. Moreover, if $\alpha$ is large, a player has convenience in removing an edge as this results in a constant increase in her eccentricity.
To deal with these issues we generalize this construction in three ways. First, we increase the number of dimensions from $2$ to a parameter $d$ so that the graph now resembles a rotated $d$-dimensional cube grid where each face has been identified with the opposite one. For each vertex $u$ we are now able to find $2^d$ other vertices that are at a distance of $k$ from $u$ and whose pairwise distances are at least $2k$.
Second, in order to get a graph with large diameter, we no longer restrict the dimensions to be equal to each other. Intuitively, instead of starting with a $d$-dimensional ``cube'', we start with a $d$-dimensional hyper-rectangle.
Finally, we ``stretch'' the graph by replacing each edge with a path of length $1 \le \ell = \Theta(\alpha)$ between its endpoints. This causes the addition of $\ell-1$ new vertices per edge. We call these new vertices ``non-intersection vertices'' to distinguish them from the already existing ``intersection vertices''. Non-intersection vertices will buy all the links of the graph and we will show that they cannot remove edges as this would result in an increase of at least $\Omega(\ell)$ in their eccentricity.
\begin{figure}[!t]
\begin{center}
\includegraphics[scale=0.45]{3d-torus.eps}
\end{center}
\caption{Graph with $d=2$ dimensions of sizes $\delta_1=15$ and $\delta_2=5$, and $\ell=2$. Intersection vertices have a bigger size than non-intersection vertices. In this example the view of the intersection vertex $(k^{*},k^{*})$ for $k=4$ is in red. Notice that the vertex $(k^{*},k^{*})$ lies on an invisible portion of the torus.}
\label{fig:3d-torus}
\end{figure}
We now give the details of the construction, which will depend on $\ell \ge 1$, $d \ge 2$, and $\delta_1,\dots,\delta_d \ge 1$, where $\delta_i$ is the ``length'' of the $i$-th dimension. Vertices will be named after their coordinates and we will interpret the $i$-th coordinate modulo $2 \delta_i$, that is $2\delta_i \equiv 0$.
The graph is built by starting from an empty graph, proceeding in the following way:
add the set of \emph{intersection vertices}, each of these vertices is a $d$-tuple $(\ell \cdot a_1, \ell \cdot a_2, \dots, \ell \cdot a_d)$ such that $a_1 \equiv a_2 \equiv \dots \equiv a_d \pmod{2}$ where each $a_i$ with $1\le i \le d$ is an integer between $0$ and $2\delta_i-1$.
Then, connect each intersection vertex $(x_1, \dots, x_d)$ to $2^d$ other vertices, using a path of length $\ell$ (so if $\ell=1$ we only need to add edges). More precisely, we connect such a vertex to $(x_1 \pm \ell, x_2 \pm \ell, \dots, x_d \pm \ell)$ for every possible choice of the $\pm$ signs. We label the $\ell-1$ non-intersection vertices on the paths by varying the coordinates of the endpoints according to the choice of $\pm$ signs, that is we traverse the path from one endpoint to another and label each non-intersection vertex by adding or subtracting $1$ from the coordinates of the previous vertex. In the following, for convenience, when we choose a vertex, we will assume that the $i$-th coordinate is between $0$ and $2\delta_i-1$.
We will consider graphs where $\delta_1 = \dots = \delta_{d-1} = \left\lceil \frac{k}{\ell} \right\rceil + 1$ and $\delta_d \ge \left\lceil \frac{k}{\ell} \right\rceil + 1$.
Let $k^*=\ell(\delta_1-1)$, $u$ be the vertex $(k^*,\dots,k^*)$, and $u^\prime$ be the vertex $(k^*+\ell, \dots, k^*+\ell)$.
Two examples of this construction, with different parameters, are shown in Figure~\ref{fig:3d-torus} and in Figure~\ref{fig:torus}.
The following result is not hard to prove:
\begin{lemma}
\label{lemma:torus_distance}
Let $x=(x_1, \dots, x_d)$ and $y=(y_1, \dots, y_d)$ be two vertices. We have $d(x, y) \ge \max_{1 \le i \le d}\min\{|x_{i}-y_{i}|,2\delta_i\ell-|x_{i}-y_{i}|\}$. If at least one of $x$ and $y$ is an intersection vertex, then the previous inequality is strict.
\end{lemma}
\begin{corollary}
\label{cor:torus_diameter}
The diameter of the graph is at least $\ell \cdot \delta_{d}$.
\end{corollary}
\begin{proof}
By Lemma \ref{lemma:torus_distance} the distance between the vertex $(0, \dots, 0)$ and any vertex whose last coordinate is $\ell \delta_d$ is at least $\ell \delta_d$.
\end{proof}
We also consider an ``open'' version of the previous graph, that is built in a similar way except that we do not treat the coordinates in a modular fashion, so $a_i$ is now between $1$ and $\delta_i$, and we connect intersection vertices (with paths) only when all their coordinates differ by exactly $\ell$. It is not hard to see that the view of each player is isomorphic to a subgraph of this ``open'' graph, and that Lemma \ref{lemma:torus_distance} becomes:
\begin{lemma}
\label{lemma:open_torus_distance}
Let $x=(x_1, \dots, x_d)$ and $y=(y_1, \dots, y_d)$ be two vertices in the ``open'' graph.
We have $d(x, y ) \ge \linebreak \max_{1 \le i \le d} |x_{i}-y_{i}|$,
If at least one of $x$ and $y$ is an intersection vertex, then the previous inequality is strict.
\end{lemma}
\begin{figure}[!t]
\centering
\includegraphics[scale=0.7]{toro.eps}
\caption{Graph with $d=2$ dimensions of sizes $\delta_1=3$ and $\delta_2=4$, and $\ell=2$. Intersection vertices have a bigger size than non-intersection vertices. The vertices on the first row (resp. column) coincide with the corresponding vertices on last row (resp. column). In this example the view of the intersection vertex $(k^{*},k^{*})$ for $k=4$ is highlighted in gray.}
\label{fig:torus}
\end{figure}
We now prove a general lemma that will be very useful in the following proofs:
\begin{lemma}
\label{lemma:shorten_distance}
Let $H$ be a graph, $u \in V(H)$, and $L = \{v_1, \dots, \linebreak v_{|L|}\} \subseteq V(H) \setminus \{ u \}$.
If $d_H(u,v_i) \ge h $ for every $1 \le i \le |L|$, $d_H(v_i, v_j) \ge 2h-2$ for every $1\le i,j \le |L|$ with $i \not= j$, and $F$ is a set of edges such that: (i) each edge in $F$ has $u$ as an endpoint, and (ii) $d_{H+F}(u,v_i) < h$ for every $1 \le i \le |L|$, then it holds $|F| \ge |L|$.
\end{lemma}
\begin{proof}
Every shortest path from $u$ to a vertex $v \in L$ in $H+F$ must use exactly one edge $(u,y) \in F$, and $d_H(v,y) \le h-2$ must hold.
For every other vertex $v^\prime \in L \setminus \{u\}$, the shortest path between $u$ and $v^\prime$ in $H+F$ cannot use the edge $(u,y)$, otherwise we would have:
\begin{multline*}
d_{H+F}(u,v^\prime) = 1+d_H(y,v^\prime) = 1+d_H(y,v^\prime)+d_H(y,v) \\ -d_H(y,v)
\ge d_H(v^\prime, v) - d_H(y,v) \ge 2h-2 - (h-2) \ge h.
\end{multline*}
This implies that $F$ must contain at least one edge for each vertex of $L$.
\end{proof}
We now define the ownership of the edges. Consider the path $\langle u=x_0,x_1,\dots, x_{\ell -1},x_\ell=u' \rangle$ from $u$ to $u'$. For $i=1,\dots,\ell-1$, vertex $x_i$ buys the edge towards $x_{i-1}$, and $x_{\ell -1}$ also buys the edge towards $u'$. The ownership of the edges of the other paths are defined symmetrically. Observe that the intersection vertices buy no edges.
Given an intersection vertex $v=(x_1, x_2, \dots, x_d)$ we define $\mathcal{F}^h(v)$ as the set of vertices reachable by traversing an edge incident to $v$, and then proceeding in the same direction for a total of $h$ steps, i.e., $\mathcal{F}^h(v)=\{ (x_1 \pm h, x_2 \pm h, \dots, x_d \pm h)\mbox{, for every possible choice of $\pm$ signs} \}$. If $h \le k$ then $|\mathcal{F}^h(v)|=2^d$ and, by Lemma \ref{lemma:torus_distance}, the distance between $v$ and any of those vertices is exactly $h$.
The following lemmas are instrumental to prove the lower bound to the PoA for \textsc{MaxNCG}\xspace as they will provide sufficient conditions for intersection and non-intersection vertices to be in equilibrium.
\begin{lemma}
\label{lemma:intersection-equilibrium}
If $d \ge \log \frac{k-1}{\alpha}$ then the intersection vertices are in equilibrium.
\end{lemma}
\begin{proof}
By symmetry, let us consider only the intersection vertex $u$.
As $u$ has not bought any edge, she can modify her strategy only by buying new edges.
Every vertex in the set $\mathcal{F}^k(u)$ is at distance $k$ from $u$, moreover, by Lemma \ref{lemma:open_torus_distance}, any two distinct vertices in $\mathcal{F}^k(u)$ have a distance of at least $2k$ in the view of $u$.
By Lemma \ref{lemma:shorten_distance}, $u$ needs to buy at least $2^d$ edges in order to reduce her eccentricity.
If she does so, she saves at most $k-1$ on the usage cost while paying at least $\alpha 2^d$, but we have $\alpha 2^d \ge k-1$.
\end{proof}
\begin{lemma}
\label{lemma:non-intersection_far_vertices}
If $k\ge \ell$, for every non-intersection vertex $v$, there is a set $L$ of $2^d$ vertices at distance $k$ from $v$ and at distance at least $2k-\ell$ between each other in the view of $v$.
\end{lemma}
\begin{proof}
By symmetry, let $v=(k^*+\ell-z, \dots, k^*+\ell-z)$ with $1<z\le \left\lfloor \frac{\ell}{2} \right\rfloor$. The nearest intersecting vertex from $v$ is $u^\prime$.
Let $L^\prime = \mathcal{F}^{k-z}(u^\prime) \setminus \{ (k^*+\ell-(k-z), \dots, k^*+\ell-(k-z) ) \}$. Any pair of vertices in $L^\prime$ differs by at least one coordinate and, by Lemma \ref{lemma:open_torus_distance}, is at distance at least $2k-2z \ge 2k-\ell$ in the view of $v$.
Let $v^\prime$ be a vertex of $L^\prime$, the shortest path from $v$ to $v^\prime$ must contain either $u$ or $u^\prime$. By Lemma \ref{lemma:open_torus_distance}, if it contains $u$ then we have: $d(v,v^\prime)=d(v,u)+d(u,v^\prime) \ge \ell - z + \ell + k - z = 2\ell +k - 2z \ge k$, otherwise it contains $u^\prime$, and we have: $d(v,v^\prime)=d(v,u^\prime)+d(u^\prime,v^\prime) \ge z+k-z = k$. Therefore $d(u^\prime, v^\prime) \ge k$.
Now consider the vertex $y=(k^*+\ell-z-k, \dots, k^*+\ell-z-k)$. We have that $y \not\in L^\prime$ and, by Lemma \ref{lemma:open_torus_distance}, $d(v, y) \ge k$ and $d(y,v^\prime) \ge 2k$.
The claim follows as we can define $L=L^\prime \cup \{ y \}$.
\end{proof}
\begin{lemma}
\label{lemma:non-intersection-increase}
If $\frac{\alpha+1}{2} \le \ell \le \alpha + 2$ and $d \ge\log (\frac{k-1}{\alpha} + 2)$, then every non-intersection vertex $v$ is in equilibrium w.r.t.\ all the strategies that increase the number of bought edges.
\end{lemma}
\begin{proof}
Notice that each non-intersection vertex has \linebreak bought at most $2$ edges, and consider a strategy that increases the number of bought edges.
If $k \le \ell - 1$ then the building cost increases by at least $\alpha$ while the usage cost decreases at most by $k-1$, but we have $\alpha \ge \ell-2 \ge k-1$.
Otherwise, $k \ge \ell$ and, by Lemma \ref{lemma:non-intersection_far_vertices}, there exists a set of at least $2^d$ vertices at distance at least $k \ge k-\frac{\ell}{2}$ from $v$, and $2k-\ell$ between each other. By Lemma \ref{lemma:shorten_distance}, $v$ needs to have at least $2^d$ incident edges in order to reduce her eccentricity by at least $\frac{\ell}{2}$.
If, in the new strategy, $v$ has less than $2^d$ incident edges then the building cost increases by at least $\alpha$, and the usage cost decreases by at most $\frac{\ell}{2}-1$, but we have $\alpha \ge \frac{\ell}{2}-1$.
If, in the new strategy, $v$ has at least $2^d$ incident edges, then the building cost increases by at least $\alpha(2^d-2)$, and the eccentricity decreases by at most $k-1$, but we have $\alpha(2^d-2) \ge k-1$.
\end{proof}
\begin{lemma}
\label{lemma:non-intersection-decrease}
If $\ell \ge \alpha$, every non-intersection vertex $v$ is in equilibrium w.r.t.\ all the strategies that decrease the number of bought edges.
\end{lemma}
\begin{proof}
By symmetry, let $v=(k^*+z, \dots, k^*+z)$ with $1 \le z < \ell$. The vertex $v$ can decrease the number of bought edges by at most $1$, thus saving $\alpha$ on the building cost.
Let $G^\prime$ be the view of $v$ where the edges incident to the vertex $v$ have been removed, $x=(k^*+z+k, \dots, k^*+z+k)$, and $x^\prime = (k^*+z-k, \dots, k^*+z-k)$.
If $x$ and $x^\prime$ are not connected in $G^\prime$, then $v$ cannot decrease the number of bought edges.
Otherwise, let $\pi$ be a shortest path between $x$ and $x^\prime$ in $G^\prime$. Let $y=(k^*+z+h, \dots, k^*+z+h)$ be the first vertex in $\pi$ such that the following vertex in $\pi$ is different from $(k^*+i, \dots, k^*+i)$ for all values of $i \in \mathbb{Z}$. Notice that $y$ must be an intersection vertex, and let $y^\prime$ be the first intersection vertex following $y$ in $\pi$. At least one of the coordinates of $y^\prime$ must be $k^*+z+h+\ell$.
We have $d_{G^\prime}(v,x) = d_{G^\prime}(v,x^\prime)=+\infty$ and, by Lemma \ref{lemma:open_torus_distance}:
\begin{multline*}
d_{G^\prime}(x,x^\prime) \ge d_{G^\prime}(x,y) + d_{G^\prime}(y, y^\prime) + d_{G^\prime}(y^\prime, x^\prime) \\ \ge k-h + \ell + k + h + \ell \ge 2k +2\ell.
\end{multline*}
By Lemma \ref{lemma:shorten_distance}, $u$ needs at least $2$ incident edges for her eccentricity to be under $k+\ell$. If $v$ decreases the number of bought edges then $v$ has at most $1$ incident edge, and her usage cost increases by at least $\ell$, but we have $\alpha \le \ell$.
\end{proof}
\begin{lemma}
\label{lemma:non-intersection-equilibrium}
If $\alpha \le \ell \le \alpha + 2$, $k \ge \ell$, and $d \ge\log (\frac{k-1}{\alpha} + 2)$, then every non-intersection vertex $v$ is in equilibrium.
\end{lemma}
\begin{proof}
By Lemma \ref{lemma:non-intersection-increase} and Lemma \ref{lemma:non-intersection-decrease} $v$ is in equilibrium w.r.t.\ all the strategies that either increase or decrease the number of bought edges.
We now show that $v$ is also in equilibrium w.r.t.\ the strategies that do not change the number of bought edges. As the building cost of $v$ remains the same, $v$ must save on her usage cost in order to change her strategy.
We will show that $v$ cannot decrease her usage even when she buys the new edges in addition to the ones already bought.
If $v$ owns only one edge then let, by symmetry, $v=(k^*+z, \dots, k^*+z)$ with $1 \le z < \ell$, and let the bought edge be towards $(k^*+z+1, \dots, k^*+z+1)$. The vertices $(k^*+z+k, \dots, k^*+z+k)$ and $(k^*+z-k, \dots, k^*+z-k)$ are at distance $k$ from $v$ and $2k$ between each other. By Lemma \ref{lemma:shorten_distance}, $v$ needs at least $2$ new incident edges to decrease her eccentricity under $k$, but she can only add one.
If $v$ owns two edges then, by symmetry, let $v=(k^* + \ell -1, \dots, k^* + \ell -1)$ so she has an edge towards $u^\prime$.
The vertices in $\mathcal{F}^{k-1}(u^\prime) \setminus { (k+\ell-k+1, \dots, k+\ell-k+1) }$ are at distance $k$ from $v$ and $2k-2$ between each other, in the view of $v$. By Lemma \ref{lemma:shorten_distance}, $v$ needs at least $2^d-1 \ge 3$ new incident edges in order to decrease her eccentricity under $k$, but she can only add two.
\end{proof}
We are now ready to prove the following:
\begin{theorem}
\label{thm:torus_poa}
If $1 < \alpha \le k \le 2^{\sqrt{\log n} - 3}$, then the PoA of \textsc{MaxNCG}\xspace is $\Omega\bigg( \frac{n }{ \alpha \cdot 2^{ (\log \frac{k}{\ell} + 3 ) \log \frac{k}{\ell}}} \bigg)$.
\end{theorem}
\begin{proof}
Fix $\ell = \left\lceil \alpha \right\rceil$, and notice that $k \ge \ell \ge 2$ holds as $k$ must be an integer.
Fix $d=\left\lceil \log\left( \frac{k}{\ell} + 2 \right) \right\rceil$, this implies $d \ge 2$. We will use the following inequalities: $\log\left(\frac{k}{\ell}\right) \le d \le \log\left(\frac{k}{\ell}\right) + 3$. Finally, as already said, we set $\delta_1, \dots, \delta_{d-1}$ to $\left\lceil \frac{k}{\ell} \right\rceil + 1$.
In order to be $\delta_d \ge \delta_1$ it suffices for $k$ to be at most $2^{\sqrt{\log n} - 3}$, as shown by the following calculations.
The number of intersection vertices of the graph is $N=2\prod_{i=1}^d \delta_i = 2(\left\lceil \frac{k}{\ell} \right\rceil + 1)^{d-1} \delta_d $ while the total number of vertices is: $n=N+2^{d-1}N(\ell-1)=N(2^{d-1}(\ell-1)+1)$ therefore $N=\frac{n}{2^{d-1}(\ell-1)+1}\ge\frac{n}{2^{d-1}\ell}$.
\begin{multline*}
\hfill \delta_d \ge \delta_1 \iff
\frac{N}{2(\left\lceil \frac{k}{\ell} \right\rceil + 1)^{d-1}} \ge \left\lceil \frac{k}{\ell} \right\rceil + 1 \hfill \\
\hfill \Longleftarrow \frac{n}{2^d\ell} \ge \left( \left\lceil \frac{k}{\ell} \right\rceil + 1 \right)^d \Longleftarrow k \le \frac{n^{\frac{1}{d}}}{2} \ell^{1-\frac{1}{d}} - 2\ell \hfill \\
\hfill \Longleftarrow 6 k \le n^\frac{1}{d} \iff d \log 6k \le \log n \hfill \\
\hfill \Longleftarrow (\log k + 3)^2 \le \log n \Longleftarrow k \le 2^{\sqrt{\log n} - 3}. \hfill
\end{multline*}
By Corollary \ref{cor:torus_diameter}, the diameter of the graph is at least:
\begin{multline*}
\ell\,\delta_d = \Omega\bigg( \frac{N \ell^{d}}{2k^{d-1}} \bigg) = \Omega\bigg(\frac{n \ell^{d-1}}{2^d k^{d-1}}\bigg) = \Omega\bigg( \frac{n \ell^{d}}{ k^{d}} \bigg) \\
= \Omega\bigg( \frac{n }{ (\frac{k}{\ell})^{d}} \bigg) = \Omega\bigg( \frac{n }{ 2^{d \log \frac{k}{\ell}}} \bigg) = \Omega\bigg( \frac{n }{ 2^{ (\log \frac{k}{\ell} + 3 ) \log \frac{k}{\ell}}} \bigg).
\end{multline*}
By Lemma \ref{lemma:intersection-equilibrium} and Proposition \ref{lemma:non-intersection-equilibrium} the graph is in equilibrium.
As every vertex in the graph owns at most $2$ edges, the total number of edges is at most $2n$, and the PoA is:
\[
\Omega\left( \frac{\alpha n + n \ell \delta_d }{\alpha n} \right) = \Omega\left( \frac{\ell \delta_d }{\alpha} \right) = \Omega\bigg( \frac{n }{ \alpha \cdot 2^{ (\log \frac{k}{\alpha} + 3 ) \log \frac{k}{\alpha}}} \bigg). \qed \]
\end{proof}
\subsection{Upper bounds for \textsc{MaxNCG}\xspace}
Given a graph $H$, we denote by $\beta_{H,h}(v)$ the \emph{ball} of radius $h$ centered at node $v$ in $H$, namely the set of vertices whose distance from $v$ in $H$ is at most $h$. When the graph $H$ is clear from the context we will drop the corresponding subscript. The following lemma shows a relation between $k$ and the number of nodes that a player sees in an equilibrium graph $G$. A similar result is shown in \cite{DHM12} for the original game.
\begin{lemma}\label{lemma:lb_balls}
Let $G$ be an equilibrium graph whose radius is greater than or equal to $k/2$, and let $N=|\beta_{G,k}(u)|$ be the number of nodes that $u$ sees in $G$. If $\alpha \le k-1$ we have that $k=O( \min \{ \sqrt[3]{N \alpha^2}, \alpha \, 4^{\sqrt{\log N}} \})$.
\end{lemma}
\begin{proof}
First, we need to prove that $k= O(\sqrt{N\alpha})$, and do so by showing that $N=\Omega(k^2/\alpha)$. For every $1\leq i\leq k/2$, let $L_i$ be the vertices of $\beta_{G,k}(u)$ whose distance from $u$ is equal to $i$. We show that $|L_i|=\Omega(i/\alpha)$. If $u$ bought the edges towards all the vertices in $L_i$ she would decrease her eccentricity by at least $k-\max\{k-i,i\}=k-k+i=i$ and increase her building cost by $\alpha|L_i|$. As $G$ is an equilibrium graph, we have that $\alpha|L_i|\geq i$, i.e., $|L_i|\geq i/\alpha$. Therefore, $N\geq \sum_{i=1}^{\lfloor k/2\rfloor}|L_i|\geq\sum_{i=1}^{\lfloor k/2\rfloor}i/\alpha=\Omega(k^2/\alpha)$.
Now we prove that $k=O(\sqrt[3]{N \alpha^2})$ by showing that $N=\Omega(k^3/\alpha^2)$. Let $h=\lfloor k/8\rfloor-1$, and let $\bar h=k-2h=k-2\lfloor k/8 \rfloor+2$. By the choice of $h$ and $\bar h$, every path of length less than or equal to $2h$ between two vertices of $\beta_{G,\bar h}(u)$ is entirely contained in $\beta_{G,k}(u)$. We select a subset of vertices in $\beta_{G,\bar h}(u)$ as {\em center points} by the following greedy algorithm. First, we unmark all vertices in $\beta_{G,\bar h}(u)$. Then we repeatedly select an unmarked vertex $x$ in $\beta_{G,\bar h}(u)$ as center point, and mark all unmarked vertices in $\beta_{G,\bar h}(u)$ whose distances in the graph induced by $\beta_{G,k}(u)$ are at most $2h$ from $x$.
Suppose that we select $l$ vertices $x_1,x_2,\dots,x_l$ as center points. By construction, every vertex in $\beta_{G,k}(u)$ has distance of at most $4h$ to some center point. If player $u$ bought the $l$ edges towards the $l$ vertices $x_1,x_2,\dots,x_l$, she would decrease her eccentricity w.r.t.\ all the vertices in $\beta_{G,k}(u)$ by at least $k-(4h+1)\geq k-4\lfloor k/8\rfloor +3>k/2$ and increase her building cost by $\alpha l$. Because $G$ is an equilibrium graph, we have $\alpha l\geq k/2$, thus $l\geq k/(2\alpha)$. By the choice of $h$ and $\bar h$, the distance in $G$ between any pair of center points is greater than or equal to $2h+1$, furthermore $\beta_{G,h}(x_i)\subseteq \beta_{G,k}(u)$ for every $i=1,\dots,l$. As a consequence, the balls of radius $h$ centered at the center points are pairwise disjoint, and thus
\begin{equation*}
N=|\beta_{G,k}(u)|\geq \sum_{i=1}^{l}|\beta_{G,h}(x_i)|= l \cdot \Omega(k^2/\alpha)=\Omega(k^3/\alpha^2).
\end{equation*}
Finally we prove that $k=O\Big(\alpha \,4^{\sqrt{\log N}}\Big)$ by showing that $N=\Omega\Big(2^{\log_4^2(k/\alpha)}\Big)$ for every value of $\alpha \le k-1$.
First, we prove the following useful claim:
\begin{claim}\label{lemma:the_growth_of_balls}
Let $i< k/5$, and let $\bar N=\min_{v \in V}|\beta_{G,i}(v)|$. Either there exists a vertex having eccentricity strictly less than $5i$ or $|\beta_{G,4i+1}(v)|\geq (\bar N i)/\alpha$ for every vertex $v$.
\end{claim}
\begin{proof}
If there is a vertex having eccentricity strictly less than $5i$, then the claim is obvious. Otherwise, for every vertex $v$, we have that the eccentricity of $v$ is greater than or equal to $5i$. Let $S$ be the set of vertices whose distance from $u$ is $3i+1$. By the choice of $i$, every path of length less than or equal to $2i$ between a pair of vertices in $\beta_{G,3i+1}(u)$ is entirely contained in $\beta_{G,k}(u)$. We select a subset of $S$, called {\em center points}, by the following greedy algorithm. First we unmark all vertices in $S$. Then we select an unmarked vertex $x \in S$ as a center point, mark all unmarked vertices in $S$ whose distance from $x$ is less than or equal to $2i$, and assign these vertices to $x$.
Suppose that we select $l$ vertices $x_1,x_2,\dots,x_l$ as center points. We prove that $l\geq i/\alpha$. If player $u$ bought the $l$ edges towards the vertices $x_1,x_2,\dots,x_l$, she would decrease her eccentricity w.r.t.\ all the vertices in $\beta_{G,k}(u)$, by at least $k-\max\{k-i,3i+1\}=k-(k-i)=i$. Because $u$ has not bought these edges, we must have $l\alpha\geq i$.
According to the greedy algorithm, the distance between any pair of center points is greater than or equal to $2i+1$; hence the balls of radius $i$ centered at the vertices $x_j$ are pairwise disjoint. Therefore,
\begin{equation*}
\left |\bigcup_{j=1}^{l}\beta_{G,i}(x_j)\right|=\sum_{j=1}^{l}|\beta_{G,i}(x_j)|\geq l \bar N\geq (\bar N i)\alpha.
\end{equation*}
For every $j=1,\dots, l$, we have $d(u,x_j)=3k+1$, so $\beta_{G,i}(x_j)\subseteq \beta_{G,4i+1}(u)$. Therefore, $|\beta_{G,4i+1}(u)|\geq (\bar N i)/\alpha$.
\end{proof}
Let $\bar N_i=\min_{v \in V}|\beta_{G,i}(v)|$. Because $G$ is a connected equilibrium and $\lceil \alpha \rceil\leq k$, $\bar N_{\lceil\alpha\rceil} \geq \lceil\alpha\rceil$. By Lemma \ref{lemma:the_growth_of_balls}, for every $i<k/5$, either there exists a vertex having eccentricity strictly less than $5i$ or $\bar N_{4i+1}\geq \bar N_{i}$. Define the numbers $a_0,a_1,\dots$ using the recurrence relation $a_i=4a_{i-1}+1$ with $a_0=\lfloor\alpha\rfloor$. By induction, $a_i\geq \alpha 4^i$. If the radius of $G$ is strictly less than $k$, then let $j$ be the least number such that the radius of $G$ is less than or equal to $5a_j$; otherwise, let $j$ be the least number such that $a_j\geq k/5$. As the radius of $G$ is greater than or equal to $k/2$, we have that $a_j=\Theta(k)$. By definition of $j$, $\bar N_{a_{i+1}}\geq (a_i\bar N_{a_i})/\alpha\geq 4^i\bar N_{a_i}$, for every $i<j$. From these inequalities we derive that $\bar N_{a_j}\geq 4^{\sum_{i=0}^{j-1}i}$. But $\bar N_{a_j}\leq N$, so $\sum_{i=1}^{j-1}i=j(j-1)/2\leq \log_4 N$. This inequality implies that $j\leq 1+\sqrt{2\log_4 N}=1+\sqrt{\log N}$. Solving the recurrence relation, $a_j=O(\alpha \, 4^j)=O(\alpha 4^{\sqrt{\log N}})$. As $a_j=\Theta(k)$, $k=O(\alpha 4^{\sqrt{\log N}})$. \qed
\end{proof}
From this, it immediately follows:
\begin{corollary}
\label{cor:complete_knowledge}
If $\alpha \le k-1$ and $k > c \cdot \min \{ n, \sqrt[3]{n \alpha^2}, \linebreak \alpha \, 4^{\sqrt{\log n}}\}$, for a suitable constant $c$, then in every equilibrium graph each player sees the whole graph, thus the set of LKEs coincides with the set of NEs.
\end{corollary}
Now, we provide an upper bound to the diameter of an equilibrium graph.
\begin{lemma}
\label{lemma:diameter-balls}
Let $G$ be an equilibrium graph of diameter $d$. If $|\beta_{G,k}(v)| \ge \gamma$ for every vertex $v$, then $ d \le \frac{3kn}{\gamma}$.
\end{lemma}
\begin{proof}
Let $\pi=(x_0,\dots,x_d)$ be a diametral path of $G$.
We select a set of vertices $C$ such that the $k$-neighborhoods of the vertices of $C$ are pairwise disjoint and cover all the nodes of $\pi$.
We must have $|C| \gamma \le n$, which implies that $|C| \le \frac{n}{\gamma}$, and thus the diameter of $G$ is at most $|C|(2k+1) = \frac{(2k+1)n}{\gamma} \le \frac{3kn}{\gamma}$.
\end{proof}
\begin{lemma}\label{lemma:UB to diameter}
If $\alpha \le k-1$ the diameter of an equilibrium graph $G$ is $O\Big( \min \{ \frac{n\alpha^2}{k^2}, \frac{k n}{2^{\log_4^2 \frac{k}{\alpha}}} \} \Big)$.
\end{lemma}
\begin{proof}
Consider a generic vertex $v$, and let $N=|\beta_{G,k}(v)|$. From Lemma \ref{lemma:lb_balls}, we have that $k \! = \! O( \min\{\sqrt[3]{N\alpha^2}, \alpha 4^{\sqrt{\log N}} \})$, which implies that $N=\Omega(\max \{ \frac{k^3}{\alpha^2}, 2^{\log_4^2 \frac{k}{\alpha}} \} )$. Now, using Lemma \ref{lemma:diameter-balls}, we have that the diameter of $G$ must be $O\Big( \min \{ \frac{n\alpha^2}{k^2}, \frac{k n}{2^{\log_4^2 \frac{k}{\alpha}}} \} \Big)$.
\end{proof}
We now derive an upper bound to the density of an equilibrium graph. We argue on the girth of the graph in a way similar to \cite{DHM12}.
\begin{lemma}\label{lemma:density}
The number of the edges of an equilibrium graph $G$ is $O(n^{1+\frac{2}{\min\{\alpha,2k\}}})$.
\end{lemma}
\begin{proof}
Let $g$ be the girth of $G$. We first show that $g \ge 2+ \min\{\alpha, 2k\}$, and then the claim follows from the fact that a graph with girth $g'$ must have at most $O(n^{1+\frac{2}{g'-2}})$ edges \cite{Boll78}. Assume by contradiction that there is a cycle $C$ of length strictly less than $2+\min\{\alpha, 2k\}$. Then consider a player $u$ that owns an edge of the cycle. Since $u$ can see the cycle, she can remove the edge. The deletion would increase the distance to any other node by at most $|C|-2$ while $u$ would save $\alpha > |C|-2$, and hence $G$ cannot be an equilibrium.
\end{proof}
We can now prove the following:
\begin{theorem}
\label{thm:UB_PoA_max}
The PoA of \textsc{MaxNCG}\xspace is
$O\big(n^{\frac{2}{\min\{\alpha,2k\}}} + \frac{n}{1+\alpha} \big)$ if $\alpha \ge k-1$ and
$O\big( n^{\frac{2}{\alpha}} + \min \{ \frac{n\alpha}{k^2}, \frac{nk}{\alpha \cdot 2^{\frac{1}{4}\log^2 \frac{k}{\alpha}}} \} \big)$
if $\alpha \le k-1$.
\end{theorem}
\begin{proof}
If $\alpha \ge k-1$ then the claim follows from Lemma~\ref{lemma:density} and from the fact that the diameter of an equilibrium graph is at most $n-1$.
Otherwise, $\alpha \le k-1$ and the claim immediately follows from Lemma~\ref{lemma:density} and Lemma \ref{lemma:UB to diameter}.
\end{proof}
\subsection{Putting all together}
\label{sec:conclusions_max}
Here we summarize our lower and upper bounds to the PoA for \textsc{MaxNCG}\xspace by showing how they combine depending on the values of $\alpha$ and $k$.
First, recall that whenever the view of the players is sufficiently large, then in every LKE, players actually have a \emph{full knowledge} of the network, and so LKEs coincides with NEs (hence the PoA is the same as in the full knowledge version of the game) as shown in Corollary~\ref{cor:complete_knowledge}. The corresponding region is shown in gray in Figure~\ref{fig:results-max}.
Concerning our three lower bounds, the first one of $\Omega\big(\frac{n}{1+\alpha}\big)$ holds for $\alpha \ge k-1$, i.e. in the regions numbered \ding{173}, \ding{174}, and \ding{177} in Figure \ref{fig:results-max}.
For $1 < \alpha \le k$ and $k \le 2^{\sqrt{\log n}-3}$ we provided a strong lower bound of $\Omega( \frac{n}{\alpha 2^{\Theta(\log^2\frac{k}{\alpha})}} )$ (regions \ding{172}, \ding{175}, \ding{176} in Figure \ref{fig:results-max}). Notice that when $k=\Theta(\alpha)$ this lower bound boils down to $\Omega(\frac{n}{\alpha})$, which is tight.
Unfortunately, if $\alpha > k$, the previous lower bound is no longer valid, instead we provided a third lower bound of $\Omega( n^\frac{1}{\Theta(k)} )$ holding for $k=o(\log n)$ (regions \ding{172}, \ding{173}, \ding{174} in Figure~\ref{fig:results-max}).
\begin{figure}[!tb]
\centering
\includegraphics[scale=1]{results-max.eps}
{\tiny$\mbox{}$\\}
\scalebox{1}{
\begin{tabular}{|c|c|c|}
\hline
\# & Lower Bound & Upper Bound \\ \hline
{\large\ding{172}} & $\!\!\Omega\bigg( \max\{ \frac{n}{\alpha 2^{\Theta(\log^2 \frac{k}{\alpha})}}, n^\frac{1}{\Theta( k)} \}\bigg)$ & $O\bigg( n^{\frac{2}{\alpha}} + \min \{\frac{n\alpha}{k^2}, \frac{nk}{\alpha 2^{\Theta(\log^2 \frac{k}{\alpha})}} \} \bigg) \!\!$ \\ \hline
{\large\ding{173}} & \multicolumn{2}{c|}{ $\Theta\bigg(\max \{\frac{n}{1+\alpha},n^{\frac{1}{\Theta(k)}} \} \bigg)$ } \\ \hline
{\large\ding{174}} & \multicolumn{2}{c|}{ $\Theta\bigg(n^{\frac{1}{\Theta(k)}} \bigg)$ } \\ \hline
{\large\ding{175}} & $\Omega\bigg(\frac{n}{\alpha 2^{\Theta(\log^2 \frac{k}{\alpha})}}\bigg)$ & $O\bigg( n^{\frac{2}{\alpha}} + \min \{\frac{n\alpha}{k^2}, \frac{nk}{\alpha 2^{\Theta(\log^2 \frac{k}{\alpha})}} \} \bigg) \!\!$ \\ \hline
{\large\ding{176}} & $\Omega\bigg(\frac{n}{\alpha 2^{\Theta(\log^2 \frac{k}{\alpha})}}\bigg)$ & $O\bigg( \min \{\frac{n\alpha}{k^2}, \frac{nk}{\alpha 2^{\Theta(\log^2 \frac{k}{\alpha})}} \} \bigg)$ \\ \hline
{\large\ding{177}} & \multicolumn{2}{c|}{ $\Theta\bigg( \frac{n}{1+\alpha} \bigg)$ } \\ \hline
{\large\ding{178}} & \multicolumn{2}{c|}{ $O\bigg( n^{\frac{2}{\alpha}} + \frac{nk}{\alpha 2^{\Theta(\log^2 \frac{k}{\alpha})}} \bigg)$ } \\ \hline
{\large\ding{179}} & \multicolumn{2}{c|}{ $O\bigg( \min\{ \frac{n\alpha}{k^2}, \frac{nk}{\alpha 2^{\Theta(\log^2 \frac{k}{\alpha})}} \} \bigg)$ } \\ \hline
\end{tabular}}
\caption{Lower and upper bounds to the PoA for \textsc{MaxNCG}\xspace. The partition in regions comes up by the combination of the various bounds that we give in Section \ref{sect:PoA}. Notice that the gray region is the set of pairs $(\alpha, k)$ such that, in every LKE, players actually have a full knowledge of the network, and so LKEs coincides with NEs.}
\label{fig:results-max}
\end{figure}
Turning to the upper bounds to the PoA, in Theorem~\ref{thm:UB_PoA_max}, we proved them by considering both the density and the diameter of an equilibrium graph. We showed that the number of edges can be at most $O(n^{1+\frac{2}{\min\{\alpha,2k\}}})$. Regarding the diameter, since for $\alpha \ge k-1$ it can be shown to be $\Omega(n)$, we considered the case $\alpha \le k-1$ where we gave an upper bound of $O(\min\{ \frac{n \alpha^2}{k^2}, \frac{nk}{2^{\Theta(\log^2 \frac{k}{\alpha})}}) \}$.
Notice that this upper bound is a minimum of two terms. Intuitively, the first one is better when $k$ is not too big w.r.t.\ $\alpha$, e.g.\ when $k=O(\alpha \operatorname{polylog}(\alpha))$. The corresponding region lies between the dashed gray curve and the line of equation $\alpha=k-1$ shown in Figure~\ref{fig:results-max}.
The bounds to the PoA that arise from the various combinations of these results are summarized in Figure~\ref{fig:results-max}. Notice that the bounds for the regions under the line $k=\alpha+1$ are essentially tight.
\section{Results for \protect\textsc{SumNCG}\xspace}
\label{sec:sum}
In this section we provide our results for \textsc{SumNCG}\xspace. Recall that in the full-knowledge version of the game the spanning star is the social optimum and has a cost of $\Theta(\alpha n + n^2)$. We start with a quite strong lower bound to the PoA.
Consider a graph similar to the one shown in Figure \ref{fig:torus}, whose construction has been described in Section \ref{sec:lower_bounds_max_m3}. In particular, we build such a graph using the following parameters: $d=2$, $\ell=2$, $\delta_1= \left\lceil \frac{k}{2} \right\rceil +1$ and $\delta_2 \ge \delta_1$ which will be specified later.
We know that $N=2 \delta_1 \delta_2$ and that $n=6 \delta_1 \delta_2$.
We want $\delta_2 \ge \delta_1 \iff \frac{n}{6\delta_1} \ge \delta_1 \iff \delta_1 \le \sqrt{\frac{n}{6}} \Longleftarrow k \le \sqrt{\frac{2n}{3}} - 4$.
We have the following:
\begin{lemma}\label{lemma:LB_sum_m3}
For $\alpha \ge 4k^3$, the previous graph is an equilibrium.
\end{lemma}
\begin{proof}
First notice that each vertex $v$ sees at most $4k^2$ other vertices. As $\alpha \ge 4k^3$ and $v$ can reduce the distance to each vertex by at most $(k-1)$, $v$ cannot increase the number of bought edges.
Suppose that the player $v$ changes her strategy from $\sigma_v$ to $\sigma^\prime_v$, let $L$ be the set of vertices that are at distance $k$ from $v$. Consider the view of $v$ before the strategy changes, where the edges of $\sigma_v$ have been removed and replaced with the edges of $\sigma^\prime_v$. Each vertex of $L$ must be at distance at most $k$ from $v$ in this new graph.
Otherwise, suppose the existence of a vertex $x \in L$ that is at distance at least $k+1$ in this new graph. When $v$ computes $\Delta(\sigma_v, \sigma^\prime_v)$ it will also consider the case where a certain number of (new) vertices $\eta$ are adjacent to $x$, therefore her usage will be at least $k \eta$ which, for a suitable value of $\eta$, is greater than the cost of $v$ in $\sigma$.
Let $v=(k^*+1, k^*+1)$, and consider the set of vertices $L= \left( \mathcal{F}^{k-1}(u) \cup \mathcal{F}^{k-1}(u^\prime) \right) \setminus \{ (k^*+3-k, k^*+3-k), (k^*+k-1,k^*+k-1) \}$.
It is easy to see that all the vertices in $L$ are at distance $k$ from $v$, and that
every vertex $x$ in the view of $v$, that is not a neighbor of $v$, has at most $2$ vertices of $L$ that are at distance at most $k-1$. Finally, every vertex $x$ in the view of $v$ has at least one vertex of $L$ at a distance at least $k$.
This suffices to conclude that every vertex $v$ is currently playing a best response and, therefore, the graph is in equilibrium.
\end{proof}
Using the above lemma we can prove the following:
\begin{theorem}
Let $\alpha \ge 4 k^3$. For any $k \le \sqrt{\frac{2n}{3}} - 4$, the PoA of \textsc{SumNCG}\xspace is $\Omega(n/k)$ if $\alpha \le n$, and $\Omega(1+\frac{n^2}{k \alpha})$ otherwise.
\label{th:LB_PoA_sum_m3}
\end{theorem}
\begin{proof}
By the above lemma, the graph is in equilibrium and has diameter $\Omega(\delta_2)=\Omega(\frac{n}{k})$. Moreover, it is easy to see that each vertex has $\Omega(n)$ vertices at distance $\Omega(\frac{n}{k})$. Since the graph has $\Theta(n)$ edges, we have that the cost of the graph is $\Omega(\alpha n +n^2/k)$ while the social optimum is a star with cost $O(\alpha n+n^2)$. The claim follows.
\end{proof}
The following theorem provides a lower bound for a different range of values of the parameters $\alpha$ and $k$.
\begin{theorem}
\label{th:LBbis_PoA_sum_m3}
If $\alpha \ge kn$ and $k \ge 2$, the PoA of \textsc{SumNCG}\xspace is $\Omega(n^{\frac{1}{2k-2}})$.
\end{theorem}
\begin{proof}
We use the same construction of Lemma \ref{lemma:lb_poa_max_dense}.
Remind that the view of each vertex $v$ is a tree of height $k$ with $q(q-1)^{i-1}$ vertices on level $i$. Therefore $v$ has to buy at least $q$ edges. Moreover, if she buys exactly $q$ edges, then she cannot improve her cost as her neighbors are the medians of the corresponding subtrees.
As $\alpha \ge kn$, $v$ cannot improve her cost by increasing the number of bought edges. The claim follows.
\end{proof}
Finally, we have:
\begin{theorem}
If $k > 1 + 2\sqrt{\alpha}$, then in every equilibrium graph each player sees the whole graph, thus the set of LKEs coincides with the set of NEs.
\end{theorem}
\begin{proof}
Let $G$ be an equilibrium with diameter at least $k$ (otherwise the claim is trivially true), and let $u$ and $v$ be two vertices such that $d_G(u,v)=k$.
By buying the edge $(u,v)$, the player $u$ could decrease the cost needed to reach the last $\left\lfloor \frac{k}{2} \right\rfloor$ vertices along the shortest path between $u$ and $v$ by at least $\frac{k^2}{4}-\frac{2}{4}$.
As $G$ is an equilibrium we must have $\frac{k^2}{4}-\frac{2}{4} \le \alpha$, which implies $k \le 2\sqrt{\alpha} + 1$.
\end{proof}
The previous results are summarized in Figure~\ref{fig:results-sum}. Notice that it remains open to establish a lower bound to the PoA for values of $k$ between $\Theta(\sqrt[3]{\alpha})$ and $\Theta(\sqrt{\alpha})$. Moreover, we plan in the future to study the corresponding upper bounds.
\begin{figure}[!tb]
\centering
\includegraphics[scale=1]{results-sum.eps}
\caption{Lower bounds to the PoA for \textsc{SumNCG}\xspace. The behavior of the PoA in the region between the two curves is open. Notice that the gray region is the set of pairs $(\alpha, k)$ such that, in every LKE, players actually have a full knowledge of the network, and so LKEs coincide with NEs.}
\label{fig:results-sum}
\end{figure}
\section{Experimental evaluation}
\label{sec:experiments}
In this section, we present an experimental study whose aim is to complement our theoretical analysis by providing insights on several features of the game.
Indeed, our asymptotic bounds depends on heterogeneous parameters, and so an empirical evidence of their tightness is required. In particular, our study concerns the behavior of best-response dynamics in \textsc{MaxNCG}\xspace, by focusing on the convergence time (i.e., the number of steps needed to form a stable network) and on the structural properties of the resulting equilibria. This way, on one hand our goal is to check whether the theoretical existence of better- and best-responce cycles is actually frequent, and on the other hand we essentially aim to verify whether stable networks have a social cost which follows the trend suggested by the analysis of the PoA we have provided.
As we will show, the answer to the first question is in the negative, while as far as the second aspect is concerned, we will provide an evidence that our analysis is actually quite accurate (at least when $\alpha$ is not too small).
\subsection{Experiments}
In our experiments, we simulated the best-response dynamics of the players for the \textsc{MaxNCG}\xspace model.
The initial configuration consists of a network randomly sampled from a certain class of graphs, as we will describe in the following.
The players play in turns, following a round-robin policy, i.e., in each round we consider all the players, one at time. When a player is considered, we compute a best-response strategy according to her local knowledge of the network, and whenever this strategy is strictly better than the current one we update the network. Then, we move to the next player. We continue this process until we attain an equilibrium, i.e., we find a round where no player is able to improve her cost.
Moreover, since the convergence is not guaranteed, whenever the execution time exceeds a given threshold, we check if the last strategy profile of the current round already appeared as the last strategy profile of any previous round. In this case, since of the round-robin policy, we can conclude that the best-response dynamics admits a cycle, hence we stop the simulation as we know that no equilibrium will ever be reached by the players.
For each starting network, the above experiment is repeated with different values of the parameters $\alpha$ and $k$ taken from the set $\{ 0.025, 0.05, 0.1, 0.2, 0.3, 0.4, 0.5, 0.7, 1, 1.5, 2, 3, 5, 7, 10 \}$ and from the set $\{ 2, 3, 4, 5, 6, 7, 10, 15, 20, 25, 30, 1000 \}$, respectively. The case $k=1000$ corresponds to the classical version of the game where every player has no restriction on her view.
In order to compensate for the variability in measurements, for each pair of the parameters $\alpha$ and $k$ we ran $20$ distinct experiments starting from different random graphs belonging to the same class. Overall, we simulated about $36\,000$ different dynamics and, after each round, we collected several different features of the current network such as: diameter, social cost, maximum/average degree, minimum/maximum/average number of bought edges, minimum/maximum/average number of vertices in the view of the players, along with others. Moreover, for each run we also collected global statistics as the number of rounds needed to reach an equilibrium (if any), and the total number of strategy changes performed by the players.
In Figures~\ref{fig:viewsize}-\ref{fig:convergence}, we will show how the mean values of those features (taken over the $20$ different starting networks) vary as a function of either $\alpha$, $k$ or $n$. We also report the corresponding $95\%$ confidence intervals, as long as they do not affect readability.
\subsection{Input classes of graphs}
We considered the following classes of graphs:
\begin{itemize}
\item \emph{Random trees}: for a given number $n$ of vertices, we picked a tree uniformly at random from the set of all possible trees on $n$ vertices. We determined the ownership of each edge $(u,v)$ of the tree by choosing between $u$ and $v$ with a fair coin toss. The above was repeated for all the values of $n$ in $\{ 20, 30, 50, 70, 100, 200 \}$. In Table~\ref{table:tree}, we report some statistics of the generated trees for the different values of $n$.
\begin{table}
\centering
\tbl{Statistics for the random trees used in the experimental evaluation.
In each row, $20$ random trees with the same number $n$ of nodes are considered. The value of $n$ is reported in the first column. The remaining columns contain the average statistics over the corresponding trees along with their $95\%$ confidence intervals.
}{
\begin{tabular}{|d{0}|d{2}|d{2}|d{2}|}
\hline
n & \multicolumn{1}{c|}{\mbox{Diameter}} & \multicolumn{1}{c|}{\mbox{Max. degree}} & \multicolumn{1}{c|}{\mbox{Max. Bought Edges}} \\ \hline
20 & 10.65 \pm 0,76 & 4.00 \pm 0,26 & 2.75 \pm 0,34 \\ \hline
30 & 13.90 \pm 1,18 & 4.30 \pm 0,31 & 3.15 \pm 0,31 \\ \hline
50 & 19.55 \pm 1,48 & 4.60 \pm 0,35 & 3.30 \pm 0,31 \\ \hline
70 & 22.10 \pm 1,57 & 5.05 \pm 0,39 & 3.50 \pm 0,32 \\ \hline
100 & 25.15 \pm 1,95 & 5.35 \pm 0,35 & 3.45 \pm 0,28 \\ \hline
200 & 43.20 \pm 3,95 & 5.30 \pm 0,31 & 3.85 \pm 0,31 \\ \hline
\end{tabular}}
\label{table:tree}
\end{table}
\item \emph{Erd\H{o}s-Rényi random graphs}: we generated random graphs according to the classical $G(n,p)$ model \cite{ER59} in which a graph of $n$ vertices is built by adding independently each (undirected) edge with a probability of $p$. The parameters $n$ and $p$ were chosen so that the resulting graph was likely to be connected. Any remaining unconnected graph was discarded and regenerated from scratch. We generated graphs with both $100$ and $200$ vertices and, for each $n$, we set the values of $p$ in order to produce three graphs with different densities. As for trees, the owner of each edge was chosen uniformly at random between its endpoints. A summary of the resulting graphs can be found in Table~\ref{table:erdos}.
\begin{table}
\centering
\tbl{Statistics for the different Erd\H{o}s-Rényi random graphs used in the experimental evaluation.
In each row, a different set of the parameters $n$ and $p$ is considered (first two columns). For each of them, $20$ different graphs have been randomly sampled. The remaining columns report the average statistics over these graphs along with their $95\%$ confidence intervals.}{
\begin{tabular}{|d{0}|d{3}|d{2}|d{2}|d{2}|d{2}|}
\hline
n & p & \multicolumn{1}{c|}{\mbox{Edges}} & \multicolumn{1}{c|}{\mbox{Diameter}} & \multicolumn{1}{c|}{\mbox{Max. degree}} & \multicolumn{1}{c|}{\mbox{Max. Bought Edges}} \\ \hline
100 & 0,060 & 301.10 \pm \ \ 7,51 & 5.30 \pm 0,22 & 12.50 \pm 0,67 & 7.90 \pm 0,43 \\ \hline
100 & 0,100 & 494.40 \pm 10,86 & 4.00 \pm 0,00 & 18.45 \pm 0,82 & 11.40 \pm 0,69 \\ \hline
100 & 0,200 & 984.35 \pm 12,23 & 3.00 \pm 0,00 & 29.90 \pm 1,26 & 18.25 \pm 0,57 \\ \hline
200 & 0,035 & 698.25 \pm 11,22 & 5.25 \pm 0,21 & 15.35 \pm 0,61 & 9.20 \pm 0,42 \\ \hline
200 & 0,050 & 992.45 \pm 10,18 & 4.20 \pm 0,19 & 19.30 \pm 0,55 & 12.50 \pm 0,65 \\ \hline
200 & 0,100 & 2005.55 \pm 12,87 & 3.00 \pm 0,00 & 32.80 \pm 1,11 & 18.95 \pm 0,54 \\ \hline
\end{tabular}}
\label{table:erdos}
\end{table}
\end{itemize}
\subsection{Computing a best-response strategy}
As we pointed out in the theoretical part, the problem of computing a best response for a player is \mbox{\sf NP}-hard. In order to deal with nontrivial values of $n$, we reduced the problem of computing a best response to the problem of computing (a variant of) the minimum dominating set of a suitable graph, and then we used the Gurobi solver \cite{gurobi} on classical integer linear programming formulation of minimum dominating set.
We now sketch the idea behind our approach.
Let $u$ be the player for which we want to compute a best-response strategy. First of all, from Proposition~\ref{prop:maxncg_best_response} we can just compute the best move of $u$ with respect to her view $H$. Now, we remove $u$ from the graph $H$, and we guess the eccentricity $h$ of $u$ in the graph resulting from a best-response move. If $H'$ denotes the $(h-1)$-th power\footnote{The $h$-th power of an undirected graph $G$ is a graph on the same vertex-set of $G$ where the edge $(u,v)$ exists iff the distance between $u$ and $v$ in $G$ is at most $h$.} of the graph $H \setminus \{u\}$, then it is easy to see that the original problem is equivalent to that of finding a minimum dominating set of $H'$ in which certain vertices are constrained to be included in the solution, namely those who bought an edge towards $u$ in the current strategy profile.
\subsection{Experimental results}
Here we discuss some interesting features of the experimental results concerning \textsc{MaxNCG}\xspace.
\paragraph{Knowledge of the network}
We start by arguing on how the size of the network known by players at equilibrium varies as a function of $k$.
This is both interesting by itself, and it will also be instrumental for further discussions regarding the quality of equilibria.
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{plots/tree100-meanviewsize.eps}
\includegraphics[width=0.5\textwidth]{plots/tree100-minviewsize.eps}
\caption{Minimum and average number of vertices in the players' view on stable networks as a function of $\alpha$ for the various values of $k$. Points correspond to mean values over $20$ different trees with $100$ vertices, while error bars represent $95\%$ confidence intervals. For the sake of readability, lines that coincide with the one for $k=1000$ are not depicted.}
\label{fig:viewsize}
\end{figure}
Let us consider Figure~\ref{fig:viewsize}, where both the mean and the minimum number of vertices in the players' view is reported. The actual value of each point is the average over the different equilibria obtained from $20$ trees with $100$ vertices each (using the same set of parameters).
The same measurements on the other classes of graphs exhibit essentially the same behavior and are not reported.
Clearly, the view of a player decreases as $\alpha$ increases, and rapidly grows as $k$ becomes larger. Intuitively, the former is due to the fact that whenever $\alpha$ is small, players are more prone to buy a large number of edges. Regarding the latter, we observe that for the interesting threshold value of $k=7$, we have that, on average, a player knows more than $99$ vertices. Moreover, the minimum size is also pretty high, since even the player who knows the least portion of the network is able to see more than $93$ vertices (on average).
\paragraph{Quality of equilibria}
We now discuss how our theoretical bounds on the PoA compare to the experimental evaluation of the quality of equilibria, i.e., the ratio between the social cost of the attained equilibrium and the social optimum.
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{plots/tree100-poa-1.eps}
\includegraphics[width=0.5\textwidth]{plots/tree100-poa-10.eps}
\caption{Quality of the stable networks as a function of $n$ for various values of $k$ on random trees. The left picture refers to $\alpha=1$ while the right one refers to $\alpha=10$. Points correspond to mean values over $20$ different trees, while error bars represent $95\%$ confidence intervals. For the sake of readability, lines that essentially coincide with the one for $k=1000$ are not depicted. Notice that for small values of $k$ the quality degrades linearly while for larger values of $k$ the PoA is almost constant.}
\label{fig:poa}
\end{figure}
First of all, notice that our bounds show that the PoA decreases as $\alpha$ and/or $k$ increase. Moreover, for fixed values of the parameters, the PoA is $\Theta(n)$ whenever $\alpha \ge 2$.
This trend can be easily recognized in Figure~\ref{fig:poa} (b), where the quality of the equilibria on experimented trees is plotted against the number of vertices for $\alpha=10$. Notice that as soon as $k$ exceeds $6$, the quality of equilibria drastically improves. This is not surprising since, as already discussed, the players have a (almost) full-knowledge of the stable networks, and hence the (almost) constant bounds to the PoA given for the classical version of \textsc{MaxNCG}\xspace hold.
Let us now focus on smaller values of $\alpha$, for example $\alpha=1$ as shown in Figure~\ref{fig:poa} (a). The general trend of the quality of equilibria is similar, except for the fact that $k \ge 5$ is now sufficient to yield full-knowledge equilibria. By contrast, our theoretical upper bound for smaller values of $k$ of
$O\bigg( n^{\frac{2}{\alpha}} + \min \{\frac{n\alpha}{k^2}, \frac{nk}{\alpha 2^{\Theta(\log^2 \frac{k}{\alpha})}} \} \bigg)$ is now $O(n^{\frac{2}{\alpha}} + n) = O(n^2)$.
This comes from the first term of the sum, which is related to the density of the equilibria. In our experiments, however, the resulting stable networks happened to be sparse, so that the usage cost prevails instead. This suggests that our density bounds might be too coarse, and this is somewhat supported by the corresponding lower bound of $\Omega\bigg( \max\{ \frac{n}{\alpha 2^{\Theta(\log^2 \frac{k}{\alpha})}}, n^\frac{1}{2k-2} \}\bigg)$, which now simplifies to $\Omega(n)$ whenever the density term is dominated by the usage cost term.
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{plots/tree100-poa-k-2.eps}
\includegraphics[width=0.5\textwidth]{plots/er-02-poa.eps}
\caption{Quality of the stable networks as a function of $k$ for various values of $n$, and $\alpha=2$. The bold red line represents the trend of our theoretical upper bound. The left picture refers to random trees while the right one refers to random graphs with $n=100$ and $p=0.2$. Points correspond to mean values over $20$ different graphs, while error bars represent $95\%$ confidence intervals.}
\label{fig:poa-k}
\end{figure}
We now examine in more detail how the quality of equilibria decreases as a function of $k$. Notice that, our theoretical upper bound to the PoA of $O\bigg( n^{\frac{2}{\alpha}} + \min \{\frac{n\alpha}{k^2}, \frac{nk}{\alpha 2^{\Theta(\log^2 \frac{k}{\alpha})}} \} \bigg)$, reduces to $f(k)=O(\frac{k}{2^{\log^2 k}})$, once $\alpha \ge 2$ and $n$ are fixed to be constant. In Figure~\ref{fig:poa-k} we report the measured quality of equilibria as a function of $k$ on both random trees (left) and graphs (right), for different values of $n$, and $\alpha=2$. Moreover, the trend of $f(k)$ is shown in red as a benchmark. For different values of $\alpha$ the trend is very similar, but the actual quality of equilibria scales down as expected.
\paragraph{Fairness of equilibria}
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{plots/maxdeg-er-100.eps}
\includegraphics[width=0.5\textwidth]{plots/maxowneddeg-er-100.eps}
\caption{Maximum degree (left) and maximum number of bought edges (right) as a function of $\alpha$ for various values of $k$. Points correspond to mean values over $20$ different random graphs with $100$ vertices and $p=0.1$. For the sake of readability, lines that essentially coincide with the one for $k=1000$ are not depicted.}
\label{fig:maxdeg}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.5\textwidth]{plots/fairness-er-100.eps}
\caption{Unfairness ratio, i.e., the ratio between the highest and the lowest of players' costs, as a function of $\alpha$ for various values of $k$. Points correspond to mean values over $20$ different random graphs with $100$ vertices and $p=0.1$. For the sake of readability, lines that coincide with the one for $k=1000$ are not depicted. Notice small values of $k$ yield more fair equilibria. }
\label{fig:fairness}
\end{figure}
We also collected statistics about the maximum degree of stable networks and the maximum number of edges bought by the players (see Figure~\ref{fig:maxdeg}, concerned with random graphs of $100$ nodes and with $p=0.1$). It is worth to notice that for $k \ge 4$ and small values of $\alpha$, we have that the maximum degree is more than $80$, while a player does not buy more than $9$ edges (for every values of $\alpha$ and $k$). This has a nice consequence in terms of fairness of stable networks, as shown in Figure~\ref{fig:fairness}, which decreases as $k$ increases. This would suggest that restricting the view of the players could help to converge towards stable networks where players' costs do not differ too much.
\paragraph{Convergence time}
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{plots/convergence-tr-100.eps}
\includegraphics[width=0.5\textwidth]{plots/convergence-verts.eps}
\caption{Number of rounds needed to converge to a stable network as a function of $\alpha$ with $n=100$ (left) and as a function of $n$ with $\alpha=2$ (right) for various values of $k$. Points correspond to mean values over $20$ different random trees.}
\label{fig:convergence}
\end{figure}
To conclude this section let us focus on the convergence time, i.e., the number of rounds needed to actually reach an equilibrium (if any). As we pointed out in the technical part, the convergence of a best-response dynamics is not guaranteed. In practice, however, it seems very common: we simulated about $36\,000$ best-response dynamics, and only encountered best-response cycles in $5$ of them.
In all the other cases convergence appeared to be pretty fast, as shown in Figure~\ref{fig:convergence} for random trees. In fact, starting from both random trees and Erd\H{o}s-Rényi graphs, for almost every combination of $\alpha$ and $k$, in more than $95\%$ of the times, at most $7$ rounds are enough to converge to a stable network. Finally, the total number of rounds increases with $n$ as one would expect, although this happens quite slowly for almost every choice of the parameters.
\section{Conclusions}
\label{sec:conclusions}
In this paper we have studied the game-theoretic and computational implications of a limited players' view (more precisely, confined to a prescribed distance radius from each player) in the autonomous formation of a (large) network. In this scenario, we developed a systematic analysis on the PoA of the two classic variants of the game, namely \textsc{MaxNCG} and \textsc{SumNCG}, along with an extensive experimental study, which, for the computational feasibility reasons that have been explained in the paper, were limited to the former game, though.
Concerning our future research, on one hand we plan to investigate some of the issues that we left (partially) open, in particular the PoA space for the \textsc{SumNCG}\xspace needs to be explored in more detail. On the other hand, we believe that our incomplete-knowledge approach deserves to be extended in several directions. As a first step in this regard, in \cite{BGLP14} we have considered three local-knowledge models usually adopted in {\sc Network Discovery}, i.e., the optimization problem of reconstructing the topology of an unknown network through a minimum number of queries at its nodes. For these models, we provided exhaustive answers to the canonical algorithmic game theoretic questions w.r.t. our LKE concept.
Besides studying new models, however, we feel that our own model has still few intriguing issues that should be investigated. Above all, it would be interesting to relax our worst-case approach, and analyze a NCG under a Bayesian perspective. Finally, we feel that the local-knowledge model could be extended to other (network) game-theoretic settings, given that the global knowledge is a recurring, still critical, assumption.
\section*{Acknowledgments}
We wish to thank Stefano Smriglio for useful discussions concerning the experimental section of the paper.
\bibliographystyle{ACM-Reference-Format-Journals}
|
\section{Introduction}
One of the most fundamental discoveries with regard to the cosmic evolution of galaxies has been the determination that a substantial fraction of the integrated Extragalactic Background Light (EBL) arises at infrared-to-millimeter wavelengths: the Cosmic Infrared Background (CIB). Quantitative observations of the CIB began with the Cosmic Background Explorer (COBE). At a low angular resolution ($0.7\deg$), COBE provided the first large-scale measurement of the spectral energy distribution (SED) of the EBL from the far-infrared to the (sub)millimeter \citep{puget96,fixsen98}. The CIB consists of the combined flux of all extragalactic sources, and contains much information about the history and formation of galaxies, and of the large scale structure of the Universe.
The observation that the cosmic density of star-formation was an order of magnitude higher at cosmological redshifts, $z\sim2-4$ \citep[e.g.,][]{madau96, lilly96}, opened the possibility that most of the CIB arose from dust re-processed UV-light from distant galaxies. These studies used the Lyman dropout technique to identify normal galaxies at high-redshift, being mostly insensitive to dust obscured star formation. Later, sensitive maps obtained with submillimeter/millimeter bolometer arrays were thus able to directly detect and identify luminous dusty star forming galaxies (DSFGs), which were soon found to contribute a fraction to the EBL at these wavelengths \citep[e.g.,][]{smail97}.
Since then, a number of groups have conducted (sub)millimeter surveys of the sky, currently yielding up to hundreds of sources in contiguous areas of the sky \citep[e.g.,][]{hughes98, barger98, eales00, bertoldi00, scott02, cowie02, voss06, bertoldi07,scott08,greve08, weiss09,austermann10,vieira10,aretxaga11,hatsukade11,scott12,mocanu13}. These blank field bolometer (sub)millimeter surveys discovered a population of luminous DSFGs at high redshift that were not accounted for in optical studies. These galaxies -- also called ``submillimeter galaxies'' (SMGs) due to the region of the electromagnetic spectrum in which they were first discovered -- have been characterised as massive starburst galaxies with typical stellar and molecular gas masses of $\sim10^{11}\ M_\sun$, typically located at $z=1-3$ \citep[e.g.,][]{chapman05} with a tail out to $z\sim6$ \citep{weiss13, riechers13}, and most likely driven by relatively bright mergers \citep{engel10}. As such, these galaxies are found to be gas/dust rich, with gas fractions typically exceeding 0.2 \citep[e.g.][]{daddi10a, tacconi10, magdis12, tacconi13,bothwell13}. Despite their large SFRs implied by the large IR luminosities ($>10^{12.0-12.5}\ L_\sun$) and significant abundance at high-redshift, these galaxies (e.g. $S_{\rm 1.2mm}>2-3$ mJy) were found to contribute only a minor fraction of the EBL at submillimeter wavelengths \citep{barger99,eales99,smail02,coppin06,knudsen08,weiss09,scott12, chen13}. Hence, questions about the properties of the population of galaxies that dominate this EBL remain.
\begin{figure*}[ht]
\centering
\includegraphics[scale=0.55]{UDF_1mm_continuum_image_v4.ps}\hspace{4mm}
\includegraphics[scale=0.55]{UDF_1mm_sensitivity_image_v4.ps}
\caption{({\it Left:}) ALMA 1.2-mm signal-to-noise continuum mosaic map obtained in the HUDF. Black and white contours show positive and negative emission, respectively. Contours are shown at $\pm2,3,4,5,8,12,20$ and $40\sigma$, with $\sigma=12.7\mu$Jy beam$^{-1}$ at the field center. The boxes show the position of the sources detected with our extraction procedure at $S/N>3.5$. The synthesized beam ($1''\times2''$) is shown in the lower left. ({\it Right:}) ALMA 1.2-mm observations primary beam (PB) pattern to represent the sensitivity obtained across the covered HUDF region. PB levels are shown by the black/white contours at levels 0.3, 0.5, 0.7 and 0.9 of the maximum. Both the signal-to-noise and PB maps are shown down to PB$=0.2$.\label{fig_map1mm}}
\end{figure*}
\begin{figure*}[ht]
\centering
\includegraphics[scale=0.55]{UDF_3mm_continuum_image_v4.ps}\hspace{4mm}
\includegraphics[scale=0.55]{UDF_3mm_sensitivity_image_v4.ps}
\caption{({\it Left:}) ALMA 3-mm signal-to-noise continuum mosaic map obtained in the HUDF. Black and white contours show the positive and negative signal, respectively. Contours are shown at $\pm2,3,4,5,8,12,20$ and $40\sigma$, with $\sigma=3.8\mu$Jy beam$^{-1}$ at the field center. The boxes show the position of the sources detected in the 1.2-mm map, with our extraction procedure at $S/N>3.5$. The synthesized beam ($2''\times3''$) is shown in the lower left. ({\it Right:}) ALMA 3-mm observations primary beam (PB) pattern. PB levels are shown by the black/white contours at levels 0.3, 0.5, 0.7 and 0.9. Both the signal-to-noise and PB maps are shown down to PB$=0.2$.\label{fig_map3mm}}
\end{figure*}
To locate and characterise the population of faint DSFGs that make up most of the EBL at (sub)millimeter wavelengths, we must overcome several observational limitations. First, the poor resolution of (sub)millimeter bolometer maps taken with single-dish telescopes, typically with beam sizes between $10-30''$, makes the identification of an optical counterpart difficult and thus limits the characterisation of submillimeter sources. In addition, this affects the number counts, since the brightest sources are seen to split into multiple components in high-resolution (sub)millimeter images \citep{younger07, wang11, smolcic12, hodge13, karim13, miettinen15}. Secondly, the sensitivity of single dish bolometer maps, typically down to $0.5-1.0$ mJy, along with confusion at the faint levels limits our view to the most luminous sources. An important approach to reach fainter galaxies has been the use of gravitational lensing enabled by massive galaxy clusters \citep[e.g.,][]{smail97, smail02, sheth04,knudsen08,noble12,johansson12, chen13}. However, these surveys suffer severely from cosmic variance, due to the small areas covered in the source plane, source confusion, and the need for accurate lens models and magnification maps. A parallel approach has been to perform stacking of the submillimeter emission using pre-selected samples of optical/infrared galaxies. This approach has successfully resolved significant amounts of the EBL at (sub)millimeter wavelengths, reaching down to sources with $S_{\rm 1.2mm}>0.1$ mJy \citep{webb04,knudsen05,greve10,decarli14}. The major limitation of this approach is that it yields average properties over a population of galaxies that must be assumed to have similar (sub)millimeter properties.
The advent of the Atacama Millimeter/submillimeter Array (ALMA) is opening up a new window for the study of the faint DSFG population. Its significantly higher angular resolution compared to single-dish telescopes ($<3''$), and the unparalleled sensitivity allow us to reach flux density levels in (sub)millimeter continuum maps even deeper than those achieved by studies of galaxy cluster fields or based on stacking analysis. Several recent studies have individually pinpointed (sub)millimeter sources down to 0.1 mJy in the 1-mm band \citep{hatsukade13,ono14,carniani15,oteo15,hatsukade16, dunlop16}. Some of these surveys have used clever approaches by taking advantage of archival data \citep{ono14,carniani15,fujimoto16}, including ALMA calibration fields \citep{oteo15}. Recently, \citet{fujimoto16} were able to reach down to a flux limit of 15$\mu$Jy at 1.2-mm, providing the deepest measurements of the number counts to date, and allowing them to resolve most of the CIB into individual sources. Despite the substantial progress, the current studies are still affected significantly by cosmic variance and are not ``blank-field'' in nature (as some of them target overdense fields). Most importantly, the lack of sufficiently deep complementary data have only permitted the characterisation of a handful of sources \citep{hatsukade15,fujimoto16, yamaguchi16}.
Using ALMA in Cycle 2, we have conducted a deep ALMA Spectroscopic Survey (ASPECS) of a region of the {\it Hubble} Ultra Deep Field (UDF), covering the full 3-mm and 1-mm bands. In this paper, we present the {\it deepest} millimeter continuum images obtained to date in a contiguous 1 arcmin$^2$ area. This is the Paper~II in the ASPECS series. A full description of the survey and spectral line search is presented in Paper~I \citep{walter16}. Measurements of the CO luminosity function and cosmic density of molecular gas are shown in Paper~III \citep{decarli16a}. A detailed analysis of the CO brightest objects is presented in Paper~IV \citep{decarli16b}. A search for {\sc [CII]} line emission is shown in Paper~V \citep{aravena16b}. This paper is organised as follows: in \S \par\noindent\hangindent 20 pt{sec_obs}, we summarise the ALMA observations and multi-wavelength ancillary data available. Here, we also present the obtained ALMA continuum maps at 1.2-mm and 3-mm. In \S \par\noindent\hangindent 20 pt{sec_results}, we present the detected sources and compute the fidelity and completeness of our extraction procedures in the 1.2-mm map. In \S \par\noindent\hangindent 20 pt{sec_counts}, we derive the number counts at 1.2-mm. In \S \par\noindent\hangindent 20 pt{sec_prop}, we characterise the multi-wavelength properties of the individually detected sources, including their typical stellar masses, SFRs and redshifts, and discuss whether our sources are starbursts or more quiescent star forming galaxies. In \S \par\noindent\hangindent 20 pt{sec_stack}, we conduct a stacking analysis to determine the average properties of the faintest population of galaxies, not detected individually by our survey. In \S \par\noindent\hangindent 20 pt{sec_ism}, we investigate the ISM properties of the individually detected sources based on measurements of the ISM masses from the 1.2-mm fluxes. We estimate their gas masses, depletion timescales and fractions. In \S \par\noindent\hangindent 20 pt{sec_cib}, we determine the contribution of both our individually-detected and stacked sample to measure the fraction of the EBL at 1.2-mm resolved by our observations. We discuss the properties of the galaxies that dominate the CIB. Finally, in \S \par\noindent\hangindent 20 pt{sec_concl}, we summarise the main results of this paper. Throughout the paper, we assume a standard $\Lambda$CDM cosmology with $H_0=70$ km s$^{-1}$ Mpc$^{-1}$, $\Omega_\Lambda=0.7$ and $\Omega_{\rm M}=0.3$.
\section{Observations}
\label{sec_obs}
\begin{table*}[t]
\begin{flushleft}
\caption{Sources detected in the ASPECS 1.2-mm continuum map. Columns: (1), (2) Source full and short names; (3), (4) Position of the 1.2-mm continuum detection in the ALMA 1.2-mm map; (5) Signal to noise ratio (SNR) of the 1.2-mm detection; (6) Flux density at 1.2-mm, corrected for PB; (7) Primary beam correction at the location of the detection in the 1.2-mm mosaic; (8) Flux density at 3.0-mm of the ALMA 1.2-mm continuum detection. Upper limits are given at the $3\sigma$ level; (9) Primary beam correction at the location of the 1.2-mm detection in the 3.0-mm map; (10) Is there an optical counterpart identification for this source? Yes or no; }\label{tab_sources}
\end{flushleft}
\begin{tabular}{lccccccccc}
\hline
IAU name & Short name & RA$_{\rm 1.2mm}$ & Dec$_{\rm 1.2mm}$ & SNR & $S_{\rm 1.2mm}$ & PB$_{\rm 1.2mm}$ & $S_{\rm 3mm}$ & PB$_{\rm 3mm}$ & OID?\\
ALMA\ldots & ASPECS\ldots & (J2000) & (J2000) & & ($\mu$Jy) & & ($\mu$Jy) & \\
(1) & (2) & (3) & (4) & (5) & (6) & (7) & (8) & (9) & (10) \\
\hline\hline
\multicolumn{10}{c}{Main sample at $>3.5\sigma$ significance}\\
\hline
MMJ033238.54-274634.6$^\dagger$ & C1 & 03:32:38.54 & $-27$:46:34.6 & 39.9 & $553\pm 14$ & 0.92 & $31.1\pm5.0$ & $0.89$ & Y \\
MMJ033239.73-274611.6$^\dagger$& C2 & 03:32:39.73 & $-27$:46:11.6 & 10.3 & $223\pm 22$ & 0.59 & $<21$ & 0.56 & Y\\
MMJ033238.03-274626.5& C3 & 03:32:38.03 & $-27$:46:26.5& 9.6 & $145\pm 12$ & 0.95 & $<12$ & 1.00 & Y\\
MMJ033236.20-274628.2& C4 & 03:32:36.20 & $-27$:46:28.2 & 6.1 & $ 87\pm 14$ & 0.89 & $<17$ & 0.68 & Y\\
MMJ033237.35-274645.7& C5 & 03:32:37.35 & $-27$:46:45.7 & 5.2 & $ 71\pm 14$ & 0.92 & $<16$ & 0.70 & Y\\
MMJ033235.47-274626.6$^\dagger$& C6 & 03:32:35.47 & $-27$:46:26.6 & 3.9 & $ 97\pm 25$ & 0.51 & $<25$ & 0.45 & Y\\
MMJ033235.75-274627.7& C7 & 03:32:35.75 & $-27$:46:27.7 & 3.7 & $ 70\pm 19$ & 0.67 & $<21$ & 0.55 & Y\\
MMJ033238.57-274648.0& C8 & 03:32:38.57 & $-27$:46:48.0 & 3.6 & $ 46\pm 13$ & 0.99 & $<18$ & 0.62 & N\\
MMJ033237.74-274603.0 & C9 & 03:32:37.74 & $-27$:46:03.0 & 3.5 & $ 55\pm 16$ & 0.80 & $<16$ & 0.70 & N\\
\hline
\multicolumn{10}{c}{Supplemetary sample at $3.0-3.5\sigma$ significance}\\
\hline
MMJ033237.36-274613.2& C10 & 03:32:37.36 & $-27$:46:13.2 & 3.3 & $ 45\pm 14$ & 0.93 & $<13$ & 0.88 & N\\
MMJ033238.77-274650.1& C11 & 03:32:38.77 & $-27$:46:50.1 & 3.2 & $ 47\pm 14$ & 0.88 & $<21$ & 0.55 & N\\
MMJ033237.42-274650.4& C12 & 03:32:37.42 & $-27$:46:50.4 & 3.2 & $ 59\pm 18$ & 0.69 & $<19$ & 0.60 & Y\\
MMJ033236.50-274647.4& C13 & 03:32:36.50 & $-27$:46:47.4 & 3.2 & $ 67\pm 21$ & 0.60 & $<22$ & 0.52 & Y\\
MMJ033236.43-274632.1& C14 & 03:32:36.43 & $-27$:46:32.1 & 3.1 & $ 46\pm 15$ & 0.85 & $<16$ & 0.73 & Y\\
MMJ033237.49-274649.3& C15 & 03:32:37.49 & $-27$:46:49.3 & 3.1 & $ 52\pm 17$ & 0.76 & $<18$ & 0.63 & N\\
MMJ033237.75-274609.6& C16 & 03:32:37.75 & $-27$:46:09.6 & 3.0 & $ 41\pm 14$ & 0.93 & $<14$ & 0.85 & N\\
\hline\hline
\end{tabular}
\flushleft
\noindent $^\dagger$ Sources ASPECS C1, C2 and C6 in this paper correspond to sources 3mm.1, 3mm.2 and 3mm.5 in Decarli et al. (Paper~IV).
\end{table*}
\subsection{ALMA observations and data reduction}
The ASPECS survey setup and data reduction steps are described in detail in Paper~I (Walter et al. 2016). Here we repeat the most relevant information for the study presented here.
ALMA band 3 and band 6 observations were obtained during Cycle-2 as part of projects 2013.1.00146.S (PI: F. Walter) and 2013.1.00718.S (PI: M. Aravena). Observations in band-3 were conducted between July 01, 2014 to January 05, 2015, and observations in band 6 were conducted between December 12, 2014 to April 21, 2015 under good weather conditions.
Observations in band 3 were performed in a single pointing in spectral scan mode, using 5 frequency tunings to cover $84.2-114.9$ GHz. Over this frequency range the ALMA half power beam width (HPBW), which corresponds to a primary beam (PB) response of 0.5, varies between $61''$ and $45''$. Observations in band 6 were performed in a 7-point mosaic, using a hexagonal pattern (Fig. \par\noindent\hangindent 20 pt{fig_map1mm}): the central pointing overlaps the other 6 pointings by about half the ALMA PB, i.e., close to Nyquist sampling. We scanned band 6 using eight frequency tunings, covering $212.0-272.0$ GHz. The ALMA PB in individual pointings ranges between $30''$ and $23''$.
Observations in bands 3 and 6 were taken with ALMA's compact array configurations, C34-2 and C34-1, respectively. The observations used between 30 and 35 antennas in each band, resulting in synthesized beam sizes of $3.6''\times2.1''$ and $1.7''\times0.9''$ from the low to high frequency ends of bands 3 and 6, respectively.
Flux calibration was performed on planets or Jupiter's moons, with passband and phase calibration determined from nearby quasars, and should be accurate within $\pm10\%$. Calibration and imaging was done using the Common Astronomy Software Application package ({\sc CASA}). The calibrated visibilities were inverted using the {\sc CASA} task \verb CLEAN \, using natural weighting. To obtain continuum maps, we collapsed along the frequency axis in the uv-plane and inverted the visibilities using the {\sc CASA} task \verb CLEAN \, using natural weighting and mosaic mode. We use the Multi-frequency Imaging Synthesis (MFS) algorithm with \verb nterms=1 \, as the joint implementation of \verb nterms>1 \, and mosaic mode are not yet available in {\sc CASA}. This implies assuming a first order polymial fit for point sources along the frequency axis, which is the best assumption for low signal to noise data (most sources with S/N$<10$) as in this paper \citep[see {\sc CASA} cookbook and ][]{rau11}. We also tested the effect of using different frequency weightings in the visibility plane, however no significant changes were seen in the final collapsed images.
In this process, we produced `clean' maps masking with tight boxes all the continuum sources previously detected in the `dirty' maps with significances above $5\sigma$, and cleaning down to a $2.5\sigma$ threshold. Given the large bandwidth covered by our observations, the contamination by line emission in the continuum map becomes negligible.
The final maps are shown in Figs. \par\noindent\hangindent 20 pt{fig_map1mm} and \par\noindent\hangindent 20 pt{fig_map3mm}. The sensitivity in each map declines with respect to the distance from the phase pointing center, and, given the smaller PB, declines particularly sharply for the 1.2-mm observations at the outskirts of the mosaicked region. We reach an rms sensitivity of $12.7\mu$Jy and $3.8\mu$Jy in the centres of the 1.2-mm and 3-mm maps, respectively. The final map average frequencies over the frequency ranges covered are 242 and 95 GHz, respectively.
Finally, we note that while source confusion for individual detections is negligible in these deep ALMA maps, it is at the level where it becomes important for stacking analyses. With an ALMA beam size at 1.2-mm of $1.7''\times0.9''$, there are $8.47\times10^6$ beams per deg$^2$. At the bottom flux bin of our number count measurements (see \S \par\noindent\hangindent 20 pt{sec_counts}), we find $1.32\times10^5$ sources per deg$^2$. This translates into one source per $\sim 64$ beams, and implies that confusion is not an issue. The same logic applies for the stacking analysis presented below (see \S \par\noindent\hangindent 20 pt{sec_stack}). The deepest stacks considered reach a $3\sigma$ level of 8 $\mu$Jy at 1.2-mm. Extrapolating the number counts to this flux level, we find about $6.0\times10^5$ sources per deg$^2$. This results in one source per $14$ beams. According to \citet{helou90}, bright source confusion becomes important at one source per 22 beams, suggesting that stacking experiments in these ALMA deep maps will be affected. However, this confusion limit depends on the slope of the number counts, and since this slope appears to flatten at these faint flux levels, it is possible that confusion would have a lesser impact at these depths, and in particular on stacking analyses.
\subsection{Multi-wavelength data}
Our ALMA observations cover a $\sim1$ arcmin$^2$ region within the deepest 4.7 arcmin$^2$ of the {\it Hubble} UDF: the eXtremely Deep Field (XDF). Available data includes {\it HST} Advanced Camera for Surveys (ACS) and Wide Field Camera 3 IR data from the HUDF09, HUDF12 and Cosmic Assembly Near-infrared Deep Extragalactic Legacy Survey (CANDELS) programs as well as public photometric and spectroscopic catalogs \citep{coe06,xu07,rhoads09,schenker13,mclure13, skelton14,bouwens14,morris15,momcheva15}. In this study, we make use of this optical and infrared coverage of the XDF, including the photometric and spectroscopic redshift information available from \citet{skelton14}. In addition to the {\it HST} coverage, a wealth of optical and infrared coverage from ground based telescopes is available in this field \citet[see ][]{skelton14}. The HUDF was also covered by the {\it Spitzer} Infrared Array Camera (IRAC) and Multiband Imaging Photometer (MIPS), as well as by the {\it Herschel} Photodetector Array Camera and Spectrometer (PACS) and the Spectral and Photometric Imaging Receiver (SPIRE) \citep{elbaz11}.
\section{Results}
\label{sec_results}
\subsection{Source detection and flux measurements}
Source detection was performed using SExtractor \citep{bertin96} in the ALMA 1.2-mm and 3-mm maps prior to PB correction. We use a minimum area of 5 pixels ($1.5''$) for detection, extracting sources down to 2.5$\sigma$, where $\sigma$ is evaluated locally for each source. Source extraction in the 1.2-mm map was performed beyond the HPBW of our mosaic, out to PB $=0.3$, however most sources are detected within PB $=0.5$, in the central region of the mosaic. Although we extract all sources down to $2.5\sigma$, we consider as individual detections only sources above $>3.5\sigma$ significance. This significance level cut corresponds to roughly $50-60\%$ fidelity of the sample (see \S~\par\noindent\hangindent 20 pt{sec_purity}). These sources are highlighted with boxes in Figs. 1 and 2, and are listed in Table~1.
Nine sources are detected in the 1.2-mm map at a significance above 3.5$\sigma$. For reference, Table~1 also lists another 7 sources with significances between $3.0-3.5\sigma$ (our supplementary sample). Given the lower significance of these sources, we choose not to use them to study the multi-wavelength properties of this population. Nevertheless we can use them to constrain the number counts of faint sources, after correcting for fidelity and completeness. Only one source is detected in the 3-mm map at the $>3.5\sigma$ significance level, corresponding to the brightest detection at 1.2-mm. For this reason, we only show the 1.2-mm detected sources in Figs.~1--2.
We compute fluxes based on 2-dimensional Gaussian fit centered at the location of the SExtractor detection. In all but one case (discussed below) the sources are unresolved at the resolution and depth of the 1.2-mm observations. We therefore list the flux as the peak flux density value at the source position delivered by the fitting routine. These fitted values are in agreement with the actual pixel values at the position of the sources. We cannot discard the possibility that sources with low significances are indeed being resolved given the relatively small beam size. It is thus unclear what fraction of the flux is being unaccounted for in individual sources.
Only the brightest source in the map is marginally spatially resolved with a measured angular size of $(0.52\pm0.14)''\times(0.43\pm0.26)''$ (PA$=49\deg$), and we record the integrated flux in Table~1. More details on this source's properties are given in Paper~IV \citep{decarli16b}. Since only one source is detected in the 3-mm map, in what follows we concentrate on characterising the properties of the 1.2-mm sources.
\begin{figure*}[ht]
\centering
\includegraphics[scale=0.4]{fig_cont_purity2.ps}
\includegraphics[scale=0.4]{fig_cont_completeness2.ps}
\caption{({\it Left:}) Fidelity ({\em top panel}) and number of detections ({\em bottom panel}) as a function of 1.2-mm flux density of the ASPECS sample (non-cumulative). The solid curve is a model for the fidelity. Our sample shows 100\% fidelity at $S_{\rm 1.2mm}\sim100\mu$Jy and 50\% fidelity at $\sim40\mu$Jy (3.0$\sigma$). ({\it Right:}) Completeness of our 1.2-mm continuum sample detection as a function of 1.2-mm flux density. The solid curve shows a model for the completeness behaviour as a function of 1.2-mm flux density. Our sample shows 100\% completeness at $S_{\rm 1.2mm}\sim300\mu$Jy and 50\% completeness at $\sim40\mu$Jy (3.0$\sigma$).}\label{fig_purity}\label{fig_completeness}
\end{figure*}
\subsection{Fidelity and completeness}\label{sec_purity}
We quantify the occurrence of spurious sources in our 1.2-mm sample by applying the detection routine explained in the previous section to the inverted `negative' map. We thus compute the fidelity $P$ of our sample as:
\begin{equation}
P({\rm S_{\rm 1.2mm}})=1-\frac{N_{\rm neg}({\rm S_{\rm 1.2mm}})}{N_{\rm pos}({\rm S_{\rm 1.2mm}})},
\end{equation}
where $N_{\rm neg}$ and $N_{\rm pos}$ are the number of negative and positive sources, respectively, detected in the map as a function of 1.2-mm flux density.
\begin{figure*}[ht]
\centering
\includegraphics[scale=0.6]{fig_ncount_v3.ps}
\caption{Number counts of ALMA 1.2mm continuum sources in the UDF compared with values from the literature. Our data have been corrected to account for completeness and fidelity in the source identification, as discussed in the text. Uncertainties in each number count measurements correspond to Poisson errors. Our measurements span almost two orders of magnitude in flux density. Filled circles represent literature measurements obtained at 1.2-mm. Open circles represent measurements from different wavelengths than 1.2-mm and converted to this wavelength. Most of the measurements from the literature at the faint levels are not blank field and are thus biased, since their observations target bright sources in the field (they measure counts around other sources). The \citet{fujimoto16} data pointing towards lower flux densities are based on lensed galaxy clusters.}\label{fig_ncounts}
\vspace{10mm}
\end{figure*}
\begin{table}[h]
\begin{flushleft}
\caption{ALMA UDF 1.2-mm number counts. Columns: (1) Flux density bin center (in units of mJy); (2) Number of entries per bin (before fidelity and completeness correction); (3) Number of sources per square degrees; (4), (5) Lower and upper uncertainties (error bars) on $N(>S_\nu)$.}\label{tab_ncounts}
\end{flushleft}
\begin{tabular}{ccccc}
\hline
log($S_\nu$) & $dN/dlog(S_\nu)$ & $N(>S_\nu)$ & $\delta N_{-}$ & $\delta N_{+}$ \\
(mJy) & (mJy$^{-1}$) & (deg$^{-2}$) & (deg$^{-2}$) & (deg$^{-2}$)\\
(1) & (2) & (3) & (4) & (5) \\
\hline\hline
$-1.49$ & 23 & 132000 & 3700 & 43000\\
$-1.24$ & 10 & 71500 & 16600 & 21500\\
$-0.99$ & 3 & 23700 & 9400 & 14700\\
$-0.74$ & 1 & 9200 & 5800 & 11900\\
$-0.24$ & 1 & 4500 & 3800 & 10400\\
\hline\hline
\end{tabular}
\end{table}
Figure \par\noindent\hangindent 20 pt{fig_purity} shows the fidelity and number of positive detections in our map as a function of 1.2-mm flux density. Not surprisingly, we find that the fidelity of our sample is a strong function of the 1.2-mm flux density. We reach 100\% fidelity at $100\mu$Jy (7.8$\sigma$) and 50\% fidelity at $40\mu$Jy ($\sim3.0\sigma$). This means that at the $3\sigma$ level, half of our sources are expected to be spurious, which motivates our choice of 3.5$\sigma$ cut for the main sample.
We parametrise the fidelity with 1.2-mm flux density as:
\begin{equation}
P(S_{\rm 1.2mm})=\frac{1}{2} {\rm erf} \{ \frac{{\rm log}_{10}(S_{\rm 1.2mm})-A}{B} \} +1.0
\end{equation}
where $A={\rm log}_{10}(42)$ and $B=1/4$, and $S_{\rm 1.2mm}$ is in units of $\mu$Jy. We use this parametrisation to compute the fidelity level or reliability of our individual detections.
We compute the completeness of our observations by running Monte Carlo simulations on our continuum map. We ingest 10 artificial point-like sources with randomly generated flux levels (between $10-300 \mu$Jy) in the ALMA map. We then run our source detection procedure to identify and compute the fraction of recovered sources (versus the input sources). Recovered artificial sources are matched with the input positions within a radius of $1''$, roughly the size of our synthesized beam. Similar to the findings of \citet{fujimoto16}, the input and recovered flux densities agree well within individual source uncertainties. We repeat this process 10 times, for a total of 100 artificial sources. Note that we do not inject all 100 sources in a single step since this would result in significant source blending in the ALMA image.
Figure \par\noindent\hangindent 20 pt{fig_completeness} shows the resulting completeness as a function of extracted 1.2-mm flux density. We find that our sample is 100\% complete at $S_{\rm 1.2mm}\sim300\mu$Jy (23$\sigma$) and 50\% complete at $\sim40\mu$Jy (3.0$\sigma$). This indicates that at the $3\sigma$ level, we recover only half of real input sources.
We parametrize the completeness with 1.2-mm flux density as:
\begin{equation}
C(S_{\rm 1.2mm})=\frac{1}{2} {\rm erf} \{ \frac{{\rm log}_{10}(S_{\rm 1.2mm})-A'}{B'} \} +1.0
\end{equation}
where $A'={\rm log}_{10}(35)$ and $B'=0.45$, and $S_{\rm 1.2mm}$ is in units of $\mu$Jy. We use this parametrisation to compute the completeness level of our individual detections.
\section{Number counts}
\label{sec_counts}
\begin{figure*}[ht]
\centering
\includegraphics[scale=0.9]{UDF_1mm_color_image_v4.ps}
\caption{HUDF multi--color image (F435W, F850LP, F105W) of the region covered by our 1-mm ALMA observations. The boxes show the position of the 1.2-mm sources detected with our extraction procedure at $S/N>3.5$. The white contour shows the coverage of our ALMA observations down to PB$=0.2$. }\label{fig_color}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[scale=1.5]{alma_cont_thumbnails_opt_part1.ps}
\caption{Multi-wavelength image thumbnails toward the ALMA 1.2-mm continuum detections ($>3.5\sigma$). From left to right, we show an optical-near infrared false color composite (F435W/F850LP/F105W), and individual images in the F850LP, F160W, IRAC channel 1 and $10"\times10"$ in size. }\label{fig_thumb}
\end{figure*}
\begin{figure*}
\centering
\figurenum{6}
\includegraphics[scale=1.5]{alma_cont_thumbnails_opt_part2.ps}
\caption{continued}
\end{figure*}
We use the sources detected in our ALMA UDF map to compute the number counts at 1.2mm. We compute the number counts ($N(S_i)$) in each flux density bin $S_i$ as:
\begin{equation}
N(S_i) = \frac{1}{A} \sum_{j=1}^{X_i} \frac{P_j}{C_j},
\end{equation}
where $A$ is the effective area of our ALMA mosaic and $X_i$ is the number of sources in each particular bin $i$. The parameters $P_j$ and $C_j$ correspond to the fidelity and completeness at the flux bin $i$. Since we are limited by the modest number of detections, we compute the cumulative number counts rather than computing differential counts by summing up each $N(S_i)$ over all measurements $>S_i$. In addition, we extend our number count measurements down to significances of $3\sigma$. While at this level there is substantial contamination and low detection rate, we can statistically correct the values for fidelity and completeness. As pointed out in the previous section, at the $3\sigma$ level we reach 50\% fidelity as well as 50\% completeness in our sample detection. This implies that these effects cancel out when we compute the number counts. Thus, while we obtain correct number counts at the $3\sigma$ level, the identification of real sources is correct only in half of the cases.
The uncertainties in the number counts are computed by including the Poissonian errors as well as flux uncertainties in each individual measurement. The uncertainties in each bin are dominated by the Poissonian errors on $X_i$, however at the lowest significance levels the flux uncertainties start to have a significant contribution.
The cumulative number counts ($N(>S_\nu)$) are shown in Fig. \par\noindent\hangindent 20 pt{fig_ncounts}. The actual measurements are listed in Table \par\noindent\hangindent 20 pt{tab_ncounts}. For comparison, we show number count measurements from the literature \citep{karim13, hatsukade13,ono14,carniani15,oteo15,simpson15,fujimoto16}. We scale the flux densities of the different studies as $S_{\rm 1.2mm}=0.4~S_{\rm 870}$, $S_{\rm 1.2mm}=0.8~S_{\rm 1.1mm}$ and $S_{\rm 1.2mm}=1.3~S_{\rm 1.3mm}$ \citep[for consistency with ][]{fujimoto16}.
Our ALMA UDF observations appear to be in general agreement with these earlier measurements, in particular with the counts obtained by \citet{carniani15} and \citet{oteo15}. However, our counts are lower by about a factor of 2 in the flux range $S_{\rm 1.2mm}=0.06-0.4$ mJy compared to other studies in the literature \citep[][]{hatsukade13,ono14,fujimoto16}. These difference could be explained by the fact that these studies might be biased as they used pointed observations toward brighter sources in the field to derive the number counts (i.e., these studies are not unbiased blank field surveys).
Another possibility is that cosmic variance does play an important role among the different analyses; e.g. the ECDFS, where the UDF resides, is believed to be underdense of submillimeter sources above $\sim3$ mJy (at 345 GHz) by a factor of $\sim2$ \citep{weiss09}. As indicated by several studies, the ECDFS appears to be underdense in other galaxy populations as well, including $BzK$ galaxies, X-ray and radio sources \citep[e.g.,][]{lehmer05,blanc08}. However, as already noted by \citet{weiss09}, the underdensity appears to be seen only in the brightest sources, given the steep slope at fainter fluxes \citep[see also ][]{karim13}. Another possibility is that the differences in number counts between studies come from scatter induced by different analysis techniques and methods. This effect was seen to be a dominant compared to statistical fluctuations in radio surveys \citep{condon07}.
\section{Multi-wavelength properties of the ALMA 1.2-mm sources}
\label{sec_prop}
\subsection{Astrometric offset}
Using the identified mm/optical counterpart positions (see below), we measure a systematic astrometric offset of the {\it HST} positions of $\approx0.3"$ to the north of the ALMA positions. To check the ALMA registration we inspected the millimeter calibrators used, finding good astrometric solutions, accurate within $0.01"$ with respect to the catalogued radio-based values. Based on the GOODS 2008 data release documentation\footnote{\url{https://archive.stsci.edu/pub/hlsp/goods/v2/}\\ \url{h_goods_v2.0_rdm.html}}, it is clear that a consistent offset ($0.32''$) was applied to the GOODS-North astrometric solution but not to the GOODS-South data. Hence, we correct the {\it HST} positions by $0.3"$ to match the ALMA millimeter registration throughout. This is consistent with results from a shallower ALMA millimeter continuum survey of the full HUDF \citep{dunlop16, rujopakarn16}.
\subsection{Identification and SED fitting}
Figure \par\noindent\hangindent 20 pt{fig_color} shows the location of the 1.2-mm continuum sources with respect to the optical galaxies in the field. Our blank-field observations encompass a significant number of optical galaxies, however this contrast the galaxies detected in the millimeter regime. Our sources do not appear to be clustered.
For each individual 1.2-mm continuum detection, we identify optical counterparts within a radius of $1''$ from the millimeter position. We choose this search radius since it is well matched to the ALMA 1.2-mm synthesized beam ($1.7''\times0.9''$). Figure \par\noindent\hangindent 20 pt{fig_thumb} presents multi-wavelength cutouts for individual detections. Seven of the continuum sources with significances $>3.5\sigma$ have an obvious counterpart in the {\it HST} images, and five of these have an available spectroscopic redshift \citep[see Table \par\noindent\hangindent 20 pt{tab_prop};][]{skelton14}. The other two millimeter detections, with lower significances in our sample ($\sim3.5-3.6\sigma$), do not show an obvious counterpart. Four out of seven sources with significances between $3.0-3.5\sigma$ do not have an optical counterpart (Table \par\noindent\hangindent 20 pt{tab_sources}), consistent with the fidelity level at this significance, and indicating that some or all of these are likely spurious millimeter detections. Another possibility would be that these are faint dusty galaxies at higher redshifts \citep[as in HDF850.1; see ][]{walter12}.
We fit the spectral energy distribution (SED) of the continuum--detected galaxies using the high-redshift extension of \texttt{MAGPHYS} \citep{dacunha08, dacunha15}. We use the available 26 broad and medium band filters in the optical and infrared regimes, from the $U$ band to {\it Spitzer} IRAC 8$\mu$m. We here also include the ALMA 1.2-mm data flux densities, however we note that the optical/infrared data has a much stronger weight given the tighter constrains in this part of the spectra. We do not include {\it Herschel} photometry in the fits since its angular resolution is very poor, being almost the size of our target field for some of the IR bands. The {\it Herschel} photometry is thus heavily blended.
\begin{figure}
\centering
\includegraphics[scale=0.3]{mstar_sfr_distr_udf_new.eps}
\caption{Distribution of stellar masses and SFRs (obtained from SED fitting) for the galaxies detected in our ALMA 1.2-mm continuum map. For comparison, the distribution of field galaxies in the relevant redshift range is shown.}\label{fig_distr}
\end{figure}
For each individual galaxy, we perform SED fits to the photometry fixed at the best available redshift. \texttt{MAGPHYS} delivers estimates for the stellar masses, star formation rate (SFR), dust mass and IR luminosity. Even though for most galaxies we do not have photometric constraints on the observed IR SED, \texttt{MAGPHYS} employs a physically-motivated prescription to balance the energy output at different wavelengths. Thus, estimates on the IR luminosity, and/or dust mass, come from constraints on the dust re-processed UV light, which is well sampled by the UV-to-infrared photometry. For some galaxies with faint optical/near-infrared fluxes or with weak constraints in the photometry, \texttt{MAGPHYS} is able to output only some of the parameters with enough accuracy (e.g., stellar masses). However all the optical counterparts of our millimeter detected sample are sufficiently bright to yield good parameters derived by \texttt{MAGPHYS}. The properties derived for individual sources detected in our ALMA 1.2-mm continuum are shown in Table \par\noindent\hangindent 20 pt{tab_prop}.
Figure \par\noindent\hangindent 20 pt{fig_distr} shows the distribution of stellar masses and SFRs of our ALMA 1.2-mm continuum sources. For comparison, we show the stellar masses and SFRs derived in the same way for field galaxies located within the field of view of our ALMA map (within PB=0.4), and selected to be in a redshift range that matches the redshifts of our ALMA continuum sources. We limit the comparison sample to sources with $m_{\rm F850LP}$ and $m_{F160W} < 27.5$ mag AB, in order to ensure good SED fits and derived properties. We find that the faint DSFG population, as revealed by our ALMA 1.2-mm sources, have higher stellar masses and SFRs than the field galaxy population at similar redshifts, yet much lower values than those found in brighter DSFGs (i.e. SMGs). Our sources show a median stellar mass of $4.0\times10^{10}\ M_\odot$ and a median SFR of $40\ M_\sun$ yr$^{-1}$, which are significantly lower than the typical values for SMGs, with stellar masses in the range $(0.8-3.0)\times10^{11}\ M_\sun$ \citep[e.g.,][]{michalowski10,hainline11,michalowski12,simpson14,dacunha15, koprowski16}, and SFRs well above $100\ M_\sun$ yr$^{-1}$ \citep[e. g.,][]{casey14}.
\begin{table*}
\begin{flushleft}
\caption{Derived properties for the ALMA UDF 1.2-mm sources. Columns: (1) Source name; (2) Best available redshift estimate. If spectroscopic, we quote three decimal places. If photometric, we quote only 2 decimal digits. References: CO based redshifts, confirmed with optical spectroscopy for C1, C2 and C6 (Walter et al. 2016, Paper~I; Decarli et al. 2016b, Paper~IV; Skelton et al. 2014). Optical redshifts for C3, C4, C5, and C7 \citep{skelton14}. Photometric redshifts for C3 and C4 from \citet{coe06} and \citet{skelton14}. (3), (4) AB magnitudes in the F850LP and F160W {\it HST} bands. Uncertainties in quoted values range between 0.01-0.05 mag; (5) Stellar mass derived through SED fitting; (6) SFR derived through SED fitting; (7) Specific SFR (SFR/$M_*$); (8) IR luminosity output from \texttt{MAGPHYS}; (9) ISM mass derived from the dust mass delivered by \texttt{MAGPHYS} and a gas-to-dust ratio $\delta_{\rm GDR}=200$; (10) ISM mass obtained from the 1.2-mm flux and the calibrations from \citet{scoville14}.}
\end{flushleft}
\begin{tabular}{cccccccccc}
\hline
ID & z$_{\rm best}$ & $m_{\rm F850LP}$ & $m_{\rm F160W}$ & log$_{10}(M_{*})$ & log$_{10}$(SFR) & log$_{10}$(sSFR) & log$_{10}(L_{\rm IR})$ & log$_{10}(M_{\rm ISM,d})$ & log$_{10}(M_{\rm ISM,1mm})$ \\
ASPECS & & (AB mag) & (AB mag) & $(M_{\odot})$ & $(M_{\odot}$ yr$^{-1})$ & (Gyr$^{-1})$ & ($L_\sun$) & $(M_{\odot})$ & $(M_{\odot})$ \\
(1) & (2) & (3) & (4) & (5) & (6) & (7) & (8) & (9) & (10) \\
\hline\hline
C1 & 2.543 & 24.0 & 23.2 & $10.36_{-0.00}^{0.12}$ & $ 2.26_{ -0.00}^{ 0.09}$ & $ 1.03_{ -0.10}^{ 0.00}$ & $12.74_{ -0.01}^{ 0.14}$ & $10.59_{ -0.08}^{ 0.06}$ & $10.69\pm 0.01$ \\
C2 & 1.552 & 24.4 & 21.7 & $11.53_{-0.12}^{0.02}$ & $ 1.65_{ -0.00}^{ 0.08}$ & $-0.88_{ -0.00}^{ 0.20}$ & $11.90_{ -0.01}^{ 0.04}$ & $10.45_{ -0.15}^{ 0.12}$ & $10.32\pm 0.04$ \\
C3 & 1.65 & 25.8 & 23.6 & $11.02_{-0.09}^{0.12}$ & $ 2.36_{ -0.34}^{ 0.08}$ & $ 0.38_{ -0.50}^{ 0.10}$ & $12.54_{ -0.32}^{ 0.04}$ & $ 10.05_{ -0.09}^{ 0.06}$ & $10.04\pm 0.05$ \\
C4 & 1.89 & 24.5 & 23.1 & $10.36_{-0.06}^{0.01}$ & $ 1.64_{ -0.35}^{ 0.00}$ & $ 0.27_{ -0.40}^{ 0.00}$ & $11.65_{ -0.33}^{ 0.00}$ & $ 10.00_{ -0.25}^{ 0.25}$ & $ 9.91\pm 0.07$ \\
C5 & 1.846 & 23.4 & 22.0 & $10.61_{-0.06}^{0.06}$ & $ 1.43_{ -0.18}^{ 0.26}$ & $-0.18_{ -0.20}^{ 0.25}$ & $11.46_{ -0.20}^{ 0.36}$ & $ 9.92_{ -0.29}^{ 0.28}$ & $ 9.83\pm 0.08$ \\
C6 & 1.088 & 22.1 & 21.1 & $10.48_{-0.10}^{0.10}$ & $ 1.41_{ -0.14}^{ 0.28}$ & $-0.07_{ -0.20}^{ 0.35}$ & $11.53_{ -0.15}^{ 0.26}$ & $ 10.10_{ -0.22}^{ 0.25}$ & $ 9.95\pm 0.11$ \\
C7 & 1.094 & 22.8 & 21.4 & $10.88_{-0.10}^{0.11}$ & $ 1.21_{ -0.37}^{ 0.38}$ & $-0.68_{ -0.34}^{ 0.35}$ & $11.49_{ -0.35}^{ 0.33}$ & $ 9.88_{ -0.40}^{ 0.30}$ & $ 9.81\pm 0.12$ \\
C8 & \ldots & $>30.6$ & $>30.6$ & \ldots & \ldots & \ldots & \ldots & \ldots & \ldots \\
C9 & \ldots & $>30.6$ & $>30.6$ & \ldots & \ldots & \ldots & \ldots & \ldots & \ldots \\
\hline \label{tab_prop}
\end{tabular}
\end{table*}
\begin{figure}[h]
\centering
\includegraphics[scale=0.4]{zdistr_udf_new.eps}
\caption{Redshift distribution for (sub)millimeter selected galaxies. The y-axis shows the number of galaxies in each bin, normalized to the total number of galaxies in each sample. The black solid line shows the redshift distribution of our ALMA UDF 1.2-mm detections ($>3\sigma$). The gray and green solid lines show the redshift distribution for the 1.2/1.4-mm selected samples of SMGs in the COSMOS \citep{miettinen15} and SPT surveys \citep{weiss13}, respectively. The dashed orange and blue lines show the 850/870-$\mu$m selected SMGs from \citet{chapman05} and from the ECDFS \citep{simpson14}.} \label{fig_zdistr}
\end{figure}
\subsection{Redshift distribution}
Since most of the galaxies detected at $>3.5\sigma$ in our sample have available spectroscopic redshifts from the various surveys of the UDF, we investigate the redshift distribution of our sample.
Figure \par\noindent\hangindent 20 pt{fig_distr} shows the redshift distribution for our ALMA continuum sources that have an optical counterpart compared with various millimeter selected samples of bright DSFGs from the literature.
We find that all the 1.2-mm continuum sources detected above $3.5\sigma$ in our sample are located in the redshift range $z=1-3$, and none are associated convincingly with a galaxy at $z>3$. This excludes the source candidates without counterparts. While this may only reflect the low number statistics due to the small area of the sky covered, it also supports the idea that the population of galaxies discovered in our deep ALMA 1.2-mm continuum map significantly differs from the population of DSFGs found in shallower but wider (sub)millimeter surveys. The DSFGs samples from the literature are found to have a median redshifts ranging from $z\sim2.1$ and $z\sim3.1$, respectively, with a possible tail extending out to $z\sim6$ \citep{chapman05,yun12, smolcic12, weiss13, riechers13, simpson14, miettinen15, strandet16, dunlop16}. We find that our faint ALMA millimeter-selected galaxies, however, have a median redshift $z=1.7\pm0.4$. The uncertainty here corresponds to the scatter in the redshifts. This median redshift is significantly lower than the typical redshift of bright DSFGs, irrespective of the nature the DSFG samples (lensed or unlensed) or the selection wavelength (870-$\mu$m or 1.2-mm). Statistically, this would not be significantly affected if the two sources without counterparts were located at $z>2$ given the small scatter in the redshift distribution.
\begin{figure*}[ht]
\centering
\includegraphics[scale=0.45]{mstar_sfr_cont_udf_new.eps}
\caption{Stellar mass versus SFR for the galaxies covered in our ALMA UDF 1.2-mm map in the two relevant redshift bins. The large yellow circles show the ALMA 1.2-mm continuum sources ($>3.5\sigma$). The small blue circles show field galaxies in either the $1.0<z<2.0$ or $2.0<z<3.0$ redshift bins. Field galaxies are restricted to be brighter than 27.5 AB mag in the F850LP and F160W bands. For comparison, the orange and magenta curves represent the best second order polynomial fits of the star formation sequence at $1.0<z<1.5$, $1.5<z<2.0$ and $2.0<z<2.5$ for the left and right panels, respectively \citep{whitaker14}.}\label{fig_msz}
\end{figure*}
While the SMG and fainter-mm source populations are obviously different as reflected by the significantly lower 1.2-mm fluxes, this is the first time that we are able to evaluate the redshift distribution of the faintest 1.2-mm emitters in a contiguous blank field (below $S_{\rm 1.2mm}=0.5$ mJy). Other studies reaching down to the faint mm flux regime, are mostly based on archival data of different individual fields where the faint mm emitters are not the main targets \citep[e.g.,][]{oteo15, carniani15, fujimoto16} or do not have the excellent deep multi wavelength coverage of the HUDF in order to address this issue.
The decline in the median redshift with decreasing flux density for millimeter selected sources was recently predicted by phenomenological models of galaxy evolution \citep{bethermin15b}. Even though the prediction does not assess the redshifts for populations with 1.2-mm flux densities below 0.2 mJy, already at this flux level they find a median redshift of $\sim2$ compared to the much higher $z\sim3$ predicted for brighter SMGs selected at 1.2-mm. By extrapolating their prediction down to a flux density cut of $\sim35\mu$Jy (our 3$\sigma$ cut), we find an expected median redshift of $\sim1.5$. This value is in good agreement with our measurements, and supports the fact that the redshift distribution of millimeter-selected galaxies is affected by the flux density cut.
\subsection{Starburst versus Main sequence}
An important result from multi-wavelength surveys in the last decade has been the determination that typical star-forming galaxies form a tight linear relationship in the SFR-$M_{\rm stars}$ plane out to $z\sim3$ \citep[e.g., ][]{brinchmann04, elbaz07, noeske07,daddi07,pannella09,karim11,rodighiero11,whitaker12}. Sources that lie close to this star formation relationship have been termed main sequence galaxies. Galaxies lying above this sequence are called starbursts, as they have excess star formation activity with respect to most galaxies in the main-sequence for the same stellar mass, or higher specific star formation rates (sSFRs). This sequence has been observed to evolve with redshift, with higher SFRs for a given stellar mass at increasing redshifts \citep{whitaker12}, and it has also been claimed to flatten at the high stellar mass end \citep{whitaker12, whitaker14,pannella15,lee15}.
Figure \par\noindent\hangindent 20 pt{fig_msz} shows the stellar mass versus SFR derived using \texttt{MAGPHYS} for all HST-detected galaxies at $1<z<3$ contained within our ALMA UDF survey area (within PB $=0.4$ of our 1.2-mm map), and restricted to be brighter than 27.5 AB mag in the F850LP and F160W bands. We show the sources detected in our 1.2-mm observations ($>3.5\sigma$), and compare with the main-sequence fit derived by Whitaker et al. (2014). We find that all the millimeter detected galaxies at $z<2$ are located within the scatter of the main sequence at $z\sim1-2$ and taking into account the uncertainties in the derived properties. Similarly, the only millimeter detection at $z>2$ (ASPECS C1) is also well within the scatter of the main sequence at $z=2.0-2.5$. We thus conclude that our faint ALMA 1.2-mm continuum sources are main-sequence galaxies at $z\sim1-3$.
\begin{figure*}[ht]
\centering
\includegraphics[scale=0.3]{stack_z01_35sigma_new.ps}
\includegraphics[scale=0.3]{stack_z12_35sigma_new.ps}
\includegraphics[scale=0.3]{stack_z23_35sigma_new.ps}\\
\vspace{3mm}
\includegraphics[scale=0.3]{stack_ms_mst1_35sigmacut_new.ps}
\includegraphics[scale=0.3]{stack_ms_mst2_35sigmacut_new.ps}
\includegraphics[scale=0.3]{stack_ms_mst3_35sigmacut_new.ps}\\
\vspace{3mm}
\includegraphics[scale=0.3]{stack_ms_sfr1_35sigmacut_new.ps}
\includegraphics[scale=0.3]{stack_ms_sfr2_35sigmacut_new.ps}
\includegraphics[scale=0.3]{stack_ms_sfr3_35sigmacut_new.ps}
\caption{Stacked 1.2-mm continuum on the location of galaxies selected as summarised in Table \par\noindent\hangindent 20 pt{tab_stack} (see also text): Galaxies selected in the redshift, stellar mass and SFR ranges are shown on the top, middle and bottom panels, respectively. Sources individually detected in the 1.2-mm map at S/N$>3.5$ are not included in the stacks. The images shown are $3.6''\times3.6''$ in size. Solid white and dashed black contours represent the positive and negative signal, respectively. Contours start at $\pm2\sigma$ in steps of $\pm1\sigma$.}\label{fig_stacking12}
\end{figure*}
Recently, \citet{hatsukade15} studied the properties of four 1.3-mm detected sources with fluxes $S_{\rm 1.3mm}>0.2$ mJy (at least two times brighter than our sources). They find that these four galaxies are in the main-sequence, with redshifts $z=1.3-1.6$. However, those sources were selected in fields where these faint millimeter emitters were not the primary target. Most of these continuum sources lie in a dense environment at $z\sim1.3$, and it is thus unclear how representative their redshift and properties is of the field population.
All the sources shown in Fig. \par\noindent\hangindent 20 pt{fig_msz} lie within the uniform sensitivity region of our 1.2-mm mosaic, within PB $=0.4$. However, there are a few of them that were not detected in the 1.2-mm continuum even though they have similar SFRs and stellar masses than the detected sources. This could partly be attributed to uncertainties in the SED fitting procedure or to the fact that some galaxies would be located at the very edges of our mosaic. However, it is also possible the non-detection of these sources could also be due to differences in the individual physical properties of these sources. For instance, galaxies with lower dust temperatures or masses would tend to have lower fluxes at 1.2-mm, or they could just be dust poor. In \S \par\noindent\hangindent 20 pt{sec_stack} below we address this issue using stacking analysis.
\begin{table*}[ht]
\label{tab_stack}
\caption{Results from the stacking analysis. Columns: (1) Sample name; (2) Selection imposed for this sample. In all cases, we excluded the individually detected sources with $>3.5\sigma$. We limited the samples to have $M_*>10^9\ M_\sun$, to be located within PB=0.4 and to lie $3.5''$ away from the five most significant 1.2-mm continuum detections to avoid contamination. Additionally, in order to reject non-star forming sources in our stacks (i.e. old passive evolving galaxies), we restricted the samples to reside above the main-sequence including its intrinsic scatter at the relevant redshift range (i.e. sources above MS-0.5 dex), using the calibrations from \citet{whitaker14}. Only sources with $m_{\rm F850LP}$ and $m_{\rm 160W}<27.5$ mag AB were included, in order to retain sources with good SED fits; (3) Median redshift of the selected sample; (4) Median SFR obtained from the optical/near-infrared photometry with \texttt{MAGPHYS}; (5) Median stellar mass obtained from the optical/near-infrared photometry with \texttt{MAGPHYS}; (6) Number of objects that entered the stack; (7) Average flux density at 1.2-mm obtained from the stacking procedure.}
\centering
\begin{tabular}{lcccccc}
\hline
Sample $^a$& Selection$^b$ & $z_{\rm med}$ $^c$ & log$_{10}$(SFR$_{\rm UV, med}$) $^d$ & log$_{10}$($M_{*, med}$) & $N_{\rm obj}$ $^e$ & $S_{\rm 1.2mm}$ $^f$\\
& & & ($M_\sun$ yr$^{-1}$) & ($M_\sun$) & & ($\mu$Jy) \\% &
(1) & (2) & (3) & (4) & (5) & (6) & (7) \\%& (8) \\
\hline
z1 & $z=0-1$ & $0.76\pm0.19$ & $0.84\pm0.52$ & $9.78\pm0.49$ & 12 & $<13$ \\%& $<-0.09$ \\
z2 & $z=1-2$ & $1.22\pm0.20$ & $0.52\pm0.63$ & $9.45\pm0.43$ & 11 & $12\pm4$ \\% & $-0.14\pm0.13$ \\
z3 & $z=2-4$ & $2.45\pm0.41$ & $0.75\pm0.36$ & $9.48\pm0.31$ & 15 & $<13$ \\%& $<-0.12$ \\
\hline
m1 & log$_{10}(M_*/M_\sun)=9.0-9.5~~$ & $1.63\pm0.80$ & $0.46\pm0.35$ & $9.25\pm0.14$ & 21 & $<8$ \\%&
m2 & log$_{10}(M_*/M_\sun)=9.5-10.0$ & $1.29\pm0.95$ & $0.93\pm0.35$ & $9.78\pm0.12$ & 12 & $11\pm3.0$ \\%
m3 & log$_{10}(M_*/M_\sun)>10.0$ & $1.10\pm0.79$ & $1.00\pm0.51$ & $10.2\pm0.22$ & 9 & $19\pm5$ \\%&
\hline
s1 & log$_{10}$(SFR)$={0.5-1.0}\ $M$_\sun$ yr$^{-1}$ & $1.67\pm0.93$ & $0.70\pm0.16$ & $9.58\pm0.36$ & 17 & $<12$ \\%& $<-0.03$ \\
s2 & log$_{10}$(SFR)$={1.0-1.5}\ $M$_\sun$ yr$^{-1}$& $2.45\pm0.78$ & $1.02\pm0.13$ & $9.81\pm0.27$ & 6 & $<15$ \\%& $<-0.06$ \\
s3 & log$_{10}$(SFR)$>1.5\ $M$_\sun$ yr$^{-1}$& $1.05\pm0.48$ & $1.73\pm0.21$ & $10.2\pm0.36$ & 5 & $25\pm8$ \\%& $0.26\pm0.14$ \\
\hline\hline
\end{tabular}
\end{table*}
\section{Stacking analysis}
\label{sec_stack}
We use the stacking analysis to investigate the nature of the fainter galaxy population not detected at the achieved sensitivity limit of our ALMA 1.2-mm mosaic. To perform the stacking, we extract smaller images, $9''\times9''$ in size, from the final clean ALMA 1.2-mm continuum mosaic, centered at the position of sources that were selected from an independent galaxy catalog (see below). Sub-images of the same size are simultaneously extracted from the PB sensitivity mosaic map. All these sub-images are then combined together, to construct a weighted average using the PB sensitivity map as the weight. The noise in this average image is then obtained from an annulus around the central position with an initial and final radius of 4 and 12 pixels, respectively (1 pixel = $0.3"$). A summary of the stacking analysis results is shown in Fig.~\par\noindent\hangindent 20 pt{fig_stacking12}, and listed in Table \par\noindent\hangindent 20 pt{tab_stack}.
\subsection{Nature of undetected galaxies}
Using stacking, we first investigate the emission from galaxies individually undetected at the $3.5\sigma$ level in the ALMA 1.2-mm continuum map as a function of redshift. If these galaxies were to follow a similar redshift distribution as the detected galaxies, then we would expect on average that the galaxies in the $1<z<2$ range would have more 1.2-mm continuum emission than those in other redshift ranges. Figure \par\noindent\hangindent 20 pt{fig_stacking12} shows the stacked emission of galaxies in 3 different redshift ranges (samples z1, z2 and z3; see Table \par\noindent\hangindent 20 pt{tab_stack}). All samples have been selected to have $M_*>10^9\ M_\sun$ and $z<4$, and sources that enter the stack were required to lie $3.5''$ away from the location of the five most significant individual continuum detections to avoid contamination. The restriction to have a relatively high stellar mass is specifically to not down weight the stack signal. To avoid including passive evolving galaxies with no star formation activity in the stacks, we only select galaxies that are located within and above the main sequence (see Fig. \par\noindent\hangindent 20 pt{fig_msz}), taking into account a conservative 0.5 dex of scatter in the main sequence relationship. The main-sequence trends as a function of redshift are taken from \citet{whitaker14}. Additionally, to limit our sample only to galaxies with good measured SED fits, we require that the sample galaxies have magnitudes brighter than 27.5 AB in the F850LP and F160W bands. Galaxies detected at the $>3.5\sigma$ level in the 1.2-mm continuum have been excluded from the stacked samples. Using this selection, we only detect 1.2-mm emission from galaxies at $1<z<2$ (the z2 sample). In all the other redshift samples, we do not find significant emission and thus place $3\sigma$ limits on the 1.2-mm flux densities (see Table \par\noindent\hangindent 20 pt{tab_stack}). This implies that most of the underlying millimeter emission that is not directly detected in our ALMA continuum map, comes from galaxies located at similar redshifts as the individually detected galaxies, which have matching redshift distribution with a median $z=1.65$.
To shed light on whether the most massive or star-forming galaxies could have underlying 1.2-mm emission, we stack on different galaxy samples split in stellar mass and SFR. We use three samples divided by stellar mass and three samples divided by SFR (see Table \par\noindent\hangindent 20 pt{tab_stack}). We apply the same restrictions than for the redshift samples, including the limit in stellar mass, the requirement that the galaxies lie within and above the main sequence and the magnitude limit in the optical/near-infrared bands. The galaxies used in these stacks are represented by blue symbols in Fig. \par\noindent\hangindent 20 pt{fig_msz} (this Fig. does not show galaxies at $z<1$ and $z>3$).
Figure \par\noindent\hangindent 20 pt{fig_stacking12} (middle and bottom panels) shows the results of this exercise. From the three stellar mass samples, only the samples m2 and m3 present a tentative detection of the stacked 1.2-mm emission. For sample m1, we place a $3\sigma$ upper limit. This indicates that less massive galaxies have fainter millimeter continuum emission. Note that the stacked detection for the m2 sample is offset from the center, being unclear the reason for this shift since we are excluding sources near the most significant 1.2-mm sources. It is possible this shift is related to the low S/N of the signal.
By stacking in samples that were selected based on their UV-SFRs (derived from SED fitting), we find a clear detection for the s3 sample, which includes all galaxies with SFR$_{\rm UV}>30\ M_\odot$ yr$^{-1}$. This is consistent with the detection of emission in the mass-selected samples m2 and m3, which have a concordantly high median UV-derived SFRs. Note that most of the galaxies individually detected at 1.2-mm comply with the s3 sample selection. Thus, the detection of stacked continuum signal in the s3 sample implies that the individually undetected galaxies are just below the detection threshold of our survey, showing on average lower millimeter emission than the individually detected galaxies. The reason for this could be due to uncertainties in the derived stellar masses and SFRs, as well as different physical properties such as lower dust content (lower dust masses).
In summary, we find that most of the millimeter continuum emission of undetected galaxies is produced by galaxies in the redshift range $z = 1-2$ (sample z2). When we make stacks on stellar mass, we obtain detections for the stellar mass ranges $10^{9.5-10.0}\ M_{\sun}$ and $>10^{10}\ M_{\sun}$ (samples m2 and m3). These stellar mass bins have median UV-derived SFRs in the range of $\sim(3-30)\ M_\sun$ yr$^{-1}$. When we explicitly consider galaxy samples with UV-derived SFRs, we only obtain a detection for galaxies with SFRs $>30 \ M_\sun$ yr$^{-1}$ (but not for the $10-30\ M_\sun$ yr$^{-1}$ bin). These stacked detections reach down to 1.2-mm continuum fluxes of $\sim$10 $\mu$Jy.
\subsection{Stacking in the 3-mm continuum}
\label{sec_3mmstack}
Since there is only one significant source in the 3-mm continuum map, we use the stacking analysis to measure the average 3-mm emission from all the sources that were detected at $>3.5\sigma$ in the 1.2-mm map. The result of this procedure is shown in Fig. \par\noindent\hangindent 20 pt{fig_stack3mm}. Including all the 1.2-mm sources in the stack, we find an average flux density of $S_{\rm 3mm, all}=12\pm3\ \mu$Jy. Masking the individually detected source in the 3-mm map, we find an average flux density of $S_{\rm 3mm, masked}=9\pm3 \mu$Jy. Using the same stacking procedure and adopting the same samples on the 1.2-mm map (i.e. stacking the 1.2-mm detected sources to obtain the average 1.2-mm flux), we find $S_{\rm 1.2mm, all}=195\pm11\ \mu$Jy and $S_{\rm 1.2mm, masked}=125\pm12\ \mu$Jy, respectively.
The ratio between these measurements can now be used to obtain an estimate of the dust emissivity index $\beta$. We use a single-component modified black body dust model in the optically thin regime of the form $S_\nu\propto (1-e^{-\tau_\nu}) B_\nu(T_{\rm d})$ (see Weiss et al. 2007), where $S_{\nu}$ is the observed flux density, $B_{\nu}$ is the Planck function, and $T_{\rm d}$ is the dust temperature. It can be shown that in the Rayleigh-Jeans (RJ) limit,
\begin{equation}
\beta = \frac{log(\frac{S(\nu_1)}{S(\nu_2)})}{log(\frac{\nu_1}{\nu_2})} - 2,
\end{equation}
where $S(\nu_1)$ and $S(\nu_2)$ are the flux densities measured at the frequencies $\nu_1$ and $\nu_2$, respectively. Note that at the observed frequencies it is valid to assume the optically thin and RJ approximations.
For the galaxy individually detected in the 1.2-mm and 3-mm maps (ASPECS C1), we find $\beta=1.3\pm0.2$. For the stack sample that includes all the sources, we find $\beta=1.1\pm0.3$. Similarly, for the masked sample we find $\beta=0.9\pm0.4$. This result suggests a significantly lower dust emissivity index for the faint population of DSFGs than what has been typically found in galaxies in the local Universe and the Milky Way, and also at high-redshift, with $\beta$ ranging from 1.5 to 2.0 \citep[e.g.,][]{chapin09,dunne11,draine11, planck11a}. Note that given the relatively small beam size of the 1.2 mm observations, we could be missing flux that could contribute to a larger $\beta$ value. Similarly, the stacked signal detected at 3-mm is weak, and its detection is thus marginal. Both issues could thus be affecting this result. Another possible cause for this low $\beta$ value is the fact that we are tracing fluxes at wavelengths that could receive contribution from free-free emission. This would tend to increase the flux at 3-mm, resulting in larger $\beta$. Finally, it is worth mentioning that due the higher CMB temperature with redshift, we would expect to see an increase in the average $\beta$ value with increasing redshift. Larger samples of faint DSFGs are needed to provide better constraints on this subject.
\begin{figure}
\centering
\includegraphics[scale=0.28]{stack_3mm_all_sn35_7sources.ps}
\includegraphics[scale=0.28]{stack_3mm_all_sn35_7sources_masked.ps}
\caption{Stacked 3-mm emission at the location of the 1.2-mm detected sources ($15"\times15"$ in size). The left panel shows the stacked map when including all sources. The right panel shows the stacked map when including all but the brightest 1.2-mm source, which was also individually detected at 3-mm. White and black contours represent positive and negative emission, respectively. The contours are shown in steps of $\pm1\sigma$ starting at $\pm2\sigma$.}\label{fig_stack3mm}
\end{figure}
\section{ISM properties}
\label{sec_ism}
\subsection{Gas masses from dust, and caveats}
\begin{figure}[t]
\centering
\includegraphics[scale=0.4]{UDF_Mgas_SFR_new.eps}
\caption{ISM mass versus SFR for the ALMA UDF 1.2-mm continuum sources, compared to different galaxy populations that have been detected in CO(1--0) or CO(2--1) from the literature. The ISM mass for the ALMA sources have been computed using the 1.2-mm continuum flux densities following the recipies from \citet{scoville14}. Literature values typically assume a CO luminosity to gas mass conversion factor of 0.8 $M_\odot$ (K km s$^{-1}$ pc$^{2}$)$^{-1}$ for local starburst galaxies and SMGs, and 3.6 or 4.6 (same units) for local spirals and main sequence galaxies at high-redshift. For the CO-based gas mass estimates in the three galaxies detected in CO line emission \citep[see Paper~IV;][]{decarli16b}, we use a conversion factor of 3.6 (same units). For clarity, the magenta lines connect the 1.2-mm continuum and CO-based gas mass estimates. The dashed and dotted lines denote the two sequences of starbursts and main-sequence galaxies defined in \citet{daddi10b}, respectively.}\label{fig_sflaw}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[scale=0.55]{UDF_tdep_fgas_z_new.eps}
\caption{Evolution of the gas depletion timescale ($t_{\rm dep}$) and the molecular gas fraction ($f_{\rm gas}$) as a function of redshift for the ALMA UDF 1.2-mm continuum sources, compared to main sequence galaxies from the literature. Stellar masses and SFRs are computed from SED fitting. The ISM mass for the ALMA sources have been computed using the ALMA 1.2-mm continuum flux density following Scoville et al. (2014). In the top panel, the blue shaded region represents the expected evolution for the gas depletion timescale, $t_{\rm dep}=1.5\times(1+z)^{\gamma}$ with $\gamma=-1.0$ to -1.5, for massive main sequence galaxies \citep{dave12, tacconi13,saintonge13}. The pink region represents the typical gas depletion timescales measured in starburst galaxies \citep[e.g.][]{aravena16a}. In the bottom panel, the blue shaded region represents the evolution of the gas fraction expected for main sequence galaxies with $M_*>10^9\ M_\sun$ following the derivation of \citet{saintonge13}.}\label{fig_tdep}
\end{figure}
A useful method to compute ISM masses in galaxies has been the use of the dust mass as a proxy for the ISM content \citep{leroy11,magdis11,magnelli12,scoville14, genzel15}. Recently, \citet{scoville14} argued that under reasonable assumptions about the dust properties, reliable ISM mass measurements can be made based on flux measurements made in the RJ tail of the dust. The method was calibrated using massive galaxies at low and high redshift and assuming a fixed gas-to-dust ratio, which is expected to be fairly constant for a relatively ample range in properties \citep[see ][ for details]{scoville14}, and assumes a fixed dust temperature of $T_{\rm d}=25$ K. Note that there is a weak dependance of this method on $T_{\rm d}$, since we are probing the RJ part of the spectrum. Following \citet{scoville14}, we compute the ISM mass in units of $10^{10} M_\sun$ as:
\begin{equation}
M_{\rm ISM}=1.2 (1+z)^{-4.8} (\frac{\nu_{\rm obs}}{350})^{-3.8} \frac{\Gamma_0}{\Gamma_{\rm RJ}} S_\nu D_{\rm L}^2,
\end{equation}
where $D_{\rm L}$ is the luminosity distance in Gpc at redshift $z$, and $S_\nu$ is the measured flux density in mJy at the observing frequency $\nu_{\rm obs}$ (in GHz). $\Gamma_{\rm RJ}$ is a correction factor that takes into account the deviation from the RJ limit as we approach higher redshifts. This factor depends on $z$, $T_{\rm d}$ and $\nu_{\rm obs}$, and becomes $\Gamma_0=0.76$ at $z=0$ for $\nu_{\rm obs}=$ 242 GHz and $T_{\rm d}=25$ K. This method to compute ISM masses assumes a dust emissivity index $\beta=1.8$, which we use throughout for consistency with other studies.
\texttt{MAGPHYS} also delivers an estimate of the dust mass ($M_{\rm d}$) using the median of the dust mass posterior probability when fitting the available photometry. From this dust mass estimate, and under the assumption of a fixed gas-to-dust ratio $(\delta_{\rm GDR})$ and that the ISM is mostly molecular, one can obtain a measurement of the gas mass as $M_{\rm gas}=\delta_{\rm GDR} M_{\rm d}$. For local galaxies it has been found that typically, $\delta_{\rm GDR}\sim72$ \citep{sandstrom13}, however metallicity-dependent variations are likely to play a significant role \citep[e.g., ][]{remyruyer14}. For the typical stellar masses of our sources ($\sim10^{10-11} M_\odot$) and assuming that local calibrations apply, we would expect metallicities close to the solar value, 12+log(O/H) $\sim9$ \citep{tremonti04}. However, since the metallicities are lower at high redshift, the typical stellar masses of our sample imply metallicities of $\sim8.4$ at $z\sim1.5$ \citep{yabe14, zahid14}. This metallicity value would translate into $\delta_{\rm GDR}\sim200$ \citep{remyruyer14}. Hence, we adopt this value to convert the dust masses obtained with \texttt{MAGPHYS} into gas mass estimates.
\citet[][; Paper~IV]{decarli16b} provide a detailed discussion of the different available methods to compute the gas masses, based on the CO measurements for four sources in the ASPECS field. From Table \par\noindent\hangindent 20 pt{tab_prop}, we find that the gas masses obtained using \texttt{MAGPHYS} SED fitting are consistent with the ISM estimates from the Scoville et al. method for the assumed $\delta_{\rm GDR}$. Decarli et al. (2016b; Paper~IV) finds that the gas estimates following Scoville et al. and the \texttt{MAGPHYS} SED fitting methods under-predict the gas masses by a factor of $\sim3-4$ compared to the CO based estimates. There are several reasons that could explain this discrepancy, including (i) a combination of high excitation and low $\alpha_{\rm CO}$ values in the CO measurements, (ii) systematics in the calibration of the dust-based measurements, and (iii) different spatial distributions of dust and molecular gas within individual galaxies (see Paper~IV for details). Another important issue is that the \citet{scoville14} calibration uses a fixed $\delta_{\rm GDR}$ value assuming solar metallicity. This assumption is reasonable for massive galaxies ($\sim10^{11}\ M_\sun$) as applied in their study, however, it may potentially underestimate the gas masses for less massive, lower metallicity galaxies, for which a higher $\delta_{\rm GDR}$ should be used.
Most importantly, perhaps, is the fact that the \citet{scoville14} calibration uses a gas to dust ratio fixed value for a solar metallicity. This assumption is reasonable for massive galaxies as applied in their study ($\sim10^{11}\ M_\sun$), however, it will likely result in lower gas masses for less massive, lower metallicity galaxies for which a higher $\delta_{\rm GDR}$ should be used.
Despite these uncertainties, the dust-based estimates constitute the only means to provide a measurement of the gas masses in our 1.2-mm continuum detected sources, given that most of them do not have CO line detections. Table \par\noindent\hangindent 20 pt{tab_prop} lists the gas masses obtained using both the Scoville et al. and the \texttt{MAGPHYS} SED fitting method. In what follows we only use the ISM masses obtained with the Scoville et al. method as a measure of the total molecular gas mass, under the assumption that most of the ISM of high-redshift galaxies is in the form of molecular gas.
\subsection{Gas depletion timescales and fractions}
Figure \par\noindent\hangindent 20 pt{fig_sflaw} shows the ISM mass (using Scoville et al. method) versus SFR (derived using SED fitting) for the galaxies detected at 1.2-mm continuum emission in our survey. For comparison, we also show the gas masses and SFRs of literature sources that have been detected in CO emission. To avoid uncertainties due to gas excitation, we only chose literature sources with low-$J$ CO measurements. We use a $^{12}$CO to gas mass conversion factor $\alpha_{\rm CO}=0.8$ K km s$^{-1}$ pc$^2$ for the samples of ultra-luminous IR galaxies (ULIRGs; Solomon et al. 1997) and both unlensed \citep{riechers11b, riechers11c,ivison11,ivison13,frayer08,thomson12,carilli11,hodge13,bothwell13,walter12,combes12,coppin10,debreuck14} and lensed DSFGs \citep{ivison10,lestrade11,swinbank10,harris10,decarli12,harris12,fu12,aravena16a}. For the samples of local spirals (Leroy et al. 2008) and main sequence galaxies \citep{daddi10a,magdis11,magnelli12}, we use $\alpha_{\rm CO}=4.6$ and 3.6 K km s$^{-1}$ pc$^2$, respectively. For reference, we also show the available CO-based gas mass estimates for the three 1.2-mm continuum sources in our sample that were detected in CO line emission \citep[C1, C2 and C6;][; Paper~IV]{decarli16b}. For these, a conversion factor of 3.6 K km s$^{-1}$ pc$^2$ has been used.
Our galaxies seem to span a significant range in ISM masses and SFRs. Two of our ALMA 1.2-mm sources appear to be aligned with the sequence formed by the local spirals and main-sequence galaxies at $z=1-2$ defined by the dashed line \citep{daddi10b}. This includes two of the CO detected galaxies, which are also detected in continuum. In particular, the 1.2-mm brightest galaxy in our sample falls into the group of main-sequence galaxies, supporting the identification of this galaxy as main sequence based on SFR--$M_*$. Only one galaxy, the third brightest in our continuum sample, is clearly located in the starburst regime. Four other sources appear to lie in between the trends of starburst or main-sequence galaxies. We remark that the gas mass values derived from the 1.2-mm fluxes could be underestimated as discussed in the previous section. This would thus imply that these four sources in our sample could belong to the trend of main sequence galaxies.
We note that the fact that the starburst and main-sequence galaxy trends in this SFR--$M_{\rm gas}$ plane appear to be well separated from each other, with virtually no source lying in between, partly relies on the use of fixed $\alpha_{\rm CO}$ factors for each particular sample. While in several cases, the $\alpha_{\rm CO}$ conversion factor has been measured directly for the literature sources, we caution that the use of a binary set of values for this parameter may artificially lead to different star formation laws for starbursts and main-sequence galaxies \citep{ivison11}. The $\alpha_{\rm CO}$ factor depends on several parameters including metallicity, gas temperature and velocity dispersion and should depend on individual galaxy properties such as the gas or SFR surface density \citep[see ][]{casey14}. Furthermore, the bi-modality might be in part caused by the pre-selection of individual sources for CO follow-up which biases the range of properties covered by targeted current observations. However, it should be pointed out that this separation is already seen when comparing the direct observables $L'_{\rm CO}$ and $L_{\rm IR}$ \citep[e.g.,][]{daddi10a,genzel10,aravena16a}.
Figure \par\noindent\hangindent 20 pt{fig_tdep} shows the implied gas depletion timescales ($t_{\rm dep}$) and gas fractions ($f_{\rm gas}$) as a function of redshift for our ALMA 1.2-mm continuum sources, compared to recent measurements of main-sequence galaxies at $z=0.5-3.0$ \citep{geach11,tacconi13,saintonge13}. Observations of massive main-sequence galaxies ($M_*>10^{10}\ M_\sun$) have shown evidence for a significant dependency of $t_{\rm dep}$ out to $z=3$ \citep{tacconi13,saintonge13,genzel15}, consistent with models of galaxy formation. These studies show a dependency of $t_{\rm dep}$ with redshift with the form $(1+z)^\gamma$, with $\gamma$ varying between -1.5 to -1.0 \citep{tacconi13}, as shown in Fig. \par\noindent\hangindent 20 pt{fig_tdep}. Recent studies, however, show that $\gamma$ can be as low as -0.3 \citep{genzel15}. Similarly, as shown in the bottom panel of Fig. \par\noindent\hangindent 20 pt{fig_tdep}, $f_{\rm gas}$ shows a significant dependency with redshift, which appears to flatten at $z>3$ \citep{saintonge13}.
The gas depletion timescales for our faint 1.2-mm sources is consistent with the ranges found for main sequence galaxies at similar redshifts. Only one galaxy has a $t_{\rm dep}$ value that puts it clearly in the range occupied by starburst galaxies. However, our galaxies present gas fractions ranging from $0.06-0.2$ for the $z\sim1.5$ sample, which significantly lower than other main sequence galaxies at similar redshifts. Only the higher redshift galaxy in our sample, ASPECS C1 at $z=2.5$, has a value of $f_{\rm gas}$ comparable to literature sources at its redshift. This implies that while most of our galaxies have measured gas depletion timescales that agree with previous studies for main sequence galaxies, they have gas fractions that are much lower than the those found for same comparison samples.
Several factors could affect the measured $t_{\rm dep}$ and $f_{\rm gas}$. This can partly be attributed to uncertainties in the derived parameters through SED fitting. However, we are using very deep multi-wavelength photometry, and thus the derived SFRs and stellar masses should be as accurate as in previous studies. This is indicated by the fact that the ranges for the location of the main sequence at different redshifts in Fig. \par\noindent\hangindent 20 pt{fig_msz} are consistent with those from the literature \citep{whitaker14}. Another possible explanation is that the gas masses computed using the 1.2-mm flux densities are being underestimated. A factor of $\sim2-3$ higher gas masses, as those derived from CO \citep[see ][; Paper~IV]{decarli16b}, would place the measured gas fractions more in line with the expected values for main sequence galaxies, while retaining high gas depletion timescales.
Additionally, our sample presents significant scatter in both plots. This scatter is unlikely caused by the possible underestimation of the gas masses where we would expect a more systematic effect. In this case, our sources present a scatter that is consistent with the typical one found in other samples studied in CO emission \citep{geach11, tacconi13, saintonge13}. Because of this scatter and the relatively narrow redshift range covered by our ALMA detections, it is hard to establish any evolutionary trend with the available data.
\section{Contribution to the EBL at 1.2-mm}
\label{sec_cib}
\subsection{Integrated intensity and fraction of the EBL}
We use the number counts at 1.2-mm derived in Section \par\noindent\hangindent 20 pt{sec_counts} to calculate the contribution to the EBL at 1.2-mm. Although our source number counts are derived from a small area of the sky, they are based in a deep contiguous blank field.
To calculate the contribution to the cosmic background at 1.2-mm from our measurements, we directly integrate the number counts, corrected for fidelity and completeness, down to the faintest flux bin ($S_{\rm 1.2mm}\sim37\mu$Jy). We obtain an integrated intensity of $7.8\pm0.4$ Jy deg$^{-2}$. The uncertainty is derived from the sum of the uncertainties of the individual detections, corrected for fidelity and completeness. However, our number counts do not extend to fluxes above 0.6 mJy. To estimate the contribution of the bright-end of the number counts, which are not traced by our survey, we use the results from from \citet{karim13} and \citet{oteo15}. While the \citet{karim13} results are measured at 870$\mu$m, we chose them since they are based on ALMA high resolution observations and thus take better into account the multiplicity and false detection rate issues seen in single-dish telescope bolometer surveys. It is a well known result from their study that bolometer surveys overpredict the number counts at the bright end (above $S_{870\mu{\rm m}}>6$ mJy). We convert their counts from 870$\mu$m to 1.2-mm using $S_{\rm 1.2mm}=0.4\times S_{\rm 870}\mu{\rm m}$, and add their contribution by integrating the values in their Table 1. Similarly, we use the \citet{oteo15} results to account for the contribution to the integrated intensity between 1.2-mm fluxes of 0.6 to 1.9 mJy, which are not covered by either the Karim et al. or our measurements. To fill this gap, we extrapolate the Oteo et al. number counts (in log-log space). By adding up the contribution of all galaxies starting at our faintest flux bin, we find that an integrated intensity of $8.6\pm0.7$ Jy deg$^{-2}$.
To compute the CIB at the frequency of our observations, we make use of the latest values derived by \citet{planck14}. By interpolating the {\it Planck} measurements (see their Table 10) over the frequency range of our observations (212-272 GHz), we find an EBL at $\sim242$ GHz of $14.2\pm0.6$ Jy deg$^{-2}$. From this, we find that our number counts recover $\sim60\pm6\%$ of the EBL at 242 GHz. Note that the EBL value at 242 GHz measured by {\it Planck} is much more precise than that measured by COBE 20 years ago, and we thus adopt this value.
In order to account for the missing contribution to the EBL, we use stacking analysis. We follow the procedure explained in \S \par\noindent\hangindent 20 pt{sec_stack}. We select the same samples (see Table \par\noindent\hangindent 20 pt{tab_stack}), but in this case we limit them to exclude all sources with a detection at the $>3\sigma$ level in order to be consistent with the faintest flux level taken into account to derive the number counts. In all cases, the samples differ by at most two sources with respect to those listed in Table \par\noindent\hangindent 20 pt{tab_stack}. Hence, we find similar results than those presented in \S \par\noindent\hangindent 20 pt{sec_stack}. We thus use the fluxes and number of objects for the m2 and m3 samples to compute the integrated intensity from the faintest, undetected sources. We find an extra contribution of $2.8\pm0.5$ Jy deg$^{-2}$ or $\sim20\pm4\%$ of the EBL at 242 GHz. Combining this to our measurement from the number counts, implies a total intensity of $11.4\pm0.8$ Jy deg$^{-2}$, which makes up $80\pm7\%$ ($\sim77--84\%$) of the EBL at 242 GHz measured by {\it Planck}.
\subsection{Nature of the sources that make up the EBL}
A critical result from this study corresponds to the properties of the galaxies that contribute to the EBL at 242 GHz. Based on our number count measurements only, we obtained an integrated intensity of $7.8\pm0.4$ Jy deg$^{-2}$. This makes up $55\pm4\%$ of the EBL measured by {\it Planck} at 242 GHz, implying that the population of galaxies that dominates this background is composed by the galaxies individually resolved by our ASPECS survey. From \S \par\noindent\hangindent 20 pt{sec_prop}, we determined that these galaxies have typical stellar masses of $\sim4\times10^{10}\ M_\sun$, SFRs of $\sim40~M_\sun$ yr$^{-1}$ at $z\sim1.7$, which corresponds to the main sequence at this redshift. This is supported by the ISM masses of these galaxies, which places them in the star-forming sequence in the $M_{\rm ISM}$ vs SFR plane. By using stacking, we find that on average the galaxies that make up another 20\% of the EBL at 242 GHz, at the faintest end, is composed by slightly less massive galaxies ($\sim(0.5-1.5)\times10^{10}\ M_\sun$) and low SFRs ($10-20\ M_\sun$ yr$^{-1}$) at similar redshifts. These findings imply that the bulk of galaxies that make up the CIB consists of faint, main-sequence galaxies at $z\sim1.7$.
Our measurements indicate that $\sim77-84\%$ of the EBL at 242 GHz can be resolved by individually detected galaxies, by those identified by stacking (in the m2+m3 samples). If we use the upper limit in the mass bin m1, we find that these galaxies could contribute up $6\%$ of the EBL at 242 GHz ($3\sigma$). This implies that up to $84\%+6\% = 90\%$ of the EBL could be identified by our observations (plus literature for the bright end), and hence only about 10\% of the EBL measured by {\it Planck} at this frequency is left unresolved. Since we have included the most massive samples in our stacking, $M_{\star}>10^9 M_\sun$, the remainder of the EBL at these frequencies would likely come from less massive galaxies ($M_{\star}<10^9 M_\sun$).
\subsection{The effect of cosmic variance}
A number of recent studies have used the archival ALMA 1.2-mm data to provide constraints on the EBL at 1.2-mm. These studies measure significantly higher integrated intensities at 1.2-mm compared to our estimates: \citet{fujimoto16} measure the number counts down to a flux limit of 15 $\mu$Jy, just below our ALMA UDF flux limit, with an integrated intensity of $\sim22$ Jy deg$^{-2}$; \citet{hatsukade13} integrated their number counts down to 0.15 mJy, obtaining an intensity of $\sim16.9$ Jy deg$^{-2}$ (converting their measurement from 1.3-mm to 1.2-mm); \citet{ono14} measures $\sim11$ Jy deg$^{-2}$ down to 0.1 mJy; similarly, \citet{carniani15} measures $\sim17$ Jy deg$^{-2}$ down to 0.1 mJy at 1.2-mm. To derive the fraction of the EBL at 1.2-mm resolved, most of these literature results use early measurements from the Far Infrared Absolute Spectrophotometer (FIRAS) on board of the COBE satellite \citep{fixsen98}, which measures an integrated intensity of $22_{-8}^{+14}$ Jy deg$^{-2}$ at this wavelength. However, the COBE spectrum of the IR background becomes highly uncertain at frequencies below 350 GHz (see Fig. 4 of Fixsen et al.), mostly due to Galactic contamination. The newer measurement from the {\it Planck} satellite has much better precision and is within the uncertainties of the COBE measurement. As such, the recent measurements from the literature imply very high resolved fractions of the EBL, in some cases even exceeding the {\it Planck} measurements at 242 GHz. We note that the EBL is a grand average of the extragalactic emission over the whole sky. Therefore measurements covering $\sim$1 arcmin$^{-2}$ or less of the sky, aiming to resolve the sources contributing to this background will be most likely highly affected by cosmic variance. If the observations were pointed to an overdense region of the sky, this will translate into a higher number of sources and higher resolved fraction of the EBL. In particular, Fig. \par\noindent\hangindent 20 pt{fig_ncounts} shows that for the flux range $0.08-0.6$ mJy our cumulative number counts are significantly below, by a factor of $\sim2$, with respect to the values derived by \citet{hatsukade13} and \citet{fujimoto16}\footnote{Over this flux range, the Fujimoto et al. results fully rely on the observations analysed by Hatsukade et al. Thus, these studies measure effectively the same number of sources.}, yet more consistent with the counts derived by \citet{oteo15} and \citet{carniani15}. This substantial difference in the number counts, possibly due to the small areas covered but also to the fact that these studies are not ``blank-field'', would explain the differences in the measured intensities and resolved fraction of the EBL between different studies. As shown in \citet{scoville13}, small scale source density variations can cover significant fractions of the sky (see their Figs. 9-11). As explained in \S \par\noindent\hangindent 20 pt{sec_counts}, the number count differences might also be due to different methods and analysis tools used. In any case, measurements on larger fields will help to elucidate the effect of small scale structure on the EBL at millimeter wavelengths.
\begin{figure}
\centering
\includegraphics[scale=0.4]{UDF_CIB_spectrum_new.eps}
\caption{Extragalactic infrared background spectral energy distribution compared to the amount of intensity resolved by our ALMA UDF observations. The shaded blue area represents the cosmic IR background revealed by the {\it Planck} satellite observations and the range on uncertainties in the measured data. Note that the uncertainty is so small that the shaded area resembles a thick line. The yellow circle shows the integrated intensity of our ASPECS observations at 242 GHz (11.4 Jy deg$^{-2}$) including both the measurement based on the number counts and the stacking analysis. The open circle shows the intensity recovered by the number count measurements only (without stacking). The green triangle shows the measurement made by \citet{fujimoto16} based on archival 1.2-mm data.}\label{fig_cib}
\end{figure}
\section{Conclusions/summary}\label{sec_concl}
Using ALMA in cycle-2, we have conducted a millimeter spectroscopic survey by scanning the full 3-mm and 1.2-mm bands over a region in the {\it Hubble} UDF. The collapsed cubes constitute the deepest continuum images ever obtained over an 1 arcmin$^2$ contiguous area of the sky. The main results of our continuum measurements can be summarised as follows:
\begin{itemize}
\item We detect nine sources with significances $>3.5\sigma$ at 1.2-mm and only one source at 3-mm. From these detections, we measure the 1.2-mm number counts over the flux density range $S_{\rm 1.2mm}=0.036-0.57$ mJy. Our number counts are similar to previous measurements, with differences within a factor of $\sim2$.
\item We measure the properties of the individually detected galaxies at S/N$>3.5$. We find that there is a large spread in stellar masses and SFRs, with median values of $4\times10^{10}\ M_\sun$ and $\sim40~M_\sun$ yr$^{-1}$, much lower than found in brighter SMGs. We find that these faint DSFGs are systematically located at lower redshifts than millimeter-selected SMGs, with a median redshift of $z=1.7$. All galaxies are consistent with being close to the main sequence at their respective redshift.
\item We use stacking analysis to estimate the average emission from samples of galaxies selected by redshift, stellar mass and SFRs. We only find detections in samples selected in the redshift range $1<z<2$, as well as in the stellar mass ranges log$(M_*/M_\sun)=9.5-10.0$ and log$(M_*/M_\sun)=10.0-10.5$, with typical SFRs of $3-10\ M_\sun$ yr$^{-1}$ . This suggests that the rest of the emission, not individually detected in our survey, comes from galaxies less massive, with lower SFRs, but at a similar redshift than the detected sources.
\item We use the 1.2-mm flux as a proxy for the ISM masses in our individually detected galaxies. We find that most of our sources are located in the star-forming trend occupied by main-sequence galaxies and local spirals, implying relatively large gas time depletion timescales, typically above 300 Myr, and a large spread in the molecular gas fractions ranging from 0.1 to 1.0. We compare these results to ISM mass estimates using CO as a tracer in \citet[][; Paper~IV]{decarli16b}.
\item Our individual detections alone are able to resolve $55\pm4\%$ of the EBL at 242 GHz measured by the {\it Planck} satellite. By adding up the integrated intensity from our number counts, to the contribution from the bright end of the number counts -- mostly composed by SMGs -- and the contribution of faint galaxies detected using stacking, we are able to resolve between 77--84\% of the CIB at 242 GHz. The typical properties of the population that makes up most of the EBL at these frequencies corresponds to that of the galaxies described in this work.
\end{itemize}
\acknowledgements
We thank the anonymous referee for her/his positive feedback and useful comments. M.A.~acknowledges partial support from FONDECYT through grant 1140099. FW, IRS, and RJI acknowledge support through ERC grants COSMIC--DAWN, DUSTYGAL, and COSMICISM, respectively. FEB and LI acknowledge Conicyt grants Basal-CATA PFB--06/2007 and Anilo ACT1417. FEB also acknowledge support from FONDECYT Regular 1141218 (FEB), and the Ministry of Economy, Development, and Tourism's Millennium Science Initiative through grant IC120009, awarded to The Millennium Institute of Astrophysics, MAS. EdC gratefully acknowledges the Australian Research Council as the recipient of a Future Fellowship (project FT150100079). DR acknowledges support from the National Science Foundation under grant number AST-1614213 to Cornell University. IRS also acknowledges support from STFC (ST/L00075X/1) and a Royal Society / Wolfson Merit award. Support for RD and BM was provided by the DFG priority program 1573 `The physics of the interstellar medium'. AK and FB acknowledge support by the Collaborative Research Council 956, sub-project A1, funded by the Deutsche Forschungsgemeinschaft (DFG). PI acknowledges Conict grants Basal-CATA PFB--06/2007 and Anilo ACT1417. RJA was supported by FONDECYT grant number 1151408. This paper makes use of the following ALMA data: ADS/JAO.ALMA\#2013.1.00146.S and ADS/JAO.ALMA\#2013.1.00718.S. ALMA is a partnership of ESO (representing its member states), NSF (USA) and NINS (Japan), together with NRC (Canada), NSC and ASIAA (Taiwan), and KASI (Republic of Korea), in cooperation with the Republic of Chile. The Joint ALMA Observatory is operated by ESO, AUI/NRAO and NAOJ. The 3mm-part of ALMA project had been supported by the German ARC.
\bibliographystyle{apj}
|
\section{Introduction}
In recent years much attention has been payed to the 24$\pm$3\,Myr \citep{Bell15} young \object{$\beta$ Pictoris} stellar association. Several studies have allowed
to increase significantly the number of confirmed members and to discover many more candidate members. Just to mention the most recent works, we refer the readers to \citet{Lepine09}; \citet{Kiss11}; \citet{Schlieder10, Schlieder12}; \citet{Shkolnik12}; \citet{Malo13, Malo14a, Malo14b}. The first comprehensive search for the rotation periods of the low-mass members of $\beta$ Pictoris was carried out by \citet{Messina10, Messina11} who retrieved a total of 38 low-mass members (i.e., spectral types from late F to M) from the earlier compilations of \citet{Zuckerman04}, \citet{Torres06}, and \citet{Kiss11}. Their study provided the rotation periods of 33 out of 38 members.\\
In the light of the mentioned studies, we have again explored up to the most recent literature and, at this time, we could finally
compile a new list of 117 stars \rm among members and candidate members with spectral types later than about F3V. Then, we have started a new rotational study on this enlarged sample. To get the photometric rotation periods of our targets, we used our own observations, archive data, and also we made use of periods from the literature.\\
As result of our photometric investigation, we obtained the rotation periods of 112 out of 117 stars. \rm Specifically,
we measured for the first time the rotation periods of 56 stars. \rm For another 27 stars, we could confirm with our analysis of
new or archived data the values reported in the literature. For 29 stars we adopted the literature values. For the remaining 5 stars, our periodogram analysis did not provide the rotation period. The results of this investigation are presented in the catalogue of photometric rotation period of the \object{$\beta$ Pictoris} association members \rm (Messina et al. 2016; Paper I) where we describe the photometric observations newly obtained, their reduction and analysis, and a detailed discussion of our results obtained for each individual star.\\
In this paper (Paper II), we focus on a sub-sample consisting of 66 members for which we know the rotation period and have one measurement at least (from the literature) of the Lithium equivalent width (EW). This sub-sample is to date the largest of any known young loose association. We will make use of it to investigate the correlation between rotation and Li depletion \rm and to compare with earlier results \rm the age of the $\beta$ Pictoris association obtained by the modeling of the Lithium Depletion Boundary (LDB), after decorrelating the Li EW from rotation. \rm
\section{Sample description}
The sample under analysis consists of 66 members with spectral
type later than F3V with one measurement, at least, of the Li EW (see Table 1). This sample, which has significantly
increased with respect to earlier studies, makes a new investigation on the age
of the $\beta$ Pictoris association with the LDB modeling method necessary. In fact, earlier studies
have made use of a smaller number of association members. \citet{Mentuch08} made use of a sample of 23 members;
\citet{Macdonald10} used a sample of only 10 members. \citet{Binks14} and
\citet{Malo14b} used, respectively, about 40 and 34 members. \\
Moreover, we also now know the rotation periods of all these 66 stars, which put us in the position to get more accurate results
and to deeply explore the mechanism of Li depletion. More specifically, we measured for the first time the rotation
periods of 30 out of 66 members. These new measurements will be presented in the mentioned
catalogue containing the photometric rotation periods of all 112 members/candidate members of the association.
For another 16 members, the rotation periods were taken from
\citet{Messina10, Messina11}. The remaining 20 rotation periods were taken from different literature sources.\\
This sample represents, to date, one of the three largest
samples ever used to investigate the correlation between rotation period and Li depletion. \rm The other samples consist of members of the \object{Pleiades} and of the \object{M34} stellar open clusters investigated by \citet{Gondoin14} using the rotation periods from \citet{Hartman10}, and of members of the \object{NGC\,2264} open cluster whose results on the Lithium-rotation connection are presented by \citet{Bouvier16}.
The earlier larger sample of Pleiades members analysed by \citet{Soderblom93} made use of the projected rotational velocity
($v\sin{i}$), which is less accurate owing to the uncertainty arising from the unknown $\sin{i}$.
\section{Li distribution}
\begin{figure*}
\begin{minipage}{20cm}
\includegraphics[scale = 0.6, trim = 0 0 0 0, clip, angle=90]{li_distri.eps}
\end{minipage}
\caption{\label{Li_distri} Distribution of Li EW versus V$-$K$_s$ color for confirmed members of the $\beta$ Pictoris association.
The meaning of symbols is given in the legend. Dotted vertical lines indicate the range of values for stars with multiple Li EW measurements. All circled symbols indicate stars hosting debris discs.}
\end{figure*}
In Fig.\,\ref{Li_distri}, we plot the distribution of Li EW for all 66 members versus the V$-$K$_s$ color. This color index is measured using the K$_s$ magnitude from the 2MASS project \citep{Cutri03} and the brightest (presumably unspotted) V magnitude from the long photometric time series available in the All Sky Automated Survey (ASAS) archive \citep{Pojmanski02}. This choice allows us to reduce the impact of photometric variability on the V-band magnitude. Among different color indexes, the V$-$K$_s$ turned out to be the most
homogeneous and that with the largest range of values, allowing an accurate representation of different masses. We applied no color correction for interstellar absorption since the members under study have distances in the range from 10\,pc to 80\,pc, \rm therefore the interstellar reddening can be considered negligible. The average uncertainty on this color is $\sigma_{\rm V-K_s}$ = 0.03\,mag. There are only 8 stars that belong to close binaries that are not resolved neither in V nor in K$_s$ magnitudes and consist of components differing by no more than 2 spectral sub-classes. For these stars the measured V$-$K$_s$ for the primary component, the one with measurement of Li EW, resulted to be redder than the expected value for the spectral type. In these cases we adopted the V$-$K$_s$ values from \citet{Pecaut13}, and corresponding to the spectral type of the primary component. \rm The Li EW measurements are retrieved from \citet{Mentuch08}, \citet{dasilva09}, \citet{Kiss11}, \citet{Binks14}, \citet{Moor13}, \citet{Malo14b}, \citet{Reid02}, \citet{Torres06}, and \citet{Desidera15}. The average uncertainty associated to these measurements is generally not larger than $\sim$40\,m\AA. The error bars associated to the measurements (in both axes) are equal to the size of the symbols plotted in Fig.\,\ref{Li_distri}. In a few cases more than one measurement is available. The range of values of the Li EW in these cases is plotted with a dotted line. Almost all measurements (for a total of 20 members) \rm in the color range 4.5 $<$ V$-$K$_s$ $<$ 5.3\,mag are upper limits.
Different symbols are used to distinguish the different nature of the considered members (see the legend in Fig.\,\ref{Li_distri}).
Detailed information on each member and the criterion to distinguish close from wide components is given in Messina et al. (Paper III, in prep.).\\
In Fig.\,\ref{Li_distri}, we note three important features. First, we note a very well defined LDB that will be used for an estimate of the age.
Second, in the color range 3.2 $<$ V$-$K$_s$ $<$ 4.5\,mag, we note, even by just a visual inspection, a scatter in the Li EW significantly larger than in the range of bluer members. We will investigate the role played by the rotation in producing this scattered depletion.
Finally, we note that in the range 2.5 $<$V$-$K$_s$ $<$ 3.0\,mag, we have only three stars that all have Li EWs significantly lower than the average. This color range corresponds to K2--K3 spectral types and is reminiscent of the Li dip observed in more evolved main-sequence stars \citep[see, e.g.,][]{Balachandran95}.
One of these stars is \object{HIP\,84586}, a SB2 spectroscopic binary whose tidal locking between the two components may have altered the internal mixing and, consequently, enhanced the Li depletion rate with respect to single stars. The other two stars are \object{TYC\,6878\,0195\,1} and \object{HIP\,11437}A, both components of wide binaries that are expected to have evolved, rotationally, as single stars. We only note that \object{HIP\,11437}A hosts a debris disc.
\section{The rotation - Li depletion correlation}
We first analyse the possible correlation between rotation and Li EW to investigate which role the rotation may play
in producing the observed Li EW dispersion among stars with similar masses. \\ \rm
This information will be used for a more accurate modeling of the LBD to infer the association age.
\begin{figure*}
\begin{minipage}{20cm}
\includegraphics[scale = 0.6, trim = 0 0 0 0, clip, angle=90]{li_fit_new.eps}
\end{minipage}
\caption{\label{li_fit} \it Top panel: \rm distribution of Li EW versus V$-$K$_s$ color in three color ranges: 0.5 $<$V$-$K$_s$ $<$ 3.4\,mag (19 stars)\rm, 3.4 $<$V$-$K$_s$ $<$ 4.5\,mag (21 stars)\rm, and 5.4 $<$V$-$K$_s$ $<$ 5.9\,mag (6 stars) \rm of $\beta$ Pictoris members with known rotation period and with overplotted (solid lines) the linear fits. \it Middle panel: \rm residuals from the fit with horizontal dotted lines indicating the 1$\sigma$ dispersion. \it Bottom panels: \rm distribution of Li EW residuals versus rotation period in the three color ranges. Solid lines are linear fits, r is the Pearson linear correlation coefficient and $\alpha$ the significance level.}
\end{figure*}
We analyse separately the following three sub-groups: blue stars with 0.5 $\le$ V$-$K$_s$ $<$ 3.4\,mag totaling 19 members, \rm red stars with 3.4 $\le$ V$-$K$_s$ $\le$ 4.5\,mag totaling 21 members, \rm and very red stars with 5.4 $<$V$-$K$_s$ $<$ 5.9\,mag totaling 6 members \rm that correspond to the onset of the Li depletion. Stars with V$-$K$_s$ $>$ 5.9\,mag have not undergone any Li depletion. In the blue stars sample, we observe a trend of increasing Li EW with increasing color (decreasing mass); in the red stars sample, we observe the opposite trend; in the very red star sample we observe again a trend of increasing Li EW with increasing color. As it is shown in the top panel of Fig.\,\ref{li_fit}, linear fits are enough for our purposes, i.e. approximating the mass dependence to outline the correlation with the rotation period. In the middle panel we plot the residuals with respect to the linear fits. Only one star (\object{HIP\,11437}A), represented by a square and significantly deviating from the trend, was excluded from the fit. Positive residuals indicate stars that are more Li rich than the average, whereas negative residuals indicate stars that are more Li-depleted than the average.
In the case of blue stars, we find a peak-to-peak amplitude of the dispersion $\sigma$ = 90\,m\AA\,\, of the residuals from the fit, whereas in the case of red stars the peak-to-peak amplitude of the dispersion is about twice as large, $\sigma$ = 180\,m\AA. Finally, in the case of very red stars we find $\sigma$ = 125\,m\AA.\,The dotted lines represent peak-to-peak amplitudes of dispersions in the three color ranges. Now, to probe any dependence of the residual dispersion on the rotation period, we computed the Pearson linear correlation coefficients (r) and significance levels\footnote{The significance
level $\alpha$ represents the probability of observing a value
of the correlation coefficient larger than $r$ for a random sample
having the same number of observations and degrees of freedom
\citep{Bevington69}} ($\alpha$) between the residuals
and the rotation period as well as the Spearman's rank correlation \citep{Spearman04}. In the case of blue stars we find $r$ = $-0.19$ with a significance level $\alpha <$ 50\%, i.e. no dependence
exists of the residuals on rotation (from Spearman's rank correlation we found $\rho$ = $-$0.04 with very small significance p-value = 0.77). On the contrary, for red stars we found $r$ = $-0.71$ with very high significance level $\alpha >$ 99.9\% (similarly, from rank correlation we found $\rho$ = $-$0.56 with high significance p-value = 0.01). Finally, for very red stars, we find $r$ = $+0.13$ with a significance level $\alpha \sim$ 80\% (from Spearman's rank correlation we found $\rho$ = 0.29 with very small significance p-value = 0.58).
Therefore, since in this regime of very red stars we have no enough data, we can not asses the existence or not of a rotation - Li depletion connection.\\
In the case of red stars (3.2 $\le$ V$-$K$_s$ $\le$ 4.5\,mag) we can state that the Li depletion is significantly correlated to rotation \rm that seems to maintain fast rotators less
depleted than slow rotators according to Eq.\,(1):\\
\begin{equation}
\delta EW(Li) = 101\pm22 - 17.7\pm3.4\times P
\end{equation}
where the Li EW is in m\AA\,\, and the rotation period P in days. The average uncertainty on the rotation period ($\Delta$P = 0.7\%) is negligible with respect to that on the Li EW.
After decorrelating the Li EW from rotation of red stars, \rm the dispersion turns to be reduced by about a factor of 2, $\sigma$ = 50 \,m\AA\,\,
and becomes to be comparable to that measured among blue stars, where no dependence on rotation was observed, and among very red stars.
This residual dispersion is larger than the uncertainty associated to the Li EW measurements.
Therefore, apart from mass and rotation, other causes must play a role in producing the observed dispersion. The intrinsic variability of the Li EW is certainly one cause. In fact, we know that spotted stars, as all our targets are, exhibit a rotational modulation of the Li EW. Specifically, the EW is larger at the rotation phases when the spots are best in view, and, consequently, the average surface temperature is lower.
In our sample 26 stars have two EW measurements. We find that the average difference between these multiple measurements is $<$$\Delta$EW$>$ = 26\,m\AA\,\,. We note only two extreme cases: \object{TX Psa} and \object{2MASS\,J05082729-2101444} both very red stars with $\Delta$EW $>$ 100\,m\AA. \rm\\
Generally, the rotational Li variability in young stars is not larger than about 5\% \citep[see Appendix in][]{Bouvier16}. \rm This corresponds in our case to a range from $\sim$5 to 25\,m\AA\,\, in absolute values, which is as the same order of the intrinsic variability we found.
\begin{figure*}
\begin{minipage}{22cm}
\includegraphics[scale = 0.3, trim = 0 0 0 0, clip, angle=90]{LDB.eps}
\includegraphics[scale = 0.3, trim = 0 0 0 0, clip, angle=90]{LDB_corrected.eps}
\end{minipage}
\caption{\label{LDB} \it Left panel: \rm distribution of Li EWs versus V$-$K$_s$ color for $\beta$ Pic members. Solid lines represent the Li EWs predicted by the models of \citet{Baraffe15} for ages in the range from 5 to 25 Myr. \it Right panel: \rm same as in the left panel, but with models arbitrarily shifted by 0.7\,mag towards redder colors to match with the observations.}
\end{figure*}
\subsection{Observed versus predicted LBD}
It is interesting to make a comparison between the measured distribution of Li EWs and the predictions of evolutionary models. In the following, we first consider the evolutionary models of \citet{Baraffe15} for solar abundances. Solar abundances have been measured in a few members, like \object{$\beta$ Pic} and \object{PZ Tel}, and generally they were adopted for the whole association \citep[see, e.g.,][]{Mentuch08}. The surface Lithium abundance is first derived from the ratio of surface Lithium abundance to initial abundance provided with the models using an initial abundance of Li$_0$ = 1$\times$10$^{-9}$. \rm Then, to make the comparison with the observations, we transformed the model Li abundance into Li EW. For this purpose, we have used the curves of growth from \citet{Zapatero02}. They are valid in the effective temperature range 2600 $<$ T$_{\rm eff}$ $<$ 4100\,K and for 1.0 $<$ A(Li) $<$ 3.4.
In the left panel of Fig.\,\ref{LDB}, we plot the measured Li EW and compare this distribution with the Li EW derived \rm from \citet{Baraffe15} models. We note a strong mismatch between observations and model predictions. The models for ages in the 10--25 Myr range, predict a substantial/complete depletion in the V$-$K$_s$ color range 3.8--4.5\,mag, whereas the Li gap is observed at colors that are redder by about 0.7\,mag. In fact, a better match is achieved if we arbitrarily shift the model colors by $\Delta$(V$-$K$_s$) = +0.7\,mag (see right panel of Fig.\,\ref{LDB}).
Moreover, assuming that the fast rotators have rotation-unaffected Li abundances we have maintained their observed Li EW, whereas for the slower rotators we have
increased the Li EW according to Equation (1).
As it is shown in the right panel of Fig.\,\ref{LDB}, we find that, even applying this V$-$K$_s$ color shift and the rotation decorrelation, we can infer different ages by looking at different part of the observed Li depletion since the slope of the observed depletion is different than that predicted by the models.
\begin{figure}[!h]
\begin{minipage}{10cm}
\includegraphics[scale = 0.3, trim = 0 0 0 0, clip, angle=90]{hr_gap.eps}
\end{minipage}
\caption{\label{hr} Absolute magnitude versus V$-$K$_s$ color of members in the Li gap. Solid lines represent the isochrones corresponding to ages from 5 (top) to 25\,Myr (bottom).}
\end{figure}
We note that the same mismatch between observations and models is found when different colors (e.g., V$-$I, J$-$H, or J$-$K) \rm or the effective temperatures are used to make the comparison.
The colors of the stars in the Li depletion gap 4.4 $<$ V$-$K$_s$ $<$ 5.4\,mag have been accurately checked.
For a few stars (either single or resolved components in wide binaries)
the V$-$K$_s$ is measured for each component. For other stars (components of unresolved binaries) the V$-$K$_s$ is derived
from the spectral type of the primary component, which is the one with measured Li.
In the first case the error on V$-$K$_s$ is $\pm$0.03\,mag, in the latter case
is $\pm$0.15\,mag.
In Fig.\,\ref{hr} we compare the absolute magnitudes of the members in the Li gap with isochrones for ages from 5 to 25 Myr taken from \citet{Baraffe15}. We note that within the gap (4.4 $<$V$-$K$_s$ $<$ 5.4\,mag) the stars seem to be either more luminous or redder \rm than expected for the most recent age estimation of 24\,Myr by \citet{Bell15}. \rm A better agreement is observed at bluer colors where stars are within the 10--25 Myr isochrones. If we correct the magnitudes of close components for binarity, we get a better agreement, but do not solve the overluminosity.
This indicates that, in order to reproduce the data, the predicted \citet{Baraffe15} colours should be shifted redward.\\
\begin{figure*}
\begin{minipage}{20cm}
\includegraphics[scale = 0.3, trim = 0 0 0 0, clip, angle=90]{Dartmouth_Li.eps}
\includegraphics[scale = 0.3, trim = 0 0 0 0, clip, angle=90]{Dartmouth_Li_corrected.eps}
\end{minipage}
\caption{\label{LDB_Dart} \it Left panel: \rm distribution of Li EWs versus V$-$K$_s$ color for $\beta$ Pic members. Solid lines represent the Li EWs predicted by the Dartmouth models for ages in the range from 20 to 30 Myr. \it Right panel: \rm same as in the left panel, but Li EW decorrelated from rotation, assuming that fast rotators have rotation-unaffected Li EW.}
\end{figure*}
All stars in the gap have a prominent magnetic activity as it is inferred from photometric variability, X-ray emission and Near-UV/Far-UV excess. Therefore, models including the effect of magnetic fields should be more adequate to describe the Li depletion pattern. For this reason, we also made use of the Dartmouth stellar evolutionary models that incorporate the effects of magnetic fields \citep{Feiden16}; \citep{Mann16}. In these models, inhibition of convection by magnetic fields cools the stellar surface temperature thereby slowing the contraction rate of young stars. Stars have a larger radius and a higher luminosity at a given age, as a result
These models were computed for solar metallicity and for an equipartition magnetic field strength in the range 2500 $<$ Bf$_{eq}$ $<$ 3000\,G. Model effective temperatures were transformed into V$-$K$_s$ colors using the empirical T$_{\rm eff}$-V$-$K$_s$ relation from \citet{Pecaut13} valid for young 5--30\,Myr stars. Li abundances were transformed into Li EW as done in the case of the Baraffe et al. models.
In the left panel of Fig.\,\ref{LDB_Dart}, we plot the measured Li EW and compare this distribution with the Li EW derived from the Dartmouth et al. models. In this case, no arbitrary color shift is needed to match with the hot side of the Li gap. A better match is achieved in the right panel of Fig.\,\ref{LDB_Dart}, when the Li EW are decorrelated from rotation, assuming again that the fast rotators have rotation-unaffected Li abundances.
The reduced chi-squares\footnote{To compute the $\chi^{2}_{\nu}$ we used the uncertainty $\sigma = \sqrt{\sigma_0^2+\sigma_v^2}$ where $\sigma_0$ is the measurement error and $\sigma_v$ is the average intrinsic variability. } computed from the fit to the hot boundary of the Li gap have their minimum (see
Fig.\,\ref{chisq}) for an age of 25$\pm$3\,Myr. The uncertainty formally represents the age interval by which the $\chi^{2}_{\nu}$ increases by one unity from the minimum value. The relatively large values of the $\chi^{2}_{\nu}$ indicate either that the EW scatter arising from the intrinsic variability (rotational modulation) is larger than our estimate of 26\,m\AA\,\,or the uncertainties associated to the measurements are in several case underestimated to some level.\\
Our age estimation is in good agreement with respect to the more recent age estimate of 23$\pm$4\,Myr and 24$\pm$3\,Myr by \citet{Mamajek14} and \citet{Bell15}, respectively, and significantly lower than the \citet{Macdonald10} estimate. \rm
\section{Conclusions}
We have retrieved from the literature the measured Li EW of 66 members of the young $\beta$ Pictoris association.
We have carried out a photometric monitoring of these members that allowed us to measure for the first time the rotation periods of 30 members. For other 16 members we retrieved the rotation periods from \citet{Messina10, Messina11}, and for the remaining members we used different sources in the literature.\\
We have explored the existence of a connection between rotation and Li depletion. After removing the mass dependence of the Li EW, using linear fits to EW versus V$-$K$_s$, we found that for stars with 0.5 $\le$ V$-$K$_s$ $\le$ 3.4\,mag, roughly corresponding to masses M $>$ 0.8\,M$_\odot$, no significant correlation is found between Li EW and rotation period. On the contrary, in the color range 3.4 $\le$ V$-$K$_s$ $\le$ 4.5\,mag, roughly corresponding to masses 0.3 $<$ M $<$ 0.8\,M$_\odot$, we find a strong correlation between the Li EW and the
rotation period, where fast rotators are much less Li-depleted than slow rotators. Finally, in the color range 5.4 $<$ V$-$K$_s$ $<$ 5.9\,mag, roughly corresponding to masses M $\sim$ 0.1\,M$_\odot$, we have some hint for an inverted correlation where fast rotators are more depleted than slow rotators. However, this correlation is currently based on only 7 stars and the significance level is high (80\%), but this is due to just one point that could be an outlier for whatever reason.\\
Interestingly, the dispersion in the 3.4 $\le$ V$-$K$_s$ $\le$ 4.5\,mag has a peak-to-peak amplitude that amounts to 180\,m\AA\,\, which is about a factor of 2 larger than that measured in the same mass range in the 5-Myr \object{NGC\,2264} open cluster \citep{Bouvier16}, and about a factor 2 smaller than that measured in the 125-Myr \object{Pleiades} open cluster \citep{Soderblom93}. Therefore, we note that the effect of rotation on the Li depletion is also age dependent and increases with age.\\
A comparison with the \citet{Baraffe15} models, shows a mismatch of about 0.7\,mag in the V$-$K$_s$ color \rm between the observed and the predicted color range where the Li gap falls. Models predict the Li gap at bluer colors than observed.
On the contrary, the Dartmouth models that incorporate the effects of magnetic fields provide a good match with the hot side of the Li depletion gap, although some mismatch on the cool side of the Li gap remains. Comparing the Dartmouth models with the hot side of the Li depletion gap, we infer an age of 25$\pm$3\,Myr, which is in very good agreement with the most recent age estimates for the $\beta$ Pictoris association. However, the relatively large values of the reduced chi-squares suggest that
either the intrinsic Li EW variability is still underestimated or some other factor, apart from rotation, plays a relevant role in producing the observed scatter among stars with similar mass.\\
\rm
\begin{figure}[!h]
\begin{minipage}{10cm}
\includegraphics[scale = 0.3, trim = 0 0 0 0, clip, angle=90]{chisq.eps}
\end{minipage}
\caption{\label{chisq} Reduced chi-squares versus age obtained from the residuals from the fit
to the hot boundary of the Li depletion gap.}
\end{figure}
{\it Acknowledgements}. Research on stellar activity at INAF- Catania Astrophysical Observatory is supported by MIUR (Ministero dell'Istruzione, dell'Universit\'a e della Ricerca). This research has made use of the Simbad database, operated at CDS (Strasbourg, France). SM thanks Jerome Bouvier for useful discussion and the anonymous Referee for useful comments that helped us to improve this paper.
\begin{table*}
\caption{Properties of the 66 members of the $\beta$ Pictoris association studied in this work}
\begin{minipage}{20cm}
\hspace{-1cm}
\scriptsize
\begin{tabular}{l l l l l l l c r r r r}
\hline
Target &
\multicolumn{3}{c}{RA } &
\multicolumn{3}{c}{DEC} &
Sp.T &
V\hspace{0.3cm} &
V$-$K$_s$ &
P \hspace{0.2cm} &
Li EW\\
& \multicolumn{3}{c}{(hh, mm, ss)} & \multicolumn{3}{c}{($^{\circ}$, $^{\prime}$, $^{\prime\prime}$)} & & (mag) & (mag) & (d)\hspace{0.2cm} & (m\AA)\\
\hline
\object{HIP\,560}& 00& 06& 50.08& $-$23& 06& 27.20& F3V& 6.15 & 0.910 & 0.224 & 87.0 \\
\object{TYC\,1186-0706-1}& 00& 23& 34.66& +20& 14& 28.75& K7.5V+M5& 10.96 & 3.623 & 7.900 & 338.0 \\
\object{GJ\,2006A}& 00& 27& 50.23& $-$32& 33& 06.42& M3.5Ve& 12.87 & 4.858 & 3.990 & 29.0 \\
\object{GJ\,2006B}& 00& 27& 50.35& $-$32& 33& 23.86& M3.5Ve& 13.16 & 5.044 & 4.910 & 26.0 \\
\object{2MASS\,J01112542+1526214}A& 01& 11& 25.42& +15& 26& 21.50& M5+M6& 14.46 & 6.252 & 0.911 & 593.0 \\
\object{2MASS\,J01351393-0712517}& 01& 35& 13.93& $-$07& 12& 51.77& M4.5V& 13.22 & 5.502 & 0.703 & 46.7 \\
\object{TYC\,1208-0468-1}& 01& 37& 39.42& +18& 35& 32.91& K3+K5& 9.85 & 3.114 & 2.803 & 440.0 \\
\object{HIP\,10679}& 02& 17& 24.74& +28& 44& 30.43& G2V& 7.75 & 1.488 & 0.777 & 160.0 \\
\object{HIP\,10680}& 02& 17& 25.28& +28& 44& 42.16& F5V& 6.95 & 1.163 & 0.240 & 140.0 \\
\object{HIP\,11437}A& 02& 27& 29.25& +30& 58& 24.60& K4& 10.12 & 3.040 & 12.500 & 248.0 \\
\object{HIP\,11437}B& 02& 27& 28.05& +30& 58& 40.53& M1& 12.44 & 4.219 & 4.660 & 220.0 \\
\object{HIP\,12545}& 02& 41& 25.90& +05& 59& 18.00& K6Ve& 10.37 & 3.301 & 4.830 & 450.0 \\
\object{GJ3305}& 04& 37& 37.30& $-$02& 29& 28.00& M1+M?& 10.59 & 4.177 & 4.890 & 140.0 \\
\object{2MASS\,J04435686+3723033}& 04& 43& 56.87& +37& 23& 03.30& M3Ve+M5?& 12.98 & 4.179 & 4.288 & 194.0 \\
\object{HIP\,23200}& 04& 59& 34.83& +01& 47& 00.68& M0.5Ve& 10.05 & 3.990 & 4.430 & 270.0 \\
\object{HIP\,23309}& 05& 00& 47.10& $-$57& 15& 25.00& M0Ve& 10.00 & 3.756 & 8.600 & 360.0 \\
\object{HIP\,23418}A& 05& 01& 58.80& +09& 59& 00.00& M3V& 11.45 & 4.780 & 1.220 & 53.0 \\
\object{BD\,-211074A}& 05& 06& 49.90& $-$21& 35& 09.00& M1.5V& 10.29 & 4.350 & 9.300 & 20.0 \\
\object{BD\,-211074B}& 05& 06& 49.90& $-$21& 35& 09.00& M2.5V& 11.67 & 4.643 & 5.400 & 20.0 \\
\object{2MASS\,J05082729-2101444}& 05& 08& 27.30& $-$21& 01& 44.40& M5.6V& 14.70 & 5.867 & 0.280 & 618.0 \\
\object{2MASS\,J05241914-1601153}& 05& 24& 19.15& $-$16& 01& 15.30& M4.5+M5& 13.50 & 5.603 & 0.401 & 217.0 \\
\object{HIP\,25486}& 05& 27& 04.76& $-$11& 54& 03.47& F7V& 6.22 & 1.294 & 0.966 & 191.0 \\
\object{2MASS\,J05335981-0221325}& 05& 33& 59.81& $-$02& 21& 32.50& M2.9V& 12.42 & 4.725 & 7.250 & 49.0 \\
\object{2MASS\,J06131330-2742054}& 06& 13& 13.31& $-$27& 42& 05.50& M3.V:& 12.09 & 5.230 & 16.9 & 28.0 \\
\object{HIP\,29964}& 06& 18& 28.20& $-$72& 02& 41.00& K4Ve& 9.80 & 2.986 & 2.670 & 420.0 \\
\object{TWA\,22}& 10& 17& 26.89& $-$53& 54& 26.50& M5& 13.99 & 6.301 & 0.830 & 510.0 \\
\object{HIP\,76629}& 15& 38& 57.50& $-$57& 42& 27.00& K0V& 7.97 & 2.118 & 4.270 & 292.0 \\
\object{2MASS\,J16430128-1754274}& 16& 43& 01.29& $-$17& 54& 27.50& M0.6& 12.50 & 3.951 & 5.140 & 300.0 \\
\object{HIP\,84586}& 17& 17& 25.50& $-$66& 57& 04.00& G5IV+K5IV& 7.23 & 2.528 & 1.680 & 250.0 \\
\object{HD\,155555C}& 17& 17& 31.29& $-$66& 57& 05.49& M3.5Ve& 12.71 & 5.081 & 4.430 & 20.0 \\
\object{TYC\,872822621}& 17& 29& 55.10& $-$54& 15& 49.00& K1V& 9.55 & 2.186 & 1.830 & 360.0 \\
\object{GSC\,08350-01924}& 17& 29& 20.67& $-$50& 14& 53.00& M3V& 12.86 & 4.766 & 1.982 & 50.0 \\
\object{V4046\,Sgr}& 18& 14& 10.50& $-$32& 47& 33.00& K5+K7& 10.44 & 3.191 & 2.420 & 440.0 \\
\object{2MASS\,J18151564-4927472}& 18& 15& 15.64& $-$49& 27& 47.20& M3V& 12.86 & 4.780 & 0.447 & 46.0 \\
\object{HIP\,89829}& 18& 19& 52.20& $-$29& 16& 33.00& G1V& 8.89 & 1.837 & 0.571 & 290.0 \\
\object{2MASS\,J18202275-1011131}A& 18& 20& 22.74& $-$10& 11& 13.62& K5Ve+K7Ve& 10.63 & 3.350 & 4.650 & 530.0 \\
\object{TYC\,907724891}& 18& 45& 37.02& $-$64& 51& 46.14& K5Ve& 9.30 & 3.204 & 0.345 & 490.0 \\
\object{TYC\,907307621}& 18& 46& 52.60& $-$62& 10& 36.00& M1Ve& 11.80 & 3.946 & 5.370 & 332.0 \\
\object{HD\,173167}& 18& 48& 06.36& $-$62& 13& 47.02& F5V& 7.28 & 1.144 & 0.250 & 107.0 \\
\object{TYC\,740800541}& 18& 50& 44.50& $-$31& 47& 47.00& K8Ve& 11.20 & 3.660 & 1.075 & 492.0 \\
\object{HIP\,92680}& 18& 53& 05.90& $-$50& 10& 50.00& K8Ve& 8.29 & 1.924 & 0.944 & 287.0 \\
\object{TYC\,687210111}& 18& 58& 04.20& $-$29& 53& 05.00& M0Ve& 11.78 & 3.762 & 0.503 & 483.0 \\
\object{2MASS\, J19102820-2319486}& 19& 10& 28.21& $-$23& 19& 48.60& M4V& 13.20 & 4.985 & 3.640 & 55.0 \\
\object{TYC\,687801951}& 19& 11& 44.70& $-$26& 04& 09.00& K4Ve& 10.27 & 2.904 & 5.650 & 320.0 \\
\object{2MASS\,J19233820-4606316}& 19& 23& 38.20& $-$46& 06& 31.60& M0V& 11.87 & 3.598 & 3.242 & 422.0 \\
\object{TYC\,744311021}& 19& 56& 04.37& $-$32& 07& 37.71& M0.0V& 11.80 & 3.954 & 11.300 & 110.0 \\
\object{2MASS\, J19560294-3207186}AB& 19& 56& 02.94& $-$32& 07& 18.70& M4V& 13.23 & 5.116 & 1.569 & 100.0 \\
\object{2MASS\,J20013718-3313139}& 20& 01& 37.18& $-$33& 13& 14.01& M1& 12.25 & 4.056 & 12.700 & 100.0 \\
\object{2MASS\,J20055640-3216591}& 20& 05& 56.41& $-$32& 16& 59.15& M2:& 11.96 & 4.022 & 8.368 & 140.0 \\
\object{HD\,191089}& 20& 09& 05.21& $-$26& 13& 26.52& F5V& 7.18 & 1.104 & 0.488 & 91.0 \\
\object{2MASS\, J20100002-2801410}AB& 20& 10& 00.03& $-$28& 01& 41.10& M2.5+M3.5& 12.80 & 4.640 & 0.470 & 46.0 \\
\object{2MASS\,J20333759-2556521}& 20& 33& 37.59& $-$25& 56& 52.20& M4.5& 14.87 & 5.993 & 0.710 & 504.0 \\
\object{HIP\,102141}A& 20& 41& 51.20& $-$32& 26& 07.00& M4Ve& 10.36 & 5.416 & 1.191 & 20.0 \\
\object{HIP\,102141}B& 20& 41& 51.10& $-$32& 26& 10.00& M4Ve& 10.36 & 5.416 & 0.781 & 20.0 \\
\object{2MASS\,J20434114-2433534}& 20& 43& 41.14& $-$24& 33& 53.19& M3.7+M4.1& 12.73 & 4.971 & 1.610 & 28.0 \\
\object{HIP\,102409}& 20& 45& 09.50& $-$31& 20& 27.00& M1Ve& 8.73 & 4.201 & 4.860 & 80.0 \\
\object{HIP\,103311}& 20& 55& 47.67& $-$17& 06& 51.04& F8V& 7.35 & 1.539 & 0.356 & 147.0 \\
\object{TYC\,634902001}& 20& 56& 02.70& $-$17& 10& 54.00& K6Ve+M2& 10.62 & 3.541 & 3.410 & 420.0 \\
\object{2MASS\,J21100535-1919573}& 21& 10& 05.36& $-$19& 19& 57.40& M2V& 11.54 & 4.344 & 3.710 & 41.0 \\
\object{2MASS\, J21103147-2710578}& 21& 10& 31.48& $-$27& 10& 57.80& M4.5V& 14.90 & 5.600 & 0.650 & 502.0 \\
\object{TYC\,9486-927-1}& 21& 25& 27.49& $-$81& 38& 27.68& M2V& 11.70 & 4.360 & 0.542 & 104.0 \\
\object{TYC\,221113091} & 22 & 00 & 41.59 & +27 & 15 & 13.60 & M0V & 11.39 & 3.666 & 1.109 & 40.0 \\
\object{TYC\,934004371}& 22& 42& 48.90& $-$71& 42& 21.00& K7Ve& 10.60 & 3.706 & 4.460 & 440.0 \\
\object{HIP\,112312}& 22& 44& 58.00& $-$33& 15& 02.00& M4Ve& 12.10 & 5.168 & 2.370 & 30.0 \\
\object{TX Psa}& 22& 45& 00.05& $-$33& 15& 25.80& M4.5Ve& 13.36 & 5.567 & 1.080 & 450.0 \\
\object{TYC\,583206661}& 23& 32& 30.90& $-$12& 15& 52.00& M0Ve& 10.54 & 3.971 & 5.680 & 185.0 \\
\hline
\end{tabular}
\end{minipage
\end{table*}
\bibliographystyle{aa}
|
\section{Introduction}
Designing materials for advanced or next-generation applications requires understanding of how properties are related to structure, thatis, identifying so-called structure-property relationships. Having such relationships guides the search for new materials with enhanced performance by identifying regions of structure and composition space that exhibit superior properties. For nuclear energy materials, a key performance metric is tolerance against radiation damage. Pyrochlores ($A_2$$B_2$O$_7$) have been extensively studied for their potential application as nuclear waste forms~\cite{Sickafus00,Begg01,Lian03a,Lian03b,Lian2006,Helean04,Sickafus07,Lumpkin07,Sattonnay13,Li12} and have been incorporated into some compositions of the SYNROC waste form~\cite{synroc}. In this context, significant effort has been directed toward understanding how the chemistry of the pyrochlore -- the nature of the $A$ and $B$ cations -- dictates the amorphization susceptibility of the compound. In particular, several experimental efforts~\cite{Ewing03,Lumpkin04,Lumpkin09,Whittle11,Lian11} have been focused on determining the critical amorphization temperature, $T_C$, the temperature at which the material recovery rate is equal to or faster than the rate of damage, as summarized in Fig.~\ref{Tc-vs-rA}. Typically, these experiments were performed in an electron microscope equipped with an ion source, such that samples were simultaneously irradiated with electrons and 1 MeV Kr ions. Though the value of $T_C$ is expected to vary depending on ion irradiation conditions~\cite{Meldrum2002}, 1 MeV Kr ion irradiation results should be comparable.
\begin{figure}
\includegraphics[width=3.5in]{Tc-vs-rA.pdf}
\caption{\label{Tc-vs-rA} {
Experimentally measured values of $T_C$, ordered as a function of $A$ cation radius, for several different pyrochlores.
}}
\end{figure}
As a consequence, a number of ``features'' -- or basic structural and energetic properties -- have been identified that provide insight into the radiation response of pyrochlores. These include the radii and electronegativities of the $A$ and $B$ cations~\cite{Lumpkin04,Lumpkin07}; the $x$ parameter, which describes how the oxygen sublattice deviates from ideality~\cite{Lian03b,Lumpkin04,Lumpkin07}; the enthalpy of formation of the pyrochlore~\cite{Helean04,Helean04b}; and the energy to disorder the pyrochlore to a disordered fluorite structure~\cite{Sickafus00,Minervini00}. Further, there has been discussion on the extent of the disordered phase field in the phase diagram and its relationship to amorphization resistance~\cite{Sickafus07}. Most of these features have been only heuristically correlated with amorphization resistance or only applied to a subset of pyrochlore chemistries. We are only aware of one attempt to quantify the relationship between these types of features and a prediction of $T_C$. In that work, Lumpkin and co-workers established a relationship between $T_C$ and lattice constants, electronegativies, disordering energetics, and oxygen positional parameter~\cite{Lumpkin07}. While their model provided a significant advance in describing the structure-property relationships of pyrochlores, here we demonstrate how, through the use of machine learning, greater insight can be extracted. In particular, while they considered the disordering energy as one of their features, they used data from atomistic potentials that does not adequately describe all of the chemistries in the experiments. Further, they did not have access to data describing the amorphous state of these compounds. Finally, modern machine learning methods, applied to materials science, offer new avenues to examine the structure-property relationships in these types of systems.
Here, we use machine learning methods to demonstrate how a set of features, for a range of pyrochlore chemistries, can be used to predict $T_C$. We use both structural parameters such as cation radius and electronegativity supplemented by energetics calculated with density functional theory (DFT) to build a database of features as a function of pyrochlore chemistry. We analyze this database, building machine learning models that predict $T_C$ as a function of pyrochlore chemistry based on a systematic collection of features. We consider pyrochlore chemistries for which experimental data exists for $T_C$, which includes pyrochlores where $B$=Ti, Zr, Hf, and Sn. We find that, when considering the full range of chemistries, the two features that best predict $T_C$ are the ratio of the radii and the difference in electronegativities of the $A$ and $B$ cations. However, to predict more subtle dependencies of $T_C$ with pyrochlore chemistry characteristic of a given B chemistry, the energies to disorder and amorphize the compound provide a better prediction of $T_C$.
As compared to Ti, Hf, or Zr, Sn is a chemically very different element. It, like Ti, is multivalent, but unlike Ti, has a much stronger prevalence to adopt a charge state other than 4+. Further, as discussed below, it has a significantly higher electronegativity than the other B cations, producing a more covalent bond. This implies that Sn pyrochlores should be less amorphization resistant~\cite{Naguib1975}. However, experiments have shown Sn pyrochlores to be more amorphization resistant than other pyrochlores~\cite{Lian2006}. This all suggests that Sn pyrochlores are electronically much more complex than the other pyrochlore families, which is one reason that we use DFT to determine the energetics of disordering and amorphization, as DFT can account for the varied valence of the Sn cations. Further, the inclusion of Sn pyrochlores in this analysis, precisely because the behavior is counter-intuitive, provides a more stringent test of the methodology.
\section*{Results}
\subsection*{DFT Energetics}
Figure~\ref{DFT-results}a provides the energetics for disorder and amorphization of a given pyrochlore, as found using DFT, as a function of the chemistry of the pyrochlore. These are ordered by $A$ cation radius. Focusing first on the energetics to disorder, there is a general trend that as the $A$ cation radius increases, the energy associated with disordering the pyrochlore to a disordered fluorite also increases, consistent with previous results using DFT~\cite{Jiang09}. This is particularly true of the $B$=Zr, Hf and Sn families of pyrochlores. For the $B$=Ti family, there is a peak in the disorder energy for the $A$=Gd composition, again consistent with previous DFT and empirical potential calculations~\cite{Minervini00,Jiang09}.
Figure~\ref{DFT-results}a highlights the apparent contradiction between experimental observations and the notion that the disordering energy correlates with amorphization resistance. If only disordering energetics dictated the response of the pyrochlore to irradiation, then one would expect that Zr pyrochlores would generally exhibit higher amorphization resistance than Ti pyrochlores (which they do) but also that Sn pyrochlores would be less resistance to amorphization than Ti pyrochlores, which they are not. Thus, other factors must also be important. We propose that the energy of the amorphous phase is one of those factors.
The energy differences between ordered pyrochlore and an amorphous structure are also provided in Fig.~\ref{DFT-results}a. In the case of the $B$=Hf and Zr families, these are again relatively monotonic with increasing $A$ cation radius. However, the behavior of the $B$=Ti and Sn families is more complex. In particular, for the $B$=Ti family, the amorphous energy is non-monotonic with $A$ cation radius, but the peak is for a different chemistry than was the disordering energy. In the $B$=Ti family, the amorphous energy is greatest for $A$=Y and generally is high for $A$=Dy and Tb. The $B$=Sn family exhibits even more complicated behavior. There is a peak in the amorphous energy for $A$=Gd and a minimum for $A$=Ho.
Finally, the shaded regions in Fig.~\ref{DFT-results}a highlight the energy gap between the disordered and amorphous states. The variation of this gap with $A$ cation radius is very different for the different families of pyrochlores. For the $B$=Zr and Hf pyrochlores, the gap slowly but steadily decreases with $A$ cation radius. For the $B$=Ti pyrochlores, the gap first increases slightly and then decreases to essentially zero for the $A$=Nd chemistry. The gap for $B$=Sn pyrochlores first decreases, then increases, and then decreases again. Further, the gap is smallest for the $B$=Ti pyrochlores and, overall, largest for the $B$=Zr and Sn pyrochlores, at least for some A chemistries.
Figure~\ref{DFT-results}b provides the volume changes between the ordered phase and both the disordered and amorphous phases, as determined from the DFT calculations. For nearly all of the cases, a transformation from the ordered to disordered phase results in a volume expansion while the formation of the amorphous phase contracts the lattice. The exceptions are the $B$=Zr and Hf pyrochlores with small $A$ cations, which exhibit very little change in volume upon disordering. On the other hand, Nd$_2$Ti$_2$O$_7$ (assumed to be cubic here) exhibits very little change upon amorphization.
\begin{figure}
\includegraphics[width=3.5in]{DFT-figure.pdf}
\caption{\label{DFT-results} {
DFT results for the (a) energetics and (b) volume changes associated with an order-to-disorder (O$\rightarrow$D, open symbols) and an order-to-amorphous (O$\rightarrow$A, closed symbols) transformation for four families of pyrochlores in which $B$=Ti (green), Zr (purple), Hf (yellow), and Sn (cyan). The shaded regions highlight the differences between the disordered and amorphous structures.
}}
\end{figure}
\subsection*{Correlation of Features with Amorphization Resistance}
The DFT results reveal that there are significant differences in the energetics of disorder and amorphization in pyrochlores as a function of both $A$ and $B$ chemistry. We use a machine learning approach to quantify the correlations between these energetics, as well as other features associated with pyrochlores, and the amorphization resistance, as characterized by $T_C$. The features considered here are $r_A/r_B$, the ratio of the ionic radii of the $A$ and $B$ cations; $\Delta X=X_B-X_A$, the difference in electronegativity of the $A$ and $B$ neutral metal atoms ($X_A$ and $X_B$, respectively); $x$, the oxygen positional parameter, which measures the deviation of the oxygen sublattice from an ideal (fluorite-like) simple cubic sutlattice; $E_{O\rightarrow D}$, the energy difference between the disordered and ordered phases; and $E_{D\rightarrow A}$, the energy difference between the amorphous and disordered phases. These features were chosen because (a) they have been shown to correlate to some degree in previous studies and (b) our DFT results indicate that the energetics depend strongly on the A and B chemistry of the pyrochlore, suggesting they may provide a strong descriptor of each compound. We did not consider the enthalpy of formation, proposed by other authors as a factor in radiation tolerance~\cite{Helean04,Helean04b}, as a feature because data was not available for all compounds.
However, before we examine the results of the machine learning model, it is instructive to examine how the selected features correlate with $T_C$. Figure~\ref{features} provides simple plots of each feature against $T_C$. The values for $T_C$, summarized in Table S1, are taken from Refs.~\cite{Ewing03,Lumpkin04,Whittle11,Lian11}. Figure~\ref{features} reveals that while there are rough correlations between $T_C$ and some of the features, there is not one feature that provides a quantitative capability of predicting $T_C$. For example, overall, $r_A/r_B$ correlates well with $T_C$ over a wide range of $B$ chemistries; however, it does not capture subtleties associated with variations in $T_C$ with a given family of pyrochlores. $\Delta X$, on the other hand, discriminates between pyrochlores with $B$=Sn and the other families but does not correlate directly with $T_C$. Similarly, $x$ shows an overall correlation with $T_C$ but again the details are lost. $E_{O\rightarrow D}$, on the other hand, seems to correlate reasonably well for pyrochlores within a given family but does not describe variations of $T_C$ between families. Finally, $E_{O\rightarrow A}$, similar to $x$ and $\Delta X$, seems to generally correlate separately for $B$=Sn pyrochlores and the other families of pyrochlores. Thus, while there are rough trends indicating some insight from each of these features, there is certainly not enough of a correlation in any case for a quantitative prediction. However, this suggests, as noted by other authors~\cite{Lumpkin07}, that combinations of these features may provide predictive capability. Hence, we use a machine learning approach to quantify this.
\begin{figure*}
\includegraphics[width=7in]{Feature_Tc_plots.pdf}
\caption{\label{features} {
Correlations between experimental measurements of $T_C$ and 5 features describing pyrochlores. (a) $x$, the oxygen positional parameter. (b) $r_A/r_B$, the ratio of cation radii. (c) $E_{O\rightarrow D}$, the energy difference between disordered fluorite and ordered pyrochlore. (d) $E_{D\rightarrow A}$, the energy difference between an amorphous structure and disordered fluorite. (e) $\Delta X$, the difference in electronegativity between the $A$ and $B$ cation. The different symbols in the figures indicate the nature of the $B$ cation (squares=Zr, circles=Ti, diamonds=Sn, triangles=Hf).
}}
\end{figure*}
\subsection*{Results of the Machine Learning Model}
We use a machine learning (ML) approach to quantify the correlations between the five features described in the previous section and $T_C$. More specifically, we employed kernel ridge regression (KRR)\cite{KRR-Witten,KRR-Muller,KRR-Hofmann}---an algorithm that works on the principle of similarity and is capable of extracting complex non-linear relationships from data in an efficient manner---with a Gaussian kernel to learn and quantify trends exhibited by $T_C$ in the feature space discussed above. A randomly selected 90\%/10\% training/test split of the available data was used for statistical learning and testing the performance of the trained model on previously unseen data. A leave-one-out cross validation is used to determine the model hyper-parameters to avoid any overfitting of the training data that may lead to poor generalizability. The trained model can subsequently be used to make an interpolative prediction of $T_C$ for a new material ($i.e.$, not used in the model training) $i$ by comparing its distance in feature space $d_i$ (suitably defined by a distance measure; in our case the Euclidean norm was used) with those of a set of reference training cases for which the $T_C$ values are known. Further details of our KRR-based ML models are provided in the Methods section.
Next, within the KRR ML model, we aim to identify the best feature combination that exhibits highest prediction performance, quantified by its ability to accurately predict $T_C$ of the test set compounds. We do this in a comprehensive manner by building KRR ML models using all possible combinations of $\Omega$ features with $\Omega$ $\in$ [2,5]. Performance of each of these models was evaluated separately on the entire data set as well as on a reduced set that only included the Ti pyrochlores. The root mean square ($rms$) errors for the $T_C$ predictions on training and test sets for various models is presented in Fig.~\ref{machine-learning}. In order to account for model prediction variability associated with randomly selected training/test splits, Fig.~\ref{machine-learning} reports the $rms$ errors averaged over 100 different randomly selected training/test splits for each of the models. The 2D models that lead to the lowest $rms$ errors on the test set data have been marked with a `$\star$' in Fig.~\ref{machine-learning}a (when taking the entire data) and Fig.~\ref{machine-learning}b (for the Ti pyrochlores). It is interesting to note that, for both cases, going beyond the best performing 2D models does not lead to a significant improvement in the model prediction performance. For instance, while the best binary feature pair ($r_A/r_B$, $\Delta X$) leads to a test set $rms$ error of 101.2 K in $T_C$, the ML models built on the best 4D and 5D (taking all 5 features considered) feature vectors only result in nominal improvements leading to $rms$ errors of 97.8K and 98K, respectively. Since, as a general rule, higher model complexity often leads to poor generalizability, in case of a comparable prediction performance, a simpler model ($i.e.$, built on a lower dimensional feature set) should always be preferred over a more complex one. Therefore, henceforth we focus our attention on the the best performing 2D models.
\begin{figure*}
\includegraphics[width=7in]{ML_Model_Results.pdf}
\caption{\label{machine-learning} {
Results from the machine learning model. (a) Results when applied to the entire set of pyrochlores and (b) results when applied only to the titanate family. The solid bars indicate the average RMS error for the training data while the hashed bars indicate the error for the test data. The indices on the abscissa indicate the features used in that particular model. Models including 2, 3, 4 or all 5 features were considered. The best 2D feature set is indicated with the star for both cases. Error bars represent the standard deviations for the rms error in predict $T_C$, computed over the 100 different training/test set splits.
}}
\end{figure*}
The superior performance exhibited by the ($r_A/r_B$, $\Delta X$) feature pair is not entirely unexpected and can be understood by looking at Fig.~\ref{features}b and e. As alluded to previously, while $r_A/r_B$ helps capture the overall $T_C$ trends among different chemistries, $\Delta X$ allows for an effective separation between different chemistries (especially, between the Sn-based compounds and rest of the dataset), while still capturing relative $T_C$ trends between these subgroups. The best performing feature pair for the titanate pyrochlores dataset, however, is constituted by $E_{O\rightarrow D}$ and $E_{D\rightarrow A}$. While the ($r_A/r_B$, $\Delta X$) feature pair performs much poorer on this subset than the overall dataset, the performance of ($r_A/r_B$, $E_{O\rightarrow D}$) feature pair is also found comparable to that of the best 2D feature pair.
While Fig.~\ref{machine-learning} captures the average performance and variability (taken over 100 different runs) for our best performing 2D models (marked with a $\star$), in Fig.~\ref{contour-plots}a-b we present parity plots comparing the experimental $T_C$ with the ML predictions using the best 2D descriptors found for the entire dataset (Fig.~\ref{contour-plots}a) and the titanates (Fig.~\ref{contour-plots}b), respectively. In each case $\sim$90$\%$ of the dataset was used for training (plotted as squares) with the remaining for testing the model performance (plotted as circles). In each case, we used four different ML runs randomly selecting training and test set splits (depicted by different colors). It can be see from the parity plots that our ML models can reasonably predict (within the error bars established in Fig.~\ref{machine-learning}), $T_C$ over the entire dataset. A couple of conclusions can be drawn from these plots. First, visually it can be seen that the model prediction performance is comparable for the training and test sets, indicating that there is no overfitting (a problem when a ML model performs very well on a training set but exhibits a poor performance on a test set). Second, despite their simplicity (given that we are only using a two-dimensional feature in each case), the ML models exhibit good predictive power and stability (predictions do not change drastically over different training/test splits). This highlights the robustness of the model.
\begin{figure}
\includegraphics[width=3.5in]{Contour-v2.pdf}
\caption{\label{contour-plots} {
(a-b) Parity plots of the machine learning results for (a) the entire set of pyrochlores and (b) the titanate family. The squares represent training data while the circles are the test set. The different colors represent different runs with different training/test set splits.
(c-d) Results from the machine learning model for (c) all of the pyrochlores considered and (d) just the titanate famliy. The size and color of the circles indicate the experimental $T_C$ while the position of the circles indicates the predicted $T_C$. The contours indicate predictions of $T_C$ for other values of the feature pairs.
}}
\end{figure}
To gain a deeper insight into the KRR model's prediction performance, we next construct contour plots for each of the two best performing 2D feature pairs discussed above. In each case, we start with a fine 2D grid in the feature space constituted by the primary features identified above, while still confining ourselves within the boundaries of the original feature space used to train the KRR models. Each point on this grid then, in principle, represents a point in the feature space, which can be used as an input for the respective trained KRR models to make predictions. That is, we map out the predicted value of $T_C$ as a function of the two features over a range of values for each of the two features. Since the ML models are interpolative, one can readily use a fine grid in the 2D feature space to visualize trends in $T_C$ versus the feature values and make predictions of $T_C$ for new chemistries.
Figure~\ref{contour-plots}c shows the best two-feature descriptor for the entire set of pyrochlores considered. Again, in this case, the two features that best correlate with $T_C$ are $r_A/r_B$ and $\Delta X$. This combination of features is able to distinguish the different $T_C$ behavior exhibited by the $B$=Sn pyrochlores and the other families of pyrochlores, by virtue of the properties of $\Delta X$.
However, as discussed above, this combination of features has an effective uncertainty of $\sim 100$ K, indicating that it cannot describe the fine features exhibited by the $B$=Ti family of pyrochlores. For example, $T_C$ is not monotonic with $A$ cation radius (see Fig.~\ref{Tc-vs-rA}). As discussed, limiting the model to just the $B$=Ti pyrochlores results in a different optimal two-feature set, namely $E_{O\rightarrow D}$ and $E_{O\rightarrow A}$, as shown in Fig.~\ref{machine-learning}b. In particular, as shown in Fig.~\ref{contour-plots}d, this set of features can describe the subtle behavior in which the $A$=Gd compound has the highest value of $T_C$, correlating with the fact that it has the highest value of $E_{O\rightarrow D}$, while the $A$=Y compound, which has values of $E_{O\rightarrow D}$ similar to the neighboring compounds, exhibits an anomalously low value of $T_C$. This is a consequence of its rather high value of $E_{D\rightarrow A}$, a consequence of the fact that Y is not a rare earth and thus the bonding associated with it is subtly different to the other elements around it.
\section*{Discussion and Conclusions}
Combining experimental results for $T_C$ for various pyrochlore compounds, DFT calculations of the energetics of disordering and amorphization, and a machine learning model, we conclude that (a) basic ionic properties such as $r_A/r_B$ and $\Delta X$ have the qualitative capability of predicting trends in $T_C$ over a wide-range of pyrochlore compounds but that (b) more quantitative predictions that capture the subtleties associated with variations in $A$ cation chemistry require knowledge of the disordering and amorphization energetics. This generalizes the previous understanding in which rough correlations between amorphization resistance and, for example, disordering energetics were hypothesized based on a few observations.
However, what is clear from the machine learning analysis is that, even with the input of DFT energetics, the predictive capabilities are still limited. Even when limited to the $B$=Ti family of pyrochlores, the model results in predictive uncertainty of 75 K. This is a consequence of many factors, including the limited amount of experimental data, the uncertainties in what experimental data that there is, and uncertainties in the DFT calculations. To determine an even better predictive model, more experimental data is required. In particular, values of $T_C$ for other families of pyrochlores would enhance the strength of the model. For example, without the $B$=Sn pyrochlores, the importance of $\Delta X$ would likely not have been revealed. Importantly, given the small data set, domain knowledge -- experience with the behavior of this system -- was important in narrowing down a set of likely relevant features.
While the feature set of $\Delta X$ and $r_A/r_B$ have the best predictive capability for distinguishing between the various families of pyrochlores, the reason why Sn pyrochlores are radiation tolerant while exhibiting such high disordering energies is found in examining the amorphization energetics. The gap between the disordering and amorphization energies for the Sn pyrochlores is typically quite large and even if, during the course of irradiation, enough energy is deposited into the lattice such that the structure becomes disordered, it is not enough to amorphize the material. The gap betwen the disordering and amorphization energies is much larger in the Sn pyrochlores than it is in the Ti family and, for some A cations, larger than for the Hf and Zr families as well. Thus, the origin of the radiation tolerance of some of the Sn pyrochlores comes from the fact that they are extremely difficult to amorphize.
The insights gained by the machine learning model apply specifically to pyrochlores and, because of the interpolative nature of these models, to the families of pyrochlores considered here. That said, the features identified as being best able to predict $T_C$ can be justified physically and thus may be applicable to other classes of complex oxides, such as $\delta$-phase~\cite{Stanek2009}, that have fluorite as the parent structure. However, other classes of complex oxides, such as spinel, which have fundamentally different crystal structures may have different dependencies on these features, or require new features to predict behavior. In particular, structural vacancies on the cation sublattice in spinel can facilitate recovery of damage in a way that is not possible in pyrochlore~\cite{Uberuaga2015}. Further, other factors, such as short-range order, which is known to occur in complex oxides~\cite{Jiang12,Shamblin2016}, may also play a role. However, we suspect that treating the disordered state as truly random captures much of the behavior of these materials, given the ability of the disordered fluorite structure to predict order-disorder temperatures in these systems~\cite{Jiang09,Li2015}.
In this work, we have used $T_C$ as a metric for relative amorphization resistance. In reality, the value of $T_C$ encompasses not only thermodynamic properties such as disordering and amorphization energetics, but also kinetic processes of defect annihilation and defect production. Thus, actually predicting $T_C$ from fundamental defect behavior would be a daunting task. However, it does provide a metric to compare the susceptibility of amorphization that has been measured for a range of pyrochlore chemistries.
Finally, this work highlights the utility of machine learning approaches in materials science. In this case, the ML model elucidates those features which provide predictive capability, providing insight into those factors which dictate amorphization resistance in pyrochlores. The model also shows that sets of two features result in optimal predictions; higher-order feature sets do not add significant value. The fact that different combinations of features provide are optimal for predictions for the entire set of pyrochlores ($r_A/r_B$ and $\Delta X$) versus the Ti family ($E_{O\rightarrow D}$ and $E_{O\rightarrow A}$) reinforces the point that the best set of features depends on the level of detail (here, the error in the predicted $T_C$) required in the prediction.
\section*{Methods}
\subsection*{Density Functional Theory}
Density functional theory (DFT) calculations were performed using the all-electron projector augmented wave method~\cite{PAW} within the local density approximation (PBE) with the VASP code~\cite{VASP}. A plane-wave cutoff of 400 eV and dense $k$-point meshes were used to ensure convergence. The lattice parameters and all atomic positions were allowed to relax, though the cells were constrained to be cubic. The disordered fluorite structure was modeled using the special quasirandom structures (SQS) approach~\cite{SQS}. The SQS structures were generated as described in Ref.~\cite{Jiang09}. The amorphous structures were created by performing {\em ab initio} molecular dynamics at a very high temperature and then quenching the structures to 0 K. For the $B$=Zr and Hf families, there is a deviation from true monotonic behavior at $A$=Tb, in contrast with previous DFT calculations~\cite{Jiang09} that used the same methodology (pseudopotentials, functional, k-point mesh, and energy cutoff). We assume that the differences from previously published results are due to changes in different versions of VASP.
\subsection*{Machine Learning Model}
We used Kernel ridge regression (KRR) with a Gaussian kernel for machine learning. KRR is a similarity-based learning algorithm, where the ML estimate of a target property (in our case the critical temperature $T_C$) of a new system $j$, is estimated by a sum of weighted kernel functions ($i.e.$, Gaussians) over the entire training set, as
\begin{equation}\label{eq:KRR}
T_{Cj}^{ML} = \sum \limits_{i=1}^N w_i \exp \left( -\frac{1}{2\sigma^2}|\textbf{d}^{ij}|^2 \right).
\end{equation}
where $i$ runs over the systems in the training dataset, and $|\textbf{d}^{ij}|^2 = ||\textbf{d}_i - \textbf{d}_j ||_2^2$, the squared Euclidean distance between the feature vectors $\textbf{d}_i$ and $\textbf{d}_j$. The coefficients $w_i$s are obtained from the training (or learning) process built on minimizing the expression $\sum \limits_{i=1}^N \bigg( T_{Ci}^{ML} - T_{Ci}^{Exp} \bigg)^2 + \lambda \sum \limits_{i=1}^N w_i^2$, with $T_{Ci}^{ML}$ being the ML estimated critical temperature, $T_{Ci}^{Exp}$ the corresponding experimental value, and the model hyper-parameters $\sigma$ and $\lambda$ are optimized within a internal cross-validation loop on a fine logarithmic grid. The explicit solution to this minimization problem is $\boldsymbol{\alpha} = (\textbf{K}+\boldsymbol{\lambda} \textbf{I} )^{-1} \textbf{P}^{DFT}$, where $\textbf{I}$ is the identity matrix, and $K_{ij} = \exp \left( -\frac{1}{2\sigma^2}|\textbf{d}^{ij}|^2 \right)$ is the kernel matrix elements of all materials in the training set. The parameters $\lambda$ and $\sigma$ are determined in an inner loop of fivefold cross validation using a logarithmically scaled fine grid.
\section*{Acknowledgments}
This work was supported by the U.S. Department of Energy, Office of Science, Basic Energy Sciences, Materials Sciences and Engineering Division.
Los Alamos National Laboratory, an affirmative action equal opportunity employer, is operated by Los Alamos National Security, LLC, for the National Nuclear Security Administration of the U.S. DOE under contract DE-AC52-06NA25396.
\section*{Author contributions}
G.P.\ developed the machine learning model and applied it to the pyrochlores. K.R.W.\ compiled the experimental data. C.J.\ constructed the DFT simulation cells. R.W.G.\ and K.E.S.\ performed the original studies that proposed the concept linking radiation tolerance to disordering. C.R.S.\ helped devise the study. B.P.U.\ wrote the main manuscript text and performed the DFT calculations. All authors reviewed the manuscript.
\section*{Additional information}
Competing financial interests: The authors declare no competing financial interests.
|
\section{Introduction}
Semiconductor coupled microcavity systems, also called photonic molecules (PMs), are attracting considerable attention as efficient test-beds for quantum and nonlinear optics
\cite{Dousse:2010uq,Abbarchi:2013fk,Gerace:2009kx,hamel2015,PhysRevB.86.045315}. The high optical nonlinearities combined with tight light confinement open up new routes for the study of light-matter
interaction in non-equilibrium driven dissipative systems. Even at the semiclassical level, already two coupled cavities --or photonic dimers-- may display a wide range of rich nonlinear dynamical phenomena
such as instabilities and bifurcations \cite{hamel2015,PhysRevB.77.125324,Rodriguez:2016fk}.
In this context, micro and nanolasers prove useful to investigate nonlinear dynamics in the low photon number regime. Either cavity quantum electrodynamics (cQED) microlasers or photonic band-gap materials can provide large enough spontaneous emission factors ($\beta$) such that strong nonlinearities may take place in an optical cavity with, ultimately, few photons. Large Purcell factors in the former, and the suppression of leaky modes in the latter can be tailored in order to obtain high $\beta$-factors. As a consequence, the saturation photon number, $s_{sat}\sim \beta ^{-1}$, may be significantly lowered \cite{PhysRevA.50.4318}. Micro and nanolasers may display rich nonlinear dynamical
behaviors even at a single laser level, such as injection locking phenomena \cite{schlottmann2016injection}, self-pulsing \cite{yu2016demonstration} and superradiant giant photon bunching \cite{Jahnke:2016vn}.
Recently, spontaneous symmetry breaking in coupled photonic crystal nanolasers has been demonstrated with $\sim 150$ intracavity photons ($\beta \sim 0.02$) \cite{hamel2015}, paving the way to explore a wealth of nonlinear dynamical phenomena in photonic dimers operating in a laser regime. These systems can also be useful for applications such as flip-flop optical memories and logic gates in active photonic circuitry \cite{liu2010}.
In this article we show, both theoretically and experimentally, the asymmetric interaction between strongly coupled modes in a photonic crystal nanolaser dimer.
We will analyze this phenomenon in the framework of "two discrete sites" Bogatov effect, in the sense that the dynamic population grating takes the form of oscillations of the population imbalance between two optical defects at the Rabi frequency (Fig.~\ref{fig1}a).
\section{The two-site Bogatov effect}
Mode competition is at the heart of multimode laser dynamics, and may occur as long as different modes share the same gain medium. For instance, mode switching and bistability
have been observed in standard microcavity systems, such as vertical cavity semiconductor lasers (VCSELs) \cite{Kawaguchi2010} and micropillar lasers \cite{Leymann2013,redlich2016mode},
and they have also been predicted in photonic crystal coupled cavities \cite{PhysRevLett.99.073902}. These mechanisms usually rely on cross-gain saturation effects.
However, mode interaction in semiconductor cavities may also result from stimulated scattering due to a dynamic carrier population grating in the
gain medium, which oscillates at the beat note frequency between adjacent cavity modes ($\sim10-100$ GHz).
Such interaction, known as the "Bogatov effect", is asymmetrical in the sense that the laser mode at the
blue side of the spectrum transfers its energy to the mode at the red side \cite{bogatov1975anomalous}. This phenomenon is a consequence of nonlinear dispersion, and it is also known to occur in free-running edge-emitters \cite{PhysRevA.69.053816,Lenstra:14},
in VCSELs supporting two orthogonal polarizations \cite{choquette1995}, and in vertical external cavity lasers (VECSELs) leading to a slow light regime \cite{PhysRevLett.105.223902}.
Yet, the observation of this effect in photonic dimers remains elusive to date.
For moderate $\beta$-factors ($\beta \sim 0.01$) as in our case,
an accurate description of the system can still be obtained in terms of mean-field equations accounting for the dynamics of two coupled complex field amplitudes in the left (L) and right (R) nanocavities filled with a quantum well gain medium \cite{hamel2015}(see Fig.~\ref{fig1}b). The evolution of the field and carrier densities are governed by
\begin{align}
\dot{a}_{L,R} &= -\kappa a_{L,R}+ \frac{\beta \gamma_{\parallel}}{2} \left(1+i\alpha\right)\left(n_{L,R}-n_0\right)a_{L,R}+ \nonumber \\
& + \left(\gamma+iK\right)a_{R,L} +F_{a_{R,L}}(t) \label{eq:aLR}\\
\dot{n}_{L,R} &= p_{L,R}-\gamma_{tot} n_{L,R}-\beta \gamma_{\parallel} \left( n_{L,R}-n_0\right) |a_{L,R}|^2 + F_{n_{R,L}}(t)\label{eq:nLR}
\end{align}
where $|a|^2$ and $n$ are normalized as the photon and carrier numbers in the cavities, respectively, $\kappa$ is the cavity loss rate, $\gamma_{\parallel}$ is the two-level radiative recombination rate, $\alpha$ the Henry factor,
$n_0$ the carrier number at transparency, $p_{L,R}$ the pump rate and $\gamma_{tot}$ is the total carrier recombination rate. The complex inter-cavity coupling constant quantifies frequency ($K$) and loss ($\gamma$) splitting as a result of the evanescent coupling, and $F_{a}$ are Langevin noise terms accounting for spontaneous emission with rate $R_{sp}= \beta B n_{L,R}^2/V_a$ where $B$ is the bimolecular radiative recombination rate and $V_a$ the volume of the active medium. The expression of $F_{n}$ is standard; it is chosen as to preserve the proper correlations between the photon and carrier numbers, see \cite{hamel2015} Supplementary Material for more details. In order to address mode interaction, we first project Eqs.~\ref{eq:aLR} and \ref{eq:nLR} on the basis of eigenmodes, $a_B=(a_L+a_R)/\sqrt{2}$ and $a_{AB}=(a_L-a_R)/\sqrt{2}$, corresponding to bonding (B) and anti-bonding (AB) modes of the photonic dimer respectively (Fig.~\ref{fig1}a).
\begin{figure}[!t]
\centering
\includegraphics[scale=0.4]{fig1.pdf}
\caption{a) Schematic representation of the eigenmodes of the coupled nanolasers and carrier population in each cavity. The dynamic population grating oscillates at the Rabi frequency and induces stimulated scattering from the blue to the red mode. b) Coupled PhC nanolasers suspended on an InP membrane. Green and orange circles: modified barrier holes. Red circles: modified holes for better beaming efficiency. c) Optical spectrum of the AB and B resonances with the corresponding far-field patterns. Grey: schematic representation of the profile of the gain curve.}
\label{fig1}
\end{figure}
Dynamic grating Bogatov effects can be described by considering a superposition of monochromatic fields, $a_{B,AB}=A_{B,AB} \exp(\pm iK t)$, which induces oscillation of the carrier number at the Rabi frequency $2K$:
\begin{eqnarray}
\bar{n} & = & \bar{n}_0+\left( \bar{N} e^{2iK t} +c.c.\right) \\
\Delta n& = & \Delta n_0+\left( \Delta N e^{2iK t} +c.c.\right)
\end{eqnarray}
where $\bar{n}=(n_L+n_R)/2$ is the mean carrier number, $\Delta{n}=(n_L-n_R)/2$ is the population imbalance between the cavities, $\bar{n}_0$ and $\Delta{n}_0$ are the steady state values, and $\bar{N}$ and $\Delta{N}$ are small amplitude complex oscillations. Replacing these expressions into Eqs.~\ref{eq:aLR} and \ref{eq:nLR}, and neglecting spontaneous emission for the moment, we obtain:
\begin{equation}
\Delta N=-\frac{\beta \gamma_\parallel}{4 \gamma_{tot}}\cdot \frac{(\bar{n}_0 -n_0)A_BA_{AB}^*}{1+\frac{\beta \gamma_\parallel}{2 \gamma_{tot}}(|A_B|^2+|A_{AB}|)^2+\frac{2iK}{\gamma_{tot}}}
\label{grating}
\end{equation}
Equation \ref{grating} stands for the discrete grating complex amplitude, and it is at the origin of asymmetric interaction between the modes (Fig.~\ref{fig1}a). The stimulated gain originating from Eq.~\ref{grating} for the bonding ($\delta {G}_B$) and the anti-bonding ($\delta G_{AB}$) modes can be approximated, in the limit of large Rabi splitting, by
\begin{equation}
\delta G_{B,AB}= \mp \frac{(\beta \gamma_\parallel)^2}{4 K}(\bar{n}_0 -n_0)\alpha|a_{AB,B}|^2
\label{stimgain}
\end{equation}
Equations \ref{stimgain} contains three important features: i) the mode interaction is anti-symmetric, meaning that the sign is reversed when changing the direction of the interaction (scattering from $a_B$ to $a_{AB}$ or viceversa);
ii) for $K>0$, i.e. blue-detuned bonding mode, the positive gain contribution is experienced by the red-detuned AB mode; and iii) this effect is proportional to the $\alpha$-factor, thus revealing the nonlinear index effect as the main mechanism for the stimulated scattering. Note that saturation term in Eq.~\ref{grating} is negligible provided that $2K/\gamma_{tot}\gg s/2s_{sat}$, where $s$ is total photon number, and $s_{sat}=\gamma_{tot}/\beta\gamma_{\parallel}$. In our case this approximation is justified since $2K/\gamma_{tot} \sim 100$ and $s$ is few times $s_{sat}$ in a normal laser operation regime.
\section{Experimental Results}
\subsection{Coupled Photonic-Crystal nanolasers}
In order to experimentally realize the discrete Bogatov effect, we have designed and fabricated two evanescently-coupled active photonic crystal (PhC) L3 cavities (three holes missing in the $\Gamma$K direction of a triangular lattice) in a semiconductor free standing membrane (Fig.~\ref{fig1}b). The size of surrounding holes has been modified to both increase beaming efficiency (band-folding technique) and control the inter-cavity coupling strength (barrier engineering).
Details on the PhC structure can be found in Refs.~\cite{hamel2015} and \cite{haddadi2014}.
Both single and coupled cavities have been etched in InP membranes containing InGaAs/InGaAsP quantum wells. The measured Q-factor at transparency is $Q\sim 4300$ ($\tau=\kappa^{-1}\approx 7 \, \mathrm{ps}$, where $\tau$ is the photon lifetime in the cavity) and the spontaneous emission factor ($\beta$) is $\sim$ 0.02 for coupled cavities. The linewidth enhancement factor is $\alpha = 7$. Two modes are observed in the coupled-cavity system: the B mode for a symmetric superposition of the single cavity modes, and the AB mode for an anti-symmetric superposition, which can be clearly identified in the far field images (Fig.~\ref{fig1}c).
\begin{figure}[!t]
\includegraphics[scale=0.2]{fig2.pdf}
\caption{a,b) Superposition of 40 experimental LL-curves in quasi-CW pumping showing the mode switching for K $\sim$ 8$\kappa$ (a) and K $\sim$ 12$\kappa$ (b). c,d) Superposition of 10 numerical LL-curves for K = 8$\kappa$ (c) and K = 12$\kappa$ (d). e) Black line: effective loss split parameter $\gamma_{eff}$. Markers: pump power at the switching point normalized by pump power at threshold plotted versus normalized mode splitting ($\Delta \omega$/$\kappa$, bottom axis) and Rabi frequency (K/$\kappa$, top axis). The corresponding wavelength splittings from the leftmost to the rightmost points are: 1.92, 2.66, 3.25, 4.06 and 4.8 nm. Blue line: fit of the experimental data using Eq.~\ref{Ps} and $\gamma_{eff}$ (see text). The star indicates the operation point of the Spontaneous Symmetry Breaking (SSB) transition reported in \cite{hamel2015}. The red numbers represent the modification of the holes in the central row of the PhC structure in order to tune the amount of evanescent coupling between the nanocavities. The inset shows a structure where central holes diameters were decreased by 15\%. }
\label{switch-quasiCW}
\end{figure}
The effective refractive index of the symmetric mode of the underlying W1 waveguide (one row of holes omitted in the $\gamma$K direction of the triangular lattice) is smaller than the index of the anti-symmetric mode. Hence the B mode is blue-detuned with respect to the AB mode \cite{haddadi2014}.
Our strategy to observe Bogatov mode interaction is to induce a mode switching by stabilizing the system close to the laser threshold at the blue most mode (B). For this, the PhC lattice parameter is tuned such that the optical resonances lie on the red slope of the gain curve (Fig.~\ref{fig1}c) (maximum at $\lambda_0$ $\sim$ 1510 nm and FWHM $\sim 63$ nm). As a result, the B mode has larger enough gain and, even though the AB mode has lower optical losses \cite{hamel2015}, the B mode experiences larger net gain. Therefore, the blue-detuned mode (B) is the lasing one close to laser threshold.
\subsection{Mode switching}
The coupled nanolasers are optically pumped at room temperature. We use optical periodic (50 kHz-repetition rate) and incoherent ($\lambda=800$ nm) pumping consisting either of a ramp for the acquisition of the output vs. input power (LL-curve) or short pulses ($\sim 100$ ps-pulse duration) to perform statistics on the output pulse energies. In both cases, the pump spot is located at the center of the coupled cavity system. The emission is collected with a N.A.=0.95 microscope objective, and its back focal plane is imaged through a lens to obtain the far-field
pattern (Fig.~\ref{fig1}c). Two single mode fibers coupled to microscope objectives are used to spatially select two regions of the far field: the center corresponds to B-mode intensity, and one of the lateral lobes to the AB-mode.
These optical signals are sent to two identical low noise (200 fW$/ \sqrt{Hz}$), 660 MHz-bandwidth avalanche photodiode (APD) detectors. This is similar to the experimental technique of Ref. \cite{hamel2015}, but now the
two detectors monitor intensity fluctuations from modes rather than individual cavities.
The optical Rabi frequency is related to the laser mode splitting $\Delta \omega$ as $K=\Delta \omega /2 +\alpha \gamma$ \cite{hamel2015}. In Fig.~\ref{switch-quasiCW}a,b we show the simultaneous output intensities of both modes as a function of the instantaneous pump power for two different Rabi mode splittings ($K=8\kappa$ and $K=12\kappa$).
This plot consists of a superposition of 40 sequences of quasi-CW pumping. It can be clearly observed that,
at threshold, lasing takes place on the B-mode, demonstrating that it has higher net gain as expected. In Fig.~\ref{switch-quasiCW}a the B mode loses stability and turns off at P $\approx$ 2.5P$_{th}$, while the AB mode turns on right after.
This reveals energy transfer from the blue-detuned to the red-detuned mode. In addition we observe that, from sequence to sequence, the instability threshold fluctuates in a range 2.1 $\leqslant$ P/P$_{th}$ $\leqslant$ 2.9.
Once the bifurcation point is crossed, the system remains stable on the AB-mode up to large values of the pumping. A similar behavior is observed for larger coupling strength (Fig.~\ref{switch-quasiCW}b), and in this case the switching point occurs at a larger pumping level, P $\sim$ 3.8P$_{th}$. Note that this is consistent with the dependence of the stimulated gain on the cavity coupling of Eq.~\ref{stimgain}, i.e. $\delta G_{AB}\sim K^{-1}$.
Next, we explore the full range of experimentally available coupling levels by extensively exploiting the barrier engineering technique. To this aim we have fabricated a number of different coupled cavity systems
with varying evanescent coupling.
Laser emission experiments on these cavities have been carried out subsequently, and the results are depicted in Fig.~\ref{switch-quasiCW}e.
Here, the pump power at the switching point is plotted as a function of the normalized frequency splitting of the laser modes.
\begin{figure}[!t]
\centering
\includegraphics[scale=0.17]{fig3.pdf}
\caption{a) Second order auto-correlation function for AB (red) and B (blue) as a function of normalized pump power. solid line: K = 8$\kappa$, dashed line: K = 12$\kappa$. b) Second order cross-correlation function as a function of normalized pump power for K = 8$\kappa$ (solid line) and K = 12$\kappa$ (dashed line).}
\label{g2_exp}
\end{figure}
\subsection{Statistics of mode fluctuations}
Our last experimental study deals with the statistics of the laser emission. Since a few hundreds of photons fill the cavities (in our case $s_{sat} \approx 150$), stochastic fluctuations play a fundamental role, which is
typical in nanolasers. We have investigated the consequences of this in the context of mode switching by performing a statistical study of the mode fluctuations.
With this aim we have implemented a short-pulse pumping scheme developed in \cite{marconi2016} which allows us to obtain second order auto and cross-correlation functions of the emitted photon number $m$:
\begin{eqnarray}
g^{(2)}(0) & \approx & 1 + \frac{(\Delta m)^2}{{\langle m \rangle}^2} \label{g2}\\
g^{(2)}_{B,AB}(0) & = & \frac{\langle m_{B} m_{AB} \rangle}{\langle m_{B} \rangle \langle m_{AB} \rangle} \label{g2cross}
\end{eqnarray}
where $\Delta m$ is the standard deviation.
Note that Eq.~\ref{g2} holds in the large photon number limit \cite{loudon}.
Figure \ref{g2_exp} depicts self and cross second order correlations functions obtained from time traces containing $10^4$ pulses, with $K=8\,\kappa$ (solid line) and $K=12\,\kappa$ (dashed line).
Before the switch, lasing occurs in the B-mode and thus $g_B^{(2)}(0) \sim $ 1 while the AB mode develops super-Poissonian fluctuations. We point out that in this close-to-threshold pump region --highly influenced by noise--,
the actual $g_{AB}^{(2)}(0)$ might be larger compared to the observed one, which is limited by the resolution of our experimental technique.
A decrease of $g_{B,AB}^{(2)}(0)$ is visible, which is the signature of a growing anti-correlation of the mode fluctuations.
The mode switching occurs as $g_{B}^{(2)}(0) \approx g_{AB}^{(2)}(0)$ which takes place at $P_s\approx 2.2P_{th}$ with a value of $g_{switch}^{(2)}(0)\approx 1.2$ for $K=8\,\kappa$, and $P_s\approx 3.5 P_{th}$ with a value of $g_{switch}^{(2)}(0)\approx 1.4$ for $K=12\,\kappa$. Note that at the switching points the modes become indistinguishable in terms of the second moments of their energy fluctuations.
Furthermore, for $K=12\,\kappa$, a "soft" region for 3 $\leqslant$ P/P$_{th}$ $\leqslant$ 4.3 is characterized by a plateau
in $g_{B,AB}^{(2)}(0) \approx 0.5$, which corresponds to both modes present in the optical spectrum. This region also exists when the coupling is smaller, but the plateau has a larger cross-correlation,
$g_{B,AB}^{(2)}(0) \approx 0.7$. This could be due to a stronger effect of spontaneous emission noise for small K as a result of a closer proximity of the switching point to the laser threshold.
The experimental results can be compared with the full solutions of the coupled field equations in presence of spontaneous emission, that we modeled as Langevin forces.
Figures~\ref{switch-quasiCW}c,d) show simulated time traces for two coupling parameters $K=8\,\kappa$ and $K=12\,\kappa$. Note that the energy transfer in the form of a switch mechanism is well reproduced by the model.
The jitter of the mode switching, in turn, is larger in the experiment, which could be explained as a consequence of other sources of noise (e.g. mechanical vibrations and thermal effects).
\section{Theoretical analysis}
We now theoretically investigate the nonlinear dynamical origin of the mode switching and explain the dependence of the switch-power upon the coupling parameter observed in Fig.~\ref{switch-quasiCW}e.
To this aim, we have performed an analytical reduction of the deterministic part of Eqs.~\ref{eq:aLR} and \ref{eq:nLR} ($F_a=F_N=0$) to a one-dimensional (1D) dynamical system able to capture the essence of the mode switching dynamics.
By exploiting the specific scaling of the parameters and
of the governing time scales, it is possible to reduce the time evolution of the system
to the dynamics of the fractional photon population imbalance between
the two modes, $h=\left(|a_B|^2-|a_{AB}|^2\right)/\left(|a_B|^2+|a_{AB}|^2\right)$.
In this framework, the B and AB solutions are represented by the two extreme values $h=\pm1$.
Under the assumption of large Rabi splitting and small population imbalance it is possible to show that the dynamics of $h$ is approximately ruled by the following 1D differential equation:
\begin{equation}
\dot{h} = -\frac{dU(h)}{dh} \label{eq:zztop}
\end{equation}
where $U(h)=-\gamma_h (h-h^3/3)$ is an effective 1D potential.
The factor $\gamma_h$ is a 1D nonlinear optical loss splitting rate which reads:
\begin{equation}
\gamma_{h} = 2\gamma-\frac{\alpha \gamma_{tot}}{2K} \left(P-1\right)\label{eq:gamz}
\end{equation}
\begin{figure}[!t]
\includegraphics[bb=20bp 0bp 630bp 451bp,clip,width=1\columnwidth]{fig4.pdf}
\caption{(a) Sketch representing the exchange of stability between the B (in blue) and the AB mode (in red) depicted in (b). The values of $\gamma_h$ corresponding to the three values of the normalized pump power $p_{1,2,3}=P_{1,2,3}/P_{th}$ are $\gamma_h=1,\,0.1$ and $-1$. (c) Bifurcation diagrams for the photon number in a single cavity in the B and AB modes as a function of pump power ($K=12\kappa$). A switching occurs from the B to the AB mode through a Hopf bifurcation at $P=3.765\,P_{th}$. (d) Blow-up around the bifurcation points and the narrow periodic bridge solution. Full and dotted lines correspond to stable and unstable solutions, respectively.}
\label{theo}
\end{figure}
Equations \ref{eq:zztop} and \ref{eq:gamz} contain the essence of the mode switching phenomenon induced by the Bogatov effect in the coupled cavity system. One of the fixed points $h=\pm1$ is stable (node) and the other one unstable (saddle), and depending on the sign of $\gamma_{h}$ the stable one will be bonding ($h=1$) or anti-bonding ($h=-1$).
This is illustrated in Fig.~\ref{theo}a, where the potential $U(h)$ is represented for three different pumping parameters: before ($\gamma_h>0$) and after ($\gamma_h<0$) the switching point, and for an intermediate pump level ($\gamma_h=0.1$).
The mode switching depicted in Fig.~\ref{theo}b is thus represented by the change of sign of $\gamma_{h}$. Setting $\gamma_{h}=0$ in Eq.~\ref{eq:gamz} allows us to find the locus of the mode switching pump power $P_s$ as a function of
the imaginary ($K$) and real ($\gamma$) parts of the coupling parameter:
\begin{equation}
P_s=\left( 1+\frac{4\gamma K}{\alpha \kappa \gamma_{tot}} \right)\cdot \frac{2\kappa}{\beta \gamma_{\parallel} N_0}+1 \label{Ps}
\end{equation}
When using Eq.~\ref{Ps} to fit the experimental results, one has to take into account the gain difference between B and AB modes, since they are spectrally located on the slope of the gain function
(see Fig.~\ref{fig1}c). A natural way to include this effect in the model is to replace $\gamma$ by an effective $\gamma_{eff}$ parameter which accounts for net optical losses (loss split minus gain difference).
Hence, the effective loss split parameter depends on mode spacing, $\gamma_{eff} (\Delta \omega)$. We consider the following phenomenological saturable function
$\gamma_{eff} (\Delta \omega)=\gamma_1+(\gamma_0-\gamma_1)/[1+(\Delta \omega-\Delta \omega_0)/\Delta \omega_{sat}]$, where $\gamma_0=\gamma_{eff}(\Delta \omega_0)$ is the "zero-detuning" loss split,
$\gamma_1$ is the large $\Delta \omega$ limit, and $\Delta \omega_{sat}$ is a saturation parameter. We fix $\gamma_0=-0.1$ for $\Delta \omega=8\kappa$ (or $K=3.3 \kappa$), for consistency with the
operation point of the symmetry breaking (SSB) transition, whose parameters have been determined in \cite{hamel2015}.
Two effects are captured by $\gamma_{eff}$($\Delta \omega$): a first-order linear increase for $\Delta \omega$ $\approx$ $\Delta \omega_0$ which comes from the non-zero slope of the gain curve (Fig.~\ref{fig1}c), and higher-order saturation terms. The latter can be attributed to saturation effects not taken into account in the original model, such as spectral hole burning \cite{agrawal}.
Our expression of $\gamma_{eff} (\Delta \omega)$ is inserted into Eq.~\ref{Ps} to fit the experimental data,
and the result is plotted in Fig.~\ref{switch-quasiCW}e. It is interesting to note that two regimes can be identified: the Rabi regime, where the nonlinear laser frequency shift is smaller than the Rabi splitting, i.e. $K/\kappa>\alpha$,
and the Josephson regime for weaker Rabi splitting, i.e. $K/\kappa<\alpha$. The SSB transition lies within the Josephson region (see Fig.~\ref{switch-quasiCW}e), where ultra fast ($\sim 150$ GHz) oscillations have been predicted
to emerge through secondary (Hopf) bifurcations \cite{hamel2015}.
In this regime, $\gamma_{eff}$ is rapidly increasing and Bogatov mode interaction can be observed as well (see point at $\Delta \omega=11\kappa$). Indeed, for $\Delta \omega=11\kappa$, also SSB can be observed after the mode switching instability. For increasing $\Delta \omega$ values we enter into the Rabi regime, where $\gamma_{eff}$ strongly saturates to $\gamma_{eff}\approx \gamma_1=0.055$. In this regime of
nearly constant $\gamma_{eff}$, the linear dependence of the switch power upon the coupling parameter K predicted in Eq.~\ref{Ps} is established.
Within this simple description, the switching point takes place at a single point $P_s$ since the switching mechanism is represented by an exchange of stability between the two solutions. First, the presence of spontaneous emission modifies this picture and induces a "transition region" in the pump parameter with a nonzero measure. Second, the single bifurcation point is a consequence of the large $K$ approximation used to derive Eq.~\ref{eq:zztop}. Relaxing this hypothesis, i.e. for finite K-values, the switching mechanism becomes more complex: a non-trivial connection of B and AB solutions in phase space comes up as the pump parameter is varied. This question can be elucidated by computing a bifurcation diagram for the full deterministic dynamical system of Eqs.~\ref{eq:aLR} and \ref{eq:nLR}, shown in Fig.~\ref{theo}c. It can be observed that the stable B-mode undergoes a supercritical Hopf bifurcation at P $\sim$ 3.765 P$_{th}$. In other words, there is a switching zone (see the inset of Fig.~\ref{theo}c) where the only stable solution is a limit cycle, which manifests itself as a cavity beating at the Rabi frequency. This limit cycle, in presence of noise, is the nonlinear dynamical representation of the soft transition region from the blue-detuned to the red-detuned mode where the coupled nanolasers operate in a dual-frequency regime. Such a fast nonlinear beating dynamically delocalizes photons in the coupled cavity system. As a result strong anti-correlated intensity fluctuations for B and AB modes are expected, which is consistent with the observed minimum of the cross-correlation function at the switching point (Fig.~\ref{g2_exp}). Hence, we conclude that the observation of a dip in the mode cross-correlation is the statistical consequence of a dual frequency operation region given by an ultrafast ($\sim0.5$ THz) noisy limit cycle rather than, for instance, bi-stable mode switching scenarios previously reported for bi-modal microlasers \cite{Leymann2013,redlich2016mode}.
\section{Conclusion}
In conclusion, we have revealed for the first time asymmetric energy transfer between eigenmodes in a semiconductor photonic dimer. We have experimentally shown that a mode switching occurs from the blue-detuned (bonding) to the red-detuned (anti-bonding) modes as the pump power is increased. We have identified the basic mechanism underlying such mode interaction as a "discrete Bogatov effect", i.e. the asymmetric stimulated light scattering induced by two-site population oscillations at the optical Rabi frequency. The predicted scaling of the onset of switching with respect to the coupling parameter in a mean-field model is shown to be in good agreement with the experimental results for different coupling strengths. We claim that this phenomenon is generic in semiconductor coupled cavities, and dominates against cross-gain saturation effects, therefore it
could be used to model nonlinear mode coupling in probabilistic theories such as bimodal birth-death models in micro/nano-laser photonic dimers. Furthermore, it should also be scalable to N-coupled cavities, such that energy can be expected to flow from the blue-most to the red-most hybrid modes of the photonic molecule with the increase of pump power.
\section*{Acknowledgments}
The authors acknowledge I. Sagnes and G. Beaudoin for the fabrication of the samples. This work is supported by a public grant overseen by the French National Research Agency (ANR) as part of the "Investissements d'Avenir" program (Labex NanoSaclay, reference: ANR-10-LABX-0035) and funding from the ANR project OPTIROC. J.J. acknowledges financial support from the Ram\'on y Cajal fellowship and project COMBINA (TEC2015-65212-C3-3-P).
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.